Merge "Fix query management in GpuProfiler" into oc-dev (511544f2) · Commits · e / os / android_frameworks_native-old

libs/vr/libdvrgraphics/gpu_profiler.cpp

+50 −2

Original line number	Diff line number	Diff line
		@@ -7,6 +7,12 @@
		namespace android {
		namespace dvr {

		namespace {

		constexpr int kMaxPendingQueries = 32;

		} // anonynmous namespace

		static int64_t AdjustTimerQueryToNs(int64_t gpu_time) { return gpu_time; }

		void GpuProfiler::TimerData::reset() {
		@@ -21,6 +27,7 @@ void GpuProfiler::TimerData::print(const char* name) const {

		// Enter a scope, records the timestamp for later matching with leave.
		void GpuProfiler::TimerData::enter(int64_t timestamp_ns) {
		entered = true;
		enter_timestamp_ns = timestamp_ns;
		}

		@@ -28,6 +35,15 @@ void GpuProfiler::TimerData::enter(int64_t timestamp_ns) {
		void GpuProfiler::TimerData::leave(int64_t timestamp_ns, const char* name,
		std::weak_ptr<int64_t> duration_ns,
		int print_period) {
		if (!entered) {
		// We got the leave event but are missing the enter. This can happen if
		// OnPendingQueryOverflow() is called, or if the calls to enter()/leave()
		// aren't properly balanced. Ignore the call but print a warning.
		ALOGW("Ignoring GpuProfiler::TimerData::leave event with no enter event");
		return;
		}
		entered = false;

		int64_t elapsed = timestamp_ns - enter_timestamp_ns;
		if (elapsed > 1000 * 1000 * 1000) {
		// More than one second, drop it as invalid data.
		@@ -50,12 +66,12 @@ GpuProfiler* GpuProfiler::Get() {

		GpuProfiler::GpuProfiler()
		: enable_gpu_tracing_(true),
		has_gl_context_(false),
		sync_with_cpu_time_(false),
		gl_timer_offset_ns_(0) {
		SyncGlTimebase();
		}

		GpuProfiler::~GpuProfiler() {}
		GpuProfiler::~GpuProfiler() { Clear(); }

		bool GpuProfiler::IsGpuProfilingSupported() const {
		// TODO(jbates) check for GL_EXT_disjoint_timer_query
		@@ -63,6 +79,9 @@ bool GpuProfiler::IsGpuProfilingSupported() const {
		}

		GLuint GpuProfiler::TryAllocateGlQueryId() {
		if (pending_gpu_queries_.size() >= kMaxPendingQueries)
		OnPendingQueryOverflow();

		GLuint query_id = 0;
		if (gl_timer_query_id_pool_.empty()) {
		glGenQueries(1, &query_id);
		@@ -95,6 +114,35 @@ void GpuProfiler::LeaveGlScope(const char* scope_name,
		}
		}

		void GpuProfiler::OnGlContextCreated() {
		has_gl_context_ = true;
		gl_timer_offset_ns_ = 0;
		SyncGlTimebase();
		}

		void GpuProfiler::OnGlContextDestroyed() {
		has_gl_context_ = false;
		Clear();
		}

		void GpuProfiler::Clear() {
		events_.clear();
		for (auto& query : pending_gpu_queries_)
		glDeleteQueries(1, &query.query_id);
		pending_gpu_queries_.clear();
		while (!gl_timer_query_id_pool_.empty()) {
		GLuint id = gl_timer_query_id_pool_.top();
		gl_timer_query_id_pool_.pop();
		glDeleteQueries(1, &id);
		}
		}

		void GpuProfiler::OnPendingQueryOverflow() {
		ALOGW("Reached limit of %d pending queries in GpuProfiler."
		" Clearing all queries.", kMaxPendingQueries);
		Clear();
		}

		void GpuProfiler::SyncGlTimebase() {
		if (!sync_with_cpu_time_) {
		return;

libs/vr/libdvrgraphics/include/private/dvr/graphics/gpu_profiler.h

+22 −1

Original line number	Diff line number	Diff line
		@@ -38,7 +38,7 @@ class GpuProfiler {
		// one of the TRACE_GPU* macros defined below.
		void SetEnableGpuTracing(bool enabled) { enable_gpu_tracing_ = enabled; }

		bool enabled() const { return enable_gpu_tracing_; }
		bool enabled() const { return enable_gpu_tracing_ && has_gl_context_; }

		// Attempt to keep the GPU times in sync with CPU times.
		void SetEnableSyncCpuTime(bool enabled) { sync_with_cpu_time_ = enabled; }
		@@ -62,6 +62,14 @@ class GpuProfiler {
		void LeaveGlScope(const char* scope_name, std::weak_ptr<int64_t> duration_ns,
		int print_period);

		// Must be called when the GL context is created. The GpuProfiler will be
		// inactive until this is called.
		void OnGlContextCreated();

		// Must be called before the GL context is destroyed. The GpuProfiler will be
		// inactive until a call to OnGlContextCreated().
		void OnGlContextDestroyed();

		private:
		// Data to queue the pending GPU timer queries that need to be polled
		// for completion.
		@@ -105,11 +113,20 @@ class GpuProfiler {
		void leave(int64_t timestamp_ns, const char* name,
		std::weak_ptr<int64_t> duration_ns, int print_period);

		bool entered = false;
		int64_t total_elapsed_ns = 0;
		int64_t enter_timestamp_ns = 0;
		int num_events = 0;
		};

		// Clear out events and free GL resources.
		void Clear();

		// Called when we detect that we've overflowed the pending query queue. This
		// shouldn't occur in practice, and probably indicates some internal
		// mismanagement of the gl query objects.
		void OnPendingQueryOverflow();

		// Synchronises the GL timebase with the CallTraceManager timebase.
		void SyncGlTimebase();

		@@ -119,6 +136,10 @@ class GpuProfiler {
		// Setting for enabling GPU tracing.
		bool enable_gpu_tracing_;

		// True if we have a GL context, false otherwise. When the GpuProfiler is
		// first created we assume no GL context.
		bool has_gl_context_;

		// Setting for synchronizing GPU timestamps with CPU time.
		bool sync_with_cpu_time_;

libs/vr/libvrflinger/compositor.cpp

+2 −0

Original line number	Diff line number	Diff line
		@@ -403,6 +403,7 @@ bool Compositor::InitializeEGL() {
		}

		load_gl_extensions();
		GpuProfiler::Get()->OnGlContextCreated();

		glEnable(BINNING_CONTROL_HINT_QCOM);
		glHint(BINNING_CONTROL_HINT_QCOM, RENDER_DIRECT_TO_FRAMEBUFFER_QCOM);
		@@ -438,6 +439,7 @@ void Compositor::Shutdown() {
		eds_renderer_.reset();

		if (context_) {
		GpuProfiler::Get()->OnGlContextDestroyed();
		eglDestroyContext(display_, context_);
		context_ = 0;
		}