| Index: gpu/command_buffer/client/gles2_implementation.cc
|
| diff --git a/gpu/command_buffer/client/gles2_implementation.cc b/gpu/command_buffer/client/gles2_implementation.cc
|
| index 22032b0147c4d3a9ba33a19a6373b8bf1327ff28..e3f765103f560c260b27b3565cdc9b0424e606e3 100644
|
| --- a/gpu/command_buffer/client/gles2_implementation.cc
|
| +++ b/gpu/command_buffer/client/gles2_implementation.cc
|
| @@ -32,6 +32,7 @@
|
| #include "gpu/command_buffer/client/gpu_control.h"
|
| #include "gpu/command_buffer/client/program_info_manager.h"
|
| #include "gpu/command_buffer/client/query_tracker.h"
|
| +#include "gpu/command_buffer/client/shared_memory_limits.h"
|
| #include "gpu/command_buffer/client/transfer_buffer.h"
|
| #include "gpu/command_buffer/client/vertex_array_object_manager.h"
|
| #include "gpu/command_buffer/common/gles2_cmd_utils.h"
|
| @@ -215,7 +216,7 @@ bool GLES2Implementation::Initialize(
|
| mapped_memory_.reset(new MappedMemoryManager(helper_, mapped_memory_limit));
|
|
|
| unsigned chunk_size = 2 * 1024 * 1024;
|
| - if (mapped_memory_limit != kNoLimit) {
|
| + if (mapped_memory_limit != SharedMemoryLimits::kNoLimit) {
|
| // Use smaller chunks if the client is very memory conscientious.
|
| chunk_size = std::min(mapped_memory_limit / 4, chunk_size);
|
| }
|
|
|