Added support for GPU Tracing on mobile devices which support it.

gpu/command_buffer/service/gpu_tracer.cc

Index: gpu/command_buffer/service/gpu_tracer.cc
diff --git a/gpu/command_buffer/service/gpu_tracer.cc b/gpu/command_buffer/service/gpu_tracer.cc
index b1c9f23e4858f5a68ac1c1e3b161c956c02d56ee..bd9bde0db41926eaaa60ca2f74ae21904d426c31 100644
--- a/gpu/command_buffer/service/gpu_tracer.cc
+++ b/gpu/command_buffer/service/gpu_tracer.cc
@@ -59,45 +59,75 @@ void TraceOutputter::Trace(const std::string& name,
   ++local_trace_id_;
 }
 
-GPUTrace::GPUTrace(const std::string& name)
-    : name_(name),
-      outputter_(NULL),
-      offset_(0),
-      end_time_(0),
-      end_requested_(false),
-      enabled_(false) {
-}
-
 GPUTrace::GPUTrace(scoped_refptr<Outputter> outputter,
                    const std::string& name,
-                   int64 offset)
+                   int64 offset,
+                   GpuTracerType tracer_type)
     : name_(name),
       outputter_(outputter),
       offset_(offset),
       start_time_(0),
       end_time_(0),
+      tracer_type_(tracer_type),
       end_requested_(false),
-      enabled_(true) {
-  glGenQueries(2, queries_);
+      discard_trace_(false) {
+  memset(queries_, 0, sizeof(queries_));
+  switch (tracer_type_) {
+    case kTracerTypeARBTimer:
+      glGenQueries(2, queries_);
+      break;
+    case kTracerTypeDisjointTimer:
+      glGenQueriesEXT(2, queries_);
+      break;
+
+    default:
+      tracer_type_ = kTracerTypeInvalid;
+  }
 }
 
 GPUTrace::~GPUTrace() {
-  if (enabled_)
-    glDeleteQueries(2, queries_);
+  switch (tracer_type_) {
+    case kTracerTypeInvalid:
+      break;
+
+    case kTracerTypeARBTimer:
+      glDeleteQueries(2, queries_);
+      break;
+    case kTracerTypeDisjointTimer:
+      glDeleteQueriesEXT(2, queries_);
+      break;
+  }
 }
 
 void GPUTrace::Start() {
   TRACE_EVENT_COPY_ASYNC_BEGIN0(
       TRACE_DISABLED_BY_DEFAULT("gpu.service"), name().c_str(), this);
-  if (enabled_) {
-    glQueryCounter(queries_[0], GL_TIMESTAMP);
+
+  switch (tracer_type_) {
+    case kTracerTypeInvalid:
+      break;
+
+    case kTracerTypeARBTimer:
+      glQueryCounter(queries_[0], GL_TIMESTAMP);
+      break;
+    case kTracerTypeDisjointTimer:
+      glQueryCounterEXT(queries_[0], GL_TIMESTAMP_EXT);
+      break;
   }
 }
 
 void GPUTrace::End() {
-  if (enabled_) {
-    glQueryCounter(queries_[1], GL_TIMESTAMP);
-    end_requested_ = true;
+  end_requested_ = true;
+  switch (tracer_type_) {
+    case kTracerTypeInvalid:
+      break;
+
+    case kTracerTypeARBTimer:
+      glQueryCounter(queries_[1], GL_TIMESTAMP);
+      break;
+    case kTracerTypeDisjointTimer:
+      glQueryCounterEXT(queries_[1], GL_TIMESTAMP_EXT);
+      break;
   }
 
   TRACE_EVENT_COPY_ASYNC_END0(
@@ -105,37 +135,75 @@ void GPUTrace::End() {
 }
 
 bool GPUTrace::IsAvailable() {
-  if (!enabled_)
-    return true;
-  else if (!end_requested_)
-    return false;
+  if (tracer_type_ != kTracerTypeInvalid) {
+    if (!end_requested_)
+      return false;
+    else if (discard_trace_)
+      return true;
+
+    GLint done = 0;
+
+    switch (tracer_type_) {
+      case kTracerTypeARBTimer:
+        glGetQueryObjectiv(queries_[1], GL_QUERY_RESULT_AVAILABLE, &done);
+        break;
+
+      case kTracerTypeDisjointTimer:
+        glGetQueryObjectivEXT(queries_[1], GL_QUERY_RESULT_AVAILABLE_EXT,
+                              &done);
+        break;
+
+      default:
+        done = 1;
+        break;
+    }
+
+    return !!done;
+  }
 
-  GLint done = 0;
-  glGetQueryObjectiv(queries_[1], GL_QUERY_RESULT_AVAILABLE, &done);
-  return !!done;
+  return true;
 }
 
 void GPUTrace::Process() {
-  if (!enabled_)
+  if (tracer_type_ == kTracerTypeInvalid)
     return;
 
   DCHECK(IsAvailable());
 
-  GLuint64 timestamp;
+  GLuint64 beg_stamp = 0;
+  GLuint64 end_stamp = 0;
 
   // TODO(dsinclair): It's possible for the timer to wrap during the start/end.
   // We need to detect if the end is less then the start and correct for the
   // wrapping.
-  glGetQueryObjectui64v(queries_[0], GL_QUERY_RESULT, &timestamp);
-  start_time_ = (timestamp / base::Time::kNanosecondsPerMicrosecond) + offset_;
+  switch (tracer_type_) {
+    case kTracerTypeARBTimer:
+      glGetQueryObjectui64v(queries_[0], GL_QUERY_RESULT, &beg_stamp);
+      glGetQueryObjectui64v(queries_[1], GL_QUERY_RESULT, &end_stamp);
+      glDeleteQueries(2, queries_);

[vmiura, 2014/08/26 23:47:26]

~GPUTrace deletes the queries too so perhaps we should remove these deletes.

[David Yen, 2014/08/27 19:47:45]

Done.

+      break;
+    case kTracerTypeDisjointTimer:
+      glGetQueryObjectui64vEXT(queries_[0], GL_QUERY_RESULT_EXT, &beg_stamp);
+      glGetQueryObjectui64vEXT(queries_[1], GL_QUERY_RESULT_EXT, &end_stamp);
+      glDeleteQueriesEXT(2, queries_);
+      break;
+
+    default:
+      return;
+  }
 
-  glGetQueryObjectui64v(queries_[1], GL_QUERY_RESULT, &timestamp);
-  end_time_ = (timestamp / base::Time::kNanosecondsPerMicrosecond) + offset_;
+  if (discard_trace_)
+    return;
 
-  glDeleteQueries(2, queries_);
+  start_time_ = (beg_stamp / base::Time::kNanosecondsPerMicrosecond) + offset_;
+  end_time_ = (end_stamp / base::Time::kNanosecondsPerMicrosecond) + offset_;
   outputter_->Trace(name(), start_time_, end_time_);
 }
 
+void GPUTrace::DiscardTrace() {
+  discard_trace_ = true;
+}
+
 GPUTracer::GPUTracer(gles2::GLES2Decoder* decoder)
     : gpu_trace_srv_category(TRACE_EVENT_API_GET_CATEGORY_GROUP_ENABLED(
           TRACE_DISABLED_BY_DEFAULT("gpu.service"))),
@@ -144,13 +212,16 @@ GPUTracer::GPUTracer(gles2::GLES2Decoder* decoder)
       decoder_(decoder),
       timer_offset_(0),
       last_tracer_source_(kTraceGroupInvalid),
-      enabled_(false),
+      tracer_type_(kTracerTypeInvalid),
       gpu_timing_synced_(false),
       gpu_executing_(false),
       process_posted_(false) {
-  if (gfx::g_driver_gl.ext.b_GL_ARB_timer_query) {
+  if (gfx::g_driver_gl.ext.b_GL_EXT_disjoint_timer_query) {
+    tracer_type_ = kTracerTypeDisjointTimer;
+    outputter_ = TraceOutputter::Create("GL_EXT_disjoint_timer_query");
+  } else if (gfx::g_driver_gl.ext.b_GL_ARB_timer_query) {
+    tracer_type_ = kTracerTypeARBTimer;
     outputter_ = TraceOutputter::Create("GL_ARB_timer_query");
-    enabled_ = true;
   }
 }
 
@@ -158,25 +229,19 @@ GPUTracer::~GPUTracer() {
 }
 
 bool GPUTracer::BeginDecoding() {
-  if (enabled_) {
-    if (*gpu_trace_dev_category) {
-      // Make sure timing is synced before tracing
-      if (!gpu_timing_synced_) {
-        CalculateTimerOffset();
-        gpu_timing_synced_ = true;
-      }
-    } else {
-      // If GPU device category is off, invalidate timing sync
-      gpu_timing_synced_ = false;

[vmiura, 2014/08/26 23:47:26]

I think we still need this part, so that we sync timing each time we start
tracing.

[David Yen, 2014/08/27 19:47:45]

Okay good point. I've moved this to the CalculateTimerOffset() function to
localize all the usage of gpu_timing_synced_.

-    }
-  }
-
   if (gpu_executing_)
     return false;
 
+  CalculateTimerOffset();
   gpu_executing_ = true;
 
   if (IsTracing()) {
+    // Reset disjoint bit for the disjoint timer.
+    if (tracer_type_ == kTracerTypeDisjointTimer) {
+      GLint disjoint_value = 0;
+      glGetIntegerv(GL_GPU_DISJOINT_EXT, &disjoint_value);
+    }
+
     // Begin a Trace for all active markers
     for (int n = 0; n < NUM_TRACER_SOURCES; n++) {
       for (size_t i = 0; i < markers_[n].size(); i++) {
@@ -272,10 +337,10 @@ const std::string& GPUTracer::CurrentName() const {
 }
 
 scoped_refptr<GPUTrace> GPUTracer::CreateTrace(const std::string& name) {
-  if (enabled_ && *gpu_trace_dev_category)
-    return new GPUTrace(outputter_, name, timer_offset_);
-  else
-    return new GPUTrace(name);
+  GpuTracerType tracer_type = *gpu_trace_dev_category ? tracer_type_ :
+                                                        kTracerTypeInvalid;
+
+  return new GPUTrace(outputter_, name, timer_offset_, tracer_type);
 }
 
 void GPUTracer::Process() {
@@ -285,7 +350,7 @@ void GPUTracer::Process() {
 }
 
 void GPUTracer::ProcessTraces() {
-  if (!enabled_) {
+  if (tracer_type_ == kTracerTypeInvalid) {
     while (!traces_.empty() && traces_.front()->IsAvailable()) {
       traces_.front()->Process();
       traces_.pop_front();
@@ -302,6 +367,19 @@ void GPUTracer::ProcessTraces() {
     return;
   }
 
+  // Check if disjoint operation has occurred, discard ongoing traces if so.
+  if (tracer_type_ == kTracerTypeDisjointTimer) {
+    GLint disjoint_value = 0;
+    glGetIntegerv(GL_GPU_DISJOINT_EXT, &disjoint_value);
+    if (disjoint_value) {
+      while (!traces_.empty()) {
+        traces_.front()->DiscardTrace();
+        traces_.front()->Process();
+        traces_.pop_front();

[vmiura, 2014/08/26 23:47:26]

Can we completely remove DiscardTrace() & Process() calls?
The destructor will delete the queries on traces_.pop_front().

That removes the need for 'discard_trace_' also.

[David Yen, 2014/08/27 19:47:45]

Yes you are right, I think I was thrown off by how the invalid case was done
above (line 354). Looking at the code calling Process() is clearly unnecessary.
I'll change them both to simply do "traces_.clear();".

+      }
+    }
+  }
+
   while (!traces_.empty() && traces_.front()->IsAvailable()) {
     traces_.front()->Process();
     traces_.pop_front();
@@ -314,23 +392,57 @@ void GPUTracer::ProcessTraces() {
 }
 
 void GPUTracer::CalculateTimerOffset() {
-  if (enabled_) {
+  if (*gpu_trace_dev_category && tracer_type_ != kTracerTypeInvalid) {
+    if (gpu_timing_synced_)
+      return;
+
     TRACE_EVENT0("gpu", "GPUTracer::CalculateTimerOffset");
 
     // NOTE(vmiura): It would be better to use glGetInteger64v, however
     // it's not available everywhere.
     GLuint64 gl_now = 0;
     GLuint query;
-    glFinish();
-    glGenQueries(1, &query);
-    glQueryCounter(query, GL_TIMESTAMP);
-    glFinish();
-    glGetQueryObjectui64v(query, GL_QUERY_RESULT, &gl_now);
+    GLint disjoint_value = 0;
+
+    switch (tracer_type_) {
+      case kTracerTypeARBTimer:
+        glFinish();
+        glGenQueries(1, &query);
+        glQueryCounter(query, GL_TIMESTAMP);
+        glFinish();
+
+        glGetQueryObjectui64v(query, GL_QUERY_RESULT, &gl_now);
+        glDeleteQueries(1, &query);
+        break;
+
+      case kTracerTypeDisjointTimer:
+        // Clear the disjoint bit before we do any queries.
+        glGetIntegerv(GL_GPU_DISJOINT_EXT, &disjoint_value);
+
+        glFinish();
+        glGenQueriesEXT(1, &query);
+        glQueryCounterEXT(query, GL_TIMESTAMP_EXT);
+        glFinish();
+
+        glGetQueryObjectui64vEXT(query, GL_QUERY_RESULT_EXT, &gl_now);
+        glDeleteQueriesEXT(1, &query);
+
+        // Discard results if disjoint operation has occurred.
+        glGetIntegerv(GL_GPU_DISJOINT_EXT, &disjoint_value);
+        if (disjoint_value)
+          return;
+
+        break;
+
+      default:
+        break;
+    }
+
     base::TimeTicks system_now = base::TimeTicks::NowFromSystemTraceTime();
 
     gl_now /= base::Time::kNanosecondsPerMicrosecond;
     timer_offset_ = system_now.ToInternalValue() - gl_now;
-    glDeleteQueries(1, &query);
+    gpu_timing_synced_ = true;
   }
 }