[Openmp-commits] [PATCH] D91746: [libomptarget] Fail on missing symbols in device image

[Jon Chesterfield via Phabricator via Openmp-commits]

JonChesterfield added a comment.

Cuda reads from _exec_mode, and writes to omptarget_device_environment, after the image has been loaded onto the gpu. Amdgpu read/writes the device image before it is loaded onto the gpu to avoid the extra round trip. This means amdgpu has an elf symbol table parsing function and makes a possibly avoidable copy of the device image (I haven't checked whether libomptarget handles a plugin mutating the image gracefully). Do we want the same optimisation for cuda?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D91746/new/

https://reviews.llvm.org/D91746