arm64 sve and risv64 v extension questions

[jli05]

• kernel/arm64/gemv_n_sve.c and kernel/arm64/gemv_n_sve_v1x3.c would divide the columns into three column blocks. Why particularly so on arm64 (in comparison with riscv64 v entension)? Why three column blocks?

• kernel/riscv64/gemv_n_vector.c would do more store and load of the y vector than kernel/riscv64/gemv_n_rvv.c. In which conditions would the former kernel be preferred? Why not just switch to the latter?

@martin-frbg @brada4

[brada4]

Go here for history deep-dive
https://en.wikipedia.org/wiki/GotoBLAS
Unless you prove it wrong (with numbers) workload is split optimally the way you observe.

[jli05]

What about the 2nd question re code using RISCV64 V extension?

[brada4]

What about you come up with numbers backing your whatabout?

[martin-frbg]

Kernels are specific to architectures and individual cpu characteristics such as cache sizes, vector register lengths - you can cross-reference the respective KERNEL files to find out what gets used where. Also they tend to be contributed by different people, and there is no requirement that a given kernel has to be the optimum imaginable (how would one prove this anyway), just that it is significantly faster than what was already implemented (while also returning correct results...)