This CL adds the GRU weights memory layout optimization with which it
will be easier to add SSE2 code in a follow up CL. The new memory
layout also improves the performance of the unoptimized code.
This CL also includes a bug fix in the GRU layer input validation.
It was a silent bug since the GRU layer of the RNN VAD has the same
input and output size. This was caught by changing memory layout of
the recurrent weights. The unit test has been adapted by removing the
unused recurrent weights (the expected result does not change).
Bug: webrtc:10480
Change-Id: Ia1551abde4cb24aa7e109c4447e0fffe7c839077
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/142177
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Reviewed-by: Gustaf Ullberg <gustaf@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#29717}
This CL adds the SSE2 optimized implementation for fully connected
(FC) layers. The change includes a weights re-alignment op done once
at construction time. It is required in order to optimize the load op
to fill 128 bit registers.
This CL also includes unit test adaptations and a benchmark test
(disabled by default).
Bug: webrtc:10480
Change-Id: I5ed87f0a629faaaf4c8bffbce1cea5557518f8c8
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/141862
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Reviewed-by: Gustaf Ullberg <gustaf@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#29712}
This reverts commit 7350a902374c796dec8ce583cfaf4b9697f3a525.
Reason for revert: possibly breaking downstream projects
Original change's description:
> RNN VAD: prepare for SIMD optimization
>
> This CL adds the boilerplate for SIMD optimization of FC and GRU layers
> in rnn.cc. The same scheme of AEC3 has been used. Unit tests for the
> optimized architectures have been added (the same unoptimized
> implementation will run).
>
> Minor changes:
> - unnecessary const removed in rnn.h
> - FC and GRU test data in the anon namespace as constexpr
>
> Bug: webrtc:10480
> Change-Id: Ifae4e970326e7e7c603d49aeaf61194b5efdabd3
> Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/141419
> Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
> Reviewed-by: Gustaf Ullberg <gustaf@webrtc.org>
> Cr-Commit-Position: refs/heads/master@{#29696}
TBR=gustaf@webrtc.org,alessiob@webrtc.org,fhernqvist@webrtc.org
Change-Id: I9ae82f4bd2d30797646fabfb5ad16bea378208b8
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: webrtc:10480
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/158893
Reviewed-by: Alessio Bazzica <alessiob@webrtc.org>
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#29699}
This CL adds the boilerplate for SIMD optimization of FC and GRU layers
in rnn.cc. The same scheme of AEC3 has been used. Unit tests for the
optimized architectures have been added (the same unoptimized
implementation will run).
Minor changes:
- unnecessary const removed in rnn.h
- FC and GRU test data in the anon namespace as constexpr
Bug: webrtc:10480
Change-Id: Ifae4e970326e7e7c603d49aeaf61194b5efdabd3
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/141419
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Reviewed-by: Gustaf Ullberg <gustaf@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#29696}
This CL has two goals: (i) avoid casting and scaling of the NN weights
for every processed feature vector and (ii) prepare for SIMD
optimizations.
Bug: webrtc:10480
Change-Id: Ice7bac5657123354714cc7c63b00abbb8a76c7d7
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/141413
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Reviewed-by: Fredrik Hernqvist <fhernqvist@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#29675}
- add test that checks that the computed VAD probability is within
tolerance *1
- speed-up some tests by reducing the input length and skipping frames
- remove unused code in test_utils
- fix some comments
*1: RnnVadTest::RnnBitExactness is replaced by
RnnVadTest::RnnVadProbabilityWithinTolerance
Bug: webrtc:10480
Change-Id: I19332d06eacffbbe671bf7749ff4c92798bdc55c
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/133910
Reviewed-by: Alex Loiko <aleloi@webrtc.org>
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#27803}
RNN implementation for the AGC2 VAD that includes a fully connected
layer and a gated recurrent unit layer.
Bug: webrtc:9076
Change-Id: Ibb8b0b4e9213f09eb9dbe118bbdc94d7e8e4f91b
Reviewed-on: https://webrtc-review.googlesource.com/72060
Reviewed-by: Patrik Höglund <phoglund@webrtc.org>
Reviewed-by: Alex Loiko <aleloi@webrtc.org>
Reviewed-by: Ivo Creusen <ivoc@webrtc.org>
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#23101}
