GPU: Implement dithering on OpenGL backend

src/core/gpu.h

@@ -181,10 +181,22 @@ protected:
     BitField<u32, bool, 28, 1> shading_enable;              // 0 - flat, 1 = gouroud
     BitField<u32, Primitive, 29, 21> primitive;
 
-    // Helper functions.
+    // Returns true if dithering should be enabled. Depends on the primitive type.
-    bool IsTextureEnabled() const { return (primitive != Primitive::Line && texture_enable); }
+    bool IsDitheringEnabled() const
-    bool IsTextureBlendingEnabled() const { return (IsTextureEnabled() && !raw_texture_enable); }
+    {
-    bool IsTransparencyEnabled() const { return transparency_enable; }
+      switch (primitive)
+      {
+        case Primitive::Polygon:
+          return shading_enable || !raw_texture_enable;
+
+        case Primitive::Line:
+          return true;
+
+        case Primitive::Rectangle:
+        default:
+          return false;
+      }
+    }
   };
 
   // TODO: Use BitField to do sign extending instead
@@ -241,6 +253,12 @@ protected:
     }
   };
 
+  // 4x4 dither matrix.
+  static constexpr s32 DITHER_MATRIX[4][4] = {{-4, +0, -3, +1},  // row 0
+                                              {+2, -2, +3, -1},  // row 1
+                                              {-3, +1, -4, +0},  // row 2
+                                              {+4, -1, +2, -2}}; // row 3
+
   void SoftReset();
 
   // Sets dots per scanline

src/core/gpu_hw.cpp

@@ -241,7 +241,7 @@ void main()
   return ss.str();
 }
 
-std::string GPU_HW::GenerateFragmentShader(HWBatchRenderMode transparency, TextureMode texture_mode)
+std::string GPU_HW::GenerateFragmentShader(HWBatchRenderMode transparency, TextureMode texture_mode, bool dithering)
 {
   const TextureMode actual_texture_mode = texture_mode & ~TextureMode::RawTextureBit;
   const bool raw_texture = (texture_mode & TextureMode::RawTextureBit) == TextureMode::RawTextureBit;
@@ -258,6 +258,16 @@ std::string GPU_HW::GenerateFragmentShader(HWBatchRenderMode transparency, Textu
   DefineMacro(ss, "PALETTE_4_BIT", actual_texture_mode == GPU::TextureMode::Palette4Bit);
   DefineMacro(ss, "PALETTE_8_BIT", actual_texture_mode == GPU::TextureMode::Palette8Bit);
   DefineMacro(ss, "RAW_TEXTURE", raw_texture);
+  DefineMacro(ss, "DITHERING", dithering);
+
+  ss << "const int[16] s_dither_values = int[16]( ";
+  for (u32 i = 0; i < 16; i++)
+  {
+    if (i > 0)
+      ss << ", ";
+    ss << DITHER_MATRIX[i / 4][i % 4];
+  }
+  ss << " );\n";
 
   ss << R"(
 in vec3 v_col0;
@@ -271,6 +281,22 @@ uniform vec2 u_transparent_alpha;
 
 out vec4 o_col0;
 
+vec4 ApplyDithering(vec4 col)
+{
+  ivec3 icol = ivec3(col.rgb * vec3(255.0, 255.0, 255.0));
+
+  // apply dither
+  ivec2 fc = ivec2(gl_FragCoord.xy) & ivec2(3, 3);
+  int offset = s_dither_values[fc.y * 4 + fc.x];
+  icol += ivec3(offset, offset, offset);
+
+  // saturate
+  icol = clamp(icol, ivec3(0, 0, 0), ivec3(255, 255, 255));
+
+  // clip to 5-bit range
+  return vec4((icol.rgb >> 3) / vec3(31.0, 31.0, 31.0), col.a);
+}
+
 #if TEXTURED
 ivec2 ApplyNativeTextureWindow(ivec2 coords)
 {
@@ -370,6 +396,10 @@ void main()
       o_col0 = vec4(v_col0, 0.0);
     #endif
   #endif
+
+  #if DITHERING
+    o_col0 = ApplyDithering(o_col0);
+  #endif
 }
 )";
 
@@ -547,13 +577,15 @@ void GPU_HW::DispatchRenderCommand(RenderCommand rc, u32 num_vertices, const u32
   const TransparencyMode transparency_mode =
     rc.transparency_enable ? m_render_state.transparency_mode : TransparencyMode::Disabled;
   const HWPrimitive rc_primitive = GetPrimitiveForCommand(rc);
+  const bool dithering_enable = rc.IsDitheringEnabled() ? m_GPUSTAT.dither_enable : false;
   if (!IsFlushed())
   {
     const u32 max_added_vertices = num_vertices + 2;
     const bool buffer_overflow = (m_batch.vertices.size() + max_added_vertices) >= MAX_BATCH_VERTEX_COUNT;
     if (buffer_overflow || rc_primitive == HWPrimitive::LineStrip || m_batch.texture_mode != texture_mode ||
         m_batch.transparency_mode != transparency_mode || m_batch.primitive != rc_primitive ||
-        m_render_state.IsTexturePageChanged() || m_render_state.IsTextureWindowChanged())
+        dithering_enable != m_batch.dithering || m_render_state.IsTexturePageChanged() ||
+        m_render_state.IsTextureWindowChanged())
     {
       FlushRender();
     }
@@ -563,6 +595,7 @@ void GPU_HW::DispatchRenderCommand(RenderCommand rc, u32 num_vertices, const u32
   m_batch.primitive = rc_primitive;
   m_batch.texture_mode = texture_mode;
   m_batch.transparency_mode = transparency_mode;
+  m_batch.dithering = dithering_enable;
 
   if (m_render_state.IsTexturePageChanged())
   {

src/core/gpu_hw.h

@@ -44,14 +44,15 @@ protected:
 
   struct HWRenderBatch
   {
-    HWPrimitive primitive;
-    TextureMode texture_mode;
     u32 texture_page_x;
     u32 texture_page_y;
     u32 texture_palette_x;
     u32 texture_palette_y;
+    HWPrimitive primitive;
+    TextureMode texture_mode;
     TransparencyMode transparency_mode;
     std::array<u8, 4> texture_window_values;
+    bool dithering;
 
     std::vector<HWVertex> vertices;
 
@@ -100,7 +101,7 @@ protected:
   }
 
   std::string GenerateVertexShader(bool textured);
-  std::string GenerateFragmentShader(HWBatchRenderMode transparency, TextureMode texture_mode);
+  std::string GenerateFragmentShader(HWBatchRenderMode transparency, TextureMode texture_mode, bool dithering);
   std::string GenerateScreenQuadVertexShader();
   std::string GenerateFillFragmentShader();
   std::string GenerateDisplayFragmentShader(bool depth_24bit, bool interlaced);

src/core/gpu_hw_opengl.cpp

@@ -121,8 +121,6 @@ void GPU_HW_OpenGL::DrawRendererStatsWindow()
     ImGui::NextColumn();
     ImGui::Text("%u", m_last_stats.num_vertices);
     ImGui::NextColumn();
-
-
   }
 
   ImGui::End();
@@ -246,13 +244,17 @@ bool GPU_HW_OpenGL::CompilePrograms()
   {
     for (u32 texture_mode = 0; texture_mode < 9; texture_mode++)
     {
-      if (!CompileProgram(m_render_programs[render_mode][texture_mode], static_cast<HWBatchRenderMode>(render_mode),
+      for (u8 dithering = 0; dithering < 2; dithering++)
-                          static_cast<TextureMode>(texture_mode)))
+      {
+        if (!CompileProgram(m_render_programs[render_mode][texture_mode][dithering],
+                            static_cast<HWBatchRenderMode>(render_mode), static_cast<TextureMode>(texture_mode),
+                            ConvertToBoolUnchecked(dithering)))
         {
           return false;
         }
       }
     }
+  }
 
   // TODO: Use string_view
   for (u8 depth_24bit = 0; depth_24bit < 2; depth_24bit++)
@@ -280,11 +282,12 @@ bool GPU_HW_OpenGL::CompilePrograms()
   return true;
 }
 
-bool GPU_HW_OpenGL::CompileProgram(GL::Program& prog, HWBatchRenderMode render_mode, TextureMode texture_mode)
+bool GPU_HW_OpenGL::CompileProgram(GL::Program& prog, HWBatchRenderMode render_mode, TextureMode texture_mode,
+                                   bool dithering)
 {
   const bool textured = texture_mode != TextureMode::Disabled;
   const std::string vs = GenerateVertexShader(textured);
-  const std::string fs = GenerateFragmentShader(render_mode, texture_mode);
+  const std::string fs = GenerateFragmentShader(render_mode, texture_mode, dithering);
   if (!prog.Compile(vs.c_str(), fs.c_str()))
     return false;
 
@@ -319,7 +322,8 @@ bool GPU_HW_OpenGL::CompileProgram(GL::Program& prog, HWBatchRenderMode render_m
 
 void GPU_HW_OpenGL::SetDrawState(HWBatchRenderMode render_mode)
 {
-  const GL::Program& prog = m_render_programs[static_cast<u32>(render_mode)][static_cast<u32>(m_batch.texture_mode)];
+  const GL::Program& prog =
+    m_render_programs[static_cast<u32>(render_mode)][static_cast<u32>(m_batch.texture_mode)][m_batch.dithering];
   prog.Bind();
 
   prog.Uniform2i(0, m_drawing_offset.x, m_drawing_offset.y);

src/core/gpu_hw_opengl.h

@@ -42,8 +42,6 @@ private:
     u32 num_vram_read_texture_updates;
   };
 
-  void DrawRendererStatistics();
-
   std::tuple<s32, s32> ConvertToFramebufferCoordinates(s32 x, s32 y);
 
   void SetMaxResolutionScale();
@@ -55,7 +53,7 @@ private:
   void CreateVertexBuffer();
 
   bool CompilePrograms();
-  bool CompileProgram(GL::Program& prog, HWBatchRenderMode render_mode, TextureMode texture_mode);
+  bool CompileProgram(GL::Program& prog, HWBatchRenderMode render_mode, TextureMode texture_mode, bool dithering);
   void SetDrawState(HWBatchRenderMode render_mode);
 
   // downsample texture - used for readbacks at >1xIR.
@@ -72,7 +70,7 @@ private:
   bool m_drawing_area_changed = true;
   bool m_show_renderer_statistics = false;
 
-  std::array<std::array<GL::Program, 9>, 4> m_render_programs;  // [render_mode][texture_mode]
+  std::array<std::array<std::array<GL::Program, 2>, 9>, 4> m_render_programs; // [render_mode][texture_mode][dithering]
   std::array<std::array<GL::Program, 2>, 2> m_display_programs;               // [depth_24][interlaced]
 
   GLStats m_stats = {};