Add MSA optimized ARGB/ABGR/BGRA/RGBA To Y/UV row functions

Add MSA optimized ARGB/ABGR/BGRA/RGBA To Y/UV row functions

R=fbarchard@google.com
BUG=libyuv:634
Performance Gain (vs C auto-vectorized)
ARGBToYJRow_MSA       - ~3.2x
ARGBToYJRow_Any_MSA   - ~2.7x
BGRAToYRow_MSA        - ~3.2x
BGRAToYRow_Any_MSA    - ~2.7x
ABGRToYRow_MSA        - ~3.2x
ABGRToYRow_Any_MSA    - ~2.6x
RGBAToYRow_MSA        - ~3.1x
RGBAToYRow_Any_MSA    - ~2.7x
ARGBToUVJRow_MSA      - ~5.5x
ARGBToUVJRow_Any_MSA  - ~4.5x
BGRAToUVRow_MSA       - ~2.1x
BGRAToUVRow_Any_MSA   - ~2.0x
ABGRToUVRow_MSA       - ~2.1x
ABGRToUVRow_Any_MSA   - ~1.9x
RGBAToUVRow_MSA       - ~2.2x
RGBAToUVRow_Any_MSA   - ~1.9x


Performance Gain (vs C non-vectorized)
ARGBToYJRow_MSA       - ~10.9x
ARGBToYJRow_Any_MSA   -  ~9.2x
BGRAToYRow_MSA        - ~10.9x
BGRAToYRow_Any_MSA    -  ~9.3x
ABGRToYRow_MSA        - ~11.0x
ABGRToYRow_Any_MSA    -  ~9.3x
RGBAToYRow_MSA        - ~10.9x
RGBAToYRow_Any_MSA    -  ~9.1x
ARGBToUVJRow_MSA      - ~12.4x
ARGBToUVJRow_Any_MSA  - ~10.5x
BGRAToUVRow_MSA       -  ~4.7x
BGRAToUVRow_Any_MSA   -  ~4.4x
ABGRToUVRow_MSA       -  ~4.7x
ABGRToUVRow_Any_MSA   -  ~4.5x
RGBAToUVRow_MSA       -  ~4.8x
RGBAToUVRow_Any_MSA   -  ~4.4x

Committed: https://chromium.googlesource.com/libyuv/libyuv/+/54ce8f23d6d82d671ec05229812842548faf14ef

[manojkumar.bhosale, 2017-01-20 10:27:01]

Description was changed from

==========
Add MSA optimized ARGB/ABGR/BGRA/RGBA To Y/UV row functions

R=fbarchard@google.com
BUG=libyuv:634
==========

to

==========
Add MSA optimized ARGB/ABGR/BGRA/RGBA To Y/UV row functions

R=fbarchard@google.com
BUG=libyuv:634

Performance Gain (vs C auto-vectorized)
ARGBToYJRow_MSA       - ~3.2x
ARGBToYJRow_Any_MSA   - ~2.7x
BGRAToYRow_MSA        - ~3.2x
BGRAToYRow_Any_MSA    - ~2.7x
ABGRToYRow_MSA        - ~3.2x
ABGRToYRow_Any_MSA    - ~2.6x
RGBAToYRow_MSA        - ~3.1x
RGBAToYRow_Any_MSA    - ~2.7x
ARGBToUVJRow_MSA      - ~5.5x
ARGBToUVJRow_Any_MSA  - ~4.5x
BGRAToUVRow_MSA       - ~1.0x
BGRAToUVRow_Any_MSA   - ~1.0x
ABGRToUVRow_MSA       - ~1.0x
ABGRToUVRow_Any_MSA   - ~1.0x
RGBAToUVRow_MSA       - ~1.0x
RGBAToUVRow_Any_MSA   - ~1.0x

Performance Gain (vs C non-vectorized)
ARGBToYJRow_MSA       - ~10.9x
ARGBToYJRow_Any_MSA   -  ~9.2x
BGRAToYRow_MSA        - ~10.9x
BGRAToYRow_Any_MSA    -  ~9.3x
ABGRToYRow_MSA        - ~11.0x
ABGRToYRow_Any_MSA    -  ~9.3x
RGBAToYRow_MSA        - ~10.9x
RGBAToYRow_Any_MSA    -  ~9.1x
ARGBToUVJRow_MSA      - ~12.4x
ARGBToUVJRow_Any_MSA  - ~10.5x
BGRAToUVRow_MSA       -  ~2.3x
BGRAToUVRow_Any_MSA   -  ~2.2x
ABGRToUVRow_MSA       -  ~2.4x
ABGRToUVRow_Any_MSA   -  ~2.3x
RGBAToUVRow_MSA       -  ~2.4x
RGBAToUVRow_Any_MSA   -  ~2.3x
==========

[manojkumar.bhosale, 2017-01-20 10:27:12]

Updated with reviewers, cc list and performance gain numbers

[fbarchard1, 2017-01-24 00:15:47]

https://codereview.chromium.org/2641153003/diff/1/source/row_msa.cc
File source/row_msa.cc (right):

https://codereview.chromium.org/2641153003/diff/1/source/row_msa.cc#newcode146
source/row_msa.cc:146: // Laods current and next row of ARGB input and averages
it to calculate U and V
Loads. Not laods

https://codereview.chromium.org/2641153003/diff/1/source/row_msa.cc#newcode2484
source/row_msa.cc:2484: for (x = 0; x < width; x += 32) {
can you make this function similar to others, doing 16 at a time using READ_ARGB
and ARGB_TO_UVJ macros?

[manojkumar.bhosale, 2017-01-25 08:35:42]

https://codereview.chromium.org/2641153003/diff/1/source/row_msa.cc
File source/row_msa.cc (right):

https://codereview.chromium.org/2641153003/diff/1/source/row_msa.cc#newcode146
source/row_msa.cc:146: // Laods current and next row of ARGB input and averages
it to calculate U and V
On 2017/01/24 00:15:47, fbarchard1 wrote:
> Loads. Not laods

Done. Sorry for typo.

https://codereview.chromium.org/2641153003/diff/1/source/row_msa.cc#newcode2484
source/row_msa.cc:2484: for (x = 0; x < width; x += 32) {
On 2017/01/24 00:15:47, fbarchard1 wrote:
> can you make this function similar to others, doing 16 at a time using
READ_ARGB
> and ARGB_TO_UVJ macros?

Actually sending a patchset to fix an issue in other 3 functions where loop
increment should be 32 instead of 16. Not creating macro in this function as it
is not reusable in other functions.

[manojkumar.bhosale, 2017-01-25 08:48:07]

Description was changed from

==========
Add MSA optimized ARGB/ABGR/BGRA/RGBA To Y/UV row functions

R=fbarchard@google.com
BUG=libyuv:634

Performance Gain (vs C auto-vectorized)
ARGBToYJRow_MSA       - ~3.2x
ARGBToYJRow_Any_MSA   - ~2.7x
BGRAToYRow_MSA        - ~3.2x
BGRAToYRow_Any_MSA    - ~2.7x
ABGRToYRow_MSA        - ~3.2x
ABGRToYRow_Any_MSA    - ~2.6x
RGBAToYRow_MSA        - ~3.1x
RGBAToYRow_Any_MSA    - ~2.7x
ARGBToUVJRow_MSA      - ~5.5x
ARGBToUVJRow_Any_MSA  - ~4.5x
BGRAToUVRow_MSA       - ~1.0x
BGRAToUVRow_Any_MSA   - ~1.0x
ABGRToUVRow_MSA       - ~1.0x
ABGRToUVRow_Any_MSA   - ~1.0x
RGBAToUVRow_MSA       - ~1.0x
RGBAToUVRow_Any_MSA   - ~1.0x

Performance Gain (vs C non-vectorized)
ARGBToYJRow_MSA       - ~10.9x
ARGBToYJRow_Any_MSA   -  ~9.2x
BGRAToYRow_MSA        - ~10.9x
BGRAToYRow_Any_MSA    -  ~9.3x
ABGRToYRow_MSA        - ~11.0x
ABGRToYRow_Any_MSA    -  ~9.3x
RGBAToYRow_MSA        - ~10.9x
RGBAToYRow_Any_MSA    -  ~9.1x
ARGBToUVJRow_MSA      - ~12.4x
ARGBToUVJRow_Any_MSA  - ~10.5x
BGRAToUVRow_MSA       -  ~2.3x
BGRAToUVRow_Any_MSA   -  ~2.2x
ABGRToUVRow_MSA       -  ~2.4x
ABGRToUVRow_Any_MSA   -  ~2.3x
RGBAToUVRow_MSA       -  ~2.4x
RGBAToUVRow_Any_MSA   -  ~2.3x
==========

to

==========
Add MSA optimized ARGB/ABGR/BGRA/RGBA To Y/UV row functions

R=fbarchard@google.com
BUG=libyuv:634

Performance Gain (vs C auto-vectorized)
ARGBToYJRow_MSA       - ~3.2x
ARGBToYJRow_Any_MSA   - ~2.7x
BGRAToYRow_MSA        - ~3.2x
BGRAToYRow_Any_MSA    - ~2.7x
ABGRToYRow_MSA        - ~3.2x
ABGRToYRow_Any_MSA    - ~2.6x
RGBAToYRow_MSA        - ~3.1x
RGBAToYRow_Any_MSA    - ~2.7x
ARGBToUVJRow_MSA      - ~5.5x
ARGBToUVJRow_Any_MSA  - ~4.5x
BGRAToUVRow_MSA       - ~2.1x
BGRAToUVRow_Any_MSA   - ~2.0x
ABGRToUVRow_MSA       - ~2.1x
ABGRToUVRow_Any_MSA   - ~1.9x
RGBAToUVRow_MSA       - ~2.2x
RGBAToUVRow_Any_MSA   - ~1.9x


Performance Gain (vs C non-vectorized)
ARGBToYJRow_MSA       - ~10.9x
ARGBToYJRow_Any_MSA   -  ~9.2x
BGRAToYRow_MSA        - ~10.9x
BGRAToYRow_Any_MSA    -  ~9.3x
ABGRToYRow_MSA        - ~11.0x
ABGRToYRow_Any_MSA    -  ~9.3x
RGBAToYRow_MSA        - ~10.9x
RGBAToYRow_Any_MSA    -  ~9.1x
ARGBToUVJRow_MSA      - ~12.4x
ARGBToUVJRow_Any_MSA  - ~10.5x
BGRAToUVRow_MSA       -  ~4.7x
BGRAToUVRow_Any_MSA   -  ~4.4x
ABGRToUVRow_MSA       -  ~4.7x
ABGRToUVRow_Any_MSA   -  ~4.5x
RGBAToUVRow_MSA       -  ~4.8x
RGBAToUVRow_Any_MSA   -  ~4.4x
==========

[manojkumar.bhosale, 2017-01-25 08:49:02]

Updated performance gains of following functions
BGRAToUVRow_MSA    
BGRAToUVRow_Any_MSA
ABGRToUVRow_MSA    
ABGRToUVRow_Any_MSA
RGBAToUVRow_MSA    
RGBAToUVRow_Any_MSA

[fbarchard1, 2017-01-26 22:25:53]

https://codereview.chromium.org/2641153003/diff/20001/source/row_msa.cc
File source/row_msa.cc (right):

https://codereview.chromium.org/2641153003/diff/20001/source/row_msa.cc#newco...
source/row_msa.cc:2485: src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
can use READ_ARGB(s, t, vec0, vec1, vec2, vec3); here yes?

shouldnt this function be same as ABGRToUVRow_MSA with different constants?

Note that I plan to do ABGRToH420 which would need an 'H' variation of this.

[manojkumar.bhosale, 2017-01-30 10:54:45]

https://codereview.chromium.org/2641153003/diff/20001/source/row_msa.cc
File source/row_msa.cc (right):

https://codereview.chromium.org/2641153003/diff/20001/source/row_msa.cc#newco...
source/row_msa.cc:2485: src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
On 2017/01/26 22:25:53, fbarchard1 wrote:
> can use READ_ARGB(s, t, vec0, vec1, vec2, vec3); here yes?
> 
> shouldnt this function be same as ABGRToUVRow_MSA with different constants?
> 
> Note that I plan to do ABGRToH420 which would need an 'H' variation of this.

AFAIK, averaging of pixels in ARGBToUVJRow_C is different from ABGRToUVRow_C. 

In ARGBToUVJRow_C, current row pixels are rounded averaged with next row pixels
and then consecutive pixels get averaged(rounded). These averaged pixels are
then used for UV calculation. 

Whereas in ABGRToUVRow_C, all 4 pixels (2 from current & next row) gets added
first, averaged(non-rounded) and used for UV calculation. Thus we cannot use
READ_ARGB macro in this code.

[fbarchard1, 2017-02-01 01:58:46]

lgtm

https://codereview.chromium.org/2641153003/diff/20001/source/row_msa.cc
File source/row_msa.cc (right):

https://codereview.chromium.org/2641153003/diff/20001/source/row_msa.cc#newco...
source/row_msa.cc:2485: src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
Looks like this has been done 4 different ways.
1. average of 4 values with rounding.  This is basically the correct version.
2. sum of 4 values, then using matrix that contains values that are 4x smaller. 
This would be equivalent to averaging and perhaps better because rounding error
can be avoided, except the coefficients for bt.601 are small and odd, and
dividing by 4 causes more error.  This is not the case for bt.709 or jpeg
coefficients.
3. average of 4 values without rounding.  On Intel and C, rounding is not free,
so this is a little faster.  But less correct and no more efficient on Arm.
4. average of rows followed by average of columns.  Efficient but wrong.  This
is used in bt.601 for efficiency.  on jpeg proper rounding was initially
implemented (average of 4 pixels), but to save code, 420 is implemented using
the same code as 422, pointing both rows to the same content for 422 on intel. 
The C code mimics the SSSE3 code doing byte averages of rows and columes and the
unittest is able to check the results exactly match.
The C code for I420 does not attempt to mimic either the intel or arm code and
the unittest checks for error.

Short term it would be best if the C reference code matches the assembly, which
helps document how the assembly works.

Historically C code performance mattered for CPUs that did not have much
optimized SIMD.  But now that Arm, Mips and Intel are all optimized for, C code
performance is secondary to correctness, and in general integer math should do
rounding.

If you keep the subsampling different, it may be worth a comment, and its likely
to change in the future.  ARGBToI420 should be using common code, like
I420ToARGB, passing a matrix parameter.

[manojkumar.bhosale, 2017-02-01 04:48:41]

https://codereview.chromium.org/2641153003/diff/20001/source/row_msa.cc
File source/row_msa.cc (right):

https://codereview.chromium.org/2641153003/diff/20001/source/row_msa.cc#newco...
source/row_msa.cc:2485: src0 = (v16u8)__msa_ld_b((v16i8*)s, 0);
On 2017/02/01 01:58:46, fbarchard1 wrote:
> Looks like this has been done 4 different ways.
> 1. average of 4 values with rounding.  This is basically the correct version.
> 2. sum of 4 values, then using matrix that contains values that are 4x
smaller. 
> This would be equivalent to averaging and perhaps better because rounding
error
> can be avoided, except the coefficients for bt.601 are small and odd, and
> dividing by 4 causes more error.  This is not the case for bt.709 or jpeg
> coefficients.
> 3. average of 4 values without rounding.  On Intel and C, rounding is not
free,
> so this is a little faster.  But less correct and no more efficient on Arm.
> 4. average of rows followed by average of columns.  Efficient but wrong.  This
> is used in bt.601 for efficiency.  on jpeg proper rounding was initially
> implemented (average of 4 pixels), but to save code, 420 is implemented using
> the same code as 422, pointing both rows to the same content for 422 on intel.

> The C code mimics the SSSE3 code doing byte averages of rows and columes and
the
> unittest is able to check the results exactly match.
> The C code for I420 does not attempt to mimic either the intel or arm code and
> the unittest checks for error.
> 
> Short term it would be best if the C reference code matches the assembly,
which
> helps document how the assembly works.
> 
> Historically C code performance mattered for CPUs that did not have much
> optimized SIMD.  But now that Arm, Mips and Intel are all optimized for, C
code
> performance is secondary to correctness, and in general integer math should do
> rounding.
> 
> If you keep the subsampling different, it may be worth a comment, and its
likely
> to change in the future.  ARGBToI420 should be using common code, like
> I420ToARGB, passing a matrix parameter.

Acknowledged.

[manojkumar.bhosale, 2017-02-01 04:59:02]

Description was changed from

==========
Add MSA optimized ARGB/ABGR/BGRA/RGBA To Y/UV row functions

R=fbarchard@google.com
BUG=libyuv:634

Performance Gain (vs C auto-vectorized)
ARGBToYJRow_MSA       - ~3.2x
ARGBToYJRow_Any_MSA   - ~2.7x
BGRAToYRow_MSA        - ~3.2x
BGRAToYRow_Any_MSA    - ~2.7x
ABGRToYRow_MSA        - ~3.2x
ABGRToYRow_Any_MSA    - ~2.6x
RGBAToYRow_MSA        - ~3.1x
RGBAToYRow_Any_MSA    - ~2.7x
ARGBToUVJRow_MSA      - ~5.5x
ARGBToUVJRow_Any_MSA  - ~4.5x
BGRAToUVRow_MSA       - ~2.1x
BGRAToUVRow_Any_MSA   - ~2.0x
ABGRToUVRow_MSA       - ~2.1x
ABGRToUVRow_Any_MSA   - ~1.9x
RGBAToUVRow_MSA       - ~2.2x
RGBAToUVRow_Any_MSA   - ~1.9x


Performance Gain (vs C non-vectorized)
ARGBToYJRow_MSA       - ~10.9x
ARGBToYJRow_Any_MSA   -  ~9.2x
BGRAToYRow_MSA        - ~10.9x
BGRAToYRow_Any_MSA    -  ~9.3x
ABGRToYRow_MSA        - ~11.0x
ABGRToYRow_Any_MSA    -  ~9.3x
RGBAToYRow_MSA        - ~10.9x
RGBAToYRow_Any_MSA    -  ~9.1x
ARGBToUVJRow_MSA      - ~12.4x
ARGBToUVJRow_Any_MSA  - ~10.5x
BGRAToUVRow_MSA       -  ~4.7x
BGRAToUVRow_Any_MSA   -  ~4.4x
ABGRToUVRow_MSA       -  ~4.7x
ABGRToUVRow_Any_MSA   -  ~4.5x
RGBAToUVRow_MSA       -  ~4.8x
RGBAToUVRow_Any_MSA   -  ~4.4x
==========

to

==========
Add MSA optimized ARGB/ABGR/BGRA/RGBA To Y/UV row functions

R=fbarchard@google.com
BUG=libyuv:634

Performance Gain (vs C auto-vectorized)
ARGBToYJRow_MSA       - ~3.2x
ARGBToYJRow_Any_MSA   - ~2.7x
BGRAToYRow_MSA        - ~3.2x
BGRAToYRow_Any_MSA    - ~2.7x
ABGRToYRow_MSA        - ~3.2x
ABGRToYRow_Any_MSA    - ~2.6x
RGBAToYRow_MSA        - ~3.1x
RGBAToYRow_Any_MSA    - ~2.7x
ARGBToUVJRow_MSA      - ~5.5x
ARGBToUVJRow_Any_MSA  - ~4.5x
BGRAToUVRow_MSA       - ~2.1x
BGRAToUVRow_Any_MSA   - ~2.0x
ABGRToUVRow_MSA       - ~2.1x
ABGRToUVRow_Any_MSA   - ~1.9x
RGBAToUVRow_MSA       - ~2.2x
RGBAToUVRow_Any_MSA   - ~1.9x


Performance Gain (vs C non-vectorized)
ARGBToYJRow_MSA       - ~10.9x
ARGBToYJRow_Any_MSA   -  ~9.2x
BGRAToYRow_MSA        - ~10.9x
BGRAToYRow_Any_MSA    -  ~9.3x
ABGRToYRow_MSA        - ~11.0x
ABGRToYRow_Any_MSA    -  ~9.3x
RGBAToYRow_MSA        - ~10.9x
RGBAToYRow_Any_MSA    -  ~9.1x
ARGBToUVJRow_MSA      - ~12.4x
ARGBToUVJRow_Any_MSA  - ~10.5x
BGRAToUVRow_MSA       -  ~4.7x
BGRAToUVRow_Any_MSA   -  ~4.4x
ABGRToUVRow_MSA       -  ~4.7x
ABGRToUVRow_Any_MSA   -  ~4.5x
RGBAToUVRow_MSA       -  ~4.8x
RGBAToUVRow_Any_MSA   -  ~4.4x

Committed:
https://chromium.googlesource.com/libyuv/libyuv/+/54ce8f23d6d82d671ec05229812...
==========

[manojkumar.bhosale, 2017-02-01 04:59:06]

Committed patchset #2 (id:20001) manually as
54ce8f23d6d82d671ec05229812842548faf14ef (presubmit successful).