codeplaysoftware
Extended Gemm interface to support mixed precision operations
Extend Gemm operator interface to support mixed precision operations, namely by decoupling matrix A and B type element_in_t from output matrix C and scalars alpha and beta type element_out_t.
Following oneMKL's spec notation for Gemm API : https://spec.oneapi.io/versions/latest/elements/oneMKL/source/domains/blas/gemm.html#onemkl-blas-gemm, this PR enables (Ta==Tb) to be set independently from (Tc==Ts). This feature has been enabled at a first stage for Ta=Tb=sycl::half and Tc=Ts=float. Thus enabling half support also enables the mixed precision case of (half, float) for gemm.
Changes include:
- Updating different Gemm kernel implementations to account for the decoupled types.
- Necessary CMake and Kernel generation scripts updates to account for the couple of types instead of single type in gemm case.
- Necessary changes to unit-tests to account for this feature in the Gemm case.
Note :
Following oneMKL expected Gemm API, Support of bfloat16-float and std::int8_t-float would be straightforward afterwards, but the additional cases of Ta==Tb==Tc while Ts (alpha & beta) is separate will require additional decoupling & re-design work..
