This change is needed by ChromeOS as it introduces -fno-omit-frame-pointer
flag (see code.google.com/p/chromium/issues/detail?id=477749). This causes
compile error for MIPS, as some MIPS optimization blocks use maximum possible
number of available registers.
Also, this change contains minor GN build fix for MIPS platform regarding the
pitch_filter_mips.c / pitch_filter_c.c file inclusion.
BUG=477749
R=andrew@webrtc.org, djordje.pesut@imgtec.com, tina.legrand@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/48139004
Patch from Ljubomir Papuga <lpapuga@mips.com>.
Cr-Commit-Position: refs/heads/master@{#9047}
The macro is defined as
#define WEBRTC_SPL_MUL_16_16_RSFT(a, b, c) \
(WEBRTC_SPL_MUL_16_16(a, b) >> (c))
where the latter macro is in C defined as
#define WEBRTC_SPL_MUL_16_16(a, b) \
((int32_t) (((int16_t)(a)) * ((int16_t)(b))))
(For definitions on ARMv7 and MIPS, see common_audio/signal_processing/include/spl_inl_{armv7,mips}.h)
The replacement consists of
- avoiding casts to int16_t if inputs already are int16_t
- adding explicit cast to <type> if result is assigned to <type> (other than int or int32_t)
- minor cleanups like remove of unnecessary parentheses and style changes
- removed commented code lines used during development
- excluded fft.c since there are neon optimizations used and a removal may cause a performance regression
BUG=3348, 3353
TESTED=locally on linux and trybots
R=kwiberg@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/48799004
Cr-Commit-Position: refs/heads/master@{#8967}
The macro is defined as
#define WEBRTC_SPL_LSHIFT_W32(a, b) ((a) << (b))
hence trivial.
The macro name may in fact mislead the user to assume a cast/truncation to int32_t is done.
- Removing usage of it.
- Some style changes.
BUG=3348, 3353
TESTED=locally on linux and trybots
R=kwiberg@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/46749005
Cr-Commit-Position: refs/heads/master@{#8918}
Pass content_browsertests in Chromium. Performance test result (lower is
better):
C version: 100%
old intrinsics Neon version (with bug): 16.5%
new intrinsics Neon version: 18.0%
asm Neon version: 23.3%
BUG=4002
R=andrew@webrtc.org, jridges@masque.com
Change-Id: Ia0a96ac237216b635fc528f67d39319cdf246281
Review URL: https://webrtc-codereview.appspot.com/46739004
Cr-Commit-Position: refs/heads/master@{#8907}
The macro is in C defined as
#define WEBRTC_SPL_MUL_16_16(a, b) ((int32_t) (((int16_t)(a)) * ((int16_t)(b))))
(For definition on ARMv7 and MIPS, see
common_audio/signal_processing/include/spl_inl_armv7.h and
common_audio/signal_processing/include/spl_inl_mips.h)
The replacement consists of
- avoiding casts to int16_t if inputs already are int16_t
- adding explicit cast to <type> if result is assigned to <type> (other than int
or int32_t)
Some other minor code cleanup also exists.
BUG=3348,3353
TESTED=locally on Mac and trybots
R=henrik.lundin@webrtc.org, kwiberg@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/42639004
Cr-Commit-Position: refs/heads/master@{#8717}
git-svn-id: http://webrtc.googlecode.com/svn/trunk@8717 4adac7df-926f-26a2-2b94-8c16560cd09d
Passed building isac_neon and modules_unittests on Android ARM64 and ARMv7.
Passed modules_unittests with following filters:
--gtest_filter=FiltersTest*
--gtest_filter=LpcMaskingModelTest*
--gtest_filter=TransformTest*
--gtest_filter=FilterBanksTest*
WebRtcIsacfix_CalculateResidualEnergyNeon is not enabled due to Issue 4224.
BUG=4002
R=andrew@webrtc.org, jridges@masque.com, kjellander@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/39979004
Patch from Zhongwei Yao <zhongwei.yao@arm.com>.
Cr-Commit-Position: refs/heads/master@{#8632}
git-svn-id: http://webrtc.googlecode.com/svn/trunk@8632 4adac7df-926f-26a2-2b94-8c16560cd09d
This intrinsics version gives bit-exact result as the current C
code. And the performance is 14% better than current assembly
neon version, 3.4 times faster than current C version. The test runs
under Cortex-a53 aarch32 mode, other cpu should give similar performance
result.
Change-Id: Icce5eaf2e17790ce44513d52b53b9f600cc16f96
BUG=4002
R=andrew@webrtc.org, jridges@masque.com
Review URL: https://webrtc-codereview.appspot.com/36689004
Patch from Zhongwei Yao <zhongwei.yao@arm.com>.
git-svn-id: http://webrtc.googlecode.com/svn/trunk@8070 4adac7df-926f-26a2-2b94-8c16560cd09d
This intrinsics version gives bit-exact result as the current assembly
neon code. And the performance is 38% better than current assembly
neon version, 5.92 times faster than current C version. The test runs
under Cortex-a53 aarch32 mode, other cpu should give similar performance
result.
BUG=4002
R=andrew@webrtc.org, jridges@masque.com
Change-Id: I257e33ef6d634a519fd71adc4f52b06dd655bd9d
Review URL: https://webrtc-codereview.appspot.com/32749004
Patch from Zhongwei Yao <zhongwei.yao@arm.com>.
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7891 4adac7df-926f-26a2-2b94-8c16560cd09d
The modification only uses the unique part of the
WebRtcIsacfix_AutocorrC function. Pass FiltersTest.AutocorrFixTest test
on both ARMv7 and ARM64, 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-a15 |
| use C as the base on each | (1.2Ghz) | (1.7Ghz) |
| CPU target | | |
|----------------------------+-----------+------------|
| Neon asm | 24% | 23% |
| Neon intrinsics (GCC 4.6) | 33% | 32% |
| Neon intrinsics (GCC 4.8) | 27% | 27% |
BUG=3850
R=andrew@webrtc.org, jridges@masque.com
Change-Id: Id6cd0671502fadbebd10b1f5493f5b16c988286f
Review URL: https://webrtc-codereview.appspot.com/27999004
Patch from Zhongwei Yao <zhongwei.yao@arm.com>.
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7802 4adac7df-926f-26a2-2b94-8c16560cd09d
Make subroutines for encoding and decoding arrays of 16-bit big-endian
integers, and in the process fix a bug: When decoding an odd number of
bytes from be16, the least significant byte of the last int16 in the
array was properly taken to be zero instead of actually being read
(since it's outside the array). However, when encoding an odd number
of bytes, the least significant byte of the last int16 in the array
was written to the output as-is instead of being taken to be zero;
thus, we encoded one byte more than we should. This was probably not
harmful, and the value was dropped at decoding anyway; nevertheless,
writing a constant zero is the safe thing to do, and this patch does
so.
R=aluebs@webrtc.org, bjornv@webrtc.org, henrik.lundin@webrtc.org, turaj@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/28569004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@7527 4adac7df-926f-26a2-2b94-8c16560cd09d
Macros should in general be avoided. WEBRTC_SPL_UMUL_32_16_RSFT16 is only used in iSAC fixed point as part of multiplying with LSB and MSB. A better approach is to have one function for that complete operation in iSAC.
This CL removes the macro and replace the operation locally.
BUG=3148, 3353
TESTED=locally on Linux and trybots
R=tina.legrand@webrtc.org, turaj@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/16349004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@6907 4adac7df-926f-26a2-2b94-8c16560cd09d
The macro is only used at four places in iSAC fixed point and the macro have been replaced at those places.
In addition, it is used in a unit test, but throws a warning treated as error (issue3674).
The macro has both MIPS and armv7 optimizations. Removing them impacts only MIPS platforms without DSP ASE. This may cause a very small increase in complexity when using iSAC fix.
The armv7 optimizations are not used anywhere, since specific ones are used inline in iSAC fix.
BUG=3348,3353,3674
TESTED=locally and trybots
R=ljubomir.papuga@gmail.com, tina.legrand@webrtc.org, turaj@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/16299004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@6871 4adac7df-926f-26a2-2b94-8c16560cd09d
Implemented functions:
- WebRtcIsacfix_CalculateResidualEnergy
- WebRtcIsacfix_Spec2Time
- WebRtcIsacfix_Time2Spec
- WebRtcIsacfix_HighpassFilterFixDec32
- WebRtcIsacfix_PCorr2Q32
Gain achieved: aprox. further 5% on top of patch#1 on ISAC encoding path.
The optimizations are bit-exact to the C code, with the excception of the
MIPS DSPr2 variant of the WebRtcIsacfix_Time2Spec function (the accuracy of
the WebRtcIsacfix_Time2Spec MIPS DSPr2 variant is same or better than C
variant). Code verification and improvement achieved have been determined
using the iSACFixtest application.
R=andrew@webrtc.org, tina.legrand@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/19749004
Patch from Ljubomir Papuga <lpapuga@mips.com>.
git-svn-id: http://webrtc.googlecode.com/svn/trunk@6749 4adac7df-926f-26a2-2b94-8c16560cd09d
To save memory in iSAC-fix, decoder operated directly on the recieved bitstream. However, this breaks constantness of input when decoder performed in-place big to little Endian conversion. Furthermore, for bit-streams with odd lengths, this meant writing outside the memory. That is because the last byte will be shifted to the Most Significat Byte which might be outside the allocated memory.
If we care about memory, the solution is to do a big-to-little Endian conversion everytime we read a Word16 from the bitstream.
BUG=845,chrome:379458
R=henrik.lundin@webrtc.org, tina.legrand@webrtc.org
Review URL: https://webrtc-codereview.appspot.com/15829004
git-svn-id: http://webrtc.googlecode.com/svn/trunk@6494 4adac7df-926f-26a2-2b94-8c16560cd09d
