Merge pull request #7539 from stenzek/batched-efb-copies

TextureCache: Deferred/batched EFB copies

Source/Android/app/src/main/java/org/dolphinemu/dolphinemu/features/settings/ui/SettingsFragmentPresenter.java

@@ -559,6 +559,7 @@ public final class SettingsFragmentPresenter
             new BooleanSetting(SettingsFile.KEY_IGNORE_FORMAT, Settings.SECTION_GFX_HACKS,
                     ignoreFormatValue);
     Setting efbToTexture = hacksSection.getSetting(SettingsFile.KEY_EFB_TEXTURE);
+    Setting deferEfbCopies = hacksSection.getSetting(SettingsFile.KEY_DEFER_EFB_COPIES);
     Setting texCacheAccuracy = gfxSection.getSetting(SettingsFile.KEY_TEXCACHE_ACCURACY);
     Setting gpuTextureDecoding = gfxSection.getSetting(SettingsFile.KEY_GPU_TEXTURE_DECODING);
     Setting xfbToTexture = hacksSection.getSetting(SettingsFile.KEY_XFB_TEXTURE);
@@ -573,6 +574,9 @@ public final class SettingsFragmentPresenter
             ignoreFormat));
     sl.add(new CheckBoxSetting(SettingsFile.KEY_EFB_TEXTURE, Settings.SECTION_GFX_HACKS,
             R.string.efb_copy_method, R.string.efb_copy_method_description, true, efbToTexture));
+    sl.add(new CheckBoxSetting(SettingsFile.KEY_DEFER_EFB_COPIES, Settings.SECTION_GFX_HACKS,
+            R.string.defer_efb_copies, R.string.defer_efb_copies_description, true,
+            deferEfbCopies));
 
     sl.add(new HeaderSetting(null, null, R.string.texture_cache, 0));
     sl.add(new SingleChoiceSetting(SettingsFile.KEY_TEXCACHE_ACCURACY,

Source/Android/app/src/main/java/org/dolphinemu/dolphinemu/features/settings/utils/SettingsFile.java

@@ -80,6 +80,7 @@ public final class SettingsFile
   public static final String KEY_SKIP_EFB = "EFBAccessEnable";
   public static final String KEY_IGNORE_FORMAT = "EFBEmulateFormatChanges";
   public static final String KEY_EFB_TEXTURE = "EFBToTextureEnable";
+  public static final String KEY_DEFER_EFB_COPIES = "DeferEFBCopies";
   public static final String KEY_TEXCACHE_ACCURACY = "SafeTextureCacheColorSamples";
   public static final String KEY_GPU_TEXTURE_DECODING = "EnableGPUTextureDecoding";
   public static final String KEY_XFB_TEXTURE = "XFBToTextureEnable";

Source/Android/app/src/main/res/values/strings.xml

@@ -207,6 +207,8 @@
     <string name="ignore_format_changes_description">Ignore any changes to the EFB format.</string>
     <string name="efb_copy_method">Store EFB Copies to Texture Only</string>
     <string name="efb_copy_method_description">Stores EFB Copies exclusively on the GPU, bypassing system memory. Causes graphical defects in a small number of games. If unsure, leave this checked.</string>
+    <string name="defer_efb_copies">Defer EFB Copies to RAM</string>
+    <string name="defer_efb_copies_description">Waits until the game synchronizes with the emulated GPU before writing the contents of EFB copies to RAM. May result in faster performance. If unsure, leave this unchecked.</string>
     <string name="texture_cache">Texture Cache</string>
     <string name="texture_cache_accuracy">Texture Cache Accuracy</string>
     <string name="texture_cache_accuracy_description">The safer the selection, the less likely the emulator will be missing any texture updates from RAM.</string>

Source/Core/Core/Analytics.cpp

@@ -317,6 +317,7 @@ void DolphinAnalytics::MakePerGameBuilder()
   builder.AddData("cfg-gfx-efb-copy-format-changes", g_Config.bEFBEmulateFormatChanges);
   builder.AddData("cfg-gfx-efb-copy-ram", !g_Config.bSkipEFBCopyToRam);
   builder.AddData("cfg-gfx-xfb-copy-ram", !g_Config.bSkipXFBCopyToRam);
+  builder.AddData("cfg-gfx-defer-efb-copies", g_Config.bDeferEFBCopies);
   builder.AddData("cfg-gfx-immediate-xfb", !g_Config.bImmediateXFB);
   builder.AddData("cfg-gfx-efb-copy-scaled", g_Config.bCopyEFBScaled);
   builder.AddData("cfg-gfx-internal-resolution", g_Config.iEFBScale);

Source/Core/Core/Config/GraphicsSettings.cpp

@@ -139,6 +139,7 @@ const ConfigInfo<bool> GFX_HACK_SKIP_XFB_COPY_TO_RAM{{System::GFX, "Hacks", "XFB
                                                      true};
 const ConfigInfo<bool> GFX_HACK_DISABLE_COPY_TO_VRAM{{System::GFX, "Hacks", "DisableCopyToVRAM"},
                                                      false};
+const ConfigInfo<bool> GFX_HACK_DEFER_EFB_COPIES{{System::GFX, "Hacks", "DeferEFBCopies"}, true};
 const ConfigInfo<bool> GFX_HACK_IMMEDIATE_XFB{{System::GFX, "Hacks", "ImmediateXFBEnable"}, false};
 const ConfigInfo<bool> GFX_HACK_COPY_EFB_SCALED{{System::GFX, "Hacks", "EFBScaledCopy"}, true};
 const ConfigInfo<bool> GFX_HACK_EFB_EMULATE_FORMAT_CHANGES{

Source/Core/Core/Config/GraphicsSettings.h

@@ -107,6 +107,7 @@ extern const ConfigInfo<bool> GFX_HACK_FORCE_PROGRESSIVE;
 extern const ConfigInfo<bool> GFX_HACK_SKIP_EFB_COPY_TO_RAM;
 extern const ConfigInfo<bool> GFX_HACK_SKIP_XFB_COPY_TO_RAM;
 extern const ConfigInfo<bool> GFX_HACK_DISABLE_COPY_TO_VRAM;
+extern const ConfigInfo<bool> GFX_HACK_DEFER_EFB_COPIES;
 extern const ConfigInfo<bool> GFX_HACK_IMMEDIATE_XFB;
 extern const ConfigInfo<bool> GFX_HACK_COPY_EFB_SCALED;
 extern const ConfigInfo<bool> GFX_HACK_EFB_EMULATE_FORMAT_CHANGES;

Source/Core/Core/ConfigLoaders/IsSettingSaveable.cpp

@@ -114,6 +114,7 @@ bool IsSettingSaveable(const Config::ConfigLocation& config_location)
       Config::GFX_HACK_SKIP_EFB_COPY_TO_RAM.location,
       Config::GFX_HACK_SKIP_XFB_COPY_TO_RAM.location,
       Config::GFX_HACK_DISABLE_COPY_TO_VRAM.location,
+      Config::GFX_HACK_DEFER_EFB_COPIES.location,
       Config::GFX_HACK_IMMEDIATE_XFB.location,
       Config::GFX_HACK_COPY_EFB_SCALED.location,
       Config::GFX_HACK_EFB_EMULATE_FORMAT_CHANGES.location,

Source/Core/DolphinQt/Config/Graphics/HacksWidget.cpp

@@ -46,10 +46,13 @@ void HacksWidget::CreateWidgets()
                                              Config::GFX_HACK_EFB_EMULATE_FORMAT_CHANGES, true);
   m_store_efb_copies = new GraphicsBool(tr("Store EFB Copies to Texture Only"),
                                         Config::GFX_HACK_SKIP_EFB_COPY_TO_RAM);
+  m_defer_efb_copies =
+      new GraphicsBool(tr("Defer EFB Copies to RAM"), Config::GFX_HACK_DEFER_EFB_COPIES);
 
   efb_layout->addWidget(m_skip_efb_cpu, 0, 0);
   efb_layout->addWidget(m_ignore_format_changes, 0, 1);
   efb_layout->addWidget(m_store_efb_copies, 1, 0);
+  efb_layout->addWidget(m_defer_efb_copies, 1, 1);
 
   // Texture Cache
   auto* texture_cache_box = new QGroupBox(tr("Texture Cache"));
@@ -109,6 +112,8 @@ void HacksWidget::CreateWidgets()
   main_layout->addStretch();
 
   setLayout(main_layout);
+
+  UpdateDeferEFBCopiesEnabled();
 }
 
 void HacksWidget::OnBackendChanged(const QString& backend_name)
@@ -129,6 +134,10 @@ void HacksWidget::OnBackendChanged(const QString& backend_name)
 void HacksWidget::ConnectWidgets()
 {
   connect(m_accuracy, &QSlider::valueChanged, [this](int) { SaveSettings(); });
+  connect(m_store_efb_copies, &QCheckBox::stateChanged,
+          [this](int) { UpdateDeferEFBCopiesEnabled(); });
+  connect(m_store_xfb_copies, &QCheckBox::stateChanged,
+          [this](int) { UpdateDeferEFBCopiesEnabled(); });
 }
 
 void HacksWidget::LoadSettings()
@@ -202,6 +211,11 @@ void HacksWidget::AddDescriptions()
       "in a small number of games.\n\nEnabled = EFB Copies to Texture\nDisabled = EFB Copies to "
       "RAM "
       "(and Texture)\n\nIf unsure, leave this checked.");
+  static const char TR_DEFER_EFB_COPIES_DESCRIPTION[] = QT_TR_NOOP(
+      "Waits until the game synchronizes with the emulated GPU before writing the contents of EFB "
+      "copies to RAM. Reduces the overhead of EFB RAM copies, provides a performance boost in many "
+      "games, at the risk of breaking those which do not safely synchronize with the emulated "
+      "GPU.\n\nIf unsure, leave this checked.");
   static const char TR_ACCUARCY_DESCRIPTION[] = QT_TR_NOOP(
       "The \"Safe\" setting eliminates the likelihood of the GPU missing texture updates "
       "from RAM.\nLower accuracies cause in-game text to appear garbled in certain "
@@ -240,6 +254,7 @@ void HacksWidget::AddDescriptions()
   AddDescription(m_skip_efb_cpu, TR_SKIP_EFB_CPU_ACCESS_DESCRIPTION);
   AddDescription(m_ignore_format_changes, TR_IGNORE_FORMAT_CHANGE_DESCRIPTION);
   AddDescription(m_store_efb_copies, TR_STORE_EFB_TO_TEXTURE_DESCRIPTION);
+  AddDescription(m_defer_efb_copies, TR_DEFER_EFB_COPIES_DESCRIPTION);
   AddDescription(m_accuracy, TR_ACCUARCY_DESCRIPTION);
   AddDescription(m_store_xfb_copies, TR_STORE_XFB_TO_TEXTURE_DESCRIPTION);
   AddDescription(m_immediate_xfb, TR_IMMEDIATE_XFB_DESCRIPTION);
@@ -248,3 +263,11 @@ void HacksWidget::AddDescriptions()
   AddDescription(m_disable_bounding_box, TR_DISABLE_BOUNDINGBOX_DESCRIPTION);
   AddDescription(m_vertex_rounding, TR_VERTEX_ROUNDING_DESCRIPTION);
 }
+
+void HacksWidget::UpdateDeferEFBCopiesEnabled()
+{
+  // We disable the checkbox for defer EFB copies when both EFB and XFB copies to texture are
+  // enabled.
+  const bool can_defer = m_store_efb_copies->isChecked() && m_store_xfb_copies->isChecked();
+  m_defer_efb_copies->setEnabled(!can_defer);
+}

Source/Core/DolphinQt/Config/Graphics/HacksWidget.h

@@ -42,8 +42,11 @@ private:
   QCheckBox* m_fast_depth_calculation;
   QCheckBox* m_disable_bounding_box;
   QCheckBox* m_vertex_rounding;
+  QCheckBox* m_defer_efb_copies;
 
   void CreateWidgets();
   void ConnectWidgets();
   void AddDescriptions();
+
+  void UpdateDeferEFBCopiesEnabled();
 };

Source/Core/VideoBackends/D3D/PSTextureEncoder.cpp

@@ -46,13 +46,8 @@ PSTextureEncoder::~PSTextureEncoder() = default;
 
 void PSTextureEncoder::Init()
 {
-  // TODO: Move this to a constant somewhere in common.
-  TextureConfig encoding_texture_config(EFB_WIDTH * 4, 1024, 1, 1, 1, AbstractTextureFormat::BGRA8,
-                                        true);
-  m_encoding_render_texture = g_renderer->CreateTexture(encoding_texture_config);
-  m_encoding_readback_texture =
-      g_renderer->CreateStagingTexture(StagingTextureType::Readback, encoding_texture_config);
-  ASSERT(m_encoding_render_texture && m_encoding_readback_texture);
+  m_encoding_render_texture = g_renderer->CreateTexture(TextureCache::GetEncodingTextureConfig());
+  ASSERT(m_encoding_render_texture);
 
   // Create constant buffer for uploading data to shaders
   D3D11_BUFFER_DESC bd = CD3D11_BUFFER_DESC(sizeof(EFBEncodeParams), D3D11_BIND_CONSTANT_BUFFER);
@@ -71,9 +66,9 @@ void PSTextureEncoder::Shutdown()
 }
 
 void PSTextureEncoder::Encode(
-    u8* dst, const EFBCopyParams& params, u32 native_width, u32 bytes_per_row, u32 num_blocks_y,
-    u32 memory_stride, const EFBRectangle& src_rect, bool scale_by_half, float y_scale, float gamma,
-    bool clamp_top, bool clamp_bottom,
+    AbstractStagingTexture* dst, const EFBCopyParams& params, u32 native_width, u32 bytes_per_row,
+    u32 num_blocks_y, u32 memory_stride, const EFBRectangle& src_rect, bool scale_by_half,
+    float y_scale, float gamma, bool clamp_top, bool clamp_bottom,
     const TextureCacheBase::CopyFilterCoefficientArray& filter_coefficients)
 {
   // Resolve MSAA targets before copying.
@@ -133,14 +128,7 @@ void PSTextureEncoder::Encode(
 
     // Copy to staging buffer
     MathUtil::Rectangle<int> copy_rect(0, 0, words_per_row, num_blocks_y);
-    m_encoding_readback_texture->CopyFromTexture(m_encoding_render_texture.get(), copy_rect, 0, 0,
-                                                 copy_rect);
-    m_encoding_readback_texture->Flush();
-    if (m_encoding_readback_texture->Map())
-    {
-      m_encoding_readback_texture->ReadTexels(copy_rect, dst, memory_stride);
-      m_encoding_readback_texture->Unmap();
-    }
+    dst->CopyFromTexture(m_encoding_render_texture.get(), copy_rect, 0, 0, copy_rect);
   }
 
   g_renderer->RestoreAPIState();

Source/Core/VideoBackends/D3D/PSTextureEncoder.h

@@ -38,9 +38,9 @@ public:
 
   void Init();
   void Shutdown();
-  void Encode(u8* dst, const EFBCopyParams& params, u32 native_width, u32 bytes_per_row,
-              u32 num_blocks_y, u32 memory_stride, const EFBRectangle& src_rect, bool scale_by_half,
-              float y_scale, float gamma, bool clamp_top, bool clamp_bottom,
+  void Encode(AbstractStagingTexture* dst, const EFBCopyParams& params, u32 native_width,
+              u32 bytes_per_row, u32 num_blocks_y, u32 memory_stride, const EFBRectangle& src_rect,
+              bool scale_by_half, float y_scale, float gamma, bool clamp_top, bool clamp_bottom,
               const TextureCacheBase::CopyFilterCoefficientArray& filter_coefficients);
 
 private:
@@ -48,7 +48,6 @@ private:
 
   ID3D11Buffer* m_encode_params = nullptr;
   std::unique_ptr<AbstractTexture> m_encoding_render_texture;
-  std::unique_ptr<AbstractStagingTexture> m_encoding_readback_texture;
   std::map<EFBCopyParams, ID3D11PixelShader*> m_encoding_shaders;
 };
 }

Source/Core/VideoBackends/D3D/TextureCache.cpp

@@ -31,8 +31,8 @@ namespace DX11
 {
 static std::unique_ptr<PSTextureEncoder> g_encoder;
 
-void TextureCache::CopyEFB(u8* dst, const EFBCopyParams& params, u32 native_width,
-                           u32 bytes_per_row, u32 num_blocks_y, u32 memory_stride,
+void TextureCache::CopyEFB(AbstractStagingTexture* dst, const EFBCopyParams& params,
+                           u32 native_width, u32 bytes_per_row, u32 num_blocks_y, u32 memory_stride,
                            const EFBRectangle& src_rect, bool scale_by_half, float y_scale,
                            float gamma, bool clamp_top, bool clamp_bottom,
                            const CopyFilterCoefficientArray& filter_coefficients)

Source/Core/VideoBackends/D3D/TextureCache.h

@@ -22,18 +22,11 @@ public:
   ~TextureCache();
 
 private:
-  u64 EncodeToRamFromTexture(u32 address, void* source_texture, u32 SourceW, u32 SourceH,
-                             bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf,
-                             const EFBRectangle& source)
-  {
-    return 0;
-  };
-
   void ConvertTexture(TCacheEntry* destination, TCacheEntry* source, const void* palette,
                       TLUTFormat format) override;
 
-  void CopyEFB(u8* dst, const EFBCopyParams& params, u32 native_width, u32 bytes_per_row,
-               u32 num_blocks_y, u32 memory_stride, const EFBRectangle& src_rect,
+  void CopyEFB(AbstractStagingTexture* dst, const EFBCopyParams& params, u32 native_width,
+               u32 bytes_per_row, u32 num_blocks_y, u32 memory_stride, const EFBRectangle& src_rect,
                bool scale_by_half, float y_scale, float gamma, bool clamp_top, bool clamp_bottom,
                const CopyFilterCoefficientArray& filter_coefficients) override;
 

Source/Core/VideoBackends/Null/TextureCache.h

@@ -25,8 +25,8 @@ public:
   {
   }
 
-  void CopyEFB(u8* dst, const EFBCopyParams& params, u32 native_width, u32 bytes_per_row,
-               u32 num_blocks_y, u32 memory_stride, const EFBRectangle& src_rect,
+  void CopyEFB(AbstractStagingTexture* dst, const EFBCopyParams& params, u32 native_width,
+               u32 bytes_per_row, u32 num_blocks_y, u32 memory_stride, const EFBRectangle& src_rect,
                bool scale_by_half, float y_scale, float gamma, bool clamp_top, bool clamp_bottom,
                const CopyFilterCoefficientArray& filter_coefficients) override
   {

Source/Core/VideoBackends/OGL/TextureCache.cpp

@@ -159,8 +159,8 @@ void main()
 
 //#define TIME_TEXTURE_DECODING 1
 
-void TextureCache::CopyEFB(u8* dst, const EFBCopyParams& params, u32 native_width,
-                           u32 bytes_per_row, u32 num_blocks_y, u32 memory_stride,
+void TextureCache::CopyEFB(AbstractStagingTexture* dst, const EFBCopyParams& params,
+                           u32 native_width, u32 bytes_per_row, u32 num_blocks_y, u32 memory_stride,
                            const EFBRectangle& src_rect, bool scale_by_half, float y_scale,
                            float gamma, bool clamp_top, bool clamp_bottom,
                            const CopyFilterCoefficientArray& filter_coefficients)

Source/Core/VideoBackends/OGL/TextureCache.h

@@ -63,8 +63,8 @@ private:
   void ConvertTexture(TCacheEntry* destination, TCacheEntry* source, const void* palette,
                       TLUTFormat format) override;
 
-  void CopyEFB(u8* dst, const EFBCopyParams& params, u32 native_width, u32 bytes_per_row,
-               u32 num_blocks_y, u32 memory_stride, const EFBRectangle& src_rect,
+  void CopyEFB(AbstractStagingTexture* dst, const EFBCopyParams& params, u32 native_width,
+               u32 bytes_per_row, u32 num_blocks_y, u32 memory_stride, const EFBRectangle& src_rect,
                bool scale_by_half, float y_scale, float gamma, bool clamp_top, bool clamp_bottom,
                const CopyFilterCoefficientArray& filter_coefficients) override;
 

Source/Core/VideoBackends/OGL/TextureConverter.cpp

@@ -46,7 +46,6 @@ struct EncodingProgram
 
 std::map<EFBCopyParams, EncodingProgram> s_encoding_programs;
 std::unique_ptr<AbstractTexture> s_encoding_render_texture;
-std::unique_ptr<AbstractStagingTexture> s_encoding_readback_texture;
 
 const int renderBufferWidth = EFB_WIDTH * 4;
 const int renderBufferHeight = 1024;
@@ -93,16 +92,11 @@ static EncodingProgram& GetOrCreateEncodingShader(const EFBCopyParams& params)
 
 void Init()
 {
-  TextureConfig config(renderBufferWidth, renderBufferHeight, 1, 1, 1, AbstractTextureFormat::BGRA8,
-                       true);
-  s_encoding_render_texture = g_renderer->CreateTexture(config);
-  s_encoding_readback_texture =
-      g_renderer->CreateStagingTexture(StagingTextureType::Readback, config);
+  s_encoding_render_texture = g_renderer->CreateTexture(TextureCache::GetEncodingTextureConfig());
 }
 
 void Shutdown()
 {
-  s_encoding_readback_texture.reset();
   s_encoding_render_texture.reset();
 
   for (auto& program : s_encoding_programs)
@@ -112,8 +106,9 @@ void Shutdown()
 
 // dst_line_size, writeStride in bytes
 
-static void EncodeToRamUsingShader(GLuint srcTexture, u8* destAddr, u32 dst_line_size,
-                                   u32 dstHeight, u32 writeStride, bool linearFilter, float y_scale)
+static void EncodeToRamUsingShader(GLuint srcTexture, AbstractStagingTexture* destAddr,
+                                   u32 dst_line_size, u32 dstHeight, u32 writeStride,
+                                   bool linearFilter, float y_scale)
 {
   FramebufferManager::SetFramebuffer(
       static_cast<OGLTexture*>(s_encoding_render_texture.get())->GetFramebuffer());
@@ -137,15 +132,14 @@ static void EncodeToRamUsingShader(GLuint srcTexture, u8* destAddr, u32 dst_line
   glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 
   MathUtil::Rectangle<int> copy_rect(0, 0, dst_line_size / 4, dstHeight);
-  s_encoding_readback_texture->CopyFromTexture(s_encoding_render_texture.get(), copy_rect, 0, 0,
-                                               copy_rect);
-  s_encoding_readback_texture->ReadTexels(copy_rect, destAddr, writeStride);
+
+  destAddr->CopyFromTexture(s_encoding_render_texture.get(), copy_rect, 0, 0, copy_rect);
 }
 
-void EncodeToRamFromTexture(u8* dest_ptr, const EFBCopyParams& params, u32 native_width,
-                            u32 bytes_per_row, u32 num_blocks_y, u32 memory_stride,
-                            const EFBRectangle& src_rect, bool scale_by_half, float y_scale,
-                            float gamma, float clamp_top, float clamp_bottom,
+void EncodeToRamFromTexture(AbstractStagingTexture* dest, const EFBCopyParams& params,
+                            u32 native_width, u32 bytes_per_row, u32 num_blocks_y,
+                            u32 memory_stride, const EFBRectangle& src_rect, bool scale_by_half,
+                            float y_scale, float gamma, float clamp_top, float clamp_bottom,
                             const TextureCacheBase::CopyFilterCoefficientArray& filter_coefficients)
 {
   g_renderer->ResetAPIState();
@@ -165,7 +159,7 @@ void EncodeToRamFromTexture(u8* dest_ptr, const EFBCopyParams& params, u32 nativ
                                   FramebufferManager::ResolveAndGetDepthTarget(src_rect) :
                                   FramebufferManager::ResolveAndGetRenderTarget(src_rect);
 
-  EncodeToRamUsingShader(read_texture, dest_ptr, bytes_per_row, num_blocks_y, memory_stride,
+  EncodeToRamUsingShader(read_texture, dest, bytes_per_row, num_blocks_y, memory_stride,
                          scale_by_half && !params.depth, y_scale);
 
   g_renderer->RestoreAPIState();

Source/Core/VideoBackends/OGL/TextureConverter.h

@@ -10,6 +10,9 @@
 #include "VideoCommon/TextureCacheBase.h"
 #include "VideoCommon/VideoCommon.h"
 
+struct EFBCopyParams;
+class AbstractStagingTexture;
+
 namespace OGL
 {
 // Converts textures between formats using shaders
@@ -21,7 +24,7 @@ void Shutdown();
 
 // returns size of the encoded data (in bytes)
 void EncodeToRamFromTexture(
-    u8* dest_ptr, const EFBCopyParams& params, u32 native_width, u32 bytes_per_row,
+    AbstractStagingTexture* dest, const EFBCopyParams& params, u32 native_width, u32 bytes_per_row,
     u32 num_blocks_y, u32 memory_stride, const EFBRectangle& src_rect, bool scale_by_half,
     float y_scale, float gamma, float clamp_top, float clamp_bottom,
     const TextureCacheBase::CopyFilterCoefficientArray& filter_coefficients);

Source/Core/VideoBackends/Software/SWTexture.h

@@ -57,6 +57,8 @@ public:
   void Unmap() override;
   void Flush() override;
 
+  void SetMapStride(size_t stride) { m_map_stride = stride; }
+
 private:
   std::vector<u8> m_data;
 };

Source/Core/VideoBackends/Software/TextureCache.h

@@ -16,8 +16,8 @@ public:
                       TLUTFormat format) override
   {
   }
-  void CopyEFB(u8* dst, const EFBCopyParams& params, u32 native_width, u32 bytes_per_row,
-               u32 num_blocks_y, u32 memory_stride, const EFBRectangle& src_rect,
+  void CopyEFB(AbstractStagingTexture* dst, const EFBCopyParams& params, u32 native_width,
+               u32 bytes_per_row, u32 num_blocks_y, u32 memory_stride, const EFBRectangle& src_rect,
                bool scale_by_half, float y_scale, float gamma, bool clamp_top, bool clamp_bottom,
                const CopyFilterCoefficientArray& filter_coefficients) override
   {