Mozart: Improve tracing and backpressure.

services/gfx/compositor/backend/gpu_output.cc

 // Copyright 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 "services/gfx/compositor/backend/gpu_output.h"
 
 #include <utility>
 
 #include "base/bind.h"
+#include "base/command_line.h"
 #include "base/location.h"
 #include "base/logging.h"
 #include "base/message_loop/message_loop.h"
+#include "base/strings/string_number_conversions.h"
 #include "base/trace_event/trace_event.h"
-#include "services/gfx/compositor/backend/gpu_rasterizer.h"
 #include "services/gfx/compositor/render/render_frame.h"
 
 namespace compositor {
 namespace {
-// Maximum number of frames to hold in the queue for rendering.
-constexpr size_t kMaxPipelineDepth = 1;
-}
-
-template <typename T>
-static void Drop(scoped_ptr<T> ptr) {}
-
-static scoped_ptr<base::MessagePump> CreateMessagePumpMojo() {
+constexpr const char* kPipelineDepthSwitch = "pipeline-depth";
+constexpr uint32_t kDefaultPipelineDepth = 2u;  // ideally should be 1
+constexpr uint32_t kMinPipelineDepth = 1u;
+constexpr uint32_t kMaxPipelineDepth = 10u;

[mikejurka, 2016/05/19 23:26:18]

why should we ever be more than 3 frames ahead?

[jeffbrown, 2016/05/20 01:09:47]

This is sort of arbitrary.  The limit only exists for experimentation.  A depth
of 10 would be ludicrous!

+
+scoped_ptr<base::MessagePump> CreateMessagePumpMojo() {
   return base::MessageLoop::CreateMessagePumpForType(
       base::MessageLoop::TYPE_DEFAULT);
 }
+}  // namespace
 
 GpuOutput::GpuOutput(
     mojo::InterfaceHandle<mojo::ContextProvider> context_provider,
     const SchedulerCallbacks& scheduler_callbacks,
     const base::Closure& error_callback)
-    : scheduler_(new VsyncScheduler(base::MessageLoop::current()->task_runner(),
-                                    scheduler_callbacks)),
-      rasterizer_delegate_(make_scoped_ptr(new RasterizerDelegate())) {
+    : compositor_task_runner_(base::MessageLoop::current()->task_runner()),
+      vsync_scheduler_(
+          new VsyncScheduler(compositor_task_runner_, scheduler_callbacks)),
+      rasterizer_thread_(new base::Thread("gpu_rasterizer")),
+      rasterizer_initialized_(true, false) {
   DCHECK(context_provider);
 
-  rasterizer_delegate_->PostInitialize(
-      std::move(context_provider), scheduler_,
-      base::MessageLoop::current()->task_runner(), error_callback);
-}
-
-GpuOutput::~GpuOutput() {
-  // Ensure destruction happens on the correct thread.
-  rasterizer_delegate_->PostDestroy(rasterizer_delegate_.Pass());
-}
-
-Scheduler* GpuOutput::GetScheduler() {
-  return scheduler_.get();
-}
-
-void GpuOutput::SubmitFrame(const scoped_refptr<RenderFrame>& frame) {
-  rasterizer_delegate_->PostFrame(frame);
-}
-
-GpuOutput::RasterizerDelegate::RasterizerDelegate() {
+  pipeline_depth_ = kDefaultPipelineDepth;
+  auto command_line = base::CommandLine::ForCurrentProcess();
+  if (command_line->HasSwitch(kPipelineDepthSwitch)) {
+    std::string str(command_line->GetSwitchValueASCII(kPipelineDepthSwitch));
+    unsigned value;
+    if (base::StringToUint(str, &value) && value >= kMinPipelineDepth &&
+        value <= kMaxPipelineDepth) {
+      pipeline_depth_ = value;
+    } else {
+      LOG(ERROR) << "Invalid value for --" << kPipelineDepthSwitch << ": \""
+                 << str << "\"";

[mikejurka, 2016/05/19 23:26:18]

is there a way to exit/crash here? should we call PostErrorCallback() here?

[jeffbrown, 2016/05/20 01:09:47]

Hmm.  Not a bad idea.

+    }
+  }
+  DVLOG(2) << "Using pipeline depth " << pipeline_depth_;
+
   base::Thread::Options options;
   options.message_pump_factory = base::Bind(&CreateMessagePumpMojo);
-
-  thread_.reset(new base::Thread("gpu_rasterizer"));
-  thread_->StartWithOptions(options);
-  task_runner_ = thread_->message_loop()->task_runner();
-}
-
-GpuOutput::RasterizerDelegate::~RasterizerDelegate() {}
-
-void GpuOutput::RasterizerDelegate::PostInitialize(
-    mojo::InterfaceHandle<mojo::ContextProvider> context_provider,
-    const scoped_refptr<VsyncScheduler>& scheduler,
-    const scoped_refptr<base::TaskRunner>& task_runner,
-    const base::Closure& error_callback) {
-  task_runner_->PostTask(
+  rasterizer_thread_->StartWithOptions(options);
+  rasterizer_task_runner_ = rasterizer_thread_->message_loop()->task_runner();
+  rasterizer_task_runner_->PostTask(
       FROM_HERE,
-      base::Bind(&RasterizerDelegate::InitializeTask, base::Unretained(this),
-                 base::Passed(std::move(context_provider)), scheduler,
-                 base::MessageLoop::current()->task_runner(), error_callback));
-}
-
-void GpuOutput::RasterizerDelegate::PostDestroy(
-    scoped_ptr<RasterizerDelegate> self) {
-  task_runner_->PostTask(
-      FROM_HERE, base::Bind(&Drop<RasterizerDelegate>, base::Passed(&self)));
-}
-
-void GpuOutput::RasterizerDelegate::PostFrame(
-    const scoped_refptr<RenderFrame>& frame) {
-  bool was_empty;
-  scoped_refptr<RenderFrame> dropped_frame;
-  {
-    std::lock_guard<std::mutex> lock(mutex_);
-    was_empty = frames_.empty();
-    if (frames_.size() == kMaxPipelineDepth) {
-      // TODO(jeffbrown): Adjust scheduler behavior to compensate.
-      LOG(ERROR) << "Renderer pipeline stalled, dropping a frame to catch up.";
-      dropped_frame = frames_.front();  // drop an old frame outside the lock
-      frames_.pop();
+      base::Bind(&GpuOutput::InitializeRasterizer, base::Unretained(this),
+                 base::Passed(std::move(context_provider))));
+  rasterizer_initialized_.Wait();
+  DCHECK(rasterizer_);
+}
+
+GpuOutput::~GpuOutput() {
+  // Ensure rasterizer destruction happens on the rasterizer thread.
+  rasterizer_task_runner_->PostTask(
+      FROM_HERE,
+      base::Bind(&GpuOutput::DestroyRasterizer, base::Unretained(this)));
+  rasterizer_thread_->Stop();

[mikejurka, 2016/05/19 23:26:18]

does ->Stop() block until the thread is done?

[jeffbrown, 2016/05/20 01:09:47]

Yes.

+  DCHECK(!rasterizer_);
+}
+
+Scheduler* GpuOutput::GetScheduler() {
+  return vsync_scheduler_.get();
+}
+
+void GpuOutput::SubmitFrame(const scoped_refptr<RenderFrame>& frame) {
+  DCHECK(frame);
+  TRACE_EVENT0("gfx", "GpuOutput::SubmitFrame");
+
+  const int64_t submit_time = MojoGetTimeTicksNow();
+  scoped_refptr<FrameData> frame_data(

[mikejurka, 2016/05/19 23:26:18]

is there a way we can avoid using a scoped_refptr here (using scoped_ptr
instead) ? 

also, like mentioned before, FrameData already inherits from
RefCountedThreadSafe -- why do we need to wrap it again?

[jeffbrown, 2016/05/20 01:09:47]

These two things go together.  There's no wrapping involved.

+      new FrameData(frame, submit_time));  // drop outside lock
+  {
+    std::lock_guard<std::mutex> lock(shared_state_.mutex);
+
+    TRACE_EVENT_FLOW_BEGIN0("gfx", "Frame Queued", frame_data.get());
+    shared_state_.current_frame_data.swap(frame_data);
+    if (frame_data && !frame_data->draw_time) {

[mikejurka, 2016/05/19 23:26:18]

what does it mean if draw_time is 0? does it mean this frame was not drawn? a
comment would be helpful.

[jeffbrown, 2016/05/20 01:09:47]

Added a comment to the header file.

+      // Dropped an undrawn frame.
+      DVLOG(2) << "Rasterizer stalled, dropping frame to catch up.";
+      TRACE_EVENT_FLOW_END1("gfx", "Frame Queued", frame_data.get(), "drawn",
+                            false);
     }
-    frames_.push(frame);
-  }
-
-  if (was_empty)
-    PostSubmit();
-}
-
-void GpuOutput::RasterizerDelegate::PostSubmit() {
-  TRACE_EVENT0("gfx", "GpuOutput::RasterizerDelegate::PostSubmit");
-  task_runner_->PostTask(FROM_HERE, base::Bind(&RasterizerDelegate::SubmitTask,
-                                               base::Unretained(this)));
-}
-
-void GpuOutput::RasterizerDelegate::InitializeTask(
-    mojo::InterfaceHandle<mojo::ContextProvider> context_provider,
-    const scoped_refptr<VsyncScheduler>& scheduler,
-    const scoped_refptr<base::TaskRunner>& task_runner,
-    const base::Closure& error_callback) {
+
+    // TODO(jeffbrown): If the draw queue is full, we should pause
+    // scheduling until the queue drains.
+    if (shared_state_.rasterizer_ready &&
+        shared_state_.drawn_frames_awaiting_finish.size() < pipeline_depth_)
+      ScheduleDrawLocked();
+  }
+}
+
+void GpuOutput::OnRasterizerReady(int64_t vsync_timebase,
+                                  int64_t vsync_interval) {
+  DCHECK(rasterizer_task_runner_->RunsTasksOnCurrentThread());
+
+  // TODO(jeffbrown): This shouldn't be hardcoded.
+  // Need to do some real tuning and possibly determine values adaptively.
+  // We should probably split the Start() method in two to separate the
+  // process of setting parameters from starting / stopping scheduling.
+  const int64_t update_phase = -vsync_interval;
+  const int64_t snapshot_phase = -vsync_interval / 6;
+  // TODO(jeffbrown): Determine the presentation phase based on queue depth.
+  const int64_t presentation_phase = vsync_interval * pipeline_depth_;
+  if (!vsync_scheduler_->Start(vsync_timebase, vsync_interval, update_phase,
+                               snapshot_phase, presentation_phase)) {
+    LOG(ERROR) << "Received invalid vsync parameters: timebase="
+               << vsync_timebase << ", interval=" << vsync_interval;
+    PostErrorCallback();
+    return;
+  }
+
+  {
+    std::lock_guard<std::mutex> lock(shared_state_.mutex);
+
+    if (shared_state_.rasterizer_ready)
+      return;
+
+    DCHECK(shared_state_.drawn_frames_awaiting_finish.empty());
+    shared_state_.rasterizer_ready = true;
+
+    if (!shared_state_.current_frame_data)
+      return;
+
+    shared_state_.current_frame_data->Recycle();
+    TRACE_EVENT_FLOW_BEGIN0("gfx", "Frame Queued",
+                            shared_state_.current_frame_data.get());
+    ScheduleDrawLocked();
+  }
+}
+
+void GpuOutput::OnRasterizerSuspended() {
+  DCHECK(rasterizer_task_runner_->RunsTasksOnCurrentThread());
+
+  vsync_scheduler_->Stop();
+
+  {
+    std::lock_guard<std::mutex> lock(shared_state_.mutex);
+
+    if (!shared_state_.rasterizer_ready)
+      return;
+
+    shared_state_.rasterizer_ready = false;
+  }
+}
+
+void GpuOutput::OnRasterizerError() {
+  DCHECK(rasterizer_task_runner_->RunsTasksOnCurrentThread());
+
+  PostErrorCallback();
+}
+
+void GpuOutput::ScheduleDrawLocked() {
+  DCHECK(shared_state_.current_frame_data);
+  DCHECK(!shared_state_.current_frame_data->draw_time);
+
+  if (shared_state_.draw_scheduled)
+    return;
+
+  shared_state_.draw_scheduled = true;
+  rasterizer_task_runner_->PostTask(
+      FROM_HERE, base::Bind(&GpuOutput::OnDraw, base::Unretained(this)));
+}
+
+void GpuOutput::OnDraw() {
+  DCHECK(rasterizer_task_runner_->RunsTasksOnCurrentThread());
+  TRACE_EVENT0("gfx", "GpuOutput::OnDraw");
+
+  scoped_refptr<FrameData> frame_data;  // used outside lock
+  {
+    std::lock_guard<std::mutex> lock(shared_state_.mutex);
+
+    DCHECK(shared_state_.draw_scheduled);
+    DCHECK(shared_state_.current_frame_data);
+    DCHECK(!shared_state_.current_frame_data->draw_time);
+
+    shared_state_.draw_scheduled = false;
+
+    if (!shared_state_.rasterizer_ready ||
+        shared_state_.drawn_frames_awaiting_finish.size() >= pipeline_depth_)
+      return;
+
+    frame_data = shared_state_.current_frame_data;
+    frame_data->draw_time = MojoGetTimeTicksNow();
+    TRACE_EVENT_FLOW_END1("gfx", "Frame Queued", frame_data.get(), "drawn",
+                          true);
+
+    TRACE_EVENT_ASYNC_BEGIN0("gfx", "Rasterize", frame_data.get());
+    shared_state_.drawn_frames_awaiting_finish.emplace(frame_data);
+  }
+
+  rasterizer_->DrawFrame(frame_data->frame);
+  frame_data->wait_time = MojoGetTimeTicksNow();
+}
+
+void GpuOutput::OnRasterizerFinishedDraw(bool presented) {
+  DCHECK(rasterizer_task_runner_->RunsTasksOnCurrentThread());
+  TRACE_EVENT0("gfx", "GpuOutput::OnRasterizerFinishedDraw");
+
+  const int64_t finish_time = MojoGetTimeTicksNow();
+  scoped_refptr<FrameData> frame_data;  // drop outside lock
+  {
+    std::lock_guard<std::mutex> lock(shared_state_.mutex);
+
+    DCHECK(shared_state_.rasterizer_ready);
+    DCHECK(!shared_state_.drawn_frames_awaiting_finish.empty());
+    size_t draw_queue_depth = shared_state_.drawn_frames_awaiting_finish.size();
+    shared_state_.drawn_frames_awaiting_finish.front().swap(frame_data);
+    shared_state_.drawn_frames_awaiting_finish.pop();
+    DCHECK(frame_data);
+    DCHECK(frame_data->draw_time);
+    TRACE_EVENT_ASYNC_END1("gfx", "Rasterize", frame_data.get(), "presented",
+                           presented);
+
+    // TODO(jeffbrown): Adjust scheduler behavior based on observed timing.
+    // Note: These measurements don't account for systematic downstream delay
+    // in the display pipeline (how long it takes pixels to actually light up).
+    if (presented) {
+      const RenderFrame::Metadata& frame_metadata =
+          frame_data->frame->metadata();
+      const mojo::gfx::composition::FrameInfo& frame_info =
+          frame_metadata.frame_info();
+      const int64_t frame_time = frame_info.frame_time;
+      const int64_t presentation_time = frame_info.presentation_time;
+      const int64_t composition_time = frame_metadata.composition_time();
+      const int64_t draw_time = frame_data->draw_time;
+      const int64_t wait_time = frame_data->wait_time;
+      const int64_t submit_time = frame_data->submit_time;
+
+      DVLOG(2) << "Presented frame: composition latency "
+               << (composition_time - frame_time) << " us, submission latency "
+               << (submit_time - composition_time) << " us, queue latency "
+               << (draw_time - submit_time) << " us, draw latency "
+               << (wait_time - draw_time) << " us, GPU latency "
+               << (finish_time - wait_time) << " us, total latency "
+               << (finish_time - frame_time) << " us, presentation time error "
+               << (finish_time - presentation_time) << " us"
+               << ", draw queue depth " << draw_queue_depth;
+    } else {
+      DVLOG(2) << "Rasterizer dropped frame.";
+    }
+
+    DCHECK(shared_state_.current_frame_data);
+    if (!shared_state_.current_frame_data->draw_time)

[mikejurka, 2016/05/19 23:26:18]

again, what does it mean if draw_time is 0? how does that differ from
"presented"?

[jeffbrown, 2016/05/20 01:09:47]

It's set in OnDraw() above.  Basically means DrawFrame was called.  It seemed
expedient to rely on the fact that time ticks are essentially always positive
after boot but I guess it wouldn't hurt to make the flag explicit so I'll do
that for clarity.

+      ScheduleDrawLocked();
+  }
+}
+
+void GpuOutput::InitializeRasterizer(
+    mojo::InterfaceHandle<mojo::ContextProvider> context_provider) {
+  DCHECK(rasterizer_task_runner_->RunsTasksOnCurrentThread());
+  DCHECK(!rasterizer_);
+  TRACE_EVENT0("gfx", "GpuOutput::InitializeRasterizer");
+
   rasterizer_.reset(new GpuRasterizer(
-      mojo::ContextProviderPtr::Create(std::move(context_provider)), scheduler,
-      task_runner, error_callback));
-}
-
-void GpuOutput::RasterizerDelegate::SubmitTask() {
-  TRACE_EVENT0("gfx", "GpuOutput::RasterizerDelegate::SubmitTask");
-  bool have_more;
-  scoped_refptr<RenderFrame> frame;
-  {
-    std::lock_guard<std::mutex> lock(mutex_);
-    DCHECK(!frames_.empty());
-    frame = frames_.front();
-    frames_.pop();
-    have_more = !frames_.empty();
-  }
-
-  if (have_more)
-    PostSubmit();
-
-  int64_t submit_time = MojoGetTimeTicksNow();
-  rasterizer_->SubmitFrame(
-      frame, base::Bind(&RasterizerDelegate::OnFrameSubmitted,
-                        base::Unretained(this), frame->frame_info().frame_time,
-                        frame->frame_info().presentation_time, submit_time));
-}
-
-void GpuOutput::RasterizerDelegate::OnFrameSubmitted(int64_t frame_time,
-                                                     int64_t presentation_time,
-                                                     int64_t submit_time,
-                                                     bool presented) {
-  TRACE_EVENT0("gfx", "GpuOutput::RasterizerDelegate::OnFrameSubmitted");
-  // TODO(jeffbrown): Adjust scheduler behavior based on observed timing.
-  // Note: These measurements don't account for systematic downstream delay
-  // in the display pipeline (how long it takes pixels to actually light up).
-  int64_t complete_time = MojoGetTimeTicksNow();
-  if (presented) {
-    DVLOG(3) << "Frame presented: submission latency "
-             << (submit_time - frame_time) << " us, rasterization latency "
-             << (complete_time - submit_time) << " us, total latency "
-             << (complete_time - frame_time) << " us, presentation time error "
-             << (complete_time - presentation_time);
-  } else {
-    DVLOG(3) << "Frame deferred.";
-  }
+      mojo::ContextProviderPtr::Create(std::move(context_provider)), this));
+  rasterizer_initialized_.Signal();
+}
+
+void GpuOutput::DestroyRasterizer() {
+  DCHECK(rasterizer_task_runner_->RunsTasksOnCurrentThread());
+  DCHECK(rasterizer_);
+  TRACE_EVENT0("gfx", "GpuOutput::DestroyRasterizer");
+
+  rasterizer_.reset();
+  rasterizer_initialized_.Reset();
+}
+
+void GpuOutput::PostErrorCallback() {
+  compositor_task_runner_->PostTask(FROM_HERE, error_callback_);
+}
+
+GpuOutput::FrameData::FrameData(const scoped_refptr<RenderFrame>& frame,
+                                int64_t submit_time)
+    : frame(frame), submit_time(submit_time), draw_time(0), wait_time(0) {}

[mikejurka, 2016/05/20 01:49:23]

shouldn't we still initialize these guys to 0?

[jeffbrown, 2016/05/20 19:31:35]

It's done in the header file.

+
+GpuOutput::FrameData::~FrameData() {}
+
+void GpuOutput::FrameData::Recycle() {
+  draw_time = 0;
+  wait_time = 0;
 }
 
 }  // namespace compositor