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

some optimizations to ocl::pyrDown, PyrLK and Canny

Browse Source
This commit is contained in:
bitwangyaoyao
2012-09-24 20:28:35 +08:00
parent 494ae1562d
commit 09359982b1
9 changed files with 1095 additions and 464 deletions
Download Patch File Download Diff File Expand all files Collapse all files

510
modules/ocl/src/kernels/pyr_down.cl
View File

@@ -43,14 +43,9 @@
//
//M*/
#pragma OPENCL EXTENSION cl_amd_printf : enable
//#pragma OPENCL EXTENSION cl_amd_printf : enable
uchar round_uchar_uchar(uchar v)
{
return v;
}
uchar round_uchar_int(int v)
{
return (uchar)((uint)v <= 255 ? v : v > 0 ? 255 : 0);
@@ -58,13 +53,7 @@ uchar round_uchar_int(int v)
uchar round_uchar_float(float v)
{
int iv = convert_int_sat_rte(v);
return round_uchar_int(iv);
}
uchar4 round_uchar4_uchar4(uchar4 v)
{
return v;
return round_uchar_int(convert_int_sat_rte(v));
}
uchar4 round_uchar4_int4(int4 v)
@@ -79,52 +68,45 @@ uchar4 round_uchar4_int4(int4 v)
uchar4 round_uchar4_float4(float4 v)
{
int4 iv = convert_int4_sat_rte(v);
return round_uchar4_int4(iv);
return round_uchar4_int4(convert_int4_sat_rte(v));
}
int idx_row_low(int y, int last_row)
{
return abs(y) % (last_row + 1);
}
int idx_row_high(int y, int last_row)
{
int i=abs_diff(y,last_row);
int j=abs_diff(i,last_row);
return j % (last_row + 1);
}
int idx_row(int y, int last_row)
{
return idx_row_low(idx_row_high(y, last_row), last_row);
}
int idx_col_low(int x, int last_col)
{
return abs(x) % (last_col + 1);
}
int idx_col_high(int x, int last_col)
{
int i=abs_diff(x,last_col);
int j=abs_diff(i,last_col);
return j % (last_col + 1);
}
int idx_col(int x, int last_col)
{
return idx_col_low(idx_col_high(x, last_col), last_col);
}
int idx_row_low(int y, int last_row)
{
{
return abs(y) % (last_row + 1);
}
int idx_row_high(int y, int last_row)
{
return abs(last_row - (int)abs(last_row - y)) % (last_row + 1);
}
int idx_row(int y, int last_row)
{
return idx_row_low(idx_row_high(y, last_row), last_row);
}
int idx_col_low(int x, int last_col)
{
return abs(x) % (last_col + 1);
}
int idx_col_high(int x, int last_col)
{
return abs(last_col - (int)abs(last_col - x)) % (last_col + 1);
}
int idx_col(int x, int last_col)
{
return idx_col_low(idx_col_high(x, last_col), last_col);
}
__kernel void pyrDown_C1_D0(__global uchar * srcData, int srcStep, int srcRows, int srcCols, __global uchar *dst, int dstStep, int dstCols)
{
const int x = get_global_id(0);
const int y = get_group_id(1);
__local float smem[256 + 4];
@@ -135,44 +117,83 @@ __kernel void pyrDown_C1_D0(__global uchar * srcData, int srcStep, int srcOffset
const int last_row = srcRows - 1;
const int last_col = srcCols - 1;
sum = 0;
sum = sum + 0.0625f * (((srcData + idx_row(src_y - 2, last_row) * srcStep))[idx_col(x, last_col)]);
sum = sum + 0.25f * (((srcData + idx_row(src_y - 1, last_row) * srcStep))[idx_col(x, last_col)]);
sum = sum + 0.375f * (((srcData + idx_row(src_y , last_row) * srcStep))[idx_col(x, last_col)]);
sum = sum + 0.25f * (((srcData + idx_row(src_y + 1, last_row) * srcStep))[idx_col(x, last_col)]);
sum = sum + 0.0625f * (((srcData + idx_row(src_y + 2, last_row) * srcStep))[idx_col(x, last_col)]);
smem[2 + get_local_id(0)] = sum;
if (src_y >= 2 && src_y < srcRows - 2 && x >= 2 && x < srcCols - 2)
{
{
sum = 0.0625f * (((srcData + (src_y - 2) * srcStep))[x]);
sum = sum + 0.25f * (((srcData + (src_y - 1) * srcStep))[x]);
sum = sum + 0.375f * (((srcData + (src_y ) * srcStep))[x]);
sum = sum + 0.25f * (((srcData + (src_y + 1) * srcStep))[x]);
sum = sum + 0.0625f * (((srcData + (src_y + 2) * srcStep))[x]);
smem[2 + get_local_id(0)] = sum;
sum = sum + 0.0625f * (((srcData + idx_row(src_y - 2, last_row) * srcStep))[idx_col(left_x, last_col)]);
sum = sum + 0.25f * (((srcData + idx_row(src_y - 1, last_row) * srcStep))[idx_col(left_x, last_col)]);
sum = sum + 0.375f * (((srcData + idx_row(src_y , last_row) * srcStep))[idx_col(left_x, last_col)]);
sum = sum + 0.25f * (((srcData + idx_row(src_y + 1, last_row) * srcStep))[idx_col(left_x, last_col)]);
if (get_local_id(0) < 2)
{
const int left_x = x - 2;
sum = 0.0625f * (((srcData + (src_y - 2) * srcStep))[left_x]);
sum = sum + 0.25f * (((srcData + (src_y - 1) * srcStep))[left_x]);
sum = sum + 0.375f * (((srcData + (src_y ) * srcStep))[left_x]);
sum = sum + 0.25f * (((srcData + (src_y + 1) * srcStep))[left_x]);
sum = sum + 0.0625f * (((srcData + (src_y + 2) * srcStep))[left_x]);
smem[get_local_id(0)] = sum;
}
if (get_local_id(0) > 253)
{
const int right_x = x + 2;
sum = 0.0625f * (((srcData + (src_y - 2) * srcStep))[right_x]);
sum = sum + 0.25f * (((srcData + (src_y - 1) * srcStep))[right_x]);
sum = sum + 0.375f * (((srcData + (src_y ) * srcStep))[right_x]);
sum = sum + 0.25f * (((srcData + (src_y + 1) * srcStep))[right_x]);
sum = sum + 0.0625f * (((srcData + (src_y + 2) * srcStep))[right_x]);
smem[4 + get_local_id(0)] = sum;
}
}
}
else
{
{
int col = idx_col(x, last_col);
sum = 0.0625f * (((srcData + idx_row(src_y - 2, last_row) * srcStep))[col]);
sum = sum + 0.25f * (((srcData + idx_row(src_y - 1, last_row) * srcStep))[col]);
sum = sum + 0.375f * (((srcData + idx_row(src_y , last_row) * srcStep))[col]);
sum = sum + 0.25f * (((srcData + idx_row(src_y + 1, last_row) * srcStep))[col]);
sum = sum + 0.0625f * (((srcData + idx_row(src_y + 2, last_row) * srcStep))[col]);
sum = sum + 0.0625f * (((srcData + idx_row(src_y - 2, last_row) * srcStep))[idx_col(right_x, last_col)]);
sum = sum + 0.25f * (((srcData + idx_row(src_y - 1, last_row) * srcStep))[idx_col(right_x, last_col)]);
sum = sum + 0.375f * (((srcData + idx_row(src_y , last_row) * srcStep))[idx_col(right_x, last_col)]);
sum = sum + 0.25f * (((srcData + idx_row(src_y + 1, last_row) * srcStep))[idx_col(right_x, last_col)]);
smem[2 + get_local_id(0)] = sum;
if (get_local_id(0) < 2)
{
const int left_x = x - 2;
col = idx_col(left_x, last_col);
sum = 0.0625f * (((srcData + idx_row(src_y - 2, last_row) * srcStep))[col]);
sum = sum + 0.25f * (((srcData + idx_row(src_y - 1, last_row) * srcStep))[col]);
sum = sum + 0.375f * (((srcData + idx_row(src_y , last_row) * srcStep))[col]);
sum = sum + 0.25f * (((srcData + idx_row(src_y + 1, last_row) * srcStep))[col]);
sum = sum + 0.0625f * (((srcData + idx_row(src_y + 2, last_row) * srcStep))[col]);
smem[get_local_id(0)] = sum;
}
if (get_local_id(0) > 253)
{
const int right_x = x + 2;
col = idx_col(right_x, last_col);
sum = 0.0625f * (((srcData + idx_row(src_y - 2, last_row) * srcStep))[col]);
sum = sum + 0.25f * (((srcData + idx_row(src_y - 1, last_row) * srcStep))[col]);
sum = sum + 0.375f * (((srcData + idx_row(src_y , last_row) * srcStep))[col]);
sum = sum + 0.25f * (((srcData + idx_row(src_y + 1, last_row) * srcStep))[col]);
sum = sum + 0.0625f * (((srcData + idx_row(src_y + 2, last_row) * srcStep))[col]);
smem[4 + get_local_id(0)] = sum;
}
}
barrier(CLK_LOCAL_MEM_FENCE);
@@ -181,9 +202,7 @@ __kernel void pyrDown_C1_D0(__global uchar * srcData, int srcStep, int srcOffset
{
const int tid2 = get_local_id(0) * 2;
sum = 0;
sum = 0.0625f * smem[2 + tid2 - 2];
sum = sum + 0.25f * smem[2 + tid2 - 1];
sum = sum + 0.375f * smem[2 + tid2 ];
sum = sum + 0.25f * smem[2 + tid2 + 1];
@@ -196,9 +215,9 @@ __kernel void pyrDown_C1_D0(__global uchar * srcData, int srcStep, int srcOffset
}
}
__kernel void pyrDown_C4_D0(__global uchar4 * srcData, int srcStep, int srcRows, int srcCols, __global uchar4 *dst, int dstStep, int dstCols)
{
{
const int x = get_global_id(0);
const int y = get_group_id(1);
__local float4 smem[256 + 4];
@@ -209,48 +228,87 @@ __kernel void pyrDown_C4_D0(__global uchar4 * srcData, int srcStep, int srcOffse
const int last_row = srcRows - 1;
const int last_col = srcCols - 1;
float4 co1 = (float4)(0.375f, 0.375f, 0.375f, 0.375f);
float4 co2 = (float4)(0.25f, 0.25f, 0.25f, 0.25f);
float4 co1 = 0.375f;//(float4)(0.375f, 0.375f, 0.375f, 0.375f);
float4 co2 = 0.25f;//(float4)(0.25f, 0.25f, 0.25f, 0.25f);
float4 co3 = 0.0625f;//(float4)(0.0625f, 0.0625f, 0.0625f, 0.0625f);
if (src_y >= 2 && src_y < srcRows - 2 && x >= 2 && x < srcCols - 2)
{
sum = co3 * convert_float4((((srcData + (src_y - 2) * srcStep / 4))[x]));
sum = sum + co2 * convert_float4((((srcData + (src_y - 1) * srcStep / 4))[x]));
sum = sum + co1 * convert_float4((((srcData + (src_y ) * srcStep / 4))[x]));
sum = sum + co2 * convert_float4((((srcData + (src_y + 1) * srcStep / 4))[x]));
sum = sum + co3 * convert_float4((((srcData + (src_y + 2) * srcStep / 4))[x]));
sum = sum + co3 * convert_float4((((srcData + idx_row(src_y - 2, last_row) * srcStep / 4))[idx_col(x, last_col)]));
sum = sum + co2 * convert_float4((((srcData + idx_row(src_y - 1, last_row) * srcStep / 4))[idx_col(x, last_col)]));
sum = sum + co1 * convert_float4((((srcData + idx_row(src_y , last_row) * srcStep / 4))[idx_col(x, last_col)]));
sum = sum + co2 * convert_float4((((srcData + idx_row(src_y + 1, last_row) * srcStep / 4))[idx_col(x, last_col)]));
smem[2 + get_local_id(0)] = sum;
if (get_local_id(0) < 2)
{
const int left_x = x - 2;
if (get_local_id(0) < 2)
{
sum = co3 * convert_float4((((srcData + (src_y - 2) * srcStep / 4))[left_x]));
sum = sum + co2 * convert_float4((((srcData + (src_y - 1) * srcStep / 4))[left_x]));
sum = sum + co1 * convert_float4((((srcData + (src_y ) * srcStep / 4))[left_x]));
sum = sum + co2 * convert_float4((((srcData + (src_y + 1) * srcStep / 4))[left_x]));
sum = sum + co3 * convert_float4((((srcData + (src_y + 2) * srcStep / 4))[left_x]));
smem[get_local_id(0)] = sum;
}
sum = sum + co3 * convert_float4((((srcData + idx_row(src_y - 2, last_row) * srcStep / 4))[idx_col(left_x, last_col)]));
sum = sum + co2 * convert_float4((((srcData + idx_row(src_y - 1, last_row) * srcStep / 4))[idx_col(left_x, last_col)]));
sum = sum + co1 * convert_float4((((srcData + idx_row(src_y , last_row) * srcStep / 4))[idx_col(left_x, last_col)]));
sum = sum + co2 * convert_float4((((srcData + idx_row(src_y + 1, last_row) * srcStep / 4))[idx_col(left_x, last_col)]));
if (get_local_id(0) > 253)
{
const int right_x = x + 2;
sum = co3 * convert_float4((((srcData + (src_y - 2) * srcStep / 4))[right_x]));
sum = sum + co2 * convert_float4((((srcData + (src_y - 1) * srcStep / 4))[right_x]));
sum = sum + co1 * convert_float4((((srcData + (src_y ) * srcStep / 4))[right_x]));
sum = sum + co2 * convert_float4((((srcData + (src_y + 1) * srcStep / 4))[right_x]));
sum = sum + co3 * convert_float4((((srcData + (src_y + 2) * srcStep / 4))[right_x]));
smem[4 + get_local_id(0)] = sum;
}
}
}
else
{
{
int col = idx_col(x, last_col);
sum = co3 * convert_float4((((srcData + idx_row(src_y - 2, last_row) * srcStep / 4))[col]));
sum = sum + co2 * convert_float4((((srcData + idx_row(src_y - 1, last_row) * srcStep / 4))[col]));
sum = sum + co1 * convert_float4((((srcData + idx_row(src_y , last_row) * srcStep / 4))[col]));
sum = sum + co2 * convert_float4((((srcData + idx_row(src_y + 1, last_row) * srcStep / 4))[col]));
sum = sum + co3 * convert_float4((((srcData + idx_row(src_y + 2, last_row) * srcStep / 4))[col]));
sum = sum + co3 * convert_float4((((srcData + idx_row(src_y - 2, last_row) * srcStep / 4))[idx_col(right_x, last_col)]));
sum = sum + co2 * convert_float4((((srcData + idx_row(src_y - 1, last_row) * srcStep / 4))[idx_col(right_x, last_col)]));
sum = sum + co1 * convert_float4((((srcData + idx_row(src_y , last_row) * srcStep / 4))[idx_col(right_x, last_col)]));
sum = sum + co2 * convert_float4((((srcData + idx_row(src_y + 1, last_row) * srcStep / 4))[idx_col(right_x, last_col)]));
smem[2 + get_local_id(0)] = sum;
if (get_local_id(0) < 2)
{
const int left_x = x - 2;
col = idx_col(left_x, last_col);
sum = co3 * convert_float4((((srcData + idx_row(src_y - 2, last_row) * srcStep / 4))[col]));
sum = sum + co2 * convert_float4((((srcData + idx_row(src_y - 1, last_row) * srcStep / 4))[col]));
sum = sum + co1 * convert_float4((((srcData + idx_row(src_y , last_row) * srcStep / 4))[col]));
sum = sum + co2 * convert_float4((((srcData + idx_row(src_y + 1, last_row) * srcStep / 4))[col]));
sum = sum + co3 * convert_float4((((srcData + idx_row(src_y + 2, last_row) * srcStep / 4))[col]));
smem[get_local_id(0)] = sum;
}
if (get_local_id(0) > 253)
{
const int right_x = x + 2;
col = idx_col(right_x, last_col);
sum = co3 * convert_float4((((srcData + idx_row(src_y - 2, last_row) * srcStep / 4))[col]));
sum = sum + co2 * convert_float4((((srcData + idx_row(src_y - 1, last_row) * srcStep / 4))[col]));
sum = sum + co1 * convert_float4((((srcData + idx_row(src_y , last_row) * srcStep / 4))[col]));
sum = sum + co2 * convert_float4((((srcData + idx_row(src_y + 1, last_row) * srcStep / 4))[col]));
sum = sum + co3 * convert_float4((((srcData + idx_row(src_y + 2, last_row) * srcStep / 4))[col]));
smem[4 + get_local_id(0)] = sum;
}
}
barrier(CLK_LOCAL_MEM_FENCE);
@@ -259,9 +317,7 @@ __kernel void pyrDown_C4_D0(__global uchar4 * srcData, int srcStep, int srcOffse
{
const int tid2 = get_local_id(0) * 2;
sum = 0;
sum = co3 * smem[2 + tid2 - 2];
sum = sum + co2 * smem[2 + tid2 - 1];
sum = sum + co1 * smem[2 + tid2 ];
sum = sum + co2 * smem[2 + tid2 + 1];
@@ -274,9 +330,9 @@ __kernel void pyrDown_C4_D0(__global uchar4 * srcData, int srcStep, int srcOffse
}
}
__kernel void pyrDown_C1_D5(__global float * srcData, int srcStep, int srcRows, int srcCols, __global float *dst, int dstStep, int dstCols)
{
{
const int x = get_global_id(0);
const int y = get_group_id(1);
__local float smem[256 + 4];
@@ -287,44 +343,83 @@ __kernel void pyrDown_C1_D5(__global float * srcData, int srcStep, int srcOffset
const int last_row = srcRows - 1;
const int last_col = srcCols - 1;
sum = 0;
sum = sum + 0.0625f * ((__global float*)((__global char*)srcData + idx_row(src_y - 2, last_row) * srcStep))[idx_col(x, last_col)];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + idx_row(src_y - 1, last_row) * srcStep))[idx_col(x, last_col)];
sum = sum + 0.375f * ((__global float*)((__global char*)srcData + idx_row(src_y , last_row) * srcStep))[idx_col(x, last_col)];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + idx_row(src_y + 1, last_row) * srcStep))[idx_col(x, last_col)];
sum = sum + 0.0625f * ((__global float*)((__global char*)srcData + idx_row(src_y + 2, last_row) * srcStep))[idx_col(x, last_col)];
smem[2 + get_local_id(0)] = sum;
if (src_y >= 2 && src_y < srcRows - 2 && x >= 2 && x < srcCols - 2)
{
{
sum = 0.0625f * ((__global float*)((__global char*)srcData + (src_y - 2) * srcStep))[x];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + (src_y - 1) * srcStep))[x];
sum = sum + 0.375f * ((__global float*)((__global char*)srcData + (src_y ) * srcStep))[x];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + (src_y + 1) * srcStep))[x];
sum = sum + 0.0625f * ((__global float*)((__global char*)srcData + (src_y + 2) * srcStep))[x];
smem[2 + get_local_id(0)] = sum;
sum = sum + 0.0625f * ((__global float*)((__global char*)srcData + idx_row(src_y - 2, last_row) * srcStep))[idx_col(left_x, last_col)];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + idx_row(src_y - 1, last_row) * srcStep))[idx_col(left_x, last_col)];
sum = sum + 0.375f * ((__global float*)((__global char*)srcData + idx_row(src_y , last_row) * srcStep))[idx_col(left_x, last_col)];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + idx_row(src_y + 1, last_row) * srcStep))[idx_col(left_x, last_col)];
if (get_local_id(0) < 2)
{
const int left_x = x - 2;
sum = 0.0625f * ((__global float*)((__global char*)srcData + (src_y - 2) * srcStep))[left_x];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + (src_y - 1) * srcStep))[left_x];
sum = sum + 0.375f * ((__global float*)((__global char*)srcData + (src_y ) * srcStep))[left_x];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + (src_y + 1) * srcStep))[left_x];
sum = sum + 0.0625f * ((__global float*)((__global char*)srcData + (src_y + 2) * srcStep))[left_x];
smem[get_local_id(0)] = sum;
}
if (get_local_id(0) > 253)
{
const int right_x = x + 2;
sum = 0.0625f * ((__global float*)((__global char*)srcData + (src_y - 2) * srcStep))[right_x];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + (src_y - 1) * srcStep))[right_x];
sum = sum + 0.375f * ((__global float*)((__global char*)srcData + (src_y ) * srcStep))[right_x];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + (src_y + 1) * srcStep))[right_x];
sum = sum + 0.0625f * ((__global float*)((__global char*)srcData + (src_y + 2) * srcStep))[right_x];
smem[4 + get_local_id(0)] = sum;
}
}
}
else
{
{
int col = idx_col(x, last_col);
sum = 0.0625f * ((__global float*)((__global char*)srcData + idx_row(src_y - 2, last_row) * srcStep))[col];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + idx_row(src_y - 1, last_row) * srcStep))[col];
sum = sum + 0.375f * ((__global float*)((__global char*)srcData + idx_row(src_y , last_row) * srcStep))[col];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + idx_row(src_y + 1, last_row) * srcStep))[col];
sum = sum + 0.0625f * ((__global float*)((__global char*)srcData + idx_row(src_y + 2, last_row) * srcStep))[col];
sum = sum + 0.0625f * ((__global float*)((__global char*)srcData + idx_row(src_y - 2, last_row) * srcStep))[idx_col(right_x, last_col)];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + idx_row(src_y - 1, last_row) * srcStep))[idx_col(right_x, last_col)];
sum = sum + 0.375f * ((__global float*)((__global char*)srcData + idx_row(src_y , last_row) * srcStep))[idx_col(right_x, last_col)];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + idx_row(src_y + 1, last_row) * srcStep))[idx_col(right_x, last_col)];
smem[2 + get_local_id(0)] = sum;
if (get_local_id(0) < 2)
{
const int left_x = x - 2;
col = idx_col(left_x, last_col);
sum = 0.0625f * ((__global float*)((__global char*)srcData + idx_row(src_y - 2, last_row) * srcStep))[col];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + idx_row(src_y - 1, last_row) * srcStep))[col];
sum = sum + 0.375f * ((__global float*)((__global char*)srcData + idx_row(src_y , last_row) * srcStep))[col];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + idx_row(src_y + 1, last_row) * srcStep))[col];
sum = sum + 0.0625f * ((__global float*)((__global char*)srcData + idx_row(src_y + 2, last_row) * srcStep))[col];
smem[get_local_id(0)] = sum;
}
if (get_local_id(0) > 253)
{
const int right_x = x + 2;
col = idx_col(right_x, last_col);
sum = 0.0625f * ((__global float*)((__global char*)srcData + idx_row(src_y - 2, last_row) * srcStep))[col];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + idx_row(src_y - 1, last_row) * srcStep))[col];
sum = sum + 0.375f * ((__global float*)((__global char*)srcData + idx_row(src_y , last_row) * srcStep))[col];
sum = sum + 0.25f * ((__global float*)((__global char*)srcData + idx_row(src_y + 1, last_row) * srcStep))[col];
sum = sum + 0.0625f * ((__global float*)((__global char*)srcData + idx_row(src_y + 2, last_row) * srcStep))[col];
smem[4 + get_local_id(0)] = sum;
}
}
barrier(CLK_LOCAL_MEM_FENCE);
@@ -333,9 +428,7 @@ __kernel void pyrDown_C1_D5(__global float * srcData, int srcStep, int srcOffset
{
const int tid2 = get_local_id(0) * 2;
sum = 0;
sum = 0.0625f * smem[2 + tid2 - 2];
sum = sum + 0.25f * smem[2 + tid2 - 1];
sum = sum + 0.375f * smem[2 + tid2 ];
sum = sum + 0.25f * smem[2 + tid2 + 1];
@@ -348,9 +441,9 @@ __kernel void pyrDown_C1_D5(__global float * srcData, int srcStep, int srcOffset
}
}
__kernel void pyrDown_C4_D5(__global float4 * srcData, int srcStep, int srcRows, int srcCols, __global float4 *dst, int dstStep, int dstCols)
{
{
const int x = get_global_id(0);
const int y = get_group_id(1);
__local float4 smem[256 + 4];
@@ -361,48 +454,87 @@ __kernel void pyrDown_C4_D5(__global float4 * srcData, int srcStep, int srcOffse
const int last_row = srcRows - 1;
const int last_col = srcCols - 1;
float4 co1 = (float4)(0.375f, 0.375f, 0.375f, 0.375f);
float4 co2 = (float4)(0.25f, 0.25f, 0.25f, 0.25f);
float4 co1 = 0.375f;//(float4)(0.375f, 0.375f, 0.375f, 0.375f);
float4 co2 = 0.25f;//(float4)(0.25f, 0.25f, 0.25f, 0.25f);
float4 co3 = 0.0625f;//(float4)(0.0625f, 0.0625f, 0.0625f, 0.0625f);
if (src_y >= 2 && src_y < srcRows - 2 && x >= 2 && x < srcCols - 2)
{
sum = co3 * ((__global float4*)((__global char4*)srcData + (src_y - 2) * srcStep / 4))[x];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + (src_y - 1) * srcStep / 4))[x];
sum = sum + co1 * ((__global float4*)((__global char4*)srcData + (src_y ) * srcStep / 4))[x];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + (src_y + 1) * srcStep / 4))[x];
sum = sum + co3 * ((__global float4*)((__global char4*)srcData + (src_y + 2) * srcStep / 4))[x];
sum = sum + co3 * ((__global float4*)((__global char4*)srcData + idx_row(src_y - 2, last_row) * srcStep / 4))[idx_col(x, last_col)];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + idx_row(src_y - 1, last_row) * srcStep / 4))[idx_col(x, last_col)];
sum = sum + co1 * ((__global float4*)((__global char4*)srcData + idx_row(src_y , last_row) * srcStep / 4))[idx_col(x, last_col)];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + idx_row(src_y + 1, last_row) * srcStep / 4))[idx_col(x, last_col)];
smem[2 + get_local_id(0)] = sum;
if (get_local_id(0) < 2)
{
const int left_x = x - 2;
if (get_local_id(0) < 2)
{
sum = co3 * ((__global float4*)((__global char4*)srcData + (src_y - 2) * srcStep / 4))[left_x];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + (src_y - 1) * srcStep / 4))[left_x];
sum = sum + co1 * ((__global float4*)((__global char4*)srcData + (src_y ) * srcStep / 4))[left_x];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + (src_y + 1) * srcStep / 4))[left_x];
sum = sum + co3 * ((__global float4*)((__global char4*)srcData + (src_y + 2) * srcStep / 4))[left_x];
smem[get_local_id(0)] = sum;
}
sum = sum + co3 * ((__global float4*)((__global char4*)srcData + idx_row(src_y - 2, last_row) * srcStep / 4))[idx_col(left_x, last_col)];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + idx_row(src_y - 1, last_row) * srcStep / 4))[idx_col(left_x, last_col)];
sum = sum + co1 * ((__global float4*)((__global char4*)srcData + idx_row(src_y , last_row) * srcStep / 4))[idx_col(left_x, last_col)];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + idx_row(src_y + 1, last_row) * srcStep / 4))[idx_col(left_x, last_col)];
if (get_local_id(0) > 253)
{
const int right_x = x + 2;
smem[get_local_id(0)] = sum;
sum = co3 * ((__global float4*)((__global char4*)srcData + (src_y - 2) * srcStep / 4))[right_x];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + (src_y - 1) * srcStep / 4))[right_x];
sum = sum + co1 * ((__global float4*)((__global char4*)srcData + (src_y ) * srcStep / 4))[right_x];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + (src_y + 1) * srcStep / 4))[right_x];
sum = sum + co3 * ((__global float4*)((__global char4*)srcData + (src_y + 2) * srcStep / 4))[right_x];
if (get_local_id(0) > 253)
{
smem[4 + get_local_id(0)] = sum;
}
}
else
{
int col = idx_col(x, last_col);
sum = co3 * ((__global float4*)((__global char4*)srcData + idx_row(src_y - 2, last_row) * srcStep / 4))[col];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + idx_row(src_y - 1, last_row) * srcStep / 4))[col];
sum = sum + co1 * ((__global float4*)((__global char4*)srcData + idx_row(src_y , last_row) * srcStep / 4))[col];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + idx_row(src_y + 1, last_row) * srcStep / 4))[col];
sum = sum + co3 * ((__global float4*)((__global char4*)srcData + idx_row(src_y + 2, last_row) * srcStep / 4))[col];
sum = sum + co3 * ((__global float4*)((__global char4*)srcData + idx_row(src_y - 2, last_row) * srcStep / 4))[idx_col(right_x, last_col)];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + idx_row(src_y - 1, last_row) * srcStep / 4))[idx_col(right_x, last_col)];
sum = sum + co1 * ((__global float4*)((__global char4*)srcData + idx_row(src_y , last_row) * srcStep / 4))[idx_col(right_x, last_col)];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + idx_row(src_y + 1, last_row) * srcStep / 4))[idx_col(right_x, last_col)];
smem[2 + get_local_id(0)] = sum;
if (get_local_id(0) < 2)
{
const int left_x = x - 2;
col = idx_col(left_x, last_col);
sum = co3 * ((__global float4*)((__global char4*)srcData + idx_row(src_y - 2, last_row) * srcStep / 4))[col];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + idx_row(src_y - 1, last_row) * srcStep / 4))[col];
sum = sum + co1 * ((__global float4*)((__global char4*)srcData + idx_row(src_y , last_row) * srcStep / 4))[col];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + idx_row(src_y + 1, last_row) * srcStep / 4))[col];
sum = sum + co3 * ((__global float4*)((__global char4*)srcData + idx_row(src_y + 2, last_row) * srcStep / 4))[col];
smem[get_local_id(0)] = sum;
}
if (get_local_id(0) > 253)
{
const int right_x = x + 2;
col = idx_col(right_x, last_col);
sum = co3 * ((__global float4*)((__global char4*)srcData + idx_row(src_y - 2, last_row) * srcStep / 4))[col];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + idx_row(src_y - 1, last_row) * srcStep / 4))[col];
sum = sum + co1 * ((__global float4*)((__global char4*)srcData + idx_row(src_y , last_row) * srcStep / 4))[col];
sum = sum + co2 * ((__global float4*)((__global char4*)srcData + idx_row(src_y + 1, last_row) * srcStep / 4))[col];
sum = sum + co3 * ((__global float4*)((__global char4*)srcData + idx_row(src_y + 2, last_row) * srcStep / 4))[col];
smem[4 + get_local_id(0)] = sum;
}
}
barrier(CLK_LOCAL_MEM_FENCE);
@@ -411,9 +543,7 @@ __kernel void pyrDown_C4_D5(__global float4 * srcData, int srcStep, int srcOffse
{
const int tid2 = get_local_id(0) * 2;
sum = 0;
sum = co3 * smem[2 + tid2 - 2];
sum = sum + co2 * smem[2 + tid2 - 1];
sum = sum + co1 * smem[2 + tid2 ];
sum = sum + co2 * smem[2 + tid2 + 1];

19
modules/ocl/src/kernels/pyrlk.cl
View File

@@ -45,6 +45,25 @@
//#pragma OPENCL EXTENSION cl_amd_printf : enable
__kernel void arithm_muls_D5 (__global float *src1, int src1_step, int src1_offset,
__global float *dst, int dst_step, int dst_offset,
int rows, int cols, int dst_step1, float scalar)
{
int x = get_global_id(0);
int y = get_global_id(1);
if (x < cols && y < rows)
{
int src1_index = mad24(y, src1_step, (x << 2) + src1_offset);
int dst_index = mad24(y, dst_step, (x << 2) + dst_offset);
float data1 = *((__global float *)((__global char *)src1 + src1_index));
float tmp = data1 * scalar;
*((__global float *)((__global char *)dst + dst_index)) = tmp;
}
}
__kernel void calcSharrDeriv_vertical_C1_D0(__global const uchar* src, int srcStep, int rows, int cols, int cn, __global short* dx_buf, int dx_bufStep, __global short* dy_buf, int dy_bufStep)
{