use new shuffle to speed up affine matrix mappts

use new shuffle to speed up affine matrix mappts

sse: 25 -> 18
neon: 95 -> 86

BUG=skia:

Committed: https://skia.googlesource.com/skia/+/e70afc9f48d00828ee6b707899a8ff542b0e8b98

[mtklein, 2015-09-10 18:57:10]

mtklein@google.com changed reviewers:
+ mtklein@google.com, reed@google.com

[mtklein, 2015-09-10 18:57:11]

FYI, going to circle back on this later. 

 Shuffling register -> register is ~25% faster with SSE.

But it's a small slowdown with NEON, where it's actually faster to reload a new
register fresh from memory (I even looked at the assembly... it's pulling from
memory -> stack, then stack -> neon register).

[reed1, 2015-09-10 18:59:31]

Does this change actually slow us down on NEON? Do we want to wait until we can
find a way to conditionalize this to only affect SSE, or is the slowdown so
small we don't care?

[mtklein, 2015-09-10 19:00:52]

Well, that's silly, the latest patch is now actually a speedup on NEON too.

I rearranged the math so that one + could happen earlier than we'd written it,
which I think shortened the critical path.

[mtklein, 2015-09-10 19:01:52]

On 2015/09/10 19:00:52, mtklein wrote:
> Well, that's silly, the latest patch is now actually a speedup on NEON too.
> 
> I rearranged the math so that one + could happen earlier than we'd written it,
> which I think shortened the critical path.

It's actually a small additional speedup for SSE too.

[reed1, 2015-09-10 19:02:18]

LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM
LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM
LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM LGTM

[mtklein, 2015-09-10 21:23:08]

The CQ bit was checked by mtklein@google.com to run a CQ dry run

[commit-bot: I haz the power, 2015-09-10 21:23:24]

Dry run: CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1333983002/40001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1333983002/40001

[commit-bot: I haz the power, 2015-09-10 21:30:49]

The CQ bit was unchecked by commit-bot@chromium.org

[commit-bot: I haz the power, 2015-09-10 21:30:49]

Dry run: This issue passed the CQ dry run.

[mtklein, 2015-09-10 21:31:39]

The CQ bit was checked by mtklein@google.com

[commit-bot: I haz the power, 2015-09-10 21:31:50]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1333983002/40001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1333983002/40001

[commit-bot: I haz the power, 2015-09-10 21:32:35]

Committed patchset #3 (id:40001) as
https://skia.googlesource.com/skia/+/e70afc9f48d00828ee6b707899a8ff542b0e8b98

[mtklein, 2015-09-10 23:30:28]

On 2015/09/10 21:32:35, commit-bot: I haz the power wrote:
> Committed patchset #3 (id:40001) as
> https://skia.googlesource.com/skia/+/e70afc9f48d00828ee6b707899a8ff542b0e8b98

Uh, so weirdly... this also made the Nexus 9 _5x_ faster and made the Ubuntu GCC
bots ~50% slower.

Crazy stuff.  Clearly I will need to follow up.

[mtklein, 2015-09-10 23:38:22]

A revert of this CL (patchset #3 id:40001) has been created in
https://codereview.chromium.org/1335003002/ by mtklein@google.com.

The reason for reverting is: Unexpected perf impact, and a whole bunch of new
images in gold (mostly invisibly different)..