Hello Dear All, Could you please tell me when to use cudaDeviceSynchronize in our code? I read some of resource that when we use cudaDeviceSynchronize, it cause our program to slow down. Previously I use cudaDeviceSynchronyce after calling kernel to transfer data from device to host In other hand I use cudaMemcpyDeviceToHost to copy back data from device to host but I still find the different of performance both of them.So which one is better? Thank you very much
Are you using multiple streams in your cuda code? If not, you do not need cudaDeviceSynchronize.
cuda stream allows you to launch multiple tasks on a GPU asynchronously. All tasks within one stream are blocking in nature. But tasks from different streams are non-blocking. One example could be, when you are running a kernel, you may want to transfer data from cpu to gpu asynchronously so that you can save the time to fetch the data for next kernel. But in your application if you want to force the synchronization after execting certain tasks, you can use cudaDeviceSynchronize. But remember it is a very expensive heavy-weight barrier, you would try to use lighter barriers and see the documentation to understand how they work and how you may possibly need those. Trying to avoid synchronizations, atomics etc. is a part of developing optimized code for GPU.
- Badges
- Science topic
- Similar topics
- Phytochemistry
- Extracts