[WIP Experiment] Implement mix-blend-mode in GLRenderer using shaders.

The current implementation of blending in GLRenderer uses Ganesh to
blend the content with the backdrop into a temporary texture,
and then uses that temporary texture by replacing the actual content.

In this experiment, blending formulas are added to the existing GL 
fragment shaders and blending is done in place, using only one
temporary texture and one pixel copy operation for storing the
backdrop pixels and using it as input for shaders. The current
implementation uses 3 temporary textures and 4 pixel copy
operations per surface per frame.

I tested the performance using animations and blending on two mobile
devices. Using shaders I got 45 fps on a Samsung Galaxy Note 10.1
tabled, versus 10 fps with the Skia based implementation. On a Nexus 5
device, it's 54 fps vs 38 fps.

Another gain in this experiment, besides performance is that rendering
defects we encountered with the skia implementation are not reproducible
using shaders: https://code.google.com/p/chromium/issues/detail?id=403217.

We can also see an image quality improvement with the shader base
implementation. Here is an example that with Chrome Canary looks pixelated,
and with this patch it looks very clear: http://codepen.io/pixelass/full/uqbjr. 
With Skia implementation (Chrome Canary), the backdrop is being read using 
content's transform, then the inverse transform is applied in order to get 
backdrop into the content's space, consequently losing quality.

[rosca, 2014-09-15 14:24:01]

rosca@adobe.com changed reviewers:
+ danakj@chromium.org

[rosca, 2014-09-15 14:27:42]

On 2014/09/15 14:24:01, Rosca wrote:
> mailto:rosca@adobe.com changed reviewers:
> + mailto:danakj@chromium.org

Hi @danakj, could you please have a high level look on this patch? What do you
think about this approach?

[rosca, 2014-09-16 18:45:38]

rosca@adobe.com changed reviewers:
+ enne@chromium.org

[danakj, 2014-09-22 15:10:25]

+enne

[rosca, 2014-09-23 21:04:44]

I updated the description with another advantage of this implementation.
@enne, could you please let us know what do you think about this approach?

[enne (OOO), 2014-09-23 21:34:24]

I saw danakj's message earlier when she CC'd me.  I will get to this when I get
a chance, but am quite swamped at the moment, sorry.

[enne (OOO), 2014-09-24 22:07:56]

This is a lot of new shader code.  Do the existing blending pixel tests cover
all of these cases?  It looks like you're also touching render pass mask
shaders, and I don't see any blending + mask tests.  Could you add some?

https://codereview.chromium.org/555133002/diff/120001/cc/output/gl_renderer.cc
File cc/output/gl_renderer.cc (right):

https://codereview.chromium.org/555133002/diff/120001/cc/output/gl_renderer.c...
cc/output/gl_renderer.cc:2735: 1,   // kScreen_Mode
Why do you need this array? Why not just blend_mode - kModulateMode along with
some compile-time asserts for sanity?

https://codereview.chromium.org/555133002/diff/120001/cc/output/gl_renderer_u...
File cc/output/gl_renderer_unittest.cc (right):

https://codereview.chromium.org/555133002/diff/120001/cc/output/gl_renderer_u...
cc/output/gl_renderer_unittest.cc:207:
EXPECT_EQ(renderer_->render_pass_program_[0][precision].program(),
Why 0? Shouldn't you test that every combination of blend mode and precision
compiles?

https://codereview.chromium.org/555133002/diff/120001/cc/output/program_bindi...
File cc/output/program_binding.h (right):

https://codereview.chromium.org/555133002/diff/120001/cc/output/program_bindi...
cc/output/program_binding.h:141: const VertexShader& vertex_shader() const {
return vertex_shader_; }
Why shadow the base class with these functions and member variables?

https://codereview.chromium.org/555133002/diff/120001/cc/output/shader.cc
File cc/output/shader.cc (right):

https://codereview.chromium.org/555133002/diff/120001/cc/output/shader.cc#new...
cc/output/shader.cc:943: shader_string.replace(
This is kind a bit sketchy.  Why can't your macro just stick these functions in
the right place in the same way that the precision macros do? Rather than
attempting to fix them up after the fact?

[rosca, 2014-09-25 10:50:40]

On 2014/09/24 22:07:56, enne wrote:
> This is a lot of new shader code.  Do the existing blending pixel tests cover
> all of these cases?  It looks like you're also touching render pass mask
> shaders, and I don't see any blending + mask tests.  Could you add some?

Thank you for comments! I will take them all into consideration.
I'm going to touch all the render pass shaders, so I'll add Layout tests and CC
pixel tests for each case.

With this patch I just wanted to hear your opinion on the approach of moving
blend modes into CC shaders.
If you are ok with this, I'm going to split the patch into smaller pieces and
send them to review.

[rosca, 2014-10-16 16:21:19]

Split into:
https://codereview.chromium.org/614153002/
https://codereview.chromium.org/658483003/.