Using openBLAS

[RajalakshmiSR]

There is a statement on the announcement page “In the next release, MKL-ML, openblas, and jemallac build options will be removed”.
Does this mean OpenBLAS GEMM functions will not be used in ONNX runtime? Or is it just the options in the build script that will be removed and ONNX Runtime can use OpenBLAS for SGEMM/DGEMM/CGEMM/ZGEMM?

[snnn]

Does this mean OpenBLAS GEMM functions will not be used in ONNX runtime?

Right. By default we use our own GEMM implementation. If a user need something different, he/she is encouraged to try the other execution providers(like DNNL) or build his own one. The default CPU execution providers only uses mlas and eigen as the math libs.

[RajalakshmiSR]

I recently added POWER10(MMA) optimized GEMM kernels in OpenBLAS. That will be faster for POWER10 processors.

[RajalakshmiSR]

The default CPU execution providers only uses mlas and eigen as the math libs.

@snnn How does onnx runtime decide to use Eigen or MLAS path for matrix multiplication specifically GEMM? (assuming we have GEMM implementation in both eigen and MLAS)

[snnn]

It prefers MLAS when MLAS has a GEMM implementation on that platform.

[RajalakshmiSR]

@snnn Is there any switch like USE_EIGEN_FOR_BLAS where I can use Eigen for sgemm instead of MLAS for some testing?

[snnn]

We don't. But welcome to contribute.