Fix BGRA/RGBA readback conversions on mac

Fix BGRA/RGBA readback conversions on mac

Reading back from a BGRA framebuffer as RGBA or vis-versa is slow on mac. On my early 2013 MBP, we see a 40% regression when using
the discrete GPU or a 13% regression on integrated.

This change adds a workaround which causes Mac to use the existing Mesa path and prefer readback from the same format as the
framebuffer.

BUG=581311
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1786813002

Committed: https://skia.googlesource.com/skia/+/b4ecabd5a2c359f66a20207ffb02a77e624560c4

[ericrk, 2016-03-11 20:43:29]

Description was changed from

==========
Fix BGRA/RGBA readback conversions on mac

BUG=581311
==========

to

==========
Fix BGRA/RGBA readback conversions on mac

BUG=581311
GOLD_TRYBOT_URL=
https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&is...
==========

[ericrk, 2016-03-11 20:57:07]

Description was changed from

==========
Fix BGRA/RGBA readback conversions on mac

BUG=581311
GOLD_TRYBOT_URL=
https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&is...
==========

to

==========
Fix BGRA/RGBA readback conversions on mac

Reading back from a BGRA framebuffer as RGBA or vis-versa is slow on mac. On my
early 2013 MBP, we see a 40% regression when using
the discrete GPU or a 13% regression on integrated.

This change adds a workaround which causes Mac to use the existing Mesa path and
prefer readback from the same format as the
framebuffer.

BUG=581311
GOLD_TRYBOT_URL=
https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&is...
==========

[ericrk, 2016-03-11 20:57:07]

ericrk@chromium.org changed reviewers:
+ bsalomon@google.com

[ericrk, 2016-03-11 20:57:25]

Patchset #1 (id:1) has been deleted

[bsalomon, 2016-03-11 22:56:53]

lgtm

[ericrk, 2016-03-11 23:04:58]

The CQ bit was checked by ericrk@chromium.org

[commit-bot: I haz the power, 2016-03-11 23:05:05]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/1786813002/20001
View timeline at
 https://chromium-cq-status.appspot.com/patch-timeline/1786813002/20001

[commit-bot: I haz the power, 2016-03-11 23:18:23]

Description was changed from

==========
Fix BGRA/RGBA readback conversions on mac

Reading back from a BGRA framebuffer as RGBA or vis-versa is slow on mac. On my
early 2013 MBP, we see a 40% regression when using
the discrete GPU or a 13% regression on integrated.

This change adds a workaround which causes Mac to use the existing Mesa path and
prefer readback from the same format as the
framebuffer.

BUG=581311
GOLD_TRYBOT_URL=
https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&is...
==========

to

==========
Fix BGRA/RGBA readback conversions on mac

Reading back from a BGRA framebuffer as RGBA or vis-versa is slow on mac. On my
early 2013 MBP, we see a 40% regression when using
the discrete GPU or a 13% regression on integrated.

This change adds a workaround which causes Mac to use the existing Mesa path and
prefer readback from the same format as the
framebuffer.

BUG=581311
GOLD_TRYBOT_URL=
https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&is...

Committed:
https://skia.googlesource.com/skia/+/b4ecabd5a2c359f66a20207ffb02a77e624560c4
==========

[commit-bot: I haz the power, 2016-03-11 23:18:24]

Committed patchset #1 (id:20001) as
https://skia.googlesource.com/skia/+/b4ecabd5a2c359f66a20207ffb02a77e624560c4

[reed1, 2016-03-12 11:06:35]

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

[reed1, 2016-03-12 11:06:36]

Is it interesting to write a bench for this, at least to document/point-to the
motivation for the checks? I'm thinking of mesa and mac.

aside: I wonder how often this actually happens, that on a mac someone would be
using bgra.

https://codereview.chromium.org/1786813002/diff/20001/src/gpu/gl/GrGLGpu.cpp
File src/gpu/gl/GrGLGpu.cpp (right):

https://codereview.chromium.org/1786813002/diff/20001/src/gpu/gl/GrGLGpu.cpp#...
src/gpu/gl/GrGLGpu.cpp:2567: // Mesa 3D takes a slow path on when reading back
BGRA from an RGBA surface and vice-versa.
Possibly we should update the comment to refer to the flag, rather than Mesa
explicitly?

[ericrk, 2016-03-14 16:55:10]

On 2016/03/12 11:06:36, reed1 wrote:
> Is it interesting to write a bench for this, at least to document/point-to the
> motivation for the checks? I'm thinking of mesa and mac.
> 
> aside: I wonder how often this actually happens, that on a mac someone would
be
> using bgra.
> 
> https://codereview.chromium.org/1786813002/diff/20001/src/gpu/gl/GrGLGpu.cpp
> File src/gpu/gl/GrGLGpu.cpp (right):
> 
>
https://codereview.chromium.org/1786813002/diff/20001/src/gpu/gl/GrGLGpu.cpp#...
> src/gpu/gl/GrGLGpu.cpp:2567: // Mesa 3D takes a slow path on when reading back
> BGRA from an RGBA surface and vice-versa.
> Possibly we should update the comment to refer to the flag, rather than Mesa
> explicitly?

Good points, thanks!

I'll put together a follow up patch to add a bench and fix the comment.

As far as how often this happens - Chrome defaults to BGRA on desktop platforms,
so we definitely use a lot of BGRA resources on mac. This investigation was
prompted due to a canvas readback regression - canvas readback is RGBA, which
was causing us to hit the BGRA > RGBA path during any readback.