Skip to content

O
opencv
Commit d513fb4c authored by Chip Kerchner's avatar Chip Kerchner Committed by Alexander Alekhin
Browse files
Options

Merge pull request #15199 from ChipKerchner:hogToHal 

Convert HOG from SSE SIMD to HAL - 35-45% faster on Power (VSX) (#15199)

* Convert SSE SIMD to HAL. 35-45% improvement for Power (VSX)

* Remove CV_NEON code. Use v_floor instead of 3 lines of code.

* Invert comparison logic to simplify code.

* Change initialization from v_load to constructor type.
parent 596b8a5d
Show whitespace changes
Inline Side-by-side
Showing with 132 additions and 377 deletions
modules/objdetect/src/hog.cpp
View file @ d513fb4c
...@@ -43,6 +43,7 @@ ...@@ -43,6 +43,7 @@
#include "precomp.hpp" #include "precomp.hpp"
#include "cascadedetect.hpp" #include "cascadedetect.hpp"
#include "opencv2/core/core_c.h" #include "opencv2/core/core_c.h"
#include "opencv2/core/hal/intrin.hpp"
#include "opencl_kernels_objdetect.hpp" #include "opencl_kernels_objdetect.hpp"
#include <cstdio> #include <cstdio>
...@@ -225,17 +226,6 @@ void HOGDescriptor::copyTo(HOGDescriptor& c) const ...@@ -225,17 +226,6 @@ void HOGDescriptor::copyTo(HOGDescriptor& c) const
c.signedGradient = signedGradient; c.signedGradient = signedGradient;
} }
#if CV_NEON
// replace of _mm_set_ps
inline float32x4_t vsetq_f32(float f0, float f1, float f2, float f3)
{
float32x4_t a = vdupq_n_f32(f0);
a = vsetq_lane_f32(f1, a, 1);
a = vsetq_lane_f32(f2, a, 2);
a = vsetq_lane_f32(f3, a, 3);
return a;
}
#endif
void HOGDescriptor::computeGradient(const Mat& img, Mat& grad, Mat& qangle, void HOGDescriptor::computeGradient(const Mat& img, Mat& grad, Mat& qangle,
Size paddingTL, Size paddingBR) const Size paddingTL, Size paddingBR) const
{ {
...@@ -257,38 +247,22 @@ void HOGDescriptor::computeGradient(const Mat& img, Mat& grad, Mat& qangle, ...@@ -257,38 +247,22 @@ void HOGDescriptor::computeGradient(const Mat& img, Mat& grad, Mat& qangle,
Mat_<float> _lut(1, 256); Mat_<float> _lut(1, 256);
const float* const lut = &_lut(0,0); const float* const lut = &_lut(0,0);
#if CV_SSE2 #if CV_SIMD128
const int indices[] = { 0, 1, 2, 3 }; v_float32x4 idx(0.0f, 1.0f, 2.0f, 3.0f);
__m128i idx = _mm_loadu_si128((const __m128i*)indices); v_float32x4 ifour = v_setall_f32(4.0);
__m128i ifour = _mm_set1_epi32(4);
float* const _data = &_lut(0, 0); float* const _data = &_lut(0, 0);
if( gammaCorrection ) if ( gammaCorrection )
for( i = 0; i < 256; i += 4 ) for ( i = 0; i < 256; i += 4)
{ {
_mm_storeu_ps(_data + i, _mm_sqrt_ps(_mm_cvtepi32_ps(idx))); v_store(_data + i, v_sqrt(idx));
idx = _mm_add_epi32(idx, ifour); idx += ifour;
} }
else else
for( i = 0; i < 256; i += 4 ) for ( i = 0; i < 256; i += 4)
{
_mm_storeu_ps(_data + i, _mm_cvtepi32_ps(idx));
idx = _mm_add_epi32(idx, ifour);
}
#elif CV_NEON
const int indices[] = { 0, 1, 2, 3 };
uint32x4_t idx = *(uint32x4_t*)indices;
uint32x4_t ifour = vdupq_n_u32(4);
float* const _data = &_lut(0, 0);
if( gammaCorrection )
for( i = 0; i < 256; i++ )
_lut(0,i) = std::sqrt((float)i);
else
for( i = 0; i < 256; i += 4 )
{ {
vst1q_f32(_data + i, vcvtq_f32_u32(idx)); v_store(_data + i, idx);
idx = vaddq_u32 (idx, ifour); idx += ifour;
} }
#else #else
if( gammaCorrection ) if( gammaCorrection )
...@@ -325,17 +299,13 @@ void HOGDescriptor::computeGradient(const Mat& img, Mat& grad, Mat& qangle, ...@@ -325,17 +299,13 @@ void HOGDescriptor::computeGradient(const Mat& img, Mat& grad, Mat& qangle,
{ {
int end = gradsize.width + 2; int end = gradsize.width + 2;
xmap -= 1, x = 0; xmap -= 1, x = 0;
#if CV_SSE2 #if CV_SIMD128
for ( ; x <= end - 4; x += 4) for ( ; x <= end - 4; x += 4)
{ {
__m128i mul_res = _mm_loadu_si128((const __m128i*)(xmap + x)); v_int32x4 mul_res = v_load(xmap + x);
mul_res = _mm_add_epi32(_mm_add_epi32(mul_res, mul_res), mul_res); // multiply by 3 mul_res += mul_res + mul_res;
_mm_storeu_si128((__m128i*)(xmap + x), mul_res); v_store(xmap + x, mul_res);
} }
#elif CV_NEON
int32x4_t ithree = vdupq_n_s32(3);
for ( ; x <= end - 4; x += 4)
vst1q_s32(xmap + x, vmulq_s32(ithree, vld1q_s32(xmap + x)));
#endif #endif
for ( ; x < end; ++x) for ( ; x < end; ++x)
xmap[x] *= 3; xmap[x] *= 3;
...@@ -366,7 +336,7 @@ void HOGDescriptor::computeGradient(const Mat& img, Mat& grad, Mat& qangle, ...@@ -366,7 +336,7 @@ void HOGDescriptor::computeGradient(const Mat& img, Mat& grad, Mat& qangle,
else else
{ {
x = 0; x = 0;
#if CV_SSE2 #if CV_SIMD128
for( ; x <= width - 4; x += 4 ) for( ; x <= width - 4; x += 4 )
{ {
int x0 = xmap[x], x1 = xmap[x+1], x2 = xmap[x+2], x3 = xmap[x+3]; int x0 = xmap[x], x1 = xmap[x+1], x2 = xmap[x+2], x3 = xmap[x+3];
...@@ -376,73 +346,34 @@ void HOGDescriptor::computeGradient(const Mat& img, Mat& grad, Mat& qangle, ...@@ -376,73 +346,34 @@ void HOGDescriptor::computeGradient(const Mat& img, Mat& grad, Mat& qangle,
T p22 = imgPtr + xmap[x+3], p20 = p02; T p22 = imgPtr + xmap[x+3], p20 = p02;
T p32 = imgPtr + xmap[x+4], p30 = p12; T p32 = imgPtr + xmap[x+4], p30 = p12;
__m128 _dx0 = _mm_sub_ps(_mm_set_ps(lut[p32[0]], lut[p22[0]], lut[p12[0]], lut[p02[0]]), v_float32x4 _dx0 = v_float32x4(lut[p02[0]], lut[p12[0]], lut[p22[0]], lut[p32[0]]) -
_mm_set_ps(lut[p30[0]], lut[p20[0]], lut[p10[0]], lut[p00[0]])); v_float32x4(lut[p00[0]], lut[p10[0]], lut[p20[0]], lut[p30[0]]);
__m128 _dx1 = _mm_sub_ps(_mm_set_ps(lut[p32[1]], lut[p22[1]], lut[p12[1]], lut[p02[1]]), v_float32x4 _dx1 = v_float32x4(lut[p02[1]], lut[p12[1]], lut[p22[1]], lut[p32[1]]) -
_mm_set_ps(lut[p30[1]], lut[p20[1]], lut[p10[1]], lut[p00[1]])); v_float32x4(lut[p00[1]], lut[p10[1]], lut[p20[1]], lut[p30[1]]);
__m128 _dx2 = _mm_sub_ps(_mm_set_ps(lut[p32[2]], lut[p22[2]], lut[p12[2]], lut[p02[2]]), v_float32x4 _dx2 = v_float32x4(lut[p02[2]], lut[p12[2]], lut[p22[2]], lut[p32[2]]) -
_mm_set_ps(lut[p30[2]], lut[p20[2]], lut[p10[2]], lut[p00[2]])); v_float32x4(lut[p00[2]], lut[p10[2]], lut[p20[2]], lut[p30[2]]);
__m128 _dy0 = _mm_sub_ps(_mm_set_ps(lut[nextPtr[x3]], lut[nextPtr[x2]], lut[nextPtr[x1]], lut[nextPtr[x0]]), v_float32x4 _dy0 = v_float32x4(lut[nextPtr[x0]], lut[nextPtr[x1]], lut[nextPtr[x2]], lut[nextPtr[x3]]) -
_mm_set_ps(lut[prevPtr[x3]], lut[prevPtr[x2]], lut[prevPtr[x1]], lut[prevPtr[x0]])); v_float32x4(lut[prevPtr[x0]], lut[prevPtr[x1]], lut[prevPtr[x2]], lut[prevPtr[x3]]);
__m128 _dy1 = _mm_sub_ps(_mm_set_ps(lut[nextPtr[x3+1]], lut[nextPtr[x2+1]], lut[nextPtr[x1+1]], lut[nextPtr[x0+1]]), v_float32x4 _dy1 = v_float32x4(lut[nextPtr[x0+1]], lut[nextPtr[x1+1]], lut[nextPtr[x2+1]], lut[nextPtr[x3+1]]) -
_mm_set_ps(lut[prevPtr[x3+1]], lut[prevPtr[x2+1]], lut[prevPtr[x1+1]], lut[prevPtr[x0+1]])); v_float32x4(lut[prevPtr[x0+1]], lut[prevPtr[x1+1]], lut[prevPtr[x2+1]], lut[prevPtr[x3+1]]);
__m128 _dy2 = _mm_sub_ps(_mm_set_ps(lut[nextPtr[x3+2]], lut[nextPtr[x2+2]], lut[nextPtr[x1+2]], lut[nextPtr[x0+2]]), v_float32x4 _dy2 = v_float32x4(lut[nextPtr[x0+2]], lut[nextPtr[x1+2]], lut[nextPtr[x2+2]], lut[nextPtr[x3+2]]) -
_mm_set_ps(lut[prevPtr[x3+2]], lut[prevPtr[x2+2]], lut[prevPtr[x1+2]], lut[prevPtr[x0+2]])); v_float32x4(lut[prevPtr[x0+2]], lut[prevPtr[x1+2]], lut[prevPtr[x2+2]], lut[prevPtr[x3+2]]);
__m128 _mag0 = _mm_add_ps(_mm_mul_ps(_dx0, _dx0), _mm_mul_ps(_dy0, _dy0)); v_float32x4 _mag0 = (_dx0 * _dx0) + (_dy0 * _dy0);
__m128 _mag1 = _mm_add_ps(_mm_mul_ps(_dx1, _dx1), _mm_mul_ps(_dy1, _dy1)); v_float32x4 _mag1 = (_dx1 * _dx1) + (_dy1 * _dy1);
__m128 _mag2 = _mm_add_ps(_mm_mul_ps(_dx2, _dx2), _mm_mul_ps(_dy2, _dy2)); v_float32x4 _mag2 = (_dx2 * _dx2) + (_dy2 * _dy2);
__m128 mask = _mm_cmpgt_ps(_mag2, _mag1); v_float32x4 mask = v_reinterpret_as_f32(_mag2 > _mag1);
_dx2 = _mm_or_ps(_mm_and_ps(_dx2, mask), _mm_andnot_ps(mask, _dx1)); _dx2 = v_select(mask, _dx2, _dx1);
_dy2 = _mm_or_ps(_mm_and_ps(_dy2, mask), _mm_andnot_ps(mask, _dy1)); _dy2 = v_select(mask, _dy2, _dy1);
mask = _mm_cmpgt_ps(_mm_max_ps(_mag2, _mag1), _mag0); mask = v_reinterpret_as_f32(v_max(_mag2, _mag1) > _mag0);
_dx2 = _mm_or_ps(_mm_and_ps(_dx2, mask), _mm_andnot_ps(mask, _dx0)); _dx2 = v_select(mask, _dx2, _dx0);
_dy2 = _mm_or_ps(_mm_and_ps(_dy2, mask), _mm_andnot_ps(mask, _dy0)); _dy2 = v_select(mask, _dy2, _dy0);
_mm_storeu_ps(dbuf + x, _dx2); v_store(dbuf + x, _dx2);
_mm_storeu_ps(dbuf + x + width, _dy2); v_store(dbuf + x + width, _dy2);
}
#elif CV_NEON
for( ; x <= width - 4; x += 4 )
{
int x0 = xmap[x], x1 = xmap[x+1], x2 = xmap[x+2], x3 = xmap[x+3];
typedef const uchar* const T;
T p02 = imgPtr + xmap[x+1], p00 = imgPtr + xmap[x-1];
T p12 = imgPtr + xmap[x+2], p10 = imgPtr + xmap[x];
T p22 = imgPtr + xmap[x+3], p20 = p02;
T p32 = imgPtr + xmap[x+4], p30 = p12;
float32x4_t _dx0 = vsubq_f32(vsetq_f32(lut[p02[0]], lut[p12[0]], lut[p22[0]], lut[p32[0]]),
vsetq_f32(lut[p00[0]], lut[p10[0]], lut[p20[0]], lut[p30[0]]));
float32x4_t _dx1 = vsubq_f32(vsetq_f32(lut[p02[1]], lut[p12[1]], lut[p22[1]], lut[p32[1]]),
vsetq_f32(lut[p00[1]], lut[p10[1]], lut[p20[1]], lut[p30[1]]));
float32x4_t _dx2 = vsubq_f32(vsetq_f32(lut[p02[2]], lut[p12[2]], lut[p22[2]], lut[p32[2]]),
vsetq_f32(lut[p00[2]], lut[p10[2]], lut[p20[2]], lut[p30[2]]));
float32x4_t _dy0 = vsubq_f32(vsetq_f32(lut[nextPtr[x0]], lut[nextPtr[x1]], lut[nextPtr[x2]], lut[nextPtr[x3]]),
vsetq_f32(lut[prevPtr[x0]], lut[prevPtr[x1]], lut[prevPtr[x2]], lut[prevPtr[x3]]));
float32x4_t _dy1 = vsubq_f32(vsetq_f32(lut[nextPtr[x0+1]], lut[nextPtr[x1+1]], lut[nextPtr[x2+1]], lut[nextPtr[x3+1]]),
vsetq_f32(lut[prevPtr[x0+1]], lut[prevPtr[x1+1]], lut[prevPtr[x2+1]], lut[prevPtr[x3+1]]));
float32x4_t _dy2 = vsubq_f32(vsetq_f32(lut[nextPtr[x0+2]], lut[nextPtr[x1+2]], lut[nextPtr[x2+2]], lut[nextPtr[x3+2]]),
vsetq_f32(lut[prevPtr[x0+2]], lut[prevPtr[x1+2]], lut[prevPtr[x2+2]], lut[prevPtr[x3+2]]));
float32x4_t _mag0 = vaddq_f32(vmulq_f32(_dx0, _dx0), vmulq_f32(_dy0, _dy0));
float32x4_t _mag1 = vaddq_f32(vmulq_f32(_dx1, _dx1), vmulq_f32(_dy1, _dy1));
float32x4_t _mag2 = vaddq_f32(vmulq_f32(_dx2, _dx2), vmulq_f32(_dy2, _dy2));
uint32x4_t mask = vcgtq_f32(_mag2, _mag1);
_dx2 = vbslq_f32(mask, _dx2, _dx1);
_dy2 = vbslq_f32(mask, _dy2, _dy1);
mask = vcgtq_f32(vmaxq_f32(_mag2, _mag1), _mag0);
_dx2 = vbslq_f32(mask, _dx2, _dx0);
_dy2 = vbslq_f32(mask, _dy2, _dy0);
vst1q_f32(dbuf + x, _dx2);
vst1q_f32(dbuf + x + width, _dy2);
} }
#endif #endif
for( ; x < width; x++ ) for( ; x < width; x++ )
...@@ -486,44 +417,40 @@ void HOGDescriptor::computeGradient(const Mat& img, Mat& grad, Mat& qangle, ...@@ -486,44 +417,40 @@ void HOGDescriptor::computeGradient(const Mat& img, Mat& grad, Mat& qangle,
// filling the result matrix // filling the result matrix
x = 0; x = 0;
#if CV_SSE2 #if CV_SIMD128
__m128 fhalf = _mm_set1_ps(0.5f), fzero = _mm_setzero_ps(); v_float32x4 fhalf = v_setall_f32(0.5f);
__m128 _angleScale = _mm_set1_ps(angleScale), fone = _mm_set1_ps(1.0f); v_float32x4 _angleScale = v_setall_f32(angleScale), fone = v_setall_f32(1.0f);
__m128i ione = _mm_set1_epi32(1), _nbins = _mm_set1_epi32(nbins), izero = _mm_setzero_si128(); v_int32x4 ione = v_setall_s32(1), _nbins = v_setall_s32(nbins), izero = v_setzero_s32();
for ( ; x <= width - 4; x += 4) for ( ; x <= width - 4; x += 4)
{ {
int x2 = x << 1; int x2 = x << 1;
__m128 _mag = _mm_loadu_ps(dbuf + x + (width << 1)); v_float32x4 _mag = v_load(dbuf + x + (width << 1));
__m128 _angle = _mm_loadu_ps(dbuf + x + width * 3); v_float32x4 _angle = v_load(dbuf + x + width * 3);
_angle = _mm_sub_ps(_mm_mul_ps(_angleScale, _angle), fhalf); _angle = (_angleScale * _angle) - fhalf;
__m128 sign = _mm_and_ps(fone, _mm_cmplt_ps(_angle, fzero)); v_int32x4 _hidx = v_floor(_angle);
__m128i _hidx = _mm_cvttps_epi32(_angle); _angle -= v_cvt_f32(_hidx);
_hidx = _mm_sub_epi32(_hidx, _mm_cvtps_epi32(sign));
_angle = _mm_sub_ps(_angle, _mm_cvtepi32_ps(_hidx));
__m128 ft0 = _mm_mul_ps(_mag, _mm_sub_ps(fone, _angle)); v_float32x4 ft0 = _mag * (fone - _angle);
__m128 ft1 = _mm_mul_ps(_mag, _angle); v_float32x4 ft1 = _mag * _angle;
__m128 ft2 = _mm_unpacklo_ps(ft0, ft1);
__m128 ft3 = _mm_unpackhi_ps(ft0, ft1);
_mm_storeu_ps(gradPtr + x2, ft2); v_store_interleave(gradPtr + x2, ft0, ft1);
_mm_storeu_ps(gradPtr + x2 + 4, ft3);
__m128i mask0 = _mm_sub_epi32(izero, _mm_srli_epi32(_hidx, 31)); v_int32x4 mask0 = _hidx >> 31;
__m128i it0 = _mm_and_si128(mask0, _nbins); v_int32x4 it0 = mask0 & _nbins;
mask0 = _mm_cmplt_epi32(_hidx, _nbins); mask0 = (_hidx >= _nbins);
__m128i it1 = _mm_andnot_si128(mask0, _nbins); v_int32x4 it1 = mask0 & _nbins;
_hidx = _mm_add_epi32(_hidx, _mm_sub_epi32(it0, it1)); _hidx += (it0 - it1);
it0 = _mm_packus_epi16(_mm_packs_epi32(_hidx, izero), izero); it0 = v_reinterpret_as_s32(v_pack(v_pack(_hidx, izero), v_reinterpret_as_s16(izero)));
_hidx = _mm_add_epi32(ione, _hidx); _hidx += ione;
_hidx = _mm_and_si128(_hidx, _mm_cmplt_epi32(_hidx, _nbins)); _hidx &= (_hidx < _nbins);
it1 = _mm_packus_epi16(_mm_packs_epi32(_hidx, izero), izero); it1 = v_reinterpret_as_s32(v_pack(v_pack(_hidx, izero), v_reinterpret_as_s16(izero)));
it0 = _mm_unpacklo_epi8(it0, it1); v_uint8x16 it2, it3;
v_zip(v_reinterpret_as_u8(it0), v_reinterpret_as_u8(it1), it2, it3);
_mm_storel_epi64((__m128i*)(qanglePtr + x2), it0); v_store_low(qanglePtr + x2, it2);
} }
#endif #endif
for( ; x < width; x++ ) for( ; x < width; x++ )
...@@ -663,31 +590,17 @@ void HOGCache::init(const HOGDescriptor* _descriptor, ...@@ -663,31 +590,17 @@ void HOGCache::init(const HOGDescriptor* _descriptor,
float bh = blockSize.height * 0.5f, bw = blockSize.width * 0.5f; float bh = blockSize.height * 0.5f, bw = blockSize.width * 0.5f;
i = 0; i = 0;
#if CV_SSE2 #if CV_SIMD128
const int a[] = { 0, 1, 2, 3 }; v_float32x4 idx(0.0f, 1.0f, 2.0f, 3.0f);
__m128i idx = _mm_loadu_si128((__m128i*)a); v_float32x4 _bw = v_setall_f32(bw), _bh = v_setall_f32(bh);
__m128 _bw = _mm_set1_ps(bw), _bh = _mm_set1_ps(bh); v_float32x4 ifour = v_setall_f32(4.0);
__m128i ifour = _mm_set1_epi32(4);
for (; i <= blockSize.height - 4; i += 4)
{
__m128 t = _mm_sub_ps(_mm_cvtepi32_ps(idx), _bh);
t = _mm_mul_ps(t, t);
idx = _mm_add_epi32(idx, ifour);
_mm_storeu_ps(_di + i, t);
}
#elif CV_NEON
const int a[] = { 0, 1, 2, 3 };
int32x4_t idx = vld1q_s32(a);
float32x4_t _bw = vdupq_n_f32(bw), _bh = vdupq_n_f32(bh);
int32x4_t ifour = vdupq_n_s32(4);
for (; i <= blockSize.height - 4; i += 4) for (; i <= blockSize.height - 4; i += 4)
{ {
float32x4_t t = vsubq_f32(vcvtq_f32_s32(idx), _bh); v_float32x4 t = idx - _bh;
t = vmulq_f32(t, t); t *= t;
idx = vaddq_s32(idx, ifour); idx += ifour;
vst1q_f32(_di + i, t); v_store(_di + i, t);
} }
#endif #endif
for ( ; i < blockSize.height; ++i) for ( ; i < blockSize.height; ++i)
...@@ -697,23 +610,15 @@ void HOGCache::init(const HOGDescriptor* _descriptor, ...@@ -697,23 +610,15 @@ void HOGCache::init(const HOGDescriptor* _descriptor,
} }
j = 0; j = 0;
#if CV_SSE2 #if CV_SIMD128
idx = _mm_loadu_si128((__m128i*)a); idx = v_float32x4(0.0f, 1.0f, 2.0f, 3.0f);
for (; j <= blockSize.width - 4; j += 4)
{ for (; j <= blockSize.height - 4; j += 4)
__m128 t = _mm_sub_ps(_mm_cvtepi32_ps(idx), _bw);
t = _mm_mul_ps(t, t);
idx = _mm_add_epi32(idx, ifour);
_mm_storeu_ps(_dj + j, t);
}
#elif CV_NEON
idx = vld1q_s32(a);
for (; j <= blockSize.width - 4; j += 4)
{ {
float32x4_t t = vsubq_f32(vcvtq_f32_s32(idx), _bw); v_float32x4 t = idx - _bw;
t = vmulq_f32(t, t); t *= t;
idx = vaddq_s32(idx, ifour); idx += ifour;
vst1q_f32(_dj + j, t); v_store(_dj + j, t);
} }
#endif #endif
for ( ; j < blockSize.width; ++j) for ( ; j < blockSize.width; ++j)
...@@ -911,7 +816,7 @@ const float* HOGCache::getBlock(Point pt, float* buf) ...@@ -911,7 +816,7 @@ const float* HOGCache::getBlock(Point pt, float* buf)
hist[h0] = t0; hist[h1] = t1; hist[h0] = t0; hist[h1] = t1;
} }
#if CV_SSE2 #if CV_SIMD128
float hist0[4], hist1[4]; float hist0[4], hist1[4];
for( ; k < C2; k++ ) for( ; k < C2; k++ )
{ {
...@@ -920,12 +825,12 @@ const float* HOGCache::getBlock(Point pt, float* buf) ...@@ -920,12 +825,12 @@ const float* HOGCache::getBlock(Point pt, float* buf)
const uchar* const h = qanglePtr + pk.qangleOfs; const uchar* const h = qanglePtr + pk.qangleOfs;
int h0 = h[0], h1 = h[1]; int h0 = h[0], h1 = h[1];
__m128 _a0 = _mm_set1_ps(a[0]), _a1 = _mm_set1_ps(a[1]); v_float32x4 _a0 = v_setall_f32(a[0]), _a1 = v_setall_f32(a[1]);
__m128 _w = _mm_mul_ps(_mm_set1_ps(pk.gradWeight), _mm_loadu_ps(pk.histWeights)); v_float32x4 w = v_setall_f32(pk.gradWeight) * v_load(pk.histWeights);
__m128 _t0 = _mm_mul_ps(_a0, _w), _t1 = _mm_mul_ps(_a1, _w); v_float32x4 _t0 = _a0 * w, _t1 = _a1 * w;
_mm_storeu_ps(hist0, _t0); v_store(hist0, _t0);
_mm_storeu_ps(hist1, _t1); v_store(hist1, _t1);
float* hist = blockHist + pk.histOfs[0]; float* hist = blockHist + pk.histOfs[0];
float t0 = hist[h0] + hist0[0]; float t0 = hist[h0] + hist0[0];
...@@ -937,31 +842,6 @@ const float* HOGCache::getBlock(Point pt, float* buf) ...@@ -937,31 +842,6 @@ const float* HOGCache::getBlock(Point pt, float* buf)
t1 = hist[h1] + hist1[1]; t1 = hist[h1] + hist1[1];
hist[h0] = t0; hist[h1] = t1; hist[h0] = t0; hist[h1] = t1;
} }
#elif CV_NEON
float hist0[4], hist1[4];
for( ; k < C2; k++ )
{
const PixData& pk = _pixData[k];
const float* const a = gradPtr + pk.gradOfs;
const uchar* const h = qanglePtr + pk.qangleOfs;
int h0 = h[0], h1 = h[1];
float32x4_t _a0 = vdupq_n_f32(a[0]), _a1 = vdupq_n_f32(a[1]);
float32x4_t _w = vmulq_f32(vdupq_n_f32(pk.gradWeight), vld1q_f32(pk.histWeights));
float32x4_t _h0 = vsetq_f32((blockHist + pk.histOfs[0])[h0], (blockHist + pk.histOfs[1])[h0], 0, 0);
float32x4_t _h1 = vsetq_f32((blockHist + pk.histOfs[0])[h1], (blockHist + pk.histOfs[1])[h1], 0, 0);
float32x4_t _t0 = vmlaq_f32(_h0, _a0, _w), _t1 = vmlaq_f32(_h1, _a1, _w);
vst1q_f32(hist0, _t0);
vst1q_f32(hist1, _t1);
(blockHist + pk.histOfs[0])[h0] = hist0[0];
(blockHist + pk.histOfs[1])[h0] = hist0[1];
(blockHist + pk.histOfs[0])[h1] = hist1[0];
(blockHist + pk.histOfs[1])[h1] = hist1[1];
}
#else #else
for( ; k < C2; k++ ) for( ; k < C2; k++ )
{ {
...@@ -985,7 +865,7 @@ const float* HOGCache::getBlock(Point pt, float* buf) ...@@ -985,7 +865,7 @@ const float* HOGCache::getBlock(Point pt, float* buf)
} }
#endif #endif
#if CV_SSE2 #if CV_SIMD128
for( ; k < C4; k++ ) for( ; k < C4; k++ )
{ {
const PixData& pk = _pixData[k]; const PixData& pk = _pixData[k];
...@@ -993,12 +873,12 @@ const float* HOGCache::getBlock(Point pt, float* buf) ...@@ -993,12 +873,12 @@ const float* HOGCache::getBlock(Point pt, float* buf)
const uchar* const h = qanglePtr + pk.qangleOfs; const uchar* const h = qanglePtr + pk.qangleOfs;
int h0 = h[0], h1 = h[1]; int h0 = h[0], h1 = h[1];
__m128 _a0 = _mm_set1_ps(a[0]), _a1 = _mm_set1_ps(a[1]); v_float32x4 _a0 = v_setall_f32(a[0]), _a1 = v_setall_f32(a[1]);
__m128 _w = _mm_mul_ps(_mm_set1_ps(pk.gradWeight), _mm_loadu_ps(pk.histWeights)); v_float32x4 w = v_setall_f32(pk.gradWeight) * v_load(pk.histWeights);
__m128 _t0 = _mm_mul_ps(_a0, _w), _t1 = _mm_mul_ps(_a1, _w); v_float32x4 _t0 = _a0 * w, _t1 = _a1 * w;
_mm_storeu_ps(hist0, _t0); v_store(hist0, _t0);
_mm_storeu_ps(hist1, _t1); v_store(hist1, _t1);
float* hist = blockHist + pk.histOfs[0]; float* hist = blockHist + pk.histOfs[0];
float t0 = hist[h0] + hist0[0]; float t0 = hist[h0] + hist0[0];
...@@ -1019,62 +899,6 @@ const float* HOGCache::getBlock(Point pt, float* buf) ...@@ -1019,62 +899,6 @@ const float* HOGCache::getBlock(Point pt, float* buf)
t0 = hist[h0] + hist0[3]; t0 = hist[h0] + hist0[3];
t1 = hist[h1] + hist1[3]; t1 = hist[h1] + hist1[3];
hist[h0] = t0; hist[h1] = t1; hist[h0] = t0; hist[h1] = t1;
// __m128 _hist0 = _mm_set_ps((blockHist + pk.histOfs[3])[h0], (blockHist + pk.histOfs[2])[h0],
// (blockHist + pk.histOfs[1])[h0], (blockHist + pk.histOfs[0])[h0]);
// __m128 _hist1 = _mm_set_ps((blockHist + pk.histOfs[3])[h1], (blockHist + pk.histOfs[2])[h1],
// (blockHist + pk.histOfs[1])[h1], (blockHist + pk.histOfs[0])[h1]);
//
// _hist0 = _mm_add_ps(_t0, _hist0);
// _hist1 = _mm_add_ps(_t1, _hist1);
//
// _mm_storeu_ps(hist0, _hist0);
// _mm_storeu_ps(hist1, _hist1);
//
// (pk.histOfs[0] + blockHist)[h0] = hist0[0];
// (pk.histOfs[1] + blockHist)[h0] = hist0[1];
// (pk.histOfs[2] + blockHist)[h0] = hist0[2];
// (pk.histOfs[3] + blockHist)[h0] = hist0[3];
//
// (pk.histOfs[0] + blockHist)[h1] = hist1[0];
// (pk.histOfs[1] + blockHist)[h1] = hist1[1];
// (pk.histOfs[2] + blockHist)[h1] = hist1[2];
// (pk.histOfs[3] + blockHist)[h1] = hist1[3];
}
#elif CV_NEON
for( ; k < C4; k++ )
{
const PixData& pk = _pixData[k];
const float* const a = gradPtr + pk.gradOfs;
const uchar* const h = qanglePtr + pk.qangleOfs;
int h0 = h[0], h1 = h[1];
float32x4_t _a0 = vdupq_n_f32(a[0]), _a1 = vdupq_n_f32(a[1]);
float32x4_t _w = vmulq_f32(vdupq_n_f32(pk.gradWeight), vld1q_f32(pk.histWeights));
float32x4_t _h0 = vsetq_f32((blockHist + pk.histOfs[0])[h0],
(blockHist + pk.histOfs[1])[h0],
(blockHist + pk.histOfs[2])[h0],
(blockHist + pk.histOfs[3])[h0]);
float32x4_t _h1 = vsetq_f32((blockHist + pk.histOfs[0])[h1],
(blockHist + pk.histOfs[1])[h1],
(blockHist + pk.histOfs[2])[h1],
(blockHist + pk.histOfs[3])[h1]);
float32x4_t _t0 = vmlaq_f32(_h0, _a0, _w), _t1 = vmlaq_f32(_h1, _a1, _w);
vst1q_f32(hist0, _t0);
vst1q_f32(hist1, _t1);
(blockHist + pk.histOfs[0])[h0] = hist0[0];
(blockHist + pk.histOfs[1])[h0] = hist0[1];
(blockHist + pk.histOfs[2])[h0] = hist0[2];
(blockHist + pk.histOfs[3])[h0] = hist0[3];
(blockHist + pk.histOfs[0])[h1] = hist1[0];
(blockHist + pk.histOfs[1])[h1] = hist1[1];
(blockHist + pk.histOfs[2])[h1] = hist1[2];
(blockHist + pk.histOfs[3])[h1] = hist1[3];
} }
#else #else
for( ; k < C4; k++ ) for( ; k < C4; k++ )
...@@ -1121,26 +945,16 @@ void HOGCache::normalizeBlockHistogram(float* _hist) const ...@@ -1121,26 +945,16 @@ void HOGCache::normalizeBlockHistogram(float* _hist) const
float* hist = &_hist[0], sum = 0.0f, partSum[4]; float* hist = &_hist[0], sum = 0.0f, partSum[4];
size_t i = 0, sz = blockHistogramSize; size_t i = 0, sz = blockHistogramSize;
#if CV_SSE2 #if CV_SIMD128
__m128 p0 = _mm_loadu_ps(hist); v_float32x4 p0 = v_load(hist);
__m128 s = _mm_mul_ps(p0, p0); v_float32x4 s = p0 * p0;
for (i = 4; i <= sz - 4; i += 4) for (i = 4; i <= sz - 4; i += 4)
{ {
p0 = _mm_loadu_ps(hist + i); p0 = v_load(hist + i);
s = _mm_add_ps(s, _mm_mul_ps(p0, p0)); s += p0 * p0;
} }
_mm_storeu_ps(partSum, s); v_store(partSum, s);
#elif CV_NEON
float32x4_t p0 = vld1q_f32(hist);
float32x4_t s = vmulq_f32(p0, p0);
for (i = 4; i <= sz - 4; i += 4)
{
p0 = vld1q_f32(hist + i);
s = vaddq_f32(s, vmulq_f32(p0, p0));
}
vst1q_f32(partSum, s);
#else #else
partSum[0] = 0.0f; partSum[0] = 0.0f;
partSum[1] = 0.0f; partSum[1] = 0.0f;
...@@ -1163,44 +977,25 @@ void HOGCache::normalizeBlockHistogram(float* _hist) const ...@@ -1163,44 +977,25 @@ void HOGCache::normalizeBlockHistogram(float* _hist) const
float scale = 1.f/(std::sqrt(sum)+sz*0.1f), thresh = (float)descriptor->L2HysThreshold; float scale = 1.f/(std::sqrt(sum)+sz*0.1f), thresh = (float)descriptor->L2HysThreshold;
i = 0, sum = 0.0f; i = 0, sum = 0.0f;
#if CV_SSE2 #if CV_SIMD128
__m128 _scale = _mm_set1_ps(scale); v_float32x4 _scale = v_setall_f32(scale);
static __m128 _threshold = _mm_set1_ps(thresh); static v_float32x4 _threshold = v_setall_f32(thresh);
__m128 p = _mm_mul_ps(_scale, _mm_loadu_ps(hist));
p = _mm_min_ps(p, _threshold);
s = _mm_mul_ps(p, p);
_mm_storeu_ps(hist, p);
for(i = 4 ; i <= sz - 4; i += 4)
{
p = _mm_loadu_ps(hist + i);
p = _mm_mul_ps(p, _scale);
p = _mm_min_ps(p, _threshold);
s = _mm_add_ps(s, _mm_mul_ps(p, p));
_mm_storeu_ps(hist + i, p);
}
_mm_storeu_ps(partSum, s); v_float32x4 p = _scale * v_load(hist);
#elif CV_NEON p = v_min(p, _threshold);
float32x4_t _scale = vdupq_n_f32(scale); s = p * p;
static float32x4_t _threshold = vdupq_n_f32(thresh); v_store(hist, p);
float32x4_t p = vmulq_f32(_scale, vld1q_f32(hist));
p = vminq_f32(p, _threshold);
s = vmulq_f32(p, p);
vst1q_f32(hist, p);
for(i = 4 ; i <= sz - 4; i += 4) for(i = 4 ; i <= sz - 4; i += 4)
{ {
p = vld1q_f32(hist + i); p = v_load(hist + i);
p = vmulq_f32(p, _scale); p *= _scale;
p = vminq_f32(p, _threshold); p = v_min(p, _threshold);
s = vaddq_f32(s, vmulq_f32(p, p)); s += p * p;
vst1q_f32(hist + i, p); v_store(hist + i, p);
} }
vst1q_f32(partSum, s); v_store(partSum, s);
#else #else
partSum[0] = 0.0f; partSum[0] = 0.0f;
partSum[1] = 0.0f; partSum[1] = 0.0f;
...@@ -1228,19 +1023,12 @@ void HOGCache::normalizeBlockHistogram(float* _hist) const ...@@ -1228,19 +1023,12 @@ void HOGCache::normalizeBlockHistogram(float* _hist) const
} }
scale = 1.f/(std::sqrt(sum)+1e-3f), i = 0; scale = 1.f/(std::sqrt(sum)+1e-3f), i = 0;
#if CV_SSE2 #if CV_SIMD128
__m128 _scale2 = _mm_set1_ps(scale); v_float32x4 _scale2 = v_setall_f32(scale);
for ( ; i <= sz - 4; i += 4)
{
__m128 t = _mm_mul_ps(_scale2, _mm_loadu_ps(hist + i));
_mm_storeu_ps(hist + i, t);
}
#elif CV_NEON
float32x4_t _scale2 = vdupq_n_f32(scale);
for ( ; i <= sz - 4; i += 4) for ( ; i <= sz - 4; i += 4)
{ {
float32x4_t t = vmulq_f32(_scale2, vld1q_f32(hist + i)); v_float32x4 t = _scale2 * v_load(hist + i);
vst1q_f32(hist + i, t); v_store(hist + i, t);
} }
#endif #endif
for ( ; i < sz; ++i) for ( ; i < sz; ++i)
...@@ -1687,7 +1475,7 @@ void HOGDescriptor::detect(const Mat& img, ...@@ -1687,7 +1475,7 @@ void HOGDescriptor::detect(const Mat& img,
double rho = svmDetector.size() > dsize ? svmDetector[dsize] : 0; double rho = svmDetector.size() > dsize ? svmDetector[dsize] : 0;
std::vector<float> blockHist(blockHistogramSize); std::vector<float> blockHist(blockHistogramSize);
#if CV_SSE2 || CV_NEON #if CV_SIMD128
float partSum[4]; float partSum[4];
#endif #endif
...@@ -1716,37 +1504,20 @@ void HOGDescriptor::detect(const Mat& img, ...@@ -1716,37 +1504,20 @@ void HOGDescriptor::detect(const Mat& img,
Point pt = pt0 + bj.imgOffset; Point pt = pt0 + bj.imgOffset;
const float* vec = cache.getBlock(pt, &blockHist[0]); const float* vec = cache.getBlock(pt, &blockHist[0]);
#if CV_SSE2 #if CV_SIMD128
__m128 _vec = _mm_loadu_ps(vec); v_float32x4 _vec = v_load(vec);
__m128 _svmVec = _mm_loadu_ps(svmVec); v_float32x4 _svmVec = v_load(svmVec);
__m128 sum = _mm_mul_ps(_svmVec, _vec); v_float32x4 sum = _svmVec * _vec;
for( k = 4; k <= blockHistogramSize - 4; k += 4 )
{
_vec = _mm_loadu_ps(vec + k);
_svmVec = _mm_loadu_ps(svmVec + k);
sum = _mm_add_ps(sum, _mm_mul_ps(_vec, _svmVec));
}
_mm_storeu_ps(partSum, sum);
double t0 = partSum[0] + partSum[1];
double t1 = partSum[2] + partSum[3];
s += t0 + t1;
#elif CV_NEON
float32x4_t _vec = vld1q_f32(vec);
float32x4_t _svmVec = vld1q_f32(svmVec);
float32x4_t sum = vmulq_f32(_svmVec, _vec);
for( k = 4; k <= blockHistogramSize - 4; k += 4 ) for( k = 4; k <= blockHistogramSize - 4; k += 4 )
{ {
_vec = vld1q_f32(vec + k); _vec = v_load(vec + k);
_svmVec = vld1q_f32(svmVec + k); _svmVec = v_load(svmVec + k);
sum = vaddq_f32(sum, vmulq_f32(_vec, _svmVec)); sum += _vec * _svmVec;
} }
vst1q_f32(partSum, sum); v_store(partSum, sum);
double t0 = partSum[0] + partSum[1]; double t0 = partSum[0] + partSum[1];
double t1 = partSum[2] + partSum[3]; double t1 = partSum[2] + partSum[3];
s += t0 + t1; s += t0 + t1;
...@@ -3582,7 +3353,7 @@ void HOGDescriptor::detectROI(const cv::Mat& img, const std::vector<cv::Point> & ...@@ -3582,7 +3353,7 @@ void HOGDescriptor::detectROI(const cv::Mat& img, const std::vector<cv::Point> &
double rho = svmDetector.size() > dsize ? svmDetector[dsize] : 0; double rho = svmDetector.size() > dsize ? svmDetector[dsize] : 0;
std::vector<float> blockHist(blockHistogramSize); std::vector<float> blockHist(blockHistogramSize);
#if CV_SSE2 || CV_NEON #if CV_SIMD128
float partSum[4]; float partSum[4];
#endif #endif
...@@ -3609,37 +3380,21 @@ void HOGDescriptor::detectROI(const cv::Mat& img, const std::vector<cv::Point> & ...@@ -3609,37 +3380,21 @@ void HOGDescriptor::detectROI(const cv::Mat& img, const std::vector<cv::Point> &
// need to divide this into 4 parts! // need to divide this into 4 parts!
const float* vec = cache.getBlock(pt, &blockHist[0]); const float* vec = cache.getBlock(pt, &blockHist[0]);
#if CV_SSE2 #if CV_SIMD128
__m128 _vec = _mm_loadu_ps(vec); v_float32x4 _vec = v_load(vec);
__m128 _svmVec = _mm_loadu_ps(svmVec); v_float32x4 _svmVec = v_load(svmVec);
__m128 sum = _mm_mul_ps(_svmVec, _vec); v_float32x4 sum = _svmVec * _vec;
for( k = 4; k <= blockHistogramSize - 4; k += 4 ) for( k = 4; k <= blockHistogramSize - 4; k += 4 )
{ {
_vec = _mm_loadu_ps(vec + k); _vec = v_load(vec + k);
_svmVec = _mm_loadu_ps(svmVec + k); _svmVec = v_load(svmVec + k);
sum = _mm_add_ps(sum, _mm_mul_ps(_vec, _svmVec)); sum += _vec * _svmVec;
} }
_mm_storeu_ps(partSum, sum); v_store(partSum, sum);
double t0 = partSum[0] + partSum[1];
double t1 = partSum[2] + partSum[3];
s += t0 + t1;
#elif CV_NEON
float32x4_t _vec = vld1q_f32(vec);
float32x4_t _svmVec = vld1q_f32(svmVec);
float32x4_t sum = vmulq_f32(_svmVec, _vec);
for( k = 4; k <= blockHistogramSize - 4; k += 4 )
{
_vec = vld1q_f32(vec + k);
_svmVec = vld1q_f32(svmVec + k);
sum = vaddq_f32(sum, vmulq_f32(_vec, _svmVec));
}
vst1q_f32(partSum, sum);
double t0 = partSum[0] + partSum[1]; double t0 = partSum[0] + partSum[1];
double t1 = partSum[2] + partSum[3]; double t1 = partSum[2] + partSum[3];
s += t0 + t1; s += t0 + t1;
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment