generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

make oclHaarDetectObjects running on more ocl platforms

Browse Source
This commit is contained in:
yao 2013-01-16 17:13:32 +08:00 committed by Andrey Kamaev
parent b5bd2cde9e
commit 56c1a7fab6
2 changed files with 344 additions and 399 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
modules/ocl/src/haar.cpp
View File

@ -883,13 +883,6 @@ CvSeq *cv::ocl::OclCascadeClassifier::oclHaarDetectObjects( oclMat &gimg, CvMemS
bool findBiggestObject = (flags & CV_HAAR_FIND_BIGGEST_OBJECT) != 0; bool findBiggestObject = (flags & CV_HAAR_FIND_BIGGEST_OBJECT) != 0;
// bool roughSearch = (flags & CV_HAAR_DO_ROUGH_SEARCH) != 0; // bool roughSearch = (flags & CV_HAAR_DO_ROUGH_SEARCH) != 0;
//the Intel HD Graphics is unsupported
if (gimg.clCxt->impl->devName.find("Intel(R) HD Graphics") != string::npos)
{
cout << " Intel HD GPU device unsupported " << endl;
return NULL;
}
//double t = 0; //double t = 0;
if( maxSize.height == 0 || maxSize.width == 0 ) if( maxSize.height == 0 || maxSize.width == 0 )
{ {
@ -937,7 +930,7 @@ CvSeq *cv::ocl::OclCascadeClassifier::oclHaarDetectObjects( oclMat &gimg, CvMemS
if( gimg.cols < minSize.width || gimg.rows < minSize.height ) if( gimg.cols < minSize.width || gimg.rows < minSize.height )
CV_Error(CV_StsError, "Image too small"); CV_Error(CV_StsError, "Image too small");
if( flags & CV_HAAR_SCALE_IMAGE ) if( (flags & CV_HAAR_SCALE_IMAGE) && gimg.clCxt->impl->devName.find("Intel(R) HD Graphics") == string::npos )
{ {
CvSize winSize0 = cascade->orig_window_size; CvSize winSize0 = cascade->orig_window_size;
//float scalefactor = 1.1f; //float scalefactor = 1.1f;

84
modules/ocl/src/kernels/haarobjectdetect_scaled2.cl
View File

@ -44,7 +44,7 @@
//M*/ //M*/
// Enter your kernel in this window // Enter your kernel in this window
#pragma OPENCL EXTENSION cl_amd_printf:enable //#pragma OPENCL EXTENSION cl_amd_printf:enable
#define CV_HAAR_FEATURE_MAX 3 #define CV_HAAR_FEATURE_MAX 3
typedef int sumtype; typedef int sumtype;
typedef float sqsumtype; typedef float sqsumtype;
@ -144,6 +144,7 @@ __kernel void gpuRunHaarClassifierCascade_scaled2(
candidate[outputoff + (lcl_id << 2) + 1] = (int4)0; candidate[outputoff + (lcl_id << 2) + 1] = (int4)0;
candidate[outputoff + (lcl_id << 2) + 2] = (int4)0; candidate[outputoff + (lcl_id << 2) + 2] = (int4)0;
candidate[outputoff + (lcl_id << 2) + 3] = (int4)0; candidate[outputoff + (lcl_id << 2) + 3] = (int4)0;
for (int scalei = 0; scalei < loopcount; scalei++) for (int scalei = 0; scalei < loopcount; scalei++)
{ {
int4 scaleinfo1; int4 scaleinfo1;
@ -155,7 +156,9 @@ __kernel void gpuRunHaarClassifierCascade_scaled2(
float factor = as_float(scaleinfo1.w); float factor = as_float(scaleinfo1.w);
float correction_t = correction[scalei]; float correction_t = correction[scalei];
int ystep = (int)(max(2.0f, factor) + 0.5f); int ystep = (int)(max(2.0f, factor) + 0.5f);
for(int grploop=get_group_id(0);grploop<totalgrp;grploop+=grpnumx){
for (int grploop = get_group_id(0); grploop < totalgrp; grploop += grpnumx)
{
int4 cascadeinfo = p[scalei]; int4 cascadeinfo = p[scalei];
int grpidy = grploop / grpnumperline; int grpidy = grploop / grpnumperline;
int grpidx = grploop - mul24(grpidy, grpnumperline); int grpidx = grploop - mul24(grpidy, grpnumperline);
@ -181,11 +184,13 @@ __kernel void gpuRunHaarClassifierCascade_scaled2(
variance_norm_factor = variance_norm_factor >= 0.f ? sqrt(variance_norm_factor) : 1.f; variance_norm_factor = variance_norm_factor >= 0.f ? sqrt(variance_norm_factor) : 1.f;
result = 1; result = 1;
nodecounter = startnode + nodecount * scalei; nodecounter = startnode + nodecount * scalei;
for(int stageloop = start_stage; stageloop < split_stage&&result; stageloop++ )
for (int stageloop = start_stage; stageloop < end_stage && result; stageloop++)
{ {
float stage_sum = 0.f; float stage_sum = 0.f;
int4 stageinfo = *(global int4 *)(stagecascadeptr + stageloop); int4 stageinfo = *(global int4 *)(stagecascadeptr + stageloop);
float stagethreshold = as_float(stageinfo.y); float stagethreshold = as_float(stageinfo.y);
for (int nodeloop = 0; nodeloop < stageinfo.x; nodeloop++) for (int nodeloop = 0; nodeloop < stageinfo.x; nodeloop++)
{ {
__global GpuHidHaarTreeNode *currentnodeptr = (nodeptr + nodecounter); __global GpuHidHaarTreeNode *currentnodeptr = (nodeptr + nodecounter);
@ -210,80 +215,21 @@ __kernel void gpuRunHaarClassifierCascade_scaled2(
stage_sum += classsum >= nodethreshold ? alpha2.y : alpha2.x; stage_sum += classsum >= nodethreshold ? alpha2.y : alpha2.x;
nodecounter++; nodecounter++;
} }
result = (stage_sum >= stagethreshold); result = (stage_sum >= stagethreshold);
} }
if (result && (ix < width) && (iy < height)) if (result && (ix < width) && (iy < height))
{ {
int queueindex = atomic_inc(lclcount); int queueindex = atomic_inc(lclcount);
lcloutindex[queueindex << 1] = (y << 16) | x; lcloutindex[queueindex << 1] = (y << 16) | x;
lcloutindex[(queueindex << 1) + 1] = as_int(variance_norm_factor); lcloutindex[(queueindex << 1) + 1] = as_int(variance_norm_factor);
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
int queuecount = lclcount[0]; int queuecount = lclcount[0];
nodecounter = splitnode + nodecount * scalei; nodecounter = splitnode + nodecount * scalei;
for(int stageloop=split_stage;stageloop<end_stage&&queuecount>0;stageloop++)
{
lclcount[0]=0;
barrier(CLK_LOCAL_MEM_FENCE);
int2 stageinfo=*(global int2*)(stagecascadeptr+stageloop);
float stagethreshold=as_float(stageinfo.y);
int perfscale=queuecount>4?3:2;
int queuecount_loop=(queuecount+(1<<perfscale)-1)>>perfscale;
int lcl_compute_win=lcl_sz>>perfscale;
int lcl_compute_win_id=(lcl_id>>(6-perfscale));
int lcl_loops=(stageinfo.x+lcl_compute_win-1)>>(6-perfscale);
int lcl_compute_id=lcl_id-(lcl_compute_win_id<<(6-perfscale));
for(int queueloop=0;queueloop<queuecount_loop&&lcl_compute_win_id<queuecount;queueloop++)
{
float stage_sum=0.f;
int temp_coord=lcloutindex[lcl_compute_win_id<<1];
float variance_norm_factor=as_float(lcloutindex[(lcl_compute_win_id<<1)+1]);
int queue_offset=mad24(((temp_coord&(int)0xffff0000)>>16),step,temp_coord&0xffff);
int tempnodecounter=lcl_compute_id;
float part_sum=0.f;
for(int lcl_loop=0;lcl_loop<lcl_loops&&tempnodecounter<stageinfo.x;lcl_loop++)
{
__global GpuHidHaarTreeNode* currentnodeptr = (nodeptr + nodecounter + tempnodecounter);
int4 info1 = *(__global int4*)(&(currentnodeptr->p[0][0]));
int4 info2 = *(__global int4*)(&(currentnodeptr->p[1][0]));
int4 info3 = *(__global int4*)(&(currentnodeptr->p[2][0]));
float4 w = *(__global float4*)(&(currentnodeptr->weight[0]));
float2 alpha2 = *(__global float2*)(&(currentnodeptr->alpha[0]));
float nodethreshold = w.w * variance_norm_factor;
info1.x +=queue_offset;
info1.z +=queue_offset;
info2.x +=queue_offset;
info2.z +=queue_offset;
float classsum = (sum[mad24(info1.y,step,info1.x)] - sum[mad24(info1.y,step,info1.z)] -
sum[mad24(info1.w,step,info1.x)] + sum[mad24(info1.w,step,info1.z)]) * w.x;
classsum += (sum[mad24(info2.y,step,info2.x)] - sum[mad24(info2.y,step,info2.z)] -
sum[mad24(info2.w,step,info2.x)] + sum[mad24(info2.w,step,info2.z)]) * w.y;
info3.x +=queue_offset;
info3.z +=queue_offset;
classsum += (sum[mad24(info3.y,step,info3.x)] - sum[mad24(info3.y,step,info3.z)] -
sum[mad24(info3.w,step,info3.x)] + sum[mad24(info3.w,step,info3.z)]) * w.z;
part_sum += classsum >= nodethreshold ? alpha2.y : alpha2.x;
tempnodecounter+=lcl_compute_win;
}
partialsum[lcl_id]=part_sum;
barrier(CLK_LOCAL_MEM_FENCE);
for(int i=0;i<lcl_compute_win&&(lcl_compute_id==0);i++)
{
stage_sum+=partialsum[lcl_id+i];
}
if(stage_sum>=stagethreshold&&(lcl_compute_id==0))
{
int queueindex=atomic_inc(lclcount);
lcloutindex[queueindex<<1]=temp_coord;
lcloutindex[(queueindex<<1)+1]=as_int(variance_norm_factor);
}
lcl_compute_win_id+=(1<<perfscale);
barrier(CLK_LOCAL_MEM_FENCE);
}
queuecount=lclcount[0];
nodecounter+=stageinfo.x;
}
if (lcl_id < queuecount) if (lcl_id < queuecount)
{ {
int temp = lcloutindex[lcl_id << 1]; int temp = lcloutindex[lcl_id << 1];
@ -297,6 +243,7 @@ __kernel void gpuRunHaarClassifierCascade_scaled2(
atomic_inc(glboutindex); atomic_inc(glboutindex);
candidate[outputoff + temp + lcl_id] = candidate_result; candidate[outputoff + temp + lcl_id] = candidate_result;
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
} }
} }
@ -308,15 +255,19 @@ __kernel void gpuscaleclassifier(global GpuHidHaarTreeNode * orinode, global Gpu
int tr_x[3], tr_y[3], tr_h[3], tr_w[3], i = 0; int tr_x[3], tr_y[3], tr_h[3], tr_w[3], i = 0;
GpuHidHaarTreeNode t1 = *(orinode + counter); GpuHidHaarTreeNode t1 = *(orinode + counter);
#pragma unroll #pragma unroll
for(i=0;i<3;i++){
for (i = 0; i < 3; i++)
{
tr_x[i] = (int)(t1.p[i][0] * scale + 0.5f); tr_x[i] = (int)(t1.p[i][0] * scale + 0.5f);
tr_y[i] = (int)(t1.p[i][1] * scale + 0.5f); tr_y[i] = (int)(t1.p[i][1] * scale + 0.5f);
tr_w[i] = (int)(t1.p[i][2] * scale + 0.5f); tr_w[i] = (int)(t1.p[i][2] * scale + 0.5f);
tr_h[i] = (int)(t1.p[i][3] * scale + 0.5f); tr_h[i] = (int)(t1.p[i][3] * scale + 0.5f);
} }
t1.weight[0] = t1.p[2][0] ? -(t1.weight[1] * tr_h[1] * tr_w[1] + t1.weight[2] * tr_h[2] * tr_w[2]) / (tr_h[0] * tr_w[0]) : -t1.weight[1] * tr_h[1] * tr_w[1] / (tr_h[0] * tr_w[0]); t1.weight[0] = t1.p[2][0] ? -(t1.weight[1] * tr_h[1] * tr_w[1] + t1.weight[2] * tr_h[2] * tr_w[2]) / (tr_h[0] * tr_w[0]) : -t1.weight[1] * tr_h[1] * tr_w[1] / (tr_h[0] * tr_w[0]);
counter += nodenum; counter += nodenum;
#pragma unroll #pragma unroll
for (i = 0; i < 3; i++) for (i = 0; i < 3; i++)
{ {
newnode[counter].p[i][0] = tr_x[i]; newnode[counter].p[i][0] = tr_x[i];
@ -325,6 +276,7 @@ __kernel void gpuscaleclassifier(global GpuHidHaarTreeNode * orinode, global Gpu
newnode[counter].p[i][3] = tr_y[i] + tr_h[i]; newnode[counter].p[i][3] = tr_y[i] + tr_h[i];
newnode[counter].weight[i] = t1.weight[i] * weight_scale; newnode[counter].weight[i] = t1.weight[i] * weight_scale;
} }
newnode[counter].left = t1.left; newnode[counter].left = t1.left;
newnode[counter].right = t1.right; newnode[counter].right = t1.right;
newnode[counter].threshold = t1.threshold; newnode[counter].threshold = t1.threshold;