Understand the document on block level APIs(https://github.com/oneapi-src/oneDNN/pull/1852)

[akote123]

@mgouicem, I am going through the document in
https://github.com/mgouicem/oneDNN/tree/mgouicem/rfcs/brgemm/rfcs/20240326-block-level-api
to understand more. There is note mentioned that "or LLM optimizations, and "pure" matrix multiplication, batch-reduction is typically not necessary unless A and B are large and require K-blocking. This can be observed in current IPEX implementations using libxsmm for FlashAttention, Multi-Head Attention and Weights-only-Quantization Matmul.",
I wanted to understand :
1.In default pytorch uses mkl for matmul computation, So does IPEX uses any heuristic to switch between mkl and libxsmm or it uses only libxsmm for matmuls?
2. As in Icelake onednn is not used for matmuls in pytorch, does brgemm is refers to implementation inside mkl?

[mgouicem]

Thanks for the interest @akote123

1. Adding @Xia-Weiwen for ipex part. I believe that libxsmm ukernels are used only for some complex fused patterns. There is ongoing work with ipex team to migrate to the new oneDNN brgemm APIs.
2. Not sure I understand the question. @jgong can better address the details about pytorch internals.