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

Fix retina ocl on NV.

Browse Source 
Previously we incorrectly assumed data was aligned in 16 bytes, which should be 32.
This commit is contained in:
pengxiao 2013-09-10 14:33:23 +08:00
parent e45f92a9e1
commit a92721b569
1 changed files with 13 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
modules/bioinspired/src/opencl/retina_kernel.cl
View File

@ -43,6 +43,9 @@
//
//M*/
//data (which is float) is aligend in 32 bytes
#define WIDTH_MULTIPLE (32 >> 2)
/////////////////////////////////////////////////////////
//*******************************************************
// basicretinafilter
@ -116,22 +119,18 @@ kernel void horizontalAnticausalFilter(
float4 result_v4 = (float4)(0), out_v4;
float result = 0;
// we assume elements_per_row is multple of 4
for(int i = 0; i < 4; ++ i, -- optr)
// we assume elements_per_row is multple of WIDTH_MULTIPLE
for(int i = 0; i < WIDTH_MULTIPLE; ++ i, -- optr)
{
if(i < elements_per_row - cols)
{
*optr = result;
}
else
if(i >= elements_per_row - cols)
{
result = *optr + _a * result;
*optr = result;
}
*optr = result;
}
result_v4.x = result;
optr -= 3;
for(int i = 1; i < elements_per_row / 4; ++i, optr -= 4)
for(int i = WIDTH_MULTIPLE / 4; i < elements_per_row / 4; ++i, optr -= 4)
{
// shift left, `offset` is type `size_t` so it cannot be negative
out_v4 = vload4(0, optr);
@ -223,23 +222,19 @@ kernel void horizontalAnticausalFilter_Irregular(
float4 buf_v4, out_v4, res_v4 = (float4)(0);
float result = 0;
// we assume elements_per_row is multple of 4
for(int i = 0; i < 4; ++ i, -- optr, -- bptr)
// we assume elements_per_row is multple of WIDTH_MULTIPLE
for(int i = 0; i < WIDTH_MULTIPLE; ++ i, -- optr, -- bptr)
{
if(i < elements_per_row - cols)
{
*optr = result;
}
else
if(i >= elements_per_row - cols)
{
result = *optr + *bptr * result;
*optr = result;
}
*optr = result;
}
res_v4.x = result;
optr -= 3;
bptr -= 3;
for(int i = 0; i < elements_per_row / 4 - 1; ++i, optr -= 4, bptr -= 4)
for(int i = WIDTH_MULTIPLE / 4; i < elements_per_row / 4; ++i, optr -= 4, bptr -= 4)
{
buf_v4 = vload4(0, bptr);
out_v4 = vload4(0, optr);