GEMM: From Pure C to SSE Optimized Micro Kernels (2014)

Re:GEMM: From Pure C to SSE Optimized Micro Kernels (2014) by gnufx on Friday March 16, 2018 @08:58AM (#16600305)

I don't know which operations are implemented for those types, but the "dnn" support in the libxsmm ("small matrix multiplication") library has I32, I16, and I8 <https://github.com/hfp/libxsmm>.

[ Parent | on HN ]

Re:GEMM: From Pure C to SSE Optimized Micro Kernels (2014) by mratsim on Friday March 16, 2018 @07:55PM (#16604838)

I did stumble upon libxsmm while looking for a BLAS library, I didn't see they had integers.

Do you know what libxsmm means by "small", there is nothing in their README/wiki. However it does say that libxsmm doesn't support 32-bit OS. My library is used in IoT devices.

[ Parent | on HN ]

Re:GEMM: From Pure C to SSE Optimized Micro Kernels (2014) by hfp on Monday March 19, 2018 @05:50AM (#16617193)

* Small Matrix Multiplication: https://github.com/hfp/libxsmm/wiki/Q&A#what-is-a-small-matr...

* Big(ger) Matrix Multiplication: https://github.com/hfp/libxsmm/wiki/Q&A#what-about-medium-si...

There is indeed no support for 32-bit. Regarding integers, LIBXSMM supports lower-precision GEMM. However, we haven't fully represented this in our usual sample code (samples/xgemm, samples/bgemm).

Put aside a tiling scheme for big matrices - the code generation for low-precision GEMM kernels is here:

* https://github.com/hfp/libxsmm/blob/master/src/template/libx...

* Generic: https://github.com/hfp/libxsmm/blob/master/src/template/libx... with https://github.com/hfp/libxsmm/blob/master/include/libxsmm_g...

* C++: https://github.com/hfp/libxsmm/blob/master/src/template/libx...

[ Parent | on HN ]