potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-29 23:33:05 +04:00
Code Packages Releases Wiki Activity

Merge branch 4.x

Browse Source
This commit is contained in:
Alexander Smorkalov
2026-04-23 11:00:34 +03:00
parent 323ea1782d 273e52ce8d
commit 7bde3f9ae9
38 changed files with 3437 additions and 293 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/src/matrix_operations.cpp
+24 -3
View File
@@ -327,6 +327,20 @@ cv::Mat cv::Mat::cross(InputArray _m) const
namespace cv
{
typedef void (*ReduceSumFunc)(const Mat& src, Mat& dst);
ReduceSumFunc getReduceCSumFunc(int sdepth, int ddepth);
ReduceSumFunc getReduceRSumFunc(int sdepth, int ddepth);
template <typename T, typename WT, typename Op>
struct ReduceR_SIMD
{
int operator()(const T*, int start, int, WT*, const Op&) const
{
return start;
}
};
template<typename T, typename ST, typename WT, class Op, class OpInit>
class ReduceR_Invoker : public ParallelLoopBody
{
@@ -350,7 +364,8 @@ public:
for( ; --height; )
{
src += srcstep;
i = range.start;
ReduceR_SIMD<T, WT, Op> simd_op;
i = simd_op(src, range.start, range.end, buf, op);
#if CV_ENABLE_UNROLLED
for(; i <= range.end - 4; i += 4 )
{
@@ -801,7 +816,10 @@ void cv::reduce(InputArray _src, OutputArray _dst, int dim, int op, int dtype)
{
if( op == REDUCE_SUM )
{
if(sdepth == CV_8U && ddepth == CV_32S)
ReduceSumFunc simd_func = getReduceRSumFunc(sdepth, ddepth);
if(simd_func)
func = (ReduceFunc)simd_func;
else if(sdepth == CV_8U && ddepth == CV_32S)
func = reduceSumR8u32s;
else if(sdepth == CV_8U && ddepth == CV_32F)
func = reduceSumR8u32f;
@@ -876,7 +894,10 @@ void cv::reduce(InputArray _src, OutputArray _dst, int dim, int op, int dtype)
{
if(op == REDUCE_SUM)
{
if(sdepth == CV_8U && ddepth == CV_32S)
ReduceSumFunc simd_func = getReduceCSumFunc(sdepth, ddepth);
if(simd_func)
func = (ReduceFunc)simd_func;
else if(sdepth == CV_8U && ddepth == CV_32S)
func = reduceSumC8u32s;
else if(sdepth == CV_8U && ddepth == CV_32F)
func = reduceSumC8u32f;
modules/core/src/matrix_transform.cpp
+54 -17
View File
@@ -973,7 +973,12 @@ static bool ocl_flip(InputArray _src, OutputArray _dst, int flipCode )
if (cn > 4)
return false;
const char * kernelName;
Size size = _src.size();
_dst.create(size, type);
UMat src = _src.getUMat(), dst = _dst.getUMat();
bool inplace = (dst.u == src.u);
String kernelName;
if (flipCode == 0)
kernelName = "arithm_flip_rows", flipType = FLIP_ROWS;
else if (flipCode > 0)
@@ -981,33 +986,65 @@ static bool ocl_flip(InputArray _src, OutputArray _dst, int flipCode )
else
kernelName = "arithm_flip_rows_cols", flipType = FLIP_BOTH;
if(inplace)
kernelName += "_inplace";
int pxPerWIy = (dev.isIntel() && (dev.type() & ocl::Device::TYPE_GPU)) ? 4 : 1;
kercn = (cn!=3 || flipType == FLIP_ROWS) ? std::max(kercn, cn) : cn;
const int TILE_SIZE = 32, BLOCK_ROWS = 8;
ocl::Kernel k(kernelName, ocl::core::flip_oclsrc,
format( "-D T=%s -D T1=%s -D DEPTH=%d -D cn=%d -D PIX_PER_WI_Y=%d -D kercn=%d",
ocl::Kernel k(kernelName.c_str(), ocl::core::flip_oclsrc,
format( "-D T=%s -D T1=%s -D DEPTH=%d -D cn=%d -D PIX_PER_WI_Y=%d -D kercn=%d -D TILE_SIZE=%d -D BLOCK_ROWS=%d%s",
kercn != cn ? ocl::typeToStr(CV_MAKE_TYPE(depth, kercn)) : ocl::vecopTypeToStr(CV_MAKE_TYPE(depth, kercn)),
kercn != cn ? ocl::typeToStr(depth) : ocl::vecopTypeToStr(depth), depth, cn, pxPerWIy, kercn));
kercn != cn ? ocl::typeToStr(depth) : ocl::vecopTypeToStr(depth), depth, cn, pxPerWIy, kercn, TILE_SIZE, BLOCK_ROWS,
inplace ? " -D INPLACE" : ""));
if (k.empty())
return false;
Size size = _src.size();
_dst.create(size, type);
UMat src = _src.getUMat(), dst = _dst.getUMat();
int cols = size.width * cn / kercn, rows = size.height;
cols = flipType == FLIP_COLS ? (cols + 1) >> 1 : cols;
rows = flipType & FLIP_ROWS ? (rows + 1) >> 1 : rows;
int work_cols = flipType == FLIP_COLS ? (cols + 1) >> 1 : cols;
int work_rows = flipType & FLIP_ROWS ? (rows + 1) >> 1 : rows;
k.args(ocl::KernelArg::ReadOnlyNoSize(src),
ocl::KernelArg::WriteOnly(dst, cn, kercn), rows, cols);
if (inplace)
{
k.args(ocl::KernelArg::ReadWriteNoSize(dst), rows, cols);
size_t maxWorkGroupSize = dev.maxWorkGroupSize();
CV_Assert(maxWorkGroupSize % 4 == 0);
int gs_cols, gs_rows;
if (flipType == FLIP_COLS)
{
gs_cols = work_cols;
gs_rows = rows;
}
else if (flipType == FLIP_ROWS)
{
gs_cols = cols;
gs_rows = work_rows;
}
else // FLIP_BOTH
{
gs_cols = cols;
gs_rows = rows;
}
size_t globalsize[2] = { (size_t)cols, ((size_t)rows + pxPerWIy - 1) / pxPerWIy },
localsize[2] = { maxWorkGroupSize / 4, 4 };
return k.run(2, globalsize, (flipType == FLIP_COLS) && !dev.isIntel() ? localsize : NULL, false);
size_t globalsize[2] = {
(size_t)divUp(gs_cols, TILE_SIZE) * TILE_SIZE,
(size_t)divUp(gs_rows, TILE_SIZE) * BLOCK_ROWS
};
size_t localsize[2] = { TILE_SIZE, BLOCK_ROWS };
return k.run(2, globalsize, localsize, false);
}
else
{
k.args(ocl::KernelArg::ReadOnlyNoSize(src),
ocl::KernelArg::WriteOnly(dst, cn, kercn), work_rows, work_cols);
size_t maxWorkGroupSize = dev.maxWorkGroupSize();
CV_Assert(maxWorkGroupSize % 4 == 0);
size_t globalsize[2] = { (size_t)work_cols, ((size_t)work_rows + pxPerWIy - 1) / pxPerWIy };
size_t localsize[2] = { maxWorkGroupSize / 4, 4 };
return k.run(2, globalsize, (flipType == FLIP_COLS) && !dev.isIntel() ? localsize : NULL, false);
}
}
#endif
modules/core/src/opencl/flip.cl
+196
View File
@@ -63,6 +63,9 @@
#endif
#define TSIZE ((int)sizeof(T1)*3)
#endif
#define LDS_STEP (TILE_SIZE + 1)
#ifndef INPLACE
__kernel void arithm_flip_rows(__global const uchar * srcptr, int src_step, int src_offset,
__global uchar * dstptr, int dst_step, int dst_offset,
@@ -183,3 +186,196 @@ __kernel void arithm_flip_cols(__global const uchar * srcptr, int src_step, int
}
}
}
#else
__kernel void arithm_flip_rows_inplace(__global uchar * srcptr, int src_step, int src_offset,
int rows, int cols)
{
int gp_x = get_group_id(0);
int gp_y = get_group_id(1);
int lx = get_local_id(0);
int ly = get_local_id(1);
__local T tile_top[TILE_SIZE * LDS_STEP];
__local T tile_bottom[TILE_SIZE * LDS_STEP];
int half_rows = (rows + 1) / 2;
int x = gp_x * TILE_SIZE + lx;
int y_top = gp_y * TILE_SIZE + ly;
int y_bottom = rows - 1 - y_top;
#pragma unroll
for (int i = 0; i < TILE_SIZE; i += BLOCK_ROWS)
{
if (x < cols && y_top + i < half_rows)
{
int curr_y_top = y_top + i;
int curr_y_bottom = rows - 1 - curr_y_top;
T val_top = loadpix(srcptr + mad24(curr_y_top, src_step, mad24(x, TSIZE, src_offset)));
T val_bottom = loadpix(srcptr + mad24(curr_y_bottom, src_step, mad24(x, TSIZE, src_offset)));
tile_top[mad24(ly + i, LDS_STEP, lx)] = val_top;
tile_bottom[mad24(ly + i, LDS_STEP, lx)] = val_bottom;
}
}
barrier(CLK_LOCAL_MEM_FENCE);
#pragma unroll
for (int i = 0; i < TILE_SIZE; i += BLOCK_ROWS)
{
if (x < cols && y_top + i < half_rows)
{
int curr_y_top = y_top + i;
int curr_y_bottom = rows - 1 - curr_y_top;
storepix(tile_bottom[mad24(ly + i, LDS_STEP, lx)],
srcptr + mad24(curr_y_top, src_step, mad24(x, TSIZE, src_offset)));
if (curr_y_top != curr_y_bottom)
{
storepix(tile_top[mad24(ly + i, LDS_STEP, lx)],
srcptr + mad24(curr_y_bottom, src_step, mad24(x, TSIZE, src_offset)));
}
}
}
}
__kernel void arithm_flip_rows_cols_inplace(__global uchar * srcptr, int src_step, int src_offset,
int rows, int cols)
{
int gp_x = get_group_id(0);
int gp_y = get_group_id(1);
int lx = get_local_id(0);
int ly = get_local_id(1);
__local T tile_first[TILE_SIZE * LDS_STEP];
__local T tile_second[TILE_SIZE * LDS_STEP];
int total_pixels = rows * cols;
int half_pixels = (total_pixels + 1) / 2;
int x_first = gp_x * TILE_SIZE + lx;
int y_first = gp_y * TILE_SIZE + ly;
int x_second = cols - 1 - x_first;
int y_second = rows - 1 - y_first;
#pragma unroll
for (int i = 0; i < TILE_SIZE; i += BLOCK_ROWS)
{
int curr_y_first = y_first + i;
int curr_y_second = rows - 1 - curr_y_first;
int linear_idx = curr_y_first * cols + x_first;
if (x_first < cols && curr_y_first < rows && linear_idx < half_pixels)
{
T val_first = loadpix(srcptr + mad24(curr_y_first, src_step, mad24(x_first, TSIZE, src_offset)));
T val_second = loadpix(srcptr + mad24(curr_y_second, src_step, mad24(x_second, TSIZE, src_offset)));
#if kercn == 2
#if cn == 1
val_first = val_first.s10;
val_second = val_second.s10;
#endif
#elif kercn == 4
#if cn == 1
val_first = val_first.s3210;
val_second = val_second.s3210;
#elif cn == 2
val_first = val_first.s2301;
val_second = val_second.s2301;
#endif
#endif
tile_first[mad24(ly + i, LDS_STEP, lx)] = val_first;
tile_second[mad24(ly + i, LDS_STEP, lx)] = val_second;
}
}
barrier(CLK_LOCAL_MEM_FENCE);
#pragma unroll
for (int i = 0; i < TILE_SIZE; i += BLOCK_ROWS)
{
int curr_y_first = y_first + i;
int curr_y_second = rows - 1 - curr_y_first;
int linear_idx = curr_y_first * cols + x_first;
if (x_first < cols && curr_y_first < rows && linear_idx < half_pixels)
{
storepix(tile_second[mad24(ly + i, LDS_STEP, lx)],
srcptr + mad24(curr_y_first, src_step, mad24(x_first, TSIZE, src_offset)));
if (linear_idx != total_pixels - 1 - linear_idx)
{
storepix(tile_first[mad24(ly + i, LDS_STEP, lx)],
srcptr + mad24(curr_y_second, src_step, mad24(x_second, TSIZE, src_offset)));
}
}
}
}
__kernel void arithm_flip_cols_inplace(__global uchar * srcptr, int src_step, int src_offset,
int rows, int cols)
{
int gp_x = get_group_id(0);
int gp_y = get_group_id(1);
int lx = get_local_id(0);
int ly = get_local_id(1);
__local T tile_left[TILE_SIZE * LDS_STEP];
__local T tile_right[TILE_SIZE * LDS_STEP];
int half_cols = (cols + 1) / 2;
int x_left = gp_x * TILE_SIZE + lx;
int y = gp_y * TILE_SIZE + ly;
int x_right = cols - 1 - x_left;
#pragma unroll
for (int i = 0; i < TILE_SIZE; i += BLOCK_ROWS)
{
if (y + i < rows && x_left < half_cols)
{
T val_left = loadpix(srcptr + mad24(y + i, src_step, mad24(x_left, TSIZE, src_offset)));
T val_right = loadpix(srcptr + mad24(y + i, src_step, mad24(x_right, TSIZE, src_offset)));
#if kercn == 2
#if cn == 1
val_left = val_left.s10;
val_right = val_right.s10;
#endif
#elif kercn == 4
#if cn == 1
val_left = val_left.s3210;
val_right = val_right.s3210;
#elif cn == 2
val_left = val_left.s2301;
val_right = val_right.s2301;
#endif
#endif
tile_left[mad24(ly + i, LDS_STEP, lx)] = val_left;
tile_right[mad24(ly + i, LDS_STEP, lx)] = val_right;
}
}
barrier(CLK_LOCAL_MEM_FENCE);
#pragma unroll
for (int i = 0; i < TILE_SIZE; i += BLOCK_ROWS)
{
if (y + i < rows && x_left < half_cols)
{
storepix(tile_right[mad24(ly + i, LDS_STEP, lx)],
srcptr + mad24(y + i, src_step, mad24(x_left, TSIZE, src_offset)));
if (x_left != x_right)
{
storepix(tile_left[mad24(ly + i, LDS_STEP, lx)],
srcptr + mad24(y + i, src_step, mad24(x_right, TSIZE, src_offset)));
}
}
}
}
#endif // INPLACE
modules/core/src/reduce.dispatch.cpp
+30
View File
@@ -0,0 +1,30 @@
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html
#include "precomp.hpp"
#include "reduce.simd.hpp"
#include "reduce.simd_declarations.hpp"
namespace cv {
typedef void (*ReduceSumFunc)(const Mat& src, Mat& dst);
ReduceSumFunc getReduceCSumFunc(int sdepth, int ddepth);
ReduceSumFunc getReduceRSumFunc(int sdepth, int ddepth);
ReduceSumFunc getReduceCSumFunc(int sdepth, int ddepth)
{
CV_INSTRUMENT_REGION();
CV_CPU_DISPATCH(getReduceCSumFunc, (sdepth, ddepth),
CV_CPU_DISPATCH_MODES_ALL);
}
ReduceSumFunc getReduceRSumFunc(int sdepth, int ddepth)
{
CV_INSTRUMENT_REGION();
CV_CPU_DISPATCH(getReduceRSumFunc, (sdepth, ddepth),
CV_CPU_DISPATCH_MODES_ALL);
}
} // namespace cv
modules/core/src/reduce.simd.hpp
+1133
View File
File diff suppressed because it is too large Load Diff