Skip to content

Preliminary HIP bfloat16 GPU support.

The purpose of this MR is to deliver bfloat16 data type support for AMD GPUs and the HIP software stack.

The changes captured herein are a work in progress in the sense that basic functionality is provided.

Performance optimizations and similar changes will be forthcoming.

The existing bfloat16 header has been enhanced with support for HIP.

Also, a GPU specific bfloat16 unit test suite has been introduced.

Edited by Rohit Santhanam

Merge request reports

Loading