GPU/SW: Specialize rasterizing functions further

scripts/gen_sw_functions.py

@@ -0,0 +1,73 @@
+texmode = [
+  "GPUTextureMode::Palette4Bit",
+  "GPUTextureMode::Palette8Bit",
+  "GPUTextureMode::Direct16Bit",
+  "GPUTextureMode::Direct16Bit",
+
+  "GPUTextureMode::RawPalette4Bit",
+  "GPUTextureMode::RawPalette8Bit",
+  "GPUTextureMode::RawDirect16Bit",
+  "GPUTextureMode::RawDirect16Bit",
+
+  "GPUTextureMode::Disabled"
+]
+
+transparentmode = [
+  "GPUTransparencyMode::HalfBackgroundPlusHalfForeground",
+  "GPUTransparencyMode::BackgroundPlusForeground",
+  "GPUTransparencyMode::BackgroundMinusForeground",
+  "GPUTransparencyMode::BackgroundPlusQuarterForeground",
+  "GPUTransparencyMode::Disabled"
+]
+
+bools = ["false", "true"]
+
+
+"""
+print("const GPU_SW_Backend::DrawRectangleFunction GPU_SW_Backend::s_rectangle_draw_functions[%d][%d][2] = {" % (len(texmode), len(transparentmode)))
+for texture in range(len(texmode)):
+  print("  { // %s" % texmode[texture])
+  for transparency in range(len(transparentmode)):
+    print("    { // %s" % transparentmode[transparency])
+    for check_mask in range(2):
+      line = "&GPU_SW_Backend::DrawRectangle<%s, %s, %s>" % (texmode[texture], transparentmode[transparency], bools[check_mask])
+      print("      %s%s" % (line, "," if check_mask == 0 else ""))
+    print("    }%s" % ("," if transparency < (len(transparentmode) - 1) else ""))
+  print("  }%s" % ("," if texture < (len(texmode) - 1) else ""))
+print("};")
+"""
+
+"""
+print("const GPU_SW_Backend::DrawTriangleFunction GPU_SW_Backend::s_triangle_draw_functions[2][%d][%d][2][2] = {" % (len(texmode), len(transparentmode)))
+for shading in range(2):
+  print("  { // shading %s" % bools[shading])
+  for texture in range(len(texmode)):
+    print("    { // %s" % texmode[texture])
+    for transparency in range(len(transparentmode)):
+      print("      { // %s" % transparentmode[transparency])
+      for dither in range(2):
+        print("        { // dither %s" % bools[dither])
+        for check_mask in range(2):
+          line = "&GPU_SW_Backend::DrawTriangle<%s, %s, %s, %s, %s>" % (bools[shading], texmode[texture], transparentmode[transparency], bools[dither], bools[check_mask])
+          print("          %s%s" % (line, "," if check_mask == 0 else ""))
+        print("        }%s" % ("," if dither == 0 else ""))
+      print("      }%s" % ("," if transparency < (len(transparentmode) - 1) else ""))
+    print("    }%s" % ("," if texture < (len(texmode) - 1) else ""))
+  print("  }%s" % ("," if shading == 0 else ""))
+print("};")
+"""
+
+print("const GPU_SW_Backend::DrawLineFunction GPU_SW_Backend::s_line_draw_functions[2][%d][2][2] = {" % (len(transparentmode)))
+for shading in range(2):
+  print("  { // shading %s" % bools[shading])
+  for transparency in range(len(transparentmode)):
+    print("      { // %s" % transparentmode[transparency])
+    for dither in range(2):
+      print("        { // dither %s" % bools[dither])
+      for check_mask in range(2):
+        line = "&GPU_SW_Backend::DrawLine<%s, %s, %s, %s>" % (bools[shading], transparentmode[transparency], bools[dither], bools[check_mask])
+        print("          %s%s" % (line, "," if check_mask == 0 else ""))
+      print("        }%s" % ("," if dither == 0 else ""))
+    print("      }%s" % ("," if transparency < (len(transparentmode) - 1) else ""))
+  print("  }%s" % ("," if shading == 0 else ""))
+print("};")

src/core/gpu_sw_backend.h

@@ -98,17 +98,13 @@ protected:
   //////////////////////////////////////////////////////////////////////////
   // Rasterization
   //////////////////////////////////////////////////////////////////////////
-  template<bool texture_enable, bool raw_texture_enable, bool transparency_enable, bool dithering_enable>
+  template<GPUTextureMode texture_mode, GPUTransparencyMode transparency_mode, bool dithering_enable, bool check_mask>
   void ShadePixel(const GPUBackendDrawCommand* cmd, u32 x, u32 y, u8 color_r, u8 color_g, u8 color_b, u8 texcoord_x,
                   u8 texcoord_y);
 
-  template<bool texture_enable, bool raw_texture_enable, bool transparency_enable>
+  template<GPUTextureMode texture_mode, GPUTransparencyMode transparency_mode, bool check_mask>
   void DrawRectangle(const GPUBackendDrawRectangleCommand* cmd);
 
-  using DrawRectangleFunction = void (GPU_SW_Backend::*)(const GPUBackendDrawRectangleCommand* cmd);
-  DrawRectangleFunction GetDrawRectangleFunction(bool texture_enable, bool raw_texture_enable,
-                                                 bool transparency_enable);
-
   //////////////////////////////////////////////////////////////////////////
   // Polygon and line rasterization ported from Mednafen
   //////////////////////////////////////////////////////////////////////////
@@ -137,31 +133,36 @@ protected:
   template<bool shading_enable, bool texture_enable>
   void AddIDeltas_DY(i_group& ig, const i_deltas& idl, u32 count = 1);
 
-  template<bool shading_enable, bool texture_enable, bool raw_texture_enable, bool transparency_enable,
+  template<bool shading_enable, GPUTextureMode texture_mode, GPUTransparencyMode transparency_mode,
-           bool dithering_enable>
+           bool dithering_enable, bool check_mask>
   void DrawSpan(const GPUBackendDrawPolygonCommand* cmd, s32 y, s32 x_start, s32 x_bound, i_group ig,
                 const i_deltas& idl);
 
-  template<bool shading_enable, bool texture_enable, bool raw_texture_enable, bool transparency_enable,
+  template<bool shading_enable, GPUTextureMode texture_mode, GPUTransparencyMode transparency_mode,
-           bool dithering_enable>
+           bool dithering_enable, bool check_mask>
   void DrawTriangle(const GPUBackendDrawPolygonCommand* cmd, const GPUBackendDrawPolygonCommand::Vertex* v0,
                     const GPUBackendDrawPolygonCommand::Vertex* v1, const GPUBackendDrawPolygonCommand::Vertex* v2);
 
+  template<bool shading_enable, GPUTransparencyMode transparency_mode, bool dithering_enable, bool check_mask>
+  void DrawLine(const GPUBackendDrawLineCommand* cmd, const GPUBackendDrawLineCommand::Vertex* p0,
+                const GPUBackendDrawLineCommand::Vertex* p1);
+
+#if 0
+  std::array<u16, VRAM_WIDTH * VRAM_HEIGHT> m_vram;
+#else
+  u16 m_vram[VRAM_WIDTH * VRAM_HEIGHT];
+#endif
+
+  using DrawRectangleFunction = void (GPU_SW_Backend::*)(const GPUBackendDrawRectangleCommand* cmd);
   using DrawTriangleFunction = void (GPU_SW_Backend::*)(const GPUBackendDrawPolygonCommand* cmd,
                                                         const GPUBackendDrawPolygonCommand::Vertex* v0,
                                                         const GPUBackendDrawPolygonCommand::Vertex* v1,
                                                         const GPUBackendDrawPolygonCommand::Vertex* v2);
-  DrawTriangleFunction GetDrawTriangleFunction(bool shading_enable, bool texture_enable, bool raw_texture_enable,
-                                               bool transparency_enable, bool dithering_enable);
-
-  template<bool shading_enable, bool transparency_enable, bool dithering_enable>
-  void DrawLine(const GPUBackendDrawLineCommand* cmd, const GPUBackendDrawLineCommand::Vertex* p0,
-                const GPUBackendDrawLineCommand::Vertex* p1);
-
   using DrawLineFunction = void (GPU_SW_Backend::*)(const GPUBackendDrawLineCommand* cmd,
                                                     const GPUBackendDrawLineCommand::Vertex* p0,
                                                     const GPUBackendDrawLineCommand::Vertex* p1);
-  DrawLineFunction GetDrawLineFunction(bool shading_enable, bool transparency_enable, bool dithering_enable);
 
-  std::array<u16, VRAM_WIDTH * VRAM_HEIGHT> m_vram;
+  static const DrawRectangleFunction s_rectangle_draw_functions[9][5][2];
+  static const DrawTriangleFunction s_triangle_draw_functions[2][9][5][2][2];
+  static const DrawLineFunction s_line_draw_functions[2][5][2][2];
 };