How can I write kernel with the Brook+, which makes the buffer to read and write?
How can I write kernel with the Brook+, which makes the buffer to read and write?
I have: image(704x576) and logical grid(22x18(edge=32)) on this image - one grid cell goups 32x32 pixel under it.
I need to proccess image, and calc (for example) number of red pixels (color.R > 200, for example) in every cell, BUT in this kernel i want mark all pixel to green (color.G = 255);
Show the code:
brook::Stream<int> img(...); // 704x576
brook::Stream<int> res(...); // 704x576
brook::Stream<int> mask(...); // edge=32 -> (704/edge)x(576/edge) -> 22x18
...
Kernel1(img, mask, res); // sart thread for every pixel
kernel void Kernel1(int frame[][], int mask[][], out int outputFrame<>
{
int xMask = vPos.x / edge;
int yMask = vPos.y / edge;
int2 maskPos = int2(xMask, yMask); // index of grid cell when current pixel is located
int color = frame[instance().xy]; // current pixel
int red = color & 0xFF; // get r-channel
if (red > 200)
mask[maskPos]++;
outputFrame = color | 0x00FF0000;
}
can i make it work?
You seems to be reading and writing from the same stream. That is not currently possible in Brook+. To make it work, you need to use a scatter and a regular output stream. That again is a limitation in Brook+. But, you can try dividing your kernel in two parts -
kernel void Kernel1(int frame<>, out int mask[][])
{
int2 vPos = instance().xy;
int xMask = vPos.x / edge;
int yMask = vPos.y / edge;
int2 maskPos = int2(xMask, yMask); // index of grid cell when current pixel is located
int color = frame; // current pixel
int red = color & 0xFF; // get r-channel
if (red > 200)
mask[maskPos]++;
}
and
kernel void Kernel2(int frame<>, out int outputFrame<>
{
outputFrame = frame | 0x00FF0000;
}
Also you need to set execution domain on Kernel1 from host side code-
// Set domain offset
Kernel1.domainOffset(uint4(0, 0, 0, 0));
Kernel1.domainSize(uint4(704, 576, 1, 1));
// call kernel
Kernel1(img, mask);
But, I think its very unlikely that this code will work as underlying hardware has some limitations in size and domain of execution for scatter use (Both has to be 64-aligned). If you can resolve these alignment constraints, it should work.
