The modification only uses the unique part of the analysis_update
function. Pass byte to byte conformance test on both ARMv7 and AArch64,
and the single function performance is similar with original assembly
version on different platforms. If not specified, the code is compiled
by GCC 4.6. The result is the "X version / C version" ratio, and the
less is better.
| run 100k times | cortex-a7 | cortex-a9 | cortex-a15 |
| use C as the base on each | (1.2Ghz) | (1.0Ghz) | (1.7Ghz) |
| CPU target | | | |
|----------------------------+-----------+-----------+------------|
| Neon asm | 15.61% | 20.15% | 14.89% |
| Neon inline asm (LLVM 3.4) | 25.98% | 33.96% | 18.18% |
| Neon intrinsics (GCC 4.6) | 22.06% | 27.01% | 19.24% |
| Neon intrinsics (GCC 4.8) | 17.28% | 18.23% | 18.55% |
| Neon intrinsics (LLVM 3.4) | 21.02% | 19.98% | 16.76% |
BUG=
R=andrew@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/28849004
Patch from Zhongwei Yao <zhongwei.yao@arm.com>.
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7596 4adac7df-926f-26a2-2b94-8c16560cd09d
The modification only uses the unique part of the synthesis_update
function. Pass byte to byte conformance test both on ARMv7 and ARMv8,
and the single function performance is similar with original assembly
version on different platforms (if not specified, the code is compiled
by GCC 4.6):
| run 100k times | cortex-a7 | cortex-a9 | cortex-a15 |
| use C as the base | (1.2Ghz) | (1.0Ghz) | (1.7Ghz) |
| (the smaller the better) | | | |
|----------------------------+-----------+-----------+------------|
| C | 100% | 100% | 100% |
| Neon asm | 15.93% | 17.01% | 12.50% |
| Neon inline asm | 27.74% | 31.41% | 14.64% |
| Neon intrinsics (GCC 4.8) | 17.84% | 14.10% | 13.84% |
| Neon intrinsics (LLVM 3.4) | 16.63% | 14.01% | 12.98% |
BUG=3580
R=andrew@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/23159004
Patch from Zhongwei Yao <zhongwei.yao@arm.com>.
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7586 4adac7df-926f-26a2-2b94-8c16560cd09d
The average speech probability for the higher band is multiplied by the quotient of the process and analyze powers, to avoid thinking that suppressed echo is speech. In order to do this both magnitudes, alanyze and process, needed to be stored. This also was used to calculate different previous STSA estimates for analyze and process.
This CL was tested on two long team member recordings (bjornv and kwiberg) and the noisiest (5) recordings from the QA set.
BUG=webrtc:3763
R=andrew@webrtc.org, bjornv@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/23799004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7437 4adac7df-926f-26a2-2b94-8c16560cd09d
The implementation of WEBRTC_SPL_RSHIFT_W16 is simply >>. This CL removes the macro usage in audio_processing and signal_processing.
Affected components:
* aecm
* agc
* nsx
Indirectly affecting (through signal_processing changes)
* codecs/cng
* codecs/isac/fix
* codecs/isac/main
BUG=3348,3353
TESTED=locally on Linux and trybots
R=kwiberg@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/28699005
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7432 4adac7df-926f-26a2-2b94-8c16560cd09d
> Remove the different block lengths in ns_core
>
> This CL has bit-exact output.
>
> What it does:
> * Remove the blockLen10Ms, as it is hardcoded to be equal to blockLen.
> * This makes outLen to be always zero, so it can be removed too.
> * It also avoids the need to have an outBuf, because it is not used, so it is also removed
> * Replaced blockLen10Ms by blockLen everywhere, since they were hardcoded to be equal.
> * We don't need to check if outLen is zero, because it always is, so it was removed.
> * Of course, the outBuf needs no initial set or copying around, because it is not used.
>
> BUG=webrtc:3811
> R=bjornv@webrtc.org, kwiberg@webrtc.org
>
> Review URL: https://webrtc-codereview.appspot.com/30539004TBR=aluebs@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/26629004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7306 4adac7df-926f-26a2-2b94-8c16560cd09d
This CL has bit-exact output.
What it does:
* Remove the blockLen10Ms, as it is hardcoded to be equal to blockLen.
* This makes outLen to be always zero, so it can be removed too.
* It also avoids the need to have an outBuf, because it is not used, so it is also removed
* Replaced blockLen10Ms by blockLen everywhere, since they were hardcoded to be equal.
* We don't need to check if outLen is zero, because it always is, so it was removed.
* Of course, the outBuf needs no initial set or copying around, because it is not used.
BUG=webrtc:3811
R=bjornv@webrtc.org, kwiberg@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/30539004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7297 4adac7df-926f-26a2-2b94-8c16560cd09d
Filled the empty analyze API, separating the noise estimation from the process API.
No formatting fixes or extra refactoring has been done, to make the review process easier.
This patch has been tested for bit-exactness over the whole QA set in every aggressiveness.
BUG=webrtc:3811
R=bjornv@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/27549004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7243 4adac7df-926f-26a2-2b94-8c16560cd09d
The macro replaced is a trivial multiplication after explicit casts to uint16_t and uint32_t. This CL replaces its use with "*" and adds explicit casts if necessary.
Affected components:
* AECMobile
* AGC
* Noise Suppression (fixed point version)
BUG=3348,3353
TESTED=locally on linux and trybots
R=kwiberg@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/27389004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7101 4adac7df-926f-26a2-2b94-8c16560cd09d
In the fixed point version of the Noise Suppression. At one place we subtract a value in the wrong Q-domain, which later may cause a divide by zero. Going through the floating point code that particular variable should be zero if this happens, which is what the old code tried to accomplish, but in an awkward way.
The bug has been there since development, so the likelihood of actually get a divide by zero is very small.
BUG=chromium:407812
R=ljubomir.papuga@gmail.com, tina.legrand@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/17229004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7035 4adac7df-926f-26a2-2b94-8c16560cd09d
The modification only uses the unique part of the spectrum (as is done for the C and asm code). It passes
byte to byte conformance test, and the single function performance
(if not specified, the code is compiled by GCC 4.6) on different
platforms:
| run 100k times | cortex-a7 | cortex-a9 | cortex-a15 |
| use C as the base on each | (1.2Ghz) | (1.0Ghz) | (1.7Ghz) |
| CPU target | | | |
|----------------------------+-----------+-----------+------------|
| C | 100% | 100% | 100% |
| Neon asm | 18% | 14% | 19% |
| Neon inline asm | 31% | 25% | 27% |
| Neon intrinsic (GCC 4.6) | 33% | 27% | 42% |
| Neon intrinscis (GCC 4.8) | 17% | 14% | 19% |
| Neon intrinsics (LLVM 3.3) | 15% | 13% | 18% |
BUG=
R=andrew@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/13739004
Patch from Joe Yu <joe.yu@arm.com>.
git-svn-id: http://webrtc.googlecode.com/svn/trunk@6920 4adac7df-926f-26a2-2b94-8c16560cd09d
Internally, it already worked on floats. This patch just changes the
signature of a bunch of functions so that floats can be passed
directly from the new and improved AudioBuffer without converting the
data to int and back again first.
(The reference data to the ApmTest.Process test had to be modified
slightly; this is because the noise suppressor comes immediately after
the echo canceller, which also works on floats. If I truncate to
integers between the two steps, ApmTest.Process doesn't complain, but
of course that's exactly the sort of thing the float conversion is
supposed to let us avoid...)
BUG=
R=aluebs@webrtc.org, bjornv@webrtc.org, tina.legrand@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/13519004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@6385 4adac7df-926f-26a2-2b94-8c16560cd09d
