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}
This reverts commit 9da3e177fd5c2236cc15fea0ee8933e1dd0d8f6d.
Reason for revert: bug in ComputePitchPeriod48kHz()
Original change's description:
> RNN VAD: pitch search optimizations (part 1)
>
> TL;DR this CL improves efficiency and includes several code
> readability improvements mainly triggered by the comments to
> patch set #10.
>
> Highlights:
> - Split `FindBestPitchPeriods()` into 12 and 24 kHz versions
> to hard-code the input size and simplify the 24 kHz version
> - Loop in `ComputePitchPeriod48kHz()` (new name for
> `RefinePitchPeriod48kHz()`) removed since the lags for which
> we need to compute the auto correlation are a few
> - `ComputePitchGainThreshold()` was only used in unit tests; it's been
> moved into the anon ns and the test removed
>
> This CL makes `ComputePitchPeriod48kHz()` is about 10% faster (measured
> with https://webrtc-review.googlesource.com/c/src/+/191320/4/modules/audio_processing/agc2/rnn_vad/pitch_search_internal_unittest.cc).
> The realtime factor has improved by about +14%.
>
> Benchmarked as follows:
> ```
> out/release/modules_unittests \
> --gtest_filter=*RnnVadTest.DISABLED_RnnVadPerformance* \
> --gtest_also_run_disabled_tests --logs
> ```
>
> Results:
>
> | baseline | this CL
> ------+----------------------+------------------------
> run 1 | 24.0231 +/- 0.591016 | 23.568 +/- 0.990788
> | 370.06x | 377.207x
> ------+----------------------+------------------------
> run 2 | 24.0485 +/- 0.957498 | 23.3714 +/- 0.857523
> | 369.67x | 380.379x
> ------+----------------------+------------------------
> run 2 | 25.4091 +/- 2.6123 | 23.709 +/- 1.04477
> | 349.875x | 374.963x
>
> Bug: webrtc:10480
> Change-Id: I9a3e9164b2442114b928de506c92a547c273882f
> Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/191320
> Reviewed-by: Per Åhgren <peah@webrtc.org>
> Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
> Cr-Commit-Position: refs/heads/master@{#32568}
TBR=alessiob@webrtc.org,peah@webrtc.org
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: webrtc:10480
Change-Id: I2a91f4f29566f872a7dfa220b31c6c625ed075db
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/192660
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Reviewed-by: Alessio Bazzica <alessiob@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32581}
This reverts commit 2f7d1c62e21e2f3786c0803c973d71b414726d8d.
Reason for revert: bug in ancestor CL https://webrtc-review.googlesource.com/c/src/+/191320
Original change's description:
> RNN VAD: pitch search optimizations (part 2)
>
> This CL brings a large improvement to the VAD by precomputing the
> energy for the sliding frame `y` in the pitch buffer instead of
> computing them twice in two different places. The realtime factor
> has improved by about +16x.
>
> There is room for additional improvement (TODOs added), but that will
> be done in a follow up CL since the change won't be bit-exact and
> careful testing is needed.
>
> Benchmarked as follows:
> ```
> out/release/modules_unittests \
> --gtest_filter=*RnnVadTest.DISABLED_RnnVadPerformance* \
> --gtest_also_run_disabled_tests --logs
> ```
>
> Results:
>
> | baseline | this CL
> ------+----------------------+------------------------
> run 1 | 23.568 +/- 0.990788 | 22.8319 +/- 1.46554
> | 377.207x | 389.367x
> ------+----------------------+------------------------
> run 2 | 23.3714 +/- 0.857523 | 22.4286 +/- 0.726449
> | 380.379x | 396.369x
> ------+----------------------+------------------------
> run 2 | 23.709 +/- 1.04477 | 22.5688 +/- 0.831341
> | 374.963x | 393.906x
>
> Bug: webrtc:10480
> Change-Id: I599a4dda2bde16dc6c2f42cf89e96afbd4630311
> Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/191484
> Reviewed-by: Per Åhgren <peah@webrtc.org>
> Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
> Cr-Commit-Position: refs/heads/master@{#32571}
TBR=alessiob@webrtc.org,peah@webrtc.org
Change-Id: I53e478d8d58912c7a5fae4ad8a8d1342a9a48091
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: webrtc:10480
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/192620
Reviewed-by: Alessio Bazzica <alessiob@webrtc.org>
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32580}
This CL brings a large improvement to the VAD by precomputing the
energy for the sliding frame `y` in the pitch buffer instead of
computing them twice in two different places. The realtime factor
has improved by about +16x.
There is room for additional improvement (TODOs added), but that will
be done in a follow up CL since the change won't be bit-exact and
careful testing is needed.
Benchmarked as follows:
```
out/release/modules_unittests \
--gtest_filter=*RnnVadTest.DISABLED_RnnVadPerformance* \
--gtest_also_run_disabled_tests --logs
```
Results:
| baseline | this CL
------+----------------------+------------------------
run 1 | 23.568 +/- 0.990788 | 22.8319 +/- 1.46554
| 377.207x | 389.367x
------+----------------------+------------------------
run 2 | 23.3714 +/- 0.857523 | 22.4286 +/- 0.726449
| 380.379x | 396.369x
------+----------------------+------------------------
run 2 | 23.709 +/- 1.04477 | 22.5688 +/- 0.831341
| 374.963x | 393.906x
Bug: webrtc:10480
Change-Id: I599a4dda2bde16dc6c2f42cf89e96afbd4630311
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/191484
Reviewed-by: Per Åhgren <peah@webrtc.org>
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32571}
TL;DR this CL improves efficiency and includes several code
readability improvements mainly triggered by the comments to
patch set #10.
Highlights:
- Split `FindBestPitchPeriods()` into 12 and 24 kHz versions
to hard-code the input size and simplify the 24 kHz version
- Loop in `ComputePitchPeriod48kHz()` (new name for
`RefinePitchPeriod48kHz()`) removed since the lags for which
we need to compute the auto correlation are a few
- `ComputePitchGainThreshold()` was only used in unit tests; it's been
moved into the anon ns and the test removed
This CL makes `ComputePitchPeriod48kHz()` is about 10% faster (measured
with https://webrtc-review.googlesource.com/c/src/+/191320/4/modules/audio_processing/agc2/rnn_vad/pitch_search_internal_unittest.cc).
The realtime factor has improved by about +14%.
Benchmarked as follows:
```
out/release/modules_unittests \
--gtest_filter=*RnnVadTest.DISABLED_RnnVadPerformance* \
--gtest_also_run_disabled_tests --logs
```
Results:
| baseline | this CL
------+----------------------+------------------------
run 1 | 24.0231 +/- 0.591016 | 23.568 +/- 0.990788
| 370.06x | 377.207x
------+----------------------+------------------------
run 2 | 24.0485 +/- 0.957498 | 23.3714 +/- 0.857523
| 369.67x | 380.379x
------+----------------------+------------------------
run 2 | 25.4091 +/- 2.6123 | 23.709 +/- 1.04477
| 349.875x | 374.963x
Bug: webrtc:10480
Change-Id: I9a3e9164b2442114b928de506c92a547c273882f
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/191320
Reviewed-by: Per Åhgren <peah@webrtc.org>
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32568}
This CL includes two changes:
1. the type for (inverted) lags and pitch periods changed from
size_t to int to reduce the chance of bugs with pitch period
manipulations
2. CheckLowerPitchPeriodsAndComputePitchGain() is optimized by
replacing an unnecessary if statement inside the loop with the
predetermined number of loops
Bug: webrtc:10480
Change-Id: I38432699254b37a2c0111279c28be8dc65b87e9b
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/139252
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Reviewed-by: Gustaf Ullberg <gustaf@webrtc.org>
Reviewed-by: Fredrik Hernqvist <fhernqvist@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#32521}
This CL is the result of running include-what-you-use tool on part
of the code base (audio target and dependencies) plus manual fixes.
bug: webrtc:8311
Change-Id: I277d281ce943c3ecc1bd45fd8d83055931743604
Reviewed-on: https://webrtc-review.googlesource.com/c/106280
Commit-Queue: Yves Gerey <yvesg@google.com>
Reviewed-by: Mirko Bonadei <mbonadei@webrtc.org>
Reviewed-by: Patrik Höglund <phoglund@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#25311}
This CL finalizes the feature extraction part for the RNN VAD adding
a class that combines a high-pass filter, LP residual computation,
pitch estimation and spectral features extraction.
This CL also includes a minor refactoring of the pitch estimation
library.
Bug: webrtc:9076
Change-Id: I918b9e143bc6dd2bf508a891446067258a68a777
Reviewed-on: https://webrtc-review.googlesource.com/75504
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Reviewed-by: Alex Loiko <aleloi@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#23235}
This CL defines SpectralFeaturesExtractor which is responsible for
computing the spectral features used as input for the RNN.
Bug: webrtc:9076
Change-Id: I5e1396b89eca9c13bb268e8419a16436a9c3450f
Reviewed-on: https://webrtc-review.googlesource.com/73760
Commit-Queue: Alessio Bazzica <alessiob@webrtc.org>
Reviewed-by: Alex Loiko <aleloi@webrtc.org>
Reviewed-by: Ivo Creusen <ivoc@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#23206}
During pitch search in the RNN VAD, we calculate auto
correlation. Before this CL, we computed kNumInvertedLags12kHz=147 dot
products of vectors with kBufSize12kHz-kMaxPitch12kHz=240
elements. This was the most time consuming step of the new VAD.
This CL makes the computation happen in frequency domain. Profiling
shows a 3x speed increase. In future, we can try using a more efficient
FFT and to reduce the FFT length to some of e.g. 400, 405, 432.
# For minimal Clang plugin check change.
TBR: kwiberg@webrtc.org
Bug: webrtc:9076
Change-Id: I688251a415869d53175a37f390f441d4e035d954
Reviewed-on: https://webrtc-review.googlesource.com/73366
Reviewed-by: Karl Wiberg <kwiberg@webrtc.org>
Reviewed-by: Alessio Bazzica <alessiob@webrtc.org>
Commit-Queue: Alex Loiko <aleloi@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#23171}
Functions to estimate pitch period and gain.
Bug: webrtc:9076
Change-Id: Icfe9430dcae11bdb96165c5bfe6e2b1d3bf848ab
Reviewed-on: https://webrtc-review.googlesource.com/70382
Commit-Queue: Alex Loiko <aleloi@webrtc.org>
Reviewed-by: Per Åhgren <peah@webrtc.org>
Cr-Commit-Position: refs/heads/master@{#23066}
