Results: RNN VAD realtime factor improved from 140x to 195x (+55x)
Test device: Pixel 2 XL
Benchmark setup: max clock speed forced on all the cores by
setting "performance" as scaling governor
Bug: webrtc:10480
Change-Id: I3e92f643853ad1fe990db909c578ce78ee826c03
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/198842
Reviewed-by: Per Åhgren <peah@webrtc.org>
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32888}
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}
Plus a few minor code readability improvements.
Bug: webrtc:10480
Change-Id: I590d8e203b1d05959a8c15373841e37abe83237e
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/195334
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Reviewed-by: Karl Wiberg <kwiberg@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32764}
This CL brings a large improvement to the RNN VAD CPU performance
by finally using `VectorMath::DotProduct()` for pitch search.
The realtime factor improved from about 390x to 570x for SSE2
(+180x, 45% faster) and to 610x for AVX2 (+235x, 60% faster).
RNN VAD benchmark results:
```
+-----+-------+------+------+
| run | none* | SSE2 | AVX2 |
+-----+-------+------+------+
| 1 | 393x | 572x | 618x |
| 2 | 388x | 568x | 607x |
| 3 | 393x | 564x | 599x |
+-----+-------+------+------+
```
*: baseline, no SIMD used for pitch search, but SSE2 used for the RNN
Results obtained as follows:
1. Force SSE2 in `DISABLED_RnnVadPerformance` for the RNN part in
order to measure the baseline correctly:
```
RnnBasedVad rnn_vad({/*sse2=*/true, /*avx2=*/true, /*neon=*/false});
```
2. Run the test:
```
$ ./out/release/modules_unittests \
--gtest_filter=*RnnVadTest*DISABLED_RnnVadPerformance* \
--gtest_also_run_disabled_tests --logs
```
Bug: webrtc:10480
Change-Id: I89a2bd420265540026944b9c0f1fdd4bfda7f475
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/195001
Reviewed-by: Gustaf Ullberg <gustaf@webrtc.org>
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32755}
In preparation for adding AVX2 code, a safe scheme to support
different SIMD optimizations is added.
Safety features:
- AVX2 kill switch to stop using it even if supported by the
architecture
- struct indicating the available CPU features propagated from
AGC2 to each component; in this way
- better control over the unit tests
- no need to propagate individual kill switches but just
set to false features that are turned off
Note that (i) this CL does not change the performance of the RNN VAD
and (ii) no AVX2 optimization is added yet.
Bug: webrtc:10480
Change-Id: I0e61f3311ecd140f38369cf68b6e5954f3dc1f5a
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/193140
Reviewed-by: Per Åhgren <peah@webrtc.org>
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32739}
- 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}
