Merge branch 'master' into shader-decomp

.github/workflows/Hydra_Build.yml

@@ -107,7 +107,7 @@ jobs:
 
     - name: Install misc packages
       run: |
-       sudo apt-get update && sudo apt install libx11-dev libgl1-mesa-glx mesa-common-dev libfuse2 libwayland-dev
+       sudo apt-get update && sudo apt install libx11-dev libgl1 libglx-mesa0 mesa-common-dev libfuse2 libwayland-dev
       
     - name: Install newer Clang
       run: |
@@ -160,7 +160,7 @@ jobs:
 
     - name: Install misc packages
       run: |
-       sudo apt-get update && sudo apt install libx11-dev libgl1-mesa-glx mesa-common-dev libfuse2 libwayland-dev
+       sudo apt-get update && sudo apt install libx11-dev libgl1 libglx-mesa0 mesa-common-dev libfuse2 libwayland-dev
        
     - name: Setup Vulkan SDK
       uses: humbletim/setup-vulkan-sdk@v1.2.0

.github/workflows/Linux_AppImage_Build.yml

@@ -24,7 +24,7 @@ jobs:
       run: git submodule update --init --recursive
 
     - name: Install misc packages
-      run: sudo apt-get update && sudo apt install libx11-dev libgl1-mesa-glx mesa-common-dev libfuse2
+      run: sudo apt-get update && sudo apt install libx11-dev libgl1 libglx-mesa0 mesa-common-dev libfuse2
 
     - name: Install newer Clang
       run: |

.github/workflows/Linux_Build.yml

@@ -24,7 +24,7 @@ jobs:
       run: git submodule update --init --recursive
 
     - name: Install misc packages
-      run: sudo apt-get update && sudo apt install libx11-dev libgl1-mesa-glx mesa-common-dev
+      run: sudo apt-get update && sudo apt install libx11-dev libgl1 libglx-mesa0 mesa-common-dev
 
     - name: Install newer Clang
       run: |

.github/workflows/Qt_Build.yml

@@ -105,8 +105,7 @@ jobs:
 
     - name: Install misc packages
       run: |
-       sudo apt-get update && sudo apt install libx11-dev libgl1-mesa-glx mesa-common-dev libfuse2 libwayland-dev libgl1-mesa-dev
+       sudo apt-get update && sudo apt install libx11-dev libgl1 libglx-mesa0 mesa-common-dev libfuse2 libwayland-dev libgl1-mesa-dev
-       sudo add-apt-repository -y ppa:savoury1/qt-6-2
        sudo apt update
        sudo apt install qt6-base-dev qt6-base-private-dev
 

CMakeLists.txt

@@ -383,7 +383,7 @@ if(ENABLE_OPENGL)
     set(RENDERER_GL_INCLUDE_FILES third_party/opengl/opengl.hpp
         include/renderer_gl/renderer_gl.hpp include/renderer_gl/textures.hpp
         include/renderer_gl/surfaces.hpp include/renderer_gl/surface_cache.hpp
-        include/renderer_gl/gl_state.hpp
+        include/renderer_gl/gl_state.hpp include/renderer_gl/gl_driver.hpp
     )
 
     set(RENDERER_GL_SOURCE_FILES src/core/renderer_gl/renderer_gl.cpp

include/PICA/pica_frag_config.hpp

@@ -17,6 +17,7 @@ namespace PICA {
 			// enable == off means a CompareFunction of Always
 			BitField<0, 3, CompareFunction> alphaTestFunction;
 			BitField<3, 1, u32> depthMapEnable;
+			BitField<4, 4, LogicOpMode> logicOpMode;
 		};
 	};
 
@@ -214,6 +215,10 @@ namespace PICA {
 				(alphaTestConfig & 1) ? static_cast<PICA::CompareFunction>(alphaTestFunction) : PICA::CompareFunction::Always;
 			outConfig.depthMapEnable = regs[InternalRegs::DepthmapEnable] & 1;
 
+			// Shows if blending is enabled. If it is not enabled, then logic ops are enabled instead
+			const bool blendingEnabled = (regs[InternalRegs::ColourOperation] & (1 << 8)) != 0;
+			outConfig.logicOpMode = blendingEnabled ? LogicOpMode::Copy : LogicOpMode(Helpers::getBits<0, 4>(regs[InternalRegs::LogicOp]));
+
 			texConfig.texUnitConfig = regs[InternalRegs::TexUnitCfg];
 			texConfig.texEnvUpdateBuffer = regs[InternalRegs::TexEnvUpdateBuffer];
 

include/PICA/regs.hpp

@@ -396,6 +396,25 @@ namespace PICA {
 		GreaterOrEqual = 7,
 	};
 
+	enum class LogicOpMode : u32 {
+		Clear = 0,
+		And = 1,
+		ReverseAnd = 2,
+		Copy = 3,
+		Set = 4,
+		InvertedCopy = 5,
+		Nop = 6,
+		Invert = 7,
+		Nand = 8,
+		Or = 9,
+		Nor = 10,
+		Xor = 11,
+		Equiv = 12,
+		InvertedAnd = 13,
+		ReverseOr = 14,
+		InvertedOr = 15,
+	};
+
 	enum class FogMode : u32 {
 		Disabled = 0,
 		Fog = 5,

include/PICA/shader_gen.hpp

@@ -26,10 +26,11 @@ namespace PICA::ShaderGen {
 		bool isSamplerEnabled(u32 environmentID, u32 lutID);
 
 		void compileFog(std::string& shader, const PICA::FragmentConfig& config);
+		void compileLogicOps(std::string& shader, const PICA::FragmentConfig& config);
 
 	  public:
 		FragmentGenerator(API api, Language language) : api(api), language(language) {}
-		std::string generate(const PICA::FragmentConfig& config);
+		std::string generate(const PICA::FragmentConfig& config, void* driverInfo = nullptr);
 		std::string getDefaultVertexShader();
 		// For when PICA shader is acceleration is enabled. Turn the PICA shader source into a proper vertex shader
 		std::string getVertexShaderAccelerated(const std::string& picaSource, const PICA::VertConfig& vertConfig, bool usingUbershader);

include/renderer_gl/gl_driver.hpp

@@ -0,0 +1,12 @@
+#pragma once
+
+// Information about our OpenGL/OpenGL ES driver that we should keep track of
+// Stuff like whether specific extensions are supported, and potentially things like OpenGL context information
+namespace OpenGL {
+	struct Driver {
+		bool supportsExtFbFetch = false;
+		bool supportsArmFbFetch = false;
+
+		bool supportFbFetch() const { return supportsExtFbFetch || supportsArmFbFetch; }
+	};
+}  // namespace OpenGL

include/renderer_gl/renderer_gl.hpp

@@ -17,6 +17,7 @@
 #include "PICA/regs.hpp"
 #include "PICA/shader_gen.hpp"
 #include "gl/stream_buffer.h"
+#include "gl_driver.hpp"
 #include "gl_state.hpp"
 #include "helpers.hpp"
 #include "logger.hpp"
@@ -143,6 +144,7 @@ class RendererGL final : public Renderer {
 	OpenGL::Program& getSpecializedShader();
 
 	PICA::ShaderGen::FragmentGenerator fragShaderGen;
+	OpenGL::Driver driverInfo;
 
 	MAKE_LOG_FUNCTION(log, rendererLogger)
 	void setupBlending();

src/core/PICA/shader_gen_glsl.cpp

@@ -5,6 +5,10 @@
 #include "PICA/pica_frag_config.hpp"
 #include "PICA/regs.hpp"
 #include "PICA/shader_gen.hpp"
+
+// We can include the driver headers here since they shouldn't have any actual API-specific code
+#include "renderer_gl/gl_driver.hpp"
+
 using namespace PICA;
 using namespace PICA::ShaderGen;
 
@@ -38,6 +42,8 @@ static constexpr const char* uniformDefinition = R"(
 
 std::string FragmentGenerator::getDefaultVertexShader() {
 	std::string ret = "";
+	// Reserve some space (128KB) in the output string to avoid too many allocations later
+	ret.reserve(128 * 1024);
 
 	switch (api) {
 		case API::GL: ret += "#version 410 core"; break;
@@ -98,7 +104,7 @@ std::string FragmentGenerator::getDefaultVertexShader() {
 	return ret;
 }
 
-std::string FragmentGenerator::generate(const FragmentConfig& config) {
+std::string FragmentGenerator::generate(const FragmentConfig& config, void* driverInfo) {
 	std::string ret = "";
 
 	switch (api) {
@@ -107,6 +113,27 @@ std::string FragmentGenerator::generate(const FragmentConfig& config) {
 		default: break;
 	}
 
+	// For GLES we need to enable & use the framebuffer fetch extension in order to emulate logic ops
+	bool emitLogicOps = api == API::GLES && config.outConfig.logicOpMode != PICA::LogicOpMode::Copy && driverInfo != nullptr;
+
+	if (emitLogicOps) {
+		auto driver = static_cast<OpenGL::Driver*>(driverInfo);
+
+		// If the driver does not support framebuffer fetch at all, don't emit logic op code
+		if (!driver->supportFbFetch()) {
+			emitLogicOps = false;
+		}
+		
+		// Figure out which fb fetch extension we have and enable it
+		else {
+			if (driver->supportsExtFbFetch) {
+				ret += "\n#extension GL_EXT_shader_framebuffer_fetch : enable\n#define fb_color fragColor\n";
+			} else if (driver->supportsArmFbFetch) {
+				ret += "\n#extension GL_ARM_shader_framebuffer_fetch : enable\n#define fb_color gl_LastFragColorARM[0]\n";
+			}
+		}
+	}
+
 	bool unimplementedFlag = false;
 	if (api == API::GLES) {
 		ret += R"(
@@ -196,10 +223,13 @@ std::string FragmentGenerator::generate(const FragmentConfig& config) {
 	}
 
 	compileFog(ret, config);
-
 	applyAlphaTest(ret, config);
 
+	if (!emitLogicOps) {
 		ret += "fragColor = combinerOutput;\n}";  // End of main function
+	} else {
+		compileLogicOps(ret, config);
+	}
 
 	return ret;
 }
@@ -781,4 +811,28 @@ void main() {
 })";
 		return ret;
 	}
+
+void FragmentGenerator::compileLogicOps(std::string& shader, const PICA::FragmentConfig& config) {
+	if (api != API::GLES) [[unlikely]] {
+		Helpers::warn("Shadergen: Unsupported API for compileLogicOps");
+		shader += "fragColor = combinerOutput;\n}"; // End of main function
+
+		return;
+	}
+	
+	shader += "fragColor = ";
+	switch (config.outConfig.logicOpMode) {
+		case PICA::LogicOpMode::Copy: shader += "combinerOutput"; break;
+		case PICA::LogicOpMode::Nop: shader += "fb_color"; break;
+		case PICA::LogicOpMode::Clear: shader += "vec4(0.0)"; break;
+		case PICA::LogicOpMode::Set: shader += "vec4(1.0)"; break;
+		case PICA::LogicOpMode::InvertedCopy: shader += "vec4(uvec4(combinerOutput * 255.0) ^ uvec4(0xFFu)) * (1.0 / 255.0)"; break;
+
+		default:
+			shader += "combinerOutput";
+			Helpers::warn("Shadergen: Unimplemented logic op mode");
+			break;
+	}
+
+	shader += ";\n}"; // End of main function
 }

src/core/renderer_gl/renderer_gl.cpp

@@ -191,6 +191,10 @@ void RendererGL::initGraphicsContextInternal() {
 
 	reset();
 
+	// Populate our driver info structure
+	driverInfo.supportsExtFbFetch = GLAD_GL_EXT_shader_framebuffer_fetch != 0;
+	driverInfo.supportsArmFbFetch = GLAD_GL_ARM_shader_framebuffer_fetch != 0;
+
 	// Initialize the default vertex shader used with shadergen
 	std::string defaultShadergenVSSource = fragShaderGen.getDefaultVertexShader();
 	defaultShadergenVs.create({defaultShadergenVSSource.c_str(), defaultShadergenVSSource.size()}, OpenGL::Vertex);
@@ -834,6 +838,10 @@ OpenGL::Program& RendererGL::getSpecializedShader() {
 	constexpr uint fsUBOBlockBinding = 2;
 
 	PICA::FragmentConfig fsConfig(regs);
+	// If we're not on GLES, ignore the logic op configuration and don't generate redundant shaders for it, since we use hw logic ops
+#ifndef USING_GLES
+	fsConfig.outConfig.logicOpMode = PICA::LogicOpMode(0);
+#endif
 
 	OpenGL::Shader& fragShader = shaderCache.fragmentShaderCache[fsConfig];
 	if (!fragShader.exists()) {