Implement geometry shaders (Battle-tested Edition)

src/citra_qt/debugger/graphics_breakpoints.cpp

@@ -36,6 +36,8 @@ QVariant BreakPointModel::data(const QModelIndex& index, int role) const {
                 {Pica::DebugContext::Event::IncomingPrimitiveBatch, tr("Incoming primitive batch")},
                 {Pica::DebugContext::Event::FinishedPrimitiveBatch, tr("Finished primitive batch")},
                 {Pica::DebugContext::Event::VertexShaderInvocation, tr("Vertex shader invocation")},
+                {Pica::DebugContext::Event::GeometryShaderInvocation,
+                 tr("Geometry shader invocation")},
                 {Pica::DebugContext::Event::IncomingDisplayTransfer,
                  tr("Incoming display transfer")},
                 {Pica::DebugContext::Event::GSPCommandProcessed, tr("GSP command processed")},

src/citra_qt/debugger/graphics_vertex_shader.cpp

@@ -522,8 +522,7 @@ void GraphicsVertexShaderWidget::Reload(bool replace_vertex_data, void* vertex_d
     info.labels.insert({entry_point, "main"});
 
     // Generate debug information
-    debug_data = Pica::g_state.vs.ProduceDebugInfo(input_vertex, num_attributes, shader_config,
-                                                   shader_setup);
+    debug_data = Pica::g_state.vs.ProduceDebugInfo(input_vertex, num_attributes, shader_config);
 
     // Reload widget state
     for (int attr = 0; attr < num_attributes; ++attr) {

src/video_core/debug_utils/debug_utils.h

@@ -39,6 +39,7 @@ class DebugContext {
         IncomingPrimitiveBatch,
         FinishedPrimitiveBatch,
         VertexShaderInvocation,
+        GeometryShaderInvocation,
         IncomingDisplayTransfer,
         GSPCommandProcessed,
         BufferSwapped,

src/video_core/pica.h

@@ -1128,7 +1128,7 @@ struct Regs {
     // Number of vertices to render
     u32 num_vertices;
 
-    INSERT_PADDING_WORDS(0x1);
+    BitField<0, 2, u32> using_geometry_shader;
 
     // The index of the first vertex to render
     u32 vertex_offset;
@@ -1176,7 +1176,11 @@ struct Regs {
         }
     } command_buffer;
 
-    INSERT_PADDING_WORDS(0x07);
+    INSERT_PADDING_WORDS(0x06);
+
+    enum class VSComMode : u32 { Shared = 0, Exclusive = 1 };
+
+    VSComMode vs_com_mode;
 
     enum class GPUMode : u32 {
         Drawing = 0,
@@ -1185,7 +1189,17 @@ struct Regs {
 
     GPUMode gpu_mode;
 
-    INSERT_PADDING_WORDS(0x18);
+    INSERT_PADDING_WORDS(0x4);
+
+    BitField<0, 4, u32> vs_outmap_total1;
+
+    INSERT_PADDING_WORDS(0x6);
+
+    BitField<0, 4, u32> vs_outmap_total2;
+
+    BitField<0, 4, u32> gsh_misc0;
+
+    INSERT_PADDING_WORDS(0xB);
 
     enum class TriangleTopology : u32 {
         List = 0,
@@ -1194,7 +1208,10 @@ struct Regs {
         Shader = 3, // Programmable setup unit implemented in a geometry shader
     };
 
-    BitField<8, 2, TriangleTopology> triangle_topology;
+    union {
+        BitField<0, 4, u32> vs_outmap_count;
+        BitField<8, 2, TriangleTopology> triangle_topology;
+    };
 
     u32 restart_primitive;
 
@@ -1213,8 +1230,10 @@ struct Regs {
         INSERT_PADDING_WORDS(0x4);
 
         union {
-            // Number of input attributes to shader unit - 1
-            BitField<0, 4, u32> num_input_attributes;
+            BitField<0, 4, u32>
+                num_input_attributes; // Number of input attributes to shader unit - 1
+            BitField<8, 4, u32> use_subdivision;
+            BitField<24, 8, u32> use_geometry_shader;
         };
 
         // Offset to shader program entry point (in words)
@@ -1267,6 +1286,8 @@ struct Regs {
             }
 
             union {
+                u32 setup;
+
                 // Index of the next uniform to write to
                 // TODO: ctrulib uses 8 bits for this, however that seems to yield lots of invalid
                 // indices
@@ -1391,7 +1412,11 @@ ASSERT_REG_POSITION(trigger_draw, 0x22e);
 ASSERT_REG_POSITION(trigger_draw_indexed, 0x22f);
 ASSERT_REG_POSITION(vs_default_attributes_setup, 0x232);
 ASSERT_REG_POSITION(command_buffer, 0x238);
+ASSERT_REG_POSITION(vs_com_mode, 0x244);
 ASSERT_REG_POSITION(gpu_mode, 0x245);
+ASSERT_REG_POSITION(vs_outmap_total1, 0x24A);
+ASSERT_REG_POSITION(vs_outmap_total2, 0x251);
+ASSERT_REG_POSITION(gsh_misc0, 0x252);
 ASSERT_REG_POSITION(triangle_topology, 0x25e);
 ASSERT_REG_POSITION(restart_primitive, 0x25f);
 ASSERT_REG_POSITION(gs, 0x280);

src/video_core/pica_state.h

@@ -20,6 +20,8 @@ struct State {
     /// Pica registers
     Regs regs;
 
+    Shader::UnitState<false> shader_units[4];
+
     Shader::ShaderSetup vs;
     Shader::ShaderSetup gs;
 
@@ -73,6 +75,14 @@ struct State {
 
     // This is constructed with a dummy triangle topology
     PrimitiveAssembler<Shader::OutputVertex> primitive_assembler;
+
+    /// Current geometry shader state
+    struct GeometryShaderState {
+        // Buffer used for geometry shader inputs
+        Shader::InputVertex buffer;
+        // The current index into the buffer
+        unsigned int index;
+    } gs_input_buffer;
 };
 
 extern State g_state; ///< Current Pica state

src/video_core/primitive_assembly.cpp

@@ -17,7 +17,6 @@ template <typename VertexType>
 void PrimitiveAssembler<VertexType>::SubmitVertex(VertexType& vtx,
                                                   TriangleHandler triangle_handler) {
     switch (topology) {
-    // TODO: Figure out what's different with TriangleTopology::Shader.
     case Regs::TriangleTopology::List:
     case Regs::TriangleTopology::Shader:
         if (buffer_index < 2) {