Skip pass-thru tev stages, upper_bound calc once

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -43,6 +43,7 @@ void RasterizerOpenGL::InitObjects() {
         auto& uniform_tev = uniform_tev_cfgs[i];
 
         std::string tev_ref_str = "tev_cfgs[" + std::to_string(i) + "]";
+        uniform_tev.enabled = glGetUniformLocation(shader.GetHandle(), (tev_ref_str + ".enabled").c_str());
         uniform_tev.color_sources = glGetUniformLocation(shader.GetHandle(), (tev_ref_str + ".color_sources").c_str());
         uniform_tev.alpha_sources = glGetUniformLocation(shader.GetHandle(), (tev_ref_str + ".alpha_sources").c_str());
         uniform_tev.color_modifiers = glGetUniformLocation(shader.GetHandle(), (tev_ref_str + ".color_modifiers").c_str());
@@ -479,6 +480,17 @@ void RasterizerOpenGL::SyncDrawState() {
         const auto& stage = tev_stages[tev_stage_idx];
         const auto& uniform_tev_cfg = uniform_tev_cfgs[tev_stage_idx];
 
+        // No need to process the tev stage if it simply passes the previous stage results through
+        if (stage.color_op == Pica::Regs::TevStageConfig::Operation::Replace &&
+            stage.alpha_op == Pica::Regs::TevStageConfig::Operation::Replace &&
+            stage.color_source1 == Pica::Regs::TevStageConfig::Source::Previous &&
+            stage.alpha_source1 == Pica::Regs::TevStageConfig::Source::Previous &&
+            stage.color_modifier1 == Pica::Regs::TevStageConfig::ColorModifier::SourceColor &&
+            stage.alpha_modifier1 == Pica::Regs::TevStageConfig::AlphaModifier::SourceAlpha &&
+            stage.GetColorMultiplier() == 1 &&
+            stage.GetAlphaMultiplier() == 1) {
+            glUniform1i(uniform_tev_cfg.enabled, 0);
+        } else {
             GLint color_srcs[3] = { (GLint)stage.color_source1.Value(), (GLint)stage.color_source2.Value(), (GLint)stage.color_source3.Value() };
             GLint alpha_srcs[3] = { (GLint)stage.alpha_source1.Value(), (GLint)stage.alpha_source2.Value(), (GLint)stage.alpha_source3.Value() };
             GLint color_mods[3] = { (GLint)stage.color_modifier1.Value(), (GLint)stage.color_modifier2.Value(), (GLint)stage.color_modifier3.Value() };
@@ -488,6 +500,7 @@ void RasterizerOpenGL::SyncDrawState() {
                                        stage.const_b / 255.0f,
                                        stage.const_a / 255.0f };
 
+            glUniform1i(uniform_tev_cfg.enabled, 1);
             glUniform3iv(uniform_tev_cfg.color_sources, 1, color_srcs);
             glUniform3iv(uniform_tev_cfg.alpha_sources, 1, alpha_srcs);
             glUniform3iv(uniform_tev_cfg.color_modifiers, 1, color_mods);
@@ -499,6 +512,7 @@ void RasterizerOpenGL::SyncDrawState() {
                         Pica::registers.tev_combiner_buffer_input.TevStageUpdatesCombinerBufferColor(tev_stage_idx),
                         Pica::registers.tev_combiner_buffer_input.TevStageUpdatesCombinerBufferAlpha(tev_stage_idx));
         }
+    }
 
     // Sync alpha testing to hw shader
     if (Pica::registers.output_merger.alpha_test.enable) {

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -35,6 +35,7 @@ public:
 private:
     /// Structure used for managing texture environment states
     struct TEVConfigUniforms {
+        GLuint enabled;
         GLuint color_sources;
         GLuint alpha_sources;
         GLuint color_modifiers;

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -60,7 +60,9 @@ void RasterizerCacheOpenGL::LoadAndBindTexture(OpenGLState &state, int texture_u
 
 void RasterizerCacheOpenGL::NotifyFlush(PAddr addr, u32 size) {
     // Flush any texture that falls in the flushed region
-    for (auto it = texture_cache.begin(); it != texture_cache.upper_bound(addr + size);) {
+    // TODO: Optimize by also inserting upper bound (addr + size) of each texture into the same map and also narrow using lower_bound
+    auto cache_upper_bound = texture_cache.upper_bound(addr + size);
+    for (auto it = texture_cache.begin(); it != cache_upper_bound;) {
         PAddr max_low_addr_bound = std::max(addr, it->first);
         PAddr min_hi_addr_bound = std::min(addr + size, it->first + it->second->size);
 

src/video_core/renderer_opengl/gl_shaders.h

@@ -144,6 +144,7 @@ uniform vec4 tev_combiner_buffer_color;
 
 struct TEVConfig
 {
+    bool enabled;
     ivec3 color_sources;
     ivec3 alpha_sources;
     ivec3 color_modifiers;
@@ -286,7 +287,11 @@ float AlphaCombine(int op, float alpha[3]) {
     return 0.0;
 }
 
-void ProcessTexEnv(int tex_env_idx) {
+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)),
@@ -300,6 +305,7 @@ void ProcessTexEnv(int tex_env_idx) {
             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;
@@ -310,13 +316,6 @@ void ProcessTexEnv(int tex_env_idx) {
         }
     }
 
-void main(void) {
-    g_combiner_buffer = tev_combiner_buffer_color;
-
-    for (int i = 0; i < NUM_TEV_STAGES; ++i) {
-        ProcessTexEnv(i);
-    }
-
     if (alphatest_func == COMPAREFUNC_NEVER) {
         discard;
     } else if (alphatest_func == COMPAREFUNC_ALWAYS) {