kineto icon indicating copy to clipboard operation
kineto copied to clipboard
verified publisher icon pytorch

Add deviceProperties for AMD devices to trace's metadata

Open aaronenyeshi opened this issue 1 year ago • 1 comments

Summary: Similar to CUDA, save the device properties to the metadata of Roctracer GPU traces.

  • Renamed CudaDeviceProperties files to DeviceProperties, and changed TARGETS' cpp_library to device_properties
  • Added gpu agnostic definitions so we can re-use existing DeviceProperties functions to return a valid devicePropertiesJson() string for AMD. Return "" for any non-CUDA or non-AMD.
  • Used ifdef for a few device properties specific to CUPTI: regsPerMultiprocessor, sharedMemPerBlockOptin, sharedMemPerMultiprocessor.
  • Clean up the TARGETS, only need 1 library and include it.

Differential Revision: D56901553

Pulled By: aaronenyeshi

aaronenyeshi avatar May 08 '24 14:05 aaronenyeshi

This pull request was exported from Phabricator. Differential Revision: D56901553

facebook-github-bot avatar May 08 '24 14:05 facebook-github-bot

This pull request was exported from Phabricator. Differential Revision: D56901553

facebook-github-bot avatar May 14 '24 15:05 facebook-github-bot

@aaronenyeshi merged this pull request in pytorch/kineto@affe44cebfcf96b0913d127975723f3405f8763c.

facebook-github-bot avatar May 14 '24 19:05 facebook-github-bot