How CudaMalloc work?

Question

I am trying to modify the imageDenosing class in CUDA SDK, I need to repeat the filter many time incase to capture the time. But my code doesn\'t work properly.

//st

Answer 1

I already answered this for you when you posted the same question previously - you need to wait for a kernel to complete before running it again - add:

cudaThreadSynchronize(); // *** wait for kernel to complete ***

after the kernel call.

Answer 2

Your kernel is running asynchronously - you need to wait for it to complete, e.g.

cudaMalloc((void **)&dst2, size);
cudaMemcpy(dst2, dst, imageW * imageH * sizeof(TColor), cudaMemcpyHostToDevice);
F1D<<<grid2, threads2>>>(dst, imageW, imageH, dst2);
cudaThreadSynchronize(); // *** wait for kernel to complete ***
cudaFree(dst2);

Answer 3

The statement

image[imageW * iy + ix] =   buffer[imageW * iy + ix];

is causing the problem. You are overwriting your input image in the kernel. So depending on thread execution order, you would be further blurring parts of the image.

Also, I don't see the purpose of

cudaMemcpy(dst2, dst, imageW*imageH*sizeof(TColor),cudaMemcpyHostToDevice);

dst looks to be device memory since you have access to it in the cuda kernal.