Add PERSISTENT_MAP usage for GpuMemoryBuffers.

cc/trees/layer_tree_settings.h

Index: cc/trees/layer_tree_settings.h
diff --git a/cc/trees/layer_tree_settings.h b/cc/trees/layer_tree_settings.h
index 0e1819b0d9578e0ec7a4e93183d4fa6b2d3b40b0..ccc858ba84f6bff72bf4178ea653180f97182290 100644
--- a/cc/trees/layer_tree_settings.h
+++ b/cc/trees/layer_tree_settings.h
@@ -74,7 +74,8 @@ class CC_EXPORT LayerTreeSettings {
   bool use_one_copy;
   bool use_zero_copy;
   bool enable_elastic_overscroll;
-  unsigned use_image_texture_target;
+  unsigned map_image_texture_target;
+  unsigned persistent_map_image_texture_target;

[reveman, 2015/05/20 14:26:07]

I don't think this mechanism where we have a different setting field for each
configuration works long term. A setting field for each combination of format
and usage would explode. Options:

1. unsigned use_image_texture_target[gfx::GpuMemoryBuffer::FORMAT_LAST +
1][gfx::GpuMemoryBuffer::USAGE_LAST + 1];

2. unsigned use_image_texture_target; and the embedder needs to figure out what
target makes sense based on its usage of the compositor. e.g. use_one_copy=true
+ use_partial_updates=true means GetImageTextureTarget(BGRA_8888,
PERSISTENT_MAP).

The second option is basically what ToT is doing but BGRA_8888 is implicit. If
you want to limit the scope of this patch then this option might be better.

[danakj, 2015/05/20 17:51:14]

I agree. I'm just maintaining the "single target" that we have now, but also
allowing us to use partial texture update. I don't think we need anything more
complicated than that for this patch. I also think we need both because the
embedder is not deciding which mode the compositor is using. This is simply
renaming the "use_image_texture_target" and providing the same simple thing for
partial updates.

[danakj, 2015/05/20 17:55:11]

OK I'm going to go with option 2 here. This will mean you cannot turn on native
buffers without breaking partial tile updates, which I was trying to avoid for
you. But this seems contentious so I'll just leave us with 1 var for now
instead.

[reveman, 2015/05/20 18:58:00]

Sounds good. We're mostly interested in using native buffers for the renderer
compositor, which I'm expecting not to be using partial updates.

[reveman, 2015/05/20 18:58:00]

Not asking you to fix the format problem. I simply like us to not move in a
direction that is known to not work long term.

We should either allow just a single format+usage configuration to be used (like
ToT) or any configuration. map_/persisten_map_ puts us somewhere in between,
which I think is confusing and makes it hard for someone to know what to do if
they like to add another usage mode and/or format.

   bool ignore_root_layer_flings;
   size_t scheduled_raster_task_limit;
   bool use_occlusion_for_tile_prioritization;