tev: Use primary color for previous source in first stage (#6921)

src/video_core/renderer_opengl/gl_shader_gen.cpp

@@ -1485,7 +1485,7 @@ 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 last_tex_env_out = vec4(0.0);\n";
+           "vec4 last_tex_env_out = rounded_primary_color;\n";
 
     for (std::size_t index = 0; index < state.tev_stages.size(); ++index) {
         WriteTevStage(out, config, static_cast<u32>(index));

src/video_core/renderer_software/sw_rasterizer.cpp

@@ -422,8 +422,8 @@ void RasterizerSoftware::ProcessTriangle(const Vertex& v0, const Vertex& v1, con
             }
 
             // Write the TEV stages.
-            WriteTevConfig(texture_color, tev_stages, primary_color, primary_fragment_color,
+            auto combiner_output = WriteTevConfig(texture_color, tev_stages, primary_color,
-                           secondary_fragment_color);
+                                                  primary_fragment_color, secondary_fragment_color);
 
             const auto& output_merger = regs.framebuffer.output_merger;
             if (output_merger.fragment_operation_mode ==
@@ -663,7 +663,7 @@ Common::Vec4<u8> RasterizerSoftware::PixelColor(u16 x, u16 y,
     return result;
 }
 
-void RasterizerSoftware::WriteTevConfig(
+Common::Vec4<u8> RasterizerSoftware::WriteTevConfig(
     std::span<const Common::Vec4<u8>, 4> texture_color,
     std::span<const Pica::TexturingRegs::TevStageConfig, 6> tev_stages,
     Common::Vec4<u8> primary_color, Common::Vec4<u8> primary_fragment_color,
@@ -676,6 +676,7 @@ void 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_buffer = {0, 0, 0, 0};
     Common::Vec4<u8> next_combiner_buffer =
         Common::MakeVec(regs.texturing.tev_combiner_buffer_color.r.Value(),
@@ -766,6 +767,8 @@ void RasterizerSoftware::WriteTevConfig(
             next_combiner_buffer.a() = combiner_output.a();
         }
     }
+
+    return combiner_output;
 }
 
 void RasterizerSoftware::WriteFog(Common::Vec4<u8>& combiner_output, float depth) const {

src/video_core/renderer_software/sw_rasterizer.h

@@ -55,7 +55,8 @@ private:
     Common::Vec4<u8> PixelColor(u16 x, u16 y, Common::Vec4<u8>& combiner_output) const;
 
     /// Emulates the TEV configuration and returns the combiner output.
-    void WriteTevConfig(std::span<const Common::Vec4<u8>, 4> texture_color,
+    Common::Vec4<u8> WriteTevConfig(
+        std::span<const Common::Vec4<u8>, 4> texture_color,
         std::span<const Pica::TexturingRegs::TevStageConfig, 6> tev_stages,
         Common::Vec4<u8> primary_color, Common::Vec4<u8> primary_fragment_color,
         Common::Vec4<u8> secondary_fragment_color);
@@ -74,9 +75,6 @@ private:
     Pica::State& state;
     const Pica::Regs& regs;
     Framebuffer fb;
-    // Kirby Blowout Blast relies on the combiner output of a previous draw
-    // in order to render the sky correctly.
-    Common::Vec4<u8> combiner_output{};
 };
 
 } // namespace SwRenderer

src/video_core/renderer_vulkan/vk_shader_gen.cpp

@@ -1527,7 +1527,7 @@ 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 last_tex_env_out = vec4(0.0);\n";
+           "vec4 last_tex_env_out = rounded_primary_color;\n";
 
     out += "vec3 color_results_1 = vec3(0.0);\n"
            "vec3 color_results_2 = vec3(0.0);\n"