diff mbox series

[FFmpeg-devel,v2,1/5] ac3enc_fixed: convert to 32-bit sample format

Message ID MQrn3cw--3-2@lynne.ee
State New
Headers show
Series [FFmpeg-devel,v2,1/5] ac3enc_fixed: convert to 32-bit sample format
Related show

Checks

Context Check Description
andriy/x86_make success Make finished
andriy/x86_make_fate fail Make fate failed
andriy/PPC64_make success Make finished
andriy/PPC64_make_fate warning Make fate failed

Commit Message

Lynne Jan. 12, 2021, 6:12 p.m. UTC
The AC3 encoder used to be a separate library called "Aften", which
got merged into libavcodec (literally, SVN commits and all).
The merge preserved as much features from the library as possible.

The code had two versions - a fixed point version and a floating
point version. FFmpeg had floating point DSP code used by other
codecs, the AC3 decoder including, so the floating-point DSP was
simply replaced with FFmpeg's own functions.
However, FFmpeg had no fixed-point audio code at that point. So
the encoder brought along its own fixed-point DSP functions,
including a fixed-point MDCT.

The fixed-point MDCT itself is trivially just a float MDCT with a
different type and each multiply being a fixed-point multiply.
So over time, it got refactored, and the FFT used for all other codecs
was templated.

Due to design decisions at the time, the fixed-point version of the
encoder operates at 16-bits of precision. Although convenient, this,
even at the time, was inadequate and inefficient. The encoder is noisy,
does not produce output comparable to the float encoder, and even
rings at higher frequencies due to the badly approximated winow function.

Enter MIPS (owned by Imagination Technologies at the time). They wanted
quick fixed-point decoding on their FPUless cores. So they contributed
patches to template the AC3 decoder so it had both a fixed-point
and a floating-point version. They also did the same for the AAC decoder.
They however, used 32-bit samples. Not 16-bits. And we did not have
32-bit fixed-point DSP functions, including an MDCT. But instead of
templating our MDCT to output 3 versions (float, 32-bit fixed and 16-bit fixed),
they simply copy-pasted their own MDCT into ours, and completely
ifdeffed our own MDCT code out if a 32-bit fixed point MDCT was selected.

This is also the status quo nowadays - 2 separate MDCTs, one which
produces floating point and 16-bit fixed point versions, and one
sort-of integrated which produces 32-bit MDCT.

MIPS weren't all that interested in encoding, so they left the encoder
as-is, and they didn't care much about the ifdeffery, mess or quality - it's
not their problem.

So the MDCT/FFT code has always been a thorn in anyone looking to clean up
code's eye.

Backstory over. Internally AC3 operates on 25-bit fixed-point coefficients.
So for the floating point version, the encoder simply runs the float MDCT,
and converts the resulting coefficients to 25-bit fixed-point, as AC3 is inherently
a fixed-point codec. For the fixed-point version, the input is 16-bit samples,
so to maximize precision the frame samples are analyzed and the highest set
bit is detected via ac3_max_msb_abs_int16(), and the coefficients are then
scaled up via ac3_lshift_int16(), so the input for the FFT is always at least 14 bits,
computed in normalize_samples(). After FFT, the coefficients are scaled up to 25 bits.

This patch simply changes the encoder to accept 32-bit samples, reusing
the already well-optimized 32-bit MDCT code, allowing us to clean up and drop
a large part of a very messy code of ours, as well as prepare for the future lavu/tx
conversion. The coefficients are simply scaled down to 25 bits during windowing,
skipping 2 separate scalings, as the hacks to extend precision are simply no longer
necessary. There's no point in running the MDCT always at 32 bits when you're
going to drop 6 bits off anyway, the headroom is plenty, and the MDCT rounds
properly.

This also makes the encoder even slightly more accurate over the float version,
as there's no coefficient conversion step necessary.

SIZE SAVINGS:
ARM32:
HARDCODED TABLES:
BASE           - 10709590
DROP  DSP      - 10702872 - diff:   -6.56KiB
DROP  MDCT     - 10667932 - diff:  -34.12KiB - both:   -40.68KiB
DROP  FFT      - 10336652 - diff: -323.52KiB - all:   -364.20KiB
SOFTCODED TABLES:
BASE           -  9685096
DROP  DSP      -  9678378 - diff:   -6.56KiB
DROP  MDCT     -  9643466 - diff:  -34.09KiB - both:   -40.65KiB
DROP  FFT      -  9573918 - diff:  -67.92KiB - all:   -108.57KiB

ARM64:
HARDCODED TABLES:
BASE           - 14641112
DROP  DSP      - 14633806 - diff:   -7.13KiB
DROP  MDCT     - 14604812 - diff:  -28.31KiB - both:   -35.45KiB
DROP  FFT      - 14286826 - diff: -310.53KiB - all:   -345.98KiB
SOFTCODED TABLES:
BASE           - 13636238
DROP  DSP      - 13628932 - diff:   -7.13KiB
DROP  MDCT     - 13599866 - diff:  -28.38KiB - both:   -35.52KiB
DROP  FFT      - 13542080 - diff:  -56.43KiB - all:    -91.95KiB

x86:
HARDCODED TABLES:
BASE           - 12367336
DROP  DSP      - 12354698 - diff:  -12.34KiB
DROP  MDCT     - 12331024 - diff:  -23.12KiB - both:   -35.46KiB
DROP  FFT      - 12029788 - diff: -294.18KiB - all:   -329.64KiB
SOFTCODED TABLES:
BASE           - 11358094
DROP  DSP      - 11345456 - diff:  -12.34KiB
DROP  MDCT     - 11321742 - diff:  -23.16KiB - both:   -35.50KiB
DROP  FFT      - 11276946 - diff:  -43.75KiB - all:    -79.25KiB

PERFORMANCE (10min random s32le):
ARM32 - before -  39.9x - 0m15.046s
ARM32 - after  -  28.2x - 0m21.525s
                       Speed:  -30%

ARM64 - before -  36.1x - 0m16.637s
ARM64 - after  -  36.0x - 0m16.727s
                       Speed: -0.5%

x86   - before - 184x -    0m3.277s
x86   - after  - 190x -    0m3.187s
                       Speed:   +3%

New patch attached.
Subject: [PATCH v2 1/5] ac3enc_fixed: convert to 32-bit sample format

The AC3 encoder used to be a separate library called "Aften", which
got merged into libavcodec (literally, SVN commits and all).
The merge preserved as much features from the library as possible.

The code had two versions - a fixed point version and a floating
point version. FFmpeg had floating point DSP code used by other
codecs, the AC3 decoder including, so the floating-point DSP was
simply replaced with FFmpeg's own functions.
However, FFmpeg had no fixed-point audio code at that point. So
the encoder brought along its own fixed-point DSP functions,
including a fixed-point MDCT.

The fixed-point MDCT itself is trivially just a float MDCT with a
different type and each multiply being a fixed-point multiply.
So over time, it got refactored, and the FFT used for all other codecs
was templated.

Due to design decisions at the time, the fixed-point version of the
encoder operates at 16-bits of precision. Although convenient, this,
even at the time, was inadequate and inefficient. The encoder is noisy,
does not produce output comparable to the float encoder, and even
rings at higher frequencies due to the badly approximated winow function.

Enter MIPS (owned by Imagination Technologies at the time). They wanted
quick fixed-point decoding on their FPUless cores. So they contributed
patches to template the AC3 decoder so it had both a fixed-point
and a floating-point version. They also did the same for the AAC decoder.
They however, used 32-bit samples. Not 16-bits. And we did not have
32-bit fixed-point DSP functions, including an MDCT. But instead of
templating our MDCT to output 3 versions (float, 32-bit fixed and 16-bit fixed),
they simply copy-pasted their own MDCT into ours, and completely
ifdeffed our own MDCT code out if a 32-bit fixed point MDCT was selected.

This is also the status quo nowadays - 2 separate MDCTs, one which
produces floating point and 16-bit fixed point versions, and one
sort-of integrated which produces 32-bit MDCT.

MIPS weren't all that interested in encoding, so they left the encoder
as-is, and they didn't care much about the ifdeffery, mess or quality - it's
not their problem.

So the MDCT/FFT code has always been a thorn in anyone looking to clean up
code's eye.

Backstory over. Internally AC3 operates on 25-bit fixed-point coefficients.
So for the floating point version, the encoder simply runs the float MDCT,
and converts the resulting coefficients to 25-bit fixed-point, as AC3 is inherently
a fixed-point codec. For the fixed-point version, the input is 16-bit samples,
so to maximize precision the frame samples are analyzed and the highest set
bit is detected via ac3_max_msb_abs_int16(), and the coefficients are then
scaled up via ac3_lshift_int16(), so the input for the FFT is always at least 14 bits,
computed in normalize_samples(). After FFT, the coefficients are scaled up to 25 bits.

This patch simply changes the encoder to accept 32-bit samples, reusing
the already well-optimized 32-bit MDCT code, allowing us to clean up and drop
a large part of a very messy code of ours, as well as prepare for the future lavu/tx
conversion. The coefficients are simply scaled down to 25 bits during windowing,
skipping 2 separate scalings, as the hacks to extend precision are simply no longer
necessary. There's no point in running the MDCT always at 32 bits when you're
going to drop 6 bits off anyway, the headroom is plenty, and the MDCT rounds
properly.

This also makes the encoder even slightly more accurate over the float version,
as there's no coefficient conversion step necessary.

SIZE SAVINGS:
ARM32:
HARDCODED TABLES:
BASE           - 10709590
DROP  DSP      - 10702872 - diff:   -6.56KiB
DROP  MDCT     - 10667932 - diff:  -34.12KiB - both:   -40.68KiB
DROP  FFT      - 10336652 - diff: -323.52KiB - all:   -364.20KiB
SOFTCODED TABLES:
BASE           -  9685096
DROP  DSP      -  9678378 - diff:   -6.56KiB
DROP  MDCT     -  9643466 - diff:  -34.09KiB - both:   -40.65KiB
DROP  FFT      -  9573918 - diff:  -67.92KiB - all:   -108.57KiB

ARM64:
HARDCODED TABLES:
BASE           - 14641112
DROP  DSP      - 14633806 - diff:   -7.13KiB
DROP  MDCT     - 14604812 - diff:  -28.31KiB - both:   -35.45KiB
DROP  FFT      - 14286826 - diff: -310.53KiB - all:   -345.98KiB
SOFTCODED TABLES:
BASE           - 13636238
DROP  DSP      - 13628932 - diff:   -7.13KiB
DROP  MDCT     - 13599866 - diff:  -28.38KiB - both:   -35.52KiB
DROP  FFT      - 13542080 - diff:  -56.43KiB - all:    -91.95KiB

x86:
HARDCODED TABLES:
BASE           - 12367336
DROP  DSP      - 12354698 - diff:  -12.34KiB
DROP  MDCT     - 12331024 - diff:  -23.12KiB - both:   -35.46KiB
DROP  FFT      - 12029788 - diff: -294.18KiB - all:   -329.64KiB
SOFTCODED TABLES:
BASE           - 11358094
DROP  DSP      - 11345456 - diff:  -12.34KiB
DROP  MDCT     - 11321742 - diff:  -23.16KiB - both:   -35.50KiB
DROP  FFT      - 11276946 - diff:  -43.75KiB - all:    -79.25KiB

PERFORMANCE (10min random s32le):
ARM32 - before -  39.9x - 0m15.046s
ARM32 - after  -  28.2x - 0m21.525s
                       Speed:  -30%

ARM64 - before -  36.1x - 0m16.637s
ARM64 - after  -  36.0x - 0m16.727s
                       Speed: -0.5%

x86   - before - 184x -    0m3.277s
x86   - after  - 190x -    0m3.187s
                       Speed:   +3%
---
 doc/encoders.texi                 |  7 ++--
 libavcodec/Makefile               |  2 +-
 libavcodec/ac3enc.c               |  1 +
 libavcodec/ac3enc.h               | 11 +++---
 libavcodec/ac3enc_fixed.c         | 59 ++++++++++++-------------------
 libavcodec/ac3enc_float.c         |  1 -
 libavcodec/ac3enc_template.c      | 21 ++++-------
 libavcodec/version.h              |  2 +-
 tests/fate/ac3.mak                |  2 +-
 tests/fate/ffmpeg.mak             |  2 +-
 tests/ref/fate/unknown_layout-ac3 |  2 +-
 tests/ref/lavf/rm                 |  2 +-
 12 files changed, 45 insertions(+), 67 deletions(-)

Comments

Andreas Rheinhardt Jan. 12, 2021, 9:24 p.m. UTC | #1
Lynne:
> The AC3 encoder used to be a separate library called "Aften", which
> got merged into libavcodec (literally, SVN commits and all).
> The merge preserved as much features from the library as possible.
> 
> The code had two versions - a fixed point version and a floating
> point version. FFmpeg had floating point DSP code used by other
> codecs, the AC3 decoder including, so the floating-point DSP was
> simply replaced with FFmpeg's own functions.
> However, FFmpeg had no fixed-point audio code at that point. So
> the encoder brought along its own fixed-point DSP functions,
> including a fixed-point MDCT.
> 
> The fixed-point MDCT itself is trivially just a float MDCT with a
> different type and each multiply being a fixed-point multiply.
> So over time, it got refactored, and the FFT used for all other codecs
> was templated.
> 
> Due to design decisions at the time, the fixed-point version of the
> encoder operates at 16-bits of precision. Although convenient, this,
> even at the time, was inadequate and inefficient. The encoder is noisy,
> does not produce output comparable to the float encoder, and even
> rings at higher frequencies due to the badly approximated winow function.
> 
> Enter MIPS (owned by Imagination Technologies at the time). They wanted
> quick fixed-point decoding on their FPUless cores. So they contributed
> patches to template the AC3 decoder so it had both a fixed-point
> and a floating-point version. They also did the same for the AAC decoder.
> They however, used 32-bit samples. Not 16-bits. And we did not have
> 32-bit fixed-point DSP functions, including an MDCT. But instead of
> templating our MDCT to output 3 versions (float, 32-bit fixed and 16-bit fixed),
> they simply copy-pasted their own MDCT into ours, and completely
> ifdeffed our own MDCT code out if a 32-bit fixed point MDCT was selected.
> 
> This is also the status quo nowadays - 2 separate MDCTs, one which
> produces floating point and 16-bit fixed point versions, and one
> sort-of integrated which produces 32-bit MDCT.
> 
> MIPS weren't all that interested in encoding, so they left the encoder
> as-is, and they didn't care much about the ifdeffery, mess or quality - it's
> not their problem.
> 
> So the MDCT/FFT code has always been a thorn in anyone looking to clean up
> code's eye.
> 
> Backstory over. Internally AC3 operates on 25-bit fixed-point coefficients.
> So for the floating point version, the encoder simply runs the float MDCT,
> and converts the resulting coefficients to 25-bit fixed-point, as AC3 is inherently
> a fixed-point codec. For the fixed-point version, the input is 16-bit samples,
> so to maximize precision the frame samples are analyzed and the highest set
> bit is detected via ac3_max_msb_abs_int16(), and the coefficients are then
> scaled up via ac3_lshift_int16(), so the input for the FFT is always at least 14 bits,
> computed in normalize_samples(). After FFT, the coefficients are scaled up to 25 bits.
> 
> This patch simply changes the encoder to accept 32-bit samples, reusing
> the already well-optimized 32-bit MDCT code, allowing us to clean up and drop
> a large part of a very messy code of ours, as well as prepare for the future lavu/tx
> conversion. The coefficients are simply scaled down to 25 bits during windowing,
> skipping 2 separate scalings, as the hacks to extend precision are simply no longer
> necessary. There's no point in running the MDCT always at 32 bits when you're
> going to drop 6 bits off anyway, the headroom is plenty, and the MDCT rounds
> properly.
> 
> This also makes the encoder even slightly more accurate over the float version,
> as there's no coefficient conversion step necessary.
> 
> SIZE SAVINGS:
> ARM32:
> HARDCODED TABLES:
> BASE           - 10709590
> DROP  DSP      - 10702872 - diff:   -6.56KiB
> DROP  MDCT     - 10667932 - diff:  -34.12KiB - both:   -40.68KiB
> DROP  FFT      - 10336652 - diff: -323.52KiB - all:   -364.20KiB
> SOFTCODED TABLES:
> BASE           -  9685096
> DROP  DSP      -  9678378 - diff:   -6.56KiB
> DROP  MDCT     -  9643466 - diff:  -34.09KiB - both:   -40.65KiB
> DROP  FFT      -  9573918 - diff:  -67.92KiB - all:   -108.57KiB
> 
> ARM64:
> HARDCODED TABLES:
> BASE           - 14641112
> DROP  DSP      - 14633806 - diff:   -7.13KiB
> DROP  MDCT     - 14604812 - diff:  -28.31KiB - both:   -35.45KiB
> DROP  FFT      - 14286826 - diff: -310.53KiB - all:   -345.98KiB
> SOFTCODED TABLES:
> BASE           - 13636238
> DROP  DSP      - 13628932 - diff:   -7.13KiB
> DROP  MDCT     - 13599866 - diff:  -28.38KiB - both:   -35.52KiB
> DROP  FFT      - 13542080 - diff:  -56.43KiB - all:    -91.95KiB
> 
> x86:
> HARDCODED TABLES:
> BASE           - 12367336
> DROP  DSP      - 12354698 - diff:  -12.34KiB
> DROP  MDCT     - 12331024 - diff:  -23.12KiB - both:   -35.46KiB
> DROP  FFT      - 12029788 - diff: -294.18KiB - all:   -329.64KiB
> SOFTCODED TABLES:
> BASE           - 11358094
> DROP  DSP      - 11345456 - diff:  -12.34KiB
> DROP  MDCT     - 11321742 - diff:  -23.16KiB - both:   -35.50KiB
> DROP  FFT      - 11276946 - diff:  -43.75KiB - all:    -79.25KiB
> 
> PERFORMANCE (10min random s32le):
> ARM32 - before -  39.9x - 0m15.046s
> ARM32 - after  -  28.2x - 0m21.525s
>                        Speed:  -30%
> 
> ARM64 - before -  36.1x - 0m16.637s
> ARM64 - after  -  36.0x - 0m16.727s
>                        Speed: -0.5%
> 
> x86   - before - 184x -    0m3.277s
> x86   - after  - 190x -    0m3.187s
>                        Speed:   +3%
> 
> New patch attached.
> 
FATE still fails with patches 1 and 2 on patchwork.

- Andreas
Lynne Jan. 12, 2021, 11:55 p.m. UTC | #2
Jan 12, 2021, 22:24 by andreas.rheinhardt@gmail.com:

> Lynne:
>
>> The AC3 encoder used to be a separate library called "Aften", which
>> got merged into libavcodec (literally, SVN commits and all).
>> The merge preserved as much features from the library as possible.
>>
>> The code had two versions - a fixed point version and a floating
>> point version. FFmpeg had floating point DSP code used by other
>> codecs, the AC3 decoder including, so the floating-point DSP was
>> simply replaced with FFmpeg's own functions.
>> However, FFmpeg had no fixed-point audio code at that point. So
>> the encoder brought along its own fixed-point DSP functions,
>> including a fixed-point MDCT.
>>
>> The fixed-point MDCT itself is trivially just a float MDCT with a
>> different type and each multiply being a fixed-point multiply.
>> So over time, it got refactored, and the FFT used for all other codecs
>> was templated.
>>
>> Due to design decisions at the time, the fixed-point version of the
>> encoder operates at 16-bits of precision. Although convenient, this,
>> even at the time, was inadequate and inefficient. The encoder is noisy,
>> does not produce output comparable to the float encoder, and even
>> rings at higher frequencies due to the badly approximated winow function.
>>
>> Enter MIPS (owned by Imagination Technologies at the time). They wanted
>> quick fixed-point decoding on their FPUless cores. So they contributed
>> patches to template the AC3 decoder so it had both a fixed-point
>> and a floating-point version. They also did the same for the AAC decoder.
>> They however, used 32-bit samples. Not 16-bits. And we did not have
>> 32-bit fixed-point DSP functions, including an MDCT. But instead of
>> templating our MDCT to output 3 versions (float, 32-bit fixed and 16-bit fixed),
>> they simply copy-pasted their own MDCT into ours, and completely
>> ifdeffed our own MDCT code out if a 32-bit fixed point MDCT was selected.
>>
>> This is also the status quo nowadays - 2 separate MDCTs, one which
>> produces floating point and 16-bit fixed point versions, and one
>> sort-of integrated which produces 32-bit MDCT.
>>
>> MIPS weren't all that interested in encoding, so they left the encoder
>> as-is, and they didn't care much about the ifdeffery, mess or quality - it's
>> not their problem.
>>
>> So the MDCT/FFT code has always been a thorn in anyone looking to clean up
>> code's eye.
>>
>> Backstory over. Internally AC3 operates on 25-bit fixed-point coefficients.
>> So for the floating point version, the encoder simply runs the float MDCT,
>> and converts the resulting coefficients to 25-bit fixed-point, as AC3 is inherently
>> a fixed-point codec. For the fixed-point version, the input is 16-bit samples,
>> so to maximize precision the frame samples are analyzed and the highest set
>> bit is detected via ac3_max_msb_abs_int16(), and the coefficients are then
>> scaled up via ac3_lshift_int16(), so the input for the FFT is always at least 14 bits,
>> computed in normalize_samples(). After FFT, the coefficients are scaled up to 25 bits.
>>
>> This patch simply changes the encoder to accept 32-bit samples, reusing
>> the already well-optimized 32-bit MDCT code, allowing us to clean up and drop
>> a large part of a very messy code of ours, as well as prepare for the future lavu/tx
>> conversion. The coefficients are simply scaled down to 25 bits during windowing,
>> skipping 2 separate scalings, as the hacks to extend precision are simply no longer
>> necessary. There's no point in running the MDCT always at 32 bits when you're
>> going to drop 6 bits off anyway, the headroom is plenty, and the MDCT rounds
>> properly.
>>
>> This also makes the encoder even slightly more accurate over the float version,
>> as there's no coefficient conversion step necessary.
>>
>> SIZE SAVINGS:
>> ARM32:
>> HARDCODED TABLES:
>> BASE           - 10709590
>> DROP  DSP      - 10702872 - diff:   -6.56KiB
>> DROP  MDCT     - 10667932 - diff:  -34.12KiB - both:   -40.68KiB
>> DROP  FFT      - 10336652 - diff: -323.52KiB - all:   -364.20KiB
>> SOFTCODED TABLES:
>> BASE           -  9685096
>> DROP  DSP      -  9678378 - diff:   -6.56KiB
>> DROP  MDCT     -  9643466 - diff:  -34.09KiB - both:   -40.65KiB
>> DROP  FFT      -  9573918 - diff:  -67.92KiB - all:   -108.57KiB
>>
>> ARM64:
>> HARDCODED TABLES:
>> BASE           - 14641112
>> DROP  DSP      - 14633806 - diff:   -7.13KiB
>> DROP  MDCT     - 14604812 - diff:  -28.31KiB - both:   -35.45KiB
>> DROP  FFT      - 14286826 - diff: -310.53KiB - all:   -345.98KiB
>> SOFTCODED TABLES:
>> BASE           - 13636238
>> DROP  DSP      - 13628932 - diff:   -7.13KiB
>> DROP  MDCT     - 13599866 - diff:  -28.38KiB - both:   -35.52KiB
>> DROP  FFT      - 13542080 - diff:  -56.43KiB - all:    -91.95KiB
>>
>> x86:
>> HARDCODED TABLES:
>> BASE           - 12367336
>> DROP  DSP      - 12354698 - diff:  -12.34KiB
>> DROP  MDCT     - 12331024 - diff:  -23.12KiB - both:   -35.46KiB
>> DROP  FFT      - 12029788 - diff: -294.18KiB - all:   -329.64KiB
>> SOFTCODED TABLES:
>> BASE           - 11358094
>> DROP  DSP      - 11345456 - diff:  -12.34KiB
>> DROP  MDCT     - 11321742 - diff:  -23.16KiB - both:   -35.50KiB
>> DROP  FFT      - 11276946 - diff:  -43.75KiB - all:    -79.25KiB
>>
>> PERFORMANCE (10min random s32le):
>> ARM32 - before -  39.9x - 0m15.046s
>> ARM32 - after  -  28.2x - 0m21.525s
>>                        Speed:  -30%
>>
>> ARM64 - before -  36.1x - 0m16.637s
>> ARM64 - after  -  36.0x - 0m16.727s
>>                        Speed: -0.5%
>>
>> x86   - before - 184x -    0m3.277s
>> x86   - after  - 190x -    0m3.187s
>>                        Speed:   +3%
>>
>> New patch attached.
>>
> FATE still fails with patches 1 and 2 on patchwork.
>

Fixed. Forgot a line. New patches attached.
Lynne Jan. 14, 2021, 1:52 a.m. UTC | #3
Jan 13, 2021, 00:55 by dev@lynne.ee:

> Jan 12, 2021, 22:24 by andreas.rheinhardt@gmail.com:
>
>> Lynne:
>>
>>> The AC3 encoder used to be a separate library called "Aften", which
>>> got merged into libavcodec (literally, SVN commits and all).
>>> The merge preserved as much features from the library as possible.
>>>
>>> The code had two versions - a fixed point version and a floating
>>> point version. FFmpeg had floating point DSP code used by other
>>> codecs, the AC3 decoder including, so the floating-point DSP was
>>> simply replaced with FFmpeg's own functions.
>>> However, FFmpeg had no fixed-point audio code at that point. So
>>> the encoder brought along its own fixed-point DSP functions,
>>> including a fixed-point MDCT.
>>>
>>> The fixed-point MDCT itself is trivially just a float MDCT with a
>>> different type and each multiply being a fixed-point multiply.
>>> So over time, it got refactored, and the FFT used for all other codecs
>>> was templated.
>>>
>>> Due to design decisions at the time, the fixed-point version of the
>>> encoder operates at 16-bits of precision. Although convenient, this,
>>> even at the time, was inadequate and inefficient. The encoder is noisy,
>>> does not produce output comparable to the float encoder, and even
>>> rings at higher frequencies due to the badly approximated winow function.
>>>
>>> Enter MIPS (owned by Imagination Technologies at the time). They wanted
>>> quick fixed-point decoding on their FPUless cores. So they contributed
>>> patches to template the AC3 decoder so it had both a fixed-point
>>> and a floating-point version. They also did the same for the AAC decoder.
>>> They however, used 32-bit samples. Not 16-bits. And we did not have
>>> 32-bit fixed-point DSP functions, including an MDCT. But instead of
>>> templating our MDCT to output 3 versions (float, 32-bit fixed and 16-bit fixed),
>>> they simply copy-pasted their own MDCT into ours, and completely
>>> ifdeffed our own MDCT code out if a 32-bit fixed point MDCT was selected.
>>>
>>> This is also the status quo nowadays - 2 separate MDCTs, one which
>>> produces floating point and 16-bit fixed point versions, and one
>>> sort-of integrated which produces 32-bit MDCT.
>>>
>>> MIPS weren't all that interested in encoding, so they left the encoder
>>> as-is, and they didn't care much about the ifdeffery, mess or quality - it's
>>> not their problem.
>>>
>>> So the MDCT/FFT code has always been a thorn in anyone looking to clean up
>>> code's eye.
>>>
>>> Backstory over. Internally AC3 operates on 25-bit fixed-point coefficients.
>>> So for the floating point version, the encoder simply runs the float MDCT,
>>> and converts the resulting coefficients to 25-bit fixed-point, as AC3 is inherently
>>> a fixed-point codec. For the fixed-point version, the input is 16-bit samples,
>>> so to maximize precision the frame samples are analyzed and the highest set
>>> bit is detected via ac3_max_msb_abs_int16(), and the coefficients are then
>>> scaled up via ac3_lshift_int16(), so the input for the FFT is always at least 14 bits,
>>> computed in normalize_samples(). After FFT, the coefficients are scaled up to 25 bits.
>>>
>>> This patch simply changes the encoder to accept 32-bit samples, reusing
>>> the already well-optimized 32-bit MDCT code, allowing us to clean up and drop
>>> a large part of a very messy code of ours, as well as prepare for the future lavu/tx
>>> conversion. The coefficients are simply scaled down to 25 bits during windowing,
>>> skipping 2 separate scalings, as the hacks to extend precision are simply no longer
>>> necessary. There's no point in running the MDCT always at 32 bits when you're
>>> going to drop 6 bits off anyway, the headroom is plenty, and the MDCT rounds
>>> properly.
>>>
>>> This also makes the encoder even slightly more accurate over the float version,
>>> as there's no coefficient conversion step necessary.
>>>
>>> SIZE SAVINGS:
>>> ARM32:
>>> HARDCODED TABLES:
>>> BASE           - 10709590
>>> DROP  DSP      - 10702872 - diff:   -6.56KiB
>>> DROP  MDCT     - 10667932 - diff:  -34.12KiB - both:   -40.68KiB
>>> DROP  FFT      - 10336652 - diff: -323.52KiB - all:   -364.20KiB
>>> SOFTCODED TABLES:
>>> BASE           -  9685096
>>> DROP  DSP      -  9678378 - diff:   -6.56KiB
>>> DROP  MDCT     -  9643466 - diff:  -34.09KiB - both:   -40.65KiB
>>> DROP  FFT      -  9573918 - diff:  -67.92KiB - all:   -108.57KiB
>>>
>>> ARM64:
>>> HARDCODED TABLES:
>>> BASE           - 14641112
>>> DROP  DSP      - 14633806 - diff:   -7.13KiB
>>> DROP  MDCT     - 14604812 - diff:  -28.31KiB - both:   -35.45KiB
>>> DROP  FFT      - 14286826 - diff: -310.53KiB - all:   -345.98KiB
>>> SOFTCODED TABLES:
>>> BASE           - 13636238
>>> DROP  DSP      - 13628932 - diff:   -7.13KiB
>>> DROP  MDCT     - 13599866 - diff:  -28.38KiB - both:   -35.52KiB
>>> DROP  FFT      - 13542080 - diff:  -56.43KiB - all:    -91.95KiB
>>>
>>> x86:
>>> HARDCODED TABLES:
>>> BASE           - 12367336
>>> DROP  DSP      - 12354698 - diff:  -12.34KiB
>>> DROP  MDCT     - 12331024 - diff:  -23.12KiB - both:   -35.46KiB
>>> DROP  FFT      - 12029788 - diff: -294.18KiB - all:   -329.64KiB
>>> SOFTCODED TABLES:
>>> BASE           - 11358094
>>> DROP  DSP      - 11345456 - diff:  -12.34KiB
>>> DROP  MDCT     - 11321742 - diff:  -23.16KiB - both:   -35.50KiB
>>> DROP  FFT      - 11276946 - diff:  -43.75KiB - all:    -79.25KiB
>>>
>>> PERFORMANCE (10min random s32le):
>>> ARM32 - before -  39.9x - 0m15.046s
>>> ARM32 - after  -  28.2x - 0m21.525s
>>>                        Speed:  -30%
>>>
>>> ARM64 - before -  36.1x - 0m16.637s
>>> ARM64 - after  -  36.0x - 0m16.727s
>>>                        Speed: -0.5%
>>>
>>> x86   - before - 184x -    0m3.277s
>>> x86   - after  - 190x -    0m3.187s
>>>                        Speed:   +3%
>>>
>>> New patch attached.
>>>
>> FATE still fails with patches 1 and 2 on patchwork.
>>
>
> Fixed. Forgot a line. New patches attached.
>

Pushed. Thanks to all who commented.
Next up will likely be the mdct15 removal, which will be around 80KiB on all platforms.

Fun fact: commit 4/5 removed MMX code. Not something you see every day.
diff mbox series

Patch

diff --git a/doc/encoders.texi b/doc/encoders.texi
index 0b1c69e982..60e763a704 100644
--- a/doc/encoders.texi
+++ b/doc/encoders.texi
@@ -151,10 +151,9 @@  the undocumented RealAudio 3 (a.k.a. dnet).
 The @var{ac3} encoder uses floating-point math, while the @var{ac3_fixed}
 encoder only uses fixed-point integer math. This does not mean that one is
 always faster, just that one or the other may be better suited to a
-particular system. The floating-point encoder will generally produce better
-quality audio for a given bitrate. The @var{ac3_fixed} encoder is not the
-default codec for any of the output formats, so it must be specified explicitly
-using the option @code{-acodec ac3_fixed} in order to use it.
+particular system. The @var{ac3_fixed} encoder is not the default codec for
+any of the output formats, so it must be specified explicitly using the option
+@code{-acodec ac3_fixed} in order to use it.
 
 @subsection AC-3 Metadata
 
diff --git a/libavcodec/Makefile b/libavcodec/Makefile
index 35318f4f4d..0546e6f6c5 100644
--- a/libavcodec/Makefile
+++ b/libavcodec/Makefile
@@ -181,7 +181,7 @@  OBJS-$(CONFIG_AC3_DECODER)             += ac3dec_float.o ac3dec_data.o ac3.o kbd
 OBJS-$(CONFIG_AC3_FIXED_DECODER)       += ac3dec_fixed.o ac3dec_data.o ac3.o kbdwin.o ac3tab.o
 OBJS-$(CONFIG_AC3_ENCODER)             += ac3enc_float.o ac3enc.o ac3tab.o \
                                           ac3.o kbdwin.o
-OBJS-$(CONFIG_AC3_FIXED_ENCODER)       += ac3enc_fixed.o ac3enc.o ac3tab.o ac3.o
+OBJS-$(CONFIG_AC3_FIXED_ENCODER)       += ac3enc_fixed.o ac3enc.o ac3tab.o ac3.o kbdwin.o
 OBJS-$(CONFIG_AC3_MF_ENCODER)          += mfenc.o mf_utils.o
 OBJS-$(CONFIG_ACELP_KELVIN_DECODER)    += g729dec.o lsp.o celp_math.o celp_filters.o acelp_filters.o acelp_pitch_delay.o acelp_vectors.o g729postfilter.o
 OBJS-$(CONFIG_AGM_DECODER)             += agm.o
diff --git a/libavcodec/ac3enc.c b/libavcodec/ac3enc.c
index b2e3b2bb4b..9dafe0ef55 100644
--- a/libavcodec/ac3enc.c
+++ b/libavcodec/ac3enc.c
@@ -2047,6 +2047,7 @@  av_cold int ff_ac3_encode_close(AVCodecContext *avctx)
     int blk, ch;
     AC3EncodeContext *s = avctx->priv_data;
 
+    av_freep(&s->mdct_window);
     av_freep(&s->windowed_samples);
     if (s->planar_samples)
     for (ch = 0; ch < s->channels; ch++)
diff --git a/libavcodec/ac3enc.h b/libavcodec/ac3enc.h
index 044564ecb4..ba62891371 100644
--- a/libavcodec/ac3enc.h
+++ b/libavcodec/ac3enc.h
@@ -30,8 +30,6 @@ 
 
 #include <stdint.h>
 
-#include "libavutil/float_dsp.h"
-
 #include "ac3.h"
 #include "ac3dsp.h"
 #include "avcodec.h"
@@ -53,6 +51,7 @@ 
 #define AC3ENC_TYPE_EAC3        2
 
 #if AC3ENC_FLOAT
+#include "libavutil/float_dsp.h"
 #define AC3_NAME(x) ff_ac3_float_ ## x
 #define MAC_COEF(d,a,b) ((d)+=(a)*(b))
 #define COEF_MIN (-16777215.0/16777216.0)
@@ -62,12 +61,13 @@  typedef float SampleType;
 typedef float CoefType;
 typedef float CoefSumType;
 #else
+#include "libavutil/fixed_dsp.h"
 #define AC3_NAME(x) ff_ac3_fixed_ ## x
 #define MAC_COEF(d,a,b) MAC64(d,a,b)
 #define COEF_MIN -16777215
 #define COEF_MAX  16777215
 #define NEW_CPL_COORD_THRESHOLD 503317
-typedef int16_t SampleType;
+typedef int32_t SampleType;
 typedef int32_t CoefType;
 typedef int64_t CoefSumType;
 #endif
@@ -141,7 +141,6 @@  typedef struct AC3Block {
     uint16_t **qmant;                           ///< quantized mantissas
     uint8_t  **cpl_coord_exp;                   ///< coupling coord exponents           (cplcoexp)
     uint8_t  **cpl_coord_mant;                  ///< coupling coord mantissas           (cplcomant)
-    uint8_t  coeff_shift[AC3_MAX_CHANNELS];     ///< fixed-point coefficient shift values
     uint8_t  new_rematrixing_strategy;          ///< send new rematrixing flags in this block
     int      num_rematrixing_bands;             ///< number of rematrixing bands
     uint8_t  rematrixing_flags[4];              ///< rematrixing flags
@@ -165,7 +164,11 @@  typedef struct AC3EncodeContext {
     AVCodecContext *avctx;                  ///< parent AVCodecContext
     PutBitContext pb;                       ///< bitstream writer context
     AudioDSPContext adsp;
+#if AC3ENC_FLOAT
     AVFloatDSPContext *fdsp;
+#else
+    AVFixedDSPContext *fdsp;
+#endif
     MECmpContext mecc;
     AC3DSPContext ac3dsp;                   ///< AC-3 optimized functions
     FFTContext mdct;                        ///< FFT context for MDCT calculation
diff --git a/libavcodec/ac3enc_fixed.c b/libavcodec/ac3enc_fixed.c
index 7818dd8c35..3b302d40df 100644
--- a/libavcodec/ac3enc_fixed.c
+++ b/libavcodec/ac3enc_fixed.c
@@ -26,12 +26,14 @@ 
  * fixed-point AC-3 encoder.
  */
 
-#define FFT_FLOAT 0
 #define AC3ENC_FLOAT 0
+#define FFT_FLOAT 0
+#define FFT_FIXED_32 1
 #include "internal.h"
 #include "audiodsp.h"
 #include "ac3enc.h"
 #include "eac3enc.h"
+#include "kbdwin.h"
 
 #define AC3ENC_TYPE AC3ENC_TYPE_AC3_FIXED
 #include "ac3enc_opts_template.c"
@@ -43,37 +45,6 @@  static const AVClass ac3enc_class = {
     .version    = LIBAVUTIL_VERSION_INT,
 };
 
-/*
- * Normalize the input samples to use the maximum available precision.
- * This assumes signed 16-bit input samples.
- */
-static int normalize_samples(AC3EncodeContext *s)
-{
-    int v = s->ac3dsp.ac3_max_msb_abs_int16(s->windowed_samples, AC3_WINDOW_SIZE);
-    v = 14 - av_log2(v);
-    if (v > 0)
-        s->ac3dsp.ac3_lshift_int16(s->windowed_samples, AC3_WINDOW_SIZE, v);
-    /* +6 to right-shift from 31-bit to 25-bit */
-    return v + 6;
-}
-
-
-/*
- * Scale MDCT coefficients to 25-bit signed fixed-point.
- */
-static void scale_coefficients(AC3EncodeContext *s)
-{
-    int blk, ch;
-
-    for (blk = 0; blk < s->num_blocks; blk++) {
-        AC3Block *block = &s->blocks[blk];
-        for (ch = 1; ch <= s->channels; ch++) {
-            s->ac3dsp.ac3_rshift_int32(block->mdct_coef[ch], AC3_MAX_COEFS,
-                                       block->coeff_shift[ch]);
-        }
-    }
-}
-
 static void sum_square_butterfly(AC3EncodeContext *s, int64_t sum[4],
                                  const int32_t *coef0, const int32_t *coef1,
                                  int len)
@@ -120,7 +91,6 @@  static av_cold void ac3_fixed_mdct_end(AC3EncodeContext *s)
     ff_mdct_end(&s->mdct);
 }
 
-
 /**
  * Initialize MDCT tables.
  *
@@ -129,9 +99,24 @@  static av_cold void ac3_fixed_mdct_end(AC3EncodeContext *s)
  */
 static av_cold int ac3_fixed_mdct_init(AC3EncodeContext *s)
 {
-    int ret = ff_mdct_init(&s->mdct, 9, 0, -1.0);
-    s->mdct_window = ff_ac3_window;
-    return ret;
+    float fwin[AC3_BLOCK_SIZE];
+
+    int32_t *iwin = av_malloc_array(AC3_WINDOW_SIZE, sizeof(*iwin));
+    if (!iwin)
+        return AVERROR(ENOMEM);
+
+    ff_kbd_window_init(fwin, 5.0, AC3_WINDOW_SIZE/2);
+
+    for (int i = 0; i < AC3_WINDOW_SIZE/2; i++)
+        iwin[AC3_WINDOW_SIZE-1-i] = lrintf(fwin[i] * (1 << 22));
+
+    s->mdct_window = iwin;
+
+    s->fdsp = avpriv_alloc_fixed_dsp(s->avctx->flags & AV_CODEC_FLAG_BITEXACT);
+    if (!s->fdsp)
+        return AVERROR(ENOMEM);
+
+    return ff_mdct_init(&s->mdct, 9, 0, -1.0);
 }
 
 
@@ -155,7 +140,7 @@  AVCodec ff_ac3_fixed_encoder = {
     .init            = ac3_fixed_encode_init,
     .encode2         = ff_ac3_fixed_encode_frame,
     .close           = ff_ac3_encode_close,
-    .sample_fmts     = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16P,
+    .sample_fmts     = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S32P,
                                                       AV_SAMPLE_FMT_NONE },
     .priv_class      = &ac3enc_class,
     .caps_internal   = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
diff --git a/libavcodec/ac3enc_float.c b/libavcodec/ac3enc_float.c
index 45bfed34f9..b17b3a2365 100644
--- a/libavcodec/ac3enc_float.c
+++ b/libavcodec/ac3enc_float.c
@@ -97,7 +97,6 @@  static void sum_square_butterfly(AC3EncodeContext *s, float sum[4],
 static av_cold void ac3_float_mdct_end(AC3EncodeContext *s)
 {
     ff_mdct_end(&s->mdct);
-    av_freep(&s->mdct_window);
 }
 
 
diff --git a/libavcodec/ac3enc_template.c b/libavcodec/ac3enc_template.c
index 0fdc95b968..de6eba71d8 100644
--- a/libavcodec/ac3enc_template.c
+++ b/libavcodec/ac3enc_template.c
@@ -91,18 +91,11 @@  static void apply_mdct(AC3EncodeContext *s)
             AC3Block *block = &s->blocks[blk];
             const SampleType *input_samples = &s->planar_samples[ch][blk * AC3_BLOCK_SIZE];
 
-#if AC3ENC_FLOAT
             s->fdsp->vector_fmul(s->windowed_samples, input_samples,
-                                s->mdct_window, AC3_WINDOW_SIZE);
-#else
-            s->ac3dsp.apply_window_int16(s->windowed_samples, input_samples,
-                                         s->mdct_window, AC3_WINDOW_SIZE);
-
-            block->coeff_shift[ch + 1] = normalize_samples(s);
-#endif
+                                 s->mdct_window, AC3_WINDOW_SIZE);
 
-            s->mdct.mdct_calcw(&s->mdct, block->mdct_coef[ch+1],
-                               s->windowed_samples);
+            s->mdct.mdct_calc(&s->mdct, block->mdct_coef[ch+1],
+                              s->windowed_samples);
         }
     }
 }
@@ -390,9 +383,6 @@  int AC3_NAME(encode_frame)(AVCodecContext *avctx, AVPacket *avpkt,
 
     apply_mdct(s);
 
-    if (!AC3ENC_FLOAT)
-        scale_coefficients(s);
-
     clip_coefficients(&s->adsp, s->blocks[0].mdct_coef[1],
                       AC3_MAX_COEFS * s->num_blocks * s->channels);
 
@@ -404,8 +394,9 @@  int AC3_NAME(encode_frame)(AVCodecContext *avctx, AVPacket *avpkt,
 
     compute_rematrixing_strategy(s);
 
-    if (AC3ENC_FLOAT)
-        scale_coefficients(s);
+#if AC3ENC_FLOAT
+    scale_coefficients(s);
+#endif
 
     return ff_ac3_encode_frame_common_end(avctx, avpkt, frame, got_packet_ptr);
 }
diff --git a/libavcodec/version.h b/libavcodec/version.h
index 1420439044..cd871f0fa0 100644
--- a/libavcodec/version.h
+++ b/libavcodec/version.h
@@ -28,7 +28,7 @@ 
 #include "libavutil/version.h"
 
 #define LIBAVCODEC_VERSION_MAJOR  58
-#define LIBAVCODEC_VERSION_MINOR 116
+#define LIBAVCODEC_VERSION_MINOR 117
 #define LIBAVCODEC_VERSION_MICRO 100
 
 #define LIBAVCODEC_VERSION_INT  AV_VERSION_INT(LIBAVCODEC_VERSION_MAJOR, \
diff --git a/tests/fate/ac3.mak b/tests/fate/ac3.mak
index 757cd51cf2..d76e22bade 100644
--- a/tests/fate/ac3.mak
+++ b/tests/fate/ac3.mak
@@ -90,7 +90,7 @@  fate-ac3-fixed-encode: tests/data/asynth-44100-2.wav
 fate-ac3-fixed-encode: SRC = $(TARGET_PATH)/tests/data/asynth-44100-2.wav
 fate-ac3-fixed-encode: CMD = md5 -i $(SRC) -c ac3_fixed -ab 128k -f ac3 -flags +bitexact -af aresample
 fate-ac3-fixed-encode: CMP = oneline
-fate-ac3-fixed-encode: REF = a1d1fc116463b771abf5aef7ed37d7b1
+fate-ac3-fixed-encode: REF = 1f548175e11a95e62ce20e442fcc8d08
 
 FATE_EAC3-$(call ALLYES, EAC3_DEMUXER EAC3_MUXER EAC3_CORE_BSF) += fate-eac3-core-bsf
 fate-eac3-core-bsf: CMD = md5pipe -i $(TARGET_SAMPLES)/eac3/the_great_wall_7.1.eac3 -c:a copy -bsf:a eac3_core -fflags +bitexact -f eac3
diff --git a/tests/fate/ffmpeg.mak b/tests/fate/ffmpeg.mak
index c6d8dc2e5c..4dfb77d250 100644
--- a/tests/fate/ffmpeg.mak
+++ b/tests/fate/ffmpeg.mak
@@ -83,7 +83,7 @@  fate-unknown_layout-pcm: CMD = md5 \
 FATE_FFMPEG-$(call ALLYES, PCM_S16LE_DEMUXER AC3_MUXER PCM_S16LE_DECODER AC3_FIXED_ENCODER) += fate-unknown_layout-ac3
 fate-unknown_layout-ac3: $(AREF)
 fate-unknown_layout-ac3: CMD = md5 -auto_conversion_filters \
-  -guess_layout_max 0 -f s16le -ac 1 -ar 44100 -i $(TARGET_PATH)/$(AREF) \
+  -guess_layout_max 0 -f s32le -ac 1 -ar 44100 -i $(TARGET_PATH)/$(AREF) \
   -f ac3 -flags +bitexact -c ac3_fixed
 
 
diff --git a/tests/ref/fate/unknown_layout-ac3 b/tests/ref/fate/unknown_layout-ac3
index d332efcec4..719a44aacf 100644
--- a/tests/ref/fate/unknown_layout-ac3
+++ b/tests/ref/fate/unknown_layout-ac3
@@ -1 +1 @@ 
-bbb7550d6d93973c10f4ee13c87cf799
+febdb165cfd6cba375aa086195e61213
diff --git a/tests/ref/lavf/rm b/tests/ref/lavf/rm
index 43ea4c7897..fc2a6564a2 100644
--- a/tests/ref/lavf/rm
+++ b/tests/ref/lavf/rm
@@ -1,2 +1,2 @@ 
-e30681d05d6f3d24108d3614600bf116 *tests/data/lavf/lavf.rm
+8dfb8d4556d61d3615e0d0012ffe540c *tests/data/lavf/lavf.rm
 346424 tests/data/lavf/lavf.rm