Merge pull request #21392 from chealy/gl-timer-query-profiler

OpenGL: Add GPU timestamp profiling with GL_EXT_disjoint_timer_query

CMakeLists.txt

@@ -771,6 +771,8 @@ add_library(Common STATIC
 	Common/GPU/OpenGL/GLRenderManager.h
 	Common/GPU/OpenGL/GLQueueRunner.cpp
 	Common/GPU/OpenGL/GLQueueRunner.h
+	Common/GPU/OpenGL/GLProfiler.cpp
+	Common/GPU/OpenGL/GLProfiler.h
 	Common/GPU/OpenGL/DataFormatGL.cpp
 	Common/GPU/OpenGL/DataFormatGL.h
 	Common/GPU/Vulkan/VulkanBarrier.cpp

Common/Common.vcxproj

@@ -418,6 +418,7 @@
     <ClInclude Include="GPU\OpenGL\GLFrameData.h" />
     <ClInclude Include="GPU\OpenGL\GLMemory.h" />
     <ClInclude Include="GPU\OpenGL\GLQueueRunner.h" />
+    <ClInclude Include="GPU\OpenGL\GLProfiler.h" />
     <ClInclude Include="GPU\OpenGL\GLRenderManager.h" />
     <ClInclude Include="GPU\OpenGL\GLSLProgram.h" />
     <ClInclude Include="GPU\OpenGL\GLCommon.h" />
@@ -885,6 +886,7 @@
     <ClCompile Include="GPU\OpenGL\GLFrameData.cpp" />
     <ClCompile Include="GPU\OpenGL\GLMemory.cpp" />
     <ClCompile Include="GPU\OpenGL\GLQueueRunner.cpp" />
+    <ClCompile Include="GPU\OpenGL\GLProfiler.cpp" />
     <ClCompile Include="GPU\OpenGL\GLRenderManager.cpp" />
     <ClCompile Include="GPU\OpenGL\GLSLProgram.cpp" />
     <ClCompile Include="GPU\OpenGL\GLDebugLog.cpp" />

Common/Common.vcxproj.filters

@@ -261,6 +261,9 @@
     <ClInclude Include="GPU\OpenGL\GLQueueRunner.h">
       <Filter>GPU\OpenGL</Filter>
     </ClInclude>
+    <ClInclude Include="GPU\OpenGL\GLProfiler.h">
+      <Filter>GPU\OpenGL</Filter>
+    </ClInclude>
     <ClInclude Include="GPU\OpenGL\GLRenderManager.h">
       <Filter>GPU\OpenGL</Filter>
     </ClInclude>
@@ -966,6 +969,9 @@
     <ClCompile Include="GPU\OpenGL\GLQueueRunner.cpp">
       <Filter>GPU\OpenGL</Filter>
     </ClCompile>
+    <ClCompile Include="GPU\OpenGL\GLProfiler.cpp">
+      <Filter>GPU\OpenGL</Filter>
+    </ClCompile>
     <ClCompile Include="GPU\OpenGL\GLRenderManager.cpp">
       <Filter>GPU\OpenGL</Filter>
     </ClCompile>

Common/GPU/OpenGL/GLFeatures.cpp

@@ -384,6 +384,7 @@ bool CheckGLExtensions() {
 	gl_extensions.ARB_explicit_attrib_location = g_set_gl_extensions.count("GL_ARB_explicit_attrib_location") != 0;
 	gl_extensions.ARB_texture_non_power_of_two = g_set_gl_extensions.count("GL_ARB_texture_non_power_of_two") != 0;
 	gl_extensions.ARB_shader_stencil_export = g_set_gl_extensions.count("GL_ARB_shader_stencil_export") != 0;
+	gl_extensions.ARB_timer_query = g_set_gl_extensions.count("GL_ARB_timer_query") != 0;
 	gl_extensions.ARB_texture_compression_bptc = g_set_gl_extensions.count("GL_ARB_texture_compression_bptc") != 0;
 	gl_extensions.ARB_texture_compression_rgtc = g_set_gl_extensions.count("GL_ARB_texture_compression_rgtc") != 0;
 	gl_extensions.KHR_texture_compression_astc_ldr = g_set_gl_extensions.count("GL_KHR_texture_compression_astc_ldr") != 0;
@@ -406,6 +407,7 @@ bool CheckGLExtensions() {
 		gl_extensions.EXT_buffer_storage = g_set_gl_extensions.count("GL_EXT_buffer_storage") != 0;
 		gl_extensions.EXT_clip_cull_distance = g_set_gl_extensions.count("GL_EXT_clip_cull_distance") != 0;
 		gl_extensions.EXT_depth_clamp = g_set_gl_extensions.count("GL_EXT_depth_clamp") != 0;
+		gl_extensions.EXT_disjoint_timer_query = g_set_gl_extensions.count("GL_EXT_disjoint_timer_query") != 0;
 		gl_extensions.APPLE_clip_distance = g_set_gl_extensions.count("GL_APPLE_clip_distance") != 0;
 
 #if defined(__ANDROID__)

Common/GPU/OpenGL/GLFeatures.h

@@ -86,6 +86,8 @@ struct GLExtensions {
 	bool KHR_texture_compression_astc_ldr;
 
 	// EXT
+	bool EXT_disjoint_timer_query;
+	bool ARB_timer_query;
 	bool EXT_texture_compression_s3tc;
 	bool EXT_swap_control_tear;
 	bool EXT_discard_framebuffer;

Common/GPU/OpenGL/GLProfiler.cpp

@@ -0,0 +1,174 @@
+#include <cstdarg>
+
+#include "Common/GPU/OpenGL/GLProfiler.h"
+#include "Common/GPU/OpenGL/GLCommon.h"
+#include "Common/GPU/OpenGL/GLFeatures.h"
+#include "Common/GPU/OpenGL/GLDebugLog.h"
+#include "Common/Log.h"
+
+// For iOS, define function pointer types and variables locally since gl3stub.h
+// content is excluded on iOS. These will remain NULL since iOS doesn't support
+// GL_EXT_disjoint_timer_query.
+#if PPSSPP_PLATFORM(IOS)
+typedef void (*PFNGLQUERYCOUNTERPROC)(GLuint id, GLenum target);
+typedef void (*PFNGLGETQUERYOBJECTUI64VPROC)(GLuint id, GLenum pname, GLuint64 *params);
+static PFNGLQUERYCOUNTERPROC glQueryCounter = nullptr;
+static PFNGLGETQUERYOBJECTUI64VPROC glGetQueryObjectui64v = nullptr;
+#endif
+
+// Constants - same values for both EXT and ARB versions
+#ifndef GL_TIMESTAMP
+#define GL_TIMESTAMP 0x8E28
+#endif
+#ifndef GL_QUERY_RESULT
+#define GL_QUERY_RESULT 0x8866
+#endif
+#ifndef GL_GPU_DISJOINT_EXT
+#define GL_GPU_DISJOINT_EXT 0x8FBB
+#endif
+
+// GLCommon.h provides access to GL functions:
+// - On GLES: function pointers from gl3stub.h, loaded via eglGetProcAddress
+// - On desktop GL: GLEW provides the functions
+
+void GLProfiler::Init() {
+	supported_ = false;
+	firstFrame_ = true;
+	numQueries_ = 0;
+	scopes_.clear();
+	scopeStack_.clear();
+
+	// Check for extension support
+	// Function pointers are declared in gl3stub.h and loaded appropriately per platform
+	if (gl_extensions.EXT_disjoint_timer_query) {
+		// GLES path - use function pointers loaded with EXT suffix
+		if (glQueryCounter && glGetQueryObjectui64v) {
+			supported_ = true;
+			INFO_LOG(Log::G3D, "GLProfiler: Using GL_EXT_disjoint_timer_query");
+		}
+	} else if (gl_extensions.ARB_timer_query) {
+		// Desktop GL path - use function pointers (no suffix)
+		if (glQueryCounter && glGetQueryObjectui64v) {
+			supported_ = true;
+			INFO_LOG(Log::G3D, "GLProfiler: Using GL_ARB_timer_query");
+		}
+	}
+
+	if (supported_) {
+		// Pre-allocate query objects
+		queries_.resize(MAX_QUERY_COUNT);
+		glGenQueries(MAX_QUERY_COUNT, queries_.data());
+		CHECK_GL_ERROR_IF_DEBUG();
+	}
+}
+
+void GLProfiler::Shutdown() {
+	if (supported_ && !queries_.empty()) {
+		glDeleteQueries((GLsizei)queries_.size(), queries_.data());
+		queries_.clear();
+	}
+	supported_ = false;
+	scopes_.clear();
+	scopeStack_.clear();
+}
+
+void GLProfiler::BeginFrame() {
+	if (!supported_) {
+		return;
+	}
+
+	// Check if profiling is enabled
+	if (enabledPtr_ && !*enabledPtr_) {
+		scopes_.clear();
+		scopeStack_.clear();
+		numQueries_ = 0;
+		return;
+	}
+
+	// Check for disjoint operation (GPU frequency changed) - GLES only
+	if (gl_extensions.EXT_disjoint_timer_query) {
+		GLint disjoint = 0;
+		glGetIntegerv(GL_GPU_DISJOINT_EXT, &disjoint);
+		if (disjoint) {
+			// Results are invalid, just clear and start fresh
+			WARN_LOG(Log::G3D, "GLProfiler: GPU disjoint detected, timing results discarded");
+			scopes_.clear();
+			scopeStack_.clear();
+			numQueries_ = 0;
+			firstFrame_ = true;
+			return;
+		}
+	}
+
+	// Read results from previous frame (guaranteed complete now)
+	if (numQueries_ > 0 && !firstFrame_) {
+		static const char * const indent[4] = { "", "  ", "    ", "      " };
+
+		if (!scopes_.empty()) {
+			INFO_LOG(Log::G3D, "OpenGL profiling events this frame:");
+		}
+
+		// Log results
+		for (auto &scope : scopes_) {
+			if (scope.endQueryId == -1) {
+				WARN_LOG(Log::G3D, "Unclosed scope: %s", scope.name);
+				continue;
+			}
+
+			GLuint64 startTime = 0, endTime = 0;
+			glGetQueryObjectui64v(queries_[scope.startQueryId], GL_QUERY_RESULT, &startTime);
+			glGetQueryObjectui64v(queries_[scope.endQueryId], GL_QUERY_RESULT, &endTime);
+
+			// Times are in nanoseconds, convert to milliseconds
+			double milliseconds = (double)(endTime - startTime) / 1000000.0;
+
+			INFO_LOG(Log::G3D, "%s%s (%0.3f ms)", indent[scope.level & 3], scope.name, milliseconds);
+		}
+	}
+
+	firstFrame_ = false;
+	scopes_.clear();
+	scopeStack_.clear();
+	numQueries_ = 0;
+}
+
+void GLProfiler::Begin(const char *fmt, ...) {
+	if (!supported_ || (enabledPtr_ && !*enabledPtr_) || numQueries_ >= MAX_QUERY_COUNT - 1) {
+		return;
+	}
+
+	GLProfilerScope scope;
+	va_list args;
+	va_start(args, fmt);
+	vsnprintf(scope.name, sizeof(scope.name), fmt, args);
+	va_end(args);
+	scope.startQueryId = numQueries_;
+	scope.endQueryId = -1;
+	scope.level = (int)scopeStack_.size();
+
+	scopeStack_.push_back(scopes_.size());
+	scopes_.push_back(scope);
+
+	glQueryCounter(queries_[numQueries_], GL_TIMESTAMP);
+	numQueries_++;
+}
+
+void GLProfiler::End() {
+	if (!supported_ || (enabledPtr_ && !*enabledPtr_) || numQueries_ >= MAX_QUERY_COUNT - 1) {
+		return;
+	}
+
+	if (scopeStack_.empty()) {
+		WARN_LOG(Log::G3D, "GLProfiler::End called without matching Begin");
+		return;
+	}
+
+	size_t scopeId = scopeStack_.back();
+	scopeStack_.pop_back();
+
+	GLProfilerScope &scope = scopes_[scopeId];
+	scope.endQueryId = numQueries_;
+
+	glQueryCounter(queries_[numQueries_], GL_TIMESTAMP);
+	numQueries_++;
+}

Common/GPU/OpenGL/GLProfiler.h

@@ -0,0 +1,53 @@
+#pragma once
+
+#include <vector>
+#include <string>
+#include <cstdint>
+
+#include "Common/GPU/OpenGL/GLCommon.h"
+#include "Common/GPU/OpenGL/gl3stub.h"
+
+// Simple scoped based profiler for OpenGL, similar to VulkanProfiler.
+// Uses GL_EXT_disjoint_timer_query (GLES) or GL_ARB_timer_query (desktop GL).
+// Put the whole thing in a FrameData to allow for overlap.
+
+struct GLProfilerScope {
+	char name[52];  // to make a struct size of 64, just because
+	int startQueryId;
+	int endQueryId;
+	int level;
+};
+
+class GLProfiler {
+public:
+	void Init();
+	void Shutdown();
+
+	void BeginFrame();
+
+	void Begin(const char *fmt, ...)
+#ifdef __GNUC__
+		__attribute__((format(printf, 2, 3)))
+#endif
+		;
+	void End();
+
+	void SetEnabledPtr(bool *enabledPtr) {
+		enabledPtr_ = enabledPtr;
+	}
+
+	bool IsSupported() const { return supported_; }
+
+private:
+	bool supported_ = false;
+	bool firstFrame_ = true;
+	bool *enabledPtr_ = nullptr;
+
+	std::vector<GLuint> queries_;
+	std::vector<GLProfilerScope> scopes_;
+	int numQueries_ = 0;
+
+	std::vector<size_t> scopeStack_;
+
+	static const int MAX_QUERY_COUNT = 1024;
+};

Common/GPU/OpenGL/GLQueueRunner.cpp

@@ -75,9 +75,12 @@ void GLQueueRunner::CreateDeviceObjects() {
 #if !PPSSPP_ARCH(X86)  // Doesn't work on AMD for some reason. See issue #17787
 	useDebugGroups_ = !gl_extensions.IsGLES && gl_extensions.VersionGEThan(4, 3);
 #endif
+
+	profiler_.Init();
 }
 
 void GLQueueRunner::DestroyDeviceObjects() {
+	profiler_.Shutdown();
 	CHECK_GL_ERROR_IF_DEBUG();
 	if (gl_extensions.ARB_vertex_array_object) {
 		glDeleteVertexArrays(1, &globalVAO_);
@@ -667,6 +670,12 @@ void GLQueueRunner::RunSteps(const std::vector<GLRStep *> &steps, GLFrameData &f
 		}
 	}
 
+	// GPU timestamp profiling - read results from previous frame and prepare for new queries
+	if (frameData.profile.enabled) {
+		profiler_.SetEnabledPtr(&frameData.profile.enabled);
+	}
+	profiler_.BeginFrame();
+
 	CHECK_GL_ERROR_IF_DEBUG();
 	size_t renderCount = 0;
 	for (size_t i = 0; i < steps.size(); i++) {
@@ -680,24 +689,34 @@ void GLQueueRunner::RunSteps(const std::vector<GLRStep *> &steps, GLFrameData &f
 		switch (step.stepType) {
 		case GLRStepType::RENDER:
 			renderCount++;
+			profiler_.Begin("RenderPass %s", step.tag);
 			if (IsVREnabled()) {
 				PreprocessStepVR(&step);
 				PerformRenderPass(step, renderCount == 1, renderCount == totalRenderCount, frameData.profile);
 			} else {
 				PerformRenderPass(step, renderCount == 1, renderCount == totalRenderCount, frameData.profile);
 			}
+			profiler_.End();
 			break;
 		case GLRStepType::COPY:
+			profiler_.Begin("Copy");
 			PerformCopy(step);
+			profiler_.End();
 			break;
 		case GLRStepType::BLIT:
+			profiler_.Begin("Blit");
 			PerformBlit(step);
+			profiler_.End();
 			break;
 		case GLRStepType::READBACK:
+			profiler_.Begin("Readback");
 			PerformReadback(step);
+			profiler_.End();
 			break;
 		case GLRStepType::READBACK_IMAGE:
+			profiler_.Begin("ReadbackImage");
 			PerformReadbackImage(step);
+			profiler_.End();
 			break;
 		case GLRStepType::RENDER_SKIP:
 			break;

Common/GPU/OpenGL/GLQueueRunner.h

@@ -7,6 +7,7 @@
 
 #include "Common/GPU/OpenGL/GLCommon.h"
 #include "Common/GPU/OpenGL/GLFrameData.h"
+#include "Common/GPU/OpenGL/GLProfiler.h"
 #include "Common/GPU/DataFormat.h"
 #include "Common/GPU/Shader.h"
 #include "Common/GPU/thin3d.h"
@@ -417,6 +418,8 @@ private:
 	bool sawOutOfMemory_ = false;
 	bool useDebugGroups_ = false;
 
+	GLProfiler profiler_;
+
 	ErrorCallbackFn errorCallback_ = nullptr;
 	void *errorCallbackUserData_ = nullptr;
 };

Common/GPU/OpenGL/gl3stub.c

@@ -38,6 +38,10 @@ GLboolean gl3stubInit() {
     FIND_PROC(glEndQuery);
     FIND_PROC(glGetQueryiv);
     FIND_PROC(glGetQueryObjectuiv);
+    // GL_EXT_disjoint_timer_query - load EXT suffix versions into non-suffix pointers
+    glQueryCounter = (void*)eglGetProcAddress("glQueryCounterEXT");
+    glGetQueryObjecti64v = (void*)eglGetProcAddress("glGetQueryObjecti64vEXT");
+    glGetQueryObjectui64v = (void*)eglGetProcAddress("glGetQueryObjectui64vEXT");
     FIND_PROC(glUnmapBuffer);
     FIND_PROC(glGetBufferPointerv);
     FIND_PROC(glDrawBuffers);
@@ -271,6 +275,10 @@ GL_APICALL void           (* GL_APIENTRY glBeginQuery) (GLenum target, GLuint id
 GL_APICALL void           (* GL_APIENTRY glEndQuery) (GLenum target);
 GL_APICALL void           (* GL_APIENTRY glGetQueryiv) (GLenum target, GLenum pname, GLint* params);
 GL_APICALL void           (* GL_APIENTRY glGetQueryObjectuiv) (GLuint id, GLenum pname, GLuint* params);
+// GL_EXT_disjoint_timer_query / GL_ARB_timer_query
+GL_APICALL void           (* GL_APIENTRY glQueryCounter) (GLuint id, GLenum target);
+GL_APICALL void           (* GL_APIENTRY glGetQueryObjecti64v) (GLuint id, GLenum pname, GLint64* params);
+GL_APICALL void           (* GL_APIENTRY glGetQueryObjectui64v) (GLuint id, GLenum pname, GLuint64* params);
 GL_APICALL GLboolean      (* GL_APIENTRY glUnmapBuffer) (GLenum target);
 GL_APICALL void           (* GL_APIENTRY glGetBufferPointerv) (GLenum target, GLenum pname, GLvoid** params);
 GL_APICALL void           (* GL_APIENTRY glDrawBuffers) (GLsizei n, const GLenum* bufs);

Common/GPU/OpenGL/gl3stub.h

@@ -409,6 +409,10 @@ extern GL_APICALL void           (* GL_APIENTRY glBeginQuery) (GLenum target, GL
 extern GL_APICALL void           (* GL_APIENTRY glEndQuery) (GLenum target);
 extern GL_APICALL void           (* GL_APIENTRY glGetQueryiv) (GLenum target, GLenum pname, GLint* params);
 extern GL_APICALL void           (* GL_APIENTRY glGetQueryObjectuiv) (GLuint id, GLenum pname, GLuint* params);
+// GL_EXT_disjoint_timer_query / GL_ARB_timer_query
+extern GL_APICALL void           (* GL_APIENTRY glQueryCounter) (GLuint id, GLenum target);
+extern GL_APICALL void           (* GL_APIENTRY glGetQueryObjecti64v) (GLuint id, GLenum pname, GLint64* params);
+extern GL_APICALL void           (* GL_APIENTRY glGetQueryObjectui64v) (GLuint id, GLenum pname, GLuint64* params);
 extern GL_APICALL GLboolean      (* GL_APIENTRY glUnmapBuffer) (GLenum target);
 extern GL_APICALL void           (* GL_APIENTRY glGetBufferPointerv) (GLenum target, GLenum pname, GLvoid** params);
 extern GL_APICALL void           (* GL_APIENTRY glDrawBuffers) (GLsizei n, const GLenum* bufs);

android/jni/Android.mk

@@ -115,6 +115,7 @@ NATIVE_FILES :=\
   $(SRC)/Common/GPU/OpenGL/GLMemory.cpp \
   $(SRC)/Common/GPU/OpenGL/GLRenderManager.cpp \
   $(SRC)/Common/GPU/OpenGL/GLQueueRunner.cpp \
+  $(SRC)/Common/GPU/OpenGL/GLProfiler.cpp \
   $(SRC)/Common/GPU/OpenGL/DataFormatGL.cpp
 
 VULKAN_FILES := \

libretro/Makefile.common

@@ -476,6 +476,7 @@ SOURCES_CXX += \
 	$(COMMONDIR)/GPU/OpenGL/GLRenderManager.cpp \
 	$(COMMONDIR)/GPU/OpenGL/GLMemory.cpp \
 	$(COMMONDIR)/GPU/OpenGL/GLQueueRunner.cpp \
+	$(COMMONDIR)/GPU/OpenGL/GLProfiler.cpp \
 	$(COMMONDIR)/GPU/OpenGL/DataFormatGL.cpp \
 	$(COMMONDIR)/GPU/Vulkan/thin3d_vulkan.cpp \
 	$(COMMONDIR)/GPU/Vulkan/VulkanQueueRunner.cpp \