GPU: Support dumping copies out to file

src/pse/gpu.cpp

@@ -3,9 +3,15 @@
 #include "common/state_wrapper.h"
 #include "dma.h"
 #include "interrupt_controller.h"
+#include "stb_image_write.h"
 #include "system.h"
 Log_SetChannel(GPU);
 
+bool GPU::DUMP_CPU_TO_VRAM_COPIES = false;
+bool GPU::DUMP_VRAM_TO_CPU_COPIES = false;
+static u32 s_cpu_to_vram_dump_id = 1;
+static u32 s_vram_to_cpu_dump_id = 1;
+
 GPU::GPU() = default;
 
 GPU::~GPU() = default;
@@ -644,6 +650,12 @@ bool GPU::HandleCopyRectangleCPUToVRAMCommand()
     return true;
   }
 
+  if (DUMP_CPU_TO_VRAM_COPIES)
+  {
+    DumpVRAMToFile(SmallString::FromFormat("cpu_to_vram_copy_%u.png", s_cpu_to_vram_dump_id++), copy_width, copy_height,
+                   sizeof(u16) * copy_width, &m_GP0_command[3], true);
+  }
+
   FlushRender();
   UpdateVRAM(dst_x, dst_y, copy_width, copy_height, &m_GP0_command[3]);
   return true;
@@ -678,6 +690,12 @@ bool GPU::HandleCopyRectangleVRAMToCPUCommand()
   for (const u32 bits : temp)
     m_GPUREAD_buffer.push_back(bits);
 
+  if (DUMP_VRAM_TO_CPU_COPIES)
+  {
+    DumpVRAMToFile(SmallString::FromFormat("vram_to_cpu_copy_%u.png", s_cpu_to_vram_dump_id++), width, height,
+                   sizeof(u16) * width, temp.data(), true);
+  }
+
   // Is this correct?
   return true;
 }
@@ -774,3 +792,26 @@ void GPU::TextureConfig::SetFromPaletteAttribute(u16 value)
   palette_attribute = value;
   page_changed = true;
 }
+
+bool GPU::DumpVRAMToFile(const char* filename, u32 width, u32 height, u32 stride, const void* buffer, bool remove_alpha)
+{
+  std::vector<u32> rgba8_buf(width * height);
+
+  const char* ptr_in = static_cast<const char*>(buffer);
+  u32* ptr_out = rgba8_buf.data();
+  for (u32 row = 0; row < height; row++)
+  {
+    const char* row_ptr_in = ptr_in;
+
+    for (u32 col = 0; col < width; col++)
+    {
+      u16 src_col;
+      std::memcpy(&src_col, row_ptr_in, sizeof(u16));
+      row_ptr_in += sizeof(u16);
+      *(ptr_out++) = RGBA5551ToRGBA8888(remove_alpha ? (src_col | u16(0x8000)) : src_col);
+    }
+
+    ptr_in += stride;
+  }
+  return (stbi_write_png(filename, width, height, 4, rgba8_buf.data(), sizeof(u32) * width) != 0);
+}

src/pse/gpu.h

@@ -48,6 +48,36 @@ protected:
       return value;
   }
 
+  // Helper/format conversion functions.
+  static constexpr u32 RGBA5551ToRGBA8888(u16 color)
+  {
+    u8 r = Truncate8(color & 31);
+    u8 g = Truncate8((color >> 5) & 31);
+    u8 b = Truncate8((color >> 10) & 31);
+    u8 a = Truncate8((color >> 15) & 1);
+
+    // 00012345 -> 1234545
+    b = (b << 3) | (b & 0b111);
+    g = (g << 3) | (g & 0b111);
+    r = (r << 3) | (r & 0b111);
+    a = a ? 255 : 0;
+
+    return ZeroExtend32(r) | (ZeroExtend32(g) << 8) | (ZeroExtend32(b) << 16) | (ZeroExtend32(a) << 24);
+  }
+
+  static constexpr u16 RGBA8888ToRGBA5551(u32 color)
+  {
+    const u16 r = Truncate16((color >> 3) & 0x1Fu);
+    const u16 g = Truncate16((color >> 11) & 0x1Fu);
+    const u16 b = Truncate16((color >> 19) & 0x1Fu);
+    const u16 a = Truncate16((color >> 31) & 0x01u);
+
+    return r | (g << 5) | (b << 10) | (a << 15);
+  }
+
+  static bool DumpVRAMToFile(const char* filename, u32 width, u32 height, u32 stride, const void* buffer,
+                             bool remove_alpha);
+
   enum class DMADirection : u32
   {
     Off = 0,
@@ -275,4 +305,8 @@ protected:
 
   std::vector<u32> m_GP0_command;
   std::deque<u32> m_GPUREAD_buffer;
+
+  // debug options
+  static bool DUMP_CPU_TO_VRAM_COPIES;
+  static bool DUMP_VRAM_TO_CPU_COPIES;
 };

src/pse/gpu_hw.cpp

@@ -191,8 +191,8 @@ out vec4 v_col0;
 void main()
 {
   // 0..+1023 -> -1..1
-  float pos_x = (float(a_pos.x) / 511.5) - 1.0;
+  float pos_x = (float(a_pos.x) / 512.0) - 1.0;
-  float pos_y = (float(a_pos.y) / -255.5) + 1.0;
+  float pos_y = (float(a_pos.y) / -256.0) + 1.0;
   gl_Position = vec4(pos_x, pos_y, 0.0, 1.0);
 
   v_col0 = a_col0;

src/pse/gpu_hw_opengl.cpp

@@ -191,32 +191,6 @@ void GPU_HW_OpenGL::SetScissor()
   glScissor(x, y, width, height);
 }
 
-inline u32 ConvertRGBA5551ToRGBA8888(u16 color)
-{
-  u8 r = Truncate8(color & 31);
-  u8 g = Truncate8((color >> 5) & 31);
-  u8 b = Truncate8((color >> 10) & 31);
-  u8 a = Truncate8((color >> 15) & 1);
-
-  // 00012345 -> 1234545
-  b = (b << 3) | (b & 0b111);
-  g = (g << 3) | (g & 0b111);
-  r = (r << 3) | (r & 0b111);
-  a = a ? 255 : 0;
-
-  return ZeroExtend32(r) | (ZeroExtend32(g) << 8) | (ZeroExtend32(b) << 16) | (ZeroExtend32(a) << 24);
-}
-
-inline u16 ConvertRGBA8888ToRGBA5551(u32 color)
-{
-  const u16 r = Truncate16((color >> 3) & 0x1Fu);
-  const u16 g = Truncate16((color >> 11) & 0x1Fu);
-  const u16 b = Truncate16((color >> 19) & 0x1Fu);
-  const u16 a = Truncate16((color >> 31) & 0x01u);
-
-  return r | (g << 5) | (b << 10) | (a << 15);
-}
-
 void GPU_HW_OpenGL::UpdateDisplay()
 {
   GPU_HW::UpdateDisplay();
@@ -246,7 +220,7 @@ void GPU_HW_OpenGL::ReadVRAM(u32 x, u32 y, u32 width, u32 height, void* buffer)
       std::memcpy(&src_col, source_row_ptr, sizeof(src_col));
       source_row_ptr += sizeof(src_col);
 
-      const u16 dst_col = ConvertRGBA8888ToRGBA5551(src_col);
+      const u16 dst_col = RGBA8888ToRGBA5551(src_col);
       std::memcpy(dst_row_ptr, &dst_col, sizeof(dst_col));
       dst_row_ptr += sizeof(dst_col);
     }
@@ -286,7 +260,7 @@ void GPU_HW_OpenGL::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void*
       std::memcpy(&src_col, source_row_ptr, sizeof(src_col));
       source_row_ptr += sizeof(src_col);
 
-      const u32 dst_col = ConvertRGBA5551ToRGBA8888(src_col);
+      const u32 dst_col = RGBA5551ToRGBA8888(src_col);
       rgba_data.push_back(dst_col);
     }
 

src/pse/pse.vcxproj

@@ -70,6 +70,9 @@
     <ProjectReference Include="..\..\dep\libsamplerate\libsamplerate.vcxproj">
       <Project>{2f2a2b7b-60b3-478c-921e-3633b3c45c3f}</Project>
     </ProjectReference>
+    <ProjectReference Include="..\..\dep\stb\stb.vcxproj">
+      <Project>{ed601289-ac1a-46b8-a8ed-17db9eb73423}</Project>
+    </ProjectReference>
     <ProjectReference Include="..\..\dep\YBaseLib\Source\YBaseLib.vcxproj">
       <Project>{b56ce698-7300-4fa5-9609-942f1d05c5a2}</Project>
     </ProjectReference>
@@ -224,7 +227,7 @@
       <PreprocessorDefinitions>ENABLE_VOODOO=1;_CRT_SECURE_NO_WARNINGS;WIN32;_DEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <SDLCheck>true</SDLCheck>
       <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
-      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)dep\stb\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <MinimalRebuild>false</MinimalRebuild>
       <LanguageStandard>stdcpp17</LanguageStandard>
@@ -245,7 +248,7 @@
       <PreprocessorDefinitions>ENABLE_VOODOO=1;_CRT_SECURE_NO_WARNINGS;WIN32;_DEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <SDLCheck>true</SDLCheck>
       <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
-      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)dep\stb\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <MinimalRebuild>false</MinimalRebuild>
       <LanguageStandard>stdcpp17</LanguageStandard>
@@ -266,7 +269,7 @@
       <PreprocessorDefinitions>ENABLE_VOODOO=1;_ITERATOR_DEBUG_LEVEL=1;_CRT_SECURE_NO_WARNINGS;WIN32;_DEBUGFAST;_DEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <SDLCheck>true</SDLCheck>
       <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
-      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)dep\stb\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <BasicRuntimeChecks>Default</BasicRuntimeChecks>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <MinimalRebuild>false</MinimalRebuild>
@@ -289,7 +292,7 @@
       <PreprocessorDefinitions>ENABLE_VOODOO=1;_ITERATOR_DEBUG_LEVEL=1;_CRT_SECURE_NO_WARNINGS;WIN32;_DEBUGFAST;_DEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <SDLCheck>true</SDLCheck>
       <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
-      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)dep\stb\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <BasicRuntimeChecks>Default</BasicRuntimeChecks>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <MinimalRebuild>false</MinimalRebuild>
@@ -311,7 +314,7 @@
       <Optimization>MaxSpeed</Optimization>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>ENABLE_VOODOO=1;_CRT_SECURE_NO_WARNINGS;WIN32;NDEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
-      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)dep\stb\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <WholeProgramOptimization>false</WholeProgramOptimization>
       <LanguageStandard>stdcpp17</LanguageStandard>
@@ -333,7 +336,7 @@
       <Optimization>MaxSpeed</Optimization>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>ENABLE_VOODOO=1;_CRT_SECURE_NO_WARNINGS;WIN32;NDEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
-      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)dep\stb\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <WholeProgramOptimization>true</WholeProgramOptimization>
       <LanguageStandard>stdcpp17</LanguageStandard>
@@ -356,7 +359,7 @@
       <Optimization>MaxSpeed</Optimization>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>ENABLE_VOODOO=1;_CRT_SECURE_NO_WARNINGS;WIN32;NDEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
-      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)dep\stb\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <WholeProgramOptimization>false</WholeProgramOptimization>
       <LanguageStandard>stdcpp17</LanguageStandard>
@@ -378,7 +381,7 @@
       <Optimization>MaxSpeed</Optimization>
       <IntrinsicFunctions>true</IntrinsicFunctions>
       <PreprocessorDefinitions>ENABLE_VOODOO=1;_CRT_SECURE_NO_WARNINGS;WIN32;NDEBUG;_CONSOLE;_LIB;%(PreprocessorDefinitions)</PreprocessorDefinitions>
-      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
+      <AdditionalIncludeDirectories>$(SolutionDir)dep\msvc\include;$(SolutionDir)dep\YBaseLib\Include;$(SolutionDir)dep\glad\include;$(SolutionDir)dep\stb\include;$(SolutionDir)src;%(AdditionalIncludeDirectories)</AdditionalIncludeDirectories>
       <MultiProcessorCompilation>true</MultiProcessorCompilation>
       <WholeProgramOptimization>true</WholeProgramOptimization>
       <LanguageStandard>stdcpp17</LanguageStandard>