The Performance Issues Caused by TARGET=RISCV64_ZVL256B

[guoyuanplct]

When compiling OpenBLAS using TARGET=RISCV64_ZVL256B, there can be significant performance issues, as shown in the two tables below. The tables show the performance when using RISCV64_GENERIC and RISCV64_ZVL256B, as well as the corresponding speedup ratios. HBMV is a Level 2 function. I am trying to offload its computation to the kernel to make specific modifications for RISC-V. However, I don't know exactly how to do it. Do you have any suggestions?

[martin-frbg]

Strange that the optimized kernels would be so much slower than their generic C counterpart - internally HBMV appears to be implemented as a series of AXPY calls, I wonder if the ZVL256B AXPY is slow too ? The only other specialty I see is that this
function appears to be implemented in single-threaded form only, but that should affect both codes equally.

[guoyuanplct]

I also find it very strange because the performance of RISCV64_ZVL256B is really good in the AXPY part.

[martin-frbg]

Then either it is some weird corner case that is holding everything up, or the CDOT/ZDOT kernels are slower, but they'd have to be dramatically slower than the simple C implementation ??

[guoyuanplct]

I’ve tested the CDOT/ZDOT kernels and they actually perform very well. I’m still looking into it to figure out what’s causing the problem.

[martin-frbg]

Unfortunately I cannot get perf to produce any useful output on my system (Banana Pi F3, seems to be a hardware issue. On x86_64, the benchmark seems to spend a bit more time in DOT than in AXPY, but the only "hot" instruction recorded is packing the result back into a complex variable. So nothing that would provide a clue. I wish I had more experience with RISCV, I'm currently stuck on a bug with NAN handling in ZSCAL and can't even get it to build a working mask with vmfne.vv

[martin-frbg]

Guessing it may be too expensive to start up the vector pipeline for very low values of K (in particular, the provided hbmv benchmark runs with k=1 unless one presets the environment variable OPENBLAS_K, which directly corresponds to running AYPY with n=1). I have not yet checked in detail where the crossover occurs, but at least for OPENBLAS_K=50, the current "_vector" kernels for AXPY and DOT appear to outperform the naive generic ones in ?HBMV.

[martin-frbg]

closing as fixed by your PR #5291