ARGBExtractAlpha 16 pixels at a time for ARM

ARGBExtractAlpha 16 pixels at a time for ARM

arm64   8     TestARGBExtractAlpha (10019 ms) <-original 64 bit code
arm64   8 x2  TestARGBExtractAlpha (7639 ms)
arm64   16    TestARGBExtractAlpha (7369 ms) <- new 64 bit code
thumb32 8     TestARGBExtractAlpha (9505 ms) <- original 32 bit code
thumb32 8 x2  TestARGBExtractAlpha (7400 ms)
thumb32 8 x2i TestARGBExtractAlpha (7266 ms) <- new 32 bit code
arm32   8     TestARGBExtractAlpha (10002 ms)

BUG=libyuv:572
TESTED=local test on nexus 9
R=harryjin@google.com, wangcheng@google.com

Committed: https://chromium.googlesource.com/libyuv/libyuv/+/6546096269c3f452effde05cb7f449428fa658d0

[fbarchard1, 2016-06-02 02:35:11]

performance varies by cpu but nexus 9 shows a substantial win
was 7278ms
now 6073ms

[fbarchard1, 2016-06-06 21:12:08]

Testing 64 vs 32 a bit more on Nexus 9
arm64   8     TestARGBExtractAlpha (10019 ms) <-original 64 bit code
arm64   8 x2  TestARGBExtractAlpha (7639 ms)
arm64   16    TestARGBExtractAlpha (7369 ms) <- new 64 bit code
thumb32 8     TestARGBExtractAlpha (9505 ms) <- original 32 bit code
thumb32 8 x2  TestARGBExtractAlpha (7400 ms)
thumb32 8 x2i TestARGBExtractAlpha (7266 ms) <- new 32 bit code
arm32   8     TestARGBExtractAlpha (10002 ms)

[fbarchard1, 2016-06-06 21:13:29]

Description was changed from

==========
16 byte at a time extract alpha

BUG=572
TESTED=local test on nexus 9
==========

to

==========
ARGBExtractAlpha 16 pixels at a time for ARM

BUG=libyuv:572
TESTED=local test on nexus 9
==========

[fbarchard1, 2016-06-06 21:15:03]

fbarchard@google.com changed reviewers:
+ nessi@chromium.org

[fbarchard1, 2016-06-06 21:15:23]

Description was changed from

==========
ARGBExtractAlpha 16 pixels at a time for ARM

BUG=libyuv:572
TESTED=local test on nexus 9
==========

to

==========
ARGBExtractAlpha 16 pixels at a time for ARM

arm64   8     TestARGBExtractAlpha (10019 ms) <-original 64 bit code
arm64   8 x2  TestARGBExtractAlpha (7639 ms)
arm64   16    TestARGBExtractAlpha (7369 ms) <- new 64 bit code
thumb32 8     TestARGBExtractAlpha (9505 ms) <- original 32 bit code
thumb32 8 x2  TestARGBExtractAlpha (7400 ms)
thumb32 8 x2i TestARGBExtractAlpha (7266 ms) <- new 32 bit code
arm32   8     TestARGBExtractAlpha (10002 ms)

BUG=libyuv:572
TESTED=local test on nexus 9
==========

[fbarchard1, 2016-06-07 02:44:42]

fbarchard@google.com changed reviewers:
+ harryjin@google.com

[fbarchard1, 2016-06-07 02:44:43]

try bots pass

[harryjin, 2016-06-07 02:52:29]

lgtm

https://codereview.chromium.org/2035573002/diff/80001/libyuv.gyp
File libyuv.gyp (right):

https://codereview.chromium.org/2035573002/diff/80001/libyuv.gyp#newcode78
libyuv.gyp:78: # '-marm',  # arm32 not thumb
arm?

[fbarchard1, 2016-06-07 03:07:40]

https://codereview.chromium.org/2035573002/diff/80001/libyuv.gyp
File libyuv.gyp (right):

https://codereview.chromium.org/2035573002/diff/80001/libyuv.gyp#newcode78
libyuv.gyp:78: # '-marm',  # arm32 not thumb
On 2016/06/07 02:52:29, harryjin wrote:
> arm?

arm code is 32 bit instructions
thumb code is 16 bit instructions for simple instructions, 32 bit for rest.

normally we compile 32 bit with -mthumb
this shows how to disable -mthumb and use -marm instead.

[harryjin, 2016-06-07 03:10:27]

lgtm

[wangcheng, 2016-06-07 16:32:18]

wangcheng@google.com changed reviewers:
+ wangcheng@google.com

[wangcheng, 2016-06-07 16:32:19]

lgtm

[fbarchard1, 2016-06-07 17:44:31]

Description was changed from

==========
ARGBExtractAlpha 16 pixels at a time for ARM

arm64   8     TestARGBExtractAlpha (10019 ms) <-original 64 bit code
arm64   8 x2  TestARGBExtractAlpha (7639 ms)
arm64   16    TestARGBExtractAlpha (7369 ms) <- new 64 bit code
thumb32 8     TestARGBExtractAlpha (9505 ms) <- original 32 bit code
thumb32 8 x2  TestARGBExtractAlpha (7400 ms)
thumb32 8 x2i TestARGBExtractAlpha (7266 ms) <- new 32 bit code
arm32   8     TestARGBExtractAlpha (10002 ms)

BUG=libyuv:572
TESTED=local test on nexus 9
==========

to

==========
ARGBExtractAlpha 16 pixels at a time for ARM

arm64   8     TestARGBExtractAlpha (10019 ms) <-original 64 bit code
arm64   8 x2  TestARGBExtractAlpha (7639 ms)
arm64   16    TestARGBExtractAlpha (7369 ms) <- new 64 bit code
thumb32 8     TestARGBExtractAlpha (9505 ms) <- original 32 bit code
thumb32 8 x2  TestARGBExtractAlpha (7400 ms)
thumb32 8 x2i TestARGBExtractAlpha (7266 ms) <- new 32 bit code
arm32   8     TestARGBExtractAlpha (10002 ms)

BUG=libyuv:572
TESTED=local test on nexus 9
R=harryjin@google.com, wangcheng@google.com

Committed:
https://chromium.googlesource.com/libyuv/libyuv/+/6546096269c3f452effde05cb7f...
==========

[fbarchard1, 2016-06-07 17:44:33]

Committed patchset #6 (id:100001) manually as
6546096269c3f452effde05cb7f449428fa658d0 (presubmit successful).

[fbarchard1, 2016-06-07 18:17:02]

FYI Tried putting subs at different points in loop
before load 7426
after load  7400
after store 7457