我想用CUDA Kernel取代两个nested for loop去处理图片,图片大小为512x512。
pseudo code:
for (i=0;i<512,++i){
for (j=0;j<512,++j){
img[i][j]
}
}
我每个block分配64个线程,然后4096个blocks。
template <typename Dtype>
__global__ void filterForward()
{
uint i = (blockIdx.x * blockDim.x) + threadIdx.x;
uint j = (blockIdx.y * blockDim.y) + threadIdx.y;
printf("%d %d\n",i,j);
dim3 threadsPerBlock(8, 8);
dim3 numBlocks(512/threadsPerBlock.x, 511/threadsPerBlock.y);
// NOLINT_NEXT_LINE(whitespace/operators)
filterForward<Dtype><<<numBlocks, threadsPerBlock>>>()
理想状态是print出来512x512行,遍历每个pixel的序号。但是只打出来192768行,说明有些组合没有遍历到。请问会是什么原因?
谢谢!
PS:GPU是TITAN Xp。是再写CAFFE。