potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-29 07:13:02 +04:00
Code Packages Releases Wiki Activity

Merge pull request #9042 from terfendail:haar_avx

Browse Source 
AVX optimized implementation of haar migrated to separate file
This commit is contained in:
Alexander Alekhin
2017-07-14 15:05:11 +00:00
committed by GitHub
parent 9439872a62 77264dcca9
commit 431e2e6d68
3 changed files with 492 additions and 372 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/objdetect/src/haar.cpp
+22 -372
View File
@@ -45,6 +45,10 @@
#include "opencv2/imgproc/imgproc_c.h"
#include "opencv2/objdetect/objdetect_c.h"
#include <stdio.h>
#include "haar.hpp"
#if CV_HAAR_FEATURE_MAX_LOCAL != CV_HAAR_FEATURE_MAX
#error CV_HAAR_FEATURE_MAX definition changed. Adjust CV_HAAR_FEATURE_MAX_LOCAL value please.
#endif
#if CV_SSE2
# if 1 /*!CV_SSE4_1 && !CV_SSE4_2*/
@@ -53,8 +57,7 @@
# endif
#endif
#if 0 /*CV_AVX*/
# define CV_HAAR_USE_AVX 1
#if CV_HAAR_USE_AVX
# if defined _MSC_VER
# pragma warning( disable : 4752 )
# endif
@@ -68,38 +71,6 @@
#define CV_ADJUST_FEATURES 1
#define CV_ADJUST_WEIGHTS 0
typedef int sumtype;
typedef double sqsumtype;
typedef struct CvHidHaarFeature
{
struct
{
sumtype *p0, *p1, *p2, *p3;
float weight;
}
rect[CV_HAAR_FEATURE_MAX];
} CvHidHaarFeature;
typedef struct CvHidHaarTreeNode
{
CvHidHaarFeature feature;
float threshold;
int left;
int right;
} CvHidHaarTreeNode;
typedef struct CvHidHaarClassifier
{
int count;
//CvHaarFeature* orig_feature;
CvHidHaarTreeNode* node;
float* alpha;
} CvHidHaarClassifier;
typedef struct CvHidHaarStageClassifier
{
int count;
@@ -420,10 +391,6 @@ icvCreateHidHaarClassifierCascade( CvHaarClassifierCascade* cascade )
#define calc_sum(rect,offset) \
((rect).p0[offset] - (rect).p1[offset] - (rect).p2[offset] + (rect).p3[offset])
#define calc_sumf(rect,offset) \
static_cast<float>((rect).p0[offset] - (rect).p1[offset] - (rect).p2[offset] + (rect).p3[offset])
CV_IMPL void
cvSetImagesForHaarClassifierCascade( CvHaarClassifierCascade* _cascade,
const CvArr* _sum,
@@ -640,129 +607,6 @@ 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};
uchar flags[8] = {0,0,0,0,0,0,0,0};
CvHidHaarTreeNode* nodes[8];
double res = 0;
uchar exitConditionFlag = 0;
for(;;)
{
float CV_DECL_ALIGNED(32) tmp[8] = {0,0,0,0,0,0,0,0};
nodes[0] = (classifier+0)->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(static_cast<float>(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_sumf(nodes[7]->feature.rect[0], p_offset),
calc_sumf(nodes[6]->feature.rect[0], p_offset),
calc_sumf(nodes[5]->feature.rect[0], p_offset),
calc_sumf(nodes[4]->feature.rect[0], p_offset),
calc_sumf(nodes[3]->feature.rect[0], p_offset),
calc_sumf(nodes[2]->feature.rect[0], p_offset),
calc_sumf(nodes[1]->feature.rect[0], p_offset),
calc_sumf(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_sumf(nodes[7]->feature.rect[1], p_offset),
calc_sumf(nodes[6]->feature.rect[1], p_offset),
calc_sumf(nodes[5]->feature.rect[1], p_offset),
calc_sumf(nodes[4]->feature.rect[1], p_offset),
calc_sumf(nodes[3]->feature.rect[1], p_offset),
calc_sumf(nodes[2]->feature.rect[1], p_offset),
calc_sumf(nodes[1]->feature.rect[1], p_offset),
calc_sumf(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_sumf(nodes[0]->feature.rect[2], p_offset) * nodes[0]->feature.rect[2].weight;
if( nodes[1]->feature.rect[2].p0 )
tmp[1] = calc_sumf(nodes[1]->feature.rect[2], p_offset) * nodes[1]->feature.rect[2].weight;
if( nodes[2]->feature.rect[2].p0 )
tmp[2] = calc_sumf(nodes[2]->feature.rect[2], p_offset) * nodes[2]->feature.rect[2].weight;
if( nodes[3]->feature.rect[2].p0 )
tmp[3] = calc_sumf(nodes[3]->feature.rect[2], p_offset) * nodes[3]->feature.rect[2].weight;
if( nodes[4]->feature.rect[2].p0 )
tmp[4] = calc_sumf(nodes[4]->feature.rect[2], p_offset) * nodes[4]->feature.rect[2].weight;
if( nodes[5]->feature.rect[2].p0 )
tmp[5] = calc_sumf(nodes[5]->feature.rect[2], p_offset) * nodes[5]->feature.rect[2].weight;
if( nodes[6]->feature.rect[2].p0 )
tmp[6] = calc_sumf(nodes[6]->feature.rect[2], p_offset) * nodes[6]->feature.rect[2].weight;
if( nodes[7]->feature.rect[2].p0 )
tmp[7] = calc_sumf(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(static_cast<float>(nodes[7]->left), static_cast<float>(nodes[6]->left),
static_cast<float>(nodes[5]->left), static_cast<float>(nodes[4]->left),
static_cast<float>(nodes[3]->left), static_cast<float>(nodes[2]->left),
static_cast<float>(nodes[1]->left), static_cast<float>(nodes[0]->left));
__m256 right = _mm256_set_ps(static_cast<float>(nodes[7]->right),static_cast<float>(nodes[6]->right),
static_cast<float>(nodes[5]->right),static_cast<float>(nodes[4]->right),
static_cast<float>(nodes[3]->right),static_cast<float>(nodes[2]->right),
static_cast<float>(nodes[1]->right),static_cast<float>(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_HAAR_USE_AVX
CV_INLINE
double icvEvalHidHaarClassifier( CvHidHaarClassifier* classifier,
double variance_norm_factor,
@@ -823,8 +667,8 @@ static int
cvRunHaarClassifierCascadeSum( const CvHaarClassifierCascade* _cascade,
CvPoint pt, double& stage_sum, int start_stage )
{
#ifdef CV_HAAR_USE_AVX
bool haveAVX = cv::checkHardwareSupport(CV_CPU_AVX);
#if CV_HAAR_USE_AVX
bool haveAVX = CV_CPU_HAS_SUPPORT_AVX;
#else
# ifdef CV_HAAR_USE_SSE
bool haveSSE2 = cv::checkHardwareSupport(CV_CPU_SSE2);
@@ -870,14 +714,14 @@ cvRunHaarClassifierCascadeSum( const CvHaarClassifierCascade* _cascade,
stage_sum = 0.0;
j = 0;
#ifdef CV_HAAR_USE_AVX
#if CV_HAAR_USE_AVX
if(haveAVX)
{
for( ; j <= ptr->count - 8; j += 8 )
{
stage_sum += icvEvalHidHaarClassifierAVX(
ptr->classifier + j,
variance_norm_factor, p_offset );
stage_sum += cv_haar_avx::icvEvalHidHaarClassifierAVX(
ptr->classifier + j,
variance_norm_factor, p_offset );
}
}
#endif
@@ -901,106 +745,20 @@ cvRunHaarClassifierCascadeSum( const CvHaarClassifierCascade* _cascade,
}
else if( cascade->isStumpBased )
{
#ifdef CV_HAAR_USE_AVX
#if CV_HAAR_USE_AVX
if(haveAVX)
{
CvHidHaarClassifier* classifiers[8];
CvHidHaarTreeNode* nodes[8];
for( i = start_stage; i < cascade->count; i++ )
{
stage_sum = 0.0;
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 )
{
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(static_cast<float>(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_sumf(nodes[7]->feature.rect[0], p_offset),
calc_sumf(nodes[6]->feature.rect[0], p_offset),
calc_sumf(nodes[5]->feature.rect[0], p_offset),
calc_sumf(nodes[4]->feature.rect[0], p_offset),
calc_sumf(nodes[3]->feature.rect[0], p_offset),
calc_sumf(nodes[2]->feature.rect[0], p_offset),
calc_sumf(nodes[1]->feature.rect[0], p_offset),
calc_sumf(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_sumf(nodes[7]->feature.rect[1], p_offset),
calc_sumf(nodes[6]->feature.rect[1], p_offset),
calc_sumf(nodes[5]->feature.rect[1], p_offset),
calc_sumf(nodes[4]->feature.rect[1], p_offset),
calc_sumf(nodes[3]->feature.rect[1], p_offset),
calc_sumf(nodes[2]->feature.rect[1], p_offset),
calc_sumf(nodes[1]->feature.rect[1], p_offset),
calc_sumf(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]);
stage_sum += cv_haar_avx::icvEvalHidHaarStumpClassifierTwoRectAVX(
cascade->stage_classifier[i].classifier + j,
variance_norm_factor, p_offset);
}
for( ; j < cascade->stage_classifier[i].count; j++ )
@@ -1018,117 +776,9 @@ cvRunHaarClassifierCascadeSum( const CvHaarClassifierCascade* _cascade,
{
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(static_cast<float>(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_sumf(nodes[7]->feature.rect[0], p_offset),
calc_sumf(nodes[6]->feature.rect[0], p_offset),
calc_sumf(nodes[5]->feature.rect[0], p_offset),
calc_sumf(nodes[4]->feature.rect[0], p_offset),
calc_sumf(nodes[3]->feature.rect[0], p_offset),
calc_sumf(nodes[2]->feature.rect[0], p_offset),
calc_sumf(nodes[1]->feature.rect[0], p_offset),
calc_sumf(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_sumf(nodes[7]->feature.rect[1], p_offset),
calc_sumf(nodes[6]->feature.rect[1], p_offset),
calc_sumf(nodes[5]->feature.rect[1], p_offset),
calc_sumf(nodes[4]->feature.rect[1], p_offset),
calc_sumf(nodes[3]->feature.rect[1], p_offset),
calc_sumf(nodes[2]->feature.rect[1], p_offset),
calc_sumf(nodes[1]->feature.rect[1], p_offset),
calc_sumf(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_sumf(nodes[0]->feature.rect[2],p_offset) * nodes[0]->feature.rect[2].weight;
if( nodes[1]->feature.rect[2].p0 )
tmp[1] = calc_sumf(nodes[1]->feature.rect[2],p_offset) * nodes[1]->feature.rect[2].weight;
if( nodes[2]->feature.rect[2].p0 )
tmp[2] = calc_sumf(nodes[2]->feature.rect[2],p_offset) * nodes[2]->feature.rect[2].weight;
if( nodes[3]->feature.rect[2].p0 )
tmp[3] = calc_sumf(nodes[3]->feature.rect[2],p_offset) * nodes[3]->feature.rect[2].weight;
if( nodes[4]->feature.rect[2].p0 )
tmp[4] = calc_sumf(nodes[4]->feature.rect[2],p_offset) * nodes[4]->feature.rect[2].weight;
if( nodes[5]->feature.rect[2].p0 )
tmp[5] = calc_sumf(nodes[5]->feature.rect[2],p_offset) * nodes[5]->feature.rect[2].weight;
if( nodes[6]->feature.rect[2].p0 )
tmp[6] = calc_sumf(nodes[6]->feature.rect[2],p_offset) * nodes[6]->feature.rect[2].weight;
if( nodes[7]->feature.rect[2].p0 )
tmp[7] = calc_sumf(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]);
stage_sum += cv_haar_avx::icvEvalHidHaarStumpClassifierAVX(
cascade->stage_classifier[i].classifier + j,
variance_norm_factor, p_offset);
}
for( ; j < cascade->stage_classifier[i].count; j++ )
@@ -1241,14 +891,14 @@ cvRunHaarClassifierCascadeSum( const CvHaarClassifierCascade* _cascade,
stage_sum = 0.0;
int k = 0;
#ifdef CV_HAAR_USE_AVX
#if CV_HAAR_USE_AVX
if(haveAVX)
{
for( ; k < cascade->stage_classifier[i].count - 8; k += 8 )
{
stage_sum += icvEvalHidHaarClassifierAVX(
cascade->stage_classifier[i].classifier + k,
variance_norm_factor, p_offset );
stage_sum += cv_haar_avx::icvEvalHidHaarClassifierAVX(
cascade->stage_classifier[i].classifier + k,
variance_norm_factor, p_offset );
}
}
#endif