I would like to assess the performance of DeepSpeech on CPU and before I do it I was thinking about building the binary which will use AVX2 and FMA, but because the process may get difficult I wanted to ask you first if you know that it is worth it. It may also help other users.
If I see it correctly current DeepSpeech binary is built supporting AVX instructions. Has anyone tried building against AVX2 and FMA yet and what are the possible performance improvement with respect to pre-built binaries?
For now I have only found the benchmarks by Lissyx: https://github.com/mozilla/DeepSpeech/issues/1023 which would indicate, that the difference between AVX build and AVX2 + FMA are marginal. Can someone confirm that this is actually the case?