gsdx ogl: add GPU timers to measure time between 2 vsync

So far only basic stuff (min/mean/max frame rendering time)

plugins/GSdx/GLLoader.cpp

@@ -100,6 +100,8 @@ PFNGLQUERYCOUNTERPROC                  glQueryCounter                      = NUL
 PFNGLGETQUERYOBJECTI64VPROC            glGetQueryObjecti64v                = NULL;
 PFNGLGETQUERYOBJECTUI64VPROC           glGetQueryObjectui64v               = NULL;
 PFNGLGETINTEGER64VPROC                 glGetInteger64v                     = NULL;
+PFNGLCREATEQUERIESPROC                 glCreateQueries                     = NULL;
+PFNGLDELETEQUERIESPROC                 glDeleteQueries                     = NULL;
 // GL4.0
 // GL4.1
 PFNGLBINDPROGRAMPIPELINEPROC           glBindProgramPipeline               = NULL;

plugins/GSdx/GLLoader.h

@@ -268,6 +268,8 @@ extern   PFNGLQUERYCOUNTERPROC                  glQueryCounter;
 extern   PFNGLGETQUERYOBJECTI64VPROC            glGetQueryObjecti64v;
 extern   PFNGLGETQUERYOBJECTUI64VPROC           glGetQueryObjectui64v;
 extern   PFNGLGETINTEGER64VPROC                 glGetInteger64v;
+extern   PFNGLCREATEQUERIESPROC                 glCreateQueries;
+extern   PFNGLDELETEQUERIESPROC                 glDeleteQueries;
 // GL4.0
 // GL4.1
 extern   PFNGLBINDPROGRAMPIPELINEPROC           glBindProgramPipeline;

plugins/GSdx/GS.cpp

@@ -1689,7 +1689,9 @@ EXPORT_C GSReplay(char* lpszCmdLine, int renderer)
 		// Ensure the rendering is complete to measure correctly the time.
 		glFinish();
 
-		if (finished > 90) {
+		if (finished >= 200) {
+			; // Nop for Nvidia Profiler
+		} else if (finished > 90) {
 			sleep(1);
 		} else {
 			unsigned long end = timeGetTime();
@@ -1724,7 +1726,7 @@ EXPORT_C GSReplay(char* lpszCmdLine, int renderer)
 		fprintf(stderr, "\n\nMean: %fms\n", mean);
 		fprintf(stderr, "Standard deviation: %fms\n", sd);
 		fprintf(stderr, "Mean by frame: %fms (%ffps)\n", mean/(float)frame_number, 1000.0f*frame_number/mean);
-		fprintf(stderr, "Standard deviatin by frame: %fms\n", sd/(float)frame_number);
+		fprintf(stderr, "Standard deviation by frame: %fms\n", sd/(float)frame_number);
 	}
 #ifdef ENABLE_OGL_DEBUG_MEM_BW
 	total_frame_nb *= 1024;
@@ -1748,4 +1750,3 @@ EXPORT_C GSReplay(char* lpszCmdLine, int renderer)
 	GSshutdown();
 }
 #endif
-

plugins/GSdx/GSDeviceOGL.cpp

@@ -71,6 +71,7 @@ GSDeviceOGL::GSDeviceOGL()
 	memset(&m_date, 0, sizeof(m_date));
 	memset(&m_shadeboost, 0, sizeof(m_shadeboost));
 	memset(&m_om_dss, 0, sizeof(m_om_dss));
+	memset(&m_profiler, 0 , sizeof(m_profiler));
 	GLState::Clear();
 
 	// Reset the debug file
@@ -94,6 +95,10 @@ GSDeviceOGL::~GSDeviceOGL()
 
 	GL_PUSH("GSDeviceOGL destructor");
 
+	if (GLLoader::in_replayer) {
+		GenerateProfilerData();
+	}
+
 	// Clean vertex buffer state
 	delete m_va;
 
@@ -139,6 +144,51 @@ GSDeviceOGL::~GSDeviceOGL()
 	m_shader = NULL;
 }
 
+void GSDeviceOGL::GenerateProfilerData()
+{
+	if (m_profiler.last_query < 3) return;
+
+	// Point to the last query
+	m_profiler.last_query--;
+
+	// Wait latest quey to get valid result
+	GLuint available = 0;
+	while (!available) {
+		glGetQueryObjectuiv(m_profiler.timer(), GL_QUERY_RESULT_AVAILABLE, &available);
+	}
+
+	uint64 time_start;
+	uint64 time_end;
+
+	uint64 min_time = 0xFFFFFFFFFFFFFFFFull;
+	uint64 max_time = 0;
+	uint64 total    = 0;
+
+	glGetQueryObjectui64v(m_profiler.timer_query[0], GL_QUERY_RESULT, &time_start);
+	for (uint32 q = 1; q < m_profiler.last_query; q++) {
+		glGetQueryObjectui64v(m_profiler.timer_query[q], GL_QUERY_RESULT, &time_end);
+		uint64 t = time_end - time_start;
+
+		min_time = std::min(t, min_time);
+		max_time = std::max(t, max_time);
+		total   += t;
+
+		time_start = time_end;
+	}
+
+	glDeleteQueries(1 << 16, m_profiler.timer_query);
+
+	// FIXME remove 1/linux_replay frame info
+	// Add fps, deviation
+	// remove glFinish
+
+	double ms = 0.000001;
+	fprintf(stderr, "\nGenerateProfilerData:\n");
+	fprintf(stderr, "Min  %f\n", (double)min_time * ms);
+	fprintf(stderr, "Mean %f\n", (double)total/(double)m_profiler.last_query * ms);
+	fprintf(stderr, "Max  %f\n", (double)max_time * ms);
+}
+
 GSTexture* GSDeviceOGL::CreateSurface(int type, int w, int h, bool msaa, int fmt)
 {
 	GL_PUSH("Create surface");
@@ -248,6 +298,11 @@ bool GSDeviceOGL::Create(GSWnd* wnd)
 		glBindFramebuffer(GL_READ_FRAMEBUFFER, m_fbo_read);
 		glReadBuffer(GL_COLOR_ATTACHMENT0);
 		glBindFramebuffer(GL_READ_FRAMEBUFFER, 0);
+
+		// Some timers to help profiling
+		if (GLLoader::in_replayer) {
+			glCreateQueries(GL_TIMESTAMP, 1 << 16, m_profiler.timer_query);
+		}
 	}
 
 	// ****************************************************************
@@ -504,6 +559,11 @@ void GSDeviceOGL::Flip()
 	#endif
 
 	m_wnd->Flip();
+
+	if (GLLoader::in_replayer) {
+		glQueryCounter(m_profiler.timer(), GL_TIMESTAMP);
+		m_profiler.last_query++;
+	}
 }
 
 void GSDeviceOGL::BeforeDraw()

plugins/GSdx/GSDeviceOGL.h

@@ -449,6 +449,13 @@ public:
 		GLuint ps;
 	} m_shadeboost;
 
+	struct {
+		uint16 last_query;
+		GLuint timer_query[1<<16];
+
+		GLuint timer() { return timer_query[last_query]; }
+	} m_profiler;
+
 	GLuint m_vs[1];
 	GLuint m_gs[1<<2];
 	GLuint m_ps_ss[1<<4];
@@ -484,6 +491,8 @@ public:
 	GSDeviceOGL();
 	virtual ~GSDeviceOGL();
 
+	void GenerateProfilerData();
+
 	static void CheckDebugLog();
 	// Used by OpenGL, so the same calling convention is required.
 	static void APIENTRY DebugOutputToFile(GLenum gl_source, GLenum gl_type, GLuint id, GLenum gl_severity, GLsizei gl_length, const GLchar *gl_message, const void* userParam);
@@ -555,7 +564,6 @@ public:
 	void SelfShaderTestRun(const string& dir, const string& file, const PSSelector& sel, int& nb_shader);
 	void SelfShaderTest();
 
-
 	void SetupIA(const void* vertex, int vertex_count, const uint32* index, int index_count, int prim);
 	void SetupPipeline(const VSSelector& vsel, const GSSelector& gsel, const PSSelector& psel);
 	void SetupCB(const VSConstantBuffer* vs_cb, const PSConstantBuffer* ps_cb);

plugins/GSdx/GSWnd.cpp

@@ -84,15 +84,17 @@ void GSWndGL::PopulateGlFunction()
 	GL_EXT_LOAD(glClientWaitSync);
 	GL_EXT_LOAD(glFlushMappedBufferRange);
 	// Query object
-	GL_EXT_LOAD(glBeginQuery);
+	GL_EXT_LOAD_OPT(glBeginQuery);
-	GL_EXT_LOAD(glEndQuery);
+	GL_EXT_LOAD_OPT(glEndQuery);
-	GL_EXT_LOAD(glGetQueryiv);
+	GL_EXT_LOAD_OPT(glGetQueryiv);
-	GL_EXT_LOAD(glGetQueryObjectiv);
+	GL_EXT_LOAD_OPT(glGetQueryObjectiv);
-	GL_EXT_LOAD(glGetQueryObjectuiv);
+	GL_EXT_LOAD_OPT(glGetQueryObjectuiv);
-	GL_EXT_LOAD(glQueryCounter);
+	GL_EXT_LOAD_OPT(glQueryCounter);
-	GL_EXT_LOAD(glGetQueryObjecti64v);
+	GL_EXT_LOAD_OPT(glGetQueryObjecti64v);
-	GL_EXT_LOAD(glGetQueryObjectui64v);
+	GL_EXT_LOAD_OPT(glGetQueryObjectui64v);
-	GL_EXT_LOAD(glGetInteger64v);
+	GL_EXT_LOAD_OPT(glGetInteger64v);
+	GL_EXT_LOAD_OPT(glCreateQueries);
+	GL_EXT_LOAD_OPT(glDeleteQueries);
 	// GL4.0
 	GL_EXT_LOAD_OPT(glBlendEquationSeparateiARB);
 	GL_EXT_LOAD_OPT(glBlendFuncSeparateiARB);