oneDNN GEMM - AshokBhat/ml GitHub Wiki
mkldnn_sgemm - GEMM routine in oneDNN
Difference with cblas_sgemm
:
row-major
matrix layout and cannot be changed.
- Matrix transpositions is controlled using
Fortran
-style char arguments: 'N' for non-transposed case and 'T' for transposed case.
mkldnn_sgemm
returns error as mkldnn_status_t, while cblas_sgemm prints errors to the screen.
Code-gen
- JIT code based on ISA support
- Once through the lifetime of the program
Cache support
- Primitive cache support in 1.1 and later
- Solves JIT overhead issue for all other primitives
Sources
See also