1

My task is to port caches' latency measurement program written in CUDA to HIP, and for that i need to get characteristics of GPU.

I'm using HIP 4.5.0, and the problem is that for some data from CUDA's deviceInfo, there're no (direct) analogues from HIP.

They are:

  1. (Max) Shared memory per multiprocessor

In CUDA, sharedMemPerMultiprocessor exists. When talking about HIP, i found maxSharedMemoryPerMultiprocessor. Are they the same? Or first one is always less or equal to the second one and thus can't be replaced?

  1. Registers per multiprocessor

In CUDA - regsPerMultiprocessor. In HIP - non-existent (according to this page). Is there's a workaround for capturing number of registers per multiprocessor in this case?

Thank you for your answers!

Robert Crovella
  • 143,785
  • 11
  • 213
  • 257
Max Azatian
  • 53
  • 7
  • Were you able to figure this out? – YouJin Aug 16 '23 at 04:20
  • @YouJin hi! sort of: 1. For `sharedMem..` the values on Nvidia side are the same, for AMD - had to take them for granted. 2. For `regsPerMultiprocessor` - switched to `regsPerSM` value. Under all tested arch's for my BA thesis, the former was equal to latter + found a formula for verification of my idea, but no sufficient proofs to confirm/reject it :( – Max Azatian Aug 24 '23 at 17:23

0 Answers0