I'm learning tensorflow 1.15 source code, I tried to find when will the outputs be copied from GPU memory back to CPU.Since ThenMemcpy() calls CUDA's cuMemcpyD2H function,I suppose it would be when ThenMemcpy() at the Stream class is called. but when I grep it in the source code, I've got no clue about it, especially during the OpKernel's Compute process and nearby functions. So what am I missing here?
0
votes
