gpu: play skp files through the command buffer

gpu/skia_runner/in_process_graphics_system.cc

+// Copyright (c) 2015 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "gpu/skia_runner/in_process_graphics_system.h"
+
+#include <iostream>
+
+#include "base/lazy_instance.h"
+#include "base/memory/discardable_memory.h"
+#include "base/memory/discardable_memory_allocator.h"
+#include "base/thread_task_runner_handle.h"
+#include "gpu/command_buffer/client/gles2_implementation.h"
+#include "gpu/command_buffer/client/gles2_lib.h"
+#include "gpu/skia_bindings/gl_bindings_skia_cmd_buffer.h"
+#include "third_party/khronos/GLES2/gl2.h"
+#include "third_party/skia/include/gpu/gl/GrGLInterface.h"
+
+namespace {
+
+// TODO(hendrikw): Replace TestDiscardableMemoryAllocator and move to base?
+class NonDiscardableMemory : public base::DiscardableMemory {
+ public:
+  NonDiscardableMemory(size_t size) : data_(size) {}

[piman, 2015/06/12 01:05:30]

nit: explicit.

[hendrikw, 2015/06/12 18:57:49]

Done.

+  bool Lock() override { return false; }
+  void Unlock() override {}
+  void* data() const override { return data_.data(); }
+
+ private:
+  mutable std::vector<uint8_t> data_;

[piman, 2015/06/12 01:05:30]

Since you don't resize the data, maybe you can use scoped_ptr<uint8_t[]>?
I'd also suggest using a const_cast<> in data() rather than using mutable here.

[hendrikw, 2015/06/12 18:57:49]

hm, I generally try to stay away from 'new' when an stl container handles it,
but in this case, I don't feel strongly one way or the other.  done.

+};
+
+class NonDiscardableMemoryAllocator : public base::DiscardableMemoryAllocator {
+ public:
+  // Overridden from DiscardableMemoryAllocator:
+  scoped_ptr<base::DiscardableMemory> AllocateLockedDiscardableMemory(
+      size_t size) override {
+    return make_scoped_ptr(new NonDiscardableMemory(size));
+  }
+};
+
+NonDiscardableMemoryAllocator g_discardable;

[piman, 2015/06/12 01:05:30]

nit: non-POD statics are disallowed by the style guide. You can use LazyInstance
for example. Or can you make it a field of InProcessGraphicsSystem?

[hendrikw, 2015/06/12 18:57:49]

Done.

+
+// Singleton used to initialize and terminate the gles2 library.
+class GLES2Initializer {
+ public:
+  GLES2Initializer() { gles2::Initialize(); }
+  ~GLES2Initializer() { gles2::Terminate(); }
+
+ private:
+  DISALLOW_COPY_AND_ASSIGN(GLES2Initializer);
+};
+
+base::LazyInstance<GLES2Initializer> g_gles2_initializer =
+    LAZY_INSTANCE_INITIALIZER;
+
+}  // namespace anonymous
+
+namespace skia_runner {
+
+void BindContext3DGLContextCallback(const GrGLInterface* gl_interface) {
+  gles2::SetGLContext(reinterpret_cast<gpu::GLInProcessContext*>(
+                          gl_interface->fCallbackData)->GetImplementation());
+}
+
+InProcessGraphicsSystem::InProcessGraphicsSystem() {
+  base::DiscardableMemoryAllocator::SetInstance(&g_discardable);
+  g_gles2_initializer.Get();
+
+  if (!gfx::GLSurface::InitializeOneOff()) {
+    LOG(ERROR) << "Unable to initialize gfx::GLSurface.";
+    return;
+  }
+
+  // Arg, we need to call base::ThreadTaskRunnerHandle::IsSet() to ensure that
+  // its lazy instance is instantiated before we create the GPU thread,
+  // otherwise shutdown order will delete the ThreadTaskRunnerHandle before the
+  // GPU thread's message loop, and when the message loop is shutdown, it will
+  // recreate ThreadTaskRunnerHandle, which will re-add a new task to the,
+  // AtExitManager, which causes a deadlock because it's already locked.
+  base::ThreadTaskRunnerHandle::IsSet();

[piman, 2015/06/12 01:05:30]

This is kinda unfortunate - it sounds like a data race... Could you fix it
closer to the source of the problem, i.e. in
gpu/command_buffer/service/in_process_command_buffer.cc? E.g. by making sure
this is called before g_default_service.Get() is called, possibly by introducing
an intermediate function called from the service_ initializer in the
InProcessCommandBuffer constructor.

+
+  // Create the in process context.
+  in_process_context_ = CreateInProcessContext();
+  if (!in_process_context_) {
+    LOG(ERROR) << "Unable to create in process context.";
+    return;
+  }
+
+  // Create and set up skia's GL bindings.
+  gles2::SetGLContext(in_process_context_->GetImplementation());
+  GrGLInterface* bindings = skia_bindings::CreateCommandBufferSkiaGLBinding();
+  if (!bindings) {
+    LOG(ERROR) << "Unable to create skia command buffer bindings.";
+    return;
+  }
+
+  bindings->fCallback = BindContext3DGLContextCallback;
+  bindings->fCallbackData =
+      reinterpret_cast<GrGLInterfaceCallbackData>(in_process_context_.get());
+
+  // Create skia's graphics context.
+  gr_context_ = skia::AdoptRef(GrContext::Create(
+      kOpenGL_GrBackend, reinterpret_cast<GrBackendContext>(bindings)));
+
+  if (!gr_context_) {
+    LOG(ERROR) << "Unable to create skia graphics context.";
+    return;
+  }
+
+  // Set skia's graphics context cache size.
+  const int kMaxGaneshResourceCacheCount = 2048;
+  const size_t kMaxGaneshResourceCacheBytes = 96 * 1024 * 1024;
+  gr_context_->setResourceCacheLimits(kMaxGaneshResourceCacheCount,
+                                      kMaxGaneshResourceCacheBytes);
+}
+
+InProcessGraphicsSystem::~InProcessGraphicsSystem() {
+}
+
+bool InProcessGraphicsSystem::IsSuccessfullyInitialized() const {
+  return gr_context_;
+}
+
+int InProcessGraphicsSystem::GetMaxTextureSize() const {
+  int max_texture_size = 0;
+  in_process_context_->GetImplementation()->GetIntegerv(GL_MAX_TEXTURE_SIZE,
+                                                        &max_texture_size);
+  return max_texture_size;
+}
+
+scoped_ptr<gpu::GLInProcessContext>
+InProcessGraphicsSystem::CreateInProcessContext() const {
+  const bool is_offscreen = true;
+  const bool share_resources = true;
+  gpu::gles2::ContextCreationAttribHelper attribs;
+  attribs.alpha_size = 8;
+  attribs.blue_size = 8;
+  attribs.green_size = 8;
+  attribs.red_size = 8;
+  attribs.depth_size = 24;
+  attribs.stencil_size = 8;
+  attribs.samples = 0;
+  attribs.sample_buffers = 0;
+  attribs.fail_if_major_perf_caveat = false;
+  attribs.bind_generates_resource = false;
+  gfx::GpuPreference gpu_preference = gfx::PreferDiscreteGpu;
+
+  scoped_ptr<gpu::GLInProcessContext> context =
+      make_scoped_ptr(gpu::GLInProcessContext::Create(
+          nullptr, nullptr, is_offscreen, gfx::kNullAcceleratedWidget,
+          gfx::Size(1, 1), nullptr, share_resources, attribs, gpu_preference,
+          gpu::GLInProcessContextSharedMemoryLimits(), nullptr, nullptr));
+
+  DCHECK(context);
+  return context.Pass();
+}
+
+}  // namespace skia_runner