GPU: Properly handle semitransparent pixels

src/common/gl_program.cpp

@@ -263,4 +263,76 @@ void Program::Uniform4f(u32 index, float x, float y, float z, float w) const
     glUniform4f(location, x, y, z, w);
 }
 
+void Program::Uniform2uiv(u32 index, const u32* v) const
+{
+  Assert(index < m_uniform_locations.size());
+  const int location = m_uniform_locations[index];
+  if (location >= 0)
+    glUniform2uiv(location, 1, v);
+}
+
+void Program::Uniform3uiv(u32 index, const u32* v) const
+{
+  Assert(index < m_uniform_locations.size());
+  const int location = m_uniform_locations[index];
+  if (location >= 0)
+    glUniform3uiv(location, 1, v);
+}
+
+void Program::Uniform4uiv(u32 index, const u32* v) const
+{
+  Assert(index < m_uniform_locations.size());
+  const int location = m_uniform_locations[index];
+  if (location >= 0)
+    glUniform4uiv(location, 1, v);
+}
+
+void Program::Uniform2iv(u32 index, const s32* v) const
+{
+  Assert(index < m_uniform_locations.size());
+  const int location = m_uniform_locations[index];
+  if (location >= 0)
+    glUniform2iv(location, 1, v);
+}
+
+void Program::Uniform3iv(u32 index, const s32* v) const
+{
+  Assert(index < m_uniform_locations.size());
+  const int location = m_uniform_locations[index];
+  if (location >= 0)
+    glUniform3iv(location, 1, v);
+}
+
+void Program::Uniform4iv(u32 index, const s32* v) const
+{
+  Assert(index < m_uniform_locations.size());
+  const int location = m_uniform_locations[index];
+  if (location >= 0)
+    glUniform4iv(location, 1, v);
+}
+
+void Program::Uniform2fv(u32 index, const float* v) const
+{
+  Assert(index < m_uniform_locations.size());
+  const int location = m_uniform_locations[index];
+  if (location >= 0)
+    glUniform2fv(location, 1, v);
+}
+
+void Program::Uniform3fv(u32 index, const float* v) const
+{
+  Assert(index < m_uniform_locations.size());
+  const int location = m_uniform_locations[index];
+  if (location >= 0)
+    glUniform3fv(location, 1, v);
+}
+
+void Program::Uniform4fv(u32 index, const float* v) const
+{
+  Assert(index < m_uniform_locations.size());
+  const int location = m_uniform_locations[index];
+  if (location >= 0)
+    glUniform4fv(location, 1, v);
+}
+
 } // namespace GL

src/common/gl_program.h

@@ -40,6 +40,15 @@ public:
   void Uniform2f(u32 index, float x, float y) const;
   void Uniform3f(u32 index, float x, float y, float z) const;
   void Uniform4f(u32 index, float x, float y, float z, float w) const;
+  void Uniform2uiv(u32 index, const u32* v) const;
+  void Uniform3uiv(u32 index, const u32* v) const;
+  void Uniform4uiv(u32 index, const u32* v) const;
+  void Uniform2iv(u32 index, const s32* v) const;
+  void Uniform3iv(u32 index, const s32* v) const;
+  void Uniform4iv(u32 index, const s32* v) const;
+  void Uniform2fv(u32 index, const float* v) const;
+  void Uniform3fv(u32 index, const float* v) const;
+  void Uniform4fv(u32 index, const float* v) const;
 
 private:
   GLuint m_program_id = 0;

src/pse/gpu_hw.cpp

@@ -176,7 +176,7 @@ in ivec2 a_pos;
 in vec4 a_col0;
 in vec2 a_tex0;
 
-out vec4 v_col0;
+out vec3 v_col0;
 #if TEXTURED
   out vec2 v_tex0;
 #endif
@@ -190,7 +190,7 @@ void main()
   float pos_y = (float(a_pos.y + u_pos_offset.y) / -256.0) + 1.0;
   gl_Position = vec4(pos_x, pos_y, 0.0, 1.0);
 
-  v_col0 = a_col0;
+  v_col0 = a_col0.rgb;
   #if TEXTURED
     v_tex0 = a_tex0;
   #endif
@@ -200,7 +200,8 @@ void main()
   return ss.str();
 }
 
-std::string GPU_HW::GenerateFragmentShader(bool textured, bool blending, TextureColorMode texture_color_mode)
+std::string GPU_HW::GenerateFragmentShader(bool textured, bool blending, bool transparent,
+                                           TextureColorMode texture_color_mode)
 {
   std::stringstream ss;
   GenerateShaderHeader(ss);
@@ -213,7 +214,8 @@ std::string GPU_HW::GenerateFragmentShader(bool textured, bool blending, Texture
   DefineMacro(ss, "PALETTE_8_BIT", textured && texture_color_mode == GPU::TextureColorMode::Palette8Bit);
 
   ss << R"(
-in vec4 v_col0;
+in vec3 v_col0;
+uniform vec2 u_transparent_alpha;
 #if TEXTURED
   in vec2 v_tex0;
   uniform sampler2D samp0;
@@ -272,13 +274,30 @@ void main()
     if (texcol == vec4(0.0, 0.0, 0.0, 0.0))
       discard;
 
+    vec3 color;
     #if BLENDING
-      o_col0 = vec4((ivec4(v_col0 * 255.0) * ivec4(texcol * 255.0)) >> 7) / 255.0;
+      color = vec3((ivec3(v_col0 * 255.0) * ivec3(texcol.rgb * 255.0)) >> 7) / 255.0;
     #else
-      o_col0 = texcol;
+      color = texcol.rgb;
+    #endif
+
+    #if TRANSPARENT
+      // Apply semitransparency. If not a semitransparent texel, destination alpha is ignored.
+      if (texcol.a != 0)
+        o_col0 = vec4(color * u_transparent_alpha.x, u_transparent_alpha.y);
+      else
+        o_col0 = vec4(color, 0.0);
+    #else
+      // Mask bit from texture.
+      o_col0 = vec4(color, texcol.a);
     #endif
   #else
-    o_col0 = v_col0;
+    #if TRANSPARENT
+      o_col0 = vec4(v_col0 * u_transparent_alpha.x, u_transparent_alpha.y);
+    #else
+      // Mask bit is cleared for untextured polygons.
+      o_col0 = vec4(v_col0, 0.0);
+    #endif
   #endif
 }
 )";

src/pse/gpu_hw.h

@@ -18,6 +18,7 @@ protected:
     u32 color;
     u16 texcoord;
     u16 padding;
+    u32 texpage;
 
     static constexpr std::tuple<u8, u8> DecodeTexcoord(u16 texcoord)
     {
@@ -74,7 +75,7 @@ protected:
   void CalcScissorRect(int* left, int* top, int* right, int* bottom);
 
   std::string GenerateVertexShader(bool textured);
-  std::string GenerateFragmentShader(bool textured, bool blending, TextureColorMode texture_color_mode);
+  std::string GenerateFragmentShader(bool textured, bool blending, bool transparent, TextureColorMode texture_color_mode);
   std::string GenerateScreenQuadVertexShader();
   std::string GenerateFillFragmentShader();
 

src/pse/gpu_hw_opengl.cpp

@@ -142,26 +142,31 @@ bool GPU_HW_OpenGL::CompilePrograms()
   for (u32 textured = 0; textured < 2; textured++)
   {
     for (u32 blending = 0; blending < 2; blending++)
+    {
+      for (u32 transparent = 0; transparent < 2; transparent++)
       {
         for (u32 format = 0; format < 3; format++)
         {
           // TODO: eliminate duplicate shaders here
-        if (!CompileProgram(m_render_programs[textured][blending][format], ConvertToBoolUnchecked(textured),
-                            ConvertToBoolUnchecked(blending), static_cast<TextureColorMode>(format)))
+          if (!CompileProgram(m_render_programs[textured][blending][transparent][format],
+                              ConvertToBoolUnchecked(textured), ConvertToBoolUnchecked(blending),
+                              ConvertToBoolUnchecked(transparent), static_cast<TextureColorMode>(format)))
           {
             return false;
           }
         }
       }
     }
+  }
 
   return true;
 }
 
-bool GPU_HW_OpenGL::CompileProgram(GL::Program& prog, bool textured, bool blending, TextureColorMode texture_color_mode)
+bool GPU_HW_OpenGL::CompileProgram(GL::Program& prog, bool textured, bool blending, bool transparent,
+                                   TextureColorMode texture_color_mode)
 {
   const std::string vs = GenerateVertexShader(textured);
-  const std::string fs = GenerateFragmentShader(textured, blending, texture_color_mode);
+  const std::string fs = GenerateFragmentShader(textured, blending, transparent, texture_color_mode);
   if (!prog.Compile(vs.c_str(), fs.c_str()))
     return false;
 
@@ -177,14 +182,16 @@ bool GPU_HW_OpenGL::CompileProgram(GL::Program& prog, bool textured, bool blendi
 
   prog.Bind();
   prog.RegisterUniform("u_pos_offset");
+  prog.RegisterUniform("u_transparent_alpha");
   prog.Uniform2i(0, 0, 0);
+  prog.Uniform2f(1, 1.0f, 0.0f);
 
   if (textured)
   {
     prog.RegisterUniform("samp0");
     prog.RegisterUniform("u_texture_page_base");
     prog.RegisterUniform("u_texture_palette_base");
-    prog.Uniform1i(1, 0);
+    prog.Uniform1i(2, 0);
   }
 
   return true;
@@ -194,16 +201,26 @@ void GPU_HW_OpenGL::SetProgram()
 {
   const GL::Program& prog =
     m_render_programs[BoolToUInt32(m_batch.texture_enable)][BoolToUInt32(m_batch.texture_blending_enable)]
-                     [static_cast<u32>(m_batch.texture_color_mode)];
+                     [BoolToUInt32(m_batch.transparency_enable)][static_cast<u32>(m_batch.texture_color_mode)];
   prog.Bind();
 
   prog.Uniform2i(0, m_drawing_offset.x, m_drawing_offset.y);
+  if (m_batch.transparency_enable)
+  {
+    static constexpr float transparent_alpha[4][2] = {{0.5f, 0.5f}, {1.0f, 1.0f}, {1.0f, 1.0f}, {0.25f, 1.0f}};
+    prog.Uniform2fv(1, transparent_alpha[static_cast<u32>(m_batch.transparency_mode)]);
+  }
+  else
+  {
+    static constexpr float disabled_alpha[2] = {1.0f, 0.0f};
+    prog.Uniform2fv(1, disabled_alpha);
+  }
 
   if (m_batch.texture_enable)
   {
     m_vram_read_texture->Bind();
-    prog.Uniform2i(2, m_batch.texture_page_x, m_batch.texture_page_y);
-    prog.Uniform2i(3, m_batch.texture_palette_x, m_batch.texture_palette_y);
+    prog.Uniform2i(3, m_batch.texture_page_x, m_batch.texture_page_y);
+    prog.Uniform2i(4, m_batch.texture_palette_x, m_batch.texture_palette_y);
   }
 }
 
@@ -228,32 +245,18 @@ void GPU_HW_OpenGL::SetScissor()
 
 void GPU_HW_OpenGL::SetBlendState()
 {
-  struct BlendVars
-  {
-    GLenum src_factor;
-    GLenum func;
-    GLenum dst_factor;
-    float color;
-  };
-  static const std::array<BlendVars, 4> blend_vars = {{
-    {GL_CONSTANT_COLOR, GL_FUNC_ADD, GL_CONSTANT_COLOR, 0.5f}, // B/2 + F/2
-    {GL_ONE, GL_FUNC_ADD, GL_ONE, -1.0f},                      // B + F
-    {GL_ONE, GL_FUNC_REVERSE_SUBTRACT, GL_ONE, -1.0f},         // B - F
-    {GL_CONSTANT_COLOR, GL_FUNC_ADD, GL_ONE, 0.25f}            // B + F/4
-  }};
-
   if (!m_batch.transparency_enable)
   {
     glDisable(GL_BLEND);
     return;
   }
 
-  const BlendVars& vars = blend_vars[static_cast<u8>(m_batch.transparency_mode)];
   glEnable(GL_BLEND);
-  glBlendFuncSeparate(vars.src_factor, vars.dst_factor, GL_ONE, GL_ZERO);
-  glBlendEquationSeparate(vars.func, GL_FUNC_ADD);
-  if (vars.color >= 0.0f)
-    glBlendColor(vars.color, vars.color, vars.color, 1.0f);
+  glBlendEquationSeparate(m_batch.transparency_mode == GPU::TransparencyMode::BackgroundMinusForeground ?
+                            GL_FUNC_REVERSE_SUBTRACT :
+                            GL_FUNC_ADD,
+                          GL_FUNC_ADD);
+  glBlendFuncSeparate(GL_ONE, GL_SRC_ALPHA, GL_ONE, GL_ZERO);
 }
 
 void GPU_HW_OpenGL::UpdateDisplay()

src/pse/gpu_hw_opengl.h

@@ -44,7 +44,7 @@ private:
   void CreateVertexBuffer();
 
   bool CompilePrograms();
-  bool CompileProgram(GL::Program& prog, bool textured, bool blending, TextureColorMode texture_color_mode);
+  bool CompileProgram(GL::Program& prog, bool textured, bool blending, bool transparent, TextureColorMode texture_color_mode);
 
   void SetProgram();
   void SetViewport();
@@ -62,7 +62,7 @@ private:
   GLuint m_vao_id = 0;
   GLuint m_attributeless_vao_id = 0;
 
-  std::array<std::array<std::array<GL::Program, 3>, 2>, 2> m_render_programs;
+  std::array<std::array<std::array<std::array<GL::Program, 3>, 2>, 2>, 2> m_render_programs;
   std::array<GL::Program, 3> m_texture_page_programs;
 
   GLStats m_stats = {};