GLES ProgramCache listens to memory pressure

gpu/command_buffer/common/constants.h

Index: gpu/command_buffer/common/constants.h
diff --git a/gpu/command_buffer/common/constants.h b/gpu/command_buffer/common/constants.h
index 7b80a33651160f7f7609b136d40eda7e344069a4..d132dcbb8c2c8eaf04c96adab5e96c35db3b1cca 100644
--- a/gpu/command_buffer/common/constants.h
+++ b/gpu/command_buffer/common/constants.h
@@ -8,6 +8,8 @@
 #include <stddef.h>
 #include <stdint.h>
 
+#include "build/build_config.h"
+
 namespace gpu {
 
 typedef int32_t CommandBufferOffset;
@@ -69,7 +71,11 @@ const int32_t kInvalidSharedMemoryId = -1;
 const int32_t kCommandBufferSharedMemoryId = 4;
 
 // The size to set for the program cache.
+#if !defined(OS_ANDROID)
 const size_t kDefaultMaxProgramCacheMemoryBytes = 6 * 1024 * 1024;
+#else
+const size_t kDefaultMaxProgramCacheMemoryBytes = 512 * 1024;

[vmiura, 2017/06/03 04:05:31]

Have we done any study of typical program cache usage?

I wonder if it would be worth using different limits if we're on low memory
Android vs default?

[ssid, 2017/06/05 19:42:45]

I looked at sizes of program cache on different websites and the sizes are in
range 4-10KiB. We could have tens of them in memory for each website.
As I said earlier we never really hit the 512KB limit on a single website in
general cases on Phones. So, 512KB can hold shaders for about 5-7 last viewed
websites. So, I reduced the limit further for low-end devices.

Are there any specific cases that we might regress on, like a shader intensive
website maybe? I could try and see if this causes any issue.

+#endif
 
 // Namespace used to separate various command buffer types.
 enum CommandBufferNamespace : int8_t {