Merge pull request #2496 from ilya-lavrenov:tapi_imgproc

This commit is contained in:
Andrey Pavlenko 2014-03-20 19:20:14 +04:00 committed by OpenCV Buildbot
commit bdfd29a0b8
3 changed files with 87 additions and 102 deletions

View File

@ -42,7 +42,6 @@
#include "precomp.hpp"
#include <limits.h>
#include <stdio.h>
#include "opencl_kernels.hpp"
/****************************************************************************************\
@ -1291,9 +1290,10 @@ static bool ocl_morphology_op(InputArray _src, OutputArray _dst, Mat kernel,
{
CV_Assert(op == MORPH_ERODE || op == MORPH_DILATE);
int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
if (_src.depth() == CV_64F && !doubleSupport)
if (depth == CV_64F && !doubleSupport)
return false;
UMat kernel8U;
@ -1324,13 +1324,14 @@ static bool ocl_morphology_op(InputArray _src, OutputArray _dst, Mat kernel,
return false;
static const char * const op2str[] = { "ERODE", "DILATE" };
String buildOptions = format("-D RADIUSX=%d -D RADIUSY=%d -D LSIZE0=%d -D LSIZE1=%d -D %s%s%s -D GENTYPE=%s -D DEPTH_%d",
anchor.x, anchor.y, (int)localThreads[0], (int)localThreads[1], op2str[op],
String buildOptions = format("-D RADIUSX=%d -D RADIUSY=%d -D LSIZE0=%d -D LSIZE1=%d -D %s%s%s"
" -D T=%s -D DEPTH_%d -D cn=%d -D T1=%s", anchor.x, anchor.y,
(int)localThreads[0], (int)localThreads[1], op2str[op],
doubleSupport ? " -D DOUBLE_SUPPORT" : "", rectKernel ? " -D RECTKERNEL" : "",
ocl::typeToStr(_src.type()), _src.depth() );
ocl::typeToStr(_src.type()), _src.depth(), cn, ocl::typeToStr(depth));
std::vector<ocl::Kernel> kernels;
for (int i = 0; i<iterations; i++)
for (int i = 0; i < iterations; i++)
{
ocl::Kernel k("morph", ocl::imgproc::morph_oclsrc, buildOptions);
if (k.empty())
@ -1341,33 +1342,27 @@ static bool ocl_morphology_op(InputArray _src, OutputArray _dst, Mat kernel,
_dst.create(src.size(), src.type());
UMat dst = _dst.getUMat();
if( iterations== 1 && src.u != dst.u)
if (iterations == 1 && src.u != dst.u)
{
Size wholesize;
Point ofs;
src.locateROI(wholesize, ofs);
int wholecols = wholesize.width, wholerows = wholesize.height;
int idxArg = 0;
idxArg = kernels[0].set(idxArg, ocl::KernelArg::ReadOnlyNoSize(src));
idxArg = kernels[0].set(idxArg, ocl::KernelArg::WriteOnlyNoSize(dst));
idxArg = kernels[0].set(idxArg, ofs.x);
idxArg = kernels[0].set(idxArg, ofs.y);
idxArg = kernels[0].set(idxArg, src.cols);
idxArg = kernels[0].set(idxArg, src.rows);
idxArg = kernels[0].set(idxArg, ocl::KernelArg::PtrReadOnly(kernel8U));
idxArg = kernels[0].set(idxArg, wholecols);
idxArg = kernels[0].set(idxArg, wholerows);
kernels[0].args(ocl::KernelArg::ReadOnlyNoSize(src), ocl::KernelArg::WriteOnlyNoSize(dst),
ofs.x, ofs.y, src.cols, src.rows, ocl::KernelArg::PtrReadOnly(kernel8U),
wholecols, wholerows);
return kernels[0].run(2, globalThreads, localThreads, false);
}
for(int i = 0; i< iterations; i++)
for (int i = 0; i < iterations; i++)
{
UMat source;
Size wholesize;
Point ofs;
if( i == 0)
if (i == 0)
{
int cols = src.cols, rows = src.rows;
src.locateROI(wholesize,ofs);
@ -1385,20 +1380,11 @@ static bool ocl_morphology_op(InputArray _src, OutputArray _dst, Mat kernel,
dst.adjustROI(-ofs.y, -wholesize.height + rows + ofs.y, -ofs.x, -wholesize.width + cols + ofs.x);
source.adjustROI(-ofs.y, -wholesize.height + rows + ofs.y, -ofs.x, -wholesize.width + cols + ofs.x);
}
source.locateROI(wholesize, ofs);
int wholecols = wholesize.width, wholerows = wholesize.height;
int idxArg = 0;
idxArg = kernels[i].set(idxArg, ocl::KernelArg::ReadOnlyNoSize(source));
idxArg = kernels[i].set(idxArg, ocl::KernelArg::WriteOnlyNoSize(dst));
idxArg = kernels[i].set(idxArg, ofs.x);
idxArg = kernels[i].set(idxArg, ofs.y);
idxArg = kernels[i].set(idxArg, source.cols);
idxArg = kernels[i].set(idxArg, source.rows);
idxArg = kernels[i].set(idxArg, ocl::KernelArg::PtrReadOnly(kernel8U));
idxArg = kernels[i].set(idxArg, wholecols);
idxArg = kernels[i].set(idxArg, wholerows);
kernels[i].args(ocl::KernelArg::ReadOnlyNoSize(source), ocl::KernelArg::WriteOnlyNoSize(dst),
ofs.x, ofs.y, source.cols, source.rows, ocl::KernelArg::PtrReadOnly(kernel8U),
wholesize.width, wholesize.height);
if (!kernels[i].run(2, globalThreads, localThreads, false))
return false;
@ -1427,13 +1413,13 @@ static void morphOp( int op, InputArray _src, OutputArray _dst,
return;
#endif
if( iterations == 0 || kernel.rows*kernel.cols == 1 )
if (iterations == 0 || kernel.rows*kernel.cols == 1)
{
_src.copyTo(_dst);
return;
}
if( !kernel.data )
if (!kernel.data)
{
kernel = getStructuringElement(MORPH_RECT, Size(1+iterations*2,1+iterations*2));
anchor = Point(iterations, iterations);
@ -1450,7 +1436,7 @@ static void morphOp( int op, InputArray _src, OutputArray _dst,
}
CV_OCL_RUN(_dst.isUMat() && _src.size() == _dst.size() && src_type == dst_type &&
_src.dims() <= 2 && (src_cn == 1 || src_cn == 4) &&
_src.dims() <= 2 && src_cn <= 4 &&
(src_depth == CV_8U || src_depth == CV_32F || src_depth == CV_64F ) &&
borderType == cv::BORDER_CONSTANT && borderValue == morphologyDefaultBorderValue() &&
(op == MORPH_ERODE || op == MORPH_DILATE),

View File

@ -43,6 +43,16 @@
#endif
#endif
#if cn != 3
#define loadpix(addr) *(__global const T *)(addr)
#define storepix(val, addr) *(__global T *)(addr) = val
#define TSIZE (int)sizeof(T)
#else
#define loadpix(addr) vload3(0, (__global const T1 *)(addr))
#define storepix(val, addr) vstore3(val, 0, (__global T1 *)(addr))
#define TSIZE ((int)sizeof(T1)*3)
#endif
#ifdef DEPTH_0
#ifdef ERODE
#define VAL 255
@ -50,16 +60,14 @@
#ifdef DILATE
#define VAL 0
#endif
#endif
#ifdef DEPTH_5
#elif defined DEPTH_5
#ifdef ERODE
#define VAL FLT_MAX
#endif
#ifdef DILATE
#define VAL -FLT_MAX
#endif
#endif
#ifdef DEPTH_6
#elif defined DEPTH_6
#ifdef ERODE
#define VAL DBL_MAX
#endif
@ -74,79 +82,70 @@
#ifdef DILATE
#define MORPH_OP(A,B) max((A),(B))
#endif
//BORDER_CONSTANT: iiiiii|abcdefgh|iiiiiii
#define ELEM(i,l_edge,r_edge,elem1,elem2) (i)<(l_edge) | (i) >= (r_edge) ? (elem1) : (elem2)
__kernel void morph(__global const uchar * restrict srcptr, int src_step, int src_offset,
// BORDER_CONSTANT: iiiiii|abcdefgh|iiiiiii
#define ELEM(i, l_edge, r_edge, elem1, elem2) (i) < (l_edge) | (i) >= (r_edge) ? (elem1) : (elem2)
__kernel void morph(__global const uchar * srcptr, int src_step, int src_offset,
__global uchar * dstptr, int dst_step, int dst_offset,
int src_offset_x, int src_offset_y,
int cols, int rows,
__constant uchar * mat_kernel,
int src_whole_cols, int src_whole_rows)
int src_offset_x, int src_offset_y, int cols, int rows,
__constant uchar * mat_kernel, int src_whole_cols, int src_whole_rows)
{
int l_x = get_local_id(0);
int l_y = get_local_id(1);
int x = get_group_id(0)*LSIZE0;
int y = get_group_id(1)*LSIZE1;
int start_x = x+src_offset_x-RADIUSX;
int end_x = x + src_offset_x+LSIZE0+RADIUSX;
int width = end_x -(x+src_offset_x-RADIUSX)+1;
int start_y = y+src_offset_y-RADIUSY;
int point1 = mad24(l_y,LSIZE0,l_x);
int point2 = point1 + LSIZE0*LSIZE1;
int tl_x = point1 % width;
int tl_y = point1 / width;
int tl_x2 = point2 % width;
int tl_y2 = point2 / width;
int cur_x = start_x + tl_x;
int cur_y = start_y + tl_y;
int cur_x2 = start_x + tl_x2;
int cur_y2 = start_y + tl_y2;
int start_addr = mad24(cur_y,src_step, cur_x*(int)sizeof(GENTYPE));
int start_addr2 = mad24(cur_y2,src_step, cur_x2*(int)sizeof(GENTYPE));
GENTYPE temp0,temp1;
__local GENTYPE LDS_DAT[2*LSIZE1*LSIZE0];
int gidx = get_global_id(0), gidy = get_global_id(1);
int l_x = get_local_id(0), l_y = get_local_id(1);
int x = get_group_id(0) * LSIZE0, y = get_group_id(1) * LSIZE1;
int start_x = x + src_offset_x - RADIUSX;
int end_x = x + src_offset_x + LSIZE0 + RADIUSX;
int width = end_x - (x + src_offset_x - RADIUSX) + 1;
int start_y = y + src_offset_y - RADIUSY;
int point1 = mad24(l_y, LSIZE0, l_x);
int point2 = point1 + LSIZE0 * LSIZE1;
int tl_x = point1 % width, tl_y = point1 / width;
int tl_x2 = point2 % width, tl_y2 = point2 / width;
int cur_x = start_x + tl_x, cur_y = start_y + tl_y;
int cur_x2 = start_x + tl_x2, cur_y2 = start_y + tl_y2;
int start_addr = mad24(cur_y, src_step, cur_x * TSIZE);
int start_addr2 = mad24(cur_y2, src_step, cur_x2 * TSIZE);
int end_addr = mad24(src_whole_rows - 1,src_step,src_whole_cols*(int)sizeof(GENTYPE));
//read pixels from src
start_addr = ((start_addr < end_addr) && (start_addr > 0)) ? start_addr : 0;
start_addr2 = ((start_addr2 < end_addr) && (start_addr2 > 0)) ? start_addr2 : 0;
__global const GENTYPE * src;
src = (__global const GENTYPE *)(srcptr+start_addr);
temp0 = src[0];
src = (__global const GENTYPE *)(srcptr+start_addr2);
temp1 = src[0];
//judge if read out of boundary
temp0= ELEM(cur_x,0,src_whole_cols,(GENTYPE)VAL,temp0);
temp0= ELEM(cur_y,0,src_whole_rows,(GENTYPE)VAL,temp0);
__local T LDS_DAT[2*LSIZE1*LSIZE0];
temp1= ELEM(cur_x2,0,src_whole_cols,(GENTYPE)VAL,temp1);
temp1= ELEM(cur_y2,0,src_whole_rows,(GENTYPE)VAL,temp1);
// read pixels from src
int end_addr = mad24(src_whole_rows - 1, src_step, src_whole_cols * TSIZE);
start_addr = start_addr < end_addr && start_addr > 0 ? start_addr : 0;
start_addr2 = start_addr2 < end_addr && start_addr2 > 0 ? start_addr2 : 0;
T temp0 = loadpix(srcptr + start_addr);
T temp1 = loadpix(srcptr + start_addr2);
// judge if read out of boundary
temp0 = ELEM(cur_x, 0, src_whole_cols, (T)(VAL),temp0);
temp0 = ELEM(cur_y, 0, src_whole_rows, (T)(VAL),temp0);
temp1 = ELEM(cur_x2, 0, src_whole_cols, (T)(VAL), temp1);
temp1 = ELEM(cur_y2, 0, src_whole_rows, (T)(VAL), temp1);
LDS_DAT[point1] = temp0;
LDS_DAT[point2] = temp1;
barrier(CLK_LOCAL_MEM_FENCE);
GENTYPE res = (GENTYPE)VAL;
for(int i=0; i<2*RADIUSY+1; i++)
for(int j=0; j<2*RADIUSX+1; j++)
T res = (T)(VAL);
for (int i = 0, sizey = 2 * RADIUSY + 1; i < sizey; i++)
for (int j = 0, sizex = 2 * RADIUSX + 1; j < sizex; j++)
{
res =
#ifndef RECTKERNEL
mat_kernel[i*(2*RADIUSX+1)+j] ?
#endif
MORPH_OP(res,LDS_DAT[mad24(l_y+i,width,l_x+j)])
MORPH_OP(res, LDS_DAT[mad24(l_y + i, width, l_x + j)])
#ifndef RECTKERNEL
:res
: res
#endif
;
}
int gidx = get_global_id(0);
int gidy = get_global_id(1);
if(gidx<cols && gidy<rows)
{
int dst_index = mad24(gidy, dst_step, dst_offset + gidx * (int)sizeof(GENTYPE));
__global GENTYPE * dst = (__global GENTYPE *)(dstptr + dst_index);
dst[0] = res;
}
if (gidx < cols && gidy < rows)
{
int dst_index = mad24(gidy, dst_step, mad24(gidx, TSIZE, dst_offset));
storepix(res, dstptr + dst_index);
}
}

View File

@ -349,28 +349,28 @@ OCL_INSTANTIATE_TEST_CASE_P(Filter, GaussianBlurTest, Combine(
Bool()));
OCL_INSTANTIATE_TEST_CASE_P(Filter, Erode, Combine(
Values(CV_8UC1, CV_8UC4, CV_32FC1, CV_32FC4, CV_64FC1, CV_64FC4),
Values(CV_8UC1, CV_8UC3, CV_8UC4, CV_32FC1, CV_32FC3, CV_32FC4, CV_64FC1, CV_64FC4),
Values(3, 5, 7),
Values(Size(0,0)),//not used
Values((BorderType)BORDER_CONSTANT),//not used
Values(1.0, 2.0, 3.0),
Bool() ) );
Bool()));
OCL_INSTANTIATE_TEST_CASE_P(Filter, Dilate, Combine(
Values(CV_8UC1, CV_8UC4, CV_32FC1, CV_32FC4, CV_64FC1, CV_64FC4),
Values(CV_8UC1, CV_8UC3, CV_8UC4, CV_32FC1, CV_32FC3, CV_32FC4, CV_64FC1, CV_64FC4),
Values(3, 5, 7),
Values(Size(0,0)),//not used
Values((BorderType)BORDER_CONSTANT),//not used
Values(1.0, 2.0, 3.0),
Bool() ) );
Bool()));
OCL_INSTANTIATE_TEST_CASE_P(Filter, MorphologyEx, Combine(
Values(CV_8UC1, CV_8UC4, CV_32FC1, CV_32FC4, CV_64FC1, CV_64FC4),
Values(CV_8UC1, CV_8UC3, CV_8UC4, CV_32FC1, CV_32FC3, CV_32FC4, CV_64FC1, CV_64FC4),
Values(3, 5, 7),
Values(Size(0,0), Size(0,1), Size(0,2), Size(0,3), Size(0,4), Size(0,5),Size(0,6)),//uses as generator of operations
Values((BorderType)BORDER_CONSTANT),//not used
Values(Size(0, 0), Size(0, 1), Size(0, 2), Size(0, 3), Size(0, 4), Size(0, 5), Size(0, 6)), // used as generator of operations
Values((BorderType)BORDER_CONSTANT),// not used
Values(1.0, 2.0, 3.0),
Bool() ) );
Bool()));
} } // namespace cvtest::ocl