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

Removed storing of zeros in local memory

Browse Source
This commit is contained in:
Alexander Karsakov 2014-06-11 15:15:15 +04:00
parent 8e54845049
commit 06fb5da7c8
2 changed files with 49 additions and 122 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

155
modules/imgproc/src/opencl/pyr_up.cl
View File

@ -78,28 +78,19 @@ __kernel void pyrUp(__global const uchar * src, int src_step, int src_offset, in
const int x = get_global_id(0);
const int y = get_global_id(1);
const int lsizex = get_local_size(0);
const int lsizey = get_local_size(1);
const int tidx = get_local_id(0);
const int tidy = get_local_id(1);
__local FT s_srcPatch[10][10];
__local FT s_dstPatch[20][16];
__local FT s_srcPatch[LOCAL_SIZE/2 + 2][LOCAL_SIZE/2 + 2];
__local FT s_dstPatch[LOCAL_SIZE/2 + 2][LOCAL_SIZE];
__global uchar * dstData = dst + dst_offset;
__global const uchar * srcData = src + src_offset;
if( tidx < 10 && tidy < 10 )
if( tidx < (LOCAL_SIZE/2 + 2) && tidy < LOCAL_SIZE/2 + 2 )
{
int srcx = mad24((int)get_group_id(0), lsizex>>1, tidx) - 1;
int srcy = mad24((int)get_group_id(1), lsizey>>1, tidy) - 1;
srcx = abs(srcx);
srcx = min(src_cols - 1, srcx);
srcy = abs(srcy);
srcy = min(src_rows - 1, srcy);
int srcx = EXTRAPOLATE(mad24((int)get_group_id(0), LOCAL_SIZE/2, tidx) - 1, src_cols);
int srcy = EXTRAPOLATE(mad24((int)get_group_id(1), LOCAL_SIZE/2, tidy) - 1, src_rows);
s_srcPatch[tidy][tidx] = convertToFT(loadpix(srcData + srcy * src_step + srcx * PIXSIZE));
}
@ -107,63 +98,30 @@ __kernel void pyrUp(__global const uchar * src, int src_step, int src_offset, in
barrier(CLK_LOCAL_MEM_FENCE);
FT sum = 0.f;
const FT evenFlag = (FT)((tidx & 1) == 0);
const FT oddFlag = (FT)((tidx & 1) != 0);
const bool eveny = ((tidy & 1) == 0);
const FT co1 = 0.75f;
const FT co2 = 0.5f;
const FT co3 = 0.125f;
if(eveny)
const FT coef1 = (tidx & 1) == 0 ? co1 : (FT) 0;
const FT coef2 = (tidx & 1) == 0 ? co3 : co2;
const FT coefy1 = (tidy & 1) == 0 ? co1 : (FT) 0;
const FT coefy2 = (tidy & 1) == 0 ? co3 : co2;
if(tidy < LOCAL_SIZE/2 + 2)
{
sum = ( evenFlag* co3 ) * s_srcPatch[1 + (tidy >> 1)][1 + ((tidx - 2) >> 1)];
sum = sum + ( oddFlag * co2 ) * s_srcPatch[1 + (tidy >> 1)][1 + ((tidx - 1) >> 1)];
sum = sum + ( evenFlag* co1 ) * s_srcPatch[1 + (tidy >> 1)][1 + ((tidx ) >> 1)];
sum = sum + ( oddFlag * co2 ) * s_srcPatch[1 + (tidy >> 1)][1 + ((tidx + 1) >> 1)];
sum = sum + ( evenFlag* co3 ) * s_srcPatch[1 + (tidy >> 1)][1 + ((tidx + 2) >> 1)];
}
s_dstPatch[2 + tidy][tidx] = sum;
if (tidy < 2)
{
sum = 0;
if (eveny)
{
sum = (evenFlag * co3 ) * s_srcPatch[lsizey-16][1 + ((tidx - 2) >> 1)];
sum = sum + ( oddFlag * co2 ) * s_srcPatch[lsizey-16][1 + ((tidx - 1) >> 1)];
sum = sum + (evenFlag * co1 ) * s_srcPatch[lsizey-16][1 + ((tidx ) >> 1)];
sum = sum + ( oddFlag * co2 ) * s_srcPatch[lsizey-16][1 + ((tidx + 1) >> 1)];
sum = sum + (evenFlag * co3 ) * s_srcPatch[lsizey-16][1 + ((tidx + 2) >> 1)];
}
sum = coef2* s_srcPatch[tidy][1 + ((tidx - 1) >> 1)];
sum = mad(coef1, s_srcPatch[tidy][1 + ((tidx ) >> 1)], sum);
sum = mad(coef2, s_srcPatch[tidy][1 + ((tidx + 2) >> 1)], sum);
s_dstPatch[tidy][tidx] = sum;
}
if (tidy > 13)
{
sum = 0;
if (eveny)
{
sum = (evenFlag * co3) * s_srcPatch[lsizey-7][1 + ((tidx - 2) >> 1)];
sum = sum + ( oddFlag * co2) * s_srcPatch[lsizey-7][1 + ((tidx - 1) >> 1)];
sum = sum + (evenFlag * co1) * s_srcPatch[lsizey-7][1 + ((tidx ) >> 1)];
sum = sum + ( oddFlag * co2) * s_srcPatch[lsizey-7][1 + ((tidx + 1) >> 1)];
sum = sum + (evenFlag * co3) * s_srcPatch[lsizey-7][1 + ((tidx + 2) >> 1)];
}
s_dstPatch[4 + tidy][tidx] = sum;
}
barrier(CLK_LOCAL_MEM_FENCE);
sum = co3 * s_dstPatch[2 + tidy - 2][tidx];
sum = sum + co2 * s_dstPatch[2 + tidy - 1][tidx];
sum = sum + co1 * s_dstPatch[2 + tidy ][tidx];
sum = sum + co2 * s_dstPatch[2 + tidy + 1][tidx];
sum = sum + co3 * s_dstPatch[2 + tidy + 2][tidx];
sum = coefy2* s_dstPatch[1 + ((tidy - 1) >> 1)][tidx];
sum = mad(coefy1, s_dstPatch[1 + ((tidy ) >> 1)][tidx], sum);
sum = mad(coefy2, s_dstPatch[1 + ((tidy + 2) >> 1)][tidx], sum);
if ((x < dst_cols) && (y < dst_rows))
storepix(convertToT(sum), dstData + y * dst_step + x * PIXSIZE);
@ -177,12 +135,12 @@ __kernel void pyrUp_unrolled(__global const uchar * src, int src_step, int src_o
const int ly = 2*get_local_id(1);
__local FT s_srcPatch[LOCAL_SIZE+2][LOCAL_SIZE+2];
__local FT s_dstPatch[2*LOCAL_SIZE+4][2*LOCAL_SIZE];
__local FT s_dstPatch[LOCAL_SIZE+2][2*LOCAL_SIZE];
__global uchar * dstData = dst + dst_offset;
__global const uchar * srcData = src + src_offset;
if( lx < (LOCAL_SIZE+2) && lx < (LOCAL_SIZE+2) )
if( lx < (LOCAL_SIZE+2) && ly < (LOCAL_SIZE+2) )
{
int srcx = mad24((int)get_group_id(0), LOCAL_SIZE, lx) - 1;
int srcy = mad24((int)get_group_id(1), LOCAL_SIZE, ly) - 1;
@ -209,17 +167,13 @@ __kernel void pyrUp_unrolled(__global const uchar * src, int src_step, int src_o
sum = co3 * s_srcPatch[1 + (ly >> 1)][1 + ((lx - 2) >> 1)];
sum = sum + co1 * s_srcPatch[1 + (ly >> 1)][1 + ((lx ) >> 1)];
sum = sum + co3 * s_srcPatch[1 + (ly >> 1)][1 + ((lx + 2) >> 1)];
s_dstPatch[2 + ly][lx] = sum;
s_dstPatch[1 + get_local_id(1)][lx] = sum;
// (x+1,y)
sum = co2 * s_srcPatch[1 + (ly >> 1)][1 + ((lx + 1 - 1) >> 1)];
sum = sum + co2 * s_srcPatch[1 + (ly >> 1)][1 + ((lx + 1 + 1) >> 1)];
s_dstPatch[2 + ly][lx+1] = sum;
// (x, y+1) (x+1, y+1)
s_dstPatch[2 + ly+1][lx] = 0.f;
s_dstPatch[2 + ly+1][lx+1] = 0.f;
s_dstPatch[1 + get_local_id(1)][lx+1] = sum;
if (ly < 1)
{
@ -227,16 +181,12 @@ __kernel void pyrUp_unrolled(__global const uchar * src, int src_step, int src_o
sum = co3 * s_srcPatch[0][1 + ((lx - 2) >> 1)];
sum = sum + co1 * s_srcPatch[0][1 + ((lx ) >> 1)];
sum = sum + co3 * s_srcPatch[0][1 + ((lx + 2) >> 1)];
s_dstPatch[ly][lx] = sum;
s_dstPatch[0][lx] = sum;
// (x+1,y)
sum = co2 * s_srcPatch[0][1 + ((lx + 1 - 1) >> 1)];
sum = sum + co2 * s_srcPatch[0][1 + ((lx + 1 + 1) >> 1)];
s_dstPatch[ly][lx+1] = sum;
// (x, y+1) (x+1, y+1)
s_dstPatch[ly+1][lx] = 0.f;
s_dstPatch[ly+1][lx+1] = 0.f;
s_dstPatch[0][lx+1] = sum;
}
if (ly > 2*LOCAL_SIZE-3)
@ -245,59 +195,40 @@ __kernel void pyrUp_unrolled(__global const uchar * src, int src_step, int src_o
sum = co3 * s_srcPatch[LOCAL_SIZE+1][1 + ((lx - 2) >> 1)];
sum = sum + co1 * s_srcPatch[LOCAL_SIZE+1][1 + ((lx ) >> 1)];
sum = sum + co3 * s_srcPatch[LOCAL_SIZE+1][1 + ((lx + 2) >> 1)];
s_dstPatch[4 + ly][lx] = sum;
s_dstPatch[LOCAL_SIZE+1][lx] = sum;
// (x+1,y)
sum = co2 * s_srcPatch[LOCAL_SIZE+1][1 + ((lx + 1 - 1) >> 1)];
sum = sum + co2 * s_srcPatch[LOCAL_SIZE+1][1 + ((lx + 1 + 1) >> 1)];
s_dstPatch[4 + ly][lx+1] = sum;
// (x, y+1) (x+1, y+1)
s_dstPatch[4 + ly+1][lx] = 0.f;
s_dstPatch[4 + ly+1][lx+1] = 0.f;
s_dstPatch[LOCAL_SIZE+1][lx+1] = sum;
}
barrier(CLK_LOCAL_MEM_FENCE);
int dst_x = 2*get_global_id(0);
int dst_y = 2*get_global_id(1);
// (x,y)
sum = co3 * s_dstPatch[2 + ly - 2][lx];
sum = sum + co2 * s_dstPatch[2 + ly - 1][lx];
sum = sum + co1 * s_dstPatch[2 + ly ][lx];
sum = sum + co2 * s_dstPatch[2 + ly + 1][lx];
sum = sum + co3 * s_dstPatch[2 + ly + 2][lx];
if ((dst_x < dst_cols) && (dst_y < dst_rows))
{
// (x,y)
sum = co3 * s_dstPatch[1 + get_local_id(1) - 1][lx];
sum = sum + co1 * s_dstPatch[1 + get_local_id(1) ][lx];
sum = sum + co3 * s_dstPatch[1 + get_local_id(1) + 1][lx];
storepix(convertToT(sum), dstData + dst_y * dst_step + dst_x * PIXSIZE);
// (x+1,y)
sum = co3 * s_dstPatch[2 + ly - 2][lx+1];
sum = sum + co2 * s_dstPatch[2 + ly - 1][lx+1];
sum = sum + co1 * s_dstPatch[2 + ly ][lx+1];
sum = sum + co2 * s_dstPatch[2 + ly + 1][lx+1];
sum = sum + co3 * s_dstPatch[2 + ly + 2][lx+1];
if ((dst_x+1 < dst_cols) && (dst_y < dst_rows))
// (x+1,y)
sum = co3 * s_dstPatch[1 + get_local_id(1) - 1][lx+1];
sum = sum + co1 * s_dstPatch[1 + get_local_id(1) ][lx+1];
sum = sum + co3 * s_dstPatch[1 + get_local_id(1) + 1][lx+1];
storepix(convertToT(sum), dstData + dst_y * dst_step + (dst_x+1) * PIXSIZE);
// (x,y+1)
sum = co3 * s_dstPatch[2 + ly+1 - 2][lx];
sum = sum + co2 * s_dstPatch[2 + ly+1 - 1][lx];
sum = sum + co1 * s_dstPatch[2 + ly+1 ][lx];
sum = sum + co2 * s_dstPatch[2 + ly+1 + 1][lx];
sum = sum + co3 * s_dstPatch[2 + ly+1 + 2][lx];
if ((dst_x < dst_cols) && (dst_y+1 < dst_rows))
// (x,y+1)
sum = co2 * s_dstPatch[1 + get_local_id(1) ][lx];
sum = sum + co2 * s_dstPatch[1 + get_local_id(1) + 1][lx];
storepix(convertToT(sum), dstData + (dst_y+1) * dst_step + dst_x * PIXSIZE);
// (x+1,y+1)
sum = co3 * s_dstPatch[2 + ly+1 - 2][lx+1];
sum = sum + co2 * s_dstPatch[2 + ly+1 - 1][lx+1];
sum = sum + co1 * s_dstPatch[2 + ly+1 ][lx+1];
sum = sum + co2 * s_dstPatch[2 + ly+1 + 1][lx+1];
sum = sum + co3 * s_dstPatch[2 + ly+1 + 2][lx+1];
if ((dst_x+1 < dst_cols) && (dst_y+1 < dst_rows))
// (x+1,y+1)
sum = co2 * s_dstPatch[1 + get_local_id(1) ][lx+1];
sum = sum + co2 * s_dstPatch[1 + get_local_id(1) + 1][lx+1];
storepix(convertToT(sum), dstData + (dst_y+1) * dst_step + (dst_x+1) * PIXSIZE);
}
}

16
modules/imgproc/src/pyramids.cpp
View File

@ -467,7 +467,7 @@ static bool ocl_pyrUp( InputArray _src, OutputArray _dst, const Size& _dsz, int
UMat dst = _dst.getUMat();
int float_depth = depth == CV_64F ? CV_64F : CV_32F;
int local_size = 8;
const int local_size = 16;
char cvt[2][50];
String buildOptions = format(
"-D T=%s -D FT=%s -D convertToT=%s -D convertToFT=%s%s "
@ -478,25 +478,21 @@ static bool ocl_pyrUp( InputArray _src, OutputArray _dst, const Size& _dsz, int
doubleSupport ? " -D DOUBLE_SUPPORT" : "",
ocl::typeToStr(depth), channels, local_size
);
size_t globalThreads[2];
size_t globalThreads[2] = { dst.cols, dst.rows };
size_t localThreads[2] = { local_size, local_size };
ocl::Kernel k;
if (ocl::Device::getDefault().isIntel() && channels == 1)
{
k.create("pyrUp_unrolled", ocl::imgproc::pyr_up_oclsrc, buildOptions);
globalThreads[0] = dst.cols/2; globalThreads[1] = dst.rows/2;
}
else
{
k.create("pyrUp", ocl::imgproc::pyr_up_oclsrc, buildOptions);
local_size = 16;
globalThreads[0] = dst.cols; globalThreads[1] = dst.rows;
}
else
k.create("pyrUp", ocl::imgproc::pyr_up_oclsrc, buildOptions);
if (k.empty())
return false;
k.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::WriteOnly(dst));
size_t localThreads[2] = {local_size, local_size};
return k.run(2, globalThreads, localThreads, false);
}