Skip to content

Comments

Add HIP backend converted from Cuda#2365

Draft
Menkib64 wants to merge 1 commit intoLeelaChessZero:masterfrom
Menkib64:hip_backend
Draft

Add HIP backend converted from Cuda#2365
Menkib64 wants to merge 1 commit intoLeelaChessZero:masterfrom
Menkib64:hip_backend

Conversation

@Menkib64
Copy link
Contributor

Code is conversion from cuda backend using hipify tool, regular expression replace, and manual fixes.

Known problems:

  • CNN networks produce NaNs in fp16 mode.
  • hipStreamWaitEvent for external event does a nullptr deference. Graph optimization disabled as workaround.
  • Kernels produce many warnings.
  • Kernels are missing AMD specific optimisations.

Code is conversion from cuda backend using hipify tool, regular
expression replace, and manual fixes.

Known problems:
* CNN networks produce NaNs in fp16 mode.
* hipStreamWaitEvent for external event does a nullptr deference. Graph
  optimization disabled as workaround.
* Kernels produce many warnings.
* Kernels are missing AMD specific optimisations.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant