GPU/Scissor: Implemented scissor testing in the hw renderer.

We're doing this in the shader since it's not possible to provide the Exclude functionality using glScissor.

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -150,6 +150,8 @@ void RasterizerOpenGL::DrawTriangles() {
     SyncFramebuffer();
     SyncDrawState();
 
+    SyncScissorTest();
+
     if (state.draw.shader_dirty) {
         SetShader();
         state.draw.shader_dirty = false;
@@ -222,6 +224,11 @@ void RasterizerOpenGL::NotifyPicaRegisterChanged(u32 id) {
         SyncDepthTest();
         break;
 
+    // Scissor test
+    case PICA_REG_INDEX(scissor_test.mode):
+        state.draw.shader_dirty = true;
+        break;
+
     // Logic op
     case PICA_REG_INDEX(output_merger.logic_op):
         SyncLogicOp();
@@ -499,6 +506,7 @@ void RasterizerOpenGL::SetShader() {
     // Update uniforms
     SyncAlphaTest();
     SyncCombinerColor();
+    SyncScissorTest();
     auto& tev_stages = Pica::g_state.regs.GetTevStages();
     for (int index = 0; index < tev_stages.size(); ++index)
         SyncTevConstColor(index, tev_stages[index]);
@@ -657,6 +665,32 @@ void RasterizerOpenGL::SyncDepthTest() {
     state.depth.write_mask = regs.output_merger.depth_write_enable ? GL_TRUE : GL_FALSE;
 }
 
+void RasterizerOpenGL::SyncScissorTest() {
+    const auto& regs = Pica::g_state.regs;
+
+    GLsizei viewport_height = (GLsizei)Pica::float24::FromRawFloat24(regs.viewport_size_y).ToFloat32() * 2;
+
+    GLsizei viewport_corner_y = -(GLsizei)static_cast<float>(regs.viewport_corner.y)
+                                + regs.framebuffer.GetHeight() - viewport_height;
+
+    // OpenGL uses different y coordinates, so negate corner offset and flip origin
+    GLint scissor_bottom = viewport_height - (GLsizei)regs.scissor_test.bottom + viewport_corner_y;
+
+    GLint scissor_top = viewport_height - (GLsizei)regs.scissor_test.top - 1;
+
+    if (uniform_block_data.data.scissor_right != regs.scissor_test.right ||
+        uniform_block_data.data.scissor_bottom != scissor_bottom ||
+        uniform_block_data.data.scissor_left != regs.scissor_test.left + 1 ||
+        uniform_block_data.data.scissor_top != scissor_top) {
+
+        uniform_block_data.data.scissor_right = regs.scissor_test.right;
+        uniform_block_data.data.scissor_bottom = scissor_bottom;
+        uniform_block_data.data.scissor_left = regs.scissor_test.left + 1;
+        uniform_block_data.data.scissor_top = scissor_top;
+        uniform_block_data.dirty = true;
+    }
+}
+
 void RasterizerOpenGL::SyncCombinerColor() {
     auto combiner_color = PicaToGL::ColorRGBA8(Pica::g_state.regs.tev_combiner_buffer_color.raw);
     if (combiner_color != uniform_block_data.data.tev_combiner_buffer_color) {

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -36,6 +36,8 @@ struct PicaShaderConfig {
         res.alpha_test_func = regs.output_merger.alpha_test.enable ?
             regs.output_merger.alpha_test.func.Value() : Pica::Regs::CompareFunc::Always;
 
+        res.scissor_test_mode = regs.scissor_test.mode;
+
         // Copy relevant TevStageConfig fields only. We're doing this manually (instead of calling
         // the GetTevStages() function) because BitField explicitly disables copies.
 
@@ -87,6 +89,7 @@ struct PicaShaderConfig {
     };
 
     Pica::Regs::CompareFunc alpha_test_func;
+    Pica::Regs::ScissorMode scissor_test_mode;
     std::array<Pica::Regs::TevStageConfig, 6> tev_stages = {};
     u8 combiner_buffer_input;
 };
@@ -191,16 +194,19 @@ private:
         GLfloat tex_coord2[2];
     };
 
-    /// Uniform structure for the Uniform Buffer Object, all members must be 16-byte aligned
+    /// Uniform structure for the Uniform Buffer Object, all members must be aligned as per the OpenGL std140 layout specification
     struct UniformData {
         // A vec4 color for each of the six tev stages
         std::array<GLfloat, 4> const_color[6];
         std::array<GLfloat, 4> tev_combiner_buffer_color;
         GLint alphatest_ref;
-        INSERT_PADDING_BYTES(12);
+        GLint scissor_right;
+        GLint scissor_bottom;
+        GLint scissor_left;
+        GLint scissor_top;
     };
 
-    static_assert(sizeof(UniformData) == 0x80, "The size of the UniformData structure has changed, update the structure in the shader");
+    static_assert(sizeof(UniformData) == 0x84, "The size of the UniformData structure has changed, update the structure in the shader");
     static_assert(sizeof(UniformData) < 16000, "UniformData structure must be less than 16kb as per the OpenGL spec");
 
     /// Reconfigure the OpenGL color texture to use the given format and dimensions
@@ -239,6 +245,9 @@ private:
     /// Syncs the depth test states to match the PICA register
     void SyncDepthTest();
 
+    /// Syncs the scissor test state to match the PICA register
+    void SyncScissorTest();
+
     /// Syncs the TEV constant color to match the PICA register
     void SyncTevConstColor(int tev_index, const Pica::Regs::TevStageConfig& tev_stage);
 

src/video_core/renderer_opengl/gl_shader_gen.cpp

@@ -327,6 +327,7 @@ std::string GenerateFragmentShader(const PicaShaderConfig& config) {
 
 in vec4 primary_color;
 in vec2 texcoord[3];
+in vec4 gl_FragCoord;
 
 out vec4 color;
 
@@ -334,6 +335,10 @@ layout (std140) uniform shader_data {
     vec4 const_color[NUM_TEV_STAGES];
     vec4 tev_combiner_buffer_color;
     int alphatest_ref;
+    int scissor_right;
+    int scissor_bottom;
+    int scissor_left;
+    int scissor_top;
 };
 
 uniform sampler2D tex[3];
@@ -341,6 +346,16 @@ uniform sampler2D tex[3];
 void main() {
 )";
 
+    // Append the scissor test
+    if (config.scissor_test_mode == Regs::ScissorMode::Include || config.scissor_test_mode == Regs::ScissorMode::Exclude) {
+        out += "if (scissor_left <= scissor_right || scissor_top >= scissor_bottom) discard;\n";
+        out += "if (";
+        // Negate the condition if we have to keep only the pixels outside the scissor box
+        if (config.scissor_test_mode == Regs::ScissorMode::Include)
+            out += "!";
+        out += "(gl_FragCoord.x >= scissor_right && gl_FragCoord.x <= scissor_left && gl_FragCoord.y >= scissor_top && gl_FragCoord.y <= scissor_bottom)) discard;\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; }";