CUDA-每次都必须分配和释放内存吗？

Question

I have a convolution kernel with CUDA which is called very often (it is used for a real time rendering). Should I cudaMalloc and cudaFree each time I want to call the kernel? I tried to store a pointer to the cudaMalloc result and proceed by just cudaMemcpy'ing things before the kernel execution but I experienced weird behavior (like empty memory after the kernel execution)

I was also thinking about using pinned memory but if I have to allocate and free it every time it could even slow the application down. How should I proceed for a kernel which gets called very often?

Answer 1

No, there is no reason to malloc/free for each kernel call. Malloc'ed memory remains valid until you free it. We have lots of code that performs many kernels on allocated memory with and without cudaMemcpy to change the contents in between.

Your problem must be elsewhere. Try to boil it down to the smallest possible example that shows the problem and post the code.

Answer 2

It sounds like what you're doing should work.

Maybe you have a bug in your kernel. Try adding cudaThreadSynchronize and cudaGetLastError calls after the kernel launches to debug.

Without more information, I can't offer you any more advice than that.