intel / ideep

Intel® Optimization for Chainer*, a Chainer module providing numpy like API and DNN acceleration using MKL-DNN.
MIT License
165 stars 90 forks source link

Remove matmul primitive caching in ideep #339

Open annop-w opened 1 month ago

annop-w commented 1 month ago

This removes aarch64 specific code branches for caching oneDNN matmul primitive in ideep. Instead, we rely on primitive caching in oneDNN. Since oneDNN v3.6 matmul primitive with ACL backend leverages the experimental stateless API in ACL and thus can be cached more efficiently.

The work to make ACL lowp matmul primitive is still in progress. Please hold off merging this PR until that work is done.