Shader recompiler: Add UBO

CMakeLists.txt

@@ -249,6 +249,7 @@ set(HEADER_FILES include/emulator.hpp include/helpers.hpp include/termcolor.hpp
                  include/audio/dsp_core.hpp include/audio/null_core.hpp include/audio/teakra_core.hpp
                  include/audio/miniaudio_device.hpp include/ring_buffer.hpp include/bitfield.hpp include/audio/dsp_shared_mem.hpp
                  include/audio/hle_core.hpp include/capstone.hpp include/audio/aac.hpp include/PICA/pica_frag_config.hpp
+                 include/PICA/pica_frag_uniforms.hpp
 )
 
 cmrc_add_resource_library(

include/PICA/pica_frag_config.hpp

@@ -23,13 +23,11 @@ namespace PICA {
 		u32 texUnitConfig;
 		u32 texEnvUpdateBuffer;
 
-		// TODO: This should probably be a uniform
-		u32 texEnvBufferColor;
-
 		// There's 6 TEV stages, and each one is configured via 5 word-sized registers
 		std::array<u32, 5 * 6> tevConfigs;
 	};
 
+	// Config used for identifying unique fragment pipeline configurations
 	struct FragmentConfig {
 		OutputConfig outConfig;
 		TextureConfig texConfig;

include/PICA/pica_frag_uniforms.hpp

@@ -0,0 +1,18 @@
+#pragma once
+#include <array>
+#include <type_traits>
+
+#include "helpers.hpp"
+
+namespace PICA {
+	struct FragmentUniforms {
+		using vec3 = std::array<float, 3>;
+		using vec4 = std::array<float, 4>;
+		static constexpr usize tevStageCount = 6;
+
+		s32 alphaReference;
+
+		alignas(16) vec4 constantColors[tevStageCount];
+		alignas(16) vec4 tevBufferColor;
+	};
+}  // namespace PICA

include/renderer_gl/renderer_gl.hpp

@@ -67,7 +67,12 @@ class RendererGL final : public Renderer {
 	OpenGL::Framebuffer screenFramebuffer;
 	OpenGL::Texture blankTexture;
 
-	std::unordered_map<PICA::FragmentConfig, OpenGL::Program> shaderCache;
+	// Cached recompiled fragment shader
+	struct CachedProgram {
+		OpenGL::Program program;
+		uint uboBinding;
+	};
+	std::unordered_map<PICA::FragmentConfig, CachedProgram> shaderCache;
 
 	OpenGL::Framebuffer getColourFBO();
 	OpenGL::Texture getTexture(Texture& tex);

src/core/PICA/shader_gen_glsl.cpp

@@ -38,8 +38,6 @@ std::string FragmentGenerator::getVertexShader(const PICARegs& regs) {
 		out vec2 v_texcoord1;
 		out vec3 v_view;
 		out vec2 v_texcoord2;
-		flat out vec4 v_textureEnvColor[6];
-		flat out vec4 v_textureEnvBufferColor;
 
 		//out float gl_ClipDistance[2];
 
@@ -103,8 +101,6 @@ std::string FragmentGenerator::generate(const PICARegs& regs) {
 		in vec2 v_texcoord1;
 		in vec3 v_view;
 		in vec2 v_texcoord2;
-		flat in vec4 v_textureEnvColor[6];
-		flat in vec4 v_textureEnvBufferColor;
 
 		out vec4 fragColor;
 		uniform sampler2D u_tex0;
@@ -115,18 +111,21 @@ std::string FragmentGenerator::generate(const PICARegs& regs) {
 		uniform sampler1DArray u_tex_lighting_lut;
 #endif
 
-		vec4 tevSources[16];
+		layout(std140) uniform FragmentUniforms {
-		vec4 tevNextPreviousBuffer;
+			int alphaReference;
+
+			vec4 constantColors[6];
+			vec4 tevBufferColor;
+		};
 	)";
 
 	// Emit main function for fragment shader
 	// When not initialized, source 13 is set to vec4(0.0) and 15 is set to the vertex colour
 	ret += R"(
 		void main() {
-			tevSources[0] = v_colour;
-			tevSources[13] = vec4(0.0);  // Previous buffer colour 
-			tevSources[15] = v_colour;   // Previous combiner
 			vec4 combinerOutput = v_colour;   // Last TEV output
+			vec4 previousBuffer = vec4(0.0);  // Previous buffer
+			vec4 tevNextPreviousBuffer = tevBufferColor;			
 	)";
 
 	ret += R"(
@@ -148,7 +147,7 @@ std::string FragmentGenerator::generate(const PICARegs& regs) {
 
 	ret += "fragColor = combinerOutput;\n";
 	ret += "}"; // End of main function
-	ret += "\n\n\n\n\n\n\n\n\n\n\n\n\n";
+	ret += "\n\n\n\n\n\n\n\n\n\n";
 
 	return ret;
 }
@@ -201,6 +200,22 @@ void FragmentGenerator::compileTEV(std::string& shader, int stage, const PICAReg
 		shader += "combinerOutput = vec4(clamp(outputColor" + std::to_string(stage) + " * " + std::to_string(tev.getColorScale()) +
 				  ".0, vec3(0.0), vec3(1.0)), clamp(outputAlpha" + std::to_string(stage) + " * " + std::to_string(tev.getAlphaScale()) +
 				  ".0, 0.0, 1.0));\n";
+
+		shader += "previousBuffer = tevNextPreviousBuffer;\n";
+
+		// Update the "next previous buffer" if necessary
+		const u32 textureEnvUpdateBuffer = regs[InternalRegs::TexEnvUpdateBuffer];
+		if (stage < 4) {
+			// Check whether to update rgb
+			if ((textureEnvUpdateBuffer & (0x100 << stage))) {
+				shader += "tevNextPreviousBuffer.rgb = combinerOutput.rgb;\n";
+			}
+
+			// And whether to update alpha
+			if ((textureEnvUpdateBuffer & (0x1000u << stage))) {
+				shader += "tevNextPreviousBuffer.a = combinerOutput.a;\n";
+			}
+		}
 	}
 }
 
@@ -308,6 +323,8 @@ void FragmentGenerator::getSource(std::string& shader, TexEnvConfig::Source sour
 		}
 
 		case TexEnvConfig::Source::Previous: shader += "combinerOutput"; break;
+		case TexEnvConfig::Source::Constant: shader += "constantColors[" + std::to_string(index) + "]"; break;
+		case TexEnvConfig::Source::PreviousBuffer: shader += "previousBuffer"; break;
 
 		default:
 			Helpers::warn("Unimplemented TEV source: %d", static_cast<int>(source));

src/core/renderer_gl/renderer_gl.cpp

@@ -5,6 +5,7 @@
 #include <cmrc/cmrc.hpp>
 
 #include "PICA/float_types.hpp"
+#include "PICA/pica_frag_uniforms.hpp"
 #include "PICA/gpu.hpp"
 #include "PICA/regs.hpp"
 #include "math_util.hpp"
@@ -413,7 +414,7 @@ void RendererGL::drawVertices(PICA::PrimType primType, std::span<const Vertex> v
 	const float depthOffset = f24::fromRaw(regs[PICA::InternalRegs::DepthOffset] & 0xffffff).toFloat32();
 	const bool depthMapEnable = regs[PICA::InternalRegs::DepthmapEnable] & 1;
 
-	// Update depth uniforms
+	// Update ubershader uniforms
 	if (usingUbershader) {
 		if (oldDepthScale != depthScale) {
 			oldDepthScale = depthScale;
@@ -429,17 +430,15 @@ void RendererGL::drawVertices(PICA::PrimType primType, std::span<const Vertex> v
 			oldDepthmapEnable = depthMapEnable;
 			glUniform1i(ubershaderData.depthmapEnableLoc, depthMapEnable);
 		}
-	}
 
-	setupTextureEnvState();
-	bindTexturesToSlots();
-
-	if (usingUbershader) {
 		// Upload PICA Registers as a single uniform. The shader needs access to the rasterizer registers (for depth, starting from index 0x48)
 		// The texturing and the fragment lighting registers. Therefore we upload them all in one go to avoid multiple slow uniform updates
 		glUniform1uiv(ubershaderData.picaRegLoc, 0x200 - 0x48, &regs[0x48]);
 	}
 
+	setupTextureEnvState();
+	bindTexturesToSlots();
+
 	if (gpu.lightingLUTDirty) {
 		updateLightingLUT();
 	}
@@ -778,6 +777,8 @@ std::optional<ColourBuffer> RendererGL::getColourBuffer(u32 addr, PICA::ColorFmt
 }
 
 OpenGL::Program& RendererGL::getSpecializedShader() {
+	constexpr uint uboBlockBinding = 2;
+
 	PICA::FragmentConfig fsConfig;
 	auto& outConfig = fsConfig.outConfig;
 	auto& texConfig = fsConfig.texConfig;
@@ -788,7 +789,6 @@ OpenGL::Program& RendererGL::getSpecializedShader() {
 
 	texConfig.texUnitConfig = regs[InternalRegs::TexUnitCfg];
 	texConfig.texEnvUpdateBuffer = regs[InternalRegs::TexEnvUpdateBuffer];
-	texConfig.texEnvBufferColor = 0;
 
 	// Set up TEV stages
 	std::memcpy(&texConfig.tevConfigs[0 * 5], &regs[InternalRegs::TexEnv0Source], 5 * sizeof(u32));
@@ -798,7 +798,9 @@ OpenGL::Program& RendererGL::getSpecializedShader() {
 	std::memcpy(&texConfig.tevConfigs[4 * 5], &regs[InternalRegs::TexEnv4Source], 5 * sizeof(u32));
 	std::memcpy(&texConfig.tevConfigs[5 * 5], &regs[InternalRegs::TexEnv5Source], 5 * sizeof(u32));
 
-	OpenGL::Program& program = shaderCache[fsConfig];
+	CachedProgram& programEntry = shaderCache[fsConfig];
+	OpenGL::Program& program = programEntry.program;
+
 	if (!program.exists()) {
 		std::string vs = fragShaderGen.getVertexShader(regs);
 		std::string fs = fragShaderGen.generate(regs);
@@ -814,8 +816,50 @@ OpenGL::Program& RendererGL::getSpecializedShader() {
 		glUniform1i(OpenGL::uniformLocation(program, "u_tex1"), 1);
 		glUniform1i(OpenGL::uniformLocation(program, "u_tex2"), 2);
 		glUniform1i(OpenGL::uniformLocation(program, "u_tex_lighting_lut"), 3);
+
+		// Allocate memory for the program UBO
+		glGenBuffers(1, &programEntry.uboBinding);
+		glBindBuffer(GL_UNIFORM_BUFFER, programEntry.uboBinding);
+		glBufferData(GL_UNIFORM_BUFFER, sizeof(PICA::FragmentUniforms), nullptr, GL_DYNAMIC_DRAW);
+
+		// Set up the binding for our UBO. Sadly we can't specify it in the shader like normal people,
+		// As it's an OpenGL 4.2 feature that MacOS doesn't support...
+		uint uboIndex = glGetUniformBlockIndex(program.handle(), "FragmentUniforms");
+		glUniformBlockBinding(program.handle(), uboIndex, uboBlockBinding);
+		glBindBufferBase(GL_UNIFORM_BUFFER, uboBlockBinding, programEntry.uboBinding);
 	}
 
+	// Upload uniform data to our shader's UBO
+	PICA::FragmentUniforms uniforms;
+	uniforms.alphaReference = Helpers::getBits<8, 8>(regs[InternalRegs::AlphaTestConfig]);
+
+	// Set up the texenv buffer color
+	const u32 texEnvBufferColor = regs[InternalRegs::TexEnvBufferColor];
+	uniforms.tevBufferColor[0] = float(texEnvBufferColor & 0xFF) / 255.0f;
+	uniforms.tevBufferColor[1] = float((texEnvBufferColor >> 8) & 0xFF) / 255.0f;
+	uniforms.tevBufferColor[2] = float((texEnvBufferColor >> 16) & 0xFF) / 255.0f;
+	uniforms.tevBufferColor[3] = float((texEnvBufferColor >> 24) & 0xFF) / 255.0f;
+
+	// Set up the constant color for the 6 TEV stages
+	for (int i = 0; i < 6; i++) {
+		static constexpr std::array<u32, 6> ioBases = {
+			PICA::InternalRegs::TexEnv0Source, PICA::InternalRegs::TexEnv1Source, PICA::InternalRegs::TexEnv2Source,
+			PICA::InternalRegs::TexEnv3Source, PICA::InternalRegs::TexEnv4Source, PICA::InternalRegs::TexEnv5Source,
+		};
+
+		auto& vec = uniforms.constantColors[i];
+		u32 base = ioBases[i];
+		u32 color = regs[base + 3];
+
+		vec[0] = float(color & 0xFF) / 255.0f;
+		vec[1] = float((color >> 8) & 0xFF) / 255.0f;
+		vec[2] = float((color >> 16) & 0xFF) / 255.0f;
+		vec[3] = float((color >> 24) & 0xFF) / 255.0f;
+	}
+
+	glBindBuffer(GL_UNIFORM_BUFFER, programEntry.uboBinding);
+	glBufferSubData(GL_UNIFORM_BUFFER, 0, sizeof(PICA::FragmentUniforms), &uniforms);
+
 	return program;
 }