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

Merge pull request #2851 from ilya-lavrenov:tapi_set_identity

Browse Source
This commit is contained in:
Alexander Alekhin
2014-07-11 10:00:15 +00:00
parent 910d8f8e0b 6dd658a0af
commit cbf63b0709
2 changed files with 48 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
modules/core/src/opencl/set_identity.cl
View File

@@ -43,20 +43,18 @@
//
//M*/
#if cn != 3
#define loadpix(addr) *(__global const T *)(addr)
#if kercn != 3
#define storepix(val, addr) *(__global T *)(addr) = val
#define TSIZE (int)sizeof(T)
#define scalar scalar_
#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)
#define scalar (T)(scalar_.x, scalar_.y, scalar_.z)
#endif
__kernel void setIdentity(__global uchar * srcptr, int src_step, int src_offset, int rows, int cols,
ST scalar_, int rowsPerWI)
ST scalar_)
{
int x = get_global_id(0);
int y0 = get_global_id(1) * rowsPerWI;
@@ -65,7 +63,35 @@ __kernel void setIdentity(__global uchar * srcptr, int src_step, int src_offset,
{
int src_index = mad24(y0, src_step, mad24(x, TSIZE, src_offset));
for (int y = y0, y1 = min(rows, y0 + rowsPerWI); y < y1; ++y, src_index += src_step)
storepix(x == y ? scalar : (T)(0), srcptr + src_index);
#if kercn == cn
#pragma unroll
for (int y = y0, i = 0, y1 = min(rows, y0 + rowsPerWI); i < rowsPerWI; ++y, ++i, src_index += src_step)
if (y < y1)
storepix(x == y ? scalar : (T)(0), srcptr + src_index);
#elif kercn == 4 && cn == 1
if (y0 < rows)
{
storepix(x == y0 >> 2 ? (T)(scalar, 0, 0, 0) : (T)(0), srcptr + src_index);
if (++y0 < rows)
{
src_index += src_step;
storepix(x == y0 >> 2 ? (T)(0, scalar, 0, 0) : (T)(0), srcptr + src_index);
if (++y0 < rows)
{
src_index += src_step;
storepix(x == y0 >> 2 ? (T)(0, 0, scalar, 0) : (T)(0), srcptr + src_index);
if (++y0 < rows)
{
src_index += src_step;
storepix(x == y0 >> 2 ? (T)(0, 0, 0, scalar) : (T)(0), srcptr + src_index);
}
}
}
}
#else
#error "Incorrect combination of cn && kercn"
#endif
}
}