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

Merge pull request #1606 from ilya-lavrenov:ocl_imgproc

Browse Source
This commit is contained in:
Andrey Pavlenko 2013-10-12 16:10:03 +04:00 committed by OpenCV Buildbot
commit a55cc60136
1 changed files with 185 additions and 357 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

408
modules/ocl/src/imgproc.cpp
View File

@ -102,12 +102,8 @@ namespace cv
static void threshold_8u(const oclMat &src, oclMat &dst, double thresh, double maxVal, int type) static void threshold_8u(const oclMat &src, oclMat &dst, double thresh, double maxVal, int type)
{ {
CV_Assert( (src.cols == dst.cols) && (src.rows == dst.rows) );
Context *clCxt = src.clCxt;
uchar thresh_uchar = cvFloor(thresh); uchar thresh_uchar = cvFloor(thresh);
uchar max_val = cvRound(maxVal); uchar max_val = cvRound(maxVal);
string kernelName = "threshold";
size_t cols = (dst.cols + (dst.offset % 16) + 15) / 16; size_t cols = (dst.cols + (dst.offset % 16) + 15) / 16;
size_t bSizeX = 16, bSizeY = 16; size_t bSizeX = 16, bSizeY = 16;
@ -128,14 +124,11 @@ namespace cv
args.push_back( make_pair(sizeof(cl_uchar), (void *)&thresh_uchar)); args.push_back( make_pair(sizeof(cl_uchar), (void *)&thresh_uchar));
args.push_back( make_pair(sizeof(cl_uchar), (void *)&max_val)); args.push_back( make_pair(sizeof(cl_uchar), (void *)&max_val));
args.push_back( make_pair(sizeof(cl_int), (void *)&type)); args.push_back( make_pair(sizeof(cl_int), (void *)&type));
openCLExecuteKernel(clCxt, &imgproc_threshold, kernelName, globalThreads, localThreads, args, src.oclchannels(), src.depth()); openCLExecuteKernel(src.clCxt, &imgproc_threshold, "threshold", globalThreads, localThreads, args, src.oclchannels(), src.depth());
} }
static void threshold_32f(const oclMat &src, oclMat &dst, double thresh, double maxVal, int type) static void threshold_32f(const oclMat &src, oclMat &dst, double thresh, double maxVal, int type)
{ {
CV_Assert( (src.cols == dst.cols) && (src.rows == dst.rows) );
Context *clCxt = src.clCxt;
float thresh_f = thresh; float thresh_f = thresh;
float max_val = maxVal; float max_val = maxVal;
int dst_offset = (dst.offset >> 2); int dst_offset = (dst.offset >> 2);
@ -143,10 +136,7 @@ namespace cv
int src_offset = (src.offset >> 2); int src_offset = (src.offset >> 2);
int src_step = (src.step >> 2); int src_step = (src.step >> 2);
string kernelName = "threshold";
size_t cols = (dst.cols + (dst_offset & 3) + 3) / 4; size_t cols = (dst.cols + (dst_offset & 3) + 3) / 4;
//size_t cols = dst.cols;
size_t bSizeX = 16, bSizeY = 16; size_t bSizeX = 16, bSizeY = 16;
size_t gSizeX = cols % bSizeX == 0 ? cols : (cols + bSizeX - 1) / bSizeX * bSizeX; size_t gSizeX = cols % bSizeX == 0 ? cols : (cols + bSizeX - 1) / bSizeX * bSizeX;
size_t gSizeY = dst.rows; size_t gSizeY = dst.rows;
@ -165,7 +155,8 @@ namespace cv
args.push_back( make_pair(sizeof(cl_float), (void *)&thresh_f)); args.push_back( make_pair(sizeof(cl_float), (void *)&thresh_f));
args.push_back( make_pair(sizeof(cl_float), (void *)&max_val)); args.push_back( make_pair(sizeof(cl_float), (void *)&max_val));
args.push_back( make_pair(sizeof(cl_int), (void *)&type)); args.push_back( make_pair(sizeof(cl_int), (void *)&type));
openCLExecuteKernel(clCxt, &imgproc_threshold, kernelName, globalThreads, localThreads, args, src.oclchannels(), src.depth());
openCLExecuteKernel(src.clCxt, &imgproc_threshold, "threshold", globalThreads, localThreads, args, src.oclchannels(), src.depth());
} }
@ -184,6 +175,7 @@ namespace cv
return thresh; return thresh;
} }
//////////////////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////////////////
/////////////////////////////// remap ////////////////////////////////////////////////// /////////////////////////////// remap //////////////////////////////////////////////////
//////////////////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////////////////
@ -199,7 +191,6 @@ namespace cv
dst.create(map1.size(), src.type()); dst.create(map1.size(), src.type());
string kernelName; string kernelName;
if ( map1.type() == CV_32FC2 && !map2.data ) if ( map1.type() == CV_32FC2 && !map2.data )
@ -240,11 +231,8 @@ namespace cv
glbSizeX = cols % blkSizeX == 0 ? cols : (cols / blkSizeX + 1) * blkSizeX; glbSizeX = cols % blkSizeX == 0 ? cols : (cols / blkSizeX + 1) * blkSizeX;
} }
else else
{
glbSizeX = dst.cols % blkSizeX == 0 ? dst.cols : (dst.cols / blkSizeX + 1) * blkSizeX; glbSizeX = dst.cols % blkSizeX == 0 ? dst.cols : (dst.cols / blkSizeX + 1) * blkSizeX;
}
size_t glbSizeY = dst.rows % blkSizeY == 0 ? dst.rows : (dst.rows / blkSizeY + 1) * blkSizeY; size_t glbSizeY = dst.rows % blkSizeY == 0 ? dst.rows : (dst.rows / blkSizeY + 1) * blkSizeY;
size_t globalThreads[3] = {glbSizeX, glbSizeY, 1}; size_t globalThreads[3] = {glbSizeX, glbSizeY, 1};
size_t localThreads[3] = {blkSizeX, blkSizeY, 1}; size_t localThreads[3] = {blkSizeX, blkSizeY, 1};
@ -271,14 +259,10 @@ namespace cv
args.push_back( make_pair(sizeof(cl_int), (void *)&cols)); args.push_back( make_pair(sizeof(cl_int), (void *)&cols));
if (src.clCxt->supportsFeature(FEATURE_CL_DOUBLE)) if (src.clCxt->supportsFeature(FEATURE_CL_DOUBLE))
{
args.push_back( make_pair(sizeof(cl_double4), (void *)&borderValue)); args.push_back( make_pair(sizeof(cl_double4), (void *)&borderValue));
}
else else
{
args.push_back( make_pair(sizeof(cl_float4), (void *)&borderFloat)); args.push_back( make_pair(sizeof(cl_float4), (void *)&borderFloat));
} }
}
if (map1.channels() == 1) if (map1.channels() == 1)
{ {
args.push_back( make_pair(sizeof(cl_mem), (void *)&dst.data)); args.push_back( make_pair(sizeof(cl_mem), (void *)&dst.data));
@ -299,14 +283,10 @@ namespace cv
args.push_back( make_pair(sizeof(cl_int), (void *)&map1.rows)); args.push_back( make_pair(sizeof(cl_int), (void *)&map1.rows));
args.push_back( make_pair(sizeof(cl_int), (void *)&cols)); args.push_back( make_pair(sizeof(cl_int), (void *)&cols));
if (src.clCxt->supportsFeature(FEATURE_CL_DOUBLE)) if (src.clCxt->supportsFeature(FEATURE_CL_DOUBLE))
{
args.push_back( make_pair(sizeof(cl_double4), (void *)&borderValue)); args.push_back( make_pair(sizeof(cl_double4), (void *)&borderValue));
}
else else
{
args.push_back( make_pair(sizeof(cl_float4), (void *)&borderFloat)); args.push_back( make_pair(sizeof(cl_float4), (void *)&borderFloat));
} }
}
openCLExecuteKernel(clCxt, &imgproc_remap, kernelName, globalThreads, localThreads, args, src.oclchannels(), src.depth()); openCLExecuteKernel(clCxt, &imgproc_remap, kernelName, globalThreads, localThreads, args, src.oclchannels(), src.depth());
} }
@ -325,7 +305,7 @@ namespace cv
int srcoffset_in_pixel = src.offset / src.elemSize(); int srcoffset_in_pixel = src.offset / src.elemSize();
int dstStep_in_pixel = dst.step1() / dst.oclchannels(); int dstStep_in_pixel = dst.step1() / dst.oclchannels();
int dstoffset_in_pixel = dst.offset / dst.elemSize(); int dstoffset_in_pixel = dst.offset / dst.elemSize();
//printf("%d %d\n",src.step1() , dst.elemSize());
string kernelName; string kernelName;
if (interpolation == INTER_LINEAR) if (interpolation == INTER_LINEAR)
kernelName = "resizeLN"; kernelName = "resizeLN";
@ -341,9 +321,8 @@ namespace cv
glbSizeX = cols % blkSizeX == 0 && cols != 0 ? cols : (cols / blkSizeX + 1) * blkSizeX; glbSizeX = cols % blkSizeX == 0 && cols != 0 ? cols : (cols / blkSizeX + 1) * blkSizeX;
} }
else else
{
glbSizeX = dst.cols % blkSizeX == 0 && dst.cols != 0 ? dst.cols : (dst.cols / blkSizeX + 1) * blkSizeX; glbSizeX = dst.cols % blkSizeX == 0 && dst.cols != 0 ? dst.cols : (dst.cols / blkSizeX + 1) * blkSizeX;
}
size_t glbSizeY = dst.rows % blkSizeY == 0 && dst.rows != 0 ? dst.rows : (dst.rows / blkSizeY + 1) * blkSizeY; size_t glbSizeY = dst.rows % blkSizeY == 0 && dst.rows != 0 ? dst.rows : (dst.rows / blkSizeY + 1) * blkSizeY;
size_t globalThreads[3] = {glbSizeX, glbSizeY, 1}; size_t globalThreads[3] = {glbSizeX, glbSizeY, 1};
size_t localThreads[3] = {blkSizeX, blkSizeY, 1}; size_t localThreads[3] = {blkSizeX, blkSizeY, 1};
@ -391,7 +370,6 @@ namespace cv
openCLExecuteKernel(clCxt, &imgproc_resize, kernelName, globalThreads, localThreads, args, src.oclchannels(), src.depth()); openCLExecuteKernel(clCxt, &imgproc_resize, kernelName, globalThreads, localThreads, args, src.oclchannels(), src.depth());
} }
void resize(const oclMat &src, oclMat &dst, Size dsize, void resize(const oclMat &src, oclMat &dst, Size dsize,
double fx, double fy, int interpolation) double fx, double fy, int interpolation)
{ {
@ -402,16 +380,11 @@ namespace cv
CV_Assert( !(dsize == Size()) || (fx > 0 && fy > 0) ); CV_Assert( !(dsize == Size()) || (fx > 0 && fy > 0) );
if (!(dsize == Size()) && (fx > 0 && fy > 0)) if (!(dsize == Size()) && (fx > 0 && fy > 0))
{
if (dsize.width != (int)(src.cols * fx) || dsize.height != (int)(src.rows * fy)) if (dsize.width != (int)(src.cols * fx) || dsize.height != (int)(src.rows * fy))
{
CV_Error(CV_StsUnmatchedSizes, "invalid dsize and fx, fy!"); CV_Error(CV_StsUnmatchedSizes, "invalid dsize and fx, fy!");
}
}
if ( dsize == Size() ) if ( dsize == Size() )
{
dsize = Size(saturate_cast<int>(src.cols * fx), saturate_cast<int>(src.rows * fy)); dsize = Size(saturate_cast<int>(src.cols * fx), saturate_cast<int>(src.rows * fy));
}
else else
{ {
fx = (double)dsize.width / src.cols; fx = (double)dsize.width / src.cols;
@ -425,12 +398,13 @@ namespace cv
resize_gpu( src, dst, fx, fy, interpolation); resize_gpu( src, dst, fx, fy, interpolation);
return; return;
} }
CV_Error(CV_StsUnsupportedFormat, "Non-supported interpolation method"); CV_Error(CV_StsUnsupportedFormat, "Non-supported interpolation method");
} }
//////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////
// medianFilter // medianFilter
void medianFilter(const oclMat &src, oclMat &dst, int m) void medianFilter(const oclMat &src, oclMat &dst, int m)
{ {
CV_Assert( m % 2 == 1 && m > 1 ); CV_Assert( m % 2 == 1 && m > 1 );
@ -450,8 +424,6 @@ namespace cv
int dstOffset = dst.offset / dst.oclchannels() / dst.elemSize1(); int dstOffset = dst.offset / dst.oclchannels() / dst.elemSize1();
Context *clCxt = src.clCxt; Context *clCxt = src.clCxt;
string kernelName = "medianFilter";
vector< pair<size_t, const void *> > args; 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_mem), (void *)&src.data));
@ -477,11 +449,12 @@ namespace cv
openCLExecuteKernel(clCxt, &imgproc_median, kernelName, globalThreads, localThreads, args, src.oclchannels(), src.depth()); openCLExecuteKernel(clCxt, &imgproc_median, kernelName, globalThreads, localThreads, args, src.oclchannels(), src.depth());
} }
else else
CV_Error(CV_StsUnsupportedFormat, "Non-supported filter length"); CV_Error(CV_StsBadArg, "Non-supported filter length");
} }
//////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////
// copyMakeBorder // copyMakeBorder
void copyMakeBorder(const oclMat &src, oclMat &dst, int top, int bottom, int left, int right, int bordertype, const Scalar &scalar) void copyMakeBorder(const oclMat &src, oclMat &dst, int top, int bottom, int left, int right, int bordertype, const Scalar &scalar)
{ {
CV_Assert(top >= 0 && bottom >= 0 && left >= 0 && right >= 0); CV_Assert(top >= 0 && bottom >= 0 && left >= 0 && right >= 0);
@ -491,42 +464,39 @@ namespace cv
(bordertype != cv::BORDER_CONSTANT) && (bordertype != cv::BORDER_CONSTANT) &&
(bordertype != cv::BORDER_REPLICATE)) (bordertype != cv::BORDER_REPLICATE))
{ {
CV_Error(CV_StsBadArg, "unsupported border type"); CV_Error(CV_StsBadArg, "Unsupported border type");
} }
} }
bordertype &= ~cv::BORDER_ISOLATED; bordertype &= ~cv::BORDER_ISOLATED;
if((bordertype == cv::BORDER_REFLECT) || (bordertype == cv::BORDER_WRAP)) if (bordertype == cv::BORDER_REFLECT || bordertype == cv::BORDER_WRAP)
{ {
CV_Assert((src.cols >= left) && (src.cols >= right) && (src.rows >= top) && (src.rows >= bottom)); CV_Assert((src.cols >= left) && (src.cols >= right) && (src.rows >= top) && (src.rows >= bottom));
} }
else if (bordertype == cv::BORDER_REFLECT_101)
if(bordertype == cv::BORDER_REFLECT_101)
{ {
CV_Assert((src.cols > left) && (src.cols > right) && (src.rows > top) && (src.rows > bottom)); CV_Assert((src.cols > left) && (src.cols > right) && (src.rows > top) && (src.rows > bottom));
} }
dst.create(src.rows + top + bottom, src.cols + left + right, src.type()); dst.create(src.rows + top + bottom, src.cols + left + right, src.type());
int srcStep = src.step1() / src.oclchannels(); int srcStep = src.step1() / src.oclchannels(), dstStep = dst.step1() / dst.oclchannels();
int dstStep = dst.step1() / dst.oclchannels(); int srcOffset = src.offset / src.elemSize(), dstOffset = dst.offset / dst.elemSize();
int srcOffset = src.offset / src.elemSize(); int depth = src.depth(), ochannels = src.oclchannels();
int dstOffset = dst.offset / dst.elemSize();
int __bordertype[] = {cv::BORDER_CONSTANT, cv::BORDER_REPLICATE, BORDER_REFLECT, BORDER_WRAP, BORDER_REFLECT_101}; int __bordertype[] = {cv::BORDER_CONSTANT, cv::BORDER_REPLICATE, BORDER_REFLECT, BORDER_WRAP, BORDER_REFLECT_101};
const char *borderstr[] = {"BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", "BORDER_WRAP", "BORDER_REFLECT_101"}; const char *borderstr[] = {"BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", "BORDER_WRAP", "BORDER_REFLECT_101"};
size_t bordertype_index; size_t bordertype_index;
for(bordertype_index = 0; bordertype_index < sizeof(__bordertype) / sizeof(int); bordertype_index++) for(bordertype_index = 0; bordertype_index < sizeof(__bordertype) / sizeof(int); bordertype_index++)
{
if (__bordertype[bordertype_index] == bordertype) if (__bordertype[bordertype_index] == bordertype)
break; break;
}
if (bordertype_index == sizeof(__bordertype) / sizeof(int)) if (bordertype_index == sizeof(__bordertype) / sizeof(int))
{
CV_Error(CV_StsBadArg, "unsupported border type"); CV_Error(CV_StsBadArg, "unsupported border type");
}
string kernelName = "copymakeborder"; string kernelName = "copymakeborder";
size_t localThreads[3] = {16, 16, 1}; size_t localThreads[3] = {16, 16, 1};
size_t globalThreads[3] = {(dst.cols + localThreads[0] - 1) / localThreads[0] *localThreads[0], size_t globalThreads[3] = { dst.cols, dst.rows, 1 };
(dst.rows + localThreads[1] - 1) / localThreads[1] *localThreads[1], 1
};
vector< pair<size_t, const void *> > args; 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_mem), (void *)&src.data));
@ -541,169 +511,30 @@ namespace cv
args.push_back( make_pair( sizeof(cl_int), (void *)&dstOffset)); args.push_back( make_pair( sizeof(cl_int), (void *)&dstOffset));
args.push_back( make_pair( sizeof(cl_int), (void *)&top)); args.push_back( make_pair( sizeof(cl_int), (void *)&top));
args.push_back( make_pair( sizeof(cl_int), (void *)&left)); args.push_back( make_pair( sizeof(cl_int), (void *)&left));
char compile_option[64];
union sc const char * const typeMap[] = { "uchar", "char", "ushort", "short", "int", "float", "double" };
{ const char * const channelMap[] = { "", "", "2", "4", "4" };
cl_uchar4 uval; std::string buildOptions = format("-D GENTYPE=%s%s -D %s",
cl_char4 cval; typeMap[depth], channelMap[ochannels],
cl_ushort4 usval; borderstr[bordertype_index]);
cl_short4 shval;
cl_int4 ival; if (src.type() == CV_8UC1 && (dst.offset & 3) == 0 && (dst.cols & 3) == 0)
cl_float4 fval;
cl_double4 dval;
} val;
switch(dst.depth())
{
case CV_8U:
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 -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_uchar) , (void *)&val.uval.s[0] ));
if(((dst.offset & 3) == 0) && ((dst.cols & 3) == 0))
{ {
kernelName = "copymakeborder_C1_D0"; kernelName = "copymakeborder_C1_D0";
globalThreads[0] = (dst.cols / 4 + localThreads[0] - 1) / localThreads[0] * localThreads[0]; globalThreads[0] = dst.cols >> 2;
}
break;
case 4:
sprintf(compile_option, "-D GENTYPE=uchar4 -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_uchar4) , (void *)&val.uval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case CV_8S:
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 -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_char) , (void *)&val.cval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=char4 -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_char4) , (void *)&val.cval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case CV_16U:
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 -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_ushort) , (void *)&val.usval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=ushort4 -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_ushort4) , (void *)&val.usval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case CV_16S:
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 -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_short) , (void *)&val.shval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=short4 -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_short4) , (void *)&val.shval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case CV_32S:
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 -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_int) , (void *)&val.ival.s[0] ));
break;
case 2:
sprintf(compile_option, "-D GENTYPE=int2 -D %s", borderstr[bordertype_index]);
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 -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_int4) , (void *)&val.ival ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case CV_32F:
val.fval.s[0] = scalar.val[0];
val.fval.s[1] = scalar.val[1];
val.fval.s[2] = scalar.val[2];
val.fval.s[3] = scalar.val[3];
switch(dst.oclchannels())
{
case 1:
sprintf(compile_option, "-D GENTYPE=float -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_float) , (void *)&val.fval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=float4 -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_float4) , (void *)&val.fval ));
break;
default:
CV_Error(CV_StsUnsupportedFormat, "unsupported channels");
}
break;
case CV_64F:
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 -D %s", borderstr[bordertype_index]);
args.push_back( make_pair( sizeof(cl_double) , (void *)&val.dval.s[0] ));
break;
case 4:
sprintf(compile_option, "-D GENTYPE=double4 -D %s", borderstr[bordertype_index]);
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");
} }
openCLExecuteKernel(src.clCxt, &imgproc_copymakeboder, kernelName, globalThreads, localThreads, args, -1, -1, compile_option); int cn = src.channels(), ocn = src.oclchannels();
int bufSize = src.elemSize1() * ocn;
AutoBuffer<uchar> _buf(bufSize);
uchar * buf = (uchar *)_buf;
scalarToRawData(scalar, buf, dst.type());
memset(buf + src.elemSize1() * cn, 0, (ocn - cn) * src.elemSize1());
args.push_back( make_pair( bufSize , (void *)buf ));
openCLExecuteKernel(src.clCxt, &imgproc_copymakeboder, kernelName, globalThreads,
localThreads, args, -1, -1, buildOptions.c_str());
} }
//////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////
@ -782,31 +613,31 @@ namespace cv
string s[3] = {"NN", "Linear", "Cubic"}; string s[3] = {"NN", "Linear", "Cubic"};
string kernelName = "warpAffine" + s[interpolation]; string kernelName = "warpAffine" + s[interpolation];
if (src.clCxt->supportsFeature(FEATURE_CL_DOUBLE)) if (src.clCxt->supportsFeature(FEATURE_CL_DOUBLE))
{ {
cl_int st; cl_int st;
coeffs_cm = clCreateBuffer(*(cl_context*)clCxt->getOpenCLContextPtr(), CL_MEM_READ_WRITE, sizeof(F) * 2 * 3, NULL, &st ); coeffs_cm = clCreateBuffer(*(cl_context*)clCxt->getOpenCLContextPtr(), CL_MEM_READ_WRITE, sizeof(F) * 2 * 3, NULL, &st );
openCLVerifyCall(st); openCLVerifyCall(st);
openCLSafeCall(clEnqueueWriteBuffer(*(cl_command_queue*)clCxt->getOpenCLCommandQueuePtr(), (cl_mem)coeffs_cm, 1, 0, sizeof(F) * 2 * 3, coeffs, 0, 0, 0)); openCLSafeCall(clEnqueueWriteBuffer(*(cl_command_queue*)clCxt->getOpenCLCommandQueuePtr(), (cl_mem)coeffs_cm, 1, 0,
sizeof(F) * 2 * 3, coeffs, 0, 0, 0));
} }
else else
{ {
cl_int st; cl_int st;
for(int m = 0; m < 2; m++) for(int m = 0; m < 2; m++)
for(int n = 0; n < 3; n++) for(int n = 0; n < 3; n++)
{
float_coeffs[m][n] = coeffs[m][n]; float_coeffs[m][n] = coeffs[m][n];
}
coeffs_cm = clCreateBuffer(*(cl_context*)clCxt->getOpenCLContextPtr(), CL_MEM_READ_WRITE, sizeof(float) * 2 * 3, NULL, &st ); coeffs_cm = clCreateBuffer(*(cl_context*)clCxt->getOpenCLContextPtr(), CL_MEM_READ_WRITE, sizeof(float) * 2 * 3, NULL, &st );
openCLSafeCall(clEnqueueWriteBuffer(*(cl_command_queue*)clCxt->getOpenCLCommandQueuePtr(), (cl_mem)coeffs_cm, 1, 0, sizeof(float) * 2 * 3, float_coeffs, 0, 0, 0)); openCLSafeCall(clEnqueueWriteBuffer(*(cl_command_queue*)clCxt->getOpenCLCommandQueuePtr(), (cl_mem)coeffs_cm,
1, 0, sizeof(float) * 2 * 3, float_coeffs, 0, 0, 0));
} }
//TODO: improve this kernel //TODO: improve this kernel
size_t blkSizeX = 16, blkSizeY = 16; size_t blkSizeX = 16, blkSizeY = 16;
size_t glbSizeX; size_t glbSizeX;
size_t cols; size_t cols;
//if(src.type() == CV_8UC1 && interpolation != 2)
if (src.type() == CV_8UC1 && interpolation != 2) if (src.type() == CV_8UC1 && interpolation != 2)
{ {
cols = (dst.cols + dst.offset % 4 + 3) / 4; cols = (dst.cols + dst.offset % 4 + 3) / 4;
@ -817,6 +648,7 @@ namespace cv
cols = dst.cols; cols = dst.cols;
glbSizeX = dst.cols % blkSizeX == 0 ? dst.cols : (dst.cols / blkSizeX + 1) * blkSizeX; glbSizeX = dst.cols % blkSizeX == 0 ? dst.cols : (dst.cols / blkSizeX + 1) * blkSizeX;
} }
size_t glbSizeY = dst.rows % blkSizeY == 0 ? dst.rows : (dst.rows / blkSizeY + 1) * blkSizeY; size_t glbSizeY = dst.rows % blkSizeY == 0 ? dst.rows : (dst.rows / blkSizeY + 1) * blkSizeY;
size_t globalThreads[3] = {glbSizeX, glbSizeY, 1}; size_t globalThreads[3] = {glbSizeX, glbSizeY, 1};
size_t localThreads[3] = {blkSizeX, blkSizeY, 1}; size_t localThreads[3] = {blkSizeX, blkSizeY, 1};
@ -840,7 +672,6 @@ namespace cv
openCLSafeCall(clReleaseMemObject(coeffs_cm)); openCLSafeCall(clReleaseMemObject(coeffs_cm));
} }
void warpPerspective_gpu(const oclMat &src, oclMat &dst, double coeffs[3][3], int interpolation) void warpPerspective_gpu(const oclMat &src, oclMat &dst, double coeffs[3][3], int interpolation)
{ {
CV_Assert( (src.oclchannels() == dst.oclchannels()) ); CV_Assert( (src.oclchannels() == dst.oclchannels()) );
@ -858,7 +689,8 @@ namespace cv
cl_int st; cl_int st;
coeffs_cm = clCreateBuffer(*(cl_context*)clCxt->getOpenCLContextPtr(), CL_MEM_READ_WRITE, sizeof(double) * 3 * 3, NULL, &st ); coeffs_cm = clCreateBuffer(*(cl_context*)clCxt->getOpenCLContextPtr(), CL_MEM_READ_WRITE, sizeof(double) * 3 * 3, NULL, &st );
openCLVerifyCall(st); openCLVerifyCall(st);
openCLSafeCall(clEnqueueWriteBuffer(*(cl_command_queue*)clCxt->getOpenCLCommandQueuePtr(), (cl_mem)coeffs_cm, 1, 0, sizeof(double) * 3 * 3, coeffs, 0, 0, 0)); openCLSafeCall(clEnqueueWriteBuffer(*(cl_command_queue*)clCxt->getOpenCLCommandQueuePtr(), (cl_mem)coeffs_cm, 1, 0,
sizeof(double) * 3 * 3, coeffs, 0, 0, 0));
} }
else else
{ {
@ -869,8 +701,10 @@ namespace cv
coeffs_cm = clCreateBuffer(*(cl_context*)clCxt->getOpenCLContextPtr(), CL_MEM_READ_WRITE, sizeof(float) * 3 * 3, NULL, &st ); coeffs_cm = clCreateBuffer(*(cl_context*)clCxt->getOpenCLContextPtr(), CL_MEM_READ_WRITE, sizeof(float) * 3 * 3, NULL, &st );
openCLVerifyCall(st); openCLVerifyCall(st);
openCLSafeCall(clEnqueueWriteBuffer(*(cl_command_queue*)clCxt->getOpenCLCommandQueuePtr(), (cl_mem)coeffs_cm, 1, 0, sizeof(float) * 3 * 3, float_coeffs, 0, 0, 0)); openCLSafeCall(clEnqueueWriteBuffer(*(cl_command_queue*)clCxt->getOpenCLCommandQueuePtr(), (cl_mem)coeffs_cm, 1, 0,
sizeof(float) * 3 * 3, float_coeffs, 0, 0, 0));
} }
//TODO: improve this kernel //TODO: improve this kernel
size_t blkSizeX = 16, blkSizeY = 16; size_t blkSizeX = 16, blkSizeY = 16;
size_t glbSizeX; size_t glbSizeX;
@ -881,12 +715,11 @@ namespace cv
glbSizeX = cols % blkSizeX == 0 ? cols : (cols / blkSizeX + 1) * blkSizeX; glbSizeX = cols % blkSizeX == 0 ? cols : (cols / blkSizeX + 1) * blkSizeX;
} }
else else
/*
*/
{ {
cols = dst.cols; cols = dst.cols;
glbSizeX = dst.cols % blkSizeX == 0 ? dst.cols : (dst.cols / blkSizeX + 1) * blkSizeX; glbSizeX = dst.cols % blkSizeX == 0 ? dst.cols : (dst.cols / blkSizeX + 1) * blkSizeX;
} }
size_t glbSizeY = dst.rows % blkSizeY == 0 ? dst.rows : (dst.rows / blkSizeY + 1) * blkSizeY; size_t glbSizeY = dst.rows % blkSizeY == 0 ? dst.rows : (dst.rows / blkSizeY + 1) * blkSizeY;
size_t globalThreads[3] = {glbSizeX, glbSizeY, 1}; size_t globalThreads[3] = {glbSizeX, glbSizeY, 1};
size_t localThreads[3] = {blkSizeX, blkSizeY, 1}; size_t localThreads[3] = {blkSizeX, blkSizeY, 1};
@ -929,9 +762,7 @@ namespace cv
Mat coeffsMat(2, 3, CV_64F, (void *)coeffsM); Mat coeffsMat(2, 3, CV_64F, (void *)coeffsM);
M.convertTo(coeffsMat, coeffsMat.type()); M.convertTo(coeffsMat, coeffsMat.type());
if (!warpInd) if (!warpInd)
{
convert_coeffs(coeffsM); convert_coeffs(coeffsM);
}
for(int i = 0; i < 2; ++i) for(int i = 0; i < 2; ++i)
for(int j = 0; j < 3; ++j) for(int j = 0; j < 3; ++j)
@ -959,9 +790,7 @@ namespace cv
Mat coeffsMat(3, 3, CV_64F, (void *)coeffsM); Mat coeffsMat(3, 3, CV_64F, (void *)coeffsM);
M.convertTo(coeffsMat, coeffsMat.type()); M.convertTo(coeffsMat, coeffsMat.type());
if (!warpInd) if (!warpInd)
{
invert(coeffsM); invert(coeffsM);
}
for(int i = 0; i < 3; ++i) for(int i = 0; i < 3; ++i)
for(int j = 0; j < 3; ++j) for(int j = 0; j < 3; ++j)
@ -972,12 +801,13 @@ namespace cv
//////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////
// integral // integral
void integral(const oclMat &src, oclMat &sum, oclMat &sqsum) void integral(const oclMat &src, oclMat &sum, oclMat &sqsum)
{ {
CV_Assert(src.type() == CV_8UC1); CV_Assert(src.type() == CV_8UC1);
if (!src.clCxt->supportsFeature(ocl::FEATURE_CL_DOUBLE) && src.depth() == CV_64F) if (!src.clCxt->supportsFeature(ocl::FEATURE_CL_DOUBLE) && src.depth() == CV_64F)
{ {
CV_Error(CV_OpenCLDoubleNotSupported, "select device don't support double"); CV_Error(CV_OpenCLDoubleNotSupported, "Select device doesn't support double");
return; return;
} }
@ -1071,6 +901,7 @@ namespace cv
} }
/////////////////////// corner ////////////////////////////// /////////////////////// corner //////////////////////////////
static void extractCovData(const oclMat &src, oclMat &Dx, oclMat &Dy, static void extractCovData(const oclMat &src, oclMat &Dx, oclMat &Dy,
int blockSize, int ksize, int borderType) int blockSize, int ksize, int borderType)
{ {
@ -1085,9 +916,8 @@ namespace cv
scale = 1. / scale; scale = 1. / scale;
} }
else else
{
scale = 1. / scale; scale = 1. / scale;
}
if (ksize > 0) if (ksize > 0)
{ {
Sobel(src, Dx, CV_32F, 1, 0, ksize, scale, 0, borderType); Sobel(src, Dx, CV_32F, 1, 0, ksize, scale, 0, borderType);
@ -1120,10 +950,10 @@ namespace cv
sprintf(borderType, "BORDER_REPLICATE"); sprintf(borderType, "BORDER_REPLICATE");
break; break;
default: default:
cout << "BORDER type is not supported!" << endl; CV_Error(CV_StsBadFlag, "BORDER type is not supported!");
} }
char build_options[150];
sprintf(build_options, "-D anX=%d -D anY=%d -D ksX=%d -D ksY=%d -D %s", std::string buildOptions = format("-D anX=%d -D anY=%d -D ksX=%d -D ksY=%d -D %s",
block_size / 2, block_size / 2, block_size, block_size, borderType); block_size / 2, block_size / 2, block_size, block_size, borderType);
size_t blockSizeX = 256, blockSizeY = 1; size_t blockSizeX = 256, blockSizeY = 1;
@ -1153,7 +983,7 @@ namespace cv
args.push_back( make_pair(sizeof(cl_int), (void *)&dst.cols)); args.push_back( make_pair(sizeof(cl_int), (void *)&dst.cols));
args.push_back( make_pair(sizeof(cl_int), (void *)&dst.step)); args.push_back( make_pair(sizeof(cl_int), (void *)&dst.step));
args.push_back( make_pair( sizeof(cl_float) , (void *)&k)); args.push_back( make_pair( sizeof(cl_float) , (void *)&k));
openCLExecuteKernel(dst.clCxt, source, kernelName, gt, lt, args, -1, -1, build_options); openCLExecuteKernel(dst.clCxt, source, kernelName, gt, lt, args, -1, -1, buildOptions.c_str());
} }
void cornerHarris(const oclMat &src, oclMat &dst, int blockSize, int ksize, void cornerHarris(const oclMat &src, oclMat &dst, int blockSize, int ksize,
@ -1168,10 +998,13 @@ namespace cv
{ {
if (!src.clCxt->supportsFeature(FEATURE_CL_DOUBLE) && src.depth() == CV_64F) if (!src.clCxt->supportsFeature(FEATURE_CL_DOUBLE) && src.depth() == CV_64F)
{ {
CV_Error(CV_OpenCLDoubleNotSupported, "select device don't support double"); CV_Error(CV_OpenCLDoubleNotSupported, "Select device doesn't support double");
return;
} }
CV_Assert(src.cols >= blockSize / 2 && src.rows >= blockSize / 2); CV_Assert(src.cols >= blockSize / 2 && src.rows >= blockSize / 2);
CV_Assert(borderType == cv::BORDER_CONSTANT || borderType == cv::BORDER_REFLECT101 || borderType == cv::BORDER_REPLICATE || borderType == cv::BORDER_REFLECT); CV_Assert(borderType == cv::BORDER_CONSTANT || borderType == cv::BORDER_REFLECT101 || borderType == cv::BORDER_REPLICATE
|| borderType == cv::BORDER_REFLECT);
extractCovData(src, dx, dy, blockSize, ksize, borderType); extractCovData(src, dx, dy, blockSize, ksize, borderType);
dst.create(src.size(), CV_32F); dst.create(src.size(), CV_32F);
corner_ocl(&imgproc_calcHarris, "calcHarris", blockSize, static_cast<float>(k), dx, dy, dst, borderType); corner_ocl(&imgproc_calcHarris, "calcHarris", blockSize, static_cast<float>(k), dx, dy, dst, borderType);
@ -1188,19 +1021,23 @@ namespace cv
if (!src.clCxt->supportsFeature(FEATURE_CL_DOUBLE) && src.depth() == CV_64F) if (!src.clCxt->supportsFeature(FEATURE_CL_DOUBLE) && src.depth() == CV_64F)
{ {
CV_Error(CV_OpenCLDoubleNotSupported, "select device don't support double"); CV_Error(CV_OpenCLDoubleNotSupported, "select device don't support double");
return;
} }
CV_Assert(src.cols >= blockSize / 2 && src.rows >= blockSize / 2); CV_Assert(src.cols >= blockSize / 2 && src.rows >= blockSize / 2);
CV_Assert(borderType == cv::BORDER_CONSTANT || borderType == cv::BORDER_REFLECT101 || borderType == cv::BORDER_REPLICATE || borderType == cv::BORDER_REFLECT); CV_Assert(borderType == cv::BORDER_CONSTANT || borderType == cv::BORDER_REFLECT101 || borderType == cv::BORDER_REPLICATE || borderType == cv::BORDER_REFLECT);
extractCovData(src, dx, dy, blockSize, ksize, borderType); extractCovData(src, dx, dy, blockSize, ksize, borderType);
dst.create(src.size(), CV_32F); dst.create(src.size(), CV_32F);
corner_ocl(&imgproc_calcMinEigenVal, "calcMinEigenVal", blockSize, 0, dx, dy, dst, borderType); corner_ocl(&imgproc_calcMinEigenVal, "calcMinEigenVal", blockSize, 0, dx, dy, dst, borderType);
} }
/////////////////////////////////// MeanShiftfiltering /////////////////////////////////////////////// /////////////////////////////////// MeanShiftfiltering ///////////////////////////////////////////////
static void meanShiftFiltering_gpu(const oclMat &src, oclMat dst, int sp, int sr, int maxIter, float eps) static void meanShiftFiltering_gpu(const oclMat &src, oclMat dst, int sp, int sr, int maxIter, float eps)
{ {
CV_Assert( (src.cols == dst.cols) && (src.rows == dst.rows) ); CV_Assert( (src.cols == dst.cols) && (src.rows == dst.rows) );
CV_Assert( !(dst.step & 0x3) ); CV_Assert( !(dst.step & 0x3) );
Context *clCxt = src.clCxt;
//Arrange the NDRange //Arrange the NDRange
int col = src.cols, row = src.rows; int col = src.cols, row = src.rows;
@ -1227,7 +1064,8 @@ namespace cv
args.push_back( make_pair( sizeof(cl_int) , (void *)&sr )); args.push_back( make_pair( sizeof(cl_int) , (void *)&sr ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&maxIter )); args.push_back( make_pair( sizeof(cl_int) , (void *)&maxIter ));
args.push_back( make_pair( sizeof(cl_float) , (void *)&eps )); args.push_back( make_pair( sizeof(cl_float) , (void *)&eps ));
openCLExecuteKernel(clCxt, &meanShift, "meanshift_kernel", globalThreads, localThreads, args, -1, -1);
openCLExecuteKernel(src.clCxt, &meanShift, "meanshift_kernel", globalThreads, localThreads, args, -1, -1);
} }
void meanShiftFiltering(const oclMat &src, oclMat &dst, int sp, int sr, TermCriteria criteria) void meanShiftFiltering(const oclMat &src, oclMat &dst, int sp, int sr, TermCriteria criteria)
@ -1251,7 +1089,6 @@ namespace cv
eps = (float)std::max(criteria.epsilon, 0.0); eps = (float)std::max(criteria.epsilon, 0.0);
meanShiftFiltering_gpu(src, dst, sp, sr, maxIter, eps); meanShiftFiltering_gpu(src, dst, sp, sr, maxIter, eps);
} }
static void meanShiftProc_gpu(const oclMat &src, oclMat dstr, oclMat dstsp, int sp, int sr, int maxIter, float eps) static void meanShiftProc_gpu(const oclMat &src, oclMat dstr, oclMat dstsp, int sp, int sr, int maxIter, float eps)
@ -1260,7 +1097,6 @@ namespace cv
CV_Assert( (src.cols == dstr.cols) && (src.rows == dstr.rows) && CV_Assert( (src.cols == dstr.cols) && (src.rows == dstr.rows) &&
(src.rows == dstsp.rows) && (src.cols == dstsp.cols)); (src.rows == dstsp.rows) && (src.cols == dstsp.cols));
CV_Assert( !(dstsp.step & 0x3) ); CV_Assert( !(dstsp.step & 0x3) );
Context *clCxt = src.clCxt;
//Arrange the NDRange //Arrange the NDRange
int col = src.cols, row = src.rows; int col = src.cols, row = src.rows;
@ -1290,7 +1126,8 @@ namespace cv
args.push_back( make_pair( sizeof(cl_int) , (void *)&sr )); args.push_back( make_pair( sizeof(cl_int) , (void *)&sr ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&maxIter )); args.push_back( make_pair( sizeof(cl_int) , (void *)&maxIter ));
args.push_back( make_pair( sizeof(cl_float) , (void *)&eps )); args.push_back( make_pair( sizeof(cl_float) , (void *)&eps ));
openCLExecuteKernel(clCxt, &meanShift, "meanshiftproc_kernel", globalThreads, localThreads, args, -1, -1);
openCLExecuteKernel(src.clCxt, &meanShift, "meanshiftproc_kernel", globalThreads, localThreads, args, -1, -1);
} }
void meanShiftProc(const oclMat &src, oclMat &dstr, oclMat &dstsp, int sp, int sr, TermCriteria criteria) void meanShiftProc(const oclMat &src, oclMat &dstr, oclMat &dstsp, int sp, int sr, TermCriteria criteria)
@ -1326,6 +1163,7 @@ namespace cv
/////////////////////////////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////////////////////////////
////////////////////////////////////////////////////hist/////////////////////////////////////////////// ////////////////////////////////////////////////////hist///////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////////////////////////// /////////////////////////////////////////////////////////////////////////////////////////////////////
namespace histograms namespace histograms
{ {
const int PARTIAL_HISTOGRAM256_COUNT = 256; const int PARTIAL_HISTOGRAM256_COUNT = 256;
@ -1336,11 +1174,8 @@ namespace cv
{ {
using namespace histograms; using namespace histograms;
Context *clCxt = mat_src.clCxt;
int depth = mat_src.depth(); int depth = mat_src.depth();
string kernelName = "calc_sub_hist";
size_t localThreads[3] = { HISTOGRAM256_BIN_COUNT, 1, 1 }; size_t localThreads[3] = { HISTOGRAM256_BIN_COUNT, 1, 1 };
size_t globalThreads[3] = { PARTIAL_HISTOGRAM256_COUNT *localThreads[0], 1, 1}; size_t globalThreads[3] = { PARTIAL_HISTOGRAM256_COUNT *localThreads[0], 1, 1};
@ -1379,6 +1214,7 @@ namespace cv
src_offset >>= dataWidth_bits; src_offset >>= dataWidth_bits;
int src_step = mat_src.step >> dataWidth_bits; int src_step = mat_src.step >> dataWidth_bits;
int datacount = tempcols * mat_src.rows; int datacount = tempcols * mat_src.rows;
args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_src.data)); args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_src.data));
args.push_back( make_pair( sizeof(cl_int), (void *)&src_step)); args.push_back( make_pair( sizeof(cl_int), (void *)&src_step));
args.push_back( make_pair( sizeof(cl_int), (void *)&src_offset)); args.push_back( make_pair( sizeof(cl_int), (void *)&src_offset));
@ -1388,16 +1224,17 @@ namespace cv
args.push_back( make_pair( sizeof(cl_int), (void *)&inc_x)); args.push_back( make_pair( sizeof(cl_int), (void *)&inc_x));
args.push_back( make_pair( sizeof(cl_int), (void *)&inc_y)); args.push_back( make_pair( sizeof(cl_int), (void *)&inc_y));
args.push_back( make_pair( sizeof(cl_int), (void *)&hist_step)); args.push_back( make_pair( sizeof(cl_int), (void *)&hist_step));
openCLExecuteKernel(clCxt, &imgproc_histogram, kernelName, globalThreads, localThreads, args, -1, depth);
openCLExecuteKernel(mat_src.clCxt, &imgproc_histogram, "calc_sub_hist", globalThreads, localThreads, args, -1, depth);
} }
if (left_col != 0 || right_col != 0) if (left_col != 0 || right_col != 0)
{ {
kernelName = "calc_sub_hist_border";
src_offset = mat_src.offset; src_offset = mat_src.offset;
localThreads[0] = 1; localThreads[0] = 1;
localThreads[1] = 256; localThreads[1] = 256;
globalThreads[0] = left_col + right_col; globalThreads[0] = left_col + right_col;
globalThreads[1] = (mat_src.rows + localThreads[1] - 1) / localThreads[1] * localThreads[1]; globalThreads[1] = mat_src.rows;
args.clear(); args.clear();
args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_src.data)); args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_src.data));
@ -1408,25 +1245,27 @@ namespace cv
args.push_back( make_pair( sizeof(cl_int), (void *)&cols)); args.push_back( make_pair( sizeof(cl_int), (void *)&cols));
args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src.rows)); args.push_back( make_pair( sizeof(cl_int), (void *)&mat_src.rows));
args.push_back( make_pair( sizeof(cl_int), (void *)&hist_step)); args.push_back( make_pair( sizeof(cl_int), (void *)&hist_step));
openCLExecuteKernel(clCxt, &imgproc_histogram, kernelName, globalThreads, localThreads, args, -1, depth);
openCLExecuteKernel(mat_src.clCxt, &imgproc_histogram, "calc_sub_hist_border", globalThreads, localThreads, args, -1, depth);
} }
} }
static void merge_sub_hist(const oclMat &sub_hist, oclMat &mat_hist) static void merge_sub_hist(const oclMat &sub_hist, oclMat &mat_hist)
{ {
using namespace histograms; using namespace histograms;
Context *clCxt = sub_hist.clCxt;
string kernelName = "merge_hist";
size_t localThreads[3] = { 256, 1, 1 }; size_t localThreads[3] = { 256, 1, 1 };
size_t globalThreads[3] = { HISTOGRAM256_BIN_COUNT *localThreads[0], 1, 1}; size_t globalThreads[3] = { HISTOGRAM256_BIN_COUNT *localThreads[0], 1, 1};
int src_step = sub_hist.step >> 2; int src_step = sub_hist.step >> 2;
vector<pair<size_t , const void *> > args; vector<pair<size_t , const void *> > args;
args.push_back( make_pair( sizeof(cl_mem), (void *)&sub_hist.data)); args.push_back( make_pair( sizeof(cl_mem), (void *)&sub_hist.data));
args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_hist.data)); args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_hist.data));
args.push_back( make_pair( sizeof(cl_int), (void *)&src_step)); args.push_back( make_pair( sizeof(cl_int), (void *)&src_step));
openCLExecuteKernel(clCxt, &imgproc_histogram, kernelName, globalThreads, localThreads, args, -1, -1);
openCLExecuteKernel(sub_hist.clCxt, &imgproc_histogram, "merge_hist", globalThreads, localThreads, args, -1, -1);
} }
void calcHist(const oclMat &mat_src, oclMat &mat_hist) void calcHist(const oclMat &mat_src, oclMat &mat_hist)
{ {
using namespace histograms; using namespace histograms;
@ -1439,6 +1278,7 @@ namespace cv
calc_sub_hist(mat_src, buf); calc_sub_hist(mat_src, buf);
merge_sub_hist(buf, mat_hist); merge_sub_hist(buf, mat_hist);
} }
///////////////////////////////////equalizeHist///////////////////////////////////////////////////// ///////////////////////////////////equalizeHist/////////////////////////////////////////////////////
void equalizeHist(const oclMat &mat_src, oclMat &mat_dst) void equalizeHist(const oclMat &mat_src, oclMat &mat_dst)
{ {
@ -1448,17 +1288,17 @@ namespace cv
calcHist(mat_src, mat_hist); calcHist(mat_src, mat_hist);
Context *clCxt = mat_src.clCxt;
string kernelName = "calLUT";
size_t localThreads[3] = { 256, 1, 1}; size_t localThreads[3] = { 256, 1, 1};
size_t globalThreads[3] = { 256, 1, 1}; size_t globalThreads[3] = { 256, 1, 1};
oclMat lut(1, 256, CV_8UC1); oclMat lut(1, 256, CV_8UC1);
vector<pair<size_t , const void *> > args;
int total = mat_src.rows * mat_src.cols; int total = mat_src.rows * mat_src.cols;
vector<pair<size_t , const void *> > args;
args.push_back( make_pair( sizeof(cl_mem), (void *)&lut.data)); args.push_back( make_pair( sizeof(cl_mem), (void *)&lut.data));
args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_hist.data)); args.push_back( make_pair( sizeof(cl_mem), (void *)&mat_hist.data));
args.push_back( make_pair( sizeof(int), (void *)&total)); args.push_back( make_pair( sizeof(int), (void *)&total));
openCLExecuteKernel(clCxt, &imgproc_histogram, kernelName, globalThreads, localThreads, args, -1, -1);
openCLExecuteKernel(mat_src.clCxt, &imgproc_histogram, "calLUT", globalThreads, localThreads, args, -1, -1);
LUT(mat_src, lut, mat_dst); LUT(mat_src, lut, mat_dst);
} }
@ -1493,12 +1333,11 @@ namespace cv
else else
{ {
cl_kernel kernel = openCLGetKernelFromSource(Context::getContext(), &imgproc_clahe, kernelName); cl_kernel kernel = openCLGetKernelFromSource(Context::getContext(), &imgproc_clahe, kernelName);
size_t wave_size = queryWaveFrontSize(kernel); int wave_size = (int)queryWaveFrontSize(kernel);
openCLSafeCall(clReleaseKernel(kernel)); openCLSafeCall(clReleaseKernel(kernel));
static char opt[20] = {0}; std::string opt = format("-D WAVE_SIZE=%d", wave_size);
sprintf(opt, "-D WAVE_SIZE=%d", (int)wave_size); openCLExecuteKernel(Context::getContext(), &imgproc_clahe, kernelName, globalThreads, localThreads, args, -1, -1, opt.c_str());
openCLExecuteKernel(Context::getContext(), &imgproc_clahe, kernelName, globalThreads, localThreads, args, -1, -1, opt);
} }
} }
@ -1522,11 +1361,10 @@ namespace cv
args.push_back( std::make_pair( sizeof(cl_int), (void *)&tilesX )); args.push_back( std::make_pair( sizeof(cl_int), (void *)&tilesX ));
args.push_back( std::make_pair( sizeof(cl_int), (void *)&tilesY )); args.push_back( std::make_pair( sizeof(cl_int), (void *)&tilesY ));
String kernelName = "transform";
size_t localThreads[3] = { 32, 8, 1 }; size_t localThreads[3] = { 32, 8, 1 };
size_t globalThreads[3] = { src.cols, src.rows, 1 }; size_t globalThreads[3] = { src.cols, src.rows, 1 };
openCLExecuteKernel(Context::getContext(), &imgproc_clahe, kernelName, globalThreads, localThreads, args, -1, -1); openCLExecuteKernel(Context::getContext(), &imgproc_clahe, "transform", globalThreads, localThreads, args, -1, -1);
} }
} }
@ -1557,6 +1395,7 @@ namespace cv
oclMat srcExt_; oclMat srcExt_;
oclMat lut_; oclMat lut_;
}; };
CLAHE_Impl::CLAHE_Impl(double clipLimit, int tilesX, int tilesY) : CLAHE_Impl::CLAHE_Impl(double clipLimit, int tilesX, int tilesY) :
clipLimit_(clipLimit), tilesX_(tilesX), tilesY_(tilesY) clipLimit_(clipLimit), tilesX_(tilesX), tilesY_(tilesY)
{ {
@ -1566,6 +1405,7 @@ namespace cv
obj.info()->addParam(obj, "clipLimit", obj.clipLimit_); obj.info()->addParam(obj, "clipLimit", obj.clipLimit_);
obj.info()->addParam(obj, "tilesX", obj.tilesX_); obj.info()->addParam(obj, "tilesX", obj.tilesX_);
obj.info()->addParam(obj, "tilesY", obj.tilesY_)) obj.info()->addParam(obj, "tilesY", obj.tilesY_))
void CLAHE_Impl::apply(cv::InputArray src_raw, cv::OutputArray dst_raw) void CLAHE_Impl::apply(cv::InputArray src_raw, cv::OutputArray dst_raw)
{ {
oclMat& src = getOclMatRef(src_raw); oclMat& src = getOclMatRef(src_raw);
@ -1605,7 +1445,6 @@ namespace cv
} }
clahe::calcLut(srcForLut, lut_, tilesX_, tilesY_, tileSize, clipLimit, lutScale); clahe::calcLut(srcForLut, lut_, tilesX_, tilesY_, tileSize, clipLimit, lutScale);
//finish();
clahe::transform(src, dst, lut_, tilesX_, tilesY_, tileSize); clahe::transform(src, dst, lut_, tilesX_, tilesY_, tileSize);
} }
@ -1643,8 +1482,8 @@ namespace cv
} }
//////////////////////////////////bilateralFilter//////////////////////////////////////////////////// //////////////////////////////////bilateralFilter////////////////////////////////////////////////////
static void
oclbilateralFilter_8u( const oclMat &src, oclMat &dst, int d, static void oclbilateralFilter_8u( const oclMat &src, oclMat &dst, int d,
double sigma_color, double sigma_space, double sigma_color, double sigma_space,
int borderType ) int borderType )
{ {
@ -1682,6 +1521,7 @@ namespace cv
int dst_step_in_pixel = dst.step / dst.elemSize(); int dst_step_in_pixel = dst.step / dst.elemSize();
int dst_offset_in_pixel = dst.offset / dst.elemSize(); int dst_offset_in_pixel = dst.offset / dst.elemSize();
int temp_step_in_pixel = temp.step / temp.elemSize(); int temp_step_in_pixel = temp.step / temp.elemSize();
// initialize color-related bilateral filter coefficients // initialize color-related bilateral filter coefficients
for( i = 0; i < 256 * cn; i++ ) for( i = 0; i < 256 * cn; i++ )
color_weight[i] = (float)std::exp(i * i * gauss_color_coeff); color_weight[i] = (float)std::exp(i * i * gauss_color_coeff);
@ -1696,21 +1536,21 @@ namespace cv
space_weight[maxk] = (float)std::exp(r * r * gauss_space_coeff); space_weight[maxk] = (float)std::exp(r * r * gauss_space_coeff);
space_ofs[maxk++] = (int)(i * temp_step_in_pixel + j); space_ofs[maxk++] = (int)(i * temp_step_in_pixel + j);
} }
oclMat oclcolor_weight(1, cn * 256, CV_32FC1, color_weight); oclMat oclcolor_weight(1, cn * 256, CV_32FC1, color_weight);
oclMat oclspace_weight(1, d * d, CV_32FC1, space_weight); oclMat oclspace_weight(1, d * d, CV_32FC1, space_weight);
oclMat oclspace_ofs(1, d * d, CV_32SC1, space_ofs); oclMat oclspace_ofs(1, d * d, CV_32SC1, space_ofs);
string kernelName = "bilateral"; string kernelName = "bilateral";
size_t localThreads[3] = { 16, 16, 1 }; size_t localThreads[3] = { 16, 16, 1 };
size_t globalThreads[3] = { (dst.cols + localThreads[0] - 1) / localThreads[0] *localThreads[0], size_t globalThreads[3] = { dst.cols, dst.rows, 1 };
(dst.rows + localThreads[1] - 1) / localThreads[1] *localThreads[1],
1
};
if ((dst.type() == CV_8UC1) && ((dst.offset & 3) == 0) && ((dst.cols & 3) == 0)) if ((dst.type() == CV_8UC1) && ((dst.offset & 3) == 0) && ((dst.cols & 3) == 0))
{ {
kernelName = "bilateral2"; kernelName = "bilateral2";
globalThreads[0] = (dst.cols / 4 + localThreads[0] - 1) / localThreads[0] * localThreads[0]; globalThreads[0] = dst.cols / 4;
} }
vector<pair<size_t , const void *> > args; vector<pair<size_t , const void *> > args;
args.push_back( make_pair( sizeof(cl_mem), (void *)&dst.data )); args.push_back( make_pair( sizeof(cl_mem), (void *)&dst.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&temp.data )); args.push_back( make_pair( sizeof(cl_mem), (void *)&temp.data ));
@ -1730,13 +1570,11 @@ namespace cv
} }
void bilateralFilter(const oclMat &src, oclMat &dst, int radius, double sigmaclr, double sigmaspc, int borderType) void bilateralFilter(const oclMat &src, oclMat &dst, int radius, double sigmaclr, double sigmaspc, int borderType)
{ {
dst.create( src.size(), src.type() ); dst.create( src.size(), src.type() );
if ( src.depth() == CV_8U ) if ( src.depth() == CV_8U )
oclbilateralFilter_8u( src, dst, radius, sigmaclr, sigmaspc, borderType ); oclbilateralFilter_8u( src, dst, radius, sigmaclr, sigmaspc, borderType );
else else
CV_Error( CV_StsUnsupportedFormat, CV_Error( CV_StsUnsupportedFormat, "Bilateral filtering is only implemented for 8uimages" );
"Bilateral filtering is only implemented for 8uimages" );
} }
} }
@ -1745,18 +1583,9 @@ namespace cv
static void convolve_run(const oclMat &src, const oclMat &temp1, oclMat &dst, string kernelName, const cv::ocl::ProgramEntry* source) static void convolve_run(const oclMat &src, const oclMat &temp1, oclMat &dst, string kernelName, const cv::ocl::ProgramEntry* source)
{ {
CV_Assert(src.depth() == CV_32FC1);
CV_Assert(temp1.depth() == CV_32F);
CV_Assert(temp1.cols <= 17 && temp1.rows <= 17);
dst.create(src.size(), src.type()); dst.create(src.size(), src.type());
CV_Assert(src.cols == dst.cols && src.rows == dst.rows); int channels = dst.oclchannels(), depth = dst.depth();
CV_Assert(src.type() == dst.type());
Context *clCxt = src.clCxt;
int channels = dst.oclchannels();
int depth = dst.depth();
size_t vector_length = 1; size_t vector_length = 1;
int offset_cols = ((dst.offset % dst.step) / dst.elemSize1()) & (vector_length - 1); int offset_cols = ((dst.offset % dst.step) / dst.elemSize1()) & (vector_length - 1);
@ -1778,15 +1607,14 @@ static void convolve_run(const oclMat &src, const oclMat &temp1, oclMat &dst, st
args.push_back( make_pair( sizeof(cl_int), (void *)&temp1.rows )); args.push_back( make_pair( sizeof(cl_int), (void *)&temp1.rows ));
args.push_back( make_pair( sizeof(cl_int), (void *)&temp1.cols )); args.push_back( make_pair( sizeof(cl_int), (void *)&temp1.cols ));
openCLExecuteKernel(clCxt, source, kernelName, globalThreads, localThreads, args, -1, depth); openCLExecuteKernel(src.clCxt, source, kernelName, globalThreads, localThreads, args, -1, depth);
} }
void cv::ocl::convolve(const oclMat &x, const oclMat &t, oclMat &y) void cv::ocl::convolve(const oclMat &x, const oclMat &t, oclMat &y)
{ {
CV_Assert(x.depth() == CV_32F); CV_Assert(x.depth() == CV_32F && t.depth() == CV_32F);
CV_Assert(t.depth() == CV_32F); CV_Assert(t.cols <= 17 && t.rows <= 17);
CV_Assert(x.type() == y.type() && x.size() == y.size());
y.create(x.size(), x.type());
string kernelName = "convolve";
convolve_run(x, t, y, kernelName, &imgproc_convolve); y.create(x.size(), x.type());
convolve_run(x, t, y, "convolve", &imgproc_convolve);
} }