Course Slides 2018

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/325581786
Digital Systems Design Course Slides
Presentation · June 2018

DOI: 10.13140/RG.2.2.26189.74720
CITATION READS
1 11,629
1 author:
Reza Sameni
Emory University
163 PUBLICATIONS 4,598 CITATIONS
SEE PROFILE
All content following this page was uploaded by Reza Sameni on 05 June 2018.
The user has requested enhancement of the downloaded file.

DIGITAL SYSTEMS DESIGN
Reza Sameni, PhD
Email: rsameni@shirazu.ac.ir
Web: www.sameni.info
School of Electrical & Computer Engineering

Shiraz University, Shiraz, Iran
Winter-Spring 2018
Revision: June 2018

Copyright Notice
 The current lecture notes are provided as handouts for the Digital
Systems Design course presented at Shiraz University for
educational purposes.
 Some images and source codes (cited within the text) have been
adopted from books, papers, datasheets, and the World Wide
Web; but may be subject to copyright.
 Please use the following for citing these notes:

R. Sameni, “Digital Systems Design Lecture Notes”, School of
Electrical & Computer Engineering, Shiraz University, Shiraz,
Iran, version 2018.
Preface
• History: The present lecture notes have been prepared for
an undergraduate course in Digital Systems Design,
presented in the School of Electrical and Computer
Engineering of Shiraz University between 2009 to 2018.
• Prerequisites: Logical Circuits, Computer Architecture,
Signals & Systems. It is highly recommended that the
students take this course simultaneously with the Digital
System Design Lab, which is designed based on the
course content.
• Evaluation: The grading is based on exercises (15%),
midterm (35%) and final (50%) examinations.
Table of Contents
• Part I: Architecture
• Part II: Electronic Design Automation
• Part III: Advanced Topics in Digital Systems Design
and Implementation
PART I
Architecture
INTRODUCTION
7
The von Neumann Architecture

The von Neumann Architecture (Bobda 2007)
8

The VN Architecture consists of:
1. A memory for storing program and data (Harvard

architectures contain two parallel accessible memories for
storing program and data separately)
2. A control unit (also called control path) featuring a program
counter that holds the address of the next instruction to be
executed.
3. An arithmetic and logic unit (also called data path) in which
instructions are executed.
9

The execution of an instruction on a VN computer can be done in five
cycles:
1. Instruction Fetch (IF): An instruction is fetched from the memory
2. Decoding (D): The meaning of the instruction is determined and the
operands are localized
3. Read Operands (R): The operands are read from the memory
4. Execute (EX): The instruction is executed with the read operands
5. Write Result (W): The result of the execution is stored back to the
memory
10
Pipelining in von Neumann Architectures

• Pipelining or instruction level parallelism (ILP) can be used to
optimize the hardware utilization as well as the performance of
programs.
• ILP does not reduce the execution latency of a single execution,
but increases the throughput of a set of instructions.
• The maximum throughput is dictated by the impact of hazards in
the computation. Hazards can be reduced, e.g., by the use of a
Harvard architecture.
An ideal VN Pipeline
11
Pipelining in von Neumann Architectures

Ideal pipelining is commonly unachievable. For example, the
Harvard architecture pipeline is as follows:
Harvard Architecture Pipelining

12
Application Specific Hardware versus von

Neumann Architectures
VN versus ASIP: Considering 5 cycles per instruction, the VN should be 15 times

faster than the ASIP to outperform its speed (Bobda 2007)
Application specific hardware have higher performance, at a cost of lower flexibility

13
Flexibility vs. Performance
Flexibility vs. performance of different architectures (Bobda 2007)

14
Applications of Reconfigurable Architectures
• Rapid prototyping
 reduced time-to-market
• In-system customization
 hardware updates and patches
• Remote reconfiguration
 via RF links for telecommunication BTS, spacecrafts, satellites,…
• Multi-modal computation
 Environment aware hardware
• Adaptive computing systems
 Machine learning applications
15
References
• Bobda, C. (2007). Introduction to reconfigurable

computing: architectures, algorithms, and applications.
Springer Science & Business Media.
PROGRAMMABLE LOGIC
DEVICES &
TECHNOLOGIES
17
PLD Technologies
Programmable Logic Devices (PLD) have a long history
(longer than conventional VN architecture CPUs):
• PROM
• Logic Chips
• SPLD: PLA & PAL
• CPLD
• FPGA
• ASIC
18
Programmable ROM (PROM)

Example 1: PROM-based Combinatorial logic
19
Programmable ROM (PROM)

Example 2: PROM-based state machine
Benefit: Any logic circuit may be implemented

Drawback: Low speed
20
Standard Logic Chips
• TTL (Transistor-Transistor Logic) Technology:

• The 74xxx-series
• CMOS Technology:
• The 4xxx-series
21
74000 Sub-series
• 74LS74: Low-power Schottky
• 74HCT74: High-speed CMOS
• 74HCT: 74LS TTL-compatible inputs
• SN74F00: Fast logic
22
Design Example Using the 74000 Series

f (a, b, c)  ab  b c
23
Design Example Using the 74000 Series
An Implementation of a 4-bit two register computer, including 6 CPU assembly instructions:

READ (read input), INCB (increment register B), MOVAB (move contents of register A to B),
MOVBA (move contents of register B to A), RETI (return from interrupt), JMP (jump).
Reference: http://en.wikipedia.org/wiki/7400_series
24
Programmable Logic Technologies

• Basic Idea: Logic functions can be realized in sum-of-
product form.
f ( x, y, z)  xy  xyz
Technologies:
• Simple PLD (SPLD)
• Complex PLD (CPLD)
25
PLA (Programmable Logic Array) Technology

The basic concept: An arbitrary sum of product generator
26
PLA (Programmable Logic Array) Technology
Example
27
PAL (Programmable Array Logic) Technology

• PLA has both AND and OR programmable gates; but PAL
has only programmable AND gates and the OR gates are
fixed
28
PAL Extra Circuitry
Macrocell
*Input – Output selection

29
CPLD (Complex Programmable Logic Device)

CPLDs can be considered as a set of PAL-like blocks with a
set of reconfigurable interconnection network
30

31
Equivalent Number of Gates

• We need a measure to compare the computation power of
different PLDs:
• Equivalent gates: total number of two input NAND gates

• Example: If 1 Macrocell ≈ 20 NAND gates, a 1000
Macrocell CPLD is roughly equivalent to 20,000 NAND
gates
32

33
FPGA (Field-Programmable Gate Array)

• FPGAs are extensions of the idea of PROMs for logic
circuit realization
34
FPGA Configurable Logic Blocks (CLB)
Two-input Look Up Table (LUT)

35

Example:
36
Three-input Look Up Table (LUT)

37
Programmed FPGA Scheme

38
Programmable Switches
• Other switching technologies: Flash-based, Anti-fuse, etc.

39
FPGA Logic Block Extra Circuitry

FPGA logic blocks require extra circuitry for sequential logic,
routing, I/O interface, etc.
40
CPLD vs. FPGA
• CPLD keeps its contents without power, also known as

non-volatile
• FPGA storage cells are volatile (lose their contents when

power is switched off)
41
PLD Programming
(a) Off-board programming (b) On-board programming using

JTAG*
* JTAG: Joint Test Action Group

42
JTAG
JTAG is a serial interface technology. The connector pins are:
•TDI: Test Data In
•TDO: Test Data Out
•TCK: Test Clock
•TMS: Test Mode Select
•TRST: Test Reset (optional)
JTAG Daisy Chain

43
Other JTAG Applications

Examples:
•Boundary Scan: the ability to set and read the values on pins without direct physical access
•Xilinx ChipScope Technology: for in-system run-time debugging
Ref: https://www.xjtag.com/about-jtag/jtag-a-technical-overview
44
ASIC (Application Specific Integrated Circuit)
• In CPLD and FPGA programmable switches consume

much space and reduce speed
• Alternatively, the chip layout can be totally customized;

but is very expensive
• Compromise: The design may be simplified by using

standard-cell or gate-array technologies
45
ASIC Standard-Cell Technology
Standard logic blocks are provided by

manufacturers as libraries that may be
connected
46
ASIC Gate-Array Technology
Only some parts of the chip are prefabricated

47
PLD Packages
• Plastic Leaded Chip Carrier (PLCC)
48
PLD Packages
• Small Outline Integrated Circuit (SOIC)
• Plastic Small Outline Package (PSOP)
49
PLD Packages
• Thin Small Outline Package
50
PLD Packages
• Pin Grid Array (PGA)
51
PLD Packages
• Ball Grid Array (BGA)
52
PLD Leading Companies

• Xilinx
• Altera
• Actel
• Lattice
• QuickLogic
53
Xilinx®
http://www.xilinx.com/
54
Altera®
http://www.altera.com/
• Altera was acquired by Intel® in 2015

55
Actel®
http://www.actel.com/
56
Lattice®
http://www.latticesemi.com/
57
QuickLogic®
http://www.quicklogic.com/
58
The FPGA Market (2013)
Reference: http://www.eetimes.com /
59
The Programmable Market Share (2009)
Reference: www.xilinx.com
60
Xilinx Revenue Breakdown (2009)
Reference: www.xilinx.com
61
Asia Pacific FPGA Market Size by

Application, 2012-2022 (USD Million)
Reference: https://www.gminsights.com/industry-analysis/field-programmable-
gate-array-fpga-market-size
62
References:
• S. Brown and Z. Vranesic, Fundamentals of Digital Logic
with Verilog Design, McGraw-Hill, 2003, Chapter 3
• B. Zeidman, Designing with FPGAs & CPLDs, CMP Books,
2002, Chapter 1
FPGA INTERNAL
ARCHITECTURE
64
FPGA Internal Architecture

Current FPGA devices consist of:
• Configurable logic
• Interconnect network
• Device-dependent peripherals and IP cores
65
Typical FPGA Architectures

(an academic classification)
• Fine Grained (homogeneous)

• Medium Grained
• Coarse Grained (heterogeneous)
 From top to bottom the logic blocks become more complex and
advanced.
 Node-Based Reconfigurable Architectures: Imagine a network
of computers and programmable devices, which can be
reconfigured on-demand
Current FPGA architectures are considered medium grain in this
classification
66
Basic FPGA Architectures

67
Actual FPGA Architectures

68

69

70
Multipurpose Logic Blocks

Logic blocks are commonly multi-purpose:
• Shift Registers
• Memory (RAM)
• Look-up-tables (LUT)
71
Configurable Logic Blocks

72
Hierarchical FPGA Architecture
The left part slices of a CLB (SLICEM) can be configured either as combinatorial logic, or can be use as 16-bit SRAM or as
shift register while right-hand slices. The SLICEL can only be configured as combinatorial logic.
73
Slicing
74
Hierarchical FPGA Architecture

• Current FPGA internal architectures have a sort of
hierarchical design, both, in their CLB and interconnection
networks:
• Xilinx Terminology: Logic Cells, Slices, Configurable Logic Blocks
• Altera Terminology: Logic Element, Logic Array Block
Why?
75
Xilinx Logic Blocks

76
Xilinx vs. Altera Logic Cells

77
Altera Logic Blocks
• Question: How do companies decide about their FPGA internal

architecture? Is it a only matter of technology or taste?
78
Interconnect Networks
79
Interconnect Networks
80
Other Peripherals within Contemporary FPGA

• Block Memories
• Digital Clock Managers
• Dedicated Adders & Multipliers
• Variety of I/O interfaces
• Embedded Processors
81
Embedded Block Memories

82
Embedded Multiplier, Adder, MAC

83
Clock Trees
84
Clock Management
Usage:
1. Jitter removal
2. Frequency synthesis
3. Phase shifting
4. Clock de-skewing
85
1. Jitter Removal
86
Jitter Specifications
87
Clock Jitter Issues

Impacts of clock jitter on signal quality and data processing:
• Analog domain:
• Critical; causes phase modulation (distortion) and directly translates
into phase noise
• Digital domain:
• Critical; when working with multiple clocks and different clock domains
• Irrelevant; in single-clock systems and if the clock jitter is far smaller
than the clock period.
Note: Clock jitter is a random variable with a probability distribution
function.
88
Practical Jitter Measurement

Using an oscilloscope:
Further Reading: Agilent® and Tektronix® clock

analysis tools
89
Jitter Calculation
90
Jitter Calculation
91
Jitter Calculation
Example 2 (Cascaded DCMs)
Assume that the input clock has 150 ps (±75 ps) of period jitter. Assume that DCM (A) uses
the CLK2X output. Use the Spartan-3 Data Sheet specification called CLKOUT_PER_JITT_2X
for the DCM output jitter, estimated here as 400 ps (±200 ps). Assume that DCM (B) uses the
CLKDV output with an integer divider value. Use the Spartan-3 Data Sheet specification called
CLKOUT_PER_JITT_DV1 for the DCM output jitter, estimated here as 300 ps (±150 ps).
Finally, assume that DCM (C) phase shifts the output from DCM (B) by 90°. Use the Spartan-3
Data Sheet specification called CLKOUT_PER_JITT_90 for the DCM output jitter, estimated
here as 300 ps (±150 ps).
92
Recommendations for Clock Jitter Minimization
• Properly Design the Power Distribution System

• Properly Design the Printed Circuit Board
• Obey Simultaneous Switching Output (SSO) Recommendations
• Place Virtual Ground Pins Around DCM Input and Output
Connections
93
2. Frequency Synthesis
94
3. Phase Shifting
95
4. Clock De-skewing
96
Clock Generator Architectures

• Delay-Locked Loop (DLL)
• Phase-Locked Loop (PLL)

97
Delay-Locked Loop (DLL)

98
Phase-Locked Loop (PLL)
Further Reading: http://pages.hmc.edu/harris/cmosvlsi/4e/lect/lect22.pdf

99
PLL vs. DLL Architecture
Further Reading:
 Basics of DLLs: https://open4tech.com/phase-and-delay-locked-loops-basics
 Control Models of PLLs and DLLs: http://pages.hmc.edu/harris/cmosvlsi/4e/lect/lect22.pdf
100
Xilinx Digital Clock Manager (DCM)
Reference: http://www.xilinx.com/support/documentation/application_notes/xapp462.pdf
101
Xilinx DCM Clock Synthesis Options

102
Xilinx DCM Functional Overview

103
Internal Clock De-skewing

104
External Clock De-skewing
Reference: http://www.xilinx.com/support/documentation/application_notes/xapp462.pdf
105
DCM Cascading
106
General Purpose I/O

• The concept of I/O Banks
107
Xilinx Series 7 I/O Banks

The 7 series Xilinx FPGAs offer:
 High-performance (HP) I/O Banks: designed to meet the
performance requirements of high-speed memory and other
chip-to-chip interfaces with voltages up to 1.8V
 High-range (HR) I/O Banks: designed to support a wider
range of I/O standards with voltages up to 3.3V
 Different I/O voltage standards: 3.3V, 2.5V, 1.8V, 1.5V,
1.35V, 1.2V, which includes LVTTL, LVCMOS, etc. standards
 Digitally-controlled impedance (DCI) and DCI cascading: in
the HP mode
108
Xilinx FPGA I/O Planning and I/O Banks

109
Digitally Controlled Impedance

Example:
110
Xilinx Series 7 I/O Buffers

• Various I/O buffers are supported on standard FPGA devices:
111
Giga-Bit Transceiver Bus

112
IP Cores
• Hard IP
• In the form of pre-implemented blocks such as microprocessor
cores, gigabit interfaces, multipliers, adders, MAC functions, etc.
Example: Xilinx PowerPC
• Soft IP
• Source-level library of high-level functions that can be integrated in
a custom design.
• Firm IP
• Libraries which have already been optimally mapped, placed, and
routed into a group of programmable logic blocks (and possibly
combined with some hard IP blocks like multipliers, etc.) and may
be integrated into a custom design.
Example: Xilinx MicroBlaze
113
Hard Embedded Processors

114
Soft Embedded Processors: Microblaze

115
Xilinx Microblaze Core Block Diagram

MicroBlaze™ is Xilinx 32-bit RISC Harvard architecture soft processor core with a
rich instruction set optimized for embedded applications.
Key Features & Benefits:

• Over 70 user configurable options
• 3-stage pipeline for optimal footprint, 5-stage pipeline for maximum performance
• Supports either PLB or AXI interface
• Big-endian or Little-endian support
• Optional Memory Management Unit (MMU)
• Optional Floating Point Unit (FPU)
• Instruction and Data-side Cache
116
FPGA with Embedded Microprocessors
• Question: Why?
Reference: http://www.gartner.com
117
Xilinx Intellectual Property (IP) Cores

• Dozens of soft IP cores are provided by Xilinx and other
vendors, which can be integrated into a custom design
118
How to Choose an FPGA?

1. Study the problem of interest.
2. Start by a preliminary system design to find a rough
estimate of the resources and system clock that you might
need. This might need some simulations, writing some
parts of the HDL code or putting together predesigned
libraries or IP cores.
3. Choose/design an appropriate (overestimated) FPGA
board for your application.
4. Proceed with the detailed design and implementation
119
Xilinx 7 Series FPGAs Overview

120
Xilinx Virtex-7 Feature Summary

121
FPGA Package Numbers

FPGA ordering information (visibly marked on the IC package)
FPGA DESIGN FLOWS
123
PLD-Based Design Flows

1. Design Idea
2. Architectural Design
3. Design Entry
4. Behavioral Simulation (Top-Down and/or Bottom-Up)
5. Register Transfer Level (RTL) Simulation/Implementation
6. Synthesis
7. Technology Mapping
8. Placement & Routing
9. FPGA/CPLD Configuration using Bitstreams
10. Final In-System Testing
11. Fully customized IC or ASIC Fabrication
12. Gate Level & Timing Simulation/Implementation For ASIC
13. Switch Level & Device Simulation/Implementation Design Only
14. Final Circuit Testing
124
FPGA vs. Microprocessor Implementation Flow
Microprocessor FPGA
Architectural design Architectural design
Choice of language (C, JAVA, etc.) Choice of language (Verilog, VHDL. etc.)
Editing programs Editing programs

Compiling programs (.DLL, .OBJ) Compiling programs
Synthesizing programs (.EDIF)
Linking programs (.EXE) Placing and routing programs (.VO, .SDF, .TTF)
Loading programs to ROM FPGA configuration with bit files
Debugging programs Debugging FPGA programs
Documenting programs/design Documenting programs/design
Delivering programs Delivering programs
125
1. Design Entry
Utilities for design entries:
• Schematic Editors
• e.g., Altium’s FPGA-ready Design Components and FPGA Generic
• Hardware Description Languages (HDL)

• e.g., Verilog, VHDL, etc.
• Finite State Machine (FSM) Editors

• e.g., ActiveHDL® FSM editor
• System Level Tools, known as HLS

• e.g., Matlab Simulink and Xilinx System Generator
126
2. Functional Simulation
• Behavioral Simulation; not necessarily implementable on hardware
• Structural Simulation; can simulate bitwise accurate models of the
final hardware
127
3. Logic Synthesis
HDL  Boolean Equations  Technology Mapping
• The output of the synthesis stage is a Netlist including all

the hardware modules and their interconnections
• Various Netlist Standards exist
• Electronic Design Interchange Format (EDIF)
• Xilinx Netlist Format (XNF)
• …
Necessity of standard tools: Consider ‘N’ vendors with

distinct standards; N2 translators are required to
interchange formats in between
128
Summary of Xilinx FPGA Design Flow

1. Synthesis: converts HDL (VHDL/Verilog) code into a gate-
level netlist, represented in the terms of the UNISIM
component library (a Xilinx library containing basic
primitives).
2. Translate: merges the incoming netlists and constraints
into a Xilinx® design file.
3. Map: fits the design into the available resources on the
target device, and optionally, places the design.
4. Place and Route: places and routes the design to the
timing constraints.
5. Generate Programming File: creates a bitstream file that
can be downloaded to the device.
PART II
Electronic Design Automation
HARDWARE DESCRIPTION
LANGUAGES
131
Hardware Description
How can we describe a hardware?
1. Schematic design tools: Visual schematic editors. e.g.,
Altium®, Protel®, OrCAD®, Xilinx PlanAhead®, etc.
2. Hardware description languages: Verilog, VHDL, etc.
3. Set of libraries and classes in software languages
4. Any other?
132
Hardware Description Examples

• Schematic editors: Altium® general FPGA design library
133

• Schematic editors: Xilinx Schematic Tools
134

• A C++ library to simulate hardware functionality
135

• Hardware Description Languages
Verilog VHDL SystemC

136
From Schematic Editors to Hardware

Description Languages
• Hardware description languages are textual means of describing a
hardware
• Text is better than pictures and Karnaugh maps; as it’s more
simple to handle and analyze for language parsers and synthesis
tools
Half-Adder
Full-Adder
137
Hardware Description Languages (HDL)

Examples of HDL languages
• VHDL
• Verilog
• SystemC
• SystemVerilog
• JHDL
• Handel-C
• Impulse C
•…
138
Hardware Description
What should a HDL look like and what features should it have?
1. Cover different levels of abstraction: transistor level, gate level,
register transfer level (RTL), system level
2. Applicable for different architectures: CPLD, FPGA, ASIC, etc.
3. Provide a unique description for all synthesizable hardware
4. Ability of accurate simulation before implementation. The language
should be able to simulate other functionalities required for
hardware description and simulation: generating synthetic
waveforms, reading/writing test vectors from/to files, setting time
bases, etc.
5. Convertible into conventional data structures such as trees and
graphs for algorithmic simplifications and optimizations
6. Existence of tools (tool chains) for translating the “hardware
description” into “hardware”
139
Current HDL Languages

• Common HDL languages support different levels of abstraction
plus additional features used for simulation, modeling, and
documentation of hardware (not necessarily synthesizable on
hardware)
• The languages can be used for hardware: design, simulation,
modeling, test, documentation
• Note: HDL languages do not generate executable codes; they
describe hardware, which are later translated into hardware by
electronic design automation (EDA) tools
140
Verilog HDL
We use Verilog HDL in this course, because
• It has all the required features of a complete HDL
• It has a rather simple syntax
• It is not as verbose as VHDL
• It is highly popular in industry (for RTL design)
Our major references:

• S. Palnitkar, Verilog HDL: A Guide to Digital Design and Synthesis, 2nd Ed.,
2003
• WF Lee, Verilog Coding for Logic Synthesis, John-Wiley, 2003
• Xilinx XST user’s manual, 2009
141
Verilog HDL History

• Verilog was created by Prabhu Goel, Phil Moorby, Chi-Lai Huang and
Douglas Warmke in 1983-1984, as a hardware modeling language.
• Verilog was originally owned by Automated Integrated Design
Systems (later renamed as Gateway Design Automation).
• Gateway Design Automation was purchased by Cadence Design
Systems in 1989.
• In 1990, Cadence put the language into the public domain, with the
intention that it should become a standard, non-proprietary language.
• Versions: Verilog-95, Verilog 2001, Verilog 2005, SystemVerilog (a
superset of Verilog 2005).
• Latest versions of Xilinx ISE® support Verilog 2005.
• Xilinx Vivado® supports Verilog 2005 and SystemVerilog.
142
Verilog HDL Syntax

Let’s start with a list of the most common digital hardware elements that we
know:
• Gates: AND, OR, NOT, XOR,…
• Electronic features/elements: Wires, buffers, tristate buffers, impedance levels,…
• Multiplexers, encoders and decoders
• Finite state machines (FSM)
• Memories: RAM, ROM, dual-port vs. single port memories
• Shift registers, Barrel shifters, etc.
• Initialization and resetting mechanisms
• Combinational logic: a combination of logical components
• Sequential logic: registers (flip-flops)
• Arithmetic units: half and full adders, multipliers, counters, timers, etc.
• Logic chips: ICs with predefined timing and digital function
• Logic circuit peripherals: I/O interface, clock management
• User-defined constraints and port mapping
These elements are from different levels of abstraction; but any HDL should
be able to “describe” them.
143
Verilog HDL Syntax

 Verilog is a free-form language (the positioning of characters on the programming
page is insignificant)
 Combinations of numbers (0,1,2,…), letters (a, b,…,z, A, B, C, …,Z),
underscore (_) and Dollar-sign ($) can be used in variables.
 Variable names are case-sensitive
 Variable names may not start with $ (Verilog system commands start with $)
 Underscores can be used between numbers as separators for better readability
 Single-line comments: // All text is considered as comment hereafter
 Single- or multi-line comment blocks: /* bla bla bla */
 Synthesis tools occasionally use comment blocks to define synthesis attributes
(user-defined properties of a block of code) in specific formats:
reg my_reg /* synthesis maxfan = 16 preserve */;
LUT4 U1 (.O(O), .I0(I0), .I1(I1), .I2(I2), .I3(I3));

// synthesis attribute INIT of U1 is "8000"
144
Nets and Registers

• The wire keyword is used to define nets (wires) and results of combinational logic
(using an assign command).
• The reg keyword is used to define registers and results of sequential logic (in an
always block). The exception is a combinational logic defined by an always block.
• Wires and registers can be defined and assigned in vector form.
145
Logic Values
• Verilog supports four logic values
Logic Value Description/Usage
0 zero, low or false
1 one, high or true
z or Z high-impedance, tristates, floating (dangling)
x or X unknown, uninitialized, collision
• Sized vs. unsized values

146
Assignments
• The assign keyword is used to connect wires and to define
single-line combinational logic.
147
Module Definition
inputs module
output
module name
T O
wire
instances
Instance name
Note: A module may not be defined inside another

module; but it can be instantiated.
148
Module Port Types

Three types of ports are available in
Verilog:
1. input: for giving input to a module
2. output: for getting outputs from a
module
3. inout: bidirectional ports which can
send/receive data depending on a
control line. Inouts ports should be
realized using tristate buffers with
appropriate control.
149
Input, Output, and Inout Port Usage in Verilog
realizes a tri-state buffer
Inout mechanism
150
Module Instance Port Mapping

• Two types of instance port mapping are supported in Verilog:
unconnected (dangling) port
Note: Port order is not important when using “by name” mapping
Note: All module ports (input, output, inout) are wires

151
Module Port Declaration

• Two forms of port declaration are possible:
Note the difference
Comments
152
Built-in and Device-Dependent Primitive Elements

• Verilog has several built-in primitive elements (switches, gates, etc.), which can
be instantiated as modules: and, nand, not, nor, or, xor, xnor, buf,
bufif0, bufif1, rtranif1, nmos, pmos, rpmos, tran, rtran, pullup,
pulldown, cmos, rnmos, tranif1, tranif0, notif0, notif1,
rtranif0, rcmos
Example:
strength levels delay parameters
• There are also device- and technology-dependent primitives:
Further Reading: http://electrosofts.com/verilog/primitives.html

153
Always Blocks
• An always block is used to define, both, combinational and sequential
logic blocks.
• Registers may only be assigned inside an always block (although they
may represent combinational logic).
• Variables assigned in an always block should all be defined as reg
Flip-flop inferred
sensitivity list
equivalent
No flip-flops inferred!
equivalent
154
D-Type Flip-Flops (A Review)

• D-type flip-flops are the basic elements used for sequential logic design.
155
Xilinx D-Type Flip-Flops

• According to Xilinx 7 Series Manual:
156
Always Blocks
 The following two pieces of code are identical (five flip-flops are inferred in total):
 We see that the always block has abbreviated the explicit declaration of five flip-
flops
 Note: All always procedures with the same sensitivity list are concurrent. They
describe parallel flip-flops, which share a common clock.
 Note: The sequence of writing wire assignments, always blocks and their internal
assignments are irrelevant; timing is manages by data-flow and state controllers,
not by code line execution orders
 Question: What issues can raise when code line sequences become irrelevant?
157
Always Block Issues (1)

Question 1: What happens if a single variable is simultaneously assigned in
multiple always blocks?
Answer: The Verilog syntax does not allow this (a register may only be assigned
in a single always block). Problem solved!...
Question: What if we really need to change a register value by two different

clock?
Answer: We need to find another way for this later (using indirect assignments).
158

Question 2: What happens if data dependency exists between two register
assignments?
vs.
Answer: Race condition; we need to find a solution.

Solution: Verilog has two different assignment operators: Blocking and Non-
blocking
159
Blocking vs. Non-Blocking Assignments

Syntactic difference:
Blocking assignment: Evaluation and assignment are somehow immediate (blocks all other
assignments and evaluations that use the same variable)
Non-blocking assignment: All assignments that use the variable are deferred until all right-hand
sides have been evaluated (end of simulation time-step)
Guideline: Blocking assignments are only used for combinational logic description. Use non-
blocking assignments for sequential register assignment.
Further Reading: http://courses.csail.mit.edu/6.111/f2007/handouts/L06.pdf
160

Question: What happens if two always blocks (with different sensitivity
lists) have data dependency between their register assignments?
Answer: Race condition; no syntactic solutions exist for this issue. Should
be avoided/resolved by proper design.
Example: Passing data between different clock domains.
161
Signal Drive Strength

• In logic circuit design, nets can have different strength levels
ranging from supply/ground (strongest) to high-impedance
(weakest).
• Verilog supports various strength levels to model the driving
strength phenomenon:
Strength level Description Keywords Degree
Supply drive Power supply connections supply0, supply1 7 (strongest)
Strong drive Default gate and assign output strength strong0, strong1 6
Pull drive Gate and assign output strength pull0, pull1 5
Large capacitor Size of trireg net capacitor large 4
Weak Gate and assign output strength weak0, weak1 3
Medium capacitor Size of trireg net capacitor medium 2
Small capacitor Size of trireg net capacitor small0, small1 1
High impedance High Impedance highz0, highz1 0 (weakest)
162
Signal Strength Collisions
supply1 strong1 pull1 weak1 highz1

supply0 x 0 0 0 0
strong0 1 x 0 0 0
pull0 1 1 x 0 0
weak0 1 1 1 x 0
highz0 1 1 1 1 z
Further Reading: http://verilog.renerta.com/source/vrg00047.htm

163
Signal Drive Strength Examples

Logic strength levels
164
Parameterized Module Design

• Verilog supports parametric module definitions
• Example 1: A parametric-length multiplexer
165
Parameterized Module Design

Example 2: A parametric full-adder
Ref: http://referencedesigner.com/tutorials/verilog/verilog_23.php
166
For-Loops in Verilog
• For-loops in their software-like
usage are not synthesizable in
Verilog.
• Question: Why?
• In synthesizable Verilog codes,
for-loops are merely used for
writing shorter scripts that
generate codes.
• We will learn alternative code
generation methods in later
sections.
167
For-Loops in VHDL
168
Hardware Description Language

(HDL) Standard Coding Techniques
• In the sequel we study standard design entries and coding styles, which
guarantee synthesizable codes for low-level implementation using EDA tools.
• Major Reference: Xilinx XST User Guide, UG627 (v 11.3) September 16, 2009.
URL: https://www.xilinx.com/support/documentation/sw_manuals/xilinx11/xst.pdf
169
Flip-Flop with Positive-Edge Clock

Verilog VHDL
170
Flip-Flop with Positive Edge Clock

with INITSTATE of the Flop Set
Verilog VHDL
Drawback: Only provides power-on

initialization, does not have run-time
resetting property.
171
Flip-Flop with Negative-Edge Clock

and Asynchronous Reset
Verilog VHDL
172

and Synchronous Set
Verilog VHDL
Note: Verilog and VHDL have if and

else in their syntax (as in software
languages); but with totally different
interpretations: “if-elses are not
executed; they are means of hardware
description.”
173

and Clock Enable
Verilog VHDL
174
4-Bit Register with Positive-Edge Clock,

Asynchronous Set, and Clock Enable
Verilog VHDL
Note: Notice the way that Verilog and VHDL

define and assign “vectors of registers” in an
abbreviated way. That’s why circuit schematics
don’t resemble the HDL code.
Question: How many flip-flops are inferred by
this piece of code?
175
Latch with Positive Gate
Verilog VHDL
Note: No flip-flops inferred.

176
Latch with Positive Gate and

Asynchronous Reset
Verilog VHDL
Important Note: The coding style

defines the inferred hardware, not
the variable names!
177
4-Bit Latch with Inverted Gate and

Asynchronous Set
Verilog VHDL
alternative form
178
Tristate Description Using Combinatorial

Process and Always
Verilog VHDL
alternative form: explicit definition of a tristate buffer

bufif0 U1(data_bus, in, data_enable_low);
179
Common Buffers
• Buffers may also be used as built-in primitives.
Gate Description
not Output inverter
buf Output buffer.
bufif0 Tri-state buffer, Active low enable.
bufif1 Tri-state buffer, Active high enable.
notif0 Tristate inverter, Low enable.
notif1 Tristate inverter, High enable.
Example:
bufif0 (weak1, pull0) #(4,5,3) (data_out, data_in, ctrl);
180
Unsigned Up-Counter with

Asynchronous Reset
Verilog VHDL
181
Unsigned Down-Counter with

Synchronous Set
Verilog VHDL
182

Asynchronous Load from Primary Input
Verilog VHDL
183
Unsigned Up-Counter with Synchronous

Load with Constant
Verilog VHDL
184

Asynchronous Reset
Verilog VHDL
185
Unsigned Up/Down-Counter with

Asynchronous Reset
Verilog VHDL
Sample applications:
• FIFO valid data counter
• Chirp signal generator
186
Signed Up-Counter with Asynchronous

Reset
Verilog VHDL
187
Signed Up-Counter with Asynchronous

Reset and Modulo Maximum
Verilog VHDL
Note: Not very practical, since

MAX should be a power of two
188
Unsigned Up Accumulator with

Asynchronous Reset
Verilog VHDL
189
Shift-Left Register with Positive-Edge

Clock, Serial In and Serial Out
Verilog VHDL
Note: If the shift register has a

synchronous parallel load, or multiple
set or reset signals, no SRL16 is
implemented.
Guideline: For better area efficiency
using built-in SRL, avoid using
sets/resets, whenever not needed.
190
Shift-Left Register with Negative-Edge Clock,

Clock Enable, Serial In and Serial Out
Verilog VHDL
191
Shift-Left Register With Positive-Edge Clock,

Asynchronous Reset, Serial In and Serial Out
Verilog VHDL
192
Shift-Left Register with Positive-Edge Clock,

Synchronous Set, Serial In and Serial Out
Verilog VHDL
193
Shift-Left Register with Positive-Edge Clock, Serial

In and Parallel Out
Verilog VHDL
194

Asynchronous Parallel Load, Serial In and Serial Out
Verilog VHDL
195

Synchronous Parallel Load, Serial In and Serial Out
Verilog VHDL
196
Shift-Left/Shift-Right Register With Positive-

Edge Clock, Serial In and Parallel Out
Verilog VHDL
197
Dynamic Shift Register With Positive-Edge

Clock, Serial In and Serial Out
Verilog VHDL
Further reading on Shift-Register applications:

https://www.xilinx.com/support/documentation/white_papers/wp271.pdf
198
Shift Registers vs. Flip-Flops

Technology Schematic on Xilinx Spartan 3
199
Shift Register Applications

Shift Registers have various applications including:
• Pipeline Compensation
• Pseudo Random Number (Noise) Generation
• Serial Frame Synchronization (in telecommunications)
• Running Average using an Adder Tree
• Running Average Using an Accumulator
• Pulse Generation and Clock Division
• Multi-stage Dividers
• Forcing the Hot State
• Pattern Generation
• FIR Filter
• FIFO
• A Complete RS-232 Receiver
Further reading : https://www.xilinx.com/support/documentation/white_papers/wp271.pdf

200
Multiplexers in Verilog
• If-Then-Else or Case can be used for multiplexers (MUXs) description.
• If one describes a MUX using a Case statement, and does not specify
all values of the selector, the result may be latches instead of a
multiplexer. When writing MUXs, one can use don’t care to describe
selector values.
• XST decides whether to infer the MUXs during the Macro Inference
step. If the MUX has several inputs that are the same, XST can decide
not to infer it. One can use the MUX_EXTRACT constraint to force XST
to infer the MUX.
• Verilog Case statements can be: full or not full; parallel or not parallel
• A Verilog Case statement is:
• Full: if all possible branches are specified
• Parallel: if it does not contain branches that can be executed
simultaneously
201
Multiplexers Full and Parallel Multiplexers Not Full But Parallel

202
Multiplexers Neither Full Nor Parallel

Note: XST automatically
determines the characteristics
of the Case statements and
generates logic using
multiplexers, priority encoders,
or latches that best implement
the exact behavior of the Case
statement.
203
MUX using IF Statements
Verilog VHDL
204
MUX using Case Statements
Verilog VHDL
205
MUX using Tristate Buffers
Verilog VHDL
206
Missing Else Statement Leading to a Latch Inference
Verilog VHDL
Caution! Unless you actually intended to

describe such a latch, add the missing
else statement. Leaving out an else
statement may also result in errors
during simulation.
207
One-Hot Decoders
Verilog VHDL
208
One-Cold Decoders
Verilog VHDL
209
No Decoder Inference (Unused Decoder Output)
Verilog VHDL
210
No Decoder Inference (Some Selector Values Unused)
Verilog VHDL
211
Priority Encoders
Verilog VHDL
Sample application: Prioritized

interrupt mechanism design
212
Logical Shifter One
Verilog VHDL
213
Logical Shifter Two (no logic shifters inferred)
Verilog VHDL
214
Logical Shifter Three (no logic shifters inferred)
Verilog VHDL
215
Unsigned Adder
Verilog VHDL
216
Unsigned Adder with Carry
Verilog VHDL
217
Unsigned Adder with Carry Out
Verilog VHDL
218
Unsigned Adder with Carry in and Carry Out
Verilog VHDL
219
Signed Adder
Verilog VHDL
220
Unsigned Subtractor
Verilog VHDL
221
Unsigned Subtractor with Borrow
Verilog VHDL
222
Unsigned Adder/Subtractor
Verilog VHDL
223
Unsigned Greater or Equal Comparator
Verilog VHDL
224
Unsigned Multiplier
Verilog VHDL
225
Sequential Complex Multipliers in Verilog
Note:
Considering that (ar + jai)(br + jbi) = (arbr - aibi) + j(arbi + aibr):
• The first two first cycles compute:
Res_real = A_real * B_real - A_imag * B_imag
• The second two cycles compute:
Res_imag = A_real * B_imag + A_imag * B_real
226
Sequential Complex Multipliers in VHDL

227
Pipelining
• Pipelining is a general technique for improving design timing and hardware
utilization efficiency by using parallel units that simultaneously process the
output of preceding stages of the pipeline.
• Implementing combinational logic using pipelines can significantly reduce the
critical path delay.
228
A Few Definitions
• (Input-Output) Latency: the
amount of time it takes to Example:
travel through the pipe.
• Critical Path: Longest
combinational path between
the output of one flip-flop to
the input of another flip-flop
(sharing a common clock)
• Throughput: The maximum
Throughput = one task every three days
rate of data flowing in or our of Latency = is input-output path dependent
a data-path (the inverse of the

critical path)
229
Pipelining Critical Paths

• Pipelining can shorten the critical path and improve the throughput
(possibly) at a cost of an increased latency between the input-output
Critical path = 8ns, Max Throughput = 125MHz, I/O Latency = 3 clocks (24ns @ fclock=125MHz)
3ns 8ns
clock
3ns 5ns 4ns
clock
New critical path = 5ns, Max Throughput = 200MHz, I/O Latency = 4 clocks (20ns @ fclock=200MHz)
(we will discuss much more about pipelining in digital systems design up to end of the course)
230
Pipelined Multiplier (Outside, Single)
Verilog VHDL
Note: This code is automatically replaced by a

four-stage pipeline multiplier, only if the
intermediate pipeline registers (pipe_1,
pipe_2 and pipe_3) are not used elsewhere
in the code. Question: Why?
231
Pipelined Multiplier (Inside, Single)
Verilog VHDL
232
Pipelined Multiplier (Outside, Shift) in Verilog

233
Pipelined Multiplier (Outside, Shift) in VHDL

234
Multiplier Adder With 2 Register Levels

on Multiplier Inputs
Verilog VHDL
235
Multiplier Adder/Subtractor With 2

Register Levels on Multiplier Inputs
Verilog VHDL
236
Multiplier Up Accumulate with Register

after Multiplication
Verilog VHDL
237
Multiplier Up/Down Accumulate

with Register after Multiplication
Verilog VHDL
238
Division by Constant Powers of 2 Dividers
Verilog VHDL
Notes:
 Dividers are supported only when the
divisor is a constant and is a power of 2.
In that case, the operator is
implemented as a shifter. Otherwise,
XST issues an error message.
 IP cores or custom code can be used
for other divisors.
Question: Why aren’t dividers built-in

primitives like adders and multipliers?
Resource Sharing (Hardware Reuse)
• The goal of resource sharing (also known as hardware
reuse or folding) is to minimize the number of operators
and the subsequent logic in the synthesized design. This
optimization is based on the principle that two similar
arithmetic resources may be implemented as one single
arithmetic operator if they are never used at the same time.
• Resource sharing is commonly handled by synthesis tools
automatically, unless if prevented by user constraints and
synthesis attributes.
• If the optimization goal is speed, disabling resource
sharing may give better results.
(we will discuss much more about resource sharing in digital systems design up to end of the course)
240
Resource Sharing Example
Verilog VHDL
241
Single-Port RAM in Read-First

Mode
Verilog VHDL
242
Single-Port RAM in Write-First

Mode in Verilog
Template 1 Template 2
243
Single-Port RAM in Write-First

Mode in VHDL
Template 1 Template 2
244
Single-Port RAM in No-Change

Mode
Verilog VHDL
245
Single-Port RAM with Asynchronous Read
Verilog VHDL
246
Single-Port RAM with Synchronous

Read (Read Through)
Verilog VHDL
247
Single-Port RAM with Enable
Verilog VHDL
248
Dual-Port RAM with Asynchronous Read
Verilog VHDL
249
Dual-Port RAM with Synchronous

Read (Read Through)
Verilog VHDL
250
Dual-Port RAM with Synchronous

Read (Read Through) and Two Clocks
Verilog VHDL
251
Dual-Port RAM with One Enable

Controlling Both Ports
Verilog VHDL
252
Dual Port RAM with Enable on Each Port
Verilog VHDL
253
Dual-Port Block RAM with Two Write Ports
Verilog VHDL
254
Multiple Write Statements
Verilog VHDL
255
Read-First Mode: Single-Port BRAM with Byte-wide

Write Enable (2 Bytes)
Verilog VHDL
256
Write-First Mode: Single-Port BRAM with Byte-Wide

Write Enable (2 Bytes)
Verilog VHDL
257
No-Change Mode: Single-Port BRAM with Byte-

Wide Write Enable (2 Bytes)
Verilog VHDL
258
Multiple-Port RAM Descriptions
Verilog VHDL
259
Block RAM with Reset Pin
Verilog VHDL
260
Block RAM with Optional Output Registers

Verilog VHDL
261
Initializing RAM Directly in HDL Code
Verilog VHDL
262
Initializing RAM Directly in HDL Code

Dual Port Block RAM Initialization in Verilog
263
Initializing RAM from an External File in Verilog
Verilog Hexadecimal/Binary in text format

264
Initializing RAM from an External File in VHDL

265
ROM with Registered Output, Example 1
Verilog VHDL
266
ROM with Registered Output, Example 2
Verilog VHDL
267
ROM with Registered Address

Verilog VHDL
268
Pipelined Distributed RAM
Verilog VHDL
269
Finite State Machines (FSM)
• Finite state machines

(automata) are used as the
backbone controllers and timing
managers of digital systems.
• FSMs can be graphically
illustrated by bubble diagrams.
• Flawless design of FSMs is
critical for a proper system
function.
• FSMs with dead-ends or
erroneous state-transitions can
result in hardware hanging or
malfunctions.
270
Finite State Machines Components

An FSM consists of:
• Inputs
• Outputs
• States and state sequences
• State transition rules
• Initial conditions (states)
• Resetting mechanism
FSM Implementation issues:

• State encoding technique
• Robust FSM design and state recovery mechanism
271
Finite State Machine (FSM) Representation

Incorporating Mealy and Moore Machines
𝑠𝑘+1 = 𝑓(𝑠𝑘 , 𝑥𝑘 )
Mealy Machine:
𝑦𝑘 = 𝑔(𝑠𝑘 , 𝑥𝑘 )
𝑠𝑘+1 = 𝑓(𝑠𝑘 , 𝑥𝑘 )
Moore Machine:
𝑦𝑘 = 𝑔(𝑠𝑘 )
Research Topic: According to the above representation, Mealy and Moore machines
can be studied from a state-space perspective. The rich literature of state-space
analysis from Control Theory can be used to study the properties of logic circuits.
272
FSM Encoding Techniques

HDL synthesis tools support various FSM encoding techniques
including:
 One-Hot
 Gray
 Compact
 Johnson
 Sequential
 Speed1
 User Defined
 Auto Encoding
273
Sample FSM Encoding Logic
The One-Hot Encoder
Ref: https://www.electronics-tutorials.ws/combination/comb_5.html
274
Sample FSM Encoding Logic
The One-Cold Encoder
Ref: https://www.electronics-tutorials.ws/combination/comb_5.html
275
FSM Implementation Issues
• State encoding methods defer in speed, area and

robustness to state transition errors.
• On FPGA, FSMs are commonly implemented using
BRAM or LUT
• Synthesis tools such as XST can add logic to an FSM
implementation that will let the state machine recover
from an invalid state. During run-time, if an FSM enters
an invalid state, this extra logic will take it back to a
known state, called a recovery state (the `reset state’ by
default). This is called a Safe FSM Implementation
276
Finite State Machine Coding Example: A Single Process
Verilog VHDL
277
Finite State Machine Coding

Example: Two Processes
Verilog VHDL
278
Finite State Machine Coding

Example: Three Processes
Verilog VHDL
Black Boxes
• A design may contain Electronic Data Interchange Format
(EDIF) or NGC files generated by synthesis tools, schematic
text editors, or any other design entry mechanism, which can be
treated as black-boxes during synthesis
• These modules must be instantiated in the code in order to be
connected to the rest of the design; but the netlist is propagated
to the final top-level netlist without being processed by the
synthesis tool.
• Synthesis tools such as XST enables one to attach specific
constraints to these Black Box instances.
• One may also have a design block for which an RTL model
exists; but the designer’s own implementation of this block is in
the form of an EDIF netlist and the RTL model is valid for
simulation purposes only.
280
Black-Box Coding Techniques
Verilog VHDL
Note: The concept of black-boxes is

similar to the notion of precompiled
static libraries in software languages,
which are bypassed by the compiler Question: Name a hardware analog for
and are linked to the rest of the code by dynamic libraries in software languages
the linker.
281
Summary
• Synthesizable HDL coding styles were reviewed in this

section.
• These guidelines are for practice and not memorization. In
practice, as far as a designer is aware of describing a
hardware and thinks logically and concurrently, one does
not need to think of the realized hardware
ADVANCED FPGA CODING
TECHNIQUES
283
Toggling a Flag with Multiple Clocks

Standard D-type Flip-Flops do not support more than a single clonk. But in practice,
there are cases where we need to change a flag using two independent clocks.
Example: handshaking mechanisms
284
Clock Speed Reduction

Apart from DCMs, various methods exist for clock speed reduction, including:
Gated-Clock; not recommended nor supported

on most FPGA devices
Standard method for clock division using FF clock enable

Standard method for clock halving using
FF clock enable
285
Mixed Clock-Edge Design

• It is possible to use both positive and negative clock edges in a
single design; but it should be avoided as much as possible
• Using mixed clock-edges does not double the clock rate; but it
rather reduces the time for combination logic result settlements
• Utilization of mixed clock-edges should be confined to phase
compensation between two signals when setup or hold-times are
not fulfilled using a single edge (commonly at FPGA I/O)
• Example:
setup time not fulfilled on posedge
setup time fulfilled on negedge
clock
data
286
Standard Resetting Mechanisms

• Although both synchronous and asynchronous reset mechanisms are supported in FPGA
designs, it is highly recommended to use a unified resetting mechanism throughout the
entire design.
• Synchronous resets with sufficient flip-flop synchronizer stages are preferred over
asynchronous resets (due to lower probability of metastability)
• Even if the original reset command is asynchronous (e.g. using a push-button or software
command), it is good practice to make an internal synchronous reset flag
Supported asynchronous reset mechanism
Generating synchronous from asynchronous reset flag
Preferred synchronous reset mechanism

287
Increasing Fan-out by HW Replication

• The maximum fan-out of a
logic circuit output is the
maximum number of gate
inputs it can drive without synthesis attributes to
avoid register
loading effects disturb its optimization and
function (switching speed merging of ce1 and
ce2
and voltage level)
• In contemporary FPGAs, Flip-Flop fan-outs
are very high (several hundreds) and only
the most frequently used signals (such as
CLOCK, RESET, CE, …) may face fan-out
issues
• The fan-out of a logic circuit may be
increased by user constraints or hardware
replication in HDL
288
Debouncing
• In digital designs, bouncing
(between 0 and 1) occurs during
manual switch transitions
• The objective of debouncing is to
avoid the mis-detection or multiple
counting of events during switch
transitions
• Debouncing can be implemented
both in hardware (analog) and
software (digital)
Reference: Arora, M. (2011). The art of hardware
architecture: Design methods and techniques
for digital circuits. Springer Science & Business
Media, Chapter 8
289
Hardware Debouncing Techniques

• Various hardware debouncing mechanisms:
RC debouncer RC debouncer with digital logic Robust RC debouncer SR debouncer

with digital logic
FPGA
IC debouncer MAX6816’s internal mechanism

290
Software Debouncing Techniques

• Software debouncing mechanisms:
ISR assembly language debouncer pseudo-code
C language debouncer pseudo-code

291
HDL Debouncing Techniques
Ref: https://eewiki.net/pages/viewpage.action?pageId=13599139
OVERVIEW OF LOGIC
SYNTHESIS METHODS*
(Optional)
293
Synthesis CAD Tools

• Electronic Design Automation (EDA) Vendors:
• Xilinx
• Synopsys
• Synplicity
• Cadence
• Altera
• Mentor Graphics
•…
294
Xilinx XST Overview

295
XST Detailed Design Flow
Note 1: XST performs a resource sharing check. This usually leads to a reduction
of the area as well as an increase in the clock frequency.
Note 2: Xilinx-Specific Netlist (NGC) = EDIF + Netlist Constraints File (NCF)

296
Logic Synthesis Approaches

1. Technology Dependent
Uses device-dependent properties during synthesis
2. Technology Independent
The most common approach; but requires a second round of
technology dependent synthesis
297
Logic Synthesis Detailed View

Logical hardware units:
• Combinational Logic
(Nodes)
• Sequential Logic
(Registers)
Note: This is where the term Register Transfer Level (RTL) comes from
298
Node Synthesis
• Two-level Logic Synthesis
• Deals with the synthesis of designs represented in two-level logic. The longest path from
input to output, in term of number of gates crossed on the path, is two.
• Two-level logic is the natural and straightforward approach to implement a Boolean
function, because each Boolean function can be represented as a sum of product terms.
• In the first level, the products are built using the AND primitives. The sums of the
resulting products are built in the second level with the OR-primitives.
• Used for CPLD
• Multi-Level Logic Synthesis

• In the multi-level synthesis, functions are represented using a multi-level logic. Those are
circuits in which the longest path from input to output goes through more than two gates.
• Used for FPGA
299
Node Representation
1. Sum of Products (SOP) Form
2. Factored Form
• a product is either a single literal or the product of two factored forms and a sum
is either a single literal or the sum of two factored forms.
• Factored forms are representative of the logic complexity.
3. Binary Decision Diagram (BDD)

• Is a rooted directed acyclic graph used to represent a Boolean function. Two
kinds of nodes exist in BDDs: variable and constant nodes.
300
Binary Decision Diagram (BDD)

Example:
301
Node Manipulation Operators

• Decomposition
• Extraction
• Factoring
• Substitution
• Collapsing (elimination): reverse of substitution
302
LUT-based Technology Mapping (Second-

Level Synthesis)
Depending on their optimization goals, these algorithms
can be classified in three categories:
1. Area Minimization: Chortle-crf, MIS-fpga, Xmap, etc.
2. Delay Minimization: FlowMap, Chortle-d, DAG-map,
MIS-pga-delay, etc.
3. Routability Maximization
PART III
Advanced Topics in Digital Design and Implementation
NUMBER
REPRESENTATION
305
Number Representation in PLD Systems

• While number representation is fully standardized and
rather automatically handled in multipurpose CPUs and
GPUs (and is rarely a concern for the designer), it is an
essential and time-taking part of most FPGA-based
designs.
• In this section, we study:
• The most common number representation standards
• Fixed-point representation issues
• Statistical analysis of truncation and rounding errors during data
acquisition (using analog-to-digital converters) and calculations
306
An Overview of Binary Number Representation

• For many reasons radix-2 has remained the
dominant number representation in digital
hardware design:
 In early technologies: the difficulty of generating high-speed
switching logic circuits with more than two distinct and
distinguishable levels of voltages.
 In current technologies: besides the simplicity of radix-2, the
huge body of literature, algorithms, codes, hardware
(transistors, gates, etc.), and engineering experience and
conventions, which already exist for radix-2 calculations makes
it too expensive to migrate to higher radixes.
307
Binary Number Representation

Number representation can be studied from various aspects,
including:
• Numbers of Interest:
• Integers
• Reals
• Sign Representation:
• Unsigned
• Signed
• Fractional Number Representation:

• Fixed-point
• Floating-point
308
Accuracy of Finite Length Binary Number

Representations
Question: How accurate is it to represent numbers (integer or
fractional) in radix-2 using finite number of bits?
Basis Representation Theorem: For a given base b, any integer x ∈ ℤ

can be uniquely represented as follows:
x = akbk + ak-1bk-1 + … + a1b1 + a0
where aj ∈ {0, 1, …, b-1} and ak ≠ 0.
Dyadic Rationals Theorem: The dyadic rational set ℙ (numbers which

can be represented as an integer divided by a power of 2), is dense in
the set of real numbers ℝ. This means that for any x ∈ ℝ, there exists
a y ∈ ℙ that is “as close as you like” to x.
Conclusion: Real numbers can be approximated in radix-2 with finite

number of bits, up to a desired level of precision.
309
Signed Binary Number Representation

Standards
The most popular signed binary number representation
standards are:
• Sign-Magnitude
• One’s-Complement
• Two’s-Complement
• Straight Offset Binary (SOB)
• Binary Coded Decimal (BCD)
• Canonical Signed Digit (CSD)
310
Sign-Magnitude Representation
The MSB is reserved for sign representation (0 for + and 1 for –). The
remaining bits are used to represent the absolute magnitude. With N bits, it
can code from –(2N-1 –1) to (2N-1 –1).
Decimal equivalent: X10 = (–1)bN-1[bN-22N-2 + bN-32N-3 + … + b12 + b0]
sign bit sign bit

bN-1 b0 bN-1 b0
0 0 1 0 1 1 0 1 = +45 1 0 1 0 1 1 0 1 = –45
magnitude bits magnitude bits
Advantage: Simple to generate and convert

Disadvantage: There are two zeros (+0 and -0); difficult to handle during
arithmetic operations
311
One’s Complement
The MSB denotes the sign (0 for + and 1 for –). With N bits, it can code from
–(2N-1 – 1) to (2N-1 – 1). Each bit corresponds to a coefficient of a power of
two in its decimal equivalent.
all bits one
Decimal equivalent: X10 = –bN-1(2N-1 – 1) + bN-22N-2 + bN-32N-3 + … + b12 + b0
sign bit sign bit

bN-1 b0 bN-1 b0
0 0 1 0 1 1 0 1 = +45 1 1 0 1 0 0 1 0 = –45
Advantage: Simple to generate and convert

Disadvantage: There are two zeros (+0 and –0); difficult to handle during
arithmetic operations
312
Two’s Complement
The MSB denotes the sign (0 for + and 1 for –). With N bits, it can code from
–2N-1 to (2N-1 – 1). Each bit corresponds to a coefficient of a power of two in
its decimal equivalent.
doesn’t fit into N bits
Decimal equivalent: X10 = –bN-12N-1 + bN-22N-2 + bN-32N-3 + … + b12 + b0
sign bit sign bit

bN-1 b0 bN-1 b0
0 0 1 0 1 1 0 1 = +45 1 1 0 1 0 0 1 1 = –45
Advantage: No repeated zeros; can code –2N-1; no sign control needed during
arithmetic operations, and several other advantages (is the most popular
signed number representation format)
Disadvantage: Slightly more difficult to read the decimal equivalent from the
binary form (for human).
313
One’s Complement vs. Two’s Complement

• 2’s complement is the most
common binary representation
used in computation
machines.
• A major property of 2’s
complement is that the binary
values are increased by one-
by-one from the most negative
to the most positive without a
break (by discarding any carry
values beyond the word
length).
• The default implementation of
arithmetic operations in Verilog
(since Verilog 2001) is in this
format.
314
Finding One and Two’s Complements

1’s Complement: Flip all the bits (0 to 1, and 1 to 0)
2’s Complement:
• Method 1: Calculate the 1’s complement, plus one
• Method 2: Subtract the number from 2N (this is where the name
2’s complement comes from)
• Method 3: Starting from the LSB, preserve all the bits as they are,
up to (and including) the right most 1. Flip all the remaining bits up
to the MSB
Note: The 2’s complement of –2N-1 can not be represented in N bits. Therefore,
during calculations, it’s 2’s complement overflows and becomes equal to itself
(just like the 2’s complement of zero)! This phenomenon can be mathematically
explained by the orbit-stabilizer theorem.
315
Properties of Two’s Complement

1. When fitting an N bit 2’s complement number into M bits (M>N),
the number should be sign extended, i.e., the left most M-N bits
should be filled with the MSB (sign bit) of the original number:
2. In arithmetic right-shifts, the number should be filled by the sign

bit from the left:
316
Properties of Two’s Complement (continued)

3. No additional circuits are required for handling the signs during
addition or subtraction (except for overflow checking). In fact, 2’s
complement numbers can be treated as unsigned numbers
during such arithmetic operations.
4. Overflow check: If two numbers with the same sign are added,
overflow occurs if and only if the result has an opposite sign.
Example:
317
Properties of Two’s Complement (continued)

5. Two’s Complement Intermediate Overflow Property: “In
successive calculation using 2’s complement arithmetic
(allowing overflows instead of saturation), if it is guaranteed that
the final result will fit in the assigned registers, then intermediate
overflows are harmless and will not affect the final answer.
Example (IIR Filter): yn = a.yn-1 + xn
Refs:
• Khan, S. A. (2011). Digital design of signal processing systems: a practical
approach. John Wiley & Sons., Section 3.5.7
• Smith, J. O. (2007). Introduction to digital filters: with audio applications (Vol. 2).
Julius Smith., P. 201
Note: Very interesting property; but I haven’t seen a rigorous statement or proof for it, yet.
Please let me know, if you find a good reference.
318
Straight Offset Binary (SOB)

2's
• Offset Binary is a binary code in which the SOB Decimal
Complement
code represents analog values between 1111 7 0111
1110 6 0110
positive and negative Full-Scale
1101 5 0101
• Using N bits, starts assigns all-zeros to –2N-1 1100 4 0100
and increments one-by-one up to 2N-1 – 1. 1011 3 0011
1010 2 0010
• Conversion to 2’s complement: Flip the MSB
1001 1 0001
to convert from SOB to 2’s complement and 1000 0 0000
vice versa. 0111 −1 1111
• Application: SOB is most common in Flash 0110 −2 1110

0101 −3 1101
Analog-to-Digital Converters (ADC) and
0100 −4 1100
Digital-to-Analog Converters (DAC) that use 0011 −5 1011
ladder comparators. 0010 −6 1010
0001 −7 1001
0000 −8 1000
319
Binary Coded Decimal (BCD)

• A class of binary encodings of
decimal numbers where each
decimal digit is represented by a
fixed number of bits (usually four
or eight).
• Special bit patterns are used for
a sign or for other indications
(e.g., error or overflow)
• Applications: whenever human

interaction is needed; such as
LCDs, 7–segments, etc.
Ref: https://en.wikipedia.org/wiki/Binary-coded_decimal
320
Canonical Signed Digit (CSD)

• CCD is a three-symbol coding system in terms of powers of two.
• It uses a sequence of (+,0,–) to code numbers. For example, the
integer 23 can be expanded as follows:
23 = + 25 – 23 – 20
In CCD, 23 is coded as (+0–00–), i.e.,
o Positive powers of two are denoted by +
o Negative powers of two are denoted by –
o Missing powers of two are denoted by 0
• CCD is popular in some digital signal processors (DSP)
Note: CCD is a non-unique number representation
Note: Statistically, the probability of a digit being zero in CCD can be shown to be close
to 66% (vs. 50% in 2's complement encoding). This property leads to more efficient
hardware implementations of add/subtract networks and multiplication by constants.
Further Reading: Khan, S. A. (2011). Digital design of signal processing systems: a practical approach. John
Wiley & Sons., Chapter 6
321
Fractional Number Representation

The most common binary representations of
fractional numbers are:
• Floating-Point: Uses an exponential representation of a
number; it is used in most CPUs and some DSP. In FPGA,
floating point units (FPUs) are provided by some vendors as
hard or soft IP
• Fixed-Point: Uses positive and negative powers of two
expansion of a number with a fixed radix point; it is commonly
used in fixed-point DSP and microcontrollers
• Mixed-Precision: Uses positive and negative powers of two
expansion of a number with a different radix point (at each point
of the computing system); it is commonly used in FPGA design
322
Floating-Point Number Representation

• The basic idea of floating point (FP) representation is to approximate
a real number in terms of a fixed number of significant digits
(significands or mantissa) scaled by an exponent of a fixed base
(e.g., 2, 10, 16, etc.).
• For example:
1.2345 = 12345 × 10 –4 exponent
significand base
• Apparently, not all real numbers can be represented in this format

(using finite number of digits). However, FP provides an
approximation with a fixed relative error throughout the real line (i.e.,
small errors for small numbers and larger errors for large numbers).
323
IEEE 754 Single-Precision Binary

Floating-Point Format
• According to IEEE 754 floating-point standard:
sign exponent (8 bits) mantissa (23 bits)
0 0 1 0 1 1 1 0 0 0 1 0 1 1 0 1 0 0 0 1 0 0 0 0 1 1 0 0 0 0 0 1
• • • • •
b31 b30 b23 b22 b0
The exponent is selected such that the
• The decimal equivalent is: left-most bit of the mantissa is always 1
(which isn’t stored in the binary form),
𝑀 making the representation unique.
𝑋10 = (−1)𝑆 × 2𝑒−𝐵 × (1 + 𝑏𝑀−𝑖 2−𝑖 )

𝑖=1
where:
• Total number of bits is 32 in single precision (binary32) and 64 in double precision (binary64)
• 𝑆 is the sign bit (b31 in single precision and b63 in double precision)
• 𝑒 is the exponent (8 bits in single precision and 11 bits in double precision)
• 𝐵 is a constant bias (equal to 127 in single precision and 1023 in double precision)
• 𝑀 is the fractional length (23 bits in single precision and 52 bits in double precision)
324
Single-Precision Binary Floating-Point

Examples
Example 1: binary floating point to decimal
0x3E200000 = (0011 1110 0010 0000 0000 0000 0000 0000)2
0 0 1 1 1 1 1 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
• • • • •
b31 b30 b23 b22 b0
0 124−127
The decimal equivalent is (−1) × 2 × 1 + 0.25 = 0.15625
the most accurate 32-bit single-
precision approximation for π
Example 2: decimal to hex/binary floating point
Scale the number in the form of ±2am, where 1≤|m|<2 and aєZ, to find the exponent
and mantissa
π (3.1415926535897932384626433832795…) ≈ 3.1415927410125732421875
0 1 0 0 0 0 0 0 0 1 0 0 1 0 0 1 0 0 0 0 1 1 1 1 1 1 0 1 1 0 1 1
• • • • •
b31 b30 b23 b22 b0
which is 0x40490FDB = (0100 0000 0100 1001 0000 1111 1101 1011)2
A nice tool: http://www.binaryconvert.com
325
Floating-Point Arithmetic
Addition/Subtraction:
1. Make the smallest exponent equal to the biggest (by right-shifting the mantissa)
2. Add/subtract the mantissas (note that the smaller ones may vanish to 0 during
the right-shifts)
Multiplication/Division:
1. Add/subtract the exponents
2. Multiply/Divide the mantissas
3. Scale and round the results
Special Values:
Floating-point representation has reserved codes for special values including: 0+, 0–,
+∞, -∞, and Not-a-Number (NaN) such as 0/0, +∞/-∞, 0×∞
Note: Due to the (implicit) leading 1 in front of the mantissa, zero needs to be defined
as a special value (when all the bits of the exponent and mantissa are zero), which is
different from epsilon (±2−127 )
326
Fixed-Point Number Representation

• Fixed-point is basically the 2’s complement representation with a fixed power-of-two
scaling factor for changing the radix point to enable fractional number
representations:
sign bit signed integer fractional part
0 0 1 0 1 1 1 0 0 0 1 0 1 1 0 1 0 0 0 1 0 0 0 0 1 1 0 0 0 0 0 1
•
bN-1 b0
• The decimal equivalent is: radix point
signed two‘s
𝑁−2 complement
𝑋10 = 2−𝑀 × (−𝑏𝑁−1 2𝑁−1 + 𝑏𝑖 2𝑖 )
𝑖=0
where:
• 𝑁 is the total number of bits
• 𝑀 is the fractional point
Note: In fixed-point systems the radix point location is assumed to be fixed throughout
the entire system. That’s where the name comes from.
327
Floating-Point vs. Fixed-Point

1. For the same number of bits, they can (almost) code the same number of real numbers.
2. Fixed-point uses all possible codes for number representation, while floating point reserves
a few codes for special values. Floating-point has a larger dynamic range (the ratio of the
largest to smallest number that are represented)
3. In fixed-point, the range of its MIN and MAX over the real line is quantized to equally
spaced numbers (therefore the approximation error is uniform from MIN to MAX); in floating-
point, the spacing of numbers is non-uniform (groups of numbers with a fixed intra-gaps but
different inter-gaps)
4. Fixed-point hardware architectures are simpler than floating-point architectures; floating-
point architectures have additional circuitry for handing special values.
fixed-point numbers
the real line

Inspired from: Izquierdo, Luis R. and
Polhill, J. Gary (2006). 'Is Your Model
Susceptible to Floating-Point Errors?'.
Journal of Artificial Societies and Social
Simulation 9(4)4
<http://jasss.soc.surrey.ac.uk/9/4/4.html>
floating-point numbers
328
The Qm.n Fixed-Point Convention

• In order to denote the total number of bits and the bits assigned to
the integer and fractional parts of a fixed-point number, various
conventions exist. For example,
 Texas Instruments’ QN format (or Q1.N) assumes 1 bit (the sign bit)
as the integer part and N bits for the fractional part.
 Matlab’s fixed-point toolbox takes the total number of bits and the
fractional length to form an fi-object.
Throughout this course, we use the Qm.n convention, where:

 m is the number of bits assigned to the integer part
 n is the number of bits assigned to the fractional part
 N = m + n is the total number of bits (including the sign)
The numbers are signed, therefore the MSB represents the sign
329
Fixed-Point Arithmetic
Addition/Subtraction:
1. Align the radix points
2. Zero pad the LSB of numbers with shorter fractional lengths
3. Sign extend the MSB of numbers with shorter integer lengths
4. Apply addition/subtraction
Multiplication/Division:
1. Apply multiplication/division as if they were integer valued (regardless of
the radix point)
2. Find the appropriate radix point by adding/subtracting the radix points
Note: Bit-growth occurs during fixed-points arithmetic, which is handled by

either:
1. increasing the number of bits,
2. truncation/rounding from the LSB or MSB (is discussed in details later), or
3. a combination of both 1 and 2
330
Bit-Growth in Fixed-Point Arithmetic

In order to guarantee that no overflow occurs during arithmetic
operations, the number of output bits should be longer than the
arithmetic operands:
1. Qm1.n1 ± Qm2.n2 = Qm.n
where m = max(m1,m2)+1 and n = max(n1,n2)
2. Qm1.n1 × Qm2.n2 = Qm.n
where m = m1+m2 and n = n1+n2
Note: During multiplication, N = N1 + N2 – 1 is generally enough. The only exception
(requiring N = N1+N2) is for signed numbers when the two most negative numbers
(–2N1-1 and –2N2-1) are multiplied together, resulting in +2(N1+N2-2), which overflows in
N = N1 + N2 – 1 bits and requires N = N1 + N2. This single bit can be saved by
either:
1. Making sure that the two operands are never equal to the most negative
numbers (this is possible when one of the operands is a known constant)
2. Approximating 2(N1+N2-2) with 2(N1+N2-2) –1! Yes, this approximation is OK in many
systems.
331
Controlling Bit-Growth in Fixed-Point Systems

It is impractical (and unnecessary) to increase the number of bits after
successive arithmetic operations. Bit growth can be controlled by
discarding either from the LSB or MSB of the arithmetic result.
• When to discard from the MSB?

• Only possible when the full-length is not utilized or the arithmetic operation
(mathematically) guarantees that no bit growths occur → results in no errors
• If the full-length is utilized → causes large sign/amplitude errors
• When to discard from the LSB?
• The right most LSB zeros can be discarded without any errors
• Truncating/rounding non-zero LSB results in relatively small errors, depending on
the number’s magnitude
• A stochastic framework is required to analyze the average truncation/rounding error
effect.
sign changing error
no errors larger errors smaller errors no errors
1 1 1 1 1 0 1 0 0 0 1 0 1 1 0 1 0 0 0 1 0 0 0 0 1 1 0 0 0 0 0 0
bN-1 b0
332
Truncation/Rounding Error Analysis

• The truncation procedure can be modeled by an operator Q(•):
yn = Q(xn) = xn + en
x: input sample (signal) y: truncated/rounded result
e: truncation/rounding error n: sample index
• The impact of truncation error depends on both the original sample (signal)
and the truncated values’ amplitudes.
• In continuous data streams, the most common approach for studying the
truncation error impact is to measure the ratio of the average data power to
the average noise power, known as the signal-to-noise ratio (SNR):
𝐸 𝑥𝑛2
SNR dB = 10log10
𝐸 𝑒𝑛2
where 𝐸{∙} denotes averaging (or stochastic expectation) over all ensembles.
Note: The calculation of the SNR requires prior assumptions regarding the input
stream and the truncation error distribution.
333
Truncation/Rounding Error SNR Calculation

Suppose that we have an m bit signed integer sequence xn, for which
we want to round the p LSB bits (to zero) and obtain yn. Assuming a
uniform distribution for xn, the probability density functions (pdf) of xn
and the error sequence en are:
rounding instead Pr(en=e) Pr(xn=x)
of truncation (+ 2-m
2-p
and – error
values) … … … …
• •-2 p-1 0
•
2p-1-1
e • •-2 m-1 0
•
2m-1-1
x
slightly biases
Error mean: 2𝑝−1 −1 towards negative
1 1 numbers truncation point
𝑒 = 𝐸 𝑒𝑛 = 𝑖∙ 𝑝 =−
𝑝−1
2 2 preserved bits (yn) discarded bits (en)
Error variance: 𝑖=−2
1 1 0 0 0 1 0 0 0 0 1 1 0 0 1 1 0 0
2𝑝−1 −1
1 2 1 22𝑝 − 1 total bits (xn)
𝜎𝑒2 = 𝐸 (𝑒𝑛 − 𝑒)2 = (𝑖 + ) ∙ 𝑝 =
𝑝−1
2 2 12
𝑖=−2
334
Truncation/Rounding Error SNR Calculation

(continued)
Similar results hold for the mean and variance of xn. Therefore the
SNR is: 2 2𝑚
𝜎𝑥 2 −1
SNR dB = 10log10 2 = 10log10 22𝑝 − 1
𝜎𝑒
which for large 𝑝 can be approximated as:
22𝑚
SNR dB ≈ 10log10 2𝑝
= 20(m − p)log10 2 ≈ 6.02(𝑚 − 𝑝)
2
Note: This is the 6dB per-bit rule of thumb: truncating each bit reduces the SNR for
about 6dB. We will find a similar rule later for ADC performance with different signal
and noise distributions.
Exercise: Derive the above equations (mean and variance of error) analytically. Do
the results change if the number is in the Qm.n format?
335
Truncation vs. Rounding

• While truncation simply discards the unnecessary bits, rounding approximates with
the closest number.
• Rounding is commonly preferred over truncation, as it is less-biased (the very
small bias is due to the representation of -2p-1 in 2’s complement).
truncation error probability rounding error probability
2-p 2-p
… …
•• 0 2p-1
• e
• •-2 p-1 0 2p-1-1
• e
Example: round(3.7) = 4; truncate(3.7) = 3;
• Truncation versus rounding in Verilog: Radix-10 equivalent trick:

[3.7 + 0.5] = 4;
[3.2 + 0.5] = 3;
336
Mixed-Precision Multiplication Examples

Example 1: Multiplication by constant powers of two: no multiplication is required; only the
radix point convention changes; no error increase
integer fraction
X integer fraction
x: 0 0 1 0 1 1 1 0 0 0 1 0 1 1 0 1
y: 0 0 1 0 1 1 1 0 0 0 1 0 1 1 0 1
xN-1 xN-P-1 x0 yN-1 yN-P y0
2.0
Example 2: Multiplication by constant non powers of two: multiplication is required; the

radix point and register length may change; error might be added due to output truncation
integer fraction
X
x: 0 0 1 0 1 1 1 0 0 0 1 0 1 1 0 1 y: 0 0 1 0 0 1 0 1 1 0 0 0 0 1 0 0 1 0 0 1
xN-1 xN-P-1 x0 yN-1 y0
3.25
3.25 0 1 1 0 1
337

Example 3: Multiplication by fractional non powers of two that can not be
represented by sum of powers of two: Unavoidable representation error, even
before multiplication
integer fraction
x: 0 0 1 0 1 1 1 0 0 0 1 0 1 1 0 1
xN-1 xN-P-1 x0 X
y: 0 1 0 0 1 1 0 0 0 1 1 1 1 0 1 0 1 0 0 0 1
yN-1 y0
3.3125 0 1 1 0 1 0 1 3.3
The closest approximation

of the constant in 7 bits
338

Rounding/truncating the coefficients in data/signal processing systems can change
the nominal performance of the system. For example, in filter design:
Sample lowpass filter designed in Matlab FDATool in double precision

floating-point (blue) and after quantization with 12-bit fixed-point (red)
339
Mixed-Precision in Digital Filters

Example 4: Discrete-time convolution 𝑦𝑛 = 𝑥𝑛 ∗ ℎ𝑛 = 𝑚 ℎ𝑚 𝑥𝑛−𝑚 : The maximum bit
growth in the output is equal to the length of the filter coefficients L1–Norm:
G = log 2 hm
m
xn yn
hn
B bits (impulse response) (B + G) bits
IIR Filter FIR Filter
Note: From Signals & Systems Theory we know that for a stable causal filter 𝑚 ℎ𝑚 = B < ∞.
Therefore “the output of a stable filter with a bounded input can always be stored in a register
of finite length without overflow”
340
Further Notes on Fixed-Point and Mixed-Precision

• Note 1: The radix point does not necessarily need to be within the range of
the register length. Example: An 8-bit register can be used to represent
fixed-point numbers with a decimal point below the LSB or above the MSB.
For example, the following are legitimate fixed-point numbers, even though
the register length is only 8 bits:
implied but not stored with
the number
0 1 1 0 1 0 1 0 x = 2 +15. (21 + 23 + 25 + 26)
1 0 1 0 1 0 1 1 x = 2 -12. (20 + 21 + 23 + 25 + 26 – 27)
• Note 2: In practice, fixed or floating-point numbers can have an arbitrary and

implicit scaling factor, which is known to the designer; but is not coded or
stored with the number. These scaling factors are only incorporated when
numbers are mapped to their corresponding physical values (voltage,
temperature, current, etc.) for user visualization or analysis. Example:
Uniform analog-to-digital convertors map their input voltage to the output
code with a constant scaling factor, which is known by the designer; but
does not affect internal FPGA calculations.
341
Coefficient Scaling and Rounding

• In order to store real numbers in finite-length registers (fixed or
floating-point), the numbers should be multiplied by appropriate
scaling factors and rounded/truncated to fit in the registers.
Examples: scaling factors
yfixed = round( 216 × yreal)

yfixed = round( 3.14 × yreal)
• When scaling a set of coefficients (time-series, filter coefficients,

etc.) to fit in N bits, the optimal performance (with minimum
quantization error) is obtained when the maximum/minimum scaled
values are equal to the maximum/minimum possible numbers (-2 N-1
and 2 N-1-1).
• Therefore, the optimal scaling factor is not necessarily a power of
two (e.g., see Matlab FDAtool’s quantization and scaling options)
342
Bit-Growth in Digital Filter Implementation*

(optional)
In digital filter implementation, the L1–Norm bit growth G = log 2 m hm is the worst-
case (most pessimistic), which does not make any assumptions on the input signal. This
formula can be relaxed (approximated) in some cases.
1. Instantaneously narrow-band signals: For signals having a dominant frequency peak
at each time instant:
𝑥𝑛 = A cos(𝜔0 𝑛 + 𝜃) → 𝑦𝑛 ≈ 𝐻 𝑒 𝑗𝜔0 A cos 𝜔0 𝑛 + 𝜑𝜔0
Bit Growth G0 = log 2 max 𝐻 𝑒 𝑗𝜔
−𝜋≤𝜔<𝜋
2. Random input signals: Using Parseval’s theorem, the output variance of a filter with
a random input is related to its input variance as follows:
𝜎𝑦2 = 𝜎𝑥2 |hm |2
𝑚
Therefore, with the following bit-growth, the probability of overflow at a filter’s output
is (almost) equal to the probability of input overflow:
Bit Growth G1 = log 2 hm 2
𝑚
343
Bit-Growth in Digital Filter Implementation*

(optional)
Example: A first-order lowpass IIR filter: 𝑦𝑛 = α𝑦𝑛−1 + 𝑥𝑛 (0 < α < 1)
The impulse response is ℎ𝑛 = α𝑛 𝑢 𝑛
1 2 1
Therefore m hm = and m hm =
1−𝛼 1−𝛼 2
xn 1 yn
𝑗𝜔
𝐻(𝑒 )=
Bit-growth analysis for α = 0.9: 1 − 𝛼𝑒 −𝑗𝜔
1
• L1-Norm: G = log 2 m hm = log 2 = 3.3219 = 4
1−𝛼
1
• Narrow-band assumption: G0 = log 2 = 3.3219 = 4
|1−𝛼|
1
• Parseval’s theorem for stochastic inputs: G1 = log 2 = 1.198 = 2
1−𝛼 2
Result: In this example the L1-norm and narrow-band assumption, both demand 4
additional bits at the output yn; but according to the output variance criterion if we are
fine with occasional overflows, adding only 2 bits is statistically OK.
ANALOG TO DIGITAL
CONVERTORS AND DIGITAL
TO ANALOG CONVERTORS
345
Analog to Digital Convertor (ADC) vs. Digital to

Analog Convertor (DAC)
ADC and DAC are integral parts of most FPGA-based signal processing systems
sampling frequency (fs) resolution bits (B)
digital signal
FPGA
analog signal x[n]
quantization and
x(t) anti-aliasing filter sample and hold
sample encoding @fs
time-domain ADC
discretization
amplitude
discretization
digital signal analog signal

FPGA
y[n] y(t)
DAC anti-imaging filter
346
The Nyquist Rate

• The Nyquist sampling theorem defines the minimum number of samples acquired from a band-limited
analog signal per unit time, in order to guarantee the reconstruction of the original signal from these
samples. It requires: fs ≥ 2B
After impulse train sampling with

fs < 2B; Nyquist rate violated
time-domain signal and its samples
After impulse train sampling with

fs > 2B; Nyquist rate fulfilled
band-limited signal in the frequency domain
Ref: https://en.wikipedia.org/wiki/Nyquist-Shannon_sampling_theorem
Further Reading: Alan V. Oppenheim, Alan S. Willsky, and S. Hamid
Nawab. Signals & Systems (2nd Ed.). Prentice-Hall, Inc., 1996
reconstructed signal
347
ADC Encoding Curve

• The mapping between the input voltage of an ADC and the output code can be
described by an encoding curve.
• In a binary encoding ADC with B bits, the input voltage range [Vmin, Vmax] is divided
into 2B segments and any input voltage within this range is approximated with one of
the nearest voltages and represented by a code.
• The ADC encoding curve may be uniform or non-uniform.
• For example, the following are two uniform encoding curves, based on rounding (left)
and truncation (right)
Question: How to quantify the performance of an ADC?

348
ADC Quantization Error Analysis

• The effect of ADC quantization error can be analyzed with a method similar to SNR
calculation due to rounding/truncation. The quantization procedure can be modeled by
a quantization operator Q(•):
yn = Q(xn) = xn + en
xn: ADC input sample (after zero-order hold), yn: quantized result, en: quantization error
• We again use the signal-to-noise ratio (SNR) as the performance measure:

2
𝐸 𝑥𝑛
SNR dB = 10log10
𝐸 𝑒𝑛2
• This analysis requires some assumptions

regarding the input signal and the quantization
error probability density functions
349
ADC Quantization Error Analysis (continued)

Quantization model: yn = Q(xn) = xn + en +Xm
Assumptions: 2B-1 +Xm - ∆
1. The signal xn is a signed real value in [-Xm, Xm)

output input
2. The quantizer is B bit and it divides [-Xm, Xm) into 2B equal code voltage
segments of length ∆= 2𝑋𝑚 /2𝐵
3. The signal xn and the quantization error en are statistically 0 0
independent (we will study the counter assumption later)
4. The quantization error samples en are independent
identically distributed (iid) with a uniform distribution
between −∆/2 and ∆/2 -Xm + ∆
-2B+1
+∞
Therefore: 𝑒=𝐸 𝑒 = −∞
𝑒𝑓𝑒 𝑒 d𝑒 = 0 -2B -Xm
+∞ ∆2 𝑓e(e)
𝜎𝑒2 = 𝐸 (𝑒 − 𝑒)2 = −∞
(𝑒 − 𝑒)2 𝑓𝑒 𝑒 d𝑒 =
12
We have calculated the denominator of the SNR equation. In 1/∆
the sequel we consider three cases for the input signal:
Sinusoidal (deterministic) signal, Gaussian distributed
stochastic signal, Uniformly distributed stochastic signal −∆/2 +∆/2 e
Quantization error probability
density function
350
ADC Quantization SNR with Sinusoidal

Input
• Sinusoidal input signals are the standard measurement method for calculating
ADC SNR.
• Assuming xn=Xmcos(ωn), we have E{xn}=0 and E{xn2} = Xm2/2.
Therefore:
𝑋 2 𝑋 2
𝑚 𝑚
𝐸 𝑥𝑛2 2 = 10log 2
SNR dB = 10log10 = 10log 10 10
𝐸 𝑒𝑛2 ∆2 4𝑋𝑚 2
12 12 × 22𝐵
or
SNR dB ≈ 6.02B + 1.76dB
Note: This is the well-known 6dB per-bit rule, which should be memorized as a
rule of thumb by any hardware engineer!
351
ADC Quantization SNR with Uniformly

Distributed Input
• We next assume that the input signal is a stochastic random variable, uniformly
distributed between –Xm and Xm: xn~U(–Xm, Xm)
• Therefore we have E{xn}=0 and E{xn2} = Xm2/3.
Therefore:
𝑋 2 𝑋 2
𝑚 𝑚
𝐸 𝑥𝑛2 3 = 10log 3
SNR dB = 10log10 = 10log 10 10
𝐸 𝑒𝑛2 ∆2 4𝑋𝑚 2
12 12 × 22𝐵
or
SNR dB ≈ 6.02B
Note: The 1.76dB is no longer there, but we still see the 6dB per-bit property.
352
ADC Quantization SNR with Gaussian

Distributed Input
• We finally assume that the input signal is a stochastic random variable, with a
Gaussian distribution xn~N(0,σx2).
• The Gaussian distribution has infinite tails and overflow at the ADC input is
unavoidable. However, the probability of overflow is reduced by controlling the input
variance σx2 relative to the ADC reference voltages –Xm and Xm.
• Let’s assume Xm=kσx. According to the Gaussian curve, for k = 1, 2, 3, and 4, the
probability of ADC input overflow is 31.73%, 4.55%, 0.26%, and 0.01%,
respectively.
• Assuming k = 4, we have E{xn}=0 and E{xn2} = Xm2/16. Therefore:
2 2
𝑋 𝑚 𝑋 𝑚
𝐸 𝑥𝑛2 16 = 10log 16
SNR dB = 10log10 = 10log 10 10
𝐸 𝑒𝑛2 ∆2 4𝑋𝑚 2
12 12 × 22𝐵
or
SNR dB ≈ 6.02B − 7.27dB
Note: We still see the 6dB per-bit property.
Note: ADC ICs commonly have an out-of-range (OTR) pin for reporting input overflow
per-sample
353
Non-ideal ADC
• Practical ADC circuitry are never ideal and do not reach
their nominal performance (SNR=6.02B + 1.76dB).
• The standard approach to measure the true performance
of an ADC is by giving it a sinusoidal input signal with an
amplitude of 1dB below full-scale (to avoid overflow) and
measuring the real SNR and the effective number of bits
(ENOB):
True SNR measured by giving a
full dynamic-range sinusoidal to
the ADC and measuring the SNR
of an acquired block of data
SNR dB − 1.76dB
ENOB =
The effective number of bits; a 6.02
real-value, always smaller
than the nominal number of
ADC bits (ENOB < B)
354
ENOB Examples
• AD9246 14-Bit, 80 MSPS/105 MSPS/125 MSPS, 1.8 V Analog-to-
Digital Converter:
355
Non-uniform ADC Encoding Curves

• Intuitively, in ADC with non-uniformly distributed inputs, many bits are “wasted” for low-
probability samples (there are profound theoretical proofs behind this intuition). One could
use the bits more efficiently by:
1. Using non-uniform ADC encoding curves: Divide –Xm and Xm into unequal segments
(assign smaller segments to higher probabilities and larger segments to lower
probability values). Example: A-law and μ-law companding algorithms used in old 8-bit
PCM digital communication systems for better use of the dynamic range
2. Making the input sequence distribution uniform: A useful theorem from random
variables:
If a random variable (RV) x with a probability density function (pdf) fX(x) and
cumulative distribution function (CDF) FX(x) passes a nonlinear memoryless
system with a characteristics u = FX(x), the output u is uniformly distributed. Also, if
a uniformly distributed RV u is given to y = FX-1(u), the output has a distribution
fX(•).
Note: This property can be used to make arbitrary RVs from uniform distributions
and vice versa in FPGA.
356
ADC SNR Improvement by Over-Sampling

• Looking back at the quantization model yn = Q(xn) = xn + en, the quantization
error samples en were assumed to be independent identically distributed (iid).
Therefore, the quantization noise has a white spectrum and its total power
E{en2} is equally distributed over the entire Nyquist-band [0,fs].
• If the signal is over-sampled beyond the Nyquist rate, the ADC SNR can be
improved by lowpass filtering the ADC outputs (in the digital domain).
• In this case, we have: SNR dB ≈ 6.02B + 1.76dB + 10log10(OSR), where OSR is
the over-sampling ration (fs/2B)
X(f) FPGA
x(t) xs[n] Lowpass

ADC Processing
Filter
-B +B f
Xs(f) Xs(f)
E(f) E’(f)
-fs-B -fs -fs+B -B +B fs-B fs fs+B f -fs-B -fs -fs+B -B +B fs-B fs fs+B f
357
ADC SNR Improvement by Over-Sampling

(continued)
• Over-sampling rule of thumb: “Each factor of two above the Nyquist rate, is equivalent to
3dB of SNR improvement (after low-pass filtering)”. Therefore, SNR improvement by
OSR is expensive!
• Question: OSR = 4 improves the SNR for 6dB, equivalent to 1 bit of higher resolution.
Does this mean that we can have a mono-bit ADC that is equivalent to a 12-bit ADC?!
Answer: Yes (to some extent)!
358
Spurious-Free Dynamic Range (SFDR)

• Looking back at the quantization model yn = Q(xn) = xn + en, the quantization
error en was assumed to be independent from xn. However, this assumption is
violated in low number of bits.
Spurs are notable components and main component

spikes of noise within a signal’s
spectrum and above the noise floor,
which do not correspond to the noise floor
spur
original signal; but are somehow SFDR
correlated with it (they move in the
spectrum as the sampling
frequency changes or as the signal
components move).
SFDR is the gap (in dB) between

the original frequency component
and the strongest spur
359
Spurious-Free Dynamic Range Improvement

Spurs are very important in practice, as they are
commonly mistaken with the original signal
previous example after dithering
components.
Note: Spurs can also occur during FPGA arithmetic
truncation/rounding
How to improve the SFDR?
1. Increase the number of ADC (quantization)
bits
2. Break the correlation between the signal and
quantization (rounding/truncation) errors by
adding dithers prior to quantization
(rounding/truncation), e.g., by using high-
thermal noise resistors in ADC inputs
Dither is a noise (at the level of the signal’s LSB)
intentionally added to the signal before
quantization to de-correlate the signal and
quantization noise
Note: Dithering improves the SFDR at a cost of decreasing the SNR (increasing the noise floor)
Note: Dithers can be generated in FPGA using linear-feedback shift registers (LFSR)
360
Further Reading on ADC and DAC* (Optional)

• ADC internal technologies: ladder, flash, delta-sigma
modulation
• Integral nonlinearity (INL)
• Clock jitter
• DAC technologies
• Contemporary FPGAs with built-in ADCs
• Quadrature ADC sampling techniques (for high speed)
• Mono-bit technologies
• ADC/DAC tradeoffs
Further reading: refer to the references on ADC/DAC in the course’s

references folder
WORD LENGTH SELECTION IN
FPGA-BASED ARITHMETIC
362
Background
• Real-world applications require the representation of real-valued
data in floating-point or fixed-point formats
• Real numbers can be approximated in these formats using the
necessary number of bits and by proper scaling
Question 1: How many bits should be used for internal calculations?

Answer: Considering that coefficient quantization and
rounding/truncation introduce additional errors to the input data, the
internal register lengths are selected to meet the minimum required
SNR (selected by the designer)
Question 2: How to choose the minimum required SNR?

Answer: It is application-dependent
363
Word-Length Selection in FPGA Designs

• The most common sources of noise in analog
and digital electronics systems are
1. Thermal noise of electronic devices and elements
2. Quantization errors in digital systems, due to number
representation in finite-length registers and
rounding/truncation
• In mixed analog digital designs (containing Thermal noise model

analog elements, ADC, DAC, FPGA, of a resistor
processors, etc.) the conventional standard is
to keep the fixed-point computational errors at
the same level or below the input analog noise
level
364
Input Word-Length Selection Procedure

How to determine the input noise level and internal register lengths?
1. Thermal noise (noise figure) calculation of all analog elements, up to the digital units
(beyond the scope of this course)
2. Calculating the ENOB of the ADC
3. Selecting the processing register lengths such that the internal FPGA quantization
errors are below (or at the same level as) the above items
Note: For pure digital processing or when the input noise level is unknown for the digital
designer, the noise level can be assumed to be half the input register LSB
Ideal System: FPGA

x0(t) Analog x(t) xs[n] ys[n]
Front-End
ADC Processing
Real System:
FPGA
x0(t) x´(t) x´s[n] y´s[n]
Analog
+ Front-End
+ ADC + Processing +
input noise front-end noise ADC quantization noise round-off error noise
365
Input Word-Length Selection Procedure

(continued)
• Note: As far as the FPGA designer is concerned, the input noise and
the analog front-end noise can usually be lumped in the ADC
quantization noise (as factors that reduce the input ENOB)
• For example, with a 16-bit ADC, the 3 LSB may fluctuate due to the
different noise factors (input noise, device thermal noise, ADC
quantization error)
A typical analog front-end with various sources of noise FPGA
Digital
Processing
Image adapted from: http://www.azcom.it/index.php/services/rf-design/analog-front-end-afe/

366
Intermediate Word-Length Selection in

FPGA Designs
Intermediate calculation word-length selection follows similar rules: “try to preserve the
signal-to-noise ratio during calculations, as much as possible”
Q3.12 rounded/truncated;
Example: Q1.15 50-tap FIR lowpass filter with SNR possibly improved
Q1.17 scaled rounded/truncated Q3.21 scaled coefficients by the filtering stage
Q1.15 with cosine/sine
ENOB=13.5 bits
I
× hi[n]
cn=cos(ωn)
|x(t)|<1 16-bit xs[n]
DDS Processing
ADC
sn=sin(ωn)
16-bit ADC with ±1V × hq[n]

Q
reference voltage and
FPGA
possible over-sampling
Q1.15 rounded/truncated
Note 1: The internal register lengths are selected according to the input noise level and ENOB, not the ADC number of bits
Note 2: The SNR can be increased due to the processing gain. For example, remember the SNR improvement due to over-
sampling noted in the previous section
ARBITRARY WAVEFORM
GENERATION
368
Waveform Generation
The calculation/generation of arbitrary functions/waveforms of the
form y = f(x) is required in many computational and signal processing
applications. We study several methods for this purpose:
• Arbitrary functions:
• Direct Implementations (functional calculation)
• Lookup-Tables & Interpolated Lookup-Tables
• Special functions:
• CORDIC machines
• Periodic functions:
• NCO and Periodic Waveform Generators
• Recursive Oscillators
• Random signal:
• LFSR
369
Direct Function Implementation

Depending on the function form, y = f(u) can be implemented using its
direct mathematical form or truncated Taylor expansion:
Example 1: y = f(u) = a·u2 + b·u + c = u·(a·u + b) + c
Requires two multipliers and two adders
u X
X y = f(u)
a +
c
b +
Example 2: y = f(u) ≈ f(a) + f'(a)·(u – a)

Requires a multiplier and two adders for a first-order approximation
Note 1: The implementation of the direct form of a function on FPGA is simplified

when the expansion coefficients are constants or powers of 2.
Note 2: The approximated Taylor expansion is only accurate for smooth functions
370
Functional Implementation by Lookup

Tables (LUT)
• In order to implement y = f(x) over a
finite domain, one may pre-calculate
and store the values of y over the
entire domain of x in a memory. The
values of x can next be used as the
address bus of the memory during
runtime.
• LUT-based implementation of
functions is applicable for arbitrary
functions (not necessarily smooth); M-bits
but requires a lot of memory when x d0
has many bits. d1
x y=f(x)
• The accuracy of this method depends
N bits … M bits
on the function form, and the number
of bits assigned to x (N) and y (M)
d2N-1
371
Functional Implementation by Interpolated

LUT
• For smooth functions, LUT- linear
approximation
based methods can be made error
more memory-efficient, if they
are combined with interpolation
(linear, quadratic, spline, etc.)
• For example, in linear
interpolation, we interpolate
between successive values of
the LUT with appropriate
weights:
𝑥 − 𝑥1 𝑦2 + (𝑥2 − 𝑥)𝑦1 (𝑥 − 𝑥1 )
𝑦≈ = 𝑦1 + (𝑦2 −𝑦1 )
(𝑥2 − 𝑥1 ) (𝑥2 − 𝑥1 )
372
Interpolated LUT Implementation

• Linear interpolated LUTs can be implemented very efficiently using a single or dual-port
LUT and minor computations.
• Idea: Suppose that x has N bits, which means that an LUT of length 2N is required for its
complete implementation. However, if one uses the P MBS bits of x (P < N) for
addressing a 2P points LUT, the N-P LSB bits of x could be used for linear interpolating
between two successive samples of the P-point LUT.
used for LUT addressing used for interpolation
xMSB xLSB
x: 0 0 1 0 1 1 1 0 0 0 1 0 1 1 0 1
xN-1 xN-P-1 x0
• Therefore, we can write 𝑦1 = 𝑓(𝑥MSB ), 𝑦2 = 𝑓(𝑥MSB + 1) and calculate the first-order

interpolation as follows:
(𝑥 − 𝑥1 ) 𝑥LSB × (𝑦2 −𝑦1 )
𝑦 = 𝑦1 + (𝑦2 −𝑦1 ) = 𝑦1 +
(𝑥2 − 𝑥1 ) 2𝑁−𝑃
This division requires only a shift (rounding)

373
Interpolated LUT Implementation Diagram

The overall block-diagram of an interpolated LUT of length N using seen as an N-bit LUT
a two-port LUT of length P for pipelined FPGA-based from outside
implementation is as follows:
Registers (M bits dual-port )
d0
x xMSB (P bits) x1 y1=f(x1)
(N bits) … (M-bits) y ≈ f(x)

Interpolator
(P bits) +1
𝑦=
xLSB x2 y2=f(x2) 𝑦1 + round[𝑥LSB (𝑦2 −𝑦1 ), N − P]
d2P-1 (M-bits)
(N-P bits)
Note: Similar ideas can be implemented using quadratic and spline interpolations. See the following
reference for further ideas and general LUT-based methods: Behrooz, P. (2000). Computer arithmetic:
Algorithms and hardware designs. Oxford University Press, Chapter 24
374
Periodic Signal Generators

An efficient method for generating periodic signals is to combine an LUT with a
numerically controlled oscillator (NCO)
LUT
NCO
1
Address 2
Increment
Sine wave
Clock (Fs) NCO 3
Increment
Accumulator +
N
LUT Address Mixer value
Example: In order to generate a sinusoidal signal with frequency f0 in a sampling rate

fs , using an LUT of length N, the NCO increment can be found as follows:
Nf0
inc =
fs
Note: As a sinusoidal signal, inc should be smaller than N/2 to fulfill the Nyquist
sampling rate.
375
Notes on Periodic Signal Generators

1. Sine and Cosines can be produced using a single two-port LUT with
¼ of initial address offset between the two ports.
2. Sine/cosine generation is precise (with no phase errors), if the

desired frequency (f0 ), sampling frequency (fs), LUT length (N) and
LUT address increment (inc) satisfy:
f0 inc
=
fs N
376
Sine Wave Generator Examples

• Example 1: We want to generate a sine wave with frequency
f0=10.7MHz at a sampling rate of fs=38.4MHz. Noting that
10.7MHz/38.4MHz = 107/384, we can have a 384-point LUT with
inc=107.
• Example 2: We want to generate a sine wave with frequency

f0=10.7MHz at a sampling rate of fs=42.8MHz. Noting that
10.7MHz/42.8MHz = 1/4, we can have a 4-point LUT, which is basically
a 4-state selector that circulates between 0,+1,0, and -1 (no LUT
needed).
• Example 3: We want to make a direct digital synthesizer (DDS) for

generating sine waves at a sampling frequency of fs=100MHz. The DDS
should be able to synthesize frequency from DC to 50MHz (Nyquist
rate), with frequency steps of Δf=100kHz. A LUT of length N=1000 is
required.
377
CORDIC Machines
• The direct implementation of arbitrary functions requires
considerable logic resources and LUT-based methods
require considerable memory.
• Classes of mathematical functions can be generated with a
combination of small-size LUTs and set of shifts and
adds/subtracts.
• The Coordinate Rotation Digital Computer (CORDIC) is
one such method
• The CORDIC machine was invented in 1956 by Jack E.
Volder to be used in B58 bomber's navigation system for
accurate real-time digital calculations
378
Volder’s CORDIC Algorithm

Volder’s original algorithm is a set of recursive multiplier-free equations:
𝑥𝑛+1 = 𝑥𝑛 − 𝑑𝑛 𝑦𝑛 2−𝑛
𝑦𝑛+1 = 𝑦𝑛 + 𝑑𝑛 𝑥𝑛 2−𝑛
𝑧𝑛+1 = 𝑧𝑛 − 𝑑𝑛 arctan 2−𝑛
where
• arctan 2−𝑛 are pre-calculated and stored in a LUT
• 𝑑𝑛 = sign(𝑧𝑛 ) (+1 if 𝑧𝑛 ≥ 0 and −1 if 𝑧𝑛 < 0)
∞ −𝑛
If |𝑧𝑛 | < 𝜃𝑚𝑎𝑥 = 𝑛=0 arctan 2 = 1.7432866 …, it can be shown that:
𝑥𝑛 𝑥0 cos 𝑧0 − 𝑦0 sin 𝑧0
lim 𝑦𝑛 = 𝐾 × 𝑥0 sin 𝑧0 + 𝑦0 cos 𝑧0
𝑛→∞
𝑧𝑛 0
∞
where 𝐾 = 𝑛=0 1 + 2−2𝑛 = 1.6467603 …
379
CORDIC Machine Principles

• The non-restoring decomposition of an arbitrary angle:
∞
𝜃= 𝑑𝑘 𝑤𝑘 , 𝑑𝑘 = ±1, 𝑤𝑘 = tan−1 (2−𝑘 )

𝑘=0
The nonrestoring algorithm:
The following algorithm converges to 𝜃:
𝑡0 = 0
𝑡𝑛+1 = 𝑡𝑛 + 𝑑𝑛 𝑤𝑛
1 if 𝑡𝑛 ≤ 𝜃
𝑑𝑛 =
−1 otherwise
Or in the reverse direction:
𝑡0 = 𝜃
𝑡𝑛+1 = 𝑡𝑛 − 𝑑𝑛 𝑤𝑛 desired angle
1 if 𝑡𝑛 ≥ 0
𝑑𝑛 =
−1 otherwise
380
The CORDIC Algorithm in Circular Rotation

Mode
• According to the restoring algorithm, for an arbitrary angle 𝜃, successive
rotations can be used to rotate from zero to 𝜃 (or from 𝜃 to 0):
𝑥𝑛+1 cos(𝑑𝑛 𝑤𝑛 ) − sin(𝑑𝑛 𝑤𝑛 ) 𝑥𝑛

𝑦𝑛+1 = 𝑦𝑛
sin(𝑑𝑛 𝑤𝑛 ) cos(𝑑𝑛 𝑤𝑛 )
or
𝑥𝑛+1 1 −𝑑𝑛 2−𝑛 𝑥𝑛
𝑦𝑛+1 = cos(𝑤𝑛 ) 𝑑𝑛 2−𝑛 1 𝑦𝑛
• The term cos 𝑤𝑛 = 1/ 1 + 2−2𝑛 is the only required multiplication, which can
be omitted, as it does not alter the rotation angles and only changes the
vector magnitudes.
• Alternatively, depending on the number of iterations 𝑃, A = 1/ 𝑃 𝑛=0 1 + 2
−2𝑛
can be compensated as a constant multiplier.
381
Alternative Forms of the CORDIC Algorithm

• Alternative modes of the CORDIC algorithm include:
Note: The implementation of

CORDIC on FPGA requires
attention in word length
selection and number
representation
Reference and further reading: Muller, Jean-Michel. Elementary functions. Birkhäuser Boston, 2006. Chapter 7
382
CORDIC Implementation on FPGA

• Before implementation, the CORDIC parameters need to
be set:
1. Choose the CORDIC mode
2. Set the input and output lengths and Qm.n data format
3. Find the required number of CORDIC iterations by
simulation, such that the calculation error is smaller than
the LSB of the selected word lengths
4. Implement the CORDIC machine using pipelining or
resource sharing (or a combination of both)
383
CORDIC Implementation on FPGA (continued)

Stage #0 Stage #1 Stage #N-1
• Pipelined: PARAMS PARAMS PARAMS
x0 x1 x2 xN-1 xN
y0 Single y1 Single y2 yN-1 Single yN

Stage Stage … Stage
z0 CORDIC z1 CORDIC z2 zN-1 CORDIC zN
Common Clock
New Data
x0 State
• Resource Shared: y0 Controller
z0 Data Ready
Stage
PARAMS
xN-1 xN
yN-1 Single yN
zN-1 Stage zN
CORDIC
384
Periodic Sequence Generation using

Feedback Shift Registers
• Consider a chain of N registers with a common clock and arbitrary initial
values (known as the seed) connected in feedback:
0 1 0 1 0 0 1 0 1 1
• The generated sequence is apparently periodic with (maximum) period N

samples (N/fs seconds)
• In FPGA, this feedback mechanism can be used to generate special
periodic sequences at a very low cost (using shift registers)
• Next, suppose that the feedback bit is a Boolean function of the
intermediate bits:
𝑥0 𝑛 = 𝑓 𝑥0 𝑛 − 1 , 𝑥1 𝑛 − 1 , … , 𝑥𝑁−1 𝑛 − 1
𝑥1 𝑛 = 𝑥0 𝑛 − 1
…
𝑥𝑁−1 𝑛 = 𝑥𝑁−2 𝑛 − 1
385
Periodic Sequence Generation using

Feedback Shift Registers (continued)
Examples:
A 16-bit Fibonacci LFSR
A 16-bit Galois LFSR

386
Pseudo Random Number Generation

using LFSR
• Linear-feedback shift register (LFSR) is a shift register whose
input bit is a linear function (e.g. XOR, XNOR, etc.) of its
previous state
• The initial value of the LFSR is called the seed
• LFSRs are deterministic FSM, as the output stream is
completely determined by its initial state and the linear function
• Since the register has a finite number of states, LFSR has a
periodic cycle. However, an LFSR with a well-chosen feedback
function can produce a sequence of bits that are pseudo-
random (have a very long period).
• An N-bit LFSR is called maximum-length, if it cycles over all 2N
possible states except 0 (from which it would not exit from)
Ref: https://en.wikipedia.org/wiki/Linear-feedback_shift_register
387
Pseudo Random Number Generation

using LFSR (continued)
• In a maximum-length LFSR The length of LFSR can be selected such
that even at the highest available flip-flop clocking speeds, the
periodicity in not observed in centuries!
• Example: A maximum-length LFSR of length 64 clocked at 1GHz, takes
(264-1)/1GHz ≈ 585 years to repeat itself!
• Moreover, with an appropriate choice of the LFSR length and the
feedback function (also known as the LFSR polynomial), the generated
sequence resembles a fully stochastic sequence, which passes all the
statistical tests of stochastic white noise.
• In this case, the periodic sequence may only be repeated by having the
initial seed.
• LFSR have profound mathematical bases with numerous applications in
coding, security, numeric computation, etc.
Ref: See the following for a nice introduction on the mathematics behind LFSR (Galois Fields):
http://inst.eecs.berkeley.edu/~cs150/sp03/handouts/15/LectureA/lec27-6up
388
Other Applications of LFSR

1. Counters: LFSR can be used as extremely efficient counters (only
requiring shift-registers and a few XOR), when the counting order is
not important. For example for FSM encoding and micro-codes
2. Cyclic Redundancy Check (CRC): LFSR can be used to generate
CRC for error detection and correction
3. Data Encryption/Decryption: LFSRs can be used for encryption of
data transmitted over public channels
Common Seed
(Private)
LFSR LFSR
Generator Generator
Binary sequence generator XOR XOR Recovered data
source public channel destination

389
Other Applications of LFSR (continued)

4. Scramblers: Scramblers are used in many communication and
storage protocols to randomize the transmitted data in order to
remove long sequences of logic zeros and ones.
Ref: Stavinov, Evgeni. 100 power tips for FPGA designers. Evgeni Stavinov, 2011
390
Pseudo Random Numbers with Arbitrary

Distributions
As noted before:
If a random variable (RV) x with a probability density function (pdf) fX(x) and
cumulative distribution function (CDF) FX(x) passes a nonlinear memoryless system
with a characteristics u = FX(x), the output u is uniformly distributed. Also, if a
uniformly distributed RV u is given to y = FX-1(u), the output has a distribution fX(•).
u ~ U(0,1) y ~ fX(x)
FX-1(u)
Random or pseudo-random
uniformly distributed variable
391
Pseudo Random Signals with Arbitrary

Spectral Color*(optional)
Alternative methods for generating signal/noise with arbitrary
spectra include:
• Frequency modulation using fast frequency sweeps (e.g.
using a Chirp signal)
• Bandpass filtering pseudo-random white noise
• Superposition of synthetic signals and noise
PIPELINING & DESIGN TIMING
IMPROVEMENT TECHNIQUES
393
Background
• The notion of pipelining was introduced before, as a means of improving
the design timing, to achieve the design constraints (clock speed)
• Different techniques for pipelining and timing improvement in FPGA
systems are presented in this section, including:
• Retiming
• Re-pipelining
• Cut-set retiming
• C-slow retiming
• Pipelining in feedback systems
References:
• Hauck, Scott, and Andre DeHon. Reconfigurable computing: the theory and
practice of FPGA-based computation. Vol. 1. Elsevier, 2010, Chapter 18
• Khan, Shoab Ahmed. Digital design of signal processing systems: a
practical approach. John Wiley & Sons, 2011, Chapter 7
394
Retiming
• Retiming consists of reducing the critical path (increasing the clock
speed) by moving the pipeline registers to an “optimal position”.
Example: In the following, each circle denotes combination logic, with
the number representing the combinational latency
Before retiming After retiming

• The objective of retiming is to automate this procedure in a systematic
manner with concise algorithms, which 1) guarantee that the circuit’s I/O
transfer function is not changed and 2) can be implemented in CAD
tools (for instance during the synthesis or technology mapping stages)
• Limitation: Retiming cannot improve the design clock speed beyond the
optimal register placement
395
Retiming (continued)
• For systematic retiming, a digital circuit is converted to a data flow graph
(DFG). Next, by using graph theory based theorems, the registers are
systematically moved across the computational nodes (combinational
logic), without changing the input/output transfer function of the original
DFG.
Delay Transfer Theorem: “without affecting the transfer function of the
system, registers can be transferred from each incoming edge of a node
of a DFG, to all outgoing edges of the same node or vice versa” [Khan,
2011, p. 304].
396
Retiming (continued)
• Retiming can also be used to merge excess registers to reduce the area
utilization.
Example:
397
Shannon Decomposition Retiming

• The Shannon decomposition can be used to improve the timing of Boolean
functions. Accordingly:
f(a0,a1,…aN-1) = ā0·f(0,a1,…aN-1) + a0·f(1,a1,…aN-1)
Example:
Note: The Shannon decomposition is specifically useful for FPGA-based designs,

which are implemented on fixed-input LUTs
398
Peripheral Retiming
• In this technique: 1) all the internal registers are shifted to the input or output of
the design; 2) the combinational logic is simplified; finally 3) the registers are
pushed to their optimal position by conventional retiming.
Example:
(1) (2)
(3)
399
Re-pipelining
• In feed-forward designs, re-pipelining adds additional registers at
the input or output and then moves these registers across the
design (by retiming) to obtain the best performance.
• The cost of re-pipelining is the additional number of registers
added to the pipeline which adds a constant clock latency
between the input and output; but the other properties of the
design are preserved.
additional registers added

before re-pipelining
400
Cut-set Retiming
• More generally, cut-set retiming permits the addition of arbitrary number of
registers in a forward path, or moving registers from the input to the output (or
vise versa) of a cut-set, while preserving the I/O transfer function.
• Reminder: In Graph theory, a cut is a virtual partitioning of the edges of a graph
into two disjoint subsets, known as cut-sets.
adding registers in feed-forward cut-sets
moving registers from cut-set output to cut-set inputs

401
Cut-set Retiming (continued)

Example 1: FIR filter retiming
Two possible cut-sets
Re-pipelining across feed-forward cut-set 2

402

Example 2: FIR filter retiming, second approach: multiple cut-set retiming
Three cut-sets with feedback paths After cut-set retiming

403

Example 3: 4-bit ripple carry adder (RCA) retiming
404

Example 4: 4-bit ripple carry adder (RCA) retiming; second approach
405

Example 5: Second-order IIR filter
406

• Cut-set retiming does not always result in an improved timing.
• Example: In a first-order IIR filter, the critical path is not changed by cut-set
retiming of the feedback loop.
407
C-Slow Retiming
• C-slow retiming consists of replicating all the registers of a synchronous design C
times, followed by moving the registers (conventional retiming), or by splitting the
circuit into C distinct parallel paths which multiplex and switch between the input data
and results.
original circuit 2-slow counterpart circuit 2-slow circuit after retiming
Note: The design interleaves between two computations (2-slow): on the first clock
cycle, it accepts the first input for the first data stream; on the second clock cycle, it
accepts the first input for the second stream, and on the third it accepts the second
input for the first stream. Due to the interleaved nature of the design, the two streams
of execution will never interfere (on odd clock cycles, the first stream of execution
accepts input; on even clock cycles, the second stream accepts input).
408
C-Slow Retiming (continued)

Example:
original circuit
2-slow counterpart circuit

409

Example (continued):
• 2-slow retiming after moving the registers to their optimal position (the critical
path is reduced from 5 to 2 time units):
• This architecture can process two parallel data paths with interleaved data
excess feed-forward registers can

be eliminated after retiming
410

• Example: A single 2-slow retimed IIR filter architecture can be used to process
the real and imaginary parts of a complex-valued digital filter by interleaving the
real and imaginary parts of the input:
yr[n] + j yi[n] = h[n]*(xr[n] + j xi[n]) = h[n]*xr[n] + j h[n]xi[n]
xr[0] xi[0] xr[1] xi[1] xr[2] xi[2] xr[3] … yr[0] yi[0] yr[1] yi[1] yr[2] yi[2] yr[3] …
411
C-Slow Retiming by Data Stream

Interleaving
• The disjoint data stream property of C-slow retiming can be
used to obtain parallel hardware threads, which interleave the
input data stream between C identical circuits, each working at
1/C of the input clock rate and finally multiplexing the results
back together. This method is referred to as unfolding in some
textbooks.
• The idea is related to loop unrolling used for optimizing for-
loops in multicore processors and GPUs
• The complementary method is hardware folding (hardware
reuse), which uses a single hardware and a scheduler (FSM
controller) to reduce the hardware size.
• Note: Systematic and ad hoc retiming and resource sharing
may additionally be used to improve the area and timing
performance of the design.
412

Interleaving (continued)
Algorithm: Any DFG can be unfolded by an unfolding factor J using the
following two steps:
S0) To unfold the graph, each node U of the original DFG is replicated J
times as U0,…, UJ-1 in the unfolded DFG.
S1) For two connected nodes U and V in the original DFG with w
delays, draw J edges such that each edge j (= 0,…, J-1) connects
node Uj to node V(j+w)%J with floor[(j+w)/J] delays.
Ref: [Khan, 2011] p. 349

413

Example: Feed-forward example
Original circuit
Unfolded system (2-fold)

414

Example: Feedback systems
Original circuit
Unfolded system (2-fold)

415

Interleaving Example* (optional)
• Example: Polyphase filter Implementation
a switch
416
Pipelining Feedback Systems by Algorithmic

Modifications
• Pipelining digital systems with feedback is a challenging issue and is not always
solved using the previous methods. In this section, we study a few techniques for
pipelining such systems by algorithmic modifications, using a simple case study.
Example: Consider a first-order recursion y[n] = a∙y[n-1] + x[n].

• Such equations appear in many applications, e.g., infinite-impulse response
(IIR) filters in signal processing
• The multiplication is problematic for pipelining, since the result of a∙y[n-1] is
needed for calculating y[n] before the next clock edge arrives
Solution?
417

Modifications (continued)
• The first-order recursion can be rewritten as follows:
y[n] = a∙y[n-1] + x[n] = a ∙ (a∙y[n-2] + x[n-1]) + x[n] = a2∙y[n-2] + a∙x[n-1] + x[n]
• This modified form requires more architecture (compared to the original form); but
it can be pipelined:
418

Modifications (continued)
• More generally:
y[n] = a∙y[n-1] + x[n] = aM∙y[n-M] + (x[n] + a ∙x[n-1] + … + aM-1∙x[n-M+1])
• This form can be pipelined as follows:
Note*: From the signal processing viewpoint,

we are using the following property of the z-
transform of the system response:
H(z) = 1/(1–az-1)
= (1 + az-1 + …+ aM-1z-M+1)/(1– aMz-M)
In other words, we are adding overlapping
zeros and poles to the transfer function, in favor
of pipelining
• This method is known as look ahead transformation in the literature.

419
Architectural Improvements by Algorithmic

Modifications*(optional)
• Replacing a system with it’s algorithmically equivalent counterpart (in favor of
architectural improvement ) is very common in digital implementations.
• Example: Consider a moving average filter (used for lowpass filtering) defined by
the input-output recursion: y[n]=x[n]+x[n-1]+…x[n-N+1]
Accordingly the impulse response and transfer functions of the system are:
h[n]=δ[n]+ δ[n-1]+… δ[n-N+1] or H(z)=1+z-1+…z-N+1
The FPGA implementation of this system requires N-input adders, which can
cause huge combinational delays for large N.
A method for improving this limitation is by using pipelined adder-trees.
Alternatively, one may use the equivalent system: y[n] = y[n-1] + x[n] – x[n-N]
We have used the fact that:
H(z) = (1 + z-1 + …+ z-N+1)
= (1– z-N)/(1– z-1)
• Cascaded Integrator Comb (CIC) also known as Hogenauer filters, which are very
common in FPGA-based designs due to their multiplier-free property, are based on
this method.
420
Further Reading
• Further reading on pipelining, folding and unfolding techniques for feed-forward
and feedback systems:
1. Khan, Shoab Ahmed. Digital design of signal processing systems: a practical
approach. John Wiley & Sons, 2011, Chapter 7.
2. Meyer-Baese, Uwe, and U. Meyer-Baese. Digital signal processing with field
programmable gate arrays. Vol. 2. Berlin: Springer, 2004, Chapter 4.
3. Hauck, Scott, and Andre DeHon. Reconfigurable computing: the theory and
practice of FPGA-based computation. Vol. 1. Elsevier, 2010, Chapter 18
METASTABILITY & MULTIPLE
CLOCK DOMAINS
422
Introduction
• Up to now, we have considered flip-flops and other logic devices as fully
deterministic elements.
• However, in reality, no two flip-flops are “exactly” the same. The (minor) deviations
between the electronic aspects and fabrication indeterminacies of these elements
result in stochastic behaviors.
• Although current FPGA vendors guarantee extremely robust behaviors and
extremely low probabilities of device failures, the consideration of the stochastic
aspects are inevitable in certain cases, including multiple clock domain
applications, which may result in metastability.
• In this section, we study some of the stochastic aspects of digital elements, such
as flip-flops and robust design methods that reduce the probability of metastability
and failure of digital systems.
Reference: M. Arora. The art of hardware architecture: Design methods and

techniques for digital circuits. Springer Science & Business Media, 2011.
423
Review of Logic Circuits Timing Parameters

• Clock period (tC): clock edge-to-edge time;
inverse of clock frequency (fC)
• Clock Skew (tskew): indeterminacy of the clock
edge arrival time
• Setup Time (tsetup): data should be stable before
clock edge
• Hold Time (thold): data should be stable after clock
edge
• Propagation Delay (tCQ): clock edge to stable
output
• Combinational delay (tlogic): combinational logic
circuit settling time
• Setup Slack (tslack): minimum data required time
minus data arrival time:
• Positive: timing met
• Negative: timing violated
• We want: tC ≥ tCQ + tlogic + tlogic + tskew Note: HIGH-to-LOW and LOW-to-HIGH
times are not necessarily the same
424
Review of Logic Circuits Timing Parameters

(continued)
• Note: All the listed parameters are stochastic in reality
(vary over time and space)
• In single clock designs, the clock frequency (fC) is
selected such that the slack requirement is met. The
maximum clock reported by synthesis tools is based on
such calculations
• In multiple clock designs, the timing cannot be
guaranteed when crossing between clock domains
• Result: The output logic is not known (HIGH, LOW, or
even a voltage in between). This is known as
metastability
425
Metastability
Metastability can occur when:
1. A flip-flop’s slack timing is
violated (high clock rate)
2. The data input to a flip-flop is
asynchronous to the clock
(leading to setup or hold-time
violations)
3. When using multiple un-
synchronized clock domains.
• During metastability tCQ becomes longer than its nominal value.

• The additional time beyond tCQ, which a metastable circuit
requires to become stable is called the settling time (tMET)
426
Metastability Examples
427
Statistical Analysis of Metastability

How often does metastability occur?
Considering tC as the FF clock period (inverse of fC), tD as the
asynchronous data period (inverse of fD), and w as the
metastability window length:
• Considering the data transition probability to be uniform over the
entire clock period and independent of the clock, the probability
of data transition during a metastable window is w/tC=w·fC
• Therefore, the rate of metastability is w·fC·fD (times per seconds)
428
Statistical Analysis of Metastability (continued)

How long does it take to recover from metastability?
• It can be shown that the electronic properties of flip-flops
eventually take it back a stable state (0 or 1)
• Assuming that a flip-flop becomes metastable at t=0, the
probability of remaining in metastability after tMET seconds has
been shown to be (approximately) exponentially decaying over
time, i.e.:
−tMET
Pr(staying metastable ≥ t MET )=e 𝜏
where 𝜏 is a device and technology dependent parameter.
• Reference: Ginosar, Ran. "Metastability and synchronizers: A tutorial." IEEE Design &
Test of Computers 28.5 (2011): 23-35.
429
Statistical Analysis of Metastability (continued)

Probability of Failure:
• If the output of a flip-flop is sampled tMET seconds after the clock
edge, the probability of failure (malfunction) is
Pr(failure) = Pr(enter metastability AND stay metastable tMET or longer)
• The above two events are statistically independent. Hence:

Pr(failure) = Pr(enter metastability)·Pr(stay metastable tMET or longer)
430
Mean Time Between Failures (MTBF) for

Metastable Flip-Flops
The industrial standard formula for Failure Rate and Mean Time Between Failures
(MTBF) of a single stage metastable flip-flop is:
1 −tMET
Failure Rate = = fD . Pr(failure) = fD . W. fC × e 𝜏
MTBF
Metastable window probability The probability of remaining in
(how often we are in a metastable window) metastability for tMET seconds
where:
• fC : system clock rate (Flip-Flop clock)
• fD : (asynchronous) input data clock rate
• W: metastability window length constant
• τ: metastability time constant
• t MET : time delay for the metastability to resolve itself
Note: W and τ are constants depending on the setup-time and hold-time of the device
(vendor and technology dependent)
431
MTBF Calculation
Example 1: Consider a 28nm ASIC high-performance CMOS with
W=20ps and 𝜏=10ps (typical values for this process technology).
Assuming fC =1GHz and fD =100MHz, we find MTBF=4x1029 years
for a single-stage synchronizer (the universe is estimated to be 1010
years old).
432
MTBF Calculation (continued)

Example 2: Suppose we want to guarantee a 1year MTBF (approximately 3×107s) on an Altera
FLEX 10K CPLD. The MTBF constants of this family of Altera devices can be seen in the table
below. In certain devices of this family tsetup= 1.6ns. For a data frequency fD = 20MHz and clock
frequency fC = 80MHz we have:
ln 3 × 107 + ln[ 80 × 106 20 × 106 1.01 × 10−13 ]
t MET = = 1.76ns
1.268 × 1010
tsetup=1.6ns
In this example the combination circuit ?
shown in the figure can have the fD=20MHz
following maximum combinational delay
to fulfil the required MTBF:
tlogic ≤ 12.5ns – 1.76ns – 1.6ns = 9.14ns
fC=80MHz
W 1/τ
Note: Due to the logarithmic form
of the equation, increasing t MET to
2.12ns increases the MTBF to 100
years.
Ref: Metastability in Altera Devices (May 1999, Available: ftp://ftp.altera.com/pub/lit_req/document/an/an042.pdf)

433
MTBF of Multistage Synchronizers

For multistage synchronizers:
1 tMET1 tMET2
MTBF = ×e 𝜏 ×e 𝜏 ×⋯
W. fC . fD
where t MET1 , t MET2 , etc. are the time delay for the metastability to resolve itself in
each synchronizer stage.
How many synchronizer stages are required? The parameters W and τ are
commonly provided by IC manufacturers; fC and fD are also known by-design. The
designer can define a desired MTBF, calculate t MET and decide about the number of
required stages to fulfil the required MTBF.
434
Metastability Guidelines
Avoiding metastability (by design):
1. Avoiding real-time data transfer between different clock domains
2. Using a single global clock instead of multiple clock domains
3. Avoiding gated clocks and using standard clock decreasing techniques (using
clock enable)
Solving metastability (by implementation):
1. Clock synchronization using DCMs
2. Using synchronizers (register chains and asynchronous FIFOs) to reduce the
probability of metastability
Note: These methods only resolve metastability; but do not solve other rate
mismatch issues, when transferring data between different clock domains. For
example, sampling a data that changes with fD=80MHz, at a clock rate of
fC=100MHz, results in regular repeated samples and sampling it at fC=60MHz
results in regular data loss (even without metastability).
435
Metastability Guidelines (continued)

Example: Using FIFOs while crossing different clock domains
FIFO
436
FIFO Size Selection

How to select the FIFO size? The overall producer data rate should
not exceed the consumer rate of processing the data.
Note: A FIFO can not overcome rate differences (no matter how deep
it is, it’ll eventually overflow if the producer’s data rate is consistently
higher than consumer’s). A FIFO can only overcome temporary
producer-consumer rate differences by buffering the excess data.
437
Applications: Metastability due to Top-

Module Asynchronous Inputs
The standard procedure for working with top-module (asynchronous) inputs is to
pass them through one or more layers of flop-flops before any internal usage.
asynchronous
input
synchronization
register chain
unpredictable
routing delays
FPGA FPGA FPGA

Not recommended Standard method Standard method
(reduced risk of metastability)
Note: The probability of metastability decreases by increasing the number of FF layers

Question: How to handle asynchronous input buses (group of asynchronous inputs)?
Answer: By placing user defined constraints over the bus routing length.
438
Applications: Metastability in Two-Way

Control/Acknowledge Systems
• Reference: Ginosar, Ran. "Metastability and synchronizers: A tutorial." IEEE

Design & Test of Computers 28.5 (2011): 23-35.
439
Flip-Flop MTBF in Xilinx FPGA Table legend:
Example: Xilinx Virtex II, metastability datasheet
Ref: https://china.xilinx.com/support/documentation/application_notes/xapp094.pdf
Note: Xilinx doesn’t seem to list the FF MTBF of its newer devices; but it reports them in Vivado® during
implementation.
440
Xilinx’s Metastability Test Circuit
Ref: Xilinx Metastability Considerations (XAPP077.pdf January 1997, Available:

http://userweb.eng.gla.ac.uk/scott.roy/DCD3/technotes.pdf)
441
Xilinx’s Metastability Test Results
Ref: Xilinx Metastability Considerations (XAPP077.pdf January 1997, Available:

http://userweb.eng.gla.ac.uk/scott.roy/DCD3/technotes.pdf)
442
Altera’s Metastability Test Circuit
Ref: ftp://ftp.altera.com/pub/lit_req/document/an/an042.pdf
443
Altera’s Metastability Test Results
Ref: Metastability in Altera Devices (May 1999, Available: ftp://ftp.altera.com/pub/lit_req/document/an/an042.pdf)

444
Further Readings on Metastability

• Kilts, Steve. Advanced FPGA design: architecture, implementation,
and optimization. John Wiley & Sons, 2007.
• Arora, Mohit. The art of hardware architecture: Design methods and
techniques for digital circuits. Springer Science & Business Media,
2011.
• Ginosar, Ran. "Metastability and synchronizers: A tutorial." IEEE
Design & Test of Computers 28.5 (2011): 23-35.
• http://www.ti.com/jp/lit/an/scza004a/scza004a.pdf
• http://userweb.eng.gla.ac.uk/scott.roy/DCD3/technotes.pdf
• https://www.altera.com/en_US/pdfs/literature/wp/wp-01082-quartus-
ii-metastability.pdf
MEMORY-MAP DESIGN IN
FPGA-BASED SYSTEMS
446
Introduction**
• Complex FPGA-based systems can contain multiple units
(modules), each having multiple operation modes that are
selected by appropriate control pins (or control bus) and give
output messages in different occasions (handshakes, error
codes, overflow flags, etc.)
• Each element of a design should have a unique address in the
system’s memory map, which can be accessed via proper
commands
• In mixed CPU-FPGA systems, the internal memory map of the
FPGA is commonly accessible by the software units
• The design of a memory map is discussed in this section by
examples
**This section is presented from industrial project source codes

447
Example: Xilinx ML605 Virtex-6 Evaluation

Board
FMC110
ML605
448
Example: Xilinx ML605 Virtex-6 Evaluation

Board (continued)
449
Memory Map
Memory map conceptual illustration

450
Accessing the Memory Map

• The internal memory map (MMap) of the FPGA and
the protocol for accessing the MM is designed and
implemented by the FPGA designer
• The MMap can be accessed through any of the I/O
ports of the FPGA board. For example:
• Ethernet
• PCI-e
• JTAG
• USB
• Etc.
451
Accessing the Memory Map (continued)

• Example: Suppose that we use the Ethernet as the access port
of the ML605 FPGA board. The FPGA board can send and
receive Ethernet packets, which can have an arbitrary format
after decoding:
FPGA Ethernet PC
Decoded Ethernet packets (arbitrary format defined by the designer):

flags address data
pN-1 p0
452
Memory Map Implementation Techniques

• Centralized: All modules have a set of input ports for
commands and output ports for handshaking and messages. All
input commands (to the FPGA) or output messages (from the
FPGA) are handles by a single module (a command or
message dispatcher), which has access to the command ports
of all modules. The only command/message interface of the
FPGA to the output world is this module.
• Distributed: There are no centralized command/message

dispatchers. A common command bus is shared between all
modules. Each module has a unique address (or address offset
with respect to the top-module, for nested modules) in the
memory map of the system. The commands/messages are
handled locally by each module.
453
Centralized Memory Map Design

data path
Module 1 Module 2 Module 3
Command
FPGA Dispatcher commands/messages/
variable parameters
Ethernet, PCI-e, JTAG, USB,

Master system USART,…
controller PC/Microcontroller/FPGA
454
Distributed Memory Map Design

data path
Module 1 Module 2 Module 3
local command Common Bus

dispatchers (commands/messages/
Bus Handler variable parameters)
FPGA
Ethernet, PCI-e, JTAG, USB,

Master system USART,…
controller PC/Microcontroller/FPGA
455
Nested Memory Maps

Module 2
local
Submodule 1 Submodule 2 command
data path bus
Module 1
Common Bus
local command (commands/messages/
dispatchers
Bus Handler variable parameters)
456
Centralized vs Distributed Memory Maps

Advantages Drawbacks
Centralized • Less prone to design errors and bus • All command/message
write conflicts (centralized command ports appear as
dispatching) input/output ports of
• Simpler for constructing the memory modules (more
map (explicit memory map addresses) complication in the top-
• No local command dispatchers module)
 Recommended for small and medium
size designs
Distributed • Simplified top-module • More prone to design
• No centralized command dispatchers errors and bus handling
required by individual modules
• Simpler for extension (similar module • More complicated memory
instances can be added to the design in map encoding/decoding
a “plug-and-play” like manner) • Each module requires a
 Recommended for complicated designs command dispatcher
with possible future extensions
DATA COMMUNICATION
METHODS & PROTOCOLS
458
Introduction
• As with other aspects of FPGA designs, data transfer
inside FPGA and between FPGA systems can be fully
customized.
• In this section we review the most common techniques
used for data transfer in FPGA designs
• The two classes of data transfer methods that we study
are:
• Stream Transfer
• Packet Transfer
459
Continuous Stream Data Transfer

• Stream Transfer: used for continuous and synchronous data
transfer between modules
• Usage: ADC, DAC, continuous data streams
• Advantage: no handshaking overheads; can use the maximum

possible throughput between two endpoints
• Disadvantage: requires synchronization; even minor
asynchrony between the sender and receiver clocks results in
metastability, data replication or data loss
• Note: depending on the processing algorithm, continuous data

streams can be up-sampled or down-sampled throughout
processing
460
Packet (Block) Data Transfer

• Packet (Block) Transfer: used for discrete data transfer
between modules
• Usage: data/message communication between
asynchronous modules
• Advantage: enables data transfer between different clock

domains; robust to minor sender/receiver clock frequency
mismatch (depending on the block size)
• Disadvantage: requires handshaking, packing overhead
(start/stop/CRC words), reduced bandwidth and
packing/unpacking hardware overheads
461
Block Processing of Streamed Data

• A common requirement in many data processing systems is the
block-wise processing of continuous data streams. Examples
include: DFT filtering, Reed-Solomon encoding, H.264
encoding, etc.
• The standard technique for implementing such algorithms is to
use a dual-buffer at the interface between the continuous data
stream and the block processor.
• As a rule if the block-wise algorithm processes a block of data
faster than the data stream is accumulated in the input buffer
(and read from the output buffer), no data loss occurs in the
input (or output) and the block processing is masked from the
outer world.
462
Block Processing of Streamed Data

(continued)
Dual-buffer implementation: When input is streamed in InBuff1, the block processor
is working on previous data written in InBuff2. When the block processor is
downloading its results in OutBuff2, the previous results are streamed from OutBuff1
to the output, etc.
Size: N samples Size: N samples
Output Buffer 1
Input Buffer 1
in read
in write mode
mode
switch
switch
switch
switch
Block Data
Continuous Processor Continuous
Output Buffer 2
Input Buffer 2
input data output data

stream @fs stream @fs
in write
in read
mode
mode
processing time < N/fs

463
The ARM Advanced Microcontroller Bus

Architecture (AMBA)
• Although on-chip data communications are rather arbitrary
(especially in FPGA-based systems), standard protocols
have been developed, which are currently adopted and
supported by many processor, FPGA and ASIC vendors.
• The Arm AMBA is an open standard for the connection,
management and communication of functional blocks in a
system-on-a-chip (SoC), including FPGA-based systems.
• The AMBA AXI4 and AXI-Lite protocols are currently used
in many Xilinx tools and IP cores
• AMBA AXI uses READY/VALID handshaking mechanisms
464
AMBA AXI4 and AXI-Lite Interfaces

• AXI4 and AXI-Lite interfaces consist
of five different channels:
• Read Address Channel
• Write Address Channel
• Read Data Channel
• Write Data Channel
• Write Response Channel
References and further reading on AXI

interface protocols:
• AXI Reference Guide,
https://www.xilinx.com/support/docu
mentation/ip_documentation/ug761_
axi_reference_guide.pdf
• AMBA® AXI™ and ACE™ Protocol
Specification,
https://www.arm.com/products/syste
m-ip/amba-specifications
• AXI4-StreamingtoStellarIP Interface,
http://www.4dsp.com/pdf/AN001_KC
705_FMC104_AXI_FFTcore_tutorial.
pdf
SCALABLE DESIGNS AND
AUTOMATIC HDL CODE
GENERATION
466
Scalable Design and Automatic HDL Code

Generation
• Verilog and VHDL have limited features for scalable and
parametric designs (such as genvar, generate, etc.)
• In this section, we will learn how to write scripts in other
languages (C, Java, Python, Matlab, etc.) to generate
synthesizable HDL codes
• These methods can be used to generate user defined HDL
libraries, Netlists and EDIF files.
• The basic idea is to open a .v or .vhd file in another
language and start writing in it with Verilog or VHDL
supported syntax, while using the flexibilities and features
of the higher level language.
467

Generation (continued)
Example 1: Matlab script for generating Running DFT Verilog code
468

Example 1 (continued): Output Verilog file
469

Example 1 (continued): Output Verilog file continued
470

Example 2: Generating Multilayer Perceptron Artificial Neural
Networks RTL codes in C# (By Pejman Torabi, Shiraz University)
471

Example 2 (continued): Generated modules
module TOP ( In1, In2, In3, In4, In5, In6, In7, In8,
Out1, Out2, Out3, Out4, Out5, Out6, Out7, Out8,
clk, en, res);
module Layer1 ( i1 ,i2, i3, i4, i5, i6, i7, i8,
w001001,w001002,w001003,w001004,w001005,w001006,w001007,w001008, B001, …
clk,en,res);
module Layer ( i1, i2, i3, i4, i5, i6, i7, i8,
w001001, w001002, w001003, w001004, w001005, w001006, w001007,w001008, B001,…
clk, en, res);
module ActFunc ( In_AF1, In_AF2, In_AF3, In_AF4, In_AF5, In_AF6, In_AF7, In_AF8,
Out_AF1, Out_AF2, Out_AF3, Out_AF4, Out_AF5, Out_AF6, Out_AF7, Out_AF8,
clk, en, res);
module Function_Interpolation (inputVal, outputVal, clk);
module mult (a, b, z, clk);
module Layer_End ( i1,i2,i3,i4,i5,i6,i7,i8,
w001001, w001002, w001003, w001004, w001005, w001006, w001007, w001008, B001,
…
clk, en, res);
module ActFunc_End (In_AF1, In_AF2, In_AF3, In_AF4, In_AF5, In_AF6, In_AF7, In_AF8,
Out_AF1, Out_AF2, Out_AF3, Out_AF4, Out_AF5, Out_AF6, Out_AF7, Out_AF8,
clk,en,res);
472

Example 2 (continued): RTL schematic of the generated codes
473

Example 3: Xilinx HEX file generation in Matlab
474

Example 3 (continued): Output HEX file
…
475

Example 4: Xilinx coefficient file generation in C
476

Example 4 (continued): Output COE file
477

Example 5: Automatic listing generation for LaTeX reports. Project reports (specifically
in LaTeX) can be automatically updated with the latest version of the source codes
478

Example 5: Output LaTeX listing
479
Further Examples
• CORDIC core generators
• LFSR generators
• Fast Fourier Transform (FFT) architecture generator
FPGA DESIGN
DOCUMENTATION
481
Hardware Documentation**
• Design documentation is a necessary and essential part
of any engineering project
• Both specific and general documentation tools and
techniques can be used for hardware documentation
• Some of these techniques and tools are reviewed in this
section by example: Doxygen, LaTeX, etc.
** This section is presented from industrial project reports

ADVANCED TOPICS*
(Optional)
483
Advanced Topics in FPGA Design

In this section some of the advanced topics in FPGA
designs are introduced by presenting a general overview:
• User Constraint File (UCF) format and options
• Fault tolerant designs and redundancy
• Robust FSM implementations
• Power analysis and power efficient designs
• Micro programmed FSM
• MicroBlaze™ Technology
• Hardware Trojans and backdoors in FPGA designs
• FPGA board PCB considerations
484
Power Analysis and Management

• Power efficient FPGA design is a critical issue in current industrial level
systems.
• Various power management techniques are reviewed in this section.
• Xilinx power analysis tools: PlanAhead Power Estimator, XPower Estimator
(XPE), XPower Analyzer (XPA)
References:
1. Stavinov, E. (2011). 100 power tips for FPGA designers. Evgeni
Stavinov.
2. Xilinx Power Solutions http://xilinx.com/power
3. Seven Steps to an Accurate Power Estimation using XPE, Xilinx White
Paper WP353
http://www.xilinx.com/support/documentation/white_papers/wp353.pdf
4. XPower User Guide, Xilinx User Guide UG440
http://www.xilinx.com/support/documentation/user_guides/ug440.pdf
View publication stats

Course Slides 2018

Uploaded by

Copyright:

Available Formats

Course Slides 2018

Uploaded by

Document Information

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

Course Slides 2018

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Digital Systems Design Course Slides

Presentation · June 2018

The user has requested enhancement of the downloaded file.

School of Electrical & Computer Engineering

Revision: June 2018

 Please use the following for citing these notes:

The von Neumann Architecture

The von Neumann Architecture

1. A memory for storing program and data (Harvard

The von Neumann Architecture

Pipelining in von Neumann Architectures

Pipelining in von Neumann Architectures

Harvard Architecture Pipelining

Application Specific Hardware versus von

VN versus ASIP: Considering 5 cycles per instruction, the VN should be 15 times

Application specific hardware have higher performance, at a cost of lower flexibility

Flexibility vs. Performance

Flexibility vs. performance of different architectures (Bobda 2007)

Applications of Reconfigurable Architectures

• Bobda, C. (2007). Introduction to reconfigurable

Programmable ROM (PROM)

Programmable ROM (PROM)

Benefit: Any logic circuit may be implemented

Standard Logic Chips

• TTL (Transistor-Transistor Logic) Technology:

Design Example Using the 74000 Series

Design Example Using the 74000 Series

An Implementation of a 4-bit two register computer, including 6 CPU assembly instructions:

Programmable Logic Technologies

PLA (Programmable Logic Array) Technology

PLA (Programmable Logic Array) Technology

PAL (Programmable Array Logic) Technology

PAL Extra Circuitry

*Input – Output selection

CPLD (Complex Programmable Logic Device)

CPLD (Complex Programmable Logic Device)

Equivalent Number of Gates

• Equivalent gates: total number of two input NAND gates

CPLD (Complex Programmable Logic Device)

FPGA (Field-Programmable Gate Array)

FPGA Configurable Logic Blocks (CLB)

Two-input Look Up Table (LUT)

FPGA Configurable Logic Blocks (CLB)

FPGA Configurable Logic Blocks (CLB)

Three-input Look Up Table (LUT)

Programmed FPGA Scheme

• Other switching technologies: Flash-based, Anti-fuse, etc.

FPGA Logic Block Extra Circuitry

CPLD vs. FPGA

• CPLD keeps its contents without power, also known as

• FPGA storage cells are volatile (lose their contents when

(a) Off-board programming (b) On-board programming using

* JTAG: Joint Test Action Group

JTAG Daisy Chain

Other JTAG Applications

ASIC (Application Specific Integrated Circuit)

• In CPLD and FPGA programmable switches consume

• Alternatively, the chip layout can be totally customized;

• Compromise: The design may be simplified by using