Move CPU feature detection to its own file.

Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common
       * SSE and SSE3... why not?

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the check (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

Once this lands and bakes a bit, I'll start following up with CLs to use it more
and to add a bunch of those little 1-2 instruction snippets we've been wanting,
e.g. cvtps2ph, cvtph2ps, ptest, pmulld, pmovzxbd, blendvps, pshufb, roundps 
(for floor) on x86, and vcvt.f32.f16, vcvt.f16.f32 on ARM.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1890483002
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Committed: https://skia.googlesource.com/skia/+/872ea29357439f05b1f6995dd300fc054733e607

Committed: https://skia.googlesource.com/skia/+/4311f016612a814282029daa4bd102053a853d82

[mtklein_C, 2016-04-13 16:15:10]

Description was changed from

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around opts/.
   - Can now detect a few new CPU features:
       * F16C -> Intel added f16<->f32 instructions between AVX and AVX2
       * FMA  -> Intel FMA instructions, added at the same time as AVX2
       * NEON_FMA -> ARM FMA instructions, quite common
       * VFP_FP16 -> ARM f16<->f32 instructions, also quite common

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the call (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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

to

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around opts/.
   - Can now detect a few new CPU features:
       * F16C -> Intel added f16<->f32 instructions between AVX and AVX2
       * FMA  -> Intel FMA instructions, added at the same time as AVX2
       * NEON_FMA -> ARM FMA instructions, quite common
       * VFP_FP16 -> ARM f16<->f32 instructions, also quite common

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the call (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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

[mtklein_C, 2016-04-13 16:21:09]

Description was changed from

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around opts/.
   - Can now detect a few new CPU features:
       * F16C -> Intel added f16<->f32 instructions between AVX and AVX2
       * FMA  -> Intel FMA instructions, added at the same time as AVX2
       * NEON_FMA -> ARM FMA instructions, quite common
       * VFP_FP16 -> ARM f16<->f32 instructions, also quite common

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the call (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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

to

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around opts/.
   - Can now detect a few new CPU features:
       * F16C -> Intel added f16<->f32 instructions between AVX and AVX2
       * FMA  -> Intel FMA instructions, added at the same time as AVX2
       * NEON_FMA -> ARM FMA instructions, quite common
       * VFP_FP16 -> ARM f16<->f32 instructions, also quite common

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the call (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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
==========

[mtklein_C, 2016-04-13 16:22:19]

Description was changed from

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around opts/.
   - Can now detect a few new CPU features:
       * F16C -> Intel added f16<->f32 instructions between AVX and AVX2
       * FMA  -> Intel FMA instructions, added at the same time as AVX2
       * NEON_FMA -> ARM FMA instructions, quite common
       * VFP_FP16 -> ARM f16<->f32 instructions, also quite common

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the call (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the call (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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
==========

[mtklein_C, 2016-04-13 16:27:01]

The CQ bit was checked by mtklein@chromium.org to run a CQ dry run

[commit-bot: I haz the power, 2016-04-13 16:27:09]

Dry run: CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1890483002/40001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1890483002/40001

[commit-bot: I haz the power, 2016-04-13 16:28:44]

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

[commit-bot: I haz the power, 2016-04-13 16:28:45]

Dry run: Try jobs failed on following builders:
  Build-Ubuntu-GCC-x86_64-Release-Trybot on client.skia.compile (JOB_FAILED,
http://build.chromium.org/p/client.skia.compile/builders/Build-Ubuntu-GCC-x86...)

[mtklein_C, 2016-04-13 16:30:13]

The CQ bit was checked by mtklein@chromium.org to run a CQ dry run

[commit-bot: I haz the power, 2016-04-13 16:30:19]

Dry run: CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1890483002/60001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1890483002/60001

[commit-bot: I haz the power, 2016-04-13 16:40:01]

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

[commit-bot: I haz the power, 2016-04-13 16:40:02]

Dry run: This issue passed the CQ dry run.

[mtklein_C, 2016-04-13 16:42:01]

Description was changed from

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the call (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the call (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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
==========

[mtklein_C, 2016-04-13 16:43:51]

Description was changed from

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the call (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common
       * SSE and SSE3... why not?

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the call (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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
==========

[mtklein_C, 2016-04-13 16:47:46]

Description was changed from

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common
       * SSE and SSE3... why not?

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the call (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common
       * SSE and SSE3... why not?

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the check (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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
==========

[mtklein_C, 2016-04-13 16:52:10]

The CQ bit was checked by mtklein@chromium.org to run a CQ dry run

[commit-bot: I haz the power, 2016-04-13 16:52:22]

Dry run: CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1890483002/80001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1890483002/80001

[mtklein_C, 2016-04-13 16:56:57]

Description was changed from

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common
       * SSE and SSE3... why not?

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the check (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

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

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common
       * SSE and SSE3... why not?

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the check (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

Once this lands and bakes a bit, I'll start following up with CLs to use it more
and to add a bunch of those little 1-2 instruction snippets we've been wanting,
e.g. cvtps2ph, cvtph2ps, ptest, pmulld, pmovzxbd, blendvps, pshufb, roundps (for
floor) on x86, and vcvt.f32.f16, vcvt.f16.f32 on ARM.

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
==========

[mtklein_C, 2016-04-13 16:57:08]

Description was changed from

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common
       * SSE and SSE3... why not?

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the check (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

Once this lands and bakes a bit, I'll start following up with CLs to use it more
and to add a bunch of those little 1-2 instruction snippets we've been wanting,
e.g. cvtps2ph, cvtph2ps, ptest, pmulld, pmovzxbd, blendvps, pshufb, roundps (for
floor) on x86, and vcvt.f32.f16, vcvt.f16.f32 on ARM.

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

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common
       * SSE and SSE3... why not?

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the check (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

Once this lands and bakes a bit, I'll start following up with CLs to use it more
and to add a bunch of those little 1-2 instruction snippets we've been wanting,
e.g. cvtps2ph, cvtph2ps, ptest, pmulld, pmovzxbd, blendvps, pshufb, roundps 
(for floor) on x86, and vcvt.f32.f16, vcvt.f16.f32 on ARM.

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
==========

[commit-bot: I haz the power, 2016-04-13 17:02:19]

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

[commit-bot: I haz the power, 2016-04-13 17:02:20]

Dry run: This issue passed the CQ dry run.

[mtklein_C, 2016-04-13 17:16:21]

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

[mtklein_C, 2016-04-13 17:16:21]

The CQ bit was checked by mtklein@chromium.org

[commit-bot: I haz the power, 2016-04-13 17:16:25]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1890483002/80001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1890483002/80001

[commit-bot: I haz the power, 2016-04-13 17:16:27]

Note for Reviewers:
The CQ is waiting for an approval.
If you believe that the CL is not ready yet, or if you would like to L-G-T-M
with comments then please uncheck the CQ checkbox.

Waiting for LGTM from valid reviewer(s) till 2016-04-13 23:16 UTC

[mtklein_C, 2016-04-13 20:30:59]

The CQ bit was checked by mtklein@chromium.org to run a CQ dry run

[commit-bot: I haz the power, 2016-04-13 20:31:07]

Dry run: CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1890483002/100001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1890483002/100001

[commit-bot: I haz the power, 2016-04-13 20:42:37]

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

[commit-bot: I haz the power, 2016-04-13 20:42:38]

Dry run: This issue passed the CQ dry run.

[mtklein_C, 2016-04-13 22:28:13]

mtklein@chromium.org changed reviewers:
+ fmalita@chromium.org, reed@google.com

[mtklein_C, 2016-04-13 22:28:14]

+Mike and Florin

I think I can take yesterday's half<->float CL in a slightly different direction
that makes it generally cheaper to do runtime CPU detection.  This is a
foundation CL, and then the next one I'll send out builds on it to add that fast
x86 float conversion.

[herb_g, 2016-04-14 15:38:34]

lgtm

Yeah!!!

https://codereview.chromium.org/1890483002/diff/100001/src/core/SkCpu.cpp
File src/core/SkCpu.cpp (right):

https://codereview.chromium.org/1890483002/diff/100001/src/core/SkCpu.cpp#new...
src/core/SkCpu.cpp:18: #include <cpuid.h>
If you have a link to the docs that describe cpuid that you used to figure out
this code could you add it here.

[f(malita), 2016-04-14 15:40:36]

lgtm

https://codereview.chromium.org/1890483002/diff/100001/src/core/SkCpu.cpp
File src/core/SkCpu.cpp (right):

https://codereview.chromium.org/1890483002/diff/100001/src/core/SkCpu.cpp#new...
src/core/SkCpu.cpp:39: if (abcd[3] & (1<<25)) { features |= SkCpu:: SSE1; }
It's a bummer we can't use the cpuid.h macros instead of magic constants.  I
guess adding an enum would be kinda redundant too.

https://codereview.chromium.org/1890483002/diff/100001/src/core/SkCpu.h
File src/core/SkCpu.h (right):

https://codereview.chromium.org/1890483002/diff/100001/src/core/SkCpu.h#newco...
src/core/SkCpu.h:62: // (Static intializers would work fine everywhere, but
Chrome really dislikes them.)
Does this mean we'll have to sell a static initialized to Chrome/MSVC?

[mtklein, 2016-04-14 15:52:27]

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

[mtklein, 2016-04-14 15:52:28]

https://codereview.chromium.org/1890483002/diff/100001/src/core/SkCpu.cpp
File src/core/SkCpu.cpp (right):

https://codereview.chromium.org/1890483002/diff/100001/src/core/SkCpu.cpp#new...
src/core/SkCpu.cpp:18: #include <cpuid.h>
On 2016/04/14 at 15:38:34, herb_g wrote:
> If you have a link to the docs that describe cpuid that you used to figure out
this code could you add it here.

Done, below.

https://codereview.chromium.org/1890483002/diff/100001/src/core/SkCpu.cpp#new...
src/core/SkCpu.cpp:39: if (abcd[3] & (1<<25)) { features |= SkCpu:: SSE1; }
On 2016/04/14 at 15:40:35, f(malita) wrote:
> It's a bummer we can't use the cpuid.h macros instead of magic constants.  I
guess adding an enum would be kinda redundant too.

Yeah, we could use them when we have cpuid.h, but that'd fork this code.  This
function more or less is the enum(s) we'd write.

https://codereview.chromium.org/1890483002/diff/100001/src/core/SkCpu.h
File src/core/SkCpu.h (right):

https://codereview.chromium.org/1890483002/diff/100001/src/core/SkCpu.h#newco...
src/core/SkCpu.h:62: // (Static intializers would work fine everywhere, but
Chrome really dislikes them.)
On 2016/04/14 at 15:40:35, f(malita) wrote:
> Does this mean we'll have to sell a static initialized to Chrome/MSVC?

Yeah, maybe, for some values of "have to" and "sell".

Skia's already got a few Windows-only static initializers, and I'm told they're
not such a problem on Windows as they are (used to be?) with other executable
formats that store the initializers in a way that's not at all disk-seek
friendly.

I'm open to continuing to look for other non-static-intializer approaches for
MSVC that still allow this to act like a constant... I just haven't found one
yet.

[mtklein, 2016-04-14 15:53:19]

The CQ bit was checked by mtklein@google.com

[mtklein, 2016-04-14 15:53:19]

The patchset sent to the CQ was uploaded after l-g-t-m from herb@google.com,
fmalita@chromium.org
Link to the patchset: https://codereview.chromium.org/1890483002/#ps120001
(title: "link")

[commit-bot: I haz the power, 2016-04-14 15:53:27]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1890483002/120001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1890483002/120001

[commit-bot: I haz the power, 2016-04-14 16:03:37]

Description was changed from

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common
       * SSE and SSE3... why not?

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the check (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

Once this lands and bakes a bit, I'll start following up with CLs to use it more
and to add a bunch of those little 1-2 instruction snippets we've been wanting,
e.g. cvtps2ph, cvtph2ps, ptest, pmulld, pmovzxbd, blendvps, pshufb, roundps 
(for floor) on x86, and vcvt.f32.f16, vcvt.f16.f32 on ARM.

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

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common
       * SSE and SSE3... why not?

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the check (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

Once this lands and bakes a bit, I'll start following up with CLs to use it more
and to add a bunch of those little 1-2 instruction snippets we've been wanting,
e.g. cvtps2ph, cvtph2ps, ptest, pmulld, pmovzxbd, blendvps, pshufb, roundps 
(for floor) on x86, and vcvt.f32.f16, vcvt.f16.f32 on ARM.

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

Committed:
https://skia.googlesource.com/skia/+/872ea29357439f05b1f6995dd300fc054733e607
==========

[commit-bot: I haz the power, 2016-04-14 16:03:38]

Committed patchset #7 (id:120001) as
https://skia.googlesource.com/skia/+/872ea29357439f05b1f6995dd300fc054733e607

[mtklein, 2016-04-15 15:40:06]

A revert of this CL (patchset #7 id:120001) has been created in
https://codereview.chromium.org/1892643003/ by mtklein@google.com.

The reason for reverting is: many unexpected GM diffs across GPU+CPU configs on
Windows (hopefully just text masks on GPU?).   seems like we pick a different
srcover variant in some places..

[mtklein, 2016-04-19 15:32:39]

On 2016/04/15 at 15:40:06, mtklein wrote:
> A revert of this CL (patchset #7 id:120001) has been created in
https://codereview.chromium.org/1892643003/ by mtklein@google.com.
> 
> The reason for reverting is: many unexpected GM diffs across GPU+CPU configs
on Windows (hopefully just text masks on GPU?).   seems like we pick a different
srcover variant in some places..

Note to self: reverting src/core/SkOpts.cpp is enough to fix the diffs on
Windows.

[mtklein, 2016-04-19 15:36:33]

On 2016/04/19 at 15:32:39, mtklein wrote:
> On 2016/04/15 at 15:40:06, mtklein wrote:
> > A revert of this CL (patchset #7 id:120001) has been created in
https://codereview.chromium.org/1892643003/ by mtklein@google.com.
> > 
> > The reason for reverting is: many unexpected GM diffs across GPU+CPU configs
on Windows (hopefully just text masks on GPU?).   seems like we pick a different
srcover variant in some places..
> 
> Note to self: reverting src/core/SkOpts.cpp is enough to fix the diffs on
Windows.

Disabling  if (abcd[2] & (1<<19)) { Init_sse41(); }  in the old code gives the
same effect as landing this CL.

[mtklein, 2016-04-19 16:05:04]

On 2016/04/19 at 15:36:33, mtklein wrote:
> On 2016/04/19 at 15:32:39, mtklein wrote:
> > On 2016/04/15 at 15:40:06, mtklein wrote:
> > > A revert of this CL (patchset #7 id:120001) has been created in
https://codereview.chromium.org/1892643003/ by mtklein@google.com.
> > > 
> > > The reason for reverting is: many unexpected GM diffs across GPU+CPU
configs on Windows (hopefully just text masks on GPU?).   seems like we pick a
different srcover variant in some places..
> > 
> > Note to self: reverting src/core/SkOpts.cpp is enough to fix the diffs on
Windows.
> 
> Disabling  if (abcd[2] & (1<<19)) { Init_sse41(); }  in the old code gives the
same effect as landing this CL.

I've fallen into a classic trap.
The problem is global initialization is in unspecified order.  I have an
implicit ordering requirement between the new Windows-only gCachedCpuFeatures
and this intiializer in SkOpts.cpp:

#if SK_ALLOW_STATIC_GLOBAL_INITIALIZERS
    static struct AutoInit {
        AutoInit() { Init(); }
    } gAutoInit;
#endif

This gAutoInit calls Init() which calls init() once, which calls
SkCpu::Supports() which calls SkCpu::RuntimeFeatures() which reads
gCachedCpuFeatures, which has not yet been initialized.  It reads as zero, so
all SkCpu::Supports() calls return false.

Simplest way to handle is probably remove that SkOpts.cpp static initializer.

[mtklein_C, 2016-04-19 20:41:50]

The CQ bit was checked by mtklein@chromium.org to run a CQ dry run

[commit-bot: I haz the power, 2016-04-19 20:41:58]

Dry run: CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1890483002/120001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1890483002/120001

[mtklein_C, 2016-04-19 20:49:49]

The CQ bit was unchecked by mtklein@chromium.org

[mtklein_C, 2016-04-19 20:49:51]

The CQ bit was checked by mtklein@chromium.org

[commit-bot: I haz the power, 2016-04-19 20:49:58]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1890483002/120001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1890483002/120001

[commit-bot: I haz the power, 2016-04-19 21:00:19]

Description was changed from

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common
       * SSE and SSE3... why not?

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the check (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

Once this lands and bakes a bit, I'll start following up with CLs to use it more
and to add a bunch of those little 1-2 instruction snippets we've been wanting,
e.g. cvtps2ph, cvtph2ps, ptest, pmulld, pmovzxbd, blendvps, pshufb, roundps 
(for floor) on x86, and vcvt.f32.f16, vcvt.f16.f32 on ARM.

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

Committed:
https://skia.googlesource.com/skia/+/872ea29357439f05b1f6995dd300fc054733e607
==========

to

==========
Move CPU feature detection to its own file.

   - Moves CPU feature detection to its own file.
   - Cleans up some redundant feature detection scattered around core/ and
opts/.
   - Can now detect a few new CPU features:
       * F16C     -> Intel f16<->f32 instructions, added between AVX and AVX2
       * FMA      -> Intel FMA instructions, added at the same time as AVX2
       * VFP_FP16 -> ARM f16<->f32 instructions, quite common
       * NEON_FMA -> ARM FMA instructions, also quite common
       * SSE and SSE3... why not?

This new internal API makes it very cheap to do fine-grained runtime CPU
feature detection.  Redundant calls to SkCpu::Supports() should be eliminated
and it's hoistable out of loops.  It compiles away entirely when we have the
appropriate instructions available at compile time.

This means we can call it to guard even a little snippet of 1 or 2 instructions
right where needed and let inlining hoist the check (if any at all) up to
somewhere that doesn't hurt performance.  I've explained how I made this work
in the private section of the new header.

Once this lands and bakes a bit, I'll start following up with CLs to use it more
and to add a bunch of those little 1-2 instruction snippets we've been wanting,
e.g. cvtps2ph, cvtph2ps, ptest, pmulld, pmovzxbd, blendvps, pshufb, roundps 
(for floor) on x86, and vcvt.f32.f16, vcvt.f16.f32 on ARM.

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

Committed:
https://skia.googlesource.com/skia/+/872ea29357439f05b1f6995dd300fc054733e607

Committed:
https://skia.googlesource.com/skia/+/4311f016612a814282029daa4bd102053a853d82
==========

[commit-bot: I haz the power, 2016-04-19 21:00:20]

Committed patchset #7 (id:120001) as
https://skia.googlesource.com/skia/+/4311f016612a814282029daa4bd102053a853d82