Videocore: Implement simple vertex caching

This gives a ~2/3 reduction in the amount of vertices that need to be
processed through the vertex loaders and the vertex shader, yielding a
good speedup.

src/video_core/command_processor.cpp

@@ -206,18 +206,42 @@ static inline void WritePicaReg(u32 id, u32 value, u32 mask) {
                 std::map<u32, u32> ranges;
             } memory_accesses;
 
+            // Simple circular-replacement vertex cache
+            // The size has been tuned for optimal balance between hit-rate and the cost of lookup
+            const size_t VERTEX_CACHE_SIZE = 32;
+            std::array<u16, VERTEX_CACHE_SIZE> vertex_cache_ids;
+            std::array<VertexShader::OutputVertex, VERTEX_CACHE_SIZE> vertex_cache;
+
+            unsigned int vertex_cache_pos = 0;
+            vertex_cache_ids.fill(-1);
+
             for (unsigned int index = 0; index < regs.num_vertices; ++index)
             {
                 unsigned int vertex = is_indexed ? (index_u16 ? index_address_16[index] : index_address_8[index]) : index;
 
+                // -1 is a common special value used for primitive restart. Since it's unknown if
+                // the PICA supports it, and it would mess up the caching, guard against it here.
+                ASSERT(vertex != -1);
+
+                bool vertex_cache_hit = false;
+                VertexShader::OutputVertex output;
+
                 if (is_indexed) {
-                    // TODO: Implement some sort of vertex cache!
                     if (g_debug_context && Pica::g_debug_context->recorder) {
                         int size = index_u16 ? 2 : 1;
                         memory_accesses.AddAccess(base_address + index_info.offset + size * index, size);
                     }
+
+                    for (unsigned int i = 0; i < VERTEX_CACHE_SIZE; ++i) {
+                        if (vertex == vertex_cache_ids[i]) {
+                            output = vertex_cache[i];
+                            vertex_cache_hit = true;
+                            break;
+                        }
+                    }
                 }
 
+                if (!vertex_cache_hit) {
                     // Initialize data for the current vertex
                     VertexShader::InputVertex input;
 
@@ -262,9 +286,9 @@ static inline void WritePicaReg(u32 id, u32 value, u32 mask) {
                                       input.attr[i][0].ToFloat32(), input.attr[i][1].ToFloat32(),
                                       input.attr[i][2].ToFloat32(), input.attr[i][3].ToFloat32());
                         } else {
-                        // TODO(yuriks): In this case, no data gets loaded and the vertex remains
-                        //              with the last value it had. This isn't currently maintained
-                        //              as global state, however, and so won't work in Cita yet.
+                            // TODO(yuriks): In this case, no data gets loaded and the vertex
+                            // remains with the last value it had. This isn't currently maintained
+                            // as global state, however, and so won't work in Citra yet.
                         }
                     }
 
@@ -284,10 +308,13 @@ static inline void WritePicaReg(u32 id, u32 value, u32 mask) {
 #endif
 
                     // Send to vertex shader
-                VertexShader::OutputVertex output = VertexShader::RunShader(input, attribute_config.GetNumTotalAttributes(), g_state.regs.vs, g_state.vs);
+                    output = VertexShader::RunShader(input, attribute_config.GetNumTotalAttributes(), g_state.regs.vs, g_state.vs);
 
                     if (is_indexed) {
-                    // TODO: Add processed vertex to vertex cache!
+                        vertex_cache[vertex_cache_pos] = output;
+                        vertex_cache_ids[vertex_cache_pos] = vertex;
+                        vertex_cache_pos = (vertex_cache_pos + 1) % VERTEX_CACHE_SIZE;
+                    }
                 }
 
                 if (Settings::values.use_hw_renderer) {