← Back to events
ActiveAIRelease

trunk/377ffb934bd31a1538eb67a5ee1bc5cba10a4660: AArch64: FP16 matmul (#198498)

Photo: PyTorch Releases

What happened

Enable dispatch to optimized oneDNN kernel for non-contigous shapes for FP16 MM on AArch64 with this: oneDNN picks up gemm:acl impl as compared to ref impl Pull Request resolved: #198498 Approved by: https://github.com/Skylion007

Summary assembled by rule from the sources below

Why it's spreading

Sources