Element ordering of inner GEMM kernels

[jerryz123]

I'm working on optimizing the inner GEMM kernels for RISC-V. I'm confused about the way the arrays are arranged once S/DGEMMKERNEL is called. The array ba[] and bb[] arguments seem to be arranged such that ba[] is row major, and bb[] is column major, turning the matrix-multiply into a series of dot products.

Furthermore, elements of ba[] and bb[] are rearranged, such that a 2x2 (or whatever size) block of elements is arranged contiguously. I guess this is to improve locality?

Is there anyway to write the inner kernel such that it receives A[] and B[] in the same arrangement, (both row major or both column major), without element reordering? The RISC-V vector implementation makes it very simple to perform GEMM if the operands are arranged in this manner, since loads and stores of long arrays stored contiguously are optimized for.

[brada4]

You can redefine anything you find in /common_macro.h

[martin-frbg]

I suspect you would need to change driver/level3/level3.c for that, basically add an "#if defined(RISCV)" branch that skips all the rearranging of the arguments. Unfortunately the code is not really documented, the best we have is https://github.com/xianyi/OpenBLAS/wiki/Developer-manual (which just gives a brief overview of code organization and a link to K.Goto's original paper).

[jerryz123]

I see, thanks. I'm starting to realize that much of the code is very optimized for packed-SIMD optimizations. Unfortunately this makes optimizing the package for a vector architecture somewhat cumbersome.