Exclusion and Difference modes using Sk4f

Exclusion and Difference modes using Sk4f

Before:
   7M	1	15.3ms	15.5ms	15.8ms	17.2ms	4%	▁█▄▁▇▂▁▁▂▁	8888	Xfermode_Exclusion
   7M	1	16.5ms	17.1ms	17.3ms	18.8ms	5%	▁█▃█▃▂▂▃▂▂	8888	Xfermode_Difference

After:
   7M	1	9.06ms	9.34ms	9.42ms	10.4ms	4%	▁▁▅▄█▁▂▁▂▃	8888	Xfermode_Exclusion
   7M	1	10.5ms	10.9ms	11ms	12ms	5%	▃▁▆█▂▁▅▂▁▃	8888	Xfermode_Difference

TBR=mtklein@google.com

Committed: https://skia.googlesource.com/skia/+/92dabe7421aa6ba4149901108d8ab72956f67eeb

[reed2, 2015-04-03 02:11:02]

The CQ bit was checked by reed@chromium.org

[commit-bot: I haz the power, 2015-04-03 02:12:32]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1060493002/20001

[commit-bot: I haz the power, 2015-04-03 02:19:19]

Committed patchset #2 (id:20001) as
https://skia.googlesource.com/skia/+/92dabe7421aa6ba4149901108d8ab72956f67eeb

[mtklein, 2015-04-03 13:23:58]

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

[mtklein, 2015-04-03 13:23:59]

https://codereview.chromium.org/1060493002/diff/20001/src/core/SkXfermode.cpp
File src/core/SkXfermode.cpp (right):

https://codereview.chromium.org/1060493002/diff/20001/src/core/SkXfermode.cpp...
src/core/SkXfermode.cpp:1208: return SkPMFloat(pma.a(), pmc.r(), pmc.g(),
pmc.b());
Boy, this is annoying.  We can make this operation faster (1 instruction) if we
stored alpha in the low position, but then we'd not get the SkPMColor /
SkPMFloat color-order alignment that makes conversion fast.

[mtklein, 2015-04-03 13:35:47]

https://codereview.chromium.org/1060493002/diff/20001/src/core/SkXfermode.cpp
File src/core/SkXfermode.cpp (right):

https://codereview.chromium.org/1060493002/diff/20001/src/core/SkXfermode.cpp...
src/core/SkXfermode.cpp:1208: return SkPMFloat(pma.a(), pmc.r(), pmc.g(),
pmc.b());
On 2015/04/03 13:23:58, mtklein wrote:
> Boy, this is annoying.  We can make this operation faster (1 instruction) if
we
> stored alpha in the low position, but then we'd not get the SkPMColor /
> SkPMFloat color-order alignment that makes conversion fast.

Well, maybe it's not so annoying.  On x86 this can be done in 2 shuffle
instructions, and it looks like Clang's doing that just fine.  Probably not
worth fussing over.

[reed1, 2015-04-03 15:21:21]

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

[reed1, 2015-04-03 15:21:22]

I wonder, would we do better with some math? In a loop, those two literals could
be in reused registers. That is 3 ops. Is that worse than what we get today?

result = alpha4 * (1, 0, 0, 0) + color4 * (0, 1, 1, 1)

Either way, likely we'll want to provide a std helper to do this sort of thing.

[mtklein, 2015-04-03 15:28:22]

On 2015/04/03 15:21:22, reed1 wrote:
> I wonder, would we do better with some math? In a loop, those two literals
could
> be in reused registers. That is 3 ops. Is that worse than what we get today?
> 
> result = alpha4 * (1, 0, 0, 0) + color4 * (0, 1, 1, 1)
> 
> Either way, likely we'll want to provide a std helper to do this sort of
thing.

Yeah, I actually tried that offline, and it is worse, both in op counts and
time.  So is an _mm_xor_ps() based approach.