Add MSA optimized remaining scale row functions

Add MSA optimized remaining scale row functions

R=fbarchard@google.com
BUG=libyuv:634
Performance Gain (vs C vectorized)
ScaleRowDown2_MSA            - ~22.3x
ScaleRowDown2_Any_MSA        - ~19.9x
ScaleRowDown2Linear_MSA      - ~31.2x
ScaleRowDown2Linear_Any_MSA  - ~29.4x
ScaleRowDown2Box_MSA         - ~20.1x
ScaleRowDown2Box_Any_MSA     - ~19.6x
ScaleRowDown4_MSA            - ~11.7x
ScaleRowDown4_Any_MSA        - ~11.2x
ScaleRowDown4Box_MSA         - ~15.1x
ScaleRowDown4Box_Any_MSA     - ~15.1x
ScaleRowDown38_MSA           - ~1x
ScaleRowDown38_Any_MSA       - ~1x
ScaleRowDown38_2_Box_MSA     - ~1.7x
ScaleRowDown38_2_Box_Any_MSA - ~1.7x
ScaleRowDown38_3_Box_MSA     - ~1.7x
ScaleRowDown38_3_Box_Any_MSA - ~1.7x
ScaleAddRow_MSA              - ~1.2x
ScaleAddRow_Any_MSA          - ~1.15x

Performance Gain (vs C non-vectorized)
ScaleRowDown2_MSA            - ~22.4x
ScaleRowDown2_Any_MSA        - ~19.8x
ScaleRowDown2Linear_MSA      - ~31.6x
ScaleRowDown2Linear_Any_MSA  - ~29.4x
ScaleRowDown2Box_MSA         - ~20.1x
ScaleRowDown2Box_Any_MSA     - ~19.6x
ScaleRowDown4_MSA            - ~11.7x
ScaleRowDown4_Any_MSA        - ~11.2x
ScaleRowDown4Box_MSA         - ~15.1x
ScaleRowDown4Box_Any_MSA     - ~15.1x
ScaleRowDown38_MSA           - ~3.2x
ScaleRowDown38_Any_MSA       - ~3.2x
ScaleRowDown38_2_Box_MSA     - ~2.4x
ScaleRowDown38_2_Box_Any_MSA - ~2.3x
ScaleRowDown38_3_Box_MSA     - ~2.9x
ScaleRowDown38_3_Box_Any_MSA - ~2.8x
ScaleAddRow_MSA              - ~8x
ScaleAddRow_Any_MSA          - ~7.46x

Committed: https://chromium.googlesource.com/libyuv/libyuv/+/288bfbefb5f7966b56c7293c727c41aa674e86fb

[manojkumar.bhosale, 2016-12-07 09:10:00]

Description was changed from

==========
Add MSA optimized remaining scale row functions

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

to

==========
Add MSA optimized remaining scale row functions

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

Performance Gain (vs C vectorized)
ScaleRowDown2_MSA            - ~22.3x
ScaleRowDown2_Any_MSA        - ~19.9x
ScaleRowDown2Linear_MSA      - ~31.2x
ScaleRowDown2Linear_Any_MSA  - ~29.4x
ScaleRowDown2Box_MSA         - ~20.1x
ScaleRowDown2Box_Any_MSA     - ~19.6x
ScaleRowDown4_MSA            - ~11.7x
ScaleRowDown4_Any_MSA        - ~11.2x
ScaleRowDown4Box_MSA         - ~15.1x
ScaleRowDown4Box_Any_MSA     - ~15.1x
ScaleRowDown38_MSA           - ~1x
ScaleRowDown38_Any_MSA       - ~1x
ScaleRowDown38_2_Box_MSA     - ~1.7x
ScaleRowDown38_2_Box_Any_MSA - ~1.7x
ScaleRowDown38_3_Box_MSA     - ~1.7x
ScaleRowDown38_3_Box_Any_MSA - ~1.7x
ScaleAddRow_MSA              - ~1.2x
ScaleAddRow_Any_MSA          - ~1.15x

Performance Gain (vs C non-vectorized)
ScaleRowDown2_MSA            - ~22.4x
ScaleRowDown2_Any_MSA        - ~19.8x
ScaleRowDown2Linear_MSA      - ~31.6x
ScaleRowDown2Linear_Any_MSA  - ~29.4x
ScaleRowDown2Box_MSA         - ~20.1x
ScaleRowDown2Box_Any_MSA     - ~19.6x
ScaleRowDown4_MSA            - ~11.7x
ScaleRowDown4_Any_MSA        - ~11.2x
ScaleRowDown4Box_MSA         - ~15.1x
ScaleRowDown4Box_Any_MSA     - ~15.1x
ScaleRowDown38_MSA           - ~3.2x
ScaleRowDown38_Any_MSA       - ~3.2x
ScaleRowDown38_2_Box_MSA     - ~2.4x
ScaleRowDown38_2_Box_Any_MSA - ~2.3x
ScaleRowDown38_3_Box_MSA     - ~2.9x
ScaleRowDown38_3_Box_Any_MSA - ~2.8x
ScaleAddRow_MSA              - ~8x
ScaleAddRow_Any_MSA          - ~7.46x
==========

[manojkumar.bhosale, 2016-12-07 09:10:00]

manojkumar.bhosale@imgtec.com changed reviewers:
+ kjellander@chromium.org

[manojkumar.bhosale, 2016-12-07 09:10:20]

Updated with reviewers, cc list and performance gain numbers

[kjellander_chromium, 2016-12-07 09:21:24]

kjellander@chromium.org changed reviewers:
- kjellander@chromium.org

[fbarchard1, 2016-12-16 18:54:06]

lgtm

https://codereview.chromium.org/2559683002/diff/1/source/scale_msa.cc
File source/scale_msa.cc (right):

https://codereview.chromium.org/2559683002/diff/1/source/scale_msa.cc#newcode234
source/scale_msa.cc:234: vec0 = __msa_hadd_u_h(src0, src0);
is hadd fast?  on some cpu's its a slow instruction and better to avoid by
rearranging with pckev/pckod and vertical adds.

https://codereview.chromium.org/2559683002/diff/1/source/scale_msa.cc#newcode301
source/scale_msa.cc:301: vec0 = __msa_hadd_u_h(src0, src0);
you're doing a lot of loads then a lot of adds, saturating each unit.
Would it be better to do 2 or 4 loads, then 2 or 4 adds, then 2 or 4 more loads,
to keep both load and math units busy at same time?

https://codereview.chromium.org/2559683002/diff/40001/include/libyuv/macros_m...
File include/libyuv/macros_msa.h (right):

https://codereview.chromium.org/2559683002/diff/40001/include/libyuv/macros_m...
include/libyuv/macros_msa.h:195: #define ST_H2(RTYPE, in0, in1, pdst, stride) \
is this is correct style for macros?  try

clang-format -style=file -i include/libyuv/macros_msa.h

[manojkumar.bhosale, 2016-12-20 09:25:56]

https://codereview.chromium.org/2559683002/diff/1/source/scale_msa.cc
File source/scale_msa.cc (right):

https://codereview.chromium.org/2559683002/diff/1/source/scale_msa.cc#newcode234
source/scale_msa.cc:234: vec0 = __msa_hadd_u_h(src0, src0);
On 2016/12/16 18:54:06, fbarchard1 wrote:
> is hadd fast?  on some cpu's its a slow instruction and better to avoid by
> rearranging with pckev/pckod and vertical adds.

pckev/pckod used strategy is not performing better than current version.

https://codereview.chromium.org/2559683002/diff/1/source/scale_msa.cc#newcode301
source/scale_msa.cc:301: vec0 = __msa_hadd_u_h(src0, src0);
On 2016/12/16 18:54:06, fbarchard1 wrote:
> you're doing a lot of loads then a lot of adds, saturating each unit.
> Would it be better to do 2 or 4 loads, then 2 or 4 adds, then 2 or 4 more
loads,
> to keep both load and math units busy at same time?

Tried another version by processing 8 (half), however that adds up more stalls
and hence degrades the overall performance. This version works better
comparatively as we have load latency of around 8-10 cycles, so 8 loads gives
better chance of instr scheduling by compiler.

https://codereview.chromium.org/2559683002/diff/40001/include/libyuv/macros_m...
File include/libyuv/macros_msa.h (right):

https://codereview.chromium.org/2559683002/diff/40001/include/libyuv/macros_m...
include/libyuv/macros_msa.h:195: #define ST_H2(RTYPE, in0, in1, pdst, stride) \
On 2016/12/16 18:54:06, fbarchard1 wrote:
> is this is correct style for macros?  try
> 
> clang-format -style=file -i include/libyuv/macros_msa.h

Already did the clang-format as above. Tried again but i am not seeing any file
updates on clang-format

[fbarchard1, 2016-12-20 18:39:17]

lgtm

https://codereview.chromium.org/2559683002/diff/60001/include/libyuv/macros_m...
File include/libyuv/macros_msa.h (right):

https://codereview.chromium.org/2559683002/diff/60001/include/libyuv/macros_m...
include/libyuv/macros_msa.h:147: #define ST_UH(...) ST_H(v8u16, __VA_ARGS__)
nit (not important) va_args versions dont seem necessary for now.  ST_H could
use v8u16 directly.

[manojkumar.bhosale, 2016-12-21 08:08:26]

Description was changed from

==========
Add MSA optimized remaining scale row functions

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

Performance Gain (vs C vectorized)
ScaleRowDown2_MSA            - ~22.3x
ScaleRowDown2_Any_MSA        - ~19.9x
ScaleRowDown2Linear_MSA      - ~31.2x
ScaleRowDown2Linear_Any_MSA  - ~29.4x
ScaleRowDown2Box_MSA         - ~20.1x
ScaleRowDown2Box_Any_MSA     - ~19.6x
ScaleRowDown4_MSA            - ~11.7x
ScaleRowDown4_Any_MSA        - ~11.2x
ScaleRowDown4Box_MSA         - ~15.1x
ScaleRowDown4Box_Any_MSA     - ~15.1x
ScaleRowDown38_MSA           - ~1x
ScaleRowDown38_Any_MSA       - ~1x
ScaleRowDown38_2_Box_MSA     - ~1.7x
ScaleRowDown38_2_Box_Any_MSA - ~1.7x
ScaleRowDown38_3_Box_MSA     - ~1.7x
ScaleRowDown38_3_Box_Any_MSA - ~1.7x
ScaleAddRow_MSA              - ~1.2x
ScaleAddRow_Any_MSA          - ~1.15x

Performance Gain (vs C non-vectorized)
ScaleRowDown2_MSA            - ~22.4x
ScaleRowDown2_Any_MSA        - ~19.8x
ScaleRowDown2Linear_MSA      - ~31.6x
ScaleRowDown2Linear_Any_MSA  - ~29.4x
ScaleRowDown2Box_MSA         - ~20.1x
ScaleRowDown2Box_Any_MSA     - ~19.6x
ScaleRowDown4_MSA            - ~11.7x
ScaleRowDown4_Any_MSA        - ~11.2x
ScaleRowDown4Box_MSA         - ~15.1x
ScaleRowDown4Box_Any_MSA     - ~15.1x
ScaleRowDown38_MSA           - ~3.2x
ScaleRowDown38_Any_MSA       - ~3.2x
ScaleRowDown38_2_Box_MSA     - ~2.4x
ScaleRowDown38_2_Box_Any_MSA - ~2.3x
ScaleRowDown38_3_Box_MSA     - ~2.9x
ScaleRowDown38_3_Box_Any_MSA - ~2.8x
ScaleAddRow_MSA              - ~8x
ScaleAddRow_Any_MSA          - ~7.46x
==========

to

==========
Add MSA optimized remaining scale row functions

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

Performance Gain (vs C vectorized)
ScaleRowDown2_MSA            - ~22.3x
ScaleRowDown2_Any_MSA        - ~19.9x
ScaleRowDown2Linear_MSA      - ~31.2x
ScaleRowDown2Linear_Any_MSA  - ~29.4x
ScaleRowDown2Box_MSA         - ~20.1x
ScaleRowDown2Box_Any_MSA     - ~19.6x
ScaleRowDown4_MSA            - ~11.7x
ScaleRowDown4_Any_MSA        - ~11.2x
ScaleRowDown4Box_MSA         - ~15.1x
ScaleRowDown4Box_Any_MSA     - ~15.1x
ScaleRowDown38_MSA           - ~1x
ScaleRowDown38_Any_MSA       - ~1x
ScaleRowDown38_2_Box_MSA     - ~1.7x
ScaleRowDown38_2_Box_Any_MSA - ~1.7x
ScaleRowDown38_3_Box_MSA     - ~1.7x
ScaleRowDown38_3_Box_Any_MSA - ~1.7x
ScaleAddRow_MSA              - ~1.2x
ScaleAddRow_Any_MSA          - ~1.15x

Performance Gain (vs C non-vectorized)
ScaleRowDown2_MSA            - ~22.4x
ScaleRowDown2_Any_MSA        - ~19.8x
ScaleRowDown2Linear_MSA      - ~31.6x
ScaleRowDown2Linear_Any_MSA  - ~29.4x
ScaleRowDown2Box_MSA         - ~20.1x
ScaleRowDown2Box_Any_MSA     - ~19.6x
ScaleRowDown4_MSA            - ~11.7x
ScaleRowDown4_Any_MSA        - ~11.2x
ScaleRowDown4Box_MSA         - ~15.1x
ScaleRowDown4Box_Any_MSA     - ~15.1x
ScaleRowDown38_MSA           - ~3.2x
ScaleRowDown38_Any_MSA       - ~3.2x
ScaleRowDown38_2_Box_MSA     - ~2.4x
ScaleRowDown38_2_Box_Any_MSA - ~2.3x
ScaleRowDown38_3_Box_MSA     - ~2.9x
ScaleRowDown38_3_Box_Any_MSA - ~2.8x
ScaleAddRow_MSA              - ~8x
ScaleAddRow_Any_MSA          - ~7.46x

Committed:
https://chromium.googlesource.com/libyuv/libyuv/+/288bfbefb5f7966b56c7293c727...
==========

[manojkumar.bhosale, 2016-12-21 08:08:31]

Committed patchset #4 (id:60001) manually as
288bfbefb5f7966b56c7293c727c41aa674e86fb (presubmit successful).