GPU: Implement CLUT cache (SW renderer only)

Fixes copyright screen in Shadow Master.
Fixes title splash in Castrol Honda Superbike Racing.

src/core/gpu.cpp

@@ -36,6 +36,7 @@ Log_SetChannel(GPU);
 
 std::unique_ptr<GPU> g_gpu;
 alignas(HOST_PAGE_SIZE) u16 g_vram[VRAM_SIZE / sizeof(u16)];
+u16 g_gpu_clut[GPU_CLUT_SIZE];
 
 const GPU::GP0CommandHandlerTable GPU::s_GP0_command_handler_table = GPU::GenerateGP0CommandHandlerTable();
 
@@ -173,7 +174,10 @@ void GPU::Reset(bool clear_vram)
   m_crtc_state.interlaced_display_field = 0;
 
   if (clear_vram)
+  {
     std::memset(g_vram, 0, sizeof(g_vram));
+    std::memset(g_gpu_clut, 0, sizeof(g_gpu_clut));
+  }
 
   // Force event to reschedule itself.
   m_crtc_tick_event->Deactivate();
@@ -224,6 +228,7 @@ void GPU::SoftReset()
   SetDrawMode(0);
   SetTexturePalette(0);
   SetTextureWindow(0);
+  InvalidateCLUT();
   UpdateDMARequest();
   UpdateCRTCConfig();
   UpdateCommandTickEvent();
@@ -308,6 +313,18 @@ bool GPU::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_displ
   sw.Do(&m_command_total_words);
   sw.Do(&m_GPUREAD_latch);
 
+  if (sw.GetVersion() < 64) [[unlikely]]
+  {
+    // Clear CLUT cache and let it populate later.
+    InvalidateCLUT();
+  }
+  else
+  {
+    sw.Do(&m_current_clut_reg_bits);
+    sw.Do(&m_current_clut_is_8bit);
+    sw.DoArray(g_gpu_clut, std::size(g_gpu_clut));
+  }
+
   sw.Do(&m_vram_transfer.x);
   sw.Do(&m_vram_transfer.y);
   sw.Do(&m_vram_transfer.width);
@@ -1443,6 +1460,27 @@ void GPU::HandleGetGPUInfoCommand(u32 value)
   }
 }
 
+void GPU::UpdateCLUTIfNeeded(GPUTextureMode texmode, GPUTexturePaletteReg clut)
+{
+  if (texmode >= GPUTextureMode::Direct16Bit)
+    return;
+
+  const bool needs_8bit = (texmode == GPUTextureMode::Palette8Bit);
+  if ((clut.bits != m_current_clut_reg_bits) || BoolToUInt8(needs_8bit) > BoolToUInt8(m_current_clut_is_8bit))
+  {
+    Log_DebugFmt("Reloading CLUT from {},{}, {}", clut.GetXBase(), clut.GetYBase(), needs_8bit ? "8-bit" : "4-bit");
+    UpdateCLUT(clut, needs_8bit);
+    m_current_clut_reg_bits = clut.bits;
+    m_current_clut_is_8bit = needs_8bit;
+  }
+}
+
+void GPU::InvalidateCLUT()
+{
+  m_current_clut_reg_bits = std::numeric_limits<decltype(m_current_clut_reg_bits)>::max(); // will never match
+  m_current_clut_is_8bit = false;
+}
+
 void GPU::ClearDisplay()
 {
   ClearDisplayTexture();
@@ -1451,10 +1489,6 @@ void GPU::ClearDisplay()
   DestroyDeinterlaceTextures();
 }
 
-void GPU::UpdateDisplay()
-{
-}
-
 void GPU::ReadVRAM(u32 x, u32 y, u32 width, u32 height)
 {
 }
@@ -1615,14 +1649,6 @@ void GPU::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 he
   }
 }
 
-void GPU::DispatchRenderCommand()
-{
-}
-
-void GPU::FlushRender()
-{
-}
-
 void GPU::SetDrawMode(u16 value)
 {
   GPUDrawModeReg new_mode_reg{static_cast<u16>(value & GPUDrawModeReg::MASK)};
@@ -1677,6 +1703,31 @@ void GPU::SetTextureWindow(u32 value)
   m_draw_mode.texture_window_changed = true;
 }
 
+void GPU::ReadCLUT(u16* dest, GPUTexturePaletteReg reg, bool clut_is_8bit)
+{
+  const u16* src_row = &g_vram[reg.GetYBase() * VRAM_WIDTH];
+  const u32 start_x = reg.GetXBase();
+  if (!clut_is_8bit)
+  {
+    // Wraparound can't happen in 4-bit mode.
+    std::memcpy(dest, &src_row[start_x], sizeof(u16) * 16);
+  }
+  else
+  {
+    if ((start_x + 256) > VRAM_WIDTH) [[unlikely]]
+    {
+      const u32 end = VRAM_WIDTH - start_x;
+      const u32 start = 256 - end;
+      std::memcpy(dest, &src_row[start_x], sizeof(u16) * end);
+      std::memcpy(dest + end, src_row, sizeof(u16) * start);
+    }
+    else
+    {
+      std::memcpy(dest, &src_row[start_x], sizeof(u16) * 256);
+    }
+  }
+}
+
 bool GPU::CompileDisplayPipelines(bool display, bool deinterlace, bool chroma_smoothing)
 {
   GPUShaderGen shadergen(g_gpu_device->GetRenderAPI(), g_gpu_device->GetFeatures().dual_source_blend,

src/core/gpu.h

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: 2019-2022 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
 // SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
 
 #pragma once
@@ -200,7 +200,7 @@ public:
   bool DumpVRAMToFile(const char* filename);
 
   // Ensures all buffered vertices are drawn.
-  virtual void FlushRender();
+  virtual void FlushRender() = 0;
 
   ALWAYS_INLINE const void* GetDisplayTextureHandle() const { return m_display_texture; }
   ALWAYS_INLINE s32 GetDisplayWidth() const { return m_display_width; }
@@ -225,6 +225,9 @@ public:
   /// Draws the current display texture, with any post-processing.
   bool PresentDisplay();
 
+  /// Reads the CLUT from the specified coordinates, accounting for wrap-around.
+  static void ReadCLUT(u16* dest, GPUTexturePaletteReg reg, bool clut_is_8bit);
+
 protected:
   TickCount CRTCTicksToSystemTicks(TickCount crtc_ticks, TickCount fractional_ticks) const;
   TickCount SystemTicksToCRTCTicks(TickCount sysclk_ticks, TickCount* fractional_ticks) const;
@@ -306,14 +309,17 @@ protected:
   void ExecuteCommands();
   void TryExecuteCommands();
   void HandleGetGPUInfoCommand(u32 value);
+  void UpdateCLUTIfNeeded(GPUTextureMode texmode, GPUTexturePaletteReg clut);
+  void InvalidateCLUT();
 
   // Rendering in the backend
   virtual void ReadVRAM(u32 x, u32 y, u32 width, u32 height);
   virtual void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color);
   virtual void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask);
   virtual void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height);
-  virtual void DispatchRenderCommand();
+  virtual void DispatchRenderCommand() = 0;
-  virtual void UpdateDisplay();
+  virtual void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) = 0;
+  virtual void UpdateDisplay() = 0;
   virtual void DrawRendererStats();
 
   ALWAYS_INLINE_RELEASE void AddDrawTriangleTicks(s32 x1, s32 y1, s32 x2, s32 y2, s32 x3, s32 y3, bool shaded,
@@ -568,6 +574,11 @@ protected:
   /// GPUREAD value for non-VRAM-reads.
   u32 m_GPUREAD_latch = 0;
 
+  // These are the bits from the palette register, but zero extended to 32-bit, so we can have an "invalid" value.
+  // If an extra byte is ever not needed here for padding, the 8-bit flag could be packed into the MSB of this value.
+  u32 m_current_clut_reg_bits = {};
+  bool m_current_clut_is_8bit = false;
+
   /// True if currently executing/syncing.
   bool m_executing_commands = false;
 
@@ -693,3 +704,4 @@ private:
 
 extern std::unique_ptr<GPU> g_gpu;
 extern u16 g_vram[VRAM_SIZE / sizeof(u16)];
+extern u16 g_gpu_clut[GPU_CLUT_SIZE];

src/core/gpu_backend.cpp

@@ -73,6 +73,12 @@ GPUBackendSetDrawingAreaCommand* GPUBackend::NewSetDrawingAreaCommand()
     AllocateCommand(GPUBackendCommandType::SetDrawingArea, sizeof(GPUBackendSetDrawingAreaCommand)));
 }
 
+GPUBackendUpdateCLUTCommand* GPUBackend::NewUpdateCLUTCommand()
+{
+  return static_cast<GPUBackendUpdateCLUTCommand*>(
+    AllocateCommand(GPUBackendCommandType::UpdateCLUT, sizeof(GPUBackendUpdateCLUTCommand)));
+}
+
 GPUBackendDrawPolygonCommand* GPUBackend::NewDrawPolygonCommand(u32 num_vertices)
 {
   const u32 size = sizeof(GPUBackendDrawPolygonCommand) + (num_vertices * sizeof(GPUBackendDrawPolygonCommand::Vertex));
@@ -309,6 +315,13 @@ void GPUBackend::HandleCommand(const GPUBackendCommand* cmd)
     }
     break;
 
+    case GPUBackendCommandType::UpdateCLUT:
+    {
+      const GPUBackendUpdateCLUTCommand* ccmd = static_cast<const GPUBackendUpdateCLUTCommand*>(cmd);
+      UpdateCLUT(ccmd->reg, ccmd->clut_is_8bit);
+    }
+    break;
+
     case GPUBackendCommandType::DrawPolygon:
     {
       DrawPolygon(static_cast<const GPUBackendDrawPolygonCommand*>(cmd));
@@ -328,6 +341,6 @@ void GPUBackend::HandleCommand(const GPUBackendCommand* cmd)
     break;
 
     default:
-      break;
+      UnreachableCode();
   }
 }

src/core/gpu_backend.h

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: 2019-2022 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
 // SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
 
 #pragma once
@@ -33,6 +33,7 @@ public:
   GPUBackendUpdateVRAMCommand* NewUpdateVRAMCommand(u32 num_words);
   GPUBackendCopyVRAMCommand* NewCopyVRAMCommand();
   GPUBackendSetDrawingAreaCommand* NewSetDrawingAreaCommand();
+  GPUBackendUpdateCLUTCommand* NewUpdateCLUTCommand();
   GPUBackendDrawPolygonCommand* NewDrawPolygonCommand(u32 num_vertices);
   GPUBackendDrawRectangleCommand* NewDrawRectangleCommand();
   GPUBackendDrawLineCommand* NewDrawLineCommand(u32 num_vertices);
@@ -60,6 +61,7 @@ protected:
   virtual void DrawLine(const GPUBackendDrawLineCommand* cmd) = 0;
   virtual void FlushRender() = 0;
   virtual void DrawingAreaChanged() = 0;
+  virtual void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) = 0;
 
   void HandleCommand(const GPUBackendCommand* cmd);
 

src/core/gpu_commands.cpp

@@ -198,6 +198,7 @@ bool GPU::HandleClearCacheCommand()
 {
   Log_DebugPrintf("GP0 clear cache");
   m_draw_mode.SetTexturePageChanged();
+  InvalidateCLUT();
   m_fifo.RemoveOne();
   AddCommandTicks(1);
   EndCommand();
@@ -347,6 +348,7 @@ bool GPU::HandleRenderPolygonCommand()
     SetDrawMode((texpage_attribute & GPUDrawModeReg::POLYGON_TEXPAGE_MASK) |
                 (m_draw_mode.mode_reg.bits & ~GPUDrawModeReg::POLYGON_TEXPAGE_MASK));
     SetTexturePalette(Truncate16(FifoPeek(2) >> 16));
+    UpdateCLUTIfNeeded(m_draw_mode.mode_reg.texture_mode, m_draw_mode.palette_reg);
   }
 
   m_counters.num_vertices += num_vertices;
@@ -371,7 +373,10 @@ bool GPU::HandleRenderRectangleCommand()
     SynchronizeCRTC();
 
   if (rc.texture_enable)
+  {
     SetTexturePalette(Truncate16(FifoPeek(2) >> 16));
+    UpdateCLUTIfNeeded(m_draw_mode.mode_reg.texture_mode, m_draw_mode.palette_reg);
+  }
 
   const TickCount setup_ticks = 16;
   AddCommandTicks(setup_ticks);

src/core/gpu_hw.cpp

@@ -3110,6 +3110,13 @@ void GPU_HW::DispatchRenderCommand()
   LoadVertices();
 }
 
+void GPU_HW::UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit)
+{
+  // Not done in HW
+  GL_INS_FMT("Reloading CLUT from {},{}, {} not implemented", reg.GetXBase(), reg.GetYBase(),
+             clut_is_8bit ? "8-bit" : "4-bit");
+}
+
 void GPU_HW::FlushRender()
 {
   const u32 base_vertex = m_batch_base_vertex;

src/core/gpu_hw.h

@@ -47,8 +47,8 @@ public:
 
   void UpdateSettings(const Settings& old_settings) override;
   void UpdateResolutionScale() override final;
-  std::tuple<u32, u32> GetEffectiveDisplayResolution(bool scaled = true) override final;
+  std::tuple<u32, u32> GetEffectiveDisplayResolution(bool scaled = true) override;
-  std::tuple<u32, u32> GetFullDisplayResolution(bool scaled = true) override final;
+  std::tuple<u32, u32> GetFullDisplayResolution(bool scaled = true) override;
 
   void UpdateDisplay() override;
 
@@ -176,6 +176,7 @@ private:
   void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask) override;
   void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height) override;
   void DispatchRenderCommand() override;
+  void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) override;
   void FlushRender() override;
   void DrawRendererStats() override;
 

src/core/gpu_sw.cpp

@@ -824,6 +824,19 @@ void GPU_SW::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32
   m_backend.PushCommand(cmd);
 }
 
+void GPU_SW::FlushRender()
+{
+}
+
+void GPU_SW::UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit)
+{
+  GPUBackendUpdateCLUTCommand* cmd = m_backend.NewUpdateCLUTCommand();
+  FillBackendCommandParameters(cmd);
+  cmd->reg.bits = reg.bits;
+  cmd->clut_is_8bit = clut_is_8bit;
+  m_backend.PushCommand(cmd);
+}
+
 std::unique_ptr<GPU> GPU::CreateSoftwareRenderer()
 {
   std::unique_ptr<GPU_SW> gpu(std::make_unique<GPU_SW>());

src/core/gpu_sw.h

@@ -40,6 +40,8 @@ protected:
   void FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color) override;
   void UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask) override;
   void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height) override;
+  void FlushRender() override;
+  void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) override;
 
   template<GPUTexture::Format display_format>
   bool CopyOut15Bit(u32 src_x, u32 src_y, u32 width, u32 height, u32 line_skip);

src/core/gpu_sw_backend.cpp

@@ -1,8 +1,8 @@
-// SPDX-FileCopyrightText: 2019-2022 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
 // SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
 
-#include "gpu.h"
 #include "gpu_sw_backend.h"
+#include "gpu.h"
 #include "system.h"
 
 #include "util/gpu_device.h"
@@ -93,10 +93,8 @@ void ALWAYS_INLINE_RELEASE GPU_SW_Backend::ShadePixel(const GPUBackendDrawComman
         const u16 palette_value =
           GetPixel((cmd->draw_mode.GetTexturePageBaseX() + ZeroExtend32(texcoord_x / 4)) % VRAM_WIDTH,
                    (cmd->draw_mode.GetTexturePageBaseY() + ZeroExtend32(texcoord_y)) % VRAM_HEIGHT);
-        const u16 palette_index = (palette_value >> ((texcoord_x % 4) * 4)) & 0x0Fu;
+        const size_t palette_index = (palette_value >> ((texcoord_x % 4) * 4)) & 0x0Fu;
-
+        texture_color.bits = g_gpu_clut[palette_index];
-        texture_color.bits =
-          GetPixel((cmd->palette.GetXBase() + ZeroExtend32(palette_index)) % VRAM_WIDTH, cmd->palette.GetYBase());
       }
       break;
 
@@ -105,9 +103,8 @@ void ALWAYS_INLINE_RELEASE GPU_SW_Backend::ShadePixel(const GPUBackendDrawComman
         const u16 palette_value =
           GetPixel((cmd->draw_mode.GetTexturePageBaseX() + ZeroExtend32(texcoord_x / 2)) % VRAM_WIDTH,
                    (cmd->draw_mode.GetTexturePageBaseY() + ZeroExtend32(texcoord_y)) % VRAM_HEIGHT);
-        const u16 palette_index = (palette_value >> ((texcoord_x % 2) * 8)) & 0xFFu;
+        const size_t palette_index = (palette_value >> ((texcoord_x % 2) * 8)) & 0xFFu;
-        texture_color.bits =
+        texture_color.bits = g_gpu_clut[palette_index];
-          GetPixel((cmd->palette.GetXBase() + ZeroExtend32(palette_index)) % VRAM_WIDTH, cmd->palette.GetYBase());
       }
       break;
 
@@ -869,9 +866,18 @@ void GPU_SW_Backend::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 wi
   }
 }
 
-void GPU_SW_Backend::FlushRender() {}
+void GPU_SW_Backend::FlushRender()
+{
+}
 
-void GPU_SW_Backend::DrawingAreaChanged() {}
+void GPU_SW_Backend::DrawingAreaChanged()
+{
+}
+
+void GPU_SW_Backend::UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit)
+{
+  GPU::ReadCLUT(g_gpu_clut, reg, clut_is_8bit);
+}
 
 GPU_SW_Backend::DrawLineFunction GPU_SW_Backend::GetDrawLineFunction(bool shading_enable, bool transparency_enable,
                                                                      bool dithering_enable)

src/core/gpu_sw_backend.h

@@ -1,8 +1,11 @@
-// SPDX-FileCopyrightText: 2019-2022 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
 // SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
 
 #pragma once
+
+#include "gpu.h"
 #include "gpu_backend.h"
+
 #include <array>
 #include <memory>
 #include <vector>
@@ -97,6 +100,7 @@ protected:
   void DrawRectangle(const GPUBackendDrawRectangleCommand* cmd) override;
   void FlushRender() override;
   void DrawingAreaChanged() override;
+  void UpdateCLUT(GPUTexturePaletteReg reg, bool clut_is_8bit) override;
 
   //////////////////////////////////////////////////////////////////////////
   // Rasterization

src/core/gpu_types.h

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: 2019-2023 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
 // SPDX-License-Identifier: (GPL-3.0 OR CC-BY-NC-ND-4.0)
 
 #pragma once
@@ -17,6 +17,7 @@ enum : u32
   VRAM_HEIGHT_MASK = VRAM_HEIGHT - 1,
   TEXTURE_PAGE_WIDTH = 256,
   TEXTURE_PAGE_HEIGHT = 256,
+  GPU_CLUT_SIZE = 256,
 
   // In interlaced modes, we can exceed the 512 height of VRAM, up to 576 in PAL games.
   GPU_MAX_DISPLAY_WIDTH = 720,
@@ -259,9 +260,10 @@ enum class GPUBackendCommandType : u8
   UpdateVRAM,
   CopyVRAM,
   SetDrawingArea,
+  UpdateCLUT,
   DrawPolygon,
   DrawRectangle,
-  DrawLine
+  DrawLine,
 };
 
 union GPUBackendCommandParameters
@@ -336,6 +338,12 @@ struct GPUBackendSetDrawingAreaCommand : public GPUBackendCommand
   GPUDrawingArea new_area;
 };
 
+struct GPUBackendUpdateCLUTCommand : public GPUBackendCommand
+{
+  GPUTexturePaletteReg reg;
+  bool clut_is_8bit;
+};
+
 struct GPUBackendDrawCommand : public GPUBackendCommand
 {
   GPUDrawModeReg draw_mode;

src/core/save_state_version.h

@@ -5,7 +5,7 @@
 #include "types.h"
 
 static constexpr u32 SAVE_STATE_MAGIC = 0x43435544;
-static constexpr u32 SAVE_STATE_VERSION = 63;
+static constexpr u32 SAVE_STATE_VERSION = 64;
 static constexpr u32 SAVE_STATE_MINIMUM_VERSION = 42;
 
 static_assert(SAVE_STATE_VERSION >= SAVE_STATE_MINIMUM_VERSION);