Special case overlapping execution and transfer times with streams

Accelerated Computing CUDA CUDA Programming and Performance
#1

Hello all,

I had a question regarding using CUDA streams. Particularly for my application, I was wondering if there was a way to reduce the amount of things I am actually transferring between CPU and GPU. In my current implementation, the flow of my code is as follows:

  1. I transfer an array of “weights” to the GPU
  2. I transfer my inputs to the GPU
  3. I launch the kernel, which calculates a dot product with the weights and the inputs, then updates the weights
  4. I transfer the updated weights and the outputs BACK to the CPU.
  5. I make this sequence of events for 1000’s of inputs

What I am particularly wondering, is there a way I could just STREAM my inputs and outputs, without copying back the “weights” every time? I really only need to copy the weights BACK from the GPU on the last iteration of my kernel call.

Does anyone have any better ideas how to organize/do this? Thanks!

#2

It sounds like you have the answer in the question. Why do you need to perform step 4 at all except at the end?

#3

I don’t think you need streams to accomplish your goal. Global memory retains its values between kernel calls, so the updated weights can stay on the GPU.

#4

Yup, you guys are right. Thanks for the help!