SkFloatConvert

SkFloatConvert

This is SkFloatToHalf and SkHalfToFloat plus a little bit of state
to know whether we have fast f16<->f32 instructions at runtime.

On my laptop with f16c instructions, xferu64_srcover_bw_N_alpha_f16 runtime drops from 1.99ms to 1.09ms.

Here's the new disassembly:

+0x90     movups              (%rdx), %xmm1
+0x93	    movq                (%rsi), %xmm2
+0x97	    vcvtph2ps           %xmm2, %xmm2
+0x9c	    movaps              %xmm1, %xmm3
+0x9f	    shufps              $255, %xmm3, %xmm3
+0xa3	    movaps              %xmm0, %xmm4
+0xa6	    subps               %xmm3, %xmm4
+0xa9	    mulps               %xmm2, %xmm4
+0xac	    addps               %xmm1, %xmm4
+0xaf	    vcvtps2ph           $0, %xmm4, %xmm4
+0xb5	    movq                %xmm4, (%rsi)
+0xb9	    addq                $16, %rdx
+0xbd	    addq                $8, %rsi
+0xc1	    decl                %ecx
+0xc3	    jne                 +0x90



BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1884683002

[mtklein_C, 2016-04-12 17:33:18]

Description was changed from

==========
SkFloatConvert

This is SkFloatToHalf and SkHalfToFloat plus a little bit of state
to know whether we have fast f16<->f32 instructions at runtime.



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

to

==========
SkFloatConvert

This is SkFloatToHalf and SkHalfToFloat plus a little bit of state
to know whether we have fast f16<->f32 instructions at runtime.



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

[mtklein_C, 2016-04-12 17:34:29]

Description was changed from

==========
SkFloatConvert

This is SkFloatToHalf and SkHalfToFloat plus a little bit of state
to know whether we have fast f16<->f32 instructions at runtime.



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

to

==========
SkFloatConvert

This is SkFloatToHalf and SkHalfToFloat plus a little bit of state
to know whether we have fast f16<->f32 instructions at runtime.

On my laptop with f16c instructions, xferu64_srcover_bw_N_alpha_f16 runtime
drops from 1.99ms to 1.09ms.



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

[mtklein_C, 2016-04-12 17:36:28]

Description was changed from

==========
SkFloatConvert

This is SkFloatToHalf and SkHalfToFloat plus a little bit of state
to know whether we have fast f16<->f32 instructions at runtime.

On my laptop with f16c instructions, xferu64_srcover_bw_N_alpha_f16 runtime
drops from 1.99ms to 1.09ms.



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

to

==========
SkFloatConvert

This is SkFloatToHalf and SkHalfToFloat plus a little bit of state
to know whether we have fast f16<->f32 instructions at runtime.

On my laptop with f16c instructions, xferu64_srcover_bw_N_alpha_f16 runtime
drops from 1.99ms to 1.09ms.

Here's the new disassembly:

+0x90     movups              (%rdx), %xmm1
+0x93	    movq                (%rsi), %xmm2
+0x97	    vcvtph2ps           %xmm2, %xmm2
+0x9c	    movaps              %xmm1, %xmm3
+0x9f	    shufps              $255, %xmm3, %xmm3
+0xa3	    movaps              %xmm0, %xmm4
+0xa6	    subps               %xmm3, %xmm4
+0xa9	    mulps               %xmm2, %xmm4
+0xac	    addps               %xmm1, %xmm4
+0xaf	    vcvtps2ph           $0, %xmm4, %xmm4
+0xb5	    movq                %xmm4, (%rsi)
+0xb9	    addq                $16, %rdx
+0xbd	    addq                $8, %rsi
+0xc1	    decl                %ecx
+0xc3	    jne                 +0x90



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

[herb_g, 2016-04-12 18:27:03]

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

[herb_g, 2016-04-12 18:27:03]

I like this a lot. Can we do this with other features like floor(round)?

[mtklein, 2016-04-12 18:34:34]

On 2016/04/12 at 18:27:03, herb wrote:
> I like this a lot. Can we do this with other features like floor(round)?

Potentially this strategy can be applied to everything SkNx.h touches.

The challenges I'm trying to work out are:

   1)  How best to keep feature-detection overhead negligible.  Here my strategy
is, cache cpuid() globally, then cache the global locally with each
SkFloatConvert instance.  I've failed so far at removing the need for that
second cache layer, but haven't given up yet.

   2)  What repercussions this would have on the usability of SkNx, especially
around how it'd affect how vectors pass in and out of function calls.

[reed1, 2016-04-12 18:47:26]

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

[reed1, 2016-04-12 18:47:26]

api lgtm

https://codereview.chromium.org/1884683002/diff/1/src/core/SkOpts.h
File src/core/SkOpts.h (right):

https://codereview.chromium.org/1884683002/diff/1/src/core/SkOpts.h#newcode24
src/core/SkOpts.h:24: extern bool has_f16c;
nit: gHas_f16c -- just to make reading it at each site slightly easier to know
its scope

[reed1, 2016-04-12 18:47:53]

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

[reed1, 2016-04-12 18:47:54]

awesome speedup!

[f(malita), 2016-04-12 22:51:52]

LGTM!

On 2016/04/12 18:34:34, mtklein wrote:
> vcvtps2ph

They must be face-rolling on the keyboard for mnemonics these days.