mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2024-11-28 21:20:18 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

increased number of rows per work-item

Browse Source
This commit is contained in:
Ilya Lavrenov 2014-06-17 19:24:25 +04:00
parent 45f21e4f12
commit c41a134394
2 changed files with 248 additions and 205 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
modules/imgproc/src/imgwarp.cpp
View File

@ -3582,7 +3582,9 @@ private:
static bool ocl_remap(InputArray _src, OutputArray _dst, InputArray _map1, InputArray _map2,
int interpolation, int borderType, const Scalar& borderValue)
{
int cn = _src.channels(), type = _src.type(), depth = _src.depth();
const ocl::Device & dev = ocl::Device::getDefault();
int cn = _src.channels(), type = _src.type(), depth = _src.depth(),
rowsPerWI = dev.isIntel() ? 4 : 1;
if (borderType == BORDER_TRANSPARENT || !(interpolation == INTER_LINEAR || interpolation == INTER_NEAREST)
|| _map1.type() == CV_16SC1 || _map2.type() == CV_16SC1)
@ -3619,12 +3621,14 @@ static bool ocl_remap(InputArray _src, OutputArray _dst, InputArray _map1, Input
static const char * const interMap[] = { "INTER_NEAREST", "INTER_LINEAR", "INTER_CUBIC", "INTER_LINEAR", "INTER_LANCZOS" };
static const char * const borderMap[] = { "BORDER_CONSTANT", "BORDER_REPLICATE", "BORDER_REFLECT", "BORDER_WRAP",
"BORDER_REFLECT_101", "BORDER_TRANSPARENT" };
String buildOptions = format("-D %s -D %s -D T=%s", interMap[interpolation], borderMap[borderType], ocl::typeToStr(type));
String buildOptions = format("-D %s -D %s -D T=%s -D rowsPerWI=%d",
interMap[interpolation], borderMap[borderType],
ocl::typeToStr(type), rowsPerWI);
if (interpolation != INTER_NEAREST)
{
char cvt[3][40];
int wdepth = std::max(CV_32F, dst.depth());
int wdepth = std::max(CV_32F, depth);
buildOptions = buildOptions
+ format(" -D WT=%s -D convertToT=%s -D convertToWT=%s"
" -D convertToWT2=%s -D WT2=%s",
@ -3653,7 +3657,7 @@ static bool ocl_remap(InputArray _src, OutputArray _dst, InputArray _map1, Input
else
k.args(srcarg, dstarg, map1arg, ocl::KernelArg::ReadOnlyNoSize(map2), scalararg);
size_t globalThreads[2] = { dst.cols, dst.rows };
size_t globalThreads[2] = { dst.cols, (dst.rows + rowsPerWI - 1) / rowsPerWI };
return k.run(2, globalThreads, NULL, false);
}

123
modules/imgproc/src/opencl/remap.cl
View File

@ -147,16 +147,21 @@ __kernel void remap_2_32FC1(__global const uchar * srcptr, int src_step, int src
ST nVal)
{
int x = get_global_id(0);
int y = get_global_id(1);
int y = get_global_id(1) * rowsPerWI;
if (x < dst_cols)
{
T scalar = convertScalar(nVal);
if (x < dst_cols && y < dst_rows)
{
int map1_index = mad24(y, map1_step, x * (int)sizeof(float) + map1_offset);
int map2_index = mad24(y, map2_step, x * (int)sizeof(float) + map2_offset);
int dst_index = mad24(y, dst_step, x * TSIZE + dst_offset);
int map1_index = mad24(y, map1_step, mad24(x, (int)sizeof(float), map1_offset));
int map2_index = mad24(y, map2_step, mad24(x, (int)sizeof(float), map2_offset));
int dst_index = mad24(y, dst_step, mad24(x, TSIZE, dst_offset));
#pragma unroll
for (int i = 0; i < rowsPerWI; ++i, ++y,
map1_index += map1_step, map2_index += map2_step, dst_index += dst_step)
if (y < dst_rows)
{
__global const float * map1 = (__global const float *)(map1ptr + map1_index);
__global const float * map2 = (__global const float *)(map2ptr + map2_index);
__global T * dst = (__global T *)(dstptr + dst_index);
@ -175,10 +180,11 @@ __kernel void remap_2_32FC1(__global const uchar * srcptr, int src_step, int src
}
else
{
int src_index = mad24(gy, src_step, gx * TSIZE + src_offset);
int src_index = mad24(gy, src_step, mad24(gx, TSIZE, src_offset));
storepix(loadpix((__global const T*)(srcptr + src_index)), dst);
}
}
}
}
__kernel void remap_32FC2(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols,
@ -187,15 +193,19 @@ __kernel void remap_32FC2(__global const uchar * srcptr, int src_step, int src_o
ST nVal)
{
int x = get_global_id(0);
int y = get_global_id(1);
int y = get_global_id(1) * rowsPerWI;
T scalar = convertScalar(nVal);
if (x < dst_cols && y < dst_rows)
if (x < dst_cols)
{
int dst_index = mad24(y, dst_step, x * TSIZE + dst_offset);
int map_index = mad24(y, map_step, x * (int)sizeof(float2) + map_offset);
T scalar = convertScalar(nVal);
int dst_index = mad24(y, dst_step, mad24(x, TSIZE, dst_offset));
int map_index = mad24(y, map_step, mad24(x, (int)sizeof(float2), map_offset));
#pragma unroll
for (int i = 0; i < rowsPerWI; ++i, ++y,
map_index += map_step, dst_index += dst_step)
if (y < dst_rows)
{
__global const float2 * map = (__global const float2 *)(mapptr + map_index);
__global T * dst = (__global T *)(dstptr + dst_index);
@ -210,10 +220,11 @@ __kernel void remap_32FC2(__global const uchar * srcptr, int src_step, int src_o
}
else
{
int src_index = mad24(gy, src_step, gx * TSIZE + src_offset);
int src_index = mad24(gy, src_step, mad24(gx, TSIZE, src_offset));
storepix(loadpix((__global const T *)(srcptr + src_index)), dst);
}
}
}
}
__kernel void remap_16SC2(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols,
@ -222,15 +233,19 @@ __kernel void remap_16SC2(__global const uchar * srcptr, int src_step, int src_o
ST nVal)
{
int x = get_global_id(0);
int y = get_global_id(1);
int y = get_global_id(1) * rowsPerWI;
T scalar = convertScalar(nVal);
if (x < dst_cols && y < dst_rows)
if (x < dst_cols)
{
int dst_index = mad24(y, dst_step, x * TSIZE + dst_offset);
int map_index = mad24(y, map_step, x * (int)sizeof(short2) + map_offset);
T scalar = convertScalar(nVal);
int dst_index = mad24(y, dst_step, mad24(x, TSIZE, dst_offset));
int map_index = mad24(y, map_step, mad24(x, (int)sizeof(short2), map_offset));
#pragma unroll
for (int i = 0; i < rowsPerWI; ++i, ++y,
map_index += map_step, dst_index += dst_step)
if (y < dst_rows)
{
__global const short2 * map = (__global const short2 *)(mapptr + map_index);
__global T * dst = (__global T *)(dstptr + dst_index);
@ -245,10 +260,11 @@ __kernel void remap_16SC2(__global const uchar * srcptr, int src_step, int src_o
}
else
{
int src_index = mad24(gy, src_step, gx * TSIZE + src_offset);
int src_index = mad24(gy, src_step, mad24(gx, TSIZE, src_offset));
storepix(loadpix((__global const T *)(srcptr + src_index)), dst);
}
}
}
}
__kernel void remap_16SC2_16UC1(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols,
@ -260,14 +276,18 @@ __kernel void remap_16SC2_16UC1(__global const uchar * srcptr, int src_step, int
int x = get_global_id(0);
int y = get_global_id(1);
T scalar = convertScalar(nVal);
if (x < dst_cols && y < dst_rows)
if (x < dst_cols)
{
int dst_index = mad24(y, dst_step, x * TSIZE + dst_offset);
int map1_index = mad24(y, map1_step, x * (int)sizeof(short2) + map1_offset);
int map2_index = mad24(y, map2_step, x * (int)sizeof(ushort) + map2_offset);
T scalar = convertScalar(nVal);
int dst_index = mad24(y, dst_step, mad24(x, TSIZE, dst_offset));
int map1_index = mad24(y, map1_step, mad24(x, (int)sizeof(short2), map1_offset));
int map2_index = mad24(y, map2_step, mad24(x, (int)sizeof(ushort), map2_offset));
#pragma unroll
for (int i = 0; i < rowsPerWI; ++i, ++y,
map1_index += map1_step, map2_index += map2_step, dst_index += dst_step)
if (y < dst_rows)
{
__global const short2 * map1 = (__global const short2 *)(map1ptr + map1_index);
__global const ushort * map2 = (__global const ushort *)(map2ptr + map2_index);
__global T * dst = (__global T *)(dstptr + dst_index);
@ -286,10 +306,11 @@ __kernel void remap_16SC2_16UC1(__global const uchar * srcptr, int src_step, int
}
else
{
int src_index = mad24(gy, src_step, gx * TSIZE + src_offset);
int src_index = mad24(gy, src_step, mad24(gx, TSIZE, src_offset));
storepix(loadpix((__global const T *)(srcptr + src_index)), dst);
}
}
}
}
#elif INTER_LINEAR
@ -301,14 +322,19 @@ __kernel void remap_16SC2_16UC1(__global const uchar * srcptr, int src_step, int
ST nVal)
{
int x = get_global_id(0);
int y = get_global_id(1);
int y = get_global_id(1) * rowsPerWI;
if (x < dst_cols && y < dst_rows)
if (x < dst_cols)
{
int dst_index = mad24(y, dst_step, x * TSIZE + dst_offset);
int map1_index = mad24(y, map1_step, x * (int)sizeof(short2) + map1_offset);
int map2_index = mad24(y, map2_step, x * (int)sizeof(ushort) + map2_offset);
int dst_index = mad24(y, dst_step, mad24(x, TSIZE, dst_offset));
int map1_index = mad24(y, map1_step, mad24(x, (int)sizeof(short2), map1_offset));
int map2_index = mad24(y, map2_step, mad24(x, (int)sizeof(ushort), map2_offset));
#pragma unroll
for (int i = 0; i < rowsPerWI; ++i, ++y,
map1_index += map1_step, map2_index += map2_step, dst_index += dst_step)
if (y < dst_rows)
{
__global const short2 * map1 = (__global const short2 *)(map1ptr + map1_index);
__global const ushort * map2 = (__global const ushort *)(map2ptr + map2_index);
__global T * dst = (__global T *)(dstptr + dst_index);
@ -350,6 +376,7 @@ __kernel void remap_16SC2_16UC1(__global const uchar * srcptr, int src_step, int
d * (u.x) * (u.y);
storepix(convertToT(dst_data), dst);
}
}
}
__kernel void remap_2_32FC1(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols,
@ -359,14 +386,19 @@ __kernel void remap_2_32FC1(__global const uchar * srcptr, int src_step, int src
ST nVal)
{
int x = get_global_id(0);
int y = get_global_id(1);
int y = get_global_id(1) * rowsPerWI;
if (x < dst_cols && y < dst_rows)
if (x < dst_cols)
{
int dst_index = mad24(y, dst_step, x * TSIZE + dst_offset);
int map1_index = mad24(y, map1_step, x * (int)sizeof(float) + map1_offset);
int map2_index = mad24(y, map2_step, x * (int)sizeof(float) + map2_offset);
int dst_index = mad24(y, dst_step, mad24(x, TSIZE, dst_offset));
int map1_index = mad24(y, map1_step, mad24(x, (int)sizeof(float), map1_offset));
int map2_index = mad24(y, map2_step, mad24(x, (int)sizeof(float), map2_offset));
#pragma unroll
for (int i = 0; i < rowsPerWI; ++i, ++y,
map1_index += map1_step, map2_index += map2_step, dst_index += dst_step)
if (y < dst_rows)
{
__global const float * map1 = (__global const float *)(map1ptr + map1_index);
__global const float * map2 = (__global const float *)(map2ptr + map2_index);
__global T * dst = (__global T *)(dstptr + dst_index);
@ -409,6 +441,7 @@ __kernel void remap_2_32FC1(__global const uchar * srcptr, int src_step, int src
d * (u.x) * (u.y);
storepix(convertToT(dst_data), dst);
}
}
}
__kernel void remap_32FC2(__global const uchar * srcptr, int src_step, int src_offset, int src_rows, int src_cols,
@ -417,13 +450,18 @@ __kernel void remap_32FC2(__global const uchar * srcptr, int src_step, int src_o
ST nVal)
{
int x = get_global_id(0);
int y = get_global_id(1);
int y = get_global_id(1) * rowsPerWI;
if (x < dst_cols && y < dst_rows)
if (x < dst_cols)
{
int dst_index = mad24(y, dst_step, x * TSIZE + dst_offset);
int map_index = mad24(y, map_step, x * (int)sizeof(float2) + map_offset);
int dst_index = mad24(y, dst_step, mad24(x, TSIZE, dst_offset));
int map_index = mad24(y, map_step, mad24(x, (int)sizeof(float2), map_offset));
#pragma unroll
for (int i = 0; i < rowsPerWI; ++i, ++y,
map_index += map_step, dst_index += dst_step)
if (y < dst_rows)
{
__global const float2 * map = (__global const float2 *)(mapptr + map_index);
__global T * dst = (__global T *)(dstptr + dst_index);
@ -464,6 +502,7 @@ __kernel void remap_32FC2(__global const uchar * srcptr, int src_step, int src_o
d * (u.x) * (u.y);
storepix(convertToT(dst_data), dst);
}
}
}
#endif