Did anyone successfully test this feature on Cuda4 ? It seems I can’t have it worked correctly, not sure if it’s a bug in my code or the feature is not fully implemented yet.
For example, I want to do a kernel call after copying a memory from 1 gpu to another. Stream 0 is created on gpu0 stream 1 is created on gpu1. Both are Fermi
cudaMemcpyAsync( mem1, mem0, size, cudaMemcpyDefault, stream0 );
cudaStreamWaitEvent(stream1, P2Pevent, 0);
Sometime it seems the memory is not ready for the kernel yet so the result is incorrect. If I add an cudaDeviceSynchronize on Gpu0 then it works fine.