Merge pull request #6193 from stenzek/readbacks

Abstract Staging Textures - VideoCommon interface for texture readbacks/uploads

Source/Core/VideoCommon/AbstractStagingTexture.cpp

@@ -0,0 +1,133 @@
+// Copyright 2017 Dolphin Emulator Project
+// Licensed under GPLv2+
+// Refer to the license.txt file included.
+
+#include <algorithm>
+#include <cstring>
+
+#include "Common/Assert.h"
+#include "Common/MsgHandler.h"
+#include "VideoCommon/AbstractStagingTexture.h"
+#include "VideoCommon/AbstractTexture.h"
+
+AbstractStagingTexture::AbstractStagingTexture(StagingTextureType type, const TextureConfig& c)
+    : m_type(type), m_config(c), m_texel_size(AbstractTexture::GetTexelSizeForFormat(c.format))
+{
+}
+
+AbstractStagingTexture::~AbstractStagingTexture() = default;
+
+void AbstractStagingTexture::CopyFromTexture(const AbstractTexture* src, u32 src_layer,
+                                             u32 src_level)
+{
+  MathUtil::Rectangle<int> src_rect = src->GetConfig().GetMipRect(src_level);
+  MathUtil::Rectangle<int> dst_rect = m_config.GetRect();
+  CopyFromTexture(src, src_rect, src_layer, src_level, dst_rect);
+}
+
+void AbstractStagingTexture::CopyToTexture(AbstractTexture* dst, u32 dst_layer, u32 dst_level)
+{
+  MathUtil::Rectangle<int> src_rect = m_config.GetRect();
+  MathUtil::Rectangle<int> dst_rect = dst->GetConfig().GetMipRect(dst_level);
+  CopyToTexture(src_rect, dst, dst_rect, dst_layer, dst_level);
+}
+
+void AbstractStagingTexture::ReadTexels(const MathUtil::Rectangle<int>& rect, void* out_ptr,
+                                        u32 out_stride)
+{
+  _assert_(m_type != StagingTextureType::Upload);
+  if (!PrepareForAccess())
+    return;
+
+  _assert_(rect.left >= 0 && static_cast<u32>(rect.right) <= m_config.width && rect.top >= 0 &&
+           static_cast<u32>(rect.bottom) <= m_config.height);
+
+  // Offset pointer to point to start of region being copied out.
+  const char* current_ptr = m_map_pointer;
+  current_ptr += rect.top * m_map_stride;
+  current_ptr += rect.left * m_texel_size;
+
+  // Optimal path: same dimensions, same stride.
+  if (rect.left == 0 && static_cast<u32>(rect.right) == m_config.width &&
+      m_map_stride == out_stride)
+  {
+    std::memcpy(out_ptr, current_ptr, m_map_stride * rect.GetHeight());
+    return;
+  }
+
+  size_t copy_size = std::min(static_cast<size_t>(rect.GetWidth() * m_texel_size), m_map_stride);
+  int copy_height = rect.GetHeight();
+  char* dst_ptr = reinterpret_cast<char*>(out_ptr);
+  for (int row = 0; row < copy_height; row++)
+  {
+    std::memcpy(dst_ptr, current_ptr, copy_size);
+    current_ptr += m_map_stride;
+    dst_ptr += out_stride;
+  }
+}
+
+void AbstractStagingTexture::ReadTexel(u32 x, u32 y, void* out_ptr)
+{
+  _assert_(m_type != StagingTextureType::Upload);
+  if (!PrepareForAccess())
+    return;
+
+  _assert_(x < m_config.width && y < m_config.height);
+  const char* src_ptr = m_map_pointer + y * m_map_stride + x * m_texel_size;
+  std::memcpy(out_ptr, src_ptr, m_texel_size);
+}
+
+void AbstractStagingTexture::WriteTexels(const MathUtil::Rectangle<int>& rect, const void* in_ptr,
+                                         u32 in_stride)
+{
+  _assert_(m_type != StagingTextureType::Readback);
+  if (!PrepareForAccess())
+    return;
+
+  _assert_(rect.left >= 0 && static_cast<u32>(rect.right) <= m_config.width && rect.top >= 0 &&
+           static_cast<u32>(rect.bottom) <= m_config.height);
+
+  // Offset pointer to point to start of region being copied to.
+  char* current_ptr = m_map_pointer;
+  current_ptr += rect.top * m_map_stride;
+  current_ptr += rect.left * m_texel_size;
+
+  // Optimal path: same dimensions, same stride.
+  if (rect.left == 0 && static_cast<u32>(rect.right) == m_config.width && m_map_stride == in_stride)
+  {
+    std::memcpy(current_ptr, in_ptr, m_map_stride * rect.GetHeight());
+    return;
+  }
+
+  size_t copy_size = std::min(static_cast<size_t>(rect.GetWidth() * m_texel_size), m_map_stride);
+  int copy_height = rect.GetHeight();
+  const char* src_ptr = reinterpret_cast<const char*>(in_ptr);
+  for (int row = 0; row < copy_height; row++)
+  {
+    std::memcpy(current_ptr, src_ptr, copy_size);
+    current_ptr += m_map_stride;
+    src_ptr += in_stride;
+  }
+}
+
+void AbstractStagingTexture::WriteTexel(u32 x, u32 y, const void* in_ptr)
+{
+  _assert_(m_type != StagingTextureType::Readback);
+  if (!PrepareForAccess())
+    return;
+
+  _assert_(x < m_config.width && y < m_config.height);
+  char* dest_ptr = m_map_pointer + y * m_map_stride + x * m_texel_size;
+  std::memcpy(dest_ptr, in_ptr, m_texel_size);
+}
+
+bool AbstractStagingTexture::PrepareForAccess()
+{
+  if (m_needs_flush)
+  {
+    if (IsMapped())
+      Unmap();
+    Flush();
+  }
+  return IsMapped() || Map();
+}

Source/Core/VideoCommon/AbstractStagingTexture.h

@@ -0,0 +1,86 @@
+// Copyright 2017 Dolphin Emulator Project
+// Licensed under GPLv2+
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <cstddef>
+#include <string>
+
+#include "Common/CommonTypes.h"
+#include "Common/MathUtil.h"
+#include "VideoCommon/TextureConfig.h"
+
+class AbstractTexture;
+
+class AbstractStagingTexture
+{
+public:
+  explicit AbstractStagingTexture(StagingTextureType type, const TextureConfig& c);
+  virtual ~AbstractStagingTexture();
+
+  const TextureConfig& GetConfig() const { return m_config; }
+  StagingTextureType GetType() const { return m_type; }
+  size_t GetTexelSize() const { return m_texel_size; }
+  bool IsMapped() const { return m_map_pointer != nullptr; }
+  char* GetMappedPointer() const { return m_map_pointer; }
+  size_t GetMappedStride() const { return m_map_stride; }
+  // Copies from the GPU texture object to the staging texture, which can be mapped/read by the CPU.
+  // Both src_rect and dst_rect must be with within the bounds of the the specified textures.
+  virtual void CopyFromTexture(const AbstractTexture* src, const MathUtil::Rectangle<int>& src_rect,
+                               u32 src_layer, u32 src_level,
+                               const MathUtil::Rectangle<int>& dst_rect) = 0;
+
+  // Wrapper for copying a whole layer of a texture to a readback texture.
+  // Assumes that the level of src texture and this texture have the same dimensions.
+  void CopyFromTexture(const AbstractTexture* src, u32 src_layer = 0, u32 src_level = 0);
+
+  // Copies from this staging texture to a GPU texture.
+  // Both src_rect and dst_rect must be with within the bounds of the the specified textures.
+  virtual void CopyToTexture(const MathUtil::Rectangle<int>& src_rect, AbstractTexture* dst,
+                             const MathUtil::Rectangle<int>& dst_rect, u32 dst_layer,
+                             u32 dst_level) = 0;
+
+  // Wrapper for copying a whole layer of a texture to a readback texture.
+  // Assumes that the level of src texture and this texture have the same dimensions.
+  void CopyToTexture(AbstractTexture* dst, u32 dst_layer = 0, u32 dst_level = 0);
+
+  // Maps the texture into the CPU address space, enabling it to read the contents.
+  // The Map call may not perform synchronization. If the contents of the staging texture
+  // has been updated by a CopyFromTexture call, you must call Flush() first.
+  // If persistent mapping is supported in the backend, this may be a no-op.
+  virtual bool Map() = 0;
+
+  // Unmaps the CPU-readable copy of the texture. May be a no-op on backends which
+  // support persistent-mapped buffers.
+  virtual void Unmap() = 0;
+
+  // Flushes pending writes from the CPU to the GPU, and reads from the GPU to the CPU.
+  // This may cause a command buffer flush depending on if one has occurred between the last
+  // call to CopyFromTexture()/CopyToTexture() and the Flush() call.
+  virtual void Flush() = 0;
+
+  // Reads the specified rectangle from the staging texture to out_ptr, with the specified stride
+  // (length in bytes of each row). CopyFromTexture must be called first. The contents of any
+  // texels outside of the rectangle used for CopyFromTexture is undefined.
+  void ReadTexels(const MathUtil::Rectangle<int>& rect, void* out_ptr, u32 out_stride);
+  void ReadTexel(u32 x, u32 y, void* out_ptr);
+
+  // Copies the texels from in_ptr to the staging texture, which can be read by the GPU, with the
+  // specified stride (length in bytes of each row). After updating the staging texture with all
+  // changes, call CopyToTexture() to update the GPU copy.
+  void WriteTexels(const MathUtil::Rectangle<int>& rect, const void* in_ptr, u32 in_stride);
+  void WriteTexel(u32 x, u32 y, const void* in_ptr);
+
+protected:
+  bool PrepareForAccess();
+
+  const StagingTextureType m_type;
+  const TextureConfig m_config;
+  const size_t m_texel_size;
+
+  char* m_map_pointer = nullptr;
+  size_t m_map_stride = 0;
+
+  bool m_needs_flush = false;
+};

Source/Core/VideoCommon/AbstractTexture.cpp

@@ -5,9 +5,11 @@
 #include <algorithm>
 
 #include "Common/Assert.h"
-
+#include "Common/MsgHandler.h"
+#include "VideoCommon/AbstractStagingTexture.h"
 #include "VideoCommon/AbstractTexture.h"
 #include "VideoCommon/ImageWrite.h"
+#include "VideoCommon/RenderBase.h"
 
 AbstractTexture::AbstractTexture(const TextureConfig& c) : m_config(c)
 {
@@ -20,93 +22,51 @@ bool AbstractTexture::Save(const std::string& filename, unsigned int level)
   // We can't dump compressed textures currently (it would mean drawing them to a RGBA8
   // framebuffer, and saving that). TextureCache does not call Save for custom textures
   // anyway, so this is fine for now.
-  _assert_(m_config.format == AbstractTextureFormat::RGBA8);
+  _assert_(!IsCompressedFormat(m_config.format));
+  _assert_(level < m_config.levels);
 
-  auto result = level == 0 ? Map() : Map(level);
+  // Determine dimensions of image we want to save.
+  u32 level_width = std::max(1u, m_config.width >> level);
+  u32 level_height = std::max(1u, m_config.height >> level);
 
-  if (!result.has_value())
+  // Use a temporary staging texture for the download. Certainly not optimal,
+  // but this is not a frequently-executed code path..
+  TextureConfig readback_texture_config(level_width, level_height, 1, 1,
+                                        AbstractTextureFormat::RGBA8, false);
+  auto readback_texture =
+      g_renderer->CreateStagingTexture(StagingTextureType::Readback, readback_texture_config);
+  if (!readback_texture)
+    return false;
+
+  // Copy to the readback texture's buffer.
+  readback_texture->CopyFromTexture(this, 0, level);
+  readback_texture->Flush();
+
+  // Map it so we can encode it to the file.
+  if (!readback_texture->Map())
+    return false;
+
+  return TextureToPng(reinterpret_cast<const u8*>(readback_texture->GetMappedPointer()),
+                      static_cast<int>(readback_texture->GetMappedStride()), filename, level_width,
+                      level_height);
+}
+
+bool AbstractTexture::IsCompressedFormat(AbstractTextureFormat format)
+{
+  switch (format)
   {
+  case AbstractTextureFormat::DXT1:
+  case AbstractTextureFormat::DXT3:
+  case AbstractTextureFormat::DXT5:
+  case AbstractTextureFormat::BPTC:
+    return true;
+
+  default:
     return false;
   }
-
-  auto raw_data = result.value();
-  return TextureToPng(raw_data.data, raw_data.stride, filename, raw_data.width, raw_data.height);
 }
 
-std::optional<AbstractTexture::RawTextureInfo> AbstractTexture::Map()
-{
-  if (m_currently_mapped)
-  {
-    Unmap();
-    m_currently_mapped = false;
-  }
-  auto result = MapFullImpl();
-
-  if (!result.has_value())
-  {
-    m_currently_mapped = false;
-    return {};
-  }
-
-  m_currently_mapped = true;
-  return result;
-}
-
-std::optional<AbstractTexture::RawTextureInfo> AbstractTexture::Map(u32 level, u32 x, u32 y,
-                                                                    u32 width, u32 height)
-{
-  _assert_(level < m_config.levels);
-
-  u32 max_level_width = std::max(m_config.width >> level, 1u);
-  u32 max_level_height = std::max(m_config.height >> level, 1u);
-
-  _assert_(width < max_level_width);
-  _assert_(height < max_level_height);
-
-  auto result = MapRegionImpl(level, x, y, width, height);
-
-  if (!result.has_value())
-  {
-    m_currently_mapped = false;
-    return {};
-  }
-
-  m_currently_mapped = true;
-  return result;
-}
-
-std::optional<AbstractTexture::RawTextureInfo> AbstractTexture::Map(u32 level)
-{
-  _assert_(level < m_config.levels);
-
-  u32 level_width = std::max(m_config.width >> level, 1u);
-  u32 level_height = std::max(m_config.height >> level, 1u);
-
-  return Map(level, 0, 0, level_width, level_height);
-}
-
-void AbstractTexture::Unmap()
-{
-}
-
-std::optional<AbstractTexture::RawTextureInfo> AbstractTexture::MapFullImpl()
-{
-  return {};
-}
-
-std::optional<AbstractTexture::RawTextureInfo>
-AbstractTexture::MapRegionImpl(u32 level, u32 x, u32 y, u32 width, u32 height)
-{
-  return {};
-}
-
-bool AbstractTexture::IsCompressedHostTextureFormat(AbstractTextureFormat format)
-{
-  // This will need to be changed if we add any other uncompressed formats.
-  return format != AbstractTextureFormat::RGBA8;
-}
-
-size_t AbstractTexture::CalculateHostTextureLevelPitch(AbstractTextureFormat format, u32 row_length)
+size_t AbstractTexture::CalculateStrideForFormat(AbstractTextureFormat format, u32 row_length)
 {
   switch (format)
   {
@@ -117,8 +77,30 @@ size_t AbstractTexture::CalculateHostTextureLevelPitch(AbstractTextureFormat for
   case AbstractTextureFormat::BPTC:
     return static_cast<size_t>(std::max(1u, row_length / 4)) * 16;
   case AbstractTextureFormat::RGBA8:
-  default:
+  case AbstractTextureFormat::BGRA8:
     return static_cast<size_t>(row_length) * 4;
+  default:
+    PanicAlert("Unhandled texture format.");
+    return 0;
+  }
+}
+
+size_t AbstractTexture::GetTexelSizeForFormat(AbstractTextureFormat format)
+{
+  switch (format)
+  {
+  case AbstractTextureFormat::DXT1:
+    return 8;
+  case AbstractTextureFormat::DXT3:
+  case AbstractTextureFormat::DXT5:
+  case AbstractTextureFormat::BPTC:
+    return 16;
+  case AbstractTextureFormat::RGBA8:
+  case AbstractTextureFormat::BGRA8:
+    return 4;
+  default:
+    PanicAlert("Unhandled texture format.");
+    return 0;
   }
 }
 

Source/Core/VideoCommon/AbstractTexture.h

@@ -5,7 +5,6 @@
 #pragma once
 
 #include <cstddef>
-#include <optional>
 #include <string>
 
 #include "Common/CommonTypes.h"
@@ -17,38 +16,27 @@ class AbstractTexture
 public:
   explicit AbstractTexture(const TextureConfig& c);
   virtual ~AbstractTexture();
+
   virtual void Bind(unsigned int stage) = 0;
-  bool Save(const std::string& filename, unsigned int level);
 
-  struct RawTextureInfo
-  {
-    const u8* data;
-    u32 stride;
-    u32 width;
-    u32 height;
-  };
-
-  std::optional<RawTextureInfo> Map();
-  std::optional<RawTextureInfo> Map(u32 level, u32 x, u32 y, u32 width, u32 height);
-  std::optional<RawTextureInfo> Map(u32 level);
-  virtual void Unmap();
-
-  virtual void CopyRectangleFromTexture(const AbstractTexture* source,
-                                        const MathUtil::Rectangle<int>& srcrect,
-                                        const MathUtil::Rectangle<int>& dstrect) = 0;
+  virtual void CopyRectangleFromTexture(const AbstractTexture* src,
+                                        const MathUtil::Rectangle<int>& src_rect, u32 src_layer,
+                                        u32 src_level, const MathUtil::Rectangle<int>& dst_rect,
+                                        u32 dst_layer, u32 dst_level) = 0;
+  virtual void ScaleRectangleFromTexture(const AbstractTexture* source,
+                                         const MathUtil::Rectangle<int>& srcrect,
+                                         const MathUtil::Rectangle<int>& dstrect) = 0;
   virtual void Load(u32 level, u32 width, u32 height, u32 row_length, const u8* buffer,
                     size_t buffer_size) = 0;
 
-  static bool IsCompressedHostTextureFormat(AbstractTextureFormat format);
-  static size_t CalculateHostTextureLevelPitch(AbstractTextureFormat format, u32 row_length);
+  bool Save(const std::string& filename, unsigned int level);
+
+  static bool IsCompressedFormat(AbstractTextureFormat format);
+  static size_t CalculateStrideForFormat(AbstractTextureFormat format, u32 row_length);
+  static size_t GetTexelSizeForFormat(AbstractTextureFormat format);
 
   const TextureConfig& GetConfig() const;
 
 protected:
-  virtual std::optional<RawTextureInfo> MapFullImpl();
-  virtual std::optional<RawTextureInfo> MapRegionImpl(u32 level, u32 x, u32 y, u32 width,
-                                                      u32 height);
-  bool m_currently_mapped = false;
-
   const TextureConfig m_config;
 };

Source/Core/VideoCommon/CMakeLists.txt

@@ -1,4 +1,5 @@
 set(SRCS
+  AbstractStagingTexture.cpp
   AbstractTexture.cpp
   AsyncRequests.cpp
   AsyncShaderCompiler.cpp

Source/Core/VideoCommon/MainBase.cpp

@@ -49,7 +49,7 @@ void VideoBackendBase::Video_CleanupShared()
 {
   // First stop any framedumping, which might need to dump the last xfb frame. This process
   // can require additional graphics sub-systems so it needs to be done first
-  g_renderer->ExitFramedumping();
+  g_renderer->ShutdownFrameDumping();
 
   Video_Cleanup();
 }

Source/Core/VideoCommon/RenderBase.cpp

@@ -43,6 +43,7 @@
 #include "Core/Movie.h"
 
 #include "VideoCommon/AVIDump.h"
+#include "VideoCommon/AbstractStagingTexture.h"
 #include "VideoCommon/AbstractTexture.h"
 #include "VideoCommon/BPMemory.h"
 #include "VideoCommon/CPMemory.h"
@@ -100,15 +101,6 @@ Renderer::Renderer(int backbuffer_width, int backbuffer_height)
 
 Renderer::~Renderer() = default;
 
-void Renderer::ExitFramedumping()
-{
-  ShutdownFrameDumping();
-  if (m_frame_dump_thread.joinable())
-    m_frame_dump_thread.join();
-
-  m_dump_texture.reset();
-}
-
 void Renderer::RenderToXFB(u32 xfbAddr, const EFBRectangle& sourceRc, u32 fbStride, u32 fbHeight,
                            float Gamma)
 {
@@ -635,14 +627,10 @@ void Renderer::Swap(u32 xfbAddr, u32 fbWidth, u32 fbStride, u32 fbHeight, const
       m_aspect_wide = flush_count_anamorphic > 0.75 * flush_total;
   }
 
-  if (IsFrameDumping() && m_last_xfb_texture)
-  {
-    FinishFrameData();
-  }
-  else
-  {
-    ShutdownFrameDumping();
-  }
+  // Ensure the last frame was written to the dump.
+  // This is required even if frame dumping has stopped, since the frame dump is one frame
+  // behind the renderer.
+  FlushFrameDump();
 
   bool update_frame_count = false;
   if (xfbAddr && fbWidth && fbStride && fbHeight)
@@ -668,10 +656,9 @@ void Renderer::Swap(u32 xfbAddr, u32 fbWidth, u32 fbStride, u32 fbHeight, const
 
       m_fps_counter.Update();
       update_frame_count = true;
+
       if (IsFrameDumping())
-      {
-        DoDumpFrame();
-      }
+        DumpCurrentFrame();
     }
 
     // Update our last xfb values
@@ -701,20 +688,16 @@ bool Renderer::IsFrameDumping()
   return false;
 }
 
-void Renderer::DoDumpFrame()
+void Renderer::DumpCurrentFrame()
 {
-  UpdateFrameDumpTexture();
+  // Scale/render to frame dump texture.
+  RenderFrameDump();
 
-  auto result = m_dump_texture->Map();
-  if (result.has_value())
-  {
-    auto raw_data = result.value();
-    DumpFrameData(raw_data.data, raw_data.width, raw_data.height, raw_data.stride,
-                  AVIDump::FetchState(m_last_xfb_ticks));
-  }
+  // Queue a readback for the next frame.
+  QueueFrameDumpReadback();
 }
 
-void Renderer::UpdateFrameDumpTexture()
+void Renderer::RenderFrameDump()
 {
   int target_width, target_height;
   if (!g_ActiveConfig.bInternalResolutionFrameDumps && !IsHeadless())
@@ -729,33 +712,99 @@ void Renderer::UpdateFrameDumpTexture()
         m_last_xfb_texture->GetConfig().width, m_last_xfb_texture->GetConfig().height);
   }
 
-  if (m_dump_texture == nullptr ||
-      m_dump_texture->GetConfig().width != static_cast<u32>(target_width) ||
-      m_dump_texture->GetConfig().height != static_cast<u32>(target_height))
+  // Ensure framebuffer exists (we lazily allocate it in case frame dumping isn't used).
+  // Or, resize texture if it isn't large enough to accommodate the current frame.
+  if (!m_frame_dump_render_texture ||
+      m_frame_dump_render_texture->GetConfig().width != static_cast<u32>(target_width) ||
+      m_frame_dump_render_texture->GetConfig().height == static_cast<u32>(target_height))
   {
-    TextureConfig config;
-    config.width = target_width;
-    config.height = target_height;
-    config.rendertarget = true;
-    m_dump_texture = g_texture_cache->CreateTexture(config);
+    // Recreate texture objects. Release before creating so we don't temporarily use twice the RAM.
+    TextureConfig config(target_width, target_height, 1, 1, AbstractTextureFormat::RGBA8, true);
+    m_frame_dump_render_texture.reset();
+    m_frame_dump_render_texture = CreateTexture(config);
+    _assert_(m_frame_dump_render_texture);
   }
-  m_dump_texture->CopyRectangleFromTexture(m_last_xfb_texture, m_last_xfb_region,
-                                           EFBRectangle{0, 0, target_width, target_height});
+
+  // Scaling is likely to occur here, but if possible, do a bit-for-bit copy.
+  if (m_last_xfb_region.GetWidth() != target_width ||
+      m_last_xfb_region.GetHeight() != target_height)
+  {
+    m_frame_dump_render_texture->ScaleRectangleFromTexture(
+        m_last_xfb_texture, m_last_xfb_region, EFBRectangle{0, 0, target_width, target_height});
+  }
+  else
+  {
+    m_frame_dump_render_texture->CopyRectangleFromTexture(
+        m_last_xfb_texture, m_last_xfb_region, 0, 0,
+        EFBRectangle{0, 0, target_width, target_height}, 0, 0);
+  }
+}
+
+void Renderer::QueueFrameDumpReadback()
+{
+  // Index 0 was just sent to AVI dump. Swap with the second texture.
+  if (m_frame_dump_readback_textures[0])
+    std::swap(m_frame_dump_readback_textures[0], m_frame_dump_readback_textures[1]);
+
+  std::unique_ptr<AbstractStagingTexture>& rbtex = m_frame_dump_readback_textures[0];
+  if (!rbtex || rbtex->GetConfig() != m_frame_dump_render_texture->GetConfig())
+  {
+    rbtex = CreateStagingTexture(StagingTextureType::Readback,
+                                 m_frame_dump_render_texture->GetConfig());
+  }
+
+  m_last_frame_state = AVIDump::FetchState(m_last_xfb_ticks);
+  m_last_frame_exported = true;
+  rbtex->CopyFromTexture(m_frame_dump_render_texture.get(), 0, 0);
+}
+
+void Renderer::FlushFrameDump()
+{
+  if (!m_last_frame_exported)
+    return;
+
+  // Ensure the previously-queued frame was encoded.
+  FinishFrameData();
+
+  // Queue encoding of the last frame dumped.
+  std::unique_ptr<AbstractStagingTexture>& rbtex = m_frame_dump_readback_textures[0];
+  rbtex->Flush();
+  if (rbtex->Map())
+  {
+    DumpFrameData(reinterpret_cast<u8*>(rbtex->GetMappedPointer()), rbtex->GetConfig().width,
+                  rbtex->GetConfig().height, static_cast<int>(rbtex->GetMappedStride()),
+                  m_last_frame_state);
+    rbtex->Unmap();
+  }
+
+  m_last_frame_exported = false;
+
+  // Shutdown frame dumping if it is no longer active.
+  if (!IsFrameDumping())
+    ShutdownFrameDumping();
 }
 
 void Renderer::ShutdownFrameDumping()
 {
+  // Ensure the last queued readback has been sent to the encoder.
+  FlushFrameDump();
+
   if (!m_frame_dump_thread_running.IsSet())
     return;
 
+  // Ensure previous frame has been encoded.
   FinishFrameData();
+
+  // Wake thread up, and wait for it to exit.
   m_frame_dump_thread_running.Clear();
   m_frame_dump_start.Set();
+  if (m_frame_dump_thread.joinable())
+    m_frame_dump_thread.join();
 }
 
 void Renderer::DumpFrameData(const u8* data, int w, int h, int stride, const AVIDump::Frame& state)
 {
-  m_frame_dump_config = FrameDumpConfig{m_last_xfb_texture, data, w, h, stride, state};
+  m_frame_dump_config = FrameDumpConfig{data, w, h, stride, state};
 
   if (!m_frame_dump_thread_running.IsSet())
   {
@@ -765,6 +814,7 @@ void Renderer::DumpFrameData(const u8* data, int w, int h, int stride, const AVI
     m_frame_dump_thread = std::thread(&Renderer::RunFrameDumps, this);
   }
 
+  // Wake worker thread up.
   m_frame_dump_start.Set();
   m_frame_dump_frame_running = true;
 }
@@ -776,7 +826,6 @@ void Renderer::FinishFrameData()
 
   m_frame_dump_done.Wait();
   m_frame_dump_frame_running = false;
-  m_frame_dump_config.texture->Unmap();
 }
 
 void Renderer::RunFrameDumps()

Source/Core/VideoCommon/RenderBase.h

@@ -34,8 +34,11 @@
 
 class AbstractRawTexture;
 class AbstractTexture;
+class AbstractStagingTexture;
 class PostProcessingShaderImplementation;
+struct TextureConfig;
 enum class EFBAccessType;
+enum class StagingTextureType;
 
 struct EfbPokeData
 {
@@ -79,6 +82,10 @@ public:
   virtual void RestoreState() {}
   virtual void ResetAPIState() {}
   virtual void RestoreAPIState() {}
+  virtual std::unique_ptr<AbstractTexture> CreateTexture(const TextureConfig& config) = 0;
+  virtual std::unique_ptr<AbstractStagingTexture>
+  CreateStagingTexture(StagingTextureType type, const TextureConfig& config) = 0;
+
   // Ideal internal resolution - multiple of the native EFB resolution
   int GetTargetWidth() const { return m_target_width; }
   int GetTargetHeight() const { return m_target_height; }
@@ -145,7 +152,7 @@ public:
   virtual void ChangeSurface(void* new_surface_handle) {}
   bool UseVertexDepthRange() const;
 
-  void ExitFramedumping();
+  void ShutdownFrameDumping();
 
 protected:
   std::tuple<int, int> CalculateTargetScale(int x, int y) const;
@@ -185,11 +192,8 @@ protected:
   u32 m_last_host_config_bits = 0;
 
 private:
-  void DoDumpFrame();
   void RunFrameDumps();
-  void ShutdownFrameDumping();
   std::tuple<int, int> CalculateOutputDimensions(int width, int height);
-  void UpdateFrameDumpTexture();
 
   PEControl::PixelFormat m_prev_efb_format = PEControl::INVALID_FMT;
   unsigned int m_efb_scale = 1;
@@ -207,7 +211,6 @@ private:
   bool m_frame_dump_frame_running = false;
   struct FrameDumpConfig
   {
-    AbstractTexture* texture;
     const u8* data;
     int width;
     int height;
@@ -215,13 +218,18 @@ private:
     AVIDump::Frame state;
   } m_frame_dump_config;
 
+  // Texture used for screenshot/frame dumping
+  std::unique_ptr<AbstractTexture> m_frame_dump_render_texture;
+  std::array<std::unique_ptr<AbstractStagingTexture>, 2> m_frame_dump_readback_textures;
+  AVIDump::Frame m_last_frame_state;
+  bool m_last_frame_exported = false;
+
+  // Tracking of XFB textures so we don't render duplicate frames.
   AbstractTexture* m_last_xfb_texture = nullptr;
   u64 m_last_xfb_id = std::numeric_limits<u64>::max();
   u64 m_last_xfb_ticks = 0;
   EFBRectangle m_last_xfb_region;
 
-  std::unique_ptr<AbstractTexture> m_dump_texture;
-
   // Note: Only used for auto-ir
   u32 m_last_xfb_width = MAX_XFB_WIDTH;
   u32 m_last_xfb_height = MAX_XFB_HEIGHT;
@@ -235,7 +243,23 @@ private:
   void DumpFrameToImage(const FrameDumpConfig& config);
 
   bool IsFrameDumping();
+
+  // Asynchronously encodes the current staging texture to the frame dump.
+  void DumpCurrentFrame();
+
+  // Fills the frame dump render texture with the current XFB texture.
+  void RenderFrameDump();
+
+  // Queues the current frame for readback, which will be written to AVI next frame.
+  void QueueFrameDumpReadback();
+
+  // Asynchronously encodes the specified pointer of frame data to the frame dump.
   void DumpFrameData(const u8* data, int w, int h, int stride, const AVIDump::Frame& state);
+
+  // Ensures all rendered frames are queued for encoding.
+  void FlushFrameDump();
+
+  // Ensures all encoded frames have been written to the output file.
   void FinishFrameData();
 };
 

Source/Core/VideoCommon/TextureCacheBase.cpp

@@ -277,9 +277,9 @@ void TextureCacheBase::ScaleTextureCacheEntryTo(TextureCacheBase::TCacheEntry* e
   std::unique_ptr<AbstractTexture> new_texture = AllocateTexture(newconfig);
   if (new_texture)
   {
-    new_texture->CopyRectangleFromTexture(entry->texture.get(),
-                                          entry->texture->GetConfig().GetRect(),
-                                          new_texture->GetConfig().GetRect());
+    new_texture->ScaleRectangleFromTexture(entry->texture.get(),
+                                           entry->texture->GetConfig().GetRect(),
+                                           new_texture->GetConfig().GetRect());
     entry->texture.swap(new_texture);
 
     auto config = new_texture->GetConfig();
@@ -406,7 +406,11 @@ TextureCacheBase::DoPartialTextureUpdates(TCacheEntry* entry_to_update, u8* pale
         dstrect.top = dst_y;
         dstrect.right = (dst_x + copy_width);
         dstrect.bottom = (dst_y + copy_height);
-        entry_to_update->texture->CopyRectangleFromTexture(entry->texture.get(), srcrect, dstrect);
+        for (u32 layer = 0; layer < entry->texture->GetConfig().layers; layer++)
+        {
+          entry_to_update->texture->CopyRectangleFromTexture(entry->texture.get(), srcrect, layer,
+                                                             0, dstrect, layer, 0);
+        }
 
         if (isPaletteTexture)
         {
@@ -1366,33 +1370,16 @@ bool TextureCacheBase::LoadTextureFromOverlappingTextures(TCacheEntry* entry_to_
         srcrect.right = (src_x + copy_width);
         srcrect.bottom = (src_y + copy_height);
 
-        if (static_cast<int>(entry->GetWidth()) == srcrect.GetWidth())
-        {
-          srcrect.right -= 1;
-        }
-
-        if (static_cast<int>(entry->GetHeight()) == srcrect.GetHeight())
-        {
-          srcrect.bottom -= 1;
-        }
-
         dstrect.left = dst_x;
         dstrect.top = dst_y;
         dstrect.right = (dst_x + copy_width);
         dstrect.bottom = (dst_y + copy_height);
 
-        if (static_cast<int>(entry_to_update->GetWidth()) == dstrect.GetWidth())
+        for (u32 layer = 0; layer < entry->texture->GetConfig().layers; layer++)
         {
-          dstrect.right -= 1;
+          entry_to_update->texture->CopyRectangleFromTexture(entry->texture.get(), srcrect, layer,
+                                                             0, dstrect, layer, 0);
         }
-
-        if (static_cast<int>(entry_to_update->GetHeight()) == dstrect.GetHeight())
-        {
-          dstrect.bottom -= 1;
-        }
-
-        entry_to_update->texture->CopyRectangleFromTexture(entry->texture.get(), srcrect, dstrect);
-
         updated_entry = true;
 
         if (tex_info.is_palette_texture)
@@ -2090,7 +2077,7 @@ std::unique_ptr<AbstractTexture> TextureCacheBase::AllocateTexture(const Texture
   }
   else
   {
-    entry = CreateTexture(config);
+    entry = g_renderer->CreateTexture(config);
     if (!entry)
       return nullptr;
 

Source/Core/VideoCommon/TextureCacheBase.h

@@ -273,8 +273,6 @@ public:
 
   void ScaleTextureCacheEntryTo(TCacheEntry* entry, u32 new_width, u32 new_height);
 
-  virtual std::unique_ptr<AbstractTexture> CreateTexture(const TextureConfig& config) = 0;
-
 protected:
   TextureCacheBase();
 

Source/Core/VideoCommon/TextureConfig.cpp

@@ -3,6 +3,7 @@
 // Refer to the license.txt file included.
 
 #include "VideoCommon/TextureConfig.h"
+#include "VideoCommon/AbstractTexture.h"
 
 #include <tuple>
 
@@ -12,7 +13,28 @@ bool TextureConfig::operator==(const TextureConfig& o) const
          std::tie(o.width, o.height, o.levels, o.layers, o.format, o.rendertarget);
 }
 
+bool TextureConfig::operator!=(const TextureConfig& o) const
+{
+  return !operator==(o);
+}
+
 MathUtil::Rectangle<int> TextureConfig::GetRect() const
 {
   return {0, 0, static_cast<int>(width), static_cast<int>(height)};
 }
+
+MathUtil::Rectangle<int> TextureConfig::GetMipRect(u32 level) const
+{
+  return {0, 0, static_cast<int>(std::max(width >> level, 1u)),
+          static_cast<int>(std::max(height >> level, 1u))};
+}
+
+size_t TextureConfig::GetStride() const
+{
+  return AbstractTexture::CalculateStrideForFormat(format, width);
+}
+
+size_t TextureConfig::GetMipStride(u32 level) const
+{
+  return AbstractTexture::CalculateStrideForFormat(format, std::max(width >> level, 1u));
+}

Source/Core/VideoCommon/TextureConfig.h

@@ -13,17 +13,37 @@
 enum class AbstractTextureFormat : u32
 {
   RGBA8,
+  BGRA8,
   DXT1,
   DXT3,
   DXT5,
-  BPTC
+  BPTC,
+  Undefined
+};
+
+enum class StagingTextureType
+{
+  Readback,  // Optimize for CPU reads, GPU writes, no CPU writes
+  Upload,    // Optimize for CPU writes, GPU reads, no CPU reads
+  Mutable    // Optimize for CPU reads, GPU writes, allow slow CPU reads
 };
 
 struct TextureConfig
 {
   constexpr TextureConfig() = default;
+  constexpr TextureConfig(u32 width_, u32 height_, u32 levels_, u32 layers_,
+                          AbstractTextureFormat format_, bool rendertarget_)
+      : width(width_), height(height_), levels(levels_), layers(layers_), format(format_),
+        rendertarget(rendertarget_)
+  {
+  }
+
   bool operator==(const TextureConfig& o) const;
+  bool operator!=(const TextureConfig& o) const;
   MathUtil::Rectangle<int> GetRect() const;
+  MathUtil::Rectangle<int> GetMipRect(u32 level) const;
+  size_t GetStride() const;
+  size_t GetMipStride(u32 level) const;
 
   u32 width = 0;
   u32 height = 0;

Source/Core/VideoCommon/VideoCommon.vcxproj

@@ -36,6 +36,7 @@
   </ImportGroup>
   <PropertyGroup Label="UserMacros" />
   <ItemGroup>
+    <ClCompile Include="AbstractStagingTexture.cpp" />
     <ClCompile Include="AbstractTexture.cpp" />
     <ClCompile Include="AsyncRequests.cpp" />
     <ClCompile Include="AsyncShaderCompiler.cpp" />
@@ -96,6 +97,7 @@
     <ClCompile Include="XFStructs.cpp" />
   </ItemGroup>
   <ItemGroup>
+    <ClInclude Include="AbstractStagingTexture.h" />
     <ClInclude Include="AbstractTexture.h" />
     <ClInclude Include="AsyncRequests.h" />
     <ClInclude Include="AsyncShaderCompiler.h" />

Source/Core/VideoCommon/VideoCommon.vcxproj.filters

@@ -191,6 +191,9 @@
     <ClCompile Include="UberShaderVertex.cpp">
       <Filter>Shader Generators</Filter>
     </ClCompile>
+    <ClCompile Include="AbstractStagingTexture.cpp">
+      <Filter>Base</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="CommandProcessor.h" />
@@ -362,6 +365,9 @@
     <ClInclude Include="UberShaderVertex.h">
       <Filter>Shader Generators</Filter>
     </ClInclude>
+    <ClInclude Include="AbstractStagingTexture.h">
+      <Filter>Base</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <Text Include="CMakeLists.txt" />