The implementations for the fully connected layer can be simlpified by
using `VectorMath:DotProduct()`. In this way, it is also possible to
remove (nearly) duplicated SIMD code, reduce the binary size and more
easily maintain the code.
This CL also forces unoptimized code for the output layer of the VAD,
which is a FC 24x1 layer. A slight improvement of the realtime has
been measured (delta ~ +5x).
Bug: webrtc:10480
Change-Id: Iee93bd59f7905ebf96275dbbfeb3c921baf4e8db
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/195580
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Reviewed-by: Ivo Creusen <ivoc@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32806}
Using `VectorMath::DotProduct()` in GatedRecurrentLayer to reuse existing
SIMD optimizations. Results:
- When SSE2/AVX2 is avilable, the GRU layer takes 40% of the unoptimized
code
- The realtime factor for the VAD improved as follows
- SSE2: from 570x to 630x
- AVX2: from 610x to 680x
This CL also improved the GRU layer benchmark by (i) benchmarking a GRU
layer havibng the same size of that used in the VAD and (ii) by prefetching
a long input sequence.
Bug: webrtc:10480
Change-Id: I9716b15661e4c6b81592b4cf7c172d90e41b5223
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/195545
Reviewed-by: Per Åhgren <peah@webrtc.org>
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32803}
Refactoring done to more easily and cleanly add SIMD optimizations and
to remove `GatedRecurrentLayer` from the RNN VAD api.
Bug: webrtc:10480
Change-Id: Ie1dffdd9b19c57c03a0b634f6818c0780456a66c
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/195445
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Reviewed-by: Jakob Ivarsson <jakobi@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32770}
