NEON Optimized RGBA->PMColor sampling in SkSwizzler

NEON Optimized RGBA->PMColor sampling in SkSwizzler

Nexus 6P Swizzle Runtime
BGRA SampleSize=2 0.16x 
RGBA SampleSize=2 0.18x
BGRA SampleSize=4 0.19x
RGBA SampleSize=4 0.21x
BGRA SampleSize=8 0.20x
RGBA SampleSize=8 0.23x

Nexus 6P RGBA PNG Decode Runtime
SampleSize=2 0.98x
SampleSize=4 0.99x
SampleSize=8 0.99x

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1680743005
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

[msarett, 2016-02-08 23:56:58]

Description was changed from

==========
Optimized RGBA->PMColor sampling in SkSwizzler

BUG=skia:
==========

to

==========
Optimized RGBA->PMColor sampling in SkSwizzler

BUG=skia:
GOLD_TRYBOT_URL=
https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&is...
==========

[msarett, 2016-02-09 00:05:57]

msarett@google.com changed reviewers:
+ scroggo@google.com

[msarett, 2016-02-09 00:05:57]

I haven't written any implementations yet.  Just want to make sure that we have
a good way to hook in first (and that we feel that the added complexity is worth
it).

I think this truly is the last swizzler optimization CL for now.  I think
sampling is particularly interesting for RGBA PNGs since that is likely what we
will see on the web.  All of the other formats are prominent in Android assets,
but I doubt that we often perform sampled decodes of assets.

And powers of 2 are particularly interesting because they are easier to optimize
and most often used.

While I think these cases may be the only fast sampled swizzles, this hopefully
sets us up to add more if we ever need to.

There is still swizzler work that could be done to benefit BGRX BMPs, 1-bit,
2-bit, 4-bit, 16-bit PNGs, BMPs, WBMPs, and 565 everything, but I think it's
fine to hold off until somebody cares.

[scroggo, 2016-02-09 13:06:10]

1gtm

https://codereview.chromium.org/1680743005/diff/1/src/codec/SkSwizzler.cpp
File src/codec/SkSwizzler.cpp (right):

https://codereview.chromium.org/1680743005/diff/1/src/codec/SkSwizzler.cpp#ne...
src/codec/SkSwizzler.cpp:968: if (1 == fSampleX && fFastProcs.fSampleSize1) {
Why not:

switch (fSampleX) {
  case 1:
    fActualProc = fFastProcs.fSampleSize;
    break;
  ...
}

if (!fActualProc) {
  fActualProc = fSlowProc;
}

(A little bit cleaner to me, but maybe a matter of preference.)

[msarett, 2016-02-09 17:51:43]

Description was changed from

==========
Optimized RGBA->PMColor sampling in SkSwizzler

BUG=skia:
GOLD_TRYBOT_URL=
https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&is...
==========

to

==========
Optimized RGBA->PMColor sampling in SkSwizzler

BUG=skia:
GOLD_TRYBOT_URL=
https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&is...
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

[msarett, 2016-02-09 18:01:45]

Patchset #4 (id:60001) has been deleted

[msarett, 2016-02-10 00:30:29]

Description was changed from

==========
Optimized RGBA->PMColor sampling in SkSwizzler

BUG=skia:
GOLD_TRYBOT_URL=
https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&is...
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

to

==========
NEON Optimized RGBA->PMColor sampling in SkSwizzler

Nexus 6P Swizzle Runtime
BGRA SampleSize=2 0.16x 
RGBA SampleSize=2 0.18x
BGRA SampleSize=4 0.19x
RGBA SampleSize=4 0.21x
BGRA SampleSize=8 0.20x
RGBA SampleSize=8 0.23x

Nexus 6P RGBA PNG Decode Runtime
SampleSize=2 0.98x
SampleSize=4 0.99x
SampleSize=8 0.99x

BUG=skia:
GOLD_TRYBOT_URL=
https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&is...
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

[msarett, 2016-02-10 00:36:18]

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

[msarett, 2016-02-10 00:36:19]

This will be the last of the PNG opts for now (see my previous comments).

I won't land this though unless we can make a bigger impact on performance.  My
guess is that total swizzle time is much shorter when we sample (at most half),
which means optimizing the swizzle has less of an impact on total decode time.

If this lands, I'll follow up with SSE.  But more likely I'll close this and
record the lessons.

https://codereview.chromium.org/1680743005/diff/1/src/codec/SkSwizzler.cpp
File src/codec/SkSwizzler.cpp (right):

https://codereview.chromium.org/1680743005/diff/1/src/codec/SkSwizzler.cpp#ne...
src/codec/SkSwizzler.cpp:968: if (1 == fSampleX && fFastProcs.fSampleSize1) {
On 2016/02/09 13:06:10, scroggo wrote:
> Why not:
> 
> switch (fSampleX) {
>   case 1:
>     fActualProc = fFastProcs.fSampleSize;
>     break;
>   ...
> }
> 
> if (!fActualProc) {
>   fActualProc = fSlowProc;
> }
> 
> (A little bit cleaner to me, but maybe a matter of preference.)

I agree this is nicer.  Done.

https://codereview.chromium.org/1680743005/diff/80001/src/codec/SkSwizzler.cpp
File src/codec/SkSwizzler.cpp (right):

https://codereview.chromium.org/1680743005/diff/80001/src/codec/SkSwizzler.cp...
src/codec/SkSwizzler.cpp:514: SkOpts::RGBA_to_rgbA_sample2((uint32_t*) dst, src
+ offset, width);
For sampling, whether or not to add "offset" here is an interesting question. 
"offset" is always non-zero if we are sampling (since we will skip the first x
pixels).

If we don't add offset, the opt functions must skip some pixels in src before
choosing the pixels we want.

If we do add offset, the opt functions must always keep the first pixel in src.

This leads to different implementations (I've done both, since I messed up the
first time :)).  I *think* this is the best way.

[mtklein, 2016-02-10 00:44:00]

I have not yet looked at the code, though I am confident it is excellent.

> Nexus 6P Swizzle Runtime
> BGRA SampleSize=2 0.16x 
> RGBA SampleSize=2 0.18x
> BGRA SampleSize=4 0.19x
> RGBA SampleSize=4 0.21x
> BGRA SampleSize=8 0.20x
> RGBA SampleSize=8 0.23x
Hot damn, good stuff.

> Nexus 6P RGBA PNG Decode Runtime
> SampleSize=2 0.98x
> SampleSize=4 0.99x
> SampleSize=8 0.99x
Oh wait, why bother?  I'd sort of only want to land this if somehow it meant we
wrote negative lines of code to support the speedup.

[msarett, 2016-02-10 13:28:33]

On 2016/02/10 00:44:00, mtklein wrote:
> I have not yet looked at the code, though I am confident it is excellent.
> 
> > Nexus 6P Swizzle Runtime
> > BGRA SampleSize=2 0.16x 
> > RGBA SampleSize=2 0.18x
> > BGRA SampleSize=4 0.19x
> > RGBA SampleSize=4 0.21x
> > BGRA SampleSize=8 0.20x
> > RGBA SampleSize=8 0.23x
> Hot damn, good stuff.
> 
> > Nexus 6P RGBA PNG Decode Runtime
> > SampleSize=2 0.98x
> > SampleSize=4 0.99x
> > SampleSize=8 0.99x
> Oh wait, why bother?  I'd sort of only want to land this if somehow it meant
we
> wrote negative lines of code to support the speedup.

Matches my thoughts.  Closing this.