Skip to content

O
opencv
Commit fb73e7b4 authored by Ilya Lavrenov's avatar Ilya Lavrenov
Browse files
Options

fixed pyrlk optical flow

parent cca2e5d8
Hide whitespace changes
Inline Side-by-side
Showing with 20 additions and 396 deletions
modules/ocl/src/pyrlk.cpp
View file @ fb73e7b4
......@@ -45,23 +45,21 @@
//
//M*/
#include "precomp.hpp"
using namespace cv;
using namespace cv::ocl;
namespace cv
{
namespace ocl
{
extern const char *pyrlk;
extern const char *pyrlk_no_image;
extern const char *operator_setTo;
extern const char *operator_convertTo;
extern const char *operator_copyToM;
extern const char *pyr_down;
}
namespace ocl
{
extern const char *pyrlk;
extern const char *pyrlk_no_image;
extern const char *pyr_down;
}
}
struct dim3
{
unsigned int x, y, z;
......@@ -88,345 +86,6 @@ static void calcPatchSize(cv::Size winSize, int cn, dim3 &block, dim3 &patch, bo
block.z = patch.z = 1;
}
///////////////////////////////////////////////////////////////////////////
//////////////////////////////// ConvertTo ////////////////////////////////
///////////////////////////////////////////////////////////////////////////
static void convert_run_cus(const oclMat &src, oclMat &dst, double alpha, double beta)
{
String kernelName = "convert_to_S";
std::stringstream idxStr;
idxStr << src.depth();
kernelName = kernelName + idxStr.str().c_str();
float alpha_f = (float)alpha, beta_f = (float)beta;
CV_DbgAssert(src.rows == dst.rows && src.cols == dst.cols);
std::vector<std::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)
{
globalThreads[0] = ((dst.cols + 4) / 4 + localThreads[0]) / localThreads[0] * localThreads[0];
}
args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&src.data ));
args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&src.cols ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&src.rows ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&srcstep_in_pixel ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&srcoffset_in_pixel ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dststep_in_pixel ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dstoffset_in_pixel ));
args.push_back( std::make_pair( sizeof(cl_float) , (void *)&alpha_f ));
args.push_back( std::make_pair( sizeof(cl_float) , (void *)&beta_f ));
openCLExecuteKernel2(dst.clCxt , &operator_convertTo, kernelName, globalThreads,
localThreads, args, dst.oclchannels(), dst.depth(), CLFLUSH);
}
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 )
{
//cout << "cv::ocl::oclMat::convertTo()" << endl;
bool noScale = fabs(alpha - 1) < std::numeric_limits<double>::epsilon()
&& fabs(beta) < std::numeric_limits<double>::epsilon();
if( rtype < 0 )
rtype = src.type();
else
rtype = CV_MAKETYPE(CV_MAT_DEPTH(rtype), src.oclchannels());
int sdepth = src.depth(), ddepth = CV_MAT_DEPTH(rtype);
if( sdepth == ddepth && noScale )
{
src.copyTo(dst);
return;
}
oclMat temp;
const oclMat *psrc = &src;
if( sdepth != ddepth && psrc == &dst )
psrc = &(temp = src);
dst.create( src.size(), rtype );
convert_run_cus(*psrc, dst, alpha, beta);
}
///////////////////////////////////////////////////////////////////////////
//////////////////////////////// setTo ////////////////////////////////////
///////////////////////////////////////////////////////////////////////////
//oclMat &operator = (const Scalar &s)
//{
// //cout << "cv::ocl::oclMat::=" << endl;
// setTo(s);
// return *this;
//}
static void set_to_withoutmask_run_cus(const oclMat &dst, const Scalar &scalar, String kernelName)
{
std::vector<std::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.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=uchar");
args.push_back( std::make_pair( sizeof(cl_uchar) , (void *)&val.uval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=uchar4");
args.push_back( std::make_pair( sizeof(cl_uchar4) , (void *)&val.uval ));
break;
default:
CV_Error(Error::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.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=char");
args.push_back( std::make_pair( sizeof(cl_char) , (void *)&val.cval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=char4");
args.push_back( std::make_pair( sizeof(cl_char4) , (void *)&val.cval ));
break;
default:
CV_Error(Error::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.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=ushort");
args.push_back( std::make_pair( sizeof(cl_ushort) , (void *)&val.usval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=ushort4");
args.push_back( std::make_pair( sizeof(cl_ushort4) , (void *)&val.usval ));
break;
default:
CV_Error(Error::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.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=short");
args.push_back( std::make_pair( sizeof(cl_short) , (void *)&val.shval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=short4");
args.push_back( std::make_pair( sizeof(cl_short4) , (void *)&val.shval ));
break;
default:
CV_Error(Error::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.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=int");
args.push_back( std::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( std::make_pair( sizeof(cl_int2) , (void *)&i2val ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=int4");
args.push_back( std::make_pair( sizeof(cl_int4) , (void *)&val.ival ));
break;
default:
CV_Error(Error::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.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=float");
args.push_back( std::make_pair( sizeof(cl_float) , (void *)&val.fval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=float4");
args.push_back( std::make_pair( sizeof(cl_float4) , (void *)&val.fval ));
break;
default:
CV_Error(Error::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.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=double");
args.push_back( std::make_pair( sizeof(cl_double) , (void *)&val.dval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=double4");
args.push_back( std::make_pair( sizeof(cl_double4) , (void *)&val.dval ));
break;
default:
CV_Error(Error::StsUnsupportedFormat, "unsupported channels");
}
break;
default:
CV_Error(Error::StsUnsupportedFormat, "unknown depth");
}
#ifdef CL_VERSION_1_2
if(dst.offset == 0 && dst.cols == dst.wholecols)
{
clEnqueueFillBuffer((cl_command_queue)dst.clCxt->oclCommandQueue(), (cl_mem)dst.data, args[0].second, args[0].first, 0, dst.step * dst.rows, 0, NULL, NULL);
}
else
{
args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dst.cols ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dst.rows ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&step_in_pixel ));
args.push_back( std::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( std::make_pair( sizeof(cl_mem) , (void *)&dst.data ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dst.cols ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dst.rows ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&step_in_pixel ));
args.push_back( std::make_pair( sizeof(cl_int) , (void *)&offset_in_pixel));
openCLExecuteKernel2(dst.clCxt , &operator_setTo, kernelName, globalThreads,
localThreads, args, -1, -1, compile_option, CLFLUSH);
#endif
}
static 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;
}
///////////////////////////////////////////////////////////////////////////
////////////////////////////////// CopyTo /////////////////////////////////
///////////////////////////////////////////////////////////////////////////
// static void copy_to_with_mask_cus(const oclMat &src, oclMat &dst, const oclMat &mask, String kernelName)
// {
// CV_DbgAssert( dst.rows == mask.rows && dst.cols == mask.cols &&
// src.rows == dst.rows && src.cols == dst.cols
// && mask.type() == CV_8UC1);
// std::vector<std::pair<size_t , const void *> > args;
// String string_types[4][7] = {{"uchar", "char", "ushort", "short", "int", "float", "double"},
// {"uchar2", "char2", "ushort2", "short2", "int2", "float2", "double2"},
// {"uchar3", "char3", "ushort3", "short3", "int3", "float3", "double3"},
// {"uchar4", "char4", "ushort4", "short4", "int4", "float4", "double4"}
// };
// char compile_option[32];
// sprintf(compile_option, "-D GENTYPE=%s", string_types[dst.oclchannels() - 1][dst.depth()].c_str());
// size_t localThreads[3] = {16, 16, 1};
// size_t globalThreads[3];
// globalThreads[0] = divUp(dst.cols, localThreads[0]) * localThreads[0];
// globalThreads[1] = divUp(dst.rows, 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();
// args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&src.data ));
// args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&dst.data ));
// args.push_back( std::make_pair( sizeof(cl_mem) , (void *)&mask.data ));
// args.push_back( std::make_pair( sizeof(cl_int) , (void *)&src.cols ));
// args.push_back( std::make_pair( sizeof(cl_int) , (void *)&src.rows ));
// args.push_back( std::make_pair( sizeof(cl_int) , (void *)&srcstep_in_pixel ));
// args.push_back( std::make_pair( sizeof(cl_int) , (void *)&srcoffset_in_pixel ));
// args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dststep_in_pixel ));
// args.push_back( std::make_pair( sizeof(cl_int) , (void *)&dstoffset_in_pixel ));
// args.push_back( std::make_pair( sizeof(cl_int) , (void *)&mask.step ));
// args.push_back( std::make_pair( sizeof(cl_int) , (void *)&mask.offset ));
// openCLExecuteKernel2(dst.clCxt , &operator_copyToM, kernelName, globalThreads,
// localThreads, args, -1, -1, compile_option, CLFLUSH);
// }
static void copyTo(const oclMat &src, oclMat &m )
{
CV_DbgAssert(!src.empty());
m.create(src.size(), src.type());
openCLCopyBuffer2D(src.clCxt, m.data, m.step, m.offset,
src.data, src.step, src.cols * src.elemSize(), src.rows, src.offset);
}
static void pyrdown_run_cus(const oclMat &src, const oclMat &dst)
{
......@@ -457,7 +116,6 @@ static 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);
}
......@@ -516,9 +174,7 @@ static void lkSparse_run(oclMat &I, oclMat &J,
releaseTexture(JTex);
}
else
{
openCLExecuteKernel2(clCxt, &pyrlk_no_image, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
}
}
void cv::ocl::PyrLKOpticalFlow::sparse(const oclMat &prevImg, const oclMat &nextImg, const oclMat &prevPts, oclMat &nextPts, oclMat &status, oclMat *err)
......@@ -527,7 +183,6 @@ void cv::ocl::PyrLKOpticalFlow::sparse(const oclMat &prevImg, const oclMat &next
{
nextPts.release();
status.release();
//if (err) err->release();
return;
}
......@@ -556,8 +211,7 @@ void cv::ocl::PyrLKOpticalFlow::sparse(const oclMat &prevImg, const oclMat &next
multiply(1.0f/(1<<maxLevel)/2.0f, temp1, temp2);
ensureSizeIsEnough(1, prevPts.cols, CV_8UC1, status);
//status.setTo(Scalar::all(1));
setTo(status, Scalar::all(1));
status.setTo(Scalar::all(1));
bool errMat = false;
if (!err)
......@@ -567,7 +221,6 @@ void cv::ocl::PyrLKOpticalFlow::sparse(const oclMat &prevImg, const oclMat &next
}
else
ensureSizeIsEnough(1, prevPts.cols, CV_32FC1, *err);
//ensureSizeIsEnough(1, prevPts.cols, CV_32FC1, err);
// build the image pyramids.
prevPyr_.resize(maxLevel + 1);
......@@ -575,19 +228,8 @@ 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);
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);
// cvtColor(nextImg, buf_, COLOR_BGR2BGRA);
// buf_.convertTo(nextPyr_[0], CV_32F);
prevImg.convertTo(prevPyr_[0], CV_32F);
nextImg.convertTo(nextPyr_[0], CV_32F);
}
for (int level = 1; level <= maxLevel; ++level)
......@@ -646,11 +288,6 @@ static void lkDense_run(oclMat &I, oclMat &J, oclMat &u, oclMat &v,
JTex = (cl_mem)J.data;
}
//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);
std::vector<std::pair<size_t , const void *> > args;
args.push_back( std::make_pair( sizeof(cl_mem), (void *)&ITex ));
......@@ -666,12 +303,10 @@ static void lkDense_run(oclMat &I, oclMat &J, oclMat &u, oclMat &v,
args.push_back( std::make_pair( sizeof(cl_int), (void *)&prevV.step ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&I.rows ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&I.cols ));
//args.push_back( std::make_pair( sizeof(cl_mem), (void *)&(*err).data ));
//args.push_back( std::make_pair( sizeof(cl_int), (void *)&(*err).step ));
if (!isImageSupported)
{
args.push_back( std::make_pair( sizeof(cl_int), (void *)&elemCntPerRow ) );
}
args.push_back( std::make_pair( sizeof(cl_int), (void *)&winSize.width ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&winSize.height ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&iters ));
......@@ -685,10 +320,7 @@ static void lkDense_run(oclMat &I, oclMat &J, oclMat &u, oclMat &v,
releaseTexture(JTex);
}
else
{
//printf("Warning: The image2d_t is not supported by the device. Using alternative method!\n");
openCLExecuteKernel2(clCxt, &pyrlk_no_image, kernelName, globalThreads, localThreads, args, I.oclchannels(), I.depth(), CLFLUSH);
}
}
void cv::ocl::PyrLKOpticalFlow::dense(const oclMat &prevImg, const oclMat &nextImg, oclMat &u, oclMat &v, oclMat *err)
......@@ -705,8 +337,7 @@ void cv::ocl::PyrLKOpticalFlow::dense(const oclMat &prevImg, const oclMat &nextI
nextPyr_.resize(maxLevel + 1);
prevPyr_[0] = prevImg;
//nextImg.convertTo(nextPyr_[0], CV_32F);
convertTo(nextImg, nextPyr_[0], CV_32F);
nextImg.convertTo(nextPyr_[0], CV_32F);
for (int level = 1; level <= maxLevel; ++level)
{
......@@ -718,10 +349,8 @@ void cv::ocl::PyrLKOpticalFlow::dense(const oclMat &prevImg, const oclMat &nextI
ensureSizeIsEnough(prevImg.size(), CV_32FC1, vPyr_[0]);
ensureSizeIsEnough(prevImg.size(), CV_32FC1, uPyr_[1]);
ensureSizeIsEnough(prevImg.size(), CV_32FC1, vPyr_[1]);
//uPyr_[1].setTo(Scalar::all(0));
//vPyr_[1].setTo(Scalar::all(0));
setTo(uPyr_[1], Scalar::all(0));
setTo(vPyr_[1], Scalar::all(0));
uPyr_[1].setTo(Scalar::all(0));
vPyr_[1].setTo(Scalar::all(0));
Size winSize2i(winSize.width, winSize.height);
......@@ -738,10 +367,8 @@ void cv::ocl::PyrLKOpticalFlow::dense(const oclMat &prevImg, const oclMat &nextI
idx = idx2;
}
//uPyr_[idx].copyTo(u);
//vPyr_[idx].copyTo(v);
copyTo(uPyr_[idx], u);
copyTo(vPyr_[idx], v);
uPyr_[idx].copyTo(u);
vPyr_[idx].copyTo(v);
clFinish((cl_command_queue)prevImg.clCxt->oclCommandQueue());
}
modules/ocl/test/test_optflow.cpp
View file @ fb73e7b4
......@@ -251,7 +251,6 @@ TEST_P(Sparse, Mat)
cv::Point2i b = nextPts_gold[i];
bool eq = std::abs(a.x - b.x) < 1 && std::abs(a.y - b.y) < 1;
//float errdiff = std::abs(err[i] - err_gold[i]);
float errdiff = 0.0f;
if (!eq || errdiff > 1e-1)
......@@ -262,12 +261,10 @@ TEST_P(Sparse, Mat)
double bad_ratio = static_cast<double>(mistmatch) / (nextPts.size());
ASSERT_LE(bad_ratio, 0.02f);
}
INSTANTIATE_TEST_CASE_P(OCL_Video, Sparse, Combine(
Values(false, true),
Values(false, true)));
INSTANTIATE_TEST_CASE_P(OCL_Video, Sparse, Combine(Bool(), Bool()));
//////////////////////////////////////////////////////
// FarnebackOpticalFlow
......
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