cc: optimize texture allocation.

cc/output/gl_renderer.cc

Index: cc/output/gl_renderer.cc
diff --git a/cc/output/gl_renderer.cc b/cc/output/gl_renderer.cc
index 6a75674271e6921f2f0658f281f12aa86425776e..d8af04ee1b2d82ab307e695bffc706c015d58ed9 100644
--- a/cc/output/gl_renderer.cc
+++ b/cc/output/gl_renderer.cc
@@ -141,6 +141,16 @@ BlendMode BlendModeFromSkXfermode(SkXfermode::Mode mode) {
   }
 }
 
+void RoundUpToPow2(gfx::RectF* rect) {
+  int w, h;

[danakj, 2016/05/04 21:24:58]

you're using ints here but it's a RectF.

[Stephen White, 2016/05/04 21:27:31]

Yep. There are implicit casts in the comparisons and the call to set_width() and
set_height(), but the multiplies are faster. :)

+  for (w = 1; w < rect->width(); w *= 2) {

[enne (OOO), 2016/05/04 21:32:22]

If you're casting to ints here for performance, might as well go all the way and
https://graphics.stanford.edu/~seander/bithacks.html#RoundUpPowerOf2 instead of
a loop.

[Stephen White, 2016/05/04 21:47:17]

Cute! I don't really care about performance here; I was just being annoying.
I've changed them to floats for clarity.

+  }
+  for (h = 1; h < rect->height(); h *= 2) {
+  }
+  rect->set_width(w);
+  rect->set_height(h);
+}
+
 // Smallest unit that impact anti-aliasing output. We use this to
 // determine when anti-aliasing is unnecessary.
 const float kAntiAliasingEpsilon = 1.0f / 1024.0f;
@@ -1032,6 +1042,9 @@ void GLRenderer::DrawRenderPassQuad(DrawingFrame* frame,
         if (dst_rect.IsEmpty()) {
           return;
         }
+        // Expand dst_rect size to the nearest power of 2, in order to get

[enne (OOO), 2016/05/04 21:32:22]

Are there limits on the size that is supported here? Like, very big textures now
becoming ~2-4 times as big?

[Stephen White, 2016/05/04 21:47:17]

If you mean, will the allocation fail due to an invalid texture dimension, I'm
pretty sure all GL texture limits are a power of 2, so no.

For overall size, these textures count against Skia's texture budget, so they
may cause other scratch textures to be evicted or reused, but they shouldn't
increase the VRAM footprint on balance.

(Note that the internal textures allocated for filter processing in Skia are
already rounded up the same way.)

+        // more cache hits in Skia's texture cache.
+        RoundUpToPow2(&dst_rect);
         filter_image = ApplyImageFilter(
             ScopedUseGrContext::Create(this, frame), resource_provider_, rect,
             dst_rect, scale, std::move(filter), contents_texture);