CUDA: Using shared memory across different kernels

Question

Is this possible?
I am trying to define a shared memory array in one kernel and then I need to use those values in a different kernel.

I tried declaring the

extern __shared__ float sharedMem[];

outside all functions and then wrote to it in one kernel and tried to access it in a different kernel. The sharedMem array is written to properly in the first kernel, but when I try to access it in the second kernel, the values are all 0. So I am guessing this won't work or I am doing something wrong.

Can someone please help me out on this?

Answer 1

You are correct, shared memory does not persist across kernel calls. Instead you must use global memory (or texture memory) and load it into shared memory in each kernel call.