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

some optimizations to ocl::pyrDown, PyrLK and Canny

Browse Source
This commit is contained in:
bitwangyaoyao
2012-09-24 20:28:35 +08:00
parent 494ae1562d
commit 09359982b1
9 changed files with 1095 additions and 464 deletions
Download Patch File Download Diff File Expand all files Collapse all files

745
modules/ocl/src/pyrlk.cpp
View File

@@ -41,7 +41,7 @@
//M*/
#include "precomp.hpp"
#include "mcwutil.hpp"
using namespace std;
using namespace cv;
using namespace cv::ocl;
@@ -59,7 +59,10 @@ namespace cv
{
///////////////////////////OpenCL kernel strings///////////////////////////
extern const char *pyrlk;
extern const char *operator_setTo;
extern const char *operator_convertTo;
extern const char *arithm_mul;
extern const char *pyr_down;
}
}
@@ -78,103 +81,6 @@ struct int2
int x, y;
};
void calcSharrDeriv_run(const oclMat& src, oclMat& dx_buf, oclMat& dy_buf, oclMat& dIdx, oclMat& dIdy, int cn)
{
Context *clCxt = src.clCxt;
string kernelName = "calcSharrDeriv_vertical";
size_t localThreads[3] = { 32, 8, 1 };
size_t globalThreads[3] = { src.cols, src.rows, 1};
vector<pair<size_t , const void *> > args;
args.push_back( make_pair( sizeof(cl_mem), (void *)&src.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src.step ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src.rows ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src.cols ));
args.push_back( make_pair( sizeof(cl_int), (void *)&cn ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&dx_buf.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&dx_buf.step ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&dy_buf.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&dy_buf.step ));
openCLExecuteKernel(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args, src.channels(), src.depth());
kernelName = "calcSharrDeriv_horizontal";
vector<pair<size_t , const void *> > args2;
args2.push_back( make_pair( sizeof(cl_int), (void *)&src.rows ));
args2.push_back( make_pair( sizeof(cl_int), (void *)&src.cols ));
args2.push_back( make_pair( sizeof(cl_int), (void *)&cn ));
args2.push_back( make_pair( sizeof(cl_mem), (void *)&dx_buf.data ));
args2.push_back( make_pair( sizeof(cl_int), (void *)&dx_buf.step ));
args2.push_back( make_pair( sizeof(cl_mem), (void *)&dy_buf.data ));
args2.push_back( make_pair( sizeof(cl_int), (void *)&dy_buf.step ));
args2.push_back( make_pair( sizeof(cl_mem), (void *)&dIdx.data ));
args2.push_back( make_pair( sizeof(cl_int), (void *)&dIdx.step ));
args2.push_back( make_pair( sizeof(cl_mem), (void *)&dIdy.data ));
args2.push_back( make_pair( sizeof(cl_int), (void *)&dIdy.step ));
openCLExecuteKernel(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args2, src.channels(), src.depth());
}
void cv::ocl::PyrLKOpticalFlow::calcSharrDeriv(const oclMat& src, oclMat& dIdx, oclMat& dIdy)
{
CV_Assert(src.rows > 1 && src.cols > 1);
CV_Assert(src.depth() == CV_8U);
const int cn = src.channels();
ensureSizeIsEnough(src.size(), CV_MAKETYPE(CV_16S, cn), dx_calcBuf_);
ensureSizeIsEnough(src.size(), CV_MAKETYPE(CV_16S, cn), dy_calcBuf_);
calcSharrDeriv_run(src, dx_calcBuf_, dy_calcBuf_, dIdx, dIdy, cn);
}
void cv::ocl::PyrLKOpticalFlow::buildImagePyramid(const oclMat& img0, vector<oclMat>& pyr, bool withBorder)
{
pyr.resize(maxLevel + 1);
Size sz = img0.size();
Mat img0Temp;
img0.download(img0Temp);
Mat pyrTemp;
oclMat o;
for (int level = 0; level <= maxLevel; ++level)
{
oclMat temp;
if (withBorder)
{
temp.create(sz.height + winSize.height * 2, sz.width + winSize.width * 2, img0.type());
}
else
{
ensureSizeIsEnough(sz, img0.type(), pyr[level]);
}
if (level == 0)
pyr[level] = img0Temp;
else
pyrDown(pyr[level - 1], pyr[level]);
if (withBorder)
copyMakeBorder(pyr[level], temp, winSize.height, winSize.height, winSize.width, winSize.width, BORDER_REFLECT_101);
sz = Size((sz.width + 1) / 2, (sz.height + 1) / 2);
if (sz.width <= winSize.width || sz.height <= winSize.height)
{
maxLevel = level;
break;
}
}
}
namespace
{
void calcPatchSize(cv::Size winSize, int cn, dim3& block, dim3& patch, bool isDeviceArch11)
@@ -199,110 +105,507 @@ namespace
}
}
struct MultiplyScalar
inline int divUp(int total, int grain)
{
MultiplyScalar(double val_, double scale_) : val(val_), scale(scale_) {}
double operator ()(double a) const
return (total + grain - 1) / grain;
}
///////////////////////////////////////////////////////////////////////////
//////////////////////////////// ConvertTo ////////////////////////////////
///////////////////////////////////////////////////////////////////////////
void convert_run_cus(const oclMat &src, oclMat &dst, double alpha, double beta)
{
string kernelName = "convert_to_S";
stringstream idxStr;
idxStr << src.depth();
kernelName += idxStr.str();
float alpha_f = (float)alpha, beta_f = (float)beta;
CV_DbgAssert(src.rows == dst.rows && src.cols == dst.cols);
vector<pair<size_t , const void *> > args;
size_t localThreads[3] = {16, 16, 1};
size_t globalThreads[3];
globalThreads[0] = (dst.cols + localThreads[0] - 1) / localThreads[0] * localThreads[0];
globalThreads[1] = (dst.rows + localThreads[1] - 1) / localThreads[1] * localThreads[1];
globalThreads[2] = 1;
int dststep_in_pixel = dst.step / dst.elemSize(), dstoffset_in_pixel = dst.offset / dst.elemSize();
int srcstep_in_pixel = src.step / src.elemSize(), srcoffset_in_pixel = src.offset / src.elemSize();
if(dst.type() == CV_8UC1)
{
return (scale * a * val);
globalThreads[0] = ((dst.cols + 4) / 4 + localThreads[0]) / localThreads[0] * localThreads[0];
}
const double val;
const double scale;
};
void callF(const oclMat& src, oclMat& dst, MultiplyScalar op, int mask)
{
Mat srcTemp;
Mat dstTemp;
src.download(srcTemp);
dst.download(dstTemp);
int i;
int j;
int k;
for(i = 0; i < srcTemp.rows; i++)
{
for(j = 0; j < srcTemp.cols; j++)
{
for(k = 0; k < srcTemp.channels(); k++)
{
((float*)dstTemp.data)[srcTemp.channels() * (i * srcTemp.rows + j) + k] = (float)op(((float*)srcTemp.data)[srcTemp.channels() * (i * srcTemp.rows + j) + k]);
}
}
}
dst = dstTemp;
args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data ));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src.cols ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src.rows ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&srcstep_in_pixel ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&srcoffset_in_pixel ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dststep_in_pixel ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dstoffset_in_pixel ));
args.push_back( make_pair( sizeof(cl_float) , (void *)&alpha_f ));
args.push_back( make_pair( sizeof(cl_float) , (void *)&beta_f ));
openCLExecuteKernel2(dst.clCxt , &operator_convertTo, kernelName, globalThreads,
localThreads, args, dst.channels(), dst.depth(), CLFLUSH);
}
static inline bool isAligned(const unsigned char* ptr, size_t size)
void convertTo( const oclMat &src, oclMat &m, int rtype, double alpha = 1, double beta = 0 );
void convertTo( const oclMat &src, oclMat &dst, int rtype, double alpha, double beta )
{
return reinterpret_cast<size_t>(ptr) % size == 0;
}
//cout << "cv::ocl::oclMat::convertTo()" << endl;
static inline bool isAligned(size_t step, size_t size)
{
return step % size == 0;
}
bool noScale = fabs(alpha - 1) < std::numeric_limits<double>::epsilon()
&& fabs(beta) < std::numeric_limits<double>::epsilon();
void callT(const oclMat& src, oclMat& dst, MultiplyScalar op, int mask)
{
if (!isAligned(src.data, 4 * sizeof(double)) || !isAligned(src.step, 4 * sizeof(double)) ||
!isAligned(dst.data, 4 * sizeof(double)) || !isAligned(dst.step, 4 * sizeof(double)))
if( rtype < 0 )
rtype = src.type();
else
rtype = CV_MAKETYPE(CV_MAT_DEPTH(rtype), src.channels());
int sdepth = src.depth(), ddepth = CV_MAT_DEPTH(rtype);
if( sdepth == ddepth && noScale )
{
callF(src, dst, op, mask);
src.copyTo(dst);
return;
}
Mat srcTemp;
Mat dstTemp;
src.download(srcTemp);
dst.download(dstTemp);
oclMat temp;
const oclMat *psrc = &src;
if( sdepth != ddepth && psrc == &dst )
psrc = &(temp = src);
int x_shifted;
int i;
int j;
for(i = 0; i < srcTemp.rows; i++)
{
const double* srcRow = (const double*)srcTemp.data + i * srcTemp.rows;
double* dstRow = (double*)dstTemp.data + i * dstTemp.rows;;
for(j = 0; j < srcTemp.cols; j++)
{
x_shifted = j * 4;
if(x_shifted + 4 - 1 < srcTemp.cols)
{
dstRow[x_shifted ] = op(srcRow[x_shifted ]);
dstRow[x_shifted + 1] = op(srcRow[x_shifted + 1]);
dstRow[x_shifted + 2] = op(srcRow[x_shifted + 2]);
dstRow[x_shifted + 3] = op(srcRow[x_shifted + 3]);
}
else
{
for (int real_x = x_shifted; real_x < srcTemp.cols; ++real_x)
{
((float*)dstTemp.data)[i * srcTemp.rows + real_x] = op(((float*)srcTemp.data)[i * srcTemp.rows + real_x]);
}
}
}
}
dst.create( src.size(), rtype );
convert_run_cus(*psrc, dst, alpha, beta);
}
void multiply(const oclMat& src1, double val, oclMat& dst, double scale = 1.0f);
void multiply(const oclMat& src1, double val, oclMat& dst, double scale)
///////////////////////////////////////////////////////////////////////////
//////////////////////////////// setTo ////////////////////////////////////
///////////////////////////////////////////////////////////////////////////
//oclMat &operator = (const Scalar &s)
//{
// //cout << "cv::ocl::oclMat::=" << endl;
// setTo(s);
// return *this;
//}
void set_to_withoutmask_run_cus(const oclMat &dst, const Scalar &scalar, string kernelName)
{
MultiplyScalar op(val, scale);
//if(src1.channels() == 1 && dst.channels() == 1)
//{
// callT(src1, dst, op, 0);
//}
//else
//{
callF(src1, dst, op, 0);
//}
vector<pair<size_t , const void *> > args;
size_t localThreads[3] = {16, 16, 1};
size_t globalThreads[3];
globalThreads[0] = (dst.cols + localThreads[0] - 1) / localThreads[0] * localThreads[0];
globalThreads[1] = (dst.rows + localThreads[1] - 1) / localThreads[1] * localThreads[1];
globalThreads[2] = 1;
int step_in_pixel = dst.step / dst.elemSize(), offset_in_pixel = dst.offset / dst.elemSize();
if(dst.type() == CV_8UC1)
{
globalThreads[0] = ((dst.cols + 4) / 4 + localThreads[0] - 1) / localThreads[0] * localThreads[0];
}
char compile_option[32];
union sc
{
cl_uchar4 uval;
cl_char4 cval;
cl_ushort4 usval;
cl_short4 shval;
cl_int4 ival;
cl_float4 fval;
cl_double4 dval;
}val;
switch(dst.depth())
{
case 0:
val.uval.s[0] = saturate_cast<uchar>(scalar.val[0]);
val.uval.s[1] = saturate_cast<uchar>(scalar.val[1]);
val.uval.s[2] = saturate_cast<uchar>(scalar.val[2]);
val.uval.s[3] = saturate_cast<uchar>(scalar.val[3]);
switch(dst.channels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=uchar");
args.push_back( make_pair( sizeof(cl_uchar) , (void *)&val.uval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=uchar4");
args.push_back( make_pair( sizeof(cl_uchar4) , (void *)&val.uval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat,"unsupported channels");
}
break;
case 1:
val.cval.s[0] = saturate_cast<char>(scalar.val[0]);
val.cval.s[1] = saturate_cast<char>(scalar.val[1]);
val.cval.s[2] = saturate_cast<char>(scalar.val[2]);
val.cval.s[3] = saturate_cast<char>(scalar.val[3]);
switch(dst.channels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=char");
args.push_back( make_pair( sizeof(cl_char) , (void *)&val.cval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=char4");
args.push_back( make_pair( sizeof(cl_char4) , (void *)&val.cval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat,"unsupported channels");
}
break;
case 2:
val.usval.s[0] = saturate_cast<ushort>(scalar.val[0]);
val.usval.s[1] = saturate_cast<ushort>(scalar.val[1]);
val.usval.s[2] = saturate_cast<ushort>(scalar.val[2]);
val.usval.s[3] = saturate_cast<ushort>(scalar.val[3]);
switch(dst.channels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=ushort");
args.push_back( make_pair( sizeof(cl_ushort) , (void *)&val.usval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=ushort4");
args.push_back( make_pair( sizeof(cl_ushort4) , (void *)&val.usval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat,"unsupported channels");
}
break;
case 3:
val.shval.s[0] = saturate_cast<short>(scalar.val[0]);
val.shval.s[1] = saturate_cast<short>(scalar.val[1]);
val.shval.s[2] = saturate_cast<short>(scalar.val[2]);
val.shval.s[3] = saturate_cast<short>(scalar.val[3]);
switch(dst.channels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=short");
args.push_back( make_pair( sizeof(cl_short) , (void *)&val.shval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=short4");
args.push_back( make_pair( sizeof(cl_short4) , (void *)&val.shval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat,"unsupported channels");
}
break;
case 4:
val.ival.s[0] = saturate_cast<int>(scalar.val[0]);
val.ival.s[1] = saturate_cast<int>(scalar.val[1]);
val.ival.s[2] = saturate_cast<int>(scalar.val[2]);
val.ival.s[3] = saturate_cast<int>(scalar.val[3]);
switch(dst.channels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=int");
args.push_back( make_pair( sizeof(cl_int) , (void *)&val.ival.s[0] ));
break;
case 2:
sprintf(compile_option, "-D GENTYPE=int2");
cl_int2 i2val;
i2val.s[0] = val.ival.s[0];
i2val.s[1] = val.ival.s[1];
args.push_back( make_pair( sizeof(cl_int2) , (void *)&i2val ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=int4");
args.push_back( make_pair( sizeof(cl_int4) , (void *)&val.ival ));
break;
default:
CV_Error(CV_StsUnsupportedFormat,"unsupported channels");
}
break;
case 5:
val.fval.s[0] = (float)scalar.val[0];
val.fval.s[1] = (float)scalar.val[1];
val.fval.s[2] = (float)scalar.val[2];
val.fval.s[3] = (float)scalar.val[3];
switch(dst.channels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=float");
args.push_back( make_pair( sizeof(cl_float) , (void *)&val.fval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=float4");
args.push_back( make_pair( sizeof(cl_float4) , (void *)&val.fval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat,"unsupported channels");
}
break;
case 6:
val.dval.s[0] = scalar.val[0];
val.dval.s[1] = scalar.val[1];
val.dval.s[2] = scalar.val[2];
val.dval.s[3] = scalar.val[3];
switch(dst.channels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=double");
args.push_back( make_pair( sizeof(cl_double) , (void *)&val.dval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=double4");
args.push_back( make_pair( sizeof(cl_double4) , (void *)&val.dval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat,"unsupported channels");
}
break;
default:
CV_Error(CV_StsUnsupportedFormat,"unknown depth");
}
#if CL_VERSION_1_2
if(dst.offset==0 && dst.cols==dst.wholecols)
{
clEnqueueFillBuffer(dst.clCxt->impl->clCmdQueue,(cl_mem)dst.data,args[0].second,args[0].first,0,dst.step*dst.rows,0,NULL,NULL);
}
else
{
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.cols ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.rows ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&step_in_pixel ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&offset_in_pixel));
openCLExecuteKernel2(dst.clCxt , &operator_setTo, kernelName, globalThreads,
localThreads, args, -1, -1,compile_option, CLFLUSH);
}
#else
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.cols ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.rows ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&step_in_pixel ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&offset_in_pixel));
openCLExecuteKernel2(dst.clCxt , &operator_setTo, kernelName, globalThreads,
localThreads, args, -1, -1,compile_option, CLFLUSH);
#endif
}
oclMat &setTo(oclMat &src, const Scalar &scalar)
{
CV_Assert( src.depth() >= 0 && src.depth() <= 6 );
CV_DbgAssert( !src.empty());
if(src.type()==CV_8UC1)
{
set_to_withoutmask_run_cus(src, scalar, "set_to_without_mask_C1_D0");
}
else
{
set_to_withoutmask_run_cus(src, scalar, "set_to_without_mask");
}
return src;
}
void arithmetic_run(const oclMat &src1, oclMat &dst, string kernelName, const char **kernelString, void *_scalar)
{
if(src1.clCxt -> impl -> double_support ==0 && src1.type() == CV_64F)
{
CV_Error(CV_GpuNotSupported,"Selected device don't support double\r\n");
return;
}
//dst.create(src1.size(), src1.type());
//CV_Assert(src1.cols == src2.cols && src2.cols == dst.cols &&
// src1.rows == src2.rows && src2.rows == dst.rows);
CV_Assert(src1.cols == dst.cols &&
src1.rows == dst.rows);
CV_Assert(src1.type() == dst.type());
CV_Assert(src1.depth() != CV_8S);
Context *clCxt = src1.clCxt;
//int channels = dst.channels();
//int depth = dst.depth();
//int vector_lengths[4][7] = {{4, 0, 4, 4, 1, 1, 1},
// {4, 0, 4, 4, 1, 1, 1},
// {4, 0, 4, 4, 1, 1, 1},
// {4, 0, 4, 4, 1, 1, 1}
//};
//size_t vector_length = vector_lengths[channels-1][depth];
//int offset_cols = (dst.offset / dst.elemSize1()) & (vector_length - 1);
//int cols = divUp(dst.cols * channels + offset_cols, vector_length);
size_t localThreads[3] = { 16, 16, 1 };
//size_t globalThreads[3] = { divUp(cols, localThreads[0]) * localThreads[0],
// divUp(dst.rows, localThreads[1]) * localThreads[1],
// 1
// };
size_t globalThreads[3] = { src1.cols,
src1.rows,
1
};
int dst_step1 = dst.cols * dst.elemSize();
vector<pair<size_t , const void *> > args;
args.push_back( make_pair( sizeof(cl_mem), (void *)&src1.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src1.step ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src1.offset ));
//args.push_back( make_pair( sizeof(cl_mem), (void *)&src2.data ));
//args.push_back( make_pair( sizeof(cl_int), (void *)&src2.step ));
//args.push_back( make_pair( sizeof(cl_int), (void *)&src2.offset ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&dst.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&dst.step ));
args.push_back( make_pair( sizeof(cl_int), (void *)&dst.offset ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src1.rows ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src1.cols ));
args.push_back( make_pair( sizeof(cl_int), (void *)&dst_step1 ));
//if(_scalar != NULL)
//{
float scalar1 = *((float *)_scalar);
args.push_back( make_pair( sizeof(float), (float *)&scalar1 ));
//}
openCLExecuteKernel2(clCxt, kernelString, kernelName, globalThreads, localThreads, args, -1, src1.depth(), CLFLUSH);
}
void multiply_cus(const oclMat &src1, oclMat &dst, float scalar)
{
arithmetic_run(src1, dst, "arithm_muls", &pyrlk, (void *)(&scalar));
}
void pyrdown_run_cus(const oclMat &src, const oclMat &dst)
{
CV_Assert(src.type() == dst.type());
CV_Assert(src.depth() != CV_8S);
Context *clCxt = src.clCxt;
string kernelName = "pyrDown";
size_t localThreads[3] = { 256, 1, 1 };
size_t globalThreads[3] = { src.cols, dst.rows, 1};
vector<pair<size_t , const void *> > args;
args.push_back( make_pair( sizeof(cl_mem), (void *)&src.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src.step ));
args.push_back( make_pair( sizeof(cl_int), (void *)&src.rows));
args.push_back( make_pair( sizeof(cl_int), (void *)&src.cols));
args.push_back( make_pair( sizeof(cl_mem), (void *)&dst.data ));
args.push_back( make_pair( sizeof(cl_int), (void *)&dst.step ));
args.push_back( make_pair( sizeof(cl_int), (void *)&dst.cols));
openCLExecuteKernel2(clCxt, &pyr_down, kernelName, globalThreads, localThreads, args, src.channels(), src.depth(), CLFLUSH);
}
void pyrDown_cus(const oclMat& src, oclMat& dst)
{
CV_Assert(src.depth() <= CV_32F && src.channels() <= 4);
dst.create((src.rows + 1) / 2, (src.cols + 1) / 2, src.type());
pyrdown_run_cus(src, dst);
}
//struct MultiplyScalar
//{
// MultiplyScalar(double val_, double scale_) : val(val_), scale(scale_) {}
// double operator ()(double a) const
// {
// return (scale * a * val);
// }
// const double val;
// const double scale;
//};
//
//void callF(const oclMat& src, oclMat& dst, MultiplyScalar op, int mask)
//{
// Mat srcTemp;
// Mat dstTemp;
// src.download(srcTemp);
// dst.download(dstTemp);
//
// int i;
// int j;
// int k;
// for(i = 0; i < srcTemp.rows; i++)
// {
// for(j = 0; j < srcTemp.cols; j++)
// {
// for(k = 0; k < srcTemp.channels(); k++)
// {
// ((float*)dstTemp.data)[srcTemp.channels() * (i * srcTemp.rows + j) + k] = (float)op(((float*)srcTemp.data)[srcTemp.channels() * (i * srcTemp.rows + j) + k]);
// }
// }
// }
//
// dst = dstTemp;
//}
//
//static inline bool isAligned(const unsigned char* ptr, size_t size)
//{
// return reinterpret_cast<size_t>(ptr) % size == 0;
//}
//
//static inline bool isAligned(size_t step, size_t size)
//{
// return step % size == 0;
//}
//
//void callT(const oclMat& src, oclMat& dst, MultiplyScalar op, int mask)
//{
// if (!isAligned(src.data, 4 * sizeof(double)) || !isAligned(src.step, 4 * sizeof(double)) ||
// !isAligned(dst.data, 4 * sizeof(double)) || !isAligned(dst.step, 4 * sizeof(double)))
// {
// callF(src, dst, op, mask);
// return;
// }
//
// Mat srcTemp;
// Mat dstTemp;
// src.download(srcTemp);
// dst.download(dstTemp);
//
// int x_shifted;
//
// int i;
// int j;
// for(i = 0; i < srcTemp.rows; i++)
// {
// const double* srcRow = (const double*)srcTemp.data + i * srcTemp.rows;
// double* dstRow = (double*)dstTemp.data + i * dstTemp.rows;;
//
// for(j = 0; j < srcTemp.cols; j++)
// {
// x_shifted = j * 4;
//
// if(x_shifted + 4 - 1 < srcTemp.cols)
// {
// dstRow[x_shifted ] = op(srcRow[x_shifted ]);
// dstRow[x_shifted + 1] = op(srcRow[x_shifted + 1]);
// dstRow[x_shifted + 2] = op(srcRow[x_shifted + 2]);
// dstRow[x_shifted + 3] = op(srcRow[x_shifted + 3]);
// }
// else
// {
// for (int real_x = x_shifted; real_x < srcTemp.cols; ++real_x)
// {
// ((float*)dstTemp.data)[i * srcTemp.rows + real_x] = op(((float*)srcTemp.data)[i * srcTemp.rows + real_x]);
// }
// }
// }
// }
//}
//
//void multiply(const oclMat& src1, double val, oclMat& dst, double scale = 1.0f);
//void multiply(const oclMat& src1, double val, oclMat& dst, double scale)
//{
// MultiplyScalar op(val, scale);
// //if(src1.channels() == 1 && dst.channels() == 1)
// //{
// // callT(src1, dst, op, 0);
// //}
// //else
// //{
// callF(src1, dst, op, 0);
// //}
//}
cl_mem bindTexture(const oclMat& mat, int depth, int channels)
{
cl_mem texture;
@@ -331,7 +634,7 @@ cl_mem bindTexture(const oclMat& mat, int depth, int channels)
#if CL_VERSION_1_2
cl_image_desc desc;
desc.image_type = CL_MEM_OBJECT_IMAGE2D;
desc.image_width = mat.cols;
desc.image_width = mat.step / mat.elemSize();
desc.image_height = mat.rows;
desc.image_depth = NULL;
desc.image_array_size = 1;
@@ -346,30 +649,35 @@ cl_mem bindTexture(const oclMat& mat, int depth, int channels)
mat.clCxt->impl->clContext,
CL_MEM_READ_WRITE,
&format,
mat.cols,
mat.step / mat.elemSize(),
mat.rows,
0,
NULL,
&err);
#endif
size_t origin[] = { 0, 0, 0 };
size_t region[] = { mat.cols, mat.rows, 1 };
size_t region[] = { mat.step / mat.elemSize(), mat.rows, 1 };
clEnqueueCopyBufferToImage(mat.clCxt->impl->clCmdQueue, (cl_mem)mat.data, texture, 0, origin, region, 0, NULL, 0);
openCLSafeCall(err);
return texture;
}
void releaseTexture(cl_mem texture)
{
openCLFree(texture);
}
void lkSparse_run(oclMat& I, oclMat& J,
const oclMat& prevPts, oclMat& nextPts, oclMat& status, oclMat* err, bool GET_MIN_EIGENVALS, int ptcount,
int level, dim3 block, dim3 patch, Size winSize, int iters)
int level, /*dim3 block, */dim3 patch, Size winSize, int iters)
{
Context *clCxt = I.clCxt;
string kernelName = "lkSparse";
size_t localThreads[3] = { 16, 16, 1 };
size_t globalThreads[3] = { 16 * ptcount, 16, 1};
size_t localThreads[3] = { 8, 32, 1 };
size_t globalThreads[3] = { 8 * ptcount, 32, 1};
int cn = I.channels();
@@ -410,7 +718,10 @@ void lkSparse_run(oclMat& I, oclMat& J,
args.push_back( make_pair( sizeof(cl_char), (void *)&calcErr ));
args.push_back( make_pair( sizeof(cl_char), (void *)&GET_MIN_EIGENVALS ));
openCLExecuteKernel(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args, I.channels(), I.depth());
openCLExecuteKernel2(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args, I.channels(), I.depth(), CLFLUSH);
releaseTexture(ITex);
releaseTexture(JTex);
}
void cv::ocl::PyrLKOpticalFlow::sparse(const oclMat& prevImg, const oclMat& nextImg, const oclMat& prevPts, oclMat& nextPts, oclMat& status, oclMat* err)
@@ -446,14 +757,15 @@ void cv::ocl::PyrLKOpticalFlow::sparse(const oclMat& prevImg, const oclMat& next
oclMat temp1 = (useInitialFlow ? nextPts : prevPts).reshape(1);
oclMat temp2 = nextPts.reshape(1);
//oclMat scalar(temp1.rows, temp1.cols, temp1.type(), Scalar(1.0f / (1 << maxLevel) / 2.0f));
//ocl::multiply(temp1, scalar, temp2);
::multiply(temp1, 1.0f / (1 << maxLevel) / 2.0f, temp2);
multiply_cus(temp1, temp2, 1.0f / (1 << maxLevel) / 2.0f);
//::multiply(temp1, 1.0f / (1 << maxLevel) / 2.0f, temp2);
ensureSizeIsEnough(1, prevPts.cols, CV_8UC1, status);
status.setTo(Scalar::all(1));
//status.setTo(Scalar::all(1));
setTo(status, Scalar::all(1));
if (err)
ensureSizeIsEnough(1, prevPts.cols, CV_32FC1, *err);
//if (err)
// ensureSizeIsEnough(1, prevPts.cols, CV_32FC1, *err);
// build the image pyramids.
@@ -462,23 +774,25 @@ void cv::ocl::PyrLKOpticalFlow::sparse(const oclMat& prevImg, const oclMat& next
if (cn == 1 || cn == 4)
{
prevImg.convertTo(prevPyr_[0], CV_32F);
nextImg.convertTo(nextPyr_[0], CV_32F);
//prevImg.convertTo(prevPyr_[0], CV_32F);
//nextImg.convertTo(nextPyr_[0], CV_32F);
convertTo(prevImg, prevPyr_[0], CV_32F);
convertTo(nextImg, nextPyr_[0], CV_32F);
}
else
{
oclMat buf_;
cvtColor(prevImg, buf_, COLOR_BGR2BGRA);
buf_.convertTo(prevPyr_[0], CV_32F);
//oclMat buf_;
// cvtColor(prevImg, buf_, COLOR_BGR2BGRA);
// buf_.convertTo(prevPyr_[0], CV_32F);
cvtColor(nextImg, buf_, COLOR_BGR2BGRA);
buf_.convertTo(nextPyr_[0], CV_32F);
// cvtColor(nextImg, buf_, COLOR_BGR2BGRA);
// buf_.convertTo(nextPyr_[0], CV_32F);
}
for (int level = 1; level <= maxLevel; ++level)
{
pyrDown(prevPyr_[level - 1], prevPyr_[level]);
pyrDown(nextPyr_[level - 1], nextPyr_[level]);
pyrDown_cus(prevPyr_[level - 1], prevPyr_[level]);
pyrDown_cus(nextPyr_[level - 1], nextPyr_[level]);
}
// dI/dx ~ Ix, dI/dy ~ Iy
@@ -487,8 +801,10 @@ void cv::ocl::PyrLKOpticalFlow::sparse(const oclMat& prevImg, const oclMat& next
{
lkSparse_run(prevPyr_[level], nextPyr_[level],
prevPts, nextPts, status, level == 0 && err ? err : 0, getMinEigenVals, prevPts.cols,
level, block, patch, winSize, iters);
level, /*block, */patch, winSize, iters);
}
clFinish(prevImg.clCxt->impl->clCmdQueue);
}
void lkDense_run(oclMat& I, oclMat& J, oclMat& u, oclMat& v,
@@ -516,10 +832,10 @@ void lkDense_run(oclMat& I, oclMat& J, oclMat& u, oclMat& v,
cl_mem ITex = bindTexture(I, I.depth(), cn);
cl_mem JTex = bindTexture(J, J.depth(), cn);
int2 halfWin = {(winSize.width - 1) / 2, (winSize.height - 1) / 2};
const int patchWidth = 16 + 2 * halfWin.x;
const int patchHeight = 16 + 2 * halfWin.y;
size_t smem_size = 3 * patchWidth * patchHeight * sizeof(int);
//int2 halfWin = {(winSize.width - 1) / 2, (winSize.height - 1) / 2};
//const int patchWidth = 16 + 2 * halfWin.x;
//const int patchHeight = 16 + 2 * halfWin.y;
//size_t smem_size = 3 * patchWidth * patchHeight * sizeof(int);
vector<pair<size_t , const void *> > args;
@@ -543,7 +859,10 @@ void lkDense_run(oclMat& I, oclMat& J, oclMat& u, oclMat& v,
args.push_back( make_pair( sizeof(cl_int), (void *)&iters ));
args.push_back( make_pair( sizeof(cl_char), (void *)&calcErr ));
openCLExecuteKernel(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args, I.channels(), I.depth());
openCLExecuteKernel2(clCxt, &pyrlk, kernelName, globalThreads, localThreads, args, I.channels(), I.depth(), CLFLUSH);
releaseTexture(ITex);
releaseTexture(JTex);
}
void cv::ocl::PyrLKOpticalFlow::dense(const oclMat& prevImg, const oclMat& nextImg, oclMat& u, oclMat& v, oclMat* err)