video_core: Use source3 when GPU_PREVIOUS is used in first stage (#7411)

src/video_core/rasterizer_accelerated.cpp

@@ -846,7 +846,7 @@ void RasterizerAccelerated::SyncTextureBorderColor(int tex_index) {
 }
 
 void RasterizerAccelerated::SyncClipPlane() {
-    const bool enable_clip1 = regs.rasterizer.clip_enable != 0;
+    const u32 enable_clip1 = regs.rasterizer.clip_enable != 0;
     const auto raw_clip_coef = regs.rasterizer.GetClipCoef();
     const Common::Vec4f new_clip_coef = {raw_clip_coef.x.ToFloat32(), raw_clip_coef.y.ToFloat32(),
                                          raw_clip_coef.z.ToFloat32(), raw_clip_coef.w.ToFloat32()};

src/video_core/renderer_software/sw_rasterizer.cpp

@@ -695,7 +695,7 @@ Common::Vec4<u8> RasterizerSoftware::WriteTevConfig(
      * with some basic arithmetic. Alpha combiners can be configured separately but work
      * analogously.
      **/
-    Common::Vec4<u8> combiner_output = primary_color;
+    Common::Vec4<u8> combiner_output = {0, 0, 0, 0};
     Common::Vec4<u8> combiner_buffer = {0, 0, 0, 0};
     Common::Vec4<u8> next_combiner_buffer =
         Common::MakeVec(regs.texturing.tev_combiner_buffer_color.r.Value(),
@@ -746,9 +746,15 @@ Common::Vec4<u8> RasterizerSoftware::WriteTevConfig(
          *       combiner_output.rgb(), but instead store it in a temporary variable until
          *       alpha combining has been done.
          **/
+        const auto source1 = tev_stage_index == 0 && tev_stage.color_source1 == Source::Previous
+                                 ? tev_stage.color_source3.Value()
+                                 : tev_stage.color_source1.Value();
+        const auto source2 = tev_stage_index == 0 && tev_stage.color_source2 == Source::Previous
+                                 ? tev_stage.color_source3.Value()
+                                 : tev_stage.color_source2.Value();
         const std::array<Common::Vec3<u8>, 3> color_result = {
-            GetColorModifier(tev_stage.color_modifier1, get_source(tev_stage.color_source1)),
+            GetColorModifier(tev_stage.color_modifier1, get_source(source1)),
-            GetColorModifier(tev_stage.color_modifier2, get_source(tev_stage.color_source2)),
+            GetColorModifier(tev_stage.color_modifier2, get_source(source2)),
             GetColorModifier(tev_stage.color_modifier3, get_source(tev_stage.color_source3)),
         };
         const Common::Vec3<u8> color_output = ColorCombine(tev_stage.color_op, color_result);

src/video_core/shader/generator/glsl_fs_shader_gen.cpp

@@ -143,7 +143,7 @@ vec4 secondary_fragment_color = vec4(0.0);
 
     out += "vec4 combiner_buffer = vec4(0.0);\n"
            "vec4 next_combiner_buffer = tev_combiner_buffer_color;\n"
-           "vec4 combiner_output = rounded_primary_color;\n";
+           "vec4 combiner_output = vec4(0.0);\n";
 
     out += "vec3 color_results_1 = vec3(0.0);\n"
            "vec3 color_results_2 = vec3(0.0);\n"
@@ -225,96 +225,75 @@ void FragmentModule::WriteScissor() {
            "gl_FragCoord.y < float(scissor_y2))) discard;\n";
 }
 
-void FragmentModule::AppendSource(Pica::TexturingRegs::TevStageConfig::Source source,
+std::string FragmentModule::GetSource(Pica::TexturingRegs::TevStageConfig::Source source,
-                                  u32 tev_index) {
+                                      u32 tev_index) {
     using Source = Pica::TexturingRegs::TevStageConfig::Source;
     switch (source) {
     case Source::PrimaryColor:
-        out += "rounded_primary_color";
+        return "rounded_primary_color";
-        break;
     case Source::PrimaryFragmentColor:
-        out += "primary_fragment_color";
+        return "primary_fragment_color";
-        break;
     case Source::SecondaryFragmentColor:
-        out += "secondary_fragment_color";
+        return "secondary_fragment_color";
-        break;
     case Source::Texture0:
-        out += "sampleTexUnit0()";
+        return "sampleTexUnit0()";
-        break;
     case Source::Texture1:
-        out += "sampleTexUnit1()";
+        return "sampleTexUnit1()";
-        break;
     case Source::Texture2:
-        out += "sampleTexUnit2()";
+        return "sampleTexUnit2()";
-        break;
     case Source::Texture3:
-        out += "sampleTexUnit3()";
+        return "sampleTexUnit3()";
-        break;
     case Source::PreviousBuffer:
-        out += "combiner_buffer";
+        return "combiner_buffer";
-        break;
     case Source::Constant:
-        out += fmt::format("const_color[{}]", tev_index);
+        return fmt::format("const_color[{}]", tev_index);
-        break;
     case Source::Previous:
-        out += "combiner_output";
+        return "combiner_output";
-        break;
     default:
-        out += "vec4(0.0)";
         LOG_CRITICAL(Render, "Unknown source op {}", source);
-        break;
+        return "vec4(0.0)";
     }
 }
 
 void FragmentModule::AppendColorModifier(
     Pica::TexturingRegs::TevStageConfig::ColorModifier modifier,
     Pica::TexturingRegs::TevStageConfig::Source source, u32 tev_index) {
+    using Source = Pica::TexturingRegs::TevStageConfig::Source;
     using ColorModifier = Pica::TexturingRegs::TevStageConfig::ColorModifier;
+    const TexturingRegs::TevStageConfig stage = config.texture.tev_stages[tev_index];
+    const bool force_source3 = tev_index == 0 && source == Source::Previous;
+    const auto color_source =
+        GetSource(force_source3 ? stage.color_source3.Value() : source, tev_index);
     switch (modifier) {
     case ColorModifier::SourceColor:
-        AppendSource(source, tev_index);
+        out += fmt::format("{}.rgb", color_source);
-        out += ".rgb";
         break;
     case ColorModifier::OneMinusSourceColor:
-        out += "vec3(1.0) - ";
+        out += fmt::format("vec3(1.0) - {}.rgb", color_source);
-        AppendSource(source, tev_index);
-        out += ".rgb";
         break;
     case ColorModifier::SourceAlpha:
-        AppendSource(source, tev_index);
+        out += fmt::format("{}.aaa", color_source);
-        out += ".aaa";
         break;
     case ColorModifier::OneMinusSourceAlpha:
-        out += "vec3(1.0) - ";
+        out += fmt::format("vec3(1.0) - {}.aaa", color_source);
-        AppendSource(source, tev_index);
-        out += ".aaa";
         break;
     case ColorModifier::SourceRed:
-        AppendSource(source, tev_index);
+        out += fmt::format("{}.rrr", color_source);
-        out += ".rrr";
         break;
     case ColorModifier::OneMinusSourceRed:
-        out += "vec3(1.0) - ";
+        out += fmt::format("vec3(1.0) - {}.rrr", color_source);
-        AppendSource(source, tev_index);
-        out += ".rrr";
         break;
     case ColorModifier::SourceGreen:
-        AppendSource(source, tev_index);
+        out += fmt::format("{}.ggg", color_source);
-        out += ".ggg";
         break;
     case ColorModifier::OneMinusSourceGreen:
-        out += "vec3(1.0) - ";
+        out += fmt::format("vec3(1.0) - {}.ggg", color_source);
-        AppendSource(source, tev_index);
-        out += ".ggg";
         break;
     case ColorModifier::SourceBlue:
-        AppendSource(source, tev_index);
+        out += fmt::format("{}.bbb", color_source);
-        out += ".bbb";
         break;
     case ColorModifier::OneMinusSourceBlue:
-        out += "vec3(1.0) - ";
+        out += fmt::format("vec3(1.0) - {}.bbb", color_source);
-        AppendSource(source, tev_index);
-        out += ".bbb";
         break;
     default:
         out += "vec3(0.0)";
@@ -326,43 +305,36 @@ void FragmentModule::AppendColorModifier(
 void FragmentModule::AppendAlphaModifier(
     Pica::TexturingRegs::TevStageConfig::AlphaModifier modifier,
     Pica::TexturingRegs::TevStageConfig::Source source, u32 tev_index) {
+    using Source = Pica::TexturingRegs::TevStageConfig::Source;
     using AlphaModifier = Pica::TexturingRegs::TevStageConfig::AlphaModifier;
+    const TexturingRegs::TevStageConfig stage = config.texture.tev_stages[tev_index];
+    const bool force_source3 = tev_index == 0 && source == Source::Previous;
+    const auto alpha_source =
+        GetSource(force_source3 ? stage.alpha_source3.Value() : source, tev_index);
     switch (modifier) {
     case AlphaModifier::SourceAlpha:
-        AppendSource(source, tev_index);
+        out += fmt::format("{}.a", alpha_source);
-        out += ".a";
         break;
     case AlphaModifier::OneMinusSourceAlpha:
-        out += "1.0 - ";
+        out += fmt::format("1.0 - {}.a", alpha_source);
-        AppendSource(source, tev_index);
-        out += ".a";
         break;
     case AlphaModifier::SourceRed:
-        AppendSource(source, tev_index);
+        out += fmt::format("{}.r", alpha_source);
-        out += ".r";
         break;
     case AlphaModifier::OneMinusSourceRed:
-        out += "1.0 - ";
+        out += fmt::format("1.0 - {}.r", alpha_source);
-        AppendSource(source, tev_index);
-        out += ".r";
         break;
     case AlphaModifier::SourceGreen:
-        AppendSource(source, tev_index);
+        out += fmt::format("{}.g", alpha_source);
-        out += ".g";
         break;
     case AlphaModifier::OneMinusSourceGreen:
-        out += "1.0 - ";
+        out += fmt::format("1.0 - {}.g", alpha_source);
-        AppendSource(source, tev_index);
-        out += ".g";
         break;
     case AlphaModifier::SourceBlue:
-        AppendSource(source, tev_index);
+        out += fmt::format("{}.b", alpha_source);
-        out += ".b";
         break;
     case AlphaModifier::OneMinusSourceBlue:
-        out += "1.0 - ";
+        out += fmt::format("1.0 - {}.b", alpha_source);
-        AppendSource(source, tev_index);
-        out += ".b";
         break;
     default:
         out += "0.0";
@@ -384,12 +356,11 @@ void FragmentModule::AppendColorCombiner(Pica::TexturingRegs::TevStageConfig::Op
         case Operation::AddSigned:
             return "color_results_1 + color_results_2 - vec3(0.5)";
         case Operation::Lerp:
-            return "color_results_1 * color_results_3 + color_results_2 * (vec3(1.0) - "
+            return "mix(color_results_2, color_results_1, color_results_3)";
-                   "color_results_3)";
         case Operation::Subtract:
             return "color_results_1 - color_results_2";
         case Operation::MultiplyThenAdd:
-            return "color_results_1 * color_results_2 + color_results_3";
+            return "fma(color_results_1, color_results_2, color_results_3)";
         case Operation::AddThenMultiply:
             return "min(color_results_1 + color_results_2, vec3(1.0)) * color_results_3";
         case Operation::Dot3_RGB:
@@ -416,11 +387,11 @@ void FragmentModule::AppendAlphaCombiner(Pica::TexturingRegs::TevStageConfig::Op
         case Operation::AddSigned:
             return "alpha_results_1 + alpha_results_2 - 0.5";
         case Operation::Lerp:
-            return "alpha_results_1 * alpha_results_3 + alpha_results_2 * (1.0 - alpha_results_3)";
+            return "mix(alpha_results_2, alpha_results_1, alpha_results_3)";
         case Operation::Subtract:
             return "alpha_results_1 - alpha_results_2";
         case Operation::MultiplyThenAdd:
-            return "alpha_results_1 * alpha_results_2 + alpha_results_3";
+            return "fma(alpha_results_1, alpha_results_2, alpha_results_3)";
         case Operation::AddThenMultiply:
             return "min(alpha_results_1 + alpha_results_2, 1.0) * alpha_results_3";
         default:

src/video_core/shader/generator/glsl_fs_shader_gen.h

@@ -41,8 +41,8 @@ private:
     /// Writes the code to emulate PICA min/max blending factors
     void WriteBlending();
 
-    /// Writes the specified TEV stage source component(s)
+    /// Returns the specified TEV stage source component(s)
-    void AppendSource(Pica::TexturingRegs::TevStageConfig::Source source, u32 tev_index);
+    std::string GetSource(Pica::TexturingRegs::TevStageConfig::Source source, u32 tev_index);
 
     /// Writes the color components to use for the specified TEV stage color modifier
     void AppendColorModifier(Pica::TexturingRegs::TevStageConfig::ColorModifier modifier,

src/video_core/shader/generator/shader_uniforms.h

@@ -86,7 +86,7 @@ struct PicaUniformsData {
 };
 
 struct VSUniformData {
-    bool enable_clip1;
+    u32 enable_clip1;
     alignas(16) Common::Vec4f clip_coef;
 };
 static_assert(sizeof(VSUniformData) == 32,

src/video_core/shader/generator/spv_fs_shader_gen.cpp

@@ -55,7 +55,7 @@ void FragmentModule::Generate() {
 
     combiner_buffer = ConstF32(0.f, 0.f, 0.f, 0.f);
     next_combiner_buffer = GetShaderDataMember(vec_ids.Get(4), ConstS32(26));
-    last_tex_env_out = rounded_primary_color;
+    combiner_output = ConstF32(0.f, 0.f, 0.f, 0.f);
 
     // Write shader bytecode to emulate PICA TEV stages
     for (u32 index = 0; index < config.texture.tev_stages.size(); ++index) {
@@ -76,7 +76,7 @@ void FragmentModule::Generate() {
         break;
     }
 
-    Id color{Byteround(last_tex_env_out, 4)};
+    Id color{Byteround(combiner_output, 4)};
     switch (config.framebuffer.logic_op) {
     case FramebufferRegs::LogicOp::Clear:
         color = ConstF32(0.f, 0.f, 0.f, 0.f);
@@ -184,12 +184,12 @@ void FragmentModule::WriteFog() {
 
     // Blend the fog
     const Id tex_env_rgb{
-        OpVectorShuffle(vec_ids.Get(3), last_tex_env_out, last_tex_env_out, 0, 1, 2)};
+        OpVectorShuffle(vec_ids.Get(3), combiner_output, combiner_output, 0, 1, 2)};
     const Id fog_color{GetShaderDataMember(vec_ids.Get(3), ConstS32(19))};
     const Id fog_factor_rgb{
         OpCompositeConstruct(vec_ids.Get(3), fog_factor, fog_factor, fog_factor)};
     const Id fog_result{OpFMix(vec_ids.Get(3), fog_color, tex_env_rgb, fog_factor_rgb)};
-    last_tex_env_out = OpVectorShuffle(vec_ids.Get(4), fog_result, last_tex_env_out, 0, 1, 2, 6);
+    combiner_output = OpVectorShuffle(vec_ids.Get(4), fog_result, combiner_output, 0, 1, 2, 6);
 }
 
 void FragmentModule::WriteGas() {
@@ -630,8 +630,7 @@ void FragmentModule::WriteLighting() {
 }
 
 void FragmentModule::WriteTevStage(s32 index) {
-    const TexturingRegs::TevStageConfig stage =
+    const TexturingRegs::TevStageConfig stage = config.texture.tev_stages[index];
-        static_cast<const TexturingRegs::TevStageConfig>(config.texture.tev_stages[index]);
 
     // Detects if a TEV stage is configured to be skipped (to avoid generating unnecessary code)
     const auto is_passthrough_tev_stage = [](const TevStageConfig& stage) {
@@ -674,18 +673,18 @@ void FragmentModule::WriteTevStage(s32 index) {
         alpha_output =
             OpFMul(f32_id, alpha_output, ConstF32(static_cast<float>(stage.GetAlphaMultiplier())));
         alpha_output = OpFClamp(f32_id, alpha_output, ConstF32(0.f), ConstF32(1.f));
-        last_tex_env_out = OpCompositeConstruct(vec_ids.Get(4), color_output, alpha_output);
+        combiner_output = OpCompositeConstruct(vec_ids.Get(4), color_output, alpha_output);
     }
 
     combiner_buffer = next_combiner_buffer;
     if (config.TevStageUpdatesCombinerBufferColor(index)) {
         next_combiner_buffer =
-            OpVectorShuffle(vec_ids.Get(4), last_tex_env_out, next_combiner_buffer, 0, 1, 2, 7);
+            OpVectorShuffle(vec_ids.Get(4), combiner_output, next_combiner_buffer, 0, 1, 2, 7);
     }
 
     if (config.TevStageUpdatesCombinerBufferAlpha(index)) {
         next_combiner_buffer =
-            OpVectorShuffle(vec_ids.Get(4), next_combiner_buffer, last_tex_env_out, 0, 1, 2, 7);
+            OpVectorShuffle(vec_ids.Get(4), next_combiner_buffer, combiner_output, 0, 1, 2, 7);
     }
 }
 
@@ -728,7 +727,7 @@ void FragmentModule::WriteAlphaTestCondition(FramebufferRegs::CompareFunc func)
     case CompareFunc::GreaterThan:
     case CompareFunc::GreaterThanOrEqual: {
         const Id alpha_scaled{
-            OpFMul(f32_id, OpCompositeExtract(f32_id, last_tex_env_out, 3), ConstF32(255.f))};
+            OpFMul(f32_id, OpCompositeExtract(f32_id, combiner_output, 3), ConstF32(255.f))};
         const Id alpha_int{OpConvertFToS(i32_id, alpha_scaled)};
         const Id alphatest_ref{GetShaderDataMember(i32_id, ConstS32(1))};
         const Id alpha_comp_ref{compare(alpha_int, alphatest_ref)};
@@ -1280,7 +1279,7 @@ Id FragmentModule::LookupLightingLUT(Id lut_index, Id index, Id delta) {
     return OpFma(f32_id, entry_g, delta, entry_r);
 }
 
-Id FragmentModule::AppendSource(TevStageConfig::Source source, s32 index) {
+Id FragmentModule::GetSource(TevStageConfig::Source source, s32 index) {
     using Source = TevStageConfig::Source;
     switch (source) {
     case Source::PrimaryColor:
@@ -1302,7 +1301,7 @@ Id FragmentModule::AppendSource(TevStageConfig::Source source, s32 index) {
     case Source::Constant:
         return GetShaderDataMember(vec_ids.Get(4), ConstS32(25), ConstS32(index));
     case Source::Previous:
-        return last_tex_env_out;
+        return combiner_output;
     default:
         LOG_CRITICAL(Render, "Unknown source op {}", source);
         return ConstF32(0.f, 0.f, 0.f, 0.f);
@@ -1311,8 +1310,11 @@ Id FragmentModule::AppendSource(TevStageConfig::Source source, s32 index) {
 
 Id FragmentModule::AppendColorModifier(TevStageConfig::ColorModifier modifier,
                                        TevStageConfig::Source source, s32 index) {
+    using Source = TevStageConfig::Source;
     using ColorModifier = TevStageConfig::ColorModifier;
-    const Id source_color{AppendSource(source, index)};
+    const TexturingRegs::TevStageConfig stage = config.texture.tev_stages[index];
+    const bool force_source3 = index == 0 && source == Source::Previous;
+    const Id source_color{GetSource(force_source3 ? stage.color_source3.Value() : source, index)};
     const Id one_vec{ConstF32(1.f, 1.f, 1.f)};
 
     const auto shuffle = [&](s32 r, s32 g, s32 b) -> Id {
@@ -1348,11 +1350,14 @@ Id FragmentModule::AppendColorModifier(TevStageConfig::ColorModifier modifier,
 
 Id FragmentModule::AppendAlphaModifier(TevStageConfig::AlphaModifier modifier,
                                        TevStageConfig::Source source, s32 index) {
+    using Source = TevStageConfig::Source;
     using AlphaModifier = TevStageConfig::AlphaModifier;
-    const Id source_color{AppendSource(source, index)};
+    const TexturingRegs::TevStageConfig stage = config.texture.tev_stages[index];
+    const bool force_source3 = index == 0 && source == Source::Previous;
+    const Id source_alpha{GetSource(force_source3 ? stage.alpha_source3.Value() : source, index)};
     const Id one_f32{ConstF32(1.f)};
 
-    const auto component = [&](s32 c) -> Id { return OpCompositeExtract(f32_id, source_color, c); };
+    const auto component = [&](s32 c) -> Id { return OpCompositeExtract(f32_id, source_alpha, c); };
 
     switch (modifier) {
     case AlphaModifier::SourceAlpha:

src/video_core/shader/generator/spv_fs_shader_gen.h

@@ -99,13 +99,13 @@ private:
     /// Lookups the lighting LUT at the provided lut_index
     [[nodiscard]] Id LookupLightingLUT(Id lut_index, Id index, Id delta);
 
-    /// Writes the specified TEV stage source component(s)
+    /// Returns the specified TEV stage source component(s)
-    [[nodiscard]] Id AppendSource(Pica::TexturingRegs::TevStageConfig::Source source, s32 index);
+    [[nodiscard]] Id GetSource(Pica::TexturingRegs::TevStageConfig::Source source, s32 index);
 
     /// Writes the color components to use for the specified TEV stage color modifier
     [[nodiscard]] Id AppendColorModifier(
         Pica::TexturingRegs::TevStageConfig::ColorModifier modifier,
-        Pica::TexturingRegs::TevStageConfig::Source source, s32 index);
+        Pica::TexturingRegs::TevStageConfig::Source source, s32 tev_index);
 
     /// Writes the alpha component to use for the specified TEV stage alpha modifier
     [[nodiscard]] Id AppendAlphaModifier(
@@ -272,7 +272,7 @@ private:
     Id secondary_fragment_color{};
     Id combiner_buffer{};
     Id next_combiner_buffer{};
-    Id last_tex_env_out{};
+    Id combiner_output{};
 
     Id color_results_1{};
     Id color_results_2{};