Hmm, a many days try to understand what is going on with my kernel code.
Probably you can help me. Are these code chuncks similar? I mean logic.
This chunk I execute on cpu
int xleft = 0, xright = 16;
int ytop = 0, ybottom = 16;
int temp = 0;
int mvlength = 100000000;
for (int j = 0; j < height; j += 16)
{
for (int i = 0; i < width; i += 16)
{
// set top and bottom range
ytop = - min(j, 16);
ybottom = min(height - 16 - j + 1, 16);
// set left and right range
xleft = - min(i, 16);
xright = min(width - 16 - i + 1, 16);
refsad = 100000000;
for (int y = ytop; y < ybottom; y++)
{
for (int x = xleft; x < xright; x++)
{
int srcidx = i + (j * width);
int index = i + x + ((j + y) * width);
// calculate SAD
//--------------------------------
for (m = 0; m < 16; m++)
{
for (n = 0; n < 16; n++)
{
temp += abs((src[srcidx + n] - ref[index + n]));
}
srcidx += width;
index += width;
}
//-------------------------------
if ((refsad >= temp) && (mvlength > abs(x) + abs(y)))
{
refsad = temp;
mvlength = abs(x) + abs(y);
refmvx = x;
refmvy = y;
refmvl = x;
}
temp = 0.0;
}
}
l++;
mvlength = 100000000;
}
}
And this as kernel
int ytop = - min(jy, 16);
int ybottom = min(height - 16 - jy + 1, 16);
// set left and right range
int xleft = - min(ix, 16);
int xright = min(width - 16 - ix + 1, 16);
int x, y;
int m, n;
int mvlength = 100000000;
sad = 100000000;
for (y = ytop; y < ybottom; y++)
{
for (x = xleft; x < xright; x++)
{
int testsad = 0;
int srcidx = ix + (jy * width);
int idx = ix + x + ((jy + y) * width);
for (m = 0; m < 16; m++)
{
for (n = 0; n < 16; n++)
{
testsad += (abs((((int)src[srcidx + n]) - ((int)ref[idx + n]))));
}
srcidx += width;
idx += width;
}
if ((sad >= testsad) && (mvlength > (abs(y) + abs(x))))
{
sad = testsad;
mvlength = (abs(y) + abs(x));
mvy = y;
mvx = x;
mvl = mvlength;
}
}
}
}
What do you think is the same logic? I have different results in mvx and mvy. Probably you see mistakes in kernel code. Because I expect absolutely the same behaviour.
I think problem exists in latest if.
if ((sad >= testsad) && (mvlength > (abs(y) + abs(x))))
If you need additional code let me know.
