Skip to content

Commit aa5caf8

Browse files
alalekalalek
committed
Merge pull request opencv#8072 from tomoaki0705:AVXcorner
2 parents 993ff7a + 07979b0 commit aa5caf8

File tree

1 file changed

+145
-11
lines changed

1 file changed

+145
-11
lines changed

‎modules/imgproc/src/corner.cpp

Lines changed: 145 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -48,12 +48,86 @@
4848
namespace cv
4949
{
5050

51+
#if CV_AVX
52+
// load three 8-packed float vector and deinterleave
53+
// probably it's better to write down somewhere else
54+
static inline void load_deinterleave(const float* ptr, __m256& a, __m256& b, __m256& c)
55+
{
56+
__m256 s0 = _mm256_loadu_ps(ptr); // a0, b0, c0, a1, b1, c1, a2, b2,
57+
__m256 s1 = _mm256_loadu_ps(ptr + 8); // c2, a3, b3, c3, a4, b4, c4, a5,
58+
__m256 s2 = _mm256_loadu_ps(ptr + 16); // b5, c5, a6, b6, c6, a7, b7, c7,
59+
__m256 s3 = _mm256_permute2f128_ps(s1, s2, 0x21); // a4, b4, c4, a5, b5, c5, a6, b6,
60+
__m256 s4 = _mm256_permute2f128_ps(s2, s2, 0x33); // c6, a7, b7, c7, c6, a7, b7, c7,
61+
62+
__m256 v00 = _mm256_unpacklo_ps(s0, s3); // a0, a4, b0, b4, b1, b5, c1, c5,
63+
__m256 v01 = _mm256_unpackhi_ps(s0, s3); // c0, c4, a1, a5, a2, a6, b2, b6,
64+
__m256 v02 = _mm256_unpacklo_ps(s1, s4); // c2, c6, a3, a7, x, x, x, x,
65+
__m256 v03 = _mm256_unpackhi_ps(s1, s4); // b3, b7, c3, c7, x, x, x, x,
66+
__m256 v04 = _mm256_permute2f128_ps(v02, v03, 0x20); // c2, c6, a3, a7, b3, b7, c3, c7,
67+
__m256 v05 = _mm256_permute2f128_ps(v01, v03, 0x21); // a2, a6, b2, b6, b3, b7, c3, c7,
68+
69+
__m256 v10 = _mm256_unpacklo_ps(v00, v05); // a0, a2, a4, a6, b1, b3, b5, b7,
70+
__m256 v11 = _mm256_unpackhi_ps(v00, v05); // b0, b2, b4, b6, c1, c3, c5, c7,
71+
__m256 v12 = _mm256_unpacklo_ps(v01, v04); // c0, c2, c4, c6, x, x, x, x,
72+
__m256 v13 = _mm256_unpackhi_ps(v01, v04); // a1, a3, a5, a7, x, x, x, x,
73+
__m256 v14 = _mm256_permute2f128_ps(v11, v12, 0x20); // b0, b2, b4, b6, c0, c2, c4, c6,
74+
__m256 v15 = _mm256_permute2f128_ps(v10, v11, 0x31); // b1, b3, b5, b7, c1, c3, c5, c7,
75+
76+
__m256 v20 = _mm256_unpacklo_ps(v14, v15); // b0, b1, b2, b3, c0, c1, c2, c3,
77+
__m256 v21 = _mm256_unpackhi_ps(v14, v15); // b4, b5, b6, b7, c4, c5, c6, c7,
78+
__m256 v22 = _mm256_unpacklo_ps(v10, v13); // a0, a1, a2, a3, x, x, x, x,
79+
__m256 v23 = _mm256_unpackhi_ps(v10, v13); // a4, a5, a6, a7, x, x, x, x,
80+
81+
a = _mm256_permute2f128_ps(v22, v23, 0x20); // a0, a1, a2, a3, a4, a5, a6, a7,
82+
b = _mm256_permute2f128_ps(v20, v21, 0x20); // b0, b1, b2, b3, b4, b5, b6, b7,
83+
c = _mm256_permute2f128_ps(v20, v21, 0x31); // c0, c1, c2, c3, c4, c5, c6, c7,
84+
}
85+
86+
// realign four 3-packed vector to three 4-packed vector
87+
static inline void v_pack4x3to3x4(const __m128i& s0, const __m128i& s1, const __m128i& s2, const __m128i& s3, __m128i& d0, __m128i& d1, __m128i& d2)
88+
{
89+
d0 = _mm_or_si128(s0, _mm_slli_si128(s1, 12));
90+
d1 = _mm_or_si128(_mm_srli_si128(s1, 4), _mm_slli_si128(s2, 8));
91+
d2 = _mm_or_si128(_mm_srli_si128(s2, 8), _mm_slli_si128(s3, 4));
92+
}
93+
94+
// separate high and low 128 bit and cast to __m128i
95+
static inline void v_separate_lo_hi(const __m256& src, __m128i& lo, __m128i& hi)
96+
{
97+
lo = _mm_castps_si128(_mm256_castps256_ps128(src));
98+
hi = _mm_castps_si128(_mm256_extractf128_ps(src, 1));
99+
}
100+
101+
// interleave three 8-float vector and store
102+
static inline void store_interleave(float* ptr, const __m256& a, const __m256& b, const __m256& c)
103+
{
104+
__m128i a0, a1, b0, b1, c0, c1;
105+
v_separate_lo_hi(a, a0, a1);
106+
v_separate_lo_hi(b, b0, b1);
107+
v_separate_lo_hi(c, c0, c1);
108+
109+
v_uint32x4 z = v_setzero_u32();
110+
v_uint32x4 u0, u1, u2, u3;
111+
v_transpose4x4(v_uint32x4(a0), v_uint32x4(b0), v_uint32x4(c0), z, u0, u1, u2, u3);
112+
v_pack4x3to3x4(u0.val, u1.val, u2.val, u3.val, a0, b0, c0);
113+
v_transpose4x4(v_uint32x4(a1), v_uint32x4(b1), v_uint32x4(c1), z, u0, u1, u2, u3);
114+
v_pack4x3to3x4(u0.val, u1.val, u2.val, u3.val, a1, b1, c1);
115+
116+
_mm256_storeu_ps(ptr, _mm256_setr_m128(_mm_castsi128_ps(a0), _mm_castsi128_ps(b0)));
117+
_mm256_storeu_ps(ptr + 8, _mm256_setr_m128(_mm_castsi128_ps(c0), _mm_castsi128_ps(a1)));
118+
_mm256_storeu_ps(ptr + 16, _mm256_setr_m128(_mm_castsi128_ps(b1), _mm_castsi128_ps(c1)));
119+
}
120+
#endif // CV_AVX
121+
51122
static void calcMinEigenVal( const Mat& _cov, Mat& _dst )
52123
{
53124
int i, j;
54125
Size size = _cov.size();
126+
#if CV_AVX
127+
bool haveAvx = checkHardwareSupport(CV_CPU_AVX);
128+
#endif
55129
#if CV_SIMD128
56-
bool simd = hasSIMD128();
130+
bool haveSimd = hasSIMD128();
57131
#endif
58132

59133
if( _cov.isContinuous() && _dst.isContinuous() )
@@ -67,8 +141,25 @@ static void calcMinEigenVal( const Mat& _cov, Mat& _dst )
67141
const float* cov = _cov.ptr<float>(i);
68142
float* dst = _dst.ptr<float>(i);
69143
j = 0;
70-
#if CV_SIMD128
71-
if( simd )
144+
#if CV_AVX
145+
if( haveAvx )
146+
{
147+
__m256 half = _mm256_set1_ps(0.5f);
148+
for( ; j <= size.width - 8; j += 8 )
149+
{
150+
__m256 v_a, v_b, v_c, v_t;
151+
load_deinterleave(cov + j*3, v_a, v_b, v_c);
152+
v_a = _mm256_mul_ps(v_a, half);
153+
v_c = _mm256_mul_ps(v_c, half);
154+
v_t = _mm256_sub_ps(v_a, v_c);
155+
v_t = _mm256_add_ps(_mm256_mul_ps(v_b, v_b), _mm256_mul_ps(v_t, v_t));
156+
_mm256_storeu_ps(dst + j, _mm256_sub_ps(_mm256_add_ps(v_a, v_c), _mm256_sqrt_ps(v_t)));
157+
}
158+
}
159+
#endif // CV_AVX
160+
161+
#if CV_SIMD128
162+
if( haveSimd )
72163
{
73164
v_float32x4 half = v_setall_f32(0.5f);
74165
for( ; j <= size.width - v_float32x4::nlanes; j += v_float32x4::nlanes )
@@ -82,7 +173,8 @@ static void calcMinEigenVal( const Mat& _cov, Mat& _dst )
82173
v_store(dst + j, (v_a + v_c) - v_sqrt(v_t));
83174
}
84175
}
85-
#endif
176+
#endif // CV_SIMD128
177+
86178
for( ; j < size.width; j++ )
87179
{
88180
float a = cov[j*3]*0.5f;
@@ -98,8 +190,11 @@ static void calcHarris( const Mat& _cov, Mat& _dst, double k )
98190
{
99191
int i, j;
100192
Size size = _cov.size();
193+
#if CV_AVX
194+
bool haveAvx = checkHardwareSupport(CV_CPU_AVX);
195+
#endif
101196
#if CV_SIMD128
102-
bool simd = hasSIMD128();
197+
bool haveSimd = hasSIMD128();
103198
#endif
104199

105200
if( _cov.isContinuous() && _dst.isContinuous() )
@@ -114,8 +209,26 @@ static void calcHarris( const Mat& _cov, Mat& _dst, double k )
114209
float* dst = _dst.ptr<float>(i);
115210
j = 0;
116211

117-
#if CV_SIMD128
118-
if( simd )
212+
#if CV_AVX
213+
if( haveAvx )
214+
{
215+
__m256 v_k = _mm256_set1_ps((float)k);
216+
217+
for( ; j <= size.width - 8; j += 8 )
218+
{
219+
__m256 v_a, v_b, v_c;
220+
load_deinterleave(cov + j * 3, v_a, v_b, v_c);
221+
222+
__m256 v_ac_bb = _mm256_sub_ps(_mm256_mul_ps(v_a, v_c), _mm256_mul_ps(v_b, v_b));
223+
__m256 v_ac = _mm256_add_ps(v_a, v_c);
224+
__m256 v_dst = _mm256_sub_ps(v_ac_bb, _mm256_mul_ps(v_k, _mm256_mul_ps(v_ac, v_ac)));
225+
_mm256_storeu_ps(dst + j, v_dst);
226+
}
227+
}
228+
#endif // CV_AVX
229+
230+
#if CV_SIMD128
231+
if( haveSimd )
119232
{
120233
v_float32x4 v_k = v_setall_f32((float)k);
121234

@@ -130,7 +243,7 @@ static void calcHarris( const Mat& _cov, Mat& _dst, double k )
130243
v_store(dst + j, v_dst);
131244
}
132245
}
133-
#endif
246+
#endif // CV_SIMD128
134247

135248
for( ; j < size.width; j++ )
136249
{
@@ -231,6 +344,9 @@ cornerEigenValsVecs( const Mat& src, Mat& eigenv, int block_size,
231344
if (tegra::useTegra() && tegra::cornerEigenValsVecs(src, eigenv, block_size, aperture_size, op_type, k, borderType))
232345
return;
233346
#endif
347+
#if CV_AVX
348+
bool haveAvx = checkHardwareSupport(CV_CPU_AVX);
349+
#endif
234350
#if CV_SIMD128
235351
bool haveSimd = hasSIMD128();
236352
#endif
@@ -268,8 +384,26 @@ cornerEigenValsVecs( const Mat& src, Mat& eigenv, int block_size,
268384
const float* dydata = Dy.ptr<float>(i);
269385
j = 0;
270386

271-
#if CV_SIMD128
272-
if (haveSimd)
387+
#if CV_AVX
388+
if( haveAvx )
389+
{
390+
for( ; j <= size.width - 8; j += 8 )
391+
{
392+
__m256 v_dx = _mm256_loadu_ps(dxdata + j);
393+
__m256 v_dy = _mm256_loadu_ps(dydata + j);
394+
395+
__m256 v_dst0, v_dst1, v_dst2;
396+
v_dst0 = _mm256_mul_ps(v_dx, v_dx);
397+
v_dst1 = _mm256_mul_ps(v_dx, v_dy);
398+
v_dst2 = _mm256_mul_ps(v_dy, v_dy);
399+
400+
store_interleave(cov_data + j * 3, v_dst0, v_dst1, v_dst2);
401+
}
402+
}
403+
#endif // CV_AVX
404+
405+
#if CV_SIMD128
406+
if( haveSimd )
273407
{
274408
for( ; j <= size.width - v_float32x4::nlanes; j += v_float32x4::nlanes )
275409
{
@@ -284,7 +418,7 @@ cornerEigenValsVecs( const Mat& src, Mat& eigenv, int block_size,
284418
v_store_interleave(cov_data + j * 3, v_dst0, v_dst1, v_dst2);
285419
}
286420
}
287-
#endif
421+
#endif // CV_SIMD128
288422

289423
for( ; j < size.width; j++ )
290424
{

0 commit comments

Comments
 (0)