Improve naive SkColorXform to half floats

Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us on my desktop.

Color Xform to Half Float (HP z620)
Original                              425us
Trans16 (not 32)                      355us
Vector Trans16                        378us
Trans16 + Keep Halfs in Vector        335us
Vector Trans16 + Keep Halfs in Vector 282us
Final                                 282us

Color Xform to Half Float (Nexus 5X)
Original                              556us
Final                                 472us

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2159993003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Committed: https://skia.googlesource.com/skia/+/6bdbf4412bd1a6fe26be1042ccf080174b13021f

[msarett, 2016-07-18 21:36:23]

Description was changed from

==========
Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us.

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

to

==========
Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2159993003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

[msarett, 2016-07-18 21:36:47]

Description was changed from

==========
Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2159993003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

to

==========
Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us on my desktop.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2159993003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

[msarett, 2016-07-18 21:45:30]

Patchset #1 (id:1) has been deleted

[msarett, 2016-07-18 21:45:35]

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

[commit-bot: I haz the power, 2016-07-18 21:45:50]

Dry run: CQ is trying da patch. Follow status at
 
https://chromium-cq-status.appspot.com/v2/patch-status/codereview.chromium.or...

[commit-bot: I haz the power, 2016-07-18 21:48:13]

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

[commit-bot: I haz the power, 2016-07-18 21:48:13]

Dry run: Try jobs failed on following builders:
  Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot on
master.client.skia (JOB_FAILED,
http://build.chromium.org/p/client.skia/builders/Test-Ubuntu-GCC-GCE-CPU-AVX2...)

[msarett, 2016-07-18 22:09:09]

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

[msarett, 2016-07-18 22:09:09]

https://codereview.chromium.org/2159993003/diff/20001/src/core/SkHalf.h
File src/core/SkHalf.h (right):

https://codereview.chromium.org/2159993003/diff/20001/src/core/SkHalf.h#newco...
src/core/SkHalf.h:68: static inline Sk4h SkFloatToVectorHalf_finite(const Sk4f&
fs) {
I'm seeing significant benefit from having a version of this that keeps the
result in the vector registers.

https://codereview.chromium.org/2159993003/diff/20001/src/opts/SkColorXform_o...
File src/opts/SkColorXform_opts.h (right):

https://codereview.chromium.org/2159993003/diff/20001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:134: #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
Would be interested in adding this to SkNx?

[mtklein, 2016-07-19 12:45:45]

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

[mtklein, 2016-07-19 12:45:46]

https://codereview.chromium.org/2159993003/diff/20001/src/core/SkHalf.h
File src/core/SkHalf.h (right):

https://codereview.chromium.org/2159993003/diff/20001/src/core/SkHalf.h#newco...
src/core/SkHalf.h:68: static inline Sk4h SkFloatToVectorHalf_finite(const Sk4f&
fs) {
On 2016/07/18 22:09:09, msarett wrote:
> I'm seeing significant benefit from having a version of this that keeps the
> result in the vector registers.

Cool cool cool.   Less assembly's always better.  Must be the old code is always
literally producing the fmov, where this new version can see it as a logical
operation?

https://codereview.chromium.org/2159993003/diff/20001/src/opts/SkColorXform_o...
File src/opts/SkColorXform_opts.h (right):

https://codereview.chromium.org/2159993003/diff/20001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:134: #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
On 2016/07/18 22:09:09, msarett wrote:
> Would be interested in adding this to SkNx?

Sure!  We've got a couple little one-off functions already floating around, so
no need to start too fancy or generalized.

// Transpose 4 Sk4h and store (256 bits total).
static inline void Sk4h_store4(void* dst, const Sk4h&, const Sk4h&, const Sk4h&,
const Sk4h&);

https://codereview.chromium.org/2159993003/diff/40001/src/core/SkHalf.h
File src/core/SkHalf.h (right):

https://codereview.chromium.org/2159993003/diff/40001/src/core/SkHalf.h#newco...
src/core/SkHalf.h:73: return vreinterpret_u16_f32(vget_low_f32(vec));
Duh, nice.

https://codereview.chromium.org/2159993003/diff/40001/src/core/SkHalf.h#newco...
src/core/SkHalf.h:96: static inline uint64_t SkFloatToHalf_finite(const Sk4f&
fs) {
Let's port over the other uses and call your new signature SkFloatToHalf_finite.
 They'll also likely benefit.

https://codereview.chromium.org/2159993003/diff/40001/src/opts/SkColorXform_o...
File src/opts/SkColorXform_opts.h (right):

https://codereview.chromium.org/2159993003/diff/40001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:132: Sk4h halfAlphas = Sk4h(0x3C00);
I feel like this is going to come up from time to time.

Let's add this as #define SK_Half1 (or really, constexpr value) to SkHalf.h?

There's also a 0x3C00000000000000 floating around here somewhere right?  
static_cast<uint64_t>(SK_Half1) << 48 ?

https://codereview.chromium.org/2159993003/diff/40001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:133: 
// Transpose the half floats back to interlaced RGBA.  There's no SkNx_ function
for this.

https://codereview.chromium.org/2159993003/diff/40001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:137: __m128i rgba0 = _mm_unpacklo_epi32(rg, ba);
Maybe just lo,hi?  Seeing rgba0 and rgba1 makes me think "ok, pixels 0 and 1"
where they're really pixels 0+1 and 2+3.  Especially when rgba0 and rgba1 mean
pixels 0 and 1 down in the portable path...

https://codereview.chromium.org/2159993003/diff/40001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:142: uint16x4x2_t rg = vzip_u16(halfReds.fVec, 
halfGreens.fVec);
Huh.  I was expecting:

    vst4_u16((uint16_t*)dst, {halfReds.fVec, halfGreens.fVec, halfBlues.fVec,
halfAlphas.fVec});

Any reason we do it ourselves?

(Side note, how do you feel about having specialized SkNx auto-cast to their
vector types.  All that .fVec is sort of annoying, eh?  They already auto-cast
from their vector types.)

https://codereview.chromium.org/2159993003/diff/40001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:150: Sk4h rgba0 = Sk4h(halfReds[0], halfGreens[0],
halfBlues[0], halfAlphas[0]);
How bad is it if you just used this strategy, i.e. convert to half first, then
swizzle?  What part of the speedup comes from swizzling half floats instead of
floats, and what part comes from doing that half float swizzle yourself with SSE
and/or NEON?  (And if you can isolate it, what part from 2 128-bit stores vs. 4
64-bit stores?)

[msarett, 2016-07-19 14:44:53]

Description was changed from

==========
Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us on my desktop.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2159993003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

to

==========
Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us on my desktop.

Color Xform to Half Float (HP z620)
Original 425us

Final    282us

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2159993003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

[msarett, 2016-07-19 15:05:47]

Description was changed from

==========
Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us on my desktop.

Color Xform to Half Float (HP z620)
Original 425us

Final    282us

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2159993003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

to

==========
Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us on my desktop.

Color Xform to Half Float (HP z620)
Original                              425us

Trans16 (not 32)                      355us
Vector Trans16                        378us
Trans16 + Keep Halfs in Vector        335us
Vector Trans16 + Keep Halfs in Vector 282us

Final                                 282us

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2159993003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

[msarett, 2016-07-19 15:11:19]

Description was changed from

==========
Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us on my desktop.

Color Xform to Half Float (HP z620)
Original                              425us

Trans16 (not 32)                      355us
Vector Trans16                        378us
Trans16 + Keep Halfs in Vector        335us
Vector Trans16 + Keep Halfs in Vector 282us

Final                                 282us

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2159993003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

to

==========
Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us on my desktop.

Color Xform to Half Float (HP z620)
Original                              425us
Trans16 (not 32)                      355us
Vector Trans16                        378us
Trans16 + Keep Halfs in Vector        335us
Vector Trans16 + Keep Halfs in Vector 282us
Final                                 282us

Color Xform to Half Float (Nexus 5X)
Original                              556us
Final                                 472us

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2159993003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

[msarett, 2016-07-19 15:23:55]

Patchset #3 (id:60001) has been deleted

[msarett, 2016-07-19 15:24:49]

https://codereview.chromium.org/2159993003/diff/20001/src/core/SkHalf.h
File src/core/SkHalf.h (right):

https://codereview.chromium.org/2159993003/diff/20001/src/core/SkHalf.h#newco...
src/core/SkHalf.h:68: static inline Sk4h SkFloatToVectorHalf_finite(const Sk4f&
fs) {
On 2016/07/19 12:45:45, mtklein wrote:
> On 2016/07/18 22:09:09, msarett wrote:
> > I'm seeing significant benefit from having a version of this that keeps the
> > result in the vector registers.
> 
> Cool cool cool.   Less assembly's always better.  Must be the old code is
always
> literally producing the fmov, where this new version can see it as a logical
> operation?

That's potentially true.

I only looked at the Intel clang disassembly.  I was having issues with:
4h -> store to uint64_t -> load to 4h -> unpack

This code was moving out of the vector registers, then reloading one lane at a
time (instead of the unpack).  It was ugly.

https://codereview.chromium.org/2159993003/diff/20001/src/opts/SkColorXform_o...
File src/opts/SkColorXform_opts.h (right):

https://codereview.chromium.org/2159993003/diff/20001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:134: #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
On 2016/07/19 12:45:45, mtklein wrote:
> On 2016/07/18 22:09:09, msarett wrote:
> > Would be interested in adding this to SkNx?
> 
> Sure!  We've got a couple little one-off functions already floating around, so
> no need to start too fancy or generalized.
> 
> // Transpose 4 Sk4h and store (256 bits total).
> static inline void Sk4h_store4(void* dst, const Sk4h&, const Sk4h&, const
Sk4h&,
> const Sk4h&);

Done.

https://codereview.chromium.org/2159993003/diff/40001/src/core/SkHalf.h
File src/core/SkHalf.h (right):

https://codereview.chromium.org/2159993003/diff/40001/src/core/SkHalf.h#newco...
src/core/SkHalf.h:96: static inline uint64_t SkFloatToHalf_finite(const Sk4f&
fs) {
On 2016/07/19 12:45:45, mtklein wrote:
> Let's port over the other uses and call your new signature
SkFloatToHalf_finite.
>  They'll also likely benefit.

Done.

https://codereview.chromium.org/2159993003/diff/40001/src/opts/SkColorXform_o...
File src/opts/SkColorXform_opts.h (right):

https://codereview.chromium.org/2159993003/diff/40001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:132: Sk4h halfAlphas = Sk4h(0x3C00);
On 2016/07/19 12:45:46, mtklein wrote:
> I feel like this is going to come up from time to time.
> 
> Let's add this as #define SK_Half1 (or really, constexpr value) to SkHalf.h?
> 
> There's also a 0x3C00000000000000 floating around here somewhere right?  
> static_cast<uint64_t>(SK_Half1) << 48 ?

SGTM.  Done.

https://codereview.chromium.org/2159993003/diff/40001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:133: 
On 2016/07/19 12:45:46, mtklein wrote:
> // Transpose the half floats back to interlaced RGBA.  There's no SkNx_
function
> for this.

Added to Nx.

https://codereview.chromium.org/2159993003/diff/40001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:137: __m128i rgba0 = _mm_unpacklo_epi32(rg, ba);
On 2016/07/19 12:45:46, mtklein wrote:
> Maybe just lo,hi?  Seeing rgba0 and rgba1 makes me think "ok, pixels 0 and 1"
> where they're really pixels 0+1 and 2+3.  Especially when rgba0 and rgba1 mean
> pixels 0 and 1 down in the portable path...

Done.

https://codereview.chromium.org/2159993003/diff/40001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:142: uint16x4x2_t rg = vzip_u16(halfReds.fVec, 
halfGreens.fVec);
On 2016/07/19 12:45:46, mtklein wrote:
> Huh.  I was expecting:
> 
>     vst4_u16((uint16_t*)dst, {halfReds.fVec, halfGreens.fVec, halfBlues.fVec,
> halfAlphas.fVec});
> 
> Any reason we do it ourselves?
> 

Duh, yup that's way better.  Think I was stuck in Intel mode haha.

> (Side note, how do you feel about having specialized SkNx auto-cast to their
> vector types.  All that .fVec is sort of annoying, eh?  They already auto-cast
> from their vector types.)

Works for me.  No strong opinion.

https://codereview.chromium.org/2159993003/diff/40001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:150: Sk4h rgba0 = Sk4h(halfReds[0], halfGreens[0],
halfBlues[0], halfAlphas[0]);
On 2016/07/19 12:45:46, mtklein wrote:
> How bad is it if you just used this strategy, i.e. convert to half first, then
> swizzle?  What part of the speedup comes from swizzling half floats instead of
> floats, and what part comes from doing that half float swizzle yourself with
SSE
> and/or NEON?  (And if you can isolate it, what part from 2 128-bit stores vs.
4
> 64-bit stores?)

Added way more performance measures to commit message.

[mtklein, 2016-07-19 15:37:47]

lgtm, with some style suggestions

https://codereview.chromium.org/2159993003/diff/80001/src/core/SkNx.h
File src/core/SkNx.h (right):

https://codereview.chromium.org/2159993003/diff/80001/src/core/SkNx.h#newcode313
src/core/SkNx.h:313: SI void Sk4h_store4(void* dst, const Sk4h& r, const Sk4h&
g, const Sk4h& b, const Sk4h& a) {
How about

auto dst64 = (uint64_t*)dst;
Sk4h(r[0], g[0], b[0], a[0]).store(dst64 + 0);
Sk4h(r[1], g[1], b[1], a[1]).store(dst64 + 1);
...

https://codereview.chromium.org/2159993003/diff/80001/src/opts/SkColorXform_o...
File src/opts/SkColorXform_opts.h (right):

https://codereview.chromium.org/2159993003/diff/80001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:129: Sk4h halfReds   =
SkFloatToHalf_finite(dstReds);
Might wanna just inline these instead of naming them?

Sk4h_store4(dst, SkFloatToHalf_finite(dstReds),
                 SkFloatToHalf_finite(dstGreens),
                 SkFloatToHalf_finite(dstBlues),
                 SK_Half1);

https://codereview.chromium.org/2159993003/diff/80001/src/opts/SkNx_neon.h
File src/opts/SkNx_neon.h (right):

https://codereview.chromium.org/2159993003/diff/80001/src/opts/SkNx_neon.h#ne...
src/opts/SkNx_neon.h:488: uint16x4x4_t rgba;
uint16x4x4_t rgba = {{
   r.fVec,
   g.fVec,
   b.fVec,
   a.fVec,
}};

might be prettier.

[msarett, 2016-07-19 15:47:20]

https://codereview.chromium.org/2159993003/diff/80001/src/core/SkNx.h
File src/core/SkNx.h (right):

https://codereview.chromium.org/2159993003/diff/80001/src/core/SkNx.h#newcode313
src/core/SkNx.h:313: SI void Sk4h_store4(void* dst, const Sk4h& r, const Sk4h&
g, const Sk4h& b, const Sk4h& a) {
On 2016/07/19 15:37:47, mtklein wrote:
> How about
> 
> auto dst64 = (uint64_t*)dst;
> Sk4h(r[0], g[0], b[0], a[0]).store(dst64 + 0);
> Sk4h(r[1], g[1], b[1], a[1]).store(dst64 + 1);
> ...

SGTM

https://codereview.chromium.org/2159993003/diff/80001/src/opts/SkColorXform_o...
File src/opts/SkColorXform_opts.h (right):

https://codereview.chromium.org/2159993003/diff/80001/src/opts/SkColorXform_o...
src/opts/SkColorXform_opts.h:129: Sk4h halfReds   =
SkFloatToHalf_finite(dstReds);
On 2016/07/19 15:37:47, mtklein wrote:
> Might wanna just inline these instead of naming them?
> 
> Sk4h_store4(dst, SkFloatToHalf_finite(dstReds),
>                  SkFloatToHalf_finite(dstGreens),
>                  SkFloatToHalf_finite(dstBlues),
>                  SK_Half1);

Woo cool.

https://codereview.chromium.org/2159993003/diff/80001/src/opts/SkNx_neon.h
File src/opts/SkNx_neon.h (right):

https://codereview.chromium.org/2159993003/diff/80001/src/opts/SkNx_neon.h#ne...
src/opts/SkNx_neon.h:488: uint16x4x4_t rgba;
On 2016/07/19 15:37:47, mtklein wrote:
> uint16x4x4_t rgba = {{
>    r.fVec,
>    g.fVec,
>    b.fVec,
>    a.fVec,
> }};
> 
> might be prettier.

Done.  I tried that first, but the compiler complained (I think I missed the
double brace)

[msarett, 2016-07-19 15:47:28]

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

[commit-bot: I haz the power, 2016-07-19 15:47:31]

Dry run: CQ is trying da patch. Follow status at
 
https://chromium-cq-status.appspot.com/v2/patch-status/codereview.chromium.or...

[commit-bot: I haz the power, 2016-07-19 16:06:09]

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

[commit-bot: I haz the power, 2016-07-19 16:06:10]

Dry run: This issue passed the CQ dry run.

[msarett, 2016-07-19 16:06:51]

The CQ bit was checked by msarett@google.com

[msarett, 2016-07-19 16:06:52]

The patchset sent to the CQ was uploaded after l-g-t-m from mtklein@google.com
Link to the patchset: https://codereview.chromium.org/2159993003/#ps100001
(title: "Prettier code")

[commit-bot: I haz the power, 2016-07-19 16:06:59]

CQ is trying da patch. Follow status at
 
https://chromium-cq-status.appspot.com/v2/patch-status/codereview.chromium.or...

[commit-bot: I haz the power, 2016-07-19 16:07:57]

Description was changed from

==========
Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us on my desktop.

Color Xform to Half Float (HP z620)
Original                              425us
Trans16 (not 32)                      355us
Vector Trans16                        378us
Trans16 + Keep Halfs in Vector        335us
Vector Trans16 + Keep Halfs in Vector 282us
Final                                 282us

Color Xform to Half Float (Nexus 5X)
Original                              556us
Final                                 472us

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2159993003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot
==========

to

==========
Improve naive SkColorXform to half floats

This should give us a good baseline to explore using SkRasterPipeline.

A particular colorxform to half float drops from 425us to 282us on my desktop.

Color Xform to Half Float (HP z620)
Original                              425us
Trans16 (not 32)                      355us
Vector Trans16                        378us
Trans16 + Keep Halfs in Vector        335us
Vector Trans16 + Keep Halfs in Vector 282us
Final                                 282us

Color Xform to Half Float (Nexus 5X)
Original                              556us
Final                                 472us

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2159993003
CQ_INCLUDE_TRYBOTS=master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Committed:
https://skia.googlesource.com/skia/+/6bdbf4412bd1a6fe26be1042ccf080174b13021f
==========

[commit-bot: I haz the power, 2016-07-19 16:07:58]

Committed patchset #4 (id:100001) as
https://skia.googlesource.com/skia/+/6bdbf4412bd1a6fe26be1042ccf080174b13021f