|Switch FFT to OpenMAX/pffft/Ne10|
|Project Member Reported by firstname.lastname@example.org, Apr 21 2014||Back to list|
This may consist of finalizing, testing and landing Kevin's old CL here: http://review.webrtc.org/1819004/ Alternately we may prefer to use the OpenMAX APIs directly. Turaj is bringing in OpenMAX now for MPLC, which has x86 support.
Apr 21 2014,
This will immediately impact: AECM (though we should probably ignore this, as it's going away) NS fixed-point Pieces we could move to OpenMAX in the future: AEC (have to profile against the current SSE optimizations, and soon-to-be-started NEON optimizations.) NS floating-point Transient suppression
May 14 2014,
Actually, we should strongly consider using pffft instead. rtoy's benchmarks show it is faster than openmax on x86. Work is ongoing to improve the NEON implementation as well.
May 31 2014,
Jun 2 2014,
Jun 23 2014,
We should consider Ne10 as well: https://github.com/projectNe10/Ne10 Yang Zhang from ARM has recently worked on improving performance: I have updated the float/int32/int16 FFT of Ne10. Currently they are faster than OpenMax DL’s on V7. I also added the intrinsic implementation for V8 Aarch 64 mode. And compared with C version, V8 intrinsic has better performance boost than V7 assembly. Next, I’m planning to investigate the non-power-of-2 FFT.
Oct 30 2014,
Jun 9 2015,
Nov 26 2015,
Oct 5 2016,
Oct 5 2016,
|► Sign in to add a comment|