Added begin/end CPU/GPU timer stats instead of elapsed.

examples/17-drawstress/drawstress.cpp

@@ -242,6 +242,11 @@ class DrawStress : public entry::AppI
 			imguiLabel("Draw calls: %d", m_dim*m_dim*m_dim);
 			imguiLabel("Avg Delta Time (1 second) [ms]: %0.4f", m_deltaTimeAvgNs/1000.0f);
 
+			imguiSeparatorLine();
+			const bgfx::Stats* stats = bgfx::getStats();
+			imguiLabel("GPU %0.6f [ms]", double(stats->gpuTimeEnd - stats->gpuTimeBegin)*1000.0/stats->gpuTimerFreq);
+			imguiLabel("CPU %0.6f [ms]", double(stats->cpuTimeEnd - stats->cpuTimeBegin)*1000.0/stats->cpuTimerFreq);
+
 			imguiEndScrollArea();
 			imguiEndFrame();
 

include/bgfx/bgfx.h

@@ -589,10 +589,12 @@ namespace bgfx
 	///
 	struct Stats
 	{
-		uint64_t cpuTime;      //!< CPU frame time.
+		uint64_t cpuTimeBegin; //!< CPU frame begin time.
+		uint64_t cpuTimeEnd;   //!< CPU frame end time.
 		uint64_t cpuTimerFreq; //!< CPU timer frequency.
 
-		uint64_t gpuTime;      //!< GPU frame time.
+		uint64_t gpuTimeBegin; //!< GPU frame begin time.
+		uint64_t gpuTimeEnd;   //!< GPU frame end time.
 		uint64_t gpuTimerFreq; //!< GPU timer frequency.
 	};
 

src/renderer_d3d11.cpp

@@ -4475,7 +4475,7 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 			DX_CHECK(device->CreateQuery(&query, &frame.m_disjoint) );
 
 			query.Query = D3D11_QUERY_TIMESTAMP;
-			DX_CHECK(device->CreateQuery(&query, &frame.m_start) );
+			DX_CHECK(device->CreateQuery(&query, &frame.m_begin) );
 			DX_CHECK(device->CreateQuery(&query, &frame.m_end) );
 		}
 
@@ -4490,7 +4490,7 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 		{
 			Frame& frame = m_frame[ii];
 			DX_RELEASE(frame.m_disjoint, 0);
-			DX_RELEASE(frame.m_start, 0);
+			DX_RELEASE(frame.m_begin, 0);
 			DX_RELEASE(frame.m_end, 0);
 		}
 	}
@@ -4506,7 +4506,7 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 
 		Frame& frame = m_frame[m_control.m_current];
 		deviceCtx->Begin(frame.m_disjoint);
-		deviceCtx->End(frame.m_start);
+		deviceCtx->End(frame.m_begin);
 	}
 
 	void TimerQueryD3D11::end()
@@ -4525,8 +4525,8 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 			ID3D11DeviceContext* deviceCtx = s_renderD3D11->m_deviceCtx;
 			Frame& frame = m_frame[m_control.m_read];
 
-			uint64_t finish;
+			uint64_t timeEnd;
-			HRESULT hr = deviceCtx->GetData(frame.m_end, &finish, sizeof(finish), D3D11_ASYNC_GETDATA_DONOTFLUSH);
+			HRESULT hr = deviceCtx->GetData(frame.m_end, &timeEnd, sizeof(timeEnd), D3D11_ASYNC_GETDATA_DONOTFLUSH);
 			if (S_OK == hr)
 			{
 				m_control.consume(1);
@@ -4540,11 +4540,13 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 				D3D11_QUERY_DATA_TIMESTAMP_DISJOINT disjoint;
 				deviceCtx->GetData(frame.m_disjoint, &disjoint, sizeof(disjoint), 0);
 
-				uint64_t start;
+				uint64_t timeBegin;
-				deviceCtx->GetData(frame.m_start, &start, sizeof(start), 0);
+				deviceCtx->GetData(frame.m_begin, &timeBegin, sizeof(timeBegin), 0);
 
 				m_frequency = disjoint.Frequency;
-				m_elapsed   = finish - start;
+				m_begin     = timeBegin;
+				m_end       = timeEnd;
+				m_elapsed   = timeEnd - timeBegin;
 
 				return true;
 			}
@@ -5505,6 +5507,10 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 		elapsed += now;
 
 		static int64_t last = now;
+
+		Stats& perfStats = _render->m_perfStats;
+		perfStats.cpuTimeBegin = last;
+
 		int64_t frameTime = now - last;
 		last = now;
 
@@ -5529,10 +5535,10 @@ BX_PRAGMA_DIAGNOSTIC_POP();
 
 		const int64_t timerFreq = bx::getHPFrequency();
 
-		Stats& perfStats   = _render->m_perfStats;
+		perfStats.cpuTimeEnd   = now;
-		perfStats.cpuTime      = frameTime;
 		perfStats.cpuTimerFreq = timerFreq;
-		perfStats.gpuTime      = m_gpuTimer.m_elapsed;
+		perfStats.gpuTimeBegin = m_gpuTimer.m_begin;
+		perfStats.gpuTimeEnd   = m_gpuTimer.m_end;
 		perfStats.gpuTimerFreq = m_gpuTimer.m_frequency;
 
 		if (_render->m_debug & (BGFX_DEBUG_IFH|BGFX_DEBUG_STATS) )

src/renderer_d3d11.h

@@ -296,10 +296,12 @@ namespace bgfx { namespace d3d11
 		struct Frame
 		{
 			ID3D11Query* m_disjoint;
-			ID3D11Query* m_start;
+			ID3D11Query* m_begin;
 			ID3D11Query* m_end;
 		};
 
+		uint64_t m_begin;
+		uint64_t m_end;
 		uint64_t m_elapsed;
 		uint64_t m_frequency;
 

src/renderer_d3d9.cpp

@@ -3318,7 +3318,7 @@ namespace bgfx { namespace d3d9
 		{
 			Frame& frame = m_frame[ii];
 			DX_CHECK(device->CreateQuery(D3DQUERYTYPE_TIMESTAMPDISJOINT, &frame.m_disjoint) );
-			DX_CHECK(device->CreateQuery(D3DQUERYTYPE_TIMESTAMP,         &frame.m_start) );
+			DX_CHECK(device->CreateQuery(D3DQUERYTYPE_TIMESTAMP,         &frame.m_begin) );
 			DX_CHECK(device->CreateQuery(D3DQUERYTYPE_TIMESTAMP,         &frame.m_end) );
 			DX_CHECK(device->CreateQuery(D3DQUERYTYPE_TIMESTAMPFREQ,     &frame.m_freq) );
 		}
@@ -3334,7 +3334,7 @@ namespace bgfx { namespace d3d9
 		{
 			Frame& frame = m_frame[ii];
 			DX_RELEASE(frame.m_disjoint, 0);
-			DX_RELEASE(frame.m_start, 0);
+			DX_RELEASE(frame.m_begin, 0);
 			DX_RELEASE(frame.m_end, 0);
 			DX_RELEASE(frame.m_freq, 0);
 		}
@@ -3349,7 +3349,7 @@ namespace bgfx { namespace d3d9
 
 		Frame& frame = m_frame[m_control.m_current];
 		frame.m_disjoint->Issue(D3DISSUE_BEGIN);
-		frame.m_start->Issue(D3DISSUE_END);
+		frame.m_begin->Issue(D3DISSUE_END);
 	}
 
 	void TimerQueryD3D9::end()
@@ -3373,14 +3373,16 @@ namespace bgfx { namespace d3d9
 			{
 				m_control.consume(1);
 
-				uint64_t timeStart;
+				uint64_t timeBegin;
-				DX_CHECK(frame.m_start->GetData(&timeStart, sizeof(timeStart), 0) );
+				DX_CHECK(frame.m_begin->GetData(&timeBegin, sizeof(timeBegin), 0) );
 
 				uint64_t freq;
 				DX_CHECK(frame.m_freq->GetData(&freq, sizeof(freq), 0) );
 
 				m_frequency = freq;
-				m_elapsed   = timeEnd - timeStart;
+				m_begin     = timeBegin;
+				m_end       = timeEnd;
+				m_elapsed   = timeEnd - timeBegin;
 
 				return true;
 			}
@@ -4084,6 +4086,10 @@ namespace bgfx { namespace d3d9
 		elapsed += now;
 
 		static int64_t last = now;
+
+		Stats& perfStats = _render->m_perfStats;
+		perfStats.cpuTimeBegin = last;
+
 		int64_t frameTime = now - last;
 		last = now;
 
@@ -4111,10 +4117,10 @@ namespace bgfx { namespace d3d9
 
 		const int64_t timerFreq = bx::getHPFrequency();
 
-		Stats& perfStats   = _render->m_perfStats;
+		perfStats.cpuTimeEnd   = now;
-		perfStats.cpuTime      = frameTime;
 		perfStats.cpuTimerFreq = timerFreq;
-		perfStats.gpuTime      = m_gpuTimer.m_elapsed;
+		perfStats.gpuTimeBegin = m_gpuTimer.m_begin;
+		perfStats.gpuTimeEnd   = m_gpuTimer.m_end;
 		perfStats.gpuTimerFreq = m_gpuTimer.m_frequency;
 
 		if (_render->m_debug & (BGFX_DEBUG_IFH|BGFX_DEBUG_STATS) )

src/renderer_d3d9.h

@@ -418,11 +418,13 @@ namespace bgfx { namespace d3d9
 		struct Frame
 		{
 			IDirect3DQuery9* m_disjoint;
-			IDirect3DQuery9* m_start;
+			IDirect3DQuery9* m_begin;
 			IDirect3DQuery9* m_end;
 			IDirect3DQuery9* m_freq;
 		};
 
+		uint64_t m_begin;
+		uint64_t m_end;
 		uint64_t m_elapsed;
 		uint64_t m_frequency;
 

src/renderer_gl.cpp

@@ -6274,6 +6274,10 @@ namespace bgfx { namespace gl
 		elapsed += now;
 
 		static int64_t last = now;
+
+		Stats& perfStats   = _render->m_perfStats;
+		perfStats.cpuTimeBegin = last;
+
 		int64_t frameTime = now - last;
 		last = now;
 
@@ -6301,10 +6305,10 @@ namespace bgfx { namespace gl
 
 		const int64_t timerFreq = bx::getHPFrequency();
 
-		Stats& perfStats   = _render->m_perfStats;
+		perfStats.cpuTimeEnd   = now;
-		perfStats.cpuTime      = frameTime;
 		perfStats.cpuTimerFreq = timerFreq;
-		perfStats.gpuTime      = elapsedGl;
+		perfStats.gpuTimeBegin = m_gpuTimer.m_begin;
+		perfStats.gpuTimeEnd   = m_gpuTimer.m_end;
 		perfStats.gpuTimerFreq = 1000000000;
 
 		if (_render->m_debug & (BGFX_DEBUG_IFH|BGFX_DEBUG_STATS) )

src/renderer_gl.h

@@ -1205,12 +1205,22 @@ namespace bgfx { namespace gl
 
 		void create()
 		{
-			GL_CHECK(glGenQueries(BX_COUNTOF(m_frame), m_frame) );
+			for (uint32_t ii = 0; ii < BX_COUNTOF(m_frame); ++ii)
+			{
+				Frame& frame = m_frame[ii];
+				GL_CHECK(glGenQueries(1, &frame.m_begin) );
+				GL_CHECK(glGenQueries(1, &frame.m_elapsed) );
+			}
 		}
 
 		void destroy()
 		{
-			GL_CHECK(glDeleteQueries(BX_COUNTOF(m_frame), m_frame) );
+			for (uint32_t ii = 0; ii < BX_COUNTOF(m_frame); ++ii)
+			{
+				Frame& frame = m_frame[ii];
+				GL_CHECK(glDeleteQueries(1, &frame.m_begin) );
+				GL_CHECK(glDeleteQueries(1, &frame.m_elapsed) );
+			}
 		}
 
 		void begin()
@@ -1220,8 +1230,13 @@ namespace bgfx { namespace gl
 				get();
 			}
 
+			Frame& frame = m_frame[m_control.m_current];
+			GL_CHECK(glQueryCounter(frame.m_begin
+					, GL_TIMESTAMP
+					) );
+
 			GL_CHECK(glBeginQuery(GL_TIME_ELAPSED
-					, m_frame[m_control.m_current]
+					, frame.m_elapsed
 					) );
 		}
 
@@ -1235,18 +1250,26 @@ namespace bgfx { namespace gl
 		{
 			if (0 != m_control.available() )
 			{
+				Frame& frame = m_frame[m_control.m_read];
+
 				GLint available;
-				GL_CHECK(glGetQueryObjectiv(m_frame[m_control.m_read]
+				GL_CHECK(glGetQueryObjectiv(frame.m_elapsed
 						, GL_QUERY_RESULT_AVAILABLE
 						, &available
 						) );
 
 				if (available)
 				{
-					GL_CHECK(glGetQueryObjectui64v(m_frame[m_control.m_read]
+					GL_CHECK(glGetQueryObjectui64v(frame.m_begin
+							, GL_QUERY_RESULT
+							, &m_begin
+							) );
+
+					GL_CHECK(glGetQueryObjectui64v(frame.m_elapsed
 							, GL_QUERY_RESULT
 							, &m_elapsed
 							) );
+					m_end = m_begin + m_elapsed;
 					m_control.consume(1);
 					return true;
 				}
@@ -1255,9 +1278,17 @@ namespace bgfx { namespace gl
 			return false;
 		}
 
+		uint64_t m_begin;
+		uint64_t m_end;
 		uint64_t m_elapsed;
 
-		GLuint m_frame[4];
+		struct Frame
+		{
+			GLuint m_begin;
+			GLuint m_elapsed;
+		};
+
+		Frame m_frame[4];
 		bx::RingBufferControl m_control;
 	};