Game Programming in C++ - Creating 3D Games

Contents
1. Cover Page
2. About This E-Book
3. Title Page
4. Copyright Page
5. Dedication Page
6. Contents at a Glance
7. Contents
8. Preface
9. Acknowledgments
10. About the Author
11. Chapter 1: Game Programming Overview
1. Setting Up a Development Environment
1. Microsoft Windows
2. Apple macOS
2. Getting This Book’s Source Code

3. Beyond the C++ Standard Library
4. The Game Loop and Game Class
1. Anatomy of a Frame
2. Implementing a Skeleton Game Class
3. Main Function
4. Basic Input Processing
5. Basic 2D Graphics
1. The Color Buffer

2. Double Buffering
3. Implementing Basic 2D Graphics
4. Drawing Walls, a Ball, and a Paddle
6. Updating the Game
1. Real Time and Game Time

2. Logic as a Function of Delta Time
3. Updating the Paddle’s Position
4. Updating the Ball’s Position
7. Game Project
8. Summary
9. Additional Reading
10. Exercises
1. Exercise 1.1
2. Exercise 1.2
12. Chapter 2: Game Objects and 2D Graphics
1. Game Objects
1. Types of Game Objects

2. Game Object Models
3. Integrating Game Objects into the Game Loop
2. Sprites
1. Loading Image Files

2. Drawing Sprites
3. Animating Sprites
3. Scrolling Backgrounds
4. Game Project
5. Summary
7. Exercises
1. Exercise 2.1
2. Exercise 2.2
3. Exercise 2.3
13. Chapter 3: Vectors and Basic Physics
1. Vectors
1. Getting a Vector between Two Points: Subtraction

2. Scaling a Vector: Scalar Multiplication
3. Combining Two Vectors: Addition
4. Determining a Distance: Length
5. Determining Directions: Unit Vectors and
Normalization
6. Converting from an Angle to a Forward Vector
7. Converting a Forward Vector to an Angle:
Arctangent
8. Determining the Angle between Two Vectors: Dot
Product
9. Calculating a Normal: Cross Product
2. Basic Movement
1. Creating a Basic MoveComponent Class

2. Creating an InputComponent Class
3. Newtonian Physics
1. Linear Mechanics Overview

2. Computing Positions with Euler Integration
3. Issues with Variable Time Steps
4. Basic Collision Detection
1. Circle-Versus-Circle Intersection
2. Creating a CircleComponent Subclass
5. Game Project
6. Summary
8. Exercises
1. Exercise 3.1
2. Exercise 3.2
3. Exercise 3.3
14. Chapter 4: Artificial Intelligence
1. State Machine Behaviors
1. Designing a State Machine

2. Basic State Machine Implementation
3. States as Classes
2. Pathfinding
1. Graphs
2. Breadth-First Search
3. Heuristics
4. Greedy Best-First Search
5. A* Search
6. Dijkstra’s Algorithm
7. Following a Path
8. Other Graph Representations
3. Game Trees
1. Minimax
2. Handling Incomplete Game Trees
3. Alpha-Beta Pruning
4. Game Project
5. Summary
7. Exercises
1. Exercise 4.1
2. Exercise 4.2
15. Chapter 5: OpenGL

1. Initializing OpenGL
1. Setting Up the OpenGL Window

2. The OpenGL Context and Initializing GLEW
3. Rendering a Frame
2. Triangle Basics
1. Why Polygons?
2. Normalized Device Coordinates
3. Vertex and Index Buffers
3. Shaders
1. Vertex Shaders
2. Fragment Shaders
3. Writing Basic Shaders
4. Loading Shaders
5. Drawing Triangles
4. Transformation Basics
1. Object Space
2. World Space
3. Transforming to World Space
5. Matrices and Transformations
1. Matrix Multiplication
2. Transforming a Point by Using a Matrix
3. Transforming to World Space, Revisited
4. Adding World Transforms to Actor
5. Transforming from World Space to Clip Space
6. Updating Shaders to Use Transform Matrices
6. Texture Mapping
1. Loading the Texture

2. Updating the Vertex Format
3. Updating the Shaders
4. Alpha Blending
7. Game Project
8. Summary
10. Exercises
1. Exercise 5.1
2. Exercise 5.2
16. Chapter 6: 3D Graphics
1. The Actor Transform in 3D
1. Transform Matrices for 3D

2. Euler Angles
3. Quaternions
4. New Actor Transform in Action
2. Loading 3D Models
1. Choosing a Model Format

2. Updating the Vertex Attributes
3. Loading a gpmesh File
3. Drawing 3D Meshes
1. Transforming to Clip Space, Revisited

2. Out with the Painter’s Algorithm, in with Z-
Buffering
3. The BasicMesh Shader
4. The MeshComponent Class
4. Lighting
1. Revisiting Vertex Attributes

2. Types of Lights
3. Phong Reflection Model
4. Implementing Lighting
5. Game Project
6. Summary
8. Exercises
1. Exercise 6.1
2. Exercise 6.2
17. Chapter 7: Audio
1. Bootstrapping Audio
1. FMOD
2. Installing FMOD
3. Creating an Audio System
4. Banks and Events
5. The SoundEvent Class
2. 3D Positional Audio
1. Setting Up a Basic Listener

2. Adding Positional Functionality to SoundEvent
3. Creating an AudioComponent to Associate Actors
with Sound Events
4. The Listener in a Third-Person Game
5. The Doppler Effect
3. Mixing and Effects
1. Buses
2. Snapshots
3. Occlusion
4. Game Project
5. Summary
7. Exercises
1. Exercise 7.1
2. Exercise 7.2
18. Chapter 8: Input Systems
1. Input Devices
1. Polling
2. Positive and Negative Edges
3. Events
4. Basic InputSystem Architecture
2. Keyboard Input
3. Mouse Input
1. Buttons and Position

2. Relative Motion
3. Scroll Wheel
4. Controller Input
1. Enabling a Single Controller

2. Buttons
3. Analog Sticks and Triggers
4. Filtering Analog Sticks in Two Dimensions
5. Supporting Multiple Controllers
5. Input Mappings
6. Game Project
7. Summary
9. Exercises
1. Exercise 8.1
2. Exercise 8.2
19. Chapter 9: Cameras
1. First-Person Camera
1. Basic First-Person Movement

2. Camera (Without Pitch)
3. Adding Pitch
4. First-Person Model
2. Follow Camera
1. Basic Follow Camera

2. Adding a Spring
3. Orbit Camera
4. Spline Camera
5. Unprojection
6. Game Project
7. Summary
9. Exercises
1. Exercise 9.1
2. Exercise 9.2
20. Chapter 10: Collision Detection
1. Geometric Types
1. Line Segments
2. Planes
3. Bounding Volumes
2. Intersection Tests
1. Contains Point Tests

2. Bounding Volume Tests
3. Line Segment Tests
4. Dynamic Objects
3. Adding Collisions to Game Code
1. The BoxComponent Class

2. The PhysWorld Class
3. Ball Collisions with SegmentCast
4. Testing Box Collisions in PhysWorld
5. Player Collision Against the Walls
4. Game Project
5. Summary
7. Exercises
1. Exercise 10.1
2. Exercise 10.2
3. Exercise 10.3
21. Chapter 11: User Interfaces
1. Font Rendering
2. UI Screens
1. The UI Screen Stack

2. The Pause Menu
3. Buttons
4. Dialog Boxes
3. HUD Elements
1. Adding an Aiming Reticule

2. Adding Radar
4. Localization
1. Working with Unicode

2. Adding a Text Map
3. Other Localization Concerns
5. Supporting Multiple Resolutions
6. Game Project
7. Summary
9. Exercises
1. Exercise 11.1
2. Exercise 11.2
3. Exercise 11.3
22. Chapter 12: Skeletal Animation
1. Foundations of Skeletal Animation
1. Skeletons and Poses

2. The Inverse Bind Pose Matrix
3. Animation Data
4. Skinning
2. Implementing Skeletal Animation
1. Drawing with Skinning Vertex Attributes

2. Loading a Skeleton
3. Loading the Animation Data
4. The Skinning Vertex Shader
5. Updating Animations
3. Game Project
4. Summary
6. Exercises
1. Exercise 12.1
2. Exercise 12.2
23. Chapter 13: Intermediate Graphics
1. Improving Texture Quality

1. Texture Sampling, Revisited
2. Mipmapping
3. Anisotropic Filtering
2. Rendering to Textures
1. Creating the Texture

2. Creating a Framebuffer Object
3. Rendering to a Framebuffer Object
4. Drawing the Mirror Texture in the HUD
3. Deferred Shading
1. Creating a G-Buffer Class

2. Writing to the G-buffer
3. Global Lighting
4. Adding Point Lights
5. Improvements and Issues
4. Game Project
5. Summary
7. Exercises
1. Exercise 13.1
2. Exercise 13.2
24. Chapter 14: Level Files and Binary Data
1. Level File Loading
1. Loading Global Properties

2. Loading Actors
3. Loading Components
2. Saving Level Files
1. Saving Global Properties

2. Saving Actors and Components
3. Binary Data
1. Saving a Binary Mesh File

2. Loading a Binary Mesh File
4. Game Project
5. Summary
7. Exercises
1. Exercise 14.1
2. Exercise 14.2
25. Appendix A: Intermediate C++ Review

26. Index
1. ii
2. iii
3. iv
4. v
5. vi
6. vii
7. viii
8. ix
9. x
10. xi
11. xii
12. xiii
13. xiv
14. xv
15. xvi
16. xvii
17. xviii
18. xix
19. xx
20. xxii
21. xxiii
22. xxiv
23. 1
24. 2
25. 3
26. 4
27. 5
28. 6
29. 7
30. 8
31. 9
32. 10
33. 11
34. 12
35. 13
36. 14
37. 15
38. 16
39. 17
40. 18
41. 19
42. 20
43. 21
44. 22
45. 23
46. 24
47. 25
48. 26
49. 27
50. 28
51. 29
52. 30
53. 31
54. 32
55. 33
56. 34
57. 35
58. 36
59. 37
60. 38
61. 39
62. 40
63. 41
64. 42
65. 43
66. 44
67. 45
68. 46
69. 47
70. 48
71. 49
72. 50
73. 51
74. 52
75. 53
76. 54
77. 55
78. 56
79. 57
80. 58
81. 59
82. 60
83. 61
84. 62
85. 63
86. 64
87. 65
88. 66
89. 67
90. 68
91. 69
92. 70
93. 71
94. 72
95. 73
96. 74
97. 75
98. 76
99. 77
100. 78
101. 79
102. 80
103. 81
104. 82
105. 83
106. 84
107. 85
108. 86
109. 87
110. 88
111. 89
112. 90
113. 91
114. 92
115. 93
116. 94
117. 95
118. 96
119. 97
120. 98
121. 99
122. 100
123. 101
124. 102
125. 103
126. 104
127. 105
128. 106
129. 107
130. 108
131. 109
132. 110
133. 111
134. 112
135. 113
136. 114
137. 115
138. 116
139. 117
140. 118
141. 119
142. 120
143. 121
144. 122
145. 123
146. 124
147. 125
148. 126
149. 127
150. 128
151. 129
152. 130
153. 131
154. 132
155. 133
156. 134
157. 135
158. 136
159. 137
160. 138
161. 139
162. 140
163. 141
164. 142
165. 143
166. 144
167. 145
168. 146
169. 147
170. 148
171. 149
172. 150
173. 151
174. 152
175. 153
176. 154
177. 155
178. 156
179. 157
180. 158
181. 159
182. 160
183. 161
184. 162
185. 163
186. 164
187. 165
188. 166
189. 167
190. 168
191. 169
192. 170
193. 171
194. 172
195. 173
196. 174
197. 175
198. 176
199. 177
200. 178
201. 179
202. 180
203. 181
204. 182
205. 183
206. 184
207. 185
208. 186
209. 187
210. 188
211. 189
212. 190
213. 191
214. 192
215. 193
216. 194
217. 195
218. 196
219. 197
220. 198
221. 199
222. 200
223. 201
224. 202
225. 203
226. 204
227. 205
228. 206
229. 207
230. 208
231. 209
232. 210
233. 211
234. 212
235. 213
236. 214
237. 215
238. 216
239. 217
240. 218
241. 219
242. 220
243. 221
244. 222
245. 223
246. 224
247. 225
248. 226
249. 227
250. 228
251. 229
252. 230
253. 231
254. 232
255. 233
256. 234
257. 235
258. 236
259. 237
260. 238
261. 239
262. 240
263. 241
264. 242
265. 243
266. 244
267. 245
268. 246
269. 247
270. 248
271. 249
272. 250
273. 251
274. 252
275. 253
276. 254
277. 255
278. 256
279. 257
280. 258
281. 259
282. 260
283. 261
284. 262
285. 263
286. 264
287. 265
288. 266
289. 267
290. 268
291. 269
292. 270
293. 271
294. 272
295. 273
296. 274
297. 275
298. 276
299. 277
300. 278
301. 279
302. 280
303. 281
304. 282
305. 283
306. 284
307. 285
308. 286
309. 287
310. 288
311. 289
312. 290
313. 291
314. 292
315. 293
316. 294
317. 295
318. 296
319. 297
320. 298
321. 299
322. 300
323. 301
324. 302
325. 303
326. 304
327. 305
328. 306
329. 307
330. 308
331. 309
332. 310
333. 311
334. 312
335. 313
336. 314
337. 315
338. 316
339. 317
340. 318
341. 319
342. 320
343. 321
344. 322
345. 323
346. 324
347. 325
348. 326
349. 327
350. 328
351. 329
352. 330
353. 331
354. 332
355. 333
356. 334
357. 335
358. 336
359. 337
360. 338
361. 339
362. 340
363. 341
364. 342
365. 343
366. 344
367. 345
368. 346
369. 347
370. 348
371. 349
372. 350
373. 351
374. 352
375. 353
376. 354
377. 355
378. 356
379. 357
380. 358
381. 359
382. 360
383. 361
384. 362
385. 363
386. 364
387. 365
388. 366
389. 367
390. 368
391. 369
392. 370
393. 371
394. 372
395. 373
396. 374
397. 375
398. 376
399. 377
400. 378
401. 379
402. 380
403. 381
404. 382
405. 383
406. 384
407. 385
408. 386
409. 387
410. 388
411. 389
412. 390
413. 391
414. 392
415. 393
416. 394
417. 395
418. 396
419. 397
420. 398
421. 399
422. 400
423. 401
424. 402
425. 403
426. 404
427. 405
428. 406
429. 407
430. 408
431. 409
432. 410
433. 411
434. 412
435. 413
436. 414
437. 415
438. 416
439. 417
440. 418
441. 419
442. 420
443. 421
444. 422
445. 423
446. 424
447. 425
448. 426
449. 427
450. 428
451. 429
452. 430
453. 431
454. 432
455. 433
456. 434
457. 435
458. 436
459. 437
460. 438
461. 439
462. 440
463. 441
464. 442
465. 443
466. 444
467. 445
468. 446
469. 447
470. 448
471. 449
472. 450
473. 451
474. 452
475. 453
476. 454
477. 455
478. 456
479. 457
480. 458
481. 459
482. 460
483. 461
484. 462
485. 463
486. 464
487. 465
488. 466
489. 467
490. 468
491. 469
492. 470
493. 471
494. 472
495. 473
496. 474
497. 475
498. 476
499. 477
500. 478
501. 479
502. 480
503. 481
504. 482
505. 483
506. 484
507. 485
508. 486
509. 487
510. 488
511. 489
512. 490
513. 491
514. 492
515. 493
516. 494
517. 495
518. 496
519. 497
520. 498
521. 499
522. 500
523. 501
524. 502
525. 503
526. 504
About This E-Book
EPUB is an open, industry-standard format for e-

books. However, support for EPUB and its many features
varies across reading devices and applications. Use your
device or app settings to customize the presentation to
your liking. Settings that you can customize often include
font, font size, single or double column, landscape or
portrait mode, and figures that you can click or tap to
enlarge. For additional information about the settings
and features on your reading device or app, visit the
device manufacturer’s Web site.
Many titles include programming code or
configuration examples. To optimize the presentation of
these elements, view the e-book in single-column,
landscape mode and adjust the font size to the smallest
setting. In addition to presenting code and
configurations in the reflowable text format, we have
included images of the code that mimic the presentation
found in the print book; therefore, where the reflowable
format may compromise the presentation of the code
listing, you will see a “Click here to view code image”
link. Click the link to view the print-fidelity code image.
To return to the previous page viewed, click the Back
button on your device or app.
Game Programming in
C++
Creating 3D Games
Sanjay Madhav
Boston • Columbus • Indianapolis • New York • San

Francisco
Amsterdam • Cape Town • Dubai • London • Madrid •
Milan
Munich • Paris • Montreal • Toronto • Delhi • Mexico
City • São
Paulo • Sydney • Hong Kong • Seoul • Singapore • Taipei
• Tokyo
Many of the designations used by manufacturers and
sellers to distinguish their products are claimed as
trademarks. Where those designations appear in this
book, and the publisher was aware of a trademark claim,
the designations have been printed with initial capital
letters or in all capitals.
The author and publisher have taken care in the
preparation of this book, but make no expressed or
implied warranty of any kind and assume no
responsibility for errors or omissions. No liability is
assumed for incidental or consequential damages in
connection with or arising out of the use of the
information or programs contained herein.
For information about buying this title in bulk
quantities, or for special sales opportunities (which may
include electronic versions; custom cover designs; and
content particular to your business, training goals,
marketing focus, or branding interests), please contact
our corporate sales department at
corpsales@pearsoned.com or (800) 382-3419.
For government sales inquiries, please contact
governmentsales@pearsoned.com.
For questions about sales outside the U.S., please contact
intlcs@pearson.com.
Visit us on the Web: informit.com/aw
Library of Congress Control Number: 2017964125
Copyright © 2018 Pearson Education, Inc.
All rights reserved. Printed in the United States of
America. This publication is protected by copyright, and
permission must be obtained from the publisher prior to
any prohibited reproduction, storage in a retrieval
system, or transmission in any form or by any means,
electronic, mechanical, photocopying, recording, or
likewise. For information regarding permissions, request
forms, and the appropriate contacts within the Pearson
Education Global Rights & Permissions Department,
please visit www.pearsoned.com/permissions/.
Overwatch™ is a trademark of Blizzard Entertainment,
Inc., in the U.S. and/or other countries.
Call of Duty® is a registered trademark of Activision
Publishing, Inc.
Uncharted ™ and Horizon: Zero Dawn™ are trademarks
of and God of War® is a registered trademark of Sony
Interactive Entertainment LLC.
Visual Studio®, Xbox®, and Windows® are registered
trademarks of Microsoft Corporation.
Apple®, iOS®, Mac®, macOS®, and Xcode® are
registered trademarks of Apple Inc.
GitHub® is a registered trademark of GitHub, Inc.
OpenGL® and the oval logo are trademarks or registered
trademarks of Silicon Graphics, Inc. in the United States
and/or other countries worldwide.
Linux® is a registered trademark of Linus Torvalds.
FMOD® is a registered trademark of Firelight
Technologies Pty, Ltd.
Pac-Man® is a registered trademark of Bandai Namco
Entertainment Inc.
Asteroids® and Pong® are registered trademarks of Atari
Interactive Inc.
PlayStation® is a registered trademark of Sony
Interactive Entertainment Inc.
Android® is a registered trademark of Google Inc.
Unreal® is a trademark or registered trademark of Epic
Games, Inc. in the United States of America and
elsewhere.
Unity® is a registered trademark of Unity Technologies.
Maya® and FBX® are registered trademarks of
Autodesk, Inc., in the USA and other countries.
Skyrim® is a registered trademark of ZeniMax Media
Inc.
Planet Coaster® is a registered trademark of Frontier
Developments Plc.
ISBN-13: 978-0-13-459720-1
ISBN-10: 0-13-459720-6
1 18
Editor-in-Chief
Mark Taub
Executive Editor
Laura Lewin
Development Editor
Michael Thurston
Managing Editor
Sandra Schroeder
Tech Editors
Josh Glazer
Brian Overland
Matt Whiting
Senior Project Editor
Lori Lyons
Production Manager
Dhayanidhi Karunanidhi
Copy Editor
Kitty Wilson
Indexer
Lisa Stumpf
Proofreader
Larry Sulky
Editorial Assistant
Courtney Martin
Cover Designer
Chuti Prasertsith
Compositor
codemantra
To my family and friends: Thanks for the support.
Contents at a Glance
Preface
Acknowledgments
About the Author
1 Game Programming Overview
2 Game Objects and 2D Graphics
3 Vectors and Basic Physics
4 Artificial Intelligence
5 OpenGL
6 3D Graphics
7 Audio
8 Input Systems
9 Cameras
10 Collision Detection
11 User Interfaces
12 Skeletal Animation
13 Intermediate Graphics
14 Level Files and Binary Data
A Intermediate C++ Review

Index
Register your copy of Game Programming in C++

on the InformIT site for convenient access to
updates and corrections as they become available.
To start the registration process, go to
informit.com/register and log in or create an
account. Enter the product ISBN 9780134597201
and click Submit. Look on the Registered Products
tab for an Access Bonus Content link next to this
product, and follow that link to access any available
bonus materials. If you would like to be notified of
exclusive offers on new editions and updates, please
check the box to receive email from us.
Contents
Preface
Acknowledgments
About the Author
1 Game Programming Overview

Setting Up a Development Environment
Microsoft Windows
Apple macOS
Getting This Book’s Source Code

Beyond the C++ Standard Library
The Game Loop and Game Class
Anatomy of a Frame
Implementing a Skeleton Game Class

Main Function
Basic Input Processing
Basic 2D Graphics
The Color Buffer
Double Buffering
Implementing Basic 2D Graphics

Drawing Walls, a Ball, and a Paddle
Updating the Game

Real Time and Game Time
Logic as a Function of Delta Time

Updating the Paddle’s Position
Updating the Ball’s Position
Game Project
Summary
Additional Reading
Exercises
Exercise 1.1
Exercise 1.2
2 Game Objects and 2D Graphics

Game Objects
Types of Game Objects
Game Object Models
Integrating Game Objects into the Game Loop
Sprites
Loading Image Files
Drawing Sprites
Animating Sprites
Scrolling Backgrounds
Game Project
Summary
Additional Reading
Exercises
Exercise 2.1
Exercise 2.2
Exercise 2.3
3 Vectors and Basic Physics
Vectors
Getting a Vector between Two Points: Subtraction
Scaling a Vector: Scalar Multiplication
Combining Two Vectors: Addition
Determining a Distance: Length
Determining Directions: Unit Vectors and Normalization
Converting from an Angle to a Forward Vector
Converting a Forward Vector to an Angle: Arctangent
Determining the Angle between Two Vectors: Dot Product
Calculating a Normal: Cross Product
Basic Movement
Creating a Basic MoveComponent Class
Creating an InputComponent Class
Newtonian Physics
Linear Mechanics Overview
Computing Positions with Euler Integration

Issues with Variable Time Steps
Basic Collision Detection

Circle-Versus-Circle Intersection
Creating a CircleComponent Subclass
Game Project
Summary
Additional Reading
Exercises
Exercise 3.1
Exercise 3.2
Exercise 3.3
4 Artificial Intelligence
State Machine Behaviors
Designing a State Machine
Basic State Machine Implementation
States as Classes
Pathfinding
Graphs
Breadth-First Search
Heuristics
Greedy Best-First Search
A* Search
Dijkstra’s Algorithm
Following a Path
Other Graph Representations
Game Trees
Minimax
Handling Incomplete Game Trees
Alpha-Beta Pruning
Game Project
Summary
Additional Reading
Exercises
Exercise 4.1
Exercise 4.2
5 OpenGL
Initializing OpenGL
Setting Up the OpenGL Window
The OpenGL Context and Initializing GLEW
Rendering a Frame
Triangle Basics
Why Polygons?
Normalized Device Coordinates
Vertex and Index Buffers
Shaders
Vertex Shaders
Fragment Shaders
Writing Basic Shaders
Loading Shaders
Drawing Triangles
Transformation Basics
Object Space
World Space
Transforming to World Space
Matrices and Transformations

Matrix Multiplication
Transforming a Point by Using a Matrix
Transforming to World Space, Revisited
Adding World Transforms to Actor
Transforming from World Space to Clip Space

Updating Shaders to Use Transform Matrices
Texture Mapping
Loading the Texture
Updating the Vertex Format
Updating the Shaders
Alpha Blending
Game Project
Summary
Additional Reading
Exercises
Exercise 5.1
Exercise 5.2
6 3D Graphics
The Actor Transform in 3D
Transform Matrices for 3D
Euler Angles
Quaternions
New Actor Transform in Action
Loading 3D Models
Choosing a Model Format
Updating the Vertex Attributes
Loading a gpmesh File
Drawing 3D Meshes
Transforming to Clip Space, Revisited
Out with the Painter’s Algorithm, in with Z-Buffering

The BasicMesh Shader
The MeshComponent Class
Lighting
Revisiting Vertex Attributes
Types of Lights
Phong Reflection Model
Implementing Lighting
Game Project
Summary
Additional Reading
Exercises
Exercise 6.1
Exercise 6.2
7 Audio
Bootstrapping Audio
FMOD
Installing FMOD
Creating an Audio System

Banks and Events
The SoundEvent Class
3D Positional Audio
Setting Up a Basic Listener
Adding Positional Functionality to SoundEvent
Creating an AudioComponent to Associate Actors with Sound Events
The Listener in a Third-Person Game
The Doppler Effect
Mixing and Effects

Buses
Snapshots
Occlusion
Game Project
Summary
Additional Reading
Exercises
Exercise 7.1
Exercise 7.2
8 Input Systems
Input Devices
Polling
Positive and Negative Edges
Events
Basic InputSystem Architecture
Keyboard Input
Mouse Input
Buttons and Position
Relative Motion
Scroll Wheel
Controller Input
Enabling a Single Controller
Buttons
Analog Sticks and Triggers
Filtering Analog Sticks in Two Dimensions
Supporting Multiple Controllers
Input Mappings
Game Project
Summary
Additional Reading
Exercises
Exercise 8.1
Exercise 8.2
9 Cameras
First-Person Camera
Basic First-Person Movement
Camera (Without Pitch)
Adding Pitch
First-Person Model
Follow Camera
Basic Follow Camera
Adding a Spring
Orbit Camera
Spline Camera
Unprojection
Game Project
Summary
Additional Reading
Exercises
Exercise 9.1
Exercise 9.2
10 Collision Detection
Geometric Types
Line Segments
Planes
Bounding Volumes
Intersection Tests
Contains Point Tests
Bounding Volume Tests
Line Segment Tests

Dynamic Objects
Adding Collisions to Game Code

The BoxComponent Class
The PhysWorld Class
Ball Collisions with SegmentCast
Testing Box Collisions in PhysWorld
Player Collision Against the Walls
Game Project
Summary
Additional Reading
Exercises
Exercise 10.1
Exercise 10.2
Exercise 10.3
11 User Interfaces
Font Rendering
UI Screens
The UI Screen Stack
The Pause Menu

Buttons
Dialog Boxes
HUD Elements
Adding an Aiming Reticule
Adding Radar
Localization
Working with Unicode
Adding a Text Map

Other Localization Concerns
Supporting Multiple Resolutions

Game Project
Summary
Additional Reading
Exercises
Exercise 11.1
Exercise 11.2
Exercise 11.3
12 Skeletal Animation
Foundations of Skeletal Animation
Skeletons and Poses
The Inverse Bind Pose Matrix

Animation Data
Skinning
Implementing Skeletal Animation

Drawing with Skinning Vertex Attributes
Loading a Skeleton
Loading the Animation Data

The Skinning Vertex Shader
Updating Animations
Game Project
Summary
Additional Reading
Exercises
Exercise 12.1
Exercise 12.2
13 Intermediate Graphics
Improving Texture Quality
Texture Sampling, Revisited
Mipmapping
Anisotropic Filtering
Rendering to Textures
Creating the Texture
Creating a Framebuffer Object
Rendering to a Framebuffer Object
Drawing the Mirror Texture in the HUD
Deferred Shading
Creating a G-Buffer Class
Writing to the G-buffer
Global Lighting
Adding Point Lights
Improvements and Issues
Game Project
Summary
Additional Reading
Exercises
Exercise 13.1
Exercise 13.2
14 Level Files and Binary Data

Level File Loading
Loading Global Properties
Loading Actors
Loading Components
Saving Level Files

Saving Global Properties
Saving Actors and Components
Binary Data
Saving a Binary Mesh File
Loading a Binary Mesh File
Game Project
Summary
Additional Reading
Exercises
Exercise 14.1
Exercise 14.2
A Intermediate C++ Review
Index
PREFACE
Today, video games are some of the most popular forms

of entertainment. Newzoo’s “Global Games Market
Report” estimates over $100 billion in revenue for games
in 2017. This staggering amount shows how popular this
field truly is. Because of the size of this market, game
programmers are in low supply and high demand.
Alongside this explosion of games, game technology has

become increasingly democratized. A single developer
can make award-winning and hit games by using one of
many popular game engines and tools. For game
designers, these tools are fantastic. So what value is there
in learning how to program games in C++?
If you take a step back, you can see that many game
engines and tools are, at their core, written in C++. This
means that C++ is ultimately the technology behind
every game created using one of these tools.
Furthermore, top-notch developers who release some of

the most popular games today—including Overwatch,
Call of Duty, and Uncharted—still predominantly use
C++ because it provides a great combination of
performance and usability. Thus, any developer who
wants to eventually work for one of these companies
needs a strong understanding of programming games—
specifically in C++.
This book dives into many of the technologies and

systems that real game developers use. The basis for
much of the material in this book is video game
programming courses taught at the University of
Southern California over the course of almost a decade.
The approach used in this book has successfully
prepared many students to make it in the video games
industry.
This book is also heavily focused on real working

implementations of code integrated into actual game
project demos. It is critical to understand how all the
various systems that go into a game work together. For
this reason, you should keep the source code handy while
working through this book.
At this writing, all the code provided with this book

works on both PC and macOS, using the Microsoft Visual
Studio 2017 and Apple Xcode 9 development
environments, respectively.
The source code for this book is available on GitHub, at

https://github.com/gameprogcpp/code. For
instructions on setting up the development environment
for this book, see Chapter 1, “Game Programming
Overview.”
WHO SHOULD READ THIS

BOOK?
This book is for you if you’re a programmer who is
comfortable with C++ and wants to learn how to
program 3D video games. For readers rusty on C++,
Appendix A, “Intermediate C++ Review,” reviews
several C++ concepts. However, if you have with
little or no prior C++ experience, you should learn
C++ before jumping into this book. (One option is
Programming Abstractions in C++ by Eric Roberts.)
This book also expects you to be familiar with some
common data structures, including dynamic arrays
(vectors), trees, and graphs, and to have some
recollection of high school-level algebra.
The topics covered in this book are applicable to readers

in academic environments, hobbyists, and junior- and
mid-level game programmers who want to expand their
knowledge of game programming. The content in this
book corresponds to a little more than a semester and a
half of material in a university setting.
HOW THIS BOOK IS ORGANIZED

This book is intended to be read linearly from
Chapter 1 through Chapter 14. However, in case you
are not interested in some specific topics, Figure P.1
shows the dependencies between the chapters.
In the first handful of chapters, the games are in 2D as

you learn core concepts. From Chapter 6 onward (with
the exception of Chapter 8), the games are in 3D.
The chapters cover the following information:

Chapter 1, “Game Programming Overview,” looks at the
fundamental concepts of game programming and how to get an
initial game up and running. It also introduces the Simple
DirectMedia Layer (SDL) library.
Chapter 2, “Game Objects and 2D Graphics,” discusses how

programmers organize the objects in their games and explores
additional 2D graphics concepts, such as flipbook animation.
Chapter 3, “Vectors and Basic Physics,” covers mathematical

vectors, which are critical tools for any game programmer. It also
explores the basics of physics, for use with both motion and
collisions.
Chapter 4, “Artificial Intelligence,” looks at approaches to make

game characters that are computer controlled, including concepts
such as state machines and pathfinding.
Chapter 5, “OpenGL,” explores how to create an OpenGL renderer,

including implementing vertex and pixel shaders. It includes a
discussion of matrices.
Figure P.1 Chapter dependencies
Chapter 6, “3D Graphics,” focuses on converting the code created so

far to work for a 3D game, including how to represent the view,
projection, and rotations.
Chapter 7, “Audio,” covers how to bootstrap an audio system using

the excellent FMOD API. It includes coverage of 3D positional
audio.
Chapter 8, “Input Systems,” discusses how to design a more robust

input system for processing keyboard, mouse, and game controller
events.
Chapter 9, “Cameras,” shows how to implement several different 3D
cameras, including a first-person camera, a follow camera, and an
orbit camera.
Chapter 10, “Collision Detection,” dives into methods of collision

detection for games, including spheres, planes, line segments, and
boxes.
Chapter 11, “User Interfaces,” looks at implementing both a menu

system and heads-up display (HUD) elements such as a radar and
aiming reticule.
Chapter 12, “Skeletal Animation,” covers how to animate characters

in 3D.
Chapter 13, “Intermediate Graphics,” explores a handful of

intermediate graphics topics, including how to implement deferred
shading.
Chapter 14, “Level Files and Binary Data,” discusses how to load
and save level files, as well as how to write binary file formats.
Appendix A, “Intermediate C++ Review,” reviews several

intermediate C++ topics used throughout the book including
memory allocation and collections.
Each chapter includes a corresponding game project

(with source code available, as mentioned),
recommended additional readings, and a couple of
exercises. These exercises generally instruct you to add
additional features to the code implemented in the
chapter.
CONVENTIONS USED IN THIS

BOOK
New terms appear in bold. Code appears in a
monospaced font. Small snippets of code
sometimes appear as standalone paragraphs:
DoSomething();
Longer code segments appear in code listings, as in

Listing P.1.
Listing P.1 Sample Code Listing
void DoSomething()
// Do the thing
ThisDoesSomething();
From time to time, some paragraphs appear as notes,

tips, sidebars, and warnings. Here is an example of each.
note
Notes contain some useful information about implementation changes or
other features that are worth noting.
tip
Tips provide hints on how to add certain additional features to your code.
warning
Warnings call out specific pitfalls that warrant caution.
SIDEBAR
Sidebars are lengthier discussions that are tangential to the main

content of the chapter. This content is interesting but isn’t crucial to
understanding the core topics of the chapter.
ACKNOWLEDGMENTS
Although this is not my first book, writing this one has

been an especially long process. I am thankful that Laura
Lewin, the executive editor on this book, was especially
patient throughout the two years this book was in
progress. I would also like to thank the rest of the team
at Pearson, including Michael Thurston, the
development editor on this book.
I would also like to acknowledge the work put in by the

technical editors on this book: Josh Glazer, Brian
Overland, and Matt Whiting. The technical reviews were
critical in making sure both that the content was correct
and that it was accessible for the target audience.
I’d also like to thank all my colleagues at the USC

Information Technology Program and especially those
who helped shape the curriculum of the games courses I
teach: Josh Glazer, Jason Gregory, Clark Kromenaker,
Mike Sheehan, and Matt Whiting. Much of the
inspiration for this book comes from that curriculum. I
would also like to thank all my excellent TAs over the
years, who are too numerous to name personally.
I would also like to thank the content creators on sites

like https://opengameart.org and https://freesound.org
for creating excellent game content released under
Creative Commons licenses. These sites were critical to
finding assets for the game projects in this book.
Finally, I’d like to thank my parents, as well my sister,

Nita, and her family. Without their support, inspiration,
and guidance, I never would have gotten here in the first
place. I’d also like to thank my friends, like Kevin, who
understood when I couldn’t go see the latest movie, go to
dinner, or really do anything social because I was
“working on my book.” Well, I guess I have time now.…
ABOUT THE AUTHOR
Sanjay Madhav is a senior lecturer at the University of

Southern California, where he teaches several
programming and video game programming courses. He
has taught at USC since 2008.
Prior to joining USC, Sanjay worked as a programmer for

several video game developers, including Electronic Arts,
Neversoft, and Pandemic Studios. His credited games
include Medal of Honor: Pacific Assault, Tony Hawk’s
Project 8, Lord of the Rings: Conquest, and The
Saboteur.
Sanjay is also the author of Game Programming

Algorithms and Techniques and co-author of
Multiplayer Game Programming. He has a B.S. and an
M.S. in computer science and is pursuing a Ph.D. in
computer science, all from USC.
CHAPTER 1
GAME PROGRAMMING
OVERVIEW
This chapter first discusses how to set up a

development environment and access the
source code for this book. Next, it covers the
core concepts behind any real-time game:
the game loop, how a game updates over
time, and the basics of game input and
output. Throughout the chapter, you will see
how to implement code for a version of the
classic game Pong.
SETTING UP A DEVELOPMENT
ENVIRONMENT
Although it’s possible to write the source code for any
program with a text editor, professional developers
typically use an integrated development
environment (IDE). The advantage of an IDE is
that it provides code completion and debugging in
addition to text editing capabilities. The code for this
book works on both Microsoft Windows and Apple
macOS, and the choice of IDE depends on the choice
of platform. For Windows, this book uses Microsoft
Visual Studio, and for macOS, it uses Apple Xcode.
The remainder of this section contains brief
instructions on setup of these environments on their
respective platforms.
Microsoft Windows
For Windows development, the most popular IDE by
far is Microsoft Visual Studio. Visual Studio also
tends to be the most popular IDE for C++ game
developers, with most PC and console developers
gravitating toward the IDE.
This book uses Microsoft Visual Studio Community 2017,

which is available as a free download at
https://www.visualstudio.com/downloads/. Installation
of Visual Studio Community 2017 requires Microsoft
Windows 7 or higher.
When you run the installer program for Visual Studio, it

asks which “workloads” it should install. Make sure to
minimally select the Game Development with C++
workload. Feel free to also select any other workloads or
options desired.
warning
THERE ARE DIFFERENT VERSIONS OF VISUAL STUDIO: There are
several other products in the Microsoft Visual Studio suite, including Visual
Studio Code and Visual Studio for Mac. Neither of these products are the
same thing as Visual Studio Community 2017, so be careful to install the
correct version!
Apple macOS
On macOS, Apple provides the free Xcode IDE for
development of programs for macOS, iOS, and other
related platforms. The code for this book works in
both Xcode 8 and 9. Note that Xcode 8 requires
macOS 10.11 El Capitan or higher, while Xcode 9
requires macOS 10.12 Sierra or higher.
To install Xcode, simply go to the Apple App Store and

search for Xcode. The first time Xcode runs, it asks if you
want to enable debugging features. Make sure to select
Yes.
GETTING THIS BOOK’S SOURCE

CODE
Most professional developers utilize source
control systems, which, among many other
features, keep a history of the source code. With such
a system, if code changes cause unexpected or
undesired behavior, it’s easy to return to a previously
known working version of code. Furthermore, source
control allows for much easier collaboration between
multiple developers.
One popular source control system is Git, originally

developed by Linus Torvalds of Linux fame. In Git, the
term repository refers to a specific project hosted
under source control. The GitHub website
(https://github.com) provides for easy creation and
management of Git repositories.
The source code for this book is available on GitHub at

https://github.com/gameprogcpp/code. If you are
unfamiliar with the Git system, you can simply click the
green Clone or Download button and choose Download
ZIP to download a compressed ZIP file that contains all
the book’s source code.
Alternatively, if you wish to use Git, you can clone the

repository via the command line, as follows:
Click here to view code image
$ git clone https://github.com/gameprogcpp/code.git
This command works out of the box in the macOS

terminal, but Windows users need to first install Git for
Windows (see https://git-for-windows.github.io).
The source code contains a separate directory (or folder)
for each chapter. For example, this chapter’s source code
is in the Chapter01 directory. In this directory, there is
a Chapter01-Windows.sln file for Microsoft Visual
Studio and a Chapter01-Mac.xcodeproj file for
Apple Xcode. Before moving forward, make sure that you
can compile the code for this chapter.
BEYOND THE C++ STANDARD

LIBRARY
The C++ Standard Library only supports text console
input and output and does not have any graphics
libraries built in. To implement graphics in a C++
program, you must use one of the many available
external libraries.
Unfortunately, many libraries are platform specific,

meaning they work on only one operating system or type
of computer. For example, the Microsoft Windows
application programming interface (API) can create
windows and other UI elements supported by the
Windows operating system. However, the Windows API
doesn’t work on Apple macOS—for obvious reasons.
Likewise, macOS has its own set of libraries for these
same features that do not work on Windows. As a game
programmer, you can’t always avoid platform-specific
libraries. For instance, game developers working with the
Sony PlayStation 4 console must use libraries provided
by Sony.
Luckily, this book sticks to cross-platform libraries,
meaning that the libraries work on many different
platforms. All the source code for this book works on
recent versions of both Windows and macOS. Although
Linux support is untested, the game projects generally
should also work on Linux.
One of the foundational libraries used in this book is

Simple DirectMedia Layer (SDL; see
https://www.libsdl.org). The SDL library is a cross-
platform game development library written in C. It
provides support for creating windows, creating basic 2D
graphics, processing input, and outputting audio, among
other features. SDL is a very lightweight library that
works on many platforms, including Microsoft Windows,
Apple macOS, Linux, iOS, and Android.
In this first chapter, the only external library needed is

SDL. Subsequent chapters use other libraries and
introduce them when needed.
THE GAME LOOP AND GAME

CLASS
One of the big differences between a game and any
other program is that a game must update many
times per second for as long as the program runs. A
game loop is a loop that controls the overall flow
for the entire game program. Like any other loop, a
game loop has code it executes on every iteration,
and it has a loop condition. For a game loop, you
want to continue looping as long as the player hasn’t
quit the game program.
Each iteration of a game loop is a frame. If a game runs

at 60 frames per second (FPS), this means the game
loop completes 60 iterations every second. Many real-
time games run at 30 or 60 FPS. By running this many
iterations per second, the game gives the illusion of
continuous motion even though it’s only updating at
periodic intervals. The term frame rate is
interchangeable with FPS; a frame rate of 60 means the
same thing as 60 FPS.
Anatomy of a Frame
At a high level, a game performs the following steps
on each frame:
1. It processes any inputs.

2. It updates the game world.
3. It generates any outputs.
Each of these three steps has more depth than may be

apparent at first glance. For instance, processing inputs
(step 1) clearly implies detecting any inputs from devices
such as a keyboard, mouse, or controller. But these
might not be the only inputs for a game. Consider a game
that supports an online multiplayer mode. In this case,
the game receives data over the Internet as an input. In
certain types of mobile games, another input might be
what’s visible to the camera, or perhaps GPS
information. Ultimately, the inputs to a game depend on
both the type of game and the platform it runs on.
Updating a game world (step 2) means going through

every object in the game world and updating it as
needed. This could be hundreds or even thousands of
objects, including characters in the game world, parts of
the user interface, and other objects that affect the game
—even if they are not visible.
For step 3, generating any outputs, the most apparent

output is the graphics. But there are other outputs, such
as audio (including sound effects, music, and dialogue).
As another example, most console games have force
feedback effects, such as the controller shaking when
something exciting happens in the game. And for an
online multiplayer game, an additional output would be
data sent to the other players over the Internet.
Consider how this style of game loop might apply to a

simplified version of the classic Namco arcade game Pac-
Man. For this simplified version of the game, assume
that the game immediately begins with Pac-Man in a
maze. The game program continues running until Pac-
Man either completes the maze or dies. In this case, the
“process inputs” phase of the game loop need only read
in the joystick input.
The “update game world” phase of the loop updates Pac-

Man based on this joystick input and then also updates
the four ghosts, pellets, and the user interface. Part of
this update code must determine whether Pac-Man runs
into any ghosts. Pac-Man can also eat any pellets or
fruits he moves over, so the update portion of the loop
also needs to check for this. Because the ghosts are fully
AI controlled, they also must update their logic. Finally,
based on what Pac-Man is doing, the UI may need to
update what data it displays.
note
This style of game loop is single-threaded, meaning it does not take
advantage of modern CPUs that can execute multiple threads
simultaneously. Making a game loop that supports multiple threads is very
complex, and not necessary for games that are smaller in scope. A good
book to learn more about multi-threaded game loops is Jason Gregory’s,
listed in the “Additional Reading” section at the end of this chapter.
The only outputs in the “generate outputs” phase of the

classic Pac-Man game are the audio and video. Listing
1.1 provides pseudocode showing what the game loop for
this simplified version of Pac-Man might look like.
Listing 1.1 Pac-Man Game Loop Pseudocode

void Game::RunLoop()
while (!mShouldQuit)
// Process Inputs
JoystickData j = GetJoystickData();
// Update Game World
UpdatePlayerPosition(j);
for (Ghost& g : mGhost)
if (g.Collides(player))
// Handle Pac-Man colliding with a ghost
else
g.Update();
// Handle Pac-Man eating pellets
// ...
// Generate Outputs
RenderGraphics();
RenderAudio();
Implementing a Skeleton Game Class

You are now ready to use your basic knowledge of the
game loop to create a Game class that contains code
to initialize and shut down the game as well as run
the game loop. If you are rusty in C++, you might
want to first review the content in Appendix A,
“Intermediate C++ Review,” as the remainder of this
book assumes familiarity with C++. In addition, it
may be helpful to keep this chapter’s completed
source code handy while reading along, as doing so
will help you understand how all the pieces fit
together.
Listing 1.2 shows the declaration of the Game class in the

Game.h header file. Because this declaration references
an SDL_Window pointer, you need to also include the
main SDL header file SDL/SDL.h. (If you wanted to
avoid including this here, you could use a forward
declaration.) Many of the member function names are
self-explanatory; for example, the Initialize function
initializes the Game class, the Shutdown function shuts
down the game, and the RunLoop function runs the
game loop. Finally, ProcessInput, UpdateGame, and
GenerateOutput correspond to the three steps of the
game loop.
Currently, the only member variables are a pointer to the

window (which you’ll create in the Initialize
function) and a bool that signifies whether the game
should continue running the game loop.
Listing 1.2 Game Declaration

class Game
public:
Game();
// Initialize the game
bool Initialize();
// Runs the game loop until the game is over
void RunLoop();
// Shutdown the game
void Shutdown();
private:
// Helper functions for the game loop
void ProcessInput();
void UpdateGame();
void GenerateOutput();
// Window created by SDL
SDL_Window* mWindow;
// Game should continue to run
bool mIsRunning;
};
With this declaration in place, you can start

implementing the member functions in Game.cpp. The
constructor simply initializes mWindow to nullptr and
mIsRunning to true.
Game::Initialize
The Initialize function returns true if
initialization succeeds and false otherwise. You
need to initialize the SDL library with the SDL_Init
function. This function takes in a single parameter, a
bitwise-OR of all subsystems to initialize. For now,
you only need to initialize the video subsystem,
which you do as follows:
int sdlResult = SDL_Init(SDL_INIT_VIDEO);
Note that SDL_Init returns an integer. If this integer is

nonzero, it means the initialization failed. In this case,
Game::Initialize should return false because
without SDL, the game cannot continue:
if (sdlResult != 0)
SDL_Log("Unable to initialize SDL: %s", SDL_GetError());
return false;
Using the SDL_Log function is a simple way to output

messages to the console in SDL. It uses the same syntax
as the C printf function, so it supports outputting
variables to printf specifiers such as %s for a C-style
string and %d for an integer. The SDL_GetError
function returns an error message as a C-style string,
which is why it’s passed in as the %s parameter in this
code.
SDL contains several different subsystems that you can

initialize with SDL_Init. Table 1.1 shows the most
commonly used subsystems; for the full list, consult the
SDL API reference at https://wiki.libsdl.org.
Table 1.1 SDL Subsystem Flags of Note
Flag Subsystem
SDL_INIT_AUDIO Audio device management,
playback, and recording
SDL_INIT_VIDEO Video subsystem for creating a

window, interfacing with
OpenGL, and 2D graphics
SDL_INIT_HAPTIC Force feedback subsystem
SDL_INIT_GAMECONTROLLER Subsystem for supporting

controller input devices
If SDL initializes successfully, the next step is to create a

window with the SDL_CreateWindow function. This is
just like the window that any other Windows or macOS
program uses. The SDL_CreateWindow function takes
in several parameters: the title of the window, the x/y
coordinates of the top-left corner, the width/height of
the window, and optionally any window creation flags:
mWindow = SDL_CreateWindow(
"Game Programming in C++ (Chapter 1)", // Window title
100, // Top left x-coordinate of window
100, // Top left y-coordinate of window
1024, // Width of window
768, // Height of window
0 // Flags (0 for no flags set)

);
As with the SDL_Init call, you should verify that

SDL_CreateWindow succeeded. In the event of failure,
mWindow will be nullptr, so add this check:
if (!mWindow)
SDL_Log("Failed to create window: %s", SDL_GetError());
return false;
As with the initialization flags, there are several possible

window creation flags, as shown in Table 1.2. As before,
you can use a bitwise-OR to pass in multiple flags.
Although many commercial games use full-screen mode,
it’s faster to debug code if the game runs in windowed
mode, which is why this book shies away from full
screen.
Table 1.2 Window Creation Flags of Note
Flag Result
SDL_WINDOW_FULLSCREEN Use full-screen mode
SDL_WINDOW_FULLSCREEN_DESKTOP Use full-screen mode at

the current desktop
resolution (and ignore
width/height parameters
to SDL_CreateWindow)
SDL_WINDOW_OPENGL Add support for the

OpenGL graphics library
SDL_WINDOW_RESIZABLE Allow the user to resize

the window
If SDL initialization and window creation succeeds,

Game::Initialize returns true.
Game::Shutdown
The Shutdown function does the opposite of
Initialize. It first destroys the SDL_Window with
SDL_DestroyWindow and then closes SDL with
SDL_Quit:
void Game::Shutdown()
SDL_DestroyWindow(mWindow);
SDL_Quit();
Game::RunLoop
The RunLoop function keeps running iterations of
the game loop until mIsRunning becomes false, at
which point the function returns. Because you have
the three helper functions for each phase of the game
loop, RunLoop simply calls these helper functions
inside the loop:
void Game::RunLoop()
while (mIsRunning)
ProcessInput();
UpdateGame();
GenerateOutput();
For now, you won’t implement these three helper

functions, which means that once in the loop, the game
won’t do anything just yet. You’ll continue to build on
this Game class and implement these helper functions
throughout the remainder of the chapter.
Main Function
Although the Game class is a handy encapsulation of
the game’s behavior, the entry point of any C++
program is the main function. You must implement
a main function (in Main.cpp) as shown in Listing
1.3.
Listing 1.3 main Implementation

int main(int argc, char** argv)
Game game;
bool success = game.Initialize();
if (success)
game.RunLoop();
game.Shutdown();
return 0;
This implementation of main first constructs an instance

of the Game class. It then calls Initialize, which
returns true if the game successfully initializes, and
false otherwise. If the game initializes, you then enter
the game loop with the call to RunLoop. Finally, once the
loop ends, you call Shutdown on the game.
With this code in place, you can now run the game
project. When you do, you see a blank window, as shown
in Figure 1.1 (though on macOS, this window may appear
black instead of white). Of course, there’s a problem: The
game never ends! Because no code changes the
mIsRunning member variable, the game loop never
ends, and the RunLoop function never returns.
Naturally, the next step is to fix this problem by allowing
the player to quit the game.
Figure 1.1 Creating a blank window
Basic Input Processing

In any desktop operating system, there are several
actions that the user can perform on application
windows. For example, the user can move a window,
minimize or maximize a window, close a window
(and program), and so on. A common way to
represent these different actions is with events.
When the user does something, the program receives
events from the operating system and can choose to
respond to these events.
SDL manages an internal queue of events that it receives

from the operating system. This queue contains events
for many different window actions, as well as events
related to input devices. Every frame, the game must poll
the queue for any events and choose either to ignore or
process each event in the queue. For some events, such
as moving the window around, ignoring the event means
SDL will just automatically handle it. But for other
events, ignoring the event means nothing will happen.
Because events are a type of input, it makes sense to

implement event processing in ProcessInput. Because
the event queue may contain multiple events on any
given frame, you must loop over all events in the queue.
The SDL_PollEvent function returns true if it finds
an event in the queue. So, a very basic implementation of
ProcessInput would keep calling SDL_PollEvent as
long as it returns true:
void Game::ProcessInput()
{
SDL_Event event;
// While there are still events in the queue
while (SDL_PollEvent(&event))
Note that the SDL_PollEvent function takes in an

SDL_Event by pointer. This stores any information
about the event just removed from the queue.
Although this version of ProcessInput makes the

game window more responsive, the player still has no
way to quit the game. This is because you simply remove
all the events from the queue and don’t respond to them.
Given an SDL_Event, the type member variable

contains the type of the event received. So, a common
approach is to create a switch based on the type inside
the PollEvent loop:
SDL_Event event;
switch (event.type)
{
// Handle different event types here
One useful event is SDL_QUIT, which the game receives

when the user tries to close the window (either by
clicking on the X or using a keyboard shortcut). You can
update the code to set mIsRunning to false when it
sees an SDL_QUIT event in the queue:
SDL_Event event;
switch (event.type)
case SDL_QUIT:
mIsRunning = false;
break;
}
Now when the game is running, clicking the X on the
window causes the while loop inside RunLoop to
terminate, which in turn shuts down the game and exits
the program. But what if you want the game to quit when
the user presses the Escape key? While you could check
for a keyboard event corresponding to this, an easier
approach is to grab the entire state of the keyboard with
SDL_GetKeyboardState, which returns a pointer to
an array that contains the current state of the keyboard:
const Uint8* state = SDL_GetKeyboardState(NULL);
Given this array, you can then query a specific key by

indexing into this array with a corresponding
SDL_SCANCODE value for the key. For example, the
following sets mIsRunning to false if the user presses
Escape:
if (state[SDL_SCANCODE_ESCAPE])
mIsRunning = false;
Combining all this yields the current version of

ProcessInput, shown in Listing 1.4. Now when
running the game, the user can quit either by closing the
window or pressing the Escape key.
Listing 1.4 Game::ProcessInput Implementation

SDL_Event event;
switch (event.type)
// If this is an SDL_QUIT event, end loop
case SDL_QUIT:
mIsRunning = false;
break;
// Get state of keyboard
const Uint8* state = SDL_GetKeyboardState(NULL);

// If escape is pressed, also end loop
if (state[SDL_SCANCODE_ESCAPE])
mIsRunning = false;
BASIC 2D GRAPHICS
Before you can implement the “generate outputs”
phase of the game loop, you need some
understanding of how 2D graphics work for games
Most displays in use today—whether televisions,

computer monitors, tablets, or smartphones—use raster
graphics, which means the display has a two-
dimensional grid of picture elements (or pixels). These
pixels can individually display different amounts of light
as well as different colors. The intensity and color of
these pixels combine to create a perception of a
continuous image for the viewer. Zooming in on a part of
a raster image makes each individual pixel discernable,
as you can see in Figure 1.2.
Figure 1.2 Zooming in on part of an image shows its
distinct pixels
The resolution of a raster display refers to the width

and height of the pixel grid. For example, a resolution of
1920×1080, commonly known as 1080p, means that
there are 1080 rows of pixels, with each row containing
1920 pixels. Similarly, a resolution of 3840×2160, known
as 4K, has 2160 rows with 3840 pixels per row.
Color displays mix colors additively to create a specific

hue for each pixel. A common approach is to mix three
colors together: red, green, and blue (abbreviated RGB).
Different intensities of these RGB colors combine to
create a range (or gamut) of colors. Although many
modern displays also support color formats other than
RGB, most video games output final colors in RGB.
Whether or not RGB values convert to something else for
display on the monitor is outside the purview of the
game programmer.
However, many games internally use a different color

representation for much of their graphics computations.
For example, many games internally support
transparency with an alpha value. The abbreviation
RGBA references RGB colors with an additional alpha
component. Adding an alpha component allows certain
objects in a game, such as windows, to have some
amount of transparency. But because few if any displays
support transparency, the game ultimately needs to
calculate a final RGB color and compute any perceived
transparency itself.
The Color Buffer

For a display to show an RGB image, it must know
the colors of each pixel. In computer graphics, the
color buffer is a location in memory containing the
color information for the entire screen. The display
can use the color buffer for drawing the contents
screen. Think of the color buffer as a two-
dimensional array, where each (x, y) index
corresponds to a pixel on the screen. In every frame
during the “generate outputs” phase of the game
loop, the game writes graphical output into the color
buffer.
The memory usage of the color buffer depends on the

number of bits that represent each pixel, called the color
depth. For example, in the common 24-bit color depth,
red, green, and blue each use 8 bits. This means there are
224, or 16,777,216, unique colors. If the game also wants
to store an 8-bit alpha value, this results in a total of 32
bits for each pixel in the color buffer. A color buffer for a
1080p (1920×1080) target resolution with 32 bits per
pixel uses 1920×1080×4 bytes, or approximately 7.9 MB.
note
Many game programmers also use the term framebuffer to reference the
location in memory that contains the color data for a frame. However, a more
precise definition of framebuffer is that it is the combination of the color buffer
and other buffers (such as the depth buffer and stencil buffer). In the interest
of clarity, this book references the specific buffers.
Some recent games use 16 bits per RGB component,

which increases the number of unique colors. Of course,
this doubles the memory usage of the color buffer, up to
approximately 16 MB for 1080p. This may seem like an
insignificant amount, given that most video cards have
several gigabytes of video memory available. But when
considering all the other memory usage of a cutting-edge
game, 8 MB here and 8 MB there quickly adds up.
Although most displays at this writing do not support 16
bits per color, some manufacturers now offer displays
that support color depths higher than 8 bits per color.
Given an 8-bit value for a color, there are two ways to

reference this value in code. One approach involves
simply using an unsigned integer corresponding to the
number of bits for each color (or channel). So, for a
color depth with 8 bits per channel, each channel has a
value between 0 and 255. The alternative approach is to
normalize the integer over a decimal range from 0.0 to
1.0.
One advantage of using a decimal range is that a value

yields roughly the same color, regardless of the
underlying color depth. For example, the normalized
RGB value (1.0, 0.0, 0.0) yields pure red whether the
maximum value of red is 255 (8 bits per color) or 65,535
(16 bits per color). However, the unsigned integer RGB
value (255, 0, 0) yields pure red only if there are 8 bits
per color. With 16 bits per color, (255, 0, 0) is nearly
black.
Converting between these two representations is

straightforward. Given an unsigned integer value, divide
it by the maximum unsigned integer value to get the
normalized value. Conversely, given a normalized
decimal value, multiply it by the maximum unsigned
integer value to get an unsigned integer value. For now,
you should use unsigned integers because the SDL
library expects them.
Double Buffering
As mentioned earlier in this chapter, games update
several times per second (at the common rates of 30
and 60 FPS). If a game updates the color buffer at
the same rate, this gives the illusion of motion, much
the way a flipbook appears to show an object in
motion when you flip through the pages.
However, the refresh rate, or the frequency at which

the display updates, may be different from the game’s
frame rate. For example, most NTSC TV displays have a
refresh rate of 59.94 Hz, meaning they refresh very
slightly less than 60 times per second. However, some
newer computer monitors support a 144 Hz refresh rate,
which is more than twice as fast.
Furthermore, no current display technology can
instantaneously update the entire screen at once. There
always is some update order—whether row by row,
column by column, in a checkerboard, and so on.
Whatever update pattern the display uses, it takes some
fraction of a second for the whole screen to update.
Suppose a game writes to the color buffer, and the

display reads from that same color buffer. Because the
timing of the game’s frame rate may not directly match
the monitor’s refresh rate, it’s very like that the display
will read from the color buffer while the game is writing
to the buffer. This can be problematic.
For example, suppose the game writes the graphical data

for frame A into the color buffer. The display then starts
reading from the color buffer to show frame A on the
screen. However, before the display finishes drawing
frame A onto the screen, the game overwrites the color
buffer with the graphical data for frame B. The display
ends up showing part of frame A and part of frame B on
the screen. Figure 1.3 illustrates this problem, known as
screen tearing.
Figure 1.3 Simulation of screen tearing with a
camera panning to the right
Eliminating screen tearing requires two changes. First,

rather than having one color buffer that the game and
display must share, you create two separate color buffers.
Then the game and display alternate between the color
buffers they use every frame. The idea is that with two
separate buffers, the game can write to one (the back
buffer) and, at the same time, the display can read from
the other one (the front buffer). After the frame
completes, the game and display swap their buffers. Due
to the use of two color buffers, the name for this
technique is double buffering.
As a more concrete example, consider the process shown

in Figure 1.4. On frame A, the game writes its graphical
output to buffer X, and the display draws buffer Y to the
screen (which is empty). When this process completes,
the game and display swap which buffers they use. Then
on frame B, the game draws its graphical output to buffer
Y, while the display shows buffer X on screen. On frame
C, the game returns to buffer X, and the display returns
to buffer Y. This swapping between the two buffers
continues until the game program closes.
Figure 1.4 Double buffering involves swapping the

buffers used by the game and display every frame
However, double buffering by itself does not eliminate

screen tearing. Screen tearing still occurs if the display is
drawing buffer X when the game wants to start writing to
X. This usually happens only if the game is updating too
quickly. The solution to this problem is to wait until the
display finishes drawing its buffer before swapping. In
other words, if the display is still drawing buffer X when
the game wants to swap back to buffer X, the game must
wait until the display finishes drawing buffer X.
Developers call this approach vertical
synchronization, or vsync, named after the signal
that monitors send when they are about to refresh the
screen.
With vertical synchronization, the game might have to
occasionally wait for a fraction of a second for the display
to be ready. This means that the game loop may not be
able to achieve its target frame rate of 30 or 60 FPS
exactly. Some players argue that this causes
unacceptable stuttering of the frame rate. Thus, the
decision on whether to enable vsync varies depending on
the game or player. A good idea is to offer vsync as an
option in the engine so that you can choose between
occasional screen tearing or occasional stuttering.
Recent advances in display technology seek to solve this

dilemma with an adaptive refresh rate that varies
based on the game. With this approach, rather than the
display notifying the game when it refreshes, the game
tells the display when to refresh. This way, the game and
display are in sync. This provides the best of both worlds
as it eliminates both screen tearing and frame rate
stuttering. Unfortunately, at this writing, adaptive
refresh technology is currently available only on certain
high-end computer monitors.
Implementing Basic 2D Graphics

SDL has a simple set of functions for drawing 2D
graphics. Because the focus of this chapter is 2D, you
can stick with these functions. Starting in Chapter 5,
“OpenGL,” you’ll switch to the OpenGL library for
graphics, as it supports both 2D and 3D.
Initialization and Shutdown

To use SDL’s graphics code, you need to construct an
SDL_Renderer via the SDL_CreateRenderer
function. The term renderer generically refers to
any system that draws graphics, whether 2D or 3D.
Because you need to reference this SDL_Renderer
object every time you draw something, first add an
mRenderer member variable to Game:
SDL_Renderer* mRenderer;
Next, in Game::Initialize, after creating the

window, create the renderer:
mRenderer = SDL_CreateRenderer(
mWindow, // Window to create renderer for
-1, // Usually -1
SDL_RENDERER_ACCELERATED | SDL_RENDERER_PRESENTVSYNC
);
The first parameter to SDL_CreateRenderer is the

pointer to the window (which you saved in mWindow).
The second parameter specifies which graphics driver to
use; this might be relevant if the game has multiple
windows. But with only a single window, the default is
-1, which means to let SDL decide. As with the other
SDL creation functions, the last parameter is for
initialization flags. Here, you choose to use an
accelerated renderer (meaning it takes advantage of
graphics hardware) and enable vertical synchronization.
These two flags are the only flags of note for
SDL_CreateRenderer.
As with SDL_CreateWindow, the

SDL_CreateRenderer function returns a nullptr if it
fails to initialize the renderer. As with initializing
SDL, Game::Initialize returns false if the
renderer fails to initialize.
To shut down the renderer, simply add a call to

SDL_DestroyRenderer in Game::Shutdown:
SDL_DestroyRenderer(mRenderer);
Basic Drawing Setup

At a high level, drawing in any graphics library for
games usually involves the following steps:
1. Clear the back buffer to a color (the game’s current

buffer).
2. Draw the entire game scene.
3. Swap the front buffer and back buffer.
First, let’s worry about the first and third steps. Because
graphics are an output, it makes sense to put graphics
drawing code in Game::GenerateOutput.
To clear the back buffer, you first need to specify a color
with SDL_SetRenderDrawColor. This function takes
in a pointer to the renderer, as well as the four RGBA
components (from 0 to 255). For example, to set the
color as blue with 100% opacity, use the following:
SDL_SetRenderDrawColor(
mRenderer,
0, // R
0, // G
255, // B
255 // A
);
Next, call SDL_RenderClear to clear the back buffer to

the current draw color:
SDL_RenderClear(mRenderer);
The next step—skipped for now—is to draw the entire

game scene.
Finally, to swap the front and back buffers, you call

SDL_RenderPresent:
SDL_RenderPresent(mRenderer);
With this code in place, if you now run the game, you’ll
see a filled-in blue window, as shown in Figure 1.5.
Figure 1.5 Game drawing a blue background
Drawing Walls, a Ball, and a Paddle

This chapter’s game project is a version of the classic
video game Pong, where a ball moves around the
screen, and the player controls a paddle that can hit
the ball. Making a version of Pong is a rite of passage
for any aspiring game developer—analogous to
making a “Hello World” program when first learning
how to program. This section explores drawing
rectangles to represent the objects in Pong. Because
these are objects in the game world, you draw them
in GenerateOuput—after clearing the back buffer
but before swapping the front and back buffers.
For drawing filled rectangles, SDL has a

SDL_RenderFillRect function. This function takes in
an SDL_Rect that represents the bounds of the rectangle
and draws a filled-in rectangle using the current draw
color. Of course, if you keep the draw color the same as
the background, you won’t see any rectangles. You
therefore need to change the draw color to white:
SDL_SetRenderDrawColor(mRenderer, 255, 255, 255, 255);
Next, to draw the rectangle, you need to specify

dimensions via an SDL_Rect struct. The rectangle has
four parameters: the x/y coordinates of the top-left
corner of the rectangle onscreen, and the width/height of
the rectangle. Keep in mind that in SDL rendering, as in
many other 2D graphics libraries, the top-left corner of
the screen is (0, 0), positive x is to the right, and positive
y is down.
For example, if you want to draw a rectangle at the top of

the screen, you can use the following declaration of an
SDL_Rect:
SDL_Rect wall{
0, // Top left x
0, // Top left y
1024, // Width
thickness // Height
};
Here, the x/y coordinates of the top-left corner are (0, 0),
meaning the rectangle will be at the top left of the screen.
You hard-code the width of the rectangle to 1024,
corresponding to the width of the window. (It’s generally
frowned upon to assume a fixed window size, as is done
here, and you’ll remove this assumption in later
chapters.) The thickness variable is const int set to
15, which makes it easy to adjust the thickness of the
wall.
Finally, you draw the rectangle with

SDL_RenderFillRect, passing in SDL_Rect by
pointer:
SDL_RenderFillRect(mRenderer, &wall);
The game then draws a wall in the top part of the screen.
You can use similar code to draw the bottom wall and the
right wall, only changing the parameters of the
SDL_Rect. For example, the bottom wall could have the
same rectangle as the top wall except that the top-left y
coordinate could be 768 - thickness.
Unfortunately, hard-coding the rectangles for the ball

and paddle does not work because both objects will
ultimately move in the UpdateGame stage of the loop.
Although it makes some sense to represent both the ball
and paddle as classes, this discussion doesn’t happen
until Chapter 2, “Game Objects and 2D Graphics.” In the
meantime, you can just use member variables to store
the center positions of both objects and draw their
rectangles based on these positions.
First, declare a simple Vector2 struct that has both x

and y components:
struct Vector2
float x;
float y;
};
For now, think of a vector (not a std::vector) as a

simple container for coordinates. Chapter 3, “Vectors
and Basic Physics,” explores the topic of vectors in much
greater detail.
Next, add two Vector2s as member variables to Game—

one for the paddle position (mPaddlePos) and one for
the ball’s position (mBallPos). The game constructor
then initializes these to sensible initial values: the ball
position to the center of the screen and the paddle
position to the center of the left side of the screen.
Armed with these member variables, you can then draw

rectangles for the ball and paddle in GenerateOutput.
However, keep in mind that the member variables
represent the center points of the paddle and ball, while
you define an SDL_Rect in terms of the top-left point.
To convert from the center point to the top-left point,
you simply subtract half the width/height from the x and
y coordinates, respectively. For example, the following
rectangle works for the ball:
SDL_Rect ball{
static_cast<int>(mBallPos.x - thickness/2),
static_cast<int>(mBallPos.y - thickness/2),
thickness,
thickness
};
The static casts here convert mBallPos.x and

mBallPos.y from floats into integers (which SDL_Rect
uses). In any event, you can make a similar calculation
for drawing the paddle, except its width and height are
different sizes.
With all these rectangles, the basic game drawing now
works, as shown in Figure 1.6. The next step is to
implement the UpdateGame phase of the loop, which
moves the ball and paddle.
Figure 1.6 A game with walls, a paddle, and a ball

drawing
UPDATING THE GAME

Most video games have some concept of time
progression. For real-time games, you measure this
progression of time in fractions of a second. For
example, a game running at 30 FPS has roughly 33
milliseconds (ms) elapse from frame to frame.
Remember that even though a game appears to
feature continuous movement, it is merely an
illusion. The game loop actually runs several times
per second, and every iteration of the game loop
updates the game in a discrete time step. So, in the
30 FPS example, each iteration of the game loop
should simulate 33ms of time progression in the
game. This section looks at how to consider this
discrete progression of time when programming a
game.
Real Time and Game Time

It is important to distinguish real time, the time
elapsing in the real world, from game time, the
time elapsing in the game’s world. Although there
often is a 1:1 correspondence between real time and
game time, this isn’t always the case. Take, for
instance, a game in a paused state. Although a great
deal of time might elapse in the real world, the game
doesn’t advance at all. It’s not until the player
unpauses the game that the game time resumes
updating.
There are many other instances where real time and

game time might diverge. For example, some games
feature a “bullet time” gameplay mechanic that reduces
the speed of the game. In this case, the game time must
update at a substantially slower rate than actual time. On
the opposite end of the spectrum, many sports games
feature sped-up time. In a football game, rather than
requiring a player to sit through 15 full minutes per
quarter, the game may update the clock twice as fast, so
each quarter takes only 7.5 minutes. And some games
may even have time advance in reverse. For example,
Prince of Persia: The Sands of Time featured a unique
mechanic where the player could rewind the game time
to a certain point.
With all these ways real time and game time might
diverge, it’s clear that the “update game” phase of the
game loop should account for elapsed game time.
Logic as a Function of Delta Time

Early game programmers assumed a specific
processor speed and, therefore, a specific frame rate.
The programmer might write the code assuming an 8
MHz processor, and if it worked properly for those
processors, the code was just fine. When assuming a
fixed frame rate, code that updates the position of an
enemy might look something like this:
// Update x position by 5 pixels
enemy.mPosition.x += 5;
If this code moves the enemy at the desired speed on an

8 MHz processor, what happens on a 16 MHz processor?
Well, because the game loop now runs twice as fast, the
enemy will now also move twice as fast. This could be the
difference between a game that’s challenging for players
and one that’s impossible. Imagine running this game on
a modern processor that is thousands of times faster. The
game would be over in a heartbeat!
To solve this issue, games use delta time: the amount of

elapsed game time since the last frame. To convert the
preceding code to using delta time, instead of thinking of
movement as pixels per frame, you should think of it as
pixels per second. So, if the ideal movement speed is 150
pixels per second, the following code is much more
flexible:
// Update x position by 150 pixels/second
enemy.mPosition.x += 150 * deltaTime;
Now the code will work well regardless of the frame rate.
At 30 FPS, the delta time is ~0.033, so the enemy will
move 5 pixels per frame, for a total of 150 pixels per
second. At 60 FPS, the enemy will move only 2.5 pixels
per frame but will still move a total of 150 pixels per
second. The movement certainly will be smoother in the
60 FPS case, but the overall per-second speed remains
the same.
Because this works across many frame rates, as a rule of

thumb, everything in the game world should update as a
function of delta time.
To help calculate delta time, SDL provides an

SDL_GetTicks member function that returns the
number of milliseconds elapsed since the SDL_Init
function call. By saving the result of SDL_GetTicks
from the previous frame in a member variable, you can
use the current value to calculate delta time.
First, you declare an mTicksCount member variable

(initializing it to zero in the constructor):
Uint32 mTicksCount;
Using SDL_GetTicks, you can then create a first

implementation of Game::UpdateGame:
void Game::UpdateGame()
// Delta time is the difference in ticks from last frame
// (converted to seconds)
float deltaTime = (SDL_GetTicks() - mTicksCount) / 1000.0f;
// Update tick counts (for next frame)
mTicksCount = SDL_GetTicks();
// TODO: Update objects in game world as function of delta time!
// ...
}
Consider what happens the very first time you call
UpdateGame. Because mTicksCount starts at zero, you
end up with some positive value of SDL_GetTicks (the
milliseconds since initialization) and divide it by
1000.0f to get a delta time in seconds. Next, you save
the current value of SDL_GetTicks in mTicksCount.
On the next frame, the deltaTime line calculates a new
delta time based on the old value of mTicksCount and
the new value. Thus, on every frame, you compute a delta
time based on the ticks elapsed since the previous frame.
Although it may seem like a great idea to allow the game

simulation to run at whatever frame rate the system
allows, in practice there can be several issues with this.
Most notably, any game that relies on physics (such as a
platformer with jumping) will have differences in
behavior based on the frame rate.
Though there are more complex solutions to this

problem, the simplest solution is to implement frame
limiting, which forces the game loop to wait until a
target delta time has elapsed. For example, suppose that
the target frame rate is 60 FPS. If a frame completes
after only 15ms, frame limiting says to wait an additional
~1.6ms to meet the 16.6ms target time.
Conveniently, SDL also provides a method for frame

limiting. For example, to ensure that at least 16ms
elapses between frames, you can add the following code
to the start of UpdateGame:

while (!SDL_TICKS_PASSED(SDL_GetTicks(), mTicksCount + 16))
You also must watch out for a delta time that’s too high.
Most notably, this happens when stepping through game
code in the debugger. For example, if you pause at a
breakpoint in the debugger for five seconds, you’ll end up
with a huge delta time, and everything will jump far
forward in the simulation. To fix this problem, you can
clamp the delta time to a maximum value (such as
0.05f). This way, the game simulation will never jump
too far forward on any one frame. This yields the version
of Game::UpdateGame in Listing 1.5. While you aren’t
updating the position of the paddle or ball just yet, you
are at least calculating the delta time value.
Listing 1.5 Game::UpdateGame Implementation

// Wait until 16ms has elapsed since last frame
while (!SDL_TICKS_PASSED(SDL_GetTicks(), mTicksCount + 16))
// Delta time is the difference in ticks from last frame

// (converted to seconds)
float deltaTime = (SDL_GetTicks() - mTicksCount) / 1000.0f;
// Clamp maximum delta time value
if (deltaTime > 0.05f)
deltaTime = 0.05f;
// TODO: Update objects in game world as function of delta time!
Updating the Paddle’s Position

In Pong, the player controls the position of the
paddle based on input. Suppose you want the W key
to move the paddle up and the S key to move the
paddle down. Pressing neither key or both keys
should mean the paddle doesn’t move at all.
You can make this concrete by using a mPaddleDir

integer member variable that’s set to 0 if the paddle
doesn’t move, -1 if if the paddle moves up (negative y),
and 1 if the paddle moves down (positive y).
Because the player controls the position of the paddle via
keyboard input, you need code in ProcessInput that
updates mPaddleDir based on the input:
mPaddleDir = 0;
if (state[SDL_SCANCODE_W])
mPaddleDir -= 1;
if (state[SDL_SCANCODE_S])
mPaddleDir += 1;
Note how you add and subtract from mPaddleDir,

which ensures that if the player presses both keys,
mPaddleDir is zero.
Next, in UpdateGame, you can add code that updates the

paddle based on delta time:
if (mPaddleDir != 0)
mPaddlePos.y += mPaddleDir * 300.0f * deltaTime;

}
Here, you update the y position of the paddle based on

the paddle direction, a speed of 300.0f pixels/second,
and delta time. If mPaddleDir is -1, the paddle will
move up, and if it’s 1, it’ll move down.
One problem is that this code allows the paddle to move

off the screen. To fix this, you can add boundary
conditions for the paddle’s y position. If the position is
too high or too low, move it back to a valid position:
if (mPaddleDir != 0)
mPaddlePos.y += mPaddleDir * 300.0f * deltaTime;
// Make sure paddle doesn't move off screen!
if (mPaddlePos.y < (paddleH/2.0f + thickness))
mPaddlePos.y = paddleH/2.0f + thickness;
else if (mPaddlePos.y > (768.0f - paddleH/2.0f - thickness))
mPaddlePos.y = 768.0f - paddleH/2.0f - thickness;

}
Here, the paddleH variable is a constant that describes

the height of the paddle. With this code in place, the
player can now move the paddle up and down, and the
paddle can’t move offscreen.
Updating the Ball’s Position

Updating the position of the ball is a bit more
complex than updating the position of the paddle.
First, the ball travels in both the x and y directions,
not just in one direction. Second, the ball needs to
bounce off the walls and paddles, which changes the
direction of travel. So you need to both represent the
velocity (speed and direction) of the ball and
perform collision detection to determine if the
ball collides with a wall.
To represent the ball’s velocity, add another Vector2

member variable called mBallVel. Initialize mBallVel
to (-200.0f, 235.0f), which means the ball starts
out moving −200 pixels/second in the x direction and
235 pixels/second in the y direction. (In other words, the
ball moves diagonally down and to the left.)
To update the position of the ball in terms of the velocity,

add the following two lines of code to UpdateGame:

mBallPos.x += mBallVel.x * deltaTime;
mBallPos.y += mBallVel.y * deltaTime;
This is like updating the paddle’s position, except now

you are updating the position of the ball in both the x
and y directions.
Next, you need code that bounces the ball off walls. The
code for determining whether the ball collides with a wall
is like the code for checking whether the paddle is
offscreen. For example, the ball collides with the top wall
if its y position is less than or equal to the height of the
ball.
The important question is: what to do when the ball

collides with the wall? For example, suppose the ball
moves upward and to the right before colliding against
the top wall. In this case, you want the ball to now start
moving downward and to the right. Similarly, if the ball
hits the bottom wall, you want the ball to start moving
upward. The insight is that bouncing off the top or
bottom wall negates the y component of the velocity, as
shown in Figure 1.7(a). Similarly, colliding with the
paddle on the left or wall on the right should negate the x
component of the velocity.
Figure 1.7 (a) The ball collides with the top wall so
starts moving down. (b) The y difference between the
ball and paddle is too large
For the case of the top wall, this yields code like the
following:
if (mBallPos.y <= thickness)
mBallVel.y *= -1;
However, there’s a key problem with this code. Suppose

the ball collides with the top wall on frame A, so the code
negates the y velocity to make the ball start moving
downward. On frame B, the ball tries to move away from
the wall, but it doesn’t move far enough. Because the ball
still collides with the wall, the code negates the y velocity
again, which means the ball starts moving upward. Then
on every subsequent frame, the code keeps negating the
ball’s y velocity, so it is forever stuck on the top wall.
To fix this issue of the ball getting stuck, you need an
additional check. You want to only negate the y velocity if
the ball collides with the top wall and the ball is moving
toward the top wall (meaning the y velocity is negative):
if (mBallPos.y <= thickness && mBallVel.y < 0.0f)
mBallVel.y *= -1;
This way, if the ball collides with the top wall but is
moving away from the wall, you do not negate the y
velocity.
The code for colliding against the bottom and right walls
is very similar to the code for colliding against the top
wall. Colliding against the paddle, however, is slightly
more complex. First, you calculate the absolute value of
the difference between the y position of the ball and the y
position of the paddle. If this difference is greater than
half the height of the paddle, the ball is too high or too
low, as shown earlier in Figure 1.7(b). You also need to
check that the ball’s x-position lines up with the paddle,
and the ball is not trying to move away from the paddle.
Satisfying all these conditions means the ball collides
with the paddle, and you should negate the x velocity:

if (
// Our y-difference is small enough
diff <= paddleH / 2.0f &&
// Ball is at the correct x-position
mBallPos.x <= 25.0f && mBallPos.x >= 20.0f &&
// The ball is moving to the left
mBallVel.x < 0.0f)
mBallVel.x *= -1.0f;
With this code complete, the ball and paddle now both
move onscreen, as in Figure 1.8. You have now
completed your simple version of Pong!
Figure 1.8 Final version of Pong
GAME PROJECT
This chapter’s game project implements the full Pong
game code constructed throughout the chapter. To
control the paddle, the player uses the W and S keys.
The game ends when the ball moves offscreen. The
code is available in the book’s GitHub repository in
the Chapter01 directory. Open Chapter01-
windows.sln in Windows and Chapter01-
mac.xcodeproj on Mac. (For instructions on how
to access the GitHub repository, consult the
instructions at the beginning of this chapter.)
SUMMARY
Real-time games update many times per second via a
loop called the game loop. Each iteration of this loop
is a frame. For example, 60 frames per second means
that there are 60 iterations of the game loop per
second. The game loop has three main phases that it
completes every frame: processing input, updating
the game world, and generating output. Input
involves not only input devices such as the keyboard
and mouse but networking data, replay data, and so
on. Outputs include graphics, audio, and force
feedback controllers.
Most displays use raster graphics, where the display

contains a grid of pixels. The size of the grid depends on
the resolution of the display. The game maintains a color
buffer that saves color data for every pixel. Most games
use double buffering, where there are two color buffers,
and the game and display alternate between using these
buffers. This helps reduce the amount of screen tearing
(that is, the screen showing parts of two frames at once).
To eliminate screen tearing, you also must enable
vertical synchronization, which means the buffers swap
only when the display is ready.
For a game to work properly at variable frame rates, you

need to write all game logic as a function of delta time—
the time interval between frames. Thus, the “update
game world” phase of the game loop should account for
delta time. You can further add frame limiting to ensure
that the frame rate does not go over some set cap.
In this chapter, you have combined all these different

techniques to create a simple version of the classic video
game Pong.
ADDITIONAL READING
Jason Gregory dedicates several pages to discussing
the different formulations of a game loop, including
how some games take better advantage of multi-core
CPUs. There are also many excellent references
online for the various libraries used; for example, the
SDL API reference is handy.
Gregory, Jason. Game Engine Architecture,

2nd edition. Boca Raton: CRC Press, 2014.
SDL API Reference.
https://wiki.libsdl.org/APIByCategory.
Accessed June 15, 2016.
EXERCISES
Both of this chapter’s exercises focus on modifying
your version of Pong. The first exercise involves
adding a second player, and the second exercise
involves adding support for multiple balls.
Exercise 1.1
The original version of Pong supported two players.
Remove the right wall onscreen and replace that wall
with a second paddle for player 2. For this second
paddle, use the I and K keys to move the paddle up
and down. Supporting a second paddle requires
duplicating all the functionality of the first paddle: a
member variable for the paddle’s position, the
direction, code to process input for player 2, code
that draws the paddle, and code that updates the
paddle. Finally, make sure to update the ball collision
code so that the ball correctly collides with both
paddles.
Exercise 1.2
Many pinball games support “multiball,” where
multiple balls are in play at once. It turns out
multiball is also fun for Pong! To support multiple
balls, create a Ball struct that contains two
Vector2s: one for the position and one for the
velocity. Next, create a std::vector<Ball>
member variable for Game to store these different
balls. Then change the code in Game::Initialize
to initialize the positions and velocities of several
balls. In Game::UpdateGame, change the ball
update code so that rather than using the individual
mBallVel and mBallPos variables, the code loops
over the std::vector for all the balls.
CHAPTER 2
GAME OBJECTS AND 2D

GRAPHICS
Most games have many different characters

and other objects, and an important decision
is how to represent these objects. This
chapter first covers different methods of
object representation. Next, it continues the
discussion of 2D graphics techniques by
introducing sprites, sprite animations, and
scrolling backgrounds. This chapter
culminates with a side-scrolling demo that
applies the covered techniques.
GAME OBJECTS
The Pong game created in Chapter 1 does not use
separate classes to represent the wall, paddles, and
ball. Instead, the Game class uses member variables
to track the position and velocity of the different
elements of the game. While this can work for a very
simple game, it’s not a scalable solution. The term
game object refers to anything in the game world
that updates, draws, or both updates and draws.
There are several methods to represent game objects.
Some games employ object hierarchies, others
employ composition, and still others utilize more
complex methods. Regardless of the implementation,
a game needs some way to track and update these
game objects.
Types of Game Objects

A common type of game object is one that’s both
updated every frame during the “update game world”
phase of the loop and drawn every frame during the
“generate outputs” phase. Any character, creature, or
otherwise movable object falls under this umbrella.
For example, in Super Mario Bros., Mario, any
enemies, and all the dynamic blocks are game objects
that the game both updates and draws.
Developers sometimes use the term static object for

game objects that draw but don’t update. These objects
are visible to the player but never need to update. An
example of this is a building in the background of a level.
In most games, a building doesn’t move or attack the
player but is visible onscreen.
A camera is an example of a game object that updates but

doesn’t draw to the screen. Another example is a
trigger, which causes something to occur based on
another object’s intersection. For instance, a horror
game might want to have zombies appear when the
player approaches a door. In this case, the level designer
would place a trigger object that can detect when the
player is near and trigger the action to spawn the
zombie. One way to implement a trigger is as an invisible
box that updates every frame to check for intersection
with the player.
Game Object Models

There are numerous game object models, or ways
to represent game objects. This section discusses
some types of game object models and the trade-offs
between these approaches.
Game Objects as a Class Hierarchy

One game object model approach is to declare game
objects in a standard object-oriented class hierarchy,
which is sometimes called a monolithic class
hierarchy because all game objects inherit from one
base class.
To use this object model, you first need a base class:
class Actor
public:
// Called every frame to update the Actor
virtual void Update(float deltaTime);
// Called every frame to draw the Actor
virtual void Draw();
};
Then, different characters have different subclasses:
class PacMan : public Actor
public:
void Update(float deltaTime) override;
void Draw() override;
};
Similarly, you could declare other subclasses of Actor.

For example, there may be a Ghost class that inherits
from Actor, and then each individual ghost could have
its own class that inherits from Ghost. Figure 2.1
illustrates this style of game object class hierarchy.
Figure 2.1 Partial class hierarchy for Pac-Man
A disadvantage of this approach is that it means that

every game object must have all the properties and
functions of the base game object (in this case, Actor).
For example, this assumes that every Actor can update
and draw. But as discussed, there may be objects that
aren’t visible, and thus calling Draw on these objects is a
waste of time.
The problem becomes more apparent as the functionality

of the game increases. Suppose many of the actors in the
game—but not all of them—need to move. In the case of
Pac-Man, the ghosts and Pac-Man need to move, but the
pellets do not. One approach is to place the movement
code inside Actor, but not every subclass will need this
code. Alternatively, you could extend the hierarchy with
a new MovingActor that exists between Actor and any
subclasses that need movement. However, this adds
more complexity to the class hierarchy.
Furthermore, having one big class hierarchy can cause

difficulties when two sibling classes later need to have
features shared between them. For instance, a game in
the vein of Grand Theft Auto might have a base
Vehicle class. From this class, it might make sense to
create two subclasses: LandVehicle (for vehicles that
traverse land) and WaterVehicle (for water-based
vehicles like boats).
But what happens if one day a designer decides to add an

amphibious vehicle? It may be tempting to create a new
subclass called AmphibiousVehicle that inherits from
both LandVehicle and WaterVehicle. However, this
requires use of multiple inheritance and, furthermore,
means that AmphibiousVehicle inherits from
Vehicle along two different paths. This type of
hierarchy, called diamond inheritance, can cause
issues because the subclass might inherit multiple
versions of a virtual function. For this reason, it’s
recommended that we avoid diamond hierarchies.
Game Objects with Components

Instead of using a monolithic hierarchy, many games
instead use a component-based game object
model. This model has become increasingly popular,
especially because the Unity game engine uses it. In
this approach, there is a game object class, but there
are no subclasses of the game object. Instead, the
game object class has-a collection of component
objects that implement needed functionality.
For example, in the monolithic hierarchy we looked at

earlier, Pinky is a subclass of Ghost, which is a subclass
of Actor. However, in a component-based model, Pinky
is a GameObject instance containing four components:
PinkyBehavior, CollisionComponent,
TransformComponent, and DrawComponent. Figure
2.2 shows this relationship.
Figure 2.2 The components that make up the ghost
Pinky
Each of these components has the specific properties and

functionality needed for that component. For example,
DrawComponent handles drawing the object to the
screen, and TransformComponent stores the position
and transformation of an object in the game world.
One way to implement a component object model is with

a class hierarchy for components. This class hierarchy
generally has a very shallow depth. Given a base
Component class, GameObject then simply has a
collection of components:
class GameObject
public:
void AddComponent(Component* comp);
void RemoveComponent(Component* comp);
private:
std::unordered_set<Component*> mComponents;
};
Notice that GameObject contains only functions for

adding and removing components. This makes systems
that track different types of components necessary. For
example, every DrawComponent might register with a
Renderer object so that when it is time to draw the
frame, the Renderer is aware of all the active
DrawComponents.
One advantage of the component-based game object

model is that it’s easier to add functionality only to the
specific game objects that require it. Any object that
needs to draw can have a DrawComponent, but objects
that don’t simply don’t have one.
However, a disadvantage of pure component systems is

that dependencies between components in the same
game object are not clear. For instance, it’s likely that the
DrawComponent needs to know about the
TransformComponent in order know where the object
should draw. This means that the DrawComponent
likely needs to ask the owning GameObject about its
TransformComponent. Depending on the
implementation, the querying can become a noticeable
performance bottleneck.
Game Objects as a Hierarchy with Components

The game object model used in this book is a hybrid
of the monolithic hierarchy and the component
object models. This is, in part, inspired by the game
object model used in Unreal Engine 4. There is a
base Actor class with a handful of virtual functions,
but each actor also has a vector of components.
Listing 2.1 shows the declaration of the Actor class,
with some getter and setter functions omitted.
Listing 2.1 Actor Declaration

class Actor
public:
// Used to track state of actor
enum State
EActive,
EPaused,
EDead
};
// Constructor/destructor
Actor(class Game* game);
virtual ~Actor();
// Update function called from Game (not overridable)
void Update(float deltaTime);
// Updates all the components attached to the actor (not overridable)
void UpdateComponents(float deltaTime);
// Any actor-specific update code (overridable)
virtual void UpdateActor(float deltaTime);
// Getters/setters
// ...
// Add/remove components
void AddComponent(class Component* component);
void RemoveComponent(class Component* component);
private:
// Actor's state
State mState;
// Transform
Vector2 mPosition; // Center position of actor
float mScale; // Uniforms scale of actor (1.0f for 100%)
float mRotation; // Rotation angle (in radians)
// Components held by this actor
std::vector<class Component*> mComponents;
class Game* mGame;
};
The Actor class has several notable features. The state

enum tracks the status of the actor. For example, Update
only updates the actor when in the EActive state. The
EDead state notifies the game to remove the actor. The
Update function calls UpdateComponents first and
then UpdateActor. UpdateComponents loops over all
the components and updates each in turn. The base
implementation of UpdateActor is empty, but Actor
subclasses can write custom behavior in an overridden
UpdateActor function.
In addition, the Actor class needs access to the Game

class for several reasons, including to create additional
actors. One approach is to make the game object a
singleton, a design pattern in which there is a single,
globally accessible instance of a class. But the singleton
pattern can cause issues if it turns out there actually need
to be multiple instances of the class. Instead of using
singletons, this book uses an approach called
dependency injection. In this approach, the actor
constructor receives a pointer to the Game class. Then an
actor can use this pointer to create another actor (or
access any other required Game functions).
As in the Chapter 1 game project, a Vector2 represents

the position of an Actor. Actors also support a scale
(which makes the actor bigger or smaller) and a rotation
(to rotate the actor). Note that the rotation is in radians,
not degrees.
Listing 2.2 contains the declaration of the Component

class. The mUpdateOrder member variable is notable. It
allows certain components to update before or after
other components, which can be useful in many
situations. For instance, a camera component tracking a
player may want to update after the movement
component moves the player. To maintain this ordering,
the AddComponent function in Actor sorts the
component vector whenever adding a new component.
Finally, note that the Component class has a pointer to
the owning actor. This is so that the component can
access the transform data and any other information it
deems necessary.
Listing 2.2 Component Declaration

class Component
public:
// Constructor
// (the lower the update order, the earlier the component updates)
Component(class Actor* owner, int updateOrder = 100);
// Destructor
virtual ~Component();
// Update this component by delta time
int GetUpdateOrder() const { return mUpdateOrder; }
protected:
// Owning actor
class Actor* mOwner;
// Update order of component
int mUpdateOrder;
};
Currently, the implementations of Actor and

Component do not account for player input devices, and
this chapter’s game project simply uses special case code
for input. Chapter 3, “Vector and Basic Physics,” revisits
how to incorporate input into the hybrid game object
model.
This hybrid approach does a better job of avoiding the

deep class hierarchies in the monolithic object model,
but certainly the depth of the hierarchy is greater than in
a pure component-based model. The hybrid approach
also generally avoids, though does not eliminate, the
issues of communication overhead between components.
This is because every actor has critical properties such as
the transform data.
Other Approaches
There are many other approaches to game object
models. Some use interface classes to declare the
different possible functionalities, and each game
object then implements the interfaces necessary to
represent it. Other approaches extend the
component model a step further and eliminate the
containing game object entirely. Instead, these
approaches use a component database that tracks
objects with a numeric identifier. Still other
approaches define objects by their properties. In
these systems, adding a health property to an object
means that it can take and receive damage.
With any game object model, each approach has

advantages and disadvantages. However, this book sticks
to the hybrid approach because it’s a good compromise
and works relatively well for games of a certain
complexity level.
Integrating Game Objects into the Game
Loop
Integrating the hybrid game object model into the
game loop requires some code, but it ultimately isn’t
that complex. First, add two std::vector of
Actor pointers to the Game class: one containing the
active actors (mActors), and one containing pending
actors (mPendingActors). You need the pending
actors vector to handle the case where, while
updating the actors (and thus looping over
mActors), you decide to create a new actor. In this
case, you cannot add an element to mActors
because you’re iterating over it. Instead, you need to
add to the mPendingActors vector and then move
these actors into mActors after you’re done
iterating.
Next, create two functions, AddActor and

RemoveActor, which take in Actor pointers. The
AddActor function adds the actor to either
mPendingActors or mActors, depending on whether
you are currently updating all mActors (denoted by the
mUpdatingActors bool):
void Game::AddActor(Actor* actor)
// If updating actors, need to add to pending

if (mUpdatingActors)
mPendingActors.emplace_back(actor);
else
mActors.emplace_back(actor);
Similarly, RemoveActor removes the actor from

whichever of the two vectors it is in.
You then need to change the UpdateGame function so

that it updates all the actors, as shown in Listing 2.3.
After computing delta time, as discussed in Chapter 1,
you then loop over every actor in mActors and call
Update on each. Next, you move any pending actors into
the main mActors vector. Finally, you see if any actors
are dead, in which case you delete them.
Listing 2.3 Game::UpdateGame Updating Actors

{
// Compute delta time (as in Chapter 1)
float deltaTime = /* ... */;
// Update all actors
mUpdatingActors = true;
for (auto actor : mActors)
actor->Update(deltaTime);
mUpdatingActors = false;
// Move any pending actors to mActors
for (auto pending : mPendingActors)
mActors.emplace_back(pending);
mPendingActors.clear();
// Add any dead actors to a temp vector

std::vector<Actor*> deadActors;
if (actor->GetState() == Actor::EDead)
deadActors.emplace_back(actor);
// Delete dead actors (which removes them from mActors)
for (auto actor : deadActors)
delete actor;
Adding and removing actors from the game’s mActors

vector also adds some complexity to the code. This
chapter’s game project has the Actor object
automatically add or remove itself from the game in its
constructor and destructor, respectively. However, this
means that code that loops over the mActors vector and
deletes the actors (such as in Game::Shutdown) must
be written carefully:
// Because ~Actor calls RemoveActor, use a different style loop
while (!mActors.empty())
delete mActors.back();
SPRITES
A sprite is a visual object in a 2D game, typically
used to represent characters, backgrounds, and other
dynamic objects. Most 2D games have dozens if not
hundreds of sprites, and for mobile games, the sprite
data accounts for much of the overall download size
of the game. Because of the prevalence of sprites in
2D games, it is important to use them as efficiently
as possible.
Each sprite has one or more image files associated with

it. There are many different image file formats, and
games use different formats based on platform and other
constraints. For example, PNG is a compressed image
format, so these files take up less space on disk. But
hardware cannot natively draw PNG files, so they take
longer to load. Some platforms recommend using
graphics hardware–friendly formats such as PVR (for
iOS) and DXT (for PC and Xbox). This book sticks with
the PNG file format because image editing programs
universally support PNGs.
Loading Image Files

For games that only need SDL’s 2D graphics, the
simplest way to load image files is to use the SDL
Image library. The first step is to initialize SDL
Image by using IMG_Init, which takes in a flag
parameter for the requested file formats. To support
PNG files, add the following call to
Game::Initialize:
IMG_Init(IMG_INIT_PNG)
Table 2.1 lists the supported file formats. Note that SDL
already supports the BMP file format without SDL
Image, which is why there is no IMG_INIT_BMP flag in
this table.
Table 2.1 SDL Image File Formats
Flag Format
IMG_INIT_JPG JPEG
IMG_INIT_PNG PNG
IMG_INIT_TIF TIFF
Once SDL Image is initialized, you can use IMG_Load to
load an image file into an SDL_Surface:
// Loads an image from a file
// Returns a pointer to an SDL_Surface if successful, otherwise nullptr
SDL_Surface* IMG_Load(
const char* file // Image file name
);
Next, SDL_CreateTextureFromSurface converts an

SDL_Surface into an SDL_Texture (which is what
SDL requires for drawing):
// Converts an SDL_Surface to an SDL_Texture
// Returns a pointer to an SDL_Texture if successful, otherwise nullptr
SDL_Texture* SDL_CreateTextureFromSurface(
SDL_Renderer* renderer, // Renderer used
SDL_Surface* surface // Surface to convert
);
The following function encapsulates this image-loading

process:
SDL_Texture* LoadTexture(const char* fileName)
// Load from file
SDL_Surface* surf = IMG_Load(fileName);
if (!surf)
SDL_Log("Failed to load texture file %s", fileName);
return nullptr;
// Create texture from surface
SDL_Texture* text = SDL_CreateTextureFromSurface(mRenderer, surf);
SDL_FreeSurface(surf);
if (!text)
SDL_Log("Failed to convert surface to texture for %s", fileName);
return nullptr;
return text;
}
An interesting question is where to store the loaded
textures. It’s very common for a game to use the same
image file multiple times for multiple different actors. If
there are 20 asteroids, and each asteroid uses the same
image file, it doesn’t make sense to load the file from disk
20 times.
A simple approach is to create a map of filenames to

SDL_Texture pointers in Game. You then create a
GetTexture function in Game that takes in the name of
a texture and returns its corresponding SDL_Texture
pointer. This function first checks to see if the texture
already exists in the map. If it does, you can simply
return that texture pointer. Otherwise, you run the code
that loads the texture from its file.
note
While a map of filenames to SDL_Texture pointers makes sense in a
simple case, consider that a game has many different types of assets—
textures, sound effects, 3D models, fonts, and so on. Therefore, writing a
more robust system to generically handle all types of assets makes sense.
But in the interest of simplicity, this book does not implement such an asset
management system.
To help split up responsibilities, you also create a

LoadData function in Game. This function is responsible
for creating all the actors in the game world. For now
these actors are hard-coded, but Chapter 14, “Level Files
and Binary Data,” adds support for loading the actors
from a level file. You call the LoadData function at the
end of Game::Initialize.
Drawing Sprites
Suppose a game has a basic 2D scene with a
background image and a character. A simple way to
draw this scene is by first drawing the background
image and then the character. This is like how a
painter would paint the scene, and hence this
approach is known as the painter’s algorithm. In
the painter’s algorithm, the game draws the sprites in
back-to-front order. Figure 2.3 demonstrates the
painter’s algorithm, first drawing the background
star field, then the moon, then any asteroids, and
finally the ship.
Figure 2.3 The painter’s algorithm applied to a

space scene
Because this book uses a game object model with

components, it makes a great deal of sense to create a
SpriteComponent class. Listing 2.4 provides the
declaration of SpriteComponent.
Listing 2.4 SpriteComponent Declaration

class SpriteComponent : public Component
public:
// (Lower draw order corresponds with further back)
SpriteComponent(class Actor* owner, int drawOrder = 100);
~SpriteComponent();
virtual void Draw(SDL_Renderer* renderer);
virtual void SetTexture(SDL_Texture* texture);
int GetDrawOrder() const { return mDrawOrder; }
int GetTexHeight() const { return mTexHeight; }
int GetTexWidth() const { return mTexWidth; }
protected:
// Texture to draw
SDL_Texture* mTexture;
// Draw order used for painter's algorithm
int mDrawOrder;
// Width/height of texture
int mTexWidth;
int mTexHeight;
};
The game implements the painter’s algorithm by

drawing sprite components in the order specified by the
mDrawOrder member variable. The SpriteComponent
constructor adds itself to a vector of sprite components
in the Game class via the Game::AddSprite function.
In Game::AddSprite, you need to ensure that

mSprites stays sorted by draw order. Because every call
to AddSprite preserves the sorted order, you can
implement this as an insertion into an already-sorted
vector:
void Game::AddSprite(SpriteComponent* sprite)
// Find the insertion point in the sorted vector
// (The first element with a higher draw order than me)
int myDrawOrder = sprite->GetDrawOrder();
auto iter = mSprites.begin();
for ( ;
iter != mSprites.end();
++iter)
if (myDrawOrder < (*iter)->GetDrawOrder())
break;
// Inserts element before position of iterator
mSprites.insert(iter, sprite);
Because this orders the sprite components by

mDrawOrder, Game::GenerateOutput can just loop
over the vector of sprite components and call Draw on
each. You put this code in between the code that clears
the back buffer and swaps the back buffer and front
buffer, replacing the code in the Chapter 1 game that
drew the wall, ball, and paddle rectangles.
As discussed in Chapter 6, “3D Graphics,” 3D games can

also use the painter’s algorithm, though there are some
drawbacks in doing so. But for 2D scenes, the painter’s
algorithm works very well.
The SetTexture function both sets the mTexture
member variable and uses SDL_QueryTexture to get
the width and height of the texture:
void SpriteComponent::SetTexture(SDL_Texture* texture)
mTexture = texture;
// Get width/height of texture
SDL_QueryTexture(texture, nullptr, nullptr,
&mTexWidth, &mTexHeight);
To draw textures, there are two different texture drawing

functions in SDL. The simpler function is
SDL_RenderCopy:
// Renders a texture to the rendering target
// Returns 0 on success, negative value on failure
int SDL_RenderCopy(
SDL_Renderer* renderer, // Render target to draw to
SDL_Texture* texture, // Texture to draw
const SDL_Rect* srcrect, // Part of texture to draw (null if whole)

const SDL_Rect* dstrect, // Rectangle to draw onto the target
);
However, for more advanced behavior (such as rotating

sprites), you can use SDL_RenderCopyEx:
// Renders a texture to the rendering target
// Returns 0 on success, negative value on failure
int SDL_RenderCopyEx(
SDL_Renderer* renderer, // Render target to draw to
SDL_Texture* texture, // Texture to draw
const SDL_Rect* srcrect, // Part of texture to draw (null if whole)
const SDL_Rect* dstrect, // Rectangle to draw onto the target
double angle, // Rotation angle (in degrees, clockwise)
const SDL_Point* center, // Point to rotate about (nullptr for center)
SDL_RenderFlip flip, // How to flip texture (usually SDL_FLIP_NONE)
);
Because actors have a rotation, and you want your

sprites to inherit this rotation, you must use
SDL_RenderCopyEx. This introduces a few
complexities to the SpriteComponent::Draw
function. First, the SDL_Rect struct’s x/y coordinates
correspond to the top-left corner of the destination.
However, the actor’s position variable specifies the
center position of the actor. So, as with the ball and
paddle in Chapter 1, you must compute the coordinates
for the top-left corner.
Second, SDL expects an angle in degrees, but Actor uses

an angle in radians. Luckily, this book’s custom math
library in the Math.h header file includes a
Math::ToDegrees function that can handle the
conversion. Finally, in SDL a positive angle is clockwise,
but this is the opposite of the unit circle (where positive
angles are counterclockwise). Thus, negate the angle to
maintain the unit circle behavior. Listing 2.5 shows the
SpriteComponent::Draw function.
Listing 2.5 SpriteComponent::Draw

Implementation
void SpriteComponent::Draw(SDL_Renderer* renderer)
if (mTexture)
SDL_Rect r;
// Scale the width/height by owner's scale

r.w = static_cast<int>(mTexWidth * mOwner->GetScale());
r.h = static_cast<int>(mTexHeight * mOwner->GetScale());
// Center the rectangle around the position of the owner
r.x = static_cast<int>(mOwner->GetPosition().x - r.w / 2);
r.y = static_cast<int>(mOwner->GetPosition().y - r.h / 2);
// Draw
SDL_RenderCopyEx(renderer,
mTexture, // Texture to draw
nullptr, // Source rectangle
&r, // Destination rectangle
-Math::ToDegrees(mOwner->GetRotation()), // (Convert angle)
nullptr, // Point of rotation
SDL_FLIP_NONE); // Flip behavior
This implementation of Draw assumes that the position

of the actor corresponds to its position onscreen. This
assumption holds only for games where the game world
exactly corresponds to the screen. This doesn’t work for a
game like Super Mario Bros. that has a game world
larger than one single screen. To handle such a case, the
code needs a camera position. Chapter 9, “Cameras,”
discusses how to implement cameras in the context of a
3D game.
Animating Sprites
Most 2D games implement sprite animation using a
technique like flipbook animation: a series of
static 2D images played in rapid succession to create
an illusion of motion. Figure 2.4 illustrates what such
a series of images for different animations for a
skeleton sprite might look like.
Figure 2.4 Series of images for a skeleton sprite
The frame rate of sprite animations can vary, but many

games choose to use 24 FPS (the traditional frame rate
used in film). This means that every second of an
animation needs 24 individual images. Some genres,
such as 2D fighting games, may use 60 FPS sprite
animations, which dramatically increases the required
number of images. Luckily, most sprite animations are
significantly shorter than 1 second in duration.
The simplest way to represent an animated sprite is with

a vector of the different images corresponding to each
frame in an animation. The AnimSpriteComponent
class, declared in Listing 2.6, uses this approach.
Listing 2.6 AnimSpriteComponent Declaration

class AnimSpriteComponent : public SpriteComponent
public:
AnimSpriteComponent(class Actor* owner, int drawOrder = 100);
// Update animation every frame (overriden from component)
// Set the textures used for animation
void SetAnimTextures(const std::vector<SDL_Texture*>& textures);
// Set/get the animation FPS
float GetAnimFPS() const { return mAnimFPS; }
void SetAnimFPS(float fps) { mAnimFPS = fps; }
private:
// All textures in the animation
std::vector<SDL_Texture*> mAnimTextures;
// Current frame displayed
float mCurrFrame;
// Animation frame rate
float mAnimFPS;
};
The mAnimFPS variable allows different animated sprites

to run at different frame rates. It also allows the
animation to dynamically speed up or slow down. For
instance, as a character gains speed, you could increase
the frame rate of the animation to further add to the
illusion of speed. The mCurrFrame variable tracks the
current frame displayed as a float, which allows you to
also keep track of how long that frame has displayed.
The SetAnimTextures function simply sets the

mAnimTextures member variable to the provided
vector and resets mCurrFrame to zero. It also calls the
SetTexture function (inherited from
SpriteComponent) and passes in the first frame of the
animation. Since this code uses the SetTexture
function from SpriteComponent, it’s unnecessary to
override the inherited Draw function.
The Update function, shown in Listing 2.7, is where

most of the heavy lifting of AnimSpriteComponent
occurs. First, update mCurrFrame as a function of the
animation FPS and delta time. Next, you make sure that
mCurrFrame remains less than the number of textures
(which means you need to wrap around back to the start
of the animation if needed). Finally, cast mCurrFrame to
an int, grab the correct texture from mAnimTextures,
and call SetTexture.
Listing 2.7 AnimSpriteComponent::Update

Implementation
void AnimSpriteComponent::Update(float deltaTime)
SpriteComponent::Update(deltaTime);
if (mAnimTextures.size() > 0)
// Update the current frame based on frame rate
// and delta time
mCurrFrame += mAnimFPS * deltaTime;
// Wrap current frame if needed
while (mCurrFrame >= mAnimTextures.size())

{
mCurrFrame -= mAnimTextures.size();
// Set the current texture
SetTexture(mAnimTextures[static_cast<int>(mCurrFrame)]);
One feature missing from AnimSpriteComponent is

better support for switching between animations.
Currently, the only way to switch an animation is to call
SetAnimTextures repeatedly. It makes more sense to
have a vector of all the different textures for all of a
sprite’s animations and then specify which images
correspond to which animation. You’ll explore this idea
further in Exercise 2.2.
SCROLLING BACKGROUNDS
A trick often used in 2D games is having a
background that scrolls by. This creates an
impression of a larger world, and infinite scrolling
games often use this technique. For now, we are
focusing on scrolling backgrounds, as opposed to
scrolling through an actual level. The easiest method
is to split the background into screen-sized image
segments, which are repositioned every frame to
create the illusion of scrolling.
As with animated sprites, it makes sense to create a

subclass of SpriteComponent for backgrounds. Listing
2.8 shows the declaration of BGSpriteComponent.
Listing 2.8 BGSpriteComponent Declaration

class BGSpriteComponent : public SpriteComponent
public:
// Set draw order to default to lower (so it's in the background)
BGSpriteComponent(class Actor* owner, int drawOrder = 10);
// Update/draw overriden from parent
void Draw(SDL_Renderer* renderer) override;
// Set the textures used for the background
void SetBGTextures(const std::vector<SDL_Texture*>& textures);
// Get/set screen size and scroll speed
void SetScreenSize(const Vector2& size) { mScreenSize = size; }
void SetScrollSpeed(float speed) { mScrollSpeed = speed; }

float GetScrollSpeed() const { return mScrollSpeed; }
private:
// Struct to encapsulate each BG image and its offset
struct BGTexture
SDL_Texture* mTexture;
Vector2 mOffset;
};
std::vector<BGTexture> mBGTextures;
Vector2 mScreenSize;
float mScrollSpeed;
};
The BGTexture struct associates each background

texture with its corresponding offset. The offsets update
every frame to create the scrolling effect. You need to
initialize the offsets in SetBGTextures, positioning
each background to the right of the previous one:
void BGSpriteComponent::SetBGTextures(const std::vector<SDL_Texture*>& texture
int count = 0;
for (auto tex : textures)
BGTexture temp;
temp.mTexture = tex;
// Each texture is screen width in offset
temp.mOffset.x = count * mScreenSize.x;
temp.mOffset.y = 0;
mBGTextures.emplace_back(temp);
count++;
This code assumes that each background image has a

width corresponding to the screen width, but it’s
certainly possible to modify the code to account for
variable sizes. The Update code updates the offsets of
each background, taking to account when one image
moves all the way off the screen. This allows the images
to infinitely repeat:
void BGSpriteComponent::Update(float deltaTime)
{
SpriteComponent::Update(deltaTime);
for (auto& bg : mBGTextures)
// Update the x offset
bg.mOffset.x += mScrollSpeed * deltaTime;
// If this is completely off the screen, reset offset to
// the right of the last bg texture
if (bg.mOffset.x < -mScreenSize.x)
bg.mOffset.x = (mBGTextures.size() - 1) * mScreenSize.x - 1;
The Draw function simply draws each background

texture using SDL_RenderCopy, making sure to adjust
the position based on the owner’s position and the offset
of that background. This achieves the simple scrolling
behavior.
Some games also implement parallax scrolling. In this

approach, you use multiple layers for the background.
Each layer scrolls at different speeds, which gives an
illusion of depth. For example, a game might have a
cloud layer and a ground layer. If the cloud layer scrolls
more slowly than the ground layer, it gives the
impression that the clouds are farther away than the
ground. Traditional animation has used this technique
for nearly a century, and it is effective. Typically, only
three layers are necessary to create a believable parallax
effect, as illustrated in Figure 2.5. Of course, more layers
add more depth to the effect.
Figure 2.5 Space scene broken into three layers to

facilitate parallax scrolling
To implement the parallax effect, attach multiple

BGSpriteComponents to a single actor and specify
different draw order values. Then you can use a different
scroll speed for each background to complete the effect.
GAME PROJECT
Unfortunately, you have not learned about enough
new topics to make a game with noticeably more
complex mechanics than the Pong clone created in
Chapter 1, “Game Programming Overview.” And it
wouldn’t be particularly interesting to just add
sprites to the previous chapter’s game. So in lieu of a
complete game, this chapter’s game project
demonstrates the new techniques covered in this
chapter. The code is available in the book’s GitHub
repository, in the Chapter02 directory. Open
Chapter02-windows.sln on Windows and
Chapter02-mac.xcodeproj on Mac. Figure 2.6
shows the game project in action. Jacob Zinman-
Jeanes created the sprite images, which are licensed
under the CC BY license.
Figure 2.6 Side-scroller project in action
The code includes an implementation of the hybrid

Actor/Component model, SpriteComponent,
AnimSpriteComponent, and parallax scrolling. It also
includes a subclass of Actor called Ship. The Ship
class contains two speed variables to control the
left/right speed and the up/down speed, respectively.
Listing 2.9 shows the declaration of Ship.
Listing 2.9 Ship Declaration

class Ship : public Actor
public:
Ship(class Game* game);
void UpdateActor(float deltaTime) override;
void ProcessKeyboard(const uint8_t* state);
float GetRightSpeed() const { return mRightSpeed; }
float GetDownSpeed() const { return mDownSpeed; }
private:
float mRightSpeed;
float mDownSpeed;
};
The Ship constructor initializes mRightSpeed and

mDownSpeed to 0, and also creates an
AnimSpriteComponent attached to the ship, with the
associated textures:
AnimSpriteComponent* asc = new AnimSpriteComponent(this);
std::vector<SDL_Texture*> anims = {
game->GetTexture("Assets/Ship01.png"),
};
asc->SetAnimTextures(anims);
The keyboard input directly affects the speed of the ship.

The game uses the W and S keys to move the ship up and
down and the A and D keys to move the ship left and
right. The ProcessKeyboard function takes in these
inputs and updates mRightSpeed and mDownSpeed as
appropriate.
The Ship::UpdateActor function implements the

ship’s movement, using techniques similar to those
shown in Chapter 1:
void Ship::UpdateActor(float deltaTime)
Actor::UpdateActor(deltaTime);
// Update position based on speeds and delta time
Vector2 pos = GetPosition();
pos.x += mRightSpeed * deltaTime;
pos.y += mDownSpeed * deltaTime;
// Restrict position to left half of screen
// ...
SetPosition(pos);
}
Movement is such a common feature for games that it

makes a lot of sense to implement it as a component, as
opposed to in the UpdateActor function. Chapter 3,
“Vectors and Basics Physics,” discusses how to create a
MoveComponent class.
The background is a generic Actor (not a subclass) that

has two BGSpriteComponents. The different scroll
speeds of these two backgrounds create a parallax effect.
All these actors, including the ship, are created in the
Game::LoadData function.
SUMMARY
There are many ways to represent game objects. The
simplest approach is to use a monolithic hierarchy
with one base class that every game object inherits
from, but this can quickly grow out of hand. With a
component-based model, you can instead define the
functionality of a game object in terms of the
components it contains. This book uses a hybrid
approach that has a shallow hierarchy of game
objects but components that implement some
behaviors, such as drawing and movement.
The very first games used 2D graphics. Although many

games today are in 3D, 2D games still are very popular.
Sprites are the primary visual building block of any 2D
game, whether animated or not. SDL supports loading
and drawing textures via a simple interface.
Many 2D games implement flipbook animation, drawing

different images in rapid succession to make the sprite
appear to animate. You can use other techniques to
implement a scrolling background layer, and you can use
the parallax effect to create the illusion of depth.
ADDITIONAL READING
Jason Gregory dedicates several pages to different
types of game object models, including the model
used at Naughty Dog. Michael Dickheiser’s book
contains an article on implementing a pure
component model.
Dickheiser, Michael, Ed. Game

Programming Gems 6. Boston: Charles
River Media, 2006.
EXERCISES
The first exercise of this chapter is a thought
experiment on the different types of game object
models. In the second exercise you add functionality
to the AnimSpriteComponent class. The final
exercise involves adding support for tile maps, a
technique for generating 2D scenes from tiles.
Exercise 2.1
Consider an animal safari game where the player can
drive around in different vehicles to observe animals
in the wild. Think about the different types of
creatures, plants, and vehicles that might exist in
such a game. How might you implement these
objects in a monolithic class hierarchy object model?
Now consider the same game but implemented with a

component game object model. How might you
implement this? Which of these two approaches seems
more beneficial for this game?
Exercise 2.2
The AnimSpriteComponent class currently
supports only a single animation, composed of all the
sprites in the vector. Modify the class to support
several different animations. Define each animation
as a range of textures in the vector. Use the
CharacterXX.png files in the
Chapter02/Assets directory for testing.
Now add support for non-looping animations. When

defining an animation as a range of textures, also allow
specification of looping or non-looping. When a non-
looping animation finishes, it should not wrap back to
the initial texture.
Exercise 2.3
One approach to generate a 2D scene is via a tile
map. In this approach, an image file (called a tile
set) contains a series of uniformly sized tiles. Many
of these tiles combine to form a 2D scene. Tiled
(http://www.mapeditor.org), which is a great
program for generating tile sets and tile maps,
generated the tile maps for this exercise. Figure 2.7
illustrates what a portion of the tile set looks like.
Figure 2.7 A portion of the tile set used in Exercise

2.3
In this case, the tile maps are in CSV files. Use the
MapLayerX.csv files in Chapter02/Assets, where
there are three different layers (Layer 1 being the closest
and Layer 3 the farthest). Tiles.png contains the tile
set. Each row in the CSV file contains a series of
numbers, like this:
-1,0,5,5,5,5
-1 means there is no image for that tile (so you should
render nothing for that tile). Every other number
references a specific tile from the tile set. The numbering
goes left to right and then up to down. So, in this tile set,
tile 8 is the leftmost tile on the second row.
Create a new component called TileMapComponent

that inherits from SpriteComponent. This class needs
a function to load and read in a tile map CSV file. Then
override Draw to draw each tile from the tile set texture.
In order to draw only part of a texture, instead of the
entire texture, use the srcrect parameter of
SDL_RenderCopyEx. This can then draw only a single
tile square from the tile set texture rather than the entire
tile set.
CHAPTER 3
VECTORS AND BASIC

PHYSICS
Vectors are a fundamental mathematical

concept that game programmers use every
day. This chapter first explores all the
different ways vectors can be used to solve
problems in games. Next, it shows how to
implement basic movement through a
MoveComponent and control it with the
keyboard via an InputComponent. This
chapter then briefly explores the basics of
Newtonian physics before ending with a
discussion of how to detect collisions
between objects. This chapter’s game project
uses some of these techniques to implement
a version of the classic game Asteroids.
VECTORS
A mathematical vector (not to be confused with
std::vector) represents both a magnitude and
direction in an n-dimensional space, with one
component per dimension. This means that a two-
dimensional (2D) vector has x and y components. To
a game programmer, a vector is one of the most
important mathematical tools. You can use vectors to
solve many different problems in games, and
understanding vectors is especially important when
you’re working with 3D games. This section covers
both the properties of vectors and how to use them in
games.
This book uses an arrow above a variable name to denote

that the variable is a vector. This book also denotes each
component of a vector with a subscript for each
dimension. For example, this is the notation for the 2D
vector :
A vector has no concept of a position. This seems

confusing, given that Chapters 1, “Game Programming
Overview,” and 2, “Game Objects and 2D Graphics,” use
a Vector2 variable to represent positions. (You’ll learn
why this is the case in a moment.)
If vectors have no position, this means that two vectors

are equivalent if they have the same magnitude (or
length) and point in the same direction. Figure 3.1
shows many vectors in a vector field. Even though the
diagram shows many vectors drawn in different
locations, because all the vectors have the same
magnitude and direction, they are equivalent.
Figure 3.1 A vector field in which all vectors are

equivalent
Even though where you draw a vector doesn’t change its

value, it often simplifies solving vector problems to draw
the vector such that its start, or tail, is at the origin. You
can then think of the arrow part of the vector (the head)
as “pointing at” a specific point in space. When drawn in
this manner, the position that the vector “points at” has
the same components as the vector. For example, if you
draw the 2D vector <1, 2> such that it starts at the origin,
its head points at (1, 2), as shown in Figure 3.2.
Figure 3.2 The 2D vector <1, 2> drawn with its tail
at the origin and its head “pointing at” (1, 2)
Because a vector can represent a direction, you often use

vectors in games to describe the orientation of an object.
The forward vector of an object is the vector that
represents the “straight ahead” direction for the object.
For example, an object facing straight down the x-axis
has a forward vector of <1, 0>.
You can compute many different vector operations. In

general, game programmers use a library to perform all
these different computations. Because of this, it’s better
to know which vector computations solve which
problems than to simply memorize the equations. The
remainder of this section explores some of the basic
vector use cases.
This book’s source code uses a custom-written vector

library in the provided Math.h header file, included in
the code for each game project from this chapter onward.
The Math.h header file declares Vector2 and Vector3
classes, as well as implementations of many operators
and member functions. Note that the x and y
components are public variables, so you can write code
like this:
Vector2 myVector;
myVector.x = 5;
myVector.y = 10;
Although the diagrams and examples in this section

almost universally use 2D vectors, almost every
operation outlined also works for 3D vectors; there is
simply one more component in 3D.
Getting a Vector between Two Points:

Subtraction
With vector subtraction, you subtract each
component of one vector from the corresponding
component of the other vector, which yields a new
vector. For example, in 2D you subtract the x
components of the vector separately from the y
components:
To visualize the subtraction of two vectors, draw the

vectors such that their tails start from the same position,
as in Figure 3.3(a). Then construct a vector from the
head of one vector to the head of the other. Because
subtraction isn’t commutative (that is, a − b is not the
same as b − a), the order is significant. A mnemonic that
helps remember the correct order is that a vector from
to is .
Figure 3.3 Vector subtraction (a) and subtraction

between two points represented as vectors (b)
You can use subtraction to construct a vector between

two points. For example, suppose that a space game
allows a ship to fire a laser at a target. You can represent
the position of the ship with point s and the position of
the target with point t. Suppose that s = (5,2) and t =
(3,5).
What if you instead thought of these points as the vectors

and , drawn with their tails at the origin and their
heads “pointing at” their respective points? As previously
discussed, the value of the x and y components of these
vectors are identical to the points. However, if they are
vectors, you can construct a vector between the two by
using subtraction, as in Figure 3.8(b). Because the laser
should point from the ship to the target, this is the
subtraction:
In the provided Math.h library, the – operator subtracts

two vectors:
Vector2 a, b;
Vector2 result = a - b;
Scaling a Vector: Scalar Multiplication

You can multiply a vector by a scalar (a single value).
To do so, simply multiply each component of the
vector by the scalar:
Multiplying a vector by a positive scalar only changes the

magnitude of the vector, while multiplying by a negative
scalar also inverts the direction of the vector (meaning
the head becomes the tail and vice versa). Figure 3.4
illustrates the result of multiplying the vector by two
different scalars.
Figure 3.4 Scalar multiplication
In the provided Math.h library, the * operator performs

scalar multiplication:
Vector2 a;
Vector2 result = 5.0f * a; // Scaled by 5
Combining Two Vectors: Addition

With vector addition, you add the components of two
vectors, which yields a new vector:
To visualize addition, draw the vectors such that the

head of one vector touches the tail of the other vector.
The result of the addition is the vector from the tail of
one vector to the head of the other, as shown in Figure
3.5.
Figure 3.5 Vector addition
Note that the order of the addition\ doesn’t change the

result. This is because vector addition is commutative,
just like addition between two real numbers:
You can use vector addition in a variety of ways. For

example, suppose the player is at point p, and the
player’s forward vector is . The point that is 150 units
“in front” of the player is then .
In the provided Math.h library, the + operator adds two

vectors:
Vector2 a, b;
Vector2 result = a + b;
Determining a Distance: Length
As mentioned earlier in this chapter, a vector
represents both a magnitude and direction. You use
two vertical bars on either side of a vector variable to
denote computing the magnitude (or length) of the
vector. For example, you write the magnitude of as
. To calculate the length of a vector, take the
square root of the sum of the squares of each
component:
This may seem very similar to the Euclidean distance

formula, and that’s because it is just a simplification of
the Euclidean distance formula! If you draw the vector
such that it starts at the origin, the formula calculates the
distance from the origin to the point the vector “points
at.”
You can use the magnitude to compute the distance

between two arbitrary points. Given points p and q, treat
them as vectors and perform vector subtraction. The
magnitude of the result of the subtraction is equal to the
distance between the two points:
The square root in this length formula is a relatively

expensive calculation. If you absolutely must know the
length, there’s no way to avoid this square root.
However, in some cases, it may seem that you need to
know the length, but you might actually be able to avoid
the square root.
For example, suppose you want to determine whether

the player is closer to object A or object B. You first
construct a vector from object A to the player, or .
Similarly, you construct a vector from object B to the
player, or . It might seem natural to calculate the
length of each vector and compare the two to figure out
which object is closer. However, you can simplify the
math somewhat. Assuming that you have no imaginary
numbers, the length of a vector must be positive. In this
case, comparing the length of these two vectors is
logically equivalent to comparing the length squared
(the square of the length) of each vector or, in other
words:
So for cases in which you merely need a relative

comparison, you can use the length squared instead of
the length:
In the provided Math.h library, the Length() member

function computes the length of a vector:
Vector2 a;
float length = a.Length();

Similarly, the LengthSquared() member function
computes the length squared.
Determining Directions: Unit Vectors and

Normalization
A unit vector is a vector with a length of one. The
notation for unit vectors is to write a “hat” above the
vector’s symbol, such as û. You can convert a vector
with a non-unit length into a unit vector through
normalization. To normalize a vector, divide each
component by the length of the vector:
In some cases, using unit vectors simplifies calculations.

However, normalizing a vector causes it to lose any of its
original magnitude information. You must be careful,
then, which vectors you normalize and when.
warning
DIVIDE BY ZERO: If a vector has zeros for all its components, the length of
this vector is also zero. In this case, the normalization formula has a division
by zero. For floating-point variables, dividing by zero yields the error value
NaN (not a number). Once a calculation has NaNs, it’s impossible to get rid
of them because any operation on NaN also yields NaN.
A common workaround for this is to make a “safe” normalize function that
first tests whether the length of the vector if close to zero. If it is, then you
simply don’t perform the division, thus avoiding the division by zero.
A good rule of thumb is to always normalize vectors if

you need only the direction. Some examples are the
direction an arrow points or the forward vector of an
actor. However, if the distance also matters, such as for a
radar showing distances of objects, then normalizing
would wipe away this information.
You usually normalize vectors such as the forward vector

(which way is an object facing?) and the up vector
(which way is up?). However, you may not want to
normalize other vectors. For example, normalizing a
gravity vector causes the loss of the magnitude of the
gravity.
The Math.h library provides two different

Normalize() functions. First, there’s a member
function that normalizes a given vector in place
(overwriting its unnormalized version):
Vector2 a;
a.Normalize(); // a is now normalized
There also is a static function that normalizes a vector

passed as a parameter and returns the normalized
vector:
Vector2 a;
Vector2 result = Vector2::Normalize(a);

Converting from an Angle to a Forward
Vector
Recall that the Actor class from Chapter 2 has a
rotation represented by an angle in radians. This
allows you to rotate the way an actor faces. Because
the rotations are in 2D for now, the angle directly
corresponds to an angle on the unit circle, as in
Figure 3.6.
Figure 3.6 Unit circle
The equation of the unit circle in terms of an angle, theta,

is:
You can directly use these equations to convert the
actor’s angle into a forward vector:
Vector3 Actor::GetForward() const
return Vector2(Math::Cos(mRotation), Math::Sin(mRotation));
Here, the Math::Cos and Math::Sin functions are

wrappers for the C++ standard library’s sine and cosine
functions. Notice that you don’t explicitly normalize the
vector. This is because the circle equation always returns
a unit length vector because the unit circle has a radius of
one.
Keep in mind that the unit circle has +y as up, whereas

SDL 2D graphics use +y down. So for the correct forward
vector in SDL, you must negate the y-component value:
Vector3 Actor::GetForward() const
// Negate y-component for SDL
return Vector2(Math::Cos(mRotation), -Math::Sin(mRotation));
}
Converting a Forward Vector to an Angle:
Arctangent
Now suppose you have a problem opposite the
problem described the previous section. Given a
forward vector, you want to convert it into an angle.
Recall that the tangent function takes in an angle and
returns the ratio between the opposite and adjacent
sides of a triangle.
Now imagine that you calculate a new “desired” forward

vector for an actor, but you need to convert this to the
angle for its rotation member variable. In this case, you
can form a right triangle with this new forward vector
and the x-axis, as shown in Figure 3.7. For this triangle,
the x component of the forward vector is the length of
the adjacent side of the triangle, and the y component of
the forward vector is the length of the opposite side of
the triangle. Given this ratio, you can then use the
arctangent function to calculate the angle θ.
Figure 3.7 Right triangle between the x-axis and a
vector
In programming, the preferred arctangent function is the

atan2 function, which takes two parameters (the
opposite length and the adjacent length) and returns an
angle in the range [–π,π]. A positive angle means the
triangle is in the first or second quadrant (a positive y
value), and a negative angle means the triangle is in the
third or fourth quadrant.
For example, suppose you want a ship to face toward an

asteroid. You can first construct the vector from the ship
to the asteroid and normalize this vector. Next, use
atan2 to convert the new forward vector to an angle.
Finally, set the ship actor’s rotation to this new angle.
Note that you must negate the y component to account
for +y down in SDL’s 2D coordinate system. This yields
the following code:
// (ship and asteroid are Actors)
Vector2 shipToAsteroid = asteroid->GetPosition() - ship->GetPosition();
shipToAsteroid.Normalize();
// Convert new forward to angle with atan2 (negate y-component for SDL)
float angle = Math::Atan2(-shipToAsteroid.y, shipToAsteroid.x);
ship->SetRotation(angle);
This arctangent approach works very well for 2D games.

However, it only works in this form for 2D games
because all objects remain on the x-y plane. For 3D
games, it’s often preferable to use the dot product
approach outlined in the next section.
Determining the Angle Between Two

Vectors: Dot Product
The dot product between two vectors results in a
single scalar value. One of the most common uses of
the dot product in games is to find the angle between
two vectors. The following equation calculates the
dot product between vectors and :
The dot product also has a relationship to the cosine of
an angle, and you can use the dot product to compute the
angle between two vectors:
This formulation, illustrated in Figure 3.8, is based on

the Law of Cosines. Given this formula, you can then
solve for θ:
Figure 3.8 Computing the angle between two unit

vectors
If the two vectors and are unit vectors, you can omit
the division because the length of each vector is one:
This is one reason it’s helpful to normalize vectors in
advance if only the direction matters.
For example, consider a player at position p with a

forward vector of . A new enemy appears at position e.
Suppose you need the angle between the original forward
vector and the vector from p to e. First, compute the
vector from p to e, using the vector representations of the
points:
Next, because only the direction matters in this case,

normalize :
Finally, use the dot product equations to determine the

angle between and :
Because the dot product can calculate an angle between

two vectors, a couple of special cases are important to
remember. If the dot product between two unit vectors is
0, it means that they are perpendicular to each other
because cos(π/2)= 0. Furthermore, a dot product of 1
means the two vectors are parallel and face the same
direction. Finally, a dot product of −1 means they are
antiparallel, meaning the vectors are parallel and face
in the opposite direction.
One drawback of using the dot product to calculate the
angle is that the arccosine function returns an angle in
the range [0, π]. This means that while the arccosine
gives the minimum angle of rotation between the two
vectors, it doesn’t say whether this rotation is clockwise
or counterclockwise.
As with multiplication between two real numbers, the

dot product is commutative, distributive over addition,
and associative:
Another useful tip is that the length squared calculation

is equivalent to taking the dot product of a vector with
itself:
The Math.h library defines a static Dot function for

both Vector2 and Vector3. For example, to find the
angle between origForward and newForward, you
could use:
float dotResult = Vector2::Dot(origForward, newForward);
float angle = Math::Acos(dotResult);

Calculating a Normal: Cross Product
A normal is a vector perpendicular to a surface.
Calculating the normal of a surface (such as a
triangle) is a very useful calculation in 3D games. For
example, the lighting models covered in Chapter 6,
“3D Graphics,” require calculating normal vectors.
Given two 3D vectors that are not parallel, there is a

single plane that contains both vectors. The cross
product finds a vector perpendicular to that plane, as in
Figure 3.9.
Figure 3.9 Cross product in a left-handed

coordinate system
The cross product does not work with 2D vectors.

However, to convert a 2D vector into a 3D vector, just
add a z component of 0.
The × symbol denotes a cross product between two

vectors:
Note that there’s technically a second vector
perpendicular to the plane in Figure 3.9: . This
illustrates an important property of the cross product.
It’s not commutative but instead anticommutative:
Using the left-hand rule is a quick way to figure out the

direction the cross product result faces. Take your left
hand and point the index finger down and then take
your middle finger and point it down , rotating your
wrist as needed. The natural position of your thumb
points in the direction of . Here, you use the left hand
because this book’s coordinate system is left-handed.
(You’ll learn more about coordinate systems in Chapter
5, “OpenGL.”) A right-handed coordinate system instead
uses the right-hand rule.
The numeric calculation of the cross product is as

follows:
A popular mnemonic to remember the cross product

calculation is “xyzzy.” This mnemonic helps you
remember the order of the subscripts for the x-
component of the cross product result:
Then, the y and z components are the subscripts rotated
in the order x → y → z → x, yielding the next two
components of the cross product result:
As with the dot product, there’s a special case to

consider. If the cross product returns the vector 〈0,0,0〉,
this means that and are collinear. Two collinear
vectors cannot form a plane, and therefore there’s no
normal for the cross product to return.
Because a triangle lies on a single plane, the cross

product can determine the normal of the triangle. Figure
3.10 shows triangle ABC. To calculate the normal, first
construct two vectors for edges of the triangle:
Figure 3.10 Normal to triangle ABC
Then take the cross product between these two vectors

and normalize the result. This yields the normal vector of
the triangle:
The Math.h library provides a static Cross function.

For example, the following calculates the cross product
between vectors a and b:
Vector3 c = Vector3::Cross(a, b);
BASIC MOVEMENT
Recall that Chapter 2’s game project overrides the
UpdateActor function for Ship (a subclass of
Actor) to make the ship move. However, movement
is such a common feature for a game that it makes
sense to instead encapsulate this behavior in a
component. This section first explores how to create
a MoveComponent class that can move actors
around the game world. You’ll leverage this class to
create asteroids that move around the screen. Next,
this section shows how to create a subclass of
MoveComponent called InputComponent that you
can hook up directly to keyboard inputs.
Creating a Basic MoveComponent Class

At a basic level, MoveComponent should allow
actors to move forward at a certain speed. To support
this, you first need a function to calculate the forward
vector for an actor, as implemented in the
“Converting from an Angle to a Forward Vector”
section, earlier in this chapter. Once you have the
forward vector of the actor, you can move forward
based on a speed (in units per second) and delta
time, as in this pseudocode:
position += GetForward() * forwardSpeed * deltaTime;
You can use a similar mechanism to update the rotation

(angle) of the actor, except in this case you don’t need a
forward vector. You just need angular speed (in rotations
per second) and delta time:
rotation += angularSpeed * deltaTime;
This way, the actor can both move forward and rotate
based on the respective speeds. To implement
MoveComponent as a subclass of Component, you first
declare the class as in Listing 3.1. It has separate speeds
to implement both forward and rotational movement, as
well as getter/setter functions for these speeds. It also
overrides the Update function, which will contain the
code that moves the actor. Note that the constructor of
MoveComponent specifies a default update order of 10.
Recall that the update order determines the order in
which the actor updates its components. Because the
default update order for other components is 100,
MoveComponent will update before most other
components do.
Listing 3.1 MoveComponent Declaration

class MoveComponent : public Component
public:
// Lower update order to update first
MoveComponent(class Actor* owner, int updateOrder = 10);
float GetAngularSpeed() const { return mAngularSpeed; }
float GetForwardSpeed() const { return mForwardSpeed; }
void SetAngularSpeed(float speed) { mAngularSpeed = speed; }
void SetForwardSpeed(float speed) { mForwardSpeed = speed; }
private:
// Controls rotation (radians/second)

float mAngularSpeed;
// Controls forward movement (units/second)
float mForwardSpeed;
};
The implementation of Update, shown in Listing 3.2,

simply converts the movement pseudocode into actual
code. Recall that a Component class can access its
owning actor through the mOwner member variable. You
use this mOwner pointer to then access the position,
rotation, and forward of the owning actor. Also note the
use of the Math::NearZero function here. This
function compares the absolute value of the parameter
with some small epsilon amount to determine whether
the value is “near” zero. In this specific case, you don’t
bother updating the rotation or position of the actor if
the corresponding speed is near zero.
Listing 3.2 MoveComponent::Update

Implementation
void MoveComponent::Update(float deltaTime)
if (!Math::NearZero(mAngularSpeed))
{
float rot = mOwner->GetRotation();
rot += mAngularSpeed * deltaTime;
mOwner->SetRotation(rot);
if (!Math::NearZero(mForwardSpeed))
Vector2 pos = mOwner->GetPosition();
pos += mOwner->GetForward() * mForwardSpeed * deltaTime;
mOwner->SetPosition(pos);
Because this chapter’s game project is a version of the

classic Asteroids game, you also need code for screen
wrapping. This means if an asteroid goes off the left side
of the screen, it will teleport to the right side of the
screen. (We omit this code here because it’s not
something desired for a generic MoveComponent.
However, the source code for this chapter does include
this modification for screen wrapping.)
With the basic MoveComponent, you can then declare

Asteroid as a subclass of Actor. Asteroid doesn’t
need an overloaded UpdateActor function to move.
Instead, you can simply construct a MoveComponent in
its constructor, along with a SpriteComponent to
display the asteroid image, as in Listing 3.3. The
constructor also sets the speed of the asteroid to a fixed
150 units/second (which in this case, corresponds to 150
pixels/second).
Listing 3.3 Asteroid Constructor

Asteroid::Asteroid(Game* game)
:Actor(game)
// Initialize to random position/orientation
Vector2 randPos = Random::GetVector(Vector2::Zero,
Vector2(1024.0f, 768.0f));
SetPosition(randPos);
SetRotation(Random::GetFloatRange(0.0f, Math::TwoPi));
// Create a sprite component, and set texture
SpriteComponent* sc = new SpriteComponent(this);
sc->SetTexture(game->GetTexture("Assets/Asteroid.png"));
// Create a move component, and set a forward speed
MoveComponent* mc = new MoveComponent(this);

mc->SetForwardSpeed(150.0f);
One other new item in this Asteroid constructor is the

use of the Random static functions. The implementation
of these functions isn’t particularly interesting: They
simply wrap the built-in C++ random number generators
to get a vector or float within a range of values. The
Random functions here make sure that each asteroid gets
a random position and orientation.
With this Asteroid class, you can then create several

asteroids in the Game::LoadData function by using the
following code:
const int numAsteroids = 20;
for (int i = 0; i < numAsteroids; i++)
new Asteroid(this);
This results in several moving asteroids onscreen (see

Figure 3.11).
Figure 3.11 Asteroids moving with move
components
Creating an InputComponent Class

The base MoveComponent is fantastic for objects
like the asteroids that the player doesn’t control.
However, if you want a ship that the player can
control with the keyboard, you have a dilemma. One
idea is to create a custom function for input in the
Ship class, as in the Chapter 2 game project.
However, because hooking up input to an actor or a
component is a common need, it’s sensible to
incorporate this into the game object model. In other
words, you want overridable functions in both Actor
and Component that subclasses can redefine for
input as needed.
To support this, you first add a virtual ProcessInput
function to Component, with an empty default
implementation:
virtual void ProcessInput(const uint8_t* keyState) {}
Then in Actor, you declare two functions: a non-virtual

ProcessInput and a virtual ActorInput function.
The idea here is that actor subclasses that want custom
input can override ActorInput but not ProcessInput
(like how there are separate Update and UpdateActor
functions):
// ProcessInput function called from Game (not overridable)
void ProcessInput(const uint8_t* keyState);
// Any actor-specific input code (overridable)
virtual void ActorInput(const uint8_t* keyState);
The Actor::ProcessInput function first checks if the

actor’s state is active. If it is, you first call
ProcessInput on all components and then call
ActorInput for any actor-overridable behavior:
void Actor::ProcessInput(const uint8_t* keyState)

{
if (mState == EActive)
for (auto comp : mComponents)
comp->ProcessInput(keyState);
ActorInput(keyState);
Finally, in Game::ProcessInput, you can loop over all

actors and call ProcessInput on each one:
mUpdatingActors = true;
actor->ProcessInput(keyState);
mUpdatingActors = false;
You set the mUpdatingActors bool to true before the
loop to handle an actor or component trying to create
another actor inside ProcessInput. In this case, you
must add to the mPendingActors vector instead of
mActors. This is the same technique used in Chapter 2
to ensure that you don’t modify mActors while iterating
over the vector.
With this glue code in place, you can then declare a

subclass of MoveComponent called InputComponent,
as in Listing 3.4. The main idea of InputComponent is
that you can set up specific keys to control both
forward/backward movement and rotation of the owning
actor. In addition, because the overridden
ProcessInput directly sets the forward/angular speeds
of MoveComponent, you need to specify the “maximum”
speeds to calculate the correct speed values based on the
keyboard input.
Listing 3.4 InputComponent Declaration

class InputComponent : public MoveComponent
public:
InputComponent(class Actor* owner);
void ProcessInput(const uint8_t* keyState) override;

// Getters/setters for private variables
// ...
private:
// The maximum forward/angular speeds
float mMaxForwardSpeed;
float mMaxAngularSpeed;
// Keys for forward/back movement
int mForwardKey;
int mBackKey;
// Keys for angular movement
int mClockwiseKey;
int mCounterClockwiseKey;
};
Listing 3.5 shows the implementation of

InputComponent::ProcessInput. You first set the
forward speed to zero and then determine the correct
forward speed based on the keys pressed. You then pass
this speed into the inherited SetForwardSpeed
function. Note that if the user presses both forward and
back keys or neither of the keys, the forward speed
becomes zero. You use similar code for setting the
angular speed.
Listing 3.5 InputComponent::ProcessInput

Implementation
void InputComponent::ProcessInput(const uint8_t* keyState)
// Calculate forward speed for MoveComponent
float forwardSpeed = 0.0f;
if (keyState[mForwardKey])
forwardSpeed += mMaxForwardSpeed;
if (keyState[mBackKey])
forwardSpeed -= mMaxForwardSpeed;
SetForwardSpeed(forwardSpeed);
// Calculate angular speed for MoveComponent
float angularSpeed = 0.0f;

if (keyState[mClockwiseKey])
angularSpeed += mMaxAngularSpeed;
if (keyState[mCounterClockwiseKey])
angularSpeed -= mMaxAngularSpeed;
SetAngularSpeed(angularSpeed);
With this in place, you can then add keyboard-controlled

movement to Ship by simply creating an
InputComponent instance. (We omit the code for the
Ship constructor here, but it essentially sets the various
InputComponent member variables for the keys and
maximum speed.) You also create a SpriteComponent
and assign it a texture. This yields a user-controllable
ship (see Figure 3.12).
Figure 3.12 Ship controlled with the keyboard
This is an excellent first pass of a more flexible system

for input. We explore input in much greater detail in
Chapter 8, “Input Systems.”
NEWTONIAN PHYSICS
Although the basic movement approach used so far
in this chapter works for some games, for movement
that more closely resembles the real world, you need
a physically accurate approach. Luckily, Isaac
Newton (among others) developed Newtonian
physics (or classical mechanics) to describe laws of
motion. Games commonly utilize Newtonian physics
because its laws hold if objects are not moving near
the speed of light and objects are larger than
quantum particles. Because games typically don’t
feature objects in those edge cases, Newtonian
physics works well.
There are several different aspects to Newtonian physics.

This book considers only the most basic: movement
without rotational forces, or linear mechanics. For
deeper discussions of other components of Newtonian
physics, consult either Ian Millington’s book, listed in
this chapter’s “Additional Reading” section, or any
college-level physics textbook.
Linear Mechanics Overview

The two cornerstones of linear mechanics are force
and mass. Force is an influence that can cause an
object to move. Because force has a magnitude and
direction, it’s natural to represent force with a vector.
Mass is a scalar that represents the quantity of
matter contained in an object. It’s common to
confuse mass with weight, but mass is independent
of any gravity, whereas weight is not. The greater the
mass of an object, the more difficult it is to change
the movement of that object.
If you apply sufficient force to an object, it will start

moving. Newton’s second law of motion
encapsulates this idea:
In this equation, F is force, m is mass, and a is

acceleration, or the rate at which the velocity of an
object increases. Because force equals mass times
acceleration, it’s also true that acceleration is force
divided by mass. This is the usual approach in games: An
arbitrary object in the game has a mass, and you can
apply forces to that object. From this, you can then
compute the acceleration of the object.
In a physics course, the typical symbolic representation

of linear mechanics is that position, velocity, and
acceleration are functions over time. Then, with calculus,
you can calculate the velocity function as the derivative
of the position function and the acceleration function as
the derivative of the velocity function.
However, this standard formulation in terms of symbolic

equations and derivatives is not particularly applicable in
a game. A game needs to apply a force to an object and,
from that force, determine the acceleration over time.
Once you have an object’s acceleration, you can compute
the change in the object’s velocity. Finally, given a
velocity, you can compute the change in the object’s
position. A game only needs to compute this in terms of
the discrete time step of delta time. It doesn’t need a
symbolic equation. This requires using integration—but
not symbolic integration. Instead, you must use
numeric integration, which approximates the
symbolic integral over a fixed time step. Although this
sounds very complex, luckily, you can accomplish
numeric integration with only a few lines of code.
Computing Positions with Euler

Integration
Numeric integration allows a game to update the
velocity based on the acceleration and then update
the position based on the velocity. However, to
compute the acceleration of an object, the game
needs to know the mass of the object as well as the
forces applied to the object.
There are multiple types of forces to consider. Some

forces, such as gravity, are constant and should apply on
every frame. Other forces may instead be impulses, or
forces that apply only for a single frame.
For example, when a character jumps, an impulse force

allows the player to get off the ground. However, the
character will eventually return to the ground because of
the constant force of gravity. Because multiple forces can
act on an object simultaneously, and forces are vectors,
adding up all the forces gives the total force applied to an
object for that frame. Dividing the sum of forces by mass
yields acceleration:
acceleration = sumOfForces / mass;
Next, you can use the Euler integration method of

numeric integration to compute the velocity and the
position:
// (Semi-Implicit) Euler Integration

// Update velocity
velocity += acceleration * deltaTime;
// Update position
position += velocity * deltaTime;
Note that force, acceleration, velocity, and position are

all represented as vectors in these calculations. Because
these calculation depend on the delta time, you can put
them in the Update function of a component that
simulates physics.
Issues with Variable Time Steps

For games that rely on physics simulation, variable
frame times (or time steps) can cause problems.
This is because the accuracy of numeric integration
depends on the size of the time step. The smaller the
time step, the more accurate the approximation.
If the frame rate varies from frame to frame, so does the

accuracy of the numeric integration. The accuracy
changing could affect the behavior in very noticeable
ways. Imagine playing Super Mario Bros., where the
distance Mario can jump varies depending on the frame
rate. The lower the frame rate, the further Mario can
jump. This is because the error amount in numeric
integration increases with a lower frame rate, which
results in an exaggerated jump arc. This means that the
game played on a slower machine allows Mario to jump
farther than the game played on a fast machine. Figure
3.13 illustrates an example where the actual simulated
arc diverges from the intended arc due to a large time
step.
Figure 3.13 The actual jump arc diverges from an

intended jump arc due to a large time step
For this reason, any game that uses physics to calculate

the positions of objects should not use a variable frame
rate—at least not for the physics simulation code.
Instead, you can use the frame-limiting approach, as in
Chapter 1, which works if the frame rate doesn’t go below
the target frame rate. A more complex alternative is to
divide the larger time step into multiple fixed-size
physics time steps.
BASIC COLLISION DETECTION

Collision detection is how a game determines
whether two objects in the game world touch each
other. In Chapter 1, you implemented a form of
collision detection to determine whether the ball
collided with the wall or paddles. However, for the
Asteroids game project in this chapter, you need
slightly more complex calculations to decide whether
the lasers the ship fires should collide with the
asteroids in the game world.
A key concept in collision detection is simplification of

the problem. For example, the asteroid image is circular
but not exactly a circle. While it is more accurate to test
collision against the actual contours of the asteroid, it’s
far more efficient to consider the asteroid as a circle for
the purposes of collision detection. If you similarly
simplify the laser into a circle, you then only need to
determine whether these two circles collide.
Circle-Versus-Circle Intersection
Two circles intersect with each other if and only if the
distance between their centers is less than or equal to
the sum of their radii. Figure 3.14 demonstrates this
between two circles. In the first case, the two circles
are far enough apart that they do not intersect. In
this case, the distance between their centers is
greater than the sum of the radii. However, in the
second case, where the circles do intersect, the
distance between their centers is less than the sum of
their radii.
Figure 3.14 Testing intersection between two
circles
You can perform this intersection test by first creating a

vector between the two centers and calculating the
magnitude of this vector. Then, compare this distance
against the sum of the circle’s radii:
Recall the discussion earlier in the chapter regarding

length versus length squared. In the case of circle
intersection, all you need is a comparison between the
distance and the sum of the radii. Because you know the
distances and the radii cannot be negative, you can
square both sides of the equation while preserving the
inequality:
>
note
The approach covered in this section also works for spheres because the
same principle applies.
Creating a CircleComponent Subclass
To support collision detection of actors, you can
create a CircleComponent and a method to test for
intersection between two circle components. You can
then add a CircleComponent to any actor that
needs collision.
First, declare CircleComponent as a subclass of

Component, as in Listing 3.6. The only member data
CircleComponent needs is a radius because the center
of the circle is simply the position of the owning actor.
Listing 3.6 CircleComponent Declaration

class CircleComponent : public Component
public:
CircleComponent(class Actor* owner);
void SetRadius(float radius) { mRadius = radius; }
float GetRadius() const;
const Vector2& GetCenter() const;
private:
float mRadius;
};
Next, declare a global Intersect function that takes in

two circle components by reference and returns true if
the two circles intersect with each other, as in Listing 3.7.
Note that the implementation directly mirrors the
equations from the previous section. You first calculate
the distance squared between the two centers and then
compare that to the sum of the radii squared.
Listing 3.7 CircleComponent Intersection

bool Intersect(const CircleComponent& a, const CircleComponent& b)
// Calculate distance squared
Vector2 diff = a.GetCenter() - b.GetCenter();
float distSq = diff.LengthSq();
// Calculate sum of radii squared
float radiiSq = a.GetRadius() + b.GetRadius();
radiiSq *= radiiSq;
return distSq <= radiiSq;
You can then create a CircleComponent just like any

other component. For example, the following two lines of
code adds a CircleComponent to an Asteroid object
(where mCircle is a member variable pointer to a
CircleComponent):
mCircle = new CircleComponent(this);
mCircle->SetRadius(40.0f);
Because each laser fired by the ship needs to check for

collision against all the asteroids, you can add a
std::vector of Asteroid pointers to Game. Then, in
Laser::UpdateActor, you can easily test for
intersection against each of these asteroids:
void Laser::UpdateActor(float deltaTime)
// Do you intersect with an asteroid?
for (auto ast : GetGame()->GetAsteroids())
{
if (Intersect(*mCircle, *(ast->GetCircle())))
// If this laser intersects with an asteroid,
// set ourselves and the asteroid to dead
SetState(EDead);
ast->SetState(EDead);
break;
The GetCircle function that is called on each asteroid

is simply a public function that returns the pointer to the
asteroid’s CircleComponent. Similarly, the mCircle
variable is the laser’s CircleComponent.
CircleComponent works well in the case of Asteroids

because you can approximate the collision of all objects
in the game with circles. However, circles don’t work well
for all types of objects, and certainly not in 3D. Chapter
10, “Collision Detection,” dives into the topic of collision
detection in much greater detail.
GAME PROJECT
This chapter’s game project implements a basic
version of the classic game Asteroids. The earlier
sections of this chapter cover most of the new code
used in the game project. The project implements
movement with MoveComponent and
InputComponent. The CircleComponent code
tests if the ship’s laser collides against asteroids. A
notable feature that’s missing in the game project is
that the asteroids do not collide with the ship
(though you will add that in Exercise 3.2). The game
project also does not implement Newtonian physics
(though you will add that in Exercise 3.3). The code
is available in the book’s GitHub repository, in the
Chapter03 directory. Open Chapter03-
mac.xcodeproj on Mac.
One game feature not covered earlier in the chapter is

how to create lasers when the player presses the
spacebar. Because detecting the spacebar input is unique
to Ship, it follows that you should override the
ActorInput function. However, if the player holds
down the spacebar (or presses it rapidly), you don’t want
to create so many lasers that it trivializes the game.
Instead, you want a cooldown where the ship can fire a
laser only once every half second. To implement this,
first create a float mLaserCooldown member
variable in Ship and initialize it to 0.0f. Next, in
ActorInput, check if the player presses the spacebar
and mLaserCooldown is less than or equal to zero. If
both conditions are met, you create the laser, set its
position and rotation to the ship’s (so it starts at the ship
and faces the direction the ship faces), and set
mLaserCooldown to 0.5f:
void Ship::ActorInput(const uint8_t* keyState)
if (keyState[SDL_SCANCODE_SPACE] && mLaserCooldown <= 0.0f)
// Create a laser and set its position/rotation to mine
Laser* laser = new Laser(GetGame());
laser->SetPosition(GetPosition());
laser->SetRotation(GetRotation());
// Reset laser cooldown (half second)
mLaserCooldown = 0.5f;
Then, override UpdateActor to decrement

mLaserCooldown by delta time:

void Ship::UpdateActor(float deltaTime)
mLaserCooldown -= deltaTime;
This way, mLaserCooldown keeps track of the amount

of time until the player can fire again. And because
ActorInput doesn’t create a laser if the timer hasn’t
run out, you make sure the player can’t fire more often
than desired. With lasers firing, you can then shoot and
destroy asteroids with the previously covered collision
code (see Figure 3.15).
Figure 3.15 Shooting lasers at asteroids

You can use a similar float variable in Laser to force the
laser to die (and be deleted from the game world) after
one second, even if the laser doesn’t collide with an
asteroid.
SUMMARY
A vector represents a magnitude and a direction. You
can use vectors for many different computations,
including creating a vector between two points (using
subtraction), calculating the distance between two
points (using subtraction and length), finding the
angle between two vectors (using the dot product),
and calculating a normal to a surface (with the cross
product).
For basic movement, this chapter shows how to create a

MoveComponent that allows actors to move in the
forward direction as well as rotate. For forward
movement, you multiply the forward vector of the actor
by movement speed and delta time. Adding this to the
current position of the actor yields the new position of
the actor after the time step. In this chapter you also
learned how to add support for overridable input
behavior in actors and components, and how to leverage
this to create an InputComponent that inherits from
MoveComponent.
In Newtonian physics, the acceleration of an object is

force applied to the object divided by the object’s mass.
You can use Euler integration to calculate the change in
velocity and position on every frame.
Finally, collision detection is how a game decides if two

objects touch each other. For some types of games, such
as this chapter’s game project, you can use circles to
represent the collisions of objects. Two circles are
considered to intersect if the distance between their
centers is less than the sum of their radii. As an
optimization, you can square both sides of this equation.
ADDITIONAL READING
Eric Lengyel provides an in-depth look at all the
different mathematical concepts used in 3D game
programming. Aspiring graphics programmers
especially should review the more advanced material
in his book. The Gaffer on Games site, maintained by
Glenn Fielder, has several articles on the basics of
physics in games, including articles on different
forms of numeric integration and why fixing a time
step is important. Finally, Ian Millington covers how
to implement Newtonian physics in games in detail.
Fielder, Glenn. Gaffer on Games. Accessed

July 18, 2016. http://gafferongames.com/.
Lengyel, Eric. Mathematics for 3D Game
Programming and Computer Graphics,
3rd edition. Boston: Cengage, 2011.
Millington, Ian. Game Physics Engine
Development, 2nd edition. Boca Raton:
CRC Press, 2010.
EXERCISES
The first exercise for this chapter is a series of short
problems that give you practice using the various
vector techniques covered in this chapter. The next
two exercises look at adding features to the chapter’s
game project.
Exercise 3.1
1. Given the vectors and , and the
scalar value s=2, calculate the following:
(a)
(b)
(c)
2. Given the triangle in Figure 3.16 and the following
points:
A=〈−,1〉
B=〈2,4〉
C=〈3,3〉
Calculate the θ using the vector operations discussed
in this chapter.
Figure 3.16 Triangle for problem 2 of Exercise 3.1
3. Suppose a 2D game features an arrow that points

from the player to the next objective.
When the game first begins, the arrow points down
the x-axis 〈1,0〉.
The player’s initial position is at (4,0).
You create a new object at the position (5,6).
a. What is the unit vector from the player’s initial
position to the new waypoint?
b. Calculate the angle of rotation between the initial
arrow direction and the vector calculated in part
(a).
c. Calculate the vector perpendicular to the plane
created by the initial arrow direction and the vector
calculated in part (a).
Exercise 3.2
Currently, the ship does not collide against asteroids
in the chapter game project. Add collision for the
ship. To do so, you first need to create a
CollisionComponent in Ship and specify a
radius. Next, in Ship::UpdateActor, you need to
test against the collision of all asteroids (much the
way the laser does). If the ship collides with an
asteroid, force it to reset in the center of the screen
with a rotation of zero.
As an extra feature, make it so the ship disappears for a

second or two after colliding with an asteroid. The ship
should then reappear in the center after this delay.
Exercise 3.3
Modify MoveComponent so that it uses Newtonian
physics. Specifically, change it to have a mass, a sum
of forces, and a velocity as member variables. Then
in Update, change the code for forward movement
so it instead calculates an acceleration from the
forces, a velocity from the acceleration, and a
position from the velocity.
Then, you need some method to set forces on the

component. One approach is to add an AddForce
function that takes in a Vector2 and adds that to the
sum of forces variable. You can additionally clear the
sum of forces on every frame after calculating
acceleration. This way, for an impulse you just call
AddForce once. And for a constant force, you simply
call AddForce for that force on every frame.
Finally, change InputComponent, Asteroid, and

Ship so that they work correctly with this new
MoveComponent that supports Newtonian physics.
CHAPTER 4
ARTIFICIAL
INTELLIGENCE
Artificial intelligence (AI) algorithms are

used to determine the actions of computer-
controlled entities in games. This chapter
covers three useful game AI techniques:
changing behaviors with state machines,
computing paths for entities to move
through the world (pathfinding), and making
decisions in two-player turn-based games
(minimax and game trees). The chapter
shows you how to apply some of these AI
techniques to create a tower defense game
project.
STATE MACHINE BEHAVIORS
For very simple games, the AI always has the same
behavior. For instance, an AI for two-player Pong
tracks the position of the ball as it moves. Because
this behavior doesn’t change throughout the game,
it’s stateless. But for more complex games, the AI
behaves differently at different points in time. In
Pac-Man, each ghost has three different behaviors:
chasing the player, scattering away (where the ghost
returns to a set “home area”), or running away from
the player. One way to represent these changes in
behaviors is with a state machine, where each
behavior corresponds to a state.
Designing a State Machine

States by themselves only partially define a state
machine. Equally important is how the state machine
decides to change, or transition between, states.
Furthermore, each state can have actions that occur
on entry or exit of the state.
When implementing a state machine for a game

character’s AI, it’s prudent to plan the different states
and how they interconnect. Take the example of a basic
guard character in a stealth game. By default, the guard
patrols on a predefined path. If the guard detects the
player while on patrol, it starts attacking the player. And,
if at any point in time, the guard receives fatal damage, it
dies. In this example, the guard AI has three different
states: Patrol, Attack, and Death.
Next, you need to define the transitions for each state.

The Death state transition is simple: When the guard
takes fatal damage, it transitions to Death. This happens
regardless of the current state. The guard enters the
Attack state if, during the Patrol state, the guard spots
the player. The state machine diagram in Figure 4.1
represents this combination of states and transitions.
Figure 4.1 Basic stealth AI state machine
Although this AI is functional, AI characters in most

stealth games are more complex. Suppose the guard
hears a suspicious sound while in the Patrol state. The
current state machine dictates that the guard continue
patrolling. Ideally, the sound should startle the guard
and cause the guard to search for the player. An
Investigate state can represent this behavior.
Furthermore, in this state machine example, the guard

always attacks when detecting the player. But for variety,
maybe the guard occasionally triggers an alarm instead.
An Alert state can represent this behavior. The Alert
state randomly transitions out to either Attack or
another new state, Alarm. Adding these refinements
makes the state machine more complex, as shown in
Figure 4.2.
Figure 4.2 More complex stealth AI state machine
From the Alert state you have two transitions: 75% and
25%. These transitions refer to the probability of the
transition. So, there’s a 75% chance that when in the
Alert state, the AI will transition to the Attack state. In
the Alarm state, the Complete transition means that after
the AI finishes triggering the alarm (perhaps by
interacting with some object in the game world), the AI
transitions into the Attack state.
Further refinements to the state machine are possible.

But the principles of designing an AI state machine are
the same regardless of the number of states. In any
event, after defining a state machine, the next step is to
implement it in code.
Basic State Machine Implementation

There are several ways to implement a state machine.
Minimally, the code must update the behavior of the
AI based on the current state, and it must support
enter and exit actions. An AIComponent class can
encapsulate this state behavior.
If there are only two states, a simple Boolean check in

Update would work, though it isn’t very robust. A more
flexible implementation is to use an enum to represent
the different states. For the state machine in Figure 4.1,
this is the enum declaration:
enum AIState
Patrol,
Death,
Attack
};
Then, you create an AIComponent class that has an

instance of AIState as member data. You also define
separate update functions for each state:
UpdatePatrol, UpdateDeath, and UpdateAttack.
The AIComponent::Update function then has a switch
on the AIState member variable and calls the update
function that corresponds to the current state:
void AIComponent::Update(float deltaTime)
switch (mState)
case Patrol:
UpdatePatrol(deltaTime);
break;
case Death:
UpdateDeath(deltaTime);
break;
case Attack:
UpdateAttack(deltaTime);
break;
default:
// Invalid
break;
}
}
You can handle the state machine transitions in a

separate ChangeState function. This way, the various
update functions can initiate a transition just by calling
ChangeState. You can implement ChangeState as
follows:
void AIComponent::ChangeState(State newState)
// Exit current state
// (Use switch to call corresponding Exit function)
// ...
mState = newState;
// Enter current state
// (Use switch to call corresponding Enter function)
// ...
Although this implementation is simple, there are issues.

First, it doesn’t scale well; adding more states reduces
the readability of both Update and ChangeState. Also,
having so many separate Update, Enter, and Exit
functions also makes the code harder to follow.
It’s also not easy to mix and match functionality between

multiple AIs. Two different AIs with different state
machines need separate enums and, therefore, separate
AI components. But many AI characters may share some
functionality. Suppose two AIs have mostly different
state machines, but both have a Patrol state. With this
basic implementation, it isn’t easy to share the Patrol
code between both AI components.
States as Classes
An alternative approach to the one just described is
to use classes to represent each state. First, define a
base class for all states called AIState:
class AIState
public:
AIState(class AIComponent* owner)
:mOwner(owner)
{ }
// State-specific behavior
virtual void Update(float deltaTime) = 0;
virtual void OnEnter() = 0;
virtual void OnExit() = 0;
// Getter for string name of state
virtual const char* GetName() const = 0;
protected:
class AIComponent* mOwner;
};
The base class includes several virtual functions to

control the state: Update updates the state, OnEnter
implements any entry transition code, and OnExit
implements any exit transition code. The GetName
function simply returns a human-readable name for the
state. You also associate AIState with a specific
AIComponent through the mOwner member variable.
Next, declare the AIComponent class, as follows:
class AIComponent : public Component
public:
AIComponent(class Actor* owner);

void ChangeState(const std::string& name);
// Add a new state to the map
void RegisterState(class AIState* state);
private:
// Maps name of state to AIState instance
std::unordered_map<std::string, class AIState*> mStateMap;
// Current state we're in
class AIState* mCurrentState;
};
Notice how AIComponent has a hash map of state

names to AIState instance pointers. It also has a
pointer to the current AIState. The RegisterState
function takes in a pointer to an AIState and adds the
state to the map:
void AIComponent::RegisterState(AIState* state)
{
mStateMap.emplace(state->GetName(), state);
The AIComponent::Update function is also

straightforward. It simply calls Update on the current
state, if it exists:
void AIComponent::Update(float deltaTime)
if (mCurrentState)
mCurrentState->Update(deltaTime);
However, the ChangeState function does several

things, as shown in Listing 4.1. First, it calls OnExit on
the current state. Next, it tries to find the state you’re
changing to in the map. If it finds this state, it changes
mCurrentState to the new state and calls OnEnter on
this new state. If it can’t find the next state in the map, it
outputs an error message and sets mCurrentState to
null.
Listing 4.1 AIComponent::ChangeState
Implementation
void AIComponent::ChangeState(const std::string& name)
// First exit the current state
if (mCurrentState)
mCurrentState->OnExit();
// Try to find the new state from the map
auto iter = mStateMap.find(name);
if (iter != mStateMap.end())
mCurrentState = iter->second;
// We're entering the new state
mCurrentState->OnEnter();
else
{
SDL_Log("Could not find AIState %s in state map", name.c_str());
mCurrentState = nullptr;
You can use this pattern by first declaring subclasses of

AIState, like this AIPatrol class:
class AIPatrol : public AIState
public:
AIPatrol(class AIComponent* owner);
// Override with behaviors for this state
void OnEnter() override;
void OnExit() override;
const char* GetName() const override
{ return "Patrol"; }
};
You then implement any special behaviors in Update,
OnEnter, and OnExit. Suppose you want AIPatrol to
change to the AIDeath state when the character dies. To
initiate the transition, you need to call ChangeState on
the owning component, passing in the name of the new
state:
void AIPatrol::Update(float deltaTime)
// Do some other updating
// ...
bool dead = /* Figure out if I'm dead */;
if (dead)
// Tell the ai component to change states
mOwner->ChangeState("Death");
On the ChangeState call, the AIComponent looks into

its state map, and if it finds a state named Death, it
transitions into this state. You can similarly declare
AIDeath and AIAttack classes to complete the basic
state machine from Figure 4.1.
To hook up the states into an AIComponent’s state map,

first create an actor and its AIComponent and then call
Register on any states you wish to add to the state
machine:
Actor* a = new Actor(this);
// Make an AIComponent
AIComponent* aic = new AIComponent(a);
// Register states with AIComponent
aic->RegisterState(new AIPatrol(aic));
aic->RegisterState(new AIDeath(aic));
aic->RegisterState(new AIAttack(aic));
To then set the AIComponent to an initial patrol state,

you call ChangeState, as follows:
aic->ChangeState("Patrol");
Overall, this approach is useful because each state’s

implementation is in a separate subclass, which means
the AIComponent remains simple. It also makes it
significantly easier to reuse the same states for different
AI characters. You simply need to register whichever
states you want with the new actor’s AIComponent.
PATHFINDING
A pathfinding algorithm finds a path between two
points, avoiding any obstacles in the way. The
complexity of this problem stems from the fact that
there might be a large set of paths between two
points, but only a small number of these paths are
the shortest. For example, Figure 4.3 shows two
potential routes between points A and B. An AI
traveling along the solid path is not particularly
intelligent because the dashed path is shorter. Thus,
you need a method to efficiently search through all
the possible paths to find one with the shortest
distance.
Figure 4.3 Two paths from A to B
Graphs
Before you can solve the pathfinding problem, you
first need a way to represent the parts of the game
world that the AI can path through. A popular choice
is the graph data structure. A graph contains a set of
nodes (also called vertices). These nodes connect to
each other via edges. These edges can be
undirected, meaning they are traversable in both
directions, or directed, meaning they are
traversable in only one direction. You might use a
directed edge for a case where the AI can jump down
from a platform but can’t jump back up. You could
represent this connection with a directed edge from
the platform to the ground.
Optionally, edges may have weights associated with

them, representing the cost of traversing the edge. In a
game, the weight minimally accounts for the distance
between the nodes. However, you might modify the
weight based on the difficulty of traversing the edge.
For example, if an edge moves over quicksand in the

game world, it should have a higher weight than an edge
of the same length that moves over concrete. A graph
without edge weights (an unweighted graph)
effectively is a graph where the weight of every edge is a
constant value. Figure 4.4 illustrates a simple undirected
and unweighted graph.
Figure 4.4 A sample graph
There are multiple ways to represent a graph in memory,

but this book uses adjacency lists. In this
representation, each node has a collection of adjacent
nodes (using std::vector). For an unweighted graph,
this adjacency list contains pointers to adjacent nodes.
The graph is then just a collection of such nodes:
struct GraphNode
// Each node has pointers to adjacent nodes
std::vector<GraphNode*> mAdjacent;
};
struct Graph
{
// A graph contains nodes
std::vector<GraphNode*> mNodes;
};
For a weighted graph, instead of a list of connected

nodes, each node stores its outgoing edges:
struct WeightedEdge
// Which nodes are connected by this edge?
struct WeightedGraphNode* mFrom;
struct WeightedGraphNode* mTo;
// Weight of this edge
float mWeight;
};
struct WeightedGraphNode
// Stores outgoing edges
std::vector<WeightedEdge*> mEdges;
};
// (A WeightedGraph has WeightedGraphNodes)
By referencing both the “from” and “to” nodes in each

edge, you can support a directed edge from node A to B
by adding an edge to node A’s mEdges vector but not to
node B’s. If you want an undirected edge, you simply add
two directed edges, one in each direction (for example,
from A to B and from B to A).
Different games represent the game world via graphs in

different manners. Partitioning a world into a grid of
squares (or hexes) is the simplest approach. This
approach is very popular for turn-based strategy games
such as Civilization or XCOM. However, for many other
types of games, it isn’t feasible to use this approach. For
simplicity, most of this section sticks with a grid of
squares. However, you will learn about other possible
representations later in this chapter.
Breadth-First Search
Suppose a game takes place in a maze designed in a
square grid. The game only allows movement in the
four cardinal directions. Because each move in the
maze is uniform in length, an unweighted graph can
represent this maze. Figure 4.5 shows a sample maze
and its corresponding graph.
Figure 4.5 A maze on a square grid and its
corresponding graph
Now imagine that a mouse AI character starts at some

square in the maze (the start node) and wants to find
the shortest path to a piece of cheese in the maze (the
goal node). One approach is to first check all squares
one move away from the start. If none of these squares
contains the cheese, you then check all squares two
moves away from the start. Repeat this process until
either the cheese is found or there are no valid moves
left. Because this algorithm only considers the further
nodes once the closer nodes are exhausted, it won’t miss
the shortest path to the cheese. This describes what
happens in a breadth-first search (BFS). The BFS
algorithm guarantees to find the shortest path when
either the edges are unweighted or every edge has the
same positive weight.
With some minor bookkeeping during BFS, it’s possible

to reconstruct the path with the minimal number of
moves. Once it’s computed, AI characters can then follow
along this path.
During BFS, each node needs to know the node visited

immediately before it. That node, called the parent
node, helps reconstruct the path after BFS completes.
While you could add this data to the GraphNode struct,
it’s better to separate the data that doesn’t change (the
graph itself) from the parents. This is because the
parents will change depending on the start and goal
nodes selected. Separating these pieces of data also
means that if you want to compute several paths
simultaneously across multiple threads, the searches
won’t interfere with each other.
To support this, first define a type of map called a

NodeToPointerMap, which simply is an unordered
map where both the key and value are GraphNode
pointers (the pointers are const because you don’t need
to modify the graph nodes):
using NodeToParentMap =
std::unordered_map<const GraphNode*, const GraphNode*>;

With this type of map, you can then implement BFS as in
Listing 4.2. The simplest way to implement BFS is with a
queue. Recall that a queue uses FIFO (first-in, first-out)
behavior when adding and removing nodes. You can add
a node to a queue via an enqueue operation and remove
a node via dequeue. To begin, you enqueue the start node
and enter a loop. In each iteration, you dequeue a node
and enqueue its neighbors. You can avoid adding the
same node multiple times to the queue by checking the
parent map. A node’s parent is null only if the node
hasn’t been enqueued before or it’s the start node.
When you use the square brackets on the outMap, one of

two things happens. If the key already exists in the map,
you can just access its parent. Otherwise, if the key does
not exist in the map, the map by default constructs a
value for that key. In this case, if you access outMap and
the node requested isn’t in the map, you initialize that
node’s parent to nullptr.
Even if no path exists between the start and the goal, the
loop will eventually terminate. This is because the
algorithm checks all nodes that are reachable from start.
Once all possibilities are exhausted, the queue becomes
empty and the loop ends.
Listing 4.2 Breadth-First Search

bool BFS(const Graph& graph, const GraphNode* start,

const GraphNode* goal, NodeToParentMap& outMap)
// Whether we found a path
bool pathFound = false;
// Nodes to consider
std::queue<const GraphNode*> q;
// Enqueue the first node
q.emplace(start);
while (!q.empty())
// Dequeue a node
const GraphNode* current = q.front();
q.pop();
if (current == goal)
pathFound = true;
break;

// Enqueue adjacent nodes that aren't already in the queue
for (const GraphNode* node : current->mAdjacent)
// If the parent is null, it hasn't been enqueued
// (except for the start node)
const GraphNode* parent = outMap[node];
if (parent == nullptr && node != start)
// Enqueue this node, setting its parent
outMap[node] = current;
q.emplace(node);
return pathFound;
Assuming that you have a Graph g, you can then run

BFS between two GraphNodes in the graph with the
following two lines:
NodeToParentMap map;
bool found = BFS(g, g.mNodes[0], g.mNodes[9], map);
If BFS succeeds, you can reconstruct the path by using

the parent pointers in the outMap. This is because the
goal’s parent points to the preceding node on the path.
Similarly, the parent of the node preceding the goal node
is two moves away from the goal. Following this chain of
parent pointers eventually leads back to the start node,
yielding a path from goal to start.
Unfortunately, you want the path in the opposite

direction: from start to goal. One solution is to reverse
the path with a stack, but a more intelligent approach is
to reverse the search. For example, instead of passing in
the mouse node as start and the cheese node as goal, do
the opposite. Then, following the parent pointers from
the goal node yields the desired path.
BFS always finds a path between the start and goal nodes
if one exists. But for weighted graphs, BFS doesn’t
guarantee to find the shortest path. This is because BFS
doesn’t look at the weight of the edges at all; every edge
traversal is equivalent. In Figure 4.6, the dashed path has
the shortest distance, but BFS returns the solid path as it
requires only two moves.
Figure 4.6 BFS finds the solid path even though the
dashed path is shorter
Another issue with BFS is that it tests nodes even if they

are in the opposite direction of the goal. By using a more
complex algorithm, it’s possible to reduce the number of
nodes you test on the way to finding the optimal
solution.
Most other pathfinding algorithms used in games have

an overall structure like BFS. On every iteration, you pick
one node to inspect next and add its neighbors to a data
structure. What changes is that different pathfinding
algorithms evaluate nodes in different orders.
Heuristics
Many search algorithms rely on a heuristic, which
is a function that approximates an expected result. In
pathfinding, the heuristic is the estimated cost from
a given node to the goal node. A heuristic can help
you more quickly find a path. For example, on each
iteration of BFS, you dequeue the next node in the
queue, even if that node sends you in a direction
pointing away from the goal. With a heuristic, you
can estimate how close you think a specific node is to
the goal and then choose to look at the “closer” nodes
first. This way, the pathfinding algorithm is likely to
terminate with fewer iterations.
The notation h(x) denotes the heuristic, where x is a

node in the graph. So, h(x) is the estimated cost from
node x to the goal node.
A heuristic function is admissible if it is always less

than or equal to the actual cost from node x to the goal. If
the heuristic occasionally overestimates the actual cost,
it’s inadmissible, and you shouldn’t use it. The A*
algorithm, discussed later in this section, requires an
admissible heuristic to guarantee the shortest path.
For a grid of squares, there are two common ways to

compute the heuristic. For example, in Figure 4.7, the
checkered node represents the goal and the solid node
represents the start. The gray squares in this figure
denote squares that are impassible.
Figure 4.7 Manhattan heuristic (left) and Euclidean
heuristic (right)
The Manhattan distance heuristic, illustrated in

Figure 4.7 (left), is akin to traveling along city blocks in a
sprawling metropolis. A building might be “five blocks
away,” but there may be multiple routes five blocks in
length. Manhattan distance assumes that diagonal
movements are invalid. If diagonal movements are valid,
Manhattan distance often overestimates the cost, making
the heuristic inadmissible.
For a 2D grid, the following formula calculates

Manhattan distance:
A second type of heuristic is Euclidean distance,

illustrated in Figure 4.7 (right). You use the standard
distance formula to calculate this heuristic, which
estimates an “as the crow flies” route. Unlike Manhattan
distance, Euclidean distance can easily work for worlds
more complex than a square grid. In 2D, the Euclidean
distance equation is as follows:
The Euclidean distance function is almost always

admissible, even in cases where the Manhattan distance
is inadmissible. This means that Euclidean distance is
usually the recommended heuristic function. However,
the Manhattan heuristic is more efficient to compute
because it doesn’t involve a square root.
The only case where a Euclidean distance heuristic

overestimates the true cost is if the game allows non-
Euclidean movement such as teleporting between two
nodes across the level.
Notice that in Figure 4.7 both heuristic h(x) functions

end up underestimating the actual cost of traveling from
the start node to the goal node. This happens because the
heuristic function knows nothing about the adjacency
lists, so it doesn’t know whether certain areas are
impassible. This is fine because the heuristic is the lower
bound of how close node x is to the goal node; the
heuristic guarantees that node x is at least that distance
away. This is more useful in a relative sense: The
heuristic can help estimate whether node A or node B is
closer to the goal node. And then you can use this
estimate to help decide whether to explore node A or
node B next.
The following section shows how to use the heuristic

function to create a more complex pathfinding
algorithm.
Greedy Best-First Search

BFS uses a queue to consider nodes in a FIFO
manner. Greedy best-first search (GBFS)
instead uses the h(x) heuristic function to decide
which node to consider next. Although this seems
like a reasonable pathfinding algorithm, GBFS
cannot guarantee a minimal path. Figure 4.8 shows
the resultant path from a sample GBFS search.
Nodes in gray are impassible. Note that the path
makes four additional moves from the start rather
than going straight down.
Figure 4.8 Greedy best-first path
note
Although GBFS does not guarantee optimality, it’s useful to understand
because it requires only a couple modifications to become A*. The A*
algorithm does guarantee the shortest path if the heuristic is admissible. So
before moving on to A*, it’s important to understand the GBFS
implementation.
Instead of using a single queue, GBFS uses two sets of

nodes during the search. The open set contains the
nodes that are under consideration. Once chosen for
evaluation, a node moves into the closed set. When a
node is in the closed set, GBFS need not investigate it
further. There’s no guarantee that a node in the open or
closed set will ultimately be on the path; these sets just
help prune nodes from consideration.
Selecting data structures for the open set and the closed
set presents an interesting dilemma. For the open set,
the two operations you need are removing the node with
the lowest cost and testing for membership. The closed
set only needs a membership test. To speed up the
membership test, you can simply use Booleans in scratch
data to track if a specific node is a member of the open
set or the closed set. And because the closed set just
needed this membership test, you don’t use an actual
collection for the closed set.
For the open set, one popular data structure is a priority

queue. However, in the interest of simplicity, this chapter
uses a vector for the open set. With a vector, you can just
use a linear search to find the element in the open set
with the lowest cost.
As with BFS, each node needs additional scratch data

during the GBFS search. Because you now have multiple
pieces of scratch data per node, it makes sense to define
a struct to encapsulate it. To use a weighted graph, the
parent is an incoming edge as opposed to a preceding
node. In addition, each node tracks its heuristic value
and its membership in the open and closed sets:
struct GBFSScratch
const WeightedEdge* mParentEdge = nullptr;
float mHeuristic = 0.0f;
bool mInOpenSet = false;
bool mInClosedSet = false;
};
Then, define a map where the key is a pointer to the node

and the value is an instance of GBFSScratch:
using GBFSMap =
std::unordered_map<const WeightedGraphNode*, GBFSScratch>;
Now you have the necessary components for a greedy

best-first search. The GBFS function takes in a
WeightedGraph, the start node, the goal node, and a
reference to a GBFSMap:
bool GBFS(const WeightedGraph& g, const WeightedGraphNode* start,
const WeightedGraphNode* goal, GBFSMap& outMap);
At the start of the GBFS function, you define a vector for

the open set:
std::vector<const WeightedGraphNode*> closedSet;
Next, you need a variable to track the current node,

which is the node under evaluation. This updates as the
algorithm progresses. Initially, current is the start
node, and you “add” it to the closed set by marking it as
closed in the scratch map:
const WeightedGraphNode* current = start;
outMap[current].mInClosedSet = true;
Next, you enter the main loop of GBFS. This main loop
does several things. First, it looks at all nodes adjacent to
the current node. It only considers nodes that aren’t
already in the closed set. These nodes have their parent
edge set to the edge incoming from the current node. For
nodes not already in the open set, the code computes the
heuristic (from the node to the goal) and adds the node
to the open set:
do
// Add adjacent nodes to open set
for (const WeightedEdge* edge : current->mEdges)
// Get scratch data for this node
GBFSScratch& data = outMap[edge->mTo];
// Add it only if it's not in the closed set
if (!data.mInClosedSet)
// Set the adjacent node's parent edge
data.mParentEdge = edge;
if (!data.mInOpenSet)
// Compute the heuristic for this node, and add to open set
data.mHeuristic = ComputeHeuristic(edge->mTo, goal);
data.mInOpenSet = true;
openSet.emplace_back(edge->mTo);
}
The ComputeHeuristic function can use any heuristic

h(x) function, such as Manhattan or Euclidean distance.
In practice, this may require additional information
stored in each node (such as the position of the node in
the world).
After processing the nodes adjacent to the current node,

you need to look at the open set. If it’s empty, this means
there are no nodes left to evaluate. This happens only if
there is no path from start to goal:
if (openSet.empty())
break; // Break out of outer loop
Alternatively, if there are still nodes in the open set, the

algorithm continues. You need to find the node in the
open set with the lowest heuristic cost and move it to the
closed set. This node becomes the new current node:
// Find lowest cost node in open set
auto iter = std::min_element(openSet.begin(), openSet.end(),
[&outMap](const WeightedGraphNode* a, const WeightedGraphNode* b)
return outMap[a].mHeuristic < outMap[b].mHeuristic;
});
// Set to current and move from open to closed
current = *iter;
openSet.erase(iter);
outMap[current].mInOpenSet = false;
To code to find the lowest element, uses the

std::min_element function from the <algorithm>
header. For its third parameter, min_element takes in a
special type of function (called a lambda expression) to
specify how to decide whether one element is less than
another. The min_element function returns an iterator
to the minimum element.
Finally, the main loop continues if the current node is

not the goal node:
} while (current != goal);

The loop terminates either when the above while
condition fails or when you hit the earlier break
statement (for when the open set is empty). You can then
figure out if GBFS found a path based on whether the
current node equals the goal node:
return (current == goal) ? true : false;
Figure 4.9 shows the first two iterations of GBFS applied

to a sample data set. In Figure 4.9(a), the start node (A2)
is in the closed set, and its adjacent nodes are in the open
set. To make the figure easy to read, it uses the
Manhattan distance heuristic. The arrows point from
children back to their parent node. The next step is to
select the node with the lowest heuristic cost, which is
the node with h = 3. This node becomes the new current
node and moves into the closed set. Figure 4.9(b) shows
the next iteration, where C2 is now the node with the
lowest cost in the open set.
Figure 4.9 Greedy best-first snapshots: (a) first
iteration and (b) second iteration
Keep in mind that just because a node in the open set has
the lowest heuristic cost doesn’t mean it’s on the optimal
path. For example, in Figure 4.9(b), the node C2 is not
on the optimal path. Unfortunately, the GBFS algorithm
still selects C2 for its path. Clearly, you need to do some
refinement to fix this issue.
Listing 4.3 shows the complete code for the greedy best-
first search function.
Listing 4.3 Greedy Best-First Search

bool GBFS(const WeightedGraph& g, const WeightedGraphNode* start,
const WeightedGraphNode* goal, GBFSMap& outMap)
std::vector<const WeightedGraphNode*> openSet;
// Set current node to start, and mark in closed set
const WeightedGraphNode* current = start;
do
// Add adjacent nodes to open set

GBFSScratch& data = outMap[edge->mTo];
// Consider it only if it's not in the closed set
// Set the adjacent node's parent edge
// Compute the heuristic for this node, and add to open set
data.mHeuristic = ComputeHeuristic(edge->mTo, goal);
openSet.emplace_back(edge->mTo);
if (openSet.empty())
{ break; }
// Find lowest cost node in open set
auto iter = std::min_element(openSet.begin(), openSet.end(),
[&outMap](const WeightedGraphNode* a, const WeightedGraphNode* b)
return outMap[a].mHeuristic < outMap[b].mHeuristic;
});
// Set to current and move from open to closed
current = *iter;
openSet.erase(iter);
outMap[current].mInOpenSet = false;
} while (current != goal);
// Did you find a path?
return (current == goal) ? true : false;
A* Search
The downside of GBFS is that it can’t guarantee an
optimal path. Luckily, with some modifications to
GBFS, you can transform it into the A* search
(pronounced “A-star”). A* adds a path-cost
component, which is the actual cost from the start
node to a given node. The notation g(x) denotes the
path-cost of a node x. When selecting a new current
node, A* selects the node with the lowest f(x) value,
which is just the sum of the g(x) path-cost and the
h(x) heuristic for that node:
There are a few conditions for A* to find an optimal path.

Of course, there must be some path between the start
and goal. Furthermore, the heuristic must be admissible
(so it can’t overestimate the actual cost). Finally, all edge
weights must be greater than or equal to zero.
To implement A*, you first define an AStarScratch

struct, as you do for GBFS. The only difference is that the
AStarScratch struct also has a float member
mActualFromStart to store the g(x) value.
There are additional differences between the GBFS code

and the A* code. When adding a node to the open set, A*
must also compute the path-cost g(x). And when
selecting the minimum node, A* selects the node with
the lowest f(x) cost. Finally, A* is pickier about which
nodes become parents, using a process called node
adoption.
In the GBFS algorithm, adjacent nodes always have their
parents set to the current node. But in A*, the g(x) path-
cost value of a node is dependent on the g(x) value of its
parent. This is because the path-cost value for node x is
simply its parent’s path-cost value plus the cost of
traversing the edge from the parent to node x. So before
assigning a new parent to a node x, A* first makes sure
the g(x) value will improve.
Figure 4.10(a) once again uses the Manhattan heuristic

function. The current node (C3) checks its adjacent
nodes. The node to its left has g = 2 and B2 as its parent.
If that node instead had C3 as its parent, it would have g
= 4, which is worse. So, A* will not change B2’s parent in
this case.
Figure 4.10 (a) Current node adoption fails; (b)

final A* path
Figure 4.10(b) shows the final path as computed by A*,

which clearly is superior to the GBFS solution.
Apart from node adoption, the code for A* ends up being

very similar to the GBFS code. Listing 4.4 shows the loop
over the adjacent nodes, which contains most of the code
changes. The only other change not shown in the text is
the code that selects the lowest-cost node in the open set
based on f(x) instead of just h(x). The game project for
this chapter provides the code for the full A*
implementation.
Listing 4.4 Loop over the Adjacent Nodes in an A*

Search
const WeightedGraphNode* neighbor = edge->mTo;
AStarScratch& data = outMap[neighbor];
// Only check nodes that aren't in the closed set
// Not in the open set, so parent must be current
data.mHeuristic = ComputeHeuristic(neighbor, goal);
// Actual cost is the parent's plus cost of traversing edge
data.mActualFromStart = outMap[current].mActualFromStart +
edge->mWeight;
openSet.emplace_back(neighbor);
else
// Compute what new actual cost is if current becomes parent
float newG = outMap[current].mActualFromStart + edge->mWeight;
if (newG < data.mActualFromStart)
// Current should adopt this node
data.mActualFromStart = newG;
}
note
Optimizing A* to run as efficiently as possible is a complex topic. One
consideration is what happens if there are a lot of ties in the open set. This is
bound to happen in a square grid, especially if you use the Manhattan
heuristic. If there are too many ties, when it’s time to select a node, you have
a high probability of selecting one that doesn’t end up on the path. This
ultimately means you need to explore more nodes in the graph, which makes
A* run more slowly.
One way to help eliminate ties is to add a weight to the heuristic function,
such as arbitrarily multiplying the heuristic by 0.75. This gives more weight to
the path-cost g(x) function over the heuristic h(x) function, which means
you’re more likely to explore nodes further from the start node.
From an efficiency standpoint, A* actually is a poor choice for grid-based
pathfinding. Other pathfinding algorithms are far more efficient for grids. One
of them is the JPS+ algorithm, outlined in Steve Rabin’s Game AI Pro 2 (see
the “Additional Reading” section). However, A* works on any graph, whereas
JPS+ works only on grids.
Dijkstra’s Algorithm
Let’s return to the maze example but now suppose
that the maze has multiple pieces of cheese in it, and
you want the mouse to move toward the closest
cheese. A heuristic could approximate which cheese
is closest, and A* could find a path to that cheese.
But there’s a chance that the cheese you select with
the heuristic isn’t actually the closest because the
heuristic is only an estimate.
In Dijkstra’s algorithm, there is a source node but no

goal node. Instead, Dijkstra’s computes the distance
from the source node to every other reachable node in
the graph. In the maze example, Dijkstra’s would find the
distance of all reachable nodes from the mouse, yielding
the actual cost of travel to every piece of cheese and
allowing the mouse to move to the closest one.
It’s possible to convert the A* code from the previous
section into Dijkstra’s. First, you remove the h(x)
heuristic component. This is equivalent to a heuristic
function of h(x) = 0, which is admissible because it’s
guaranteed to be less than or equal to the actual cost.
Next, you remove the goal node and make the loop
terminate only when the open set is empty. This then
computes the g(x) path-cost for every node reachable
from the start.
The original formulation of the algorithm by Edsger

Dijkstra is slightly different. But the approach proposed
in this section is functionally equivalent to the original.
(AI textbooks sometimes call this approach uniform
cost search). Interestingly, the invention of Dijkstra’s
algorithm predates GBFS and A*. However, games
usually prefer heuristic-guided approaches such as A*
because they generally search far fewer nodes than
Dijkstra’s.
Following a Path
Once the pathfinding algorithm generates a path, the
AI needs to follow it. You can abstract the path as a
sequence of points. The AI then just moves from
point to point in this path. You can implement this in
a subclass of MoveComponent called
NavComponent. Because MoveComponent can
already move an actor forward, NavComponent only
needs to rotate the actor to face the correct direction
as the actor moves along the path.
First, the TurnTo function in NavComponent rotates
the actor to face a point:
void NavComponent::TurnTo(const Vector2& pos)
// Vector from me to pos
Vector2 dir = pos - mOwner->GetPosition();
// New angle is just atan2 of this dir vector
// (Negate y because +y is down on screen)
float angle = Math::Atan2(-dir.y, dir.x);
mOwner->SetRotation(angle);
Next, NavComponent has a mNextPoint variable that

tracks the next point in the path. The Update function
tests whether the actor reaches mNextPoint:
void NavComponent::Update(float deltaTime)
// If you've reached the next point, advance along path
Vector2 diff = mOwner->GetPosition() - mNextPoint;

if (diff.Length() <= 2.0f)
mNextPoint = GetNextPoint();
TurnTo(mNextPoint);
// This moves the actor forward
MoveComponent::Update(deltaTime);
This assumes that the GetNextPoint function returns

the next point on the path. Assuming that the actor starts
at the first point on the path, initializing mNextPoint to
the second point as well as setting a linear speed gets the
actor moving along the path.
There’s one issue with updating the movement along the

path in this way: It assumes that the actor is not moving
so fast that it jumps too far past a node in one step. If
this happens, the distance between the two will never be
close enough, and the actor will get lost.
Other Graph Representations

For a game with real-time action, non-player
characters (NPCs) usually don’t move from square to
square on a grid. This makes it more complex to
represent the world with a graph. This section
discusses two alternative approaches: using path
nodes and using navigation meshes.
Path nodes (also called waypoint graphs) became

popular with the advent of first-person shooter (FPS)
games in the early 1990s. With this approach, a designer
places path nodes at locations in the game world that the
AI can path to. These path nodes directly translate into
nodes in the graph.
Typically, you generate the edges between path nodes

automatically. The algorithm works as follows: For each
path node, test whether there are obstructions between it
and nearby nodes. Any paths without obstructions yield
edges. A line segment cast or similar collision test can
determine if there are obstructions. Chapter 10,
“Collision Detection,” covers how to implement line
segment casts.
The primary drawback of using path nodes is that the AI

can only move to locations on the nodes or edges. This is
because even if path nodes form a triangle, there is no
guarantee that the interior of the triangle is a valid
location. There may be an obstruction in the way, so the
pathfinding algorithm must assume that any location not
on a node or an edge is invalid.
In practice, this means that either there’s a lot of space in

the world that’s off-limits to AI or you need many path
nodes. The first is undesirable because it results in less
believable behavior from the AI, and the second is simply
inefficient. The more nodes and more edges there are,
the longer it takes a pathfinding algorithm to arrive at a
solution. This presents a trade-off between performance
and accuracy.
Other games use a navigation mesh (or nav mesh). In

this approach, each node in the graph corresponds to a
convex polygon. Adjacent nodes are any adjacent convex
polygons. This means that a handful of convex polygons
can represent entire regions in the game world. With a
navigation mesh, the AI can safely travel to any location
inside a convex polygon node. This means the AI has
improved maneuverability. Figure 4.11 compares the
path node and navigation mesh representations of a
location in a game.
Figure 4.11 Path node (a) and navigation mesh (b)

representations of a room
Navigation meshes also better support characters of

different sizes. Suppose a game has both cows and
chickens walking around a farm. Given that chickens are
smaller than cows, there are some areas that are
accessible to chickens but not to cows. Therefore, a path
node network designed for chickens won’t work correctly
for cows. This means that if the game uses path nodes, it
needs two separate graphs: one for each type of creature.
In contrast, each node in a navigation mesh is a convex
polygon, so it’s possible to calculate whether a character
fits in a specific area. Therefore, the game can use a
single navigation mesh for both chickens and cows.
Most games that use navigation meshes automatically

generate them. This is useful because designers can
change a level without worrying much about the effect on
AI pathing. However, navigation mesh generation
algorithms are complex. Luckily, there are open source
libraries that implement nav mesh generation. The most
popular, Recast, generates a navigation mesh given the
triangle geometry of a 3D level. See the “Additional
Reading” section at the end of this chapter for more
information on Recast.
GAME TREES
Games such as tic-tac-toe or chess are very different
from most real-time games. First, the game has two
players, and each player alternates taking a turn.
Second, the game is adversarial, meaning that the
two players are playing against each other. The AI
needs for these types of games are very different
from those of real-time games. These types of games
require some representation of the overall game
state, and this state informs the AI’s decisions. One
approach is to use a tree called a game tree. In a
game tree, the root node represents the current state
of the game. Each edge represents a move in the
game and leads to a new game state.
Figure 4.12 shows a game tree for an in-progress game of

tic-tac-toe. Starting at the root node, the current player
(called the max player) can select from three different
moves. After the max player makes a move, the game
state transitions to a node in the first level of the tree.
The opponent (called the min player) then selects a
move leading to the second level of the tree. This process
repeats until reaching a leaf node, which represents an
end state of the game.
Figure 4.12 Partial game tree for tic-tac-toe
In tic-tac-toe, there are only three outcomes: win, lose, or

draw. The numeric values assigned to the leaf nodes in
Figure 4.12 reflect these outcomes. These values are
scores from the perspective of the max player: 1 means
the max player wins, -1 means the min player wins, and 0
means a tie.
Different games have different state representations. For

tic-tac-toe, the state is simply a 2D array of the board:
struct GameState
enum SquareState { Empty, X, O };
SquareState mBoard[3][3];
};
A game tree node stores both a list of children as well as

the game state at that node:
struct GTNode
// Children nodes
std::vector<GTNode*> mChildren;
// State of game at this node
GameState mState;
};
To generate a complete game tree, you set the root node
to the current game state and create children for each
possible first move. Then you repeat this process for each
node in the first level and continue until all moves are
exhausted.
The size of a game tree grows exponentially based on the

number of potential moves. For tic-tac-toe, the upper
bound of the game tree is 9!, or 362,880 nodes. This
means it’s possible to generate and evaluate a complete
game tree for tic-tac-toe. But for chess, a complete game
tree would have 10120 nodes, which makes it impossible
to fully evaluate (both in terms of time and space
complexity). For now, let’s assume we have a complete
game tree. Later, we’ll discuss how to manage an
incomplete tree.
Minimax
The minimax algorithm evaluates a two-player
game tree to determine the best move for the current
player. Minimax assumes that each player will make
the choice most beneficial to herself. Because scores
are from the perspective of the max player, this
means the max player tries to maximize her score,
while the min player strives to minimize the score of
the max player.
For example, in Figure 4.12 the max player (X in this

case) has three possible moves. If max selects either top-
mid or bottom-mid, the min player (O) can win with
bottom-right. The min player would take this winning
play when available. Thus, the max player selects
bottom-right to maximize her potential final score.
If the max player selects bottom-right, the min player

can select either top-mid or bottom-mid. The choice here
is between a score of 1 or 0. Because the min player aims
to minimize the max player’s score, min selects bottom-
mid. This means the game ends in a tie, which is the
expected result of a game of tic-tac-toe where both
players play optimally.
The implementation of minimax in Listing 4.5 uses a

separate function for the min and max players’ behavior.
Both functions first test if the node is a leaf node, in
which case the GetScore function computes the score.
Next, both functions determine the best possible subtree
using recursion. For the max player, the best subtree
yields the highest value. Likewise, the min player finds
the subtree with the lowest value.
Listing 4.5 MaxPlayer and MinPlayer Functions

float MaxPlayer(const GTNode* node)
// If this is a leaf, return score
if (node->mChildren.empty())
return GetScore(node->mState);
}
// Find the subtree with the maximum value
float maxValue = -std::numeric_limits<float>::infinity();
for (const GTNode* child : node->mChildren)
maxValue = std::max(maxValue, MinPlayer(child));
return maxValue;
float MinPlayer(const GTNode* node)
// If this is a leaf, return score
if (node->mChildren.empty())
return GetScore(node->mState);
// Find the subtree with the minimum value
float minValue = std::numeric_limits<float>::infinity();
for (const GTNode* child : node->mChildren)

{
minValue = std::min(minValue, MaxPlayer(child));
return minValue;
Calling MaxPlayer on the root node returns the best

possible score for the max player. However, this doesn’t
specify which next move is optimal, which the AI player
also wants to know. The code for determining the best
move is in a separate MinimaxDecide function, given in
Listing 4.6. MinimaxDecide resembles the MaxPlayer
function, except it tracks which child yields the best
value.
Listing 4.6 MinimaxDecide Implementation

const GTNode* MinimaxDecide(const GTNode* root)
// Find the subtree with the maximum value, and save the choice
const GTNode* choice = nullptr;
for (const GTNode* child : root->mChildren)

{
float v = MinPlayer(child);
if (v > maxValue)
maxValue = v;
choice = child;
return choice;
Handling Incomplete Game Trees

As mentioned earlier in this chapter, it’s not always
viable to generate a complete game tree. Luckily, it’s
possible to modify the minimax code to account for
incomplete game trees. First, the functions must
operate on a game state as opposed to a node. Next,
rather than iterate over child nodes, the code iterates
over the next possible moves from a given state.
These modifications mean the minimax algorithm
generates the tree during execution rather than
beforehand.
If the tree is too large, such as in chess, it’s still not

possible to generate the entire tree. Much as how an
expert chess player can see only eight moves ahead, the
AI needs to limit the depth of its game tree. This means
the code treats some nodes as leaves even though they
are not terminal states of the game.
To make informed decisions, minimax needs to know

how good these nonterminal states are. But unlike with
terminal states, it’s impossible to know the exact score.
Thus, the scoring function needs a heuristic component
that approximates the quality of nonterminal states. This
also means that scores are now ranges of values, unlike
the {-1, 0, 1} ternary choice for tic-tac-toe.
Importantly, adding the heuristic component means

minimax cannot guarantee to make the best decision.
The heuristic tries to approximate the quality of a game
state, but it’s unknown how accurate this approximation
is. With an incomplete game tree, it’s possible that the
move selected by minimax is suboptimal and eventually
leads to a loss.
Listing 4.7 provides the MaxPlayerLimit function.

(You would need to modify the other functions
similarly.) This code assumes that GameState has three
member functions: IsTerminal, GetScore, and
GetPossibleMoves. IsTerminal returns true if the
state is an end state. GetScore returns either the
heuristic for nonterminal states or the score for terminal
states. GetPossibleMoves returns a vector of the game
states that are one move after the current state.
Listing 4.7 MaxPlayerLimit Implementation
float MaxPlayerLimit(const GameState* state, int depth)
// If this is terminal or we've gone max depth
if (depth == 0 || state->IsTerminal())
return state->GetScore();
// Find the subtree with the max value
for (const GameState* child : state->GetPossibleMoves())
maxValue = std::max(maxValue, MinPlayer(child, depth - 1));
return maxValue;
The heuristic function varies depending on the game. For

example, a simple chess heuristic might count the
number of pieces each player has, weighting the pieces
by power. However, a drawback of such a simple
heuristic is that sometimes sacrificing a piece in the short
term is better for the long term. Other heuristics might
look at control of the board’s center, the safety of the
king, or the mobility of the queen. Ultimately, several
different factors affect the heuristic.
More complex heuristics require more calculations. Most

games institute some sort of time limit for AI moves. For
example, a chess game AI might have only 10 seconds to
decide its next move. This makes it necessary to strike a
balance between the depth explored and heuristic
complexity.
Alpha-Beta Pruning
Alpha-beta pruning is an optimization of the
minimax algorithm that, on average, reduces the
number of nodes evaluated. In practice, this means
it’s possible to increase the maximum depth explored
without increasing the computation time.
Figure 4.13 shows a game tree simplified by alpha-beta

pruning. Assuming a left-to-right order of evaluation for
siblings, the max player first inspects subtree B. The min
player then sees the leaf with value 5, which means the
min player has a choice between 5 and other values. If
these other values are greater than 5, the min player
obviously selects 5. This means that the upper bound of
subtree B is 5, but the lower bound is negative infinity.
The min player continues and sees the leaf with value 0
and selects this leaf because the min player wants the
minimum possible score.
Figure 4.13 A game tree simplified by alpha-beta

pruning
Control returns to the max player function, which now

knows that subtree B has a value of 0. Next, the max
player inspects subtree C. The min player first sees the
leaf with value -3. As before, this means the upper bound
of subtree C is -3. However, you already know that
subtree B has a value of 0, which is better than -3. This
means that there’s no way subtree C can be better for the
max player than subtree B. Alpha-beta pruning
recognizes this and, as a result, does not inspect any
other children of C.
Alpha-beta pruning adds two additional variables, called
alpha and beta. Alpha is the best score guaranteed for
the max player at the current level or higher. Conversely,
beta is the best score guaranteed for the min player at
the current level or higher. In other words, alpha and
beta are the lower and upper bounds of the score.
Initially, alpha is negative infinity and beta is positive

infinity—the worst possible values for both players.
AlphaBetaDecide, in Listing 4.8, initializes alpha and
beta to these values and then recurses by calling
AlphaBetaMin.
Listing 4.8 AlphaBetaDecide Implementation

const GameState* AlphaBetaDecide(const GameState* root, int maxDepth)
const GameState* choice = nullptr;
// Alpha starts at negative infinity, beta at positive infinity
float beta = std::numeric_limits<float>::infinity();
for (const GameState* child : root->GetPossibleMoves())
float v = AlphaBetaMin(child, maxDepth - 1, maxValue, beta);
if (v > maxValue)
{
maxValue = v;
choice = child;
return choice;
The implementation of AlphaBetaMax, shown in

Listing 4.9, builds on MaxPlayerLimit. If on any
iteration the max value is greater than or equal to beta, it
means the score can be no better than the previous upper
bound. This makes it unnecessary to test the remaining
siblings, and so the function returns. Otherwise, the code
increases the alpha lower bound if the max value is
greater than alpha.
Listing 4.9 AlphaBetaMax Implementation

float AlphaBetaMax(const GameState* node, int depth, float alpha,
float beta)
if (depth == 0 || node->IsTerminal())
{
return node->GetScore();
for (const GameState* child : node->GetPossibleMoves())
maxValue = std::max(maxValue,
AlphaBetaMin(child, depth - 1, alpha, beta));
if (maxValue >= beta)
return maxValue; // Beta prune
alpha = std::max(maxValue, alpha); // Increase lower bound
return maxValue;
Similarly, AlphaBetaMin, shown in Listing 4.10, checks

whether the min value is less than or equal to alpha. In
this case, the score can be no better than the lower
bound, so the function returns. Then the code decreases
the beta upper bound as necessary.
Listing 4.10 AlphaBetaMin Implementation
float AlphaBetaMin(const GameState* node, int depth, float alpha,
float beta)
if (depth == 0 || node->IsTerminal())
return node->GetScore();
float minValue = std::numeric_limits<float>::infinity();
for (const GameState* child : node->GetPossibleMoves())
minValue = std::min(minValue,
AlphaBetaMax(child, depth - 1, alpha, beta));
if (minValue <= alpha)
return minValue; // Alpha prune
beta = std::min(minValue, beta); // Decrease upper bound
}
return minValue;
Note that the order of evaluation for children affects the

number of nodes pruned. This means that even with a
consistent depth limit, different starting states yield
different execution times. This can be problematic if the
AI has a fixed time limit; an incomplete search means
the AI has no idea which move to take. One solution is
iterative deepening, which runs the algorithm
multiple times at increasing depth limits. For example,
first run alpha-beta pruning with a depth limit of three,
which yields some baseline move. Then run with a depth
limit of four, then five, and so on, until time runs out. At
this point, the code returns the move from the previous
iteration. This guarantees that some move is always
available, even when time runs out.
GAME PROJECT
This chapter’s game project, shown in Figure 4.14, is
a tower defense game. In this style of game, the
enemies try to move from the start tile on the left to
an end tile on the right. Initially, the enemies move
in a straight line from left to right. However, the
player can build towers on squares in the grid, even
where the path is, which causes the path to redirect
around these towers as needed. The code is available
in the book’s GitHub repository, in the Chapter04
directory. Open Chapter04-windows.sln on
Windows and Chapter04-mac.xcodeproj on
Mac.
Figure 4.14 Chapter 4 game project
Use the mouse to click on and select tiles. After selecting

a tile, use the B key to build a tower. The enemy airplanes
path around the towers using the A* pathfinding
algorithm. Each new tower built changes the path as
necessary. To ensure that the player can’t fully block in
the enemies, when the player requests to build a tower,
the code first ensures that a path would still exist for the
enemies. If a tower would completely block the path, the
game doesn’t let the player build it.
As a simplification, the Tile class in the game project

contains all the graph information, as well as the scratch
data used by the A* search. The code that creates all the
tiles and initializes the graph is in the constructor of the
Grid class. The Grid class also contains the FindPath
function that runs the actual A* search.
For completeness, the source code for this chapter also

includes the versions of the search and minimax
algorithms covered in the text in a separate Search.cpp
file. It also includes the implementation of AIState and
AIComponent, even though no actors in the game
project use these features.
SUMMARY
Artificial intelligence is a deep topic with many
different sub-areas. Using state machines is an
effective way to give behaviors to AI-controlled
characters in a game. While a switch is the simplest
implementation of a state machine, the state design
pattern adds flexibility by making each state a
separate class.
Pathfinding algorithms find the shortest path between

two points in the game world. First, you formulate a
graph representation for the game world. For a square
grid, this is simple, but other games use path nodes or
navigation meshes. For unweighted graphs, breadth-first
search (BFS) guarantees to find the shortest path if one
exists. But for weighted graphs, you need other
algorithms, such as A* or Dijkstra’s, to find the shortest
path.
For two-player adversarial turn-based games such as
checkers or chess, a game tree represents the sequence of
possible moves from the current game state. The
minimax algorithm assumes that the current player aims
to maximize his or her score, and the opponent aims to
minimize the current player’s score. Alpha-beta pruning
optimizes minimax, though for most games the tree must
have a depth limit.
ADDITIONAL READING
Many resources cover AI techniques. Stuart Russell
and Peter Norvig’s book is a popular AI text that
covers many techniques, though only some are
applicable to games. Mat Buckland’s book, although
dated, covers many useful game AI topics. Steve
Rabin’s Game AI Pro series has many interesting
articles written by different game AI developers.
For navigation meshes, Stephen Pratt’s in-depth web

article covers the steps to generate a navigation mesh
from level geometry. The Recast project provides an
open source implementation of both navigation mesh
generation and pathfinding algorithms.
Buckland, Mat. Programming Game AI by

Example. Plano: Wordware Publishing,
2005.
Mononen, Mikko. “Recast Navigation Mesh
Toolkit.” Accessed July 7, 2017.
https://github.com/recastnavigation.
Pratt, Stephen. “Study: Navigation Mesh
Generation.” Accessed July 7, 2017.
http://critterai.org/projects/nmgen_study/index.html.
Rabin, Steve, Ed. Game AI Pro 3: Collected
Wisdom of Game AI Professionals. Boca
Raton: CRC Press, 2017.
Russell, Stuart, and Peter Norvig. Artificial
Intelligence: A Modern Approach, 3rd
edition. Upper Saddle River: Pearson,
2009.
EXERCISES
The two exercises for this chapter implement
techniques not used in this chapter’s game project.
The first looks at state machines, and the second uses
alpha-beta pruning for a four-in-a-row game.
Exercise 4.1
Given this chapter’s game project code, update either
the Enemy or Tower class (or both!) to use an AI
state machine. First, consider which behaviors the AI
should have and design the state machine graph.
Next, use the provided AIComponent and AIState
base classes to implement these behaviors.
Exercise 4.2
In a four-in-a-row game, players have a vertical grid
of six rows and seven columns. The two players take
turns putting a piece at the top of a column, and then
the piece slides down to the lowest free position in
the column. The game continues until one player gets
four in a row horizontally, vertically, or diagonally.
The starting code in Exercises/4.2 allows the human

player to click to make a move. In the starting code, the
AI randomly selects a move from the set of valid moves.
Modify the AI code to instead use alpha-beta pruning
with a depth cutoff.
CHAPTER 5
OPENGL
This chapter provides an in-depth

introduction on how to use OpenGL for
graphics in games. It covers many topics,
including initializing OpenGL, using
triangles, writing shader programs, using
matrices for transforms, and adding support
for textures. The game project for this
chapter converts the game project from
Chapter 3, “Vectors and Basic Physics,” to
use OpenGL for all its graphics rendering.
INITIALIZING OPENGL
Although the SDL renderer supports 2D graphics, it
does not support 3D. Thus, to switch to 3D, which is
used in every subsequent chapter in this book, you
need to switch from SDL 2D graphics to a different
library that supports both 2D and 3D graphics.
This book uses the OpenGL library. OpenGL is an

industry-standard library for cross-platform 2D/3D
graphics that’s been around for 25 years. Unsurprisingly,
the library has been around so long that it has evolved in
many ways over the years. The set of functions the
original version of OpenGL used is very different from
the set in modern OpenGL. This book uses functions
defined up to and including OpenGL 3.3.
warning
OLDER VERSIONS OF OPENGL ARE VERY DIFFERENT: Be careful when
consulting any online OpenGL references, as many refer to older versions of
OpenGL.
The goal of this chapter is to convert the game project

from Chapter 3 from SDL graphics to OpenGL graphics.
You need to take a lot of steps to get there. This section
walks through the steps of configuring and initializing
OpenGL and a helper library called GLEW.
Setting Up the OpenGL Window

To use OpenGL, you must drop usage of the
SDL_Renderer from the earlier chapters. You
therefore need to remove all references to
SDL_Renderer, including the mRenderer variable
in Game, the call to SDL_CreateRenderer, and any
calls to the SDL functions in GenerateOuput. This
also means that the SpriteComponent code (which
relies on SDL_Renderer) won’t work without
changes. For now, all the code in
Game::GenerateOutput is commented out until
OpenGL is up and running.
In SDL, when you create a window, you can request a

window for OpenGL usage by passing in the
SDL_WINDOW_OPENGL flag as the final parameter of the
SDL_CreateWindow call:
mWindow = SDL_CreateWindow("Game Programming in C++ (Chapter 5)", 100, 100,
1024, 768, SDL_WINDOW_OPENGL);
Prior to creating the OpenGL window, you can request

attributes such as the version of OpenGL, the color
depth, and several other parameters. To configure these
parameters, you use the SDL_GL_SetAttribute
function:
// Set OpenGL window's attributes (use prior to creating the window)

// Returns 0 if successful, otherwise a negative value
SDL_GL_SetAttribute(
SDL_GLattr attr, // Attribute to set
int value // Value for this attribute
);
There are several different attributes in the SDL_GLattr

enum, but this chapter uses only some of them. To set
the attributes, you add the code in Listing 5.1 prior to the
call of SDL_CreateWindow inside
Game::Initialize. This code sets several attributes.
First, it requests the core OpenGL profile.
note
There are three main profiles supported by OpenGL: core, compatibility, and
ES. The core profile is the recommended default profile for a desktop
environment. The only difference between the core and compatibility profiles
is that the compatibility profile allows the program to call OpenGL functions
that are deprecated (no longer intended for use). The OpenGL ES profile is
for mobile development.
Listing 5.1 Requesting OpenGL Attributes

// Use the core OpenGL profile
SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK,
SDL_GL_CONTEXT_PROFILE_CORE);
// Specify version 3.3
SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 3);
SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 3);
// Request a color buffer with 8-bits per RGBA channel
SDL_GL_SetAttribute(SDL_GL_RED_SIZE, 8);
SDL_GL_SetAttribute(SDL_GL_GREEN_SIZE, 8);
SDL_GL_SetAttribute(SDL_GL_BLUE_SIZE, 8);
SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
// Enable double buffering
SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
// Force OpenGL to use hardware acceleration
SDL_GL_SetAttribute(SDL_GL_ACCELERATED_VISUAL, 1);
The next two attributes request OpenGL version 3.3.

Although there are newer versions of OpenGL, the 3.3
version supports all the required features for this book
and has a feature set closely aligned with the ES profile.
Thus, most of the code in this book should also work on
current mobile devices.
The next attributes specify the bit depth of each channel.

In this case, the program requests 8 bits per RGBA
channel, for a total of 32 bits per pixel. The second-to-
last attribute asks to enable double buffering. The final
attribute asks to run OpenGL with hardware
acceleration. This means that the OpenGL rendering will
run on graphics hardware (a GPU).
The OpenGL Context and Initializing

GLEW
Once the OpenGL attributes are set and you’ve
created the window, the next step is to create an
OpenGL context. Think of a context as the “world”
of OpenGL that contains every item that OpenGL
knows about, such as the color buffer, any images or
models loaded, and any other OpenGL objects.
(While it is possible to have multiple contexts in one
OpenGL program, this book sticks to one.)
To create the context, first add the following member

variable to Game:
SDL_GLContext mContext;
Next, immediately after creating the SDL window with

SDL_CreateWindow, add the following line of code,
which creates an OpenGL context and saves it in the
member variable:
mContext = SDL_GL_CreateContext(mWindow);
As with creating and deleting the window, you need to

delete the OpenGL context in the destructor. To do this,
add the following line of code to Game::Shutdown,
right before the call to SDL_DeleteWindow:
SDL_GL_DeleteContext(mContext);
Although the program now creates an OpenGL context,

there is one final hurdle you must pass to gain access to
the full set of OpenGL 3.3 features. OpenGL supports
backward compatibility with an extension system.
Normally, you must query any extensions you want
manually, which is tedious. To simplify this process, you
can use an open source library called the OpenGL
Extension Wrangler Library (GLEW). With one simple
function call, GLEW automatically initializes all
extension functions supported by the current OpenGL
context’s version. So in this case, GLEW initializes all
extension functions supported by OpenGL 3.3 and
earlier.
To initialize GLEW, you add the following code

immediately after creating the OpenGL context:
// Initialize GLEW
glewExperimental = GL_TRUE;
if (glewInit() != GLEW_OK)
{
SDL_Log("Failed to initialize GLEW.");
return false;
// On some platforms, GLEW will emit a benign error code,
// so clear it
glGetError();
The glewExperimental line prevents an initialization

error that may occur when using the core context on
some platforms. Furthermore, because some platforms
emit a benign error code when initializing GLEW, the
call to glGetError clears this error code.
note
Some old PC machines with integrated graphics (from 2012 or earlier) may
have issues running OpenGL version 3.3. In this case, you can try two
things: updating to newer graphics drivers or requesting OpenGL version 3.1.
Rendering a Frame
You now need to convert the clear, draw scene, and
swap buffers process in Game::GenerateOutput
to use OpenGL functions:
// Set the clear color to gray
glClearColor(0.86f, 0.86f, 0.86f, 1.0f);

// Clear the color buffer
glClear(GL_COLOR_BUFFER_BIT);
// TODO: Draw the scene
// Swap the buffers, which also displays the scene
SDL_GL_SwapWindow(mWindow);
This code first sets the clear color to 86% red, 86% green,
86% blue, and 100% alpha, which yields a gray color. The
glClear call with the GL_COLOR_BUFFER_BIT
parameter clears the color buffer to the specified color.
Finally, the SDL_GL_SwapWindow call swaps the front
buffer and back buffer. At this point, running the game
yields a gray screen because you aren’t drawing the
SpriteComponents yet.
TRIANGLE BASICS
The graphical needs of 2D and 3D games couldn’t
seem more different. As discussed in Chapter 2,
“Game Objects and 2D Graphics,” most 2D games
use sprites for their 2D characters. On the other
hand, a 3D game features a simulated 3D
environment that you somehow flatten into a 2D
image that you show onscreen.
Early 2D games could simply copy sprite images into the
desired locations of the color buffer. This process, called
blitting, was efficient on sprite-based consoles such as
the Nintendo Entertainment System (NES). However,
modern graphical hardware is inefficient at blitting but is
very efficient at polygonal rendering. Because of this,
nearly all modern games, whether 2D or 3D, ultimately
use polygons for their graphical needs.
Why Polygons?
There are many ways a computer could simulate a
3D environment. Polygons are popular in games for a
multitude of reasons. Compared to other 3D graphics
techniques, polygons do not require as many
calculations at runtime. Furthermore, polygons are
scalable: A game running on less-powerful hardware
could simply use 3D models with fewer polygons.
And, importantly, you can represent most 3D objects
with polygons.
Triangles are the polygon of choice for most games.

Triangles are the simplest polygon, and you need only
three points (or vertices) to create a triangle.
Furthermore, a triangle can only lie on a single plane. In
other words, the three points of a triangle must be
coplanar. Finally, triangles tessellate easily, meaning
it’s relatively simple to break any complex 3D object into
many triangles. The remainder of this chapter talks
about triangles, but the techniques discussed here also
work for other polygons (such as quads), provided that
they maintain the coplanar property.
2D games use triangles to represent sprites by drawing a

rectangle and filling in the rectangle with colors from an
image file. We discuss this in much greater detail later in
the chapter.
Normalized Device Coordinates

To draw a triangle, you must specify the coordinates
of its three vertices. Recall that in SDL, the top-left
corner of the screen is (0, 0), positive x is to the right,
and positive y is down. More generally, a
coordinate space specifies where the origin is and
in which direction its coordinates increase. The
basis vectors of the coordinate space are the
direction in which the coordinates increase.
An example of a coordinate space from basic geometry is

a Cartesian coordinate system (see Figure 5.1). In a
2D Cartesian coordinate system, the origin (0, 0) has a
specific point (usually the center), positive x is to the
right, and positive y is up.
Figure 5.1 A point drawn at (2, 3) in a Cartesian
coordinate system
Normalized device coordinates (NDC) is the

default coordinate system used with OpenGL. Given an
OpenGL window, the center of the window is the origin
in normalized device coordinates. Furthermore, the
bottom-left corner is (–1, –1), and the top-right corner is
(1, 1). This is regardless of the width and height of the
window (hence normalized device coordinates).
Internally, the graphics hardware then converts these
NDC into the corresponding pixels in the window.
For example, to draw a square with sides of unit length

in the center of the window, you need two triangles. The
first triangle has the vertices (–0.5, 0.5), (0.5, 0.5), and
(0.5, –0.5), and the second triangle has the vertices (0.5,
–0.5), (–0.5, –0.5), and (–0.5, 0.5). Figure 5.2 illustrates
this square. Keep in mind that if the length and width of
the window are not uniform, a square in normalized
device coordinates will not look like a square onscreen.
Figure 5.2 A square drawn in 2D normalized device

coordinates
In 3D, the z component of normalized device coordinates

also ranges from [–1, 1], with a positive z value going into
the screen. For now, we stick with a z value of zero. We’ll
explore 3D in much greater detail in Chapter 6, “3D
Graphics.”
Vertex and Index Buffers

Suppose you have a 3D model comprised of many
triangles. You need some way to store the vertices of
these triangles in memory. The simplest approach is
to directly store the coordinates of each triangle in a
contiguous array or buffer. For example, assuming
3D coordinates, the following array contains the
vertices of the two triangles shown in Figure 5.2:
float vertices[] = {
-0.5f, 0.5f, 0.0f,
0.5f, 0.5f, 0.0f,
0.5f, -0.5f, 0.0f,
0.5f, -0.5f, 0.0f,
-0.5f, -0.5f, 0.0f,
-0.5f, 0.5f, 0.0f,
};
Even in this simple example, the array of vertices has

some duplicate data. Specifically, the coordinates (–0.5,
0.5, 0.0) and (0.5, –0.5, 0.0) appear twice. If there were
a way to remove these duplicates, you would cut the
number of values stored in the buffer by 33%. Rather
than having 12 values, you would have only 8. Assuming
single-precision floats that use 4 bytes each, you’d save
24 bytes of memory by removing the duplicates. This
might seem insignificant, but imagine a much larger
model with 20,000 triangles. In this case, the amount of
memory wasted due to duplicate coordinates would be
high.
The solution to this issue has two parts. First, you create
a vertex buffer that contains only the unique
coordinates used by the 3D geometry. Then, to specify
the vertices of each triangle, you index into this vertex
buffer (much like indexing into an array). The aptly
named index buffer contains the indices for each
individual triangle, in sets of three. For this example’s
sample square, you’d need the following vertex and index
buffers:
float vertexBuffer[] = {
-0.5f, 0.5f, 0.0f, // vertex 0
0.5f, 0.5f, 0.0f, // vertex 1
0.5f, -0.5f, 0.0f, // vertex 2
-0.5f, -0.5f, 0.0f // vertex 3
};
unsigned short indexBuffer[] = {
0, 1, 2,
2, 3, 0
};
For example, the first triangle has the vertices 0, 1, and 2,

which corresponds to the coordinates (–0.5, 0.5, 0.0),
(0.5, 0.5, 0.0), and (0.5, –0.5, 0.0). Keep in mind that
the index is the vertex number, not the floating-point
element (for example, vertex 1 instead of “index 2” of the
array). Also note that this code uses an unsigned short
(typically 16 bits) for the index buffer, which reduces the
memory footprint of the index buffer. You can use
smaller bit size integers to save memory in the index
buffer.
In this example, the vertex/index buffer combination

uses 12 × 4 + 6 × 2, or 60 total bytes. On the other hand,
if you just used the original vertices, you’d need 72 bytes.
While the savings in this example is only 20%, a more
complex model would save much more memory by using
the vertex/index buffer combination.
To use the vertex and index buffers, you must let

OpenGL know about them. OpenGL uses a vertex
array object to encapsulate a vertex buffer, an index
buffer, and the vertex layout. The vertex layout
specifies what data you store for each vertex in the
model. For now, assume the vertex layout is a 3D
position (you can just use a z component of 0.0f if you
want something 2D). Later in this chapter you’ll add
other data to each vertex.
Because any model needs a vertex array object, it makes

sense to encapsulate its behavior in a VertexArray
class. Listing 5.2 shows the declaration of this class.
Listing 5.2 VertexArray Declaration

class VertexArray
{
public:
VertexArray(const float* verts, unsigned int numVerts,
const unsigned int* indices, unsigned int numIndices);
~VertexArray();
// Activate this vertex array (so we can draw it)
void SetActive();
unsigned int GetNumIndices() const { return mNumIndices; }
unsigned int GetNumVerts() const { return mNumVerts; }
private:
// How many vertices in the vertex buffer?
unsigned int mNumVerts;
// How many indices in the index buffer
unsigned int mNumIndices;
// OpenGL ID of the vertex buffer
unsigned int mVertexBuffer;
// OpenGL ID of the index buffer
unsigned int mIndexBuffer;
// OpenGL ID of the vertex array object

unsigned int mVertexArray;
};
The constructor for VertexArray takes in pointers to

the vertex and index buffer arrays so that it can hand off
the data to OpenGL (which will ultimately load the data
on the graphics hardware). Note that the member data
contains several unsigned integers for the vertex buffer,
index buffer, and vertex array object. This is because
OpenGL does not return pointers to objects that it
creates. Instead, you merely get back an integral ID
number. Keep in mind that the ID numbers are not
unique across different types of objects. It’s therefore
very possible to have an ID of 1 for both the vertex and
index buffers because OpenGL considers them different
types of objects.
The implementation of the VertexArray constructor is

complex. First, create the vertex array object and store its
ID in the mVertexArray member variable:
glGenVertexArrays(1, &mVertexArray);
glBindVertexArray(mVertexArray);
Once you have a vertex array object, you can create a

vertex buffer:

glGenBuffers(1, &mVertexBuffer);
glBindBuffer(GL_ARRAY_BUFFER, mVertexBuffer);
The GL_ARRAY_BUFFER parameter to glBindBuffer

means that you intend to use the buffer as a vertex
buffer.
Once you have a vertex buffer, you need to copy the

verts data passed into the VertexArray constructor
into this vertex buffer. To copy the data, use
glBufferData, which takes several parameters:
glBufferData(
GL_ARRAY_BUFFER, // The active buffer type to write to
numVerts * 3 * sizeof(float), // Number of bytes to copy
verts, // Source to copy from (pointer)
GL_STATIC_DRAW // How will we use this data?
);
Note that you don’t pass in the object ID to

glBufferData; instead, you specify a currently bound
buffer type to write to. In this case, GL_ARRAY_BUFFER
means use the vertex buffer just created.
For the second parameter, you pass in the number of

bytes, which is the amount of data for each vertex
multiplied by the number of vertices. For now, you can
assume that each vertex contains three floats for (x, y, z).
The usage parameter specifies how you want to use the

buffer data. A GL_STATIC_DRAW usage means you only
want to load the data once and use it frequently for
drawing.
Next, create an index buffer. This is very similar to

creating the vertex buffer, except you instead specify the
GL_ELEMENT_ARRAY_BUFFER type, which corresponds
to an index buffer:
glGenBuffers(1, &mIndexBuffer);
glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, mIndexBuffer);
Then copy the indices data into the index buffer:
glBufferData(
GL_ELEMENT_ARRAY_BUFFER, // Index buffer
numIndices * sizeof(unsigned int), // Size of data
indices, GL_STATIC_DRAW);
Note that the type here is

GL_ELEMENT_ARRAY_BUFFER, and the size is the
number of indices multiplied by an unsigned int because
that’s the type used for indices here.
Finally, you must specify a vertex layout, also called the

vertex attributes. As mentioned earlier, the current
layout is a position with three float values.
To enable the first vertex attribute (attribute 0), use

glEnableVertexAttribArray:
glEnableVertexAttribArray(0);
You then use glVertexAttribPointer to specify the

size, type, and format of the attribute:
glVertexAttribPointer(
0, // Attribute index (0 for first one)
3, // Number of components (3 in this case)
GL_FLOAT, // Type of the components
GL_FALSE, // (Only used for integral types)
sizeof(float) * 3, // Stride (usually size of each vertex)
0 // Offset from start of vertex to this attribute
);
The first two parameters are 0 and 3 because the

position is attribute 0 of the vertex, and there are three
components (x, y, z). Because each component is a float,
you specify the GL_FLOAT type. The fourth parameter is
only relevant for integral types, so here you set it to
GL_FALSE. Finally, the stride is the byte offset between
consecutive vertices’ attributes. But assuming you don’t
have padding in the vertex buffer (which you usually
don’t), the stride is just the size of the vertex. Finally, the
offset is 0 because this is the only attribute. For
additional attributes, you have to pass in a nonzero value
for the offset.
The VertexArray’s destructor destroys the vertex

buffer, index buffer, and vertex array object:
VertexArray::~VertexArray()
glDeleteBuffers(1, &mVertexBuffer);
glDeleteBuffers(1, &mIndexBuffer);
glDeleteVertexArrays(1, &mVertexArray);
Finally, the SetActive function calls

glBindVertexArray, which just specifies which vertex
array you’re currently using:.
void VertexArray::SetActive()
glBindVertexArray(mVertexArray);
The following code in Game::InitSpriteVerts

allocates an instance of VertexArray and saves it in a
member variable of Game called mSpriteVerts:
mSpriteVerts = new VertexArray(vertexBuffer, 4, indexBuffer, 6);
The vertex and index buffer variables here are the arrays
for the sprite quad. In this case, there are 4 vertices in
the vertex buffer and 6 indices in the index buffer
(corresponding to the 2 triangles in the quad). You will
use this member variable later in this chapter to draw
sprites, as all sprites will ultimately use the same
vertices.
SHADERS
In a modern graphics pipeline, you don’t simply feed
in the vertex/index buffers and have triangles draw.
Instead, you specify how you want to draw the
vertices. For example, should the triangles be a fixed
color, or should they use a color from a texture? Do
you want to perform lighting calculations for every
pixel you draw?
Because there are many techniques you may want to use

to display the scene, there is no truly one-size-fits-all
method. To allow for more customization, graphics APIs
including OpenGL support shader programs—small
programs that execute on the graphics hardware to
perform specific tasks. Importantly, shaders are separate
programs, with their own separate main functions.
note
Shader programs do not use the C++ programming language. This book
uses the GLSL programming language for shader programs. Although GLSL
superficially looks like C, there are many semantics specific to GLSL. Rather
than present all the details of GLSL at once, this book introduces the
concepts as needed.
Because shaders are separate programs, you write them

in separate files. Then in your C++ code, you need to tell
OpenGL when to compile and load these shader
programs and specify what you want OpenGL to use
these shader programs for.
Although you can use several different types of shaders

in games, this book focuses on the two most important
ones: the vertex shader and the fragment (or pixel)
shader.
Vertex Shaders
A vertex shader program runs once for every
vertex of every triangle drawn. The vertex shader
receives the vertex attribute data as an input. The
vertex shader can then modify these vertex attributes
as it sees fit. While it may seem unclear why you’d
want to modify vertex attributes, it’ll become more
apparent as this chapter continues.
Given that triangles have three vertices, you can think of

a vertex shader as running three times per triangle.
However, if you use vertex and index buffers, then you
will invoke the vertex shader less often because some
triangles share vertices. This is an additional advantage
of using a vertex and index buffer instead just a vertex
buffer. Note that if you draw the same model multiple
times per frame, the vertex shader calls for each time you
draw it are independent of each other.
Fragment Shaders
After the vertices of a triangle have gone through the
vertex shader, OpenGL must determine which pixels
in the color buffer correspond to the triangle. This
process of converting the triangle into pixels is
rasterization. There are many different
rasterization algorithms, but today’s graphics
hardware does rasterization for us.
The job of a fragment shader (or pixel shader) is to

determine the color of each pixel, and so the fragment
shader program executes at least once for every pixel.
This color may take into account properties of the
surface, such as textures, colors, and materials. If the
scene has any lighting, the fragment shader might also
do lighting calculations. Because there are so many
potential calculations, the average 3D game has a lot
more code in the fragment shader than in the vertex
shader.
Writing Basic Shaders

Although you could load in the shader programs
from hard-coded strings in C++ code, it’s much
better to put them in separate files. This book uses
the .vert extension for vertex shader files and the
.frag extension for fragment shader files.
Because these source files are in a different programming

language, they are in the Shaders subdirectory for the
chapter. For example, Chapter05/Shaders contains
the source files for the shaders in this chapter.
The Basic.vert File

Basic.vert contains the vertex shader code.
Remember that this code is not C++ code.
Every GLSL shader file first must specify the version of

the GLSL programming language used. The following
line represents the version of GLSL corresponding to
OpenGL 3.3:
#version 330
Next, because this is a vertex shader, you must specify

the vertex attributes for each vertex. These attributes
should match the attributes of the vertex array object
created earlier, and are the input to the vertex shader.
However, in GLSL the main function does not receive
any parameters. Instead, the shader inputs look like
global variables, marked with a special in keyword.
For now, you only have one input variable—the 3D

position. The following line declares this input variable:
in vec3 inPosition;
The type of inPosition variable is vec3, which

corresponds to a vector of three floating-point values.
This will contain the x, y, and z components
corresponding to the vertex’s position. You can access
each component of the vec3 via dot syntax; for example,
inPosition.x accesses the x component of the vector.
As with a C/C++ program, a shader program has a main

function as its entry point:
void main()
// TODO: Shader code goes here
Note that the main function here returns void. GLSL

also uses global variables to define the outputs of the
shader. In this case, you’ll use a built-in variable called
gl_Position to store the vertex position output of the
shader.
For now, the vertex shader directly copies the vertex

position from inPosition to gl_Position. However,
gl_Position expects four components: the normal (x,
y, z) coordinates plus a fourth component called the w
component. We’ll look at what this w represents later
in this chapter. For now, assume that w is always 1.0. To
convert inPosition from vec3 to vec4, you can use
the following syntax:
gl_Position = vec4(inPosition, 1.0);
Listing 5.3 shows the complete Basic.vert code, which

simply copies along the vertex position without any
modification.
Listing 5.3 Basic.vert Code

// Request GLSL 3.3
#version 330
// Any vertex attributes go here

// For now, just a position.
in vec3 inPosition;
void main()
// Directly pass along inPosition to gl_Position
gl_Position = vec4(inPosition, 1.0);
The Basic.frag File

The job of the fragment shader is to compute an
output color for the current pixel. For Basic.frag,
you’ll hard-code a blue output color for all pixels.
As with the vertex shader, the fragment shader always

begin with a #version line. Next, you declare a global
variable to store the output color, using the out variable
specifier:
out vec4 outColor;
The outColor variable is a vec4 corresponding to the

four components of the RGBA color buffer.
Next, you declare the entry point of the fragment shader

program. Inside this function, you set outColor to the
desired color for the pixel. The RGBA value of blue is
(0.0, 0.0, 1.0, 1.0), which means you use the following
assignment:
outColor = vec4(0.0, 0.0, 1.0, 1.0);
Listing 5.4 gives the full source code for Basic.frag.
Listing 5.4 Basic.frag Code

// Request GLSL 3.3
#version 330
// This is output color to the color buffer
out vec4 outColor;
void main()
// Set to blue
outColor = vec4(0.0, 0.0, 1.0, 1.0);
}
Loading Shaders
Once you have the separate shader files written, you
must load in these shaders in the game’s C++ code to
let OpenGL know about them. At a high level, you
need to follow these steps:
1. Load and compile the vertex shader.

2. Load and compile the fragment shader.
3. Link the two shaders together into a “shader
program.”
There are many steps to loading a shader, so it is a good

idea to declare a separate Shader class, as in Listing 5.5.
Listing 5.5 Initial Shader Declaration

class Shader
public:
Shader();
~Shader();
// Load the vertex/fragment shaders with the given names
bool Load(const std::string& vertName,
const std::string& fragName);
// Set this as the active shader program

void SetActive();
private:
// Tries to compile the specified shader
bool CompileShader(const std::string& fileName,
GLenum shaderType, GLuint& outShader);
// Tests whether shader compiled successfully
bool IsCompiled(GLuint shader);
// Tests whether vertex/fragment programs link
bool IsValidProgram();
// Store the shader object IDs
GLuint mVertexShader;
GLuint mFragShader;
GLuint mShaderProgram;
};
Note how the member variables here correspond to

shader object IDs. They have object IDs much like the
vertex and index buffers. (GLuint is simply OpenGL’s
version of unsigned int.)
You declare CompileShader, IsCompiled, and

IsValidProgram in the private section because they
are helper functions used by Load. This reduces the code
duplication in Load.
The Function
CompileShader takes three parameters: the name
of the shader file to compile, the type of shader, and
a reference parameter to store the ID of the shader.
The return value is a bool that denotes whether
CompileShader succeeded.

CompileShader, which has several steps. First, create
an ifstream to load in the file. Next, use a string
stream to load the entire contents of the file into a single
string, contents, and get the C-style string pointer with
the c_str function.
Next, the glCreateShader function call creates an

OpenGL shader object corresponding to the shader (and
saves this ID in outShader). The shaderType
parameter can be GL_VERTEX_SHADER,
GL_FRAGMENT_SHADER, or a few other shader types.
The glShaderSource call specifies the string

containing the shader source code, and
glCompileShader compiles the code. You then use the
IsCompiled helper function (implemented in a
moment) to validate that the shader compiles.
In the event of any errors, including being unable to load

the shader file or failing to compile it, CompileShader
outputs an error message and returns false.
Listing 5.6 Shader::CompileShader
Implementation
bool Shader::CompileShader(const std::string& fileName,
GLenum shaderType,
GLuint& outShader)
// Open file
std::ifstream shaderFile(fileName);
if (shaderFile.is_open())
// Read all the text into a string
std::stringstream sstream;
sstream << shaderFile.rdbuf();
std::string contents = sstream.str();
const char* contentsChar = contents.c_str();
// Create a shader of the specified type
outShader = glCreateShader(shaderType);
// Set the source characters and try to compile
glShaderSource(outShader, 1, &(contentsChar), nullptr);

glCompileShader(outShader);
if (!IsCompiled(outShader))
SDL_Log("Failed to compile shader %s", fileName.c_str());
return false;
else
SDL_Log("Shader file not found: %s", fileName.c_str());
return false;
return true;
The Function
The IsCompiled function, shown in Listing 5.7,
validates whether a shader object compiled, and if it
didn’t, it outputs the compilation error message. This
way, you can get some information about why a
shader fails to compile.
Listing 5.7 Shader::IsCompiled Implementation
bool Shader::IsCompiled(GLuint shader)
GLint status;
// Query the compile status
glGetShaderiv(shader, GL_COMPILE_STATUS, &status);
if (status != GL_TRUE)
char buffer[512];
memset(buffer, 0, 512);
glGetShaderInfoLog(shader, 511, nullptr, buffer);
SDL_Log("GLSL Compile Failed:\n%s", buffer);
return false;
return true;
The glGetShaderiv function queries the compilation

status, which the function returns as an integral status
code. If this status is not GL_TRUE, there was an error. In
the event of an error, you can get a human-readable
compile error message with glGetShaderInfoLog.
The Function
The Load function in Listing 5.8 takes in the
filenames of both the vertex and fragment shaders
and then tries to compile and link these shaders
together.
As shown in Listing 5.8, you compile both the vertex and

fragment shaders using CompileShader and then save
their objects IDs in mVertexShader and
mFragShader, respectively. If either of the
CompileShader calls fail, Load returns false.
Listing 5.8 Shader::Load Implementation

bool Shader::Load(const std::string& vertName,
const std::string& fragName)
// Compile vertex and fragment shaders
if (!CompileShader(vertName, GL_VERTEX_SHADER, mVertexShader) ||
!CompileShader(fragName, GL_FRAGMENT_SHADER, mFragShader))
return false;
}
// Now create a shader program that
// links together the vertex/frag shaders
mShaderProgram = glCreateProgram();
glAttachShader(mShaderProgram, mVertexShader);
glAttachShader(mShaderProgram, mFragShader);
glLinkProgram(mShaderProgram);
// Verify that the program linked successfully
if (!IsValidProgram())
return false;
return true;
After you’ve compiled both the fragment and vertex

shader, you link them together in a third object, called a
shader program. When it’s time to draw an object,
OpenGL uses the currently active shader program to
render the triangles.
You create a shader program with glCreateProgram,

which returns the object ID to the new shader program.
Next, use glAttachShader to add the vertex and
fragment shaders to the combined shader program. Then
use glLinkProgram to link together all attached
shaders into the final shader program.
As with shader compilation, figuring out whether the link

was successful requires additional function calls, which
you can place in the IsValidProgram helper function.
The Function
The code for IsValidProgram is very similar to the
code for IsCompiled. There are only two
differences. First, instead of calling
glGetShaderiv, call glGetProgramiv:
glGetProgramiv(mShaderProgram, GL_LINK_STATUS, &status);
Next, instead of calling glGetShaderInfoLog, call

glGetProgramInfoLog:
glGetProgramInfoLog(mShaderProgram, 511, nullptr, buffer);
The SetActive Function

The SetActive function sets a shader program as
the active one:
void Shader::SetActive()
glUseProgram(mShaderProgram);
OpenGL uses the active shader when drawing triangles.
The Unload Function

The Unload function simply deletes the shader
program, the vertex shader, and the pixel shader:
void Shader::Unload()
glDeleteProgram(mShaderProgram);
glDeleteShader(mVertexShader);
glDeleteShader(mFragShader);
Adding a Shader to the Game

With the Shader class, you can now add a Shader
pointer as a member variable to Game:
class Shader* mSpriteShader;

This variable is called mSpriteShader because,
ultimately, you’ll use it to draw sprites. The
LoadShaders function loads in the shader files and sets
the shader as active:
bool Game::LoadShaders()
mSpriteShader = new Shader();
if (!mSpriteShader->Load("Shaders/Basic.vert", "Shaders/Basic.frag"))
return false;
mSpriteShader->SetActive();
You call LoadShaders in Game::Initialize

immediately after finishing initialization of OpenGL and
GLEW (and before you create the mSpriteVerts vertex
array object).
After you’ve created simple vertex and pixel shaders and

loaded in triangles, you can finally try to draw some
triangles.
Drawing Triangles
As mentioned earlier, you can draw sprites with
triangles by drawing rectangles onscreen. You’ve
already loaded in the unit square vertices and a basic
shader that can draw blue pixels. As before, you want
to draw sprites in the Draw function in
SpriteComponent.
First, you change the declaration of

SpriteComponent::Draw so that it takes in Shader*
instead of SDL_Renderer*. Next, draw a quad with a
call to glDrawElements:
void SpriteComponent::Draw(Shader* shader)
glDrawElements(
GL_TRIANGLES, // Type of polygon/primitive to draw
6, // Number of indices in index buffer
GL_UNSIGNED_INT, // Type of each index
nullptr // Usually nullptr
);
The first parameter to glDrawElements specifies the

type of element you’re drawing (in this case, triangles).
The second parameter is the number of indices in the
index buffer; in this case, because the index buffer for the
unit square has six elements, you pass in 6 as the
parameter. The third parameter is the type of each index,
established earlier as unsigned int. The last
parameter is nullptr.
The glDrawElements call requires both an active

vertex array object and an active shader. On every frame,
you need to activate both the sprite vertex array object
and shader before drawing any SpriteComponents.
You do this in the Game::GenerateOutput function,
as shown in Listing 5.9. Once you’ve set the shader and
vertex array as active, you call Draw once for each sprite
in the scene.
Listing 5.9 Game::GenerateOutput Attempting to

Draw Sprites
void Game::GenerateOutput()
// Set the clear color to gray
// Clear the color buffer
glClear(GL_COLOR_BUFFER_BIT);
// Set sprite shader and vertex array objects active
mSpriteShader ->SetActive();
mSpriteVerts ->SetActive();
// Draw all sprites
for (auto sprite : mSprites)
sprite ->Draw(mSpriteShader);
// Swap the buffers
return true;
What happens when you run this code now? Well, first,
the fragment shader only writes out a blue color. So it’s
reasonable to expect that you’d see blue squares for each
SpriteComponent. However, there’s another issue: For
every sprite, you use the same sprite verts. These sprite
verts define a unit square in normalized device
coordinates. This means that for every
SpriteComponent, you merely draw the same unit
square in NDC. Thus, if you run the game right now,
you’ll see only a gray background and a rectangle, as in
Figure 5.3.
Figure 5.3 Drawing many NDC unit squares (even

though it looks like one rectangle)
It may seem like the solution is to define different vertex

arrays for each sprite. However, it turns out that with
only this one vertex array, you can draw whichever
sprites you want to. The key is to take advantage the
vertex shader’s ability to transform vertex attributes.
TRANSFORMATION BASICS
Suppose a game has 10 asteroids moving around.
You could represent these 10 asteroids individually
with different vertex array objects. However, you
need these asteroids to show up in different locations
onscreen. This means the triangles you draw for each
asteroid need different normalized device
coordinates.
A naïve idea is to create 10 different vertex buffers, 1 for

each of the 10 asteroids, and recompute the vertex
positions in these vertex buffers as needed. But this is
wasteful both in terms of memory usage and in terms of
computation. Changing vertices in vertex buffers and
resubmitting them to OpenGL is not efficient.
Instead, think of a sprite in an abstract sense. Every

sprite is ultimately just a rectangle. Different sprites may
have different locations on the screen, different sizes, or
different rotations, but they’re still rectangles.
Thinking of it this way, a more efficient solution is to

have a single vertex buffer for the rectangle and just
reuse it. Every time you draw the rectangle, you may
have a position offset, scale, or rotation. But given the
NDC unit square, you can change, or transform, it such
that it is an arbitrary rectangle with an arbitrary position,
scale, and/or orientation.
This same concept of reusing a single vertex buffer for a

type of object also extends to 3D. For example, a game
taking place in the forest might have hundreds of trees,
many of which are only slight variations of each other.
It’s inefficient to have a separate vertex buffer for every
single instance of the same tree. Instead, you could
create a single tree vertex buffer, and the game could
draw many instances of this same tree with some
variation in position, scale, and orientation.
Object Space
When you create a 3D object (such as in a 3D
modeling program), you generally don’t express
vertex positions in normalized device coordinates.
Instead, the positions are relative to an arbitrary
origin of the object itself. This origin is often in the
center of the object, but it does not have to be. This
coordinate space relative to the object itself is object
space, or model space.
As discussed earlier in this chapter, defining a coordinate

space requires knowing both the origin of the coordinate
space and the direction in which the various components
increase (the basis vectors). For example, some 3D
modeling programs use +y as up, whereas others use +z
as up. These different basis vectors define different
object spaces for the objects. Figure 5.4 illustrates a 2D
square where the center of the square is its object space
origin, +y moves up, and +x moves right.
Figure 5.4 A quad relative to its object space origin
Now imagine a game that takes place in an office

building. You’d need models for computer monitors,
keyboards, desks, office chairs, and so on. You’d create
each of these individual models in its own object space,
which means each object’s vertex positions are relative to
that model’s unique object space origin.
At runtime, you load each unique model into its own

vertex array object (VAO). For example, you might have
a VAO for the monitor, one for the keyboard, and so on.
When it’s time to render the scene, each vertex of each
object you draw goes to the vertex shader. If you just
directly passed along the vertex positions, as in
Basic.vert, then you’re saying that these vertex
positions are in normalized device coordinates.
This is a problem because the coordinates for the models

are not in NDC but instead are relative to each object’s
unique object space. Passing through the vertex positions
as is would yield garbage output.
World Space
To solve the problem with different objects having
different object space coordinates, you first define a
coordinate space for the game world itself. This
coordinate space, called world space, has its own
origin and basis vectors. For the game in the office
building, the origin of world space might be in the
center of the building on the ground floor.
Much as an office planner might place the desks and

chairs at different positions and orientations in the
office, you can think of the objects in the game as having
arbitrary positions, scales, or orientations relative to the
world space origin. For example, if there are five
instances of the same desk placed in the office, each of
these instances needs information describing how the
object appears in world space.
When you draw each instance of the desk, you use the
same vertex array object for each desk. However, each
instance now needs some additional information,
specifying how you want to transform the object space
coordinates into world space. You can send this extra
data to the vertex shader when drawing an instance,
which allows the vertex shader to adjust the vertex
positions as needed. Of course, the graphics hardware
ultimately needs the coordinates in NDC to draw them,
so you still have an additional step after transforming the
vertices into world space. For now, let’s look at how to
transform vertices from their object space into world
space.
Transforming to World Space

When transforming between coordinate spaces, you
need to know whether the basis vectors between the
two coordinate spaces are the same. For example,
consider the point (0, 5) in object space. If you define
object space with +y as up, this means that the point
(0, 5) is five units “above” the origin. However, if you
choose to define world space such that +y is to the
right, (0, 5) is instead five units to the right.
For now, assume that the basis vectors in object and

world space are the same. Because the game currently is
2D, you can assume that +y is up and +x is to the right
for both object space and world space.
note
The 2D coordinate system used here is different from the SDL coordinate
system where +y is down! This means the code for Actor::GetForward
no longer negates the y component. Furthermore, if you use atan2 for any
calculations, you no longer negate the first parameter.
Now consider a unit square centered around the object

space origin, as in Figure 5.4. Assume that the world
space origin is the center of the game window. The goal
is to take the unit square centered around its object
space origin and express it as a rectangle with an
arbitrary position, scale, or orientation relative to the
world space origin.
For example, suppose one instance of the rectangle
should appear in world space such that it’s double in size
and is 50 units to the right of the world space origin. You
can accomplish this by applying mathematical
operations to each vertex of the rectangle.
One approach is to use algebra equations to compute the

correct vertex positions. Although you ultimately won’t
approach it in this manner, this is a useful bridge to
understanding the preferred solution. This chapter
focuses on 2D coordinate systems, though the same
method outlined here would also work in 3D (just with
an additional z component).
Translation
Translation takes a point and translates, or moves,
it by an offset. Given the point (x, y), you can
translate it by the offset (a, b) by using the following
equations:
For example, you could translate the point (1, 3) by the

offset (20, 15) as follows:
If you apply the same translation to every vertex of a

triangle, you translate the entire triangle.
Scale
When applied to each vertex in a triangle, scale
increases or decreases the size of the triangle. In a
uniform scale, you scale each component by the
same scale factor, s:
So you can uniformly scale (1, 3) by 5 as follows:
Scaling each vertex in the triangle by 5 would quintuple

the size of the triangle.
In a non-uniform scale, there are separate scale

factors (sx, sy) for each component:
For the example of transforming a unit square, a non-

uniform scale results in a rectangle instead of a square.
Rotation
Recall the discussion of the unit circle from Chapter
4, “Vectors and Basic Physics.” The unit circle begins
at the point (1, 0). A rotation of 90˚, or radians, is
counterclockwise to the point (0, 1), a rotation of
180˚, or π radians, is the point (–1, 0), and so on.
This is technically a rotation about the z-axis, even
though you don’t draw the z-axis in a typical unit
circle diagram.
Using sine and cosine, you can rotate an arbitrary point

(x, y) by the angle θ as follows:
Notice that both equations depend on the original x and

y values. For example, rotating (5, 0) by 270˚ is as
follows:
As with the unit circle, the angle θ represents a

counterclockwise rotation.
Keep in mind that this is a rotation about the origin.

Given a triangle centered around the object space origin,
rotating each vertex would rotate the triangle about the
origin.
Combining Transformations
Although the preceding equations apply each
transformation independently, it’s common to
require multiple transformations on the same vertex.
For example, you might want to both translate and
rotate a quad. It’s important to combine these
transformations in the correct order.
Suppose a triangle has the following points:

This original triangle points straight up, as in Figure
5.5(a). Now suppose you want to translate the triangle by
(5, 0) and rotate it by 90˚. If you rotate first and
translate second, you get this:
This results in the triangle rotated so that it points to the

left and translated to the right, as in Figure 5.5(b).
If you reverse the order of the transformations so that

you evaluate the translation first, you end up with this
calculation:
In the case of translation first, rotation second, you end
up with a triangle still facing to the left but positioned
several units above the origin, as in Figure 5.5(c). This
happens because you first move the triangle to the right,
and then you rotate about the origin. Usually, this
behavior is undesirable.
Figure 5.5 (a) Initial triangle, (b) rotating then

translating, and (c) translating then rotating
Because the order of transformations matter, it’s

important to have a consistent order. For the
transformation from object space to world space, always
apply the transformations in the order scale, then
rotation, then translation. Keeping this in mind, you
could combine all three separate equations for scale,
rotation, and translation into one set of equations to
scale by (sx, sy), rotate by θ, and translate by (a, b):
Issues with Combining Equations

The combined equations derived in the previous
section may seem like a solution to the problem:
Take an arbitrary vertex in object space, apply the
equations to each component, and you now have the
vertex transformed into world space with an
arbitrary scale, rotation, and position.
However, as alluded to earlier, this only transforms the

vertices from object space to world space. Because world
space is not normalized to device coordinates, you still
have more transformations to apply in the vertex shader.
These additional transformations typically do not have
equations as simple as the equations covered thus far.
This is especially because the basis vectors between these
different coordinate spaces might be different.
Combining these additional transformations into one
equation would become unnecessarily complex.
The solution to these issues is to not use separate

equations for each component. Instead, you use matrices
to describe the different transformations, and you can
easily combine these transformations with matrix
multiplication.
MATRICES AND
TRANSFORMATIONS
A matrix is a grid of values, with 2×2 columns. For
example, you could write a 2×2 matrix as follows,
with a through d representing individual values in
the matrix:
You use matrices to represent transformations in

computer graphics. All the transformations from the
preceding section have corresponding matrix
representations. If you are experienced in linear algebra,
you might recall that matrices can be used to solve
systems of linear equations. Thus, it’s natural that you
can represent the system of equations in the previous
section as matrices.
This section explores some of the basic use cases of

matrices in game programming. As with vectors, it’s
most important to understand how and when to use
these matrices in code. This book’s custom Math.h
header file defines Matrix3 and Matrix4 classes, along
with operators, member functions, and static functions
that implement all the necessary features.
Matrix Multiplication
Much as with scalars, you can multiply two matrices
together. Suppose you have the following matrices:
The result of the multiplication C = AB is:
In other words, the top-left element of C is the dot

product of the first row of A with the first column of B.
Matrix multiplication does not require matrices to have

identical dimensions, but the number of columns in the
left matrix must be equal to the number of rows in the
right matrix. For instance, the following multiplication is
also valid:
Matrix multiplication is not commutative, though it is
associative:
Transforming a Point by Using a Matrix

A key aspect of transformations is that you can
represent an arbitrary point as a matrix. For
example, you can represent the point p=(x, y) as a
single row (called a row vector):
You can instead represent p as a single column (called a

column vector):
Either representation works, but it’s important to

consistently use one approach. This is because whether
the point is a row or a column determines whether the
point appears on the left- or right-hand side of the
multiplication.
Suppose you have a transformation matrix T:
With matrix multiplication, you can transform the point

p by this matrix, yielding the transformed point (x′,y′).
However, whether p is a single row or a single column
gives different results when multiplied by T.
If p is a row, the multiplication is as follows:
But if p is a column, the multiplication would yield the

following:
This gives two different values for x′ and y′, but only one
is the correct answer—because the definition of a
transform matrix relies on whether you’re using row
vectors or column vectors.
Whether to use row or column vectors is somewhat

arbitrary. Most linear algebra textbooks use column
vectors. However, in computer graphics there is a history
of using either row or column vectors, depending on the
resource and graphics API. This book uses row vectors,
mainly because the transformations apply in a left-to-
right order for a given point. For example, when using
row vectors, the following equation transforms the point
q by matrix T first and then by matrix R:
You can switch between row and column vectors by

taking the transpose of each transform matrix. The
transpose of the matrix rotates the matrix such that the
first row of the original matrix becomes the first column
of the result:
If you wanted to switch the equation to transform q
using column vectors, you would calculate as follows:
The matrices in the remainder of this book assume that

you are using row vectors. However, a simple transpose
of these matrices converts them to work with column
vectors.
Finally, the identity matrix is a special type of matrix

represented by an uppercase l. An identity matrix always
has an equal number of rows and columns. All values in
the identity matrix are 0, except for the diagonal, which
is all 1s. For example, the 3×3 identity matrix is as
follows:
Any arbitrary matrix multiplied by the identity matrix

does not change. In other words:
Transforming to World Space, Revisited
You can represent the scale, rotation, and translation
transformations with matrices. To combine the
transformations, instead of deriving a combined
equation, multiply the matrices together. Once you
have a combined world transform matrix, you can
transform every vertex of the object by this world
transform matrix.
As before, let’s focus on 2D transformations first.
Scale Matrix
You can use a 2×2 scale matrix to apply the scale
transformation:
For example, this would scale (1,3) by (5,2):

Rotation Matrix
A 2D rotation matrix represents a rotation (about the
z-axis) by angle θ:
So, you can rotate (0,3) by 90˚ with the following:

Translation Matrices
You can represent 2D scale and rotation matrices
with 2×2 matrices. However, there’s no way to write
a generic 2D translation matrix of size 2×2. The only
way to express the translation T(a,b) is with a 3×3
matrix:
However, you can’t multiply a 1×2 matrix representing a

point by a 3×3 matrix because the 1×2 matrix doesn’t
have enough columns. The only way you can multiply
these together is if you add an additional column to the
row vector, making it a 1×3 matrix. This requires adding
an extra component to the point. Homogenous
coordinates use n+1 components to represent an n-
dimensional space. So, for a 2D space, homogeneous
coordinates use three components.
Although it might seem reasonable to call this third

component the z component, it’s a misnomer. That’s
because you’re not representing a 3D space, and you
want to reserve the z component for 3D spaces. Thus this
special homogeneous coordinate is the w component.
You use w for both 2D and 3D homogeneous
coordinates. So, a 2D point represented in homogeneous
coordinates is (x,y,w), while a 3D point represented in
homogeneous coordinates is (x,y,z,w).
For now, you will only use a value of 1 for the w

component. For example, you can represent the point p=
(x,y) with the homogeneous coordinate (x,y,1). To
understand how homogeneous coordinates work,
suppose you wish to translate the point (1,3) by (20,15).
First, you represent the point as a homogeneous
coordinate with a w component of 1 and then you
multiply the point by the translation matrix:
Note that, in this calculation, the w component remains

1. However, you’ve translated the x and y components by
the desired amount.
Combining Transformations
As mentioned earlier, you can combine multiple
transform matrices by multiplying them together.
However, you can’t multiply a 2×2 matrix with a 3×3
matrix. Thus, you must represent the scale and
rotation transforms with 3×3 matrices that work
with homogeneous coordinates:
Now that you’ve represented the scale, rotation, and

translation matrices as 3×3 matrices, you can multiply
them together into one combined transform matrix. This
combined matrix that transforms from object space to
world space is the world transform matrix. To
compute the world transform matrix, multiply the scale,
rotation, and translation matrices in the following order:
This order of multiplication corresponds to the order in
which you wish to apply the transformations (scale, then
rotate, then translate). You can then pass this world
transform matrix to the vertex shader and use it to
transform every vertex of an object by its world
transform matrix.
Adding World Transforms to Actor

Recall that the declaration of the Actor class already
has a Vector2 for position, a float for scale, and a
float for the angle rotation. You now must combine
these different attributes into a world transform
matrix.
First, add two member variables to Actor, a Matrix4

and a bool:
Matrix4 mWorldTransform;
bool mRecomputeWorldTransform;
The mWorldTransform variable obviously stores the

world transform matrix. The reason you use a Matrix4
here instead of a Matrix3 is because the vertex layout
assumes that all vertices have a z component (even
though in 2D, you don’t actually need the z component).
Since the homogenous coordinates for 3D are (x, y, z, w),
you need a 4×4 matrix.
The Boolean tracks whether you need to recalculate the

world transform matrix. The idea is that you want to
recalculate the world transform only if the actor’s
position, scale, or rotation changes. In each of the setter
functions for the position, scale, and rotation of the
actor, you set mRecomputeWorldTransform to true.
This way, whenever you change these component
properties, you’ll be sure to compute the world transform
again.
You also initialize mRecomputeWorldTransform to

true in the constructor, which guarantees to compute
the world transform at least once for each actor.
Next, implement a CreateWorldTransform function,

as follows:
void Actor::ComputeWorldTransform()
if (mRecomputeWorldTransform)
mRecomputeWorldTransform = false;
// Scale, then rotate, then translate
mWorldTransform = Matrix4::CreateScale(mScale);
mWorldTransform *= Matrix4::CreateRotationZ(mRotation);
mWorldTransform *= Matrix4::CreateTranslation(
Vector3(mPosition.x, mPosition.y, 0.0f));
Note that you use various Matrix4 static functions to

create the component matrices. CreateScale creates a
uniform scale matrix, CreateRotationZ creates a
rotation matrix about the z-axis, and
CreateTranslation creates a translation matrix.
You call ComputeWorldTransform in

Actor::Update, both before you update any
components and after you call UpdateActor (in case it
changes in the interim):
void Actor::Update(float deltaTime)
if (mState == EActive)
ComputeWorldTransform();
UpdateComponents(deltaTime);
UpdateActor(deltaTime);
Next, add a call to ComputeWorldTransform in

Game::Update to make sure any “pending” actors
(actors created while updating other actors) have their
world transform calculated in the same frame where
they’re created:
// In Game::Update (move any pending actors to mActors)
for (auto pending : mPendingActors)
pending ->ComputeWorldTransform();
mActors.emplace_back(pending);
It would be nice to have a way to notify components

when their owner’s world transform gets updated. This
way, the component can respond as needed. To support
this, first add a virtual function declaration to the base
Component class:
virtual void OnUpdateWorldTransform() { }
Next, call OnUpdateWorldTransform on each of the

actor’s components inside the
ComputeWorldTransform function. Listing 5.10 shows
the final version of ComputeWorldTransform.
Listing 5.10 Actor::ComputeWorldTransform

Implementation
void Actor::ComputeWorldTransform()
if (mRecomputeWorldTransform)
mRecomputeWorldTransform = false;
mWorldTransform *= Matrix4::CreateRotationZ(mRotation);
mWorldTransform *= Matrix4::CreateTranslation(
Vector3(mPosition.x, mPosition.y, 0.0f));
// Inform components world transform updated

for (auto comp : mComponents)
comp ->OnUpdateWorldTransform();
For now, you won’t implement

OnUpdateWorldTransform for any components.
However, you will use it for some components in
subsequent chapters.
Although actors now have world transform matrices, you

aren’t using the matrices in the vertex shader yet.
Therefore, running the game with the code as discussed
so far would just yield the same visual output as in
Figure 5.3. Before you can use the world transform
matrices in the shader, we need to discuss one other
transformation.
Transforming from World Space to Clip

Space
With the world transform matrix, you can transform
vertices into world space. The next step is to
transform the vertices into clip space, which is the
expected output for the vertex shader. Clip space is
a close relative of normalized device coordinates. The
only difference is that clip space also has a w
component. This was why you created a vec4 to save
the vertex position in the gl_Position variable.
The view-projection matrix transforms from world

space to clip space. As might be apparent from the name,
the view-projection matrix has two component matrices:
the view and the projection. The view accounts for how
a virtual camera sees the game world, and the
projection specifies how to convert from the virtual
camera’s view to clip space. Chapter 6, “3D Graphics,”
talks about both matrices in much greater detail. For
now, because the game is 2D, you can use a simple view-
projection matrix.
Recall that in normalized device coordinates, the

bottom-left corner of the screen is (–1, –1) and the top-
right corner of the screen is (1, 1). Now consider a 2D
game that does not have scrolling. A simple way to think
of the game world is in terms of the window’s resolution.
For example, if the game window is 1024×768, why not
make the game world that big, also?
In other words, consider a view of world space such that

the center of the window is the world space origin, and
there’s a 1:1 ratio between a pixel and a unit in world
space. In this case, moving up by 1 unit in world space is
the same as moving up by 1 pixel in the window.
Assuming a 1024×768 resolution, this means that the
bottom-left corner of the window corresponds to (–512,
–384) in world space, and the top-right corner of the
window corresponds to (512, 384), as in Figure 5.6.
Figure 5.6 The view of a world where the screen
resolution is 1024×768 and there’s a 1:1 ratio
between a pixel and a unit in world space
With this view of the world, it’s not too difficult to

convert from world space into clip space. Simply divide
the x-coordinate by width / 2 and divide the y-
coordinate by height / 2. In matrix form, assuming 2D
homogeneous coordinates, this simple view projection
matrix is as follows:
For example, given the 1024×768 resolution and the
point (256,192) in world space, if you multiply the point
by SimpleViewProjection, you get this:
The reason this works is that you normalize the range [–

512, 512] of the x-axis to [–1, 1], and the range of [–384,
384] on the y-axis [–1, 1], just as with normalized device
coordinates!
Combining the SimpleViewProjection matrix with the

world transform matrix, you can transform an arbitrary
vertex v from its object space into clip space with this:
This is precisely what you will calculate in the vertex

shader for every single vertex, at least until
SimpleViewProjection has outlived its usefulness.
Updating Shaders to Use Transform
Matrices
In this section, you’ll create a new vertex shader file
called Transform.vert. It starts initially as a copy
of the Basic.vert shader from Listing 5.3. As a
reminder, you write this shader code in GLSL, not
C++.
First, you declare two new global variables in

Transform.vert with the type specifier uniform. A
uniform is a global variable that typically stays the
same between numerous invocations of the shader
program. This contrasts in and out variables, which will
change every time the shader runs (for example, once per
vertex or pixel). To declare a uniform variable, use the
keyword uniform, followed by the type, followed by the
variable name.
In this case, you need two uniforms for the two different
matrices. You can declare these uniforms as follows:
uniform mat4 uWorldTransform;
uniform mat4 uViewProj;
Here, the mat4 type corresponds to a 4×4 matrix, which

is needed for a 3D space with homogeneous coordinates.
Then you change the code in the vertex shader’s main

function. First, convert the 3D inPosition into
homogeneous coordinates:
vec4 pos = vec4(inPosition, 1.0);
Remember that this position is in object space. So you

next multiply it by the world transform matrix to
transform it into world space, and then multiply it by the
view-projection matrix to transform it into clip space:
gl_Position = pos * uWorldTransform * uViewProj;
These changes yield the final version of

Transform.vert, shown in Listing 5.11.
Listing 5.11 Transform.vert Vertex Shader

#version 330
// Uniforms for world transform and view-proj
// Vertex attributes
in vec3 inPosition;
void main()
{
You then change the code in Game::LoadShaders to

use the Transform.vert vertex shader instead of
Basic.vert:
if (!mSpriteShader ->Load("Shaders/Transform.vert", "Shaders/Basic.frag"))
return false;
Now that you have uniforms in the vertex shader for the
world transform and view-projection matrices, you need
a way to set these uniforms from C++ code. OpenGL
provides functions to set uniform variables in the active
shader program. It makes sense to add wrappers for
these functions to the Shader class. For now, you can
add a function called SetMatrixUniform, shown in
Listing 5.12, to Shader.
Listing 5.12 Shader::SetMatrixUniform

Implementation
void Shader::SetMatrixUniform(const char* name, const Matrix4& matrix)
// Find the uniform by this name
GLuint loc = glGetUniformLocation(mShaderProgram, name);
// Send the matrix data to the uniform
glUniformMatrix4fv(
loc, // Uniform ID
1, // Number of matrices (only 1 in this case)
GL_TRUE, // Set to TRUE if using row vectors
matrix.GetAsFloatPtr() // Pointer to matrix data
);
Notice that SetMatrixUniform takes in a name as a

string literal, as well as a matrix. The name corresponds
to the variable name in the shader file. So, for
uWorldTransform, the parameter would be
"uWorldTransform". The second parameter is the
matrix to send to the shader program for that uniform.
In the implementation of SetMatrixUniform, you get

the location ID of the uniform with
glGetUniformLocation. Technically, you don’t have
to query the ID every single time you update the same
uniform because the ID doesn’t change during execution.
You could improve the performance of this code by
caching the values of specific uniforms.
Next, the glUniformMatrix4fv function assigns a

matrix to the uniform. The third parameter of this
function must be set to GL_TRUE when using row
vectors. The GetAsFloatPtr function is simply a
helper function in Matrix4 that returns a float*
pointer to the underlying matrix.
note
OpenGL has a newer approach to setting uniforms, called uniform buffer
objects (abbreviated UBOs). With UBOs, you can group together multiple
uniforms in the shader and send them all at once. For shader programs with
many uniforms, this generally is more efficient than individually setting each
uniform’s value.
With uniform buffer objects, you can split up uniforms into multiple groups.
For example, you may have a group for uniforms that update once per frame
and uniforms that update once per object. The view-projection won’t change
more than once per frame, while every actor will have a different world
transform matrix. This way, you can update all per-frame uniforms in just one
function call at the start of the frame. Likewise, you can update all per-object
uniforms separately for each object. To implement this, you must change
how you declare uniforms in the shader and how you mirror that data in the
C++ code.
However, at this writing, some hardware still has spotty support for UBOs.
Specifically, the integrated graphics chips of some laptops don’t fully support
uniform buffer objects. On other hardware, UBOs may even run more slowly
than uniforms set the old way. Because of this, this book does not use
uniform buffer objects. However, the concept of buffer objects is prevalent in
other graphics APIs, such as DirectX 11 and higher.
Now that you have a way to set the vertex shader’s matrix
uniforms, you need to set them. Because the simple view-
projection won’t change throughout the course of the
program, you only need to set it once. However, you need
to set the world transform matrix once for each sprite
component you draw because each sprite component
draws with the world transform matrix of its owning
actor.
In Game::LoadShaders, add the following two lines to

create and set the view-projection matrix to the simple
view projection, assuming a screen width of 1024×768:
Matrix4 viewProj = Matrix4::CreateSimpleViewProj(1024.f, 768.f);
mShader.SetMatrixUniform("uViewProj", viewProj);
The world transform matrix for SpriteComponent is a

little more complex. The actor’s world transform matrix
describes the position, scale, and orientation of the actor
in the game world. However, for a sprite, you also want
to scale the size of the rectangle based on the size of the
texture. For example, if an actor has a scale of 1.0f, but
the texture image corresponding to its sprite is 128×128,
you need to scale up the unit square to 128×128. For
now, assume that you have a way to load in the textures
(as you did in SDL) and that the sprite component knows
the dimensions of these textures via the mTexWidth and
mTexHeight member variables.

SpriteComponent::Draw (for now). First, create a
scale matrix to scale by the width and height of the
texture. You then multiply this by the owning actor’s
world transform matrix to create the desired world
transform matrix for the sprite. Next, call
SetMatrixUniform to set the uWorldTransform in
the vertex shader program. Finally, you draw the
triangles as before, with glDrawElements.
Listing 5.13 Current Implementation of

SpriteComponent::Draw
void SpriteComponent::Draw(Shader* shader)
// Scale the quad by the width/height of texture
Matrix4 scaleMat = Matrix4::CreateScale(
static_cast<float>(mTexWidth),
static_cast<float>(mTexHeight),
1.0f);
Matrix4 world = scaleMat * mOwner ->GetWorldTransform();
// Set world transform
shader ->SetMatrixUniform("uWorldTransform", world);
// Draw quad
glDrawElements(GL_TRIANGLES, 6, GL_UNSIGNED_INT, nullptr);

}
With the world transform and view-projection matrices

added to the shader, you now can see the individual
sprite components in the world at arbitrary positions,
scales, and rotations, as in Figure 5.7. Of course, all the
rectangles are just a solid color for now because
Basic.frag just outputs blue. This is the last thing to
fix to achieve feature parity with the SDL 2D rendering
from the previous chapters.
Figure 5.7 Drawing sprite components with

different world transform matrices
TEXTURE MAPPING
Texture mapping is a technique for rendering a
texture (image) on the face of a triangle. It allows
you to use colors from a texture when drawing a
triangle instead of using just a solid color.
To use texture mapping, you need an image file. Next,

you need to decide how to apply textures to each triangle.
If you have just the sprite rectangle, it makes sense that
the top-left corner of the rectangle should correspond to
the top-left corner of the texture. However, you can use
texture mapping with arbitrary 3D objects in the game.
For example, to correctly apply a texture to a character’s
face, you need to know which parts of the texture should
correspond to which triangles.
To support this, you need an additional vertex attribute

for every vertex in the vertex buffer. Previously, the
vertex attributes only stored a 3D position in each vertex.
For texture mapping, each vertex also needs a texture
coordinate that specifies the location in the texture that
corresponds to that vertex.
Texture coordinates typically are normalized

coordinates. In OpenGL, the coordinates are such that
the bottom-left corner of the texture is (0, 0) and the
top-right corner is (1, 1), as shown in Figure 5.8. The U
component defines the right direction of the texture,
and the V component defines the up direction of the
texture. Thus, many use the term UV coordinates as a
synonym for texture coordinates.
Figure 5.8 UV coordinates for a texture in OpenGL
Because OpenGL specifies the bottom left of the texture

as its origin, it also expects the image pixel data format
as one row at a time, starting at the bottom row.
However, a big issue with this is that most image file
formats store their data starting at the top row. Not
accounting for this discrepancy results in textures that
appear upside down. There are multiple ways to solve
this problem: invert the V-component, load the image
upside down, or store the image on disk upside down.
This book simply inverts the V component—that is,
assumes that the top-left corner is (0, 0). This
corresponds to the texture coordinate system that
DirectX uses.
Each vertex of a triangle has its own separate UV

coordinates. Once you know the UV coordinates for each
vertex of a triangle, you can fill in every pixel in the
triangle by blending (or interpolating) the texture
coordinate, based on the distance from each of the three
vertices. For example, a pixel exactly in the center of the
triangle corresponds to a UV coordinate that’s the
average of the three vertices’ UV coordinates, as in
Figure 5.9.
Figure 5.9 Texture mapping applied to a triangle
Recall that a 2D image is just a grid of pixels with

different colors. So, once you have a texture coordinate
for a specific pixel, you need to convert this UV
coordinate to correspond to a specific pixel in the
texture. This “pixel in the texture” is a texture pixel, or
texel. The graphics hardware uses a process called
sampling to select a texel corresponding to a specific
UV coordinate.
One complication of using normalized UV coordinates is
that two slightly different UV coordinates may end up
closest to the same texel in the image file. The idea of
selecting the texel closest to a UV coordinate and using
that for the color is called nearest-neighbor filtering.
However, nearest-neighbor filtering has some issues.

Suppose you map a texture to a wall in a 3D world. As
the player gets closer to the wall, the wall appears larger
and larger onscreen. This looks like zooming in on an
image file in a paint program, and the texture appears
blocky or pixelated because each individual texel is very
large onscreen.
To solve this pixelation, you can instead use bilinear

filtering. With bilinear filtering, you select a color based
on the blending of each texel neighboring the nearest
neighbor. If you use bilinear filtering for the wall
example, as the player gets closer, the wall seems to blur
instead of appearing pixelated. Figure 5.10 shows a
comparison between nearest-neighbor and bilinear
filtering of part of the star texture.
Figure 5.10 Nearest-neighbor filtering (left) and

bilinear filtering (right)
We explore the idea of improving the quality of textures
further in Chapter 13, “Intermediate Graphics.” For now,
let’s enable bilinear filtering for all textures.
To use texture mapping in OpenGL, there are three

things you need to do:
Load image files (textures) and create OpenGL texture objects.
Update the vertex format to include texture coordinates.
Update the shaders to use the textures.
Loading the Texture

Although you can use the SDL Image library to load
images for OpenGL, the Simple OpenGL Image
Library (SOIL) is a little easier to use. SOIL can read
in several file formats, including PNG, BMP, JPG,
TGA, and DDS. As it’s designed to work with
OpenGL, it slots in easily with the other OpenGL
code needed for creating a texture object.
Listing 5.14 gives the declaration of a Texture class that

encapsulates loading in a texture file and using it with
OpenGL. The names of the functions and member
variables are mostly self-explanatory; for example, Load
loads the texture from the file. For member variables,
you have a width and height of the texture and an
OpenGL texture ID.
Listing 5.14 Texture Declaration

class Texture
public:
Texture();
~Texture();
bool Load(const std::string& fileName);
void Unload();
void SetActive();
int GetWidth() const { return mWidth; }
int GetHeight() const { return mHeight; }
private:
// OpenGL ID of this texture
unsigned int mTextureID;
// Width/height of the texture
int mWidth;
int mHeight;
};
The implementation of Load contains the bulk of the
Texture class code. You first declare a local variable to
store the number of channels and then call
SOIL_load_image to load in the texture:
int channels = 0;
unsigned char* image = SOIL_load_image(
fileName.c_str(), // Name of file
&mWidth, // Stores width
&mHeight, // Stores height
&channels, // Stores number of channels
SOIL_LOAD_AUTO // Type of image file, or auto for any
);
If SOIL fails to load the image file, SOIL_load_image

returns nullptr. Therefore, you should add a check to
make sure the image loaded.
Then you need to determine whether the image is RGB

or RGBA. You can assume that this is based on the
number of channels (three means RGB, four means
RGBA):
int format = GL_RGB;

if (channels == 4)
format = GL_RGBA;
Next, use glGenTextures to create an OpenGL texture

object (saving the ID in mTextureID) and
glBindTexture to set the texture as active:
glGenTextures(1, &mTextureID);
glBindTexture(GL_TEXTURE_2D, mTextureID);
The GL_TEXTURE_2D target passed to glBindTexture

is by far the most common texture target, but there are
other choices for advanced types of textures.
Once you have an OpenGL texture object, the next step is

to copy the raw image data into it with the
glTexImage2D function, which takes quite a few
parameters:
glTexImage2D(
GL_TEXTURE_2D, // Texture target
0, // Level of detail (for now, assume 0)

format, // Color format OpenGL should use
mWidth, // Width of texture
mHeight, // Height of texture
0, // Border – "this value must be 0"
format, // Color format of input data
GL_UNSIGNED_BYTE,// Bit depth of input data
// Unsigned byte specifies 8-bit channels
image // Pointer to image data
);
Once you’ve copied the image data to OpenGL, you can

tell SOIL to free the image from memory:
SOIL_free_image_data(image);
Finally, use the glTexParameteri function to enable

bilinear filtering:
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
For now, don’t worry about the parameters passed to

glTexParameteri. (They are discussed further in
Chapter 13, “Intermediate Graphics.”)
Listing 5.15 shows the final version of Texture::Load.
Listing 5.15 Texture::Load Implementation

bool Texture::Load(const std::string& fileName)
int channels = 0;
unsigned char* image = SOIL_load_image(fileName.c_str(),
&mWidth, &mHeight, &channels, SOIL_LOAD_AUTO);
if (image == nullptr)
SDL_Log("SOIL failed to load image %s: %s",
fileName.c_str(), SOIL_last_result());
return false;
int format = GL_RGB;
if (channels == 4)
format = GL_RGBA;
}
glTexImage2D(GL_TEXTURE_2D, 0, format, mWidth, mHeight, 0, format,
GL_UNSIGNED_BYTE, image);
SOIL_free_image_data(image);
// Enable bilinear filtering
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
return true;
The Texture::Unload and Texture::SetActive

functions are each just a single line. Unload deletes the
texture object, and SetActive calls glBindTexture:
void Texture::Unload()
{
glDeleteTextures(1, &mTextureID);
void Texture::SetActive()
You can then load textures into a map in Game, much as

you did for SDL_Texture previously. The
Game::GetTexture function then returns a Texture*
for the requested texture. Next, SpriteComponent
needs a Texture* member variable instead of
SDL_Texture*.
Finally, in SpriteComponent::Draw, add a call to

SetActive on mTexture, right before drawing the
vertices. This means you can now set a different active
texture for each sprite component you draw:
// In SpriteComponent::Draw...
// Set current texture
mTexture ->SetActive();
// Draw quad
glDrawElements(GL_TRIANGLES, 6,
GL_UNSIGNED_INT, nullptr);
Updating the Vertex Format

To use texture mapping, the vertices need to have
texture coordinates, so you need to update the sprite
VertexArray:
float vertices[] = {
-0.5f, 0.5f, 0.f, 0.f, 0.f, // top left
0.5f, 0.5f, 0.f, 1.f, 0.f, // top right
0.5f, -0.5f, 0.f, 1.f, 1.f, // bottom right
-0.5f, -0.5f, 0.f, 0.f, 1.f // bottom left
};
Remember that the V texture coordinate is flipped to

account for OpenGL’s idiosyncrasy with how it expects
the image data.
For each vertex, the first three floating-point values are

the position and the next two floating-point values are
the texture coordinates. Figure 5.11 shows the memory
layout of this new vertex format.
Figure 5.11 Vertex memory layout with position and
texture coordinates
Because you’re changing the vertex layout, you must

change code in the VertexArray constructor. For
simplicity, assume that all vertices must have a 3D
position and 2D texture coordinates. (You’ll change this
in subsequent chapters.)
Because the size of each vertex has changed, you need to

update the glBufferData call to specify that each
vertex now has five floats per vertex:
glBufferData(GL_ARRAY_BUFFER, numVerts * 5 * sizeof(float),
verts, GL_STATIC_DRAW);
Because the index buffer is still the same, the

glBufferData call for that doesn’t change. However,
you must change vertex attribute 0 to specify that the
stride of the vertex is now five floats:
glVertexAttribPointer(0, 3, GL_FLOAT, GL_FALSE,
sizeof(float) * 5, // The stride is now 5 floats
0); // Vertex position is still offset 0
This only fixes the position vertex attribute. However,

because you now have a second vertex attribute for the
texture coordinate, you must enable vertex attribute 1
and specify its format:
glVertexAttribPointer(
1, // Vertex attribute index
2, // Number of components (2 because UV)
GL_FLOAT, // Type of each component
GL_FALSE, // Not used for GL_FLOAT
sizeof(float) * 5, // Stride (usually size of each vertex)
reinterpret_cast<void*>(sizeof(float) * 3) // Offset pointer
);
The last parameter to this glVertexAttribPointer

call is rather ugly. OpenGL needs to know the number of
bytes from the start of a vertex to this attribute. That’s
where the sizeof(float) * 3 comes from. However,
OpenGL wants this as an offset pointer. Thus, you must
use reinterpret_cast to coerce the type into a
void* pointer.
tip
If you use a struct in C++ code to represent the format of the vertices, you
can use the offsetof macro to determine the offsets to a vertex attribute
rather than manually computing them. This is especially helpful if there is
padding between vertex elements.
Updating the Shaders

Because the vertex format now uses texture
coordinates, you should create two new shaders:
Sprite.vert (initially a copy of
Transform.vert) and Sprite.frag (initially a
copy of Basic.frag).
The Shader
There previously was only one vertex attribute, so
you could just declare position as an in variable, and
GLSL knew which vertex attribute it corresponded
to. However, now that there are multiple vertex
attributes, you must specify which attribute slot
corresponds to which in variable. This changes the
variable declarations to the following:
layout(location=0) in vec3 inPosition;
layout(location=1) in vec2 inTexCoord;

The layout directive specifies which attribute slot
corresponds to which in variable. Here, you specify that
there’s a 3D vector of floats for vertex attribute slot 0 and
a 2D vector of floats for vertex attribute slot 1. This
corresponds to the slot numbers in the
glVertexAttribPointer calls.
Next, although the texture coordinates are an input to

the vertex shader (because it’s in the vertex layout), the
fragment shader also needs to know the texture
coordinates. This is because the fragment shader needs
to know the texture coordinates to determine the color at
the pixel. Luckily, you can pass data from the vertex
shader to the fragment shader by declaring a global out
variable in the vertex shader:
out vec2 fragTexCoord;
Then, inside the vertex shader’s main function, add the

following line, which copies the texture coordinates
directly from the vertex input variable to the output
variable:
fragTexCoord = inTexCoord;
The reason this ultimately works is that OpenGL

automatically interpolates vertex shader outputs across
the face of the triangle. Therefore, even though a triangle
has only three vertices, any arbitrary pixel on the face of
a triangle will know its corresponding texture
coordinates in the fragment shader.
For completeness, Listing 5.16 gives the full source code

for Sprite.vert.
Listing 5.16 Sprite.vert Implementation

#version 330
// Uniforms for world transform and view-proj
// Attribute 0 is position, 1 is tex coords.
layout(location = 0) in vec3 inPosition;
layout(location = 1) in vec2 inTexCoord;
// Add texture coordinate as output
out vec2 fragTexCoord;
void main()
// Convert position to homogeneous coordinates

// Transform position to world space, then clip space
// Pass along the texture coordinate to frag shader
The Sprite.frag Shader

As a rule, any out variables in the vertex shader
should have a corresponding in in the fragment
shader. The name and type of the in variable in the
fragment shader must have the same name and type
as the corresponding out variable in the vertex
shader:
in vec4 fragTexCoord;
Next, you need to add a uniform for a texture sampler

(that can get the color from a texture given a texture
coordinate):
uniform sampler2D uTexture;
The sampler2D type is a special type that can sample

2D textures. Unlike with the world transform and view-
projection uniforms in the vertex shader, you currently
don’t need any code in C++ to bind this sampler
uniform. This is because you currently bind only one
texture at a time, so OpenGL automatically knows that
the only texture sampler in the shader corresponds to the
active texture.
Finally, replace the outColor assignment in the main

function with the following:
outColor = texture(uTexture, fragTexCoord);
This samples the color from the texture, using the texture
coordinates received from the vertex shader (after the
coordinates interpolate across the face of the triangle).
Listing 5.17 shows the full source code for

Sprite.frag.
Listing 5.17 Sprite.frag Implementation

#version 330
// Tex coord input from vertex shader
in vec2 fragTexCoord;
// Output color
out vec4 outColor;

// For texture sampling
void main()
// Sample color from texture
outColor = texture(uTexture, fragTexCoord);
You then change the code in Game:LoadShaders to

now load Sprite.vert and Sprite.frag. The
previous code inside the various actors that set textures
on SpriteComponents now also successfully loads the
textures with SOIL. With this code, you can now draw
sprites with texture mapping, as shown in Figure 5.12.
Unfortunately, there’s still one last issue to fix. Right
now, the code is drawing black for pixels that should be
transparent.
Figure 5.12 Texture-mapped sprites
Alpha Blending
Alpha blending determines how to blend pixels
with transparency (an alpha channel less than 1).
Alpha blending uses an equation in the following
form to calculate the pixel color:
In this equation, the source color is the color for the

new source you’re drawing (from the fragment shader),
and the destination color is the color that’s already in the
color buffer. You can customize the alpha blending
function by specifying the factor parameters.
To get the desired alpha blending result for
transparency, you set the source factor to the alpha of the
pixel you’re drawing (the source alpha) and the
destination factor to one minus source alpha:
For example, suppose you have 8 bits per color, and the
color buffer at some pixel is red. In this case, this is the
destination color:
Next, say that you want to draw a pixel that’s blue; this is
the source color:
Now suppose the source alpha is zero, meaning the pixel

is completely transparent. In this case, our equation
evaluates to the following:
This is the result you want for a fully transparent pixel.
You completely ignore the source color if the alpha is
zero and just use whatever color is already in the color
buffer.
To enable this in code, add the following to

Game::GenerateOuput, right before drawing all the
sprites:
glEnable(GL_BLEND);
glBlendFunc(
GL_SRC_ALPHA, // srcFactor is srcAlpha
GL_ONE_MINUS_SRC_ALPHA // dstFactor is 1 - srcAlpha
);
The glEnable call says to turn on color buffer blending

(which is disabled by default). Then you use
glBlendFunc to specify the srcFactor and
dstFactor that you want.
With alpha blending in place, the sprites now look

correct, as shown in Figure 5.13. The 2D OpenGL
rendering code now has feature parity with the previous
use of SDL 2D rendering. It took a lot of work to get here,
but the advantage is that the game code now has the
foundations for 3D graphics support, which is the topic
of Chapter 6.
Figure 5.13 Texture-mapped sprites, with alpha
blending
GAME PROJECT
This chapter’s game project demonstrates all the
code to convert the game code from SDL graphics to
OpenGL. It converts the Asteroids game project from
Chapter 3 to instead use OpenGL. The controls are
the same as in Chapter 3: WASD to move the ship and
spacebar to fire lasers. The code is available in the
book’s GitHub repository, in the Chapter05
Mac.
SUMMARY
Because graphics hardware is optimized for
polygons, 2D and 3D games internally use polygons
(usually triangles) to represent all graphical objects
in the world. Even a 2D sprite that you might think of
as an image is a rectangle with a texture mapped to
it. To send triangles to the graphics hardware, you
must declare the attributes of each vertex and create
a vertex and an index buffer.
All modern graphics APIs expect a programmer to use

vertex and fragment (pixel) shaders to specify how
polygons should render. You write these shaders as
separate programs in a shader programming language
(not C++). The vertex shader minimally outputs the
vertex position in clip space, while the fragment shader
determines the final color at a pixel.
Transformations allow you to draw several instances of

the same object without needing separate vertex and
index buffers for each instance. Object space is the
coordinate space relative to the origin of an object, while
world space is the coordinate space relative to the game
world.
Games use matrices to represent transformations, and

there are several different matrices for transformations,
such as scale, rotation, and translation. Combining these
transformations in the order scale, rotate, and translate
yields a world transform matrix that can transform from
object space to world space. To convert from world space
to clip space, you use the view-projection matrix. For 2D
games, you can simplify this by making 1 unit in world
space equivalent to 1 pixel in the window.
Texture mapping applies part of a texture to the face of a

triangle. To implement this, you need texture (UV)
coordinates as a vertex attribute. In the fragment shader,
you sample the texture color from the UV coordinate.
This can either be based on the color of the texture pixel
(texel) nearest to the UV coordinate, or it can be based
on a bilinear filtering that considers nearby texels.
Finally, even though the task seems trivial, displaying

sprites in OpenGL requires a lot of code. First, you must
initialize OpenGL and GLEW. Next, to render any
triangles, you must create a vertex array object, specify a
vertex layout, write a vertex and pixel shader, and write
code to load these shader programs. To transform the
vertices from object space to clip space, you have to use
uniforms to specify the world transform and view-
projection matrices. To add texture mapping, you have to
load an image, change the vertex layout to include UV
coordinates, and update the shaders to sample from the
texture.
ADDITIONAL READING
There are many excellent online references for
aspiring OpenGL developers. The official OpenGL
reference pages are useful for finding out what the
parameters for each function do. Of all the OpenGL
tutorial sites, one of the best is Learn OpenGL. For
an extensive look at the graphical techniques used in
game development, Real-Time Rendering by
Thomas Akenine-Moller et al. is a definitive
reference.
Akenine-Moller, Thomas, Eric Haines, and

Naty Hoffman. Real-Time Rendering, 3rd
edition. Natick: A K Peters, 2008.
Learn OpenGL. Accessed November 24,
2017. http://learnopengl.com/.
OpenGL Reference Pages. Accessed
November 24, 2017.
https://www.opengl.org/sdk/docs/man/.
EXERCISES
The exercises for this chapter involve making some
modifications to this chapter’s game project to gain
more experience using various OpenGL functions.
Exercise 5.1
Modify the background clear color so that it
smoothly changes between colors. For example,
starting from black, smoothly change over several
seconds to blue. Then select another color (such as
red) and smoothly change over several seconds to
this other color. Think about how you can use
deltaTime in Game::Update to facilitate this
smooth transition.
Exercise 5.2
Modify the sprite vertices so that each vertex also has
an RGB color associated with it. This is known as a
vertex color. Update the vertex shader to take the
vertex color as an input and pass it to the fragment
shader. Then change the fragment shader so that
rather than simply drawing the color sampled from
the texture, it averages the color between the vertex
color and texture color.
CHAPTER 6
3D GRAPHICS
This chapter covers how to switch from a 2D

environment to a full 3D game, which
requires several changes. The Actor
transform, including 3D rotations, becomes
more complex. In addition, you need to load
and draw 3D models. Finally, most 3D games
have some type of lighting applied to the
scene. The game project for this chapter
demonstrates all these 3D techniques.
THE ACTOR TRANSFORM IN 3D
The representation of the Actor transform you have
used so far in this book works for 2D graphics.
However, supporting a fully 3D world requires some
modification. Most obviously, the position Vector2
becomes Vector3. But this brings up an important
question: Which directions are x, y, and z in the
world? Most 2D games use a coordinate system
where x is the horizontal direction and y is the
vertical direction. But even in 2D, +y might be up or
down, depending on the implementation. Adding a
third component increases the possible
representations. It’s an arbitrary decision, but it
should be consistent. In this book, +x is forward, +y
is to the right, and +z is up. Figure 6.1 illustrates this
coordinate system.
Figure 6.1 3D coordinate system used in this book
If you take your left hand and pretend that the thumb is
up, the index finger is forward, and the middle finger is
right, you see that it matches the coordinate system in
Figure 6.1 perfectly. Thus, this type of coordinate system
is a left-handed coordinate system. It would be
right-handed if +y were instead to the left.
Transform Matrices for 3D

Using 3D coordinates means that the homogenous
coordinates are now (x,y,z,w). Recall that you need
the w component for translation matrices to work.
With 3D coordinates, the translation matrices
become 4×4 matrices. This modification is simple for
translation and scale.
A 4×4 translation matrix translates by the offset (a,b,c):
Similarly, the scale matrix can scale by up to three

factors:
However, rotations in 3D are not as simple.
Euler Angles
Representing rotations in 3D is more complex than
in 2D. Previously, an actor only needed one float for
rotation. This represented a rotation about the z-axis
because that’s the only rotation possible in 2D. But in
3D, it’s valid to rotate about any of the three
coordinate axes. One approach for 3D rotations is
Euler angles, where there are three angles (yaw,
pitch, and roll) representing rotation about each axis.
The names yaw, pitch, and roll come from airplane
terminology. Yaw is a rotation about the up axis,
pitch is a rotation about the side axis, and roll is a
rotation about the forward axis. In the coordinate
from Figure 6.1, yaw is rotation about +z, pitch is
rotation about +y, and roll is rotation about +x.
When you have three different rotation angles, you can

combine them by creating a separate rotation matrix for
each Euler angle. You then multiply these three matrices,
and the order of multiplication affects the final rotation
of the object. One common approach is roll, then pitch,
then yaw:
However, there isn’t any one “correct” order to apply

these Euler angles. You simply must pick one order and
stick with it.
With Euler angles, it’s difficult to derive an arbitrary

rotation. Suppose a spaceship faces down the +x
(forward) axis in object space. You want to rotate the
ship so it points toward an arbitrary object at position P.
Achieving this new orientation may require some
combination of yaw, pitch, and roll, and computing these
separate angles is not straightforward.
Furthermore, suppose you have an object with an initial

Euler angle orientation, and you also have a target Euler
angle orientation. You want to smoothly transition, or
interpolate between, these two orientations over some
period. You can interpolate Euler angles by interpolating
each angle separately. However, in various situations,
this interpolation will not look correct because when you
interpolate each component separately, you may
encounter singularities where the interpolation appears
at odd orientations.
Although it is possible to use Euler angles in a game, for

arbitrary rotations there’s another choice that tends to
work better.
Quaternions
Many games use quaternions instead of Euler
angles. The formal mathematical definition of a
quaternion is complex. For the purposes of this book,
think of a quaternion as a method to represent a
rotation about an arbitrary axis (not just x, y, or z).
Basic Definition
3D graphics use unit quaternions, which are
quaternions with a magnitude of one. A quaternion
has both a vector and a scalar component. This book
uses the following notation to represent a quaternion
as its vector and scalar components:
The calculation of the vector and scalar components
depends on the normalized axis of rotation, â, and the
angle of rotation, θ:
This equation works only with a normalized axis of

rotation. An unnormalized axis yields a non-unit
quaternion and causes shearing (non-uniform
stretching) of objects in the game.
To crystalize the use of quaternions, consider the earlier

problem of rotating a ship to face an arbitrary object.
Remember that with Euler angles, it’s difficult to
calculate the exact yaw, pitch, and roll angles. However,
quaternions make this problem easier. Initially, the ship
is at position S with an initial facing down the x-axis. Say
that you want to rotate the ship to instead face an
arbitrary point P. First, compute the vector from the ship
to the new point and normalize this vector:
Next, compute the axis of rotation between the original
facing and the new facing, using the cross product, and
normalize this vector:
Then compute the angle of rotation using the dot product

and arccosine:
Finally, plug in this axis and angle to create the

quaternion representing the rotation of the ship such
that it faces point P. This works wherever P is in the 3D
world.
One edge case is that if NewFacing is parallel to the

original facing, the cross product yields a vector of all
zeros. This vector has a length of zero, and so dividing by
zero to normalize the vector corrupts the axis of rotation.
Thus, any code doing such calculations needs to verify
that NewFacing is not parallel to the original facing. If
they are parallel, this means the object already faces in
the NewFacing direction. In this case, the quaternion is
just the identity quaternion, which applies no rotation. If
the vectors are antiparallel, then you must rotate by π
radians about up.
Combining Rotations
Another common operation is applying an additional
rotation to an existing quaternion. Given two
quaternions, p and q, the Grassmann product is
the rotation of q followed by p:
Note that even though the multiplication has p to the left

of q, the rotation applies in a right-to-left order. Also,
because the Grassmann product uses a cross product, it
isn’t commutative. So, swapping p and q will reverse the
order of rotation.
Much like a matrix, a quaternion has an inverse. For unit

quaternions, the inverse of a quaternion is the negation
of its vector component:
Because there’s an inverse, there is also an identity

quaternion, defined as follows:
Rotating a Vector by a Quaternion
To rotate a 3D vector, v, by a quaternion, first
represent v as the following quaternion, r:
Next, compute r′ with two Grassmann products:
The rotated vector is then simply the vector component

of r′:
Spherical Linear Interpolation

Quaternions support a more accurate form of
interpolation called spherical linear
interpolation (Slerp). The Slerp equation takes in
two quaternions, a and b, as well as the fractional
value in the range [0, 1] from a to b. For example, the
following yields a quaternion rotation 25% of the way
from a to b:
In the interest of brevity, this section omits the

calculations for Slerp.
Quaternion-to-Rotation Matrix
Because you ultimately still need a world transform
matrix, you need to eventually convert the
quaternion rotation into a matrix. The conversion of
a quaternion into a matrix has quite a few terms:
Quaternions in Code
As with vectors and matrices, for quaternions the
custom Math.h header file has a Quaternion class.
Listing 6.1 shows the most useful functions. Because
the order of multiplication of quaternions often
confuses game programmers (for example, to rotate
p followed by q, you multiply q by p), instead of using
the multiplication operator, the Math.h library
declares a Concatenate function. This function
simply takes in the quaternions in the order many
expect—so the rotation “p followed by q” is as
follows:
Quaternion result = Quaternion::Concatenate(q, p);
Listing 6.1 Quaternion Functions of Note

class Quaternion
public:
// Functions/data omitted
// ...
// Construct the quaternion from an axis and angle
explicit Quaternion(const Vector3& axis, float angle);
// Spherical Linear Interpolation

static Quaternion Slerp(const Quaternion& a, const Quaternion& b, float f)
// Concatenate (rotate by q FOLLOWED BY p, uses Grassmann product pq)
static Quaternion Concatenate(const Quaternion& q, const Quaternion& p);
// v = (0, 0, 0); s = 1
static const Quaternion Identity;
};
// In Matrix4...
// Create Matrix4 from Quaternion
static Matrix4 CreateFromQuaternion(const class Quaternion& q);
// In Vector3...
// Transform a Vector3 by a Quaternion
static Vector3 Transform(const Vector3& v, const class Quaternion& q);
New Actor Transform in Action

With the rotation question settled, the Actor class
transformation now has a Vector3 for position, a
Quaternion for rotation, and a float for scale:
Vector3 mPosition;
Quaternion mRotation;
float mScale;
With this new transform representation, the code for
calculating the world transform matrix in
ComputeWorldTransform changes to this:
mWorldTransform *= Matrix4::CreateFromQuaternion(mRotation);
mWorldTransform *= Matrix4::CreateTranslation(mPosition);
Getting the forward vector of an actor now requires

transforming the initial forward vector (+x) by the
rotation quaternion:
Vector3 GetForward() const
return Vector3::Transform(Vector3::UnitX, mRotation);
You then need to fix any code that applied rotations

using a single angle, such as
MoveComponent::Update. To keep things simple for
now, MoveComponent only rotates about the +z axis
(yaw). This updated code is in Listing 6.2, which first
gets the existing quaternion rotation of the owning actor.
It next creates a new quaternion representing the
additional rotation to apply. Finally, it concatenates the
original rotation with the new quaternion to get the final
rotation quaternion.
Listing 6.2 MoveComponent::Update

Implementation with Quaternions
void MoveComponent::Update(float deltaTime)
if (!Math::NearZero(mAngularSpeed))
Quaternion rot = mOwner->GetRotation();
float angle = mAngularSpeed * deltaTime;
// Create quaternion for incremental rotation
// (Rotate about up axis)
Quaternion inc(Vector3::UnitZ, angle);
// Concatenate old and new quaternion
rot = Quaternion::Concatenate(rot, inc);
mOwner->SetRotation(rot);
}
// Updating position based on forward speed stays the same
// ...
LOADING 3D MODELS
For sprite-based games, every sprite draws with a
single quad, which means it’s okay to hard-code the
vertex and index buffers. However, for a full 3D
game, there are a lot of other triangular meshes. For
example, a first-person shooter has enemy meshes,
weapon meshes, character meshes, meshes for the
environment, and so on. An artist creates these
models in a 3D modeling program such as Blender or
Autodesk Maya. The game then needs code to load
these models into vertex and index buffers.
Choosing a Model Format

Before you can use 3D models, you need to decide
how to store the models in files. One idea is to choose
a modeling program and add support for loading that
program’s proprietary file format. However, doing so
has several drawbacks. First, the feature set of a 3D
modeling program is significantly higher than that of
a game. Modeling programs support many other
types of geometry, including NURBS, quads, and n-
gons. Modeling programs also support complex
lighting and rendering techniques, including ray-
tracing. No game will replicate all these features.
Furthermore, most modeling files have a great deal of

data that’s unnecessary at runtime. For example, the file
format might store the undo history of the model.
Clearly, the game doesn’t need access to this when
running. All this extra information means that a
modeling file format is large, and loading it at runtime is
a performance hit.
Furthermore, modeling file formats are opaque and,

depending on the format, may not have any
documentation. So outside of reverse-engineering the file
format, it might not even be possible to load it into the
game.
Finally, choosing one modeling format ties the game

directly to one specific modeling program. What if a new
artist wants to use an entirely different modeling
program? Using a proprietary format makes this difficult
if there’s no easy conversion process.
Exchange formats aim to work across several

modeling programs. Among the most popular formats
are FBX and COLLADA, which many different modeling
programs support. Even though SDKs exist for loading
these formats, the formats still suffer from having way
more data than needed for a game at runtime.
It’s useful to consider how a commercial engine such as

Unity or Epic’s Unreal Engine works. While both engines
support importing file formats such as FBX into their
editors, the runtime doesn’t use these formats. Instead,
on import there’s a conversion process into an internal
engine format. The game runtime then loads models in
this internal format.
Other engines provide exporter plugins for popular

modeling programs. The exporter plugin converts the
modeling program’s format into a custom format
designed for the game runtime’s consumption.
In the spirit of independence, this book uses a custom

file format. While a binary file format is more efficient
(and what most real games use), for simplicity, this
book’s model file format is a JSON (JavaScript Object
Notation) text format. This makes it very easy to
manually edit a model file and validate that the model
file loads properly. You will eventually explore how to
use a binary format in Chapter 14, “Level Files and
Binary Data.”
Listing 6.3 shows the representation of a cube in the

gpmesh file format this book uses. The first entry
specifies the version, which currently is 1. The next line
specifies the vertex format for the model. Recall that in
Chapter 5, “OpenGL,” you used three floats for position
and two floats for texture coordinates as the vertex
format. The PosNormTex format specified here adds
three floats for the vertex normal in between the position
and texture coordinates. For now, don’t worry about
what a vertex normal is; we revisit this topic in the
lighting discussion later in this chapter.
Listing 6.3 Cube.gpmesh
"version":1,
"vertexformat":"PosNormTex",
"shader":"BasicMesh",
"textures":[
"Assets/Cube.png"
],
"vertices":[
[1.0,1.0,-1.0,0.57,0.57,-0.57,0.66,0.33],
[1.0,-1.0,-1.0,0.57,-0.57,-0.57,0.66,0.0],
[-1.0,-1.0,-1.0,-0.57,-0.57,-0.57,1.0,0.33],
[-1.0,1.0,-1.0,-0.57,0.57,-0.57,0.66,0.66],
[1.0,0.99,1.0,0.57,0.57,0.57,0.33,0.33],
[0.99,-1.0,1.0,0.57,-0.57,0.57,0.0,0.0],
[-1.0,-1.0,1.0,-0.57,-0.57,0.57,0.66,0.33],
[-1.0,1.0,1.0,-0.57,0.57,0.57,0.33,0.66]
],
"indices":[
[1,3,0],
[7,5,4],
[4,1,0],
[5,2,1],
[2,7,3],
[0,7,4],
[1,2,3],
[7,6,5],
[4,5,1],
[5,6,2],
[2,6,7],
[0,3,7]
The shader entry specifies which shader program you

should use for drawing the model. (You’ll define the
BasicMesh shader program later in this chapter.) Next,
the textures array specifies a list of textures associated
with the model.
The final two elements, vertices and indices, specify

the vertex and index buffers for the model. Each row in
vertices is one individual vertex, while each row in
indices is one triangle.
Of course, a model file format isn’t particularly useful if

there’s no way to create a model of that format in a
modeling program. To solve this, the GitHub code
repository provides two exporters in the Exporter
directory. One is an export script for the Blender
modeling program, and it supports the basic style of
mesh used throughout most of this book. The other is an
exporter plug-in for Epic Unreal Engine, which can
export not only meshes but also the animation data used
in Chapter 12, “Skeletal Animation.” The code for these
exporters is very specific to Blender and Unreal,
respectively, so we omit the discussion here. However,
interested readers can peruse the code in the repository.
Each exporter also includes a text file with instructions
on how to use each exporter with its respective program.
Updating the Vertex Attributes

Because the gpmesh file uses three vertex attributes
(position, normal, and texture coordinates) for each
vertex, let’s assume that all meshes use this format
for now. This means even the quad mesh will need
normals. Figure 6.2 shows this new vertex layout.
Figure 6.2 Vertex layout with a position, normal,
and texture coordinates
Every vertex array will use the new vertex layout, so the
constructor for VertexArray changes to specify this
new layout. Most notably, the size of each vertex is now
eight floats, and you add an attribute for the normal:
// Position is 3 floats
glVertexAttribPointer(0, 3, GL_FLOAT, GL_FALSE, 8 * sizeof(float), 0);
// Normal is 3 floats
glVertexAttribPointer(1, 3, GL_FLOAT, GL_FALSE, 8 * sizeof(float),
reinterpret_cast<void*>(sizeof(float) * 3));
// Texture coordinates is 2 floats
glVertexAttribPointer(2, 2, GL_FLOAT, GL_FALSE, 8 * sizeof(float),

reinterpret_cast<void*>(sizeof(float) * 6))
Next, you change the Sprite.vert to also reference the

new vertex layout:
// Attribute 0 is position, 1 is normal, 2 is tex coords
layout(location = 1) in vec3 inNormal;
Finally, the quad created in

Game::CreateSpriteVerts adds three additional
floats for the normals. (They can be zero because they
aren’t used by the sprite shader program.) With these
changes, sprites still draw correctly with the new vertex
layout.
Loading a gpmesh File

Because the gpmesh format is in JSON, there are
many libraries you can use to parse the JSON. This
book uses RapidJSON (http://rapidjson.org), which
supports efficient reading of JSON files. As with
textures in Chapter 5, here you encapsulate mesh
loading in a Mesh class. Listing 6.4 shows the
declaration of Mesh.
Listing 6.4 Mesh Declaration
class Mesh
public:
Mesh();
~Mesh();
// Load/unload mesh
bool Load(const std::string& fileName, class Game* game);
void Unload();
// Get the vertex array associated with this mesh
class VertexArray* GetVertexArray() { return mVertexArray; }
// Get a texture from specified index
class Texture* GetTexture(size_t index);
// Get name of shader
const std::string& GetShaderName() const { return mShaderName; }
// Get object space bounding sphere radius
float GetRadius() const { return mRadius; }
private:
// Textures associated with this mesh

std::vector<class Texture*> mTextures;
// Vertex array associated with this mesh
class VertexArray* mVertexArray;
// Name of shader specified by mesh
std:string mShaderName;
// Stores object space bounding sphere radius
float mRadius;
};
As before, there are a constructor and a destructor as

well as Load and Unload. However, notice that Load
also takes in a pointer to Game. This is so that Mesh can
access any textures associated with the meshes because
the game has a map of the loaded textures.
The member data of Mesh contains a vector of texture

pointers (one for each texture specified in the gpmesh
file), a VertexArray pointer (for the vertex/index
buffer), and a radius for the object space bounding
sphere. This bounding sphere radius computes as the
mesh file loads. The radius is simply the distance
between the object space origin and the point farthest
away from the origin. Computing this on load means that
later, any collision components that require an object
space radius have access to the data. Chapter 10,
“Collision Detection,” covers collision in detail. As a
performance improvement, you could instead compute
this radius in the gpmesh exporter.
The implementation of Mesh::Load is lengthy but not

especially interesting. It constructs two temporary
vectors: one for all the vertices and one for all the
indices. When it’s finished reading in all the values
through the RapidJSON library, it constructs a
VertexArray object. To see the full implementation of
Mesh::Load, open this chapter’s game project in the
Chapter06 directory on GitHub.
You also create a map of loaded meshes and a GetMesh

function in Game. As with textures, GetMesh determines
whether a mesh is already in the map or you need to load
it from disk.
DRAWING 3D MESHES
Once 3D meshes are loading, the next step is to draw
them. However, there are a lot of topics to touch on
before the 3D meshes start showing up.
Before we dive into these topics, it’s time for some

housekeeping. The amount of rendering-specific code in
Game has grown to the point that it’s difficult to separate
out what’s related to rendering and what isn’t. Adding
3D mesh drawing is only going to compound the issue.
To solve this, it makes sense to now create a separate
Renderer class that encapsulates all rendering code.
This is the same code that was previously in Game, just
moved into a separate class. Listing 6.5 provides an
abbreviated declaration of Renderer.
Listing 6.5 Abbreviated Renderer Declaration

class Renderer
public:
Renderer();
~Renderer();
// Initialize and shutdown renderer
bool Initialize(float screenWidth, float screenHeight);
void Shutdown();
// Unload all textures/meshes
void UnloadData();
// Draw the frame
void Draw();
void AddSprite(class SpriteComponent* sprite);
void RemoveSprite(class SpriteComponent* sprite);
class Texture* GetTexture(const std::string& fileName);

class Mesh* GetMesh(const std::string& fileName);
private:
bool LoadShaders();
void CreateSpriteVerts();
// Member data omitted
// ...
};
The Game class then constructs and initializes an

instance of Renderer in Game::Initialize. Note
that the Initialize function takes in the width/height
of the screen and saves these parameters in member
variables. Next, Game::GenerateOutput calls Draw
on the renderer instance. The map of loaded textures,
map of loaded meshes, and vector of
SpriteComponents also move over to Renderer. This
requires some changes throughout the codebase.
However, none of this code is new; it’s just moved. From
this point forward, all rendering-related code will go in
Renderer instead of Game.
Transforming to Clip Space, Revisited

Recall that with the OpenGL 2D rendering
implemented in Chapter 5, the simple view-
projection matrix scales down world space
coordinates into clip-space coordinates. For a 3D
game, this type of view-projection matrix is
insufficient. Instead, you need to decompose the
view-projection matrix into separate view and
projection matrices.
View Matrix
The view matrix represents the position and
orientation of the camera, or “eye” in the world.
Chapter 9, “Cameras,” covers several different
implementations of cameras, but for now let’s keep it
simple. At a minimum, a look-at matrix represents
the position and orientation of the camera.
In the typical construction of a look-at matrix, there are

three parameters: the position of the eye, the target
position the eye “looks at,” and the up direction. Given
these parameters, you first compute four different
vectors:
These vectors then define the elements of the look-at
matrix as follows:
A quick way to make the camera move is to create an

actor for the camera. The position of this actor
represents the eye position. The target position is then
some point in front of the camera actor. For the up
direction, +z works if the actor can’t flip upside down
(which currently is the case). Pass in these parameters to
Matrix4::CreateLookAt, and you have a valid view
matrix.
For example, if the camera actor is mCameraActor, the

following code constructs a view matrix:
// Location of camera
Vector3 eye = mCameraActor->GetPosition();
// Point 10 units in front of camera
Vector3 target = mCameraActor->GetPosition() +
mCameraActor->GetForward() * 10.0f;
Matrix4 view = Matrix4::CreateLookAt(eye, target, Vector3::UnitZ);
Projection Matrix
The projection matrix determines how the 3D
world flattens into the 2D world drawn onscreen.
Two types of projection matrices are common in 3D
games: orthographic and perspective.
In an orthographic projection, objects farther away

from the camera are the same size as objects closer to the
camera. This means that players will not be able to
perceive whether objects are closer or farther away from
the camera. Most 2D games use an orthographic
projection. Figure 6.3 illustrates a scene rendered with
an orthographic projection.
Figure 6.3 (a) Top-down view of an orthographic
projection and (b) the resulting 2D image onscreen
In a perspective projection, objects farther away from

the camera are smaller than closer ones. Thus, players
perceive that there’s depth in the scene. Most 3D games
use this sort of projection, and you’ll use it for the game
project in this chapter. Figure 6.4 shows the same 3D
scene as Figure 6.3, except this time using a perspective
projection.
Figure 6.4 (a) Top-down view of a perspective
projection and (b) the resulting 2D image onscreen
Each of these projections has a near plane and a far

plane. The near plane is typically very close to the
camera. Anything between the camera and the near
plane is not visible onscreen. This is why games have
objects partially disappear if the camera gets too close to
them. Similarly, the far plane is far away from the
camera, and anything past it is not visible. Games
sometimes allow the player to reduce the “draw distance”
to improve performance. This is often just a matter of
pulling in the far plane.
The orthographic projection matrix has four parameters:

the width of the view, the height of the view, the distance
to the near plane, and the distance to the far plane. Given
these parameters, this is the orthographic projection
matrix:
Note that this orthographic projection matrix is like the

SimpleViewProjection matrix in Chapter 5, except there
are additional terms accounting for the near and far
planes.
The perspective projection has an additional parameter

called the horizontal field of view (FOV). This is the
horizontal angle around the camera that’s visible in the
projection. Changing the field of view determines how
much of the 3D world is visible. This is the perspective
matrix:
Note that the perspective matrix changes the w
component of the homogeneous coordinate. The
perspective divide divides each component of the
transformed vertex by the w component, so that the w
component becomes 1 again. This is what reduces the
size of objects as they get farther from the camera.
OpenGL automatically does a perspective divide behind
the scenes.
note
We omit the derivation of the orthographic and perspective matrices here.
Both types of protection matrices have helper functions
in the Math.h library. You can use
Matrix4::CreateOrtho for an orthographic matrix
and Matrix4::CreatePerspectiveFOV for a
perspective matrix.
Calculating the View-Projection

The view-projection matrix is just the product of the
separate view and projection matrices:
The vertex shader then uses this view-projection matrix

to transform vertex positions from world space into clip
space.
Out with the Painter’s Algorithm, in with Z-

Buffering
Chapter 2, “Game Objects and 2D Graphics,”
introduces the painter’s algorithm. Recall that the
painter’s algorithm draws objects from back to front.
While this is great for 2D games, it faces
complications in 3D.
Painter’s Algorithm Blues

A major problem with the painter’s algorithm for 3D
games is that the back-to-front ordering is not static.
As the camera moves and rotates through the scene,
which object is in front or behind changes. To use the
painter’s algorithm in a 3D scene, you must sort all
the triangles in the scene from back to front,
potentially every frame. For even a slightly complex
scene, this perpetual sorting is a performance
bottleneck.
It gets even worse for a split-screen game. If player A and

player B face each other, the back- to-front ordering is
different for each player. To solve this, you must sort on
a per-view basis.
Another issue is that the painter’s algorithm can result in

a massive amount of overdraw, or writing the color of a
single pixel more than once per frame. This happens all
the time in the painter’s algorithm, as objects in the back
of the scene have pixels overwritten by closer objects. In
a modern 3D game, the process of calculating the final
color of a pixel is one of the most expensive parts of the
rendering pipeline. This is because the fragment shader
contains code for texturing, lighting, and many other
advanced techniques. Every pixel overdrawn is an
execution of the fragment shader wasted. Thus, 3D
games aim to eliminate as much overdraw as possible.
Finally, there’s an issue with overlapping triangles. Look

at the three triangles in Figure 6.5. Which one is the
farthest back? The answer is that no one triangle is the
farthest back. In this case, the only way the painter’s
algorithm can draw these triangles correctly is by
splitting one triangle in half, which isn’t ideal. For these
reasons, 3D games do not use the painter’s algorithm for
most objects.
Figure 6.5 Overlapping triangles painter’s
algorithm failure case
Z-Buffering
Z-buffering (or depth buffering) uses an additional
memory buffer during the rendering process. This
buffer, known as the z-buffer (or depth buffer),
stores data for each pixel in the scene, much like the
color buffer. But while the color buffer stores color
information, the z-buffer stores the distance from the
camera, or depth, at each pixel. Collectively, the set
of buffers that graphically represent the frame
(including the color buffer, z-buffer, and others) is
the frame buffer.
At the start of a frame, you need to clear the z-buffer

(much like the color buffer). Instead of clearing to a
color, you clear each pixel to the maximum depth value
in normalized device coordinates, which is 1.0. During
rendering, before drawing a pixel, z-buffering computes
the depth at the pixel. If the depth at the pixel is less than
the current depth value stored in the z-buffer (meaning
it’s closer), that pixel draws to the color buffer. The z-
buffer then updates the depth value for that pixel.
Figure 6.6 shows a visualization of the z-buffer for a

scene. Because the sphere is closer than the cube, its z-
buffer values are closer to zero (and thus closer to black).
The first object drawn in a frame will always have all its
pixels’ color and depth information written into the color
and z-buffers, respectively. But when drawing the second
object, only pixels with depths closer than the values in
the z-buffer draw. Listing 6.6 gives pseudocode for the z-
buffer algorithm.
Figure 6.6 A sample scene (left) and its

corresponding z-buffer (right)
Listing 6.6 Z-Buffering Pseudocode

// zBuffer[x][y] grabs depth at that pixel

foreach MeshComponent m in scene
foreach Pixel p in m
float depth = p.Depth()
if zBuffer[p.x][p.y] < depth
p.draw
endif
endfor
endfor
With z-buffering, drawing the scene in any arbitrary

order will look correct if there are no objects with
transparency. This is not to say that the order is
irrelevant. For example, drawing a scene back-to-front
yields the same amount of overdraw as the painter’s
algorithm. Conversely, drawing a scene front-to-back
yields zero overdraw. But the gain of z-buffering is that
any arbitrary order works. And because z-buffering is on
a per-pixel basis, not a per-object or per-triangle basis, it
will work even for the overlapping triangles in Figure 6.5.
Luckily, z-buffering is not something a graphics

programmer implements anymore; you merely need to
enable it. OpenGL supports depth buffering with
minimal effort. (OpenGL uses the term depth buffer
instead of z-buffer.) First, prior to creating the OpenGL
context, you need to request a depth buffer (24-bit is a
typical size):
SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 24);
Then, the following call enables depth buffering:
glEnable(GL_DEPTH_TEST);
The glClear function handles clearing the depth buffer.

One call can clear both the color and depth buffers:
glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
Although z-buffering works well, there are some issues.

For one, transparent objects don’t work with z-buffering
as specified. Suppose a game has semi-transparent
water, and under this water is a rock. In z-buffering,
drawing the water first writes to the z-buffer and
prevents the rock from drawing, since the rock has a
higher depth.
The solution to this dilemma is to render opaque objects

first, using z-buffering. Then disable depth buffer writes
and render transparent objects back-to-front. As the
pixels render, you should test the depth at each pixel to
ensure that transparent pixels behind opaque objects
don’t draw. Although this means the transparent objects
render with the painter’s algorithm, the number of
transparent objects is hopefully very small.
Although you don’t need to use transparent 3D objects in

this book, remember that sprite rendering uses alpha
blending to support textures with transparency. Because
this doesn’t play nicely with z-buffering, you must
disable alpha blending for 3D objects and then reenable
it for sprites. Likewise, sprites must render with z-
buffering disabled.
This naturally leads to rendering in two phases: First,

render all 3D objects with alpha blending disabled and z-
buffering enabled. Then render all sprites with alpha
blending enabled and z-buffering disabled. When you do
this, all 2D sprites appear on top of the 3D scene. This is
okay because a 3D game uses 2D sprites only for UI or
HUD elements.
The BasicMesh Shader

Recall that earlier in this chapter, you modified the
Sprite.vert shader file to include support for
vertex normals in the vertex layout. It turns out that
this modified code for the sprite vertex shader and
the original code for the Sprite.frag shader file
from Chapter 5 also works for full 3D meshes. You
set the view-projection matrix uniform to a different
value for 3D meshes, but the actual vertex/fragment
shader code is fine as is. So for now, the
BasicMesh.vert/BasicMesh.frag shader files
are just copies of the Sprite.vert/Sprite.frag
shader files.
Next, add a Shader* member variable for the mesh
shader to Renderer, as well as separate Matrix4
variables for the view and projection matrices.
Renderer::InitShaders then loads the BasicMesh
shader (with code very similar to code for loading the
sprite shader) and initializes the view and projection
matrices. You initialize the view matrix to a look-at
matrix facing down the x-axis and the projection matrix
to a perspective matrix:
mMeshShader->SetActive();
// Set the view-projection matrix
mView = Matrix4::CreateLookAt(
Vector3::Zero, // Camera position
Vector3::UnitX, // Target position
Vector3::UnitZ // Up
);
mProjection = Matrix4::CreatePerspectiveFOV(
Math::ToRadians(70.0f), // Horizontal FOV
mScreenWidth, // Width of view
mScreenHeight, // Height of view
25.0f, // Near plane distance
10000.0f // Far plane distance

);
mMeshShader->SetMatrixUniform("uViewProj", mView * mProjection);
For simplicity, we assume here that all meshes use the

same shader (ignoring the shader property stored in the
gpmesh file). In Exercise 6.1 you add support for
different mesh shaders.
In any event, now that there’s a shader for meshes, the

next step is to create a MeshComponent class to draw
3D meshes.
The MeshComponent Class

Recall that all the code for transforming vertices
from object space into clip space is in the vertex
shader. The code for filling in the color of each pixel
is in the fragment shader. This means that the
MeshComponent class doesn’t need to do much for
drawing.
Listing 6.7 provides the declaration of MeshComponent.

Note that unlike SpriteComponent, MeshComponent
does not have a draw order variable. This is because the
order doesn’t matter because 3D meshes use z-buffering.
The only member data is a pointer to the associated
mesh and a texture index. Because a gpmesh can have
multiple associated textures, the index determines which
texture to use when drawing the MeshComponent.
Listing 6.7 MeshComponent Declaration
class MeshComponent : public Component
public:
MeshComponent(class Actor* owner);
~MeshComponent();
// Draw this mesh component with provided shader
virtual void Draw(class Shader* shader);
// Set the mesh/texture index used by mesh component
virtual void SetMesh(class Mesh* mesh);
void SetTextureIndex(size_t index);
protected:
class Mesh* mMesh;
size_t mTextureIndex;
};
The Renderer then has a vector of MeshComponent

pointers and functions to add and remove these
components. The constructor and destructor of
MeshComponent call these add/remove functions.
The Draw function, shown in Listing 6.8, first sets the
world transform matrix uniform. MeshComponent
directly uses the owning actor’s world transform matrix
because there’s no extra scale needed, as there is in
SpriteComponent. Next, the code activates the texture
and vertex array associated with the mesh. Finally,
glDrawElements draws the triangles. The index buffer
size isn’t hard-coded here because different meshes have
different numbers of indices.
Listing 6.8 MeshComponent::Draw Implementation

void MeshComponent::Draw(Shader* shader)
if (mMesh)
// Set the world transform
shader->SetMatrixUniform("uWorldTransform",
mOwner->GetWorldTransform());
// Set the active texture
Texture* t = mMesh->GetTexture(mTextureIndex);
if (t) { t->SetActive(); }
// Set the mesh's vertex array as active
VertexArray* va = mMesh->GetVertexArray();
va->SetActive();
// Draw
glDrawElements(GL_TRIANGLES, va->GetNumIndices(),
Finally, the Renderer needs code that draws all the

mesh components. After clearing the frame buffer, the
Renderer first draws all meshes with depth buffering
enabled and alpha blending disabled. It next draws all
sprites in the same manner as before. After drawing
everything, the Renderer swaps the front and back
buffers. Listing 6.9 shows only the new code for
rendering meshes. The code recalculates the view-
projection matrix every frame to account for a moving
camera.
Listing 6.9 Drawing MeshComponents in

Renderer::Draw
// Enable depth buffering/disable alpha blend
glDisable(GL_BLEND);
// Set the basic mesh shader active
mMeshShader->SetActive();
// Update view-projection matrix
mMeshShader->SetMatrixUniform("uViewProj", mView * mProjection);
for (auto mc : mMeshComps)
mc->Draw(mMeshShader);
Because MeshComponent is just like any other

component, you can attach it to an arbitrary actor and
draw meshes for the actor. Figure 6.7 shows
MeshComponent in action, drawing a sphere and a cube
mesh.
Figure 6.7 Drawing a simple scene with
MeshComponent
LIGHTING
So far, the mesh fragment shader directly uses the
texture color as the final color for a pixel. However,
without any contrast, the scene looks dull. To
approximate concepts such as the sun or light bulbs
or simply to add variety to the scene, you need
lighting.
Revisiting Vertex Attributes

Lighting meshes require more vertex attributes than
just the vertex position and UV (texture) coordinates.
They also need vertex normals. You added this vertex
attribute earlier in this chapter. However, the
concept of a vertex normal requires further
explanation. It seems almost nonsensical because a
normal is a vector perpendicular to a surface, but a
single point is not a surface, so how can there be a
normal to a point?
You can compute a vertex normal by averaging the

normals of the triangles that contain that vertex, as in
Figure 6.8(a). This works well for smooth models but
doesn’t quite work for hard edges. For example,
rendering a cube with averaged vertex normals yields
rounded corners. To solve this, the artist creates multiple
vertices for the corners of the cube, and each vertex on
the corner has a different normal. Figure 6.8(b) shows a
cube authored in this manner.
Figure 6.8 (a) Averaged vertex normals. (b) Vertex

A on the cube uses one of three different normals,
depending on the face
Remember that all vertex attributes interpolate across

the triangle when sent to the fragment shader. This
means that any arbitrary pixel on the face of a triangle
has a normal value that’s the interpolation between the
triangle’s three vertex normals.
Types of Lights
While there are many potential choices, a handful of
light types consistently see use in 3D games. Some
lights globally affect the entire scene, whereas other
lights affect only the area around the light.
Ambient Light
Ambient light is a uniform amount of light applied
to every single object in a scene. The amount of
ambient light might differ for different levels in a
game, depending on the time of day. A level set at
night will have a much darker and cooler ambient
light than a level set in the daytime, which will be
brighter and warmer.
Because it provides an even amount of lighting, ambient

light does not light different sides of objects differently.
It is a global amount of light uniformly applied to every
part of every object in a scene. This is akin to the sun on
a cloudy day in nature, as in Figure 6.9(a).
Figure 6.9 Examples in nature of ambient light (a)
and directional light (b)
In code, the simplest representation of ambient light is

an RGB color value that represents both the color and
intensity of the light. For example, (0.2, 0.2, 0.2) is
darker than (0.5, 0.5, 0.5).
Directional Light
A directional light is a light emitted from a specific
direction. Like ambient light, directional light affects
an entire scene. However, because a directional light
comes from a specific direction, it illuminates one
side of objects while leaving the other side in
darkness. An example of a directional light is the sun
on a sunny day. The direction of the light depends on
where the sun is at that time of day. The side facing
the sun is bright, while the other side is dark. Figure
6.9(b) shows a directional light at Yellowstone
National Park. (Note that in a game, shadowing is
not a property of a directional light itself. Instead,
computing shadows requires additional
calculations.)
Games that use directional lights often have only one

directional light for the entire level, representing either
the sun or the moon. But this isn’t always the case. For
example, an approximation of the lighting of a sports
stadium at night could use multiple directional lights.
In code, a directional light needs both an RGB color

value (as with ambient light) and a normalized vector for
the direction of the light.
Point Light
A point light exists at a specific point and emanates
in all directions from that point. Because it starts at a
specific point, a point light also illuminates only one
side of an object. Usually, a point light also has a
radius of influence. For example, think of a light bulb
in a dark room, as in Figure 6.10(a). There’s visible
light in the area immediately around the light, but it
slowly dissipates until it no longer adds light. The
point light doesn’t go on infinitely.
Figure 6.10 A light build point light (a) and (b) a

spotlight
In code, a point light should have an RGB color, a

position of the light, and a falloff radius that
determines how much the light value decreases as the
distance from the light increases.
Spotlight
A spotlight is much like a point light, except that
instead of traveling in all directions, it’s focused in a
cone. To simulate a spotlight, you need all the
parameters of a point light and additionally the angle
of the cone. A classic example of a spotlight is a
theater spotlight, but another example is a flashlight
in the dark. Figure 6.10(b) illustrates a spotlight.
Phong Reflection Model

To simulate lights, not only do you need their
associated data, you also need to calculate how the
lights affect the objects in the scene. A tried-and-true
method for approximating light is a bidirectional
reflectance distribution function (BRDF),
which is a function that approximates how light
bounces off surfaces. There are many different types
of BRDFs, but a classic one is the Phong reflection
model.
The Phong model is a local lighting model because it

doesn’t calculate secondary reflections of light. In other
words, the reflection model lights each object as if it’s the
only object in the entire scene. In the real world, shining
a red light on a white wall will fill the rest of the room
with a reddish color. However, this does not happen in
the Phong model.
The Phong model divides light into three distinct

components: ambient, diffuse, and specular. Figure 6.11
illustrates these components. All three components
consider the color of the surface as well as the color of
the light affecting the surface.
Figure 6.11 Phong reflection model
The ambient component is the overall illumination of

the scene. Thus, it makes sense to directly tie the
ambient component to the ambient light. Because
ambient light applies evenly to the entire scene, the
ambient component is independent of any other lights
and the camera.
The diffuse component is the primary reflection of

light off the surface. Any directional lights, point lights,
or spotlights affecting the object affect the diffuse
component. The diffuse component calculation uses both
the normal of the surface and a vector from the surface to
the light. The position of the camera does not affect the
diffuse component.
The final component in the Phong model is the

specular component. This approximates shiny
reflections off a surface. An object with a high degree of
specularity, such as a polished metal object, has stronger
highlights than an object painted in matte black. As with
the diffuse component, the specular component depends
on both the light vector and the normal of the surface.
However, specularity also depends on the position of the
camera. This is because looking at a shiny object from
different angles changes the perceived reflections.
Figure 6.12 shows the Phong reflection from a side view.

Computing Phong reflection requires a series of
calculations that include several variables:
—Normalized surface normal
—Normalized vector from the surface to the light
—Normalized vector from the surface to the camera (eye) position
—Normalized reflection of – about
α—Specular power (which determines the shininess of the object)
In addition, there are colors for the lights:
ka—Ambient color
kd—Diffuse color
ks—Specular color
Figure 6.12 Diagram of Phong reflection
calculations (vectors not to scale)
In the Phong reflection model, you calculate the light

applied to a surface as follows:
Note that the diffuse and specular component computes

for all lights in the scene, but there’s only one ambient
component. The test ensures that the light affects
only surfaces facing the light.
In any event, the Phong equation as described here yields

a color for all lights in the scene. The final color of the
surface is the light color multiplied by the color of the
surface. Because both the light and surface color are RGB
values, you use component-wise multiplication.
A more complex implementation separates the surface

color into separate ambient, diffuse, and specular colors.
In this implementation, the equations change to multiply
each separate color by each component rather than only
one multiplication at the end.
One remaining question is how frequently to compute

the BRDF. There are three common options: once per
surface (flat shading), once per vertex (Gouraud
shading), or once per pixel (Phong shading).
Although per-pixel lighting is more computationally
expensive, modern graphics hardware can easily handle
it. It’s conceivable that some games might choose other
types of shading for aesthetic reasons, but this chapter
sticks to per-pixel lighting.
Implementing Lighting
This section covers how to add ambient and
directional lights to the game. Implementing this
requires changes to both the vertex and fragment
shaders. The BasicMesh.vert/.frag shaders are
a starting point for the new Phong.vert/.frag
shaders. (Remember that this shader code is in
GLSL, not C++.) You’ll then change it so all meshes
use this new Phong shader.
Because the lighting is per pixel, the Phong fragment

shader needs several additional uniforms: a camera
position, an ambient light color, and several variables for
a directional light (see Listing 6.10).
Listing 6.10 Phong.frag Lighting Uniforms

// Create a struct for directional light
struct DirectionalLight
// Direction of light
vec3 mDirection;
// Diffuse color
vec3 mDiffuseColor;
// Specular color
vec3 mSpecColor;
};
// Uniforms for lighting
// Camera position (in world space)
uniform vec3 uCameraPos;
// Ambient light level
uniform vec3 uAmbientLight;
// Specular power for this surface
uniform float uSpecPower;
// Directional Light (only one for now)

uniform DirectionalLight uDirLight;
Note the declaration of a DirectionalLight struct.

GLSL supports struct declarations, much like those in
C/C++. Next, you declare a corresponding
DirectionalLight struct in the C++ code and add
two member variables to Renderer for the ambient light
and directional light.
Back in C++, the glUniform3fv and glUniform1f

functions set 3D vector and float uniforms, respectively.
You create two new functions in Shader,
SetVectorUniform and SetFloatUniform, to call
these functions. The implementation of both these
functions is like that of the SetMatrixUniform
function in Chapter 5.
A new function in Renderer called

SetLightUniforms handles setting the new uniform
values:
void Renderer::SetLightUniforms(Shader* shader)
// Camera position is from inverted view
Matrix4 invView = mView;
invView.Invert();
shader->SetVectorUniform("uCameraPos", invView.GetTranslation());
// Ambient light
shader->SetVectorUniform("uAmbientLight", mAmbientLight);
// Directional light
shader->SetVectorUniform("uDirLight.mDirection", mDirLight.mDirection);
shader->SetVectorUniform("uDirLight.mDiffuseColor",
mDirLight.mDiffuseColor);
shader->SetVectorUniform("uDirLight.mSpecColor", mDirLight.mSpecColor);
Note that this function uses dot notation to reference

specific members of the uDirLight struct.
Extracting the camera position from the view matrix

requires inverting the view matrix. After inverting the
view matrix, the first three components of the fourth row
(returned by the GetTranslation member function)
correspond to the world space position of the camera.
Next, you update the gpmesh file format so that you can
specify the specular power of a mesh’s surface with the
specularPower property. You then update the
Mesh::Load code to read in this property, and set the
uSpecPower uniform in MeshComponent::Draw right
before drawing the mesh.
Back in GLSL, you must make some changes to the

vertex shader in Phong.vert. Both the camera position
and the directional light’s direction are in world space.
However, the gl_Position computed in the vertex
shader is in clip space. Getting the correct vector from
the surface to the camera requires a position in world
space. Furthermore, the input vertex normals are in
object space, but they also need to be in world space.
This means the vertex shader must compute both the
world space normal and the world space position and
send these to the fragment shader via out variables:
// Normal (in world space)
out vec3 fragNormal;
// Position (in world space)
out vec3 fragWorldPos;
Similarly, you declare fragNormal and fragWorldPos

as in variables in the fragment shader. Next, the main
function of the vertex shader, in Listing 6.11, computes
both fragNormal and fragWorldPos. The .xyz
syntax, known as a swizzle, is a shorthand for extracting
the x, y, and z components from a 4D vector and creating
a new 3D vector with those values. This effectively
converts between a vec4 and vec3.
The code also converts the normal into homogenous

coordinates so that multiplication by the world
transform matrix works. However, the w component
here is 0 instead of 1. This is because the normal is not a
position, so translating the normal makes no sense.
Setting the w component to 0 means the translation
component of the world transform matrix zeros out in
the multiplication.
Listing 6.11 Phong.vert Main Function

void main()
// Transform position to world space
pos = pos * uWorldTransform;
// Save world position
fragWorldPos = pos.xyz;
// Transform to clip space
gl_Position = pos * uViewProj;
// Transform normal into world space (w = 0)
fragNormal = (vec4(inNormal, 0.0f) * uWorldTransform).xyz;

The fragment shader, in Listing 6.12, computes the

Phong reflection model as outlined in the equations in
the previous section. Note that you must normalize
fragNormal because OpenGL interpolates the vertex
normal across the face of the triangle. Interpolating two
normalized vectors does not guarantee a normalized
vector at each step of the interpolation, so you must
renormalize it.
Because the directional light emits from a direction, the

vector from the surface to the light is just the negation of
the light’s direction vector. The fragment shader uses a
handful of new GLSL functions. The dot function
computes a dot product, reflect computes a reflection
vector, max selects the maximum of two values, and pow
computes a power. The clamp function restricts the
value of each component of the passed in vector to the
range specified. In this case, the valid light values are
from 0.0 (no light) to 1.0 (maximum light of that color).
The final color is the texture color multiplied by the
Phong light.
One edge case occurs when the dot product between R

and V is negative. In this case, the specular component
might end up negative, or effectively remove light from
the scene. The max function call prevents this because if
the dot product is negative, max selects 0.
Listing 6.12 Phong.frag Main Function

void main()
// Surface normal
vec3 N = normalize(fragNormal);
// Vector from surface to light
vec3 L = normalize(-uDirLight.mDirection);
// Vector from surface to camera
vec3 V = normalize(uCameraPos - fragWorldPos);
// Reflection of -L about N
vec3 R = normalize(reflect(-L, N));
// Compute phong reflection
vec3 Phong = uAmbientLight;
float NdotL = dot(N, L);
if (NdotL > 0)
vec3 Diffuse = uDirLight.mDiffuseColor * NdotL;

vec3 Specular = uDirLight.mSpecColor *
pow(max(0.0, dot(R, V)), uSpecPower);
Phong += Diffuse + Specular;
// Final color is texture color times phong light (alpha = 1)
outColor = texture(uTexture, fragTexCoord) * vec4(Phong, 1.0f);
Figure 6.13 shows the Phong shader in action, lighting

the sphere and cube from Figure 6.10. This figure uses
the following light values:
Ambient light—Dark gray (0.2, 0.2, 0.2)
Directional direction—Down and to the left (0, -0.7, -0.7)
Directional diffuse color—Green (0, 1, 0)
Directional specular color—Bright green (0.5, 1, 0.5)

Figure 6.13 Phong shader in action
In Figure 6.13, the specular power of the sphere is 10.0

and the specular power of the cube is 100.0f to make the
sphere shinier than the cube.
GAME PROJECT
This chapter’s game project implements most of the
topics covered: mesh loading, a MeshComponent,
and the Phong shader. Figure 6.14 shows the final
version of this chapter’s game project. The code is
available in the book’s GitHub repository, in the
windows.sln on Windows and Chapter06-
The LoadData function in Game instantiates several

different actors for the objects in the world. A simple
CameraActor allows the camera to move through the
world. Use the W and S keys to move forward and back
and the A and D keys to yaw the camera. (Chapter
9 discusses more complex cameras. The current camera
is a simple version of a first-person camera.)
The sprite elements on the screen, such as health and

radar, don’t do anything yet. They are onscreen just to
demonstrate that sprite rendering still works. Chapter 11,
“User Interfaces,” shows how to implement some UI
features.
SUMMARY
This chapter covers the process of transitioning from
a 2D game world to a 3D game world. Actors now
have a transform with a 3D position and a
quaternion for rotations about an arbitrary axis.
A 3D scene also needs a more complex view-projection

matrix. Using a look-at matrix is a preferred method to
create a view matrix. A projection matrix can be either
orthographic or perspective, but only a perspective
projection gives the scene depth. 3D games also shy away
from using the painter’s algorithm and instead use z-
buffering to determine which pixels are visible.
The simple gpmesh file format contains enough

information to create vertex and index buffers for 3D
models at runtime—without including the extraneous
data of a complex file format. The MeshComponent class
can attach to any actor and implements 3D mesh
rendering (through the shaders).
A game might use many types of lights. Ambient and

directional lights affect the entire scene, while point and
spotlights affect only certain areas. One method to
approximate how light affects the scene is the Phong
reflection model. The Phong model has three
components: ambient, diffuse, and specular.
ADDITIONAL READING
Rendering is a highly specialized area of game
programming, and excelling in rendering requires a
strong foundation in mathematics. There are many
excellent resources available. Thomas Akenine-
Moller’s book, although somewhat dated, is a
popular reference for rendering programmers—and
an updated fourth edition is forthcoming. Although
this book uses OpenGL, there are alternative
graphics APIs. For PC and Xbox, the DirectX API
dominates. Frank Luna’s book covers how to use
DirectX 11. Finally, Matt Pharr’s text is an excellent
overview of a realistic lighting technique called
physically based rendering.

Luna, Frank. Introduction to 3D Game
Programming with DirectX 11. Dulles:
Mercury Learning and Information, 2012.
Pharr, Matt, Wenzel Jakob, and Greg
Humphreys. Physically Based Rendering:
From Theory to Implementation, 3rd
edition. Cambridge: Elsevier, 2017.
EXERCISES
This chapter’s exercises involve adding
improvements to the game project. In the first
exercise you add support for different meshes
rendering with different shaders. In the second
exercise you add point lights, which provide a great
deal of flexibility in the lighting of the game.
Exercise 6.1
Modify the mesh rendering code so that it’s possible
to draw different meshes with different shaders. This
means storing the different mesh shaders in a map
and ensuring that each shader has its uniforms set
properly.
However, constantly switching between shaders is

inefficient. To solve this, group mesh components by
their corresponding shader. For example, if there are 10
meshes drawn with the BasicMesh shader and 5 drawn
with the Phong shader, the code shouldn’t alternate
between the two shaders repeatedly. Instead, draw all
meshes that use BasicMesh first and then draw all
meshes that use Phong.
To test this, modify the gpmesh files so that some draw

with BasicMesh and others draw with Phong. Remember
that because the gpmesh files are JSON, you can use any
text editor to edit them.
Exercise 6.2
Because a point light affects a limited radius, these
lights can add a lot to a scene. Modify the Phong
shader so that it also supports a maximum of four
point lights in the scene. Create a struct for point
lights much like the struct for directional lights. This
struct needs a position of the light, diffuse color,
specular color, specular power, and a radius of
influence. Then create an array of point lights as a
uniform. (Arrays work in GLSL just like in C/C++.)
The Phong equations are the same, except the code now
needs to consider all lights for specular and diffuse. In
addition, a point light should affect a pixel only if the
pixel is within that light’s radius. To test this, create
different point lights at different positions and with
different colors.
CHAPTER 7
AUDIO
Though sometimes overlooked, audio is an

important part of games. Whether it’s to
provide audio cues for gameplay situations
or enhance the overall atmosphere, quality
sound adds a lot to games.
Leveraging the powerful FMOD API, this
chapter covers how to bootstrap an audio
system that goes well beyond simply playing
sound files. Covered topics include using
sound events, adding positional 3D audio,
mixing sounds, and adding effects.
BOOTSTRAPPING AUDIO
A rudimentary game audio system loads and plays
back standalone sound files (such as WAV or OGG
files) as needed. Although this approach is functional
—and might be perfectly acceptable for a simple 2D
game—it has limitations. In many cases, a single
game action does not correspond to a single sound
file. Suppose a game features a character that runs
around. Every time the character’s foot hits the
ground, a footstep sound should play. If there were
only a single footstep sound file played repeatedly, it
would quickly become repetitive.
At the very least, rather than having a single footstep

sound file, you might want 10 different sound files for
variety. Each time the player takes a step, the game
might want to randomly pick one of these 10 footstep
sound files. Or maybe the player can walk on different
surfaces, and a footstep on grass sounds different than a
footstep on concrete. In this case, the game needs a way
to choose from the correct set of footstep sounds based
on the surface the player is walking on.
Another consideration is that the game can only play a

limited number of sounds simultaneously. You use
sound channels to track the sounds that are playing,
and there is some limit on the number of channels.
Imagine a game where there are several enemies
onscreen at any one time. If each enemy individually
plays footstep sounds, it not only overwhelms the ear of
the player but may also take up all available channels.
Certain sounds, such as the player character attacking
the enemy, are far more important than an enemy’s
footstep. Different sounds might therefore need different
priorities.
Now consider a 3D game with a fireplace. Imagine that

as the player character moves through the game world,
the fireplace sound plays at the same volume level from
all speakers. It doesn’t matter if the player stands right
next to the fireplace or is hundreds of feet away; the
sound plays at the same level. Not only is this annoying,
but it’s unrealistic. The game needs to account for the
distance between the player and the fireplace and
calculate a volume based on this distance.
So even though games need sound files to play audio,

they also need additional information to play these
sounds correctly. Ideally, the decision-making power of
what sounds “correct” shouldn’t be the audio
programmer’s. Much like how a 3D artist creates models
in a specialized modeling program, ideally sound
designers construct dynamic sounds using external tools
designed for their skill sets.
FMOD
Designed by Firelight Technologies, FMOD
(https://fmod.com) is a popular sound engine for
video games. FMOD supports any realistic game
platform, including Windows, Mac, Linux, iOS,
Android, HTML5, and every modern console. The
current version of FMOD has two distinct
components: FMOD Studio, which is an external
authoring tool for sound designers, and the FMOD
API (application programming interface), which
integrates into games that use FMOD.
The FMOD Studio tool grants a lot of power to sound

designers and can implement many of the features
discussed earlier. A sound event can correspond to one
or more sound files, and these events can have
parameters that dynamically drive the behavior of the
sound events. FMOD Studio also allows designers to
control how the different sounds mix together. For
example, the designer can place music and sound effects
on separate tracks and then adjust the volume of the
tracks separately.
note
This chapter doesn’t cover how to use FMOD Studio, but there are excellent
references available on the official FMOD website, among other places. For
interested readers, the FMOD Studio project file used for the audio content in
this chapter is in the GitHub repository, in the FMODStudio/Chapter07
directory.
The FMOD API has two parts. The FMOD Low Level API
is the foundation for FMOD. It contains functionality to
load and play sounds, manage channels, update sounds
in a 3D environment, add digital effects to sound, and
more. It’s possible to use the Low Level API by itself, but
then any events created in FMOD Studio are not usable.
Supporting FMOD Studio requires the FMOD Studio
API, which builds on the Low Level API. However, using
the FMOD Studio API does not preclude an audio
programmer from accessing the Low Level API if needed.
For the most part, this chapter uses the FMOD Studio
API.
Installing FMOD
Because of FMOD’s licensing terms, the book’s
source code on GitHub does not include the FMOD
library and header files. Luckily, FMOD is free to
download and has very favorable licensing terms for
commercial projects. (See the FMOD site for details.)
To download the FMOD library, go to the FMOD site
(https://fmod.com) and create an account.
Once you have an account on the FMOD website, click

the Download link. From here, find the downloads for
the FMOD Studio API. Make sure you select the version
1.09.09 from the Version dropdown. (Version 1.10.x or
newer may not work with this chapter’s code.) Next,
select Windows if you’re developing on Windows or Mac
if you’re developing on Mac.
On Windows, run the installer and choose the default

installation directories. Choosing a different directory
doesn’t work out of the box because the Visual Studio
project file points directly to the default directories.
However, if you really want to install the FMOD API to
another directory, you can change the project file (which
means changing the include directories, the library
directories, and the post-build step that copies DLL files
to the executable directory).
On Mac, the FMOD API download is a DMG package file.

Open this package file and copy all its contents into the
External/FMOD directory in your copy of the book’s
source code. You should end up with an
External/FMOD/FMOD Programmers API directory
after copying.
To make sure the installation worked properly, try

opening the Chapter07/Chapter07-Windows.sln
file on a PC or Chapter07-mac.xcodeproj on a Mac
and make sure you can compile and run the code.
note
With the exception of Chapter 8, “Input Systems,” every chapter after this
one also uses the audio code from this chapter. Therefore, it’s important to
ensure that you install FMOD properly, or none of the subsequent chapters’
projects will run.
Creating an Audio System

Much as the Renderer class is separate from the
Game, it’s sensible to create a new AudioSystem
class that handles audio. This helps ensure that the
FMOD API calls aren’t all over the codebase.
Listing 7.1 shows the initial declaration of

AudioSystem. The declarations of the Initialize,
Shutdown, and Update functions are standard at this
point. The member variables include pointers to the
FMOD Studio system as well as the Low Level API
system. You’ll mostly use the mSystem pointer, but this
listing includes a mLowLevelSystem pointer as well.
Listing 7.1 Initial AudioSystem Declaration

class AudioSystem
public:
AudioSystem(class Game* game);
~AudioSystem();
bool Initialize();
void Shutdown();
void Update(float deltaTime);
private:
class Game* mGame;
// FMOD studio system
FMOD::Studio::System* mSystem;
// FMOD Low-level system (in case needed)
FMOD::System* mLowLevelSystem;
};
The header fmod_studio.hpp defines the FMOD
Studio API types. However, to avoid this include,
AudioSystem.h instead creates forward declarations of
the FMOD types. This way, you only need to include the
FMOD header in AudioSystem.cpp.
Initializing FMOD, which is handled in

AudioSystem::Initialize, involves several steps.
First, call Debug_Initialize to set up error logging:
FMOD::Debug_Initialize(
FMOD_DEBUG_LEVEL_ERROR, // Log only errors
FMOD_DEBUG_MODE_TTY // Output to stdout
);
The first parameter to Debug_Initialize controls the

verbosity of the logging messages. (The default is quite
verbose.) The second parameter specifies where to write
log messages. In this case, log messages write to stdout.
For games that have custom debug code, it’s also
possible to declare a custom callback function for all
FMOD log messages.
note
Initializing debug logging is relevant only if you’re using the logging build of
FMOD, as is the case in this chapter. Enabling error logging is extremely
useful during development, but a shipped version of a game shouldn’t
include logging.
Next, construct an instance of an FMOD Studio system
with this code:
FMOD_RESULT result;
result = FMOD::Studio::System::create(&mSystem);
if (result != FMOD_OK)
SDL_Log("Failed to create FMOD system: %s",
FMOD_ErrorString(result));
return false;
Note that the function call returns an FMOD_RESULT. An

FMOD function always returns a result value to let the
caller know if everything went okay. The
FMOD_ErrorString function converts the error code
into a human-readable message. In this case, if the
system fails to create, AudioSystem::Initialize
returns false.
After constructing the system, the next step is to call

initialize on the FMOD system:
result = mSystem->initialize(
512, // Max number of concurrent sounds
FMOD_STUDIO_INIT_NORMAL, // Use default settings
FMOD_INIT_NORMAL, // Use default settings
nullptr // Usually null
);
// Validate result == FMOD_OK...
The first parameter here specifies the maximum number

of channels. The next two parameters can adjust the
behavior of both the FMOD Studio and FMOD Low Level
APIs. For now, stick to the default parameters. You use
the last parameter if you want to use extra driver data,
but because you usually don’t, this parameter usually is
nullptr.
note
FMOD uses a naming convention in which member functions begin with a
lowercase letter. This is different from this book’s naming convention, which
uses an uppercase letter for the first letter of a member function.
Finally, you grab and save the Low Level system pointer
to complete initialization:
mSystem ->getLowLevelSystem(&mLowLevelSystem);
For now, AudioSystem’s Shutdown and Update

functions each make a single function call. Shutdown
calls mSystem ->release(), while Update calls
mSystem ->update(). FMOD requires calling the
update function once per frame. This function performs
actions such as updating 3D audio calculations.
As with Renderer, you then add an AudioSystem

pointer as a member variable to Game:
class AudioSystem* mAudioSystem;
Game::Initialize then creates and calls

mAudioSystem ->Initialize(), UpdateGame calls
mAudioSystem ->Update(deltTime), and
Shutdown calls mAudioSystem ->Shutdown.
For convenience, a Game::GetAudioSystem function

returns the AudioSystem pointer.
With these functions, FMOD now initializes and updates.

Of course, no sounds are playing yet.
Banks and Events

In FMOD Studio, events correspond to sounds
played in the game. An event can have multiple
associated sound files, parameters, information
about the event’s timing, and so on. Rather than play
back sound files directly, the game plays these
events.
A bank is a container for events, sample data, and
streaming data. Sample data is the raw audio data that
events reference. This data comes from the sound files
that the sound designer imports into FMOD Studio (such
as WAV or OGG files). At runtime, sample data is either
preloaded or loaded on demand. However, an event
cannot play until its associated sample data is in
memory. Most in-game sound effects use sample data.
Streaming data is sample data that streams into
memory in small pieces at a time. Events using
streaming data can start playing without preloading the
data. Music and dialogue files typically use streaming
data.
A sound designer creates one or more banks in FMOD

Studio. Then the game runtime needs to load in these
banks. After it loads the banks, the events contained
within are accessible.
There are two different classes associated with events in

FMOD. EventDescription contains information
about an event, such as its associated sample data,
volume settings, parameters, and so on.
EventInstance is an active instance of an event, and it
is what plays the event. In other words,
EventDescription is like a type of event, while
EventInstance is an instance of that type. For
example, if there’s an explosion event, it will globally
have one EventDescription, but it can have any
number of EventInstances based on the number of
active explosion instances.
To track loaded banks and events, you add two maps to
the private data in AudioSystem:
// Map of loaded banks
std::unordered_map<std::string, FMOD::Studio::Bank*> mBanks;
// Map of event name to EventDescription
std::unordered_map<std::string, FMOD::Studio::EventDescription*> mEvents;
Both maps have strings for their keys. The string in

mBanks is the filename of the bank, while the string in
mEvents is the name assigned by FMOD for the event.
FMOD events have names in the form of a path—for
example, event:/Explosion2D.
Loading/Unloading Banks
Loading a bank minimally requires calling the
loadBank function on the mSystem object.
However, this does not load the sample data and
does not give easy access to the event descriptions. It
makes sense to create a new function in
AudioSystem called LoadBank, as shown in Listing
7.2, that does a bit more than the minimum
loadBank call. Once the bank loads, you add the
bank to the mBanks map. You then load the sample
data for the bank. Then use getEventCount and
getEventList to get the list of all event
descriptions in the bank. Finally, you add each of
these event descriptions to the mEvents map so they
are easily accessible.
Listing 7.2 AudioSystem::LoadBank

Implementation
void AudioSystem::LoadBank(const std::string& name)
// Prevent double-loading
if (mBanks.find(name) != mBanks.end())
return;
// Try to load bank
FMOD::Studio::Bank* bank = nullptr;
FMOD_RESULT result = mSystem ->loadBankFile(
name.c_str(), // File name of bank
FMOD_STUDIO_LOAD_BANK_NORMAL, // Normal loading
&bank // Save pointer to bank
);
const int maxPathLength = 512;
if (result == FMOD_OK)
// Add bank to map
mBanks.emplace(name, bank);
// Load all non-streaming sample data
bank ->loadSampleData();
// Get the number of events in this bank
int numEvents = 0;
bank ->getEventCount(&numEvents);
if (numEvents > 0)
// Get list of event descriptions in this bank
std::vector<FMOD::Studio::EventDescription*> events(numEvents);
bank ->getEventList(events.data(), numEvents, &numEvents);
char eventName[maxPathLength];
for (int i = 0; i < numEvents; i++)
FMOD::Studio::EventDescription* e = events[i];
// Get the path of this event (like event:/Explosion2D)

e ->getPath(eventName, maxPathLength, nullptr);
// Add to event map
mEvents.emplace(eventName, e);
Similarly, you create an AudioSystem::UnloadBank

function. This function first removes all the bank’s events
from the mEvents banks, unloads the sample data,
unloads the bank, and removes the bank from the
mBanks map.
For easy cleanup, you also create an

AudioSystem::UnloadAllBanks function. This
function just unloads all banks and clears out mEvents
and mBanks.
Every FMOD Studio project has two default bank files

named "Master Bank.bank" and "Master
Bank.strings.bank". The FMOD Studio runtime
does not have access to any other banks or events unless
it loads in the two master banks first. Because the master
banks always exist, you load them in
AudioSystem::Initialize with the following code:
// Load the master banks (strings first)
LoadBank("Assets/Master Bank.strings.bank");
LoadBank("Assets/Master Bank.bank");
Note how the code loads the master strings bank first.
The master strings bank is a special bank that contains
the human-readable names of all events and other data
in the FMOD Studio project. If you don’t load this bank,
the names are inaccessible in code. Without the names,
the code needs to use GUIDs (globally unique IDs) to
access all the FMOD Studio data. This means that,
technically, loading the master strings bank is optional,
but loading the strings makes the AudioSystem easier
to implement.
Creating and Playing Event Instances

Given an FMOD EventDescription, the
createInstance member function creates an
FMOD EventInstance for that event. Once you
have an EventInstance, the start function
begins playing it. So, a first pass of a PlayEvent
function in AudioSystem might look like this:
void AudioSystem::PlayEvent(const std::string& name)
{
// Make sure event exists
auto iter = mEvents.find(name);
if (iter != mEvents.end())
// Create instance of event
FMOD::Studio::EventInstance* event = nullptr;
iter ->second ->createInstance(&event);
if (event)
// Start the event instance
event ->start();
// Release schedules destruction of the event
// instance when it stops.
// (Non-looping events automatically stop.)
event ->release();
Although this version of PlayEvent is simple to use, it

does not expose much FMOD functionality. For example,
if the event is a looping event, there’s no way to stop the
event. There’s also no way to set any event parameters,
change the volume of the event, and so on.
It might be tempting to return the EventInstance

pointer directly from PlayEvent. Then, the caller can
access all the FMOD member functions. However, this is
not ideal because it exposes FMOD API calls outside the
audio system. This means that any programmer who
wants to simply play and stop sounds would need some
knowledge of the FMOD API.
Exposing the raw pointer also might be dangerous

because of the way FMOD cleans up memory for event
instances. After calling the release function, FMOD
destroys the event sometime after the event stops. If a
caller has access to the EventInstance pointer,
dereferencing it after destruction might cause a memory
access violation. Skipping the release call is also not a
great idea because then the system will leak memory
over time. Therefore, you need a more robust solution.
The SoundEvent Class

Rather than directly return an EventInstance
pointer from PlayEvent, you can track each active
event instance via an integer ID. Next, you can create
a new class called SoundEvent that allows
manipulation of the active events, using the integer
IDs to reference events. PlayEvent then returns an
instance of SoundEvent.
To track event instances, AudioSystem needs a new
map of unsigned integers to event instances:
std::unordered_map<unsigned int,
FMOD::Studio::EventInstance*> mEventInstances;
You also add a static sNextID variable that’s initialized

to 0. Each time PlayEvent creates an event instance, it
increments sNextID and adds the event instance to the
map with that new ID. Then PlayEvent returns a
SoundEvent with the associated ID, as in Listing 7.3.
(The declaration of SoundEvent is forthcoming.)
Listing 7.3 AudioSystem::PlayEvent

Implementation with Event IDs
SoundEvent AudioSystem::PlayEvent(const std::string& name)
unsigned int retID = 0;
auto iter = mEvents.find(name);
if (iter != mEvents.end())
// Create instance of event

FMOD::Studio::EventInstance* event = nullptr;
iter->second->createInstance(&event);
if (event)
// Start the event instance
event->start();
// Get the next id, and add to map
sNextID++;
retID = sNextID;
mEventInstances.emplace(retID, event);
return SoundEvent(this, retID);
Because sNextID is an unsigned int, IDs start

repeating after more than 4 billion calls to PlayEvent.
This should not be an issue, but it’s something to keep in
mind.
Note that PlayEvent no longer calls release on the

event instance. Instead, AudioSystem::Update now
handles cleaning up event instances that are no longer
needed. Every frame, Update checks the playback state
of each event instance in the map by using
getPlayBackState. It releases any event instances in
the stopped state and then removes them from the map.
This assumes that stopping an event means freeing it is
okay. A caller who wants to keep an event around can
pause it instead of stopping it. Listing 7.4 shows the
implementation of Update.
Listing 7.4 AudioSystem::Update Implementation

with Event IDs
void AudioSystem::Update(float deltaTime)
// Find any stopped event instances
std::vector<unsigned int> done;
for (auto& iter : mEventInstances)
FMOD::Studio::EventInstance* e = iter.second;
// Get the state of this event
FMOD_STUDIO_PLAYBACK_STATE state;
e->getPlaybackState(&state);
if (state == FMOD_STUDIO_PLAYBACK_STOPPED)
{
// Release the event and add id to done
e->release();
done.emplace_back(iter.first);
// Remove done event instances from map
for (auto id : done)
mEventInstances.erase(id);
// Update FMOD
mSystem->update();
Next, you add a GetEventInstance helper function to

AudioSystem that takes in an ID. If the ID exists in the
map, this function returns the corresponding
EventInstance pointer. Otherwise,
GetEventInstance returns nullptr. To prevent
every class from accessing event instances,
GetEventInstance is in the protected section of
AudioSystem. But because SoundEvent needs access
to this function, SoundEvent is declared as a friend of
AudioSystem.
Listing 7.5 gives the declaration of SoundEvent. Most

notably, its member data includes a pointer to the
AudioSystem and the ID. Note that while the default
constructor is public, the constructor with parameters is
protected. Because AudioSystem is a friend of
SoundEvent, only AudioSystem has access to this
constructor. This ensures that only AudioSystem can
assign IDs to SoundEvents. The rest of the functions in
SoundEvent are wrappers for various event instance
functionality, such as pausing sound events, changing
their volume, and setting event parameters.
Listing 7.5 SoundEvent Declaration

class SoundEvent
public:
SoundEvent();
// Returns true if associated FMOD event instance exists
bool IsValid();
// Restart event from beginning
void Restart();
// Stop this event
void Stop(bool allowFadeOut = true);
// Setters
void SetPaused(bool pause);
void SetVolume(float value);
void SetPitch(float value);
void SetParameter(const std::string& name, float value);
// Getters
bool GetPaused() const;
float GetVolume() const;
float GetPitch() const;
float GetParameter(const std::string& name);
protected:
// Make this constructor protected and AudioSystem a friend
// so that only AudioSystem can access this constructor.
friend class AudioSystem;
SoundEvent(class AudioSystem* system, unsigned int id);
private:
class AudioSystem* mSystem;
unsigned int mID;

};
The implementations for most of the SoundEvent

member functions have very similar syntax. They call
GetEventInstance to get an EventInstance pointer
and then call some function on the EventInstance.
For example, the implementation of
SoundEvent::SetPaused is as follows:
void SoundEvent::SetPaused(bool pause)
auto event = mSystem ?
mSystem->GetEventInstance(mID) : nullptr;
if (event)
event->setPaused(pause);
Note how the code validates that both the mSystem and
event pointer are non-null. This ensures that even if the
ID is not in the map, the function will not crash.
Similarly, the SoundEvent::IsValid function returns
true only if mSystem is non-null and the ID is in the
event instance map in AudioSystem.
With this code hooked up, it’s now possible to control

events after they start playing. For example, the
following starts playing an event called Music and saves
the SoundEvent in mMusicEvent:
mMusicEvent = mAudioSystem ->PlayEvent("event:/Music");
Elsewhere, you can toggle the pause state of the music

event with this:
mMusicEvent.SetPaused(!mMusicEvent.GetPaused());
With the addition of SoundEvent, you now have a

reasonable integration of FMOD for 2D audio.
3D POSITIONAL AUDIO
For 3D games, most sound effects are positional.
This means that an object in the world, such as a
fireplace, emits a sound. The game has a listener,
or a virtual microphone, that picks up this sound. For
example, if the listener faces the fireplace, it should
sound like the fireplace is in front. Similarly, if the
listener has his or her back to the fireplace, the
fireplace should sound like it’s behind.
Positional audio also means that as the listener gets

farther away from a sound, the volume of the sound
decreases, or attenuates. A falloff function describes
how the volume of the sound attenuates as the listener
gets farther away. In FMOD Studio, 3D sound events can
have user-configurable falloff functions.
The effect of positional audio is most apparent in a

surround sound configuration where there are more
than two speakers as output devices. For example, the
common 5.1 configuration (see Figure 7.1) features
front-left, front-center, front-right, back-left, and back-
right speakers as well as a subwoofer (or LFE) for low
frequency sounds. For the example of the in-game
fireplace, if the player faces the fireplace on screen, he or
she expects the sound to come out of the front speakers.
Figure 7.1 A 5.1 surround sound configuration

Luckily, FMOD has built-in support for positional audio.
To integrate this into a game, you must provide position
and orientation data for both the listener and any active
3D event instances. There are three parts to this: setting
up the listener, adding positional functionality to
SoundEvent, and creating an AudioComponent to
associate actors with sound events.
Setting Up a Basic Listener

A common approach is to use the camera as the
listener. In this case, the position of the listener is the
position of the camera in the world, and the
orientation of the listener is the orientation of the
camera. This approach works great for games with a
first-person camera, such as this chapter’s game
project. However, third-person cameras have
additional issues to consider, as discussed later in
this section.
A trap to watch out for when using any 3D positional

audio library (not just FMOD) is that the library may use
a different coordinate system than the game. For
example, FMOD uses a left-handed coordinate system
with +z forward, +x right, and +y up. However, our game
uses a left-handed coordinate system with +x forward,
+y right, +z up. So, when passing position and directions
from the game to FMOD, you must convert the
coordinates. This just involves switching around some
components when converting between a Vector3 and
FMOD’s vector type, FMOD_VECTOR. To help with this,
declare a VecToFMOD helper function:
FMOD_VECTOR VecToFMOD(const Vector3& in)
// Convert from our coordinates (+x forward, +y right, +z up)
// to FMOD (+z forward, +x right, +y up)
FMOD_VECTOR v;
v.x = in.y;
v.y = in.z;
v.z = in.x;
return v;
Next, you add a function called SetListener to

AudioSystem. This function, as shown in Listing 7.6,
takes in the view matrix and sets the listener’s position,
forward, and up vectors from the view. This means the
same code that sets the renderer’s view matrix can also
call SetListener. This process involves a little bit of
math. Recall that the view matrix transforms from world
space to view space. However, the listener requires a
world space position and orientation.
Extracting this information from the view matrix

requires several steps. First, you invert the view matrix.
Given this inverted view matrix, the first three
components of the fourth row (returned by
GetTranslation) correspond to the world space
position of the camera. The first three components of the
third row (returned by GetZAxis) correspond to the
forward vector, and the first three components of the
second row (returned by GetYAxis) correspond to the
up vector. You use VecToFMOD on all three of these
vectors to convert them to the FMOD coordinate system.
Listing 7.6 AudioSystem::SetListener

Implementation
void AudioSystem::SetListener(const Matrix4& viewMatrix)
// Invert the view matrix to get the correct vectors
Matrix4 invView = viewMatrix;
invView.Invert();
FMOD_3D_ATTRIBUTES listener;
// Set position, forward, up
listener.position = VecToFMOD(invView.GetTranslation());
// In the inverted view, third row is forward
listener.forward = VecToFMOD(invView.GetZAxis());
// In the inverted view, second row is up
listener.up = VecToFMOD(invView.GetYAxis());
// Set velocity to zero (fix if using Doppler effect)
listener.velocity = {0.0f, 0.0f, 0.0f};
// Send to FMOD (0 = only one listener)
mSystem->setListenerAttributes(0, &listener);
Note that SetListener currently sets the velocity

parameter of FMOD_3D_ATTRIBUTES to all zeros. The
velocity parameter matters only when enabling the
Doppler effect on sound events, as discussed later in this
section.
Adding Positional Functionality to

SoundEvent
Each EventInstance has 3D attributes that
describe its world position and orientation. It makes
sense to integrate this into the existing SoundEvent
class with two new functions, Is3D and
Set3DAttributes, both in Listing 7.7.
When you create a sound event in FMOD Studio, the

event can be 2D or 3D. The Is3D function returns true
if the event is 3D, and it returns false otherwise.
The Set3DAttributes function takes in a world

transform matrix and converts it into FMOD’s 3D
attributes. This makes it simple to pass in the world
transform matrix of an Actor to update the position and
orientation of the event. Note that this function does not
need to invert the matrix because the matrix is already in
world space. However, it’s still necessary to convert
between the game and FMOD coordinate systems.
Listing 7.7 SoundEvent’s Is3D and

Set3DAttributes Implementation
bool SoundEvent::Is3D() const
bool retVal = false;
auto event = mSystem ? mSystem->GetEventInstance(mID) : nullptr;
if (event)
// Get the event description
FMOD::Studio::EventDescription* ed = nullptr;
event->getDescription(&ed);
if (ed)
ed->is3D(&retVal); // Is this 3D?

}
return retVal;
void SoundEvent::Set3DAttributes(const Matrix4& worldTrans)
auto event = mSystem ? mSystem->GetEventInstance(mID) : nullptr;
if (event)
FMOD_3D_ATTRIBUTES attr;
// Set position, forward, up
attr.position = VecToFMOD(worldTrans.GetTranslation());
// In world transform, first row is forward
attr.forward = VecToFMOD(worldTrans.GetXAxis());
// Third row is up
attr.up = VecToFMOD(worldTrans.GetZAxis());
// Set velocity to zero (fix if using Doppler effect)
attr.velocity = { 0.0f, 0.0f, 0.0f };
event->set3DAttributes(&attr);
}
Creating an AudioComponent to Associate

Actors with Sound Events
The premise behind an AudioComponent class is to
associate sound events with specific actors. This way,
when the actor moves, AudioComponent can
update the associated event’s 3D attributes.
Furthermore, if an actor dies, any sound events
associated with the actor can stop.
Listing 7.8 gives the declaration of AudioComponent.

Note that it has two different std::vector collections:
one for 2D events and one for 3D events. The only
member functions not inherited from Component are
PlayEvent and StopAllEvents.
Listing 7.8 AudioComponent Declaration

class AudioComponent : public Component
AudioComponent(class Actor* owner, int updateOrder = 200);
~AudioComponent();

void OnUpdateWorldTransform() override;
SoundEvent PlayEvent(const std::string& name);
void StopAllEvents();
private:
std::vector<SoundEvent> mEvents2D;
std::vector<SoundEvent> mEvents3D;
};
The AudioComponent::PlayEvent function first calls

PlayEvent on the AudioSystem. You then check
whether the event is 3D to determine which of the two
vectors should store the SoundEvent. Finally, if the
event is 3D, call Set3DAttributes on it:
SoundEvent AudioComponent::PlayEvent(const std::string& name)
SoundEvent e = mOwner->GetGame()->GetAudioSystem()->PlayEvent(name);
// Is this 2D or 3D?
if (e.Is3D())
mEvents3D.emplace_back(e);
// Set initial 3D attributes
e.Set3DAttributes(mOwner->GetWorldTransform());
else
mEvents2D.emplace_back(e);
return e;
The AudioComponent::Update function (omitted

here) removes any of the events in mEvents2D or
mEvents3D that are no longer valid. (IsValid returns
false.)
Next, you add an override of

OnUpdateWorldTransform. Recall that every time the
owning actor computes its world transform matrix, it
notifies each component by calling this function. For the
AudioComponent, it needs to update the 3D attributes
of any 3D events in mEvents3D every time the world
transform changes:
void AudioComponent::OnUpdateWorldTransform()
{
Matrix4 world = mOwner->GetWorldTransform();
for (auto& event : mEvents3D)
if (event.IsValid())
event.Set3DAttributes(world);
Finally, AudioComponent::StopAllEvents (also

omitted here) simply calls stop on every event in both
vectors and clears out the vectors. The destructor of
AudioComponent calls this function, but there may be
other situations in which a game wants to just stop the
sound events for an actor.
With these additions, you can attach an

AudioComponent to an actor and play sound events on
the audio component. The AudioComponent then
automatically updates the 3D attributes of the associated
events as needed.
The Listener in a Third-Person Game

The listener directly using the camera position and
orientation works great for first-person games where
the camera is from the perspective of the player’s
character. However, things are not quite as simple
for a third-person game where the camera follows
the player character. Figure 7.2 illustrates the side
view of a third-person game. The player character is
at position P, and the camera is at position C.
Position A represents a sound effect right next to the
player character. Position B is a sound effect close to
the camera.
Figure 7.2 Sound effects in a third-person game
Now suppose the listener uses the camera position and

orientation, as in the preceding code. In this case, both
sounds A and B will sound like they’re in front. This is
good because both sound effects are visible onscreen, so
you as the player should perceive the sounds in front.
However, sound B will sound closer than sound A. This
seems weird because you expect sounds right next to the
player to be louder. And even if there were no sound B,
any sounds right next to (or even on) the player will
always have some attenuation applied to them, which
may be frustrating for the sound designer.
If instead the listener uses the player position and

orientation, then sound A will be louder than sound B.
However, sound B then sounds like it’s behind because
it’s positioned behind the player. This is very weird
because the sound is onscreen, so you expect it to sound
like it’s in front.
What you effectively want is an attenuation based on the

player position but an orientation based on the camera.
Guy Somberg describes a great solution to this problem
(his book is listed in the “Additional Reading” section at
the end of the chapter), which involves just a little bit of
vector math. Given the player at position P, the camera
at position C, and a sound at position S, first compute
two vectors—one vector from camera to sound and the
other vector from player to sound:
The length of the PlayerToSound vector is the desired

distance for attenuation. The normalized
CameraToSound vector is the correct heading. Scalar
multiplying the normalized CameraToSound vector by
the length of PlayerToSound yields a virtual position for
the sound:
This virtual position, illustrated in Figure 7.3, yields both
the correct attenuation and the correct orientation of the
sound. The listener itself then directly uses the camera,
as before.
Figure 7.3 Sound effects in a third-person game

with virtual positions
Note that this approach may be untenable if true world

positions of sounds are necessary for other calculations
(such as for occlusion, discussed later in this chapter).
The Doppler Effect

Imagine standing on a street corner. While the police
car is approaching, the pitch of the siren sound is
increased. Conversely, after the police car passes, the
pitch of the sound decreases. This is the Doppler
effect in action, and it is illustrated in Figure 7.4.
Figure 7.4 Doppler effect
The Doppler effect (or Doppler shift) occurs because

sound waves take time to travel through the air. As the
police car gets closer, each sound wave starts closer,
which means the waves arrive closer together. This
causes a perceived increase in frequency, leading to the
heightened pitch. The true pitch of the sound is audible
when the car is right next to the listener. Finally, as the
car travels off, the opposite effect occurs: The sound
waves arrive farther apart, yielding a lower pitch. The
Doppler effect applies to all types of waves, but sound
waves are the most easily observable.
In a game, the Doppler effect can create more realistic

sounds for objects such as vehicles. FMOD can
automatically calculate Doppler pitch shifts; it just
requires passing in the correct velocities both in
setListenerAttributes and set3DAttributes.
This means the game likely needs a more correct physics-
based movement approach with forces, as briefly
discussed in Chapter 3, “Vectors and Basic Physics.”
There also are some additional Doppler parameters

accessible through the Low Level API. The
set3DSettings function sets these parameters:
mLowLevelSystem->set3DSettings(
1.0f, // Doppler scale, 1 = normal, higher exaggerates effect
50.0f, // How many game units = 1 meter (our game is ~50)
1.0f // (Not for Doppler, leave at 1)
);
MIXING AND EFFECTS

One of the advantages of digitized sounds is that
manipulation during playback is easy. You’ve already
manipulated sounds as they play to account for the
sound’s position relative to the listener. The term
digital signal processing (DSP) refers to
computational manipulation of a signal. For audio,
adjusting the volume or pitch of the signal is a type of
DSP.
Two other common DSP effects in games are reverb and

equalization. Reverb simulates sound bouncing in an
enclosed area. For example, sound effects while inside a
cave have an echo because of waves bouncing off the
walls. Equalization, on the other hand, tries to
normalize the volume levels of sounds into a set range.
FMOD Studio allows configuration of chains of DSP

effects. In other words, a sound can pass through
multiple stages that modify the signal prior to output.
Although each sound event can have its own DSP chain,
a more common approach is to group sounds into types.
Then, different groups can have different effects applied
to them.
Buses
In FMOD Studio, a bus is a grouping of sounds. For
example, you might have a bus for sound effects, a
bus for music, and a bus for dialogue. Each bus can
individually have different DSP effects attached to it,
and at runtime you can adjust buses. For instance,
many games offer separate volume sliders for
different categories of sound. This is straightforward
to implement with buses.
By default, every project has a master bus, specified by

the root path bus:/. However, a sound designer can add
any number of additional buses. So, much as with
loading in event descriptions on bank load, you can load
in buses at the same time. First, you add a map of buses
to AudioSystem:
std::unordered_map<std::string, FMOD::Studio::Bus*> mBuses;

Then, when loading in a bank, call getBusCount and
getBusList on the bank to get the list of buses to add
to mBuses. (This is very similar to the code for event
descriptions, so this chapter omits that code.)
Next, add functions to AudioSystem to control the

buses:
float GetBusVolume(const std::string& name) const;
bool GetBusPaused(const std::string& name) const;
void SetBusVolume(const std::string& name, float volume);
void SetBusPaused(const std::string& name, bool pause);
The implementations of these functions are similar—and

aren’t surprising. For example, SetVolume is as follows:
void AudioSystem::SetBusVolume(const std::string& name, float volume)
auto iter = mBuses.find(name);
if (iter != mBuses.end())
iter->second->setVolume(volume);
}
In this chapter’s game project, there are three buses in

all: master, SFX, and music. The sound effects, including
footsteps, the fire loop, and the explosion sound, go
through the SFX bus, while the background music goes
through the music bus.
Snapshots
In FMOD, snapshots are special types of events
that control buses. Because they’re just events, they
use the same event interface that already exists, and
the existing PlayEvent function works with them.
The only difference is that their paths begin with
snapshot:/ instead of event:/.
Note that the game project in this chapter uses a

snapshot to enable reverb on the SFX bus. Use the R key
to enable or disable reverb.
Occlusion
Imagine living in a small apartment when there’s a
party next door. The music at the party is very loud
and travels through your wall. You’ve heard the song
before, but it sounds different when listening
through the wall. The bass is more dominant, and it’s
tough to hear the high-frequency parts. This is
sound occlusion, as illustrated in Figure 7.5(a).
Figure 7.5 Sound occlusion (a) and testing for

occlusion (b)
Sound occlusion occurs when a sound does not have a

direct path from emitter to listener. Instead, the sound
must travel through some material to reach the listener.
The predominant result of sound occlusion is a low-
pass filter, which means a reduction in volume of
higher-frequency sounds.
Implementing occlusion involves two separate tasks:

detection of occlusion and modification of occluded
sounds. One approach for detection is to draw line
segments between the emitter and an arc around the
listener, as in Figure 7.5(b). If all line segments can reach
the listener without hitting any objects, there is no
occlusion. If only some line segments reach, there’s
partial occlusion, and if none reach, there’s full
occlusion. This style of detection requires the collision
calculations covered in Chapter 10, “Collision Detection.”
Modifying occluded sounds in FMOD is simple.

However, it requires calls into the Low Level API. First,
when initializing FMOD, you enable software low-pass
filtering:
result = mSystem->initialize(
512, // Max number of concurrent sounds
FMOD_STUDIO_INIT_NORMAL, // Use default settings
FMOD_INIT_CHANNEL_LOWPASS, // Initialize low-pass filter
nullptr // Usually null
);
Next, each event instance affected by occlusion needs to

set occlusion parameters. For example, the following
code enables occlusion for event:
// Flush commands to ensure channel group is available
mSystem->flushCommands();
// Get channel group from event
FMOD::ChannelGroup* cg = nullptr;
event->getChannelGroup(&cg);
// Set occlusion factor - occFactor ranges
// from 0.0 (no occlusion) to 1.0 (full occlusion)
cg->set3DOcclusion(occFactor, occFactor);
GAME PROJECT
This chapter’s game project demonstrates most of
the audio features covered in this chapter. The code
mac.xcodeproj on Mac. The FMOD Studio project
corresponding to this chapter’s content is in
FMODStudio/Chapter07.
A music track plays in the background. As the player

walks around, a footstep event triggers. The sphere emits
a positional fire loop sound.
As before, use WASD to move around. The following keys

provide additional behavior:
E—Play the explosion (2D) sound
M—Pause/unpause the music event
R—Enable/disable reverb on the SFX bus (via a snapshot)
1—Set footstep parameter to default
2—Set footstep parameter to grass
−—Reduce master bus volume
+—Increase master bus volume
All the corresponding function calls for these behaviors

are in Game::HandleKeyPress.
The sound files used in this chapter come from
https://opengameart.org and http://freesound.org, both
great websites for finding quality sounds for games.
SUMMARY
Most games require audio systems that go beyond
simply playing sound files. Using the FMOD API, this
chapter shows how to implement a production-
quality sound system into the game. The audio
system loads in banks and plays back events. The
SoundEvent class tracks outstanding event
instances and allows manipulation of these
instances.
Positional audio simulates sounds in a 3D environment.

By setting the properties of the listener and every 3D
event instance, the audio behaves as it would in a real 3D
environment. While a first-person game can directly use
the camera orientation and position for the listener, a
third-person game is more complex. For fast-moving
objects, the Doppler effect shifts the pitch of the sound as
it approaches or leaves.
Mixing adds more control to the sound environment.

Buses group different sounds into independently
controllable categories. Snapshots can also dynamically
change the buses at runtime, such as enabling DSP
effects like reverb. Finally, occlusion simulates sounds
traveling through surfaces.
ADDITIONAL READING
Until recently, it was difficult to find references for
aspiring game audio programmers. However, Guy
Somberg’s excellent book has articles from many
experienced developers. This book provides the most
complete coverage of game audio currently available.
Somberg, Guy, Ed. Game Audio

Programming: Principles and Practices.
Boca Raton: CRC Press, 2016.
EXERCISES
This chapter’s exercises build on the audio features
implemented in the chapter. In the first exercise you
add support for the Doppler effect, while in the
second exercise you implement virtual positions for a
third-person listener.
Exercise 7.1
Adjust the listener and event instance attribute code
so that it correctly sets the velocity parameters. Then
make the sphere actor (created in
Game::LoadData) move quickly back and forth to
test the Doppler effect. Use set3DSettings to
adjust the intensity of the effect as needed. The
Doppler effect should be perceptible for the fire loop
audio sound once it’s working correctly.
Exercise 7.2
Implement virtual positions for event instances as
per the third-person listener formulas in this
chapter. Replace the CameraActor class in the
Chapter 7 game project with the CameraActor class
in Exercise/7.2 on GitHub. This version of the
CameraActor implements a basic third-person
camera for testing purposes.
CHAPTER 8
INPUT SYSTEMS
This chapter takes an in-depth look at a wide

variety of input devices for games, including
the keyboard, mouse, and controller. It
explores how to integrate these devices into a
cohesive system that all actors and
components in the game can interact with
for their input needs.
INPUT DEVICES
Without input, games would be a static form of
entertainment, much like film or television. The fact
that a game responds to the keyboard, mouse,
controller, or another input device is what enables
interactivity. You query these input devices for their
current state during the “process input” phase of the
game loop, and this affects the game world during
the “update game world” phase of the game loop.
Some input devices yield only Boolean values. For

example, for the keyboard you can check the state of each
key, and this state is true or false, depending on whether
the key is down or up. There’s no way for us to discern
whether a key is “half pressed” because the input device
simply doesn’t detect this.
Other input devices give a range of values. For example,

most joysticks yield a range of values in two axes that you
can use to determine how far the user has moved the
joystick in a specific direction.
Many of the devices used in games are composite,

meaning they combine multiple types of inputs into one.
For example, a typical controller might have two
joysticks and triggers that yield a range of values, as well
as other buttons that yield only Boolean values.
Similarly, the movement of the mouse or scroll wheel
might be some range, but the mouse buttons may be
Boolean.
Polling
Earlier in this book, you used the
SDL_GetKeyboardState function to get the
Boolean state of every key on the keyboard. With the
additions in Chapter 3, “Vectors and Basic Physics,”
you then passed this keyboard state to every actor’s
ProcessInput function, which in turn passes it to
every component’s ProcessInput function. Then,
in these functions you can query the state of a
specific key to decide whether to perform an action,
such as moving the player character forward when
pressing the W key. Because you’re checking the value
of a specific key on every frame, this approach is
considered polling the state of the key.
Input systems designed around polling are conceptually

simple to understand, and for this reason many game
developers prefer to use a polling approach. It works
especially well for things like character movement
because you need to know the state of some input device
on every frame and update the character movement
based on that. And, in fact, you will stick to this basic
polling approach for most of the input needs in the code
for this book.
Positive and Negative Edges

Consider a game where pressing the spacebar causes
a character to jump. On every frame, you check the
state of the spacebar. Suppose the spacebar is up for
the first three frames, and then the player presses the
spacebar prior to frame 4. The player continues to
hold the spacebar down until prior to frame 6 and
then releases it. You can draw this as a graph, as in
Figure 8.1, where the x-axis corresponds to the time
at each frame and the y-axis corresponds to the
binary value for that frame. On frame 4, the spacebar
changes from 0 to 1, and on frame 6, the spacebar
changes back from 1 to 0. The frame where the input
changes from 0 to 1 is a positive edge (or rising
edge), and the frame where the input changes from 1
to 0 is a negative edge (or falling edge).
Figure 8.1 Graph of the spacebar polled over nine

frames
Now consider what would happen if the process input for

the character simply said the following (in pseudocode):
if (spacebar == 1)
character.jump()
For the sample input in Figure 8.1, this code would call
the character.jump() function twice: once on frame
4 and once on frame 5. And if the player held the button
for 10 frames instead of 2, then you’d call
character.jump() 10 times. Clearly, you don’t want
the character to jump every frame when the spacebar
value is 1. Instead, you should only call
character.jump() on the frame where the spacebar
has a positive edge. For the input graph in Figure 8.1,
this is on frame 4. This way, for every press of the
spacebar, regardless of how long the player holds the
spacebar, the character jumps only once. In this case,
you want pseudocode like this:
if (spacebar has positive edge)
character.jump()
The “has positive edge” term in the pseudocode means

that on the last frame the key was 0, and on this frame
the key is 1. But with the current method of using
SDL_GeyKeyboardState to get the state of the
keyboard on the current frame, it might not be apparent
how to implement this. If you add a variable called
spacebarLast that you initialize to 0, you can use this
variable to track the value in the last frame. Then you
initiate the jump only if the value in the last frame is 0
and the value in this frame is 1:
if (spacebar == 1 and spacebarLast == 0)
character.jump()
spacebarLast = spacebar
Consider what happens in the case of the example in

Figure 8.1. On frame 3, you set spacebarLast to the
current value of spacebar, or 0. Then, on frame 4,
spacebar is 1 while spacebarLast is 0, so you trigger
character.jump(). After this, spacebarLast
becomes the current value of spacebar, or 1. Then on
frame 5, both spacebar and spacebarLast are 1, so the
character doesn’t jump.
You could use this pattern throughout the code.

However, it would be nice to have a system that tracks
values of keys on the previous frame automatically. That
way, you could easily ask the system whether a key has a
positive edge or negative edge, which might reduce the
burden for other programmers on the team.
If you generalize the approach of storing the value of the

input last frame and comparing it to the value this frame,
there are four possible results, as shown in Table 8.1. If
both values are 0, the button state is None. Similarly, if
both values are 1, this means the player holds the key for
consecutive frames, or the button state is Held. Finally,
if the values are different, it’s either a positive edge or
negative edge, which you denote with the button states
Pressed and Released, respectively.
Table 8.1 Four Possible Input States, Given the

Value in the Last Frame and in the Current Frame
Last Frame Current Frame Button State
0 0 None
0 1 Pressed
1 0 Released
1 1 Held
Consider how you might use this for a game where the
player can hold a key to charge up an attack. On the
frame on which you detect the Pressed state of the key,
you begin charging the attack. Then as long as the key’s
state on subsequent frames remains Held, you continue
to charge the attack. Finally, when the key’s state
becomes Released, it means the player let go of the key,
and you can now execute the attack with the appropriate
charge level.
But for actions such as just moving forward if W is 1,

you’d rather just use the old approach, where you check
the value of the input on that frame. In this chapter’s
input system, you will give the option of either querying
this basic value or querying for the different input states.
Events
Recall from Chapter 1, “Game Programming
Overview,” that SDL generates different events that
the program can optionally respond to. Currently,
you respond to the SDL_Quit event, which occurs
when the player tries to close the window.
Game::ProcessInput checks every frame if there
are events in the queue and can selectively choose to
respond to them.
SDL also generates events for input devices. For

example, every time the player presses a key on the
keyboard, SDL generates an SDL_KEYDOWN event
(corresponding to the Pressed button state).
Conversely, every time the player releases a key, it
generates an SDL_KEYUP event (corresponding to the
Released state). If you only care about positive and
negative edges, then this is a very quick way to set up
code to respond to these actions.
However, for the case of pressing W to move forward, this

means you need extra code to track whether W is held
because you only get the negative and positive edges
from the SDL events. Although you can certainly design
an input system entirely based around events, this
chapter uses SDL events only when required (such as for
mouse wheel scrolling).
There is one subtle relationship between SDL events and

the various polling functions. The keyboard state you get
from SDL_GetKeyboardState updates only after
calling SDL_PollEvents in the message pump loop.
This means you can delineate when the state data
changes between frames because you know where the
code calls SDL_PollEvents. This comes in handy when
implementing an input system that saves the data for the
previous frame.
Basic InputSystem Architecture

Before diving into each of the different input devices,
let’s consider a structure for an input system.
Currently you let actors and components know about
the current keyboard state via ProcessInput.
However, this mechanism means that
ProcessInput currently cannot access the mouse
or controller without directly calling SDL functions.
While this works for a simple game (and it’s the
approach largely used outside this chapter), it’s
better if the programmers writing the code for actors
and components do not need much specific
knowledge of SDL functions. Furthermore, some
SDL input functions return the difference in state
between calls of the function. If you call those
functions more than once during one frame, you’ll
get values of zero after the first call.
To solve this problem, you can have the InputSystem
class populate data in a helper class called InputState.
You can then pass this InputState by const reference
to actors/components via their ProcessInput function.
You can also add several helper functions to
InputState to make it easy to query whatever state the
actor/component cares about.
Listing 8.1 shows the initial declaration of the relevant

pieces. First, declare a ButtonState enum to
correspond to the four different states outlined in Table
8.1. Next, declare an InputState struct (which
currently has no members). Finally, you declare
InputSystem, which contains
Initialize/Shutdown functions (much like Game). It
also has a PrepareForUpdate function that is called
before SDL_PollEvents, and then an Update function
that is called after polling events. The GetState
function returns a const reference to the InputState
it holds as member data.
Listing 8.1 Basic InputSystem Declarations

enum ButtonState
ENone,
EPressed,
EReleased,
EHeld
};
// Wrapper that contains current state of input
struct InputState
KeyboardState Keyboard;
};
class InputSystem
public:
bool Initialize();
void Shutdown();
// Called right before SDL_PollEvents loop
void PrepareForUpdate();
// Called right after SDL_PollEvents loop
void Update();
const InputState& GetState() const { return mState; }
private:
InputState mState;
};
To integrate this code into the game, you add an

InputSystem pointer to the member data of Game
called mInputSystem. Game::Initialize allocates
and initializes InputSystem and Game::Shutdown
shuts down and deletes it.
Next, you change the declaration of ProcessInput in

both Actor and Component to the following:
void ProcessInput(const InputState& state);
Recall that in Actor, ProcessInput is not overridable

because it calls ProcessInput on all the attached
components. However, actors also have an overridable
ActorInput function for any input specific to that
actor. So, you similarly change the declaration of
ActorInput to take in a constant InputState
reference.
Finally, the implementation of Game::ProcessInput

changes to the following outline of steps:
mInputSystem->PrepareForUpdate();
// SDL_PollEvent loop...
mInputSystem->Update();
const InputState& state = mInputSystem->GetState();
// Process any keys here as desired...
// Send state to all actor's ProcessInput...
With the InputSystem in place, you now have the

basics needed to add support for several input devices.
For each of these devices, you need to add a new class to
encapsulate the state and add an instance of this class to
the InputState struct.
KEYBOARD INPUT
Recall that the SDL_GetKeyboardState function
returns a pointer to the keyboard state. Notably, the
return value of SDL_GetKeyboardState does not
change throughout the lifetime of the application, as
it points to internal SDL data. Therefore, to track the
current state of the keyboard, you merely need a
single pointer that you initialize once. However,
because SDL overwrites the current keyboard state
when you call SDL_PollEvents, you need a
separate array to save the previous frame state.
This leads naturally to the member data in the

declaration of KeyboardState, shown in Listing 8.2.
You have a pointer that points to the current state and an
array for the previous state. The size of the array
corresponds to the size of the buffer that SDL uses for
keyboard scan codes. For the member functions of
KeyboardState, you provide both a method to get the
basic current value of a key (GetKeyValue) and one
that returns one of the four button states
(GetKeyState). Finally, you make InputSystem a
friend of KeyboardState. This makes it easy for
InputSystem to directly manipulate the member data
of KeyboardState.
Listing 8.2 KeyboardState Declaration

class KeyboardState
{
public:
// Friend so InputSystem can easily update it
friend class InputSystem;
// Get just the boolean true/false value of key
bool GetKeyValue(SDL_Scancod keyCode) const;
// Get a state based on current and previous frame
ButtonState GetKeyState(SDL_Scancode keyCode) const;
private:
// Current state
const Uint8* mCurrState;
// State previous frame
Uint8 mPrevState[SDL_NUM_SCANCODES];
};
Next, you add a KeyboardState instance called

Keyboard to the member data of InputState:
struct InputState
{
KeyboardState Keyboard;
};
Next, you need to add code to both Initialize and

PrepareForUpdate within InputSystem. In
Initialize, you need to first set the mCurrState
pointer and then also zero out the memory of
mPrevState (because before the game starts, the keys
have no previous state). You get the current state pointer
from SDL_GetKeyboardState, and you can clear the
memory with memset:
// (In InputSystem::Initialize...)
// Assign current state pointer
mState.Keyboard.mCurrState = SDL_GetKeyboardState(NULL);
// Clear previous state memory
memset(mState.Keyboard.mPrevState, 0,
SDL_NUM_SCANCODES);
Then in PrepareForUpdate, you need to copy all the

“current” data to the previous buffer. Remember that
when you call PrepareForUpdate, the “current” data is
stale from the previous frame. This is because you call
PrepareForUpdate when you’re on a new frame but
haven’t called SDL_PollEvents yet. This is critical
because SDL_PollEvents is what updates the internal
SDL keyboard state data (which you’re pointing to with
mCurrState). So, before SDL overwrites the current
state, use memcpy to copy from the current buffer to the
previous buffer:
// (In InputSystem::PrepareForUpdate...)
memcpy(mState.Keyboard.mPrevState,
mState.Keyboard.mCurrState,
SDL_NUM_SCANCODES);
Next, you need to implement the member functions in

KeyboardState. GetKeyValue is straightforward. It
simply indexes into the mCurrState buffer and returns
true if the value is 1 and false if the value is 0.
The GetKeyState function, shown in Listing 8.3, is

slightly more complex. It uses both the current frame’s
and previous frame’s key state to determine which of the
four button states to return. This simply maps the entries
in Table 8.1 into source code.
Listing 8.3 KeyboardState::GetKeyState

Implementation
ButtonState KeyboardState::GetKeyState(SDL_Scancode keyCode) const
if (mPrevState[keyCode] == 0)
if (mCurrState[keyCode] == 0)
{ return ENone; }
else
{ return EPressed; }
else // Prev state must be 1
if (mCurrState[keyCode] == 0)
{ return EReleased; }
else
{ return EHeld; }
With this KeyboardState code, you can still access the

value of a key with the GetKeyValue function. For
example, the following checks if the current value of the
spacebar is true:
if (state.Keyboard.GetKeyValue(SDL_SCANCODE_SPACE))
However, the advantage of the InputState object is

that you can also query the button state of a key. For
example, the following code in Game::ProcessInput
detects if the button state of the Escape key is
EReleased, and it exits only at that point:
if (state.Keyboard.GetKeyState(SDL_SCANCODE_ESCAPE)
== EReleased)
mIsRunning = false;
This means that initially pressing Escape does not

immediately quit the game, but releasing the key causes
the game to quit.
MOUSE INPUT
For mouse input, there are three main types of input
to focus on: button input, movement of the mouse,
and movement of the scroll wheel. The button input
code is like the keyboard code except that the
number of buttons is significantly smaller. The
movement input is a little more complex because
there are two modes of input (absolute and relative).
Ultimately, you can still poll the mouse input with a
single function call per frame. However, for the scroll
wheel, SDL only reports the data via an event, so you
must add some code to InputSystem to also
process certain SDL events.
By default, SDL shows the system’s mouse cursor (at

least on platforms that have a system mouse cursor).
However, you can enable or disable the cursor by using
SDL_ShowCursor, passing in SDL_TRUE to enable it
and SDL_FALSE to disable it. For example, this disables
the cursor:
SDL_ShowCursor(SDL_FALSE);
Buttons and Position

For querying both the position of the mouse and the
state of its buttons, you use a single call to
SDL_GetMouseState. The return value of this
function is a bitmask of the button state, and you
pass in two integers by address to get the x/y
coordinates of the mouse, like this:
int x = 0, y = 0;
Uint32 buttons = SDL_GetMouseState(&x, &y);
note
For the position of the mouse, SDL uses the SDL 2D coordinate system. This
means that the top-left corner is (0, 0), positive x is to the right, and positive y
is down. However, you can easily convert these coordinates to whichever
other system you prefer.
For example, to convert to the simple view-projection coordinate system from
Chapter 5, “OpenGL,” you can use the following two lines of code:
x = x - screenWidth/2;
y = screenHeight/2 - y;
Because the return value of SDL_GetMouseState is a

bitmask, you need to use a bitwise-AND along with the
correct bit value to find out if a specific button is up or
down. For example, given the buttons variable
populated from SDL_GetMouseState, the following
statement is true if the left mouse button is down:
bool leftIsDown = (buttons & SDL_BUTTON(SDL_BUTTON_LEFT)) == 1;
The SDL_BUTTON macro shifts a bit based on the

requested button, and the bitwise-AND returns 1 if the
button is down and 0 if it’s up. Table 8.2 shows the
button constants corresponding to the five different
mouse buttons that SDL supports.
Table 8.2 SDL Mouse Button Constants

Button Constant
Left SDL_BUTTON_LEFT
Right SDL_BUTTON_RIGHT
Middle SDL_BUTTON_MIDDLE
Mouse button 4 SDL_BUTTON_X1
Mouse button 5 SDL_BUTTON_X2
You now have enough knowledge to create the initial

declaration of MouseState, which is shown in Listing
8.4. You save a 32-bit unsigned integer for both the
previous and current buttons’ bitmasks and a Vector2
for the current mouse position. Listing 8.4 omits the
implementations of the button functions because they
are almost identical to the functions for the keyboard
keys. The only difference is that these functions use the
bitmask as outlined earlier.
Listing 8.4 Initial MouseState Declaration

class MouseState
public:

// For mouse position
const Vector2& GetPosition() const { return mMousePos; }
// For buttons
bool GetButtonValue(int button) const;
ButtonState GetButtonState(int button) const;
private:
// Store mouse position
Vector2 mMousePos;
// Store button data
Uint32 mCurrButtons;
Uint32 mPrevButtons;
};
Next, you add a MouseState instance called Mouse to

InputState. Then, in InputSystem, add the
following to PrepareForUpdate, which copies the
current button state to the previous state:
mState.Mouse.mPrevButtons = mState.Mouse.mCurrButtons;
In Update, you call SDL_GetMouseState to update all
the MouseState members:
int x = 0, y = 0;
mState.Mouse.mCurrButtons = SDL_GetMouseState(&x, &y);
mState.Mouse.mMousePos.x = static_cast<float>(x);
mState.Mouse.mMousePos.y = static_cast<float>(y);
With these changes, you can now access basic mouse

information from InputState. For example, to
determine if the left mouse button is in state EPressed,
you use the following:
if (state.Mouse.GetButtonState(SDL_BUTTON_LEFT) == EPressed)
Relative Motion
SDL supports two different modes for detecting
mouse movement. In the default mode, SDL reports
the current coordinates of the mouse. However,
sometimes you instead want to know the relative
change of the mouse between frames. For example,
in many first-person games on PC, you can use the
mouse to rotate the camera. The speed of the
camera’s rotation depends on how fast the player
moves the mouse. In this case, exact coordinates of
the mouse aren’t useful, but the relative movement
between frames is.
You could approximate the relative movement between

frames by saving the position of the mouse on the
previous frame. However, SDL supports a relative
mouse mode that instead reports the relative movement
between calls to the SDL_GetRelativeMouseState
function. The big advantage of SDL’s relative mouse
mode is that it hides the mouse, locks the mouse to the
window, and centers the mouse on every frame. This
way, it’s not possible for the player to accidentally move
the mouse cursor out of the window.
To enable relative mouse mode, call the following:

SDL_SetRelativeMouseMode(SDL_TRUE);
Similarly, to disable relative mouse mode, pass in

SDL_FALSE as the parameter.
Once you’ve enabled relative mouse mode, instead of

using SDL_GetMouseState, you use
SDL_GetRelativeMouseState.
To support this in InputSystem, you first add a

function that can enable or disable relative mouse mode:
void InputSystem::SetRelativeMouseMode(bool value)
SDL_bool set = value ? SDL_TRUE : SDL_FALSE;
SDL_SetRelativeMouseMode(set);
mState.Mouse.mIsRelative = value;
You save the state of the relative mouse mode in a

Boolean variable in MouseState that you initialize to
false.
Next, change the code in InputSystem::Update so

that if you’re in relative mouse mode, you use the correct
function to grab the position and buttons of the mouse:
int x = 0, y = 0;
if (mState.Mouse.mIsRelative)
mState.Mouse.mCurrButtons = SDL_GetRelativeMouseState(&x, &y);
else
{
mState.Mouse.mCurrButtons = SDL_GetMouseState(&x, &y);
mState.Mouse.mMousePos.x = static_cast<float>(x);
mState.Mouse.mMousePos.y = static_cast<float>(y);
With this code, you can now enable relative mouse mode
and access the relative mouse position via MouseState.
Scroll Wheel
For the scroll wheel, SDL does not provide a function
to poll the current state of the wheel. Instead, SDL
generates the SDL_MOUSEWHEEL event. To support
this in the input system, then, you must first add
support for passing SDL events to InputSystem.
You can do this via a ProcessEvent function, and
then you update the event polling loop in
Game::ProcessInput to pass the mouse wheel
event to the input system:
SDL_Event event;
switch (event.type)
{
case SDL_MOUSEWHEEL:
mInputSystem->ProcessEvent(event);
break;
// Other cases omitted ...
Next, in MouseState add the following member

variable:
Vector2 mScrollWheel;
You use a Vector2 object because SDL reports scrolling

in both the vertical and horizontal directions, as many
mouse wheels support scrolling in both directions.
You then need to make changes to InputSystem. First,

implement ProcessEvent to read in the scroll wheel’s
x/y values from the event.wheel struct, as in Listing
8.5.
Listing 8.5 InputSystem::ProcessEvent

Implementation for the Scroll Wheel
void InputSystem::ProcessEvent(SDL_Event& event)

{
switch (event.type)
case SDL_MOUSEWHEEL:
mState.Mouse.mScrollWheel = Vector2(
static_cast<float>(event.wheel.x),
static_cast<float>(event.wheel.y));
break;
default:
break;
Next, because the mouse wheel event only triggers on

frames where the scroll wheel moves, you need to make
sure to reset the mScrollWheel variable during
PrepareForUpdate:
mState.Mouse.mScrollWheel = Vector2::Zero;
This ensures that if the scroll wheel moves on frame 1

but doesn’t move on frame 2, you don’t erroneously
report the same scroll value on frame 2.
With this code, you can now access the scroll wheel state
every frame with the following:
Vector2 scroll = state.Mouse.GetScrollWheel();
CONTROLLER INPUT
For numerous reasons, detecting controller input in
SDL is more complex than for the keyboard and
mouse. First, a controller has a much greater variety
of sensors than a keyboard or mouse. For example, a
standard Microsoft Xbox controller has two analog
joysticks, a directional pad, four standard face
buttons, three special face buttons, two bumper
buttons, and two triggers—which is a lot of different
sensors to get data from.
Furthermore, while PC/Mac users have only a single

keyboard or mouse, it’s possible to have multiple
controllers connected. Finally, controllers support hot
swapping, which means it’s possible to plug and unplug
controllers while a program is running. Combined, these
elements add complexity to handling controller input.
note
Depending on the controller and your platform, you may need to first install a
driver for your controller in order for SDL to detect it.
Before you can use a controller, you must first initialize
the SDL subsystem that handles controllers. To enable it,
simply add the SDL_INIT_GAMECONTROLLER flag to the
SDL_Init call in Game::Initialize:
SDL_Init(SDL_INIT_VIDEO | SDL_INIT_AUDIO | SDL_INIT_GAMECONTROLLER);
Enabling a Single Controller

For now, assume that you’re using only a single
controller and that this controller is plugged in when
the game starts. To initialize the controller, you need
to use the SDL_GameControllerOpen function.
This function returns a pointer to an
SDL_Controller struct upon successful
initialization or nullptr if it fails. You can then use
the SDL_Controller* variable to query the state of
the controller.
For this single controller, you first add an

SDL_Controller* pointer called mController to the
InputState member data. Then, add the following call
to open controller 0:
mController = SDL_GameControllerOpen(0);
To disable a controller, you can call

SDL_GameControllerClose, which takes the
SDL_GameController pointer as its parameter.
tip
By default, SDL supports a handful of common controllers, such as the
Microsoft Xbox controller. You can find controller mappings that specify the
button layouts of many other controllers. The
SDL_GameControllerAddMappingsFromFile function can load
controller mappings from a supplied file. A community-maintained mapping
file is available on GitHub at
https://github.com/gabomdq/SDL_GameControllerDB.
Because you do not want to assume that the player has a

controller, you must be vigilant to null check
mController wherever you might want to access it in
code.
Buttons
Game controllers in SDL support many different
buttons. SDL uses a naming convention that mirrors
the button names of a Microsoft Xbox controller. For
example, the names of the face buttons are A, B, X,
and Y. Table 8.3 lists the different button constants
defined by SDL, where * is a wildcard that denotes
multiple possible values.
Table 8.3 SDL Controller Button Constants
Button Constant
A, B, X, or Y SDL_CONTROLLER_BUTTON_* (replace * with A,

B, X, or Y)
Back SDL_CONTROLLER_BACK
Start SDL_CONTROLLER_START
Pressing SDL_CONTROLLER_BUTTON_*STICK (replace *

left/right with LEFT or RIGHT)
stick
Left/right SDL_CONTROLLER_BUTTON_*SHOULDER (replace

shoulder * with LEFT or RIGHT)
Directional SDL_CONTROLLER_BUTTON_DPAD_* (replace *

pad with UP, DOWN, LEFT, or RIGHT)
Note that the left and right stick buttons are for when the
user physically presses in the left/right stick. Some
games use pressing in the right stick for sprinting, for
example.
SDL does not have a mechanism to query the state of all

controller buttons simultaneously. Instead, you must
individually query each button via the
SDL_GameControllerGetButton function.
However, you can take advantage of the fact that the

enum for the controller button names defines an
SDL_CONTROLLER_BUTTON_MAX member that is the
number of buttons the controller has. Thus, the first pass
of the ControllerState class, shown in Listing 8.6,
contains arrays for both the current and previous button
states. The code also has a Boolean so the game code can
determine whether there’s a controller connected.
Finally, the class has declarations for the now-standard
button value/state functions.
Listing 8.6 Initial ControllerState Declaration

class ControllerState
public:
// For buttons
bool GetButtonValue(SDL_GameControllerButton button) const;
ButtonState GetButtonState(SDL_GameControllerButton button)
const;
bool GetIsConnected() const { return mIsConnected; }
private:
// Current/previous buttons
Uint8 mCurrButtons[SDL_CONTROLLER_BUTTON_MAX];
Uint8 mPrevButtons[SDL_CONTROLLER_BUTTON_MAX];
// Is this controlled connected?
bool mIsConnected;
};
Then add an instance of ControllerState to

InputState:
ControllerState Controller;
Next, back in InputSystem::Initialize, after you

try to open controller 0, set the mIsConnected variable
based on whether the mController pointer is non-null.
You also clear out the memory for both mCurrButtons
and mPrevButtons:
mState.Controller.mIsConnected = (mController != nullptr);
memset(mState.Controller.mCurrButtons, 0,
SDL_CONTROLLER_BUTTON_MAX);
memset(mState.Controller.mPrevButtons, 0,
As with the keyboard, the code in PrepareForUpdate

then copies the button states from current to previous:
memcpy(mState.Controller.mPrevButtons,
mState.Controller.mCurrButtons,
Finally, in Update, loop over the mCurrButtons array

and set the value of each element to the result of the
SDL_GameControllerGetButton call that queries the
state of that button:
for (int i = 0; i < SDL_CONTROLLER_BUTTON_MAX; i++)
mState.Controller.mCurrButtons[i] =
SDL_GameControllerGetButton(mController,
SDL_GameControllerButton(i));
With this code, you can then query the state of a specific
game controller button, using a pattern like the keyboard
and mouse buttons. For example, this code checks if the
A button on the controller has a positive edge this frame:
if (state.Controller.GetButtonState(SDL_CONTROLLER_BUTTON_A) == EPressed)
Analog Sticks and Triggers

SDL supports a total of six axes. Each analog stick
has two axes: one in the x direction and one in the y
direction. Furthermore, each of the triggers has a
single axis. Table 8.4 shows the list of axes. (Once
again, * denotes a wildcard.)
Table 8.4 SDL Controller Axis Constants
Button Constant
Left analog SDL_CONTROLLER_AXIS_LEFT* (replace * with X

stick or Y)
Right analog SDL_CONTROLLER_AXIS_RIGHT* (replace * with

stick X or Y)
Left/right SDL_CONTROLLER_AXIS_TRIGGER* (replace *

triggers with LEFT or RIGHT)
For triggers, the value ranges from 0 to 32,767, with 0

meaning there is no pressure on the trigger. For the
analog stick axes, the value ranges from -32,768 to
32,767, with 0 representing centered. A positive y-axis
value corresponds to down on the analog stick, and a
positive x-axis value corresponds to right.
However, an issue with continuous input such as these

axes is that the ranges specified by the API are
theoretical. Each individual device has its own
imprecisions. You can observe this behavior by releasing
one of the analog sticks, which returns the stick to its
center. You might reasonably expect that because the
stick is at rest, the values reported for the stick’s x- and
y-axes are zero. However, in practice the values will be
around zero but rarely precisely zero. Conversely, if the
player slams the stick all the way to the right, the value
reported by the stick’s x-axis will be near the maximum
value but rarely precisely the maximum value.
This is problematic for games for two reasons. First, it

may cause phantom inputs, where the player isn’t
touching an input axis but the game reports that
something is happening. For example, suppose the
player completely puts the controller down on a table.
The player should rightfully expect that his or her
character in game will not move around. However, if the
issue isn’t handled, the game will detect some value of
input to the axis and move the character.
Furthermore, many games have the character move

based on how far the analog stick is moved in one
direction—so that slightly moving the stick might cause
the character to slowly walk, whereas moving the stick all
the way in a direction might cause the character to
sprint. However, if you only make the player sprint when
the axis reports the maximum value, the player will
never sprint.
To solve this issue, code that processes the input from an

axis should filter the value. Specifically, you want to
interpret values close to zero as zero and values close to
the minimum or maximum as the minimum or
maximum. Furthermore, it’s convenient for users of the
input system if you convert the integral ranges into a
normalized floating-point range. For the axes that yield
both positive and negative values, this means a range
between −1.0 and 1.0.
Figure 8.2 shows an example of such a filter for a single

axis. The numbers above the line are the integral values
before filtering, and the numbers below the line are the
floating-point values after filtering. The area near zero
that you interpreted as 0.0 is called a dead zone.
Figure 8.2 A sample filter for an axis, with the input

values above and the output values below
Listing 8.7 shows the implementation the

InputSystem::Filter1D function, which the input
system uses to filter one-dimensional axes such as the
triggers. First, you declare two constants for the dead
zone and maximum value. Note that deadZone here is
250—which is less than in Figure 8.2 because this value
works better for the triggers (but you could make the
constants parameters or user configurable, if desired).
Next, the code takes the absolute value of the input by

using a ternary operator. If this value is less than the
dead zone constant, you simply return 0.0f. Otherwise,
you convert the input to a fractional value representing
where it lands in between the dead zone and the
maximum value. For example, an input halfway between
deadZone and maxValue is 0.5f.
Then you ensure that the sign of this fractional value

matches the sign of the original input. Finally, you clamp
the value to the range of -1.0 to 1.0 to account for the
cases where the input is greater than the maximum value
constant. The implementation Math::Clamp is in the
custom Math.h header file.
Listing 8.7 Filter1D Implementation

float InputSystem::Filter1D(int input)
// A value < dead zone is interpreted as 0%
const int deadZone = 250;
// A value > max value is interpreted as 100%
const int maxValue = 30000;
float retVal = 0.0f;
// Take absolute value of input
int absValue = input > 0 ? input : -input;
// Ignore input within dead zone

if (absValue > deadZone)
// Compute fractional value between dead zone and max value
retVal = static_cast<float>(absValue - deadZone) /
(maxValue - deadZone);
// Make sure sign matches original value
retVal = input > 0 ? retVal : -1.0f * retVal;
// Clamp between -1.0f and 1.0f
retVal = Math::Clamp(retVal, -1.0f, 1.0f);
return retVal;
Using the Filter1D function, an input value of 5000

returns 0.0f, and a value of -19000 returns -0.5f.
The Filter1D function works well when you only need

a single axis, such as for one of the triggers. However,
because the analog sticks really are two different axes in
concert, it’s usually preferable to instead filter them in
two dimensions, as discussed in the next section.
For now, you can add two floats to ControllerState

for the left and right triggers:
float mLeftTrigger;
float mRightTrigger;
Next, in InputSystem::Update use the

SDL_GameControllerGetAxis function to read in the
values of both triggers and call the Filter1D function
on this value to convert it to a range of 0.0 to 1.0
(because triggers cannot be negative). For example, the
following sets the mLeftTrigger member:
mState.Controller.mLeftTrigger =
Filter1D(SDL_GameControllerGetAxis(mController,
SDL_CONTROLLER_AXIS_TRIGGERLEFT));
You then add GetLeftTrigger() and

GetRightTrigger() functions to access these. For
example, the following code gets the value of the left
trigger:
float left = state.Controller.GetLeftTrigger();

Filtering Analog Sticks in Two Dimensions
A common control scheme for an analog stick is that
the orientation of the stick corresponds to the
direction in which the player’s character moves. For
example, pressing the stick up and to the left would
cause the character onscreen to also move in that
direction. To implement this, you should interpret
the x- and y-axes together.
Although it is tempting to apply the Filter1D function

to the x- and y-axes independently, doing so can cause an
interesting issue. If the player moves the stick all the way
up, interpreting it as a normalized vector yields <0.0,
1.0>. On the other hand, if the player moves the stick all
the way up and to the right, the normalized vector is
<1.0, 1.0>. The length of these two vectors is different,
which is a problem if you use the length to dictate the
speed at which the character moves: The character could
move faster diagonally than straight in one direction!
Although you could just normalize vectors with a length

greater than one, interpreting each axis independently
still ultimately means you’re interpreting the dead zone
and maximum values as a square. A better approach is to
interpret them as concentric circles, as shown in Figure
8.3. The square border represents the raw input values,
the inner circle represents the dead zone, and the outer
circle represents the maximum values.
Figure 8.3 Filtering in two dimensions
Listing 8.8 gives the code for Filter2D, which takes in

both the x- and y-axes for the analog stick and filters in
two dimensions. You first create a 2D vector and then
determine the length of that vector. Lengths less than the
dead zone result in Vector2::Zero. For lengths
greater than the dead zone, you determine the fractional
value between the dead zone and max and set the length
of the vector to this fractional value.
Listing 8.8 InputSystem::Filter2D

Implementation
Vector2 InputSystem::Filter2D(int inputX, int inputY)

{
const float deadZone = 8000.0f;
const float maxValue = 30000.0f;
// Make into 2D vector
Vector2 dir;
dir.x = static_cast<float>(inputX);
dir.y = static_cast<float>(inputY);
float length = dir.Length();
// If length < deadZone, should be no input
if (length < deadZone)
dir = Vector2::Zero;
else
{
// Calculate fractional value between
// dead zone and max value circles
float f = (length - deadZone) / (maxValue - deadZone);

// Clamp f between 0.0f and 1.0f
f = Math::Clamp(f, 0.0f, 1.0f);
// Normalize the vector, and then scale it to the
// fractional value
dir *= f / length;
return dir;
Next, add two Vector2s to ControllerState for the

left and right sticks, respectively. You can then add code
in InputSystem::Update to grab the values of the two
axes for each stick and then run Filter2D to get the
final analog stick value. For example, the following code
filters the left stick and saves the result in the controller
state:
x = SDL_GameControllerGetAxis(mController,
SDL_CONTROLLER_AXIS_LEFTX);
y = -SDL_GameControllerGetAxis(mController,
SDL_CONTROLLER_AXIS_LEFTY);
mState.Controller.mLeftStick = Filter2D(x, y);
Note that this code negates the y-axis value. This is

because SDL reports the y-axis in the SDL coordinate
system where +y is down. Thus, to get the expected
values in the game’s coordinate system, you must negate
the value.
You can then access the value of the left stick via
InputState with code like this:
Vector2 leftStick = state.Controller.GetLeftStick();
Supporting Multiple Controllers

Supporting multiple local controllers is more
complex than supporting one. This section briefly
touches on the different pieces of code needed to
support it, though it does not fully implement this
code. First, to initialize all connected controllers at
startup, you need to rewrite the controller detection
code to loop over all joysticks and see which ones are
controllers. You can then open each one individually,
with code roughly like this:
for (int i = 0; i < SDL_NumJoysticks(); ++i)
{
// Is this joystick a controller?
if (SDL_IsGameController(i))
{
// Open this controller for use
SDL_GameController* controller = SDL_GameControllerOpen(i);
// Add to vector of SDL_GameController* pointers
Next, you change InputState to contain several

ControllerStates instead of just one. You also
update all the functions in InputSystem to support
each of these different controllers.
To support hot swapping (adding/removing controllers

while the game is running), SDL generates two different
events for adding and removing controllers:
SDL_CONTROLLERDEVICEADDED and
SDL_CONTROLLERDEVICEREMOVED. Consult the SDL
documentation for further information about these
events (see
https://wiki.libsdl.org/SDL_ControllerDeviceEvent).
INPUT MAPPINGS
The way you currently use the data from
InputState, the code assumes that specific input
devices and keys map directly to actions. For
example, if you want the player character to jump on
the positive edge of a spacebar, you add code like this
to ProcessInput:
bool shouldJump = state.Keyboard.GetKeyState(SDL_SCANCODE_SPACE)
== Pressed;
Although this works, ideally you’d like to instead define

an abstract “Jump” action. Then, you want some
mechanism that allows the game code to specify that
“Jump” corresponds to the spacebar key. To support this,
you want a map between these abstract actions and the
{device, button} pair corresponding to this abstract
action. (You will actually work on implementing this in
Exercise 8.2.)
You could further enhance this system by allowing for

multiple bindings to the same abstract action. This
means you could bind both the spacebar and the A
button on the controller to “Jump.”
Another advantage of defining such abstract actions is

that doing so makes it easier for AI-controlled characters
to perform the same action. Rather than needing some
separate code path for the AI, you could update the AI
character such that it generates a “Jump” action when
the AI wants to jump.
Another improvement to this system allows for the

definition of a movement along an axis, such as a
“ForwardAxis” action that corresponds to the W and S
keys or one of the controller axes. You can then use this
action to specify movement of characters in the game.
Finally, with these types of mappings, you can add a

mechanism to load mappings from a file. This makes it
easy for designers or users to configure the mappings
without modifying the code.
GAME PROJECT
This chapter’s game project adds a full
implementation of the InputSystem from this
chapter to the game project from Chapter 5. This
includes all the code for the keyboard, mouse, and
controller. Recall that the Chapter 5 project uses 2D
movement (so position is a Vector2). The code is
In this chapter’s project, the game controller moves the

spaceship. The left stick affects the direction in which the
ship travels, and the right stick rotates the direction the
ship faces. The right trigger fires a laser. This is a control
scheme popularized by “twin stick shooter” games.
With the input system already returning 2D axes for the

left/right stick, implementing the twin stick–style
controls does not require too much code. First, in
Ship::ActorInput, you add the following lines of
code to grab both the left and right sticks and save them
in member variables:
if (state.Controller.GetIsConnected())
mVelocityDir = state.Controller.GetLeftStick();
if (!Math::NearZero(state.Controller.GetRightStick().Length()))
mRotationDir = state.Controller.GetRightStick();
You add the NearZero check for the right stick to make
sure that if the player releases the right stick completely,
the ship doesn’t snap back to an initial angle of zero.
Next, in Ship::UpdateActor, add the following code

to move the actor based on the direction of the velocity, a
speed, and delta time:
pos += mVelocityDir * mSpeed * deltaTime;

SetPosition(pos);
Note that this code reduces the speed based on how far
you move the left stick in a direction because
mVelocityDir can have a length less than one in this
case.
Finally, you add the following code (also in

UpdateActor) to rotate the actor based on the
mRotationDir, using the atan2 approach:
float angle = Math::Atan2(mRotationDir.y, mRotationDir.x);
SetRotation(angle);
Again, this code compiles because the Actor class in this

chapter’s project harkens back to the 2D actor class that
used a single float for the angle, as opposed to the
quaternion rotation used in 3D.
Figure 8.4 shows what the game looks like with the ship
moving around.
Figure 8.4 Ship moving around in the Chapter 8
game project
SUMMARY
Many different input devices are used for games. A
device might report either a single Boolean value or a
range of inputs. For a key/button that reports a
simple on/off state, it’s useful to consider the
difference between the value in this frame and the
value in the last frame. This way, you can detect the
positive or negative edge of the input, corresponding
to a “pressed” or “released” state.
SDL provides support for the most common input

devices including the keyboard, mouse, and controller.
For each of these devices, you add data in an
InputState struct that you then pass to each actor’s
ProcessInput function. This way, actors can query the
input state for not only the current values of inputs but
also negative and positive edges.
For devices that give a range of values, such as the

triggers or analog sticks, you typically need to filter this
data. This is because even when the device is at rest, the
device may give spurious signals. The filtering
implemented in this chapter ensures that input less than
some dead zone is ignored and also ensures that you
detect the maximum input even when the input is only
“almost” the maximum.
This chapter’s game project takes advantage of the new

controller input functionality to add support for twin-
stick shooter–style movement.
ADDITIONAL READING
Bruce Dawson covers how to record input and then
play it back, which is very useful for testing. The
Oculus SDK documentation covers how to interface
with Oculus VR touch controllers. Finally, Mick West
explores how to measure input lag, which is the
amount of time it takes a game to detect inputs from
controllers. Input lag is generally not the fault of the
input code, but West’s material is interesting
nonetheless.
Dawson, Bruce. “Game Input Recording

and Playback.” Game Programming Gems
2, edited by Mark DeLoura. Cengage
Learning, 2001.
Oculus PC SDK. Accessed November 29,
2017.
https://developer.oculus.com/documentation/pcsdk/latest/.
West, Mick. “Programming
Responsiveness.” Gamasutra. Accessed
November 29, 2017.
http://www.gamasutra.com/view/feature/1942/programming_responsiven
print=1.
EXERCISES
In this chapter’s exercises you will improve the input
system. In the first exercise you add support for
multiple controllers. In the second exercise you add
input mappings.
Exercise 8.1
Recall that to support multiple controllers, you need
to have multiple ControllerState instances in
the InputState struct. Add code to support a
maximum of four controllers simultaneously. On
initialization, change the code to detect any
connected controllers and enable them individually.
Then change the Update code so that it updates up
to all four controllers instead of just a single one.
Finally, investigate the events that SDL sends when the

user connects/disconnects controllers and add support
to dynamically add and remove controllers.
Exercise 8.2
Add support for basic input mappings for actions. To
do this, create a text file format that maps actions to
both a device and a button/key on that device. For
example, an entry in this text file to specify that the
“Fire” action corresponds to the A button on the
controller might look like this:
Fire,Controller,A
Then parse this data in the InputSystem and save it

into a map. Next, add a generic
GetMappedButtonState function to InputState
that takes in the action name and returns the
ButtonState from the correct device. The signature of
this function is roughly the following:
ButtonState GetMappedButtonState(const std::string& actionName);

CHAPTER 9
CAMERAS
The camera determines the player’s point of

view in a 3D game world, and there are many
different types of cameras. This chapter
covers the implementation of four cameras:
a first-person camera, a follow camera, an
orbit camera, and a spline camera that
follows paths. And because the camera often
dictates the movement of the player
character, this chapter also covers how to
update movement code for different types of
cameras.
FIRST-PERSON CAMERA
A first-person camera shows the game world
from the perspective of a character moving through
the world. This type of camera is popular in first-
person shooters such as Overwatch but also sees use
in some role-playing games like Skyrim or narrative-
based games such as Gone Home. Some designers
feel that a first-person camera is the most immersive
type of camera for a video game.
Even though it’s tempting to think of a camera as just a

view, the camera also informs the player how the player
character moves around the world. This means the
camera and movement system implementations depend
on each other. The typical controls for a first-person
shooter on PC use both the keyboard and mouse. The
W/S keys move forward and backward, while the A/D
keys strafe the character (that is, move left and right).
Moving the mouse left and right rotates the character
about the up axis, but moving the mouse up and down
pitches only the view, not the character.
Basic First-Person Movement

Implementing movement is easier than working with
the view, so this is a good starting point. You create a
new actor called FPSActor that implements first-
person movement. The forward/back movement in
MoveComponent already works in the 3D world,
based on the changes made in Chapter 6, “3D
Graphics.” Implementing strafing requires just a few
updates. First, you create a GetRight function in
Actor, which is like GetForward (just using the y-
axis instead):
Vector3 Actor::GetRight() const
// Rotate right axis using quaternion rotation
return Vector3::Transform(Vector3::UnitY, mRotation);
Next, you add a new variable in MoveComponent called

mStrafeSpeed that affects the speed at which the
character strafes. In Update, you simply use the right
vector of the actor to adjust the position based on the
strafe speed:
if (!Math::NearZero(mForwardSpeed) || !Math::NearZero(mStrafeSpeed))
Vector3 pos = mOwner->GetPosition();
pos += mOwner->GetForward() * mForwardSpeed * deltaTime;

// Update position based on strafe
pos += mOwner->GetRight() * mStrafeSpeed * deltaTime;
mOwner->SetPosition(pos);
Then in FPSActor::ActorInput, you can detect the

A/D keys and adjust the strafe speed as needed. Now the
character can move with standard first-person WASD
controls.
The left/right rotation also already exists in

MoveComponent via the angular speed. So, the next task
is to convert mouse left/right movements to angular
speed. First, the game needs to enable relative mouse
mode via SDL_RelativeMouseMode. Recall from
Chapter 8, “Input Systems,” that relative mouse mode
reports the change in (x, y) values per frame, as opposed
to absolute (x, y) coordinates. (Note that in this chapter,
you will directly use SDL input functions rather than the
input system created in Chapter 8.)
Converting the relative x movement into an angular

speed only requires a few calculations, shown in Listing
9.1. First, SDL_GetRelativeMouseState retrieves the
(x, y) motion. The maxMouseSpeed constant is an
expected maximum amount of relative motion possible
per frame, though this might be an in-game setting.
Similarly, maxAngularSpeed converts the motion into
a rotation per second. You then take the reported x value,
divide by maxMouseSpeed, and multiply by
maxAngularSpeed. This yields an angular speed that’s
sent to the MoveComponent.
Listing 9.1 FPS Angular Speed Calculation from the

Mouse
// Get relative movement from SDL
int x, y;
Uint32 buttons = SDL_GetRelativeMouseState(&x, &y);
// Assume mouse movement is usually between -500 and +500
const int maxMouseSpeed = 500;
// Rotation/sec at maximum speed
const float maxAngularSpeed = Math::Pi * 8;
float angularSpeed = 0.0f;
if (x != 0)
// Convert to approximately [-1.0, 1.0]
angularSpeed = static_cast<float>(x) / maxMouseSpeed;
// Multiply by rotation/sec
angularSpeed *= maxAngularSpeed;
}
mMoveComp->SetAngularSpeed(angularSpeed);
Camera (Without Pitch)

The first step to implement a camera is to create a
subclass of Component called CameraComponent.
All the different types of cameras in this chapter will
subclass from CameraComponent, so any common
camera functionality can go in this new component.
The declaration of CameraComponent is like that of
any other component subclass. For now, the only
new function is a protected function called
SetViewMatrix, which simply forwards the view
matrix to the renderer and audio system:
void CameraComponent::SetViewMatrix(const Matrix4& view)
// Pass view matrix to renderer and audio system
Game* game = mOwner->GetGame();
game->GetRenderer()->SetViewMatrix(view);
game->GetAudioSystem()->SetListener(view);
For the FPS camera specifically, you create a subclass of

CameraComponent called FPSCamera, which has an
overridden Update function. Listing 9.2 shows the code
for Update. For now, Update uses the same logic as the
basic camera actor introduced in Chapter 6. The camera
position is the owning actor’s position, the target point is
an arbitrary point in the forward direction of the owning
actor, and the up vector is the z-axis. Finally,
Matrix4::CreateLookAt creates the view matrix.
Listing 9.2 FPSCamera::Update Implementation

(Without Pitch)
void FPSCamera::Update(float deltaTime)
// Camera position is owner position
Vector3 cameraPos = mOwner->GetPosition();
// Target position 100 units in front of owner
Vector3 target = cameraPos + mOwner->GetForward() * 100.0f;
// Up is just unit z
Vector3 up = Vector3::UnitZ;
// Create look at matrix, set as view
Matrix4 view = Matrix4::CreateLookAt(cameraPos, target, up);
SetViewMatrix(view);
}
Adding Pitch
Recall from Chapter 6 that yaw is rotation about the
up axis and pitch is rotation about the side axis (in
this case, the right axis). Incorporating pitch into the
FPS camera requires a few changes. The camera still
starts with the forward vector from the owner, but
you apply an additional rotation to account for the
pitch. Then, you derive a target from this view
forward. To implement this, you add three new
member variables to FPSCamera:
// Rotation/sec speed of pitch
float mPitchSpeed;
// Maximum pitch deviation from forward
float mMaxPitch;
// Current pitch
float mPitch;
The mPitch variable represents the current (absolute)

pitch of the camera, while mPitchSpeed is the current
rotation/second in the pitch direction. Finally, the
mMaxPitch variable is the maximum the pitch can
deviate from the forward vector in either direction. Most
first-person games limit the total amount the player can
pitch the view up or down. The reason for this limitation
is that the controls seem odd if the player faces straight
up. In this case, you can use 60° (converted to radians)
as the default maximum pitch value.
Next, you modify FPSCamera::Update to take into

account the pitch, as in Listing 9.3. First, the current
pitch value updates based on the pitch speed and delta
time. Second, you clamp the pitch to make sure it does
not exceed +/- the maximum pitch. Recall from Chapter
6 that a quaternion can represent an arbitrary rotation.
Thus, you can construct a quaternion representing this
pitch. Note that this rotation is about the owner’s right
axis. (It’s not just the y-axis because the pitch axis
changes depending on the owner’s yaw.)
The view forward is then the owner’s forward vector,

transformed by the pitch quaternion. You use this view
forward to determine the target position that’s “in front”
of the camera. You also rotate the up vector by the pitch
quaternion. Then you construct the look-at matrix from
these vectors. The camera position is still the owner’s
position.
Listing 9.3 FPSCamera::Update Implementation

(with Pitch Added)
void FPSCamera::Update(float deltaTime)
{
// Call parent update (doesn't do anything right now)
CameraComponent::Update(deltaTime);
// Camera position is owner position
// Update pitch based on pitch speed
mPitch += mPitchSpeed * deltaTime;
// Clamp pitch to [-max, +max]
mPitch = Math::Clamp(mPitch, -mMaxPitch, mMaxPitch);
// Make a quaternion representing pitch rotation,
// which is about owner's right vector
Quaternion q(mOwner->GetRight(), mPitch);
// Rotate owner forward by pitch quaternion
Vector3 viewForward = Vector3::Transform(
mOwner->GetForward(), q);
// Target position 100 units in front of view forward
Vector3 target = cameraPos + viewForward * 100.0f;
// Also rotate up by pitch quaternion
Vector3 up = Vector3::Transform(Vector3::UnitZ, q);

// Create look at matrix, set as view
Finally, FPSActor updates the pitch speed based on the

relative y motion of the mouse. This requires code in
ProcessInput that is almost identical to the code you use
to update the angular speed based on the x motion from
Listing 9.1. With this in place, the first-person camera
now pitches without adjusting the pitch of the owning
actor.
First-Person Model
Although it’s not strictly part of the camera, most
first-person games also incorporate a first-person
model. This model may have parts of an animated
character, such as arms, feet, and so on. If the player
carries a weapon, then when the player pitches up,
the weapon appears to also aim up. You want the
weapon model to pitch up even though the player
character remains flat with the ground.
You can implement this with a separate actor for the

first-person model. Then every frame, FPSActor
updates the first-person model position and rotation.
The position of the first-person model is the position of
the FPSActor with an offset. This offset places the first-
person model a little to the right of the actor. The
rotation of the model starts with the rotation of the
FPSActor but then has an additional rotation applied
for the view pitch. Listing 9.4 shows the code for this.
Listing 9.4 Updating Position and Rotation of the First-

Person Model
// Update position of FPS model relative to actor position
const Vector3 modelOffset(Vector3(10.0f, 10.0f, -10.0f));
Vector3 modelPos = GetPosition();
modelPos += GetForward() * modelOffset.x;
modelPos += GetRight() * modelOffset.y;
modelPos.z += modelOffset.z;
mFPSModel->SetPosition(modelPos);
// Initialize rotation to actor rotation
Quaternion q = GetRotation();
// Rotate by pitch from camera
q = Quaternion::Concatenate(q,
Quaternion(GetRight(), mCameraComp->GetPitch()));
mFPSModel->SetRotation(q);
Figure 9.1 demonstrates the first-person camera with a

first-person model. The aiming reticule is just a
SpriteComponent positioned in the center of the
screen.
Figure 9.1 First-person camera with first-person

model
FOLLOW CAMERA
A follow camera is a camera that follows behind a
target object. This type of camera is popular in many
games, including racing games where the camera
follows behind a car and third-person
action/adventure games such as Horizon Zero
Dawn. Because follow cameras see use in many
different types of games, there is a great deal of
variety in their implementation. This section focuses
on a follow camera tracking a car.
As was the case with the first-person character, you’ll

create a new actor called FollowActor to correspond to
the different style of movement when the game uses a
follow camera. The movement controls are W/S to move
the car forward and A/D to rotate the car left/right. The
normal MoveComponent supports both types of
movements, so it doesn’t require any changes here.
Basic Follow Camera

With a basic follow camera, the camera always
follows a set distance behind and above the owning
actor. Figure 9.2 gives the side view of this basic
follow camera. The camera is a set horizontal
distance HDist behind the car and a set vertical
distance VDist above the car. The target point of the
camera is not the car itself but a point TargetDist in
front of the car. This causes the camera to look at a
point a little in front of the car rather than directly at
the car itself.
Figure 9.2 Basic follow camera tracking a car
To compute the camera position, you use vector addition

and scalar multiplication. The camera position is HDist
units behind the owner and VDist units above the owner,
yielding the following equation:
OwnerForward and OwnerUp in this equation are the

owner’s forward and up vectors, respectively.
Similarly, TargetPos is just a point TargetDist units in

front of the owner:
In code, you declare a new subclass of

CameraComponent called FollowCamera. It has
member variables for the horizontal distance
(mHorzDist), vertical distance (mVertDist), and target
distance (mTargetDist). First, you create a function to
compute the camera position (using the previous
equation):
Vector3 FollowCamera::ComputeCameraPos() const
// Set camera position behind and above owner
cameraPos -= mOwner->GetForward() * mHorzDist;
cameraPos += Vector3::UnitZ * mVertDist;
return cameraPos;
Next, the FollowCamera::Update function uses this

camera position as well as a computed target position to
create the view matrix:
void FollowCamera::Update(float deltaTime)
// Target is target dist in front of owning actor
Vector3 target = mOwner->GetPosition() +
mOwner->GetForward() * mTargetDist;
// (Up is just UnitZ since we don't flip the camera)
Matrix4 view = Matrix4::CreateLookAt(GetCameraPos(), target,
Vector3::UnitZ);
Although this basic follow camera successfully tracks the

car as it moves through the game world, it appears very
rigid. Because the camera is always a set distance from
the target, it’s difficult to get a sense of speed.
Furthermore, when the car turns, it almost seems like
the world—not the car—is turning. So even though the
basic follow camera is a good starting point, it’s not a
very polished solution.
One simple change that improves the sense of speed is to

make the horizontal follow distance a function of the
speed of the owner. Perhaps at rest the horizontal
distance is 350 units, but when moving at max speed it
increases to 500. This makes it easier to perceive the
speed of the car, but the camera still seems stiff when the
car is turning. To solve the rigidity of the basic follow
camera, you can add springiness to the camera.
Adding a Spring
Rather than having the camera position instantly
changing to the position as per the equation, you can
have the camera adjust to this position over the
course of several frames. To accomplish this, you can
separate the camera position into an “ideal” camera
position and an “actual” camera position. The ideal
camera position is the position derived from the
basic follow camera equations, while the actual
camera position is what the view matrix uses.
Now, imagine that there’s a spring connecting the ideal

camera and the actual camera. Initially, both cameras are
at the same location. As the ideal camera moves, the
spring stretches and the actual camera also starts to
move—but at a slower rate. Eventually, the spring
stretches completely, and the actual camera moves just
as quickly as the ideal camera. Then, when the ideal
camera stops, the spring eventually compresses back to
its steady state. At this point, the ideal camera and actual
camera are at the same point again. Figure 9.3 visualizes
this idea of a spring connecting the ideal and actual
cameras.
Figure 9.3 A spring connecting the ideal and actual

cameras
Implementing a spring requires a few more member

variables in FollowCamera. A spring constant
(mSpringConstant) represents the stiffness of the
spring, with a higher value being stiffer. You also must
track the actual position (mActualPos) and the velocity
(mVelocity) of the camera from frame to frame, so you
add two vector member variables for these.
Listing 9.5 gives the code for FollowCamera::Update

with a spring. First, you compute a spring dampening
based on the spring constant. Next, the ideal position is
simply the position from the previously implemented
ComputeCameraPos function. You then compute the
difference between the actual and ideal positions and
compute an acceleration of the camera based on this
distance and a dampening of the old velocity. Next, you
compute the velocity and acceleration of the camera by
using the Euler integration technique introduced in
Chapter 3, “Vectors and Basic Physics.” Finally, the
target position calculation remains the same, and the
CreateLookAt function now uses the actual position as
opposed to the ideal one.
Listing 9.5 FollowCamera::Update Implementation

(with Spring)
void FollowCamera::Update(float deltaTime)

// Compute dampening from spring constant
float dampening = 2.0f * Math::Sqrt(mSpringConstant);
// Compute ideal position
Vector3 idealPos = ComputeCameraPos();
// Compute difference between actual and ideal

Vector3 diff = mActualPos - idealPos;
// Compute acceleration of spring
Vector3 acel = -mSpringConstant * diff -
dampening * mVelocity;
// Update velocity
mVelocity += acel * deltaTime;
// Update actual camera position
mActualPos += mVelocity * deltaTime;
// Target is target dist in front of owning actor

// Use actual position here, not ideal
Matrix4 view = Matrix4::CreateLookAt(mActualPos, target,
Vector3::UnitZ);
A big advantage of using a spring camera is that when

the owning object turns, the camera takes a moment to
catch up to the turn. This means that the side of the
owning object is visible as it turns. This gives a much
better sense that the object, not the world, is turning.
Figure 9.4 shows the spring follow camera in action.
Figure 9.4 Spring follow camera following a car as
it turns
The red sports car model used here is “Racing Car” by

Willy Decarpentrie, licensed under CC Attribution and
downloaded from https://sketchfab.com.
Finally, to make sure the camera starts out correctly at

the beginning of the game, you create a SnapToIdeal
function that’s called when the FollowActor first
initializes:
void FollowCamera::SnapToIdeal()
// Set actual position to ideal
mActualPos = ComputeCameraPos();
// Zero velocity
mVelocity = Vector3::Zero;
// Compute target and view
Matrix4 view = Matrix4::CreateLookAt(mActualPos, target,
Vector3::UnitZ);
}
ORBIT CAMERA
An orbit camera focuses on a target object and
orbits around it. This type of camera might be used
in a builder game such as Planet Coaster, as it allows
the player to easily see the area around an object.
The simplest implementation of an orbit camera
stores the camera’s position as an offset from the
target rather than as an absolute world space
position. This takes advantage of the fact that
rotations always rotate about the origin. So, if the
camera position is an offset from the target object,
any rotations are effectively about the target object.
In this section, you’ll create an OrbitActor as well as

an OrbitCamera class. A typical control scheme uses
the mouse for both yaw and pitch around the object. The
input code that converts relative mouse movement into
rotation values is like the code covered in the “First-
Person Camera” section, earlier in this chapter. However,
you add a restriction that the camera rotates only when
the player is holding down the right mouse button (since
this is a typical control scheme). Recall that the
SDL_GetRelativeMouseState function returns the
state of the buttons. The following conditional tests
whether the player is holding the right mouse button:

if (buttons & SDL_BUTTON(SDL_BUTTON_RIGHT))
The OrbitCamera class requires the following member

variables:
// Offset from target
Vector3 mOffset;
// Up vector of camera
Vector3 mUp;
// Rotation/sec speed of pitch
float mPitchSpeed;
// Rotation/sec speed of yaw
float mYawSpeed;
The pitch speed (mPitchSpeed) and yaw speed

(mYawSpeed) simply track the current rotations per
second of the camera for each type of rotation. The
owning actor can update these speeds as needed, based
on the mouse rotation. In addition, the OrbitCamera
needs to track the offset of the camera (mOffset), as
well as the up vector of the camera (mUp). The up vector
is needed because the orbit camera allows full 360-
degree rotations in both yaw and pitch. This means the
camera could flip upside down, so you can’t universally
pass in (0, 0, 1) as up. Instead, you must update the up
vector as the camera rotates.
The constructor for OrbitCamera initializes

mPitchSpeed and mYawSpeed both to zero. The
mOffset vector can initialize to any value, but here you
initialize it to 400 units behind the object (-400, 0, 0).
The mUp vector initializes to world up (0, 0, 1).

OrbitCamera::Update. First, you create a quaternion
representing the amount of yaw to apply this frame,
which is about the world up vector. You use this
quaternion to transform both the camera offset and up.
Next, you compute the camera forward vector from the
new offset. The cross product between the camera
forward and camera yields the camera right vector. You
then use this camera right vector to compute the pitch
quaternion and transform both the camera offset and up
by this quaternion, as well.
Listing 9.6 OrbitCamera::Update Implementation

void OrbitCamera::Update(float deltaTime)
// Create a quaternion for yaw about world up
Quaternion yaw(Vector3::UnitZ, mYawSpeed * deltaTime);

// Transform offset and up by yaw
mOffset = Vector3::Transform(mOffset, yaw);
mUp = Vector3::Transform(mUp, yaw);
// Compute camera forward/right from these vectors
// Forward owner.position - (owner.position + offset)
// = -offset
Vector3 forward = -1.0f * mOffset;
forward.Normalize();
Vector3 right = Vector3::Cross(mUp, forward);
right.Normalize();

// Create quaternion for pitch about camera right
Quaternion pitch(right, mPitchSpeed * deltaTime);
// Transform camera offset and up by pitch
mOffset = Vector3::Transform(mOffset, pitch);
mUp = Vector3::Transform(mUp, pitch);
// Compute transform matrix
Vector3 target = mOwner->GetPosition();

Vector3 cameraPos = target + mOffset;
Matrix4 view = Matrix4::CreateLookAt(cameraPos, target, mUp);
For the look-at matrix, the target position of the camera

is simply the owner’s position, the camera position is the
owner’s position plus the offset, and the up is the camera
up. This yields the final orbited camera. Figure 9.5
demonstrates the orbit camera with the car as the target.
Figure 9.5 Orbit camera focused on the car
SPLINE CAMERA
A spline is a mathematical representation of a curve
specified by a series of points on the curve. Splines
are popular in games because they enable an object
to smoothly move along a curve over some period.
This can be very useful for a cutscene camera
because the camera can follow a predefined spline
path. This type of camera also sees use in games like
God of War, where the camera follows along a set
path as the player progresses through the world.
The Catmull-Rom spline is a type of spline that’s

relatively simple to compute, and it is therefore used
frequently in games and computer graphics. This type of
spline minimally requires four control points, named P0
through P3. The actual curve runs from P1 to P2, while P0
is a control point prior to the curve and P3 is a control
point after the curve. For best results, you can space
these control points roughly evenly along the curve—and
you can approximate this with Euclidean distance.
Figure 9.6 illustrates a Catmull-Rom spline with four
control points.
Figure 9.6 Catmull-Rom spline

Given these four control points, you can express the
position between P1 and P2 as the following parametric
equation, where t = 0 is at P1 and t = 1 is at P2:
Although the Catmull-Rom spline equation has only four

control points, you can extend the spline to any arbitrary
number of control points. This works provided that there
still is one point before the path and one point after the
path because those control points are not part of the
path. In other words, you need n + 2 points to represent
a curve of n points. You can then take any sequence of
four neighboring points and substitute them into the
spline equation.
To implement a camera that follows a spline path, you

first create a struct to define a spline. The only member
data Spline needs is a vector of the control points:
struct Spline
// Control points for spline
// (Requires n + 2 points where n is number
// of points in segment)
std::vector<Vector3> mControlPoints;
// Given spline segment where startIdx = P1,
// compute position based on t value
Vector3 Compute(size_t startIdx, float t) const;
size_t GetNumPoints() const { return mControlPoints.size(); }
};
The Spline::Compute function applies the spline

equation given a start index corresponding to P1 and a t
value in the range [0.0, 1.0]. It also performs boundary
checks to make sure startIdx is a valid index, as shown
in Listing 9.7.
Listing 9.7 Spline::Compute Implementation

Vector3 Spline::Compute(size_t startIdx, float t) const
// Check if startIdx is out of bounds
if (startIdx >= mControlPoints.size())
{ return mControlPoints.back(); }
else if (startIdx == 0)
{ return mControlPoints[startIdx]; }
else if (startIdx + 2 >= mControlPoints.size())

{ return mControlPoints[startIdx]; }
// Get p0 through p3
Vector3 p0 = mControlPoints[startIdx - 1];
Vector3 p1 = mControlPoints[startIdx];
Vector3 p2 = mControlPoints[startIdx + 1];
Vector3 p3 = mControlPoints[startIdx + 2];
// Compute position according to Catmull-Rom equation
Vector3 position = 0.5f * ((2.0f * p1) + (-1.0f * p0 + p2) * t +
(2.0f * p0 - 5.0f * p1 + 4.0f * p2 - p3) * t * t +
(-1.0f * p0 + 3.0f * p1 - 3.0f * p2 + p3) * t * t * t);
return position;
The SplineCamera class then needs a Spline in its

member data. In addition, it tracks the current index
corresponding to P1, the current t value, a speed, and
whether the camera should move along the path:
// Spline path camera follows

Spline mPath;
// Current control point index and t
size_t mIndex;
float mT;
// Amount t changes/sec
float mSpeed;
// Whether to move the camera along the path
bool mPaused;
The spline camera updates by first increasing the t value

as a function of speed and delta time. If the t value is
greater than or equal to 1.0, P1 advances to the next point
on the path (assuming that there are enough points on
the path). Advancing P1 also means you must subtract 1
from the t value. If the spline has no more points, the
spline camera pauses.
For the camera calculations, the position of the camera is

simply the point computed from the spline. To compute
the target point, you increase t by a small delta to
determine the direction the spline camera is moving.
Finally, the up vector stays at (0, 0, 1), which assumes
that you do not want the spline to flip upside down.
Listing 9.8 gives the code for SplineCamera::Update,
and Figure 9.7 shows the spline camera in action.
Figure 9.7 Spline camera in a game
Listing 9.8 SplineCamera::Update Implementation

void SplineCamera::Update(float deltaTime)
// Update t value
if (!mPaused)
mT += mSpeed * deltaTime;
// Advance to the next control point if needed.

// This assumes speed isn't so fast that you jump past
// multiple control points in one frame.
if (mT >= 1.0f)
// Make sure we have enough points to advance the path
if (mIndex < mPath.GetNumPoints() - 3)
mIndex++;
mT = mT - 1.0f;
else
// Path's done, so pause
mPaused = true;
// Camera position is the spline at the current t/index
Vector3 cameraPos = mPath.Compute(mIndex, mT);

// Target point is just a small delta ahead on the spline
Vector3 target = mPath.Compute(mIndex, mT + 0.01f);
// Assume spline doesn't flip upside-down
const Vector3 up = Vector3::UnitZ;
UNPROJECTION
Given a point in world space, to transform it into clip
space, you first multiply by the view matrix followed
by the projection matrix. Imagine that the player in a
first-person shooter wants to fire a projectile based
on the screen position of the aiming reticule. In this
case, the aiming reticule position is a coordinate in
screen space, but to correctly fire the projectile, you
need a position in world space. An unprojection is
a calculation that takes in a screen space coordinate
and converts it into a world space coordinate.
Assuming the screen space coordinate system described

in Chapter 5, “OpenGL,” the center of the screen is (0, 0),
the top-left corner is (-512, 384), and the bottom-right
corner is (512, -384). The first step to calculating an
unprojection is converting a screen space coordinate into
a normalized device coordinate with a range of [-1, 1] for
both the x and y components:
However, the issue is that any single (x, y) coordinate can

correspond to any z coordinate in the range [0, 1], where
0 is a point on the near plane (right in front of the
camera), and 1 is a point on the far plane (the maximum
distance you can see from the camera). So, to correctly
perform the unprojection, you also need a z component
in the range [0, 1]. You then represent this as a
homogenous coordinate:
Now you construct an unprojection matrix, which is

simply the inverse of the view-projection matrix:
When multiplying the NDC point by the unprojection

matrix, the w component changes. However, you need to
renormalize the w component (setting it back to 1) by
dividing each component by w. This yields the following
calculation for the point in world space:
You add a function for an unprojection into the

Renderer class because it’s the only class with access to
both the view and projection matrices. Listing 9.9
provides the implementation for Unproject. In this
code, the TransformWithPerspDiv function does the
w component renormalization.
Listing 9.9 Renderer::Unproject Implementation

Vector3 Renderer::Unproject(const Vector3& screenPoint) const
// Convert screenPoint to device coordinates (between -1 and +1)
Vector3 deviceCoord = screenPoint;
deviceCoord.x /= (mScreenWidth) * 0.5f;
deviceCoord.y /= (mScreenHeight) * 0.5f;
// Transform vector by unprojection matrix
Matrix4 unprojection = mView * mProjection;
unprojection.Invert();
return Vector3::TransformWithPerspDiv(deviceCoord, unprojection);
You can use Unproject to calculate a single world space

position. However, it some cases, it’s more useful to
construct a vector in the direction of the screen space
point, as it gives opportunities for other useful features.
One such feature is picking, which is the capability to
click to select an object in the 3D world. Figure 9.8
illustrates picking with a mouse cursor.
Figure 9.8 Picking with a vector in the direction of

the screen space coordinate of the mouse
To construct a direction vector, you use Unproject

twice, once for a start point and once for the end point.
Then simply use vector subtraction and normalize this
vector, as in the implementation of
Renderer::GetScreenDirection in Listing 9.10.
Note how the function computes both the start point of
the vector in world space and the direction.
Listing 9.10 Renderer::GetScreenDirection
Implementation
void Renderer::GetScreenDirection(Vector3& outStart,
Vector3& outDir) const
// Get start point (in center of screen on near plane)
Vector3 screenPoint(0.0f, 0.0f, 0.0f);
outStart = Unproject(screenPoint);
// Get end point (in center of screen, between near and far)
screenPoint.z = 0.9f;
Vector3 end = Unproject(screenPoint);
// Get direction vector
outDir = end - outStart;
outDir.Normalize();
GAME PROJECT
different cameras discussed in the chapter, as well as
the unprojection code. The code is available in the
Mac.
The camera starts out in first-person mode. To switch

between the different cameras, use the 1 through 4 keys:
1—Enable first-person camera mode
2—Enable follow camera mode
3—Enable orbit camera mode
4—Enable spline camera mode and restart the spline path
Depending on the camera mode, the character has

different controls, summarized below:
First-person—Use W/S to move forward and back, A/D to strafe, and

the mouse to rotate
Follow—Use W/S to move forward and back and use A/D to rotate
(yaw)
Orbit camera mode—Hold down the right mouse button and move
the mouse to rotate
Spline camera mode—No controls (moves automatically)
In addition, in any camera mode, you can left-click to

compute the unprojection. This positions two spheres—
one at the “start” position of the vector and one at the
“end” position.
SUMMARY
This chapter shows how to implement many
different types of cameras. The first-person camera
presents the world from the perspective of a
character moving through it. A typical first-person
control scheme uses the WASD keys for movement
and the mouse for rotation. Moving the mouse left
and right rotates the character, while moving the
mouse up and down pitches the view. You can
additionally use the first-person view pitch to orient
a first-person model.
A basic follow camera follows rigidly behind an object.

However, this camera does not look polished when
rotating because it’s difficult to discern if the character or
the world is rotating. An improvement is to incorporate a
spring between “ideal” and “actual” camera positions.
This adds smoothness to the camera that’s especially
noticeable when turning.
An orbit camera rotates around an object, typically with

mouse or joystick control. To implement orbiting, you
represent the camera as an offset from the target object.
Then, you can apply both yaw and pitch rotations by
using quaternions and some vector math to yield the
final view.
A spline is a curve defined by points on the curve. Splines

are popular for cutscene cameras. The Catmull-Rom
spline requires a minimum of n + 2 points to represent a
curve of n points. By applying the Catmull-Rom spline
equations, you can create a camera that follows along
this spline path.
Finally, an unprojection has many uses, such as selecting

or picking objects with the mouse. To compute an
unprojection, you first transform a screen space point
into normalized device coordinates. You then multiply by
the unprojection matrix, which is simply the inverse of
the view-projection matrix.
ADDITIONAL READING
There are not many books dedicated to the topic of
game cameras. However, Mark Haigh-Hutchinson,
the primary programmer for the Metroid Prime
camera system, provides an overview of many
different techniques relevant for game cameras.
Haigh-Hutchinson, Mark. Real-Time

Cameras. Burlington: Morgan Kaufmann,
2009.
EXERCISES
In this chapter’s exercises, you will add features to
some of the cameras. In the first exercise, you add
mouse support to the follow camera, and in the
second exercise, you add features to the spline
camera.
Exercise 9.1
Many follow cameras have support for user-
controlled rotation of the camera. For this exercise,
add code to the follow camera implementation that
allows the user to rotate the camera. When the player
holds down the right mouse button, apply an
additional pitch and yaw rotation to the camera.
When the player releases the right mouse button, set
the pitch/yaw rotation back to zero.
The code for the rotation is like the rotation code for the
orbit camera. Furthermore, as with the orbit camera, the
code can no longer assume that the z-axis is up. When
the player releases the mouse button, the camera won’t
immediately snap back to the original orientation
because of the spring. However, this is aesthetically
pleasing, so there’s no reason to change this behavior!
Exercise 9.2
Currently, the spline camera goes in only one
direction on the path and stops upon reaching the
end. Modify the code so that when the spline hits the
end of the path, it starts moving backward.
CHAPTER 10
COLLISION DETECTION
You use collision detection to determine

whether objects in the game world intersect
with each other. While earlier chapters
discuss some basic ways to check for
collisions, this chapter takes a deeper look at
the topic. This chapter first introduces the
basic geometric types commonly used in
games and then covers the calculation of
intersection between these types. It
concludes with a discussion of how to
incorporate collisions into game behavior.
GEOMETRIC TYPES
Collision detection for games utilizes several
different concepts from geometry and linear algebra.
This section covers some fundamental geometric
types commonly used in games, such as line
segments, planes, and boxes. Each geometric type
discussed here has a corresponding declaration in
the Collision.h header file included in this
chapter’s game project.
Line Segments
A line segment comprises start and end points:
struct LineSegment
Vector3 mStart;
Vector3 mEnd;
};
To calculate any arbitrary point on a line segment, you

can use the following parametric equation, where Start
and End are the start and end points, and t is the
parameter:
For convenience, you can add a member function to
LineSegment that returns a point on the line segment
given a t value:
Vector3 LineSegment::PointOnSegment(float t) const
return mStart + (mEnd - mStart) * t;
The parametric representation of a line segment easily

expands to define a ray or a line. A ray follows the above
equation, but the bounds for t are as follows:
Similarly, a line has an unbounded t:
Line segments and rays are versatile primitives for many

different types of collision detection in games. For
example, firing a bullet in a straight line or testing for
landing on a ground can use a line segment. You can also
use line segments for aiming reticules (as in Chapter 11,
“User Interfaces”), sound occlusion tests (as in Chapter
7, “Audio”), or mouse picking (as in Chapter 9,
“Cameras”).
Another useful operation is finding the minimal distance

between a line segment and an arbitrary point. Imagine
that a line segment starts at point A and ends at point B.
Given an arbitrary point C, you want to find the minimal
distance between the segment and point C. There are
three different cases to consider, shown in Figure 10.1.
Figure 10.1 Three cases for minimal distance

between a point and a line segment
In the first case, shown in Figure 10.1(a), the angle

between AB and AC is greater than 90°. You can test this
by using the dot product because a negative dot product
between two vectors means they form an obtuse angle. If
this is true, the minimal distance between C and the line
segment is the length of the vector AC.
In the second case, shown in Figure 10.1(b), the angle

between BA and BC is greater than 90°. As in the first
case, you can test this with a dot product. If this is true,
the minimal distance is the length of BC.
In the final case, shown in Figure 10.1(c), you draw a new

line segment from AB to C that’s perpendicular to AB.
The distance of this new line segment is the minimal
distance between C and AB. To figure out this line
segment, you first need to compute the vector p.
You already know the direction of p, as it’s in the same

direction as the normalized AB. To figure out the
distance of p, you apply a property of the dot product
called the scalar projection. Given a unit vector and a
non-unit vector, extend (or contract) the unit vector such
that a right triangle forms with the non-unit vector. The
dot product then returns the length of this extended unit
vector.
In this example, the length of p is the dot product

between AC and a normalized AB:
The vector p is then the scalar multiplication between the

length of p and the normalized AB:
Using some algebraic manipulation—and remembering

that the length squared of a vector is the same as a dot
product with itself—you can simplify p as follows:
Finally, you construct the vector from p to AC, and the

length of this vector is the minimum distance from AB to
C:
Remember that because the distance must be positive in

this case, you can square both sides of the equation to
instead get the minimal distance squared from AB to C:
This way you avoid the expensive square root operation.
For the most part, you will calculate the distance squared
instead of distance throughout this chapter. Listing 10.1
gives the code for this MinDistSq function.
Listing 10.1 LineSegment::MinDistSq

Implementation
float LineSegment::MinDistSq(const Vector3& point) const
// Construct vectors
Vector3 ab = mEnd - mStart;
Vector3 ba = -1.0f * ab;
Vector3 ac = point - mStart;
Vector3 bc = point - mEnd;
// Case 1: C projects prior to A
if (Vector3::Dot(ab, ac) < 0.0f)
return ac.LengthSq();
// Case 2: C projects after B
else if (Vector3::Dot(ba, bc) < 0.0f)

{
return bc.LengthSq();
// Case 3: C projects onto line
else
// Compute p
float scalar = Vector3::Dot(ac, ab)
/ Vector3::Dot(ab, ab);
Vector3 p = scalar * ab;
// Compute length squared of ac - p
return (ac - p).LengthSq();
Planes
A plane is a flat, two-dimensional surface that
extends infinitely, much as a line is a one-
dimensional object that extends infinitely. In a game,
you may use a plane as an abstraction for the ground
or walls. The equation of a plane is as follows:
where P an arbitrary point on the plane, n is the normal
to the plane, and d is the signed minimal distance
between the plane and the origin.
In code, a typical test is whether a point lies on the plane

(and thus satisfies the plane equation). Because of this,
the definition of the Plane struct only stores the normal
and d:
struct Plane
Vector3 mNormal;
float mD;
};
By definition, a triangle lies on a single plane. So, given a

triangle, you can derive the equation of that plane. You
compute the normal to the triangle with the cross
product, which corresponds to the normal of the plane.
You already know an arbitrary point on the plane, as
well, because all three of the vertices of the triangle are
on the plane. Given this normal and point, you can then
solve for d, as in Listing 10.2.
Listing 10.2 Constructing a Plane from Three Points

Plane::Plane(const Vector3& a, const Vector3& b, const Vector3& c)

{
// Compute vectors from a to b and a to c
Vector3 ab = b - a;
Vector3 ac = c - a;
// Cross product and normalize to get normal
mNormal = Vector3::Cross(ab, ac);
mNormal.Normalize();
// d = -P dot n
mD = -Vector3::Dot(a, mNormal);
Finding the minimum distance between an arbitrary

point C and the plane is simpler than for a line segment,
though it also uses the scalar projection property of the
dot product. Figure 10.2 illustrates the calculation,
showing the plane from a side view.
Figure 10.2 Calculations for minimum distance
between point C and a plane
You already know the normal of the plane n and the

minimal distance d between the origin and the plane.
You need to calculate the scalar projection of C onto the
normal n, which is simply the dot product:
Then, the difference between d and this scalar projection

yields a signed distance between C and the plane:
A negative value means C is below the plane (facing away

from the normal), while a positive value means C is
above. The signed distance calculation translates to the
following code:
float Plane::SignedDist(const Vector3& point) const
return Vector3::Dot(point, mNormal) - mD;
Bounding Volumes
Modern 3D games have characters and objects
drawn with thousands of triangles. When
determining whether two objects collide, it’s not
efficient to test all the triangles comprising the
object. For this reason, games use simplified
bounding volumes, such as boxes or spheres.
When deciding whether two objects intersect, the
game uses the simplified collision for calculations.
This yields greatly improved efficiency.
Spheres
The simplest representation of the bounds of a 3D
object is a sphere. The definition of a sphere only
requires the position of the center of the sphere and a
radius:
struct Sphere
{
Vector3 mCenter;
float mRadius;
};
As illustrated in Figure 10.3, bounding spheres fit around

some objects better than others. For example, a sphere
around a humanoid character has a lot of empty space
between the character and the bounds of the sphere.
Having loose bounds for an object increases the number
of false positive collisions, meaning the bounding
volumes for two objects intersect but the objects
themselves do not. For instance, if a first-person shooter
used bounding spheres for humanoids, players could
shoot way to the left or right of a character, and the game
would count that as a hit.
Figure 10.3 Bounding spheres for different objects
However, the advantage of using bounding spheres is

that intersection calculations are extremely efficient.
Furthermore, rotation has no effect on a sphere, so a
bounding sphere works regardless of the rotation of the
underlying 3D object. And for some objects, like balls,
spheres perfectly express the bounds.
Axis-Aligned Bounding Boxes
In 2D, an axis-aligned bounding box (AABB) is
a rectangle where the edges are parallel to either the
x-axis or y-axis. Similarly, in 3D, an AABB is a
rectangular prism where every face of the prism is
parallel to one of the coordinate axis planes.
You can define an AABB by two points: a minimum point

and a maximum point. In 2D, the minimum point
corresponds to the bottom-left point, while the
maximum corresponds to the top-right point. In other
words, the minimum point has the minimal x and y
values for the box, and the maximum point has the
maximal x and y values for the box. This carries over
directly to 3D, where the minimum point has the
minimal x, y, and z values, and the maximum is likewise
the maximal x, y, and z values. This translates to the
following struct:
struct AABB
Vector3 mMin;
Vector3 mMax;
};
One useful operation with an AABB is constructing it

from a series of points. For example, when loading in a
model, you have a sequence of vertices, and you can use
this sequence of vertices to define an AABB for the
model. To do this, you can create a new function called
UpdateMinMax that takes in a point and updates min
and max, accounting for this point:
void AABB::UpdateMinMax(const Vector3& point)
// Update each component separately
mMin.x = Math::Min(mMin.x, point.x);
mMin.y = Math::Min(mMin.y, point.y);
mMin.z = Math::Min(mMin.z, point.z);
mMax.x = Math::Max(mMax.x, point.x);
mMax.y = Math::Max(mMax.y, point.y);
mMax.z = Math::Max(mMax.z, point.z);
Because you do not know where the new point is in

relation to all the other points, you must test every
component independently to decide which components
of min and max should update.
Then, given a container of points, you first initialize min

and max of the AABB to the first point in the container.
For each remaining point, you simply call
UpdateMinMax:
// Assume points is a std::vector<Vector3>
AABB box(points[0], points[0]);
for (size_t i = 1; i < points.size(); i++)
box.UpdateMinMax(points[i]);
Because an AABB must keep its sides parallel to the

coordinate planes, rotating an object does not rotate the
AABB. Instead, it changes the dimensions of the AABB,
as in Figure 10.4. In some cases, it may be desirable not
to compute AABB rotations. For example, most
humanoid characters in games only rotate about the up
axis. If you make the AABB for the character wide
enough, rotating the character does not change the AABB
enough to warrant rotating the AABB (though watch out
for animations that move the character too much).
However, for other objects, it is necessary to compute a
rotated AABB.
Figure 10.4 AABBs for different orientations of a
character
One way to compute an AABB after rotation is to first

construct the eight points representing the corners of the
AABB. These points are simply all possible permutations
of the min and max x, y, and z components. Then you
rotate each point individually and use the
UpdateMinMax function to create a new AABB from
these rotated points. Note that this process, shown in
Listing 10.3, does not compute the minimal possible
AABB of the underlying object after rotation. Thus, the
game should save the original object space AABB to
avoid error propagation after multiple rotations.
Listing 10.3 AABB::Rotate Implementation

void AABB::Rotate(const Quaternion& q)
// Construct the 8 points for the corners of the box
std::array<Vector3, 8> points;
// Min point is always a corner

points[0] = mMin;
// Permutations with 2 min and 1 max
points[1] = Vector3(mMax.x, mMin.y, mMin.z);
points[2] = Vector3(mMin.x, mMax.y, mMin.z);
points[3] = Vector3(mMin.x, mMin.y, mMax.z);
// Permutations with 2 max and 1 min
points[4] = Vector3(mMin.x, mMax.y, mMax.z);
points[5] = Vector3(mMax.x, mMin.y, mMax.z);
points[6] = Vector3(mMax.x, mMax.y, mMin.z);
// Max point corner
points[7] = Vector3(mMax);
// Rotate first point
Vector3 p = Vector3::Transform(points[0], q);

// Reset min/max to first point rotated
mMin = p;
mMax = p;
// Update min/max based on remaining points, rotated
for (size_t i = 1; i < points.size(); i++)
p = Vector3::Transform(points[i], q);
UpdateMinMax(p);
Oriented Bounding Boxes

An oriented bounding box (OBB) does not have
the parallel restrictions of an AABB. This means that
an OBB maintains the tightness of its bounds,
regardless of the rotation of the underlying object, as
in Figure 10.5. One way to represent an OBB is with a
center point, a quaternion for the rotation, and the
extents (width, height, and depth) of the box:
struct OBB
Vector3 mCenter;
Vector3 mExtents;
};
Figure 10.5 An oriented bounding box for a
humanoid character that’s rotated
Although it’s tempting to use OBBs, the downside is that

collision computations are far more expensive with OBBs
than with AABBs.
Capsules
A capsule is a line segment with a radius:
struct Capsule
LineSegment mSegment;
float mRadius;
};
Capsules are popularly used for representing humanoid

characters in a game, as in Figure 10.6. A capsule can
also represent a sphere moving over a set period because
there’s a start point and an end point for the sphere’s
movement, and the sphere of course has a radius.
Figure 10.6 A capsule for a humanoid character
Convex Polygons
Sometimes, a game may need bounds for an object
that are more accurate than any of the basic shapes.
For a 2D game, the object might have bounds
represented as a convex polygon. Recall that a
polygon is convex if all its interior angles are less
than 180°.
You can represent a convex polygon as a collection of

vertices:
struct ConvexPolygon
{
// Vertices have a clockwise ordering
std::vector<Vector2> mVertices;
};
These vertices should have a set ordering, such as

clockwise or counterclockwise along the polygon’s edge.
Without an ordering, intersections are more difficult to
compute.
Note that this representation assumes that the developer

correctly uses it, and it does no testing to make sure the
polygon is convex and has vertices in a clockwise order.
INTERSECTION TESTS
Once the game is using geometric types to represent
game objects, the next step is to test for intersections
between these objects. This section looks at a series
of useful tests. First, it explores whether an object
contains a point. Then, it looks at intersections
between different types of bounding volumes. Next,
it looks at intersections between a line segment and
other objects. Finally, this section covers how to
handle dynamically moving objects.
Contains Point Tests

Testing whether a shape contains a point is useful by
itself. For example, you might use this type of test for
finding out whether a player is inside a region of the
game world. In addition, some shape intersection
algorithms rely on finding the point closest to an
object and then figuring out if that point is inside the
object. This section considers a point to be
“contained” by a shape even if it’s technically on one
of the edges of that shape.
Sphere Contains Point Tests

To figure out if a sphere contains a point, first find
the distance between the point and the center of the
sphere. If this distance is less than or equal to the
radius, then the sphere contains the point.
Because the distance and radius are both positive values,

you can optimize this comparison by squaring both sides
of the inequality. This way, you avoid the expensive
square root operation and add only one multiplication,
making it far more efficient:
bool Sphere::Contains(const Vector3& point) const
// Get distance squared between center and point
float distSq = (mCenter - point).LengthSq();
return distSq <= (mRadius * mRadius);
}
AABB Contains Point Tests
Given a 2D axis-aligned box, a point is outside the
box if any of the following cases are true: The point is
to the left of the box, the point is to the right of the
box, the point is above the box, or the point is below
the box. If none of these cases are true, then the box
must contain the point.
You can check for this by simply comparing components

of the point to the min and max points of the box. For
example, a point is to the left of the box if its x
component is less than min.x.
This concept easily extends to a 3D AABB. However,

instead of having four checks—one for each side of a 2D
box—you have six checks because there are six faces for a
3D AABB:
bool AABB::Contains(const Vector3& point) const
bool outside = point.x < mMin.x ||
point.y < mMin.y ||
point.z < mMin.z ||
point.x > mMax.x ||
point.y > mMax.y ||
point.z > mMax.z;

// If none of these are true, the point is inside the box
return !outside;
Capsule Contains Point Tests

For testing whether a capsule contains a point, you
first compute the minimum distance squared
between the point and the line segment. To do so,
you can use the existing
LineSegment::MinDistSq function declared
earlier. You know the capsule contains the point if
that minimal distance squared is less than or equal to
the radius squared:
bool Capsule::Contains(const Vector3& point) const
// Get minimal dist. sq. between point and line segment
float distSq = mSegment.MinDistSq(point);
return distSq <= (mRadius * mRadius);
Convex Polygon Contains Point (2D) Tests

There are multiple ways to test whether a 2D polygon
contains a point. One of the simplest approaches is to
construct vectors from the point to each pair of
adjacent vertices. Then you can use the dot product
and arccosine to find the angle formed by these
vectors. If the sum of all these angles is close to 360°,
then the point is inside the polygon. Otherwise, the
point is outside the polygon. Figure 10.7 illustrates
this concept.
Figure 10.7 Angle summation test for whether a

convex polygon contains a point
The code for this type of test, shown in Listing 10.4,

relies on the fact that two adjacent vertices are also at
adjacent indices in the convex polygon’s vector.
Listing 10.4 ConvexPolygon::Contains

Implementation
bool ConvexPolygon::Contains(const Vector2& point) const
float sum = 0.0f;
Vector2 a, b;
for (size_t i = 0; i < mVertices.size() - 1; i++)
// From point to first vertex
a = mVertices[i] - point;
a.Normalize();
// From point to second vertex
b = mVertices[i + 1] - point;
b.Normalize();
// Add angle to sum
sum += Math::Acos(Vector2::Dot(a, b));
// Compute angle for last vertex and first vertex
a = mVertices.back() - point;
a.Normalize();
b = mVertices.front() - point;
b.Normalize();
sum += Math::Acos(Vector2::Dot(a, b));
// Return true if approximately 2pi
return Math::NearZero(sum - Math::TwoPi);
Unfortunately, this angle summation approach is not

terribly efficient because it requires several square roots
and arccosine calculations. Other, more complex,
methods are more efficient. One such method is to draw
an infinite ray starting at the point and to then count the
number of edges the ray intersects. If the ray intersects
an odd number of edges, the point is inside the polygon;
otherwise, it’s outside. This ray method works for both
convex and concave polygons.
Bounding Volume Tests

It’s very common to compute intersection tests
between different bounding volumes. For example,
imagine that both the player and the wall use AABBs
for collision. When the player tries to move forward,
you can test if the player’s bounding volume
intersects with the wall bounding volume. If they
intersect, then you can fix the player’s position so
that they no longer intersect. (You will see how to do
this later in this chapter.) This section doesn’t cover
all possible intersections between the different types
of bounding volumes discussed earlier, but it touches
on some important ones.
Sphere Versus Sphere Test

Two spheres intersect if the distance between their
centers is less than or equal to the sum of their radii.
As with the sphere contains point test, you can
square both sides of the inequality for efficiency,
using the following function:
bool Intersect(const Sphere& a, const Sphere& b)
float distSq = (a.mCenter - b.mCenter).LengthSq();
float sumRadii = a.mRadius + b.mRadius;
return distSq <= (sumRadii * sumRadii);
AABB Versus AABB Test

The logic for testing AABB intersection is like the
logic for whether an AABB contains a point. You test
for the cases where the two AABBs cannot intersect.
If none of these tests are true, then the AABBs must
intersect. For 2D AABBs, boxes A and B do not
intersect if A is to the left of B, A is to the right of B, A
is above B, or A is below B. You test this by
leveraging the min and max points, as before. For
example, A is to the left of B if the max.x of A is less
than the min.x of B. Figure 10.8 illustrates these
tests for 2D AABBs.
Figure 10.8 Four cases where two 2D AABBs do not

intersect
As before, when switching from 2D AABBs to 3D AABBs,

you must add two more checks, for a total of six:
bool Intersect(const AABB& a, const AABB& b)
bool no = a.mMax.x < b.mMin.x ||

a.mMax.y < b.mMin.y ||
a.mMax.z < b.mMin.z ||
b.mMax.x < a.mMin.x ||
b.mMax.y < a.mMin.y ||
b.mMax.z < a.mMin.z;
// If none of these are true, they must intersect
return !no;
This form of AABB intersection is an application of the

separating axis theorem, which states that if two
convex objects A and B do not intersect, then there must
exist an axis that separates A from B. In the AABB case,
you’re testing the three coordinate axes to see if there is
separation between the boxes on any of these axes. If the
AABBs have separation on any coordinate axis, then, by
the separating axis theorem, they cannot intersect. You
can extend this approach to oriented bounding boxes, as
discussed in Exercise 10.3 at the end of this chapter, and,
in fact, to any set of convex objects.
Sphere Versus AABB Test

For sphere versus AABB intersection, you first need
to calculate the minimum distance squared between
the center of the sphere and the box. The algorithm
for finding the minimum distance between a point
and an AABB tests each component individually. For
each component, there are three cases: The point’s
component is less than min, the point’s component is
between min and max, or the point’s component is
greater than max. In the middle case, the distance
between the point and box for that axis is zero. In the
other two cases, the distance between the point and
the box for the axis is the distance to the closest edge
(either min or max). Figure 10.9 illustrates this for a
2D AABB.
Figure 10.9 Distance between points and AABB
You can express this with multiple Math::Max function

calls. For example, the distance in the x direction is as
follows:
float dx = Math::Max(mMin.x - point.x, 0.0f);
dx = Math::Max(dx, point.x - mMax.x);

This works because if point.x < min.x, then min.x
- point.x is the largest of the three values and the
delta for the x-axis. Otherwise, if min.x < point.x <
max.x, then zero is the highest. Finally, if point.x >
max.x, then point.x - max.x is the highest. Once
you have the delta for all three axes, you then use the
distance formula to compute the final distance squared
between the point and the AABB:
float AABB::MinDistSq(const Vector3& point) const
// Compute differences for each axis
float dx = Math::Max(mMin.x - point.x, 0.0f);
dx = Math::Max(dx, point.x - mMax.x);
float dy = Math::Max(mMin.y - point.y, 0.0f);
dy = Math::Max(dy, point.y - mMax.y);
float dz = Math::Max(mMin.z - point.z, 0.0f);
dz = Math::Max(dy, point.z - mMax.z);
// Distance squared formula
return dx * dx + dy * dy + dz * dz;
}
Once you have the MinDistSq function, you can
implement sphere versus AABB intersection. You find
the minimal distance squared between the center of the
sphere and the AABB. If it is less than or equal to the
radius squared, then the sphere and AABB intersect:
bool Intersect(const Sphere& s, const AABB& box)
float distSq = box.MinDistSq(s.mCenter);
return distSq <= (s.mRadius * s.mRadius);
Capsule Versus Capsule Test

Intersecting two capsules is conceptually
straightforward. Because both capsules are line
segments with radii, you first find the minimal
distance squared between these line segments. If the
distance squared is less than or equal to the sum of
radii squared, then the two capsules intersect:
bool Intersect(const Capsule& a, const Capsule& b)
float distSq = LineSegment::MinDistSq(a.mSegment,
b.mSegment);
float sumRadii = a.mRadius + b.mRadius;
return distSq <= (sumRadii * sumRadii);
Unfortunately, computing the minimal distance between

two line segments is complex due to several edge cases.
This chapter doesn’t go into those details, but its source
code provides an implementation of MinDistSq for two
line segments.
Line Segment Tests

As mentioned earlier, line segments are versatile in
collision detection. This chapter’s game project uses
line segment tests for testing whether the ball
projectiles collide against objects. This section looks
at a few key tests between line segments and other
objects. For these tests, you want to know not only if
the line segment intersects but the first such point of
intersection.
This section relies heavily on the previously defined

parametric equation of a line segment:
The approach for most line segment intersection tests is

to first treat the segment as an infinite line—because if
the infinite line does not intersect with the object, then
there’s no way the line segment will. Once you solve for
the infinite line intersection, you then verify that t is
within the [0, 1] bounds for the line segment.
Line Segment Versus Plane Test

To find the point of intersection between a line
segment and a plane, you want to find whether there
exists a t such that L(t) is a point on the plane:
You can solve this with some algebraic manipulation.

First substitute for L(t):
Because the dot product is distributive over addition, you

can rewrite this as follows:
Finally, you solve for t:

Note that there’s a potential for division by zero if the dot
product in the denominator evaluates to zero. This will
happen only in the case where the line is perpendicular
to the normal of the plane, meaning the line is parallel to
the plane. In this case, the line and plane intersect only if
the line is entirely on the plane.
After you calculate the value of t, you then test whether

it’s within the bounds of the line segment, as in Listing
10.5. The Intersect function here returns the t value
by reference, and the caller can use this to figure out the
point of intersection if needed.
Listing 10.5 Line Segment Versus Plane Intersection

bool Intersect(const LineSegment& l, const Plane& p, float& outT)
// First test if there's a solution for t
float denom = Vector3::Dot(l.mEnd - l.mStart,
p.mNormal);
if (Math::NearZero(denom))
{
// The only way they intersect if start/end are
// points on the plane (P dot N) == d
if (Math::NearZero(Vector3::Dot(l.mStart, p.mNormal) - p.mD))
outT = 0.0f;
return true;
else
{ return false; }
else
float numer = -Vector3::Dot(l.mStart, p.mNormal) - p.mD;
outT = numer / denom;
// Validate t is within bounds of the line segment
if (outT >= 0.0f && outT <= 1.0f)
return true;
else
{
return false;
Line Segment Versus Sphere Test

To find the point of intersection between a line
segment and a sphere, you find if there’s a t value
such that the distance between the line and the
center of the sphere C is equal to the radius of the
sphere r:
To simplify this equation, you introduce substitutions:

To solve for t, you need some method to extract it from
inside the length operation. To do this, you square both
sides of the equation and replace the length squared
operation with a dot product:
Because the dot product is distributive over vector

addition, you can apply the FOIL (first, outside, inside,
last) distribution rule:
You then rewrite this in the quadratic form:

Finally, you apply the quadratic equation to solve for t:
The discriminant of the quadratic equation (the value

under the radical) tells you the number and types of
solutions to the equation. A negative discriminant means
the solutions are imaginary. For the purposes of a game,
you can assume that none of the objects have imaginary
positions. Thus, you know that a negative discriminant
means the line does not intersect with the sphere.
Otherwise, there can be one or two solutions to the
quadratic equation. A discriminant of zero means there’s
one solution because the line is tangential to the sphere.
A discriminant greater than zero means the line has two
points of intersection with the sphere. Figure 10.10
illustrates these three possibilities.
Figure 10.10 Possible discriminant values for line
versus sphere intersection
Once you have solutions for t, you then validate that t is

within the range [0, 1]. Because there are two possible
solutions, you give preference to the lower value of t,
which represents the first intersection. But if the line
segment begins inside the sphere and exits the sphere,
the higher value of t represents the point of intersection.
Listing 10.6 gives this code for line segment versus
sphere intersection. Note that this function returns
false if the sphere wholly contains the line segment.
Listing 10.6 Line Segment Versus Sphere Intersection

bool Intersect(const LineSegment& l, const Sphere& s, float& outT)
// Compute X, Y, a, b, c as per equations
Vector3 X = l.mStart - s.mCenter;
Vector3 Y = l.mEnd - l.mStart;
float a = Vector3::Dot(Y, Y);
float b = 2.0f * Vector3::Dot(X, Y);

float c = Vector3::Dot(X, X) - s.mRadius * s.mRadius;
// Compute discriminant
float disc = b * b - 4.0f * a * c;
if (disc < 0.0f)
return false;
else
disc = Math::Sqrt(disc);
// Compute min and max solutions of t
float tMin = (-b - disc) / (2.0f * a);
float tMax = (-b + disc) / (2.0f * a);
// Check whether either t is within bounds of segment
if (tMin >= 0.0f && tMin <= 1.0f)
outT = tMin;
return true;
else if (tMax >= 0.0f && tMax <= 1.0f)

{
outT = tMax;
return true;
else
return false;
Line Segment Versus AABB Test

One approach for line segment versus AABB
intersection is to construct planes for each edge of
the box. In 2D, this yields four planes for the four
different sides. Because the planes are infinite,
simply intersecting with a side plane does not mean
the segment intersects the box. In Figure 10.11(a),
the line segment intersects the top plane at P1 and
the left plane at P2. But because the box contains
neither of these points, these points do not intersect
the box. However, in Figure 10.11(b), the segment
intersects the left plane at P3. Because the box
contains P3, this is a point of intersection.
Figure 10.11 Intersection with the side planes but
not the box (a), intersection with the box (b), and
intersection with the box at two points (c)
Sometimes, the line segment may have multiple points of

intersection, as in Figure 10.11(c). Both P4 and P5
intersect the box. In this case, the intersection should
return the point closest to the start point, or the one at
the lowest t value in the parametric formulation of the
line segment.
For the segment tests versus each plane, recall that the
equation for line segment versus plane intersection is as
follows:
However, because each plane is parallel to a coordinate

axis (or, in 3D, a coordinate plane), you can optimize this
equation because the normal of each plane will always
have zeros for two components and a one for the third
component. Thus, two of the three dot product
components will always evaluate to zero.
For example, the normal of the left side plane points

directly to the left or right; the direction doesn’t matter
for the purposes of the intersection test. In 2D, this is as
follows:
Because the min point for the box is on its left plane, the
d value is as follows:
Similarly, the dot products in the segment versus plane

intersection equation also simplify to their x
components. This means the final equation solving for
intersection against the left plane is as follows:
The equations for the other side planes have similar

derivations. For 3D, there are a total of six planes to test.
Listing 10.7 shows a helper function that encapsulates
the behavior of testing a single side plane. Note that if
the segment intersects with the plane, the function adds
the t value to a supplied std::vector. The intersection
function uses this std::vector to test all possible t
values in order, from earliest to latest planar
intersection.
Listing 10.7 Line Segment Versus AABB Helper

Function
bool TestSidePlane(float start, float end, float negd,
std::vector<float>& out)
float denom = end - start;
if (Math::NearZero(denom))
return false;
else
float numer = -start + negd;
float t = numer / denom;
// Test that t is within bounds
if (t >= 0.0f && t <= 1.0f)
out.emplace_back(t);
return true;
else
{
return false;
The Intersect function, shown in Listing 10.8, uses

the TestSidePlane function to test the six different
side planes of the AABB against the line segment. Each
point of planar intersection has its t value stored in the
tValues vector. Then, you sort this vector in ascending
order and return the first intersection point contained by
the AABB. If the AABB contains none of these points, the
function returns false.
Listing 10.8 Line Segment Versus AABB Intersection

bool Intersect(const LineSegment& l, const AABB& b, float& outT)
// Vector to save all possible t values
std::vector<float> tValues;
// Test the x planes
TestSidePlane(l.mStart.x, l.mEnd.x, b.mMin.x, tValues);
TestSidePlane(l.mStart.x, l.mEnd.x, b.mMax.x, tValues);

// Test the y planes
TestSidePlane(l.mStart.y, l.mEnd.y, b.mMin.y, tValues);
TestSidePlane(l.mStart.y, l.mEnd.y, b.mMax.y, tValues);
// Test the z planes
TestSidePlane(l.mStart.z, l.mEnd.z, b.mMin.z, tValues);
TestSidePlane(l.mStart.z, l.mEnd.z, b.mMax.z, tValues);
// Sort the t values in ascending order
std::sort(tValues.begin(), tValues.end());
// Test if the box contains any of these points of intersection
Vector3 point;
for (float t : tValues)
point = l.PointOnSegment(t);
if (b.Contains(point))
outT = t;
return true;
//None of the intersections are within bounds of box
return false;
By testing each side of the box independently, you can

modify the code to return which side intersects the line
segment. This is useful if an object needs to bounce off
the box (such as a ball bouncing as in this chapter’s game
project). While not shown here, this requires associating
each call to TestSidePlane with a side of the box.
Then, you add that side (or normal to the side) as a
reference parameter that Intersect can write to.
You can optimize the segment versus AABB intersection

by instead using slabs, which are infinite areas bounded
by two planes. However, grasping this approach requires
additional mathematical backing. It’s one of the many
topics discussed in Christer Ericson’s book, listed in the
“Additional Reading” section at the end of this chapter.
Dynamic Objects
The intersection tests covered thus far are
instantaneous tests. In a game, this means that
you test whether two objects intersect on the current
frame. Although this might be sufficient for simple
games, in practice there are issues.
Consider the case where a character fires a bullet at piece
of paper. Suppose you use a bounding sphere for the
bullet and a box for the paper. On each frame, you test
whether the bullet intersects with the paper. Because the
bullet travels quickly, it’s unlikely that there’s one
specific frame where the bullet exactly intersects with the
paper. This means that instantaneous tests will miss the
intersection, as in Figure 10.12.
Figure 10.12 Instantaneous tests on frames 0 and 1

miss the collision between the bullet and paper
For the specific example of a bullet, you might solve this

by representing the bullet as a line segment. The start
point of the line segment is the position of the bullet in
the last frame, and the end is the position of the bullet
this frame. This way, you can detect whether the bullet
intersects with the paper at any point between the last
frame and this frame. However, this works only because
the bullet is very small. You can’t represent larger objects
with a line segment.
For some types of moving objects, such as two moving
spheres, you can solve directly for the time of
intersection. However, this doesn’t work well for cases
such as boxes that rotate between two frames. For other
types of moving objects, you might try to sample the
intersection at multiple points between frames. The term
continuous collision detection (CCD) can reference
either directly solving the point of intersection or
sampling the intersection.
To get a taste of how to solve directly for a time of

intersection, consider the case of intersection between
two moving spheres. This intersection, called swept-
sphere intersection, also happens to commonly come
up in video game companies’ interview questions.
For each sphere, you have the center positions during the
last frame and during this frame. You can represent
these positions using the same parametric equation as
for line segments, where the position last frame is t = 0
and the position this frame is t = 1. For sphere P, P0 is
the position last frame and P1 is the position this frame.
Similarly, the sphere Q has the positions Q0 and Q1. So,
these are the parametric equations for the positions of
spheres P and Q:
You want to solve for the value of t where the distance

between the two spheres is equal to the sum of their
radii:
You now proceed in a manner similar to that used to test
for line segment versus sphere intersection. You square
both sides and replace the length squared with a dot
product:
Then, you factor the terms and make substitutions:
Finally, you distribute the dot product over addition,

rewrite in the quadratic form, and solve the quadratic
equation:
As with line segment versus sphere, you use the
discriminant to determine whether any real solutions
exist. However, for swept-sphere intersection you only
care about the first point of intersection, which is the
lower of the two t values. As before, you must validate
that t is within the range [0, 1]. Listing 10.9 gives the
code for swept-sphere intersection. The function returns
t by reference, so the caller can use this to determine the
position of the spheres at the time of intersection.
Listing 10.9 Swept-Sphere Intersection

bool SweptSphere(const Sphere& P0, const Sphere& P1,
const Sphere& Q0, const Sphere& Q1, float& outT)
// Compute X, Y, a, b, and c
Vector3 X = P0.mCenter - Q0.mCenter;
Vector3 Y = P1.mCenter - P0.mCenter -

(Q1.mCenter - Q0.mCenter);
float a = Vector3::Dot(Y, Y);
float b = 2.0f * Vector3::Dot(X, Y);
float sumRadii = P0.mRadius + Q0.mRadius;
float c = Vector3::Dot(X, X) - sumRadii * sumRadii;
// Solve discriminant
float disc = b * b - 4.0f * a * c;
if (disc < 0.0f)
return false;
else
disc = Math::Sqrt(disc);
// We only care about the smaller solution
outT = (-b - disc) / (2.0f * a);
if (outT >= 0.0f && outT <= 0.0f)
return true;
}
else
return false;
ADDING COLLISIONS TO GAME

CODE
The preceding sections discuss geometry objects
used for collision and how to detect intersections
between these objects. This section explores how to
incorporate these techniques into game code. A new
BoxComponent class adds AABBs for actors, and a
PhysWorld class tracks boxes and detects
intersections as needed. The character movement
and new projectile firing code then leverage this new
collision functionality.
The BoxComponent Class

The BoxComponent class declaration is much like
the declaration of other components. However,
instead of overriding the Update function, it
overrides the OnUpdateWorldTransform function.
Recall that the owning actor calls
OnUpdateWorldTransform whenever it
recomputes the world transform.
The member data of BoxComponent has two instances

of the AABB struct: one AABB for the object space
bounds, and one AABB for the world space bounds. The
object space bounds shouldn’t change after initialization
of the BoxComponent, but the world space bounds
change whenever the owning actor’s world transform
changes. Finally, BoxComponent has a Boolean for
whether you want the BoxComponent to rotate based on
the world rotation. This way, you can choose whether or
not an actor’s BoxComponent rotates when the actor
rotates. Listing 10.10 shows the declaration of
BoxComponent.
Listing 10.10 BoxComponent Declaration

class BoxComponent : public Component
public:
BoxComponent(class Actor* owner);
~BoxComponent();
void OnUpdateWorldTransform() override;
void SetObjectBox(const AABB& model) { mObjectBox = model; }

const AABB& GetWorldBox() const { return mWorldBox; }
void SetShouldRotate(bool value) { mShouldRotate = value; }
private:
AABB mObjectBox;
AABB mWorldBox;
bool mShouldRotate;
};
To get the object space bounds of mesh files, the Mesh

class also adds an AABB in its member data. Then, when
loading in a gpmesh file, Mesh calls
AABB::UpdateMinMax on each vertex, ultimately
yielding an object space AABB. Then actors using a mesh
can grab the mesh’s object space bounds and pass these
bounds into the actor’s BoxComponent:
Mesh* mesh = GetGame()->GetRenderer()->GetMesh("Assets/Plane.gpmesh");
// Add collision box
BoxComponent* bc = new BoxComponent(this);
bc->SetObjectBox(mesh->GetBox());
To convert from object bounds to world bounds, you

need to apply scale, rotation, and translation to the
bounds. As when constructing the world transform
matrix, the order is important because rotation is about
the origin. Listing 10.11 gives the code for
OnUpdateWorldTransform. To scale the box, you
multiply min and max by the scale of the owning actor.
To rotate the box, you use the previously discussed
AABB::Rotate function, passing in the quaternion of
the owning actor. You do this rotation only if
mShouldRotate is true (which it is by default). To
translate the box, you add the position of the owning
actor to both min and max.
Listing 10.11
BoxComponent::OnUpdateWorldTransform
Implementation
void BoxComponent::OnUpdateWorldTransform()
// Reset to object space box
mWorldBox = mObjectBox;
// Scale
mWorldBox.mMin *= mOwner->GetScale();
mWorldBox.mMax *= mOwner->GetScale();
// Rotate
if (mShouldRotate)
{
mWorldBox.Rotate(mOwner->GetRotation());
// Translate
mWorldBox.mMin += mOwner->GetPosition();
mWorldBox.mMax += mOwner->GetPosition();
The PhysWorld Class

Much as you have separate Renderer and
AudioSystem classes, it’s sensible to create a
PhysWorld class for the physics world. You add a
PhysWorld pointer to Game and initialize it in
Game::Initialize.
PhysWorld has a vector of BoxComponent pointers and

corresponding public AddBox and RemoveBox
functions, as shown in the skeleton declaration in Listing
10.12. Then the constructor and destructor for
BoxComponent can call AddBox and RemoveBox,
respectively. This way, PhysWorld will have a vector of
all box components, much like how Renderer has a
vector of all sprite components.
Listing 10.12 Skeleton PhysWorld Declaration

class PhysWorld
public:
PhysWorld(class Game* game);
// Add/remove box components from world
void AddBox(class BoxComponent* box);
void RemoveBox(class BoxComponent* box);
// Other functions as needed
// ...
private:
class Game* mGame;
std::vector<class BoxComponent*> mBoxes;
};
Now that PhysWorld tracks all the box components in

the world, the next step is to add support for collision
tests against these boxes. You define a function called
SegmentCast that takes in a line segment and returns
true if the segment intersects with a box. In addition, it
returns by reference information about the first such
collision:
bool SegmentCast(const LineSegment& l, CollisionInfo& outColl);
The CollisionInfo struct contains the position of

intersection, the normal at the intersection, and pointers
to both the BoxComponent and Actor of the collision:
struct CollisionInfo
// Point of collision
Vector3 mPoint;
// Normal at collision
Vector3 mNormal;
// Component collided with
class BoxComponent* mBox;
// Owning actor of component
class Actor* mActor;
};
Because the segment potentially intersects with multiple

boxes, SegmentCast assumes that the closest
intersection is the most important one. Because the
vector of box components has no ordering,
SegmentCast can’t simply return after the first
intersection. Instead, the function needs to test against
all boxes and return the intersection result with the
lowest t value, as shown in Listing 10.13. This works
because the lowest t value intersection is the one closest
to the start of the line segment. SegmentCast uses the
segment versus AABB intersection function discussed
earlier but now modified to also return the normal of the
side the line segment intersects with.
Listing 10.13 PhysWorld::SegmentCast

Implementation
bool PhysWorld::SegmentCast(const LineSegment& l, CollisionInfo& outColl)
bool collided = false;
// Initialize closestT to infinity, so first
// intersection will always update closestT
float closestT = Math::Infinity;
Vector3 norm;
// Test against all boxes
for (auto box : mBoxes)
float t;
// Does the segment intersect with the box?
if (Intersect(l, box->GetWorldBox(), t, norm))
// Is this closer than previous intersection?
if (t < closestT)
outColl.mPoint = l.PointOnSegment(t);
outColl.mNormal = norm;
outColl.mBox = box;
outColl.mActor = box->GetOwner();
collided = true;
return collided;
Ball Collisions with SegmentCast

In the chapter game project, you use SegmentCast
to determine whether the ball projectile the player
fires hits something. If it does, you want the ball to
bounce off the normal of the surface. This means that
once the ball hits the surface, you must rotate it to
face an arbitrary direction.
You first add a helper function to Actor that uses the

dot product, cross product, and quaternions to adjust the
actor’s rotation to face the desired direction. Listing
10.14 shows the implementation of this helper function,
RotateToNewForward.
Listing 10.14 Actor::RotateToNewForward

Implementation
void Actor::RotateToNewForward(const Vector3& forward)
// Figure out difference between original (unit x) and new
float dot = Vector3::Dot(Vector3::UnitX, forward);
float angle = Math::Acos(dot);
// Are we facing down X?
if (dot > 0.9999f)
{ SetRotation(Quaternion::Identity); }
// Are we facing down -X?
else if (dot < -0.9999f)

{ SetRotation(Quaternion(Vector3::UnitZ, Math::Pi)); }
else
// Rotate about axis from cross product
Vector3 axis = Vector3::Cross(Vector3::UnitX, forward);
axis.Normalize();
SetRotation(Quaternion(axis, angle));
Next, you create a BallActor class and attach to it a

new MoveComponent subclass called BallMove, which
implements the movement code specific to BallActor.
The BallMove::Update function, shown in Listing
10.15, first constructs a line segment in the direction the
ball is traveling. If this segment intersects with anything
in the world, you want it to bounce off the surface. You
reflect the movement direction off the surface with
Vector3::Reflect and then use
RoateToNewForward to tell the ball to rotate to face
this new direction.
Listing 10.15 Using SegmentCast for Ball Movement

void BallMove::Update(float deltaTime)

{
// Construct segment in direction of travel
const float segmentLength = 30.0f;
Vector3 start = mOwner->GetPosition();
Vector3 dir = mOwner->GetForward();
Vector3 end = start + dir * segmentLength;
LineSegment ls(start, end);
// Test segment vs world
PhysWorld* phys = mOwner->GetGame()->GetPhysWorld();
PhysWorld::CollisionInfo info;
if (phys->SegmentCast(ls, info))
// If we collided, reflect the direction about the normal
dir = Vector3::Reflect(dir, info.mNormal);
mOwner->RotateToNewForward(dir);
// Base class update moves based on forward speed
MoveComponent::Update(deltaTime);
}
One thing to watch out for is what happens when you

add a BoxComponent to the player, as you will do later
in this section. You clearly don’t want the ball to collide
against the player because the player fires the ball!
Luckily, you can leverage the fact that the
CollisionInfo from SegmentCast has a pointer to
the actor owning the box component. Thus, if you save a
pointer to the player somewhere, you can make sure a
ball doesn’t collide against the player.
Testing Box Collisions in PhysWorld

Although not used in this chapter’s game project,
some game features may require testing all the boxes
in the physics world against each other. A naïve
implementation is to perform pairwise intersection
tests between all combinations of boxes in the world.
This basic approach, shown in Listing 10.16, uses an
O(n2) algorithm because it tests every box against
every other box. The TestPairwise function takes
in a user-supplied function, f, and calls f on every
intersection between boxes.
Listing 10.16 PhysWorld::TestPairwise

Implementation
void PhysWorld::TestPairwise(std::function<void(Actor*, Actor*)> f)

{
// Naive implementation O(n^2)
for (size_t i = 0; i < mBoxes.size(); i++)
// Don't need to test vs. itself and any previous i values
for (size_t j = i + 1; j < mBoxes.size(); j++)
BoxComponent* a = mBoxes[i];
BoxComponent* b = mBoxes[j];
if (Intersect(a->GetWorldBox(), b->GetWorldBox()))
// Call supplied function to handle intersection
f(a->GetOwner(), b->GetOwner());
Although TestPairwise is conceptually simple, it ends

up making a lot of unnecessary calls to Intersect. It
treats two boxes on opposite sides of the world the same
as two boxes right next to each other. In the case of this
chapter’s game project, there are 144 boxes.
TestPairwise makes over 10,000 calls to the
Intersect function given these 144 boxes.
You can optimize this by observing that two 2D axis-

aligned boxes do not intersect unless they overlap on
both coordinate axes. For example, the interval [min.x,
max.x] from one box must overlap with the [min.x,
max.x] interval from another box if the two boxes are to
intersect. The sweep-and-prune method takes
advantage of this observation to reduce the number of
box intersection tests. The sweep-and-prune method
involves selecting an axis and testing only boxes that
have overlapping intervals along that axis.
Figure 10.13 illustrates a handful of AABBs and considers

their intervals along the x-axis. Box A’s and Box B’s x
intervals overlap, so they may potentially intersect.
Similarly, Box B’s and Box C’s x intervals overlap, so they
may intersect. However, Box A’s and Box C’s x intervals
do not overlap, so they cannot intersect. Similarly, Box
D’s x interval overlaps with none of the other boxes, so it
can’t intersect with any of them. In this case, the sweep-
and-prune algorithm calls Intersect only on the pairs
(A, B) and (B, C) instead of on all six possible
combinations.
Figure 10.13 AABB intervals along the x-axis
Listing 10.17 gives the code for the sweep-and-prune

method along the x-axis. You first sort the vector of
boxes by their minimum x-value. Then, for every box,
you grab the maximum x value and save it in max. In the
inner loop, you only consider boxes whose min.x is less
than max. Once the inner loop hits the first box with a
min.x greater than max, there are no other boxes that
overlap along the x-axis with the outer loop box. This
means there are no other possible intersections against
the outer loop box, so you break to the next iteration of
the outer loop.
Listing 10.17 PhysWorld::TestSweepAndPrune

Implementation
void PhysWorld::TestSweepAndPrune(std::function<void(Actor*, Actor*)> f)

{
// Sort by min.x
std::sort(mBoxes.begin(), mBoxes.end(),
[](BoxComponent* a, BoxComponent* b) {
return a->GetWorldBox().mMin.x <
b->GetWorldBox().mMin.x;
});
for (size_t i = 0; i < mBoxes.size(); i++)
// Get max.x for box[i]
BoxComponent* a = mBoxes[i];
float max = a->GetWorldBox().mMax.x;
for (size_t j = i + 1; j < mBoxes.size(); j++)
BoxComponent* b = mBoxes[j];
// If box[j] min.x is past the max.x bounds of box[i],
// then there aren't any other possible intersections
// against box[i]
if (b->GetWorldBox().mMin.x > max)
{
break;
else if (Intersect(a->GetWorldBox(), b->GetWorldBox()))
f(a->GetOwner(), b->GetOwner());
In this chapter’s game project, TestSweepAndPrune

cuts down the number of calls to Intersect roughly by
half in comparison to TestPairwise. The complexity of
this algorithm is on average O(n log n). Even though
sweep-and-prune requires a sort, it is generally more
efficient than the naïve pairwise test—unless there are
very few boxes. Some implementations of the sweep-
and-prune algorithm do the pruning along all three axes,
as in Exercise 10.2. This requires maintaining multiple
sorted vectors. The advantage of testing all three axes is
that after you prune for all three axes, the remaining set
of boxes must intersect with each other.
Sweep-and-prune is one of a category of techniques

called broadphase techniques. The broadphase tries to
eliminate as many collisions as possible before the
narrowphase, which tests the individual pairs of
collisions. Other techniques use grids, cells, or trees.
Player Collision Against the Walls

Recall that MoveComponent uses the
mForwardSpeed variable to move the character
forward or backward. However, the current
implementation allows the player to move through
walls. To fix this, you can add a BoxComponent to
each wall (encapsulated by PlaneActor), as well as
a BoxComponent to the player. Because you only
want to test the player against every PlaneActor,
you don’t use TestSweepAndPrune. Instead, you
can just make a vector of PlaneActor pointers in
Game and access this from the player’s code.
The basic idea is that every frame, you test the player’s
collision against every PlaneActor. If the AABBs
intersect, you adjust the player’s position so that it no
longer collides with the wall. To understand this
calculation, it helps to visualize the problem in 2D.
Figure 10.14 illustrates a player’s AABB colliding with a

platform AABB. You calculate two differences per axis.
For example, dx1 is the difference between the player’s
max.x and the platform’s min.x. Conversely, dx2 is the
difference between the player’s min.x and the platform’s
max.x. Whichever of these differences has the lowest
absolute value is the minimum overlap between the
two AABBs. In Figure 10.14, the minimum overlap is dy1.
If you then add dy1 to the player’s y position, the player
stands exactly on top of the platform. Thus, to correctly
fix the collision, you can just adjust the position in the
axis of the minimum overlap.
Figure 10.14 Calculating minimum overlap in 2D
In 3D, the principle is the same, except there are now six
difference values because there are three axes. The
FPSActor::FixCollisions function as shown in
Listing 10.18 implements this minimum overlap test.
Importantly, because changing the position of the player
changes the player’s BoxComponent, in between each
intersect we must recompute the world bounds of the
BoxComponent. You then call this function from
UpdateActor, which means you call it after the
MoveComponent updates the player’s position every
frame.
Listing 10.18 FPSActor::FixCollisions

Implementation
void FPSActor::FixCollisions()
// Need to recompute my world transform to update world box
const AABB& playerBox = mBoxComp->GetWorldBox();
auto& planes = GetGame()->GetPlanes();
for (auto pa : planes)
// Do we collide with this PlaneActor?
const AABB& planeBox = pa->GetBox()->GetWorldBox();
if (Intersect(playerBox, planeBox))
// Calculate all our differences

float dx1 = planeBox.mMin.x - playerBox.mMax.x;
float dx2 = planeBox.mMax.x - playerBox.mMin.x;
float dy1 = planeBox.mMin.y - playerBox.mMax.y;
float dy2 = planeBox.mMax.y - playerBox.mMin.y;
float dz1 = planeBox.mMin.z - playerBox.mMax.z;
float dz2 = planeBox.mMax.z - playerBox.mMin.z;
// Set dx to whichever of dx1/dx2 have a lower abs
float dx = (Math::Abs(dx1) < Math::Abs(dx2)) ? dx1 : dx2;
// Ditto for dy
float dy = (Math::Abs(dy1) < Math::Abs(dy2)) ? dy1 : dy2;
// Ditto for dz
float dz = (Math::Abs(dz1) < Math::Abs(dz2)) ? dz1 : dz2;
// Whichever is closest, adjust x/y position
if (Math::Abs(dx) <= Math::Abs(dy) &&
Math::Abs(dx) <= Math::Abs(dz))
pos.x += dx;
}
else if (Math::Abs(dy) <= Math::Abs(dx) &&
Math::Abs(dy) <= Math::Abs(dz))
pos.y += dy;
else
pos.z += dz;
// Need to set position and update box component
SetPosition(pos);
mBoxComp->OnUpdateWorldTransform();
Because you also use PlaneActor instances for the

ground beneath the player, you can also leverage this
code, with modification, to test whether the player lands
on the platforms. In Exercise 10.1, you’ll explore adding
jumping to the player.
GAME PROJECT
This chapter’s game project implements all the
different types of intersections discussed in this
chapter, as well as BoxComponent and PhysWorld.
It also uses SegmentCast for the ball projectile and
implements fixing the player colliding against the
walls. The result is a first-person shooting gallery, as
shown in Figure 10.15. The code is available in the
directory. Open Chapter10-windows.sln in
Mac.
The controls for this game project use the FPS-style

controls implemented in Chapter 9. Recall that W/S move
forward and back, A/D strafe, and the mouse rotates the
character. In addition, clicking the left mouse button
now fires a ball projectile in the direction of the vector
derived from an unprojection (also discussed in Chapter
9). The ball projectile uses SegmentCast to test whether
it intersects with a wall or target. In either case, the ball
reflects its facing direction based on the normal of the
surface, causing it to bounce. If the ball hits a target, the
game plays a ding sound.
SUMMARY
This chapter provides an in-depth introduction to
collision detection techniques in games. Games
might use many different geometric types for
collision detection. A line segment has a start point
and an end point. The representation of a plane is its
normal and distance to the origin. Spheres are
simple bounding volumes but may cause many false
negatives for characters of different shapes. Axis-
aligned bounding boxes have sides aligned with the
axes, while oriented bounding boxes do not have this
restriction.
For intersection tests, this chapter covers many different

types of intersections. Contains point tests can say
whether a shape contains a point. You can also test
whether two bounding volumes (such as two AABBs)
intersect with each other. This chapter also covers tests
for whether a line segment intersects with objects
including planes, spheres, and other boxes. For moving
objects, you may need to use a form of continuous
collision detection to ensure that the game doesn’t miss
collisions that occur between frames.
Finally, this chapter covers how to integrate collision

detection into game code. The BoxComponent class has
both object space bounds (derived from the mesh) and
world space bounds that are updated based on the
owning actor. PhysWorld keeps track of all box
components in the world, and SegmentCast tests a line
segment against all the boxes. For collisions between
pairs of boxes, it’s more efficient to use the sweep-and-
prune broadphase algorithm. Sweep-and-prune exploits
the fact that two boxes cannot intersect if their intervals
along a coordinate axis do not overlap. This chapter
shows how to use both segment cast and box component
collisions to implement some game-specific features,
such as the ball bouncing off objects or the player
colliding with the wall.
ADDITIONAL READING
Christer Ericson provides extremely detailed
coverage of collision detection, covering both the
mathematical bases of the algorithms and usable
implementations. Ian Millington doesn’t have as
much coverage of collision detection algorithms but
explains how to incorporate collision in the context
of physics engine movement, which is something this
chapter does not discuss in detail.
Ericson, Christer. Real-time Collision
Detection. San Francisco: Morgan
Kaufmann, 2005.
Millington, Ian. Game Physics Engine
Development, 2nd edition. Boca Raton:
CRC Press, 2010.
EXERCISES
In this chapter’s first exercise, you add jumping to
the chapter’s game project. In the second exercise,
you improve upon the sweep-and-prune
implementation covered in this chapter. In the last
exercise you implement OBB versus OBB
intersection between oriented bounding boxes.
Exercise 10.1
Add jumping to the player character. The ground
objects already have corresponding axis-aligned
bounding boxes. To implement jumping, select a key
(such as the spacebar). When the player presses the
jump key, set an additional velocity in the positive z
direction. Similarly, add a negative z acceleration for
gravity that slows down the jump velocity. After the
player hits the apex of the jump, he or she starts
falling. While they player is falling, you can detect in
FixCollisions whether the player lands on top of
a PlaneActor (because you know that the top is
dz2). While the player is on the ground, disable
gravity and set the z velocity back to zero.
To help keep the code modular, it’s recommended that

you use a simple state machine to represent the different
states of the character: on the ground, jumping, and
falling. As an additional feature, experiment with
transitioning from the “on ground” state to the “falling”
state. While in the “on ground” state, keep making
downward SegmentCasts to detect whether the player
has walked off the platform. If that happens, switch from
“on ground” to “falling.”
Exercise 10.2
Change the SweepAndPrune function to sweep and
prune across all three coordinate axes. Have
PhysWorld maintain three vectors of boxes and
change it so that AddBox and RemoveBox touch all
three vectors. Then sort each vector by its
corresponding axis.
The sweep-and-prune code should then test along each

axis independently and create a map of pairs of
overlapping boxes along that axis. Once all three axes are
complete, the code should compare the overlapping box
vectors. The boxes that overlap along all three axes are
the only boxes that intersect with each other.
Exercise 10.3
Implement OBB versus OBB intersection in a new
Intersect function. As with AABBs, use the
separating axis approach (that is, figure out whether
they cannot intersect and then logically not the
result). However, whereas there are 3 axes to test for
AABBs, for OBBs there are a total of 15 different axes
to test.
To implement this, first compute the 8 different corners

of both OBBs. Each OBB has a total of 3 local axes
corresponding to each side of the box. You can compute
these by using vector subtraction between the correct set
of points and normalizing the vectors. Because each box
has three local axes, that yields the first 6 potential
separating axes. The other 9 vectors are the
combinations of cross products between the two OBBs’
local axes. For example, OBB A’s up vector crosses OBB
B’s up, right, and forward vectors.
To determine the interval of the box along an axis,

compute the dot product of each corner of that box
versus the axis. The lowest dot product result is the
minimum value of the interval and, similarly, the highest
dot product result is the maximum value. Then
determine whether the [min, max] intervals of both
boxes separate along the axis. If any of the 15 axes
separate, then the boxes cannot intersect. Otherwise,
they must intersect.
CHAPTER 11
USER INTERFACES
Most games include UI elements such as a

menu system and an in-game heads-up
display (HUD). The menu system allows the
player to perform actions such as starting
and pausing the game. The HUD includes
elements that give information to the player
during gameplay. This can include elements
such as an aiming reticule or a radar.
This chapter looks at core systems needed to
implement user interfaces, including text
rendering with a font, a system for UI
screens, and localization for different
languages. The chapter also explores
implementations of certain HUD elements.
FONT RENDERING
In the TrueType font format, straight line segments
and Bézier curves form the outlines of individual
characters (or glyphs). The SDL TTF library
provides support for loading and rendering TrueType
fonts. After initializing the library, the basic process
is to load fonts at specific point sizes. Then SDL TTF
takes in a string and renders the string to a texture,
using the glyphs from the font. Once this texture
exists, the game can render the texture just like any
other 2D sprite.
Much as in other systems, the Game class initializes SDL

TTF in Game::Initialize. The TTF_Init function
returns 0 if successful, and -1 in the event of an error.
Similarly, Game::Shutdown calls TTF_Quit to shut
down the library.
Next, you declare a Font class to encapsulate any font-

specific functionality, as shown in Listing 11.1. The Load
function loads in a font from the specified file, and
Unload frees all font data. The RenderText function
takes in the provided string, color, and point size and
creates a texture that contains the text.
Listing 11.1 Font Declaration

class Font
public:
Font();
~Font();
// Load/unload from a file
void Unload();
// Given string and this font, draw to a texture
class Texture* RenderText(const std::string& text,
const Vector3& color = Color::White,
int pointSize = 30);
private:
// Map of point sizes to font data
std::unordered_map<int, TTF_Font*> mFontData;
};
The TTF_OpenFont function loads a font from a .ttf

file at a specific point size and returns a pointer to the
TTF_Font data corresponding to the font at that size.
This means to support different-sized text in the game,
you must call TTF_OpenFont multiple times. The
Font::Load function, shown in Listing 11.2, first
creates a vector of the desired point sizes, and then it
loops over this vector, calling TTF_OpenFont once per
size and adding each TTF_Font to the mFontData map.
Listing 11.2 Font::Load Implementation

bool Font::Load(const std::string& fileName)
// Support these font sizes
std::vector<int> fontSizes = {
8, 9, 10, 11, 12, 14, 16, 18, 20, 22, 24, 26, 28,
30, 32, 34, 36, 38, 40, 42, 44, 46, 48, 52, 56,
60, 64, 68, 72
};
// Call TTF_OpenFont once per every font size
for (auto& size : fontSizes)
TTF_Font* font = TTF_OpenFont(fileName.c_str(), size);
if (font == nullptr)
SDL_Log("Failed to load font %s in size %d", fileName.c_str(),

size);
return false;
mFontData.emplace(size, font);
return true;
As with other resources, you want to keep track of loaded

fonts in a central place. In this case, the Game class adds
a map where the key is the filename of the font, and the
value is a Font pointer. You then add a corresponding
GetFont function. As with GetTexture and similar
functions, GetFont first tries to find the data in its map
and, if it fails, it loads the font file and adds it to the map.
The Font::RenderText function, shown in Listing

11.3, creates a texture given a text string, using a font of
the appropriate size. First, you convert the Vector3
color into an SDL_Color, where each component ranges
from 0 to 255. Next, you look in the mFontData map to
find the TTF_Font data corresponding to the font at the
requested point size.
Next, you call the TTF_RenderText_Blended

function, which takes in a TTF_Font*, the string of text
to render, and a color. The Blended suffix means that
the font will draw with alpha transparency around the
glyphs. Unfortunately, TTF_RenderText_Blended
returns a pointer to an SDL_Surface. OpenGL cannot
directly draw an SDL_Surface.
Recall that in Chapter 5, “OpenGL,” you created the

Texture class to encapsulate a texture loaded for
OpenGL usage. You can add a
Texture::CreateFromSurface function to convert
an SDL_Surface into a Texture. (This chapter omits
the implementation of CreateFromSurface, but check
the source code for the gameproject.) Once the
SDL_Surface is converted into a Texture object, you
can free the surface.
Listing 11.3 Font::RenderText Implementation

Texture* Font::RenderText(const std::string& text,
const Vector3& color, int pointSize)
Texture* texture = nullptr;
// Convert to SDL_Color
SDL_Color sdlColor;
sdlColor.r = static_cast<Uint8>(color.x * 255);
sdlColor.g = static_cast<Uint8>(color.y * 255);

sdlColor.b = static_cast<Uint8>(color.z * 255);
sdlColor.a = 255;
// Find the font data for this point size
auto iter = mFontData.find(pointSize);
if (iter != mFontData.end())
TTF_Font* font = iter->second;
// Draw this to a surface (blended for alpha)
SDL_Surface* surf = TTF_RenderText_Blended(font, text.c_str(),
sdlColor);
if (surf != nullptr)
// Convert from surface to texture
texture = new Texture();
texture->CreateFromSurface(surf);
SDL_FreeSurface(surf);
else
{
SDL_Log("Point size %d is unsupported", pointSize);
return texture;
Because creating a texture is somewhat expensive, the UI

code does not call RenderText every frame. Instead, it
calls RenderText only when the text string changes and
saves the resulting texture. Then on every frame, the UI
code can draw the texture that contains the rendered
text. For maximum efficiency, you could even render
each letter in the alphabet to separate textures and then
stich these letter textures together to form words.
UI SCREENS
Because a UI system might be used for many things,
including the HUD and menus, flexibility is an
important feature. Although there are data-driven
systems that utilize tools such as Adobe Flash, this
chapter instead focuses on a code-driven
implementation. However, many of the ideas
presented here can still apply to a more data-driven
system.
It’s useful to think of the UI as containing different

layers. For example, during gameplay, the heads-up
display (HUD) shows information relevant to the
player, such as health or a score. If the player pauses the
game, the game might show a menu that lets the player
choose between different options. While the game shows
the pause menu, you may still want the HUD elements to
be visible under the pause menu.
Now suppose that one of the options in the pause menu

is to quit the game. When the player selects this option,
you might want to have the game show a confirmation
dialog box that asks whether the player truly wants to
quit. The player might still see parts of both the HUD
and pause menu onscreen under this dialog box.
During this sequence, the player can typically interact

with only the topmost layer of the UI. This naturally
leads to the idea of using a stack to represent the
different layers of the UI. You can implement the idea of
a single UI layer with the UIScreen class. Each type of
UI screen, such as the pause menu or HUD, is a subclass
of UIScreen. After drawing the game world, the game
draws all UI screens on the stack in a bottom-up order.
At any point in time, only the UIScreen on top of the UI
stack might receive input events.
Listing 11.4 shows the first iteration of the base

UIScreen class. Notice that it has several virtual
functions that subclasses can override: Update for
updating the UI screen’s state, Draw for drawing it, and
the two input functions to handle different types of
incoming input. You can also keep track of the state of a
specific UI screen; in this case, you only need two states
for whether the screen is active or closing.
The UI screen also might have a title, so the member

data contains a pointer to the Font, a pointer to the
Texture that contains the rendered title, and a position
for the title onscreen. Subclasses can then call the
SetTitle function, which uses Font::RenderText to
set the mTitle member.
Finally, because a UIScreen is not an Actor, you

cannot attach any types of components to it. Thus, the
UIScreen class doesn’t use the drawing functionality
from SpriteComponent. Instead, you need a new
helper function called DrawTexture that draws a
texture at the specified position onscreen. Every UI
screen can then call DrawTexture as needed.
Listing 11.4 Initial UIScreen Declaration

class UIScreen
public:
UIScreen(class Game* game);
virtual ~UIScreen();
// UIScreen subclasses can override these

virtual void Draw(class Shader* shader);
virtual void ProcessInput(const uint8_t* keys);
virtual void HandleKeyPress(int key);
// Tracks if the UI is active or closing
enum UIState { EActive, EClosing };
// Set state to closing
void Close();
// Get state of UI screen
UIState GetState() const { return mState; }
// Change the title text
void SetTitle(const std::string& text,
const Vector3& color = Color::White,
int pointSize = 40);
protected:
// Helper to draw a texture
void DrawTexture(class Shader* shader, class Texture* texture,
const Vector2& offset = Vector2::Zero,
float scale = 1.0f);
class Game* mGame;
// For the UI screen's title text

class Font* mFont;
class Texture* mTitle;
Vector2 mTitlePos;
// State
UIState mState;
};
The UI Screen Stack

Adding the UI screen stack to the game requires
connections in several places. First, you add a
std::vector of UIScreen pointers to the Game
class for the UI stack. You don’t just use
std::stack here because you need to iterate over
the entire UI stack, which is not possible in
std::stack. You also add functions to push a new
UIScreen onto the stack (PushUI) and a function
to get the entire stack by reference:
// UI stack for game
std::vector<class UIScreen*> mUIStack;
// Returns entire stack by reference
const std::vector<class UIScreen*>& GetUIStack();
// Push specified UIScreen onto stack

void PushUI(class UIScreen* screen);
Then the constructor of UIScreen calls PushUI and

passes in its this pointer as the screen. This means that
simply dynamically allocating a UIScreen (or subclass
of UIScreen) automatically adds the UIScreen to the
stack.
Updating the UI screens on the stack happens in

UpdateGame, after updating all the actors in the world.
This requires looping over the entire UI screen stack and
calling Update on any active screens:
for (auto ui : mUIStack)
if (ui->GetState() == UIScreen::EActive)
ui->Update(deltaTime);
After updating all the UI screens, you also delete any

screens whose state is EClosing.
Drawing the UI screens must happen in Renderer.

Recall that Renderer::Draw first draws all the 3D
mesh components by using the mesh shader and then
draws all the sprite components by using the sprite
shader. Because the UI comprises several textures, it’s
natural to draw them by using the same shader that
sprites use. So, after drawing all sprite components, the
Renderer gets the UI stack from the Game object and
calls Draw on each UIScreen:
for (auto ui : mGame->GetUIStack())
ui->Draw(mSpriteShader);
For testing purposes, you can create a subclass of

UIScreen called HUD. You can create an instance of HUD
in Game::LoadData and save it in an mHUD member
variable:
mHUD = new HUD(this);
Because the constructor of HUD calls the constructor of

UIScreen, this automatically adds the object to the
game’s UI stack. For now, HUD doesn’t draw any
elements to the screen or otherwise override any
behavior of UIScreen. (Later in this chapter, you’ll
learn about supporting different functionality in the
HUD.)
Handling input for the UI stack is a bit trickier. In most

cases, a specific input action such as clicking the mouse
should affect the game or the UI—but not both
simultaneously. Thus, you first need a way to decide
whether to route the input to the game or to the UI.
To implement this, you first add the mGameState

variable to Game that supports three different states:
gameplay, paused, and quit. In the gameplay state, all
input actions route to the game world, which means you
pass the input to each actor. On the other hand, in the
paused state, all input actions go to the UI screen at the
top of the UI stack. This means that
Game::ProcessInput must call ProcessInput on
either each actor or the UI screen, depending on this
state:
if (mGameState == EGameplay)
if (actor->GetState() == Actor::EActive)
actor->ProcessInput(state);
}
else if (!mUIStack.empty())
mUIStack.back()->ProcessInput(state);
You can also extend this behavior so that the UI screen

on top of the stack can decide whether it wants to
process the input. If the UI screen decides it doesn’t want
to process the input, it can forward the input to the next
topmost UI on the stack.
Similarly, when responding to SDL_KEYDOWN and

SDL_MOUSEBUTTON events, you either send the event to
the game world or to the UI screen at the top of the stack
(via the HandleKeyPress function).
Because you added mGameState to track the state of the

game, you also make changes to the game loop. The
condition of the game loop changes to keep looping as
long as the game is not in the EQuit state. You further
update the game loop such that you call Update on all
the actors in the world only if the game state is
EGameplay. This way, the game doesn’t continue to
update the objects in the game world while in the paused
state.
The Pause Menu

Once the game has support for a paused state, you
can add a pause menu. First, you declare
PauseMenu as a subclass of UIScreen. The
constructor of PauseMenu sets the game state to
paused and sets the title text of the UI screen:
PauseMenu::PauseMenu(Game* game)
:UIScreen(game)
mGame->SetState(Game::EPaused);
SetTitle("PAUSED");
The destructor sets the game’s state back to gameplay:

PauseMenu::~PauseMenu()
mGame->SetState(Game::EGameplay);
}
Finally, the HandleKeyPress function closes the pause
menu if the player presses the Escape key:
void PauseMenu::HandleKeyPress(int key)
UIScreen::HandleKeyPress(key);
if (key == SDLK_ESCAPE)
Close();
This leads to the game deleting the PauseMenu instance,

which calls the PauseMenu destructor, which sets the
game state back to gameplay.
To show the pause menu, you need to construct a new

PauseMenu object because the constructor
automatically adds the UIScreen to the stack. You
create the pause menu in Game::HandleKeyPress to
have it appear when the player presses the Escape key.
The overall flow is that while in the gameplay state, the

player can press Escape to see the pause menu.
Constructing the pause menu object causes the game to
enter the pause state, which means actors don’t update.
Then, if the player presses Escape when in the pause
menu, you delete the pause menu and return to the
gameplay state. Figure 11.1 shows the game paused with
this simple version of the pause menu (which isn’t really
a menu yet because it has no buttons).
Figure 11.1 Game showing the basic pause menu
Buttons
Most menus in games also have buttons that the
player can interact with. For example, a pause menu
might have buttons for resuming the game, quitting
the game, configuring options, and so on. Because
different UI screens may need buttons, it makes
sense to add this support to the base UIScreen
class.
To encapsulate buttons, you can declare a Button class,

as shown in Listing 11.5. You can assume that every
button has a text name, and so it also needs a pointer to
the Font that renders this text. In addition, the button
has a position onscreen as well as dimensions (width and
height). Finally, when the player clicks the button, some
action should occur, depending on the button clicked.
To customize a button’s action, Button uses the

std::function class to encapsulate a callback
function. This function can be a standalone function or,
more likely, a lambda expression. When declaring a
Button, the constructor takes in this function. Then,
when the code detects a button click, you can call this
function. This way, any arbitrary button you create for
any arbitrary menu can call a corresponding arbitrary
function.
Listing 11.5 Button Declaration

class Button
public:
// Constructor takes in a name, font,
// callback function, and position/dimensions of button
Button(const std::string& name, class Font* font,

std::function<void()> onClick,
const Vector2& pos, const Vector2& dims);
~Button();
// Set the name of the button, and generate name texture
void SetName(const std::string& name);
// Returns true if the point is within the button's bounds
bool ContainsPoint(const Vector2& pt) const;
// Called when button is clicked
void OnClick();
// Getters/setters
// ...
private:
std::function<void()> mOnClick;
std::string mName;
class Texture* mNameTex;
class Font* mFont;
Vector2 mPosition;
Vector2 mDimensions;
bool mHighlighted;
};
Button has a ContainsPoint function that, given a

point, returns true if the point is within the 2D bounds
of the button. This function uses the same approach as in
Chapter 10, “Collision Detection”: You test for the four
cases where the point is not within the bounds. If none of
these four cases are true, then the button must contain
the point:
bool Button::ContainsPoint(const Vector2& pt) const
bool no = pt.x < (mPosition.x - mDimensions.x / 2.0f) ||
pt.x > (mPosition.x + mDimensions.x / 2.0f) ||
pt.y < (mPosition.y - mDimensions.y / 2.0f) ||
pt.y > (mPosition.y + mDimensions.y / 2.0f);
return !no;
The Button::SetName function uses the previously

discussed RenderText function to create the texture for
the button’s name, stored in mNameTex. The OnClick
function simply calls the mOnClick handler, if it exists:
void Button::OnClick()
if (mOnClick)
mOnClick();
You then add additional member variables to UIScreen

to support buttons: a vector of Button pointers and two
textures for the buttons. One texture is for unselected
buttons, and the other is for selected buttons. Having
different textures makes it easier for the player to
differentiate between the selected and unselected
buttons.
Next, you add a helper function to make it easy to create

new buttons:
void UIScreen::AddButton(const std::string& name,
std::function<void()> onClick)
Vector2 dims(static_cast<float>(mButtonOn->GetWidth()),
static_cast<float>(mButtonOn->GetHeight()));
Button* b = new Button(name, mFont, onClick, mNextButtonPos, dims);
mButtons.emplace_back(b);
// Update position of next button
// Move down by height of button plus padding
mNextButtonPos.y -= mButtonOff->GetHeight() + 20.0f;
The mNextButtonPos variable allows the UIScreen to

control where the buttons draw. You could certainly add
more customization with more parameters, but using the
provided code is a simple way to get a vertical list of
buttons.
Next, you add code in UIScreen::DrawScreen to

draw the buttons. For each button, you first draw the
button texture (which is either mButtonOn or
mButtonOff, depending on whether the button is
selected). Next, you draw the text for the button:
for (auto b : mButtons)
// Draw background of button
Texture* tex = b->GetHighlighted() ? mButtonOn : mButtonOff;
DrawTexture(shader, tex, b->GetPosition());

// Draw text of button
DrawTexture(shader, b->GetNameTex(), b->GetPosition());
You also want the player to use the mouse to select and
click on buttons. Recall that the game uses a relative
mouse mode so that that mouse movement turns the
camera. To allow the player to highlight and click on
buttons, you need to disable this relative mouse mode.
You can leave responsibility for this to the PauseMenu
class; in the constructor, it disables relative mouse mode,
and then it reenables it in the destructor. This way, when
the player returns to gameplay, the mouse can once
again rotate the camera.
The UIScreen::ProcessInput function, shown in

Listing 11.6, handles highlighting buttons with the
mouse. You first get the position of the mouse and
convert it to the simple screen space coordinates where
the center of the screen is (0, 0). You get the width and
height of the screen from the renderer. You then loop
over all the buttons in the mButtons vector and use the
ContainsPoint function to determine whether the
mouse cursor is within the bounds of the button. If the
button contains the mouse cursor, its state is set to
highlighted.
Listing 11.6 UIScreen::ProcessInput

Implementation
void UIScreen::ProcessInput(const uint8_t* keys)
// Are there buttons?
if (!mButtons.empty())
// Get position of mouse
int x, y;
SDL_GetMouseState(&x, &y);
// Convert to (0,0) center coordinates (assume 1024x768)
Vector2 mousePos(static_cast<float>(x), static_cast<float>(y));
mousePos.x -= mGame->GetRenderer()->GetScreenWidth() * 0.5f;
mousePos.y = mGame->GetRenderer()->GetScreenHeight() * 0.5f
- mousePos.y;
// Highlight any buttons
for (auto b : mButtons)
if (b->ContainsPoint(mousePos))
{
b->SetHighlighted(true);
else
b->SetHighlighted(false);
The mouse clicks are routed through

UIScreen::HandleKeyPress. Because
ProcessInput already determines which buttons are
highlighted by the mouse, HandleKeyPress just calls
the OnClick function on any highlighted buttons.
Using all this code, you can add buttons to PauseMenu.

For now, you can add two buttons—one to resume the
game and one to quit the game:
AddButton("Resume", [this]() {
Close();
});
AddButton("Quit", [this]() {
mGame->SetState(Game::EQuit);
});
The lambda expressions passed to AddButton define

what happens when the player clicks the button. When
the player clicks Resume, the pause menu closes, and
when the player clicks Quit, the game ends. Both lambda
expressions capture the this pointer so that they can
access members of PauseMenu. Figure 11.2 shows the
pause menu with these buttons.
Figure 11.2 Pause menu with buttons
Dialog Boxes
For certain menu actions, such as quitting the game,
it’s preferable to show the player a confirmation
dialog box. This way, if the player clicks on the first
button by mistake, he or she can still correct the
mistake. Using a UI screen stack makes it easy to
transfer control from one UI screen (such as the
pause menu) to a dialog box. In fact, you can
implement the dialog box with all the existing
UIScreen functionality. To do this, you can make a
new subclass of UIScreen called DialogBox.
The DialogBox constructor takes in a string for the text

shown, as well as a function to execute when the user
clicks OK:
DialogBox::DialogBox(Game* game, const std::string& text,
std::function<void()> onOK)
:UIScreen(game)
// Adjust positions for dialog box
mBGPos = Vector2(0.0f, 0.0f);
mTitlePos = Vector2(0.0f, 100.0f);
mNextButtonPos = Vector2(0.0f, 0.0f);
// Set background texture
mBackground = mGame->GetRenderer()->GetTexture("Assets/DialogBG.png");
SetTitle(text, Vector3::Zero, 30);
// Setup buttons
AddButton("OK", [onOK]() {
onOK();
});
AddButton("Cancel", [this]() {
Close();
});
The constructor first initializes some of the position

member variables for both the title and buttons. Notice
that you also use a new member of UIScreen,
mBackground, which is a texture for a background that
appears behind the UIScreen. In UIScreen::Draw,
you draw the background (if it exists) prior to drawing
anything else.
Finally, DialogBox sets up both OK and Cancel buttons.

You could add additional parameters to DialogBox so
that the user can configure the text of the buttons, as well
as callbacks for both buttons. However, for now you can
just use the OK and Cancel text and assume that the
Cancel button simply closes the dialog box.
Because DialogBox is also a UIScreen, you can add it

to the UI stack by dynamically allocating an instance of
DialogBox. In the case of the pause menu, you change
the Quit button so that it creates a dialog box to confirm
that the user wants to quit:
AddButton("Quit", [this]() {
new DialogBox(mGame, "Do you want to quit?",
[this]() {
mGame->SetState(Game::EQuit);
});
});
Figure 11.3 shows this dialog box for quitting the game.
Figure 11.3 Quit dialog box

note
You could use the UI system as described to this point to also create a main
menu screen. However, this also requires adding additional states to the
Game class because the game can no longer spawn all the objects in the
game world immediately; instead, you need to wait until the player
progresses past the main menu.
HUD ELEMENTS
The types of elements in the HUD vary depending on
the game. Such elements include showing hit points
or ammo count, a score, or an arrow that points to
the next objective. This section looks at two types of
elements that are common for first-person games: a
crosshair (or aiming reticule) and a radar that
shows target positions.
Adding an Aiming Reticule

Most first-person games have some sort of aiming
reticule (such as a crosshair) in the middle of the
screen. As the player aims at different objects, the
reticule might change appearance, with different
textures. For example, if the player can pick up an
object, the reticule might change to a hand. For
games that allow the player to shoot, the reticule
might change colors. If you implement the color
change with a change of texture, then there isn’t a
difference between these two behaviors.
In this case, you can implement a reticule that changes to

red when the player aims at one of the target objects in
the game. To do this, you need to add member variables
to HUD for the different textures, as well as a Boolean for
whether the player is aiming at an enemy:
// Textures for crosshair
class Texture* mCrosshair;
class Texture* mCrosshairEnemy;
// Whether crosshair targets an enemy
bool mTargetEnemy;
To track what the targets are, you create a new

component called TargetComponent. Then, you create
a vector of TargetComponent pointers as a member
variable in HUD:
std::vector<class TargetComponent*> mTargetComps;
You then add AddTarget and RemoveTarget functions

that can add and remove from mTargetComps. You call
these functions in the TargetComponent constructor
and destructor, respectively.
Next, you create an UpdateCrosshair function, as

shown in Listing 11.7. This function gets called by
HUD::Update. You first reset mTargetEnemy back to
false. Next, you use the GetScreenDirection
function first described in Chapter 9, “Cameras.” Recall
that this function returns a normalized vector of the
camera’s current facing in the world. You use this vector
and a constant to construct a line segment, and you use
the SegmentCast function from Chapter 10 to
determine the first actor that intersects with the
segment.
Next, you see if this actor has a TargetComponent. The

way you can check this for now is to find if any
TargetComponent in mTargetComps has an owner
corresponding to the actor that you collide against. You
can optimize this significantly after you implement a
method to figure out which components an actor has;
you will do this in Chapter 14, “Level Files and Binary
Data.”
Listing 11.7 HUD::UpdateCrosshair

Implementation
void HUD::UpdateCrosshair(float deltaTime)
// Reset to regular cursor
mTargetEnemy = false;
// Make a line segment
const float cAimDist = 5000.0f;

Vector3 start, dir;
mGame->GetRenderer()->GetScreenDirection(start, dir);
LineSegment l(start, start + dir * cAimDist);
// Segment cast
PhysWorld::CollisionInfo info;
if (mGame->GetPhysWorld()->SegmentCast(l, info))
// Check if this actor has a target component
for (auto tc : mTargetComps)
if (tc->GetOwner() == info.mActor)
mTargetEnemy = true;
break;
Drawing the crosshair texture is straightforward. In

HUD::Draw, you simply check the value of
mTargetEnemy and draw the corresponding texture in
the center of the screen. You pass in 2.0f as the scale of
the texture, as well:
Texture* cross = mTargetEnemy ? mCrosshairEnemy : mCrosshair;
DrawTexture(shader, cross, Vector2::Zero, 2.0f);
This way, as the player moves the aiming reticule to

target an object, the reticule changes to the red crosshair
texture, as shown in Figure 11.4.
Figure 11.4 Red aiming reticule when aiming at a

target
Adding Radar
A game may have a radar that displays nearby
enemies (or other objects) within a certain radius of
the player. You can represent these enemies on the
radar with blips (which look like dots or circles on
the radar). This way, the player can get a sense of
whether there are enemies around. Some games
always show enemies on the radar, while others show
enemies only under certain conditions (such as if the
enemy recently fired a weapon). However, these
conditions would only be an extension of a basic
approach that shows all enemies.
There are two parts to implementing a working radar.

First, you need to track the actors that should appear on
the radar. Then, on every frame you must update the
blips on the radar based on the position of the actors
relative to the player. The most basic approach is to
represent the blips with a Vector2 offset from the
center of the radar, but you could also add other
properties of blips, such as different textures.
You can leverage existing code and say that any actor
that has a TargetComponent should also appear on the
radar.
For this basic radar, you must add a few member

variables to HUD:
// 2D offsets of blips relative to radar
std::vector<Vector2> mBlips;
// Adjust range of radar and radius
float mRadarRange;
float mRadarRadius;
The mBlips vector tracks the 2D offsets of the blips

relative to the radar center. When updating the radar,
you update mBlips. This way, drawing the radar simply
means drawing the background and then the blip
textures at the required offset.
Finally, the mRadarRange and mRadarRadius

variables are parameters for the radar. The range is how
far the radar sees in the world. For example, a range of
2000 means the radar has a range of 2000 units in world
space. So, for every target that’s within mRadarRange of
the player, you create a blip on the radar. The radius
variable is the radius of the 2D radar drawn onscreen.
Suppose a game has a radar with a range of 50 units.

Now imagine that there’s an object 25 units directly in
front of the player. Because the object positions are in
3D, you need to convert both the position of the player
and the object into 2D coordinates for the onscreen
radar. In the case of a z-up world, this means the radar
acts like a projection of the player and game objects on
the x-y plane. This means your radar ignores the z
components of both the player and the objects it tracks.
Because up on the radar usually denotes forward in

world space, and your world is +x forward, just ignoring
the z component is not enough. For both the player and
any actors on the radar, you need to convert their (x, y, z)
coordinates to the 2D vector (y, x) for the radar offsets.
Once both the player’s and the object’s positions are in

2D radar coordinates, you can construct a vector from
the player to the object, which for clarity is the vector .
The length of determines whether the object is within
the range of the radar. For the previous example with a
range of 50 units and an object 25 units in front, the
length of is less than the maximum range. This means
the object should appear on the radar halfway between
the center of the radar and the edge. You can convert
to a scale relative to the radius of the radar by first
dividing by the maximum range of the radar and then
multiplying by the scale of the radar, saving the result in
a new vector :
However, most radars rotate as the player rotates so that

up on the radar always corresponds to forward in the
game world. This means you can’t just directly use as
the offset of the radar blip. Instead, you need to figure
out the angle between the x-y projection of the player’s
facing vector and the world forward (unit x). Because you
want the angle on the x-y plane, you can compute this
angle θ with the atan2 function and construct a 2D
rotation matrix given θ. Recall that given row vectors, the
2D rotation matrix is as follows:
Once you have the rotation matrix, the final blip offset is
simply rotated by this matrix:
Listing 11.8 shows the code that computes the positions

of all the blips. You loop over all the target components
and test whether the owning actor is in range of the
radar. If so, you compute the blip offset by using the
preceding equations.
Listing 11.8 HUD::UpdateRadar Implementation

void HUD::UpdateRadar(float deltaTime)
// Clear blip positions from last frame
mBlips.clear();
// Convert player position to radar coordinates (x forward, z up)
Vector3 playerPos = mGame->GetPlayer()->GetPosition();
Vector2 playerPos2D(playerPos.y, playerPos.x);
// Ditto for player forward
Vector3 playerForward = mGame->GetPlayer()->GetForward();

Vector2 playerForward2D(playerForward.x, playerForward.y);
// Use atan2 to get rotation of radar
float angle = Math::Atan2(playerForward2D.y, playerForward2D.x);
// Make a 2D rotation matrix
Matrix3 rotMat = Matrix3::CreateRotation(angle);
// Get positions of blips
for (auto tc : mTargetComps)
Vector3 targetPos = tc->GetOwner()->GetPosition();
Vector2 actorPos2D(targetPos.y, targetPos.x);
// Calculate vector between player and target
Vector2 playerToTarget = actorPos2D - playerPos2D;
// See if within range
if (playerToTarget.LengthSq() <= (mRadarRange * mRadarRange))
// Convert playerToTarget into an offset from

// the center of the on-screen radar
Vector2 blipPos = playerToTarget;
blipPos *= mRadarRadius/mRadarRange;
// Rotate blipPos
blipPos = Vector2::Transform(blipPos, rotMat);
mBlips.emplace_back(blipPos);
Drawing the radar is then just a matter of first drawing

the background and then looping through each blip and
drawing it as an offset of the center of the radar:
const Vector2 cRadarPos(-390.0f, 275.0f);
DrawTexture(shader, mRadar, cRadarPos, 1.0f);
// Blips
for (const Vector2& blip : mBlips)
DrawTexture(shader, mBlipTex, cRadarPos + blip, 1.0f);

}
Figure 11.5 shows this radar in a game. The dots on the

radar each correspond to a target actor in the game
world. The arrow in the middle of the radar is just an
extra texture drawn to show where the player is, but this
always just draws at the center of the radar.
Figure 11.5 Radar in a game
Other extensions to the radar could include having

different styles of blips depending on whether the enemy
is above or below the player. Switching between such
styles involves considering the z components of both the
player and the object.
LOCALIZATION
Localization is the process of converting a game
from one region or locale to a different one. The
most common items to localize include any voice-
over dialogue as well as any text shown onscreen. For
example, a game developer working in English may
want to localize to Chinese if releasing the game in
China. The biggest expense of localization is in the
content: Someone must translate all the text and
dialogue, and in the case of dialogue, different actors
must speak the lines in different languages.
However, part of the responsibility of localization falls on

the programmer. In the case of the UI, the game needs
some system to easily show text from different locales
onscreen. This means you can’t hard-code strings such as
“Do you want to quit?” throughout the code. Instead, you
minimally need a map to convert between a key like
"QuitText" and the actual text shown onscreen.
Working with Unicode

One issue with localizing text is that each ASCII
character only has 7 bits of information (though it’s
stored internally as 1 byte). Only 7 bits of
information means there is a total of 128 characters.
Of these characters, 52 are letters (upper- and
lowercase English), and the rest of the characters are
numbers and other symbols. ASCII does not contain
any glyphs from other languages.
To deal with this issue, a consortium of many different
companies introduced the Unicode standard in the
1980s. At this writing, the current version of Unicode
supports over 100,000 different glyphs, including glyphs
in many different languages, as well as emojis.
Because a single byte can’t represent more than 256

distinct values, Unicode must use a different byte
encoding. There are several different byte encodings,
including ones where each character is 2 bytes or 4 bytes.
However, arguably the most popular encoding is UTF-8,
in which each character in a string has a variable length
between 1 and 4 bytes. Within a string, some characters
may be only 1 byte, others may be 2 bytes, others may be
3, and others may be 4.
Although this seems more complex to handle than a

fixed number of bytes for each character, the beauty of
UTF-8 is that it’s fully backward compatible with ASCII.
This means that an ASCII sequence of bytes directly
corresponds to the same UTF-8 sequence of bytes. Think
of ASCII as a special case of UTF-8 in which each
character in the UTF-8 string is 1 byte. The backward
compatibility is likely the reason UTF-8 is the default
encoding for the World Wide Web, as well as for file
formats such as JSON.
Unfortunately, C++ does not have great built-in support

for Unicode. For example, the std::string class is
intended only for ASCII characters. However, you can
use the std::string class to store a UTF-8 string. The
catch is that if the string is UTF-8 encoded, the length
member function no longer guarantees to specify the
number of glyphs (or letters) in the string. Instead, the
length is the number of bytes stored in the string
object.
Luckily, both the RapidJSON library and SDL TTF

support UTF-8 encoding. This, combined with storing
the UTF-8 strings in std::string, means you can add
support for UTF-8 strings without much additional code.
Adding a Text Map

In Game, you add a member variable called
mTextMap that’s a std::unordered_map with key
and value types of std::string. This map converts
a key such as "QuitText" to the displayed text “Do
You Want to Quit?”
You can define this map in a simple JSON file format, as

shown in Listing 11.9. Every language has its own version
of this JSON file, which makes it easy to switch between
different languages.
Listing 11.9 English.gptext Text Map File

"TextMap":{
"PauseTitle": "PAUSED",
"ResumeButton": "Resume",
"QuitButton": "Quit",
"QuitText": "Do you want to quit?",
"OKButton": "OK",
"CancelButton": "Cancel"
You then add a LoadText function to Game that parses

in a gptext file and populates the mTextMap. (This
function calls various RapidJSON functions to parse in
the file, but we omit its implementation here for the sake
of brevity.)
You likewise implement a GetText function in a game

and, given a key, it returns the associated text. This just
performs a find operation on the mTextMap.
You then make two modifications to

Font::RenderText. First, rather than directly
rendering the text string it takes in as a parameter, you
have that text string looked up in the text map:
const std::string& actualText = mGame->GetText(textKey);
Next, instead of calling TTF_RenderText_Blended,

you call TTF_RenderUTF8_Blended, which has the
same syntax but takes in a UTF-8 encoded string instead
of an ASCII string:
SDL_Surface* surf = TTF_RenderUTF8_Blended(font,
actualText.c_str(), sdlColor);
Finally, any code that previously used a hard-coded text

string instead uses the text key. For example, the title
text for the pause menu is no longer "PAUSED" but
"PauseTitle". This way, when you eventually call
RenderText, the correct text loads from the map.
tip
If the game code has finalized English text, a quick hack to localize the text is
to use the finalized English text as the text key. This way, you don’t have to
track down every single non-localized string usage in the code. However, this
can be dangerous if someone later changes the English strings in the code,
thinking that this will change the text onscreen!
To demonstrate the functionality of this code, you can

create a Russian.gptext file with a Russian
translation of the strings in Listing 11.9. Figure 11.6
shows the Russian version of the pause menu with the
Do You Want to Quit? dialog box.
Figure 11.6 Pause menu in Russian
Other Localization Concerns

The code presented in this section works only if the
TrueType font file supports all needed glyphs. In
practice, it’s common for a font file to include only a
subset of glyphs. Some languages, such as Chinese,
typically have a dedicated font file for the language.
To solve this issue, you could add a font entry to the
gptext file. When populating the mTextMap, you
can also load the correct font. Then, the rest of the UI
code needs to be sure to use this correct font.
Some issues with localization aren’t clear at first glance.

For example, German text is typically 20% longer than
the equivalent English text. This means that if a UI
element barely fits the English text, it likely won’t fit the
German text. Although this usually is a content issue, it
may be an issue if the UI code assumes certain paddings
or text sizes. One way to circumvent this is to always
query the size of the rendered font texture and scale
down the size of the text if it won’t fit within the needed
extents.
Finally, in some instances, content beyond text or

dialogue might need a localization process. For example,
in Germany, it is illegal to sell a product containing
symbols associated with the Third Reich. So, a game set
in World War II may show swastikas or other symbols in
the English version, but the German version must
replace these symbols with alternatives such as the iron
cross. As another example, some games also have
content restrictions in China (such as not showing too
much blood). However, this type of issue usually can be
solved without additional help from a programmer, since
the artists can simply create alternative content for these
regions.
SUPPORTING MULTIPLE
RESOLUTIONS
For PC and mobile games, it’s very common to have
players with different screen resolutions. On a PC,
common monitor resolutions include 1080p
(1920×1080), 1440p (2560×1440), and 4K
(3840×2160). On mobile platforms, there are a
staggering number of different device resolutions.
Although the Renderer class currently supports
creating the window at different resolutions, the UI
code in this chapter assumes a fixed resolution.
One way to support multiple resolutions is to avoid using

specific pixel locations, or absolute coordinates, for UI
elements. An example of using an absolute coordinate is
placing a UI element precisely at the coordinate (1900,
1000) and assuming that this corresponds to the bottom-
right corner.
Instead, you could use relative coordinates, where the

coordinates are relative to a specific part of the screen,
called an anchor. For example, in relative coordinates,
you can say that you want to place an element at (–100,–
100) relative to the bottom-right corner. This means the
element would appear at (1820, 980) on a 1080p screen,
while it would appear at (1580, 950) on a 1680×1050
screen (see Figure 11.7). You can express coordinates
relative to key points on the screen (usually the corners
or the center of the screen) or even relative to other UI
elements. To implement this, you need to be able to
specify the anchor points and relative coordinates for UI
elements and then calculate the absolute coordinates
dynamically at runtime.
Figure 11.7 A UI element positioned relative to the
bottom-right corner of the screen
Another refinement is to scale the size of UI elements

depending on the resolution. This is useful because at
very high resolutions, the UI might become too small and
unusable. At higher resolutions, you could scale the size
of the UI elements or even make the UI scale an option
for the player.
GAME PROJECT
features discussed in this chapter except for
supporting multiple resolutions. The Game class has
a UI stack, along with a UIScreen class, a
PauseMenu class, and a DialogBox class. The HUD
demonstrates both the aiming reticule and the radar.
The code also implements text localization. The code
In the game, use the standard first-person controls

(WASD plus mouse look) to move around the world. Use
the Escape key to enter the pause menu and the mouse
controls to select and click on buttons in the menu.
During gameplay, use the 1 and 2 key to switch between
English (1) and Russian (2) text. Pressing these keys
while the pause menu is showing doesn’t do anything
because the UI screen absorbs the game input.
SUMMARY
This chapter provides a high-level overview of the
challenges involved in implementing the user
interface in code. Using the SDL TTF library is a
convenient way to render fonts, as it can load in
TrueType fonts and then render the text to a texture.
In the UI stack system, you represent each unique UI
screen as an element on the UI stack. At any point in
time, only the topmost screen on the UI might
receive input from the player. You can extend this
system to support buttons as well as dialog boxes.
The HUD might contain many different elements,

depending on the game. An aiming reticule that changes
based on aiming at an object requires use of collision
detection to determine what object the player is aiming
at. If the player aims at a target object, the HUD can
draw a different texture. For a radar, you can project the
player and any enemy objects onto the x/y plane and use
these converted coordinates to determine where to draw
a blip on the radar.
Finally, the UI needs code to handle text for different

locales. A simple map can convert between text keys and
values. Using UTF-8 encoding of these text values makes
it relatively painless to use. The RapidJSON library can
load in JSON files encoded in UTF-8, and SDL TTF
supports rendering UTF-8 strings.
ADDITIONAL READING
Desi Quintans’ short article gives examples of good
and bad game UI, from a design perspective. Luis
Sempé, a UI programmer for games including Deus
Ex: Human Revolution, has written the only book
solely dedicated to programming UIs for games. (In
the interest of full disclosure, I worked with the
author many years ago.) Finally, Joel Spolsky’s book
is for UI design in general, but it provides insight
into how to create an effective UI.
Quintans, Desi. “Game UI by Example: A

Crash Course in the Good and the Bad.”
https://gamedevelopment.tutsplus.com/tutorials/
game-ui-by-example-a-crash-course-in-the-
good-and-the-bad--gamedev-3943.
Accessed September 10, 2017.
Sempé, Luis. User Interface Programming
for Games. Self-published, 2014.
Spolsky, Joel. User Interface Design for
Programmers. Berkeley: Apress, 2001.
EXERCISES
In this chapter’s exercises, you explore adding a main
menu as well as making changes to the game’s HUD.
Exercise 11.1
Create a main menu. To support this, the game class
needs a new state called EMainMenu. The game
should first start in this state and display a UI screen
with the menu options Start and Quit. If the player
clicks Start, the game should switch to gameplay. If
the player clicks Quit, the menu should show a dialog
box confirming that the player wants to quit.
To add further functionality, consider spawning the

actors only when first entering the gameplay state from
the main menu. In addition, change the pause menu so
that the Quit option deletes all actors and returns to the
main menu rather than immediately quitting the game.
Exercise 11.2
Modify the radar so that it uses different blip
textures, depending on whether the actor is above or
below the player. Use the provided BlipUp.png and
BlipDown.png textures to show these different
states. Testing this feature may require changing the
positions of some of the target actors in order to
more clearly distinguish the height.
Exercise 11.3
Implement an onscreen 2D arrow that points to a
specific actor. Create a new type of actor called
ArrowTarget and place it somewhere in the game
world. Then, in the HUD, compute the vector from
the player to the ArrowTarget. Use the angle
between this and the player’s forward on the x-y
plane to determine the angle to rotate the onscreen
2D arrow. Finally, add code to
UIScreen::DrawTexture to supporting rotating a
texture (with a rotation matrix).
CHAPTER 12
SKELETAL ANIMATION
Animating characters for a 3D game is very

different from animating characters for a 2D
game. This chapter looks at skeletal
animation, the most common animation
used in 3D games. This chapter first goes
over the mathematical foundations of the
approach and then dives into the
implementation details.
FOUNDATIONS OF SKELETAL
ANIMATION
As described in Chapter 2, “Game Objects and 2D
Graphics,” for 2D animation, games use a sequence
of image files to yield the illusion of an animated
character. A naïve solution for animating 3D
characters is similar: Construct a sequence of 3D
models and render those different models in rapid
succession. Although this solution conceptually
works, it’s not a very practical approach.
Consider a character model composed of 15,000

triangles, which is a conservative number for a modern
game. Assuming only 10 bytes per vertex, the total
memory usage of this one model might be around 50 to
100 KB. A two-second animation running at 30 frames
per second would need a total of 60 different models.
This means the total memory usage for this single
animation would be 3 to 6 MB. Now imagine that the
game uses several different animations and several
different character models. The memory usage for the
game’s models and animations will quickly become too
high.
In addition, if a game has 20 different humanoid

characters, chances are their movements during an
animation such as running are largely the same. If you
use the naïve solution just described, each of these 20
characters needs a different model set for its animations.
This also means that artists need to manually author
these different model sets and animations for each
character.
Because of these issues, most 3D games instead take

inspiration from anatomy: vertebrates, like humans,
have bones. Attached to these bones are muscles, skin,
and other tissue. Bones are rigid, while the other tissues
are not. Thus, given the position of a bone, it’s possible to
derive the position of the tissue attached to the bone.
Similarly, in skeletal animation, the character has an

underlying rigid skeleton. This skeleton is what the
animator animates. Then, each vertex in the model has
an association with one or more bones in the skeleton.
When the animation moves the bones, the vertices
deform around the associated bones (much the way your
skin stretches when you move around). This means that
there only needs to be a single 3D model for a character,
regardless of the number of animations for the model.
note
Because skeletal animation has bones and vertices that deform along the
bones, some call this technique skinned animation. The “skin” in this case
is the model’s vertices.
Similarly, the terms bone and joint, though different in the context of anatomy,
are interchangeable terms in the context of skeletal animation.
An advantage of skeletal animation is that the same

skeleton can work for several different characters. For
example, it’s common in a game for all humanoid
characters to share the same skeleton. This way, the
animator creates one set of animations for the skeleton,
and all characters can then use those animations.
Furthermore, many popular 3D model authoring

programs such as Autodesk Maya and Blender support
skeletal animations. Thus, artists can use these tools to
author the skeletons and animations for characters.
Then, as with 3D models, you can write exporter plugins
to export into the preferred format for the game code. As
with 3D models this book uses a JSON-based file format
for the skeleton and animations. (As a reminder, the
book’s code on GitHub includes an exporter for Epic
Games’s Unreal Engine, in the Exporter directory.)
The remainder of this section looks at the high-level

concepts and mathematics that drive skeletal animation.
The subsequent section then dives into the details of
implementing skeletal animation in game code.
Skeletons and Poses

The usual representation of a skeleton is as a
hierarchy (or tree) of bones. The root bone is the
base of the hierarchy and has no parent bone. Every
other bone in the skeleton has a single parent bone.
Figure 12.1 illustrates a simple skeletal hierarchy for
a humanoid character. The spine bone is a child of
the root bone, and then in turn the left and right hip
bones are children of the spine bone.
Figure 12.1 Character with a basic skeleton with
some bones labeled
This bone hierarchy attempts to emulate anatomy. For

example, if a human rotates her shoulder, the rest of the
arm follows that rotation. With a game skeleton, you
might represent this by saying the shoulder bone is the
parent of the elbow bone, the elbow bone is the parent of
the wrist bone, and the wrist bone is the parent of the
finger bones.
Given a skeleton, a pose represents a configuration of

the skeleton. For example, if a character waves hello in
an animation, there is one pose during the animation
where the character’s hand bone is raised up to wave. An
animation is then just a sequence of poses the skeleton
transitions between over time.
The bind pose is the default pose of the skeleton, prior

to applying any animation. Another term for bind pose is
t-pose because, typically, the character’s body forms a T
shape in bind pose, as in Figure 12.1. You author a
character’s model so that it looks like this bind pose
configuration.
The reason the bind pose usually looks like a T because it

makes it easier to associate bones to vertices, as
discussed later in this chapter.
In addition to specifying the parent/child relationships

of the bones in the skeleton, you also must specify each
bone’s position and orientation. Recall that in a 3D
model, each vertex has a position relative to the object
space origin of the model. In the case of a humanoid
character, a common placement of the object space
origin is between the feet of the character in bind pose.
It’s not accidental that this also corresponds to the
typical placement of the root bone of the skeleton.
For each bone in the skeleton, you can describe its

position and orientation in two ways. A global pose is
relative to the object space origin. Conversely, a local
pose is relative to a parent bone. Because the root bone
has no parent, its local pose and global pose are
identical. In other words, the position and orientation of
the root bone is always relative to the object space origin.
Suppose that you store local pose data for all the bones.
One way to represent position and orientation is with a
transform matrix. Given a point in the bone’s coordinate
space, this local pose matrix would transform the point
into the parent’s coordinate space.
If each bone has a local pose matrix, then given the

parent/child relationships of the hierarchy, you can
always calculate the global pose matrix for any bone. For
example, the parent of the spine is the root bone, so its
local pose matrix is its position and orientation relative
to the root bone. As established, the root bone’s local
pose matrix corresponds to its global pose matrix. So
multiplying the local pose matrix of the spine by the root
bone’s global pose matrix yields the global pose matrix
for the spine:
With the spine’s global pose matrix, given a point in the

spine’s coordinate space, you could transform it into
object space.
Similarly, to compute the global pose matrix of the left

hip, whose parent is the spine, the calculation is as
follows:
Because you can always convert from local poses to

global poses, it may seem reasonable to store only local
poses. However, by storing some information in global
form, you can reduce the number of calculations
required every frame.
Although storing bone poses with matrices can work,

much as with actors, you may want to separate the bone
position and orientation into a vector for the translation
and a quaternion for the rotation. The main reason to do
this is that quaternions allow for more accurate
interpolation of the rotation of a bone during an
animation. You can omit a scale for the bones because
scaling bones typically only sees use for cartoon-style
characters who can stretch in odd ways.
You can combine the position and orientation into the

following BoneTransform struct:
struct BoneTransform
Vector3 mTranslation;
// Convert to matrix
Matrix4 ToMatrix() const;
};
The ToMatrix function converts the transform into a

matrix. This just creates rotation and translation
matrices from the member data and multiplies these
matrices together. This function is necessary because
even though many intermediate calculations directly use
the quaternion and vector variables, ultimately the
graphics code and shaders need matrices.
To define the overall skeleton, for each bone you need to
know the name of the bone, the parent of the bone, and
its bone transform. For the bone transform, you
specifically store the local pose (the transform from the
parent) when the overall skeleton is in the bind pose.
One way to store these bones is in an array. Index 0 of

the array corresponds to the root bone, and then each
subsequent bone references its parent by an index
number. For the example in Figure 12.2, the spine bone,
stored in index 1, has a parent index of 0 because the
root bone is its parent. Similarly, the hip bone, stored in
index 2, has a parent of index 1.
Figure 12.2 Representation of skeleton as an array

of bones
This leads to the following Bone struct that contains the

local bind pose transform, a bone name, and a parent
index:
struct Bone
BoneTransform mLocalBindPose;
std::string mName;
int mParent;
};
Then, you define a std::vector of bones that you can

fill in based on the skeleton. The root bone sets its parent
index to -1, but every other bone has a parent indexing
into the array. To simplify later calculations, parents
should be at earlier indices in the array than their
children bones. For example, because the left hip is a
child of spine, it should never be the case that left hip has
a lower index than spine.
The JSON-based file format used to store the skeleton

data directly mirrors this representation. Listing 12.1
gives a snippet of a skeleton file, showing the first two
bones: root and pelvis.
Listing 12.1 The Beginning of a Skeleton Data File

"version":1,
"bonecount":68,
"bones":[
"name":"root",
"parent":-1,
"bindpose":{
"rot":[0.000000,0.000000,0.000000,1.000000],
"trans":[0.000000,0.000000,0.000000]
},
"name":"pelvis",
"parent":0,
"bindpose":{
"rot":[0.001285,0.707106,-0.001285,-0.707106],
"trans":[0.000000,-1.056153,96.750603]
},
// ...
The Inverse Bind Pose Matrix

With the local bind pose information stored in the
skeleton, you can easily compute a global bind pose
matrix for every bone by using matrix multiplication,
as shown earlier. Given a point in a bone’s coordinate
space, multiplying by this global bind pose matrix
yields that point transformed into object space. This
assumes that the skeleton is in the bind pose.
The inverse bind pose matrix for a bone is simply the

inverse of the global bind pose matrix. Given a point in
object space, multiplying it by the inverse bind pose
matrix yields that point transformed into the bone’s
coordinate space. This is actually very useful because the
model’s vertices are in object space, and the models’
vertices are in the bind pose configuration. Thus, the
inverse bind pose matrix allows you to transform a
vertex from the model into a specific bone’s coordinate
space (in bind pose).
For example, you can compute the spine bone’s global

bind pose matrix with this:
Its inverse bind pose matrix is then simply as follows:
The simplest way to compute the inverse bind pose

matrix is in two passes. First, you compute each bone’s
global bind pose matrix using the multiplication
procedure from the previous section. Second, you invert
each of these matrices to get the inverse bind pose.
Because the inverse bind pose matrix for each bone
never changes, you can compute these matrices when
loading the skeleton.
Animation Data
Much the way you describe the bind pose of a
skeleton in terms of local poses for each of the bones,
you can describe any arbitrary pose. More formally,
the current pose of a skeleton is just the set of local
poses for each bone. An animation is then simply a
sequence of poses played over time. As with the bind
pose, you can convert these local poses into global
pose matrices for each bone, as needed.
You can store this animation data as a 2D dynamic array

of bone transforms. In this case, the row corresponds to
the bone, and the column corresponds to the frame of
the animation.
One issue with storing animation on a per-frame basis is

that the frame rate of the animation may not correspond
to the frame rate of the game. For example, the game
may update at 60 FPS, but the animation may update at
30 FPS. If the animation code tracks the duration of the
animation, then every frame, you can update this by
delta time. However, it will sometimes be the case that
the game needs to show the animation between two
different frames. To support this, you can add a static
Interpolate function to BoneTransform:

BoneTransform BoneTransform::Interpolate(const BoneTransform& a,
const BoneTransform& b, float f)
BoneTransform retVal;
retVal.mRotation = Quaternion::Slerp(a.mRotation, b.mRotation, f);
retVal.mTranslation = Vector3::Lerp(a.mTranslation,
b.mTranslation, f);
return retVal;
Then, if the game must show a state between two

different frames, you can interpolate the transforms of
each bone to get the current local pose.
Skinning
Skinning involves associating vertices in the 3D
model with one or more bones in the corresponding
skeleton. (This is different from the term skinning in
a non-animation context.) Then, when drawing a
vertex, the position and orientation of any associated
bones influence the position of the vertex. Because
the skinning of a model does not change during the
game, the skinning information is an attribute of
each vertex.
In a typical implementation of skinning, each vertex can
have associations with up to four different bones. Each of
these associations has a weight, which designates how
much each of the four bones influences the vertex. These
weights must sum to one. For example, the spine and left
hip bone might influence a vertex on the lower-left part
of the torso of the character. If the vertex is closer to the
spine, it might have a weight of 0.7 for the spline bone
and 0.3 for the hip bone. If a vertex has only one bone
that influences it, as is common, then that one bone
simply has a weight of 1.0.
For the moment, don’t worry about how to add these

additional vertex attributes for both the bones and
skinning weights. Instead, consider the example of a
vertex that has only one bone influencing it. Remember
that the vertices stored in the vertex buffer are in object
space, while the model is in bind pose. But if you want to
draw the model in an arbitrary pose, P, you must then
transform each vertex from object space bind pose into
object space in the current pose, P.
To make this example concrete, suppose that the sole

bone influence of vertex v is the spine bone. You already
know the inverse bind pose matrix for the spine from
earlier calculations. In addition, from the animation
data, you can calculate the spine’s global pose matrix for
the current pose, P. To transform v into object space of
the current pose, you first transform it into the local
space of the spine in bind pose. Then you transform it
into object space of the current pose. Mathematically, it
looks like this:
Now, suppose that v instead has two bone influences:

The spine has a weight of 0.75, and the left hip has a
weight of 0.25. To calculate v in the current pose in this
case, you need to calculate each bone’s current pose
vertex position separately and then interpolate between
them, using these weights:
You could similarly extend the calculation for a vertex

with four different bone influences.
Some bones, such as the spine, influence hundreds of

vertices on the character model. Recalculating the
multiplication between the spine’s inverse bind pose
matrix and the current pose matrix for each of these
vertices is redundant. On a single frame, the result of this
multiplication will never change. The solution is to create
an array of matrices called the matrix palette. Each
index in this array contains the result of the
multiplication between the inverse bind pose matrix and
the current pose matrix for the bone with the
corresponding index.
For example, if the spine is at index 1 in the bone array,
then index 1 of the matrix palette contains the following:
Any vertex that’s influenced by the spine can then use the
precomputed matrix from the palette. For the case of the
vertex solely influenced by the spine, its transformed
position is as follows:
Using this matrix palette saves thousands of extra matrix

multiplications per frame.
IMPLEMENTING SKELETAL
ANIMATION
With the mathematical foundations established, you
can now add skeletal animation support to the game.
First, you add support for the additional vertex
attributes that a skinned model needs (bone
influences and weights), and then you draw the
model in bind pose. Next, you add support for
loading the skeleton and compute the inverse bind
pose for each bone. Then, you can calculate the
current pose matrices of an animation and save the
matrix palette. This allows you to draw the model in
the first frame of an animation. Finally, you add
support for updating the animation based on delta
time.
Drawing with Skinning Vertex Attributes

Although drawing a model with different vertex
attributes seems straightforward, several pieces of
code written in Chapter 6, “3D Graphics,” assume a
single vertex layout. Recall that to this point, all 3D
models have used a vertex layout with a position, a
normal, and texture coordinates. To add support for
the new skinning vertex attributes, you need to make
a nontrivial number of changes.
First, you create a new vertex shader called

Skinned.vert. Recall that you write shaders in GLSL,
not C++. You don’t need a new fragment shader in this
case because you still want to light the pixels with the
Phong fragment shader from Chapter 6. Initially,
Skinned.vert is just a copy of Phong.vert. Recall
that the vertex shader must specify the expected vertex
layout of each incoming vertex. Thus, you must change
the declaration of the vertex layout in Skinned.vert to
the following:
layout(location = 1) in vec3 inNormal;
layout(location = 2) in uvec4 inSkinBones;
layout(location = 3) in vec4 inSkinWeights;

This set of declarations says that you expect the vertex

layout to have three floats for position, three floats for
the normal, four unsigned integers for the bones that
influence the vertex, four floats for the weights of these
bone influences, and two floats for the texture
coordinates.
The previous vertex layout—with position, normal, and

texture coordinates—uses single-precision floats (4 bytes
each) for all the values. Thus, the old vertex layout has a
size of 32 bytes. If you were to use single-precision floats
for the skinning weights and full 32-bit integers for the
skinned bones, this would add an additional 32 bytes,
doubling the size of each vertex in memory.
Instead, you can limit the number of bones in a model to

256. This means you only need a range of 0 to 255 for
each bone influence—or a single byte each. This reduces
the size of inSkinBones from 16 bytes to 4 bytes. In
addition, you can specify that the skinning weights will
also be in a range of 0 to 255. OpenGL can then
automatically convert this 0–255 range to a normalized
floating-point range of 0.0–1.0. This reduces the size of
inSkinWeights to 4 bytes, as well. This means that, in
total, the size of each vertex will be the original 32 bytes,
plus an additional 8 bytes for the skinned bones and
weights. Figure 12.3 illustrates this layout.
Figure 12.3 Vertex layout with bone influences and
weights
To reduce the memory usage of inSkinBones and

inSkinWeights, you don’t need to make any further
changes to the shader code. Instead, you need to specify
the expected sizes of these attributes when defining the
vertex array attributes in your C++ code. Recall from
Chapter 5, “OpenGL,” that the definition of the vertex
array attributes occurs in the VertexArray constructor.
To support different types of vertex layouts, you add a
new enum to the declaration of the VertexArray class
in VertexArray.h:
enum Layout
PosNormTex,
PosNormSkinTex
};
Then, you modify the VertexArray constructor so that
it takes in a Layout as a parameter. Then, in the code
for the constructor you check the layout to determine
how to define the vertex array attributes. For the case of
PosNormTex, you use the previously written vertex
attribute code. Otherwise, if the layout is
PosNormSkinTex, you define the layout as in Listing
12.2.
Listing 12.2 Declaring Vertex Attributes in the

VertexArray Constructor
if (layout == PosNormTex)
{ /* From Chapter 6... */ }
else if (layout == PosNormSkinTex)
// Position is 3 floats
glVertexAttribPointer(0, 3, GL_FLOAT, GL_FALSE, vertexSize, 0);
// Normal is 3 floats
glVertexAttribPointer(1, 3, GL_FLOAT, GL_FALSE, vertexSize,

// Skinning bones (keep as ints)
glVertexAttribIPointer(2, 4, GL_UNSIGNED_BYTE, vertexSize,
// Skinning weights (convert to floats)
glVertexAttribPointer(3, 4, GL_UNSIGNED_BYTE, GL_TRUE, vertexSize,
reinterpret_cast<void*>(sizeof(float) * 6 + 4));
// Texture coordinates
glVertexAttribPointer(4, 2, GL_FLOAT, GL_FALSE, vertexSize,
reinterpret_cast<void*>(sizeof(float) * 6 + 8));
The declarations for the first two attributes, position and

normal, are the same as in Chapter 6. Recall that the
parameters to glVertexAttribPointer are the
attribute number, the number of elements in the
attribute, the type of the attribute (in memory), whether
OpenGL should normalize the value, the size of each
vertex (or stride), and the byte offset from the start of the
vertex to that attribute. So, both the position and normal
are three float values.
Next, you define the skinning bones and weight

attributes. For the bones, you use glVertexAttribIPointer,
which is for values that are integers in the shader.
Because the definition of inSkinBones uses four
unsigned integers, you must use the AttribI function
instead of the regular Attrib version. Here, you specify
that each integer is an unsigned byte (from 0 to 255). For
the weights, you specify that each is stored in memory as
an unsigned byte, but you want to convert these
unsigned bytes to a normalized float value from 0.0 to
1.0.
Finally, the declaration of the texture coordinates is the

same as in Chapter 6, except that they have a different
offset because they appear later in the vertex layout.
Once you have defined the vertex attributes, the next

step is to update the Mesh file loading code to load in a
gpmesh file with skinning vertex attributes. (This chapter
omits the code for file loading in the interest of brevity.
But as always, the source code is available in this
chapter’s corresponding game project.)
Next, you declare a SkeletalMeshComponent class

that inherits from MeshComponent, as in Listing 12.3.
For now, the class does not override any behavior from
the base MeshComponent. So, the Draw function for
now simply calls MeshComponent::Draw. This will
change when you begin playing animations.
Listing 12.3 SkeletalMeshComponent Declaration
class SkeletalMeshComponent : public MeshComponent
public:
SkeletalMeshComponent(class Actor* owner);
// Draw this mesh component
void Draw(class Shader* shader) override;
};
Then, you need to make changes to the Renderer class

to separate meshes and skeletal meshes. Specifically, you
create a separate std::vector of
SkeletalMeshComponent pointers. Then, you change
the Renderer::AddMesh and RemoveMesh function to
add a given mesh to either the normal
MeshComponent* vector or the one for
SkeletalMeshComponent pointers. (To support this,
you add a mIsSkeletal member variable to
MeshComponent that says whether the mesh is skeletal.)
Next, you load the skinning vertex shader and the Phong
fragment shaders in Renderer::LoadShader and save
the resulting shader program in a mSkinnedShader
member variable.
Finally, in Renderer::Draw, after drawing the regular
meshes, you draw all the skeletal meshes. The code is
almost identical to the regular mesh drawing code from
Chapter 6, except you use the skeletal mesh shader:
// Draw any skinned meshes now
mSkinnedShader->SetActive();
// Update view-projection matrix
mSkinnedShader->SetMatrixUniform("uViewProj", mView * mProjection);
// Update lighting uniforms
SetLightUniforms(mSkinnedShader);
for (auto sk : mSkeletalMeshes)
if (sk->GetVisible())
sk->Draw(mSkinnedShader);
With all this code in place, you can now draw a model
with skinning vertex attributes, as in Figure 12.4. The
character model used in this chapter is the Feline
Swordsman model created by Pior Oberson. The model
file is CatWarrior.gpmesh in the Assets directory for
this chapter’s game project.
Figure 12.4 Drawing the Feline Swordsman model

in bind pose
The character faces to the right because the bind pose of

the model faces down the +y axis, whereas this book’s
game uses a +x axis as forward. However, the animations
all rotate the model to face toward the +x axis. So, once
you begin playing the animations, the model will face in
the correct direction.
Loading a Skeleton
Now that the skinned model is drawing, the next step
is to load the skeleton. The gpskel file format simply
defines the bones, their parents, and the local pose
transform for every bone in bind pose. To
encapsulate the skeleton data, you can declare a
Skeleton class, as shown in Listing 12.4.
Listing 12.4 Skeleton Declaration

class Skeleton
public:
// Definition for each bone in the skeleton
struct Bone
BoneTransform mLocalBindPose;
std::string mName;
int mParent;
};
// Load from a file
// Getter functions
size_t GetNumBones() const { return mBones.size(); }

const Bone& GetBone(size_t idx) const { return mBones[idx]; }
const std::vector<Bone>& GetBones() const { return mBones; }
const std::vector<Matrix4>& GetGlobalInvBindPoses() const
{ return mGlobalInvBindPoses; }
protected:
// Computes the global inverse bind pose for each bone
// (Called when loading the skeleton)
void ComputeGlobalInvBindPose();
private:
// The bones in the skeleton
std::vector<Bone> mBones;
// The global inverse bind poses for each bone
std::vector<Matrix4> mGlobalInvBindPoses;
};
In the member data of Skeleton, you store both a

std::vector for all of the bones and a std::vector
for the global inverse bind pose matrices. The Load
function is not particularly notable, as it just parses in
the gpmesh file and converts it to the vector of bones
format discussed earlier in the chapter. (As with the
other JSON file loading code, this chapter omits the code
in this case, though it is available with the project code in
the book’s GitHub repository.)
If the skeleton file loads successfully, the function then

calls the ComputeGlobalInvBindPose function,
which uses matrix multiplication to calculate the global
inverse bind pose matrix for every bone. You use the
two-pass approach discussed earlier in the chapter: First,
you calculate each bone’s global bind pose matrix, and
then you invert each of these matrices to yield the inverse
bind pose matrix for each bone. Listing 12.5 gives the
implementation of ComputeGlobalInvBindPose.
Listing 12.5 ComputeGlobalInvBindPose

Implementation
void Skeleton::ComputeGlobalInvBindPose()
// Resize to number of bones, which automatically fills identity
mGlobalInvBindPoses.resize(GetNumBones());
// Step 1: Compute global bind pose for each bone
// The global bind pose for root is just the local bind pose
mGlobalInvBindPoses[0] = mBones[0].mLocalBindPose.ToMatrix();

// Each remaining bone's global bind pose is its local pose
// multiplied by the parent's global bind pose
for (size_t i = 1; i < mGlobalInvBindPoses.size(); i++)
Matrix4 localMat = mBones[i].mLocalBindPose.ToMatrix();
mGlobalInvBindPoses[i] = localMat *
mGlobalInvBindPoses[mBones[i].mParent];
// Step 2: Invert each matrix
for (size_t i = 0; i < mGlobalInvBindPoses.size(); i++)
mGlobalInvBindPoses[i].Invert();
Using the familiar pattern for loading in data files, you

can add an unordered_map of Skeleton pointers to the
Game class, as well as code to load a skeleton into the
map and retrieve it from the map.
Finally, because each SkeletalMeshComponent also

needs to know its associated skeleton, you add a
Skeleton pointer to the member data of
SkeletalMeshComponent. Then when creating the
SkeletalMeshComponent object, you also assign the
appropriate skeleton to it.
Unfortunately, adding the Skeleton code does not

make any visible difference over just drawing the
character model in bind pose. To see anything change,
you need to do more work.
Loading the Animation Data

The animation file format this book uses is also
JSON. It first contains some basic information, such
as the number of frames and duration (in seconds) of
the animation, as well as the number of bones in the
associated skeleton. The remainder of the file is local
pose information for the bones in the model during
the animation. The file organizes the data into
tracks, which contain pose information for each
bone on each frame. (The term tracks comes from
time-based editors such as video and sound editors.)
If the skeleton has 10 bones and the animation has
50 frames, then there are 10 tracks, and each track
has 50 poses for that bone. Listing 12.6 shows the
basic layout of this gpanim data format.
Listing 12.6 The Beginning of an Animation Data File

{
"version":1,
"sequence":{
"frames":19,
"duration":0.600000,
"bonecount":68,
"tracks":[
"bone":0,
"transforms":[
"rot":[-0.500199,0.499801,-0.499801,0.500199],
"trans":[0.000000,0.000000,0.000000]
},
"rot":[-0.500199,0.499801,-0.499801,0.500199],
"trans":[0.000000,0.000000,0.000000]
},
// Additional transforms up to frame count
// ...
],
// Additional tracks for each bone
// ...
This format does not guarantee that every bone has a

track, which is why each track begins with a bone index.
In some cases, bones such as the fingers don’t need to
have any animation applied to them. In such a case, the
bone simply would not have a track. However, if a bone
has a track, it will have a local pose for every single frame
in the animation.
Also, the animation data for each track contains an extra

frame at the end that’s a duplicate of the first frame. So
even though the example above says there are 19 frames
with a duration of 0.6 seconds, frame 19 is actually a
duplicate of frame 0. So, there are really only 18 frames,
with a rate in this case of exactly 30 FPS. This duplicate
frame is included because it makes looping slightly easier
to implement.
As is the case for the skeleton, you declare a new class

called Animation to store the loaded animation data.
Listing 12.7 shows the declaration of the Animation
class. The member data contains the number of bones,
the number of frames in the animation, the duration of
the animation, and the tracks containing the pose
information for each bone. As is the case with the other
JSON-based file formats, this chapter omits the code for
loading the data from the file. However, the data stored
in the Animation class clearly mirrors the data in the
gpanim file.
Listing 12.7 Animation Declaration

class Animation
public:
size_t GetNumBones() const { return mNumBones; }
size_t GetNumFrames() const { return mNumFrames; }
float GetDuration() const { return mDuration; }
float GetFrameDuration() const { return mFrameDuration; }
// Fills the provided vector with the global (current) pose matrices
// for each bone at the specified time in the animation.
void GetGlobalPoseAtTime(std::vector<Matrix4>& outPoses,

const class Skeleton* inSkeleton, float inTime) const;
private:
// Number of bones for the animation
size_t mNumBones;
// Number of frames in the animation
size_t mNumFrames;
// Duration of the animation in seconds
float mDuration;
// Duration of each frame in animation
float mFrameDuration;
// Transform information for each frame on the track
// Each index in the outer vector is a bone, inner vector is a frame
std::vector<std::vector<BoneTransform>> mTracks;
};
The job of the GetGlobalPoseAtTime function is to

compute the global pose matrices for each bone in the
skeleton at the specified inTime. It writes these global
pose matrices to the provided outPoses
std::vector of matrices. For now, you can ignore the
inTime parameter and just hard-code the function so
that it uses frame 0. This way, you can first get the game
to draw the first frame of the animation properly. The
“Updating Animations” section, later in this chapter,
circles back to GetGlobalPoseAtTime and how to
properly implement it.
To compute the global pose for each bone, you follow the
same approach discussed before. You first set the root
bone’s global pose, and then each other bone’s global
pose is its local pose multiplied by its parent’s global
pose. The first index of mTracks corresponds to the
bone index, and the second index corresponds to the
frame in the animation. So, this first version of
GetGlobalPoseAtTime hard-codes the second index
to 0 (the first frame of the animation), as shown in
Listing 12.8.
Listing 12.8 First Version of GetGlobalPoseAtTime

void Animation::GetGlobalPoseAtTime(std::vector<Matrix4>& outPoses,
const Skeleton* inSkeleton, float inTime) const
// Resize the outPoses vector if needed
if (outPoses.size() != mNumBones)
outPoses.resize(mNumBones);

// For now, just compute the pose for every bone at frame 0
const int frame = 0;
// Set the pose for the root
// Does the root have a track?
if (mTracks[0].size() > 0)
// The global pose for the root is just its local pose
outPoses[0] = mTracks[0][frame].ToMatrix();
else
outPoses[0] = Matrix4::Identity;
const std::vector<Skeleton::Bone>& bones = inSkeleton->GetBones();
// Now compute the global pose matrices for every other bone
for (size_t bone = 1; bone < mNumBones; bone++)
Matrix4 localMat; // Defaults to identity
if (mTracks[bone].size() > 0)
{
localMat = mTracks[bone][frame].ToMatrix();
outPoses[bone] = localMat * outPoses[bones[bone].mParent];
Note that because not every bone has a track,

GetGlobalPoseAtTime must first check that the bone
has a track. If it doesn’t, the local pose matrix for the
bone remains the identity matrix.
Next, you use the common pattern of creating a map for

your data and a corresponding get function that caches
the data in the map. This time, the map contains
Animation pointers, and you add it to Game.
Now you need to add functionality to the

SkeletalMeshComponent class. Recall that for each
bone, the matrix palette stores the inverse bind pose
matrix multiplied by the current pose matrix. Then when
calculating the position of a vertex with skinning, you use
this palette. Because the SkeletalMeshComponent
class tracks the current playback of an animation and has
access to the skeleton, it makes sense to store the palette
here. You first declare a simple struct for the
MatrixPalette, as follows:
const size_t MAX_SKELETON_BONES = 96;
struct MatrixPalette
Matrix4 mEntry[MAX_SKELETON_BONES];
};
You set a constant for the maximum number of bones to

96, but you could go as high as 256 because your bone
indices can range from 0 to 255.
You then add member variables to

SkeletalMeshComponent to track the current
animation, the play rate of the animation, the current
time in the animation, and the current matrix palette:
// Matrix palette
MatrixPalette mPalette;
// Animation currently playing
class Animation* mAnimation;
// Play rate of animation (1.0 is normal speed)
float mAnimPlayRate;
// Current time in the animation
float mAnimTime;
Next, you create a ComputeMatrixPalette function,

as shown in Listing 12.9, that grabs the global inverse
bind pose matrices as well as the global current pose
matrices. Then for each bone, you multiply these
matrices together, yielding the matrix palette entry.
Listing 12.9 ComputeMatrixPalette

Implementation
void SkeletalMeshComponent::ComputeMatrixPalette()
const std::vector<Matrix4>& globalInvBindPoses =
mSkeleton->GetGlobalInvBindPoses();
std::vector<Matrix4> currentPoses;
mAnimation->GetGlobalPoseAtTime(currentPoses, mSkeleton,
mAnimTime);
// Setup the palette for each bone
for (size_t i = 0; i < mSkeleton->GetNumBones(); i++)
{
// Global inverse bind pose matrix times current pose matrix
mPalette.mEntry[i] = globalInvBindPoses[i] * currentPoses[i];
Finally, you create a PlayAnimation function that

takes in an Animation pointer as well as the play rate of
the animation. This sets the new member variables, calls
ComputeMatrixPalette, and returns the duration of
the animation:
float SkeletalMeshComponent::PlayAnimation(const Animation* anim,
float playRate)
mAnimation = anim;
mAnimTime = 0.0f;
mAnimPlayRate = playRate;
if (!mAnimation) { return 0.0f; }
ComputeMatrixPalette();
return mAnimation->GetDuration();
}
Now you can load the animation data, compute the pose
matrices for frame 0 of the animation, and calculate the
matrix palette. However, the current pose of the
animation still won’t show up onscreen because the
vertex shader needs modification.
The Skinning Vertex Shader

Recall from Chapter 5 that the vertex shader
program’s responsibility is to transform a vertex
from object space into clip space. Thus, for skeletal
animation, you must update the vertex shader so that
it also accounts for bone influences and the current
pose. First, you add a new uniform declaration for
the matrix palette to Skinned.vert:
uniform mat4 uMatrixPalette[96];
Once the vertex shader has a matrix palette, you can then
apply the skinning calculations from earlier in the
chapter. Remember that because each vertex has up to
four different bone influences, you must calculate four
different positions and blend between them based on the
weight of each bone. You do this before transforming the
point into world space because the skinned vertex is still
in object space (just not in the bind pose).
Listing 12.10 shows the main function for the skinning
vertex shader program. Recall that inSkinBones and
inSkinWeights are the four bone indices and the four
bone weights. The accessors for x, y, and so on are simply
accessing the first bone, the second bone, and so on.
Once you calculate the interpolated skinned position of
the vertex, you transform the point to world space and
then projection space.
Listing 12.10 Skinned.vert Main Function

void main()
// Skin the position
vec4 skinnedPos = (pos * uMatrixPalette[inSkinBones.x]) * inSkinWeights.x;
skinnedPos += (pos * uMatrixPalette[inSkinBones.y]) * inSkinWeights.y;
skinnedPos += (pos * uMatrixPalette[inSkinBones.z]) * inSkinWeights.z;
skinnedPos += (pos * uMatrixPalette[inSkinBones.w]) * inSkinWeights.w;
// Transform position to world space

skinnedPos = skinnedPos * uWorldTransform;
// Save world position
fragWorldPos = skinnedPos.xyz;
// Transform to clip space
gl_Position = skinnedPos * uViewProj;
// Skin the vertex normal
vec4 skinnedNormal = vec4(inNormal, 0.0f);
skinnedNormal =
(skinnedNormal * uMatrixPalette[inSkinBones.x]) * inSkinWeights.x
+ (skinnedNormal * uMatrixPalette[inSkinBones.y]) * inSkinWeights.y
+ (skinnedNormal * uMatrixPalette[inSkinBones.z]) * inSkinWeights.z
+ (skinnedNormal * uMatrixPalette[inSkinBones.w]) * inSkinWeights.w;
// Transform normal into world space (w = 0)
fragNormal = (skinnedNormal * uWorldTransform).xyz;
Similarly, you also need to skin the vertex normals; if you

don’t, the lighting will not look correct as the character
animates.
Then, back in the C++ code for

SkeletalMeshComponent::Draw, you need to make
sure the SkeletalMeshComponent copies the matrix
palette data to the GPU with the following:
shader->SetMatrixUniforms("uMatrixPalette", &mPalette.mEntry[0],
MAX_SKELETON_BONES);
The SetMatrixUniforms function on the shader takes

in the name of the uniform, a pointer to a Matrix4, and
the number of matrices to upload.
You now have everything in place to draw the first frame

of an animation. Figure 12.5 shows the first frame of the
CatActionIdle.gpanim animation. This and other
animations in this chapter are also by Pior Oberson.
Figure 12.5 A character in the first frame of the
“action idle” animation
Updating Animations
The final step to get a working skeletal animation
system is to update the animation every frame, based
on delta time. You need to change the Animation
class so that it correctly gets the pose based on the
time in the animation, and you need to add an
Update function to SkeletalMeshComponent.
For the GetGlobalPoseAtTime function, in Listing

12.11, you can no longer hard-code it to only use frame 0
of the animation. Instead, based on the duration of each
frame and the current time, you figure out the frame
before the current time (frame) and the frame after the
current time (nextFrame). You then calculate a value
from 0.0 to 1.0 that specifies where exactly between the
two frames you are (pct). This way, you can account for
the animation and game frame rates being different.
Once you have this fractional value, you compute the
global poses mostly the same as before. However, now
instead of directly using a BoneTransform for a frame,
you interpolate between the bone transforms of frame
and nextFrame to figure out the correct in-between
pose.
Listing 12.11 Final Version of GetGlobalPoseAtTime

void Animation::GetGlobalPoseAtTime(std::vector<Matrix4>& outPoses,
const Skeleton* inSkeleton, float inTime) const
if (outPoses.size() != mNumBones)
outPoses.resize(mNumBones);
// Figure out the current frame index and next frame
// (This assumes inTime is bounded by [0, AnimDuration]
size_t frame = static_cast<size_t>(inTime / mFrameDuration);
size_t nextFrame = frame + 1;

// Calculate fractional value between frame and next frame
float pct = inTime / mFrameDuration - frame;
// Setup the pose for the root
if (mTracks[0].size() > 0)
// Interpolate between the current frame's pose and the next frame
BoneTransform interp = BoneTransform::Interpolate(mTracks[0][frame],
mTracks[0][nextFrame], pct);
outPoses[0] = interp.ToMatrix();
else
outPoses[0] = Matrix4::Identity;
const std::vector<Skeleton::Bone>& bones = inSkeleton->GetBones();
// Now setup the poses for the rest
for (size_t bone = 1; bone < mNumBones; bone++)
{
Matrix4 localMat; // (Defaults to identity)
if (mTracks[bone].size() > 0)
BoneTransform interp =
BoneTransform::Interpolate(mTracks[bone][frame],
mTracks[bone][nextFrame], pct);
localMat = interp.ToMatrix();
outPoses[bone] = localMat * outPoses[bones[bone].mParent];
Then, in SkeletalMeshComponent, you add an

Update function:
void SkeletalMeshComponent::Update(float deltaTime)
if (mAnimation && mSkeleton)
mAnimTime += deltaTime * mAnimPlayRate;

// Wrap around anim time if past duration
while (mAnimTime > mAnimation->GetDuration())
{ mAnimTime -= mAnimation->GetDuration(); }
// Recompute matrix palette
ComputeMatrixPalette();
Here, all you do is update mAnimTime based on delta

time and the animation play rate. You also wrap
mAnimTime around as the animation loops. This works
correctly even when transitioning from the last frame of
the animation to the first because, as mentioned earlier,
the animation data duplicates the first frame at the end
of the track.
Finally, Update calls ComputeMatrixPalette. This

function uses GetGlobalPoseAtTime to calculate the
new matrix palette for this frame.
Because SkeletalMeshComponent is a component,

the owning actor calls Update every frame. Then in the
“generate outputs” phase of the game loop, the
SkeletalMeshComponent draws with this new matrix
palette as usual, which means the animation now
updates onscreen!
GAME PROJECT
This chapter’s game project implements skeletal
animation as described in this chapter. It includes
the SkeletalMeshComponent, Animation, and
Skeleton classes, as well as the skinned vertex
shader. The code is available in the book’s GitHub
repository, in the Chapter12 directory. Open
Chapter12-windows.sln in Windows and
Chapter12-mac.xcodeproj on Mac.
This chapter’s game project goes back to the follow

camera discussed in Chapter 9, “Cameras,” to make the
character visible. The FollowActor class has a
SkeletalMeshComponent component, and it thus uses
the animation code. The player can use the WASD keys to
move the character around. When the character is
standing still, an idle animation plays. When the player
moves the character, a running animation plays (see
Figure 12.6). Currently, the transition between the two
animations is not smooth, but you will change that in
Exercise 12.2.
Figure 12.6 Character running through the game
world
SUMMARY
This chapter provides a comprehensive overview of
skeletal animation. In skeletal animation, a character
has a rigid skeleton that animates, and vertices act
like a skin that deforms with this skeleton. The
skeleton contains a hierarchy of bones, and every
bone except for the root has a parent bone.
The bind pose is the initial pose of the skeleton, prior to

any animations. You can store a local transform for each
bone in bind pose, which describes the position and
orientation of a bone relative to its parent. A global
transform instead describes the position and orientation
of a bone relative to object space. You can convert a local
transform into a global one by multiplying the local pose
by the global pose of its parent. The root bone’s local
pose and global pose are identical.
The inverse bind pose matrix is the inverse of each

bone’s global bind pose matrix. This matrix transforms a
point in object space while in bind pose into the bone’s
coordinate space while in bind pose.
An animation is a sequence of poses played back over

time. As with bind pose, you can construct a global pose
matrix for the current pose for each bone. These current
pose matrices can transform a point in a bone’s
coordinate space while in bind pose into object space for
the current pose.
The matrix palette stores the multiplication of the

inverse bind pose matrix and the current pose matrix for
each bone. When computing the object space position of
a skinned vertex, you use the matrix palette entries for
any bones that influence the vertex.
ADDITIONAL READING
Jason Gregory takes an in-depth look at more
advanced topics in animation systems, such as
blending animations, compressing animation data,
and inverse kinematics.

EXERCISES
In this chapter’s exercises you will add features to the
animation system. In Exercise 12.1 you add support
for getting the position of a bone in the current pose,
and in Exercise 12.2 you add blending when
transitioning between two animations.
Exercise 12.1
It’s useful for a game to get the position of a bone as
an animation plays. For example, if a character holds
an object in his hand, you need to know the position
of the bone as the animation changes. Otherwise, the
character will no longer hold the item properly!
Because the SkeletalMeshComponent knows the

progress in the animation, the code for this system needs
to go in here. First, add as a member variable a
std::vector to store the current pose matrices. Then,
when the code calls GetGlobalPoseAtTime, save the
current pose matrices in this member variable.
Next, add a function called GetBonePosition that

takes the name of a bone and returns the object space
position of the bone in the current pose. This is easier
than it sounds because if you multiply a zero vector by
the current pose matrix for a bone, you get the object
space position of that bone in the current pose. This
works because a zero vector here means it is exactly at
the origin of the bone’s local space, and then the current
pose matrix transforms it back to object space.
Exercise 12.2
Currently,
SkeletalMeshComponent::PlayAnimation
instantly switches to a new animation. This does not
look very polished, and you can address this issue by
adding blending to the animations. First, add an
optional blend time parameter to PlayAnimation,
which represents the duration of the blend. To blend
between multiple animations, you must track each
animation and animation time separately. If you
limit blending to only two animations, you just need
to duplicate those member variables.
Then, to blend between the animations, when you call

GetGlobalPoseAtTime, you need to do so for both
active animations. You need to get the bone transforms
of every bone for each animation, interpolate these bone
transforms to get the final transforms, and then convert
these to the pose matrices to get the blended current
pose.
CHAPTER 13
INTERMEDIATE
GRAPHICS
There are a multitude of different graphical

techniques used in games, which is why
there are entire volumes and book series on
the topic. This chapter explores a handful of
intermediate graphics concepts: how to
improve texture quality, rendering to
textures, and a different method for lighting
the scene, called deferred shading.
IMPROVING TEXTURE QUALITY
Recall from Chapter 5, “OpenGL,” that bilinear
filtering can improve the visual quality of a texture as
it gets larger on the screen. For example, suppose a
wall has a texture on it. As the player gets closer to
the wall, the size of the texture becomes larger
onscreen. Without bilinear filtering, the texture will
look pixelated. However, bilinear filtering makes the
image look smoother (although slightly blurry).
Also recall from Chapter 5 that images are just 2D grids

of pixels, and each of these “texture pixels” is called a
texel. Another way to look at the enlargement effect is
that as the wall texture becomes larger onscreen, the size
of every texel becomes larger onscreen. In other words,
the ratio between a texel from the texture and a pixel
onscreen decreases.
For example, if every 1 texel corresponds to 2 pixels

onscreen, then that ratio is 1:2. Texel density is this
ratio between pixels onscreen and texels. Ideally, you
want the texel density to be as close to 1:1 as possible. As
the density decreases, the image quality decreases.
Ultimately, the texture appears either too pixelated (if
using nearest-neighbor filtering) or too blurry (if using
bilinear filtering).
If the texel density becomes too high, this means that

each pixel onscreen corresponds to multiple texels in the
texture. For example, a 10:1 texel density means that
every pixel onscreen corresponds to 10 texels.
Ultimately, each of these pixels needs to choose a single
color to display. This means that the texture will appear
to have texels missing when viewed onscreen; this is
called a sampling artifact. In graphics, the term
artifact refers to a graphical glitch that’s a result of a
graphics algorithm.
Figure 13.1 illustrates the different graphical artifacts

caused by varying texel densities. Figure 13.1(a) shows a
star texture at a texel density of roughly 1:1, meaning the
texture appears onscreen with exactly the same ratio as
the original image file. Figure 13.1(b) shows part of the
star at a texel density of 1:5, which makes the edges
appear blurry. Finally, Figure 13.1(c) shows the texture
with a texel density of 5:1, which causes the edges of the
star to disappear; to make the image easier to see, the
figure illustrates it larger than the actual size.
Figure 13.1 Star texture bilinear filtered with

varying texel densities: (a) 1:1, (b) 1:5, (c) 5:1
Texture Sampling, Revisited

To understand why a high texel density causes texels
to appear to be missing, we need to look more closely
at how texture sampling works in general. Recall that
textures use UV coordinates (also called texture
coordinates) in the range of (0, 0) for the top-left
corner and (1, 1) for the bottom-right corner.
Suppose you have a texture that’s a 4×4 square of
texels. In this case, the UV coordinate for the center
of the top-left texel is (0.125, 0.125). Similarly, the
exact center of the texture corresponds to the UV
coordinate (0.5, 0.5), as in Figure 13.2(a).
Now suppose you have a texel density of 1:2, and you

draw the region of the texture from (0, 0) to (0.5, 0.5).
This means that the top one-fourth of the texture
appears at two times the size onscreen. When drawing
this in the fragment shader, each fragment (pixel) gets a
UV coordinate corresponding to the center of the pixel.
For example, the top-left pixel in Figure 13.2(b) is
sampling from the texture with a UV coordinate of
(0.0625, 0.0625). However, in the original image, no
texel’s center directly corresponds to this coordinate.
This is where a filtering algorithm comes in: It helps
select what color to draw for these in-between UV
coordinates.
In nearest-neighbor filtering, you simply select the

texel whose center is the closest to that UV coordinate.
So, because the top-left coordinate of (0.0625, 0.0625) is
closest to the white texel at (0.125, 0.125), nearest-
neighbor filtering selects white for that pixel. The result
of this is that every texel is resized proportionally to the
texel density, as in Figure 13.2(b). More plainly, in
nearest-neighbor filtering, increasing the size of the
texture onscreen increases the perceived size of each
texel, making the image look pixelated.
In bilinear filtering, you find the four texel centers

closest to a UV coordinate, and the sampled color at a UV
coordinate is the weighted average between these four
nearest texels. This yields a smoother transition as the
image magnifies, though the image will appear blurry if
magnified too much. Figure 13.2(c) illustrates bilinear
filtering. Notice that there are fewer neighboring pixels
with the same color, but instead the colors blend
together.
Figure 13.2 (a) Original texture; texture at 2x

magnification with (b) nearest-neighbor filtering and
(c) bilinear filtering
To understand how to calculate the weighted average in

bilinear filtering, remember that you can treat a color as
a 3D value and interpolate between colors the same way
you interpolate other values. You then decompose the
bilinear interpolation into the two separate axes’
interpolations. Consider a point P that’s nearest the four
texels A, B, C, and D, as in Figure 13.3. First, you
compute the interpolation between the colors at A and B
in the u direction and, similarly, the interpolation
between C and D in the u direction. This yields colors at
the two points R and S, shown in Figure 13.3. Finally,
you interpolate the colors at R and S in the v direction,
which yields the final color at P.
Figure 13.3 Bilinear interpolation of P relative to

texels A, B, C, and D
Given the texture coordinates for A, B, C, D, and P, you

can calculate this bilinear interpolation with the
following set of equations:
In these equations, uFactor determines the weighting in
the u component direction, and vFactor determines the
weighting in the v component direction. You then use
these weightings to first calculate the colors at R and S
and then, finally, the color at P.
These bilinear filtering calculations automatically occur

on the graphics card if the texture has bilinear filtering
enabled. And although this sounds like a lot of
calculations to do for every fragment that samples the
texture, modern graphics hardware can rapidly perform
millions of such calculations per second.
As you’ve seen, magnifying the texture too much causes

the image to appear either pixelated or blurry, depending
on the technique used. The issue with reducing the size
of the texture is that there aren’t enough texture samples
to maintain all the information stored in the texture.
Returning to the example texture, if you reduce the size
of the image by a factor of two, the filtering loses details
from the texture, as in Figure 13.4(b). You no longer see
the border, as in the original image. This example is
especially dramatic because you have only four pixels left
after the size reduction.
Figure 13.4 (a) Original texture; (b) texture bilinear
filtered to half size
Mipmapping
In mipmapping, rather than having a single source
texture, you generate a series of additional textures,
called mipmaps, that are at lower resolutions than
the source texture. For example, if the source texture
has a resolution of 256×256, you may generate
mipmaps of 128×128, 64×64, and 32×32. Then,
when it’s time to draw the texture onscreen, the
graphics hardware can select the mipmap texture
that yields a texel density closest to 1:1. While
mipmapping doesn’t improve texture quality when
you’re magnifying a texture to a resolution higher
than the original resolution, it greatly improves the
quality when you’re reducing the size of a texture.
The main reason for the quality improvement is that you

generate the mipmap textures only once—at the time the
texture is loaded. This means that you can use more
expensive algorithms that generate high-quality
mipmaps (such as using a box filter). Thus, sampling
from these high-quality mipmaps with a texel density
close to 1:1 will look much better than sampling from the
original texture with some higher texel density, such as
4:1.
Figure 13.5 illustrates sample mipmaps for the star

texture. The highest-resolution texture is the original
texture at 256×256, and the remaining textures are auto-
generated mipmaps. Note how even the smallest
mipmap maintains the border of the texture, which was
missing previously when you directly sampled from the
256×256 texture at a low texel density.
Figure 13.5 Star texture with three mipmap
textures
Much as texture sampling can use nearest-neighbor

filtering or bilinear filtering, there are two different
approaches to applying mipmaps. In nearest-neighbor
mipmapping, you simply select the mipmap that gives
the texel density closest to 1:1. Although this works well
in many cases, in some instances (such as with a floor
texture), it may cause banding at the borders where the
mipmap texture (or mip level) changes. In trilinear
filtering, you sample the two mip levels closest to a 1:1
texel density separately (with bilinear filtering), and the
final color is a blend between these two samples. This is
“trilinear” because it blends now in three dimensions—
the UV coordinates of the texture samples as well as the
mip-level blend.
Enabling mipmapping for a texture in OpenGL is

straightforward. After loading a texture with the code
from Chapter 5, you simply add a call to
glGenerateMipmap:
glGenerateMipmap(GL_TEXTURE_2D);
This automatically generates appropriate mip levels,

using a high-quality filtering algorithm.
When setting texture parameters, you can set both the

minimization filter (what happens when the texture
becomes smaller onscreen) and the magnification filter
(what happens when the texture becomes larger
onscreen). This is what the GL_TEXTURE_MIN_FILTER
and GL_ TEXTURE_MAG_FILTER parameters reference.
After you’ve generated the mipmaps, you then change

the texture parameter for the min filter to use
mipmapping. To do trilinear filtering, you use these
texture parameters:
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER,
GL_LINEAR_MIPMAP_LINEAR);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER,
GL_LINEAR);
Note that you still use GL_LINEAR as the filtering

function for magnification because mipmaps do not help
with texel density lower than 1:1. To instead use nearest-
neighbor mipmapping for minification, you would pass
in GL_LINEAR_MIPMAP_NEAREST as the final
parameter to the GL_TEXTURE_MIN_FILTER call.
Another advantage of mipmapping is that it improves the

rendering performance due to the way that texture
caching works. Much like a CPU cache, the graphics card
has a cache for its memory. Small mip levels are very
cache friendly, which means the overall rendering
performance increases.
Anisotropic Filtering
Although mipmapping greatly reduces sampling
artifacts in most instances, textures viewed at
oblique angles relative to the camera will appear very
blurry. This is noticeable especially with floor
textures, as shown in Figure 13.6(b). Anisotropic
filtering mitigates this by sampling additional
points on the texture when it is viewed at an oblique
angle. For example, 16x anisotropic filtering means
that there are 16 different samples for the texel color.
The graphics hardware performs the anisotropic

calculations, using a series of mathematical functions.
This chapter does not cover these functions, but you can
consult OpenGL Extensions Registry in the “Additional
Reading” section at the end of this chapter for more
information.
Although the newest specifications of OpenGL include

anisotropic filtering as a default feature, anisotropic
filtering is an extension in OpenGL 3.3. This means you
should verify that the graphics hardware supports
anisotropy before enabling the feature. For the most
part, this is academic because every graphics card made
in the past decade supports anisotropic filtering. But in
general, it is a good idea to test whether an OpenGL
extension is available before using said extension.
To turn on anisotropic filtering, you set the texture to use

mipmapping and then add the following lines of code:
if (GLEW_EXT_texture_filter_anisotropic)
// Get the maximum anisotropy value
GLfloat largest;
glGetFloatv(GL_MAX_TEXTURE_MAX_ANISOTROPY_EXT, &largest);
// Enable it
glTexParameterf(GL_TEXTURE_2D, GL_TEXTURE_MAX_ANISOTROPY_EXT,
largest);
}
This code tests whether anisotropic filtering is available,

and if it is, you ask OpenGL for the maximum anisotropy
value. Then you set the texture parameter to use
anisotropic filtering.
Figure 13.6 shows the ground from this chapter’s game

project with different settings. Figure 13.6(a) shows the
ground using only bilinear filtering; note how the ground
has many sampling artifacts on the edges of the bricks.
Figure 13.6(b) shows trilinear filtering enabled; this is an
improvement, but the distant ground is blurry. Finally,
Figure 13.6(c) shows both trilinear filtering and
anisotropic filtering enabled, which yields the best visual
quality of the three choices.
Figure 13.6 Viewing the ground with different

filtering methods: (a) bilinear filtering, (b) trilinear
filtering, and (c) trilinear and anisotropic filtering
RENDERING TO TEXTURES
To this point, you’ve always drawn polygons directly
to the color buffer. However, this color buffer isn’t
special; it’s just a 2D image that you write colors to at
specific coordinates. It turns out you can also draw
the scene to any arbitrary texture, or render-to-
texture. Although this may seem unnecessary, there
are many reasons you may want to render to a
texture.
For example, a racing game might have a car with a

rearview mirror. If you want the mirror to look accurate,
you might render the game world from the perspective of
the rearview mirror to a texture and then draw the
texture on the mirror in the scene. Furthermore, some
graphical techniques use textures as temporary storage
before computing the final output to the color buffer.
This section explores how to render to a texture and then

display this texture on the screen. This will require some
changes to the overall rendering code, which previously
assumed that everything writes directly to the color
buffer. You also need to add support for rendering the
scene from the perspectives of different cameras.
note
For high-quality reflections, such as for a large mirror, you must render the
scene from the perspective of the surface. However, if the game scene
contains many surfaces that need low-quality reflections, rendering the
scene from the perspective of each of these surfaces is too expensive. In this
case, you can instead generate a single reflection map of the entire scene.
Then, for every low-quality reflective surface, you sample from this reflection
map to give the illusion of a reflection. Although the quality is significantly
lower than when rendering from the perspective of the reflective surface, it is
sufficient for surfaces that only need low-quality reflections.
This book does not cover how to implement reflection maps, but you can
consult the “Additional Reading” section at the end of this chapter for further
information on the topic.
Creating the Texture

To render to a texture, you first need to create a
texture. You can add a new function to the Texture
class to support creating a texture for rendering. The
code for creating a texture, shown in Listing 13.1, is
like the code for creating textures from Chapter 5.
However, rather than assuming that you want an
RGBA format (which will result in 8 bits per
component and 32 bits per pixel), you use a
parameter to specify the format. Second, the texture
has no initial data, which is why the last parameter to
glTexImage2D is nullptr. If this last parameter is
nullptr, then the second and third-to-last
parameters are ignored. Finally, you purposefully do
not enable mipmapping or bilinear filtering on the
texture. You want the sampled data from the texture
to precisely match the actual output.
Listing 13.1 Creating a Texture for Rendering

void Texture::CreateForRendering(int width, int height,
unsigned int format)
{
mWidth = width;
mHeight = height;
// Create the texture id
// Set the image width/height with null initial data
glTexImage2D(GL_TEXTURE_2D, 0, format, mWidth, mHeight, 0, GL_RGB,
GL_FLOAT, nullptr);
// For a texture we'll render to, just use nearest neighbor
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
Creating a Framebuffer Object

Much the way that OpenGL uses a vertex array object
to contain all information about vertices (including
the vertex buffer, vertex format, and index buffer), a
framebuffer object (FBO) contains all information
about a framebuffer. The FBO includes any textures
associated with the framebuffer, an associated depth
buffer (if it exists), and other parameters. You can
then select which framebuffer to use for rendering.
OpenGL provides a default framebuffer object with
ID 0, which is the framebuffer that you’ve been
drawing to up to this point. However, you can also
create additional framebuffers and switch to other
framebuffers as needed.
For now, you will use a custom framebuffer object for a

rearview mirror that you display in the HUD onscreen.
First, you must add two new member variables to the
Renderer class:
// Framebuffer object for the mirror
unsigned int mMirrorBuffer;
// Texture for the mirror
class Texture* mMirrorTexture;
You store the ID of the framebuffer object you create in

mMirrorBuffer and the texture object associated with
the framebuffer in mMirrorTexture.
Next, you need a function that creates and configures the

mirror framebuffer object, as shown in Listing 13.2.
Several steps are necessary here. First,
glGenFrameBuffers creates the framebuffer object
and stores the ID in mMirrorBuffer. The
glBindFrameBuffer call then sets this framebuffer as
active. The next several lines of
CreateMirrorTexture create a depth buffer and
attach it to the current framebuffer object. This way,
when rendering for the mirror, you still have a depth
buffer to ensure that further objects appear behind closer
objects.
Then you create the mirror texture, with a width and

height one-quarter the size of the screen. You don’t use
the full screen size because you want the mirror to take
up only part of the screen. You request a GL_RGB format
for the texture because the mirror will contain the color
output of the scene from the perspective of the mirror.
Next, the glFramebufferTexture call associates the

mirror texture with the framebuffer object. Note how you
specify GL_COLOR_ATTACHMENT0 as the second
parameter. This says that the mirror texture corresponds
to the first color output of the fragment shader. Right
now, your fragment shader writes only one output, but as
you’ll see later in this chapter, it’s possible to write
multiple outputs from the fragment shader.
The glDrawBuffers call then says that for this

framebuffer object, you want to be able to draw to the
texture in the GL_COLOR_ATTACHMENT0 slot (which is
the mirror texture). Finally, the glCheckFrameBuffer
status call verifies that everything worked properly. If
there was an issue, you delete the framebuffer object and
mirror texture and return false.
Listing 13.2 Creating the Mirror Framebuffer

bool Renderer::CreateMirrorTarget()
int width = static_cast<int>(mScreenWidth) / 4;
int height = static_cast<int>(mScreenHeight) / 4;
// Generate a framebuffer for the mirror texture
glGenFramebuffers(1, &mMirrorBuffer);
glBindFramebuffer(GL_FRAMEBUFFER, mMirrorBuffer);
// Create the texture we'll use for rendering
mMirrorTexture = new Texture();
mMirrorTexture->CreateForRendering(width, height, GL_RGB);
// Add a depth buffer to this target
GLuint depthBuffer;
glGenRenderbuffers(1, &depthBuffer);
glBindRenderbuffer(GL_RENDERBUFFER, depthBuffer);
glRenderbufferStorage(GL_RENDERBUFFER, GL_DEPTH_COMPONENT, width, height
glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT,
GL_RENDERBUFFER, depthBuffer);

// Attach mirror texture as the output target for the framebuffer
glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0,
mMirrorTexture->GetTextureID(), 0);
// Set the list of buffers to draw to for this framebuffer
GLenum drawBuffers[] = { GL_COLOR_ATTACHMENT0 };
glDrawBuffers(1, drawBuffers);
// Make sure everything worked
if (glCheckFramebufferStatus(GL_FRAMEBUFFER) != GL_FRAMEBUFFER_COMPLETE)
// If it didn't work, delete the framebuffer,
// unload/delete the texture and return false
glDeleteFramebuffers(1, &mMirrorBuffer);
mMirrorTexture->Unload();
delete mMirrorTexture;
mMirrorTexture = nullptr;
return false;
}
return true;
In Renderer::Initialize, you add a call to

CreateMirrorTarget and verify that the function
returns true. Similarly, in Renderer::Shutdown, you
delete the mirror framebuffer and mirror textures (using
the same code that runs if the glCheckFrameBuffer
call says the framebuffer is not complete).
Rendering to a Framebuffer Object

To support a mirror, you need to render the 3D scene
twice: once from the perspective of the mirror and
once from the perspective of the normal camera.
Each time you render the scene is a render pass. To
assist with drawing the 3D scene multiple times, you
can create a Draw3DScene function, the skeleton of
which is in Listing 13.3.
The Draw3DScene function takes in the ID of the

framebuffer, the view matrix, the projection matrix, and
the scale of the viewport. The viewport size lets OpenGL
know the actual size of the framebuffer target that it’s
writing to. So, you need a viewport scale parameter here
so that the normal framebuffer can use the full screen
width and height, but the mirror can use its one-fourth
size. You use the glViewport call to set the viewport to
the correct size based on the screen width/height and the
scale.
The code for drawing meshes is the same as in Chapter 6,
“3D Graphics,” and the code for drawing skinned meshes
is the same as in Chapter 12, “Skeletal Animation.” Other
than the viewport code, the only other difference is that
before drawing anything, the glBindFramebuffer call
sets the active framebuffer to the requested one.
Listing 13.3 Renderer::Draw3DScene Helper

Function
void Renderer::Draw3DScene(unsigned int framebuffer,
const Matrix4& view, const Matrix4& proj,
float viewportScale)
// Set the current framebuffer
glBindFramebuffer(GL_FRAMEBUFFER, framebuffer);
// Set viewport size based on scale
glViewport(0, 0,
static_cast<int>(mScreenWidth * viewPortScale),
static_cast<int>(mScreenHeight * viewPortScale)
);

// Clear color buffer/depth buffer
glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
// Draw mesh components
// (Same code as Chapter 6)
// ...
// Draw any skinned meshes now
// ...
You then change the code in Renderer::Draw to call

Draw3DScene twice, as in Listing 13.4. First, you draw
using the mirror’s view and rendering to the mirror
framebuffer, and then you draw using the normal
camera’s view and rendering to the default framebuffer.
Finally, you draw the sprites and UI screens using the
code from Chapters 6 and 12.
Listing 13.4 Renderer::Draw Updated to Render

Both Mirror and Default Passes
void Renderer::Draw()
// Draw to the mirror texture first (viewport scale of 0.25)
Draw3DScene(mMirrorBuffer, mMirrorView, mProjection, 0.25f);
// Now draw the normal 3D scene to the default framebuffer
Draw3DScene(0, mView, mProjection);
// Draw all sprite components
// ...
// Draw any UI screens
// ...
// Swap the buffers
Here, mMirrorView is a separate view matrix for the

mirror. The specifics of the mirror view aren’t anything
new. You can create a MirrorCamera class that uses a
basic follow camera, as in Chapter 9, “Cameras.”
However, the mirror camera is in front of the character,
facing behind the character. This MirrorCamera then
attaches to the player actor and updates mMirrorView.
Drawing the Mirror Texture in the HUD

Now that the drawing code is writing to the mirror
texture, you can use it just like any other texture and
draw it onscreen. Because the mirror in this case is
just a HUD element, you can leverage the existing
DrawTexture functionality in UIScreen.
However, drawing with the existing code results in a

mirror that has a flipped y value from what is expected.
This is because, internally, OpenGL places the UV origin
at the bottom-left corner of the image instead of in the
top-left corner (as is more typical). Luckily, this is easy
enough to fix: When drawing the texture, you already
create a scale matrix. If you negate the y-axis of this scale
matrix, it will flip the texture in the y direction. To
support this, you add a new flipY bool as an optional
parameter to UIScreen::DrawTexture, as shown in
Listing 13.5. You default flipY to false because the
existing UI textures don’t need their y-axis flipped.
Listing 13.5 Adding a flipY Option to

UIScreen::DrawTexture
void UIScreen::DrawTexture(class Shader* shader, class Texture* texture,
const Vector2& offset, float scale, bool flipY)
// Scale the quad by the width/height of texture
// and flip the y if we need to
float yScale = static_cast<float>(texture->GetHeight()) * scale;
if (flipY) { yScale *= -1.0f; }
Matrix4 scaleMat = Matrix4::CreateScale(
static_cast<float>(texture->GetWidth()) * scale,
yScale,
1.0f);
// Translate to position on screen
Matrix4 transMat = Matrix4::CreateTranslation(
Vector3(offset.x, offset.y, 0.0f));
// Set world transform
Matrix4 world = scaleMat * transMat;

shader->SetMatrixUniform("uWorldTransform", world);
// Set current texture
texture->SetActive();
// Draw quad
Finally, you add two lines to HUD::Draw to display the

mirror texture in the bottom-left corner of the screen,
with a scale of 1.0 and flipY set to true:
Texture* mirror = mGame->GetRenderer()->GetMirrorTexture();
DrawTexture(shader, mirror, Vector2(-350.0f, -250.0f), 1.0f, true);
Figure 13.7 shows the mirror in action. Notice that the

main view shows the normal perspective, which faces in
the direction of the Feline Swordsman, but the mirror in
the bottom left renders the scene in the opposite
direction.
Figure 13.7 Game with a rearview mirror in the
bottom left
DEFERRED SHADING
Recall that the Phong lighting implemented in
Chapter 6 performs the lighting calculations for each
fragment when drawing a mesh. The pseudocode for
this type of lighting calculation is as follows:
foreach Mesh m in Scene
foreach Pixel p to draw from m
if p passes depth test
foreach Light li that effects p
color = Compute lighting equation(li, p)

Write color to framebuffer
This method of performing lighting calculations, called

forward rendering, works well with a small number of
lights. For example, the game currently has only one
directional light, so forward rendering works perfectly
fine. However, consider a game that takes place at night
in a city. For such a game, a single directional light won’t
yield a believable nightscape. Instead, you would want
dozens of lights for street lights, car headlights, lights
inside buildings, and so on. Unfortunately, forward
rendering doesn’t scale well in this case. You need to
compute lighting equations on the order of O(m · p · li),
which means adding several more lights increases the
amount of lighting calculations significantly.
An alternative approach is to create a series of textures,

collectively called the G-buffer, to store information
about the visible surfaces in the scene. This G-buffer
might contain the diffuse color (albedo), specular power,
and normals of visible surfaces in the scene. You then
render the scene in two passes. First, you go through
every mesh and render the properties of their surfaces to
the G-buffer. Then, in the second pass, you loop through
every light and compute the lighting equations based on
these lights and what is in the G-buffer. The following
pseudocode accomplishes this:
foreach Mesh m in Scene

foreach Pixel p1 to draw from m
if p passes depth test
Write surface properties of p1 to G-buffer
foreach Light li in the scene
foreach Pixel p2 affected by li
s = surface properties from the G-buffer at p2
color = Compute lighting equation (l, s)
Write color to framebuffer
Note how the complexity of this two-pass approach is

O(m · p1 + li · p2). This means that you can support far
more lights in the scene than with forward rendering.
Because there are two passes, and the shading of the
fragment onscreen doesn’t occur until the second pass,
this technique is called deferred shading (or
deferred rendering).
Implementing deferred shading requires several steps.

First, you must set up a framebuffer object that supports
multiple output textures. Then, you must create
fragment shaders that write surface properties to the G-
buffer. Next, you draw a quad that covers the entire
screen and samples from the G-buffer to output the
result of global lighting (such as directional and ambient
light). Finally, you calculate the lighting for each non-
global light (such as point lights or spotlights).
Creating a G-Buffer Class

Because the framebuffer object for the G-buffer is far
more complex than the one for the mirror in the
preceding section, it makes sense to encapsulate the
FBO and all its associated textures into a new
GBuffer class. Listing 13.6 shows the declaration of
GBuffer. You declare an enum that defines the
types of data stored in the different G-buffer
textures. The G-buffer in this chapter stores the
diffuse color, the normals, and the world position of
each surface.
note
Storing the world position in the G-buffer makes your later calculations
simpler—but at the expense of increased memory and rendering bandwidth
usage.
It’s possible to reconstruct the world position at a pixel from the depth buffer
and the view-projection matrix, which eliminates the need for the world
position in the G-buffer. Consult Phil Djonov’s article in the “Additional
Reading” section at the end of the chapter to learn how to do these
calculations.
One surface property missing from this G-buffer is the

specular power. This means you currently cannot
calculate the specular component of the Phong reflection
model; in Exercise 13.1 you will fix this.
In the member data for GBuffer, you store the

framebuffer object ID as well as a vector of the textures
that serve as render targets.
Listing 13.6 GBuffer Declaration
class GBuffer
public:
// Different types of data stored in the G-buffer
enum Type
EDiffuse = 0,
ENormal,
EWorldPos,
NUM_GBUFFER_TEXTURES
};
GBuffer();
~GBuffer();
// Create/destroy the G-buffer
bool Create(int width, int height);
void Destroy();

// Get the texture for a specific type of data
class Texture* GetTexture(Type type);
// Get the framebuffer object ID
unsigned int GetBufferID() const { return mBufferID; }
// Setup all the G-buffer textures for sampling
void SetTexturesActive();
private:
// Textures associated with G-buffer
std::vector<class Texture*> mTextures;
// Framebuffer object ID
unsigned int mBufferID;
};
For the member functions of GBuffer, most of the work

occurs in the Create function, which creates a G-buffer
of the specified width and height. Listing 13.7 gives the
truncated code for this function. The Create function
first creates a framebuffer object and adds a depth buffer
target, as was done in Listing 13.2.
Listing 13.7 GBuffer::Create Implementation

bool GBuffer::Create(int width, int height)
// Create the framebuffer object and save in mBufferID
// ...
// Add a depth buffer to this target
// ...
// Create textures for each output in the G-buffer
for (int i = 0; i < NUM_GBUFFER_TEXTURES; i++)
Texture* tex = new Texture();
// We want three 32-bit float components for each texture
tex->CreateForRendering(width, height, GL_RGB32F);
mTextures.emplace_back(tex);
// Attach this texture to a color output
glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0 + i,
tex->GetTextureID(), 0);
// Create a vector of the color attachments

std::vector<GLenum> attachments;
attachments.emplace_back(GL_COLOR_ATTACHMENT0 + i);
// Set the list of buffers to draw to
glDrawBuffers(static_cast<GLsizei>(attachments.size()),
attachments.data());
// Make sure everything worked
if (glCheckFramebufferStatus(GL_FRAMEBUFFER) !=
GL_FRAMEBUFFER_COMPLETE)
Destroy();
return false;
return true;
Next, you loop over each type of texture desired in the G-

buffer and create one Texture instance for each type of
data (because they are separate render targets). Note
that you request the GL_RGB32F format for each texture.
This means there are three components per texel, and
each of these components is a 32-bit single-precision
floating-point value. You then attach each texture to a
corresponding color attachment slot with the
glFramebufferTexture call. The code takes
advantage of the fact that the OpenGL definitions for the
color attachments are consecutive numbers.
note
Although GL_RGB32F yields a lot of precision for the values in the G-buffer,
the trade-off is that the G-buffer takes up a significant amount of graphics
memory. Three GL_RGB32F textures at a resolution of 1024×768 (your
screen resolution) takes up 27 MB of memory on the GPU. To reduce
memory usage, many games instead use GL_RGB16F (three half-precision
floats), which would cut the memory usage in half.
You could further optimize the memory usage with other tricks. For example,
because a normal is unit length, given the x and y components and the sign
of the z component, you can solve for the z component. This means you
could store the normals in GL_RG16F format (two half-precision floats) and
later derive the z component. In the interest of simplicity, this chapter does
not implement these optimizations, but you should know that many
commercial games use such tricks.
You then create a vector of all the different color

attachments and call glDrawBuffers to set the texture
attachments for the G-buffer. Finally, you validate that
creating the G-buffer succeeds. If it doesn’t, the
Destroy function deletes all associated textures and
destroys the framebuffer object.
Next, you add a GBuffer pointer to the member data of

Renderer:
class GBuffer* mGBuffer;
Then in Renderer::Initialize, you create the

GBuffer object and set it to the width/height of the
screen:
mGBuffer = new GBuffer();
int width = static_cast<int>(mScreenWidth);
int height = static_cast<int>(mScreenHeight);
if (!mGBuffer->Create(width, height))
SDL_Log("Failed to create G-buffer.");
return false;
In Renderer::Shutdown, you add code that calls the

Destroy member function on mGBuffer.
Writing to the G-buffer

Now that you have a G-buffer, you need to write data
into it. Recall that mesh rendering currently uses the
Phong fragment shader to write final (fully lit) colors
to the default framebuffer. However, this is
antithetical to the approach of deferred shading. You
need to create a new fragment shader that writes
surface properties into the G-buffer.
Another difference is that every previous fragment

shader wrote only a single output value. However,
fragment shaders can have multiple output values, or
multiple render targets. This means that writing to
each texture in the G-buffer is just a matter of writing to
each of the correct outputs. In fact, the GLSL code for the
main function of the fragment shader is relatively simple
compared to the code for fragment shaders you’ve seen
earlier in this book. You sample the diffuse color from
the texture and simply pass along the normal and world
position directly to the G-buffer.
Listing 13.8 gives the full GLSL code for

GBufferWrite.frag. Note that you declare three
different out values for the three different G-buffer
textures. You also specify layout locations for each of the
outputs; these numbers correspond to the color
attachment indices specified when creating the G-buffer.
Listing 13.8 GBufferWrite.frag Shader

#version 330
// Inputs from vertex shader
in vec2 fragTexCoord; // Tex coord
in vec3 fragNormal; // Normal (in world space)

in vec3 fragWorldPos; // Position (in world space)
// This corresponds to the outputs to the G-buffer
layout(location = 0) out vec3 outDiffuse;
layout(location = 1) out vec3 outNormal;
layout(location = 2) out vec3 outWorldPos;
// Diffuse texture sampler
void main()
// Diffuse color is from texture
outDiffuse = texture(uTexture, fragTexCoord).xyz;
// Pass normal/world position directly along
outNormal = fragNormal;
outWorldPos = fragWorldPos;
You then change the shader loading code for the

mMeshShader and mSkinnedShader to use
GBufferWrite.frag as the fragment shader, instead
of the previous Phong.frag.
Finally, in Renderer::Draw, you remove the call to

Draw3DScene, which draws to the default framebuffer.
You instead want to draw to the G-buffer:
Draw3DScene(mGBuffer->GetBufferID(), mView, mProjection, 1.0f, false);
The last Boolean parameter is new; it specifies that

Draw3DScene should not set any lighting constants on
the mesh shaders. This makes sense because the
GBufferWrite.frag shader doesn’t have any lighting
constants to set in the first place!
Running the game at this point would yield an entirely

black window other than the UI elements. This is
because although you’re writing surface properties to the
G-buffer, you aren’t drawing anything to the default
framebuffer based on these surface properties. However,
by using a graphics debugger such as RenderDoc (see the
sidebar “Graphics Debuggers”), you can view the output
to the different textures in the G-buffer. Figure 13.8
shows a visualization of the output to the different
components of the G-buffer, including the depth buffer.
Figure 13.8 Output to the different components of
the G-buffer
GRAPHICS DEBUGGERS
One difficulty with writing increasingly complex graphics code is that it is

more difficult to debug than normal C++ code. With C++ code, if there is
an issue, you can place a breakpoint and step through the execution of
the code. However, if the game does not show the correct graphics
output, it could be one of several issues. It might be that you’re calling
the wrong OpenGL functions, or the data passed to the shaders is
wrong, or the GLSL shader code is wrong.
This difficulty in determining the source of problems led to the creation of

graphics debuggers. There are several graphics debuggers available,
some of which are proprietary to specific types of graphics hardware or
consoles. At a minimum, these debuggers allow you to capture a frame
of graphics data and step through the commands executed to see how
the output to the framebuffer changes. They also allow you to view all
the data sent to the GPU, including vertex data, textures, and shader
constants. Some even allow you to step through the execution of a
vertex or pixel shader to see where it goes wrong.
For Windows and Linux, the best graphics debugger that supports
OpenGL is RenderDoc (https://renderdoc.org), an open source tool
created by Baldur Karlsson. In addition to OpenGL, it supports
debugging for Vulkan as well as Microsoft Direct3D 11 and 12 (the latter
two only on Windows). Unfortunately, at this writing, RenderDoc has no
macOS support.
For macOS users, Intel Graphics Performance Analyzers (GPA) is a
great alternative. See https://software.intel.com/en-us/gpa.
Global Lighting
Now that the game is writing surface properties to
the G-buffer, the next step is to use these properties
to display a fully lit scene. This section focuses on
global lights such as the ambient and a global
directional light. The basic premise is to draw a quad
the size of the screen to the default framebuffer. For
each fragment in this quad, you sample surface
properties from the G-buffer. Then, using these
surface properties, you can compute the same Phong
lighting equations from Chapter 6 to light the
fragment.
First, you create a vertex and fragment shader in GLSL

for global lighting from the G-buffer. Because you’re
ultimately drawing a quad to the screen, the vertex
shader is identical to the sprite vertex shader from
Chapter 5. The fragment shader, shown in Listing 13.9,
has some differences from the Phong fragment shader.
First, the only input from the vertex shader is the texture
coordinates. This is because the normal and world
positions at the fragment are in the G-buffer. Next, you
add three sampler2D uniforms for the three different
textures in the G-buffer (diffuse color, normal, and world
position). In the main function for the fragment shader,
you sample the diffuse color, normal, and world position
from the G-buffer textures. This, combined with the
directional light uniforms (as in Chapter 6), gives all the
information needed to light the fragment with the
ambient and diffuse components of the Phong reflection
model. You cannot calculate the specular component
because this depends on the specular power of each
surface, and you currently do not store the specular
information in the G-buffer. (In Exercise 13.1 you explore
adding the specular component.)
After calculating the Phong ambient and diffuse

component, you multiply the diffuse color of the surface
(from the G-buffer) to compute the final color at the
pixel.
Listing 13.9 GBufferGlobal.frag Shader

#version 330
// Inputs from vertex shader
in vec2 fragTexCoord; // Tex coord
layout(location = 0) out vec4 outColor;
// Different textures from G-buffer
uniform sampler2D uGDiffuse;
uniform sampler2D uGNormal;

uniform sampler2D uGWorldPos;
// Lighting uniforms (as in Chapter 6)
// ...
void main()
// Sample diffuse color, normal, world position from G-buffer
vec3 gbufferDiffuse = texture(uGDiffuse, fragTexCoord).xyz;
vec3 gbufferNorm = texture(uGNormal, fragTexCoord).xyz;
vec3 gbufferWorldPos = texture(uGWorldPos, fragTexCoord).xyz;
// Calculate Phong lighting (as in Chapter 6, minus specular)
// ...
// Final color is diffuse color times phong light
outColor = vec4(gbufferDiffuse * Phong, 1.0);
With the global lighting vertex and fragment shader code

written, the next step is to load these shaders in the
Renderer class. You create a Shader* member variable
called mGGlobalShader and instantiate it in the
LoadShader function. In this code, shown in Listing
13.10, you first load the vertex and fragment shader files.
Then, you set some of the uniforms for the shader.
The SetIntUniform calls associate each of the three

sampler2D uniforms in the fragment shader with a
specific texture index. The first SetMatrixUniform call
sets the view-projection matrix to be identical to the
sprite view-projection matrix (because you’re drawing a
quad). The second call sets the world transform to scale
the quad to the entire screen and invert the y-axis (to
solve the inverted y problem, as when drawing the
mirror texture to the screen).
Listing 13.10 Loading the G-buffer Global Lighting

Shader
mGGlobalShader = new Shader();
if (!mGGlobalShader->Load("Shaders/GBufferGlobal.vert",
"Shaders/GBufferGlobal.frag"))
return false;
// For the GBuffer, we need to associate each sampler with an index

mGGlobalShader->SetActive();
mGGlobalShader->SetIntUniform("uGDiffuse", 0);
mGGlobalShader->SetIntUniform("uGNormal", 1);
mGGlobalShader->SetIntUniform("uGWorldPos", 2);
// The view projection is just the sprite one
mGGlobalShader->SetMatrixUniform("uViewProj", spriteViewProj);
// The world transform scales to the screen and flips y
Matrix4 gbufferWorld = Matrix4::CreateScale(mScreenWidth,
-mScreenHeight, 1.0f);
mGGlobalShader->SetMatrixUniform("uWorldTransform", gbufferWorld);
Next, you add a function to the GBuffer class that binds

each texture in the G-buffer to a corresponding texture
index:
void GBuffer::SetTexturesActive()
mTextures[i]->SetActive(i);
}
Here, the SetActive function called on each texture

takes in an index, which corresponds to the indices set
on the sampler2D uniforms in GLSL.
The final step is to add a function to Renderer that

draws the G-buffer quad using the global lighting
shader. You create a new DrawFromGBuffer function,
as shown in Listing 13.11. Because the first step in
Renderer::Draw is now to draw the scene to the G-
buffer, DrawFromGBuffer is now the first code that
draws to the default framebuffer. You need to disable
depth testing for the quad, as you don’t want it to affect
the depth buffer. You then set the G-buffer shader and
sprite quad vertices as active and call the
SetTexturesActive function to activate all the G-
buffer textures. You then use the SetLightUniforms
function, created in Chapter 6, to set all the directional
light uniforms in the G-buffer shader. Finally, you draw
the quad, which invokes your G-buffer fragment shader
for every fragment onscreen.
Listing 13.11 Renderer::DrawFromGBuffer

Implementation
void Renderer::DrawFromGBuffer()
{
// Disable depth testing for the global lighting pass
glDisable(GL_DEPTH_TEST);
// Activate global G-buffer shader
mGGlobalShader->SetActive();
// Activate sprite verts quad
mSpriteVerts->SetActive();
// Set the G-buffer textures to sample
mGBuffer->SetTexturesActive();
// Set the lighting uniforms
SetLightUniforms(mGGlobalShader, mView);
// Draw the triangles for the quad
Next, you change the code at the start of

Renderer::Draw to first draw the 3D scene to the G-
buffer, change the framebuffer to the default and finally
call DrawFromGBuffer. After this, you render the
sprites and UI screens as before:

// Draw the 3D scene to the G-buffer
Draw3DScene(mGBuffer->GetBufferID(), mView, mProjection, false);
// Set the framebuffer back to zero (screen's framebuffer)
glBindFramebuffer(GL_FRAMEBUFFER, 0);
// Draw from the GBuffer
DrawFromGBuffer();
// Draw Sprite/UI as before
// ...
With the global lighting shader code in place, the

rendering code now draws the entire scene fully lit once
again. Figure 13.9 shows the scene output. Note that
because you are no longer calculating the specular
component of the Phong lighting equation, the scene
looks darker than before—even with a slightly higher
ambient light value than you had previously. However,
you can still see the entire scene, and other than the
darkness, it looks like the forward-rendered scene. Also,
note that the mirror still works properly, even though
you still use forward rendering for the mirror (and
because of the higher ambient light, the mirror looks
brighter than before).
Figure 13.9 Scene with global lights calculated via
deferred shading
Adding Point Lights

Recall that one of the main reasons to use deferred
shading is that it scales very well as the number of
lights in the scene increases. This section discusses
how to add support for many non-global lights.
Suppose the game has 100 different point lights. You

could create a uniform array in the shader that stores all
the information about these point lights, including
position, color, radius, and so on. Then in the
GBufferGlobal.frag shader code, you could loop
over these point lights. Using the G-buffer sampled
world position, you could figure out whether a fragment
is within range of a point light and, if so, compute the
Phong equation for this.
Although this approach could work, there are some
issues with it. You need to test every fragment against
every point light, even for lights that are nowhere near
the fragment. This means a lot of conditional checks in
the shader code, which is expensive.
The solution to these problems is to instead use light

geometry, or meshes that represent the lights. Because
a point light has a radius, its corresponding light
geometry is a sphere placed in the world. Drawing this
sphere will then trigger a fragment shader call for every
fragment the sphere touches. Using the world position
information from the G-buffer, you can compute the
intensity of the light to the fragment.
Adding a PointLightComponent Class

For point lights, you can create a component so that
it’s easy to attach to any actor to move the light. First,
you declare the PointLightComponent class, as
shown in Listing 13.12. For simplicity, you make its
member variables public. The diffuse color is simply
the diffuse color of the point light. The inner and
outer radius variables determine the area of
influence of the point light. The outer radius is the
maximum distance from which the point light affects
an object. The inner radius is the radius at which
the point light applies its full intensity of light.
Anything inside the inner radius has the full diffuse
color, while the color intensity falls off when
approaching the outer radius. The point light doesn’t
affect anything past the outer radius.
Listing 13.12 PointLightComponent Declaration
class PointLightComponent
public:
PointLightComponent(class Actor* owner);
~PointLightComponent();
// Draw this point light as geometry
void Draw(class Shader* shader, class Mesh* mesh);
// Diffuse color
Vector3 mDiffuseColor;
// Radius of light
float mInnerRadius;
float mOuterRadius;
};
You then add a vector of PointLightComponent

pointers to the Renderer class called mPointLights.
The constructor for PointLightComponent adds the
light to mPointLights, and the destructor removes the
light from the vector.
Point Light Fragment Shader

The next step is to create a
GBufferPointLight.frag fragment shader file.
As in the GBufferGlobal.frag shader, you need
to declare three different sampler2D uniforms for
the three different G-buffer textures. Unlike with the
global lighting shader, however, you need to store
information about a specific point light. You declare
a PointLight struct and add a uPointLight
uniform for this. You also add a uniform,
uScreenDimensions, that stores the width/height
of the screen:
// Additional uniforms for GBufferPointLight.frag
struct PointLight
// Position of light
vec3 mWorldPos;
// Diffuse color
vec3 mDiffuseColor;
// Radius of the light
float mInnerRadius;
float mOuterRadius;
};
uniform PointLight uPointLight;
// Stores width/height of screen
uniform vec2 uScreenDimensions;
The shader’s main function, given in Listing 13.13, is

different from the global light shader in several ways.
With the quad you drew for global lighting, you could
simply use the texture coordinates of the quad to sample
correctly into the G-buffer. However, using the texture
coordinates from the point light’s sphere mesh would not
yield a correct UV coordinate to sample into the G-
buffer. Instead, you can use gl_FragCoord, which is a
built-in GLSL variable that contains the position in
screen space of the fragment. In this instance, you only
care about the x and y coordinates. However, because UV
coordinates are in the range [0, 1], you need to divide the
screen space coordinates by the dimensions of the
screen. The division operator in this case is a
component-wise division.
Once you have the correct UV coordinates, you use them

to sample the diffuse, normal, and world positions from
the G-buffer. Next, you compute the N and L vectors,
much as in the previous Phong fragment shader.
Listing 13.13 GBufferPointLight.frag Main
Function
void main()
// Calculate the coordinate to sample into the G-buffer
vec2 gbufferCoord = gl_FragCoord.xy / uScreenDimensions;
// Sample from G-buffer
vec3 gbufferDiffuse = texture(uGDiffuse, gbufferCoord).xyz;
vec3 gbufferNorm = texture(uGNormal, gbufferCoord).xyz;
vec3 gbufferWorldPos = texture(uGWorldPos, gbufferCoord).xyz;
// Calculate normal and vector from surface to light
vec3 N = normalize(gbufferNorm);
vec3 L = normalize(uPointLight.mWorldPos - gbufferWorldPos);
// Compute Phong diffuse component for the light
vec3 Phong = vec3(0.0, 0.0, 0.0);
float NdotL = dot(N, L);
if (NdotL > 0)
{
// Get the distance between the light and the world pos
float dist = distance(uPointLight.mWorldPos, gbufferWorldPos);
// Use smoothstep to compute value in range [0,1]
// between inner/outer radius
float intensity = smoothstep(uPointLight.mInnerRadius,
uPointLight.mOuterRadius, dist);
// The diffuse color of the light depends on intensity
vec3 DiffuseColor = mix(uPointLight.mDiffuseColor,
vec3(0.0, 0.0, 0.0), intensity);
Phong = DiffuseColor * NdotL;
// Final color is texture color times phong light
outColor = vec4(gbufferDiffuse * Phong, 1.0);
However, when computing the diffuse color, you first

calculate the distance between the point light’s center
and the fragment’s world position. Then, the
smoothstep function calculates a value in the range [0,
1]. The function returns 0 for distances less than or equal
to the inner radius and 1 for distances greater than or
equal to the outer radius. Distances in between yield
some value in between. The smoothstep function uses
a Hermite function (a type of polynomial) to calculate
this in-between value. The resulting value corresponds to
an intensity of the diffuse light; the value 0 means full
intensity because the fragment is within the inner radius,
whereas the value 1 means the point light should not
affect the fragment.
You then compute the applied DiffuseColor based on

the intensity value. Here, the mix function performs a
linear interpolation between the point light’s diffuse
color and pure black. Remember that you do not
calculate the specular component of the Phong reflection
here because you currently do not have access to the
specular power in the G-buffer.
It’s important to understand that because point light

rendering occurs after the global light G-buffer
calculations, each fragment in the framebuffer already
has a color. You don’t want the point light shader to
overwrite the colors that are already there. For example,
if a fragment’s world position says it’s out of range of the
point light, the shader will return black. If you just set
the fragment to black, you lose all the color that was
already there from the global lighting pass.
Instead, you want to add the output of the point light

shader to whatever color is already there. Adding black
to the color does not change any of the RGB values,
which means it preserves the existing light. On the other
hand, if you add a green value, that makes the fragment
greener. Adding the output color to the existing color
doesn’t require any changes to the fragment shader code
itself; instead, you can do this on the C++ side of things.
Drawing Point Lights

You need to add some glue code to Renderer and
PointLightComponent before you can draw the
point lights in DrawFromGBuffer. First, you add a
new shader member variable called
mGPointLightShader. You then load this shader
in LoadShaders. For the vertex shader, you use the
BasicMesh.vert shader from Chapter 6 because
the point light’s sphere mesh doesn’t need any
special behavior. For the fragment shader, you use
the GBufferPointLight.frag shader.
As with the global lighting shader, you need to set the

uniforms for the different samplers to bind them to
specific G-buffer textures. You also set the
uScreenDimensions uniform to the width and height
of the screen.
You also add a mPointLightMesh member variable

that simply points to the mesh you want to use for the
point lights. You load the mesh when initializing the
Renderer and save it in the variable; the mesh in
question is a sphere.
Now you add additional code to DrawFromGBuffer,

shown in Listing 13.14. This code goes after all the code
that drew the full-screen quad that applied global
lighting. The first part of this code copies the depth
buffer from the G-buffer to the default framebuffer’s
depth buffer. Because you’re drawing the 3D scene to the
G-buffer, its depth buffer contains the actual depth
information for every fragment. Because you want to
depth test the point light spheres, you need to copy over
this information to the default depth buffer.
Listing 13.14 Drawing Point Lights in

Renderer::DrawFromGBuffer
// Copy depth buffer from G-buffer to default framebuffer
glBindFramebuffer(GL_READ_FRAMEBUFFER, mGBuffer->GetBufferID());
int width = static_cast<int>(mScreenWidth);
int height = static_cast<int>(mScreenHeight);
glBlitFramebuffer(0, 0, width, height,
0, 0, width, height,
GL_DEPTH_BUFFER_BIT, GL_NEAREST);
// Enable depth test, but disable writes to depth buffer
glDepthMask(GL_FALSE);

// Set the point light shader and mesh as active
mGPointLightShader->SetActive();
mPointLightMesh->GetVertexArray()->SetActive();
// Set the view-projection matrix
mGPointLightShader->SetMatrixUniform("uViewProj",
mView * mProjection);
// Set the G-buffer textures for sampling
mGBuffer->SetTexturesActive();
// The point light color should add to existing color
glEnable(GL_BLEND);
glBlendFunc(GL_ONE, GL_ONE);
// Draw the point lights
for (PointLightComponent* p : mPointLights)
p->Draw(mGPointLightShader, mPointLightMesh);
Next, you reenable the depth test (because you disabled

it when drawing the full-screen quad for the global
lighting), but you disable the depth mask. This means
that when you try to draw fragments for each point
light’s sphere, they need to pass the depth test, but these
fragments do not write new depth values to the depth
buffer. This ensures that the point light sphere meshes
do not interfere with the existing depth buffer values.
Because you’re disabling depth buffer writes here, you
add a corresponding call to the beginning of
Draw3DScene that reenables writes to the depth buffer.
(Otherwise, you can’t clear the depth buffer!)
Then you activate the shader for the point lights as well
as the corresponding point light mesh. You need to set
the view-projection matrix just as for any other object
rendered in the world to make sure the point light has
the correct location onscreen. You also need to bind the
G-buffer textures to their respective slots.
Because you want to add to the colors already in the

color buffer, you enable blending. The blend function
with GL_ONE as both parameters says that you just want
to directly add the two colors, without considering the
alpha values or any other parameters.
Finally, you loop over all the point lights and call the
Draw function on each point light. The code for
PointLightComponent::Draw, shown in Listing
13.15, doesn’t look that much different from the code for
drawing any other mesh. For the world transform
matrix, you need to scale based on the outer radius of the
light. You divide by the radius of the mesh because the
point light mesh does not have a unit radius. The
translation is just based on the position of the light,
which comes from the owning actor.
Furthermore, you need to set the different uniforms for

this specific point light, which isn’t different from how
you’ve set the uniforms before. Finally, the
glDrawElements call draws the light geometry for the
point light, which is your sphere mesh. You don’t need to
set the vertex array as active because the Renderer does
this before calling Draw.
Once you draw all the point light meshes, for every
fragment you calculate the contribution of the point light
to the color of the fragment. You then add this additional
light color to the already existing color from the global
lighting pass.
Listing 13.15 PointLightComponent::Draw

Implementation
void PointLightComponent::Draw(Shader* shader, Mesh* mesh)
// Scale world transform to the outer radius (divided by
// the mesh radius) and positioned to the world position
Matrix4 scale = Matrix4::CreateScale(mOwner->GetScale() *
mOuterRadius / mesh->GetRadius());
Matrix4 trans = Matrix4::CreateTranslation(mOwner->GetPosition());

Matrix4 worldTransform = scale * trans;
shader->SetMatrixUniform("uWorldTransform", worldTransform);
// Set point light shader constants
shader->SetVectorUniform("uPointLight.mWorldPos", mOwner->GetPosition()
shader->SetVectorUniform("uPointLight.mDiffuseColor", mDiffuseColor);
shader->SetFloatUniform("uPointLight.mInnerRadius", mInnerRadius);
shader->SetFloatUniform("uPointLight.mOuterRadius", mOuterRadius);
// Draw the sphere
glDrawElements(GL_TRIANGLES, mesh->GetVertexArray()->GetNumIndices(),
To demonstrate the point light rendering, this chapter’s

game project creates several point lights with different
colors along the floor. Figure 13.10 illustrates these point
lights, powered by deferred shading.
Figure 13.10 Numerous point lights in the game
project
Improvements and Issues

Although deferred shading is a very powerful
rendering technique used by many modern games, it
isn’t perfect. One problem is that it can’t handle
partially transparent objects such as windows.
Because the G-buffer can only store a single surface’s
properties, drawing such an object into the G-buffer
would overwrite the objects behind it. The solution
for this case is to draw transparent objects in a
separate pass after drawing the rest of the scene.
Also, for some types of games, the overhead of setting up

the G-buffer and rendering to multiple targets is not
worth it. If a game takes place largely during the day or
has a very small number of lights, the cost of the deferred
shading setup every frame might be higher than the cost
of a forward-rendering approach. Many virtual-reality
games, which need a very high frame rate, use forward
rendering for this reason.
Another issue is that the light geometry has many edge

cases to consider and fix. For example, if the point light
sphere partially intersects with a wall, the point light will
affect both sides of the wall in the current approach.
Also, if you create a very big point light but place the
camera inside the light, you don’t see the effect of that
light. To fix these light geometry issues, you need to use a
stencil buffer, which is a different type of output buffer.
GAME PROJECT
This chapter’s game project provides the full
implementation of deferred shading. In addition, it
uses both mipmapping and anisotropic aliasing to
improve texture quality. The project includes the
mirror texture that’s forward rendered. The code is
There are no changes to the controls or character from

the previous chapter. The player still uses the WASD keys
to move the character around. To demonstrate the point
lights, several point lights are provided in
Game::LoadData.
SUMMARY
This chapter covers a handful of intermediate
graphics techniques. First, it looks at how texture
filtering works—both nearest-neighbor filtering and
bilinear filtering. Mipmapping can reduce sampling
artifacts when reducing the size of textures because it
generates several lower-resolution textures.
However, for oblique surfaces, mipmapping may
appear blurry. In this case, anisotropic filtering
improves the quality of the textures.
Another powerful technique is rendering the scene to

textures. OpenGL allows creation of arbitrary
framebuffer objects associated with textures. Then, you
can choose to draw the 3D scene to this texture. One use
of this technique is to draw a high-quality reflection,
such as for a mirror.
Finally, this chapter explores deferred shading, which is

a two-pass approach to lighting. In the first pass, you
write the object’s surface properties, such as diffuse
color, normals, and world position, into a G-buffer. In
the second pass, you read from the G-buffer to calculate
lighting equations. For lights with limited ranges, such as
point lights, you render lighting geometry to ensure that
the light affects only the fragments in range. Deferred
shading is an excellent approach when there are many
lights in the scene, though there are some issues, such as
the inability to handle partially transparent objects.
ADDITIONAL READING
As mentioned in Chapter 6, Real-Time Rendering by
Thomas Akenine-Moller et al. is the go-to book when
it comes to rendering techniques and games. Jason
Zink et al. give a good overview of many techniques,
including deferred shading, even though the book
focuses on Direct3D 11 instead of OpenGL. Matt
Pharr et al. cover physically based rendering, which
is a newer technique games use to achieve more
realistic lighting. Wolfgang Engel’s books are always
on the cutting edge of what graphics programmers in
the video game industry are using. Phil Djonov
discusses how to eliminate the need for the world
position in the G-buffer. Finally, sometimes to
understand how various OpenGL extensions work,
you need to read the official registry.

Djonov, Phil. “Deferred Shading Tricks.”
Shiny Pixels. Accessed November 26, 2017.
http://vec3.ca/code/graphics/deferred-
shading-tricks/.
Engel, Wolfgang, ed. GPU Zen: Advanced
Rendering Techniques. Encinitas: Black Cat
Publishing, 2017.
Khronos Group. OpenGL Extensions
Registry. Accessed October 16, 2017.
https://github.com/KhronosGroup/OpenGL-
Registry.
Pharr, Matt, Wenzel Jakob, and Greg
Humphreys. Physically Based Rendering:
From Theory to Implementation, 3rd
edition. Cambridge: Elsevier, 2017.
Zink, Jason, Matt Pettineo, and Jack
Hoxley. Practical Rendering and
Computation with Direct3D 11. Boca
Raton: CRC Press, 2012.
EXERCISES
In this chapter’s exercises, you explore improving the
deferred shading techniques covered in the latter half
of the chapter.
Exercise 13.1
Add support for the specular component to both the
global G-buffer lighting (the directional light) and
the point lights. To do this, first you need a new
texture in the G-buffer that stores the specular power
of the surface. Add this new texture to the relevant
parts of code (both in C++ and in GLSL).
Next, change the PointLightComponent class, the

PointLightComponent::Draw function, and the
shader code for point lights and the global light. For the
point lights, use the intensity to interpolate the specular
color, as is done for the diffuse color. Calculate the
specular component according to the Phong equations,
as before.
Exercise 13.2
Adding a new type of light to deferred shading
requires a new type of light geometry. Add support
for spotlights. To do so, you need to create a
SpotLightComponent as well as a corresponding
shader to draw these lights after the point lights.
Use the provided SpotLight.gpmesh file (which is a

cone) as the mesh for spotlights. A spotlight should have
parameters like those of a point light, but it also needs a
variable for its angle. To be able to change the angle, the
mesh needs to also scale non-uniformly. The default
mesh has a half angle of 30 degrees.
CHAPTER 14
LEVEL FILES AND

BINARY DATA
This chapter explores how to load and save

JSON-based level files representing the game
world. These level files store global
properties as well as properties of all the
actors and components in the game.
In addition, this chapter explores the trade-
offs of using text-based file formats versus
binary file formats. As an example, it
discusses an implementation of a binary
mesh file format.
LEVEL FILE LOADING
To this point, this book hasn’t used a data-driven
approach to the placement of objects in the game
world. Instead, the Game::LoadData function code
dictates the actors and components in the game, as
well as global properties, such as the ambient light.
The current approach has several disadvantages,
most notably that even small changes, such as
placement of a cube in a level, requires recompilation
of the source code. A designer who wants to change
the placement of objects in the game shouldn’t have
to change the C++ source code.
The solution is to create a separate data file for the level.

This data file should be able to specify which actors the
level contains and which properties and, optionally,
adjust the components of these actors. This level file
should also include any needed global properties.
For a 2D game, using a basic text file works perfectly

well. You can simply define different ASCII characters
for different objects in the world and create a text grid of
these objects. This makes the level file look like ASCII
art. Unfortunately, this approach doesn’t work very well
for a 3D game because each object in the game world
could be at some arbitrary 3D coordinate. Furthermore,
in the game object model used in this book, actors can
have components, so you may need to also save
properties of each attached component.
For all the reasons just listed, you need a file format
that’s more structured. As with the rest of the book, in
this chapter you once again use a text-based JSON
format for data. However, this chapter also explores the
trade-offs that any text format makes, as well as
techniques needed for binary file formats.
This section explores building up a JSON-level file

format. You start with global properties and slowly add
additional features to the file so that the
Game::LoadData function has barely any code other
than a function call that specifies the level file to load.
Unlike earlier chapters, this chapter explores the usage
of the RapidJSON library to parse in the JSON file.
Loading Global Properties

The only global properties the game world really has
are the lighting properties—the ambient light and the
global directional light. With such a limited number
of properties, this is a good starting point for
defining the JSON level file format. Listing 14.1
shows how you might specify the global lighting
properties in the level file.
Listing 14.1 Level with Global Lighting Properties

(Level0.gplevel)
{
"version": 1,
"globalProperties": {
"ambientLight": [0.2, 0.2, 0.2],
"directionalLight": {
"direction": [0.0, -0.707, -0.707],
"color": [0.78, 0.88, 1.0]
Listing 14.1 shows several constructs that you commonly

encounter in a level file. First, at its core, a JSON
document is a dictionary of key/value pairs (or
properties) called a JSON object. The key name is in
quotes, and then the value follows the colon. Values can
be of several types. The basic types are strings, numbers,
and Booleans. The complex types are arrays and JSON
objects. For this file, the globalProperties key
corresponds to a JSON object. This JSON object then has
two keys: one for the ambient light and one for the
directional light. The ambientLight key corresponds to
an array of three numbers. Similarly, the
directionalLight key corresponds to another JSON
object, with two additional keys.
This nesting of JSON objects and properties drives the
implementation of the parsing code. Specifically, you can
see common operations where, given a JSON object and
a key name, you want to read in a value. And in your C++
code, the types you have are far more varied than the
JSON format, so you should add code to assist with
parsing.
To parse these global properties in code, you begin by

declaring a LevelLoader class. Because loading the
level from a file affects the state of the game, but not the
level loader itself, you declare the LoadLevel function
as a static function, as follows:
class LevelLoader
public:
// Load the level -- returns true if successful
static bool LoadLevel(class Game* game, const std::string& fileName);
};
Note that in addition to the filename, the LoadLevel

function takes in the pointer to the Game object. This is
necessary because creating or modifying anything
requires access to the game.
The first step in LoadLevel is to load and parse the

level file into a rapidjson::Document. The most
efficient approach is to first load the entire file into
memory and then pass this buffer to the Parse member
function of the Document. Because loading a JSON file
into a Document is a common operation, it makes sense
to create a helper function. This way, gpmesh, gpanim,
and any other asset types that need to load in a JSON file
can also reuse this function.
Listing 14.2 shows the implementation of LoadJSON.

This function is also a static function. It takes in the
filename and a reference to the output document. The
first step loads the file into an ifstream. Note that you
load the file in binary mode instead of text mode. This is
for efficiency purposes because all you need to do is load
the entire file into a character buffer (array) and pass
that buffer directly to RapidJSON. You also use the
std::ios::ate flag to specify that the stream should
start at the end of the file.
If the file loads successfully, you use the tellg function

to get the current position of the file stream. Because the
stream is at the end of the file, this corresponds to the
size of the entire file. Next, you have the seekg call set
the stream back to the beginning of the file. You then
create a vector with enough space to fit the entire file
plus a null terminator and have the read function read
the file into the vector. Finally, you call the Parse
function on outDoc to parse the JSON file.
Listing 14.2 LevelLoader::LoadJSON
Implementation
bool LevelLoader::LoadJSON(const std::string& fileName,
rapidjson::Document& outDoc)
// Load the file from disk into an ifstream in binary mode,
// loaded with stream buffer at the end (ate)
std::ifstream file(fileName, std::ios::in |
std::ios::binary | std::ios::ate);
if (!file.is_open())
SDL_Log("File %s not found", fileName.c_str());
return false;
// Get the size of the file
std::ifstream::pos_type fileSize = file.tellg();
// Seek back to start of file
file.seekg(0, std::ios::beg);

// Create a vector of size + 1 (for null terminator)
std::vector<char> bytes(static_cast<size_t>(fileSize) + 1);
// Read in bytes into vector
file.read(bytes.data(), static_cast<size_t>(fileSize));
// Load raw data into RapidJSON document
outDoc.Parse(bytes.data());
if (!outDoc.IsObject())
SDL_Log("File %s is not valid JSON", fileName.c_str());
return false;
return true;
You then call LoadJSON at the start of LoadLevel:
rapidjson::Document doc;
if (!LoadJSON(fileName, doc))
{
SDL_Log("Failed to load level %s", fileName.c_str());
return false;
Given a JSON object, you need to read in keys and

extract their corresponding values. You shouldn’t assume
that a given key will always be there, so you should first
validate that the key exists and matches the expected
type. If it does, you read in the value. You can implement
this behavior in another class with a static function
called JsonHelper. Listing 14.3 shows the
JsonHelper::GetInt function. It tries to find the
property, validates that it matches the expected type, and
then returns true if successful.
Listing 14.3 JsonHelper::GetInt Implementation

bool JsonHelper::GetInt(const rapidjson::Value& inObject,
const char* inProperty, int& outInt)
// Check if this property exists
auto itr = inObject.FindMember(inProperty);
if (itr == inObject.MemberEnd())
{
return false;
// Get the value type, and check it's an integer
auto& property = itr->value;
if (!property.IsInt())
return false;
// We have the property
outInt = property.GetInt();
return true;
You can then use the GetInt function in LoadLevel to

validate that the loaded file’s version matches the
expected version:
int version = 0;
if (!JsonHelper::GetInt(doc, "version", version) ||
version != LevelVersion)
SDL_Log("Incorrect level file version for %s", fileName.c_str());
return false;
Here, the JSON object in question is the overall

document (the root JSON object). You first make sure
that GetInt returns a value and, if it does, you check
that its value matches the expected value (a const called
LevelVersion).
You also add similar functions to JsonHelper to extract

other basic types: GetFloat, GetBool, and
GetString. However, where this paradigm really
becomes powerful is for non-basic types. Specifically,
many properties in this game are of type Vector3 (such
as ambientLight), so having a GetVector3 function
is very useful. The overall construction of the function is
still the same, except you need to validate that the
property is an array with three members that are floats.
You can similarly declare a GetQuaternion function.
Ambient and Directional Lights

With the helper functions in place, you can create a
function to load in the global properties. Because the
global properties are varied and may not necessarily
need the same class types, you must manually query
the specific properties you need. The
LoadGlobalProperties function in Listing 14.4
demonstrates how to load the ambient light and
directional light properties. Notice that for the most
part, you call the helper functions you’ve created for
these properties.
Note that you can access a property as a

rapidjson::Value& directly through operator[].
The dirObj["directionalLight"] call gets the
value with the directionalLight key name, and then
the IsObject() call validates that the type of the value
is a JSON object.
Another interesting pattern for the directional light is to

have direct access to variables you want to set. In this
case, you do not need to add any conditional checks on
the GetVector3 calls. This is because if the property
requested does not exist, the Get functions guarantee
not to change the variable. If you have direct access to a
variable and don’t care if the property is unset, then this
reduces the amount of code.
Listing 14.4
LevelLoader::LoadGlobalProperties
Implementation
void LevelLoader::LoadGlobalProperties(Game* game,

const rapidjson::Value& inObject)
// Get ambient light
Vector3 ambient;
if (JsonHelper::GetVector3(inObject, "ambientLight", ambient))
game->GetRenderer()->SetAmbientLight(ambient);
// Get directional light
const rapidjson::Value& dirObj = inObject["directionalLight"];
if (dirObj.IsObject())
DirectionalLight& light = game->GetRenderer()->GetDirectionalLight();
// Set direction/color, if they exist
JsonHelper::GetVector3(dirObj, "direction", light.mDirection);
JsonHelper::GetVector3(dirObj, "color", light.mDiffuseColor);
}
You then add a call to LoadGlobalProperties in
LoadLevel, immediately after the validation code for
the level file version:
// Handle any global properties
const rapidjson::Value& globals = doc["globalProperties"];
if (globals.IsObject())
LoadGlobalProperties(game, globals);
You can then add a call to LoadLevel in

Game::LoadData, which loads in the
Level0.gplevel file:
LevelLoader::LoadLevel(this, "Assets/Level0.gplevel");
Because you’re now loading in the light properties from

the level file, you can also remove the code in LoadData
that hard-coded the ambient light and directional light.
Loading Actors
Loading in the actors means the JSON file needs an
array of actors, and each actor has property
information for that actor. However, you need some
way to specify which type of Actor you need
(because there are subclasses). In addition, you want
to avoid having a long set of conditional checks in the
level loading code to determine which Actor
subclass to allocate.
As before, it helps to first visualize what the data might

look like. Listing 14.5 shows one method to specify the
actors in the JSON file. This example only shows actors
of type TargetActor, but the type can easily specify any
other Actor subclass. Note that in addition to the type
are any other properties to specify for that actor. Here,
the only properties set are position and rotation, but
these could conceivably be any property the actor has.
Listing 14.5 Level with Actors (Level1.gplevel)

// Version and global properties
// ...
"actors": [
"type": "TargetActor",
"properties": {
"position": [1450.0, 0.0, 100.0]
},
"properties": {
"position": [0.0, -1450.0, 200.0],
"rotation": [0.0, 0.0, 0.7071, 0.7071]
},
"properties": {
"position": [0.0, 1450.0, 200.0],
"rotation": [0.0, 0.0, -0.7071, 0.7071]
}
Assuming for a moment that you have a method to
construct an actor of a specific type, you also need to be
able to load properties for the actor. The simplest
approach is to create a virtual LoadProperties
function in the base Actor class, shown in Listing 14.6.
Listing 14.6 Actor::LoadProperties Function

void Actor::LoadProperties(const rapidjson::Value& inObj)
// Use strings for different states
std::string state;
if (JsonHelper::GetString(inObj, "state", state))
if (state == "active")
SetState(EActive);
else if (state == "paused")
SetState(EPaused);
}
else if (state == "dead")
SetState(EDead);
// Load position, rotation, and scale, and compute transform
JsonHelper::GetVector3(inObj, "position", mPosition);
JsonHelper::GetQuaternion(inObj, "rotation", mRotation);
JsonHelper::GetFloat(inObj, "scale", mScale);
Then, for some subclass of Actor, you can override the

LoadProperties function to load any additional
properties, as needed:
void SomeActor::LoadProperties(const rapidjson::Value& inObj)
// Load base actor properties
Actor::LoadProperties(inObj);

// Load any of my custom properties
// ...
Now that you have a way to load properties, the next step
is to solve the issue of constructing an actor of the correct
type. One approach is to create a map where the key is
the string name of the actor type, and the value is a
function that can dynamically allocate an actor of that
type. The key is straightforward because it’s just a string.
For the value, you can make a static function that
dynamically allocates an actor of a specific type. To avoid
having to declare a separate function in each subclass of
Actor, you can instead create a template function like
this in the base Actor class:
template <typename T>
static Actor* Create(class Game* game, const rapidjson::Value& inObj)
// Dynamically allocate actor of type T
T* t = new T(game);
// Call LoadProperties on new actor
t->LoadProperties(inObj);
return t;
Because it’s templated on a type, it can dynamically

allocate an object of the specified type and then call
LoadProperties to set any parameters of the actor
type, as needed.
Then, back in LevelLoader, you need to create the

map. The key type is std::string, but for the value,
you need a function that matches the signature of the
Actor::Create function. For this, you can once again
use the std::function helper class to define the
signature.
First, you use an alias declaration (which is like a

typedef) to create an ActorFunc type specifier:
using ActorFunc = std::function<
class Actor*(class Game*, const rapidjson::Value&)
>;
The template parameters to std::function specify

that the function returns an Actor* and takes in two
parameters: Game* and rapidjson::Value&.
Next, you declare the map as a static variable in

LevelLoader:
static std::unordered_map<std::string, ActorFunc> sActorFactoryMap;
Then in LevelLoader.cpp, you construct the

sActorFactoryMap to fill in the different actors you
can create:
std::unordered_map<std::string, ActorFunc> LevelLoader::sActorFactoryMap
{ "Actor", &Actor::Create<Actor> },
{ "BallActor", &Actor::Create<BallActor> },
{ "FollowActor", &Actor::Create<FollowActor> },
{ "PlaneActor", &Actor::Create<PlaneActor> },
{ "TargetActor", &Actor::Create<TargetActor> },
};
This initialization syntax sets up entries in the map with

a key as the specified string name and the value as the
address of an Actor::Create function, templated to
create the specific type of Actor subclass. Note that you
don’t call the various create functions here. Instead, you
just get the memory address of a function and save it for
later use.
With the map set up, you can now create a LoadActors
function, as in Listing 14.7. Here, you loop over the
actors array in the JSON file and get the type string for
the actor. You use this type to then look up in
sActorFactoryMap. If you find the type, you call the
function stored as the value in the map (iter-
>second), which in turn calls the correct version of
Actor::Create. If you don’t find the type, you have a
helpful debug log message output.
Listing 14.7 LevelLoader::LoadActors

Implementation
void LevelLoader::LoadActors(Game* game, const rapidjson::Value& inArray)
// Loop through array of actors
for (rapidjson::SizeType i = 0; i < inArray.Size(); i++)
const rapidjson::Value& actorObj = inArray[i];
if (actorObj.IsObject())
// Get the type
std::string type;
if (JsonHelper::GetString(actorObj, "type", type))
// Is this type in the map?
auto iter = sActorFactoryMap.find(type);
if (iter != sActorFactoryMap.end())
// Construct with function stored in map
Actor* actor = iter->second(game, actorObj["properties"]);
else
SDL_Log("Unknown actor type %s", type.c_str());
You then add a call to LoadActors inside LoadLevel,

immediately after loading in the global properties:

const rapidjson::Value& actors = doc["actors"];
if (actors.IsArray())
LoadActors(game, actors);
With this code, you’re now loading in actors and setting

their properties. However, you are not yet able to adjust
properties of components nor add additional
components in the level file.
Loading Components
Loading data for components involves many of the
same patterns as for actors. However, there is one
key difference. Listing 14.8 shows a snippet of the
declaration of two different actors with their
components property set. The base Actor type
does not have any existing components attached to
it. So in this case, the MeshComponent type means
that you must construct a new MeshComponent for
the actor. However, the TargetActor type already
has a MeshComponent, as one is created in the
constructor for TargetActor. In this case, the
properties specified should update the existing
component rather than create a new one. This means
the code for loading components needs to handle
both cases.
Listing 14.8 Actors with Components in JSON (Excerpt

from the Full File)
"actors": [
"type": "Actor",
"properties": {
"position": [0.0, 0.0, 0.0],
"scale": 5.0
},
"components": [
"type": "MeshComponent",
"properties": { "meshFile": "Assets/Sphere.gpmesh" }
},
"properties": { "position": [1450.0, 0.0, 100.0] },
"components": [
"type": "MeshComponent",
"properties": { "meshFile": "Assets/Sphere.gpmesh" }
To determine whether Actor already has a component

of a specific type, you need a way to search through an
actor’s component vector by type. While you might be
able to use the built-in type information in C++, it’s more
common for game programmers to use their own type
information (and disable the built-in functionality). This
is mainly because of the well-documented downsides of
the built-in C++ Run-type type information (RTTI) not
obeying the “you only pay for what you use” rule.
There are many ways to implement your own type

information; this chapter shows a simple approach. First,
you declare a TypeID enum in the Component class,
like so:
enum TypeID
TComponent = 0,
TAudioComponent,
TBallMove,
// Other types omitted
// ...
NUM_COMPONENT_TYPES
};
Then, you add a virtual function called GetType that

simply returns the correct TypeID based on the
component. For example, the implementation of
MeshComponent::GetType is as follows:
TypeID GetType() const override { return TMeshComponent; }
Next, you add a GetComponentOfType function to

Actor that loops through the mComponents vector and
returns the first component that matches the type:
Component* GetComponentOfType(Component::TypeID type)

{
Component* comp = nullptr;
for (Component* c : mComponents)
if (c->GetType() == type)
comp = c;
break;
return comp;
The disadvantage of this approach is that every time you

create a new Component subclass, you must remember
to add an entry to the TypeID enum and implement the
GetType function. You could automate this somewhat
by using macros or templates, but the code here does not
do so for the sake of readability and understanding.
Note that this system also assumes that you won’t have
multiple components of the same type attached to one
actor. If you wanted to have multiple components of the
same type, then GetComponentOfType would
potentially have to return a collection of components
rather than just a single pointer.
Also, the type information does not give inheritance

information; you can’t figure out
SkeletalMeshComponent is a subclass of
MeshComponent, as GetType for
SkeletalMeshComponent just returns
TSkeletalMeshComponent. To support inheritance
information, you would need an approach that saves
some hierarchy information as well.
With the basic type system in place, you can move on to

more familiar steps. As with Actor, you need to create a
virtual LoadProperties function in the base
Component class and then override it for any subclasses,
as needed. The implementations in the various
subclasses are not necessarily straightforward. Listing
14.9 shows the implementation of LoadProperties for
MeshComponent. Recall that MeshComponent has an
mMesh member variable that’s a pointer to the vertex
data to draw. You don’t want to specify the vertex
directly in the JSON file; instead, you want to reference
the gpmesh file. The code first checks for the meshFile
property and then gets the corresponding mesh from the
renderer.
Listing 14.9 MeshComponent::LoadProperties

Implementation
void MeshComponent::LoadProperties(const rapidjson::Value& inObj)
Component::LoadProperties(inObj);
std::string meshFile;
if (JsonHelper::GetString(inObj, "meshFile", meshFile))
SetMesh(mOwner->GetGame()->GetRenderer()->GetMesh(meshFile));
int idx;
if (JsonHelper::GetInt(inObj, "textureIndex", idx))
mTextureIndex = static_cast<size_t>(idx);
JsonHelper::GetBool(inObj, "visible", mVisible);
JsonHelper::GetBool(inObj, "isSkeletal", mIsSkeletal);
}
The next step is to add a static templated Create
function for Component, which is very similar to the one
in Actor except that the parameters are different. (It
takes in Actor* as the first parameter instead of
Game*.)
You then need a map in LevelLoader. You use

std::function again to create a helper type called
ComponentFunc:
using ComponentFunc = std::function<
class Component*(class Actor*, const rapidjson::Value&)
>;
Then, you declare the map. However, unlike with the

sActorFactoryMap, which has only a single value, in
this case, you need a pair of values. The first element in
the pair is an integer corresponding to the TypeID of the
component, and the second element is the
ComponentFunc:
static std::unordered_map<std::string,
std::pair<int, ComponentFunc>> sComponentFactoryMap;
Then, in LevelLoader.cpp, you instantiate the

sComponentFactoryMap:
std::unordered_map<std::string, std::pair<int, ComponentFunc>>
LevelLoader::sComponentFactoryMap
{ "AudioComponent",
{ Component::TAudioComponent, &Component::Create<AudioComponent>}
},
{ "BallMove",
{ Component::TBallMove, &Component::Create<BallMove> }
},
// Other components omitted
// ...
};
You then implement a LoadComponents helper

function in LevelLoader, as shown in Listing 14.10. As
with LoadActors, it takes in an array of the
components to load and loops through this array. You
then use the sComponentFactoryMap to find the
component type. If it is found, you then check if the actor
already has a component of the type. The iter->
second.first accesses the first element of the value
pair, which corresponds to the type ID. If the actor
doesn’t already have a component of the requested type,
then you create one by using the function stored in the
second element of the value pair (iter-
>second.second). If the component already exists, you
can then just directly call LoadProperties on it.
Listing 14.10 LevelLoader::LoadComponents

Implementation
void LevelLoader::LoadComponents(Actor* actor,
const rapidjson::Value& inArray)
// Loop through array of components
for (rapidjson::SizeType i = 0; i < inArray.Size(); i++)
const rapidjson::Value& compObj = inArray[i];
if (compObj.IsObject())
// Get the type
std::string type;
if (JsonHelper::GetString(compObj, "type", type))
{
auto iter = sComponentFactoryMap.find(type);
if (iter != sComponentFactoryMap.end())
// Get the typeid of component
Component::TypeID tid = static_cast<Component::TypeID>
(iter->second.first);
// Does the actor already have a component of this type?
Component* comp = actor->GetComponentOfType(tid);
if (comp == nullptr)
// It's a new component, call function from map
comp = iter->second.second(actor, compObj["properties"]);
else
// It already exists, just load properties
comp->LoadProperties(compObj["properties"]);
else
{
SDL_Log("Unknown component type %s", type.c_str());
Finally, you add code in LoadActors that accesses the

components property, if it exists, and calls
LoadComponents on it:
// Construct with function stored in map
Actor* actor = iter->second(game, actorObj["properties"]);
// Get the actor's components
if (actorObj.HasMember("components"))
const rapidjson::Value& components = actorObj["components"];
if (components.IsArray())
LoadComponents(actor, components);
}
With all this code in place, you can now load the entire
level from a file, including the global properties, actors,
and any components associated with each actor.
SAVING LEVEL FILES

Saving to a level file is conceptually simpler than
loading from a file. First, you write the global
properties for the level. Then, you loop through every
actor in the game and every component attached to
every actor. For each of these, you need to write out
the relevant properties.
The implementation details are bit involved because the

RapidJSON interface is slightly more complicated for
creating JSON files than for reading in files. However,
overall you can use techniques like those used for loading
the level file.
First, you create helper Add functions in JsonHelper so

that you can quickly add additional properties to an
existing JSON object. For example, the AddInt function
has the following syntax:
void JsonHelper::AddInt(rapidjson::Document::AllocatorType& alloc,

rapidjson::Value& inObject, const char* name, int value)
rapidjson::Value v(value);
inObject.AddMember(rapidjson::StringRef(name), v, alloc);
The last three parameters are identical to the parameters

of the GetInt function, except the Value is now not
const. The first parameter is an allocator that
RapidJSON uses when needing to allocate memory.
Every call to AddMember requires an allocator, so you
must pass one in. You can get the default allocator just
from a Document object, but you could conceivably use a
different allocator if desired. You then create a Value
object to encapsulate the integer and use the AddMember
function to add a value with the specified name to
inObject.
The rest of the Add functions are similar, except for

AddVector3 and AddQuaternion, for which you must
first create an array and then add float values to that
array. (You’ll see this array syntax when looking at the
global properties.)
You then create a skeleton for the

LevelLoader::SaveLevel function, as shown in
Listing 14.11. First, you create the RapidJSON document
and make an object for its root via SetObject. Next,
you add the version integer. Then, you use the
StringBuffer and PrettyWriter to create a pretty-
printed output string of the JSON file. Finally, you use a
standard std::ofstream to write out the string to a
file.
Listing 14.11 LevelLoader::SaveLevel

Implementation
void LevelLoader::SaveLevel(Game* game,
const std::string& fileName)
// Create the document and root object
rapidjson::Document doc;
doc.SetObject();
// Write the version
JsonHelper::AddInt(doc.GetAllocator(), doc, "version", LevelVersion);
// Create the rest of the file (TODO)
// ...
// Save JSON to string buffer

rapidjson::StringBuffer buffer;
// Use PrettyWriter for pretty output (otherwise use Writer)
rapidjson::PrettyWriter<rapidjson::StringBuffer> writer(buffer);
doc.Accept(writer);
const char* output = buffer.GetString();
// Write output to file
std::ofstream outFile(fileName);
if (outFile.is_open())
outFile << output;
For now, this function only writes out the version to the
output file. But with this skeleton code, you can start
adding the remaining output.
Saving Global Properties

Next, you need to add a SaveGlobalProperties
function to LevelLoader. We omit the
implementation here, as it’s very similar to the other
functions written thus far. You simply need to add
the properties for the ambient light and the
directional light object.
Once this function is complete, you integrate it into your

SaveLevel function as follows:
rapidjson::Value globals(rapidjson::kObjectType);
SaveGlobalProperties(doc.GetAllocator(), game, globals);
doc.AddMember("globalProperties", globals, doc.GetAllocator());
Saving Actors and Components

To be able to save actors and components, you need a
way to get a string name of the type, given an Actor
or Component pointer. You already have a TypeID
for components, so to get a corresponding string, you
need only declare a constant array of the different
names in Component. You declare this array in
Component.h as follows:
static const char* TypeNames[NUM_COMPONENT_TYPES];
And then in Component.cpp, you fill in the array. It’s

important that you maintain the same ordering as the
TypeID enum:
const char* Component::TypeNames[NUM_COMPONENT_TYPES] = {
"Component",
"AudioComponent",
"BallMove",
// Rest omitted
// ...
};
By maintaining the ordering, you make it easy to get the

name of a component, given the type, using a snippet like
this:
Component* comp = /* points to something */;
const char* name = Component::TypeNames[comp->GetType()];
To do the same thing for the Actor and its subclasses,

you need to add a TypeID enum to Actor as well. This
is essentially the same as the code for TypeIDs in
components earlier in this chapter, so we omit it here.
You then need to create a virtual SaveProperties

function in both Actor and Component and then
override it in every subclass that needs to do so. This
ends up playing out very similarly to the
LoadProperties functions written when loading in the
level files. As an example, Listing 14.12 shows the
implementation of Actor::SaveProperties. Note
that you liberally use the Add functions in
LevelLoader, and you need to pass in the allocator
because all the Add functions need it.
Listing 14.12 Actor::SaveProperties

Implementation
void Actor::SaveProperties(rapidjson::Document::AllocatorType& alloc,
rapidjson::Value& inObj) const
std::string state = "active";
if (mState == EPaused)
state = "paused";
else if (mState == EDead)
state = "dead";
JsonHelper::AddString(alloc, inObj, "state", state);
JsonHelper::AddVector3(alloc, inObj, "position", mPosition);
JsonHelper::AddQuaternion(alloc, inObj, "rotation", mRotation);
JsonHelper::AddFloat(alloc, inObj, "scale", mScale);
With all these pieces in place, you can then add

SaveActors and SaveComponents functions to
LevelLoader. Listing 14.13 shows the SaveActors
function. First, you get the vector of actors from the
game by const reference. Then, you loop through every
actor and create a new JSON object for it. You then add
the string for the type by using the TypeID and
TypeNames functionality. Next, you create a JSON
object for the properties and call the actor’s
SaveProperties function. You then create an array for
the components before calling SaveComponents.
Finally, you add the actor’s JSON object into the JSON
array of actors.
Listing 14.13 LevelLoader::SaveActors

Implementation
void LevelLoader::SaveActors(rapidjson::Document::AllocatorType& alloc,
Game* game, rapidjson::Value& inArray)

{
const auto& actors = game->GetActors();
for (const Actor* actor : actors)
// Make a JSON object
rapidjson::Value obj(rapidjson::kObjectType);
// Add type
AddString(alloc, obj, "type", Actor::TypeNames[actor->GetType()]);
// Make object for properties
rapidjson::Value props(rapidjson::kObjectType);
// Save properties
actor->SaveProperties(alloc, props);
// Add the properties to the JSON object
obj.AddMember("properties", props, alloc);
// Save components
rapidjson::Value components(rapidjson::kArrayType);
SaveComponents(alloc, actor, components);
obj.AddMember("components", components, alloc);

// Add actor to inArray
inArray.PushBack(obj, alloc);
You similarly implement a SaveComponents function.

With all this code implemented, you can now save all the
actors and components to the file. For testing purposes,
pressing the R key in this chapter’s game project saves to
the Assets/Save.gplevel level file.
note
With some work, you could create a single serialize function that both load
and saves properties. This way, you could avoid having to update two
different functions every time you add a new property to an actor or a
component.
While this code will save almost everything in the game,

it doesn’t quite fully capture the current state of the game
at a specific point in time. For example, it does not save
the state of any active FMOD sound events. To
implement this, you would need to ask FMOD for the
current timestamp of the sound events, and then when
loading the game from the file, you would need to restart
the sound events with those timestamps. It takes some
additional work to go from saving a level file to being
usable as a save file for the player.
BINARY DATA
You’ve used JSON file formats throughout this book:
for meshes, animations, skeletons, text localization,
and now for level loading. The advantages of using a
text-based file format are numerous. Text files are
easy for humans to look at, find errors in, and (if
needed) manually edit. Text files also play very nicely
with source control systems such as Git because it’s
very easy to see what changed in a file between two
revisions. During development, it’s also easier to
debug loading of assets if they are text files.
However, the disadvantage of using text-based file

formats is that they are inefficient, both in terms of disk
and memory usage as well as in terms of performance at
runtime. Formats such as JSON or XML take up a lot of
space on disk simply because of the formatting
characters they use, such as braces and quotation marks.
On top of this, parsing text-based files at runtime is slow,
even with high-performance libraries such as
RapidJSON. For example, on my computer, it takes
about three seconds to load in the
CatWarrior.gpmesh file in a debug build. Clearly, this
would lead to slow load times for a larger game.
For the best of both worlds, you may want to use text
files during development (at least for some members of
the team) and then binary files in optimized builds. This
section explores how to create a binary mesh file format.
To keep things simple, in the code that loads in the
gpmesh JSON format, you will first check if a
corresponding gpmesh.bin file exists. If it does, you’ll
load that in instead of the JSON file. If it doesn’t exist,
the game will create the binary version file so that next
time you run the game, you can load the binary version
instead of the text version.
Note that one potential downside of this approach is that

it may lead to bugs that occur only with the binary
format but not the text one. To avoid this, it’s important
that you continue to use both formats throughout
development. If one of the two formats becomes stale,
then there’s a greater chance that format will stop
working.
Saving a Binary Mesh File

With any binary file format, an important step is to
decide on a layout for the file. Most binary files begin
with some sort of header that defines the contents
of the file as well as any specific size information
that’s needed to read in the rest of the file. In the case
of a mesh file format, you want the header to store
information about the version, the number of
vertices and indices, and so on. Listing 14.14 shows
the MeshBinHeader struct that defines the layout of
the header. In this example, the header is not
packed (reduced in size as much as possible), but it
gives the general idea of what you might want to
store in a header.
Listing 14.14 MeshBinHeader Struct

struct MeshBinHeader
// Signature for file type
char mSignature[4] = { 'G', 'M', 'S', 'H' };
// Version
uint32_t mVersion = BinaryVersion;
// Vertex layout type
VertexArray::Layout mLayout = VertexArray::PosNormTex;
// Info about how many of each you have
uint32_t mNumTextures = 0;
uint32_t mNumVerts = 0;
uint32_t mNumIndices = 0;
// Box/radius of mesh, used for collision
AABB mBox{ Vector3::Zero, Vector3::Zero };
float mRadius = 0.0f;
};
The mSignature field is a special 4-byte magic number

that specifies the file type. Most popular binary file types
have some sort of signature. The signature helps you
figure out what a file type is from its first few bytes
without knowing anything other than the signature to
look for. The rest of the data is information you need to
reconstruct the mesh data from the file.
After the header is the main data section of the file. In

this case, there are three main things to store: the
filenames for associated textures, the vertex buffer data,
and the index buffer data.
With the file format decided on, you can then create the
SaveBinary function, as shown in Listing 14.15. This
function takes in a lot of parameters because there’s a lot
of information needed to create the binary file. In total,
you need the filename, a pointer to the vertex buffer, the
number of vertices, the layout of these vertices, a pointer
to the index buffer, the number of indices, a vector of the
texture names, the bounding box of the mesh, and the
radius of the mesh. With all these parameters, you can
save the file.
Listing 14.15 Mesh::SaveBinary Implementation

void Mesh::SaveBinary(const std::string& fileName, const void* verts,
uint32_t numVerts, VertexArray::Layout,
const uint32_t* indices, uint32_t numIndices,
const std::vector<std::string>& textureNames,
const AABB& box, float radius)
{
// Create header struct
MeshBinHeader header;
header.mLayout = layout;
header.mNumTextures =
static_cast<unsigned>(textureNames.size());
header.mNumVerts = numVerts;
header.mNumIndices = numIndices;
header.mBox = box;
header.mRadius = radius;
// Open binary file for writing
std::ofstream outFile(fileName, std::ios::out
| std::ios::binary);
if (outFile.is_open())
// Write the header
outFile.write(reinterpret_cast<char*>(&header), sizeof(header));
// For each texture, we need to write the size of the name,
// followed by the string, followed by a null terminator

for (const auto& tex : textureNames)
uint16_t nameSize = static_cast<uint16_t>(tex.length()) + 1;
outFile.write(reinterpret_cast<char*>(&nameSize),
sizeof(nameSize));
outFile.write(tex.c_str(), nameSize - 1);
outFile.write("\0", 1);
// Figure out number of bytes for each vertex, based on layout
unsigned vertexSize = VertexArray::GetVertexSize(layout);
// Write vertices
outFile.write(reinterpret_cast<const char*>(verts),
numVerts * vertexSize);
// Write indices
outFile.write(reinterpret_cast<const char*>(indices),
numIndices * sizeof(uint32_t));
}
The code in Listing 14.15 does quite a lot. First, you
create an instance of the MeshBinHeader struct and fill
in all its members. Next, you create a file for output and
open it in binary mode. If this file successfully opens, you
can write to it.
Then you write the header of the file with the write
function call. The first parameter write expects is a
char pointer, so in many cases it’s necessary to cast a
different pointer to a char*. This requires a
reinterpret_cast because a MeshBinHeader*
cannot directly convert to a char*. The second
parameter to write is the number of bytes to write to the
file. Here, you use sizeof to specify the number of bytes
corresponding to the size of MeshBinHeader. In other
words, you are writing sizeof(header) bytes starting
at the address of header. This is a quick way to just
write the entire struct in one fell swoop.
warning
WATCH OUT FOR ENDIANNESS: The order in which a CPU platform saves
values larger than 1 byte is called endianness. The method used here to
read and write MeshBinHeader will not work if the endianness of the
platform that writes out the gpmesh.bin file is different from the endianness
of the platform that reads the gpmesh.bin file.
Although most platforms today are little endian, endianness can still be a
potential issue with code of this style.
Next, you loop through all the texture names and write
each of them to the file. For each filename, you first write
the number of characters in the filename (plus one for
the null terminator) and then write the string itself. Note
that this code assumes that a filename can’t be larger
than 64 KB, which should be a safe assumption. The
reason you write the number of characters and the name
is for loading. The header only stores the number of
textures and not the size of each string. Without storing
the number of characters, at load time you would have
no way of knowing how many bytes to read for the
filename.
After writing all the filenames, you then write all the
vertex and index buffer data directly to the file. You don’t
need to include the sizes here because they already
appear in the header. For the vertex data, the number of
bytes is the number of vertices times the size of each
vertex. Luckily, you can use a VertexArray helper
function to get the size of each vertex based on layout.
For the index data, you have a fixed size (32-bit indices),
so the total number of bytes is easier to calculate.
Then in Mesh::Load, if the binary file doesn’t exist, the

code loads the JSON file and creates the corresponding
binary file.
Loading a Binary Mesh File

Loading a binary mesh file is like writing to it but in
reverse. The steps are to load in the header, check the
validity of the header, load in the textures, load in the
vertex and index data, and finally create the actual
VertexArray (which will upload the data to the
GPU via OpenGL). Listing 14.16 shows the outline of
the code for Mesh::LoadBinary.
Listing 14.16 Mesh::LoadBinary Outline
void Mesh::LoadBinary(const std::string& filename,
Renderer* renderer)
std::ifstream inFile(fileName, /* in/binary flags ... */);
if (inFile.is_open())
MeshBinHeader header;
inFile.read(reinterpret_cast<char*>(&header), sizeof(header));
// Validate the header signature and version
char* sig = header.mSignature;
if (sig[0] != 'G' || sig[1] != 'M' || sig[2] != 'S' ||
sig[3] != 'H' || header.mVersion != BinaryVersion)
return false;
// Read in the texture file names (omitted)

// ...
// Read in vertices/indices
unsigned vertexSize = VertexArray::GetVertexSize(header.mLayout);
char* verts = new char[header.mNumVerts * vertexSize];
uint32_t* indices = new uint32_t[header.mNumIndices];
inFile.read(verts, header.mNumVerts * vertexSize);
inFile.read(reinterpret_cast<char*>(indices),
header.mNumIndices * sizeof(uint32_t));
// Now create the vertex array
mVertexArray = new VertexArray(verts, header.mNumVerts,
header.mLayout, indices, header.mNumIndices);
// Delete verts/indices
delete[] verts;
delete[] indices;
mBox = header.mBox;
mRadius = header.mRadius;

return true;
return false;
First, you open the file for reading in binary mode. Next,
you read in the header via the read function. Much as
with write, read takes in a char* for where to write
and the number of bytes to read from the file. Next, you
verify that the signature and version in the header match
what is expected; if they don’t, you can’t load the file.
After this, you read in all the texture filenames and load
them, though we omit that code from Listing 14.16 to
save space. Next, you allocate memory to store the vertex
and index buffers, and you use read to grab the data
from the file. Once you have the vertex and index data,
you can construct the VertexArray object and pass in
all the information it needs. You need to make sure to
clean up the memory and set the mBox and mRadius
members before returning.
Note that LoadBinary returns false if the file fails to

load. This way, the Mesh::Load code first tries to load
the binary file. If it succeeds, that’s it. Otherwise, it can
proceed using the JSON parsing code from before:
bool Mesh::Load(const std::string& fileName, Renderer* renderer)
mFileName = fileName;
// Try loading the binary file first
if (LoadBinary(fileName + ".bin", renderer))
return true;
// ...
With the switch to binary mesh file loading, the

performance improves significantly in debug mode. The
CatWarrior.gpmesh.bin file now loads in one second
as opposed to three—meaning a 3x performance gain
over the JSON version! This is great because you’ll spend
most of your development time running in debug mode.
Unfortunately, in an optimized build, the performance of

both the JSON and binary path is almost identical. This
could be due to several factors, including the RapidJSON
library being very optimized or other aspects being the
primary overhead, such as transferring the data to the
GPU or loading in the textures.
On the disk space side of things, you save space. While

the JSON version of the Feline Swordsman is around 6.5
MB on disk, the binary version is only 2.5 MB.
GAME PROJECT
This chapter’s game project implements the systems
discussed in this chapter. Everything loads from a
gplevel file, and pressing the R key saves the
current state of the world into
Assets/Saved.gplevel. The project also
implements the binary saving and loading of mesh
files in the .gpmesh.bin format. The code is
Figure 14.1 shows the game project in action. Notice that

it looks identical to the game project from Chapter 13,
“Intermediate Graphics.” However, the entire contents of
the game world now load directly from the
Assets/Level3.gplevel file, which was in turn
created by saving the level file. The first time the game
runs, it creates a binary mesh file for every mesh loaded.
Subsequent runs load meshes from the binary files
instead of JSON.
SUMMARY
This chapter explores how to create level files in
JSON. Loading from a file requires several systems.
First, you create helper functions that wrap the
functionality of the RapidJSON library to easily be
able to write the game’s types to JSON. You then add
code to set global properties, load in actors, and load
in components associated with the actors. To do this,
you need to add some type information to
components, as well as maps that associate names of
types to a function that can dynamically allocate that
type. You also need to create virtual
LoadProperties functions in both Component
and Actor.
You also need to create code to save the game world to
JSON, and you create helper functions to assist with this
process. At a high level, saving the file requires saving all
the global properties first and then looping through all
the actors and components to write their properties. As
with file loading, you have to create virtual
SaveProperties functions in both Component and
Actor.
Finally, this chapter discusses the trade-offs involved in

using a text-based file format instead of a binary one.
While a text format is often more convenient to use in
development, it comes at a cost of inefficiency—both in
performance and disk usage. This chapter explores how
to design a binary file format for mesh files, which
involves writing and reading from files in binary mode.
ADDITIONAL READING
There are no books devoted specifically to level files
or binary data. However, the classic Game
Programming Gems series has some articles on the
topic. Bruno Sousa’s article discusses how to use
resource files, which are files that combine several
files into one. Martin Brownlow’s article discusses
how to create a save-anywhere system. Finally, David
Koenig’s article looks at how to improve the
performance of loading files.
Brownlow, Martin. “Save Me Now!” Game

Programming Gems 3. Ed. Dante Treglia.
Hingham: Charles River Media, 2002.
Koenig, David L. “Faster File Loading with
Access Based File Reordering.” Game
Programming Gems 6. Ed. Mike
Dickheiser. Rockland: Charles River Media,
2006.
Sousa, Bruno. “File Management Using
Resource Files.” Game Programming Gems
2. Ed. Mark DeLoura. Hingham: Charles
River Media, 2001.
EXERCISES
In this chapter’s first exercise, you need to reduce the
size of the JSON files created by SaveLevel. In the
second exercise you convert the Animation file
format to binary.
Exercise 14.1
One issue with the SaveLevel code is that you write
every property for every actor and all its components.
However, for a specific subclass like TargetActor,
few if any of the properties or components change
after construction.
To solve this problem, when it’s time to save the level,

you can create a temporary TargetActor and write out
the JSON object for that actor by using the normal
writing techniques. This JSON object serves as the
template, as it’s the state of TargetActor when it’s
originally spawned. Then, for each TargetActor to
save in the level, compare its JSON object to the
template one and write only the properties and
components that are different.
You can then use this process for all the different types of
actors. To assist with this, RapidJSON provides
overloaded comparison operators. Two
rapidjson::Values are equal only if they have the
same type and contents. This way, you can eliminate
setting at least most of the components (because they
won’t change). It will require a bit more work to do this
on a granular (per-property) level.
Exercise 14.2
Applying the same binary file techniques you used
for the mesh files, create a binary file format for the
animation files. Because all the tracks of bone
transforms are the same size, you can use a format
where, after writing the header, you write the ID for
each track followed by the entire track information.
For a refresher on the animation file format, refer to
Chapter 12, “Skeletal Animation.”
APPENDIX A
INTERMEDIATE C++
REVIEW
This appendix provides a quick review of

intermediate C++ concepts used throughout
the book. The concepts roughly correspond
to typical topics in Computer Science 1 and 2
courses. If you are rusty with C++, you
should spend extra time reviewing this
material.
REFERENCES, POINTERS, AND
ARRAYS
Although references, pointers, and arrays might
seem like separate concepts, they are closely related.
Furthermore, because pointers are often stumbling
points for C++ programmers, it’s worthwhile to
spend some time reviewing their intricacies.
References
A reference is a variable that refers to another
variable that already exists. To denote a variable as a
reference, add an & immediately after the type. For
example, here is how you can declare r as a reference
to the already existing integer i:
int i = 20;
int& r = i; // r refers to i
By default, functions pass parameters by value (pass-

by-value), meaning that when you call a function, the
parameter copies to a new variable. When passing by
value, modifications to parameters do not persist beyond
the function call. For example, here is an (incorrect)
implementation of a Swap function that swaps two
integers:
void Swap(int a, int b)
int temp = a;
a = b;
b = temp;
The problem with Swap is that a and b are copies of the

parameters, which means the function cannot truly swap
the parameters as desired. To solve this problem, you
should instead declare the parameters of Swap as
references to integers:
void Swap(int& a, int& b)
// (The body of the function is identical to the previous)
When passing a parameter by reference (pass-by-

reference), any changes made to that parameter within
the function will persist after the function ends.
One caveat is that because a and b are now references to

integers, they must reference existing variables. You
can’t pass in temporary values for the parameters. For
example, Swap(50,100) is invalid because 50 and 100
are not declared variables.
warning
PASS-BY-VALUE IS DEFAULT: By default, all parameters in C++, even
objects, pass by value. In contrast, languages like Java and C# default to
passing objects by reference.
Pointers
To understand pointers, it first helps to remember
the way computers store variables in memory.
During program execution, entering a function
automatically allocates memory for local variables in
a segment of memory called the stack. This means
that all local variables in a function have memory
addresses known to the C++ program.
Table A.1 shows code snippets and possible locations of

their variables in memory. Notice that each variable has
an associated memory address. The table shows the
memory addresses in hexadecimal simply because that’s
the typical notation for memory addresses.
Table A.1 Variable Storage
Code Variable Memory Address Value
int x = 50; x 0xC230 50
int y = 100; y 0xC234 100
int z = 200; z 0xC238 200

The address-of operator (also &) queries the address of
a variable. To get the address of a variable, place a & in
front of the variable. For example, given the code in
Table A.1, the following code outputs the value 0xC234:
std::cout << &y;
A pointer is a variable that stores an integral value

corresponding to a memory address. The following line
declares the pointer p that stores the memory address of
the variable y:
int* p = &y;
The * after the type signifies a pointer. Table A.2 shows

the pointer p in action. Note that, just like any other
variable, p has both a memory address and a value. But
because p is a pointer, its value corresponds to the
memory address of y.
Table A.2 Variable Storage (with Pointers)
int x = 50; x 0xC230 50
int y = 100; y 0xC234 100
int z = 200; z 0xC238 200

int* p = &y; p 0xC23C 0xC234
The * operator also dereferences a pointer.

Dereferencing a pointer accesses the memory “pointed
to” by the pointer. For example, the last line in Table A.3
changes the value of y to 42. This is because
dereferencing p goes to the memory address 0xC234,
which corresponds to the location of y in memory. Thus,
writing the value 42 at this memory address overwrites
the value of y.
Table A.3 Variable Storage (with Dereferencing)
int x = 50; x 0xC230 50
int y = 100; y 0xC234 42
int z = 200; z 0xC238 200
int* p = &y; p 0xC23C 0xC234
*p = 42;
Unlike references, which must refer to something,

pointers can point to nothing. A pointer that points to
nothing is a null pointer. To initialize a pointer as null,
use the nullptr keyword, as in the following code:
char* ptr = nullptr;

Dereferencing a null pointer crashes the program. The
error message varies depending on the operating system,
but typically “access violation” or “segmentation fault”
errors occur when dereferencing a null pointer.
Arrays
An array is a collection of multiple elements of the
same type. The following code declares an array of 10
integers called a and then sets the first element in
the array (index 0) to 50:
int a[10];
a[0] = 50;
By default, the elements in an array are uninitialized.

While you could manually initialize each element in an
array, it’s more convenient to use either the initializer
syntax or a loop. The initializer syntax uses braces, like
this:
int fib[5] = { 0, 1, 1, 2, 3 };
Alternatively, you could use a loop. The following

initializes each of the 50 elements in array to 0:
int array[50];
for (int i = 0; i < 50; i++)
array[i] = 0;
warning
ARRAYS DON’T BOUND CHECK: Requesting invalid indices can lead to
memory corruption and other errors. Several tools exist to help find bad
memory accesses, such as the AddressSanitizer tool available in Xcode.
C++ stores arrays contiguously in memory. This means

that the data for index 0 is right next to the data for
index 1, which is right next to index 2, and so on. Table
A.4 shows an example of a five-element array in memory.
Keep in mind that the variable array (without the
subscript) references the memory address of index 0 (in
this case, 0xF2E0). Because of this, you can pass a
single-dimensional array to a function via a pointer.
Table A.4 An Array in Memory
Code Variable Memory Value

Address
array[0] 0xF2E0 2
int array[5] = {
2, 4, 6, 8, 10 array[1] 0xF2E4 4
}; array[2] 0xF2E8 6
array[3] 0xF2EC 8
array[4] 0xF2F0 10
A POINTER BY ANY OTHER NAME...
The C programming language (the precursor to C++) does not support

references. Thus, the concept of passing by reference does not exist in
C. Instead of using references, you must use pointers. For example, in
C, you would write the Swap function as follows:
void Swap(int* a, int* b)

{
int temp = *a;
*a = *b;
*b = temp;
}
Calling this version of Swap then requires the address-of operator:
int x = 20;
int y = 37;
Swap(&x, &y);
At program execution time, there is no difference between how

references and pointers work. However, keep in mind that a reference
must refer to something, whereas a pointer can be nullptr.
In C++, many developers prefer passing by reference over passing by

pointer. This is because passing by pointer implies that nullptr is a
valid parameter. However, for stylistic reasons, this book typically passes
dynamically allocated objects by pointer, even in cases where
references would work.
You also can declare multidimensional arrays. For

example, the following code creates a 2D array of floats
with four rows and four columns:
float matrix[4][4];
To pass a multidimensional array into a function, you

must explicitly specify the dimensions, like this:
void InvertMatrix(float m[4][4])
// Code here...
Dynamic Memory Allocation

As discussed earlier, memory allocation for local
variables is automatic in C++. These variables end up
in memory on the stack. This is great for temporary
variables and function parameters. However, local
variables are sometimes not enough.
First, the stack has a limited amount of memory

available—typically much less than the amount of
memory an average program might want to use. For
example, the Microsoft Visual C++ compiler has a
default stack size of 1 MB. Such a small amount of
memory won’t be enough for all but the simplest games.
Second, local variables have a fixed lifetime. They are

only available from the point of declaration until the end
of the containing scope. This scope is typically within a
function, as global variables are stylistically undesirable.
In dynamic memory allocation, the programmer

controls allocation and deallocation of variables in
memory. Dynamic allocations go into the heap, which is
a separate part of memory. The heap is much larger in
size compared to the stack (several gigabytes on current
machines), and data on the heap persists until either the
programmer deletes the data or the program ends.
Recall that in C++, the new and delete operators

allocate and deallocate memory on the heap. The new
operator allocates memory for the requested type of
variable, and for classes and structs, it calls the
constructor. The delete operator performs the
opposite: It calls the destructor for class/struct types and
deallocates the memory for the variable.
For example, this code dynamically allocates memory for

a single int variable:
int* dynamicInt = new int;
To free the memory of a dynamically allocated variable,

use delete:
delete dynamicInt;
Forgetting to delete a dynamically allocated variable

causes a memory leak, meaning that the memory is
unusable for the remaining life span of the program. For
programs that run for a long time, small memory leaks
accumulate over time and eventually cause the heap to
run out of memory. If the heap runs out of memory, the
program will almost always crash soon afterward.
Of course, dynamically allocating just a single integer
doesn’t take advantage of all the available memory on the
heap. You can also dynamically allocate arrays:
char* dynArray = new char[4*1024*1024];
dynArray[0] = 32; // Set the first element to 32
Note that when dynamically allocating an array, you

place square brackets immediately after the type and
specify the size there. Unlike with a statically allocated
array, with a dynamically allocated array, you can specify
a size at runtime.
To delete a dynamically allocated array, use delete[]:
delete[] dynArray;
ASSORTED CLASS TOPICS

Recall that C++ supports object-oriented
programming through classes. This section assumes
familiarity with the basics of classes in C++: classes
versus objects, how to declare a class with member
variables and functions, the constructor, and
inheritance and polymorphism. It instead focuses on
certain topics that cause some issues when using
classes in C++.
References, const, and Classes
Passing objects by value to functions is inefficient.
This is because copying the object can be expensive,
especially if the object has a lot of data. Thus, a best
practice is to pass objects by reference.
However, one issue with references is that they allow the

function to modify the parameter. For example, suppose
an Intersects function takes in two Circle objects
and returns whether these circles intersect. If the
function takes in these circles by reference, it could
choose to modify the centers or radii of the circles.
The solution is to instead use a constant (const)

reference. A const reference guarantees that the
function can only read from the reference but not write
to it. So, a more correct declaration of Intersects uses
const references:
bool Intersects(const Circle& a, const Circle& b);
You can also mark member functions as const

member functions to guarantee that a member
function does not modify member data. For example, a
GetRadius function for Circle shouldn’t modify
member data, which means it should be a const
member function. To denote that a member function is
const, add the const keyword immediately after the
closing parenthesis of the function declaration, as in
Listing A.1.
Listing A.1 Circle Class with const Member

Function
class Circle
public:
float GetRadius() const { return mRadius }
// Other functions omitted
// ...
private:
Point mCenter;
float mRadius;
};
To summarize, the best practices when it comes to

references, const, and classes are as follows:
Pass non-basic types by reference, const reference, or pointers, to

avoid making copies.
Pass by const reference when a function does not need to modify a

reference parameter.
Mark member functions that don’t modify data as const.

Dynamic Allocation of Classes
Just as with any other type, you can dynamically
allocate classes. Listing A.2 shows a declaration for a
Complex class that encapsulates a real part and an
imaginary part.
Listing A.2 A Complex Class

class Complex
public:
Complex(float real, float imaginary)
: mReal(real)
, mImaginary(imaginary)
{ }
private:
float mReal;
float mImaginary;
};
Notice how the constructor for Complex takes two

parameters. To dynamically allocate an instance of
Complex, you must pass in these parameters:
Complex* c = new Complex(1.0f, 2.0f);
As with dynamic allocation of other types, the new

operator returns a pointer to the dynamically allocated
object. Given a pointer to an object, the -> operator
accesses any public members. For example, if the
Complex class had a public Negate member function
that takes no parameters, the following would call the
function on the object c:
c->Negate();
You can also dynamically allocate arrays of objects. This

works only if the class has a default constructor (a
constructor that takes in no parameters). This is because
there’s no way to specify constructor parameters when
dynamically allocating an array. If you don’t define any
constructors for a class, C++ automatically creates a
default constructor for you. However, if you declare a
constructor that takes in parameters, then C++ will not
automatically create a default constructor. In this case, if
you want a default constructor you must declare it
yourself. In the case of Complex, because you declared a
non-default constructor, there is no default constructor.
Destructors
Suppose you needed to dynamically allocate arrays of
integers several times throughout a program. Rather
than manually write this code repeatedly, it might
make sense to encapsulate this functionality inside a
DynamicArray class, as in Listing A.3.
Listing A.3 A Basic Declaration for DynamicArray

class DynamicArray
public:
// Constructor takes in size of element
DynamicArray(int size)
: mSize(size)
, mArray(nullptr)
mArray = new int[mSize];
// At function used to access an index
int& At(int index) { return mArray[index]; }
private:
int* mArray;
int mSize;
};
With this DynamicArray class, you could create a

dynamic array with 50 elements by using the following
code:
DynamicArray scores(50);
However, as previously discussed, every call to new must

have a matching call to delete. In this case,
DynamicArray dynamically allocates an array in its
constructor, but there’s no matching delete[]
anywhere. This means that when the scores object goes
out of scope, there is a memory leak.
The solution is to use another special member function

called the destructor. The destructor is a member
function that automatically runs when destroying the
object. For objects allocated on the stack, this happens
when the object goes out of scope. For dynamically
allocated objects, the delete call on the object also
invokes the destructor.
The destructor always has the same name as the class,

except prefaced with a tilde (~). So for DynamicArray,
this is the destructor:
DynamicArray::~DynamicArray()
{
delete[] mArray;
If you add this destructor, when scores goes out of

scope, the destructor deallocates mArray, eliminating
the memory leak.
The Copy Constructor

The copy constructor is a special constructor that
creates an object as a copy of another object of the
same type. For example, suppose you declare the
following Complex object:
Complex c1 = Complex(5.0f, 3.5f);
You could then instantiate a second instance of Complex

as a copy of c1:
Complex c2(c1);
In most cases, C++ provides a copy constructor

implementation if the programmer doesn’t declare one.
This default copy constructor directly copies all member
data from the original object to the new object. For
Complex, this works perfectly fine; for example, it
means that c2.mReal and c2.mImaginary directly
copy from the corresponding members of c1.
However, for classes with pointers to data, such as

DynamicArray, directly copying member data doesn’t
give the desired result. Suppose you run the following
code:
DynamicArray array(50);
DynamicArray otherArray(array);
With the default copy constructor, you directly copy the

mArray pointers rather than copying the underlying
dynamically allocated array. This means that if you next
modify otherArray, you also modify array at the
same time! Figure A.1 illustrates this problematic
behavior, called a shallow copy.
Figure A.1 A shallow copy of mArray

If the default copy constructor is insufficient, as is the
case for DynamicArray, you must declare a custom
copy constructor:
DynamicArray(const DynamicArray& other)
: mSize(other.mSize)
, mArray(nullptr)
// Dynamically allocate my own data
// Copy from other's data
for (int i = 0; i < mSize; i++)
mArray[i] = other.mArray[i];
Note that the only parameter to the copy constructor is a

const reference to another instance of the class. The
implementation in this case dynamically allocates a new
array and then copies over the data from the other
DynamicArray. This is as a deep copy because the two
objects now have separate underlying dynamically
allocated arrays.
In general, classes that dynamically allocate data should

implement the following member functions:
A destructor to free the dynamically allocated memory
A copy constructor to implement deep copies
An assignment operator (discussed in the next section), also to

implement deep copies
If it’s necessary to implement any of these three

functions, then you should implement all three. This
problem is so common in C++ that developers coined the
term rule of three to remember it.
note
In the C++11 standard, the rule of three expands to the rule of five, as there
are two additional special functions (the move constructor and the move
assignment operator). While this book does use some C++11 features, it
does not use these additional functions.
Operator Overloading
C++ gives programmers the ability to specify the
behavior of built-in operators for custom types. For
example, you can define how the arithmetic
operators work for the Complex class. In the case of
addition, you can declare the + operator as follows:
friend Complex operator+(const Complex& a, const Complex& b)
return Complex(a.mReal + b.mReal,

a.mImaginary + b.mImaginary);
Here, the friend keyword means that operator+ is a

standalone function that can access Complex’s private
data. This is a typical declaration signature for binary
operators.
After overloading the + operator, you can use it to add

two complex objects, like so:
Complex result = c1 + c2;
You can also override binary comparison operators. The

only difference is that these operators return a bool. For
example, the following code overloads the == operator:
friend bool operator==(const Complex& a, const Complex& b)
return (a.mReal == b.mReal) &&
(a.mImaginary == b.mImaginary);
You can also overload the = operator (or assignment

operator). As with the copy constructor, if you don’t
specify an assignment operator, C++ gives you a default
assignment operator that performs a shallow copy. So,
you usually only need to overload the assignment
operator in the “rule of three” case.
There’s one big difference between an assignment

operator and a copy constructor. With a copy
constructor, you’re constructing a new object as a copy of
an existing one. With an assignment operator, you’re
overwriting an already existing instance of an object.
For example, the following code invokes the assignment
operator for DynamicArray on the third line because
the first line previously constructed a1:
DynamicArray a1(50);
DynamicArray a2(75);
a1 = a2;
Because the assignment operator overwrites an already

existing instance with new values, it needs to deallocate
any previously dynamically allocated data. For example,
this is the correct implementation of the assignment
operator for DynamicArray:
DynamicArray& operator=(const DynamicArray& other)
{
// Delete existing data
delete[] mArray;
// Copy from other
mSize = other.mSize;
for (int i = 0; i < mSize; i++)
mArray[i] = other.mArray[i];
// By convention, return *this
return *this;
Note that the assignment operator is a member function

of the class, not a standalone friend function. Also, by
convention, the assignment operator returns a reference
to the reassigned object. This allows for code (albeit ugly
code) that chains assignments, such as the following:
a = b = c;
You can override nearly every operator in C++, including

the subscript [] operator, new, and delete. However,
with great power comes great responsibility. You should
try to overload an operator only when it’s clear what the
operator does. It makes sense that the + operator does
addition, but you should avoid reassigning the meaning
of an operator.
For example, some math libraries override | and ^ for

the dot and cross products, even though | and ^ are the
bitwise OR and bitwise XOR for integral types.
Overusing operator overloading in this manner leads to
code that is difficult to understand. Amusingly, the C++
library itself breaks this best practice: Streams overload
the >> and << operators for input and output (which are
bitshift operators for integral types).
COLLECTIONS
A collection provides a way to store elements of
data. The C++ Standard Library (STL) provides
many different collections, and so it’s important to
understand when to utilize which collections. This
section discusses the most commonly used
collections.
Big-O Notation
Big-O notation describes the rate at which an
algorithm scales as the problem size scales. This rate
is also known as the time complexity of the
algorithm. You can use Big-O to understand the
relative scaling of specific operations on collections.
For example, an operation with a Big-O of O(1)
means that regardless of the number of elements in
the collection, the operation will always take the
same amount of time. On the other hand, a Big-O of
O(n) means that the time complexity is a linear
function of the number of elements.
Table A.5 lists the most common time complexities, from

the fastest to the slowest. Algorithms that are
exponential or slower are too slow to see real use beyond
very small problem sizes.
Table A.5 Common Time Complexities in Big-O

Notation (Fastest to Slowest)
Big-O Described Examples

As
O(1) Constant Insertion into the front of a linked list,

array indexing
O(log Logarithmic Binary search (given an already sorted

n) collection)
O(n) Linear Linear search
O(n log “n log n” Merge sort, quick sort (average case)

n)
O(n2) Quadratic Insertion sort, bubble sort
O(2n) Exponential Integer factorization
O(n!) Factorial Brute-forcing the traveling salesperson

problem
Although Big-O notation says how an algorithm scales,

for certain problem sizes, algorithms with worse time
complexity may perform better. For example, a quick
sort has an average time complexity of O(n log n), while
an insertion sort has a time complexity of O(n2).
However, for small problem sizes (such as n<20), the
insertion sort has a faster execution time because it does
not use recursion. Thus, it’s as important to consider the
actual execution performance of an algorithm as its
specific use case.
Vector
A vector is a dynamic array that automatically
resizes based on the number of elements in the
collection. To insert elements into a vector, use the
push_back (or emplace_back) member function.
This adds an element to the end (back) of the vector.
For example, the following code declares a vector of
floats and then adds three elements at the end of the
vector:
// #include <vector> to use std::vector
std::vector<float> vecOfFloats;
vecOfFloats.push_back(5.0f); // Contents: { 5.0f }
vecOfFloats.push_back(7.5f); // Contents: { 5.0f, 7.5f }

vecOfFloats.push_back(10.0f); // Contents: { 5.0f, 7.5f, 10.0f }
Once the vector has elements, you can use array

subscript notation to access specific elements in the
vector. So given the vector from the preceding snippet,
vecOfFloats[2] accesses the third element in the
vector, yielding 10.0f.
In the long run, insertion into the back of a vector

averages out to O(1). However, because a vector exists in
one contiguous block of memory, as in Figure A.2,
insertion at an arbitrary position in the vector is O(n).
Because of this, you should avoid arbitrary insertion into
a vector. But an advantage of this contiguous memory
layout is that accessing an element at an index is O(1).
Figure A.2 The internal memory layout of a vector

is contiguous, as with arrays
Linked List
A linked list is a collection that stores each element
at a separate location in memory and links them
together with pointers. The std::list collection
allows for insertion to both the front and the back of
the list. Use the push_front (or emplace_front)
function to insert into the front, and push_back (or
emplace_back) for the back. The following code
creates a linked list of integers and inserts a handful
of elements:
// #include <list> to use std::list
std::list<int> myList;
myList.push_back(4);
myList.push_front(2);
Figure A.3 illustrates myList after completing all the

insertions. Note that, by definition, the elements in the
linked list are not next to each other in memory. One
advantage of a linked list is that insertion to either end of
the list is O(1). If you have a pointer to an element in the
list, you can also insert before or after that element in
O(1) time.
Figure A.3 myList with elements inserted into it
However, one disadvantage of a linked list is that

accessing the nth element of the list is O(n). For this
reason, the implementation of std::list does not
allow indexing via array subscripting.
EFFICIENCY: LINKED LIST OR VECTOR?
In cases where each individual element in the collection is small (less

than 64 bytes), a vector almost always outperforms a linked list. This is
because of the way a CPU accesses memory.
Reading values from memory is very slow for the CPU, so when it needs
to read a value from memory, it also loads neighboring values into a
high-speed cache. Because elements in a vector are contiguous in
memory, accessing an element at a specific index also loads its
neighboring indices into the cache.
However, because elements in a linked list are not contiguous, loading

one element also loads unrelated memory into the cache. Thus,
operations such as looping over an entire collection are much more
efficient with a vector than with a linked list, even though both operations
have a time complexity of O(n).
Queues
A queue exhibits first-in, first-out (FIFO)
behavior, much like waiting in a line at a store. With
a queue, you cannot remove elements in any
arbitrary order. With a queue, you must remove
elements in the same order in which they were
added. Although many books use enqueue to
reference insertion into a queue and dequeue to
reference removal from a queue, the implementation
of std::queue uses push (or emplace) for
insertion and pop for removal. To access the element
at the front of the queue, use front.
The following code inserts three elements into a queue

and then removes each element from the queue,
outputting the values:
// #include <queue> to use std::queue
std::queue<int> myQueue;
myQueue.push(10);
myQueue.push(20);
myQueue.push(30);
for (int i = 0; i < 3; i++)
std::cout << myQueue.front() << ' ';
myQueue.pop();
Because queues operate in a FIFO manner, the above

code outputs the following:
10 20 30
The std::queue implementation guarantees O(1) time

complexity for insertion, accessing the front element,
and removal.
Stack
A stack exhibits last-in, first-out (LIFO)
behavior. For example, if you add the elements A, B,
and C to a stack, you can only remove them in the
order C, B, A. You use the push (or emplace)
function to add an element onto the stack and the
pop function to remove an element from the stack.
The top function accesses the element on the “top”
of the stack. The following code shows std::stack
in action:
// Include <stack> to use std::stack
std::stack<int> myStack;
myStack.push(10);
myStack.push(20);
myStack.push(30);
for (int i = 0; i < 3; i++)
std::cout << myStack.top() << ' ';
myStack.pop();
Because of the LIFO behavior of a stack, the above code

outputs the following:
30 20 10
As with queue, the major operations for std::stack
all have constant time complexity.
Maps
A map is an ordered collection of {key, value} pairs,
sorted by key. Each key in the map must be unique.
Because a map has both a key type and a value type,
you must specify both types when declaring a map.
The recommended way to add an element to a map is
with the emplace function, which takes in the key
and values as parameters. For example, the following
code creates a std::map of months, where the key is
the number of the month and the value is the string
name of the month:
// #include <map> to use std::map
std::map<int, std::string> months;
months.emplace(1, "January");
months.emplace(2, "February");
months.emplace(3, "March");
// ...
The easiest way to access an element from a map is by

using the [] operator and passing in the key. For
example, the following would output February:
std::cout << months[2];
However, this syntax works as expected only if the key is

in the map. To determine if a key is in a map, use the
find function. This returns an iterator to the element, if
found. (We discuss iterators in a moment.)
Internally, the std::map implementation uses a

balanced binary search tree. This means that std::map
can find an element by key in O(log n) (logarithmic)
time. Insertion and removal from the map are also
logarithmic. Furthermore, due to the binary search tree,
looping over the contents of a map is in ascending order
of the keys.
Hash Maps
While a regular map maintains an ascending order of
the keys, a hash map is unordered. In exchange for
the lack of ordering, insertion, removal, and search
are all O(1). Thus, in cases where you need a map but
don’t need ordering, a hash map yields better
performance than a regular map.
The C++ hash map std::unordered_map has the

same functions as std::map, just without any
guaranteed ordering. To use the hash map class, use
#include <unordered_map>.
Iterators, Auto, and Range-Based For

Loops
For looping over all the elements in a vector you can
use the same syntax as for looping over an array.
However, many of the other C++ STL collections,
such as list and map, do not support this array
syntax.
One way to loop over these other containers is with an

iterator, which is an object that helps traverse the
collection. Every C++ STL collection supports iterators.
Each collection has a begin function that returns an
iterator to the first element and an end function that
returns an iterator to the last element. The type of the
iterator is the type of the collection followed by
::iterator. For example, the following code creates a
list and then uses an iterator to loop over each element in
the list:
std::list<int> numbers;
numbers.emplace_back(2);
for (std::list<int>::iterator iter = numbers.begin();
iter != numbers.end();
++iter)
std::cout << *iter << std::endl;

}
Note that the iterator is dereferenced with *, the same

way a pointer is dereferenced. The syntax for looping
over other collections with iterators is similar.
In the case of a map, the iterator actually points to a

std::pair. So, given an iterator to an element in a
map, you must use first and second to access the key
and value, respectively. Returning to the months map
from earlier, you can get an iterator to an element and
output its data with the following code:
// Get an iterator to the element with the key 2
std::map<int, std::string> iter = months.find(2);
if (iter != months.end()) // This is only true if found
std::cout << iter->first << std::endl; // Outputs 2
std::cout << iter->second << std::endl; // Outputs February
Typing out the long type names for iterators is annoying.

C++11 provides the auto keyword to help reduce this
pain. auto tells the compiler to deduce the type of a
variable for you, based on the assigned value. For
example, because the begin function returns an iterator
of a very specific type, auto can deduce that correct type.
There is no performance penalty for using auto, though
some programmers find the code harder to understand.
Using auto, you can rewrite the list loop as follows:
// auto is deduced to be std::list<int>::iterator
for (auto iter = numbers.begin();
iter != numbers.end();
++iter)
std::cout << *iter << std::endl;
The code in this book uses auto only when it provides a

benefit in readability.
Even with auto, the code for looping via iterators is

clunky. Many other programming languages provide a
foreach construct for looping over collections. C++11
has a similar construct called a range-based for loop.
To loop over the numbers list with a range-based for
loop, use the following syntax:
for (int i : numbers)
// i stores the element for the current loop iteration
std::cout << i << std::endl;
This loop makes a copy of each element in the list as it

iterates over it. However, you can also pass by reference
if you want to modify elements in the collection.
Similarly, you can use const references.
You can also use auto for the type when writing a range-
based for loop. However, as with using an explicit type,
this makes a copy of each element. However, you can
also use const and & with auto, if needed.
One disadvantage of a range-based for loop is that you

can’t add or remove elements in the collection during the
loop. So, if you need this behavior, you must use another
type of loop.
ADDITIONAL READING
There are many excellent resources available online
to help you learn and practice the fundamentals of
C++. One such website is LearnCPP.com, which
contains a very in-depth progression of topics. If you
prefer traditional books, you should see Stephen
Prata’s book, which provides coverage of the basics.
Eric Roberts’s book covers both the fundamentals of
C++ and relevant data structures.
Both of Scott Meyers’s books are great resources for best

practices in C++. They are short reads that provide many
tips on how to achieve maximum effectiveness from C++
code.
There is also a great deal of information available on the

C++ Standard Library. Bjarne Stroustrup, the creator of
C++, devotes a large section of his book to the C++
collection implementations.
LearnCpp.com. Last modified April 28,

2016. http://www.learncpp.com.
Meyers, Scott. Effective C++, 3rd edition.
Boston: Addison-Wesley, 2005.
Meyers, Scott. Effective Modern C++.
Sebastopol: O’Reilly Media, 2014.
Prata, Stephen. C++ Primer Plus, 6th
edition. Upper Saddle River: Addison-
Wesley, 2012.
Roberts, Eric. Programming Abstractions
in C++. Boston: Pearson, 2014.
Stroustrup, Bjarne. The C++ Programming
Language, 4th edition. Upper Saddle River:
Pearson, 2013.
INDEX
SYMBOLS
∗ operator, 63
+ operator, 64
= operator, 469
== operator, 469
− operator, 62
NUMBERS
2D coordinate systems, 151
2D graphics, 14–15
drawing, 19–21
filtering analog sticks, 267–269
implementing, 18–21
Initialize function, 19
scrolling backgrounds, 51–53
Shutdown function, 19
sprites, 42
animating, 48–50
drawing, 44–48
2D transformations
rotation matrix, 157–158
scale matrix, 157
translation matrices, 158–159
3D
Actor transform, 184, 189–190
BasicMesh shader, 203–204
calculating view-projection matrix, 200
Euler angles, 185–186
lighting. See lighting
MeshComponent, 204–206
quaternions, 186–187
in code, 188–189
combining rotations, 187
quaternion-to-rotation matrix, 188
rotating vectors, 188
spherical linear interpolation (Slerp), 188
transform matrices, 184–185
transforming clip space
projection matrix, 197–200
view matrix, 196–197
updating vertex attributes, 193–194
z-buffering, 201–203
3D AABB, 308
3D coordinate system, 184
3D meshes, drawing, 195–196
3D models
choosing formats, 191–193
loading, 190
vertex and index buffers, 133–138
3D positional audio, 233–234
setting up listeners, 234–236
4x4 matrices, 184–185
A
A*, 106, 111–113
optimizing, 113
AABB (axis-aligned bounding box), 303–306
AABB contains point tests, 308
AABB versus AABB test, 311–312
acceleration, 79
Actor transform, 3D, 184, 189–190
Actor::LoadProperties, 436–437
actors
adding to world transforms, 159–161
associating with sound events, AudioComponent,
237–238
jumping, 248–251
loading in level files, 435–439
saving in level files, 446–448
Actor::SaveProperties, 446–447
adaptive refresh rate, 18
AddButton, 349
adding
an aiming reticule, 352–354
flipY option to UIScreen::DrawTexture, 406
pitch to first-person cameras, 278–280
point lights, 418
PointLightComponent class, 419
positional functionality to SoundEvent, 236–237
radar, 354–358
shaders to games, 146
springs to follow cameras, 283–285
text maps, 359–361
world transforms, to actors, 159–161
AddInt, 444
addition, vector addition, 63–64
AddSprite, 46
AddTarget, 352
adjacency lists, 99
admissible, heuristics, 104
adversarial, 116
AI (artificial intelligence), 91
designing state machines, 92–93
game trees, 116–118
alpha-beta pruning, 121–124
incomplete game trees, 120–121
minimax, 118–119
graphs, 98–100
BFS (breadth-first search), 100–103
implementing state machines, 93–95
pathfinding 98
Dijkstra’s Algorithm, 113–114
following paths, 114–115
heuristics, 104–105
navigation meshes, 115–116
path nodes, 115
state machine behaviors, 92
states, as classes, 95–98
AIComponent class, 93–95
AIComponent::Update function, 94, 96
AIDeath state, 97
aiming reticule, adding, 352–354
AIPatrol class, 97
AIState, 95
Alarm state, 93
Alert state, 93
alpha, 121
alpha blending, 178–180
alpha values, 14–15
AlphaBetaDecide, 121–122
AlphaBetaMax, 122–123
AlphaBetaMin, 123
ambient light, 207–208
loading global properties, 434–435
analog sticks, 264–267
filtering, 267–269
anchors, 361
angle summation, 309
angles
converting from forward vectors, 67–68
converting to forward vectors, 66–67
determining between two vectors, dot product, 68–70
animating sprites, 48–50
animation, skeletal animation. See skeletal animation
Animation class, 381–382
animation data, 371–372
loading for skeletal animation, 380–385
animations, updating, 386–389
AnimSpriteComponent, 49, 50, 53–54
anisotropic filtering, 399–400
API (application programming interface), 3
Apple macOS, setting up development environments, 2
application programming interface (API), 3
architecture, InputSystem, 251–253
arctangent, 67–68
arrays, 460–461, 462
artificial intelligence. See AI (artificial intelligence)
ASCII, 358–359
AStarScratch, 111
Asteroid constructor, 74–75
Asteroids, 84–86
atan2 function, 67
Attack state, 93
attenuates, 233
attributes
OpenGL, 129
skinning vertex attributes, drawing with, 374–378
vertex attributes, updating, 193–194, 206
audio
3D positional audio, 233–234
setting up listeners, 234–236
audio systems, creating, 224–226
AudioComponent, creating to associate actors with
sound events, 237–238
banks, loading/unloading, 227–229
bootstrapping, 222
buses, 242–243
Doppler effect, 240–241
effects, 241–242
equalization, 242
FMOD, 222–223
banks and events, 226–227
event instances, 229–230
installing, 223–224
listeners, in third-person games, 239–240
mixing, 241–242
occlusion, 243–244
positional functionality, adding to SoundEvent,
236–237
reverb, 242
snapshots, 243
SoundEvent class, 230–233
AudioComponent, creating to associate actors with
sound events, 237–238
AudioComponent::PlayEvent, 238
AudioComponent::StopAllEvents, 238
AudioComponent::Update, 238
AudioSystem::PlayEvent Implementation with Event
IDs, 230–231
AudioSystem::SetListener, 235–236
AudioSystem::UnloadAllBanks, 228
AudioSystem::Update, 231
auto, 475–476
axis, filters, 265
axis-aligned bounding box. See AABB (axis-aligned
bounding box)
B
back buffer, 17
backgrounds, scrolling backgrounds, 51–53
ball collisions
with SegmentCast, 327–328
testing in PhysWorld, 329–331
BallActor, 328
balls
drawing, 20–23
updating position of, 28–30
banks
FMOD, 226–227
loading/unloading, 227–229
basic input processing, 11–13
Basic.frag file, 141
BasicMesh.vert shader, 421
Basic.vert file, writing basic shaders, 139–140
beta, 121
BGSpriteComponent, 51
BGTexture, 51
bidirectional light, 214
bidirectional reflectance distribution function (BRDF),
209
Big-O notation, 470–471
bilinear filtering, 169, 394, 395–397, 400
bilinear interpolation, 396
binary data, 448–449
binary mesh files, loading, 452–454
saving, binary mesh files, 449–452
binary mesh files
loading, 452–454
saving, 449–452
bind pose, 368
Feline Swordsman, 378
blank windows, creating, 10–11
Blended, 339
blitting, 131
bone hierarchy, 367–368
Bone struct, 370
bones, 366
BoneTransform, 369, 371–372
bootstrapping, audio, 222
bounding spheres, 303
bounding volume tests, 310
capsule versus capsule test, 313–314
sphere versus AABB test, 312–313
sphere versus sphere tests, 311
bounding volumes, 302–303
capsules, 306–307
convex polygons, 307
OBB (oriented bounding box), 306
spheres, 303
BoxComponent class, 324–325
BRDF (bidirectional reflectance distribution function),
209
breadth-first search (BFS), 100–103
broadphase techniques, 331
buses, 242–243
Button class, 346–350
buttons
controllers, 262–264
mouse input, 256–258
user interfaces, 346–350
Button::SetName, 347
C
C++ Standard Library, 3
calculating
lighting, 407–408
normal vectors, 70–72
points on line segments, 298
view-projection matrix, 200
weighted averages, in bilinear filtering, 396
world transform matrix, 189
camera position, computing, 282
CameraComponent, 277–278
cameras, 34
first-person camera, 276
adding pitch, 278–280
cameras without pitch, 277–278
first-person models, 280–281
movement, 276–277
follow cameras, 281–283
adding springs, 283–285
orbit cameras, 286–288
spline cameras, 289–292
unprojection, 292–294
without pitch, 277–278
capsule contains point tests, 309
capsules, 306–307
Catmull-Rom spline, 289
CCD (continuous collision detection), 321
ChangeState function, 94, 96, 98
channels, 15, 222
character.jump(), 249
characters, jumping, 248–251
choosing 3D model formats, 191–193
CircleComponent subclass, creating, 83–84
circle-versus-circle intersection, 82
class hierarchies, game objects, 34–36
classes, 464–465
dynamic allocation of, 465
states as, 95–98
clip space, transforming
with projection matrix (3D), 197–200
with view matrix (3D), 196–197
from world space, 161–163
closed set, GBFS (greedy best-first search), 106
collections, 470
collision detection, 81
ball collisions
with SegmentCast, 327–328
testing in PhysWorld, 329–331
bounding volumes, 302–303
capsules, 306–307
spheres, 303
BoxComponent class, 324–325
CircleComponent subclass, creating, 83–84
circle-versus-circle intersection, 82
dynamic objects, 321–323
geometric types, 298
line segments, 298–301
planes, 301–302
intersection tests. See intersection tests
line segment tests
line segment versus AABB test, 318–320
line segment versus plane test, 314–315
line segment versus sphere test, 315–317
PhysWorld class, 325–327
player collision, against walls, 331–333
CollisionInfo, 326
color buffer, 15–16
color depth, 15
column vectors, 155, 156
combining
equations, 154
rotation, for 3D, 187
transformations, 152–153, 159
vectors, 63–64
compatibility profiles, OpenGL, 129
CompileShader, 142–143
Complete transition, Alarm state, 93
Component class, 440
component declaration, 39–40
component-based game object model, 36–38
ComponentFunc, 442
components
game objects, 36–38
as hierarchy, 38–40
loading, in level files, 439–444
saving, in level files, 446–448
composite, 248
ComputeGlobalInvBindPose, 379
ComputeMatrixPalette, 384, 389
computing
camera position, 282
positions, with Euler integration, 80
const, 464–465
contains point tests, 308
convex polygon contains point (2D) tests, 309–310
sphere contains point tests, 308
ContainsPoint, 347
context, OpenGL, 130
continuous collision detection (CCD), 321
controller input, 261
buttons, 262–264
disabling controllers, 262
enabling single controllers, 261–262
multiple controllers, 269–270
ControllerState, 263, 267
converting
from angles to forward vectors, 66–67
forward vectors, to angles, 67–68
coordinate systems
3D coordinate system, 184
left-handed coordinate system, 184
UV coordinates, 395
coordinates, NDC (normalized device coordinates),
132–133
copy constructor, 467–468
core profile, OpenGL, 129
Create, 410
CreateMirrorTarget, 404
CreateMirrorTexture, 402
CreateRotationZ, 160
CreateScale, 160
CreateTranslation, 160
CreateWorldTransform, 160
cross product, 70–72
crosshair texture, drawing, 353
cross-platform libraries, 3
D
dead zones, 265
Death state transition, 92
debug logging, initializing, 225
Debug_Initialize, 225
deferred shading, 407–408, 424–425
G-buffer
creating, 408–411
writing to, 412–414
global lighting, 414–418
point lights, 418–419
drawing, 421–424
deleting, dynamically allocated arrays, 463
delta time, 24–28
dependency injection, 39
depth buffer, 201
depth buffering. See z-buffering
designing, state machines, AI (artificial intelligence),
92–93
Destroy, 411
destructors, 466–467
development environments, setting up, 2
Apple macOS, 2
Microsoft Windows, 2
dialog boxes, 349–352
diamond inheritance, 36
DiffuseColor, 421
digital signal processing (DSP), 241
directed, 98
directed edges, 100
directional light, 208
directions, vectors, 65–66
disabling controllers, 262
distance, vectors, 64–65
Doppler effect, 240–241
Dot function, 70
dot product, 68–70
double buffering, 16–18
Draw function, 52
Draw3DScene, 404–405, 413
DrawComponent, 36
DrawFromGBuffer, 416–417
drawing
3D meshes, 195–196
balls, 20–23
crosshair texture, 353
mirror texture, in HUD, 406–407
paddles, 20–23
point lights, 421–424
radar, 357
with skinning vertex attributes, 374–378
sprites, 44–48
textures, 47
triangles, 146–148
transformations, 148–149
walls, 20–23
DrawScreen, 348
DSP (digital signal processing), 241
dynamic memory allocation, 462–463
dynamic memory allocation of classes, 465
dynamic objects, collision detection, 321–323
DynamicArray, 466, 467
E
edges, 98
effects, audio, 241–242
encapsulating buttons, 346
endianness, 451
equalization, 242
equations, combining, 154
ES profiles, OpenGL, 129
Escape key, 256
Euclidean distance heuristic, 105
Euler angles, 185–186
Euler integration, 80
event instances, FMOD, 229–230
event processing, 11–12
EventInstances, 229–230
events, 11
FMOD, 226–227
input devices, 251
exchange formats, 191
extensions
.frag extension, 139
.vert extension, 139
F
falloff function, 233
falloff radius, 208
FBO (framebuffer objects)
creating, 402–404
rendering, 404–405
Feline Swordsman, bind pose, 378
FIFO (first-in, first out), 473
Filter1D, 265–267
Filter2D, 268–269
filtering, 400
bilinear filtering, 394, 395–397, 400
nearest-neighbor filtering, 395
trilinear filtering, 398, 399, 400
filters, for axis, 265
first-in, first-out (FIFO), 473
first-person camera, 276
adding pitch, 278–280
cameras without pitch, 277–278
movement, 276–277
first-person movement, 276–277
FixCollisions, 332–333
flat shading, 211
flipY option, adding to UIScreen::DrawTexture, 406
FMOD, 222–223
banks and events, 226–227
event instances, 229–230
installing, 223–224
occlusion, 244
positional audio, 234
snapshots, 243
FMOD Studio, 223
buses, 242–243
DSP (digital signal processing), 242
FMOD_VECTOR, 234–235
FMOD_ErrorString, 225
follow cameras, 281–283
adding springs, 283–285
FollowCamera, 282
FollowCamera::Update, 283, 284–285
following paths, AI (artificial intelligence), 114–115
font rendering, 338–340
Font::Load, 338–339
Font::RenderText, 339, 340, 359
For loops, 475–476
force, 79–80
force feedback, 5
formats, 3D model formats, choosing, 191–193
forward rendering, 407–408
forward vectors
converting from angles, 66–67
converting to angles, 67–68
FPS (frames per second), 4
FPS angular speed, 277
FPSActor::FixCollisions, 332–333
FPSCamera, 278
FPSCamera::Update, 279
.frag extension, 139
fragment shaders, 139, 412
frame buffer, 201
frame limiting, 25–26
frame rate, 4
framebuffer, 15
framebuffer objects
creating, 402–404
rendering, 404–405
frames, 4–6
rendering, in OpenGL, 131
frames per second (FPS), 4
front buffer, 17
G
game classes
main function, 10–11
skeleton game class, 6–7
Initialize function, 7–9
RunLoop function, 9–10
Game Declaration, skeleton game class, 7
game loops, 4
frames, 4–6
integrating game objects into, 40–42
single-threaded game loop, 5
game object models, 40
game objects as class hierarchies, 34–36
game objects as hierarchy with components, 38–40
game objects with components, 36–38
game objects, 34
as class hierarchies, 34–36
with components, 36–38
as hierarchy with components, 38–40
integrating, into game loops, 40–42
types of, 34
game projects
3D graphics, 216
Asteroids, 84–86
audio, 244–245
cameras, 295
collision detection, 333–334
convertingAstroids game to OpenGL, 180
intermediate graphics, 425
level files and binary data, 454–455
moving spaceships, 271–272
skeletal animation, 389–390
sprites, 53–55
tower defense game, 124
user interfaces, 362
game time, 24
game trees, 116–118
AI (artificial intelligence)
minimax, 118–119
Game::GenerateOutput function, 147
Game::LoadData, 430
Game::LoadShaders, 166
games
adding shaders to, 146
Asteroids, 84–86
Pac-Man game, pseudocode, 6
Pong. See Pong
tic-tac-toe, game trees, 116–117
updating, 23
delta time, 24–28
game time, 24
real time, 24
GameState, 120
gamut, 14
GBFS (greedy best-first search), 105–111
G-buffer, 408
creating, 408–411
writing to, 412–414
GBuffer class, creating, 408–411
GBuffer::Create, 410
GBufferGlobal.frag shader, 415–416, 418
GBufferPointLight.frag shader, 419–421
GBufferWrite.frag, 412
geometric types, 298
planes, 301–302
GetCircle function, 84
GetComponentOfType, 441
GetEventInstance, 232
GetFont, 339
GetGlobalPoseAtTime, 382–383, 387–388
GetInt, JsonHelper, 433
GetKeyState, 254, 255
GetKeyValue, 254, 255
GetLeftTrigger(), 267
GetMesh, 195
GetName function, 95
GetNextPoint function, 115
GetPossibleMoves, 120
GetRightTrigger(), 267, 276
GetScore function, 118, 120
GetType, 441
GetVector3, 434
GL_ARRAY_BUFFER, 136
GL_ELEMENT_ARRAY_BUFFER, 136
GL_LINEAR, 399
GL_RGB, 402
GL_RGB16F, 411
GL_RGB32F, 410–411
GL_STATIC_DRAW, 136
GL_TEXTURE_2D, 171
glAttachShader, 145
glBindBuffer, 136
glBindFramebuffer, 402, 404
glBindTexture, 173
glBlendFunc, 179
glBufferData, 174
glCheckFrameBuffer, 403, 404
glClear, 131, 202
glCreateProgram, 145
glDrawBuffers, 403
glDrawElements, 146–147
glEnable, 179
GLEW, initializing, 130–131
glFramebufferTexture, 403
glGenerateMipmap, 399
glGenFrameBuffers, 402
glGenTextures, 171
glGetError, 130
glGetShaderiv, 144
properties, loading, 430–431
global pose, 368
global properties
loading, 430–434
saving, 446
globally unique IDs (GUIDs), 229
GLSL, 138
Basic.vert file, 139–140
glTexImage2D, 171
glVertexAttribIPointer, 376
glVertexAttribPointer, 376
goal nodes, 100
Gouraud shading, 211
gpmesh files, loading, 194–195
graphics
2D graphics. See 2D graphics
intermediate graphics. See intermediate graphics
graphics debuggers, 414
GraphNode, 100
graphs, AI (artificial intelligence), 98–100
Grassmann product, 187
greedy best-first search), 105–111
GUIDs (globally unique IDs), 229
H
HandleKeyPress, 344
hash maps, 475
heads-up display. See HUD (heads-up display)
A*, 111–113
GBFS (greedy best-first search), 105–111
hierarchies, game objects with components, 38–40
high-quality reflections, 401
homogenous coordinates, 158
horizontal field of view (FOV), 199
hot swapping, 261, 270
HUD (heads-up display), 337
drawing, mirror texture, 406–407
UI screen stack, 342–344
UI screens, 340–342
HUD elements, 352
adding
an aiming reticule, 352–354
radar, 354–358
HUD::Draw, 353
HUD::UpdateCrosshair, 353
HUD::UpdateRadar, 356–357
I
IDE (integrated development environment), 2
image files, loading, 43–44
IMG_Load, 43
implementing
lighting, 212–216
drawing with skinning vertex attributes, 374–378
state machines, AI (artificial intelligence), 93–95
improving texture quality, 394
inadmissible, heuristics, 104
index buffers, 133–138
inheritance, diamond inheritance, 36
Initialize function
2D graphics, 19
initializing
debug logging, 225
GLEW, 130–131
OpenGL, 128
input devices, 248
controller input, 261
analog sticks. See analog sticks
buttons, 262–264
enabling single controllers, 261–262
events, 251
keyboards, 253–256
mouse, 256
buttons and position, 256–258
relative motion, 258–259
scroll wheels, 260–261
polling, 248
positive and negative edges, 248–251
triggers, 264
input mappings, 270–271
InputComponent class, creating, 76–79
InputSystem, architecture, 251–253
InputSystem::Filter2D, 268–269
InputSystem::Initialize, 263
installing
FMOD, 223–224
Xcode, 2
instantaneous tests, 321
integrated development environment (IDE), 2
integrating game objects into game loops, 40–42
intermediate graphics
deferred shading, 407–408
drawing mirror texture in HUD, 406–407
improving texture quality, 394
mipmapping, 397–399
rendering to textures, 400–401
creating framebuffer objects, 402–404
creating textures, 401–402
texture sampling, 395–397
interpolate, 186
Intersect function, 83, 315, 319–320
intersection, circle-versus-circle intersection, 82
intersection tests, 307
bounding volume tests, 310
contains point tests, 308
inverse bind pose matrix, 371
IsCompiled, 143–144
IsTerminal, 120
IsValidProgram, 145
iterative deepening, 124
iterators, 475–476
J
joints, 366
JSON (JavaScript Object Notation), 191, 448–449
level files
loading actors, 435–439
loading components, 439–444
saving, 444–446
saving actors and components, 446–448
saving global properties, 446
JsonHelper, 433, 444
JsonHelper::GetInt, 433
jumping characters, spacebar, 248–251
K
keyboards, 248
input, 253–256
KeyboardState, 254
L
last-in, first-out (LIFO), 474
left-handed coordinate system, 184
length, vectors, 64–65
Length () function, 65
LengthSquared() function, 65
level files
loading, 430
actors, 435–439
components, 439–444
global properties, 430–434
saving, 444–446
actors, 446–448
global properties, 446
LevelLoader, 437
LevelLoader class, 431
LevelLoader::LoadActors, 438–439
LevelLoader::LoadComponents, 443–444
LevelLoader::LoadGlobalProperties, 434–435
LevelLoader::LoadJSON, 431–433
LevelLoader::SaveActors, 447–448
LevelLoader:SaveLevel, 445
LevelLoader::SaveLevel, 445
libraries
3D positional audio library, 234–235
C++ Standard Library, 3
OpenGL. See OpenGL
SDL (Simple DirectMedia Layer) library, 4
SDL TTF library, 338
LIFO (last-in, first-out), 474
lighting, 206
ambient light, 207–208
bidirectional light, 214
deferred shading. See deferred shading
directional light, 208
Phong reflection model, 209–211
point lights, 208
adding, 418–419
drawing, 421–424
spotlight, 209
vertex attributes, 206–207
lighting calculations, 407–408
line segment tests, 314
linear mechanics, 79–80
LineSegment::MinDistSq, 300–301, 309
linked lists, 472
listeners, 233
setting up for 3D positional audio, 234–236
in third-person games, 239–240
listings
AABB::Rotate Implementation, 305–306
Abbreviated Render Declaration, 195–196
Actor Declaration, 38–39
Actor::ComputeWorldTransform Implementation, 161
Actor::LoadProperties Function, 436–437
Actor::RotateToNewForward, 327–328
Actors with Components in JSON (Excerpt from the
Full File), 440
Actor::SaveProperties Implementation, 447
Adding a flipY Option to UIScreen::DrawTexture, 406
AIComponent::ChangeState Implementation, 96
AlphaBetaDecide Implementation, 122
AlphaBetaMin Implementation, 123
AlphaBetaMax Implementation, 123
Animation Declaration, 381–382
AnimSpriteComponent Declaration, 49
AnimSpriteComponent::Update Implementation, 50
Asteroid Constructor, 74–75
AudioComponent Declaration, 237
AudioSystem::LoadBank Implementation, 227–228
AudioSystem::PlayEvent Implementation with Event
IDs, 230–231
AudioSystem::SetListener Implementation, 235–236
AudioSystem::Update Implementation with Event IDs,
231
Basic InputSystem Declarations, 252
Basic.frag Code, 141
Basic.vert Code, 140
The Beginning of a Skeleton Data, 370
The Beginning of an Animation Data, 380–381
BGSpriteComponent Declaration, 51
BoxComponent Declaration, 324
BoxComponent::OnUpdateWorldTransform
Implementation, 325
Breadth-First Search, 102
Button Declaration, 346
Circle Class with const Member, 464
CircleComponent Declaration, 83
CircleComponent Intersection, 83
Component Declaration, 39–40
ComputeGlobalInvBindPose, 379–380
ComputeMatrixPalette, 384
Constructing a Plane from Three Points, 301
ConvexPolygon::Contains Implementation, 310
Creating a Texture for Rendering, 401–402
Creating the Mirror Framebuffer, 403–404
Cube.gpmesh, 192
Current Implementation of SpriteComponent::Draw,
166–167
Declaring Vertex Attributes in the VertexArray
Constructor, 375–376
directional light, loading global properties, 434–435
Drawing MeshComponents in Renderer::Draw, 205
Drawing Point Lights in Renderer::DrawFromGBuffer,
422
English.gptext Text Map File, 359
Filter1D Implementation, 266
Final Version of GetGlobalPoseAtTime, 387–388
First Version of GetGlobalPoseAtTime, 382–383
FollowCamera::Update Implementation (with Spring),
284–285
Font Declaration, 338
Font::Load Implementation, 339
Font::RenderText Implementation, 340
FPS Angular Speed Calculation from the Mouse, 277
FPSActor::FixCollisions, 332–333
FPSCamera::Update Implementation (with Pitch
Added), 279
FPSCamera::Update Implementation (Without Pitch),
278
Game Declaration, 7
Game::GenerateOutput Attempting to Draw Sprites,
147
Game::ProcessInput Implementation, 13
Game::UpdateGame Implementation, 26
Game::UpdateGame Updating Actors, 41–42
GBuffer Declaration, 409
GBuffer::Create Implementation, 410
GBufferGlobal.frag Shader, 415–416
GBufferPointLight.frag Main Function, 420–421
GBufferWrite.frag shader, 412
Greedy Best-First Search, 110–111
HUD::UpdateCrosshair, 353
HUD::UpdateRadar Implementation, 356–357
Initial AudioSystem Declaration, 224
Initial ControllerState, 263
Initial MouseState Declaration, 258
Initial Shader Declaration, 142
Initial UIScreen Declaration, 341–342
InputComponent Declaration, 77
InputComponent::ProcessInput Implementation,
77–78
InputSystem::Filter2D, 268–269
InputSystem::ProcessEvent Implementation for the
Scroll Wheel, 260
JsonHelper::GetInt Implementation, 433
KeyboardState Declaration, 254
KeyboardState::GetKeyState, 255
Level with Actors (Level1.gplevel), 435–436
Level with Global Lighting Properties, 430–431
LevelLoader::LoadActors Implementation, 438–439
LevelLoader::LoadComponents Implementation,
443–444
LevelLoader::LoadGlobalProperties, 434–435
LevelLoader::LoadJSON, 432
LevelLoader::SaveActors Implementation, 447–448
LevelLoader:SaveLevel Implementation, 445
Line Segment Versus AABB Helper Function, 319
Line Segment Versus AABB Intersection, 320
Line Segment Versus Plane Intersection, 315
Line Segment Versus Sphere Intersection, 317
LineSegment::MinDistSq Implementation, 300–301
Loading the G-buffer Global Lighting Shader, 416
Loop over the Adjacent Nodes in an A* Search, 112–113
main Implementation, 10
MaxPlayer and MinPlayer Functions, 118–119
MaxPlayerLimit Implementation, 120
Mesh Declaration, 194
MeshBinHeader Struct, 449–450
MeshComponent Declaration, 204
MeshComponent::Draw Implementation, 205
MeshComponent::LoadProperties Implementation,
442
Mesh::LoadBinary Outline, 452–453
Mesh::SaveBinary Implementation, 450–451
MoveComponent Declaration, 73
MoveComponent::Update Implementation, 74
MoveComponent::Update Implementation with
Quaternions, 190
OrbitCamera::Update Implementation, 287–288
Pac-Man Game Loop Pseudocode, 6
Phong.frag Lighting Uniforms, 212
Phong.frag Main Function, 215
Phong.vert Main Function, 214
PhysWorld::SegmentCast, 326–327
PhysWorld::TestPairWise, 329
PhysWorld::TestSweepAndPrune, 330–331
PointLightComponent Declaration, 419
PointLightComponent::Draw Implementation,
423–424
Quaternion Functions of Note, 189
Renderer::Draw Updated to Render Both Mirror and
Default Passes, 405
Renderer::Draw3DScene Helper Function, 404–405
Renderer::DrawFromGBuffer Implementation, 417
Renderer::GetScreenDirection Implementation, 294
Renderer::Unproject Implementation, 293
Requesting OpenGL Attributes, 129
Shader::CompileShader Implementation, 143
Shader::IsCompiled Implementation, 144
Shader::Load Implementation, 144–145
Shader::SetMatrixUniform, 165
Ship Declaration, 54
SkeletalMeshComponent Declaration, 377
Skeleton Declaration, 378–379
Skeleton PhysWorld Declaration, 325–326
Skinned.vert Main Function, 385–386
SoundEvent Declaration, 232
SoundEvent's Is3D and Set3DAttributes
Implementation, 236–237
SplineCamera::Update, 291–292
Spline::Compute Implementation, 290
SpriteComponent Declaration, 45–46
SpriteComponent::Draw Implementation, 48
Sprite.frag Implementation, 177
Sprite.vert Implementation, 176
Swept-Sphere Intersection, 323
Texture Declaration, 170–171
Texture::Load Implementation, 172–173
Transform.vert Vertex Shader, 164
UIScreen::ProcessInput, 348–349
Updating Position and Rotation of the First-Person
Model, 280
Using SegmentCast for Ball Movement, 328
VertexArray Declaration, 135
Z-Buffering Pseudocode, 202
Load function, 144–145, 338–339, 379
LoadActors, 438
LoadBank, 227–228
LoadBinary, 452–453
LoadComponents, 443–444
LoadData function, 44
LoadGlobalProperties, 434
loading
3D models, 190
animation data, skeletal animation, 380–385
banks, 227–229
binary mesh files, 452–454
G-buffer global lighting shader, 416
gpmesh files, 194–195
image files, 43–44
level files, 430
actors, 435–439
shaders
adding shaders to games, 146
IsValidProgram, 145
Load, 144–145
OpenGL, 141–142
SetActive, 145
Unload, 146
skeletons, 378–380
textures, 170–173
LoadJSON, 431–433
LoadLevel function, 431
LoadProperties, 436, 441
LoadShaders, 146, 415
LoadText, 359
local lighting model, 209
local pose, 368
localization, 358, 361
text maps, adding, 359–361
Unicode, 358–359
look-at matrix, 196
low-pass filter, occlusion, 243
low-quality reflections, 401
M
Mac, installing FMOD, 223
macros
offsetof macro, 175
SDL_BUTTON macro, 257
main function, 10–11
Manhattan distance heuristics, 104–105
maps, 474–475
hash maps, 475
mass, 79–80
Math::Clamp, 266
Math::Cos, 67
Math.h library
Dot function, 70
Normalize(), 66
Math::Max, 313
Math::NearZero function, 73–74
Math::Sin, 67
Math::ToDegrees, 47
matrices
look-at matrix, 196
transformations, matrix multiplication, 154–155
transformations and, 154
transforming points, 155–157
view matrix (3D), 196–197
view-projection matrix, calculating, 200
matrix multiplication, 154–155, 156
matrix palettes, 385
max players, 116
maxAngularSpeed, 277
maxMouseSpeed, 277
MaxPlayer, 119
MaxPlayerLimit function, 120
mCurrFrame, 50
membership tests, 106
memory allocation, 462–463
memory usage, reducing, 375
Mesh, 194–195
MeshBinHeader, 449–450
MeshBinHeader*, 451
MeshComponent, 439
3D, 204–206
MeshComponent::GetType, 441
MeshComponent::LoadProperties, 442
meshes, drawing 3D meshes, 195–196
Mesh::Load, 195, 452
Mesh::LoadBinary, 452–453
Mesh::SaveBinary, 450–451
Microsoft Visual Studio Community 2017, 2
Microsoft Windows
API (application programming interface), 3
setting up development environments, 2
min players, 116
MinDistSq, 300–301, 313
minimax, 118–119
MinimaxDecide, 119
mirror framebuffer, creating, 403–404
mirror texture, drawing in HUD, 406–407
MirrorCamera, 405
mirrors, rendering framebuffer objects, 404–405
mixing audio, 241–242
monolithic hierarchy, 36
mouse, 256
buttons and position, 256–258
FPS angular speed, 277
relative motion, 258–259
scroll wheels, 260–261
MouseState, 257–258
MoveComponent, 190, 276
creating, 73–75
movement, 270
first-person camera, 276–277
InputComponent class, creating, 76–79
MoveComponent class, creating, 73–75
multiplying vectors, 62–63
N
narrowphase techniques, 331
nav mesh, 115–116
NavComponent, 114
NDC (normalized device coordinates), 132–133
nearest-neighbor filtering, 395
nearest-neighbor mipmapping, 398
negative edges, 248–251
Newtonian physics, 79
node adoption, 111
nodes, 98
goal nodes, 100
parent nodes, 100
path nodes, 115
start nodes, 100
NodeToPointerMap, 101
non-player characters (NPCs), 115
normal vectors, calculating, 70–72
normalization, unit vectors, 65–66
Normalize(), 66
normalized device coordinates (NDC), 132–133
NPCs (non-player characters), 115
numeric integration, 80–81
O
object space, 149–150
occlusion, 243–244
offsetof macro, 175
OnClick, 347, 349
OnEnter, 95
OnExit, 95
OnUpdateWorldTransform, 324–325
open set, GBFS (greedy best-first search), 106
OpenGL
alpha blending, 178–180
anisotropic filtering, 399
context, 130
initializing, 128
loading shaders, 141–142
IsValidProgram, 145
Load, 144–145
SetActive, 145
Unload, 146
mipmapping, 399
rendering frames, 131
requesting attributes, 129
setting up OpenGL window, 128–129
texture mapping, 167–170
updating vertex format, 173–175
UBOs (uniform buffer objects), 165
vertex array object, 134–135
writing basic shaders, 139–141
operator overloading, 468–470
operator+, 469
optimizing A*, 113
orbit cameras, 286–288
OrbitCamera class, 287
OrbitCamera::Update, 287–288
oriented bounding box (OBB), 306
orthographic projection, 197
outMap, 101
overdraw, 200
overloading operators, 468–470
P
Pac-Man game
partial class hierarchy, 36
pseudocode, 6
state machine behaviors, 92
paddles
drawing, 20–23
updating position of, 26–28
painter’s algorithm, issues with in 3D, 200–201
parallax effect, 53
parameters, sound events, 223
parent nodes, 100
pass by value, 459
path nodes, 115
path-cost component, 111
pathfinding, AI (artificial intelligence), 98
following paths, 114–115
path nodes, 115
Patrol state, 92, 94–95
pause menu, 344–345
perspective divide, 199
perspective projection, 197
phantom inputs, 265
Phong reflection model, 209–211
Phong shading, 211
physics, Newtonian physics, 79
PhysWorld class, 325–327
testing ball collisions, 329–331
PhysWorld::TestPairWise, 329
pitch, 185
pixels, 14
PlaneActor, 331, 333
planes, 301–302
platform specific libraries, 3
player collision, against walls, 331–333
PlayEvent, 229, 230, 238
playing event instances, 229–230
PNG files, 43
point light fragment shaders, 419–421
point lights, 208
adding, 418–419
drawing, 421–424
pointers, 459–460
PointLightComponent class, adding, 419
PointLightComponent::Draw, 423–424
points, transforming, with matrices, 155–157
polling, 248
polygons, 131–132
Pong
drawing walls, balls, and paddles, 20–23
updating
ball’s position, 28–30
paddle’s position, 26–28
poses, skeletons and, 367–370
position of
balls, updating, 28–30
mouse, 256–258
paddles, updating, 26–28
positional functionality, adding to SoundEvent, 236–237
positions, computing with Euler integration, 80
positive edges, 248–251
PosNormTex format, 191–192
PrepareForUpdate, 255, 264
printf function, 8
processing, basic input processing, 11–13
ProcessInput, 11–12, 13, 251, 343
InputComponent class, 76
ProcessKeyboard function, 54
projection matrix (3D), transforming, clip space,
197–200
properties, loading global properties, 430–434
Q
quality, improving texture quality, 394
quaternions, 186–187
in code, 188–189
combining, rotations, 187
MoveComponent, 190
rotating vectors, 188
queues, 473–474
quit dialog box, 351
R
radar
adding, 354–358
drawing, 357
Random function, 75
range-based For loops, 475–476
RapidJSON, 194, 430, 449
saving level files, 444–446
raster graphics, 14
real time, 24
red, green, blue (RGB), 14
reducing memory usage, 375
references, 458–459, 464–465
reflections, 401
refresh rate, 16
RegisterState function, 96
relative motion, mouse input, 258–259
RemoteTarget, 352
RemoveActor, 41
Renderer, abbreviated declaration, 195–196
renderer, 19
Renderer::Draw, 343, 377, 405
Renderer::DrawFromGBuffer, 417
Renderer::GetScreenDirection, 294
Renderer::Initialize, 404, 411
Renderer::Shutdown, 404, 411
rendering
framebuffer objects, 404–405
frames, OpenGL, 131
to textures, 400–401
RenderText, 338, 340, 347
repositories, 3
requesting OpenGL, attributes, 129
resolution, 14
supporting multiple resolutions, 361–362
reticule, adding, aiming reticule, 352–354
reverb, 242
RGB (red, green, and blue), 14
RGBA, 14–15
roll, 185
RotateToNewForward, 327–328
rotating vectors, quaternions, 188
rotation, 152
combining for 3D, 187
Euler angles, 3D, 185–186
quaternions, 3D, 186–187
row vectors, 155
S
sample data, 226–227
SaveActors, 447–448
SaveBinary, 450–451
SaveComponents, 448
SaveLevel, 445, 446
saving
binary mesh files, 449–452
level files, 444–446
actors, 446–448
global properties, 446
scalar multiplication, 62–63
scale, 151–152
scale matrix, 157
scaling vectors, 62–63
screen tearing, 16–17
scroll wheels, mouse input, 260–261
scrolling backgrounds, 51–53
SDL (Simple DirectMedia Layer) library, 4
input devices, 251
SDL 2D coordinate system, mouse position, 257
SDL controller axis constants, 264
SDL Image, 43
SDL image file formats, 43
SDL subsystem flags, 8
SDL TTF library, 338
SDL_BUTTON macro, 257
SDL_CreateRenderer, 19
SDL_CreateTextureFromSurface, 43
SDL_CreateWindow function, 8, 128
SDL_DestroyRenderer, 19
SDL_Event, 12
SDL_GameControllerAddMappingsFromFile, 262
SDL_GameControllerGetAxis, 267
SDL_GameControllerGetButton, 264
SDL_GameControllerOpen, 261
SDL_GetKeyboardState, 13, 248, 253
SDL_GetMouseState, 256–258
SDL_GetRelativeMouseState, 259, 277
SDL_GetTicks, 25
SDL_INIT_AUDIO, 8
SDL_INIT_GAMECONTROLLER, 8
SDL_INIT_HAPTIC, 8
SDL_INIT_VIDEO, 8
SDL_Log function, 8
SDL_PollEvent function, 11–12
SDL_QueryTexture, 46
SDL_QUIT, 12
SDL_Rect, 21
SDL_RenderClear, 20
SDL_RenderCopy, 47
SDL_RenderCopyEx, 47
SDL_Renderer, 128
SDL_RenderFillRect, 20, 22
SDL_RenderPresent, 20
SDL_SCANCODE, 13
SDL_SetRenderDrawColor, 20
SDL_ShowCursor, 256
SDL_Surface, 43, 339
SDL_Texture, 44
SDL_WINDOW_FULLSCREEN, 9
SDL_WINDOW_FULLSCREEN_DESKTOP, 9
SDL_WINDOW_OPENGL, 9
SDL_WINDOW_RESIZABLE, 9
SegmentCast, 326
ball collisions, 327–328
Set3DAttributes, 236–237
set3DSettings, 241
SetActive, 137, 145, 416
SetAnimTextures function, 50
SetIntUniform, 415–416
SetListener, 235, 236
SetMatrixUniform, 164, 165, 386
setMatrixUniform, 415–416
SetName, 347
SetTexture function, 46
SetTexturesActive, 416–417
setting up
development environments, 2
Apple macOS, 2
Microsoft Windows, 2
listeners, for 3D positional audio, 234–236
OpenGL window, 128–129
SetViewMatrix, 277–278
SetVolume, 242
shader programs, 145
shaders, 138
BasicMesh.vert shader, 421
fragment shaders, 139
loading, 141–142
IsValidProgram, 145
Load, 144–145
SetActive, 145
Unload, 146
point light fragment shaders, 419–421
skinning vertex shaders, 385–387
updating, 175–176
Sprite.frag shader, 176–178
to use transform matrices, 163–167
vertex shaders, 138–139
writing basic shaders, 139
Ship::UpdateActor function, 54
Shutdown function, 2D graphics, 19
Simple DirectMedia Layer. See SDL
Simple OpenGL Image Library (SOIL), 170
SimpleViewProjection matrix, 163
single controllers, enabling, 261–262
single-threaded game loop, 5
animation data, 371–372
drawing with skinning vertex attributes, 374–378
loading animation data, 380–385
loading skeletons, 378–380
skeletons and poses, 367–370
skinning, 372–373
skinning vertex shader, 385–387
updating, 386–389
skeletal hierarchy, 367–368
SkeletalMeshComponent, 376–377, 380, 383–384, 441
SkeletalMeshComponent::Draw, 386
SkeletalMeshComponent::Update, 388–389
skeletons
loading, 378–380
poses and, 367–370
skinned animation. See skeletal animation
Skinned.vert, 374, 385–386
skinning, 372–373
skinning vertex attributes, drawing with, 374–378
skinning vertex shaders, 385–387
Slerp (spherical linear interpolation), 188
smoothstep function, 421
SnapToIdeal, 285
SOIL (Simple OpenGL Image Library), 170, 172
sound effects, third-person games, 239–240
sound events, 223
sound occlusion, 243–244
SoundEvent class, 230–233
adding positional functionality to, 236–237
SoundEvent::IsValid, 233
sounds. See audio
source code, 3
source control systems, 3
space
clip space. See clip space, 161–163
object space, 149–150
world space, 150
transforming, 150–151
spheres, 303
spline cameras, 289–292
SplineCamera class, 290
SplineCamera::Update, 291–292
Spline::Compute, 290
splines, Catmull-Rom spline, 289
spotlight, 209
springs, adding to follow cameras, 283–285
SpriteComponent, 45–46, 166
SpriteComponent::Draw, 47–48, 166–167, 173
sprites, 42, 148
animating, 48–50
drawing, 44–48
texture-mapped sprites, 178
Sprite.vert shader, 175–176
stacks, 474
start nodes, 100
state machine behaviors, AI (artificial intelligence), 92
state machines
designing, 92–93
transitions, 94
states, as classes, 95–98
static objects, 34
std::function, 346, 437, 442
std::map, 475
std::pair, 476
std::queue, 474
std::string, 359, 437
std::unordered_map, 475
std::vector, 319, 370
strafe speed, 276
streaming data, 226–227
subtraction, vectors, 61–62
surround sound, 234
Swap, 462
sweep-and-prune, 331
swept-sphere intersection, 323
swizzle, 213–214
T
TargetActor, 435
TargetComponent, 353, 354
testing ball collisions in PhysWorld, 329–331
TestPairWise, 329
tests
contains point tests. See contains point tests
instantaneous tests, 321
intersection tests. See intersection tests
line segment tests. See line segment tests
TestSweepAndPrune, 330–331
texel density, 394
texels, 394
text, localization, 361
text maps, adding, 359–361
texture coordinates, 168, 395
updating, vertex format, 173–175
texture quality, improving, 394
texture sampling, 395–397
Texture::Load, 172–173
texture-mapped sprites, 178
textures
drawing, 47
loading, 170–173
rendering to, 400–401
Texture::SetActive, 173
Texture::Unload, 173
third-person games, listeners, 239–240
tic-tac-toe, game trees, 116–117
ToMatrix, 369
tracking
event instances, 230
loaded banks and events, 227
tracks, 380
transform matrices
3D, 184–185
updating shaders, 163–167
transformation matrix, 155
combining, 152–153, 159
matrices, 154
matrix multiplication, 154–155
TransformComponent, 37–38
transforming
clip space
projection matrix (3D), 197–200
view matrix (3D), 196–197
points, with matrices, 155–157
from world space, to clip space, 161–163
world space, 150–151, 157
rotation, 152
scale, 151–152
scale matrix, 157
translation, 151
Transform.vert, 163
TransformWithPerspDiv, 293
transitions, 92
state machines, 94
translation, 151
triangles, 131–132
drawing, 146–148
fragment shaders, 139
normalized device coordinates (NDC), 132–133
vertex and index buffers, 133–138
triggers, 34, 264
trilinear filtering, 398, 399, 400
TTF_OpenFont, 338
TTF_Quit, 338
TTF_RenderText_Blended, 339
turning on anisotropic filtering, 400
TurnTo function, 114
types of game objects, 34
U
UBOs (uniform buffer objects), 165
UIScreen, 340–342
UIScreen::DrawScreen, 348
UIScreen::DrawTexture, 406
UIScreen::ProcessInput, 348–349
undirected, 98
Unicode, 358–359
uniform buffer objects (UBOs), 165
uniform cost search, 114
unit quaternions, 186
unit vectors, 65–66
Unload function, 146
unloading banks, 227–229
Unproject, 294
unprojection, 292–294
unweighted graphs, 98–99
Update function, 39, 50, 388–389
UpdateActor, 39
UpdateComponents, 39
UpdateCrosshair, 353
UpdateGame, 25, 26, 41–42
UpdateMinMax, 304, 305
updating
animations, skeletal animation, 386–389
balls, position of, 28–30
games, 23
delta time, 24–28
game time, 24
real time, 24
paddles, position of, 26–28
shaders
Sprite.vert shader, 175–176
to use transform matrices, 163–167
vertex attributes, 193–194
vertex format, 173–175
user interfaces (UI)
buttons, 346–350
dialog boxes, 349–352
font rendering, 338–340
localization, 358, 361
text maps, 359–361
Unicode, 358–359
pause menu, 344–345
UTF-8, 358
UV coordinates, 395
V
VAO (vertex array object), 150
variable time steps, 80–81
VecToFMOD, 234–235
vector addition, 63–64
vector subtraction, 61–62
vectors, 59–61, 471–472
column vectors, 155, 156
combining, 63–64
converting forward vectors to angles, 67–68
converting from angles to forward vectors, 66–67
determining angles between, dot product, 68–70
determining directions, 65–66
determining distance, length, 64–65
normal vectors, calculating, 70–72
rotating, with quaternions, 188
row vectors, 155
scaling, 62–63
velocity, 79–80
.vert extension, 139
vertex array object (VAO), 150
vertex, creating, 163
vertex attributes
lighting, 206–207
updating, 193–194
vertex buffers, 133–138
vertex format, updating, 173–175
vertex normal, 206
VertexArray class, 135, 452
VertexArray constructor, 375–376
vertical synchronization (vsync), 17
vertices, 98
view matrix (3D), transforming, clip space, 196–197
view-projection matrix, 162
calculating, 200
Visual Studio, 2
vsync (vertical synchronization), 17
W
walls
drawing, 20–23
player collision against, 331–333
waypoint graphs, 115
weighted graphs, 99
weights, edges, 98–99
Window Creation flags, 9
Windows, FMOD, installing, 223
windows, creating blank windows, 10–11
world space, 150
combining transformations, 152–153
transforming, 150–151, 157
to clip space, 161–163
rotation, 152
scale, 151–152
scale matrix, 157
translation, 151
world transform matrix, calculating, 189
world transforms, adding to actors, 159–161
writing
basic shaders, 139
to G-buffer, 412–414
X
Xcode, installing, 2
.xyz syntax, 213–214
Y
yaw, 185
Z
z-buffer, 201
z-buffering, 201–203
Code Snippets

Game Programming in C++ - Creating 3D Games

Uploaded by

Copyright:

Available Formats

Game Programming in C++ - Creating 3D Games

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Game Programming in C++ - Creating 3D Games

Uploaded by

Copyright:

Available Formats

Contents

1. Setting Up a Development Environment

2. Getting This Book’s Source Code

1. The Color Buffer

1. Real Time and Game Time

12. Chapter 2: Game Objects and 2D Graphics

1. Types of Game Objects

1. Loading Image Files

1. Getting a Vector between Two Points: Subtraction

1. Creating a Basic MoveComponent Class

1. Linear Mechanics Overview

4. Basic Collision Detection

14. Chapter 4: Artificial Intelligence

1. State Machine Behaviors

1. Designing a State Machine

15. Chapter 5: OpenGL

1. Setting Up the OpenGL Window

5. Matrices and Transformations

1. Loading the Texture

16. Chapter 6: 3D Graphics

1. The Actor Transform in 3D

1. Transform Matrices for 3D

1. Choosing a Model Format

1. Transforming to Clip Space, Revisited

1. Revisiting Vertex Attributes

17. Chapter 7: Audio

1. Setting Up a Basic Listener

3. Mixing and Effects

18. Chapter 8: Input Systems

1. Buttons and Position

1. Enabling a Single Controller

1. Basic First-Person Movement

1. Basic Follow Camera

20. Chapter 10: Collision Detection

1. Contains Point Tests

1. The BoxComponent Class

21. Chapter 11: User Interfaces

1. The UI Screen Stack

1. Adding an Aiming Reticule

1. Working with Unicode

22. Chapter 12: Skeletal Animation

1. Foundations of Skeletal Animation

1. Skeletons and Poses

2. Implementing Skeletal Animation

1. Drawing with Skinning Vertex Attributes

23. Chapter 13: Intermediate Graphics

1. Improving Texture Quality

1. Creating the Texture

1. Creating a G-Buffer Class

24. Chapter 14: Level Files and Binary Data

1. Level File Loading

1. Loading Global Properties

2. Saving Level Files

1. Saving Global Properties

1. Saving a Binary Mesh File

25. Appendix A: Intermediate C++ Review

EPUB is an open, industry-standard format for e-

Boston • Columbus • Indianapolis • New York • San

1 Game Programming Overview