Merge pull request #1187 from bunnei/shader-gen

GLSL Fragment Shader Generation

src/video_core/CMakeLists.txt

@@ -1,6 +1,7 @@
 set(SRCS
             renderer_opengl/gl_rasterizer.cpp
             renderer_opengl/gl_rasterizer_cache.cpp
+            renderer_opengl/gl_shader_gen.cpp
             renderer_opengl/gl_shader_util.cpp
             renderer_opengl/gl_state.cpp
             renderer_opengl/renderer_opengl.cpp
@@ -21,8 +22,8 @@ set(HEADERS
             renderer_opengl/gl_rasterizer.h
             renderer_opengl/gl_rasterizer_cache.h
             renderer_opengl/gl_resource_manager.h
+            renderer_opengl/gl_shader_gen.h
             renderer_opengl/gl_shader_util.h
-            renderer_opengl/gl_shaders.h
             renderer_opengl/gl_state.h
             renderer_opengl/pica_to_gl.h
             renderer_opengl/renderer_opengl.h

src/video_core/pica.h

@@ -317,6 +317,7 @@ struct Regs {
         };
 
         union {
+            u32 sources_raw;
             BitField< 0, 4, Source> color_source1;
             BitField< 4, 4, Source> color_source2;
             BitField< 8, 4, Source> color_source3;
@@ -326,6 +327,7 @@ struct Regs {
         };
 
         union {
+            u32 modifiers_raw;
             BitField< 0, 4, ColorModifier> color_modifier1;
             BitField< 4, 4, ColorModifier> color_modifier2;
             BitField< 8, 4, ColorModifier> color_modifier3;
@@ -335,6 +337,7 @@ struct Regs {
         };
 
         union {
+            u32 ops_raw;
             BitField< 0, 4, Operation> color_op;
             BitField<16, 4, Operation> alpha_op;
         };
@@ -348,6 +351,7 @@ struct Regs {
         };
 
         union {
+            u32 scales_raw;
             BitField< 0, 2, u32> color_scale;
             BitField<16, 2, u32> alpha_scale;
         };

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -8,6 +8,8 @@
 #include <glad/glad.h>
 
 #include "common/color.h"
+#include "common/file_util.h"
+#include "common/make_unique.h"
 #include "common/math_util.h"
 #include "common/microprofile.h"
 #include "common/profiler.h"
@@ -19,7 +21,7 @@
 #include "video_core/pica.h"
 #include "video_core/utils.h"
 #include "video_core/renderer_opengl/gl_rasterizer.h"
-#include "video_core/renderer_opengl/gl_shaders.h"
+#include "video_core/renderer_opengl/gl_shader_gen.h"
 #include "video_core/renderer_opengl/gl_shader_util.h"
 #include "video_core/renderer_opengl/pica_to_gl.h"
 
@@ -38,36 +40,6 @@ RasterizerOpenGL::RasterizerOpenGL() : last_fb_color_addr(0), last_fb_depth_addr
 RasterizerOpenGL::~RasterizerOpenGL() { }
 
 void RasterizerOpenGL::InitObjects() {
-    // Create the hardware shader program and get attrib/uniform locations
-    shader.Create(GLShaders::g_vertex_shader_hw, GLShaders::g_fragment_shader_hw);
-    attrib_position = glGetAttribLocation(shader.handle, "vert_position");
-    attrib_color = glGetAttribLocation(shader.handle, "vert_color");
-    attrib_texcoords = glGetAttribLocation(shader.handle, "vert_texcoords");
-
-    uniform_alphatest_enabled = glGetUniformLocation(shader.handle, "alphatest_enabled");
-    uniform_alphatest_func = glGetUniformLocation(shader.handle, "alphatest_func");
-    uniform_alphatest_ref = glGetUniformLocation(shader.handle, "alphatest_ref");
-
-    uniform_tex = glGetUniformLocation(shader.handle, "tex");
-
-    uniform_tev_combiner_buffer_color = glGetUniformLocation(shader.handle, "tev_combiner_buffer_color");
-
-    const auto tev_stages = Pica::g_state.regs.GetTevStages();
-    for (unsigned tev_stage_index = 0; tev_stage_index < tev_stages.size(); ++tev_stage_index) {
-        auto& uniform_tev_cfg = uniform_tev_cfgs[tev_stage_index];
-
-        std::string tev_ref_str = "tev_cfgs[" + std::to_string(tev_stage_index) + "]";
-        uniform_tev_cfg.enabled = glGetUniformLocation(shader.handle, (tev_ref_str + ".enabled").c_str());
-        uniform_tev_cfg.color_sources = glGetUniformLocation(shader.handle, (tev_ref_str + ".color_sources").c_str());
-        uniform_tev_cfg.alpha_sources = glGetUniformLocation(shader.handle, (tev_ref_str + ".alpha_sources").c_str());
-        uniform_tev_cfg.color_modifiers = glGetUniformLocation(shader.handle, (tev_ref_str + ".color_modifiers").c_str());
-        uniform_tev_cfg.alpha_modifiers = glGetUniformLocation(shader.handle, (tev_ref_str + ".alpha_modifiers").c_str());
-        uniform_tev_cfg.color_alpha_op = glGetUniformLocation(shader.handle, (tev_ref_str + ".color_alpha_op").c_str());
-        uniform_tev_cfg.color_alpha_multiplier = glGetUniformLocation(shader.handle, (tev_ref_str + ".color_alpha_multiplier").c_str());
-        uniform_tev_cfg.const_color = glGetUniformLocation(shader.handle, (tev_ref_str + ".const_color").c_str());
-        uniform_tev_cfg.updates_combiner_buffer_color_alpha = glGetUniformLocation(shader.handle, (tev_ref_str + ".updates_combiner_buffer_color_alpha").c_str());
-    }
-
     // Create sampler objects
     for (size_t i = 0; i < texture_samplers.size(); ++i) {
         texture_samplers[i].Create();
@@ -78,29 +50,25 @@ void RasterizerOpenGL::InitObjects() {
     vertex_buffer.Create();
     vertex_array.Create();
 
-    // Update OpenGL state
     state.draw.vertex_array = vertex_array.handle;
     state.draw.vertex_buffer = vertex_buffer.handle;
-    state.draw.shader_program = shader.handle;
-
     state.Apply();
 
-    // Set the texture samplers to correspond to different texture units
-    glUniform1i(uniform_tex, 0);
-    glUniform1i(uniform_tex + 1, 1);
-    glUniform1i(uniform_tex + 2, 2);
-
     // Set vertex attributes
-    glVertexAttribPointer(attrib_position, 4, GL_FLOAT, GL_FALSE, sizeof(HardwareVertex), (GLvoid*)offsetof(HardwareVertex, position));
-    glVertexAttribPointer(attrib_color, 4, GL_FLOAT, GL_FALSE, sizeof(HardwareVertex), (GLvoid*)offsetof(HardwareVertex, color));
-    glVertexAttribPointer(attrib_texcoords, 2, GL_FLOAT, GL_FALSE, sizeof(HardwareVertex), (GLvoid*)offsetof(HardwareVertex, tex_coord0));
-    glVertexAttribPointer(attrib_texcoords + 1, 2, GL_FLOAT, GL_FALSE, sizeof(HardwareVertex), (GLvoid*)offsetof(HardwareVertex, tex_coord1));
-    glVertexAttribPointer(attrib_texcoords + 2, 2, GL_FLOAT, GL_FALSE, sizeof(HardwareVertex), (GLvoid*)offsetof(HardwareVertex, tex_coord2));
-    glEnableVertexAttribArray(attrib_position);
-    glEnableVertexAttribArray(attrib_color);
-    glEnableVertexAttribArray(attrib_texcoords);
-    glEnableVertexAttribArray(attrib_texcoords + 1);
-    glEnableVertexAttribArray(attrib_texcoords + 2);
+    glVertexAttribPointer(GLShader::ATTRIBUTE_POSITION, 4, GL_FLOAT, GL_FALSE, sizeof(HardwareVertex), (GLvoid*)offsetof(HardwareVertex, position));
+    glEnableVertexAttribArray(GLShader::ATTRIBUTE_POSITION);
+
+    glVertexAttribPointer(GLShader::ATTRIBUTE_COLOR, 4, GL_FLOAT, GL_FALSE, sizeof(HardwareVertex), (GLvoid*)offsetof(HardwareVertex, color));
+    glEnableVertexAttribArray(GLShader::ATTRIBUTE_COLOR);
+
+    glVertexAttribPointer(GLShader::ATTRIBUTE_TEXCOORD0, 2, GL_FLOAT, GL_FALSE, sizeof(HardwareVertex), (GLvoid*)offsetof(HardwareVertex, tex_coord0));
+    glVertexAttribPointer(GLShader::ATTRIBUTE_TEXCOORD1, 2, GL_FLOAT, GL_FALSE, sizeof(HardwareVertex), (GLvoid*)offsetof(HardwareVertex, tex_coord1));
+    glVertexAttribPointer(GLShader::ATTRIBUTE_TEXCOORD2, 2, GL_FLOAT, GL_FALSE, sizeof(HardwareVertex), (GLvoid*)offsetof(HardwareVertex, tex_coord2));
+    glEnableVertexAttribArray(GLShader::ATTRIBUTE_TEXCOORD0);
+    glEnableVertexAttribArray(GLShader::ATTRIBUTE_TEXCOORD1);
+    glEnableVertexAttribArray(GLShader::ATTRIBUTE_TEXCOORD2);
+
+    SetShader();
 
     // Create textures for OGL framebuffer that will be rendered to, initially 1x1 to succeed in framebuffer creation
     fb_color_texture.texture.Create();
@@ -150,61 +118,15 @@ void RasterizerOpenGL::InitObjects() {
 }
 
 void RasterizerOpenGL::Reset() {
-    const auto& regs = Pica::g_state.regs;
-
     SyncCullMode();
     SyncBlendEnabled();
     SyncBlendFuncs();
     SyncBlendColor();
-    SyncAlphaTest();
     SyncLogicOp();
     SyncStencilTest();
     SyncDepthTest();
 
-    // TEV stage 0
-    SyncTevSources(0, regs.tev_stage0);
-    SyncTevModifiers(0, regs.tev_stage0);
-    SyncTevOps(0, regs.tev_stage0);
-    SyncTevColor(0, regs.tev_stage0);
-    SyncTevMultipliers(0, regs.tev_stage0);
-
-    // TEV stage 1
-    SyncTevSources(1, regs.tev_stage1);
-    SyncTevModifiers(1, regs.tev_stage1);
-    SyncTevOps(1, regs.tev_stage1);
-    SyncTevColor(1, regs.tev_stage1);
-    SyncTevMultipliers(1, regs.tev_stage1);
-
-    // TEV stage 2
-    SyncTevSources(2, regs.tev_stage2);
-    SyncTevModifiers(2, regs.tev_stage2);
-    SyncTevOps(2, regs.tev_stage2);
-    SyncTevColor(2, regs.tev_stage2);
-    SyncTevMultipliers(2, regs.tev_stage2);
-
-    // TEV stage 3
-    SyncTevSources(3, regs.tev_stage3);
-    SyncTevModifiers(3, regs.tev_stage3);
-    SyncTevOps(3, regs.tev_stage3);
-    SyncTevColor(3, regs.tev_stage3);
-    SyncTevMultipliers(3, regs.tev_stage3);
-
-    // TEV stage 4
-    SyncTevSources(4, regs.tev_stage4);
-    SyncTevModifiers(4, regs.tev_stage4);
-    SyncTevOps(4, regs.tev_stage4);
-    SyncTevColor(4, regs.tev_stage4);
-    SyncTevMultipliers(4, regs.tev_stage4);
-
-    // TEV stage 5
-    SyncTevSources(5, regs.tev_stage5);
-    SyncTevModifiers(5, regs.tev_stage5);
-    SyncTevOps(5, regs.tev_stage5);
-    SyncTevColor(5, regs.tev_stage5);
-    SyncTevMultipliers(5, regs.tev_stage5);
-
-    SyncCombinerColor();
-    SyncCombinerWriteFlags();
+    SetShader();
 
     res_cache.FullFlush();
 }
@@ -221,6 +143,11 @@ void RasterizerOpenGL::DrawTriangles() {
     SyncFramebuffer();
     SyncDrawState();
 
+    if (state.draw.shader_dirty) {
+        SetShader();
+        state.draw.shader_dirty = false;
+    }
+
     glBufferData(GL_ARRAY_BUFFER, vertex_batch.size() * sizeof(HardwareVertex), vertex_batch.data(), GL_STREAM_DRAW);
     glDrawArrays(GL_TRIANGLES, 0, (GLsizei)vertex_batch.size());
 
@@ -272,6 +199,7 @@ void RasterizerOpenGL::NotifyPicaRegisterChanged(u32 id) {
     // Alpha test
     case PICA_REG_INDEX(output_merger.alpha_test):
         SyncAlphaTest();
+        state.draw.shader_dirty = true;
         break;
 
     // Stencil test
@@ -290,117 +218,57 @@ void RasterizerOpenGL::NotifyPicaRegisterChanged(u32 id) {
         SyncLogicOp();
         break;
 
-    // TEV stage 0
+    // TEV stages
     case PICA_REG_INDEX(tev_stage0.color_source1):
-        SyncTevSources(0, regs.tev_stage0);
-        break;
     case PICA_REG_INDEX(tev_stage0.color_modifier1):
-        SyncTevModifiers(0, regs.tev_stage0);
-        break;
     case PICA_REG_INDEX(tev_stage0.color_op):
-        SyncTevOps(0, regs.tev_stage0);
-        break;
-    case PICA_REG_INDEX(tev_stage0.const_r):
-        SyncTevColor(0, regs.tev_stage0);
-        break;
     case PICA_REG_INDEX(tev_stage0.color_scale):
-        SyncTevMultipliers(0, regs.tev_stage0);
-        break;
-
-    // TEV stage 1
     case PICA_REG_INDEX(tev_stage1.color_source1):
-        SyncTevSources(1, regs.tev_stage1);
-        break;
     case PICA_REG_INDEX(tev_stage1.color_modifier1):
-        SyncTevModifiers(1, regs.tev_stage1);
-        break;
     case PICA_REG_INDEX(tev_stage1.color_op):
-        SyncTevOps(1, regs.tev_stage1);
-        break;
-    case PICA_REG_INDEX(tev_stage1.const_r):
-        SyncTevColor(1, regs.tev_stage1);
-        break;
     case PICA_REG_INDEX(tev_stage1.color_scale):
-        SyncTevMultipliers(1, regs.tev_stage1);
-        break;
-
-    // TEV stage 2
     case PICA_REG_INDEX(tev_stage2.color_source1):
-        SyncTevSources(2, regs.tev_stage2);
-        break;
     case PICA_REG_INDEX(tev_stage2.color_modifier1):
-        SyncTevModifiers(2, regs.tev_stage2);
-        break;
     case PICA_REG_INDEX(tev_stage2.color_op):
-        SyncTevOps(2, regs.tev_stage2);
-        break;
-    case PICA_REG_INDEX(tev_stage2.const_r):
-        SyncTevColor(2, regs.tev_stage2);
-        break;
     case PICA_REG_INDEX(tev_stage2.color_scale):
-        SyncTevMultipliers(2, regs.tev_stage2);
-        break;
-
-    // TEV stage 3
     case PICA_REG_INDEX(tev_stage3.color_source1):
-        SyncTevSources(3, regs.tev_stage3);
-        break;
     case PICA_REG_INDEX(tev_stage3.color_modifier1):
-        SyncTevModifiers(3, regs.tev_stage3);
-        break;
     case PICA_REG_INDEX(tev_stage3.color_op):
-        SyncTevOps(3, regs.tev_stage3);
-        break;
-    case PICA_REG_INDEX(tev_stage3.const_r):
-        SyncTevColor(3, regs.tev_stage3);
-        break;
     case PICA_REG_INDEX(tev_stage3.color_scale):
-        SyncTevMultipliers(3, regs.tev_stage3);
-        break;
-
-    // TEV stage 4
     case PICA_REG_INDEX(tev_stage4.color_source1):
-        SyncTevSources(4, regs.tev_stage4);
-        break;
     case PICA_REG_INDEX(tev_stage4.color_modifier1):
-        SyncTevModifiers(4, regs.tev_stage4);
-        break;
     case PICA_REG_INDEX(tev_stage4.color_op):
-        SyncTevOps(4, regs.tev_stage4);
+    case PICA_REG_INDEX(tev_stage4.color_scale):
+    case PICA_REG_INDEX(tev_stage5.color_source1):
+    case PICA_REG_INDEX(tev_stage5.color_modifier1):
+    case PICA_REG_INDEX(tev_stage5.color_op):
+    case PICA_REG_INDEX(tev_stage5.color_scale):
+    case PICA_REG_INDEX(tev_combiner_buffer_input):
+        state.draw.shader_dirty = true;
         break;
-    case PICA_REG_INDEX(tev_stage4.const_r):
-        SyncTevColor(4, regs.tev_stage4);
+    case PICA_REG_INDEX(tev_stage0.const_r):
+        SyncTevConstColor(0, regs.tev_stage0);
         break;
-    case PICA_REG_INDEX(tev_stage4.color_scale):
-        SyncTevMultipliers(4, regs.tev_stage4);
+    case PICA_REG_INDEX(tev_stage1.const_r):
+        SyncTevConstColor(1, regs.tev_stage1);
         break;
-
-    // TEV stage 5
-    case PICA_REG_INDEX(tev_stage5.color_source1):
-        SyncTevSources(5, regs.tev_stage5);
+    case PICA_REG_INDEX(tev_stage2.const_r):
+        SyncTevConstColor(2, regs.tev_stage2);
         break;
-    case PICA_REG_INDEX(tev_stage5.color_modifier1):
-        SyncTevModifiers(5, regs.tev_stage5);
+    case PICA_REG_INDEX(tev_stage3.const_r):
+        SyncTevConstColor(3, regs.tev_stage3);
         break;
-    case PICA_REG_INDEX(tev_stage5.color_op):
-        SyncTevOps(5, regs.tev_stage5);
+    case PICA_REG_INDEX(tev_stage4.const_r):
+        SyncTevConstColor(4, regs.tev_stage4);
         break;
     case PICA_REG_INDEX(tev_stage5.const_r):
-        SyncTevColor(5, regs.tev_stage5);
-        break;
-    case PICA_REG_INDEX(tev_stage5.color_scale):
-        SyncTevMultipliers(5, regs.tev_stage5);
+        SyncTevConstColor(5, regs.tev_stage5);
         break;
 
     // TEV combiner buffer color
     case PICA_REG_INDEX(tev_combiner_buffer_color):
         SyncCombinerColor();
         break;
-
-    // TEV combiner buffer write flags
-    case PICA_REG_INDEX(tev_combiner_buffer_input):
-        SyncCombinerWriteFlags();
-        break;
     }
 }
 
@@ -592,6 +460,41 @@ void RasterizerOpenGL::ReconfigureDepthTexture(DepthTextureInfo& texture, Pica::
     state.Apply();
 }
 
+void RasterizerOpenGL::SetShader() {
+    PicaShaderConfig config = PicaShaderConfig::CurrentConfig();
+    std::unique_ptr<PicaShader> shader = Common::make_unique<PicaShader>();
+
+    // Find (or generate) the GLSL shader for the current TEV state
+    auto cached_shader = shader_cache.find(config);
+    if (cached_shader != shader_cache.end()) {
+        current_shader = cached_shader->second.get();
+
+        state.draw.shader_program = current_shader->shader.handle;
+        state.Apply();
+    } else {
+        LOG_DEBUG(Render_OpenGL, "Creating new shader");
+
+        shader->shader.Create(GLShader::GenerateVertexShader().c_str(), GLShader::GenerateFragmentShader(config).c_str());
+
+        state.draw.shader_program = shader->shader.handle;
+        state.Apply();
+
+        // Set the texture samplers to correspond to different texture units
+        glUniform1i(PicaShader::Uniform::Texture0, 0);
+        glUniform1i(PicaShader::Uniform::Texture1, 1);
+        glUniform1i(PicaShader::Uniform::Texture2, 2);
+
+        current_shader = shader_cache.emplace(config, std::move(shader)).first->second.get();
+    }
+
+    // Update uniforms
+    SyncAlphaTest();
+    SyncCombinerColor();
+    auto& tev_stages = Pica::g_state.regs.GetTevStages();
+    for (int index = 0; index < tev_stages.size(); ++index)
+        SyncTevConstColor(index, tev_stages[index]);
+}
+
 void RasterizerOpenGL::SyncFramebuffer() {
     const auto& regs = Pica::g_state.regs;
 
@@ -712,9 +615,7 @@ void RasterizerOpenGL::SyncBlendColor() {
 
 void RasterizerOpenGL::SyncAlphaTest() {
     const auto& regs = Pica::g_state.regs;
-    glUniform1i(uniform_alphatest_enabled, regs.output_merger.alpha_test.enable);
-    glUniform1i(uniform_alphatest_func, (GLint)regs.output_merger.alpha_test.func.Value());
-    glUniform1f(uniform_alphatest_ref, regs.output_merger.alpha_test.ref / 255.0f);
+    glUniform1i(PicaShader::Uniform::AlphaTestRef, regs.output_merger.alpha_test.ref);
 }
 
 void RasterizerOpenGL::SyncLogicOp() {
@@ -744,56 +645,14 @@ void RasterizerOpenGL::SyncDepthTest() {
     state.depth.write_mask = regs.output_merger.depth_write_enable ? GL_TRUE : GL_FALSE;
 }
 
-void RasterizerOpenGL::SyncTevSources(unsigned stage_index, const Pica::Regs::TevStageConfig& config) {
-    GLint color_srcs[3] = { (GLint)config.color_source1.Value(),
-                            (GLint)config.color_source2.Value(),
-                            (GLint)config.color_source3.Value() };
-    GLint alpha_srcs[3] = { (GLint)config.alpha_source1.Value(),
-                            (GLint)config.alpha_source2.Value(),
-                            (GLint)config.alpha_source3.Value() };
-
-    glUniform3iv(uniform_tev_cfgs[stage_index].color_sources, 1, color_srcs);
-    glUniform3iv(uniform_tev_cfgs[stage_index].alpha_sources, 1, alpha_srcs);
-}
-
-void RasterizerOpenGL::SyncTevModifiers(unsigned stage_index, const Pica::Regs::TevStageConfig& config) {
-    GLint color_mods[3] = { (GLint)config.color_modifier1.Value(),
-                            (GLint)config.color_modifier2.Value(),
-                            (GLint)config.color_modifier3.Value() };
-    GLint alpha_mods[3] = { (GLint)config.alpha_modifier1.Value(),
-                            (GLint)config.alpha_modifier2.Value(),
-                            (GLint)config.alpha_modifier3.Value() };
-
-    glUniform3iv(uniform_tev_cfgs[stage_index].color_modifiers, 1, color_mods);
-    glUniform3iv(uniform_tev_cfgs[stage_index].alpha_modifiers, 1, alpha_mods);
-}
-
-void RasterizerOpenGL::SyncTevOps(unsigned stage_index, const Pica::Regs::TevStageConfig& config) {
-    glUniform2i(uniform_tev_cfgs[stage_index].color_alpha_op, (GLint)config.color_op.Value(), (GLint)config.alpha_op.Value());
-}
-
-void RasterizerOpenGL::SyncTevColor(unsigned stage_index, const Pica::Regs::TevStageConfig& config) {
-    auto const_color = PicaToGL::ColorRGBA8(config.const_color);
-    glUniform4fv(uniform_tev_cfgs[stage_index].const_color, 1, const_color.data());
-}
-
-void RasterizerOpenGL::SyncTevMultipliers(unsigned stage_index, const Pica::Regs::TevStageConfig& config) {
-    glUniform2i(uniform_tev_cfgs[stage_index].color_alpha_multiplier, config.GetColorMultiplier(), config.GetAlphaMultiplier());
-}
-
 void RasterizerOpenGL::SyncCombinerColor() {
     auto combiner_color = PicaToGL::ColorRGBA8(Pica::g_state.regs.tev_combiner_buffer_color.raw);
-    glUniform4fv(uniform_tev_combiner_buffer_color, 1, combiner_color.data());
+    glUniform4fv(PicaShader::Uniform::TevCombinerBufferColor, 1, combiner_color.data());
 }
 
-void RasterizerOpenGL::SyncCombinerWriteFlags() {
-    const auto& regs = Pica::g_state.regs;
-    const auto tev_stages = regs.GetTevStages();
-    for (unsigned tev_stage_index = 0; tev_stage_index < tev_stages.size(); ++tev_stage_index) {
-        glUniform2i(uniform_tev_cfgs[tev_stage_index].updates_combiner_buffer_color_alpha,
-                    regs.tev_combiner_buffer_input.TevStageUpdatesCombinerBufferColor(tev_stage_index),
-                    regs.tev_combiner_buffer_input.TevStageUpdatesCombinerBufferAlpha(tev_stage_index));
-    }
+void RasterizerOpenGL::SyncTevConstColor(int stage_index, const Pica::Regs::TevStageConfig& tev_stage) {
+    auto const_color = PicaToGL::ColorRGBA8(tev_stage.const_color);
+    glUniform4fv(PicaShader::Uniform::TevConstColors + stage_index, 1, const_color.data());
 }
 
 void RasterizerOpenGL::SyncDrawState() {
@@ -824,12 +683,6 @@ void RasterizerOpenGL::SyncDrawState() {
         }
     }
 
-    // Skip processing TEV stages that simply pass the previous stage results through
-    const auto tev_stages = regs.GetTevStages();
-    for (unsigned tev_stage_index = 0; tev_stage_index < tev_stages.size(); ++tev_stage_index) {
-        glUniform1i(uniform_tev_cfgs[tev_stage_index].enabled, !IsPassThroughTevStage(tev_stages[tev_stage_index]));
-    }
-
     state.Apply();
 }
 

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -4,15 +4,104 @@
 
 #pragma once
 
+#include <cstddef>
+#include <cstring>
+#include <memory>
 #include <vector>
+#include <unordered_map>
 
 #include "common/common_types.h"
+#include "common/hash.h"
 
+#include "video_core/pica.h"
 #include "video_core/hwrasterizer_base.h"
 #include "video_core/renderer_opengl/gl_rasterizer_cache.h"
 #include "video_core/renderer_opengl/gl_state.h"
 #include "video_core/shader/shader_interpreter.h"
 
+/**
+ * This struct contains all state used to generate the GLSL shader program that emulates the current
+ * Pica register configuration. This struct is used as a cache key for generated GLSL shader
+ * programs. The functions in gl_shader_gen.cpp should retrieve state from this struct only, not by
+ * directly accessing Pica registers. This should reduce the risk of bugs in shader generation where
+ * Pica state is not being captured in the shader cache key, thereby resulting in (what should be)
+ * two separate shaders sharing the same key.
+ */
+struct PicaShaderConfig {
+    /// Construct a PicaShaderConfig with the current Pica register configuration.
+    static PicaShaderConfig CurrentConfig() {
+        PicaShaderConfig res;
+        const auto& regs = Pica::g_state.regs;
+
+        res.alpha_test_func = regs.output_merger.alpha_test.enable ?
+            regs.output_merger.alpha_test.func.Value() : Pica::Regs::CompareFunc::Always;
+
+        // Copy relevant TevStageConfig fields only. We're doing this manually (instead of calling
+        // the GetTevStages() function) because BitField explicitly disables copies.
+
+        res.tev_stages[0].sources_raw = regs.tev_stage0.sources_raw;
+        res.tev_stages[1].sources_raw = regs.tev_stage1.sources_raw;
+        res.tev_stages[2].sources_raw = regs.tev_stage2.sources_raw;
+        res.tev_stages[3].sources_raw = regs.tev_stage3.sources_raw;
+        res.tev_stages[4].sources_raw = regs.tev_stage4.sources_raw;
+        res.tev_stages[5].sources_raw = regs.tev_stage5.sources_raw;
+
+        res.tev_stages[0].modifiers_raw = regs.tev_stage0.modifiers_raw;
+        res.tev_stages[1].modifiers_raw = regs.tev_stage1.modifiers_raw;
+        res.tev_stages[2].modifiers_raw = regs.tev_stage2.modifiers_raw;
+        res.tev_stages[3].modifiers_raw = regs.tev_stage3.modifiers_raw;
+        res.tev_stages[4].modifiers_raw = regs.tev_stage4.modifiers_raw;
+        res.tev_stages[5].modifiers_raw = regs.tev_stage5.modifiers_raw;
+
+        res.tev_stages[0].ops_raw = regs.tev_stage0.ops_raw;
+        res.tev_stages[1].ops_raw = regs.tev_stage1.ops_raw;
+        res.tev_stages[2].ops_raw = regs.tev_stage2.ops_raw;
+        res.tev_stages[3].ops_raw = regs.tev_stage3.ops_raw;
+        res.tev_stages[4].ops_raw = regs.tev_stage4.ops_raw;
+        res.tev_stages[5].ops_raw = regs.tev_stage5.ops_raw;
+
+        res.tev_stages[0].scales_raw = regs.tev_stage0.scales_raw;
+        res.tev_stages[1].scales_raw = regs.tev_stage1.scales_raw;
+        res.tev_stages[2].scales_raw = regs.tev_stage2.scales_raw;
+        res.tev_stages[3].scales_raw = regs.tev_stage3.scales_raw;
+        res.tev_stages[4].scales_raw = regs.tev_stage4.scales_raw;
+        res.tev_stages[5].scales_raw = regs.tev_stage5.scales_raw;
+
+        res.combiner_buffer_input =
+            regs.tev_combiner_buffer_input.update_mask_rgb.Value() |
+            regs.tev_combiner_buffer_input.update_mask_a.Value() << 4;
+
+        return res;
+    }
+
+    bool TevStageUpdatesCombinerBufferColor(unsigned stage_index) const {
+        return (stage_index < 4) && (combiner_buffer_input & (1 << stage_index));
+    }
+
+    bool TevStageUpdatesCombinerBufferAlpha(unsigned stage_index) const {
+        return (stage_index < 4) && ((combiner_buffer_input >> 4) & (1 << stage_index));
+    }
+
+    bool operator ==(const PicaShaderConfig& o) const {
+        return std::memcmp(this, &o, sizeof(PicaShaderConfig)) == 0;
+    };
+
+    Pica::Regs::CompareFunc alpha_test_func;
+    std::array<Pica::Regs::TevStageConfig, 6> tev_stages = {};
+    u8 combiner_buffer_input;
+};
+
+namespace std {
+
+template <>
+struct hash<PicaShaderConfig> {
+    size_t operator()(const PicaShaderConfig& k) const {
+        return Common::ComputeHash64(&k, sizeof(PicaShaderConfig));
+    }
+};
+
+} // namespace std
+
 class RasterizerOpenGL : public HWRasterizer {
 public:
 
@@ -45,20 +134,24 @@ public:
     /// Notify rasterizer that a 3DS memory region has been changed
     void NotifyFlush(PAddr addr, u32 size) override;
 
-private:
-    /// Structure used for managing texture environment states
-    struct TEVConfigUniforms {
-        GLuint enabled;
-        GLuint color_sources;
-        GLuint alpha_sources;
-        GLuint color_modifiers;
-        GLuint alpha_modifiers;
-        GLuint color_alpha_op;
-        GLuint color_alpha_multiplier;
-        GLuint const_color;
-        GLuint updates_combiner_buffer_color_alpha;
+    /// OpenGL shader generated for a given Pica register state
+    struct PicaShader {
+        /// OpenGL shader resource
+        OGLShader shader;
+
+        /// Fragment shader uniforms
+        enum Uniform : GLuint {
+            AlphaTestRef = 0,
+            TevConstColors = 1,
+            Texture0 = 7,
+            Texture1 = 8,
+            Texture2 = 9,
+            TevCombinerBufferColor = 10,
+        };
     };
 
+private:
+
     /// Structure used for storing information about color textures
     struct TextureInfo {
         OGLTexture texture;
@@ -129,6 +222,9 @@ private:
     /// Reconfigure the OpenGL depth texture to use the given format and dimensions
     void ReconfigureDepthTexture(DepthTextureInfo& texture, Pica::Regs::DepthFormat format, u32 width, u32 height);
 
+    /// Sets the OpenGL shader in accordance with the current PICA register state
+    void SetShader();
+
     /// Syncs the state and contents of the OpenGL framebuffer to match the current PICA framebuffer
     void SyncFramebuffer();
 
@@ -156,27 +252,12 @@ private:
     /// Syncs the depth test states to match the PICA register
     void SyncDepthTest();
 
-    /// Syncs the specified TEV stage's color and alpha sources to match the PICA register
-    void SyncTevSources(unsigned stage_index, const Pica::Regs::TevStageConfig& config);
-
-    /// Syncs the specified TEV stage's color and alpha modifiers to match the PICA register
-    void SyncTevModifiers(unsigned stage_index, const Pica::Regs::TevStageConfig& config);
-
-    /// Syncs the specified TEV stage's color and alpha combiner operations to match the PICA register
-    void SyncTevOps(unsigned stage_index, const Pica::Regs::TevStageConfig& config);
-
-    /// Syncs the specified TEV stage's constant color to match the PICA register
-    void SyncTevColor(unsigned stage_index, const Pica::Regs::TevStageConfig& config);
-
-    /// Syncs the specified TEV stage's color and alpha multipliers to match the PICA register
-    void SyncTevMultipliers(unsigned stage_index, const Pica::Regs::TevStageConfig& config);
+    /// Syncs the TEV constant color to match the PICA register
+    void SyncTevConstColor(int tev_index, const Pica::Regs::TevStageConfig& tev_stage);
 
     /// Syncs the TEV combiner color buffer to match the PICA register
     void SyncCombinerColor();
 
-    /// Syncs the TEV combiner write flags to match the PICA register
-    void SyncCombinerWriteFlags();
-
     /// Syncs the remaining OpenGL drawing state to match the current PICA state
     void SyncDrawState();
 
@@ -213,21 +294,11 @@ private:
     std::array<SamplerInfo, 3> texture_samplers;
     TextureInfo fb_color_texture;
     DepthTextureInfo fb_depth_texture;
-    OGLShader shader;
+
+    std::unordered_map<PicaShaderConfig, std::unique_ptr<PicaShader>> shader_cache;
+    const PicaShader* current_shader = nullptr;
+
     OGLVertexArray vertex_array;
     OGLBuffer vertex_buffer;
     OGLFramebuffer framebuffer;
-
-    // Hardware vertex shader
-    GLuint attrib_position;
-    GLuint attrib_color;
-    GLuint attrib_texcoords;
-
-    // Hardware fragment shader
-    GLuint uniform_alphatest_enabled;
-    GLuint uniform_alphatest_func;
-    GLuint uniform_alphatest_ref;
-    GLuint uniform_tex;
-    GLuint uniform_tev_combiner_buffer_color;
-    TEVConfigUniforms uniform_tev_cfgs[6];
 };

src/video_core/renderer_opengl/gl_resource_manager.h

@@ -71,7 +71,7 @@ public:
     /// Creates a new internal OpenGL resource and stores the handle
     void Create(const char* vert_shader, const char* frag_shader) {
         if (handle != 0) return;
-        handle = ShaderUtil::LoadShaders(vert_shader, frag_shader);
+        handle = GLShader::LoadProgram(vert_shader, frag_shader);
     }
 
     /// Deletes the internal OpenGL resource

src/video_core/renderer_opengl/gl_shader_gen.cpp

@@ -0,0 +1,388 @@
+// Copyright 2015 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "video_core/pica.h"
+#include "video_core/renderer_opengl/gl_rasterizer.h"
+#include "video_core/renderer_opengl/gl_shader_gen.h"
+
+using Pica::Regs;
+using TevStageConfig = Regs::TevStageConfig;
+
+namespace GLShader {
+
+/// Detects if a TEV stage is configured to be skipped (to avoid generating unnecessary code)
+static bool IsPassThroughTevStage(const TevStageConfig& stage) {
+    return (stage.color_op             == TevStageConfig::Operation::Replace &&
+            stage.alpha_op             == TevStageConfig::Operation::Replace &&
+            stage.color_source1        == TevStageConfig::Source::Previous &&
+            stage.alpha_source1        == TevStageConfig::Source::Previous &&
+            stage.color_modifier1      == TevStageConfig::ColorModifier::SourceColor &&
+            stage.alpha_modifier1      == TevStageConfig::AlphaModifier::SourceAlpha &&
+            stage.GetColorMultiplier() == 1 &&
+            stage.GetAlphaMultiplier() == 1);
+}
+
+/// Writes the specified TEV stage source component(s)
+static void AppendSource(std::string& out, TevStageConfig::Source source,
+        const std::string& index_name) {
+    using Source = TevStageConfig::Source;
+    switch (source) {
+    case Source::PrimaryColor:
+        out += "primary_color";
+        break;
+    case Source::PrimaryFragmentColor:
+        // HACK: Until we implement fragment lighting, use primary_color
+        out += "primary_color";
+        break;
+    case Source::SecondaryFragmentColor:
+        // HACK: Until we implement fragment lighting, use zero
+        out += "vec4(0.0)";
+        break;
+    case Source::Texture0:
+        out += "texture(tex[0], texcoord[0])";
+        break;
+    case Source::Texture1:
+        out += "texture(tex[1], texcoord[1])";
+        break;
+    case Source::Texture2:
+        out += "texture(tex[2], texcoord[2])";
+        break;
+    case Source::PreviousBuffer:
+        out += "combiner_buffer";
+        break;
+    case Source::Constant:
+        ((out += "const_color[") += index_name) += ']';
+        break;
+    case Source::Previous:
+        out += "last_tex_env_out";
+        break;
+    default:
+        out += "vec4(0.0)";
+        LOG_CRITICAL(Render_OpenGL, "Unknown source op %u", source);
+        break;
+    }
+}
+
+/// Writes the color components to use for the specified TEV stage color modifier
+static void AppendColorModifier(std::string& out, TevStageConfig::ColorModifier modifier,
+        TevStageConfig::Source source, const std::string& index_name) {
+    using ColorModifier = TevStageConfig::ColorModifier;
+    switch (modifier) {
+    case ColorModifier::SourceColor:
+        AppendSource(out, source, index_name);
+        out += ".rgb";
+        break;
+    case ColorModifier::OneMinusSourceColor:
+        out += "vec3(1.0) - ";
+        AppendSource(out, source, index_name);
+        out += ".rgb";
+        break;
+    case ColorModifier::SourceAlpha:
+        AppendSource(out, source, index_name);
+        out += ".aaa";
+        break;
+    case ColorModifier::OneMinusSourceAlpha:
+        out += "vec3(1.0) - ";
+        AppendSource(out, source, index_name);
+        out += ".aaa";
+        break;
+    case ColorModifier::SourceRed:
+        AppendSource(out, source, index_name);
+        out += ".rrr";
+        break;
+    case ColorModifier::OneMinusSourceRed:
+        out += "vec3(1.0) - ";
+        AppendSource(out, source, index_name);
+        out += ".rrr";
+        break;
+    case ColorModifier::SourceGreen:
+        AppendSource(out, source, index_name);
+        out += ".ggg";
+        break;
+    case ColorModifier::OneMinusSourceGreen:
+        out += "vec3(1.0) - ";
+        AppendSource(out, source, index_name);
+        out += ".ggg";
+        break;
+    case ColorModifier::SourceBlue:
+        AppendSource(out, source, index_name);
+        out += ".bbb";
+        break;
+    case ColorModifier::OneMinusSourceBlue:
+        out += "vec3(1.0) - ";
+        AppendSource(out, source, index_name);
+        out += ".bbb";
+        break;
+    default:
+        out += "vec3(0.0)";
+        LOG_CRITICAL(Render_OpenGL, "Unknown color modifier op %u", modifier);
+        break;
+    }
+}
+
+/// Writes the alpha component to use for the specified TEV stage alpha modifier
+static void AppendAlphaModifier(std::string& out, TevStageConfig::AlphaModifier modifier,
+        TevStageConfig::Source source, const std::string& index_name) {
+    using AlphaModifier = TevStageConfig::AlphaModifier;
+    switch (modifier) {
+    case AlphaModifier::SourceAlpha:
+        AppendSource(out, source, index_name);
+        out += ".a";
+        break;
+    case AlphaModifier::OneMinusSourceAlpha:
+        out += "1.0 - ";
+        AppendSource(out, source, index_name);
+        out += ".a";
+        break;
+    case AlphaModifier::SourceRed:
+        AppendSource(out, source, index_name);
+        out += ".r";
+        break;
+    case AlphaModifier::OneMinusSourceRed:
+        out += "1.0 - ";
+        AppendSource(out, source, index_name);
+        out += ".r";
+        break;
+    case AlphaModifier::SourceGreen:
+        AppendSource(out, source, index_name);
+        out += ".g";
+        break;
+    case AlphaModifier::OneMinusSourceGreen:
+        out += "1.0 - ";
+        AppendSource(out, source, index_name);
+        out += ".g";
+        break;
+    case AlphaModifier::SourceBlue:
+        AppendSource(out, source, index_name);
+        out += ".b";
+        break;
+    case AlphaModifier::OneMinusSourceBlue:
+        out += "1.0 - ";
+        AppendSource(out, source, index_name);
+        out += ".b";
+        break;
+    default:
+        out += "0.0";
+        LOG_CRITICAL(Render_OpenGL, "Unknown alpha modifier op %u", modifier);
+        break;
+    }
+}
+
+/// Writes the combiner function for the color components for the specified TEV stage operation
+static void AppendColorCombiner(std::string& out, TevStageConfig::Operation operation,
+        const std::string& variable_name) {
+    out += "clamp(";
+    using Operation = TevStageConfig::Operation;
+    switch (operation) {
+    case Operation::Replace:
+        out += variable_name + "[0]";
+        break;
+    case Operation::Modulate:
+        out += variable_name + "[0] * " + variable_name + "[1]";
+        break;
+    case Operation::Add:
+        out += variable_name + "[0] + " + variable_name + "[1]";
+        break;
+    case Operation::AddSigned:
+        out += variable_name + "[0] + " + variable_name + "[1] - vec3(0.5)";
+        break;
+    case Operation::Lerp:
+        // TODO(bunnei): Verify if HW actually does this per-component, otherwise we can just use builtin lerp
+        out += variable_name + "[0] * " + variable_name + "[2] + " + variable_name + "[1] * (vec3(1.0) - " + variable_name + "[2])";
+        break;
+    case Operation::Subtract:
+        out += variable_name + "[0] - " + variable_name + "[1]";
+        break;
+    case Operation::MultiplyThenAdd:
+        out += variable_name + "[0] * " + variable_name + "[1] + " + variable_name + "[2]";
+        break;
+    case Operation::AddThenMultiply:
+        out += "min(" + variable_name + "[0] + " + variable_name + "[1], vec3(1.0)) * " + variable_name + "[2]";
+        break;
+    default:
+        out += "vec3(0.0)";
+        LOG_CRITICAL(Render_OpenGL, "Unknown color combiner operation: %u", operation);
+        break;
+    }
+    out += ", vec3(0.0), vec3(1.0))"; // Clamp result to 0.0, 1.0
+}
+
+/// Writes the combiner function for the alpha component for the specified TEV stage operation
+static void AppendAlphaCombiner(std::string& out, TevStageConfig::Operation operation,
+        const std::string& variable_name) {
+    out += "clamp(";
+    using Operation = TevStageConfig::Operation;
+    switch (operation) {
+    case Operation::Replace:
+        out += variable_name + "[0]";
+        break;
+    case Operation::Modulate:
+        out += variable_name + "[0] * " + variable_name + "[1]";
+        break;
+    case Operation::Add:
+        out += variable_name + "[0] + " + variable_name + "[1]";
+        break;
+    case Operation::AddSigned:
+        out += variable_name + "[0] + " + variable_name + "[1] - 0.5";
+        break;
+    case Operation::Lerp:
+        out += variable_name + "[0] * " + variable_name + "[2] + " + variable_name + "[1] * (1.0 - " + variable_name + "[2])";
+        break;
+    case Operation::Subtract:
+        out += variable_name + "[0] - " + variable_name + "[1]";
+        break;
+    case Operation::MultiplyThenAdd:
+        out += variable_name + "[0] * " + variable_name + "[1] + " + variable_name + "[2]";
+        break;
+    case Operation::AddThenMultiply:
+        out += "min(" + variable_name + "[0] + " + variable_name + "[1], 1.0) * " + variable_name + "[2]";
+        break;
+    default:
+        out += "0.0";
+        LOG_CRITICAL(Render_OpenGL, "Unknown alpha combiner operation: %u", operation);
+        break;
+    }
+    out += ", 0.0, 1.0)";
+}
+
+/// Writes the if-statement condition used to evaluate alpha testing
+static void AppendAlphaTestCondition(std::string& out, Regs::CompareFunc func) {
+    using CompareFunc = Regs::CompareFunc;
+    switch (func) {
+    case CompareFunc::Never:
+        out += "true";
+        break;
+    case CompareFunc::Always:
+        out += "false";
+        break;
+    case CompareFunc::Equal:
+    case CompareFunc::NotEqual:
+    case CompareFunc::LessThan:
+    case CompareFunc::LessThanOrEqual:
+    case CompareFunc::GreaterThan:
+    case CompareFunc::GreaterThanOrEqual:
+    {
+        static const char* op[] = { "!=", "==", ">=", ">", "<=", "<", };
+        unsigned index = (unsigned)func - (unsigned)CompareFunc::Equal;
+        out += "int(last_tex_env_out.a * 255.0f) " + std::string(op[index]) + " alphatest_ref";
+        break;
+    }
+
+    default:
+        out += "false";
+        LOG_CRITICAL(Render_OpenGL, "Unknown alpha test condition %u", func);
+        break;
+    }
+}
+
+/// Writes the code to emulate the specified TEV stage
+static void WriteTevStage(std::string& out, const PicaShaderConfig& config, unsigned index) {
+    auto& stage = config.tev_stages[index];
+    if (!IsPassThroughTevStage(stage)) {
+        std::string index_name = std::to_string(index);
+
+        out += "vec3 color_results_" + index_name + "[3] = vec3[3](";
+        AppendColorModifier(out, stage.color_modifier1, stage.color_source1, index_name);
+        out += ", ";
+        AppendColorModifier(out, stage.color_modifier2, stage.color_source2, index_name);
+        out += ", ";
+        AppendColorModifier(out, stage.color_modifier3, stage.color_source3, index_name);
+        out += ");\n";
+
+        out += "vec3 color_output_" + index_name + " = ";
+        AppendColorCombiner(out, stage.color_op, "color_results_" + index_name);
+        out += ";\n";
+
+        out += "float alpha_results_" + index_name + "[3] = float[3](";
+        AppendAlphaModifier(out, stage.alpha_modifier1, stage.alpha_source1, index_name);
+        out += ", ";
+        AppendAlphaModifier(out, stage.alpha_modifier2, stage.alpha_source2, index_name);
+        out += ", ";
+        AppendAlphaModifier(out, stage.alpha_modifier3, stage.alpha_source3, index_name);
+        out += ");\n";
+
+        out += "float alpha_output_" + index_name + " = ";
+        AppendAlphaCombiner(out, stage.alpha_op, "alpha_results_" + index_name);
+        out += ";\n";
+
+        out += "last_tex_env_out = vec4("
+            "clamp(color_output_" + index_name + " * " + std::to_string(stage.GetColorMultiplier()) + ".0, vec3(0.0), vec3(1.0)),"
+            "clamp(alpha_output_" + index_name + " * " + std::to_string(stage.GetAlphaMultiplier()) + ".0, 0.0, 1.0));\n";
+    }
+
+    if (config.TevStageUpdatesCombinerBufferColor(index))
+        out += "combiner_buffer.rgb = last_tex_env_out.rgb;\n";
+
+    if (config.TevStageUpdatesCombinerBufferAlpha(index))
+        out += "combiner_buffer.a = last_tex_env_out.a;\n";
+}
+
+std::string GenerateFragmentShader(const PicaShaderConfig& config) {
+    std::string out = R"(
+#version 330
+#extension GL_ARB_explicit_uniform_location : require
+
+#define NUM_TEV_STAGES 6
+
+in vec4 primary_color;
+in vec2 texcoord[3];
+
+out vec4 color;
+)";
+
+    using Uniform = RasterizerOpenGL::PicaShader::Uniform;
+    out += "layout(location = " + std::to_string((int)Uniform::AlphaTestRef) + ") uniform int alphatest_ref;\n";
+    out += "layout(location = " + std::to_string((int)Uniform::TevConstColors) + ") uniform vec4 const_color[NUM_TEV_STAGES];\n";
+    out += "layout(location = " + std::to_string((int)Uniform::Texture0) + ") uniform sampler2D tex[3];\n";
+    out += "layout(location = " + std::to_string((int)Uniform::TevCombinerBufferColor) + ") uniform vec4 tev_combiner_buffer_color;\n";
+
+    out += "void main() {\n";
+    out += "vec4 combiner_buffer = tev_combiner_buffer_color;\n";
+    out += "vec4 last_tex_env_out = vec4(0.0);\n";
+
+    // Do not do any sort of processing if it's obvious we're not going to pass the alpha test
+    if (config.alpha_test_func == Regs::CompareFunc::Never) {
+        out += "discard; }";
+        return out;
+    }
+
+    for (size_t index = 0; index < config.tev_stages.size(); ++index)
+        WriteTevStage(out, config, (unsigned)index);
+
+    if (config.alpha_test_func != Regs::CompareFunc::Always) {
+        out += "if (";
+        AppendAlphaTestCondition(out, config.alpha_test_func);
+        out += ") discard;\n";
+    }
+
+    out += "color = last_tex_env_out;\n}";
+
+    return out;
+}
+
+std::string GenerateVertexShader() {
+    std::string out = "#version 330\n";
+    out += "layout(location = " + std::to_string((int)ATTRIBUTE_POSITION)  + ") in vec4 vert_position;\n";
+    out += "layout(location = " + std::to_string((int)ATTRIBUTE_COLOR)     + ") in vec4 vert_color;\n";
+    out += "layout(location = " + std::to_string((int)ATTRIBUTE_TEXCOORD0) + ") in vec2 vert_texcoord0;\n";
+    out += "layout(location = " + std::to_string((int)ATTRIBUTE_TEXCOORD1) + ") in vec2 vert_texcoord1;\n";
+    out += "layout(location = " + std::to_string((int)ATTRIBUTE_TEXCOORD2) + ") in vec2 vert_texcoord2;\n";
+
+    out += R"(
+out vec4 primary_color;
+out vec2 texcoord[3];
+
+void main() {
+    primary_color = vert_color;
+    texcoord[0] = vert_texcoord0;
+    texcoord[1] = vert_texcoord1;
+    texcoord[2] = vert_texcoord2;
+    gl_Position = vec4(vert_position.x, -vert_position.y, -vert_position.z, vert_position.w);
+}
+)";
+
+    return out;
+}
+
+} // namespace GLShader

src/video_core/renderer_opengl/gl_shader_gen.h

@@ -0,0 +1,27 @@
+// Copyright 2015 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <string>
+
+#include "video_core/renderer_opengl/gl_rasterizer.h"
+
+namespace GLShader {
+
+/**
+ * Generates the GLSL vertex shader program source code for the current Pica state
+ * @returns String of the shader source code
+ */
+std::string GenerateVertexShader();
+
+/**
+ * Generates the GLSL fragment shader program source code for the current Pica state
+ * @param config ShaderCacheKey object generated for the current Pica state, used for the shader
+ *               configuration (NOTE: Use state in this struct only, not the Pica registers!)
+ * @returns String of the shader source code
+ */
+std::string GenerateFragmentShader(const PicaShaderConfig& config);
+
+} // namespace GLShader

src/video_core/renderer_opengl/gl_shader_util.cpp

@@ -8,9 +8,9 @@
 #include "common/logging/log.h"
 #include "video_core/renderer_opengl/gl_shader_util.h"
 
-namespace ShaderUtil {
+namespace GLShader {
 
-GLuint LoadShaders(const char* vertex_shader, const char* fragment_shader) {
+GLuint LoadProgram(const char* vertex_shader, const char* fragment_shader) {
 
     // Create the shaders
     GLuint vertex_shader_id = glCreateShader(GL_VERTEX_SHADER);
@@ -65,6 +65,7 @@ GLuint LoadShaders(const char* vertex_shader, const char* fragment_shader) {
     GLuint program_id = glCreateProgram();
     glAttachShader(program_id, vertex_shader_id);
     glAttachShader(program_id, fragment_shader_id);
+
     glLinkProgram(program_id);
 
     // Check the program
@@ -87,4 +88,4 @@ GLuint LoadShaders(const char* vertex_shader, const char* fragment_shader) {
     return program_id;
 }
 
-}
+} // namespace GLShader

src/video_core/renderer_opengl/gl_shader_util.h

@@ -6,8 +6,22 @@
 
 #include <glad/glad.h>
 
-namespace ShaderUtil {
+namespace GLShader {
 
-GLuint LoadShaders(const char* vertex_file_path, const char* fragment_file_path);
+enum Attributes {
+    ATTRIBUTE_POSITION,
+    ATTRIBUTE_COLOR,
+    ATTRIBUTE_TEXCOORD0,
+    ATTRIBUTE_TEXCOORD1,
+    ATTRIBUTE_TEXCOORD2,
+};
 
-}
+/**
+ * Utility function to create and compile an OpenGL GLSL shader program (vertex + fragment shader)
+ * @param vertex_shader String of the GLSL vertex shader program
+ * @param fragment_shader String of the GLSL fragment shader program
+ * @returns Handle of the newly created OpenGL shader object
+ */
+GLuint LoadProgram(const char* vertex_shader, const char* fragment_shader);
+
+} // namespace

src/video_core/renderer_opengl/gl_shaders.h

@@ -1,337 +0,0 @@
-// Copyright 2014 Citra Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#pragma once
-
-namespace GLShaders {
-
-const char g_vertex_shader[] = R"(
-#version 150 core
-
-in vec2 vert_position;
-in vec2 vert_tex_coord;
-out vec2 frag_tex_coord;
-
-// This is a truncated 3x3 matrix for 2D transformations:
-// The upper-left 2x2 submatrix performs scaling/rotation/mirroring.
-// The third column performs translation.
-// The third row could be used for projection, which we don't need in 2D. It hence is assumed to
-// implicitly be [0, 0, 1]
-uniform mat3x2 modelview_matrix;
-
-void main() {
-    // Multiply input position by the rotscale part of the matrix and then manually translate by
-    // the last column. This is equivalent to using a full 3x3 matrix and expanding the vector
-    // to `vec3(vert_position.xy, 1.0)`
-    gl_Position = vec4(mat2(modelview_matrix) * vert_position + modelview_matrix[2], 0.0, 1.0);
-    frag_tex_coord = vert_tex_coord;
-}
-)";
-
-const char g_fragment_shader[] = R"(
-#version 150 core
-
-in vec2 frag_tex_coord;
-out vec4 color;
-
-uniform sampler2D color_texture;
-
-void main() {
-    color = texture(color_texture, frag_tex_coord);
-}
-)";
-
-const char g_vertex_shader_hw[] = R"(
-#version 150 core
-
-#define NUM_VTX_ATTR 7
-
-in vec4 vert_position;
-in vec4 vert_color;
-in vec2 vert_texcoords[3];
-
-out vec4 o[NUM_VTX_ATTR];
-
-void main() {
-    o[2] = vert_color;
-    o[3] = vec4(vert_texcoords[0].xy, vert_texcoords[1].xy);
-    o[5] = vec4(0.0, 0.0, vert_texcoords[2].xy);
-
-    gl_Position = vec4(vert_position.x, -vert_position.y, -vert_position.z, vert_position.w);
-}
-)";
-
-// TODO: Create a shader constructor and cache that builds this program with minimal conditionals instead of using tev_cfg uniforms
-const char g_fragment_shader_hw[] = R"(
-#version 150 core
-
-#define NUM_VTX_ATTR 7
-#define NUM_TEV_STAGES 6
-
-#define SOURCE_PRIMARYCOLOR           0x0
-#define SOURCE_PRIMARYFRAGMENTCOLOR   0x1
-#define SOURCE_SECONDARYFRAGMENTCOLOR 0x2
-#define SOURCE_TEXTURE0               0x3
-#define SOURCE_TEXTURE1               0x4
-#define SOURCE_TEXTURE2               0x5
-#define SOURCE_TEXTURE3               0x6
-#define SOURCE_PREVIOUSBUFFER         0xd
-#define SOURCE_CONSTANT               0xe
-#define SOURCE_PREVIOUS               0xf
-
-#define COLORMODIFIER_SOURCECOLOR         0x0
-#define COLORMODIFIER_ONEMINUSSOURCECOLOR 0x1
-#define COLORMODIFIER_SOURCEALPHA         0x2
-#define COLORMODIFIER_ONEMINUSSOURCEALPHA 0x3
-#define COLORMODIFIER_SOURCERED           0x4
-#define COLORMODIFIER_ONEMINUSSOURCERED   0x5
-#define COLORMODIFIER_SOURCEGREEN         0x8
-#define COLORMODIFIER_ONEMINUSSOURCEGREEN 0x9
-#define COLORMODIFIER_SOURCEBLUE          0xc
-#define COLORMODIFIER_ONEMINUSSOURCEBLUE  0xd
-
-#define ALPHAMODIFIER_SOURCEALPHA         0x0
-#define ALPHAMODIFIER_ONEMINUSSOURCEALPHA 0x1
-#define ALPHAMODIFIER_SOURCERED           0x2
-#define ALPHAMODIFIER_ONEMINUSSOURCERED   0x3
-#define ALPHAMODIFIER_SOURCEGREEN         0x4
-#define ALPHAMODIFIER_ONEMINUSSOURCEGREEN 0x5
-#define ALPHAMODIFIER_SOURCEBLUE          0x6
-#define ALPHAMODIFIER_ONEMINUSSOURCEBLUE  0x7
-
-#define OPERATION_REPLACE         0
-#define OPERATION_MODULATE        1
-#define OPERATION_ADD             2
-#define OPERATION_ADDSIGNED       3
-#define OPERATION_LERP            4
-#define OPERATION_SUBTRACT        5
-#define OPERATION_MULTIPLYTHENADD 8
-#define OPERATION_ADDTHENMULTIPLY 9
-
-#define COMPAREFUNC_NEVER              0
-#define COMPAREFUNC_ALWAYS             1
-#define COMPAREFUNC_EQUAL              2
-#define COMPAREFUNC_NOTEQUAL           3
-#define COMPAREFUNC_LESSTHAN           4
-#define COMPAREFUNC_LESSTHANOREQUAL    5
-#define COMPAREFUNC_GREATERTHAN        6
-#define COMPAREFUNC_GREATERTHANOREQUAL 7
-
-in vec4 o[NUM_VTX_ATTR];
-out vec4 color;
-
-uniform bool alphatest_enabled;
-uniform int alphatest_func;
-uniform float alphatest_ref;
-
-uniform sampler2D tex[3];
-
-uniform vec4 tev_combiner_buffer_color;
-
-struct TEVConfig
-{
-    bool enabled;
-    ivec3 color_sources;
-    ivec3 alpha_sources;
-    ivec3 color_modifiers;
-    ivec3 alpha_modifiers;
-    ivec2 color_alpha_op;
-    ivec2 color_alpha_multiplier;
-    vec4 const_color;
-    bvec2 updates_combiner_buffer_color_alpha;
-};
-
-uniform TEVConfig tev_cfgs[NUM_TEV_STAGES];
-
-vec4 g_combiner_buffer;
-vec4 g_last_tex_env_out;
-vec4 g_const_color;
-
-vec4 GetSource(int source) {
-    if (source == SOURCE_PRIMARYCOLOR) {
-        return o[2];
-    } else if (source == SOURCE_PRIMARYFRAGMENTCOLOR) {
-        // HACK: Until we implement fragment lighting, use primary_color
-        return o[2];
-    } else if (source == SOURCE_SECONDARYFRAGMENTCOLOR) {
-        // HACK: Until we implement fragment lighting, use zero
-        return vec4(0.0, 0.0, 0.0, 0.0);
-    } else if (source == SOURCE_TEXTURE0) {
-        return texture(tex[0], o[3].xy);
-    } else if (source == SOURCE_TEXTURE1) {
-        return texture(tex[1], o[3].zw);
-    } else if (source == SOURCE_TEXTURE2) {
-        // TODO: Unverified
-        return texture(tex[2], o[5].zw);
-    } else if (source == SOURCE_TEXTURE3) {
-        // TODO: no 4th texture?
-    } else if (source == SOURCE_PREVIOUSBUFFER) {
-        return g_combiner_buffer;
-    } else if (source == SOURCE_CONSTANT) {
-        return g_const_color;
-    } else if (source == SOURCE_PREVIOUS) {
-        return g_last_tex_env_out;
-    }
-
-    return vec4(0.0);
-}
-
-vec3 GetColorModifier(int factor, vec4 color) {
-    if (factor == COLORMODIFIER_SOURCECOLOR) {
-        return color.rgb;
-    } else if (factor == COLORMODIFIER_ONEMINUSSOURCECOLOR) {
-        return vec3(1.0) - color.rgb;
-    } else if (factor == COLORMODIFIER_SOURCEALPHA) {
-        return color.aaa;
-    } else if (factor == COLORMODIFIER_ONEMINUSSOURCEALPHA) {
-        return vec3(1.0) - color.aaa;
-    } else if (factor == COLORMODIFIER_SOURCERED) {
-        return color.rrr;
-    } else if (factor == COLORMODIFIER_ONEMINUSSOURCERED) {
-        return vec3(1.0) - color.rrr;
-    } else if (factor == COLORMODIFIER_SOURCEGREEN) {
-        return color.ggg;
-    } else if (factor == COLORMODIFIER_ONEMINUSSOURCEGREEN) {
-        return vec3(1.0) - color.ggg;
-    } else if (factor == COLORMODIFIER_SOURCEBLUE) {
-        return color.bbb;
-    } else if (factor == COLORMODIFIER_ONEMINUSSOURCEBLUE) {
-        return vec3(1.0) - color.bbb;
-    }
-
-    return vec3(0.0);
-}
-
-float GetAlphaModifier(int factor, vec4 color) {
-    if (factor == ALPHAMODIFIER_SOURCEALPHA) {
-        return color.a;
-    } else if (factor == ALPHAMODIFIER_ONEMINUSSOURCEALPHA) {
-        return 1.0 - color.a;
-    } else if (factor == ALPHAMODIFIER_SOURCERED) {
-        return color.r;
-    } else if (factor == ALPHAMODIFIER_ONEMINUSSOURCERED) {
-        return 1.0 - color.r;
-    } else if (factor == ALPHAMODIFIER_SOURCEGREEN) {
-        return color.g;
-    } else if (factor == ALPHAMODIFIER_ONEMINUSSOURCEGREEN) {
-        return 1.0 - color.g;
-    } else if (factor == ALPHAMODIFIER_SOURCEBLUE) {
-        return color.b;
-    } else if (factor == ALPHAMODIFIER_ONEMINUSSOURCEBLUE) {
-        return 1.0 - color.b;
-    }
-
-    return 0.0;
-}
-
-vec3 ColorCombine(int op, vec3 color[3]) {
-    if (op == OPERATION_REPLACE) {
-        return color[0];
-    } else if (op == OPERATION_MODULATE) {
-        return color[0] * color[1];
-    } else if (op == OPERATION_ADD) {
-        return min(color[0] + color[1], 1.0);
-    } else if (op == OPERATION_ADDSIGNED) {
-        return clamp(color[0] + color[1] - vec3(0.5), 0.0, 1.0);
-    } else if (op == OPERATION_LERP) {
-        return color[0] * color[2] + color[1] * (vec3(1.0) - color[2]);
-    } else if (op == OPERATION_SUBTRACT) {
-        return max(color[0] - color[1], 0.0);
-    } else if (op == OPERATION_MULTIPLYTHENADD) {
-        return min(color[0] * color[1] + color[2], 1.0);
-    } else if (op == OPERATION_ADDTHENMULTIPLY) {
-        return min(color[0] + color[1], 1.0) * color[2];
-    }
-
-    return vec3(0.0);
-}
-
-float AlphaCombine(int op, float alpha[3]) {
-    if (op == OPERATION_REPLACE) {
-        return alpha[0];
-    } else if (op == OPERATION_MODULATE) {
-        return alpha[0] * alpha[1];
-    } else if (op == OPERATION_ADD) {
-        return min(alpha[0] + alpha[1], 1.0);
-    } else if (op == OPERATION_ADDSIGNED) {
-        return clamp(alpha[0] + alpha[1] - 0.5, 0.0, 1.0);
-    } else if (op == OPERATION_LERP) {
-        return alpha[0] * alpha[2] + alpha[1] * (1.0 - alpha[2]);
-    } else if (op == OPERATION_SUBTRACT) {
-        return max(alpha[0] - alpha[1], 0.0);
-    } else if (op == OPERATION_MULTIPLYTHENADD) {
-        return min(alpha[0] * alpha[1] + alpha[2], 1.0);
-    } else if (op == OPERATION_ADDTHENMULTIPLY) {
-        return min(alpha[0] + alpha[1], 1.0) * alpha[2];
-    }
-
-    return 0.0;
-}
-
-void main(void) {
-    g_combiner_buffer = tev_combiner_buffer_color;
-
-    for (int tex_env_idx = 0; tex_env_idx < NUM_TEV_STAGES; ++tex_env_idx) {
-        if (tev_cfgs[tex_env_idx].enabled) {
-            g_const_color = tev_cfgs[tex_env_idx].const_color;
-
-            vec3 color_results[3] = vec3[3](GetColorModifier(tev_cfgs[tex_env_idx].color_modifiers.x, GetSource(tev_cfgs[tex_env_idx].color_sources.x)),
-                                            GetColorModifier(tev_cfgs[tex_env_idx].color_modifiers.y, GetSource(tev_cfgs[tex_env_idx].color_sources.y)),
-                                            GetColorModifier(tev_cfgs[tex_env_idx].color_modifiers.z, GetSource(tev_cfgs[tex_env_idx].color_sources.z)));
-            vec3 color_output = ColorCombine(tev_cfgs[tex_env_idx].color_alpha_op.x, color_results);
-
-            float alpha_results[3] = float[3](GetAlphaModifier(tev_cfgs[tex_env_idx].alpha_modifiers.x, GetSource(tev_cfgs[tex_env_idx].alpha_sources.x)),
-                                              GetAlphaModifier(tev_cfgs[tex_env_idx].alpha_modifiers.y, GetSource(tev_cfgs[tex_env_idx].alpha_sources.y)),
-                                              GetAlphaModifier(tev_cfgs[tex_env_idx].alpha_modifiers.z, GetSource(tev_cfgs[tex_env_idx].alpha_sources.z)));
-            float alpha_output = AlphaCombine(tev_cfgs[tex_env_idx].color_alpha_op.y, alpha_results);
-
-            g_last_tex_env_out = vec4(min(color_output * tev_cfgs[tex_env_idx].color_alpha_multiplier.x, 1.0), min(alpha_output * tev_cfgs[tex_env_idx].color_alpha_multiplier.y, 1.0));
-        }
-
-        if (tev_cfgs[tex_env_idx].updates_combiner_buffer_color_alpha.x) {
-            g_combiner_buffer.rgb = g_last_tex_env_out.rgb;
-        }
-
-        if (tev_cfgs[tex_env_idx].updates_combiner_buffer_color_alpha.y) {
-            g_combiner_buffer.a = g_last_tex_env_out.a;
-        }
-    }
-
-    if (alphatest_enabled) {
-        if (alphatest_func == COMPAREFUNC_NEVER) {
-            discard;
-        } else if (alphatest_func == COMPAREFUNC_ALWAYS) {
-
-        } else if (alphatest_func == COMPAREFUNC_EQUAL) {
-            if (g_last_tex_env_out.a != alphatest_ref) {
-                discard;
-            }
-        } else if (alphatest_func == COMPAREFUNC_NOTEQUAL) {
-            if (g_last_tex_env_out.a == alphatest_ref) {
-                discard;
-            }
-        } else if (alphatest_func == COMPAREFUNC_LESSTHAN) {
-            if (g_last_tex_env_out.a >= alphatest_ref) {
-                discard;
-            }
-        } else if (alphatest_func == COMPAREFUNC_LESSTHANOREQUAL) {
-            if (g_last_tex_env_out.a > alphatest_ref) {
-                discard;
-            }
-        } else if (alphatest_func == COMPAREFUNC_GREATERTHAN) {
-            if (g_last_tex_env_out.a <= alphatest_ref) {
-                discard;
-            }
-        } else if (alphatest_func == COMPAREFUNC_GREATERTHANOREQUAL) {
-            if (g_last_tex_env_out.a < alphatest_ref) {
-                discard;
-            }
-        }
-    }
-
-    color = g_last_tex_env_out;
-}
-)";
-
-}

src/video_core/renderer_opengl/gl_state.h

@@ -65,6 +65,7 @@ public:
         GLuint vertex_array; // GL_VERTEX_ARRAY_BINDING
         GLuint vertex_buffer; // GL_ARRAY_BUFFER_BINDING
         GLuint shader_program; // GL_CURRENT_PROGRAM
+        bool shader_dirty;
     } draw;
 
     OpenGLState();

src/video_core/renderer_opengl/renderer_opengl.cpp

@@ -21,9 +21,44 @@
 #include "video_core/debug_utils/debug_utils.h"
 #include "video_core/renderer_opengl/gl_rasterizer.h"
 #include "video_core/renderer_opengl/gl_shader_util.h"
-#include "video_core/renderer_opengl/gl_shaders.h"
 #include "video_core/renderer_opengl/renderer_opengl.h"
 
+static const char vertex_shader[] = R"(
+#version 150 core
+
+in vec2 vert_position;
+in vec2 vert_tex_coord;
+out vec2 frag_tex_coord;
+
+// This is a truncated 3x3 matrix for 2D transformations:
+// The upper-left 2x2 submatrix performs scaling/rotation/mirroring.
+// The third column performs translation.
+// The third row could be used for projection, which we don't need in 2D. It hence is assumed to
+// implicitly be [0, 0, 1]
+uniform mat3x2 modelview_matrix;
+
+void main() {
+    // Multiply input position by the rotscale part of the matrix and then manually translate by
+    // the last column. This is equivalent to using a full 3x3 matrix and expanding the vector
+    // to `vec3(vert_position.xy, 1.0)`
+    gl_Position = vec4(mat2(modelview_matrix) * vert_position + modelview_matrix[2], 0.0, 1.0);
+    frag_tex_coord = vert_tex_coord;
+}
+)";
+
+static const char fragment_shader[] = R"(
+#version 150 core
+
+in vec2 frag_tex_coord;
+out vec4 color;
+
+uniform sampler2D color_texture;
+
+void main() {
+    color = texture(color_texture, frag_tex_coord);
+}
+)";
+
 /**
  * Vertex structure that the drawn screen rectangles are composed of.
  */
@@ -207,7 +242,7 @@ void RendererOpenGL::InitOpenGLObjects() {
     glClearColor(Settings::values.bg_red, Settings::values.bg_green, Settings::values.bg_blue, 0.0f);
 
     // Link shaders and get variable locations
-    program_id = ShaderUtil::LoadShaders(GLShaders::g_vertex_shader, GLShaders::g_fragment_shader);
+    program_id = GLShader::LoadProgram(vertex_shader, fragment_shader);
     uniform_modelview_matrix = glGetUniformLocation(program_id, "modelview_matrix");
     uniform_color_texture = glGetUniformLocation(program_id, "color_texture");
     attrib_position = glGetAttribLocation(program_id, "vert_position");