Skip to content

O
opencv
Commit 95967461 authored by Vadim Pisarevsky's avatar Vadim Pisarevsky
Browse files
Options

restored SSE2 and added AVX optimization of the old haar face detector

parent dea52eb7
Hide whitespace changes
Inline Side-by-side
Showing with 427 additions and 103 deletions
modules/core/src/system.cpp
View file @ 95967461
......@@ -178,7 +178,7 @@ struct HWFeatures
f.have[CV_CPU_SSE4_1] = (cpuid_data[2] & (1<<19)) != 0;
f.have[CV_CPU_SSE4_2] = (cpuid_data[2] & (1<<20)) != 0;
f.have[CV_CPU_POPCNT] = (cpuid_data[2] & (1<<23)) != 0;
f.have[CV_CPU_AVX] = (cpuid_data[2] & (1<<28)) != 0;
f.have[CV_CPU_AVX] = (((cpuid_data[2] & (1<<28)) != 0)&&((cpuid_data[2] & (1<<27)) != 0));//OS uses XSAVE_XRSTORE and CPU support AVX
}
return f;
......
modules/objdetect/src/haar.cpp
View file @ 95967461
......@@ -43,26 +43,23 @@
#include "precomp.hpp"
#include <stdio.h>
/*
#if CV_SSE2
#if CV_SSE2 || CV_SSE3
# if !CV_SSE4_1 && !CV_SSE4_2
# define _mm_blendv_pd(a, b, m) _mm_xor_pd(a, _mm_and_pd(_mm_xor_pd(b, a), m))
# define _mm_blendv_ps(a, b, m) _mm_xor_ps(a, _mm_and_ps(_mm_xor_ps(b, a), m))
# endif
#endif
#if defined CV_ICC
# if defined CV_AVX
# if CV_AVX
# define CV_HAAR_USE_AVX 1
# else
# if defined CV_SSE2 || defined CV_SSE4_1 || defined CV_SSE4_2
# if CV_SSE2 || CV_SSE3
# define CV_HAAR_USE_SSE 1
# else
# define CV_HAAR_NO_SIMD 1
# endif
# endif
#endif
*/
/* these settings affect the quality of detection: change with care */
#define CV_ADJUST_FEATURES 1
#define CV_ADJUST_WEIGHTS 0
......@@ -636,34 +633,163 @@ cvSetImagesForHaarClassifierCascade( CvHaarClassifierCascade* _cascade,
}
//AVX version icvEvalHidHaarClassifier. Process 8 CvHidHaarClassifiers per call. Check AVX support before invocation!!
#ifdef CV_HAAR_USE_AVX
CV_INLINE
double icvEvalHidHaarClassifierAVX( CvHidHaarClassifier* classifier,
double variance_norm_factor, size_t p_offset )
{
int CV_DECL_ALIGNED(32) idxV[8] = {0,0,0,0,0,0,0,0};
char flags[8] = {0,0,0,0,0,0,0,0};
CvHidHaarTreeNode* nodes[8];
double res = 0;
char exitConditionFlag = 0;
for(;;)
{
float CV_DECL_ALIGNED(32) tmp[8] = {0,0,0,0,0,0,0,0};
nodes[0] = classifier ->node + idxV[0];
nodes[1] = (classifier+1)->node + idxV[1];
nodes[2] = (classifier+2)->node + idxV[2];
nodes[3] = (classifier+3)->node + idxV[3];
nodes[4] = (classifier+4)->node + idxV[4];
nodes[5] = (classifier+5)->node + idxV[5];
nodes[6] = (classifier+6)->node + idxV[6];
nodes[7] = (classifier+7)->node + idxV[7];
__m256 t = _mm256_set1_ps(variance_norm_factor);
t = _mm256_mul_ps(t, _mm256_set_ps(nodes[7]->threshold,nodes[6]->threshold,nodes[5]->threshold,nodes[4]->threshold,nodes[3]->threshold,nodes[2]->threshold,nodes[1]->threshold,nodes[0]->threshold));
__m256 offset = _mm256_set_ps(calc_sum(nodes[7]->feature.rect[0],p_offset), calc_sum(nodes[6]->feature.rect[0],p_offset), calc_sum(nodes[5]->feature.rect[0],p_offset),
calc_sum(nodes[4]->feature.rect[0],p_offset), calc_sum(nodes[3]->feature.rect[0],p_offset), calc_sum(nodes[2]->feature.rect[0],p_offset), calc_sum(nodes[1]->feature.rect[0],
p_offset),calc_sum(nodes[0]->feature.rect[0],p_offset));
__m256 weight = _mm256_set_ps(nodes[7]->feature.rect[0].weight, nodes[6]->feature.rect[0].weight, nodes[5]->feature.rect[0].weight,
nodes[4]->feature.rect[0].weight, nodes[3]->feature.rect[0].weight, nodes[2]->feature.rect[0].weight, nodes[1]->feature.rect[0].weight, nodes[0]->feature.rect[0].weight);
__m256 sum = _mm256_mul_ps(offset, weight);
offset = _mm256_set_ps(calc_sum(nodes[7]->feature.rect[1],p_offset),calc_sum(nodes[6]->feature.rect[1],p_offset),calc_sum(nodes[5]->feature.rect[1],p_offset),
calc_sum(nodes[4]->feature.rect[1],p_offset),calc_sum(nodes[3]->feature.rect[1],p_offset),calc_sum(nodes[2]->feature.rect[1],p_offset),calc_sum(nodes[1]->feature.rect[1],p_offset),
calc_sum(nodes[0]->feature.rect[1],p_offset));
weight = _mm256_set_ps(nodes[7]->feature.rect[1].weight, nodes[6]->feature.rect[1].weight, nodes[5]->feature.rect[1].weight, nodes[4]->feature.rect[1].weight,
nodes[3]->feature.rect[1].weight, nodes[2]->feature.rect[1].weight, nodes[1]->feature.rect[1].weight, nodes[0]->feature.rect[1].weight);
sum = _mm256_add_ps(sum, _mm256_mul_ps(offset,weight));
if( nodes[0]->feature.rect[2].p0 )
tmp[0] = calc_sum(nodes[0]->feature.rect[2],p_offset) * nodes[0]->feature.rect[2].weight;
if( nodes[1]->feature.rect[2].p0 )
tmp[1] = calc_sum(nodes[1]->feature.rect[2],p_offset) * nodes[1]->feature.rect[2].weight;
if( nodes[2]->feature.rect[2].p0 )
tmp[2] = calc_sum(nodes[2]->feature.rect[2],p_offset) * nodes[2]->feature.rect[2].weight;
if( nodes[3]->feature.rect[2].p0 )
tmp[3] = calc_sum(nodes[3]->feature.rect[2],p_offset) * nodes[3]->feature.rect[2].weight;
if( nodes[4]->feature.rect[2].p0 )
tmp[4] = calc_sum(nodes[4]->feature.rect[2],p_offset) * nodes[4]->feature.rect[2].weight;
if( nodes[5]->feature.rect[2].p0 )
tmp[5] = calc_sum(nodes[5]->feature.rect[2],p_offset) * nodes[5]->feature.rect[2].weight;
if( nodes[6]->feature.rect[2].p0 )
tmp[6] = calc_sum(nodes[6]->feature.rect[2],p_offset) * nodes[6]->feature.rect[2].weight;
if( nodes[7]->feature.rect[2].p0 )
tmp[7] = calc_sum(nodes[7]->feature.rect[2],p_offset) * nodes[7]->feature.rect[2].weight;
sum = _mm256_add_ps(sum,_mm256_load_ps(tmp));
__m256 left = _mm256_set_ps(nodes[7]->left,nodes[6]->left,nodes[5]->left,nodes[4]->left,nodes[3]->left,nodes[2]->left,nodes[1]->left,nodes[0]->left);
__m256 right = _mm256_set_ps(nodes[7]->right,nodes[6]->right,nodes[5]->right,nodes[4]->right,nodes[3]->right,nodes[2]->right,nodes[1]->right,nodes[0]->right);
_mm256_store_si256((__m256i*)idxV,_mm256_cvttps_epi32(_mm256_blendv_ps(right, left,_mm256_cmp_ps(sum, t, _CMP_LT_OQ ))));
for(int i = 0; i < 8; i++)
{
if(idxV[i]<=0)
{
if(!flags[i])
{
exitConditionFlag++;
flags[i]=1;
res+=((classifier+i)->alpha[-idxV[i]]);
}
idxV[i]=0;
}
}
if(exitConditionFlag==8)
return res;
}
}
#endif
CV_INLINE
double icvEvalHidHaarClassifier( CvHidHaarClassifier* classifier,
double variance_norm_factor,
size_t p_offset )
{
int idx = 0;
do
/*#if CV_HAAR_USE_SSE && !CV_HAAR_USE_AVX
if(cv::checkHardwareSupport(CV_CPU_SSE2))//based on old SSE variant. Works slow
{
double CV_DECL_ALIGNED(16) temp[2];
__m128d zero = _mm_setzero_pd();
do
{
CvHidHaarTreeNode* node = classifier->node + idx;
__m128d t = _mm_set1_pd((node->threshold)*variance_norm_factor);
__m128d left = _mm_set1_pd(node->left);
__m128d right = _mm_set1_pd(node->right);
double _sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
_sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
if( node->feature.rect[2].p0 )
_sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
__m128d sum = _mm_set1_pd(_sum);
t = _mm_cmplt_sd(sum, t);
sum = _mm_blendv_pd(right, left, t);
_mm_store_pd(temp, sum);
idx = (int)temp[0];
}
while(idx > 0 );
}
else
#endif*/
{
CvHidHaarTreeNode* node = classifier->node + idx;
double t = node->threshold * variance_norm_factor;
do
{
CvHidHaarTreeNode* node = classifier->node + idx;
double t = node->threshold * variance_norm_factor;
double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
if( node->feature.rect[2].p0 )
sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
if( node->feature.rect[2].p0 )
sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
idx = sum < t ? node->left : node->right;
idx = sum < t ? node->left : node->right;
}
while( idx > 0 );
}
while( idx > 0 );
return classifier->alpha[-idx];
}
static int
cvRunHaarClassifierCascadeSum( const CvHaarClassifierCascade* _cascade,
CvPoint pt, double& stage_sum, int start_stage )
{
#ifdef CV_HAAR_USE_AVX
bool haveAVX = false;
if(cv::checkHardwareSupport(CV_CPU_AVX))
if(_xgetbv(_XCR_XFEATURE_ENABLED_MASK)&0x6)// Check if the OS will save the YMM registers
{
haveAVX = true;
}
#else
#ifdef CV_HAAR_USE_SSE
bool haveSSE2 = cv::checkHardwareSupport(CV_CPU_SSE2);
#endif
#endif
int p_offset, pq_offset;
int i, j;
double mean, variance_norm_factor;
......@@ -702,10 +828,20 @@ cvRunHaarClassifierCascadeSum( const CvHaarClassifierCascade* _cascade,
{
stage_sum = 0.0;
#ifdef CV_HAAR_USE_AVX
if(haveAVX)
{
for( ; j < cascade->stage_classifier[i].count-8; j+=8 )
{
stage_sum += icvEvalHidHaarClassifierAVX(
cascade->stage_classifier[i].classifier+j,
variance_norm_factor, p_offset );
}
}
#endif
for( j = 0; j < ptr->count; j++ )
{
stage_sum += icvEvalHidHaarClassifier( ptr->classifier + j,
variance_norm_factor, p_offset );
stage_sum += icvEvalHidHaarClassifier( ptr->classifier + j, variance_norm_factor, p_offset );
}
if( stage_sum >= ptr->threshold )
......@@ -723,99 +859,287 @@ cvRunHaarClassifierCascadeSum( const CvHaarClassifierCascade* _cascade,
}
else if( cascade->isStumpBased )
{
for( i = start_stage; i < cascade->count; i++ )
{
#ifndef CV_HAAR_USE_SSE
stage_sum = 0.0;
#else
__m128d stage_sum = _mm_setzero_pd();
#endif
if( cascade->stage_classifier[i].two_rects )
{
for( j = 0; j < cascade->stage_classifier[i].count; j++ )
{
CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
CvHidHaarTreeNode* node = classifier->node;
#ifndef CV_HAAR_USE_SSE
double t = node->threshold*variance_norm_factor;
double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
stage_sum += classifier->alpha[sum >= t];
#else
// ayasin - NHM perf optim. Avoid use of costly flaky jcc
__m128d t = _mm_set_sd(node->threshold*variance_norm_factor);
__m128d a = _mm_set_sd(classifier->alpha[0]);
__m128d b = _mm_set_sd(classifier->alpha[1]);
__m128d sum = _mm_set_sd(calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight +
calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight);
t = _mm_cmpgt_sd(t, sum);
stage_sum = _mm_add_sd(stage_sum, _mm_blendv_pd(b, a, t));
#endif
}
}
else
{
for( j = 0; j < cascade->stage_classifier[i].count; j++ )
{
CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
CvHidHaarTreeNode* node = classifier->node;
#ifndef CV_HAAR_USE_SSE
double t = node->threshold*variance_norm_factor;
double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
if( node->feature.rect[2].p0 )
sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
stage_sum += classifier->alpha[sum >= t];
#else
// ayasin - NHM perf optim. Avoid use of costly flaky jcc
__m128d t = _mm_set_sd(node->threshold*variance_norm_factor);
__m128d a = _mm_set_sd(classifier->alpha[0]);
__m128d b = _mm_set_sd(classifier->alpha[1]);
double _sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
_sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
if( node->feature.rect[2].p0 )
_sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
__m128d sum = _mm_set_sd(_sum);
t = _mm_cmpgt_sd(t, sum);
stage_sum = _mm_add_sd(stage_sum, _mm_blendv_pd(b, a, t));
#endif
}
}
#ifndef CV_HAAR_USE_SSE
if( stage_sum < cascade->stage_classifier[i].threshold )
#else
__m128d i_threshold = _mm_set_sd(cascade->stage_classifier[i].threshold);
if( _mm_comilt_sd(stage_sum, i_threshold) )
#endif
return -i;
}
}
else
#ifdef CV_HAAR_USE_AVX
if(haveAVX)
{
CvHidHaarClassifier* classifiers[8];
CvHidHaarTreeNode* nodes[8];
for( i = start_stage; i < cascade->count; i++ )
{
stage_sum = 0.0;
int j = 0;
float CV_DECL_ALIGNED(32) buf[8];
if( cascade->stage_classifier[i].two_rects )
{
for( ; j <= cascade->stage_classifier[i].count-8; j+=8 )
{
//__m256 stage_sumPart = _mm256_setzero_ps();
classifiers[0] = cascade->stage_classifier[i].classifier + j;
nodes[0] = classifiers[0]->node;
classifiers[1] = cascade->stage_classifier[i].classifier + j + 1;
nodes[1] = classifiers[1]->node;
classifiers[2] = cascade->stage_classifier[i].classifier + j + 2;
nodes[2]= classifiers[2]->node;
classifiers[3] = cascade->stage_classifier[i].classifier + j + 3;
nodes[3] = classifiers[3]->node;
classifiers[4] = cascade->stage_classifier[i].classifier + j + 4;
nodes[4] = classifiers[4]->node;
classifiers[5] = cascade->stage_classifier[i].classifier + j + 5;
nodes[5] = classifiers[5]->node;
classifiers[6] = cascade->stage_classifier[i].classifier + j + 6;
nodes[6] = classifiers[6]->node;
classifiers[7] = cascade->stage_classifier[i].classifier + j + 7;
nodes[7] = classifiers[7]->node;
__m256 t = _mm256_set1_ps(variance_norm_factor);
t = _mm256_mul_ps(t, _mm256_set_ps(nodes[7]->threshold,nodes[6]->threshold,nodes[5]->threshold,nodes[4]->threshold,nodes[3]->threshold,nodes[2]->threshold,nodes[1]->threshold,nodes[0]->threshold));
__m256 offset = _mm256_set_ps(calc_sum(nodes[7]->feature.rect[0],p_offset), calc_sum(nodes[6]->feature.rect[0],p_offset), calc_sum(nodes[5]->feature.rect[0],p_offset),
calc_sum(nodes[4]->feature.rect[0],p_offset), calc_sum(nodes[3]->feature.rect[0],p_offset), calc_sum(nodes[2]->feature.rect[0],p_offset), calc_sum(nodes[1]->feature.rect[0],
p_offset),calc_sum(nodes[0]->feature.rect[0],p_offset));
__m256 weight = _mm256_set_ps(nodes[7]->feature.rect[0].weight, nodes[6]->feature.rect[0].weight, nodes[5]->feature.rect[0].weight,
nodes[4]->feature.rect[0].weight, nodes[3]->feature.rect[0].weight, nodes[2]->feature.rect[0].weight, nodes[1]->feature.rect[0].weight, nodes[0]->feature.rect[0].weight);
__m256 sum = _mm256_mul_ps(offset, weight);
offset = _mm256_set_ps(calc_sum(nodes[7]->feature.rect[1],p_offset),calc_sum(nodes[6]->feature.rect[1],p_offset),calc_sum(nodes[5]->feature.rect[1],p_offset),
calc_sum(nodes[4]->feature.rect[1],p_offset),calc_sum(nodes[3]->feature.rect[1],p_offset),calc_sum(nodes[2]->feature.rect[1],p_offset),calc_sum(nodes[1]->feature.rect[1],p_offset),
calc_sum(nodes[0]->feature.rect[1],p_offset));
weight = _mm256_set_ps(nodes[7]->feature.rect[1].weight, nodes[6]->feature.rect[1].weight, nodes[5]->feature.rect[1].weight, nodes[4]->feature.rect[1].weight,
nodes[3]->feature.rect[1].weight, nodes[2]->feature.rect[1].weight, nodes[1]->feature.rect[1].weight, nodes[0]->feature.rect[1].weight);
sum = _mm256_add_ps(sum, _mm256_mul_ps(offset,weight));
__m256 alpha0 = _mm256_set_ps(classifiers[7]->alpha[0],classifiers[6]->alpha[0],classifiers[5]->alpha[0],classifiers[4]->alpha[0],classifiers[3]->alpha[0],
classifiers[2]->alpha[0],classifiers[1]->alpha[0],classifiers[0]->alpha[0]);
__m256 alpha1 = _mm256_set_ps(classifiers[7]->alpha[1],classifiers[6]->alpha[1],classifiers[5]->alpha[1],classifiers[4]->alpha[1],classifiers[3]->alpha[1],
classifiers[2]->alpha[1],classifiers[1]->alpha[1],classifiers[0]->alpha[1]);
_mm256_store_ps(buf, _mm256_blendv_ps(alpha0, alpha1, _mm256_cmp_ps(t, sum, _CMP_LE_OQ )));
stage_sum+=(buf[0]+buf[1]+buf[2]+buf[3]+buf[4]+buf[5]+buf[6]+buf[7]);
}
for( ; j < cascade->stage_classifier[i].count; j++ )
{
CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
CvHidHaarTreeNode* node = classifier->node;
double t = node->threshold*variance_norm_factor;
double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
stage_sum += classifier->alpha[sum >= t];
}
}
else
{
for( ; j <= (cascade->stage_classifier[i].count)-8; j+=8 )
{
float CV_DECL_ALIGNED(32) tmp[8] = {0,0,0,0,0,0,0,0};
classifiers[0] = cascade->stage_classifier[i].classifier + j;
nodes[0] = classifiers[0]->node;
classifiers[1] = cascade->stage_classifier[i].classifier + j + 1;
nodes[1] = classifiers[1]->node;
classifiers[2] = cascade->stage_classifier[i].classifier + j + 2;
nodes[2]= classifiers[2]->node;
classifiers[3] = cascade->stage_classifier[i].classifier + j + 3;
nodes[3] = classifiers[3]->node;
classifiers[4] = cascade->stage_classifier[i].classifier + j + 4;
nodes[4] = classifiers[4]->node;
classifiers[5] = cascade->stage_classifier[i].classifier + j + 5;
nodes[5] = classifiers[5]->node;
classifiers[6] = cascade->stage_classifier[i].classifier + j + 6;
nodes[6] = classifiers[6]->node;
classifiers[7] = cascade->stage_classifier[i].classifier + j + 7;
nodes[7] = classifiers[7]->node;
__m256 t = _mm256_set1_ps(variance_norm_factor);
t = _mm256_mul_ps(t, _mm256_set_ps(nodes[7]->threshold,nodes[6]->threshold,nodes[5]->threshold,nodes[4]->threshold,nodes[3]->threshold,nodes[2]->threshold,nodes[1]->threshold,nodes[0]->threshold));
__m256 offset = _mm256_set_ps(calc_sum(nodes[7]->feature.rect[0],p_offset), calc_sum(nodes[6]->feature.rect[0],p_offset), calc_sum(nodes[5]->feature.rect[0],p_offset),
calc_sum(nodes[4]->feature.rect[0],p_offset), calc_sum(nodes[3]->feature.rect[0],p_offset), calc_sum(nodes[2]->feature.rect[0],p_offset), calc_sum(nodes[1]->feature.rect[0],
p_offset),calc_sum(nodes[0]->feature.rect[0],p_offset));
__m256 weight = _mm256_set_ps(nodes[7]->feature.rect[0].weight, nodes[6]->feature.rect[0].weight, nodes[5]->feature.rect[0].weight,
nodes[4]->feature.rect[0].weight, nodes[3]->feature.rect[0].weight, nodes[2]->feature.rect[0].weight, nodes[1]->feature.rect[0].weight, nodes[0]->feature.rect[0].weight);
__m256 sum = _mm256_mul_ps(offset, weight);
offset = _mm256_set_ps(calc_sum(nodes[7]->feature.rect[1],p_offset),calc_sum(nodes[6]->feature.rect[1],p_offset),calc_sum(nodes[5]->feature.rect[1],p_offset),
calc_sum(nodes[4]->feature.rect[1],p_offset),calc_sum(nodes[3]->feature.rect[1],p_offset),calc_sum(nodes[2]->feature.rect[1],p_offset),calc_sum(nodes[1]->feature.rect[1],p_offset),
calc_sum(nodes[0]->feature.rect[1],p_offset));
weight = _mm256_set_ps(nodes[7]->feature.rect[1].weight, nodes[6]->feature.rect[1].weight, nodes[5]->feature.rect[1].weight, nodes[4]->feature.rect[1].weight,
nodes[3]->feature.rect[1].weight, nodes[2]->feature.rect[1].weight, nodes[1]->feature.rect[1].weight, nodes[0]->feature.rect[1].weight);
sum = _mm256_add_ps(sum, _mm256_mul_ps(offset,weight));
if( nodes[0]->feature.rect[2].p0 )
tmp[0] = calc_sum(nodes[0]->feature.rect[2],p_offset) * nodes[0]->feature.rect[2].weight;
if( nodes[1]->feature.rect[2].p0 )
tmp[1] = calc_sum(nodes[1]->feature.rect[2],p_offset) * nodes[1]->feature.rect[2].weight;
if( nodes[2]->feature.rect[2].p0 )
tmp[2] = calc_sum(nodes[2]->feature.rect[2],p_offset) * nodes[2]->feature.rect[2].weight;
if( nodes[3]->feature.rect[2].p0 )
tmp[3] = calc_sum(nodes[3]->feature.rect[2],p_offset) * nodes[3]->feature.rect[2].weight;
if( nodes[4]->feature.rect[2].p0 )
tmp[4] = calc_sum(nodes[4]->feature.rect[2],p_offset) * nodes[4]->feature.rect[2].weight;
if( nodes[5]->feature.rect[2].p0 )
tmp[5] = calc_sum(nodes[5]->feature.rect[2],p_offset) * nodes[5]->feature.rect[2].weight;
if( nodes[6]->feature.rect[2].p0 )
tmp[6] = calc_sum(nodes[6]->feature.rect[2],p_offset) * nodes[6]->feature.rect[2].weight;
if( nodes[7]->feature.rect[2].p0 )
tmp[7] = calc_sum(nodes[7]->feature.rect[2],p_offset) * nodes[7]->feature.rect[2].weight;
sum = _mm256_add_ps(sum, _mm256_load_ps(tmp));
__m256 alpha0 = _mm256_set_ps(classifiers[7]->alpha[0],classifiers[6]->alpha[0],classifiers[5]->alpha[0],classifiers[4]->alpha[0],classifiers[3]->alpha[0],
classifiers[2]->alpha[0],classifiers[1]->alpha[0],classifiers[0]->alpha[0]);
__m256 alpha1 = _mm256_set_ps(classifiers[7]->alpha[1],classifiers[6]->alpha[1],classifiers[5]->alpha[1],classifiers[4]->alpha[1],classifiers[3]->alpha[1],
classifiers[2]->alpha[1],classifiers[1]->alpha[1],classifiers[0]->alpha[1]);
__m256 outBuf = _mm256_blendv_ps(alpha0, alpha1, _mm256_cmp_ps(t, sum, _CMP_LE_OQ ));
outBuf = _mm256_hadd_ps(outBuf, outBuf);
outBuf = _mm256_hadd_ps(outBuf, outBuf);
_mm256_store_ps(buf, outBuf);
stage_sum+=(buf[0]+buf[4]);//(buf[0]+buf[1]+buf[2]+buf[3]+buf[4]+buf[5]+buf[6]+buf[7]);
}
for( ; j < cascade->stage_classifier[i].count; j++ )
{
CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
CvHidHaarTreeNode* node = classifier->node;
double t = node->threshold*variance_norm_factor;
double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
if( node->feature.rect[2].p0 )
sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
stage_sum += classifier->alpha[sum >= t];
}
}
if( stage_sum < cascade->stage_classifier[i].threshold )
return -i;
}
}
else
#endif
#ifdef CV_HAAR_USE_SSE && !CV_HAAR_USE_AVX //old SSE optimization
if(haveSSE2)
{
for( i = start_stage; i < cascade->count; i++ )
{
__m128d stage_sum = _mm_setzero_pd();
if( cascade->stage_classifier[i].two_rects )
{
for( j = 0; j < cascade->stage_classifier[i].count; j++ )
{
CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
CvHidHaarTreeNode* node = classifier->node;
// ayasin - NHM perf optim. Avoid use of costly flaky jcc
__m128d t = _mm_set_sd(node->threshold*variance_norm_factor);
__m128d a = _mm_set_sd(classifier->alpha[0]);
__m128d b = _mm_set_sd(classifier->alpha[1]);
__m128d sum = _mm_set_sd(calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight +
calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight);
t = _mm_cmpgt_sd(t, sum);
stage_sum = _mm_add_sd(stage_sum, _mm_blendv_pd(b, a, t));
}
}
else
{
for( j = 0; j < cascade->stage_classifier[i].count; j++ )
{
CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
CvHidHaarTreeNode* node = classifier->node;
// ayasin - NHM perf optim. Avoid use of costly flaky jcc
__m128d t = _mm_set_sd(node->threshold*variance_norm_factor);
__m128d a = _mm_set_sd(classifier->alpha[0]);
__m128d b = _mm_set_sd(classifier->alpha[1]);
double _sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
_sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
if( node->feature.rect[2].p0 )
_sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
__m128d sum = _mm_set_sd(_sum);
t = _mm_cmpgt_sd(t, sum);
stage_sum = _mm_add_sd(stage_sum, _mm_blendv_pd(b, a, t));
}
}
__m128d i_threshold = _mm_set1_pd(cascade->stage_classifier[i].threshold);
if( _mm_comilt_sd(stage_sum, i_threshold) )
return -i;
}
}
else
#endif
{
for( i = start_stage; i < cascade->count; i++ )
{
stage_sum = 0.0;
if( cascade->stage_classifier[i].two_rects )
{
for( j = 0; j < cascade->stage_classifier[i].count; j++ )
{
CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
CvHidHaarTreeNode* node = classifier->node;
double t = node->threshold*variance_norm_factor;
double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
stage_sum += classifier->alpha[sum >= t];
}
}
else
{
for( j = 0; j < cascade->stage_classifier[i].count; j++ )
{
CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier + j;
CvHidHaarTreeNode* node = classifier->node;
double t = node->threshold*variance_norm_factor;
double sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight;
sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight;
if( node->feature.rect[2].p0 )
sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight;
stage_sum += classifier->alpha[sum >= t];
}
}
if( stage_sum < cascade->stage_classifier[i].threshold )
return -i;
}
}
}
else
{
for( i = start_stage; i < cascade->count; i++ )
{
stage_sum = 0.0;
for( j = 0; j < cascade->stage_classifier[i].count; j++ )
{
stage_sum += icvEvalHidHaarClassifier(
cascade->stage_classifier[i].classifier + j,
variance_norm_factor, p_offset );
}
int j = 0;
#ifdef CV_HAAR_USE_AVX
if(haveAVX)
{
for( ; j < cascade->stage_classifier[i].count-8; j+=8 )
{
stage_sum += icvEvalHidHaarClassifierAVX(
cascade->stage_classifier[i].classifier+j,
variance_norm_factor, p_offset );
}
}
#endif
for(; j < cascade->stage_classifier[i].count; j++ )
{
stage_sum += icvEvalHidHaarClassifier(
cascade->stage_classifier[i].classifier + j,
variance_norm_factor, p_offset );
}
if( stage_sum < cascade->stage_classifier[i].threshold )
return -i;
}
}
//_mm256_zeroupper();
return 1;
}
CV_IMPL int
cvRunHaarClassifierCascade( const CvHaarClassifierCascade* _cascade,
CvPoint pt, int start_stage )
......
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