Vadim Pirogov
Vadim Pirogov
Upon further investigation the rootcase is in block size. Reassigning to @aaraujom for analysis and fix.
Hi @affranchi, Thank you for the report. Matmul should be always relying on GEMM implementation. The fact that you see `ref:any` with 50x50 bias might actually be a bug.
@nSircombe, does this issue looks like something your team will be able to help with?
I would expect the issue to be reproducible with TBB threading in regular builds for Arm as well. The conda forge packages were created with TBB threading as this is...
Thanks for investigation, @nSircombe! Keeping this defect open in case someone wants to jump in.
Thanks for the proposal, @maayaneh! Were you able to evaluate performance loss with the proposed sequence? Implementation PoC from the comments: https://github.com/oneapi-src/oneDNN/compare/master...maayaneh:oneDNN:master
We analyzed top contributors to oneDNN binary size and there are several factors that have the biggest impact: 1. GEMM autogenerated kernels 2. Templated simple reorder implementations 3. Growth in...
We promoted changes addressing item (2) from the list, these are targeting oneDNN v2.5 release.
[Build options](https://github.com/oneapi-src/oneDNN/blob/rfcs/rfcs/20210510-binary-size-reduction/ISA_extension.md) that allow limit ISA-specific optimizations in the library landed into the main branch. The feature is targeting oneDNN v2.5.
The implementation that is not prone to overflow for Intel AVX512/Intel AVX2 instruction sets is expected to be 2x slower than the current one and requires non-trivial investment. This is...