GPU: Add hardware texture cache

src/common/gsvector_neon.h

@@ -2998,6 +2998,10 @@ public:
     return GSVector4i(vsetq_lane_s32(high, vsetq_lane_s32(low, vdupq_n_s32(0), 0), 1));
   }
 
+  ALWAYS_INLINE GSVector2 xy() const { return GSVector2(vget_low_s32(v4s)); }
+
+  ALWAYS_INLINE GSVector2 zw() const { return GSVector2(vget_high_s32(v4s)); }
+
 #define VECTOR4_SHUFFLE_4(xs, xn, ys, yn, zs, zn, ws, wn)                                                              \
   ALWAYS_INLINE GSVector4 xs##ys##zs##ws() const                                                                       \
   {                                                                                                                    \

src/common/gsvector_nosimd.h

@@ -2289,6 +2289,10 @@ public:
     return GSVector4i(static_cast<s32>(F64[0]), static_cast<s32>(F64[1]), 0, 0);
   }
 
+  ALWAYS_INLINE GSVector2 xy() const { return GSVector2(x, y); }
+
+  ALWAYS_INLINE GSVector2 zw() const { return GSVector2(z, w); }
+
 #define VECTOR4_SHUFFLE_4(xs, xn, ys, yn, zs, zn, ws, wn)                                                              \
   ALWAYS_INLINE GSVector4 xs##ys##zs##ws() const { return GSVector4(F32[xn], F32[yn], F32[zn], F32[wn]); }
 

src/common/gsvector_sse.h

@@ -2365,6 +2365,10 @@ public:
 
   ALWAYS_INLINE GSVector4i f64toi32() const { return GSVector4i(_mm_cvttpd_epi32(_mm_castps_pd(m))); }
 
+  ALWAYS_INLINE GSVector2 xy() const { return GSVector2(m); }
+
+  ALWAYS_INLINE GSVector2 zw() const { return GSVector2(_mm_shuffle_ps(m, m, _MM_SHUFFLE(3, 2, 3, 2))); }
+
 #define VECTOR4_SHUFFLE_4(xs, xn, ys, yn, zs, zn, ws, wn)                                                              \
   ALWAYS_INLINE GSVector4 xs##ys##zs##ws() const                                                                       \
   {                                                                                                                    \

src/core/CMakeLists.txt

@@ -51,6 +51,8 @@ add_library(core
   gpu_hw.h
   gpu_hw_shadergen.cpp
   gpu_hw_shadergen.h
+  gpu_hw_texture_cache.cpp
+  gpu_hw_texture_cache.h
   gpu_shadergen.cpp
   gpu_shadergen.h
   gpu_sw.cpp
@@ -108,8 +110,6 @@ add_library(core
   spu.h
   system.cpp
   system.h
-  texture_replacements.cpp
-  texture_replacements.h
   timers.cpp
   timers.h
   timing_event.cpp

src/core/core.vcxproj

@@ -47,6 +47,7 @@
     <ClCompile Include="gpu_backend.cpp" />
     <ClCompile Include="gpu_commands.cpp" />
     <ClCompile Include="gpu_hw_shadergen.cpp" />
+    <ClCompile Include="gpu_hw_texture_cache.cpp" />
     <ClCompile Include="gpu_shadergen.cpp" />
     <ClCompile Include="gpu_sw.cpp" />
     <ClCompile Include="gpu_sw_backend.cpp" />
@@ -81,7 +82,6 @@
     <ClCompile Include="sio.cpp" />
     <ClCompile Include="spu.cpp" />
     <ClCompile Include="system.cpp" />
-    <ClCompile Include="texture_replacements.cpp" />
     <ClCompile Include="timers.cpp" />
     <ClCompile Include="timing_event.cpp" />
   </ItemGroup>
@@ -124,6 +124,7 @@
     <ClInclude Include="gdb_server.h" />
     <ClInclude Include="gpu_backend.h" />
     <ClInclude Include="gpu_hw_shadergen.h" />
+    <ClInclude Include="gpu_hw_texture_cache.h" />
     <ClInclude Include="gpu_shadergen.h" />
     <ClInclude Include="gpu_sw.h" />
     <ClInclude Include="gpu_sw_backend.h" />
@@ -161,7 +162,6 @@
     <ClInclude Include="sio.h" />
     <ClInclude Include="spu.h" />
     <ClInclude Include="system.h" />
-    <ClInclude Include="texture_replacements.h" />
     <ClInclude Include="timers.h" />
     <ClInclude Include="timing_event.h" />
     <ClInclude Include="types.h" />

src/core/core.vcxproj.filters

@@ -46,7 +46,6 @@
     <ClCompile Include="cpu_recompiler_code_generator_aarch32.cpp" />
     <ClCompile Include="gpu_backend.cpp" />
     <ClCompile Include="gpu_sw_backend.cpp" />
-    <ClCompile Include="texture_replacements.cpp" />
     <ClCompile Include="multitap.cpp" />
     <ClCompile Include="host.cpp" />
     <ClCompile Include="game_database.cpp" />
@@ -67,6 +66,7 @@
     <ClCompile Include="justifier.cpp" />
     <ClCompile Include="gdb_server.cpp" />
     <ClCompile Include="gpu_sw_rasterizer.cpp" />
+    <ClCompile Include="gpu_hw_texture_cache.cpp" />
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="types.h" />
@@ -116,7 +116,6 @@
     <ClInclude Include="gpu_types.h" />
     <ClInclude Include="gpu_backend.h" />
     <ClInclude Include="gpu_sw_backend.h" />
-    <ClInclude Include="texture_replacements.h" />
     <ClInclude Include="multitap.h" />
     <ClInclude Include="host.h" />
     <ClInclude Include="achievements.h" />
@@ -140,6 +139,7 @@
     <ClInclude Include="justifier.h" />
     <ClInclude Include="gdb_server.h" />
     <ClInclude Include="gpu_sw_rasterizer.h" />
+    <ClInclude Include="gpu_hw_texture_cache.h" />
   </ItemGroup>
   <ItemGroup>
     <None Include="gpu_sw_rasterizer.inl" />

src/core/fullscreen_ui.cpp

@@ -4502,23 +4502,45 @@ void FullscreenUI::DrawDisplaySettingsPage()
 
   MenuHeading(FSUI_CSTR("Texture Replacements"));
 
-  DrawToggleSetting(bsi, FSUI_CSTR("Enable VRAM Write Texture Replacement"),
+  ActiveButton(FSUI_CSTR("The texture cache is currently experimental, and may cause rendering errors in some games."),
-                    FSUI_CSTR("Enables the replacement of background textures in supported games."),
+               false, false, ImGuiFullscreen::LAYOUT_MENU_BUTTON_HEIGHT_NO_SUMMARY, g_large_font);
-                    "TextureReplacements", "EnableVRAMWriteReplacements", false);
+
-  DrawToggleSetting(bsi, FSUI_CSTR("Preload Replacement Textures"),
+  DrawToggleSetting(bsi, FSUI_CSTR("Enable Texture Cache"),
-                    FSUI_CSTR("Loads all replacement texture to RAM, reducing stuttering at runtime."),
+                    FSUI_CSTR("Enables caching of guest textures, required for texture replacement."), "GPU",
-                    "TextureReplacements", "PreloadTextures", false);
+                    "EnableTextureCache", false);
   DrawToggleSetting(bsi, FSUI_CSTR("Use Old MDEC Routines"),
                     FSUI_CSTR("Enables the older, less accurate MDEC decoding routines. May be required for old "
                               "replacement backgrounds to match/load."),
                     "Hacks", "UseOldMDECRoutines", false);
 
-  DrawToggleSetting(bsi, FSUI_CSTR("Dump Replaceable VRAM Writes"),
+  const bool texture_cache_enabled = GetEffectiveBoolSetting(bsi, "GPU", "EnableTextureCache", false);
-                    FSUI_CSTR("Writes textures which can be replaced to the dump directory."), "TextureReplacements",
+  DrawToggleSetting(bsi, FSUI_CSTR("Enable Texture Replacements"),
+                    FSUI_CSTR("Enables loading of replacement textures. Not compatible with all games."),
+                    "TextureReplacements", "EnableTextureReplacements", false, texture_cache_enabled);
+  DrawToggleSetting(
+    bsi, FSUI_CSTR("Enable Texture Dumping"),
+    FSUI_CSTR("Enables dumping of textures to image files, which can be replaced. Not compatible with all games."),
+    "TextureReplacements", "DumpTextures", false, texture_cache_enabled);
+  DrawToggleSetting(bsi, FSUI_CSTR("Dump Replaced Textures"),
+                    FSUI_CSTR("Dumps textures that have replacements already loaded."), "TextureReplacements",
+                    "DumpReplacedTextures", false, texture_cache_enabled);
+
+  DrawToggleSetting(bsi, FSUI_CSTR("Enable VRAM Write Texture Replacement"),
+                    FSUI_CSTR("Enables the replacement of background textures in supported games."),
+                    "TextureReplacements", "EnableVRAMWriteReplacements", false);
+
+  DrawToggleSetting(bsi, FSUI_CSTR("Enable VRAM Write Dumping"),
+                    FSUI_CSTR("Writes backgrounds that can be replaced to the dump directory."), "TextureReplacements",
                     "DumpVRAMWrites", false);
-  DrawToggleSetting(bsi, FSUI_CSTR("Set VRAM Write Dump Alpha Channel"),
+
-                    FSUI_CSTR("Clears the mask/transparency bit in VRAM write dumps."), "TextureReplacements",
+  DrawToggleSetting(bsi, FSUI_CSTR("Preload Replacement Textures"),
-                    "DumpVRAMWriteForceAlphaChannel", true);
+                    FSUI_CSTR("Loads all replacement texture to RAM, reducing stuttering at runtime."),
+                    "TextureReplacements", "PreloadTextures", false,
+                    ((texture_cache_enabled &&
+                      GetEffectiveBoolSetting(bsi, "TextureReplacements", "EnableTextureReplacements", false)) ||
+                     GetEffectiveBoolSetting(bsi, "TextureReplacements", "EnableVRAMWriteReplacements", false)));
+
+  DrawFolderSetting(bsi, FSUI_CSTR("Textures Directory"), "Folders", "Textures", EmuFolders::Textures);
 
   EndMenuButtons();
 }
@@ -7239,7 +7261,6 @@ TRANSLATE_NOOP("FullscreenUI", "Clear Settings");
 TRANSLATE_NOOP("FullscreenUI", "Clear Shaders");
 TRANSLATE_NOOP("FullscreenUI", "Clears a shader from the chain.");
 TRANSLATE_NOOP("FullscreenUI", "Clears all settings set for this game.");
-TRANSLATE_NOOP("FullscreenUI", "Clears the mask/transparency bit in VRAM write dumps.");
 TRANSLATE_NOOP("FullscreenUI", "Close");
 TRANSLATE_NOOP("FullscreenUI", "Close Game");
 TRANSLATE_NOOP("FullscreenUI", "Close Menu");
@@ -7319,7 +7340,8 @@ TRANSLATE_NOOP("FullscreenUI", "Downsampling");
 TRANSLATE_NOOP("FullscreenUI", "Downsampling Display Scale");
 TRANSLATE_NOOP("FullscreenUI", "Duck icon by icons8 (https://icons8.com/icon/74847/platforms.undefined.short-title)");
 TRANSLATE_NOOP("FullscreenUI", "DuckStation is a free simulator/emulator of the Sony PlayStation(TM) console, focusing on playability, speed, and long-term maintainability.");
-TRANSLATE_NOOP("FullscreenUI", "Dump Replaceable VRAM Writes");
+TRANSLATE_NOOP("FullscreenUI", "Dump Replaced Textures");
+TRANSLATE_NOOP("FullscreenUI", "Dumps textures that have replacements already loaded.");
 TRANSLATE_NOOP("FullscreenUI", "Emulation Settings");
 TRANSLATE_NOOP("FullscreenUI", "Emulation Speed");
 TRANSLATE_NOOP("FullscreenUI", "Enable 8MB RAM");
@@ -7338,6 +7360,10 @@ TRANSLATE_NOOP("FullscreenUI", "Enable Rewinding");
 TRANSLATE_NOOP("FullscreenUI", "Enable SDL Input Source");
 TRANSLATE_NOOP("FullscreenUI", "Enable Subdirectory Scanning");
 TRANSLATE_NOOP("FullscreenUI", "Enable TTY Logging");
+TRANSLATE_NOOP("FullscreenUI", "Enable Texture Cache");
+TRANSLATE_NOOP("FullscreenUI", "Enable Texture Dumping");
+TRANSLATE_NOOP("FullscreenUI", "Enable Texture Replacements");
+TRANSLATE_NOOP("FullscreenUI", "Enable VRAM Write Dumping");
 TRANSLATE_NOOP("FullscreenUI", "Enable VRAM Write Texture Replacement");
 TRANSLATE_NOOP("FullscreenUI", "Enable XInput Input Source");
 TRANSLATE_NOOP("FullscreenUI", "Enable debugging when supported by the host's renderer API. Only for developer use.");
@@ -7345,6 +7371,9 @@ TRANSLATE_NOOP("FullscreenUI", "Enable/Disable the Player LED on DualSense contr
 TRANSLATE_NOOP("FullscreenUI", "Enables alignment and bus exceptions. Not needed for any known games.");
 TRANSLATE_NOOP("FullscreenUI", "Enables an additional 6MB of RAM to obtain a total of 2+6 = 8MB, usually present on dev consoles.");
 TRANSLATE_NOOP("FullscreenUI", "Enables an additional three controller slots on each port. Not supported in all games.");
+TRANSLATE_NOOP("FullscreenUI", "Enables caching of guest textures, required for texture replacement.");
+TRANSLATE_NOOP("FullscreenUI", "Enables dumping of textures to image files, which can be replaced. Not compatible with all games.");
+TRANSLATE_NOOP("FullscreenUI", "Enables loading of replacement textures. Not compatible with all games.");
 TRANSLATE_NOOP("FullscreenUI", "Enables smooth scrolling of menus in Big Picture UI.");
 TRANSLATE_NOOP("FullscreenUI", "Enables the older, less accurate MDEC decoding routines. May be required for old replacement backgrounds to match/load.");
 TRANSLATE_NOOP("FullscreenUI", "Enables the replacement of background textures in supported games.");
@@ -7626,7 +7655,6 @@ TRANSLATE_NOOP("FullscreenUI", "Selects the view that the game list will open to
 TRANSLATE_NOOP("FullscreenUI", "Serial");
 TRANSLATE_NOOP("FullscreenUI", "Session: {}");
 TRANSLATE_NOOP("FullscreenUI", "Set Input Binding");
-TRANSLATE_NOOP("FullscreenUI", "Set VRAM Write Dump Alpha Channel");
 TRANSLATE_NOOP("FullscreenUI", "Sets a threshold for discarding precise values when exceeded. May help with glitches in some games.");
 TRANSLATE_NOOP("FullscreenUI", "Sets a threshold for discarding the emulated depth buffer. May help in some games.");
 TRANSLATE_NOOP("FullscreenUI", "Sets the fast forward speed. It is not guaranteed that this speed will be reached on all systems.");
@@ -7705,10 +7733,12 @@ TRANSLATE_NOOP("FullscreenUI", "Temporarily disables all enhancements, useful wh
 TRANSLATE_NOOP("FullscreenUI", "Test Unofficial Achievements");
 TRANSLATE_NOOP("FullscreenUI", "Texture Filtering");
 TRANSLATE_NOOP("FullscreenUI", "Texture Replacements");
+TRANSLATE_NOOP("FullscreenUI", "Textures Directory");
 TRANSLATE_NOOP("FullscreenUI", "The SDL input source supports most controllers.");
 TRANSLATE_NOOP("FullscreenUI", "The XInput source provides support for XBox 360/XBox One/XBox Series controllers.");
 TRANSLATE_NOOP("FullscreenUI", "The audio backend determines how frames produced by the emulator are submitted to the host.");
 TRANSLATE_NOOP("FullscreenUI", "The selected memory card image will be used in shared mode for this slot.");
+TRANSLATE_NOOP("FullscreenUI", "The texture cache is currently experimental, and may cause rendering errors in some games.");
 TRANSLATE_NOOP("FullscreenUI", "This game has no achievements.");
 TRANSLATE_NOOP("FullscreenUI", "This game has no leaderboards.");
 TRANSLATE_NOOP("FullscreenUI", "Threaded Rendering");
@@ -7763,7 +7793,7 @@ TRANSLATE_NOOP("FullscreenUI", "When playing a multi-disc game and using per-gam
 TRANSLATE_NOOP("FullscreenUI", "When this option is chosen, the clock speed set below will be used.");
 TRANSLATE_NOOP("FullscreenUI", "Widescreen Rendering");
 TRANSLATE_NOOP("FullscreenUI", "Wireframe Rendering");
-TRANSLATE_NOOP("FullscreenUI", "Writes textures which can be replaced to the dump directory.");
+TRANSLATE_NOOP("FullscreenUI", "Writes backgrounds that can be replaced to the dump directory.");
 TRANSLATE_NOOP("FullscreenUI", "Yes, {} now and risk memory card corruption.");
 TRANSLATE_NOOP("FullscreenUI", "\"Challenge\" mode for achievements, including leaderboard tracking. Disables save state, cheats, and slowdown functions.");
 TRANSLATE_NOOP("FullscreenUI", "\"PlayStation\" and \"PSX\" are registered trademarks of Sony Interactive Entertainment Europe Limited. This software is not affiliated in any way with Sony Interactive Entertainment.");

src/core/gpu.cpp

@@ -1555,8 +1555,8 @@ void GPU::SetDrawMode(u16 value)
   if (new_mode_reg.bits == m_draw_mode.mode_reg.bits)
     return;
 
-  m_draw_mode.texture_page_changed |= ((new_mode_reg.bits & GPUDrawModeReg::TEXTURE_PAGE_MASK) !=
+  m_draw_mode.texture_page_changed |= ((new_mode_reg.bits & GPUDrawModeReg::TEXTURE_MODE_AND_PAGE_MASK) !=
-                                       (m_draw_mode.mode_reg.bits & GPUDrawModeReg::TEXTURE_PAGE_MASK));
+                                       (m_draw_mode.mode_reg.bits & GPUDrawModeReg::TEXTURE_MODE_AND_PAGE_MASK));
   m_draw_mode.mode_reg.bits = new_mode_reg.bits;
 
   if (m_GPUSTAT.draw_to_displayed_field != new_mode_reg.draw_to_displayed_field)

src/core/gpu_commands.cpp

@@ -2,9 +2,9 @@
 // SPDX-License-Identifier: CC-BY-NC-ND-4.0
 
 #include "gpu.h"
+#include "gpu_hw_texture_cache.h"
 #include "interrupt_controller.h"
 #include "system.h"
-#include "texture_replacements.h"
 
 #include "common/assert.h"
 #include "common/log.h"
@@ -532,9 +532,9 @@ void GPU::FinishVRAMWrite()
                      m_vram_transfer.height, sizeof(u16) * m_vram_transfer.width, m_blit_buffer.data(), true);
     }
 
-    if (g_settings.texture_replacements.ShouldDumpVRAMWrite(m_vram_transfer.width, m_vram_transfer.height))
+    if (GPUTextureCache::ShouldDumpVRAMWrite(m_vram_transfer.width, m_vram_transfer.height))
     {
-      TextureReplacements::DumpVRAMWrite(m_vram_transfer.width, m_vram_transfer.height,
+      GPUTextureCache::DumpVRAMWrite(m_vram_transfer.width, m_vram_transfer.height,
                                      reinterpret_cast<const u16*>(m_blit_buffer.data()));
     }
 

src/core/gpu_hw.cpp

@@ -194,6 +194,8 @@ GPU_HW::GPU_HW() : GPU()
 
 GPU_HW::~GPU_HW()
 {
+  GPUTextureCache::Shutdown();
+
   if (m_sw_renderer)
   {
     m_sw_renderer->Shutdown();
@@ -262,6 +264,8 @@ bool GPU_HW::Initialize()
   m_clamp_uvs = ShouldClampUVs(m_texture_filtering) || ShouldClampUVs(m_sprite_texture_filtering);
   m_compute_uv_range = m_clamp_uvs;
   m_allow_sprite_mode = ShouldAllowSpriteMode(m_resolution_scale, m_texture_filtering, m_sprite_texture_filtering);
+  m_use_texture_cache = g_settings.gpu_texture_cache;
+  m_texture_dumping = m_use_texture_cache && g_settings.texture_replacements.dump_textures;
 
   CheckSettings();
 
@@ -282,13 +286,27 @@ bool GPU_HW::Initialize()
     return false;
   }
 
+  if (m_use_texture_cache)
+  {
+    if (!GPUTextureCache::Initialize())
+    {
+      ERROR_LOG("Failed to initialize texture cache, disabling.");
+      m_use_texture_cache = false;
+    }
+  }
+
   UpdateDownsamplingLevels();
+
   RestoreDeviceContext();
   return true;
 }
 
 void GPU_HW::Reset(bool clear_vram)
 {
+  // Texture cache needs to be invalidated before we load, otherwise we dump black.
+  if (m_use_texture_cache)
+    GPUTextureCache::Invalidate();
+
   if (m_batch_vertex_ptr)
     UnmapGPUBuffer(0, 0);
 
@@ -364,6 +382,7 @@ bool GPU_HW::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_di
   else if (sw.IsReading())
   {
     // Need to update the VRAM copy on the GPU with the state data.
+    // Would invalidate the TC, but base DoState() calls Reset().
     UpdateVRAMOnGPU(0, 0, VRAM_WIDTH, VRAM_HEIGHT, g_vram, VRAM_WIDTH * sizeof(u16), false, false, VRAM_SIZE_RECT);
   }
 
@@ -373,10 +392,12 @@ bool GPU_HW::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_di
     DebugAssert(!m_batch_vertex_ptr && !m_batch_index_ptr);
     ClearVRAMDirtyRectangle();
     SetFullVRAMDirtyRectangle();
+    UpdateVRAMReadTexture(true, false);
+    ClearVRAMDirtyRectangle();
     ResetBatchVertexDepth();
   }
 
-  return true;
+  return GPUTextureCache::DoState(sw, !m_use_texture_cache);
 }
 
 void GPU_HW::RestoreDeviceContext()
@@ -401,7 +422,8 @@ void GPU_HW::UpdateSettings(const Settings& old_settings)
   const bool clamp_uvs = ShouldClampUVs(m_texture_filtering) || ShouldClampUVs(m_sprite_texture_filtering);
   const bool framebuffer_changed = (m_resolution_scale != resolution_scale || m_multisamples != multisamples ||
                                     g_settings.IsUsingAccurateBlending() != old_settings.IsUsingAccurateBlending() ||
-                                    m_pgxp_depth_buffer != g_settings.UsingPGXPDepthBuffer());
+                                    m_pgxp_depth_buffer != g_settings.UsingPGXPDepthBuffer() ||
+                                    (!old_settings.gpu_texture_cache && g_settings.gpu_texture_cache));
   const bool shaders_changed =
     (m_resolution_scale != resolution_scale || m_multisamples != multisamples ||
      m_true_color != g_settings.gpu_true_color || prev_force_progressive_scan != m_force_progressive_scan ||
@@ -468,6 +490,8 @@ void GPU_HW::UpdateSettings(const Settings& old_settings)
   m_clamp_uvs = clamp_uvs;
   m_compute_uv_range = m_clamp_uvs;
   m_allow_sprite_mode = ShouldAllowSpriteMode(resolution_scale, m_texture_filtering, m_sprite_texture_filtering);
+  m_use_texture_cache = g_settings.gpu_texture_cache;
+  m_texture_dumping = m_use_texture_cache && g_settings.texture_replacements.dump_textures;
   m_batch.sprite_mode = (m_allow_sprite_mode && m_batch.sprite_mode);
 
   const bool depth_buffer_changed = (m_pgxp_depth_buffer != g_settings.UsingPGXPDepthBuffer());
@@ -521,6 +545,23 @@ void GPU_HW::UpdateSettings(const Settings& old_settings)
       UpdateDepthBufferFromMaskBit();
   }
 
+  if (m_use_texture_cache && !old_settings.gpu_texture_cache)
+  {
+    if (!GPUTextureCache::Initialize())
+    {
+      ERROR_LOG("Failed to initialize texture cache, disabling.");
+      m_use_texture_cache = false;
+    }
+  }
+  else if (!m_use_texture_cache && old_settings.gpu_texture_cache)
+  {
+    GPUTextureCache::Shutdown();
+  }
+  else if (m_use_texture_cache)
+  {
+    GPUTextureCache::UpdateSettings(old_settings);
+  }
+
   if (g_settings.gpu_downsample_mode != old_settings.gpu_downsample_mode ||
       (g_settings.gpu_downsample_mode == GPUDownsampleMode::Box &&
        g_settings.gpu_downsample_scale != old_settings.gpu_downsample_scale))
@@ -728,6 +769,9 @@ void GPU_HW::AddWrittenRectangle(const GSVector4i rect)
 {
   m_vram_dirty_write_rect = m_vram_dirty_write_rect.runion(rect);
   SetTexPageChangedOnOverlap(m_vram_dirty_write_rect);
+
+  if (m_use_texture_cache)
+    GPUTextureCache::AddWrittenRectangle(rect);
 }
 
 void GPU_HW::AddDrawnRectangle(const GSVector4i rect)
@@ -735,13 +779,22 @@ void GPU_HW::AddDrawnRectangle(const GSVector4i rect)
   // Normally, we would check for overlap here. But the GPU's texture cache won't actually reload until the page
   // changes, or it samples a larger region, so we can get away without doing so. This reduces copies considerably in
   // games like Mega Man Legends 2.
-  m_vram_dirty_draw_rect = m_vram_dirty_draw_rect.runion(rect);
+  if (m_current_draw_rect.rcontains(rect))
+    return;
+
+  m_current_draw_rect = m_current_draw_rect.runion(rect);
+  m_vram_dirty_draw_rect = m_vram_dirty_draw_rect.runion(m_current_draw_rect);
+
+  if (m_use_texture_cache)
+    GPUTextureCache::AddDrawnRectangle(m_current_draw_rect, m_clamped_drawing_area);
 }
 
 void GPU_HW::AddUnclampedDrawnRectangle(const GSVector4i rect)
 {
   m_vram_dirty_draw_rect = m_vram_dirty_draw_rect.runion(rect);
   SetTexPageChangedOnOverlap(m_vram_dirty_draw_rect);
+  if (m_use_texture_cache)
+    GPUTextureCache::AddDrawnRectangle(rect, rect);
 }
 
 void GPU_HW::SetTexPageChangedOnOverlap(const GSVector4i update_rect)
@@ -749,9 +802,9 @@ void GPU_HW::SetTexPageChangedOnOverlap(const GSVector4i update_rect)
   // the vram area can include the texture page, but the game can leave it as-is. in this case, set it as dirty so the
   // shadow texture is updated
   if (!m_draw_mode.IsTexturePageChanged() && m_batch.texture_mode != BatchTextureMode::Disabled &&
-      (m_draw_mode.mode_reg.GetTexturePageRectangle().rintersects(update_rect) ||
+      (GetTextureRect(m_draw_mode.mode_reg.texture_page, m_draw_mode.mode_reg.texture_mode).rintersects(update_rect) ||
        (m_draw_mode.mode_reg.IsUsingPalette() &&
-        m_draw_mode.palette_reg.GetRectangle(m_draw_mode.mode_reg.texture_mode).rintersects(update_rect))))
+        GetPaletteRect(m_draw_mode.palette_reg, m_draw_mode.mode_reg.texture_mode).rintersects(update_rect))))
   {
     m_draw_mode.SetTexturePageChanged();
   }
@@ -882,6 +935,8 @@ void GPU_HW::ClearFramebuffer()
       g_gpu_device->ClearDepth(m_vram_depth_texture.get(), m_pgxp_depth_buffer ? 1.0f : 0.0f);
   }
   ClearVRAMDirtyRectangle();
+  if (m_use_texture_cache)
+    GPUTextureCache::Invalidate();
   m_last_depth_z = 1.0f;
 }
 
@@ -986,7 +1041,7 @@ bool GPU_HW::CompilePipelines(Error* error)
   const u32 active_texture_modes =
     m_allow_sprite_mode ? NUM_TEXTURE_MODES :
                           (NUM_TEXTURE_MODES - (NUM_TEXTURE_MODES - static_cast<u32>(BatchTextureMode::SpriteStart)));
-  const u32 total_vertex_shaders = (m_allow_sprite_mode ? 5 : 3);
+  const u32 total_vertex_shaders = (m_allow_sprite_mode ? 7 : 3);
   const u32 total_fragment_shaders =
     ((needs_rov_depth ? 2 : 1) * 5 * 5 * active_texture_modes * 2 * (1 + BoolToUInt32(!true_color)) *
      (1 + BoolToUInt32(!m_force_progressive_scan)) * (1 + BoolToUInt32(needs_rov_depth)));
@@ -1013,7 +1068,7 @@ bool GPU_HW::CompilePipelines(Error* error)
   // vertex shaders - [textured/palette/sprite]
   // fragment shaders - [depth_test][render_mode][transparency_mode][texture_mode][check_mask][dithering][interlacing]
   static constexpr auto destroy_shader = [](std::unique_ptr<GPUShader>& s) { s.reset(); };
-  DimensionalArray<std::unique_ptr<GPUShader>, 2, 2, 2> batch_vertex_shaders{};
+  DimensionalArray<std::unique_ptr<GPUShader>, 2, 3, 2> batch_vertex_shaders{};
   DimensionalArray<std::unique_ptr<GPUShader>, 2, 2, 2, NUM_TEXTURE_MODES, 5, 5, 2> batch_fragment_shaders{};
   ScopedGuard batch_shader_guard([&batch_vertex_shaders, &batch_fragment_shaders]() {
     batch_vertex_shaders.enumerate(destroy_shader);
@@ -1022,7 +1077,7 @@ bool GPU_HW::CompilePipelines(Error* error)
 
   for (u8 textured = 0; textured < 2; textured++)
   {
-    for (u8 palette = 0; palette < 2; palette++)
+    for (u8 palette = 0; palette < 3; palette++)
     {
       if (palette && !textured)
         continue;
@@ -1034,7 +1089,7 @@ bool GPU_HW::CompilePipelines(Error* error)
 
         const bool uv_limits = ShouldClampUVs(sprite ? m_sprite_texture_filtering : m_texture_filtering);
         const std::string vs = shadergen.GenerateBatchVertexShader(
-          textured != 0, palette != 0, uv_limits, !sprite && force_round_texcoords, m_pgxp_depth_buffer);
+          textured != 0, palette == 1, palette == 2, uv_limits, !sprite && force_round_texcoords, m_pgxp_depth_buffer);
         if (!(batch_vertex_shaders[textured][palette][sprite] =
                 g_gpu_device->CreateShader(GPUShaderStage::Vertex, shadergen.GetLanguage(), vs, error)))
         {
@@ -1203,6 +1258,8 @@ bool GPU_HW::CompilePipelines(Error* error)
                    static_cast<BatchTextureMode>(texture_mode) == BatchTextureMode::Palette8Bit ||
                    static_cast<BatchTextureMode>(texture_mode) == BatchTextureMode::SpritePalette4Bit ||
                    static_cast<BatchTextureMode>(texture_mode) == BatchTextureMode::SpritePalette8Bit);
+                const bool page_texture =
+                  (static_cast<BatchTextureMode>(texture_mode) == BatchTextureMode::PageTexture);
                 const bool sprite = (static_cast<BatchTextureMode>(texture_mode) >= BatchTextureMode::SpriteStart);
                 const bool uv_limits = ShouldClampUVs(sprite ? m_sprite_texture_filtering : m_texture_filtering);
                 const bool use_shader_blending = (render_mode == static_cast<u8>(BatchRenderMode::ShaderBlend));
@@ -1216,7 +1273,9 @@ bool GPU_HW::CompilePipelines(Error* error)
                     std::span<const GPUPipeline::VertexAttribute>(vertex_attributes, NUM_BATCH_VERTEX_ATTRIBUTES);
 
                 plconfig.vertex_shader =
-                  batch_vertex_shaders[BoolToUInt8(textured)][BoolToUInt8(palette)][BoolToUInt8(sprite)].get();
+                  batch_vertex_shaders[BoolToUInt8(textured)][page_texture ? 2 : BoolToUInt8(palette)]
+                                      [BoolToUInt8(sprite)]
+                                        .get();
                 plconfig.fragment_shader =
                   batch_fragment_shaders[BoolToUInt8(depth_test && needs_rov_depth)][render_mode]
                                         [use_shader_blending ? transparency_mode :
@@ -1836,19 +1895,26 @@ void GPU_HW::UnmapGPUBuffer(u32 used_vertices, u32 used_indices)
 }
 
 ALWAYS_INLINE_RELEASE void GPU_HW::DrawBatchVertices(BatchRenderMode render_mode, u32 num_indices, u32 base_index,
-                                                     u32 base_vertex)
+                                                     u32 base_vertex, const GPUTextureCache::Source* texture)
 {
   // [depth_test][transparency_mode][render_mode][texture_mode][dithering][interlacing][check_mask]
-  const u8 texture_mode = static_cast<u8>(m_batch.texture_mode) +
+  const u8 texture_mode = texture ? static_cast<u8>(BatchTextureMode::PageTexture) :
-                          ((m_batch.texture_mode != BatchTextureMode::Disabled && m_batch.sprite_mode) ?
+                                    (static_cast<u8>(m_batch.texture_mode) +
+                                     ((m_batch.texture_mode < BatchTextureMode::PageTexture && m_batch.sprite_mode) ?
                                         static_cast<u8>(BatchTextureMode::SpriteStart) :
-                             0);
+                                        0));
   const u8 depth_test = BoolToUInt8(m_batch.use_depth_buffer);
   const u8 check_mask = BoolToUInt8(m_batch.check_mask_before_draw);
   g_gpu_device->SetPipeline(m_batch_pipelines[depth_test][static_cast<u8>(m_batch.transparency_mode)][static_cast<u8>(
     render_mode)][texture_mode][BoolToUInt8(m_batch.dithering)][BoolToUInt8(m_batch.interlacing)][check_mask]
                               .get());
 
+  if (m_use_texture_cache && texture_mode != static_cast<u8>(BatchTextureMode::Disabled))
+  {
+    g_gpu_device->SetTextureSampler(0, texture ? texture->texture : m_vram_read_texture.get(),
+                                    g_gpu_device->GetNearestSampler());
+  }
+
   GL_INS_FMT("Texture mode: {}", s_batch_texture_modes[texture_mode]);
   GL_INS_FMT("Transparency mode: {}", s_transparency_modes[static_cast<u8>(m_batch.transparency_mode)]);
   GL_INS_FMT("Render mode: {}", s_batch_render_modes[static_cast<u8>(render_mode)]);
@@ -2199,7 +2265,7 @@ void GPU_HW::ComputePolygonUVLimits(BatchVertex* vertices, u32 num_vertices)
   for (u32 i = 0; i < num_vertices; i++)
     vertices[i].SetUVLimits(min_u, max_u, min_v, max_v);
 
-  if (m_texpage_dirty != 0)
+  if (ShouldCheckForTexPageOverlap())
     CheckForTexPageOverlap(GSVector4i(min).upl32(GSVector4i(max)).u16to32());
 }
 
@@ -2620,7 +2686,7 @@ void GPU_HW::LoadVertices()
           const u32 tex_right = tex_left + quad_width;
           const u32 uv_limits = BatchVertex::PackUVLimits(tex_left, tex_right - 1, tex_top, tex_bottom - 1);
 
-          if (rc.texture_enable && m_texpage_dirty != 0)
+          if (rc.texture_enable && ShouldCheckForTexPageOverlap())
           {
             CheckForTexPageOverlap(GSVector4i(static_cast<s32>(tex_left), static_cast<s32>(tex_top),
                                               static_cast<s32>(tex_right), static_cast<s32>(tex_bottom)));
@@ -2801,7 +2867,7 @@ void GPU_HW::LoadVertices()
   }
 }
 
-bool GPU_HW::BlitVRAMReplacementTexture(const TextureReplacements::ReplacementImage* tex, u32 dst_x, u32 dst_y,
+bool GPU_HW::BlitVRAMReplacementTexture(const GPUTextureCache::TextureReplacementImage* tex, u32 dst_x, u32 dst_y,
                                         u32 width, u32 height)
 {
   if (!m_vram_replacement_texture || m_vram_replacement_texture->GetWidth() < tex->GetWidth() ||
@@ -2844,7 +2910,7 @@ bool GPU_HW::BlitVRAMReplacementTexture(const TextureReplacements::ReplacementIm
 
 ALWAYS_INLINE_RELEASE void GPU_HW::CheckForTexPageOverlap(GSVector4i uv_rect)
 {
-  DebugAssert(m_texpage_dirty != 0 && m_batch.texture_mode != BatchTextureMode::Disabled);
+  DebugAssert((m_texpage_dirty != 0 || m_texture_dumping) && m_batch.texture_mode != BatchTextureMode::Disabled);
 
   if (m_texture_window_active)
   {
@@ -2871,6 +2937,34 @@ ALWAYS_INLINE_RELEASE void GPU_HW::CheckForTexPageOverlap(GSVector4i uv_rect)
     m_current_uv_rect = new_uv_rect;
 
     bool update_drawn = false, update_written = false;
+    if (m_texpage_dirty & TEXPAGE_DIRTY_PAGE_RECT)
+    {
+      DebugAssert(!(m_texpage_dirty & (TEXPAGE_DIRTY_DRAWN_RECT | TEXPAGE_DIRTY_WRITTEN_RECT)));
+      DebugAssert(m_batch.texture_mode == BatchTextureMode::PageTexture &&
+                  m_batch.texture_cache_key.page < NUM_VRAM_PAGES);
+
+      if (GPUTextureCache::AreSourcePagesDrawn(m_batch.texture_cache_key, m_current_uv_rect))
+      {
+        // UVs intersect with drawn area, can't use TC
+        if (m_batch_index_count > 0)
+        {
+          FlushRender();
+          EnsureVertexBufferSpaceForCurrentCommand();
+        }
+
+        // We need to swap the dirty tracking over to drawn/written.
+        const GSVector4i page_rect = GetTextureRect(m_batch.texture_cache_key.page, m_batch.texture_cache_key.mode);
+        m_texpage_dirty = (m_vram_dirty_draw_rect.rintersects(page_rect) ? TEXPAGE_DIRTY_DRAWN_RECT : 0) |
+                          (m_vram_dirty_write_rect.rintersects(page_rect) ? TEXPAGE_DIRTY_WRITTEN_RECT : 0);
+        m_compute_uv_range = (ShouldCheckForTexPageOverlap() || m_clamp_uvs);
+        m_batch.texture_mode = static_cast<BatchTextureMode>(m_draw_mode.mode_reg.texture_mode.GetValue());
+      }
+      else
+      {
+        // Page isn't drawn, we're done.
+        return;
+      }
+    }
     if (m_texpage_dirty & TEXPAGE_DIRTY_DRAWN_RECT)
     {
       DebugAssert(!m_vram_dirty_draw_rect.eq(INVALID_RECT));
@@ -2905,6 +2999,11 @@ ALWAYS_INLINE_RELEASE void GPU_HW::CheckForTexPageOverlap(GSVector4i uv_rect)
   }
 }
 
+bool GPU_HW::ShouldCheckForTexPageOverlap() const
+{
+  return (m_texpage_dirty != 0);
+}
+
 ALWAYS_INLINE bool GPU_HW::IsFlushed() const
 {
   return (m_batch_index_count == 0);
@@ -3003,8 +3102,13 @@ void GPU_HW::UpdateSoftwareRenderer(bool copy_vram_from_hw)
 {
   const bool current_enabled = (m_sw_renderer != nullptr);
   const bool new_enabled = g_settings.gpu_use_software_renderer_for_readbacks;
+  const bool use_thread = !g_settings.gpu_texture_cache;
   if (current_enabled == new_enabled)
+  {
+    if (m_sw_renderer)
+      m_sw_renderer->SetThreadEnabled(use_thread);
     return;
+  }
 
   if (!new_enabled)
   {
@@ -3015,7 +3119,7 @@ void GPU_HW::UpdateSoftwareRenderer(bool copy_vram_from_hw)
   }
 
   std::unique_ptr<GPU_SW_Backend> sw_renderer = std::make_unique<GPU_SW_Backend>();
-  if (!sw_renderer->Initialize(true))
+  if (!sw_renderer->Initialize(use_thread))
     return;
 
   // We need to fill in the SW renderer's VRAM with the current state for hot toggles.
@@ -3080,7 +3184,17 @@ void GPU_HW::FillVRAM(u32 x, u32 y, u32 width, u32 height, u32 color)
   GL_INS_FMT("Dirty draw area before: {}", m_vram_dirty_draw_rect);
 
   const GSVector4i bounds = GetVRAMTransferBounds(x, y, width, height);
+
+  // If TC is enabled, we have to update local memory.
+  if (m_use_texture_cache && !IsInterlacedRenderingEnabled())
+  {
+    AddWrittenRectangle(bounds);
+    GPU_SW_Rasterizer::FillVRAM(x, y, width, height, color, false, 0);
+  }
+  else
+  {
     AddUnclampedDrawnRectangle(bounds);
+  }
 
   GL_INS_FMT("Dirty draw area after: {}", m_vram_dirty_draw_rect);
 
@@ -3126,6 +3240,8 @@ void GPU_HW::ReadVRAM(u32 x, u32 y, u32 width, u32 height)
     return;
   }
 
+  // TODO: Only read if it's in the drawn area
+
   // Get bounds with wrap-around handled.
   GSVector4i copy_rect = GetVRAMTransferBounds(x, y, width, height);
 
@@ -3177,7 +3293,17 @@ void GPU_HW::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, b
 {
   GL_SCOPE_FMT("UpdateVRAM({},{} => {},{} ({}x{})", x, y, x + width, y + height, width, height);
 
-  if (m_sw_renderer)
+  // TODO: Handle wrapped transfers... break them up or something
+  const GSVector4i bounds = GetVRAMTransferBounds(x, y, width, height);
+  DebugAssert(bounds.right <= static_cast<s32>(VRAM_WIDTH) && bounds.bottom <= static_cast<s32>(VRAM_HEIGHT));
+  AddWrittenRectangle(bounds);
+
+  // We want to dump *before* the write goes through, otherwise we dump bad data.
+  if (m_use_texture_cache)
+  {
+    GPUTextureCache::WriteVRAM(x, y, width, height, data, set_mask, check_mask, bounds);
+  }
+  else if (m_sw_renderer)
   {
     const u32 num_words = width * height;
     GPUBackendUpdateVRAMCommand* cmd = m_sw_renderer->NewUpdateVRAMCommand(num_words);
@@ -3192,10 +3318,6 @@ void GPU_HW::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, b
     m_sw_renderer->PushCommand(cmd);
   }
 
-  const GSVector4i bounds = GetVRAMTransferBounds(x, y, width, height);
-  DebugAssert(bounds.right <= static_cast<s32>(VRAM_WIDTH) && bounds.bottom <= static_cast<s32>(VRAM_HEIGHT));
-  AddWrittenRectangle(bounds);
-
   if (check_mask)
   {
     // set new vertex counter since we want this to take into consideration previous masked pixels
@@ -3203,7 +3325,7 @@ void GPU_HW::UpdateVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, b
   }
   else
   {
-    const TextureReplacements::ReplacementImage* rtex = TextureReplacements::GetVRAMReplacement(width, height, data);
+    const GPUTextureCache::TextureReplacementImage* rtex = GPUTextureCache::GetVRAMReplacement(width, height, data);
     if (rtex && BlitVRAMReplacementTexture(rtex, x * m_resolution_scale, y * m_resolution_scale,
                                            width * m_resolution_scale, height * m_resolution_scale))
     {
@@ -3283,7 +3405,26 @@ void GPU_HW::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32
 {
   GL_SCOPE_FMT("CopyVRAM({}x{} @ {},{} => {},{}", width, height, src_x, src_y, dst_x, dst_y);
 
-  if (m_sw_renderer)
+  // masking enabled, oversized, or overlapping
+  const bool use_shader =
+    (m_GPUSTAT.IsMaskingEnabled() || ((src_x % VRAM_WIDTH) + width) > VRAM_WIDTH ||
+     ((src_y % VRAM_HEIGHT) + height) > VRAM_HEIGHT || ((dst_x % VRAM_WIDTH) + width) > VRAM_WIDTH ||
+     ((dst_y % VRAM_HEIGHT) + height) > VRAM_HEIGHT);
+  const GSVector4i src_bounds = GetVRAMTransferBounds(src_x, src_y, width, height);
+  const GSVector4i dst_bounds = GetVRAMTransferBounds(dst_x, dst_y, width, height);
+
+  // If we're copying a region that hasn't been drawn to, and we're using the TC, we can do it in local memory.
+  if (m_use_texture_cache && !GPUTextureCache::IsRectDrawn(src_bounds))
+  {
+    GL_INS("Performed in local memory.");
+    GPUTextureCache::CopyVRAM(src_x, src_y, dst_x, dst_y, width, height, m_GPUSTAT.set_mask_while_drawing,
+                              m_GPUSTAT.check_mask_before_draw, src_bounds, dst_bounds);
+    UpdateVRAMOnGPU(dst_bounds.left, dst_bounds.top, dst_bounds.width(), dst_bounds.height(),
+                    &g_vram[dst_bounds.top * VRAM_WIDTH + dst_bounds.left], VRAM_WIDTH * sizeof(u16), false, false,
+                    dst_bounds);
+    return;
+  }
+  else if (m_sw_renderer)
   {
     GPUBackendCopyVRAMCommand* cmd = m_sw_renderer->NewCopyVRAMCommand();
     FillBackendCommandParameters(cmd);
@@ -3296,16 +3437,8 @@ void GPU_HW::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32
     m_sw_renderer->PushCommand(cmd);
   }
 
-  // masking enabled, oversized, or overlapping
-  const bool use_shader =
-    (m_GPUSTAT.IsMaskingEnabled() || ((src_x % VRAM_WIDTH) + width) > VRAM_WIDTH ||
-     ((src_y % VRAM_HEIGHT) + height) > VRAM_HEIGHT || ((dst_x % VRAM_WIDTH) + width) > VRAM_WIDTH ||
-     ((dst_y % VRAM_HEIGHT) + height) > VRAM_HEIGHT);
-  const GSVector4i src_bounds = GetVRAMTransferBounds(src_x, src_y, width, height);
-  const GSVector4i dst_bounds = GetVRAMTransferBounds(dst_x, dst_y, width, height);
   const bool intersect_with_draw = m_vram_dirty_draw_rect.rintersects(src_bounds);
   const bool intersect_with_write = m_vram_dirty_write_rect.rintersects(src_bounds);
-
   if (use_shader || IsUsingMultisampling())
   {
     if (intersect_with_draw || intersect_with_write)
@@ -3343,6 +3476,7 @@ void GPU_HW::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32
     g_gpu_device->SetViewportAndScissor(dst_bounds_scaled);
     g_gpu_device->SetPipeline(
       m_vram_copy_pipelines[BoolToUInt8(m_GPUSTAT.check_mask_before_draw && m_write_mask_as_depth)].get());
+    g_gpu_device->SetTextureSampler(0, m_vram_read_texture.get(), g_gpu_device->GetNearestSampler());
     g_gpu_device->PushUniformBuffer(&uniforms, sizeof(uniforms));
     g_gpu_device->Draw(3, 0);
     RestoreDeviceContext();
@@ -3362,7 +3496,8 @@ void GPU_HW::CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32
       UpdateVRAMReadTexture(intersect_with_draw, intersect_with_write);
   }
 
-  if (intersect_with_draw)
+  // We don't have it in local memory, so TC can't read it.
+  if (intersect_with_draw || m_use_texture_cache)
   {
     AddUnclampedDrawnRectangle(dst_bounds);
   }
@@ -3398,27 +3533,30 @@ void GPU_HW::DispatchRenderCommand()
 {
   const GPURenderCommand rc{m_render_command.bits};
 
-  BatchTextureMode texture_mode = BatchTextureMode::Disabled;
+  // TODO: avoid all this for vertex loading, only do when the type of draw changes
+  BatchTextureMode texture_mode = rc.IsTexturingEnabled() ? m_batch.texture_mode : BatchTextureMode::Disabled;
+  GPUTextureCache::SourceKey texture_cache_key = m_batch.texture_cache_key;
   if (rc.IsTexturingEnabled())
   {
     // texture page changed - check that the new page doesn't intersect the drawing area
-    if (m_draw_mode.IsTexturePageChanged())
+    if (m_draw_mode.IsTexturePageChanged() || texture_mode == BatchTextureMode::Disabled)
     {
       m_draw_mode.ClearTexturePageChangedFlag();
 
-#if 0
+      // start by assuming we can use the TC
-      if (!m_vram_dirty_draw_rect.eq(INVALID_RECT) || !m_vram_dirty_write_rect.eq(INVALID_RECT))
+      bool use_texture_cache = m_use_texture_cache;
-      {
-        GL_INS_FMT("VRAM DIRTY: {} {}", m_vram_dirty_draw_rect, m_vram_dirty_write_rect);
-        GL_INS_FMT("PAGE RECT: {}", m_draw_mode.mode_reg.GetTexturePageRectangle());
-        if (m_draw_mode.mode_reg.IsUsingPalette())
-          GL_INS_FMT("PALETTE RECT: {}", m_draw_mode.palette_reg.GetRectangle(m_draw_mode.mode_reg.texture_mode));
-      }
-#endif
 
+      // check that the palette isn't in a drawn area
       if (m_draw_mode.mode_reg.IsUsingPalette())
       {
-        const GSVector4i palette_rect = m_draw_mode.palette_reg.GetRectangle(m_draw_mode.mode_reg.texture_mode);
+        const GSVector4i palette_rect =
+          GetPaletteRect(m_draw_mode.palette_reg, m_draw_mode.mode_reg.texture_mode, use_texture_cache);
+        if (!use_texture_cache || GPUTextureCache::IsRectDrawn(palette_rect))
+        {
+          if (use_texture_cache)
+            GL_INS_FMT("Palette at {} is in drawn area, can't use TC", palette_rect);
+          use_texture_cache = false;
+
           const bool update_drawn = palette_rect.rintersects(m_vram_dirty_draw_rect);
           const bool update_written = palette_rect.rintersects(m_vram_dirty_write_rect);
           if (update_drawn || update_written)
@@ -3430,33 +3568,61 @@ void GPU_HW::DispatchRenderCommand()
             UpdateVRAMReadTexture(update_drawn, update_written);
           }
         }
+      }
 
-      const GSVector4i page_rect = m_draw_mode.mode_reg.GetTexturePageRectangle();
+      m_compute_uv_range = (m_clamp_uvs || m_texture_dumping);
-      GSVector4i::storel(m_current_texture_page_offset, page_rect);
 
-      u8 new_texpage_dirty = m_vram_dirty_draw_rect.rintersects(page_rect) ? TEXPAGE_DIRTY_DRAWN_RECT : 0;
+      const GPUTextureMode gpu_texture_mode =
-      new_texpage_dirty |= m_vram_dirty_write_rect.rintersects(page_rect) ? TEXPAGE_DIRTY_WRITTEN_RECT : 0;
+        (m_draw_mode.mode_reg.texture_mode == GPUTextureMode::Reserved_Direct16Bit) ? GPUTextureMode::Direct16Bit :
+                                                                                      m_draw_mode.mode_reg.texture_mode;
+      const GSVector4i page_rect = GetTextureRect(m_draw_mode.mode_reg.texture_page, m_draw_mode.mode_reg.texture_mode);
 
-      if (new_texpage_dirty != 0)
+      // TODO: This will result in incorrect global-space UVs when the texture page wraps around.
+      // Need to deal with it if it becomes a problem.
+      m_current_texture_page_offset[0] = static_cast<s32>(m_draw_mode.mode_reg.GetTexturePageBaseX());
+      m_current_texture_page_offset[1] = static_cast<s32>(m_draw_mode.mode_reg.GetTexturePageBaseY());
+
+      if (use_texture_cache)
       {
-        GL_INS("Texpage is in dirty area, checking UV ranges");
+        texture_mode = BatchTextureMode::PageTexture;
-        m_texpage_dirty = new_texpage_dirty;
+        texture_cache_key =
-        m_compute_uv_range = true;
+          GPUTextureCache::SourceKey(m_draw_mode.mode_reg.texture_page, m_draw_mode.palette_reg, gpu_texture_mode);
-        m_current_uv_rect = INVALID_RECT;
+
+        const bool is_drawn = GPUTextureCache::IsRectDrawn(page_rect);
+        if (is_drawn)
+          GL_INS_FMT("Texpage [{}] {} is drawn in TC, checking UV ranges", texture_cache_key.page, page_rect);
+
+        m_texpage_dirty =
+          (is_drawn ? TEXPAGE_DIRTY_PAGE_RECT : 0) | (m_texture_dumping ? TEXPAGE_DIRTY_ONLY_UV_RECT : 0);
+        m_compute_uv_range |= ShouldCheckForTexPageOverlap();
       }
       else
       {
-        m_compute_uv_range = m_clamp_uvs;
+        texture_mode = static_cast<BatchTextureMode>(gpu_texture_mode);
-        if (m_texpage_dirty)
+        m_texpage_dirty = (m_vram_dirty_draw_rect.rintersects(page_rect) ? TEXPAGE_DIRTY_DRAWN_RECT : 0) |
-          GL_INS("Texpage is no longer dirty");
+                          (m_vram_dirty_write_rect.rintersects(page_rect) ? TEXPAGE_DIRTY_WRITTEN_RECT : 0);
-        m_texpage_dirty = 0;
+        if (m_texpage_dirty & TEXPAGE_DIRTY_DRAWN_RECT)
+          GL_INS_FMT("Texpage {} is in dirty DRAWN area {}", page_rect, m_vram_dirty_draw_rect);
+        if (m_texpage_dirty & TEXPAGE_DIRTY_WRITTEN_RECT)
+          GL_INS_FMT("Texpage {} is in dirty WRITTEN area {}", page_rect, m_vram_dirty_write_rect);
+
+        // Current UV rect _must_ be cleared here, because we're only check for texpage intersection when it grows in
+        // size, a switch from a non-contained page to a contained page would go undetected otherwise.
+        if (m_texpage_dirty != 0)
+        {
+          m_compute_uv_range = true;
+          m_current_uv_rect = INVALID_RECT;
+        }
+      }
     }
   }
 
-    texture_mode = (m_draw_mode.mode_reg.texture_mode == GPUTextureMode::Reserved_Direct16Bit) ?
+  DebugAssert((rc.IsTexturingEnabled() && (texture_mode == BatchTextureMode::PageTexture &&
-                     BatchTextureMode::Direct16Bit :
+                                           texture_cache_key.mode == m_draw_mode.mode_reg.texture_mode) ||
-                     static_cast<BatchTextureMode>(m_draw_mode.mode_reg.texture_mode.GetValue());
+               texture_mode == static_cast<BatchTextureMode>(m_draw_mode.mode_reg.texture_mode.GetValue())) ||
-  }
+              (!rc.IsTexturingEnabled() && texture_mode == BatchTextureMode::Disabled));
+  DebugAssert(!(m_texpage_dirty & TEXPAGE_DIRTY_PAGE_RECT) || texture_mode == BatchTextureMode::PageTexture ||
+              !rc.IsTexturingEnabled());
 
   // has any state changed which requires a new batch?
   // Reverse blending breaks with mixed transparent and opaque pixels, so we have to do one draw per polygon.
@@ -3464,12 +3630,16 @@ void GPU_HW::DispatchRenderCommand()
   const GPUTransparencyMode transparency_mode =
     rc.transparency_enable ? m_draw_mode.mode_reg.transparency_mode : GPUTransparencyMode::Disabled;
   const bool dithering_enable = (!m_true_color && rc.IsDitheringEnabled()) ? m_GPUSTAT.dither_enable : false;
+  if (!IsFlushed())
+  {
     if (texture_mode != m_batch.texture_mode || transparency_mode != m_batch.transparency_mode ||
         (transparency_mode == GPUTransparencyMode::BackgroundMinusForeground && !m_allow_shader_blend) ||
-      dithering_enable != m_batch.dithering)
+        dithering_enable != m_batch.dithering ||
+        (texture_mode == BatchTextureMode::PageTexture && m_batch.texture_cache_key != texture_cache_key))
     {
       FlushRender();
     }
+  }
 
   EnsureVertexBufferSpaceForCurrentCommand();
 
@@ -3512,6 +3682,7 @@ void GPU_HW::DispatchRenderCommand()
     m_batch.texture_mode = texture_mode;
     m_batch.transparency_mode = transparency_mode;
     m_batch.dithering = dithering_enable;
+    m_batch.texture_cache_key = texture_cache_key;
 
     if (m_draw_mode.IsTextureWindowChanged())
     {
@@ -3577,10 +3748,21 @@ void GPU_HW::FlushRender()
     return;
 
 #ifdef _DEBUG
-  GL_SCOPE_FMT("Hardware Draw {}", ++s_draw_number);
+  GL_SCOPE_FMT("Hardware Draw {}: {}", ++s_draw_number, m_current_draw_rect);
 #endif
 
   GL_INS_FMT("Dirty draw area: {}", m_vram_dirty_draw_rect);
+  if (m_compute_uv_range)
+    GL_INS_FMT("UV rect: {}", m_current_uv_rect);
+
+  const GPUTextureCache::Source* texture = nullptr;
+  if (m_batch.texture_mode == BatchTextureMode::PageTexture)
+  {
+    texture = LookupSource(m_batch.texture_cache_key, m_current_uv_rect,
+                           m_batch.transparency_mode != GPUTransparencyMode::Disabled ?
+                             GPUTextureCache::PaletteRecordFlags::HasSemiTransparentDraws :
+                             GPUTextureCache::PaletteRecordFlags::None);
+  }
 
   if (m_batch_ubo_dirty)
   {
@@ -3589,21 +3771,24 @@ void GPU_HW::FlushRender()
     m_batch_ubo_dirty = false;
   }
 
+  m_current_draw_rect = INVALID_RECT;
+  m_current_uv_rect = INVALID_RECT;
+
   if (m_wireframe_mode != GPUWireframeMode::OnlyWireframe)
   {
     if (NeedsShaderBlending(m_batch.transparency_mode, m_batch.texture_mode, m_batch.check_mask_before_draw) ||
         m_rov_active || (m_use_rov_for_shader_blend && m_pgxp_depth_buffer))
     {
-      DrawBatchVertices(BatchRenderMode::ShaderBlend, index_count, base_index, base_vertex);
+      DrawBatchVertices(BatchRenderMode::ShaderBlend, index_count, base_index, base_vertex, texture);
     }
     else if (NeedsTwoPassRendering())
     {
-      DrawBatchVertices(BatchRenderMode::OnlyOpaque, index_count, base_index, base_vertex);
+      DrawBatchVertices(BatchRenderMode::OnlyOpaque, index_count, base_index, base_vertex, texture);
-      DrawBatchVertices(BatchRenderMode::OnlyTransparent, index_count, base_index, base_vertex);
+      DrawBatchVertices(BatchRenderMode::OnlyTransparent, index_count, base_index, base_vertex, texture);
     }
     else
     {
-      DrawBatchVertices(m_batch.GetRenderMode(), index_count, base_index, base_vertex);
+      DrawBatchVertices(m_batch.GetRenderMode(), index_count, base_index, base_vertex, texture);
     }
   }
 
@@ -3623,6 +3808,8 @@ void GPU_HW::UpdateDisplay()
 
   GL_SCOPE("UpdateDisplay()");
 
+  GPUTextureCache::Compact();
+
   if (g_settings.debugging.show_vram)
   {
     if (IsUsingMultisampling())

src/core/gpu_hw.h

@@ -4,7 +4,7 @@
 #pragma once
 
 #include "gpu.h"
-#include "texture_replacements.h"
+#include "gpu_hw_texture_cache.h"
 
 #include "util/gpu_device.h"
 
@@ -38,6 +38,7 @@ public:
     Palette4Bit,
     Palette8Bit,
     Direct16Bit,
+    PageTexture,
     Disabled,
 
     SpritePalette4Bit,
@@ -52,6 +53,11 @@ public:
                 static_cast<u8>(BatchTextureMode::Palette8Bit) == static_cast<u8>(GPUTextureMode::Palette8Bit) &&
                 static_cast<u8>(BatchTextureMode::Direct16Bit) == static_cast<u8>(GPUTextureMode::Direct16Bit));
 
+  static constexpr GSVector4i VRAM_SIZE_RECT = GSVector4i::cxpr(0, 0, VRAM_WIDTH, VRAM_HEIGHT);
+  static constexpr GSVector4i INVALID_RECT =
+    GSVector4i::cxpr(std::numeric_limits<s32>::max(), std::numeric_limits<s32>::max(), std::numeric_limits<s32>::min(),
+                     std::numeric_limits<s32>::min());
+
   GPU_HW();
   ~GPU_HW() override;
 
@@ -83,6 +89,8 @@ private:
   {
     TEXPAGE_DIRTY_DRAWN_RECT = (1 << 0),
     TEXPAGE_DIRTY_WRITTEN_RECT = (1 << 1),
+    TEXPAGE_DIRTY_PAGE_RECT = (1 << 2),
+    TEXPAGE_DIRTY_ONLY_UV_RECT = (1 << 3),
   };
 
   static_assert(GPUDevice::MIN_TEXEL_BUFFER_ELEMENTS >= (VRAM_WIDTH * VRAM_HEIGHT));
@@ -116,6 +124,8 @@ private:
     bool use_depth_buffer = false;
     bool sprite_mode = false;
 
+    GPUTextureCache::SourceKey texture_cache_key = {};
+
     // Returns the render mode for this batch.
     BatchRenderMode GetRenderMode() const;
   };
@@ -140,11 +150,6 @@ private:
     u32 num_uniform_buffer_updates;
   };
 
-  static constexpr GSVector4i VRAM_SIZE_RECT = GSVector4i::cxpr(0, 0, VRAM_WIDTH, VRAM_HEIGHT);
-  static constexpr GSVector4i INVALID_RECT =
-    GSVector4i::cxpr(std::numeric_limits<s32>::max(), std::numeric_limits<s32>::max(), std::numeric_limits<s32>::min(),
-                     std::numeric_limits<s32>::min());
-
   /// Returns true if a depth buffer should be created.
   GPUTexture::Format GetDepthBufferFormat() const;
 
@@ -169,7 +174,8 @@ private:
   void DeactivateROV();
   void MapGPUBuffer(u32 required_vertices, u32 required_indices);
   void UnmapGPUBuffer(u32 used_vertices, u32 used_indices);
-  void DrawBatchVertices(BatchRenderMode render_mode, u32 num_indices, u32 base_index, u32 base_vertex);
+  void DrawBatchVertices(BatchRenderMode render_mode, u32 num_indices, u32 base_index, u32 base_vertex,
+                         const GPUTextureCache::Source* texture);
 
   u32 CalculateResolutionScale() const;
   GPUDownsampleMode GetDownsampleMode(u32 resolution_scale) const;
@@ -186,6 +192,7 @@ private:
   void SetTexPageChangedOnOverlap(const GSVector4i update_rect);
 
   void CheckForTexPageOverlap(GSVector4i uv_rect);
+  bool ShouldCheckForTexPageOverlap() const;
 
   bool IsFlushed() const;
   void EnsureVertexBufferSpace(u32 required_vertices, u32 required_indices);
@@ -217,7 +224,7 @@ private:
 
   void UpdateVRAMOnGPU(u32 x, u32 y, u32 width, u32 height, const void* data, u32 data_pitch, bool set_mask,
                        bool check_mask, const GSVector4i bounds);
-  bool BlitVRAMReplacementTexture(const TextureReplacements::ReplacementImage* tex, u32 dst_x, u32 dst_y, u32 width,
+  bool BlitVRAMReplacementTexture(const GPUTextureCache::TextureReplacementImage* tex, u32 dst_x, u32 dst_y, u32 width,
                                   u32 height);
 
   /// Expands a line into two triangles.
@@ -290,6 +297,9 @@ private:
   bool m_texture_window_active : 1 = false;
   bool m_rov_active : 1 = false;
 
+  bool m_use_texture_cache : 1 = false;
+  bool m_texture_dumping : 1 = false;
+
   u8 m_texpage_dirty = 0;
 
   BatchConfig m_batch;
@@ -300,8 +310,9 @@ private:
 
   // Bounding box of VRAM area that the GPU has drawn into.
   GSVector4i m_vram_dirty_draw_rect = INVALID_RECT;
-  GSVector4i m_vram_dirty_write_rect = INVALID_RECT;
+  GSVector4i m_vram_dirty_write_rect = INVALID_RECT; // TODO: Don't use in TC mode, should be kept at zero.
   GSVector4i m_current_uv_rect = INVALID_RECT;
+  GSVector4i m_current_draw_rect = INVALID_RECT;
   s32 m_current_texture_page_offset[2] = {};
 
   std::unique_ptr<GPUPipeline> m_wireframe_pipeline;

src/core/gpu_hw_shadergen.cpp

@@ -55,13 +55,14 @@ void GPU_HW_ShaderGen::WriteBatchUniformBuffer(std::stringstream& ss)
                        false);
 }
 
-std::string GPU_HW_ShaderGen::GenerateBatchVertexShader(bool textured, bool palette, bool uv_limits,
+std::string GPU_HW_ShaderGen::GenerateBatchVertexShader(bool textured, bool palette, bool page_texture, bool uv_limits,
                                                         bool force_round_texcoords, bool pgxp_depth)
 {
   std::stringstream ss;
   WriteHeader(ss);
   DefineMacro(ss, "TEXTURED", textured);
   DefineMacro(ss, "PALETTE", palette);
+  DefineMacro(ss, "PAGE_TEXTURE", page_texture);
   DefineMacro(ss, "UV_LIMITS", uv_limits);
   DefineMacro(ss, "FORCE_ROUND_TEXCOORDS", force_round_texcoords);
   DefineMacro(ss, "PGXP_DEPTH", pgxp_depth);
@@ -69,7 +70,22 @@ std::string GPU_HW_ShaderGen::GenerateBatchVertexShader(bool textured, bool pale
 
   WriteBatchUniformBuffer(ss);
 
-  if (textured)
+  if (textured && page_texture)
+  {
+    if (uv_limits)
+    {
+      DeclareVertexEntryPoint(
+        ss, {"float4 a_pos", "float4 a_col0", "uint a_texcoord", "uint a_texpage", "float4 a_uv_limits"}, 1, 1,
+        {{"nointerpolation", "float4 v_uv_limits"}}, false, "", UsingMSAA(), UsingPerSampleShading(),
+        m_disable_color_perspective);
+    }
+    else
+    {
+      DeclareVertexEntryPoint(ss, {"float4 a_pos", "float4 a_col0", "uint a_texcoord", "uint a_texpage"}, 1, 1, {},
+                              false, "", UsingMSAA(), UsingPerSampleShading(), m_disable_color_perspective);
+    }
+  }
+  else if (textured)
   {
     if (uv_limits)
     {
@@ -127,10 +143,11 @@ std::string GPU_HW_ShaderGen::GenerateBatchVertexShader(bool textured, bool pale
   v_col0 = a_col0;
   #if TEXTURED
     v_tex0 = float2(uint2(a_texcoord & 0xFFFFu, a_texcoord >> 16));
-    #if !PALETTE
+    #if !PALETTE && !PAGE_TEXTURE
       v_tex0 *= u_resolution_scale;
     #endif
 
+    #if !PAGE_TEXTURE
       // base_x,base_y,palette_x,palette_y
       v_texpage.x = (a_texpage & 15u) * 64u;
       v_texpage.y = ((a_texpage >> 4) & 1u) * 256u;
@@ -138,6 +155,7 @@ std::string GPU_HW_ShaderGen::GenerateBatchVertexShader(bool textured, bool pale
         v_texpage.z = ((a_texpage >> 16) & 63u) * 16u;
         v_texpage.w = ((a_texpage >> 22) & 511u);
       #endif
+    #endif
 
     #if UV_LIMITS
       v_uv_limits = a_uv_limits * 255.0;
@@ -146,7 +164,7 @@ std::string GPU_HW_ShaderGen::GenerateBatchVertexShader(bool textured, bool pale
         // Add 0.5 to the upper bounds when upscaling, to work around interpolation differences.
         // Limited to force-round-texcoord hack, to avoid breaking other games.
         v_uv_limits.zw += 0.5;
-      #elif !PALETTE
+      #elif !PAGE_TEXTURE && !PALETTE
         // Treat coordinates as being in upscaled space, and extend the UV range to all "upscaled"
         // pixels. This means 1-pixel-high polygon-based framebuffer effects won't be downsampled.
         // (e.g. Mega Man Legends 2 haze effect)
@@ -707,6 +725,7 @@ std::string GPU_HW_ShaderGen::GenerateBatchFragmentShader(
   const bool textured = (texture_mode != GPU_HW::BatchTextureMode::Disabled);
   const bool palette =
     (texture_mode == GPU_HW::BatchTextureMode::Palette4Bit || texture_mode == GPU_HW::BatchTextureMode::Palette8Bit);
+  const bool page_texture = (texture_mode == GPU_HW::BatchTextureMode::PageTexture);
   const bool shader_blending = (render_mode == GPU_HW::BatchRenderMode::ShaderBlend);
   const bool use_dual_source = (!shader_blending && !use_rov && m_supports_dual_source_blend &&
                                 ((render_mode != GPU_HW::BatchRenderMode::TransparencyDisabled &&
@@ -725,6 +744,7 @@ std::string GPU_HW_ShaderGen::GenerateBatchFragmentShader(
   DefineMacro(ss, "PALETTE", palette);
   DefineMacro(ss, "PALETTE_4_BIT", texture_mode == GPU_HW::BatchTextureMode::Palette4Bit);
   DefineMacro(ss, "PALETTE_8_BIT", texture_mode == GPU_HW::BatchTextureMode::Palette8Bit);
+  DefineMacro(ss, "PAGE_TEXTURE", page_texture);
   DefineMacro(ss, "DITHERING", dithering);
   DefineMacro(ss, "DITHERING_SCALED", m_scaled_dithering);
   DefineMacro(ss, "INTERLACING", interlacing);
@@ -804,6 +824,8 @@ uint2 FloatToIntegerCoords(float2 coords)
   return uint2((UPSCALED == 0 || FORCE_ROUND_TEXCOORDS != 0) ? roundEven(coords) : floor(coords));
 }
 
+#if !PAGE_TEXTURE
+
 float4 SampleFromVRAM(TEXPAGE_VALUE texpage, float2 coords)
 {
   #if PALETTE
@@ -855,11 +877,43 @@ float4 SampleFromVRAM(TEXPAGE_VALUE texpage, float2 coords)
   #endif
 }
 
+#else
+
+float4 SampleFromPageTexture(float2 coords)
+{
+  // Cached textures.
+#if UPSCALED == 0 || FORCE_ROUND_TEXCOORDS != 0
+  float2 fpart = coords - roundEven(coords);
+#else
+  float2 fpart = frac(coords);
 #endif
+  uint2 icoord = ApplyTextureWindow(FloatToIntegerCoords(coords));
+  coords = (float2(icoord) + fpart) * (1.0f / 256.0f);
+  return SAMPLE_TEXTURE(samp0, coords);
+}
+
+#endif
+
+#endif // TEXTURED
 )";
 
   const u32 num_fragment_outputs = use_rov ? 0 : (use_dual_source ? 2 : 1);
-  if (textured)
+  if (textured && page_texture)
+  {
+    if (uv_limits)
+    {
+      DeclareFragmentEntryPoint(ss, 1, 1, {{"nointerpolation", "float4 v_uv_limits"}}, true, num_fragment_outputs,
+                                use_dual_source, m_write_mask_as_depth, UsingMSAA(), UsingPerSampleShading(), false,
+                                m_disable_color_perspective, shader_blending && !use_rov, use_rov);
+    }
+    else
+    {
+      DeclareFragmentEntryPoint(ss, 1, 1, {}, true, num_fragment_outputs, use_dual_source, m_write_mask_as_depth,
+                                UsingMSAA(), UsingPerSampleShading(), false, m_disable_color_perspective,
+                                shader_blending && !use_rov, use_rov);
+    }
+  }
+  else if (textured)
   {
     if (texture_filtering != GPUTextureFilter::Nearest)
       WriteBatchTextureFilter(ss, texture_filtering);
@@ -905,7 +959,17 @@ float4 SampleFromVRAM(TEXPAGE_VALUE texpage, float2 coords)
 
   #if TEXTURED
     float4 texcol;
-    #if TEXTURE_FILTERING
+    #if PAGE_TEXTURE
+      #if UV_LIMITS
+        texcol = SampleFromPageTexture(clamp(v_tex0, v_uv_limits.xy, v_uv_limits.zw));
+      #else
+        texcol = SampleFromPageTexture(v_tex0);
+      #endif
+      if (VECTOR_EQ(texcol, TRANSPARENT_PIXEL_COLOR))
+        discard;
+
+      ialpha = 1.0;
+    #elif TEXTURE_FILTERING
       FilteredSampleFromVRAM(v_texpage, v_tex0, v_uv_limits, texcol, ialpha);
       if (ialpha < 0.5)
         discard;
@@ -1687,3 +1751,33 @@ std::string GPU_HW_ShaderGen::GenerateBoxSampleDownsampleFragmentShader(u32 fact
 
   return ss.str();
 }
+
+std::string GPU_HW_ShaderGen::GenerateReplacementMergeFragmentShader(bool semitransparent)
+{
+  std::stringstream ss;
+  WriteHeader(ss);
+  DefineMacro(ss, "SEMITRANSPARENT", semitransparent);
+  DeclareUniformBuffer(ss, {"float4 u_src_rect"}, true);
+  DeclareTexture(ss, "samp0", 0);
+  DeclareFragmentEntryPoint(ss, 0, 1);
+
+  ss << R"(
+{
+  float2 coords = u_src_rect.xy + v_tex0 * u_src_rect.zw;
+  float4 color = SAMPLE_TEXTURE(samp0, coords);
+  o_col0.rgb = color.rgb;
+
+  // Alpha processing.
+  #if SEMITRANSPARENT
+    // Map anything not 255 to 1 for semitransparent, otherwise zero for opaque.
+    o_col0.a = (color.a <= 0.95f) ? 1.0f : 0.0f;
+    o_col0.a = VECTOR_EQ(color, float4(0.0, 0.0, 0.0, 0.0)) ? 0.0f : o_col0.a;
+  #else
+    // Leave (0,0,0,0) as 0000 for opaque replacements for cutout alpha.
+    o_col0.a = color.a;
+  #endif
+}
+)";
+
+  return ss.str();
+}

src/core/gpu_hw_shadergen.h

@@ -15,8 +15,8 @@ public:
                    bool supports_dual_source_blend, bool supports_framebuffer_fetch);
   ~GPU_HW_ShaderGen();
 
-  std::string GenerateBatchVertexShader(bool textured, bool palette, bool uv_limits, bool force_round_texcoords,
+  std::string GenerateBatchVertexShader(bool textured, bool palette, bool page_texture, bool uv_limits,
-                                        bool pgxp_depth);
+                                        bool force_round_texcoords, bool pgxp_depth);
   std::string GenerateBatchFragmentShader(GPU_HW::BatchRenderMode render_mode, GPUTransparencyMode transparency,
                                           GPU_HW::BatchTextureMode texture_mode, GPUTextureFilter texture_filtering,
                                           bool uv_limits, bool force_round_texcoords, bool dithering, bool interlacing,
@@ -36,6 +36,8 @@ public:
   std::string GenerateAdaptiveDownsampleCompositeFragmentShader();
   std::string GenerateBoxSampleDownsampleFragmentShader(u32 factor);
 
+  std::string GenerateReplacementMergeFragmentShader(bool semitransparent);
+
 private:
   ALWAYS_INLINE bool UsingMSAA() const { return m_multisamples > 1; }
   ALWAYS_INLINE bool UsingPerSampleShading() const { return m_multisamples > 1 && m_per_sample_shading; }

src/core/gpu_hw_texture_cache.h

@@ -0,0 +1,147 @@
+// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
+// SPDX-License-Identifier: CC-BY-NC-ND-4.0
+
+#pragma once
+
+#include "gpu_types.h"
+
+class GPUTexture;
+class RGBA8Image;
+class StateWrapper;
+
+struct Settings;
+
+//////////////////////////////////////////////////////////////////////////
+// Texture Cache
+//////////////////////////////////////////////////////////////////////////
+namespace GPUTextureCache {
+
+/// 4 pages in C16 mode, 2+4 pages in P8 mode, 1+1 pages in P4 mode.
+static constexpr u32 MAX_PAGE_REFS_PER_SOURCE = 6;
+
+static constexpr u32 MAX_PAGE_REFS_PER_WRITE = 32;
+
+enum class PaletteRecordFlags : u32
+{
+  None = 0,
+  HasSemiTransparentDraws = (1 << 0),
+};
+IMPLEMENT_ENUM_CLASS_BITWISE_OPERATORS(PaletteRecordFlags);
+
+using HashType = u64;
+using TextureReplacementImage = RGBA8Image;
+
+struct Source;
+struct HashCacheEntry;
+
+template<typename T>
+struct TList;
+template<typename T>
+struct TListNode;
+
+template<typename T>
+struct TList
+{
+  TListNode<T>* head;
+  TListNode<T>* tail;
+};
+
+template<typename T>
+struct TListNode
+{
+  // why inside itself? because we have 3 lists
+  T* ref;
+  TList<T>* list;
+  TListNode<T>* prev;
+  TListNode<T>* next;
+};
+
+struct SourceKey
+{
+  u8 page;
+  GPUTextureMode mode;
+  GPUTexturePaletteReg palette;
+
+  SourceKey() = default;
+  ALWAYS_INLINE constexpr SourceKey(u8 page_, GPUTexturePaletteReg palette_, GPUTextureMode mode_)
+    : page(page_), mode(mode_), palette(palette_)
+  {
+  }
+  ALWAYS_INLINE constexpr SourceKey(const SourceKey& k) : page(k.page), mode(k.mode), palette(k.palette) {}
+
+  ALWAYS_INLINE bool HasPalette() const { return (mode < GPUTextureMode::Direct16Bit); }
+
+  ALWAYS_INLINE SourceKey& operator=(const SourceKey& k)
+  {
+    page = k.page;
+    mode = k.mode;
+    palette.bits = k.palette.bits;
+    return *this;
+  }
+
+  ALWAYS_INLINE bool operator==(const SourceKey& k) const { return (std::memcmp(&k, this, sizeof(SourceKey)) == 0); }
+  ALWAYS_INLINE bool operator!=(const SourceKey& k) const { return (std::memcmp(&k, this, sizeof(SourceKey)) != 0); }
+};
+static_assert(sizeof(SourceKey) == 4);
+
+// TODO: Pool objects
+struct Source
+{
+  SourceKey key;
+  u32 num_page_refs;
+  GPUTexture* texture;
+  HashCacheEntry* from_hash_cache;
+  GSVector4i texture_rect;
+  GSVector4i palette_rect;
+  HashType texture_hash;
+  HashType palette_hash;
+  GSVector4i active_uv_rect;
+  PaletteRecordFlags palette_record_flags;
+
+  std::array<TListNode<Source>, MAX_PAGE_REFS_PER_SOURCE> page_refs;
+  TListNode<Source> hash_cache_ref;
+};
+
+bool Initialize();
+void UpdateSettings(const Settings& old_settings);
+bool DoState(StateWrapper& sw, bool skip);
+void Shutdown();
+
+void Invalidate();
+
+void AddWrittenRectangle(const GSVector4i rect, bool update_vram_writes = false);
+void AddDrawnRectangle(const GSVector4i rect, const GSVector4i clip_rect);
+
+void CopyVRAM(u32 src_x, u32 src_y, u32 dst_x, u32 dst_y, u32 width, u32 height, bool set_mask, bool check_mask,
+              const GSVector4i src_bounds, const GSVector4i dst_bounds);
+void WriteVRAM(u32 x, u32 y, u32 width, u32 height, const void* data, bool set_mask, bool check_mask,
+                const GSVector4i bounds);
+void UpdateVRAMTrackingState();
+
+const Source* LookupSource(SourceKey key, const GSVector4i uv_rect, PaletteRecordFlags flags);
+
+bool IsPageDrawn(u32 page_index);
+bool IsPageDrawn(u32 page_index, const GSVector4i rect);
+bool IsRectDrawn(const GSVector4i rect);
+bool AreSourcePagesDrawn(SourceKey key, const GSVector4i rect);
+
+void InvalidatePageSources(u32 pn);
+void InvalidatePageSources(u32 pn, const GSVector4i rc);
+void DestroySource(Source* src);
+
+void Compact();
+
+void DecodeTexture(GPUTextureMode mode, const u16* page_ptr, const u16* palette, u32* dest, u32 dest_stride, u32 width,
+                   u32 height);
+HashType HashPartialPalette(GPUTexturePaletteReg palette, GPUTextureMode mode, u32 min, u32 max);
+HashType HashRect(const GSVector4i rc);
+
+void SetGameID(std::string game_id);
+void ReloadTextureReplacements();
+
+// VRAM Write Replacements
+const TextureReplacementImage* GetVRAMReplacement(u32 width, u32 height, const void* pixels);
+void DumpVRAMWrite(u32 width, u32 height, const void* pixels);
+bool ShouldDumpVRAMWrite(u32 width, u32 height);
+
+} // namespace GPUTextureCache

src/core/gpu_sw.cpp

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: CC-BY-NC-ND-4.0
 
 #include "gpu_sw.h"
+#include "gpu_hw_texture_cache.h"
 #include "settings.h"
 #include "system.h"
 
@@ -70,7 +71,11 @@ bool GPU_SW::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_di
   m_backend.Sync(true);
 
   // ignore the host texture for software mode, since we want to save vram here
-  return GPU::DoState(sw, nullptr, update_display);
+  if (!GPU::DoState(sw, nullptr, update_display))
+    return false;
+
+  // need to still call the TC, to toss any data in the state
+  return GPUTextureCache::DoState(sw, true);
 }
 
 void GPU_SW::Reset(bool clear_vram)

src/core/gpu_types.h

@@ -26,7 +26,15 @@ enum : u32
   GPU_MAX_DISPLAY_WIDTH = 720,
   GPU_MAX_DISPLAY_HEIGHT = 576,
 
-  DITHER_MATRIX_SIZE = 4
+  DITHER_MATRIX_SIZE = 4,
+
+  VRAM_PAGE_WIDTH = 64,
+  VRAM_PAGE_HEIGHT = 256,
+  VRAM_PAGES_WIDE = VRAM_WIDTH / VRAM_PAGE_WIDTH,
+  VRAM_PAGES_HIGH = VRAM_HEIGHT / VRAM_PAGE_HEIGHT,
+  VRAM_PAGE_X_MASK = 0xf,  // 16 pages wide
+  VRAM_PAGE_Y_MASK = 0x10, // 2 pages high
+  NUM_VRAM_PAGES = VRAM_PAGES_WIDE * VRAM_PAGES_HIGH,
 };
 
 enum : s32
@@ -61,6 +69,11 @@ enum class GPUTextureMode : u8
 
 IMPLEMENT_ENUM_CLASS_BITWISE_OPERATORS(GPUTextureMode);
 
+ALWAYS_INLINE static constexpr bool TextureModeHasPalette(GPUTextureMode mode)
+{
+  return (mode < GPUTextureMode::Direct16Bit);
+}
+
 enum class GPUTransparencyMode : u8
 {
   HalfBackgroundPlusHalfForeground = 0,
@@ -169,7 +182,7 @@ static constexpr s32 TruncateGPUVertexPosition(s32 x)
 union GPUDrawModeReg
 {
   static constexpr u16 MASK = 0b1111111111111;
-  static constexpr u16 TEXTURE_PAGE_MASK = UINT16_C(0b0000000000011111);
+  static constexpr u16 TEXTURE_MODE_AND_PAGE_MASK = UINT16_C(0b0000000110011111);
 
   // Polygon texpage commands only affect bits 0-8, 11
   static constexpr u16 POLYGON_TEXPAGE_MASK = 0b0000100111111111;
@@ -177,11 +190,9 @@ union GPUDrawModeReg
   // Bits 0..5 are returned in the GPU status register, latched at E1h/polygon draw time.
   static constexpr u32 GPUSTAT_MASK = 0b11111111111;
 
-  static constexpr std::array<u32, 4> texture_page_widths = {
-    {TEXTURE_PAGE_WIDTH / 4, TEXTURE_PAGE_WIDTH / 2, TEXTURE_PAGE_WIDTH, TEXTURE_PAGE_WIDTH}};
-
   u16 bits;
 
+  BitField<u16, u8, 0, 5> texture_page;
   BitField<u16, u8, 0, 4> texture_page_x_base;
   BitField<u16, u8, 4, 1> texture_page_y_base;
   BitField<u16, GPUTransparencyMode, 5, 2> transparency_mode;
@@ -197,15 +208,6 @@ union GPUDrawModeReg
 
   /// Returns true if the texture mode requires a palette.
   ALWAYS_INLINE bool IsUsingPalette() const { return (bits & (2 << 7)) == 0; }
-
-  /// Returns a rectangle comprising the texture page area.
-  ALWAYS_INLINE_RELEASE GSVector4i GetTexturePageRectangle() const
-  {
-    const u32 base_x = GetTexturePageBaseX();
-    const u32 base_y = GetTexturePageBaseY();
-    return GSVector4i(base_x, base_y, base_x + texture_page_widths[static_cast<u8>(texture_mode.GetValue())],
-                      base_y + TEXTURE_PAGE_HEIGHT);
-  }
 };
 
 union GPUTexturePaletteReg
@@ -217,17 +219,8 @@ union GPUTexturePaletteReg
   BitField<u16, u16, 0, 6> x;
   BitField<u16, u16, 6, 9> y;
 
-  ALWAYS_INLINE u32 GetXBase() const { return static_cast<u32>(x) * 16u; }
+  ALWAYS_INLINE constexpr u32 GetXBase() const { return static_cast<u32>(x) * 16u; }
-  ALWAYS_INLINE u32 GetYBase() const { return static_cast<u32>(y); }
+  ALWAYS_INLINE constexpr u32 GetYBase() const { return static_cast<u32>(y); }
-
-  /// Returns a rectangle comprising the texture palette area.
-  ALWAYS_INLINE_RELEASE GSVector4i GetRectangle(GPUTextureMode mode) const
-  {
-    static constexpr std::array<u32, 4> palette_widths = {{16, 256, 0, 0}};
-    const u32 base_x = GetXBase();
-    const u32 base_y = GetYBase();
-    return GSVector4i(base_x, base_y, base_x + palette_widths[static_cast<u8>(mode)], base_y + 1);
-  }
 };
 
 struct GPUTextureWindow
@@ -238,6 +231,119 @@ struct GPUTextureWindow
   u8 or_y;
 };
 
+ALWAYS_INLINE static constexpr u32 VRAMPageIndex(u32 px, u32 py)
+{
+  return ((py * VRAM_PAGES_WIDE) + px);
+}
+ALWAYS_INLINE static constexpr GSVector4i VRAMPageRect(u32 px, u32 py)
+{
+  return GSVector4i::cxpr(px * VRAM_PAGE_WIDTH, py * VRAM_PAGE_HEIGHT, (px + 1) * VRAM_PAGE_WIDTH,
+                          (py + 1) * VRAM_PAGE_HEIGHT);
+}
+ALWAYS_INLINE static constexpr GSVector4i VRAMPageRect(u32 pn)
+{
+  // TODO: Put page rects in a LUT instead?
+  return VRAMPageRect(pn % VRAM_PAGES_WIDE, pn / VRAM_PAGES_WIDE);
+}
+
+ALWAYS_INLINE static constexpr u32 VRAMCoordinateToPage(u32 x, u32 y)
+{
+  return VRAMPageIndex(x / VRAM_PAGE_WIDTH, y / VRAM_PAGE_HEIGHT);
+}
+
+ALWAYS_INLINE static constexpr u32 VRAMPageStartX(u32 pn)
+{
+  return (pn % VRAM_PAGES_WIDE) * VRAM_PAGE_WIDTH;
+}
+
+ALWAYS_INLINE static constexpr u32 VRAMPageStartY(u32 pn)
+{
+  return (pn / VRAM_PAGES_WIDE) * VRAM_PAGE_HEIGHT;
+}
+
+ALWAYS_INLINE static constexpr u8 GetTextureModeShift(GPUTextureMode mode)
+{
+  return ((mode < GPUTextureMode::Direct16Bit) ? (2 - static_cast<u8>(mode)) : 0);
+}
+
+ALWAYS_INLINE static constexpr u32 ApplyTextureModeShift(GPUTextureMode mode, u32 vram_width)
+{
+  return vram_width << GetTextureModeShift(mode);
+}
+
+ALWAYS_INLINE static GSVector4i ApplyTextureModeShift(GPUTextureMode mode, const GSVector4i rect)
+{
+  return rect.sll32(GetTextureModeShift(mode));
+}
+
+ALWAYS_INLINE static constexpr u32 TexturePageCountForMode(GPUTextureMode mode)
+{
+  return ((mode < GPUTextureMode::Direct16Bit) ? (1 + static_cast<u8>(mode)) : 4);
+}
+
+ALWAYS_INLINE static constexpr u32 TexturePageWidthForMode(GPUTextureMode mode)
+{
+  return TEXTURE_PAGE_WIDTH >> GetTextureModeShift(mode);
+}
+
+ALWAYS_INLINE static constexpr bool TexturePageIsWrapping(GPUTextureMode mode, u32 pn)
+{
+  return ((VRAMPageStartX(pn) + TexturePageWidthForMode(mode)) > VRAM_WIDTH);
+}
+
+ALWAYS_INLINE static constexpr u32 PalettePageCountForMode(GPUTextureMode mode)
+{
+  return (mode == GPUTextureMode::Palette4Bit) ? 1 : 4;
+}
+
+ALWAYS_INLINE static constexpr u32 PalettePageNumber(GPUTexturePaletteReg reg)
+{
+  return VRAMCoordinateToPage(reg.GetXBase(), reg.GetYBase());
+}
+
+ALWAYS_INLINE static constexpr GSVector4i GetTextureRect(u32 pn, GPUTextureMode mode)
+{
+  u32 left = VRAMPageStartX(pn);
+  u32 top = VRAMPageStartY(pn);
+  u32 right = left + TexturePageWidthForMode(mode);
+  u32 bottom = top + VRAM_PAGE_HEIGHT;
+  if (right > VRAM_WIDTH) [[unlikely]]
+  {
+    left = 0;
+    right = VRAM_WIDTH;
+  }
+  if (bottom > VRAM_HEIGHT) [[unlikely]]
+  {
+    top = 0;
+    bottom = VRAM_HEIGHT;
+  }
+
+  return GSVector4i::cxpr(left, top, right, bottom);
+}
+
+/// Returns the maximum index for a paletted texture.
+ALWAYS_INLINE static constexpr u32 GetPaletteWidth(GPUTextureMode mode)
+{
+  return (mode == GPUTextureMode::Palette4Bit ? 16 : ((mode == GPUTextureMode::Palette8Bit) ? 256 : 0));
+}
+
+/// Returns a rectangle comprising the texture palette area.
+ALWAYS_INLINE static constexpr GSVector4i GetPaletteRect(GPUTexturePaletteReg palette, GPUTextureMode mode,
+                                                         bool clamp_instead_of_wrapping = false)
+{
+  const u32 width = GetPaletteWidth(mode);
+  u32 left = palette.GetXBase();
+  u32 top = palette.GetYBase();
+  u32 right = left + width;
+  u32 bottom = top + 1;
+  if (right > VRAM_WIDTH) [[unlikely]]
+  {
+    right = VRAM_WIDTH;
+    left = clamp_instead_of_wrapping ? left : 0;
+  }
+  return GSVector4i::cxpr(left, top, right, bottom);
+}
+
 // 4x4 dither matrix.
 static constexpr s32 DITHER_MATRIX[DITHER_MATRIX_SIZE][DITHER_MATRIX_SIZE] = {{-4, +0, -3, +1},  // row 0
                                                                               {+2, -2, +3, -1},  // row 1

src/core/hotkeys.cpp

@@ -7,12 +7,12 @@
 #include "cpu_pgxp.h"
 #include "fullscreen_ui.h"
 #include "gpu.h"
+#include "gpu_hw_texture_cache.h"
 #include "host.h"
 #include "imgui_overlays.h"
 #include "settings.h"
 #include "spu.h"
 #include "system.h"
-#include "texture_replacements.h"
 
 #include "util/gpu_device.h"
 #include "util/input_manager.h"
@@ -22,8 +22,8 @@
 #include "common/file_system.h"
 #include "common/timer.h"
 
-#include "IconsFontAwesome5.h"
 #include "IconsEmoji.h"
+#include "IconsFontAwesome5.h"
 #include "fmt/format.h"
 
 #include <cmath>
@@ -441,7 +441,7 @@ DEFINE_HOTKEY("ReloadTextureReplacements", TRANSLATE_NOOP("Hotkeys", "Graphics")
                 {
                   Host::AddKeyedOSDMessage("ReloadTextureReplacements",
                                            TRANSLATE_STR("OSDMessage", "Texture replacements reloaded."), 10.0f);
-                  TextureReplacements::Reload();
+                  GPUTextureCache::ReloadTextureReplacements();
                 }
               })
 
@@ -554,7 +554,8 @@ DEFINE_HOTKEY("AudioCDAudioMute", TRANSLATE_NOOP("Hotkeys", "Audio"), TRANSLATE_
                 {
                   g_settings.cdrom_mute_cd_audio = !g_settings.cdrom_mute_cd_audio;
                   Host::AddIconOSDMessage(
-                    "AudioControlHotkey", g_settings.cdrom_mute_cd_audio ? ICON_EMOJI_MUTED_SPEAKER : ICON_EMOJI_MEDIUM_VOLUME_SPEAKER,
+                    "AudioControlHotkey",
+                    g_settings.cdrom_mute_cd_audio ? ICON_EMOJI_MUTED_SPEAKER : ICON_EMOJI_MEDIUM_VOLUME_SPEAKER,
                     g_settings.cdrom_mute_cd_audio ? TRANSLATE_STR("OSDMessage", "CD Audio Muted.") :
                                                      TRANSLATE_STR("OSDMessage", "CD Audio Unmuted."),
                     2.0f);

src/core/save_state_version.h

@@ -6,7 +6,7 @@
 #include "common/types.h"
 
 static constexpr u32 SAVE_STATE_MAGIC = 0x43435544;
-static constexpr u32 SAVE_STATE_VERSION = 72;
+static constexpr u32 SAVE_STATE_VERSION = 73;
 static constexpr u32 SAVE_STATE_MINIMUM_VERSION = 42;
 
 static_assert(SAVE_STATE_VERSION >= SAVE_STATE_MINIMUM_VERSION);

src/core/settings.cpp

@@ -228,6 +228,7 @@ void Settings::Load(SettingsInterface& si, SettingsInterface& controller_si)
       si.GetStringValue("GPU", "ForceVideoTiming", GetForceVideoTimingName(DEFAULT_FORCE_VIDEO_TIMING_MODE)).c_str())
       .value_or(DEFAULT_FORCE_VIDEO_TIMING_MODE);
   gpu_widescreen_hack = si.GetBoolValue("GPU", "WidescreenHack", false);
+  gpu_texture_cache = si.GetBoolValue("GPU", "EnableTextureCache", false);
   display_24bit_chroma_smoothing = si.GetBoolValue("GPU", "ChromaSmoothing24Bit", false);
   gpu_pgxp_enable = si.GetBoolValue("GPU", "PGXPEnable", false);
   gpu_pgxp_culling = si.GetBoolValue("GPU", "PGXPCulling", true);
@@ -433,16 +434,43 @@ void Settings::Load(SettingsInterface& si, SettingsInterface& controller_si)
   debugging.show_mdec_state = si.GetBoolValue("Debug", "ShowMDECState");
   debugging.show_dma_state = si.GetBoolValue("Debug", "ShowDMAState");
 
+  texture_replacements.enable_texture_replacements =
+    si.GetBoolValue("TextureReplacements", "EnableTextureReplacements", false);
   texture_replacements.enable_vram_write_replacements =
     si.GetBoolValue("TextureReplacements", "EnableVRAMWriteReplacements", false);
   texture_replacements.preload_textures = si.GetBoolValue("TextureReplacements", "PreloadTextures", false);
+  texture_replacements.dump_textures = si.GetBoolValue("TextureReplacements", "DumpTextures", false);
+  texture_replacements.dump_replaced_textures = si.GetBoolValue("TextureReplacements", "DumpReplacedTextures", true);
   texture_replacements.dump_vram_writes = si.GetBoolValue("TextureReplacements", "DumpVRAMWrites", false);
-  texture_replacements.dump_vram_write_force_alpha_channel =
+
+  texture_replacements.config.dump_texture_pages = si.GetBoolValue("TextureReplacements", "DumpTexturePages", false);
+  texture_replacements.config.dump_full_texture_pages =
+    si.GetBoolValue("TextureReplacements", "DumpFullTexturePages", false);
+  texture_replacements.config.dump_texture_force_alpha_channel =
+    si.GetBoolValue("TextureReplacements", "DumpTextureForceAlphaChannel", false);
+  texture_replacements.config.dump_vram_write_force_alpha_channel =
     si.GetBoolValue("TextureReplacements", "DumpVRAMWriteForceAlphaChannel", true);
-  texture_replacements.dump_vram_write_width_threshold =
+  texture_replacements.config.dump_c16_textures = si.GetBoolValue("TextureReplacements", "DumpC16Textures", false);
-    si.GetIntValue("TextureReplacements", "DumpVRAMWriteWidthThreshold", 128);
+  texture_replacements.config.reduce_palette_range = si.GetBoolValue("TextureReplacements", "ReducePaletteRange", true);
-  texture_replacements.dump_vram_write_height_threshold =
+  texture_replacements.config.convert_copies_to_writes =
-    si.GetIntValue("TextureReplacements", "DumpVRAMWriteHeightThreshold", 128);
+    si.GetBoolValue("TextureReplacements", "ConvertCopiesToWrites", false);
+  texture_replacements.config.replacement_scale_linear_filter =
+    si.GetBoolValue("TextureReplacements", "ReplacementScaleLinearFilter", true);
+
+  texture_replacements.config.max_vram_write_splits = si.GetUIntValue("TextureReplacements", "MaxVRAMWriteSplits", 0u);
+  texture_replacements.config.max_vram_write_coalesce_width =
+    si.GetUIntValue("TextureReplacements", "MaxVRAMWriteCoalesceWidth", 0u);
+  texture_replacements.config.max_vram_write_coalesce_height =
+    si.GetUIntValue("TextureReplacements", "MaxVRAMWriteCoalesceHeight", 0u);
+
+  texture_replacements.config.texture_dump_width_threshold =
+    si.GetUIntValue("TextureReplacements", "DumpTextureWidthThreshold", 16);
+  texture_replacements.config.texture_dump_height_threshold =
+    si.GetUIntValue("TextureReplacements", "DumpTextureHeightThreshold", 16);
+  texture_replacements.config.vram_write_dump_width_threshold =
+    si.GetUIntValue("TextureReplacements", "DumpVRAMWriteWidthThreshold", 128);
+  texture_replacements.config.vram_write_dump_height_threshold =
+    si.GetUIntValue("TextureReplacements", "DumpVRAMWriteHeightThreshold", 128);
 
 #ifdef __ANDROID__
   // Android users are incredibly silly and don't understand that stretch is in the aspect ratio list...
@@ -526,6 +554,7 @@ void Settings::Save(SettingsInterface& si, bool ignore_base) const
   si.SetStringValue("GPU", "WireframeMode", GetGPUWireframeModeName(gpu_wireframe_mode));
   si.SetStringValue("GPU", "ForceVideoTiming", GetForceVideoTimingName(gpu_force_video_timing));
   si.SetBoolValue("GPU", "WidescreenHack", gpu_widescreen_hack);
+  si.SetBoolValue("GPU", "EnableTextureCache", gpu_texture_cache);
   si.SetBoolValue("GPU", "ChromaSmoothing24Bit", display_24bit_chroma_smoothing);
   si.SetBoolValue("GPU", "PGXPEnable", gpu_pgxp_enable);
   si.SetBoolValue("GPU", "PGXPCulling", gpu_pgxp_culling);
@@ -677,16 +706,41 @@ void Settings::Save(SettingsInterface& si, bool ignore_base) const
     si.SetBoolValue("Debug", "ShowDMAState", debugging.show_dma_state);
   }
 
+  si.SetBoolValue("TextureReplacements", "EnableTextureReplacements", texture_replacements.enable_texture_replacements);
   si.SetBoolValue("TextureReplacements", "EnableVRAMWriteReplacements",
                   texture_replacements.enable_vram_write_replacements);
   si.SetBoolValue("TextureReplacements", "PreloadTextures", texture_replacements.preload_textures);
   si.SetBoolValue("TextureReplacements", "DumpVRAMWrites", texture_replacements.dump_vram_writes);
+  si.SetBoolValue("TextureReplacements", "DumpTextures", texture_replacements.dump_textures);
+  si.SetBoolValue("TextureReplacements", "DumpReplacedTextures", texture_replacements.dump_replaced_textures);
+
+  si.SetBoolValue("TextureReplacements", "DumpTexturePages", texture_replacements.config.dump_texture_pages);
+  si.SetBoolValue("TextureReplacements", "DumpFullTexturePages", texture_replacements.config.dump_full_texture_pages);
+  si.SetBoolValue("TextureReplacements", "DumpTextureForceAlphaChannel",
+                  texture_replacements.config.dump_texture_force_alpha_channel);
+
   si.SetBoolValue("TextureReplacements", "DumpVRAMWriteForceAlphaChannel",
-                  texture_replacements.dump_vram_write_force_alpha_channel);
+                  texture_replacements.config.dump_vram_write_force_alpha_channel);
-  si.SetIntValue("TextureReplacements", "DumpVRAMWriteWidthThreshold",
+  si.SetBoolValue("TextureReplacements", "DumpC16Textures", texture_replacements.config.dump_c16_textures);
-                 texture_replacements.dump_vram_write_width_threshold);
+  si.SetBoolValue("TextureReplacements", "ReducePaletteRange", texture_replacements.config.reduce_palette_range);
-  si.SetIntValue("TextureReplacements", "DumpVRAMWriteHeightThreshold",
+  si.SetBoolValue("TextureReplacements", "ConvertCopiesToWrites", texture_replacements.config.convert_copies_to_writes);
-                 texture_replacements.dump_vram_write_height_threshold);
+  si.SetBoolValue("TextureReplacements", "ReplacementScaleLinearFilter",
+                  texture_replacements.config.replacement_scale_linear_filter);
+
+  si.SetUIntValue("TextureReplacements", "MaxVRAMWriteSplits", texture_replacements.config.max_vram_write_splits);
+  si.SetUIntValue("TextureReplacements", "MaxVRAMWriteCoalesceWidth",
+                  texture_replacements.config.max_vram_write_coalesce_width);
+  si.GetUIntValue("TextureReplacements", "MaxVRAMWriteCoalesceHeight",
+                  texture_replacements.config.max_vram_write_coalesce_height);
+
+  si.SetUIntValue("TextureReplacements", "DumpTextureWidthThreshold",
+                  texture_replacements.config.texture_dump_width_threshold);
+  si.SetUIntValue("TextureReplacements", "DumpTextureHeightThreshold",
+                  texture_replacements.config.texture_dump_height_threshold);
+  si.SetUIntValue("TextureReplacements", "DumpVRAMWriteWidthThreshold",
+                  texture_replacements.config.vram_write_dump_width_threshold);
+  si.SetUIntValue("TextureReplacements", "DumpVRAMWriteHeightThreshold",
+                  texture_replacements.config.vram_write_dump_height_threshold);
 }
 
 void Settings::Clear(SettingsInterface& si)
@@ -716,6 +770,146 @@ void Settings::Clear(SettingsInterface& si)
   si.ClearSection("TextureReplacements");
 }
 
+bool Settings::TextureReplacementSettings::Configuration::operator==(const Configuration& rhs) const
+{
+  return (dump_texture_pages == rhs.dump_texture_pages && dump_full_texture_pages == rhs.dump_full_texture_pages &&
+          dump_texture_force_alpha_channel == rhs.dump_texture_force_alpha_channel &&
+          dump_vram_write_force_alpha_channel == rhs.dump_vram_write_force_alpha_channel &&
+          dump_c16_textures == rhs.dump_c16_textures && reduce_palette_range == rhs.reduce_palette_range &&
+          convert_copies_to_writes == rhs.convert_copies_to_writes &&
+          replacement_scale_linear_filter == rhs.replacement_scale_linear_filter &&
+          max_vram_write_splits == rhs.max_vram_write_splits &&
+          max_vram_write_coalesce_width == rhs.max_vram_write_coalesce_width &&
+          max_vram_write_coalesce_height == rhs.max_vram_write_coalesce_height &&
+          texture_dump_width_threshold == rhs.texture_dump_width_threshold &&
+          texture_dump_height_threshold == rhs.texture_dump_height_threshold &&
+          vram_write_dump_width_threshold == rhs.vram_write_dump_width_threshold &&
+          vram_write_dump_height_threshold == rhs.vram_write_dump_height_threshold);
+}
+
+bool Settings::TextureReplacementSettings::Configuration::operator!=(const Configuration& rhs) const
+{
+  return !operator==(rhs);
+}
+
+bool Settings::TextureReplacementSettings::operator==(const TextureReplacementSettings& rhs) const
+{
+  return (enable_texture_replacements == rhs.enable_texture_replacements &&
+          enable_vram_write_replacements == rhs.enable_vram_write_replacements &&
+          preload_textures == rhs.preload_textures && dump_textures == rhs.dump_textures &&
+          dump_replaced_textures == rhs.dump_replaced_textures && dump_vram_writes == rhs.dump_vram_writes &&
+          config == rhs.config);
+}
+
+bool Settings::TextureReplacementSettings::operator!=(const TextureReplacementSettings& rhs) const
+{
+  return !operator==(rhs);
+}
+
+std::string Settings::TextureReplacementSettings::Configuration::ExportToYAML(bool comment) const
+{
+  static constexpr const char CONFIG_TEMPLATE[] = R"(# DuckStation Texture Replacement Configuration
+# This file allows you to set a per-game configuration for the dumping and
+# replacement system, avoiding the need to use the normal per-game settings
+# when moving files to a different computer. It also allows for the definition
+# of texture aliases, for reducing duplicate files.
+#
+# All options are commented out by default. If an option is commented, the user's
+# current setting will be used instead. If an option is defined in this file, it
+# will always take precedence over the user's choice.
+
+# Enables texture page dumping mode.
+# Instead of tracking VRAM writes and attempting to identify the "real" size of
+# textures, create sub-rectangles from pages based on how they are drawn. In
+# most games, this will lead to significant duplication in dumps, and reduce
+# replacement reliability. However, some games are incompatible with write
+# tracking, and must use page mode.
+{}DumpTexturePages: {}
+
+# Dumps full texture pages instead of sub-rectangles.
+# 256x256 pages will be dumped/replaced instead.
+{}DumpFullTexturePages: {}
+
+# Enables the dumping of direct textures (i.e. C16 format).
+# Most games do not use direct textures, and when they do, it is usually for
+# post-processing or FMVs. Ignoring C16 textures typically reduces garbage/false
+# positive texture dumps, however, some games may require it.
+{}DumpC16Textures: {}
+
+# Reduces the size of palettes (i.e. CLUTs) to only those indices that are used.
+# This can help reduce duplication and improve replacement reliability in games
+# that use 8-bit textures, but do not reserve or use the full 1x256 region in
+# video memory for storage of the palette. When replacing textures dumped with
+# this option enabled, CPU usage on the GPU thread does increase trivially,
+# however, generally it is worthwhile for the reliability improvement. Games
+# that require this option include Metal Gear Solid.
+{}ReducePaletteRange: {}
+
+# Converts VRAM copies to VRAM writes, when a copy of performed into a previously
+# tracked VRAM write. This is required for some games that construct animated
+# textures by copying and replacing small portions of the texture with the parts
+# that are animated. Generally this option will cause duplication when dumping,
+# but it is required in some games, such as Final Fantasy VIII.
+{}ConvertCopiesToWrites: {}
+
+# Determines the maximum number of times a VRAM write/upload can be split, before
+# it is discarded and no longer tracked. This is required for games that partially
+# overwrite texture data, such as Gran Turismo.
+{}MaxVRAMWriteSplits: {}
+
+# Determines the maximum size of an incoming VRAM write that will be merged with
+# another write to the left/above of the incoming write. Needed for games that
+# upload textures one line at a time. These games will log "tracking VRAM write
+# of Nx1" repeatedly during loading. If the upload size is 1, then you can set
+# the corresponding maximum coalesce dimension to 1 to merge these uploads,
+# which should enable these textures to be dumped/replaced.
+{}MaxVRAMWriteCoalesceWidth: {}
+{}MaxVRAMWriteCoalesceHeight: {}
+
+# Determines the minimum size of a texture that will be dumped. Textures with a
+# width/height smaller than this value will be ignored.
+{}DumpTextureWidthThreshold: {}
+{}DumpTextureHeightThreshold: {}
+
+# Determines the minimum size of a VRAM write that will be dumped, in background
+# dumping mode. Uploads smaller than this size will be ignored.
+{}DumpVRAMWriteWidthThreshold: {}
+{}DumpVRAMWriteHeightThreshold: {}
+
+# Enables the use of a bilinear filter when scaling replacement textures.
+# If more than one replacement texture in a 256x256 texture page has a different
+# scaling over the native resolution, or the texture page is not covered, a
+# bilinear filter will be used to resize/stretch the replacement texture, and/or
+# the original native data.
+{}ReplacementScaleLinearFilter: {}
+
+# Use this section to define replacement aliases. One line per replacement
+# texture, with the key set to the source ID, and the value set to the filename
+# which should be loaded as a replacement. For example, without the newline,
+# or keep the multi-line separator.
+#Aliases:
+  # Alias-Texture-Name: Path-To-Texture
+  #  texupload-P4-AAAAAAAAAAAAAAAA-BBBBBBBBBBBBBBBB-64x256-0-192-64x64-P0-14: |
+  #    texupload-P4-BBBBBBBBBBBBBBBB-BBBBBBBBBBBBBBBB-64x256-0-64-64x64-P0-13.png
+  #  texupload-P4-AAAAAAAAAAAAAAAA-BBBBBBBBBBBBBBBB-64x256-0-192-64x64-P0-14: mytexture.png
+)";
+
+  const std::string_view comment_str = comment ? "#" : "";
+  return fmt::format(CONFIG_TEMPLATE, comment_str, dump_texture_pages, // DumpTexturePages
+                     comment_str, dump_full_texture_pages,             // DumpFullTexturePages
+                     comment_str, dump_c16_textures,                   // DumpC16Textures
+                     comment_str, reduce_palette_range,                // ReducePaletteRange
+                     comment_str, convert_copies_to_writes,            // ConvertCopiesToWrites
+                     comment_str, max_vram_write_splits,               // MaxVRAMWriteSplits
+                     comment_str, max_vram_write_coalesce_width,       // MaxVRAMWriteCoalesceWidth
+                     comment_str, max_vram_write_coalesce_height,      // MaxVRAMWriteCoalesceHeight
+                     comment_str, texture_dump_width_threshold,        // DumpTextureWidthThreshold
+                     comment_str, texture_dump_height_threshold,       // DumpTextureHeightThreshold
+                     comment_str, vram_write_dump_width_threshold,     // DumpVRAMWriteWidthThreshold
+                     comment_str, vram_write_dump_height_threshold,    // DumpVRAMWriteHeightThreshold
+                     comment_str, replacement_scale_linear_filter);    // ReplacementScaleLinearFilter
+}
+
 void Settings::FixIncompatibleSettings(bool display_osd_messages)
 {
   if (g_settings.disable_all_enhancements)
@@ -2043,7 +2237,6 @@ bool EmuFolders::EnsureFoldersExist()
   result = FileSystem::EnsureDirectoryExists(Covers.c_str(), false) && result;
   result = FileSystem::EnsureDirectoryExists(Dumps.c_str(), false) && result;
   result = FileSystem::EnsureDirectoryExists(Path::Combine(Dumps, "audio").c_str(), false) && result;
-  result = FileSystem::EnsureDirectoryExists(Path::Combine(Dumps, "textures").c_str(), false) && result;
   result = FileSystem::EnsureDirectoryExists(GameIcons.c_str(), false) && result;
   result = FileSystem::EnsureDirectoryExists(GameSettings.c_str(), false) && result;
   result = FileSystem::EnsureDirectoryExists(InputProfiles.c_str(), false) && result;

src/core/settings.h

@@ -117,6 +117,7 @@ struct Settings
   bool gpu_force_round_texcoords : 1 = false;
   bool gpu_accurate_blending : 1 = false;
   bool gpu_widescreen_hack : 1 = false;
+  bool gpu_texture_cache : 1 = false;
   bool gpu_pgxp_enable : 1 = false;
   bool gpu_pgxp_culling : 1 = true;
   bool gpu_pgxp_texture_correction : 1 = true;
@@ -239,20 +240,46 @@ struct Settings
   // texture replacements
   struct TextureReplacementSettings
   {
+    struct Configuration
+    {
+      constexpr Configuration() = default;
+
+      bool dump_texture_pages : 1 = false;
+      bool dump_full_texture_pages : 1 = false;
+      bool dump_texture_force_alpha_channel : 1 = false;
+      bool dump_vram_write_force_alpha_channel : 1 = true;
+      bool dump_c16_textures : 1 = false;
+      bool reduce_palette_range : 1 = true;
+      bool convert_copies_to_writes : 1 = false;
+      bool replacement_scale_linear_filter = true;
+
+      u32 max_vram_write_splits = 0;
+      u32 max_vram_write_coalesce_width = 0;
+      u32 max_vram_write_coalesce_height = 0;
+      u32 texture_dump_width_threshold = 16;
+      u32 texture_dump_height_threshold = 16;
+
+      u32 vram_write_dump_width_threshold = 128;
+      u32 vram_write_dump_height_threshold = 128;
+
+      bool operator==(const Configuration& rhs) const;
+      bool operator!=(const Configuration& rhs) const;
+
+      std::string ExportToYAML(bool comment) const;
+    };
+
+    bool enable_texture_replacements : 1 = false;
     bool enable_vram_write_replacements : 1 = false;
     bool preload_textures : 1 = false;
 
+    bool dump_textures : 1 = false;
+    bool dump_replaced_textures : 1 = true;
     bool dump_vram_writes : 1 = false;
-    bool dump_vram_write_force_alpha_channel : 1 = true;
-    u32 dump_vram_write_width_threshold = 128;
-    u32 dump_vram_write_height_threshold = 128;
 
-    ALWAYS_INLINE bool AnyReplacementsEnabled() const { return enable_vram_write_replacements; }
+    Configuration config;
 
-    ALWAYS_INLINE bool ShouldDumpVRAMWrite(u32 width, u32 height)
+    bool operator==(const TextureReplacementSettings& rhs) const;
-    {
+    bool operator!=(const TextureReplacementSettings& rhs) const;
-      return dump_vram_writes && width >= dump_vram_write_width_threshold && height >= dump_vram_write_height_threshold;
-    }
   } texture_replacements;
 
   bool bios_tty_logging : 1 = false;
@@ -345,8 +372,6 @@ struct Settings
     DEFAULT_DMA_HALT_TICKS = 100,
     DEFAULT_GPU_FIFO_SIZE = 16,
     DEFAULT_GPU_MAX_RUN_AHEAD = 128,
-    DEFAULT_VRAM_WRITE_DUMP_WIDTH_THRESHOLD = 128,
-    DEFAULT_VRAM_WRITE_DUMP_HEIGHT_THRESHOLD = 128,
   };
 
   void Load(SettingsInterface& si, SettingsInterface& controller_si);

src/core/system.cpp

@@ -16,6 +16,7 @@
 #include "game_database.h"
 #include "game_list.h"
 #include "gpu.h"
+#include "gpu_hw_texture_cache.h"
 #include "gte.h"
 #include "host.h"
 #include "host_interface_progress_callback.h"
@@ -30,7 +31,6 @@
 #include "save_state_version.h"
 #include "sio.h"
 #include "spu.h"
-#include "texture_replacements.h"
 #include "timers.h"
 
 #include "scmversion/scmversion.h"
@@ -1794,7 +1794,7 @@ bool System::BootSystem(SystemBootParameters parameters, Error* error)
 
   // Texture replacement preloading.
   // TODO: Move this and everything else below OnSystemStarted().
-  TextureReplacements::SetGameID(s_running_game_serial);
+  GPUTextureCache::SetGameID(s_running_game_serial);
 
   // Good to go.
   s_state = State::Running;
@@ -1969,8 +1969,6 @@ void System::DestroySystem()
 
   ClearMemorySaveStates();
 
-  TextureReplacements::Shutdown();
-
   PCDrv::Shutdown();
   SIO::Shutdown();
   MDEC::Shutdown();
@@ -1984,6 +1982,7 @@ void System::DestroySystem()
   CPU::Shutdown();
   Bus::Shutdown();
   TimingEvents::Shutdown();
+  GPUTextureCache::Shutdown();
   ClearRunningGame();
 
   // Restore present-all-frames behavior.
@@ -4078,7 +4077,7 @@ void System::UpdateRunningGame(const std::string_view path, CDImage* image, bool
   }
 
   if (!booting)
-    TextureReplacements::SetGameID(s_running_game_serial);
+    GPUTextureCache::SetGameID(s_running_game_serial);
 
   if (booting)
     Achievements::ResetHardcoreMode(true);
@@ -4389,6 +4388,7 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
         g_settings.gpu_downsample_mode != old_settings.gpu_downsample_mode ||
         g_settings.gpu_downsample_scale != old_settings.gpu_downsample_scale ||
         g_settings.gpu_wireframe_mode != old_settings.gpu_wireframe_mode ||
+        g_settings.gpu_texture_cache != old_settings.gpu_texture_cache ||
         g_settings.display_deinterlacing_mode != old_settings.display_deinterlacing_mode ||
         g_settings.display_24bit_chroma_smoothing != old_settings.display_24bit_chroma_smoothing ||
         g_settings.display_crop_mode != old_settings.display_crop_mode ||
@@ -4404,7 +4404,8 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
         g_settings.display_line_start_offset != old_settings.display_line_start_offset ||
         g_settings.display_line_end_offset != old_settings.display_line_end_offset ||
         g_settings.rewind_enable != old_settings.rewind_enable ||
-        g_settings.runahead_frames != old_settings.runahead_frames)
+        g_settings.runahead_frames != old_settings.runahead_frames ||
+        g_settings.texture_replacements.config != old_settings.texture_replacements.config)
     {
       g_gpu->UpdateSettings(old_settings);
       if (IsPaused())
@@ -4455,13 +4456,6 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
       UpdateMemorySaveStateSettings();
     }
 
-    if (g_settings.texture_replacements.enable_vram_write_replacements !=
-          old_settings.texture_replacements.enable_vram_write_replacements ||
-        g_settings.texture_replacements.preload_textures != old_settings.texture_replacements.preload_textures)
-    {
-      TextureReplacements::Reload();
-    }
-
     if (g_settings.audio_backend != old_settings.audio_backend ||
         g_settings.emulation_speed != old_settings.emulation_speed ||
         g_settings.fast_forward_speed != old_settings.fast_forward_speed ||
@@ -4576,7 +4570,9 @@ void System::WarnAboutUnsafeSettings()
   LargeString messages;
   auto append = [&messages](const char* icon, std::string_view msg) { messages.append_format("{} {}\n", icon, msg); };
 
-  if (!g_settings.disable_all_enhancements && ImGuiManager::IsShowingOSDMessages())
+  if (!g_settings.disable_all_enhancements)
+  {
+    if (ImGuiManager::IsShowingOSDMessages())
     {
       if (g_settings.cpu_overclock_active)
       {
@@ -4603,7 +4599,8 @@ void System::WarnAboutUnsafeSettings()
       }
       if (g_settings.gpu_force_video_timing != ForceVideoTimingMode::Disabled)
       {
-      append(ICON_FA_TV, TRANSLATE_SV("System", "Force frame timings is enabled. Games may run at incorrect speeds."));
+        append(ICON_FA_TV,
+               TRANSLATE_SV("System", "Force frame timings is enabled. Games may run at incorrect speeds."));
       }
       if (!g_settings.IsUsingSoftwareRenderer())
       {
@@ -4626,6 +4623,16 @@ void System::WarnAboutUnsafeSettings()
       }
     }
 
+    // Always display TC warning.
+    if (g_settings.gpu_texture_cache)
+    {
+      append(
+        ICON_FA_PAINT_ROLLER,
+        TRANSLATE_SV("System",
+                     "Texture cache is enabled. This feature is experimental, some games may not render correctly."));
+    }
+  }
+
   if (g_settings.disable_all_enhancements)
   {
     append(ICON_EMOJI_WARNING, TRANSLATE_SV("System", "Safe mode is enabled."));

src/core/texture_replacements.cpp

@@ -1,335 +0,0 @@
-// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
-// SPDX-License-Identifier: CC-BY-NC-ND-4.0
-
-#include "texture_replacements.h"
-#include "gpu_types.h"
-#include "host.h"
-#include "settings.h"
-
-#include "common/bitutils.h"
-#include "common/file_system.h"
-#include "common/hash_combine.h"
-#include "common/log.h"
-#include "common/path.h"
-#include "common/string_util.h"
-#include "common/timer.h"
-
-#include "fmt/format.h"
-#include "xxhash.h"
-#if defined(CPU_ARCH_X86) || defined(CPU_ARCH_X64)
-#include "xxh_x86dispatch.h"
-#endif
-
-#include <cinttypes>
-#include <tuple>
-#include <unordered_map>
-#include <vector>
-
-LOG_CHANNEL(TextureReplacements);
-
-namespace TextureReplacements {
-namespace {
-struct VRAMReplacementHash
-{
-  u64 low;
-  u64 high;
-
-  TinyString ToString() const;
-  bool ParseString(std::string_view sv);
-
-  bool operator<(const VRAMReplacementHash& rhs) const { return std::tie(low, high) < std::tie(rhs.low, rhs.high); }
-  bool operator==(const VRAMReplacementHash& rhs) const { return low == rhs.low && high == rhs.high; }
-  bool operator!=(const VRAMReplacementHash& rhs) const { return low != rhs.low || high != rhs.high; }
-};
-
-struct VRAMReplacementHashMapHash
-{
-  size_t operator()(const VRAMReplacementHash& hash) const;
-};
-} // namespace
-
-using VRAMWriteReplacementMap = std::unordered_map<VRAMReplacementHash, std::string, VRAMReplacementHashMapHash>;
-using TextureCache = std::unordered_map<std::string, ReplacementImage>;
-
-static bool ParseReplacementFilename(const std::string& filename, VRAMReplacementHash* replacement_hash,
-                                     ReplacmentType* replacement_type);
-
-static std::string GetSourceDirectory();
-static std::string GetDumpDirectory();
-
-static VRAMReplacementHash GetVRAMWriteHash(u32 width, u32 height, const void* pixels);
-static std::string GetVRAMWriteDumpFilename(u32 width, u32 height, const void* pixels);
-
-static void FindTextures(const std::string& dir);
-
-static const ReplacementImage* LoadTexture(const std::string& filename);
-static void PreloadTextures();
-static void PurgeUnreferencedTexturesFromCache();
-
-static std::string s_game_id;
-
-// TODO: Check the size, purge some when it gets too large.
-static TextureCache s_texture_cache;
-
-static VRAMWriteReplacementMap s_vram_write_replacements;
-} // namespace TextureReplacements
-
-size_t TextureReplacements::VRAMReplacementHashMapHash::operator()(const VRAMReplacementHash& hash) const
-{
-  size_t hash_hash = std::hash<u64>{}(hash.low);
-  hash_combine(hash_hash, hash.high);
-  return hash_hash;
-}
-
-TinyString TextureReplacements::VRAMReplacementHash::ToString() const
-{
-  return TinyString::from_format("{:08X}{:08X}", high, low);
-}
-
-bool TextureReplacements::VRAMReplacementHash::ParseString(std::string_view sv)
-{
-  if (sv.length() != 32)
-    return false;
-
-  std::optional<u64> high_value = StringUtil::FromChars<u64>(sv.substr(0, 16), 16);
-  std::optional<u64> low_value = StringUtil::FromChars<u64>(sv.substr(16), 16);
-  if (!high_value.has_value() || !low_value.has_value())
-    return false;
-
-  low = low_value.value();
-  high = high_value.value();
-  return true;
-}
-
-void TextureReplacements::SetGameID(std::string game_id)
-{
-  if (s_game_id == game_id)
-    return;
-
-  s_game_id = game_id;
-  Reload();
-}
-
-const TextureReplacements::ReplacementImage* TextureReplacements::GetVRAMReplacement(u32 width, u32 height,
-                                                                                     const void* pixels)
-{
-  const VRAMReplacementHash hash = GetVRAMWriteHash(width, height, pixels);
-
-  const auto it = s_vram_write_replacements.find(hash);
-  if (it == s_vram_write_replacements.end())
-    return nullptr;
-
-  return LoadTexture(it->second);
-}
-
-void TextureReplacements::DumpVRAMWrite(u32 width, u32 height, const void* pixels)
-{
-  const std::string filename = GetVRAMWriteDumpFilename(width, height, pixels);
-  if (filename.empty())
-    return;
-
-  RGBA8Image image;
-  image.SetSize(width, height);
-
-  const u16* src_pixels = reinterpret_cast<const u16*>(pixels);
-
-  for (u32 y = 0; y < height; y++)
-  {
-    for (u32 x = 0; x < width; x++)
-    {
-      image.SetPixel(x, y, VRAMRGBA5551ToRGBA8888(*src_pixels));
-      src_pixels++;
-    }
-  }
-
-  if (g_settings.texture_replacements.dump_vram_write_force_alpha_channel)
-  {
-    for (u32 y = 0; y < height; y++)
-    {
-      for (u32 x = 0; x < width; x++)
-        image.SetPixel(x, y, image.GetPixel(x, y) | 0xFF000000u);
-    }
-  }
-
-  INFO_LOG("Dumping {}x{} VRAM write to '{}'", width, height, Path::GetFileName(filename));
-  if (!image.SaveToFile(filename.c_str())) [[unlikely]]
-    ERROR_LOG("Failed to dump {}x{} VRAM write to '{}'", width, height, filename);
-}
-
-void TextureReplacements::Shutdown()
-{
-  s_texture_cache.clear();
-  s_vram_write_replacements.clear();
-  s_game_id.clear();
-}
-
-// TODO: Organize into PCSX2-style.
-std::string TextureReplacements::GetSourceDirectory()
-{
-  return Path::Combine(EmuFolders::Textures, s_game_id);
-}
-
-std::string TextureReplacements::GetDumpDirectory()
-{
-  return Path::Combine(EmuFolders::Dumps, Path::Combine("textures", s_game_id));
-}
-
-TextureReplacements::VRAMReplacementHash TextureReplacements::GetVRAMWriteHash(u32 width, u32 height,
-                                                                               const void* pixels)
-{
-  XXH128_hash_t hash = XXH3_128bits(pixels, width * height * sizeof(u16));
-  return {hash.low64, hash.high64};
-}
-
-std::string TextureReplacements::GetVRAMWriteDumpFilename(u32 width, u32 height, const void* pixels)
-{
-  if (s_game_id.empty())
-    return {};
-
-  const VRAMReplacementHash hash = GetVRAMWriteHash(width, height, pixels);
-  const std::string dump_directory(GetDumpDirectory());
-  std::string filename(Path::Combine(dump_directory, fmt::format("vram-write-{}.png", hash.ToString())));
-
-  if (FileSystem::FileExists(filename.c_str()))
-    return {};
-
-  if (!FileSystem::EnsureDirectoryExists(dump_directory.c_str(), false))
-    return {};
-
-  return filename;
-}
-
-void TextureReplacements::Reload()
-{
-  s_vram_write_replacements.clear();
-
-  if (g_settings.texture_replacements.AnyReplacementsEnabled())
-    FindTextures(GetSourceDirectory());
-
-  if (g_settings.texture_replacements.preload_textures)
-    PreloadTextures();
-
-  PurgeUnreferencedTexturesFromCache();
-}
-
-void TextureReplacements::PurgeUnreferencedTexturesFromCache()
-{
-  TextureCache old_map = std::move(s_texture_cache);
-  s_texture_cache = {};
-
-  for (const auto& it : s_vram_write_replacements)
-  {
-    auto it2 = old_map.find(it.second);
-    if (it2 != old_map.end())
-    {
-      s_texture_cache[it.second] = std::move(it2->second);
-      old_map.erase(it2);
-    }
-  }
-}
-
-bool TextureReplacements::ParseReplacementFilename(const std::string& filename, VRAMReplacementHash* replacement_hash,
-                                                   ReplacmentType* replacement_type)
-{
-  const std::string_view file_title = Path::GetFileTitle(filename);
-  if (!file_title.starts_with("vram-write-"))
-    return false;
-
-  const std::string_view hashpart = file_title.substr(11);
-  if (!replacement_hash->ParseString(hashpart))
-    return false;
-
-  const std::string_view file_extension = Path::GetExtension(filename);
-  bool valid_extension = false;
-  for (const char* test_extension : {"png", "jpg", "webp"})
-  {
-    if (StringUtil::EqualNoCase(file_extension, test_extension))
-    {
-      valid_extension = true;
-      break;
-    }
-  }
-
-  *replacement_type = ReplacmentType::VRAMWrite;
-  return valid_extension;
-}
-
-void TextureReplacements::FindTextures(const std::string& dir)
-{
-  FileSystem::FindResultsArray files;
-  FileSystem::FindFiles(dir.c_str(), "*", FILESYSTEM_FIND_FILES | FILESYSTEM_FIND_RECURSIVE, &files);
-
-  for (FILESYSTEM_FIND_DATA& fd : files)
-  {
-    if (fd.Attributes & FILESYSTEM_FILE_ATTRIBUTE_DIRECTORY)
-      continue;
-
-    VRAMReplacementHash hash;
-    ReplacmentType type;
-    if (!ParseReplacementFilename(fd.FileName, &hash, &type))
-      continue;
-
-    switch (type)
-    {
-      case ReplacmentType::VRAMWrite:
-      {
-        auto it = s_vram_write_replacements.find(hash);
-        if (it != s_vram_write_replacements.end())
-        {
-          WARNING_LOG("Duplicate VRAM write replacement: '{}' and '{}'", it->second, fd.FileName);
-          continue;
-        }
-
-        s_vram_write_replacements.emplace(hash, std::move(fd.FileName));
-      }
-      break;
-    }
-  }
-
-  INFO_LOG("Found {} replacement VRAM writes for '{}'", s_vram_write_replacements.size(), s_game_id);
-}
-
-const TextureReplacements::ReplacementImage* TextureReplacements::LoadTexture(const std::string& filename)
-{
-  auto it = s_texture_cache.find(filename);
-  if (it != s_texture_cache.end())
-    return &it->second;
-
-  RGBA8Image image;
-  if (!image.LoadFromFile(filename.c_str()))
-  {
-    ERROR_LOG("Failed to load '{}'", Path::GetFileName(filename));
-    return nullptr;
-  }
-
-  INFO_LOG("Loaded '{}': {}x{}", Path::GetFileName(filename), image.GetWidth(), image.GetHeight());
-  it = s_texture_cache.emplace(filename, std::move(image)).first;
-  return &it->second;
-}
-
-void TextureReplacements::PreloadTextures()
-{
-  static constexpr float UPDATE_INTERVAL = 1.0f;
-
-  Common::Timer last_update_time;
-  u32 num_textures_loaded = 0;
-  const u32 total_textures = static_cast<u32>(s_vram_write_replacements.size());
-
-#define UPDATE_PROGRESS()                                                                                              \
-  if (last_update_time.GetTimeSeconds() >= UPDATE_INTERVAL)                                                            \
-  {                                                                                                                    \
-    Host::DisplayLoadingScreen("Preloading replacement textures...", 0, static_cast<int>(total_textures),              \
-                               static_cast<int>(num_textures_loaded));                                                 \
-    last_update_time.Reset();                                                                                          \
-  }
-
-  for (const auto& it : s_vram_write_replacements)
-  {
-    UPDATE_PROGRESS();
-
-    LoadTexture(it.second);
-    num_textures_loaded++;
-  }
-
-#undef UPDATE_PROGRESS
-}

src/core/texture_replacements.h

@@ -1,30 +0,0 @@
-// SPDX-FileCopyrightText: 2019-2024 Connor McLaughlin <stenzek@gmail.com>
-// SPDX-License-Identifier: CC-BY-NC-ND-4.0
-
-#pragma once
-
-#include "types.h"
-
-#include "util/image.h"
-
-#include <string>
-
-namespace TextureReplacements {
-
-using ReplacementImage = RGBA8Image;
-
-enum class ReplacmentType
-{
-  VRAMWrite,
-};
-
-void SetGameID(std::string game_id);
-
-void Reload();
-
-const ReplacementImage* GetVRAMReplacement(u32 width, u32 height, const void* pixels);
-void DumpVRAMWrite(u32 width, u32 height, const void* pixels);
-
-void Shutdown();
-
-} // namespace TextureReplacements

src/duckstation-qt/CMakeLists.txt

@@ -143,6 +143,7 @@ set(SRCS
   setupwizarddialog.cpp
   setupwizarddialog.h
   setupwizarddialog.ui
+  texturereplacementsettingsdialog.ui
 )
 
 set(TS_FILES

src/duckstation-qt/duckstation-qt.vcxproj

@@ -339,6 +339,9 @@
     <QtUi Include="selectdiscdialog.ui">
       <FileType>Document</FileType>
     </QtUi>
+    <QtUi Include="texturereplacementsettingsdialog.ui">
+      <FileType>Document</FileType>
+    </QtUi>
     <None Include="translations\duckstation-qt_es-es.ts" />
     <None Include="translations\duckstation-qt_tr.ts" />
   </ItemGroup>

src/duckstation-qt/duckstation-qt.vcxproj.filters

@@ -284,6 +284,7 @@
     <QtUi Include="audiostretchsettingsdialog.ui" />
     <QtUi Include="controllerbindingwidget_justifier.ui" />
     <QtUi Include="selectdiscdialog.ui" />
+    <QtUi Include="texturereplacementsettingsdialog.ui" />
   </ItemGroup>
   <ItemGroup>
     <ResourceCompile Include="duckstation-qt.rc" />

src/duckstation-qt/graphicssettingswidget.cpp

@@ -5,13 +5,21 @@
 #include "qtutils.h"
 #include "settingswindow.h"
 #include "settingwidgetbinder.h"
+#include "ui_texturereplacementsettingsdialog.h"
 
 #include "core/game_database.h"
 #include "core/gpu.h"
 #include "core/settings.h"
 
+#include "util/ini_settings_interface.h"
 #include "util/media_capture.h"
 
+#include "common/error.h"
+
+#include <QtCore/QDir>
+#include <QtWidgets/QDialog>
+#include <QtWidgets/QFileDialog>
+#include <QtWidgets/QInputDialog>
 #include <algorithm>
 
 static QVariant GetMSAAModeValue(uint multisamples, bool ssaa)
@@ -232,26 +240,50 @@ GraphicsSettingsWidget::GraphicsSettingsWidget(SettingsWindow* dialog, QWidget*
 
   // Texture Replacements Tab
 
-  SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.vramWriteReplacement, "TextureReplacements",
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.enableTextureCache, "GPU", "EnableTextureCache", false);
-                                               "EnableVRAMWriteReplacements", false);
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.useOldMDECRoutines, "Hacks", "UseOldMDECRoutines", false);
+
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.enableTextureReplacements, "TextureReplacements",
+                                               "EnableTextureReplacements", false);
   SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.preloadTextureReplacements, "TextureReplacements",
                                                "PreloadTextures", false);
-  SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.useOldMDECRoutines, "Hacks", "UseOldMDECRoutines", false);
+
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.enableTextureDumping, "TextureReplacements", "DumpTextures",
+                                               false);
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.dumpReplacedTextures, "TextureReplacements",
+                                               "DumpReplacedTextures", true);
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.vramWriteReplacement, "TextureReplacements",
+                                               "EnableVRAMWriteReplacements", false);
   SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.vramWriteDumping, "TextureReplacements", "DumpVRAMWrites",
                                                false);
-  SettingWidgetBinder::BindWidgetToBoolSetting(sif, m_ui.setVRAMWriteAlphaChannel, "TextureReplacements",
-                                               "DumpVRAMWriteForceAlphaChannel", true);
-  SettingWidgetBinder::BindWidgetToIntSetting(sif, m_ui.minDumpedVRAMWriteWidth, "TextureReplacements",
-                                              "DumpVRAMWriteWidthThreshold",
-                                              Settings::DEFAULT_VRAM_WRITE_DUMP_WIDTH_THRESHOLD);
-  SettingWidgetBinder::BindWidgetToIntSetting(sif, m_ui.minDumpedVRAMWriteHeight, "TextureReplacements",
-                                              "DumpVRAMWriteHeightThreshold",
-                                              Settings::DEFAULT_VRAM_WRITE_DUMP_HEIGHT_THRESHOLD);
 
+  if (!m_dialog->isPerGameSettings())
+  {
+    SettingWidgetBinder::BindWidgetToFolderSetting(sif, m_ui.texturesDirectory, m_ui.texturesDirectoryBrowse,
+                                                   tr("Select Textures Directory"), m_ui.texturesDirectoryOpen,
+                                                   m_ui.texturesDirectoryReset, "Folders", "Textures",
+                                                   Path::Combine(EmuFolders::DataRoot, "textures"));
+  }
+  else
+  {
+    m_ui.tabTextureReplacementsLayout->removeWidget(m_ui.texturesDirectoryGroup);
+    delete m_ui.texturesDirectoryGroup;
+    m_ui.texturesDirectoryGroup = nullptr;
+    m_ui.texturesDirectoryBrowse = nullptr;
+    m_ui.texturesDirectoryOpen = nullptr;
+    m_ui.texturesDirectoryReset = nullptr;
+    m_ui.texturesDirectoryLabel = nullptr;
+    m_ui.texturesDirectory = nullptr;
+  }
+
+  connect(m_ui.enableTextureCache, &QCheckBox::checkStateChanged, this,
+          &GraphicsSettingsWidget::onEnableTextureCacheChanged);
+  connect(m_ui.enableTextureReplacements, &QCheckBox::checkStateChanged, this,
+          &GraphicsSettingsWidget::onEnableAnyTextureReplacementsChanged);
   connect(m_ui.vramWriteReplacement, &QCheckBox::checkStateChanged, this,
           &GraphicsSettingsWidget::onEnableAnyTextureReplacementsChanged);
-  connect(m_ui.vramWriteDumping, &QCheckBox::checkStateChanged, this,
+  connect(m_ui.textureReplacementOptions, &QPushButton::clicked, this,
-          &GraphicsSettingsWidget::onEnableVRAMWriteDumpingChanged);
+          &GraphicsSettingsWidget::onTextureReplacementOptionsClicked);
 
   // Debugging Tab
 
@@ -274,8 +306,8 @@ GraphicsSettingsWidget::GraphicsSettingsWidget(SettingsWindow* dialog, QWidget*
   onMediaCaptureBackendChanged();
   onMediaCaptureAudioEnabledChanged();
   onMediaCaptureVideoEnabledChanged();
+  onEnableTextureCacheChanged();
   onEnableAnyTextureReplacementsChanged();
-  onEnableVRAMWriteDumpingChanged();
   onShowDebugSettingsChanged(QtHost::ShouldShowDebugOptions());
 
   // Rendering Tab
@@ -534,23 +566,26 @@ GraphicsSettingsWidget::GraphicsSettingsWidget(SettingsWindow* dialog, QWidget*
        "separate two pairs from each other.</b><br>For example: \"compression_level = 4 : joint_stereo = 1\""));
 
   // Texture Replacements Tab
-
+  dialog->registerWidgetHelp(m_ui.enableTextureCache, tr("Enable Texture Cache"), tr("Unchecked"),
-  dialog->registerWidgetHelp(m_ui.vramWriteReplacement, tr("Enable VRAM Write Replacement"), tr("Unchecked"),
+                             tr("Enables caching of guest textures, required for texture replacement."));
-                             tr("Enables the replacement of background textures in supported games. <strong>This is "
-                                "not general texture replacement.</strong>"));
-  dialog->registerWidgetHelp(m_ui.preloadTextureReplacements, tr("Preload Texture Replacements"), tr("Unchecked"),
-                             tr("Loads all replacement texture to RAM, reducing stuttering at runtime."));
   dialog->registerWidgetHelp(m_ui.useOldMDECRoutines, tr("Use Old MDEC Routines"), tr("Unchecked"),
                              tr("Enables the older, less accurate MDEC decoding routines. May be required for old "
                                 "replacement backgrounds to match/load."));
-  dialog->registerWidgetHelp(m_ui.setVRAMWriteAlphaChannel, tr("Set Alpha Channel"), tr("Checked"),
+
-                             tr("Clears the mask/transparency bit in VRAM write dumps."));
+  dialog->registerWidgetHelp(m_ui.enableTextureReplacements, tr("Enable Texture Replacements"), tr("Unchecked"),
+                             tr("Enables loading of replacement textures. Not compatible with all games."));
+  dialog->registerWidgetHelp(m_ui.preloadTextureReplacements, tr("Preload Texture Replacements"), tr("Unchecked"),
+                             tr("Loads all replacement texture to RAM, reducing stuttering at runtime."));
+  dialog->registerWidgetHelp(
+    m_ui.enableTextureDumping, tr("Enable Texture Dumping"), tr("Unchecked"),
+    tr("Enables dumping of textures to image files, which can be replaced. Not compatible with all games."));
+  dialog->registerWidgetHelp(m_ui.dumpReplacedTextures, tr("Dump Replaced Textures"), tr("Unchecked"),
+                             tr("Dumps textures that have replacements already loaded."));
+
+  dialog->registerWidgetHelp(m_ui.vramWriteReplacement, tr("Enable VRAM Write Replacement"), tr("Unchecked"),
+                             tr("Enables the replacement of background textures in supported games."));
   dialog->registerWidgetHelp(m_ui.vramWriteDumping, tr("Enable VRAM Write Dumping"), tr("Unchecked"),
                              tr("Writes backgrounds that can be replaced to the dump directory."));
-  dialog->registerWidgetHelp(m_ui.minDumpedVRAMWriteWidth, tr("Dump Size Threshold"), tr("128px"),
-                             tr("Determines the threshold that triggers a VRAM write to be dumped."));
-  dialog->registerWidgetHelp(m_ui.minDumpedVRAMWriteHeight, tr("Dump Size Threshold"), tr("128px"),
-                             tr("Determines the threshold that triggers a VRAM write to be dumped."));
 
   // Debugging Tab
 
@@ -1097,19 +1132,120 @@ void GraphicsSettingsWidget::onMediaCaptureAudioEnabledChanged()
   m_ui.audioCaptureArguments->setEnabled(enabled);
 }
 
+void GraphicsSettingsWidget::onEnableTextureCacheChanged()
+{
+  const bool tc_enabled = m_dialog->getEffectiveBoolValue("GPU", "EnableTextureCache", false);
+  m_ui.enableTextureReplacements->setEnabled(tc_enabled);
+  m_ui.enableTextureDumping->setEnabled(tc_enabled);
+}
+
 void GraphicsSettingsWidget::onEnableAnyTextureReplacementsChanged()
 {
   const bool any_replacements_enabled =
-    m_dialog->getEffectiveBoolValue("TextureReplacements", "EnableVRAMWriteReplacements", false);
+    (m_dialog->getEffectiveBoolValue("TextureReplacements", "EnableVRAMWriteReplacements", false) ||
+     (m_dialog->getEffectiveBoolValue("GPU", "EnableTextureCache", false) &&
+      m_dialog->getEffectiveBoolValue("TextureReplacements", "EnableTextureReplacements", false)));
   m_ui.preloadTextureReplacements->setEnabled(any_replacements_enabled);
 }
 
-void GraphicsSettingsWidget::onEnableVRAMWriteDumpingChanged()
+void GraphicsSettingsWidget::onTextureReplacementOptionsClicked()
 {
-  const bool enabled = m_dialog->getEffectiveBoolValue("TextureReplacements", "DumpVRAMWrites", false);
+  QDialog dlg(QtUtils::GetRootWidget(this));
-  m_ui.setVRAMWriteAlphaChannel->setEnabled(enabled);
+
-  m_ui.minDumpedVRAMWriteWidth->setEnabled(enabled);
+  Ui::TextureReplacementSettingsDialog dlgui;
-  m_ui.minDumpedVRAMWriteHeight->setEnabled(enabled);
+  dlgui.setupUi(&dlg);
-  m_ui.vramWriteDumpThresholdLabel->setEnabled(enabled);
+  dlgui.icon->setPixmap(QIcon::fromTheme(QStringLiteral("image-fill")).pixmap(32, 32));
-  m_ui.vramWriteDumpThresholdSeparator->setEnabled(enabled);
+
+  constexpr Settings::TextureReplacementSettings::Configuration default_replacement_config;
+  SettingsInterface* const sif = m_dialog->getSettingsInterface();
+
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, dlgui.dumpTexturePages, "TextureReplacements", "DumpTexturePages",
+                                               default_replacement_config.dump_texture_pages);
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, dlgui.dumpFullTexturePages, "TextureReplacements",
+                                               "DumpFullTexturePages",
+                                               default_replacement_config.dump_full_texture_pages);
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, dlgui.dumpC16Textures, "TextureReplacements", "DumpC16Textures",
+                                               default_replacement_config.dump_c16_textures);
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, dlgui.reducePaletteRange, "TextureReplacements",
+                                               "ReducePaletteRange", default_replacement_config.reduce_palette_range);
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, dlgui.convertCopiesToWrites, "TextureReplacements",
+                                               "ConvertCopiesToWrites",
+                                               default_replacement_config.convert_copies_to_writes);
+  SettingWidgetBinder::BindWidgetToIntSetting(sif, dlgui.maxVRAMWriteSplits, "TextureReplacements",
+                                              "MaxVRAMWriteSplits", default_replacement_config.max_vram_write_splits);
+  SettingWidgetBinder::BindWidgetToIntSetting(sif, dlgui.maxVRAMWriteCoalesceWidth, "TextureReplacements",
+                                              "MaxVRAMWriteCoalesceWidth",
+                                              default_replacement_config.max_vram_write_coalesce_width);
+  SettingWidgetBinder::BindWidgetToIntSetting(sif, dlgui.maxVRAMWriteCoalesceHeight, "TextureReplacements",
+                                              "MaxVRAMWriteCoalesceHeight",
+                                              default_replacement_config.max_vram_write_coalesce_height);
+  SettingWidgetBinder::BindWidgetToIntSetting(sif, dlgui.minDumpedTextureWidth, "TextureReplacements",
+                                              "DumpTextureWidthThreshold",
+                                              default_replacement_config.texture_dump_width_threshold);
+  SettingWidgetBinder::BindWidgetToIntSetting(sif, dlgui.minDumpedTextureHeight, "TextureReplacements",
+                                              "DumpTextureHeightThreshold",
+                                              default_replacement_config.texture_dump_height_threshold);
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, dlgui.setTextureDumpAlphaChannel, "TextureReplacements",
+                                               "DumpTextureForceAlphaChannel",
+                                               default_replacement_config.dump_texture_force_alpha_channel);
+
+  SettingWidgetBinder::BindWidgetToIntSetting(sif, dlgui.minDumpedVRAMWriteWidth, "TextureReplacements",
+                                              "DumpVRAMWriteWidthThreshold",
+                                              default_replacement_config.vram_write_dump_width_threshold);
+  SettingWidgetBinder::BindWidgetToIntSetting(sif, dlgui.minDumpedVRAMWriteHeight, "TextureReplacements",
+                                              "DumpVRAMWriteHeightThreshold",
+                                              default_replacement_config.vram_write_dump_height_threshold);
+  SettingWidgetBinder::BindWidgetToBoolSetting(sif, dlgui.setVRAMWriteAlphaChannel, "TextureReplacements",
+                                               "DumpVRAMWriteForceAlphaChannel",
+                                               default_replacement_config.dump_vram_write_force_alpha_channel);
+
+  dlgui.dumpFullTexturePages->setEnabled(
+    m_dialog->getEffectiveBoolValue("TextureReplacements", "DumpTexturePages", false));
+  connect(dlgui.dumpTexturePages, &QCheckBox::checkStateChanged, this, [this, full_cb = dlgui.dumpFullTexturePages]() {
+    full_cb->setEnabled(m_dialog->getEffectiveBoolValue("TextureReplacements", "DumpTexturePages", false));
+  });
+  connect(dlgui.closeButton, &QPushButton::clicked, &dlg, &QDialog::accept);
+  connect(dlgui.exportButton, &QPushButton::clicked, &dlg, [&dlg, &dlgui]() {
+    Settings::TextureReplacementSettings::Configuration config;
+
+    config.dump_texture_pages = dlgui.dumpTexturePages->isChecked();
+    config.dump_full_texture_pages = dlgui.dumpFullTexturePages->isChecked();
+    config.dump_c16_textures = dlgui.dumpC16Textures->isChecked();
+    config.reduce_palette_range = dlgui.reducePaletteRange->isChecked();
+    config.convert_copies_to_writes = dlgui.convertCopiesToWrites->isChecked();
+    config.max_vram_write_splits = dlgui.maxVRAMWriteSplits->value();
+    config.max_vram_write_coalesce_width = dlgui.maxVRAMWriteCoalesceWidth->value();
+    config.max_vram_write_coalesce_height = dlgui.maxVRAMWriteCoalesceHeight->value();
+    config.texture_dump_width_threshold = dlgui.minDumpedTextureWidth->value();
+    config.texture_dump_height_threshold = dlgui.minDumpedTextureHeight->value();
+    config.dump_texture_force_alpha_channel = dlgui.setTextureDumpAlphaChannel->isChecked();
+    config.vram_write_dump_width_threshold = dlgui.minDumpedVRAMWriteWidth->value();
+    config.vram_write_dump_height_threshold = dlgui.minDumpedVRAMWriteHeight->value();
+    config.dump_vram_write_force_alpha_channel = dlgui.setTextureDumpAlphaChannel->isChecked();
+
+    QInputDialog idlg(&dlg);
+    idlg.resize(600, 400);
+    idlg.setWindowTitle(tr("Texture Replacement Configuration"));
+    idlg.setInputMode(QInputDialog::TextInput);
+    idlg.setOption(QInputDialog::UsePlainTextEditForTextInput);
+    idlg.setLabelText(tr("Texture Replacement Configuration (config.yaml)"));
+    idlg.setTextValue(QString::fromStdString(config.ExportToYAML(false)));
+    idlg.setOkButtonText(tr("Save"));
+    if (idlg.exec())
+    {
+      const QString path = QFileDialog::getSaveFileName(&dlg, tr("Save Configuration"), QString(),
+                                                        tr("Configuration Files (config.yaml)"));
+      if (path.isEmpty())
+        return;
+
+      Error error;
+      if (!FileSystem::WriteStringToFile(QDir::toNativeSeparators(path).toUtf8().constData(),
+                                         idlg.textValue().toStdString(), &error))
+      {
+        QMessageBox::critical(&dlg, tr("Write Failed"), QString::fromStdString(error.GetDescription()));
+      }
+    }
+  });
+
+  dlg.exec();
 }

src/duckstation-qt/graphicssettingswidget.h

@@ -39,8 +39,9 @@ private Q_SLOTS:
   void onMediaCaptureVideoAutoResolutionChanged();
   void onMediaCaptureAudioEnabledChanged();
 
+  void onEnableTextureCacheChanged();
   void onEnableAnyTextureReplacementsChanged();
-  void onEnableVRAMWriteDumpingChanged();
+  void onTextureReplacementOptionsClicked();
 
 private:
   static constexpr int TAB_INDEX_RENDERING = 0;

src/duckstation-qt/graphicssettingswidget.ui

@@ -1054,9 +1054,9 @@
      </widget>
      <widget class="QWidget" name="tabTextureReplacements">
       <attribute name="title">
-       <string>Texture Replacements</string>
+       <string>Texture Replacement</string>
       </attribute>
-      <layout class="QVBoxLayout" name="verticalLayout_2">
+      <layout class="QVBoxLayout" name="tabTextureReplacementsLayout">
        <property name="leftMargin">
         <number>0</number>
        </property>
@@ -1067,31 +1067,85 @@
         <number>0</number>
        </property>
        <item>
-        <widget class="QGroupBox" name="groupBox_5">
+        <widget class="QGroupBox" name="groupBox_12">
          <property name="title">
           <string>General Settings</string>
          </property>
-         <layout class="QFormLayout" name="formLayout_8">
+         <layout class="QGridLayout" name="gridLayout_9">
-          <item row="0" column="0" colspan="2">
+          <item row="1" column="0">
-           <layout class="QGridLayout" name="gridLayout_6">
+           <widget class="QCheckBox" name="enableTextureCache">
-            <item row="0" column="0">
-             <widget class="QCheckBox" name="vramWriteReplacement">
             <property name="text">
-               <string>Enable VRAM Write Replacement</string>
+             <string>Enable Texture Cache</string>
             </property>
            </widget>
           </item>
-            <item row="0" column="1">
+          <item row="0" column="0" colspan="2">
+           <widget class="QLabel" name="label_4">
+            <property name="text">
+             <string>The texture cache is currently experimental, and may cause rendering errors in some games.</string>
+            </property>
+            <property name="wordWrap">
+             <bool>true</bool>
+            </property>
+           </widget>
+          </item>
+          <item row="1" column="1">
+           <widget class="QCheckBox" name="useOldMDECRoutines">
+            <property name="text">
+             <string>Use Old MDEC Routines</string>
+            </property>
+           </widget>
+          </item>
+         </layout>
+        </widget>
+       </item>
+       <item>
+        <widget class="QGroupBox" name="groupBox_5">
+         <property name="title">
+          <string>Texture Replacement</string>
+         </property>
+         <layout class="QFormLayout" name="formLayout_8">
+          <item row="0" column="0" colspan="2">
+           <layout class="QGridLayout" name="gridLayout_6" columnstretch="1,1">
+            <item row="0" column="0">
+             <widget class="QCheckBox" name="enableTextureReplacements">
+              <property name="text">
+               <string>Enable Texture Replacements</string>
+              </property>
+             </widget>
+            </item>
+            <item row="1" column="0">
              <widget class="QCheckBox" name="preloadTextureReplacements">
               <property name="text">
                <string>Preload Texture Replacements</string>
               </property>
              </widget>
             </item>
-            <item row="1" column="0">
+            <item row="2" column="1">
-             <widget class="QCheckBox" name="useOldMDECRoutines">
+             <layout class="QHBoxLayout" name="horizontalLayout_4"/>
+            </item>
+            <item row="0" column="1">
+             <layout class="QHBoxLayout" name="horizontalLayout_6">
+              <item>
+               <widget class="QCheckBox" name="enableTextureDumping">
                 <property name="text">
-               <string>Use Old MDEC Routines</string>
+                 <string>Enable Texture Dumping</string>
+                </property>
+               </widget>
+              </item>
+              <item>
+               <widget class="QToolButton" name="textureReplacementOptions">
+                <property name="icon">
+                 <iconset theme="settings-3-line"/>
+                </property>
+               </widget>
+              </item>
+             </layout>
+            </item>
+            <item row="1" column="1">
+             <widget class="QCheckBox" name="dumpReplacedTextures">
+              <property name="text">
+               <string>Dump Replaced Textures</string>
               </property>
              </widget>
             </item>
@@ -1103,72 +1157,68 @@
        <item>
         <widget class="QGroupBox" name="groupBox_4">
          <property name="title">
-          <string>VRAM Write Dumping</string>
+          <string>VRAM Write (Background) Replacement</string>
          </property>
          <layout class="QFormLayout" name="formLayout_9">
           <item row="0" column="0" colspan="2">
            <layout class="QGridLayout" name="gridLayout_7">
             <item row="0" column="0">
+             <widget class="QCheckBox" name="vramWriteReplacement">
+              <property name="text">
+               <string>Enable VRAM Write Replacement</string>
+              </property>
+             </widget>
+            </item>
+            <item row="0" column="1">
              <widget class="QCheckBox" name="vramWriteDumping">
               <property name="text">
                <string>Enable VRAM Write Dumping</string>
               </property>
              </widget>
             </item>
-            <item row="0" column="1">
+           </layout>
-             <widget class="QCheckBox" name="setVRAMWriteAlphaChannel">
+          </item>
+         </layout>
+        </widget>
+       </item>
+       <item>
+        <widget class="QGroupBox" name="texturesDirectoryGroup">
+         <property name="title">
+          <string>Textures Directory</string>
+         </property>
+         <layout class="QGridLayout" name="gridLayout_10">
+          <item row="1" column="2">
+           <widget class="QPushButton" name="texturesDirectoryOpen">
             <property name="text">
-               <string>Set Alpha Channel</string>
+             <string>Open...</string>
             </property>
            </widget>
           </item>
-           </layout>
+          <item row="0" column="0" colspan="4">
-          </item>
+           <widget class="QLabel" name="texturesDirectoryLabel">
-          <item row="1" column="0">
-           <widget class="QLabel" name="vramWriteDumpThresholdLabel">
             <property name="text">
-             <string>Dump Size Threshold:</string>
+             <string>Directory to load replacement textures from, and save dumps to.</string>
             </property>
            </widget>
           </item>
           <item row="1" column="1">
-           <layout class="QHBoxLayout" name="horizontalLayout_6" stretch="1,0,1">
+           <widget class="QPushButton" name="texturesDirectoryBrowse">
-            <item>
-             <widget class="QSpinBox" name="minDumpedVRAMWriteWidth">
-              <property name="suffix">
-               <string>px</string>
-              </property>
-              <property name="minimum">
-               <number>1</number>
-              </property>
-              <property name="maximum">
-               <number>1024</number>
-              </property>
-             </widget>
-            </item>
-            <item>
-             <widget class="QLabel" name="vramWriteDumpThresholdSeparator">
             <property name="text">
-               <string>x</string>
+             <string>Browse...</string>
             </property>
            </widget>
           </item>
-            <item>
+          <item row="1" column="0">
-             <widget class="QSpinBox" name="minDumpedVRAMWriteHeight">
+           <widget class="QLineEdit" name="texturesDirectory"/>
-              <property name="suffix">
+          </item>
-               <string>px</string>
+          <item row="1" column="3">
-              </property>
+           <widget class="QPushButton" name="texturesDirectoryReset">
-              <property name="minimum">
+            <property name="text">
-               <number>1</number>
+             <string>Reset</string>
-              </property>
-              <property name="maximum">
-               <number>512</number>
             </property>
            </widget>
           </item>
          </layout>
-          </item>
-         </layout>
         </widget>
        </item>
        <item>

src/duckstation-qt/mainwindow.cpp

@@ -1814,6 +1814,7 @@ void MainWindow::updateEmulationActions(bool starting, bool running, bool cheevo
   m_ui.menuCheats->setDisabled(cheevos_challenge_mode);
   m_ui.actionCPUDebugger->setDisabled(cheevos_challenge_mode);
   m_ui.actionMemoryScanner->setDisabled(cheevos_challenge_mode);
+  m_ui.actionReloadTextureReplacements->setDisabled(starting || !running);
   m_ui.actionDumpRAM->setDisabled(starting || !running || cheevos_challenge_mode);
   m_ui.actionDumpVRAM->setDisabled(starting || !running || cheevos_challenge_mode);
   m_ui.actionDumpSPURAM->setDisabled(starting || !running || cheevos_challenge_mode);
@@ -2095,6 +2096,8 @@ void MainWindow::connectSignals()
   connect(m_ui.actionCPUDebugger, &QAction::triggered, this, &MainWindow::openCPUDebugger);
   SettingWidgetBinder::BindWidgetToBoolSetting(nullptr, m_ui.actionEnableGDBServer, "Debug", "EnableGDBServer", false);
   connect(m_ui.actionOpenDataDirectory, &QAction::triggered, this, &MainWindow::onToolsOpenDataDirectoryTriggered);
+  connect(m_ui.actionOpenTextureDirectory, &QAction::triggered, this, &MainWindow::onToolsOpenTextureDirectoryTriggered);
+  connect(m_ui.actionReloadTextureReplacements, &QAction::triggered, g_emu_thread, &EmuThread::reloadTextureReplacements);
   connect(m_ui.actionMergeDiscSets, &QAction::triggered, m_game_list_widget, &GameListWidget::setMergeDiscSets);
   connect(m_ui.actionShowGameIcons, &QAction::triggered, m_game_list_widget, &GameListWidget::setShowGameIcons);
   connect(m_ui.actionGridViewShowTitles, &QAction::triggered, m_game_list_widget, &GameListWidget::setShowCoverTitles);
@@ -2815,6 +2818,15 @@ void MainWindow::onToolsOpenDataDirectoryTriggered()
   QtUtils::OpenURL(this, QUrl::fromLocalFile(QString::fromStdString(EmuFolders::DataRoot)));
 }
 
+void MainWindow::onToolsOpenTextureDirectoryTriggered()
+{
+  QString dir = QString::fromStdString(EmuFolders::Textures);
+  if (s_system_valid && !s_current_game_serial.isEmpty())
+    dir = QStringLiteral("%1" FS_OSPATH_SEPARATOR_STR "%2").arg(dir).arg(s_current_game_serial);
+
+  QtUtils::OpenURL(this, QUrl::fromLocalFile(dir));
+}
+
 void MainWindow::onSettingsTriggeredFromToolbar()
 {
   if (s_system_valid)

src/duckstation-qt/mainwindow.h

@@ -178,6 +178,7 @@ private Q_SLOTS:
   void onToolsCoverDownloaderTriggered();
   void onToolsMediaCaptureToggled(bool checked);
   void onToolsOpenDataDirectoryTriggered();
+  void onToolsOpenTextureDirectoryTriggered();
   void onSettingsTriggeredFromToolbar();
 
   void onGameListRefreshComplete();

src/duckstation-qt/mainwindow.ui

@@ -208,6 +208,8 @@
     <property name="title">
      <string>&amp;Tools</string>
     </property>
+    <addaction name="actionOpenDataDirectory"/>
+    <addaction name="separator"/>
     <addaction name="actionMemoryCardEditor"/>
     <addaction name="actionCoverDownloader"/>
     <addaction name="separator"/>
@@ -215,7 +217,8 @@
     <addaction name="separator"/>
     <addaction name="actionMediaCapture"/>
     <addaction name="separator"/>
-    <addaction name="actionOpenDataDirectory"/>
+    <addaction name="actionOpenTextureDirectory"/>
+    <addaction name="actionReloadTextureReplacements"/>
    </widget>
    <addaction name="menuSystem"/>
    <addaction name="menuSettings"/>
@@ -917,6 +920,16 @@
     <string>Media Ca&amp;pture</string>
    </property>
   </action>
+  <action name="actionOpenTextureDirectory">
+   <property name="text">
+    <string>Open Texture Directory...</string>
+   </property>
+  </action>
+  <action name="actionReloadTextureReplacements">
+   <property name="text">
+    <string>Reload Texture Replacements</string>
+   </property>
+  </action>
  </widget>
  <resources>
   <include location="resources/duckstation-qt.qrc"/>

src/duckstation-qt/qthost.cpp

@@ -19,6 +19,7 @@
 #include "core/game_list.h"
 #include "core/gdb_server.h"
 #include "core/gpu.h"
+#include "core/gpu_hw_texture_cache.h"
 #include "core/host.h"
 #include "core/imgui_overlays.h"
 #include "core/memory_card.h"
@@ -1351,6 +1352,18 @@ void EmuThread::clearInputBindStateFromSource(InputBindingKey key)
   InputManager::ClearBindStateFromSource(key);
 }
 
+void EmuThread::reloadTextureReplacements()
+{
+  if (!isOnThread())
+  {
+    QMetaObject::invokeMethod(this, "reloadTextureReplacements", Qt::QueuedConnection);
+    return;
+  }
+
+  if (System::IsValid())
+    GPUTextureCache::ReloadTextureReplacements();
+}
+
 void EmuThread::runOnEmuThread(std::function<void()> callback)
 {
   callback();

src/duckstation-qt/qthost.h

@@ -199,6 +199,7 @@ public Q_SLOTS:
   void reloadPostProcessingShaders();
   void updatePostProcessingSettings();
   void clearInputBindStateFromSource(InputBindingKey key);
+  void reloadTextureReplacements();
 
 private Q_SLOTS:
   void stopInThread();

src/duckstation-qt/texturereplacementsettingsdialog.ui

@@ -0,0 +1,380 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<ui version="4.0">
+ <class>TextureReplacementSettingsDialog</class>
+ <widget class="QDialog" name="TextureReplacementSettingsDialog">
+  <property name="geometry">
+   <rect>
+    <x>0</x>
+    <y>0</y>
+    <width>646</width>
+    <height>587</height>
+   </rect>
+  </property>
+  <property name="windowTitle">
+   <string>Texture Replacement Settings</string>
+  </property>
+  <layout class="QVBoxLayout" name="verticalLayout">
+   <item>
+    <layout class="QHBoxLayout" name="horizontalLayout_5">
+     <property name="bottomMargin">
+      <number>10</number>
+     </property>
+     <item>
+      <widget class="QLabel" name="icon">
+       <property name="minimumSize">
+        <size>
+         <width>32</width>
+         <height>32</height>
+        </size>
+       </property>
+       <property name="maximumSize">
+        <size>
+         <width>32</width>
+         <height>16777215</height>
+        </size>
+       </property>
+       <property name="alignment">
+        <set>Qt::AlignmentFlag::AlignLeading|Qt::AlignmentFlag::AlignLeft|Qt::AlignmentFlag::AlignTop</set>
+       </property>
+      </widget>
+     </item>
+     <item>
+      <widget class="QLabel" name="label_5">
+       <property name="text">
+        <string>&lt;html&gt;&lt;head/&gt;&lt;body&gt;&lt;p&gt;&lt;span style=&quot; font-weight:700;&quot;&gt;Texture Replacement Settings&lt;/span&gt;&lt;br/&gt;These settings fine-tune the behavior of the texture replacement system. You can also export a game-specific configuration file. Each of the options is explained in the configuration file, and at &lt;a href=&quot;https://github.com/stenzek/duckstation/wiki/Texture-Replacement&quot;&gt;&lt;span style=&quot; text-decoration: underline; color:#0078d4;&quot;&gt;https://github.com/stenzek/duckstation/wiki/Texture-Replacement&lt;/span&gt;&lt;/a&gt;.&lt;/p&gt;&lt;/body&gt;&lt;/html&gt;</string>
+       </property>
+       <property name="textFormat">
+        <enum>Qt::TextFormat::RichText</enum>
+       </property>
+       <property name="alignment">
+        <set>Qt::AlignmentFlag::AlignLeading|Qt::AlignmentFlag::AlignLeft|Qt::AlignmentFlag::AlignTop</set>
+       </property>
+       <property name="wordWrap">
+        <bool>true</bool>
+       </property>
+      </widget>
+     </item>
+    </layout>
+   </item>
+   <item>
+    <widget class="QGroupBox" name="groupBox">
+     <property name="title">
+      <string>Texture Dumping Mode</string>
+     </property>
+     <layout class="QGridLayout" name="gridLayout">
+      <item row="1" column="1">
+       <widget class="QCheckBox" name="reducePaletteRange">
+        <property name="text">
+         <string>Reduce Palette Range</string>
+        </property>
+       </widget>
+      </item>
+      <item row="0" column="0">
+       <widget class="QCheckBox" name="dumpTexturePages">
+        <property name="text">
+         <string>Dump Texture Pages</string>
+        </property>
+       </widget>
+      </item>
+      <item row="1" column="0">
+       <widget class="QCheckBox" name="dumpC16Textures">
+        <property name="text">
+         <string>Dump C16 Textures</string>
+        </property>
+       </widget>
+      </item>
+      <item row="0" column="1">
+       <widget class="QCheckBox" name="dumpFullTexturePages">
+        <property name="text">
+         <string>Dump Full Texture Pages</string>
+        </property>
+       </widget>
+      </item>
+      <item row="2" column="0" colspan="2">
+       <widget class="QLabel" name="label">
+        <property name="text">
+         <string>The texture dumping system can either operate in page mode, or write-tracking mode. Replacements can be loaded from either dump method.</string>
+        </property>
+        <property name="wordWrap">
+         <bool>true</bool>
+        </property>
+       </widget>
+      </item>
+     </layout>
+    </widget>
+   </item>
+   <item>
+    <widget class="QGroupBox" name="groupBox_3">
+     <property name="title">
+      <string>Write Tracking Options</string>
+     </property>
+     <layout class="QFormLayout" name="formLayout_3">
+      <item row="0" column="0">
+       <widget class="QLabel" name="label_2">
+        <property name="text">
+         <string>Maximum Write Coalesce Size:</string>
+        </property>
+       </widget>
+      </item>
+      <item row="0" column="1">
+       <layout class="QHBoxLayout" name="horizontalLayout_8" stretch="1,0,1">
+        <item>
+         <widget class="QSpinBox" name="maxVRAMWriteCoalesceWidth">
+          <property name="suffix">
+           <string>px</string>
+          </property>
+          <property name="minimum">
+           <number>0</number>
+          </property>
+          <property name="maximum">
+           <number>1024</number>
+          </property>
+          <property name="value">
+           <number>0</number>
+          </property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QLabel" name="maxVRAMWriteCoalesceSeparator">
+          <property name="text">
+           <string>x</string>
+          </property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QSpinBox" name="maxVRAMWriteCoalesceHeight">
+          <property name="suffix">
+           <string>px</string>
+          </property>
+          <property name="minimum">
+           <number>0</number>
+          </property>
+          <property name="maximum">
+           <number>512</number>
+          </property>
+          <property name="value">
+           <number>0</number>
+          </property>
+         </widget>
+        </item>
+       </layout>
+      </item>
+      <item row="1" column="0">
+       <widget class="QLabel" name="label_4">
+        <property name="text">
+         <string>Maximum Write Splits:</string>
+        </property>
+       </widget>
+      </item>
+      <item row="1" column="1">
+       <widget class="QSpinBox" name="maxVRAMWriteSplits">
+        <property name="minimum">
+         <number>0</number>
+        </property>
+        <property name="maximum">
+         <number>32</number>
+        </property>
+       </widget>
+      </item>
+      <item row="2" column="0" colspan="2">
+       <widget class="QCheckBox" name="convertCopiesToWrites">
+        <property name="text">
+         <string>Convert Copies To Writes</string>
+        </property>
+       </widget>
+      </item>
+     </layout>
+    </widget>
+   </item>
+   <item>
+    <widget class="QGroupBox" name="groupBox_4">
+     <property name="title">
+      <string>Texture Dumping Options</string>
+     </property>
+     <layout class="QFormLayout" name="formLayout">
+      <item row="0" column="0">
+       <widget class="QLabel" name="vramWriteDumpThresholdLabel_2">
+        <property name="text">
+         <string>Dump Size Threshold:</string>
+        </property>
+       </widget>
+      </item>
+      <item row="0" column="1">
+       <layout class="QHBoxLayout" name="horizontalLayout_7" stretch="1,0,1,0">
+        <item>
+         <widget class="QSpinBox" name="minDumpedTextureWidth">
+          <property name="suffix">
+           <string>px</string>
+          </property>
+          <property name="minimum">
+           <number>1</number>
+          </property>
+          <property name="maximum">
+           <number>1024</number>
+          </property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QLabel" name="vramWriteDumpThresholdSeparator_2">
+          <property name="text">
+           <string>x</string>
+          </property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QSpinBox" name="minDumpedTextureHeight">
+          <property name="suffix">
+           <string>px</string>
+          </property>
+          <property name="minimum">
+           <number>1</number>
+          </property>
+          <property name="maximum">
+           <number>512</number>
+          </property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QCheckBox" name="setTextureDumpAlphaChannel">
+          <property name="text">
+           <string>Set Alpha Channel</string>
+          </property>
+         </widget>
+        </item>
+       </layout>
+      </item>
+      <item row="1" column="0" colspan="2">
+       <widget class="QLabel" name="label_8">
+        <property name="text">
+         <string>Determines the minimum size of a texture that will be dumped. Textures with a size smaller than this value will be ignored.</string>
+        </property>
+        <property name="wordWrap">
+         <bool>true</bool>
+        </property>
+       </widget>
+      </item>
+     </layout>
+    </widget>
+   </item>
+   <item>
+    <widget class="QGroupBox" name="groupBox_2">
+     <property name="title">
+      <string>Background Dumping Options</string>
+     </property>
+     <layout class="QFormLayout" name="formLayout_2">
+      <item row="0" column="0">
+       <widget class="QLabel" name="vramWriteDumpThresholdLabel">
+        <property name="text">
+         <string>Dump Size Threshold:</string>
+        </property>
+       </widget>
+      </item>
+      <item row="0" column="1">
+       <layout class="QHBoxLayout" name="horizontalLayout_6" stretch="1,0,1,0">
+        <item>
+         <widget class="QSpinBox" name="minDumpedVRAMWriteWidth">
+          <property name="suffix">
+           <string>px</string>
+          </property>
+          <property name="minimum">
+           <number>1</number>
+          </property>
+          <property name="maximum">
+           <number>1024</number>
+          </property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QLabel" name="vramWriteDumpThresholdSeparator">
+          <property name="text">
+           <string>x</string>
+          </property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QSpinBox" name="minDumpedVRAMWriteHeight">
+          <property name="suffix">
+           <string>px</string>
+          </property>
+          <property name="minimum">
+           <number>1</number>
+          </property>
+          <property name="maximum">
+           <number>512</number>
+          </property>
+         </widget>
+        </item>
+        <item>
+         <widget class="QCheckBox" name="setVRAMWriteAlphaChannel">
+          <property name="text">
+           <string>Set Alpha Channel</string>
+          </property>
+         </widget>
+        </item>
+       </layout>
+      </item>
+      <item row="1" column="0" colspan="2">
+       <widget class="QLabel" name="label_9">
+        <property name="text">
+         <string>Determines the minimum size of a VRAM write that will be dumped, in background dumping mode. Uploads smaller than this size will be ignored.</string>
+        </property>
+        <property name="wordWrap">
+         <bool>true</bool>
+        </property>
+       </widget>
+      </item>
+     </layout>
+    </widget>
+   </item>
+   <item>
+    <spacer name="verticalSpacer">
+     <property name="orientation">
+      <enum>Qt::Orientation::Vertical</enum>
+     </property>
+     <property name="sizeHint" stdset="0">
+      <size>
+       <width>20</width>
+       <height>40</height>
+      </size>
+     </property>
+    </spacer>
+   </item>
+   <item>
+    <layout class="QHBoxLayout" name="horizontalLayout">
+     <item>
+      <spacer name="horizontalSpacer">
+       <property name="orientation">
+        <enum>Qt::Orientation::Horizontal</enum>
+       </property>
+       <property name="sizeHint" stdset="0">
+        <size>
+         <width>198</width>
+         <height>20</height>
+        </size>
+       </property>
+      </spacer>
+     </item>
+     <item>
+      <widget class="QPushButton" name="exportButton">
+       <property name="text">
+        <string>Export...</string>
+       </property>
+      </widget>
+     </item>
+     <item>
+      <widget class="QPushButton" name="closeButton">
+       <property name="text">
+        <string>Close</string>
+       </property>
+       <property name="default">
+        <bool>true</bool>
+       </property>
+      </widget>
+     </item>
+    </layout>
+   </item>
+  </layout>
+ </widget>
+ <resources/>
+ <connections/>
+</ui>