Fixing laggy chrome on a multitude of accelerated canvases

third_party/WebKit/Source/core/html/HTMLCanvasElement.cpp

Index: third_party/WebKit/Source/core/html/HTMLCanvasElement.cpp
diff --git a/third_party/WebKit/Source/core/html/HTMLCanvasElement.cpp b/third_party/WebKit/Source/core/html/HTMLCanvasElement.cpp
index 3b5b5c4d0c1f043cee3d8dff903fc75d1bff9dfe..e80957a7740de3e4abc1a9d3561315107e91e774 100644
--- a/third_party/WebKit/Source/core/html/HTMLCanvasElement.cpp
+++ b/third_party/WebKit/Source/core/html/HTMLCanvasElement.cpp
@@ -80,6 +80,20 @@ const int MaxCanvasArea = 32768 * 8192; // Maximum canvas area in CSS pixels
 // In Skia, we will also limit width/height to 32767.
 const int MaxSkiaDim = 32767; // Maximum width/height in CSS pixels.
 
+// We estimate the max limit of GPU allocated memory for canvases before Chrome becomes laggy by setting the

[Justin Novosad, 2015/12/08 22:15:27]

There is no hard rule for this in Blink, but we generally respect 80col for
multi-line comments

+// total allocated memory for accelerated canvases to be equivalent to memory used by 80 accelerated
+// canvases, each has a size of 1000*500 and 2d context.
+// Each such canvas occupies 4000000 = 1000 * 500 * 2 * 4 bytes, where 2 is the gpuBufferCount in
+// ImageBuffer::updateGPUMemoryUsage() and 4 means four bytes per pixel per buffer.
+#if !OS(ANDROID)
+const int MaxTotalGPUMemoryUsage = 4000000 * 80;
+#else
+// We estimate that the max limit for android phones is half of that for desktops based on local
+// experimental results on Android One; though a very slight lagginess on Android One is still expected,
+// other Android devices should work fine.
+const int MaxTotalGPUMemoryUsage = 4000000 * 40;

[Justin Novosad, 2015/12/08 22:15:27]

This limit feels excessive

+#endif
+
 // A default value of quality argument for toDataURL and toBlob
 // It is in an invalid range (outside 0.0 - 1.0) so that it will not be misinterpreted as a user-input value
 const int UndefinedQualityValue = -1.0;
@@ -614,6 +628,12 @@ bool HTMLCanvasElement::shouldAccelerate(const IntSize& size) const
     if (!Platform::current()->canAccelerate2dCanvas())
         return false;
 
+    // When GPU allocated memory runs low (due to having created too many accelerated canvases), the
+    // compositor starves and browser becomes laggy. Thus, we should stop allocating more GPU memory to

[Justin Novosad, 2015/12/08 22:15:27]

wrapping

+    // new canvases created when the current memory usage exceeds the threshold.
+    if (ImageBuffer::getTotalGPUMemoryUsage() >= MaxTotalGPUMemoryUsage)
+        return false;
+
     return true;
 }
 
@@ -762,7 +782,6 @@ void HTMLCanvasElement::updateExternallyAllocatedMemory() const
     if (m_copiedImage)
         bufferCount++;
 
-    // Four bytes per pixel per buffer.

[Justin Novosad, 2015/12/08 22:15:27]

Why remove this comment?

     Checked<intptr_t, RecordOverflow> checkedExternallyAllocatedMemory = 4 * bufferCount;
     if (is3D())
         checkedExternallyAllocatedMemory += m_context->externallyAllocatedBytesPerPixel();