From 1987169c8164bd9d6fbf9995ca016fe186c24101 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Henrik=20Rydg=C3=A5rd?= <hrydgard@gmail.com>
Date: Wed, 17 May 2023 18:35:13 +0200
Subject: [PATCH] OpenGL: When possible, avoid rebinding vertex arrays between
 glDrawArrays

Profitable optimization in DrawArrays-heavy games like GTA.
---
 Common/GPU/OpenGL/GLFrameData.h       |  1 +
 Common/GPU/OpenGL/GLQueueRunner.cpp   |  9 +++++----
 Common/GPU/OpenGL/GLQueueRunner.h     |  4 ++--
 Common/GPU/OpenGL/GLRenderManager.cpp | 11 ++++++++---
 4 files changed, 16 insertions(+), 9 deletions(-)

diff --git a/Common/GPU/OpenGL/GLFrameData.h b/Common/GPU/OpenGL/GLFrameData.h
index 571b3d8abe..7d8d61c6c3 100644
--- a/Common/GPU/OpenGL/GLFrameData.h
+++ b/Common/GPU/OpenGL/GLFrameData.h
@@ -39,6 +39,7 @@ struct GLQueueProfileContext {
 	bool enabled;
 	double cpuStartTime;
 	double cpuEndTime;
+	int drawArraysRebindsAvoided;
 };
 
 
diff --git a/Common/GPU/OpenGL/GLQueueRunner.cpp b/Common/GPU/OpenGL/GLQueueRunner.cpp
index 98c2095865..9c0e19e7ef 100644
--- a/Common/GPU/OpenGL/GLQueueRunner.cpp
+++ b/Common/GPU/OpenGL/GLQueueRunner.cpp
@@ -651,7 +651,7 @@ retry_depth:
 	currentReadHandle_ = fbo->handle;
 }
 
-void GLQueueRunner::RunSteps(const std::vector<GLRStep *> &steps, bool skipGLCalls, bool keepSteps, bool useVR) {
+void GLQueueRunner::RunSteps(const std::vector<GLRStep *> &steps, bool skipGLCalls, bool keepSteps, bool useVR, GLQueueProfileContext &profile) {
 	if (skipGLCalls) {
 		if (keepSteps) {
 			return;
@@ -713,9 +713,9 @@ void GLQueueRunner::RunSteps(const std::vector<GLRStep *> &steps, bool skipGLCal
 			if (IsVREnabled()) {
 				GLRStep vrStep = step;
 				PreprocessStepVR(&vrStep);
-				PerformRenderPass(vrStep, renderCount == 1, renderCount == totalRenderCount);
+				PerformRenderPass(vrStep, renderCount == 1, renderCount == totalRenderCount, profile);
 			} else {
-				PerformRenderPass(step, renderCount == 1, renderCount == totalRenderCount);
+				PerformRenderPass(step, renderCount == 1, renderCount == totalRenderCount, profile);
 			}
 			break;
 		case GLRStepType::COPY:
@@ -791,7 +791,7 @@ static void EnableDisableVertexArrays(uint32_t prevAttr, uint32_t newAttr) {
 	}
 }
 
-void GLQueueRunner::PerformRenderPass(const GLRStep &step, bool first, bool last) {
+void GLQueueRunner::PerformRenderPass(const GLRStep &step, bool first, bool last, GLQueueProfileContext &profile) {
 	CHECK_GL_ERROR_IF_DEBUG();
 
 	PerformBindFramebufferAsRenderTarget(step);
@@ -1216,6 +1216,7 @@ void GLQueueRunner::PerformRenderPass(const GLRStep &step, bool first, bool last
 						// Compatible draws.
 						offset = diff / layout->stride;
 						rebind = false;
+						profile.drawArraysRebindsAvoided++;
 					}
 				}
 				if (rebind) {
diff --git a/Common/GPU/OpenGL/GLQueueRunner.h b/Common/GPU/OpenGL/GLQueueRunner.h
index 9cdbfeee01..1aa7398eaa 100644
--- a/Common/GPU/OpenGL/GLQueueRunner.h
+++ b/Common/GPU/OpenGL/GLQueueRunner.h
@@ -357,7 +357,7 @@ public:
 
 	void RunInitSteps(const std::vector<GLRInitStep> &steps, bool skipGLCalls);
 
-	void RunSteps(const std::vector<GLRStep *> &steps, bool skipGLCalls, bool keepSteps, bool useVR);
+	void RunSteps(const std::vector<GLRStep *> &steps, bool skipGLCalls, bool keepSteps, bool useVR, GLQueueProfileContext &profile);
 
 	void CreateDeviceObjects();
 	void DestroyDeviceObjects();
@@ -382,7 +382,7 @@ private:
 	void InitCreateFramebuffer(const GLRInitStep &step);
 
 	void PerformBindFramebufferAsRenderTarget(const GLRStep &pass);
-	void PerformRenderPass(const GLRStep &pass, bool first, bool last);
+	void PerformRenderPass(const GLRStep &pass, bool first, bool last, GLQueueProfileContext &profile);
 	void PerformCopy(const GLRStep &pass);
 	void PerformBlit(const GLRStep &pass);
 	void PerformReadback(const GLRStep &pass);
diff --git a/Common/GPU/OpenGL/GLRenderManager.cpp b/Common/GPU/OpenGL/GLRenderManager.cpp
index 7d094a1d51..7cfcce91cd 100644
--- a/Common/GPU/OpenGL/GLRenderManager.cpp
+++ b/Common/GPU/OpenGL/GLRenderManager.cpp
@@ -193,7 +193,11 @@ std::string GLRenderManager::GetGpuProfileString() const {
 	const GLQueueProfileContext &profile = frameData_[curFrame].profile;
 
 	float cputime_ms = 1000.0f * (profile.cpuEndTime - profile.cpuStartTime);
-	return StringFromFormat("CPU time to run the list: %0.2f ms", cputime_ms);
+	return StringFromFormat(
+		"CPU time to run the list: %0.2f ms\n"
+		"Avoided DrawArrays rebinds: %d",
+		cputime_ms,
+		profile.drawArraysRebindsAvoided);
 }
 
 void GLRenderManager::BindFramebufferAsRenderTarget(GLRFramebuffer *fb, GLRRenderPassAction color, GLRRenderPassAction depth, GLRRenderPassAction stencil, uint32_t clearColor, float clearDepth, uint8_t clearStencil, const char *tag) {
@@ -429,17 +433,18 @@ bool GLRenderManager::Run(GLRRenderThreadTask &task) {
 
 	if (frameData.profile.enabled) {
 		frameData.profile.cpuStartTime = time_now_d();
+		frameData.profile.drawArraysRebindsAvoided = 0;
 	}
 
 	if (IsVREnabled()) {
 		int passes = GetVRPassesCount();
 		for (int i = 0; i < passes; i++) {
 			PreVRFrameRender(i);
-			queueRunner_.RunSteps(task.steps, skipGLCalls_, i < passes - 1, true);
+			queueRunner_.RunSteps(task.steps, skipGLCalls_, i < passes - 1, true, frameData.profile);
 			PostVRFrameRender();
 		}
 	} else {
-		queueRunner_.RunSteps(task.steps, skipGLCalls_, false, false);
+		queueRunner_.RunSteps(task.steps, skipGLCalls_, false, false, frameData.profile);
 	}
 
 	if (frameData.profile.enabled) {