search

drowe67 / LPCNet

Experimental Neural Net speech coding for FreeDV
BSD 3-Clause "New" or "Revised" License
68 stars 25 forks source link

SSE support (for testing) #24

Closed jg1uaa closed 4 years ago

jg1uaa commented 4 years ago

For performance reason, LPCNet (lpcnet_dec) requires AVX FMA instruction. But cheap Intel processor (Pentium/Celeron) does not support AVX still now.

I tried SSE-vectorized LPCNet on OpenBSD-6.6/amd64 (clang 8.0.1) with AMD's A10-7860K processor PC, its performance barely achieved with enabling -msse4.1 clang option. Old SSE (SSE3 or former) SSE shown lower performance.

Testing code is at my repository, https://github.com/jg1uaa/LPCNet/ . Can you evaluate SSE-vectorize is usable or not?

drowe67 commented 4 years ago

Hi @jg1uaa - nice work :+1:

Sure, I will check it out.

Can you please raise a pull request for your fork against https://github.com/drowe67/LPCNet? That way I can easily see what you've changed. We can maybe write some simple bench marks to compare with AVX.

hobbes1069 commented 4 years ago

I believe the pull request was accepted. Closing