guest_gpu: remove memory-unmap submission deadlock + remove legacy agc buffering

Extracted from 3db2b3c5c5e1a26a861df7ebcacd9ccb8c484420 in KytyPS5/KytyPS5#147.

.github/workflows/build-linux.yml

@@ -1,82 +0,0 @@
-name: Build KytyPS5 (Linux)
-
-on:
-  workflow_dispatch:
-  push:
-    branches: [ main, master ]
-  pull_request:
-
-jobs:
-  build:
-    runs-on: ubuntu-24.04
-
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-        with:
-          submodules: recursive
-
-      - name: Install build dependencies
-        shell: bash
-        run: |
-          sudo apt-get update
-          sudo apt-get install --no-install-recommends --yes \
-            clang \
-            glslang-tools \
-            libgl1-mesa-dev \
-            libx11-dev \
-            libxcursor-dev \
-            libxext-dev \
-            libxfixes-dev \
-            libxi-dev \
-            libxrandr-dev \
-            libxss-dev \
-            lld \
-            ninja-build
-
-      - name: Install Qt 6.10.3
-        uses: jurplel/install-qt-action@v4
-        with:
-          version: "6.10.3"
-          host: linux
-          target: desktop
-          arch: linux_gcc_64
-          cache: true
-
-      - name: Verify toolchain
-        shell: bash
-        run: |
-          git --version
-          cmake --version
-          ninja --version
-          clang++ --version
-          ld.lld --version
-          glslangValidator --version
-          echo "Qt6_DIR=$Qt6_DIR"
-
-      - name: Configure
-        shell: bash
-        run: |
-          cmake -S src -B _Build/linux \
-            -G Ninja \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DCMAKE_C_COMPILER=clang \
-            -DCMAKE_CXX_COMPILER=clang++ \
-            -DCMAKE_PREFIX_PATH="$Qt6_DIR"
-
-      - name: Build
-        shell: bash
-        run: |
-          cmake --build _Build/linux --target launcher --parallel
-
-      - name: Install
-        shell: bash
-        run: |
-          cmake --install _Build/linux --prefix _Build/linux/install
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          name: KytyPS5-Linux
-          path: _Build/linux/install/**
-          if-no-files-found: error

.github/workflows/build.yml

@@ -1,4 +1,4 @@
-name: Build KytyPS5
+name: Build and Release KytyPS5
 
 on:
   workflow_dispatch:
@@ -7,7 +7,8 @@ on:
   pull_request:
 
 jobs:
-  build:
+  windows:
+    name: Build KytyPS5 (Windows)
     runs-on: windows-2022
 
     steps:
@@ -22,11 +23,31 @@ jobs:
       - name: Setup Ninja
         uses: seanmiddleditch/gha-setup-ninja@v5
 
-      - name: Install glslang
+      - name: Locate vcpkg
+        id: vcpkg
         shell: pwsh
         run: |
           $vcpkgRoot = Split-Path (Get-Command vcpkg).Source
-          vcpkg install glslang[tools,opt]:x64-windows
+          $portHash = (Get-FileHash "$vcpkgRoot\ports\glslang\vcpkg.json" -Algorithm SHA256).Hash
+          "root=$vcpkgRoot" |
+            Out-File -FilePath $env:GITHUB_OUTPUT -Encoding utf8 -Append
+          "port_hash=$($portHash.ToLowerInvariant())" |
+            Out-File -FilePath $env:GITHUB_OUTPUT -Encoding utf8 -Append
+
+      - name: Cache glslang
+        id: glslang_cache
+        uses: actions/cache@v4
+        with:
+          path: ${{ steps.vcpkg.outputs.root }}\installed\x64-windows
+          key: glslang-${{ runner.os }}-${{ runner.arch }}-${{ steps.vcpkg.outputs.port_hash }}-tools-opt
+
+      - name: Install glslang
+        shell: pwsh
+        run: |
+          $vcpkgRoot = "${{ steps.vcpkg.outputs.root }}"
+          if ("${{ steps.glslang_cache.outputs.cache-hit }}" -ne "true") {
+            vcpkg install glslang[tools,opt]:x64-windows
+          }
           "$vcpkgRoot\installed\x64-windows\tools\glslang" |
             Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
 
@@ -62,52 +83,323 @@ jobs:
       - name: Build
         shell: cmd
         run: |
-          cmake --build _Build/windows --target launcher --parallel
+          cmake --build _Build/windows --target launcher audio_out2_port_tests virtual_memory_allocation_tests --parallel
+
+      - name: Test
+        shell: cmd
+        run: |
+          ctest --test-dir _Build/windows --output-on-failure -R "^(audio_out2_port|virtual_memory_allocation)$"
 
       - name: Install
         shell: cmd
         run: |
           cmake --install _Build/windows --prefix _Build/windows/install
 
-      - name: Upload Artifacts
+      - name: Upload Windows artifacts
         uses: actions/upload-artifact@v4
         with:
-          name: KytyPS5
+          name: KytyPS5-Windows-x64
           path: _Build/windows/install/**
           if-no-files-found: error
 
+  macos:
+    name: Build KytyPS5 (macOS)
+    runs-on: macos-15
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          submodules: recursive
+
+      - name: Select Xcode 26
+        shell: bash
+        run: |
+          xcode_path="$(ls -d /Applications/Xcode_26*.app | tail -n 1)"
+          test -n "$xcode_path"
+          sudo xcode-select --switch "$xcode_path/Contents/Developer"
+          xcodebuild -version
+
+      - name: Install build dependencies
+        shell: bash
+        run: brew install glslang ninja
+
+      - name: Install Qt 6.10.3
+        uses: jurplel/install-qt-action@v4
+        with:
+          version: "6.10.3"
+          host: mac
+          target: desktop
+          arch: clang_64
+          cache: true
+
+      - name: Verify toolchain
+        shell: bash
+        run: |
+          git --version
+          cmake --version
+          ninja --version
+          clang++ --version
+          glslangValidator --version
+          echo "Host architecture: $(uname -m)"
+          echo "QT_ROOT_DIR=$QT_ROOT_DIR"
+
+      - name: Configure
+        shell: bash
+        run: |
+          cmake -S src -B _Build/macos \
+            -G Ninja \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_OSX_ARCHITECTURES=x86_64 \
+            -DCMAKE_C_COMPILER=clang \
+            -DCMAKE_CXX_COMPILER=clang++ \
+            -DCMAKE_PREFIX_PATH="$QT_ROOT_DIR"
+
+      - name: Build
+        shell: bash
+        run: |
+          cmake --build _Build/macos \
+            --target launcher audio_out2_port_tests virtual_memory_allocation_tests \
+            --parallel
+
+      - name: Test
+        shell: bash
+        run: |
+          ctest --test-dir _Build/macos --output-on-failure \
+            -R '^(audio_out2_port|virtual_memory_allocation)$'
+
+      - name: Install
+        shell: bash
+        run: |
+          cmake --install _Build/macos --prefix _Build/macos/install
+
+      - name: Bundle MoltenVK
+        shell: bash
+        env:
+          MOLTENVK_VERSION: v1.4.2
+          MOLTENVK_SHA256: f95765a6229cb7b915990a2890ce12ebe36a730b021545d3d52ae69ce4c4024e
+        run: |
+          archive="$RUNNER_TEMP/MoltenVK-macos.tar"
+          package="$RUNNER_TEMP/MoltenVK"
+          curl --fail --location --retry 3 \
+            --output "$archive" \
+            "https://github.com/KhronosGroup/MoltenVK/releases/download/$MOLTENVK_VERSION/MoltenVK-macos.tar"
+          echo "$MOLTENVK_SHA256  $archive" | shasum -a 256 --check
+          tar -xf "$archive" -C "$RUNNER_TEMP"
+          install -m 755 \
+            "$package/MoltenVK/dynamic/dylib/macOS/libMoltenVK.dylib" \
+            _Build/macos/install/libMoltenVK.dylib
+          install -m 644 "$package/LICENSE" _Build/macos/install/LICENSE.MoltenVK
+          codesign --force --sign - --timestamp=none _Build/macos/install/libMoltenVK.dylib
+
+      - name: Verify artifacts
+        shell: bash
+        run: |
+          file _Build/macos/install/launcher
+          file _Build/macos/install/kyty_emulator
+          file _Build/macos/install/libMoltenVK.dylib
+          lipo _Build/macos/install/launcher -verify_arch x86_64
+          lipo _Build/macos/install/kyty_emulator -verify_arch x86_64
+          lipo _Build/macos/install/libMoltenVK.dylib -verify_arch x86_64
+          codesign --verify --strict _Build/macos/install/kyty_emulator
+          codesign --verify --strict _Build/macos/install/libMoltenVK.dylib
+
+      - name: Upload macOS artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: KytyPS5-macOS-x86_64
+          path: _Build/macos/install/**
+          if-no-files-found: error
+
+  linux:
+    name: Build KytyPS5 (Linux)
+    runs-on: ubuntu-24.04
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          submodules: recursive
+
+      - name: Install build dependencies
+        shell: bash
+        run: |
+          sudo apt-get update
+          sudo apt-get install --no-install-recommends --yes \
+            clang \
+            glslang-tools \
+            libasound2-dev \
+            libdbus-1-dev \
+            libgl1-mesa-dev \
+            libpulse-dev \
+            libudev-dev \
+            libwayland-dev \
+            libx11-dev \
+            libxcursor-dev \
+            libxext-dev \
+            libxfixes-dev \
+            libxi-dev \
+            libxkbcommon-dev \
+            libxrandr-dev \
+            libxss-dev \
+            lld \
+            ninja-build \
+            wayland-protocols
+
+      - name: Install Qt 6.10.3
+        uses: jurplel/install-qt-action@v4
+        with:
+          version: "6.10.3"
+          host: linux
+          target: desktop
+          arch: linux_gcc_64
+          cache: true
+
+      - name: Verify toolchain
+        shell: bash
+        run: |
+          git --version
+          cmake --version
+          ninja --version
+          clang++ --version
+          ld.lld --version
+          glslangValidator --version
+          echo "Qt6_DIR=$Qt6_DIR"
+
+      - name: Configure
+        shell: bash
+        run: |
+          mkdir -p _Build
+          cmake -S src -B _Build/linux \
+            -G Ninja \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_C_COMPILER=clang \
+            -DCMAKE_CXX_COMPILER=clang++ \
+            -DCMAKE_PREFIX_PATH="$Qt6_DIR" 2>&1 | tee _Build/configure.log
+          exit "${PIPESTATUS[0]}"
+
+      - name: Verify SDL2 backends
+        shell: bash
+        run: |
+          status=0
+          for feature in SDL_ALSA SDL_PULSEAUDIO SDL_WAYLAND SDL_X11 SDL_LIBUDEV SDL_DBUS; do
+            if grep -qE "^--   ${feature} +\\(Wanted: ON\\): ON" _Build/configure.log; then
+              echo "ok   ${feature}"
+            else
+              echo "FAIL ${feature} is not enabled"
+              status=1
+            fi
+          done
+          exit "$status"
+
+      - name: Build
+        shell: bash
+        run: |
+          cmake --build _Build/linux \
+            --target launcher page_manager_tests memory_tracker_tests \
+              audio_out2_port_tests virtual_memory_allocation_tests \
+            --parallel
+
+      - name: Test
+        shell: bash
+        run: |
+          ctest --test-dir _Build/linux --output-on-failure \
+            -R '^(audio_out2_port|page_manager|memory_tracker|virtual_memory_allocation)$'
+
+      - name: Install
+        shell: bash
+        run: |
+          cmake --install _Build/linux --prefix _Build/linux/install
+
+      - name: Verify artifacts
+        shell: bash
+        run: |
+          file _Build/linux/install/launcher
+          file _Build/linux/install/kyty_emulator
+          file _Build/linux/install/kyty_emulator | grep -q "ELF 64-bit LSB .*x86-64"
+          ldd _Build/linux/install/kyty_emulator > /dev/null
+          readelf -d _Build/linux/install/launcher | grep -q 'RPATH.*\$ORIGIN/lib'
+          while IFS= read -r -d '' binary; do
+            while read -r dependency; do
+              test -e "_Build/linux/install/lib/$dependency"
+            done < <(
+              readelf -d "$binary" |
+                sed -n 's/.*Shared library: \[\(libQt6[^]]*\|libicu[^]]*\)\].*/\1/p'
+            )
+          done < <(
+            find _Build/linux/install/launcher _Build/linux/install/plugins \
+              -type f \( -name launcher -o -name '*.so' \) -print0
+          )
+          _Build/linux/install/kyty_emulator --help > /dev/null
+
+      - name: Upload Linux artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: KytyPS5-Linux-x86_64
+          path: _Build/linux/install/**
+          if-no-files-found: error
+
   release:
-    if: github.event_name == 'push'
-    needs: build
+    name: Release KytyPS5
+    if: github.event_name == 'push' && github.repository == 'KytyPS5/KytyPS5'
+    needs: [windows, macos, linux]
     runs-on: ubuntu-latest
     permissions:
       contents: write
 
     steps:
-      - name: Download build
+      - name: Download builds
         uses: actions/download-artifact@v4
         with:
-          name: KytyPS5
-          path: KytyPS5
+          path: artifacts
 
       - name: Set release name
         shell: bash
         run: |
           echo "RELEASE_NAME=KytyPS5-$(date -u +'%Y-%m-%d')-${GITHUB_SHA::7}" >> "$GITHUB_ENV"
 
-      - name: Package build
+      - name: Package builds
         shell: bash
-        run: zip -r "$RELEASE_NAME.zip" KytyPS5
+        run: |
+          windows_dir="artifacts/KytyPS5-Windows-x64"
+          macos_dir="artifacts/KytyPS5-macOS-x86_64"
+          linux_dir="artifacts/KytyPS5-Linux-x86_64"
+          test -d "$windows_dir"
+          test -d "$macos_dir"
+          test -d "$linux_dir"
+          chmod a+x \
+            "$macos_dir/launcher" \
+            "$macos_dir/kyty_emulator" \
+            "$macos_dir/libMoltenVK.dylib" \
+            "$linux_dir/launcher" \
+            "$linux_dir/kyty_emulator"
+          (
+            cd "$windows_dir"
+            zip -r "$GITHUB_WORKSPACE/$RELEASE_NAME-Windows-x64.zip" .
+          )
+          (
+            cd "$macos_dir"
+            zip -r "$GITHUB_WORKSPACE/$RELEASE_NAME-macOS-x86_64.zip" .
+          )
+          tar -C "$linux_dir" -czf \
+            "$GITHUB_WORKSPACE/$RELEASE_NAME-Linux-x86_64.tar.gz" .
 
       - name: Create release
         shell: bash
         env:
           GH_TOKEN: ${{ github.token }}
         run: |
+          assets=(
+            "$RELEASE_NAME-Windows-x64.zip"
+            "$RELEASE_NAME-macOS-x86_64.zip"
+            "$RELEASE_NAME-Linux-x86_64.tar.gz"
+          )
           if gh release view "$RELEASE_NAME" --repo "$GITHUB_REPOSITORY" > /dev/null 2>&1; then
-            gh release upload "$RELEASE_NAME" "$RELEASE_NAME.zip" --clobber --repo "$GITHUB_REPOSITORY"
+            gh release upload "$RELEASE_NAME" "${assets[@]}" \
+              --clobber \
+              --repo "$GITHUB_REPOSITORY"
           else
-            gh release create "$RELEASE_NAME" "$RELEASE_NAME.zip" \
+            gh release create "$RELEASE_NAME" "${assets[@]}" \
               --generate-notes \
               --repo "$GITHUB_REPOSITORY" \
               --target "$GITHUB_SHA" \

.gitignore

@@ -2,4 +2,5 @@
 .vs/
 .idea/
 build/
-_Build/vscode-clang/
+_Build/vscode-clang/
+_Build/

README.md

@@ -1,15 +1,16 @@
 # KytyPS5
 
-[![Windows Build](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml/badge.svg)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
-[![Linux Build](https://github.com/KytyPS5/KytyPS5/actions/workflows/build-linux.yml/badge.svg)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build-linux.yml)
-[![Platform](https://img.shields.io/badge/platform-Windows%20x64-0078D4.svg)](#system-requirements)
+[![Build KytyPS5 (Windows)](https://img.shields.io/github/actions/workflow/status/KytyPS5/KytyPS5/build.yml?branch=main&event=push&label=Build%20KytyPS5%20%28Windows%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
+[![Build KytyPS5 (Linux)](https://img.shields.io/github/actions/workflow/status/KytyPS5/KytyPS5/build.yml?branch=main&event=push&label=Build%20KytyPS5%20%28Linux%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
+[![Build KytyPS5 (macOS)](https://img.shields.io/github/actions/workflow/status/KytyPS5/KytyPS5/build.yml?branch=main&event=push&label=Build%20KytyPS5%20%28macOS%29)](https://github.com/KytyPS5/KytyPS5/actions/workflows/build.yml)
+[![Platform](https://img.shields.io/badge/platform-Windows%20x64%20%7C%20Linux%20x64%20%7C%20macOS%20x86__64-0078D4.svg)](#system-requirements)
 [![Status](https://img.shields.io/badge/status-early%20development-orange.svg)](#current-status)
 [![License](https://img.shields.io/badge/license-GPL--2.0-blue.svg)](LICENSE)
 
-KytyPS5 is a free and open-source PlayStation 5 emulator written in C++ for Windows. It is based on
-a heavily modified version of [Kyty](https://github.com/InoriRus/Kyty). The project is in an early
-stage of development, so compatibility is limited and behavior may change significantly between
-builds.
+KytyPS5 is a free and open-source PlayStation 5 emulator written in C++ for Windows and Linux,
+with experimental macOS support. It is based on a heavily modified version of
+[Kyty](https://github.com/InoriRus/Kyty). The project is in an early stage of development, so
+compatibility is limited and behavior may change significantly between builds.
 
 > [!IMPORTANT]
 > KytyPS5 is not affiliated with Sony Interactive Entertainment or PlayStation. The project does
@@ -23,7 +24,12 @@ KytyPS5 can boot 2D games and a selection of 3D games, including titles built wi
 
 Development is focused on compatibility and boot reliability.
 
-Linux support is planned, but Windows is the only supported platform at this time.
+Windows is the primary platform and receives the most testing. Linux builds and runs; see
+[Building on Linux](#building-on-linux).
+
+macOS support is experimental. The emulator is built for x86-64 and runs on Apple Silicon under
+Rosetta 2, with Vulkan provided by MoltenVK. A small number of titles have been verified in-game
+on Apple Silicon hardware; see [Building on macOS](#building-on-macos).
 
 ## Bugs and Issues
 
@@ -45,7 +51,7 @@ graphical glitches, low compatibility, and poor performance.
   </tr>
   <tr>
     <td align="center">
-      <strong>Minecraft Legends</strong><br>
+      <strong>Neptunia ReVerse</strong><br>
       <img src="docs/screenshots/ps5-04.png" width="300" alt="Minecraft Legends running in KytyPS5">
     </td>
     <td align="center">
@@ -53,15 +59,28 @@ graphical glitches, low compatibility, and poor performance.
       <img src="docs/screenshots/ps5-05.png" width="300" alt="SILENT HILL: The Short Message running in KytyPS5">
     </td>
   </tr>
+  <tr>
+    <td align="center">
+      <strong>Hellboy</strong><br>
+      <img src="docs/screenshots/ps5-02.png" width="300" alt="Disgaea 6 running in KytyPS5">
+    </td>
+    <td align="center">
+      <strong>Paleo Pines</strong><br>
+      <img src="docs/screenshots/ps5-06.png" width="300" alt="Dreaming Sarah running in KytyPS5">
+    </td>
+  </tr>
 </table>
 
+<p align="center"><em>And many more...</em></p>
+
 ## Contributing
 
 Testing games and submitting detailed bug reports are useful ways to contribute. Search existing
 issues first, then use the **Game Emulation Bug Report** template and attach the complete log file.
 
-Code contributions should be focused, build successfully on Windows, and include relevant tests
-where practical. Because KytyPS5 is still evolving quickly, consider opening an issue before
+Code contributions should be focused, build successfully on the platforms they touch, and include
+relevant tests where practical. Windows is the primary target, so a change that alters shared code
+should not regress it; changes confined to a platform's own code paths only need to build there. Because KytyPS5 is still evolving quickly, consider opening an issue before
 starting a large change.
 
 ### Formatting
@@ -96,11 +115,12 @@ the Vulkan/SPIR-V validation rules.
 
 ### System requirements
 
-- Windows 10 version 1803
-- A 64-bit x86 processor
-- A Vulkan 1.3-capable GPU with current drivers
+- Windows 10 version 1803, a current Linux distribution, or macOS on Apple Silicon
+- A 64-bit x86 processor (on macOS, an Apple Silicon processor with Rosetta 2)
+- A Vulkan 1.3-capable GPU with current drivers (on macOS, Vulkan is provided by the bundled
+  MoltenVK)
 
-### Build requirements
+### Build requirements (Windows)
 
 - Git
 - CMake 3.12 or newer
@@ -134,11 +154,98 @@ cmake --install _Build/windows --prefix _Build/windows/install
 The finished application and its runtime dependencies will be placed in
 `_Build/windows/install`.
 
+### Building on Linux
+
+Install the toolchain and the libraries the bundled SDL2 needs. Without the audio, Wayland and
+udev development packages SDL2 quietly configures itself without those backends, and the resulting
+build has no working sound and no gamepad hotplug:
+
+```bash
+sudo apt-get install --no-install-recommends \
+  clang lld ninja-build cmake git glslang-tools \
+  libgl1-mesa-dev libx11-dev libxcursor-dev libxext-dev libxfixes-dev \
+  libxi-dev libxrandr-dev libxss-dev libxkbcommon-dev \
+  libasound2-dev libpulse-dev libudev-dev libdbus-1-dev libwayland-dev wayland-protocols
+```
+
+Qt 6 (Concurrent, Network, Widgets) is also required — either the distribution packages
+(`qt6-base-dev`) or an official Qt installation.
+
+```bash
+git submodule update --init --recursive
+
+cmake -S src -B _Build/linux -G Ninja -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ \
+  -DCMAKE_PREFIX_PATH="$Qt6_DIR"
+
+cmake --build _Build/linux --target launcher --parallel
+cmake --install _Build/linux --prefix _Build/linux/install
+```
+
+The install step copies the Qt libraries and plugins next to the binaries, so
+`_Build/linux/install` runs without a matching system Qt.
+
+As on Windows, the MSVC compiler is not used; Clang is required. `cl.exe` is rejected at configure
+time.
+
+Note that the CMake source root is `src`, not the repository root.
+
+### Building on macOS
+
+macOS builds target x86-64 and run under Rosetta 2 on Apple Silicon, so the PS5's x86-64 game
+code executes through the same translation layer as the emulator itself. Prebuilt archives are
+attached to releases; the steps below are for building from source.
+
+Requirements:
+
+- An Apple Silicon Mac with Rosetta 2 installed (`softwareupdate --install-rosetta`)
+- Xcode (or the Command Line Tools)
+- Homebrew packages: `brew install cmake ninja glslang`
+- Qt 6 (Concurrent, Network, Widgets) with x86-64 support. The official Qt installation is
+  universal and works; Homebrew's Qt is arm64-only and will not link
+
+```bash
+git submodule update --init --recursive
+
+cmake -S src -B _Build/macos -G Ninja -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_OSX_ARCHITECTURES=x86_64 \
+  -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ \
+  -DCMAKE_PREFIX_PATH="$Qt6_DIR"
+
+cmake --build _Build/macos --target launcher --parallel
+cmake --install _Build/macos --prefix _Build/macos/install
+```
+
+The build re-signs `kyty_emulator` with the JIT entitlements it needs to execute translated
+guest code; no manual signing step is required.
+
+Vulkan comes from MoltenVK. Download `MoltenVK-macos.tar` from the
+[MoltenVK releases](https://github.com/KhronosGroup/MoltenVK/releases), then copy
+`MoltenVK/dynamic/dylib/macOS/libMoltenVK.dylib` next to `kyty_emulator` and ad-hoc sign it:
+
+```bash
+codesign --force --sign - _Build/macos/install/libMoltenVK.dylib
+```
+
+Release archives already include a signed `libMoltenVK.dylib`.
+
+### Regression tests
+
+Build every regression executable and run the registered tests with:
+
+```powershell
+cmake --build _Build/windows --target kyty_tests
+ctest --test-dir _Build/windows --output-on-failure
+```
+
+Use `_Build/linux` instead of `_Build/windows` for a Linux build.
+
 ### Visual Studio Code
 
 A ready-made Visual Studio Code setup is included in [`.vscode`](.vscode). It configures CMake
 Tools to build the project with Ninja and `clang-cl` and provides launch profiles for both
-`launcher.exe` and `kyty_emulator.exe`.
+`launcher.exe` and `kyty_emulator.exe`. It is Windows-only: VS Code settings cannot select a
+compiler per platform, so on Linux configure from the command line as shown above.
 
 Before using it:
 
@@ -160,6 +267,10 @@ To use the graphical launcher:
 .\_Build\windows\install\launcher.exe
 ```
 
+```bash
+./_Build/linux/install/launcher
+```
+
 On first launch, add one or more game folders in the global settings. The launcher searches those
 folders recursively for game directories containing `eboot.bin`. Select a detected game and run it
 from the game list.
@@ -170,7 +281,19 @@ The emulator can also be started directly with a legally obtained game directory
 .\_Build\windows\install\kyty_emulator.exe --game "D:\Games\ExampleGame"
 ```
 
-Run `kyty_emulator.exe --help` to see the available graphics, logging, validation, profiling, and
+```bash
+./_Build/linux/install/kyty_emulator --game "/games/ExampleGame"
+```
+
+On macOS, point SDL at the MoltenVK library explicitly; the hardened runtime prevents it from
+being picked up from the executable's directory:
+
+```bash
+cd _Build/macos/install
+SDL_VULKAN_LIBRARY="$PWD/libMoltenVK.dylib" ./kyty_emulator --game "/games/ExampleGame"
+```
+
+Run `kyty_emulator --help` to see the available graphics, logging, validation, profiling, and
 debugging options.
 
 ### AI Use

src/CMakeLists.txt

@@ -22,6 +22,8 @@ set(CMAKE_CXX_SCAN_FOR_MODULES OFF)
 include(utils.cmake)
 include(CTest)
 
+option(KYTY_ENABLE_CLANG_TIDY "Run clang-tidy checks during builds" OFF)
+
 set(KYTY_THIRD_PARTY_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../3rdparty")
 
 set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
@@ -129,7 +131,7 @@ if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.0.0)
 		common
 		#launcher
 	)
-	list(APPEND KYTY_CLANG_TYDY 
+	list(APPEND KYTY_CLANG_TIDY
 		kyty_emulator
 		#common
 		launcher
@@ -156,16 +158,24 @@ file(GLOB kyty_emulator_src CONFIGURE_DEPENDS
 	graphics/host_gpu/*.h
 	graphics/host_gpu/renderer/*.cpp
 	graphics/host_gpu/renderer/*.h
+	graphics/host_gpu/renderer/cache/*.cpp
+	graphics/host_gpu/renderer/cache/*.h
+	graphics/host_gpu/renderer/image/*.cpp
+	graphics/host_gpu/renderer/image/*.h
+	graphics/host_gpu/renderer/pipeline/*.cpp
+	graphics/host_gpu/renderer/pipeline/*.h
 	graphics/shader/*.cpp
 	graphics/shader/*.h
 	graphics/shader/recompiler/*.cpp
 	graphics/shader/recompiler/*.h
-	graphics/shader/recompiler/shaderIR/*.cpp
-	graphics/shader/recompiler/shaderIR/*.h
-	graphics/shader/recompiler/spirvEmitter/*.cpp
-	graphics/shader/recompiler/spirvEmitter/*.h
-	graphics/host_gpu/objects/*.cpp
-	graphics/host_gpu/objects/*.h
+	graphics/shader/recompiler/cfg/*.cpp
+	graphics/shader/recompiler/cfg/*.h
+	graphics/shader/recompiler/decompiler/*.cpp
+	graphics/shader/recompiler/decompiler/*.h
+	graphics/shader/recompiler/emitter/*.cpp
+	graphics/shader/recompiler/emitter/*.h
+	graphics/shader/recompiler/ir/*.cpp
+	graphics/shader/recompiler/ir/*.h
 	graphics/presentation/*.cpp
 	graphics/presentation/*.h
 	graphics/presentation/window/*.cpp
@@ -249,6 +259,26 @@ endif()
 
 set(kyty_emulator_link_libraries common Vulkan::Headers spirv-tools-opt spirv-tools SDL2-static xxhash FFmpeg::ffmpeg fmt::fmt nlohmann_json::nlohmann_json LibAtrac9)
 
+# Linux system libraries required by the static FFmpeg archive.
+if(LINUX)
+	find_package(Threads REQUIRED)
+	list(APPEND kyty_emulator_link_libraries m ${CMAKE_DL_LIBS} Threads::Threads)
+
+	# Optional FFmpeg dependencies.
+	find_package(ZLIB)
+	if(ZLIB_FOUND)
+		list(APPEND kyty_emulator_link_libraries ZLIB::ZLIB)
+	endif()
+	find_library(KYTY_BZ2_LIBRARY bz2)
+	if(KYTY_BZ2_LIBRARY)
+		list(APPEND kyty_emulator_link_libraries ${KYTY_BZ2_LIBRARY})
+	endif()
+	find_library(KYTY_LZMA_LIBRARY lzma)
+	if(KYTY_LZMA_LIBRARY)
+		list(APPEND kyty_emulator_link_libraries ${KYTY_LZMA_LIBRARY})
+	endif()
+endif()
+
 set(inc_headers
 	${CMAKE_CURRENT_SOURCE_DIR}
 	${KYTY_THIRD_PARTY_DIR}/SDL2/include
@@ -282,6 +312,19 @@ function(add_kyty_full_emulator_test target source)
 		target_link_libraries(${target} onecore)
 		add_custom_command(TARGET ${target} POST_BUILD COMMAND ${CMAKE_COMMAND} -E copy_if_different "${KYTY_THIRD_PARTY_DIR}/winpthread/bin/libwinpthread-1.dll" $<TARGET_FILE_DIR:${target}>/libwinpthread-1.dll)
 	endif()
+	# The macOS x86_64 guest address space needs its .zerofill segments anchored
+	# by linker flags, or the kernel kills the binary on load (posix_spawn EIO).
+	configure_macos_guest_address_space(${target})
+endfunction()
+
+function(configure_macos_guest_address_space target)
+	if(APPLE AND (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR
+			(NOT CMAKE_OSX_ARCHITECTURES AND CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|AMD64)$")))
+		target_sources(${target} PRIVATE kernel/macosGuestAddressSpace.cpp)
+		target_compile_definitions(${target} PRIVATE KYTY_LINKED_GUEST_ADDRESS_SPACE=1)
+		target_link_options(${target} PRIVATE
+			-Wl,-ld_classic,-no_pie,-no_fixup_chains,-no_huge,-pagezero_size,0x40000,-segaddr,SYSTEM_MANAGED,0x40000,-segaddr,SYSTEM_RESERVED,0x7ffffc000,-segaddr,USER_AREA,0x7000000000,-image_base,0x700000000000)
+	endif()
 endfunction()
 
 add_kyty_full_emulator_test(shader_cfg_tests ../tests/shaderCfgTests.cpp)
@@ -289,8 +332,9 @@ add_kyty_full_emulator_test(shader_cfg_tests ../tests/shaderCfgTests.cpp)
 add_executable(scalar_provenance_tests EXCLUDE_FROM_ALL
 	../tests/ScalarProvenanceTests.cpp
 	graphics/host_gpu/hostMemory.cpp
-	graphics/shader/recompiler/ScalarProvenance.cpp
-	graphics/shader/recompiler/SrtWalker.cpp
+	graphics/shader/recompiler/ir/ReadLaneElimination.cpp
+	graphics/shader/recompiler/ir/ScalarProvenance.cpp
+	graphics/shader/recompiler/ir/SrtWalker.cpp
 )
 target_link_libraries(scalar_provenance_tests fmt::fmt)
 target_include_directories(scalar_provenance_tests PRIVATE ${inc_headers})
@@ -301,6 +345,11 @@ add_executable(page_manager_tests EXCLUDE_FROM_ALL
 )
 target_include_directories(page_manager_tests PRIVATE ${inc_headers})
 
+add_executable(bit_array_tests EXCLUDE_FROM_ALL
+	../tests/BitArrayTests.cpp
+)
+target_include_directories(bit_array_tests PRIVATE ${inc_headers})
+
 add_executable(memory_tracker_tests EXCLUDE_FROM_ALL
 	../tests/MemoryTrackerTests.cpp
 	graphics/host_gpu/pageManager.cpp
@@ -308,7 +357,6 @@ add_executable(memory_tracker_tests EXCLUDE_FROM_ALL
 )
 target_link_libraries(memory_tracker_tests fmt::fmt common)
 target_include_directories(memory_tracker_tests PRIVATE ${inc_headers})
-target_compile_definitions(memory_tracker_tests PRIVATE KYTY_MEMORY_TRACKER_TESTS=1)
 
 add_executable(shader_vertex_metadata_tests EXCLUDE_FROM_ALL
 	../tests/ShaderVertexMetadataTests.cpp
@@ -322,9 +370,9 @@ add_executable(shader_stage_runtime_tests EXCLUDE_FROM_ALL
 	graphics/guest_gpu/gpu_format.cpp
 	graphics/host_gpu/hostMemory.cpp
 	graphics/shader/shaderStageRuntime.cpp
-	graphics/shader/recompiler/ResourceMaterialization.cpp
-	graphics/shader/recompiler/ScalarProvenance.cpp
-	graphics/shader/recompiler/SrtWalker.cpp
+	graphics/shader/recompiler/ir/ResourceMaterialization.cpp
+	graphics/shader/recompiler/ir/ScalarProvenance.cpp
+	graphics/shader/recompiler/ir/SrtWalker.cpp
 )
 target_link_libraries(shader_stage_runtime_tests fmt::fmt)
 target_include_directories(shader_stage_runtime_tests PRIVATE ${inc_headers})
@@ -333,24 +381,32 @@ add_executable(resource_tracking_tests EXCLUDE_FROM_ALL
 	../tests/ResourceTrackingTests.cpp
 	graphics/guest_gpu/gpu_format.cpp
 	graphics/host_gpu/hostMemory.cpp
-	graphics/shader/recompiler/ScalarProvenance.cpp
-	graphics/shader/recompiler/SrtWalker.cpp
-	graphics/shader/recompiler/SrtPatcher.cpp
-	graphics/shader/recompiler/ResourceTracking.cpp
-	graphics/shader/recompiler/ResourceMaterialization.cpp
-	graphics/shader/recompiler/ShaderInfoCollection.cpp
-	graphics/shader/recompiler/BindingLayout.cpp
+	graphics/shader/recompiler/ir/ScalarProvenance.cpp
+	graphics/shader/recompiler/ir/SrtWalker.cpp
+	graphics/shader/recompiler/ir/SrtPatcher.cpp
+	graphics/shader/recompiler/ir/ResourceTracking.cpp
+	graphics/shader/recompiler/ir/ResourceMaterialization.cpp
+	graphics/shader/recompiler/ir/ShaderInfoCollection.cpp
+	graphics/shader/recompiler/ir/BindingLayout.cpp
 )
 target_link_libraries(resource_tracking_tests fmt::fmt)
 target_include_directories(resource_tracking_tests PRIVATE ${inc_headers})
 
 add_executable(resource_mutex_tests EXCLUDE_FROM_ALL
 	../tests/ResourceMutexTests.cpp
-	graphics/host_gpu/renderer/resourceMutex.cpp
+	graphics/host_gpu/renderer/cache/resourceMutex.cpp
 )
 target_link_libraries(resource_mutex_tests common)
 target_include_directories(resource_mutex_tests PRIVATE ${inc_headers})
 
+add_executable(audio_out2_port_tests EXCLUDE_FROM_ALL
+	../tests/AudioOut2PortTests.cpp
+	libs/libAudio2.cpp
+	loader/timer.cpp
+)
+target_link_libraries(audio_out2_port_tests common fmt::fmt)
+target_include_directories(audio_out2_port_tests PRIVATE ${inc_headers})
+
 add_executable(event_queue_lifetime_tests EXCLUDE_FROM_ALL
 	../tests/EventQueueLifetimeTests.cpp
 	kernel/eventQueue.cpp
@@ -392,13 +448,30 @@ add_kyty_full_emulator_test(virtual_memory_allocation_tests ../tests/VirtualMemo
 target_compile_definitions(virtual_memory_allocation_tests PRIVATE
 	KYTY_VIRTUAL_MEMORY_ALLOCATION_TESTS=1)
 
+# These tests use exceptions.
+if(NOT KYTY_CLANG_CL)
+	foreach(kyty_exception_test scalar_provenance_tests resource_tracking_tests
+			virtual_memory_allocation_tests)
+		target_compile_options(${kyty_exception_test} PRIVATE -fexceptions)
+	endforeach()
+endif()
+
 if(BUILD_TESTING)
+	add_test(NAME shader_cfg COMMAND $<TARGET_FILE:shader_cfg_tests>)
+	add_test(NAME scalar_provenance COMMAND $<TARGET_FILE:scalar_provenance_tests>)
 	add_test(NAME image_page_table COMMAND $<TARGET_FILE:image_page_table_tests>)
 	add_test(NAME memory_tracker COMMAND $<TARGET_FILE:memory_tracker_tests>)
 	add_test(NAME page_manager COMMAND $<TARGET_FILE:page_manager_tests>)
+	add_test(NAME bit_array COMMAND $<TARGET_FILE:bit_array_tests>)
+	add_test(NAME shader_vertex_metadata COMMAND $<TARGET_FILE:shader_vertex_metadata_tests>)
+	add_test(NAME shader_stage_runtime COMMAND $<TARGET_FILE:shader_stage_runtime_tests>)
+	add_test(NAME resource_tracking COMMAND $<TARGET_FILE:resource_tracking_tests>)
 	add_test(NAME resource_mutex COMMAND $<TARGET_FILE:resource_mutex_tests>)
 	add_test(NAME event_queue_lifetime COMMAND $<TARGET_FILE:event_queue_lifetime_tests>)
+	add_test(NAME audio_out2_port COMMAND $<TARGET_FILE:audio_out2_port_tests>)
 	add_test(NAME shader_recompiler_compute COMMAND $<TARGET_FILE:shader_recompiler_compute_tests>)
+	add_test(NAME virtual_memory_allocation
+		COMMAND $<TARGET_FILE:virtual_memory_allocation_tests>)
 	add_test(NAME command_scheduler_timeline
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --scheduler-only)
 	add_test(NAME stream_buffer_ring
@@ -408,28 +481,47 @@ if(BUILD_TESTING)
 	add_test(NAME gpu_tiler
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --gpu-tiler-only)
 
+	add_test(NAME texture_cache_layered_image
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --layered-image-only)
+	add_test(NAME texture_cache_image_views
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --image-view-cache-only)
+	add_test(NAME texture_cache_storage_sampled
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --storage-sampled-only)
+	add_test(NAME texture_cache_depth_readback
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --depth-readback-only)
+	add_test(NAME buffer_cache_dirty_gc
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --buffer-cache-gc-only)
+
 	if(WIN32)
+		# These tests still depend on the Windows multisample-depth path.
 		add_test(NAME texture_cache_image_overlap
 			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --image-overlap-only)
 		add_test(NAME texture_cache_htile_clear
 			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --htile-clear-only)
-		add_test(NAME texture_cache_layered_image
-			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --layered-image-only)
-		add_test(NAME texture_cache_image_views
-			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --image-view-cache-only)
-		add_test(NAME texture_cache_storage_sampled
-			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --storage-sampled-only)
-		add_test(NAME texture_cache_depth_readback
-			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --depth-readback-only)
 		add_test(NAME buffer_cache_ranges
 			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --buffer-cache-range-only)
-		add_test(NAME buffer_cache_dirty_gc
-			COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --buffer-cache-gc-only)
 	endif()
+
+	add_custom_target(kyty_tests DEPENDS
+		shader_cfg_tests
+		scalar_provenance_tests
+		image_page_table_tests
+		memory_tracker_tests
+		page_manager_tests
+		bit_array_tests
+		shader_vertex_metadata_tests
+		shader_stage_runtime_tests
+		resource_tracking_tests
+		resource_mutex_tests
+		event_queue_lifetime_tests
+		shader_recompiler_compute_tests
+		virtual_memory_allocation_tests
+	)
 endif()
 
 
 add_executable(kyty_emulator main.cpp ${kyty_emulator_src})
+configure_macos_guest_address_space(kyty_emulator)
 
 target_link_libraries(kyty_emulator ${kyty_emulator_link_libraries})
 if (WIN32)
@@ -444,6 +536,8 @@ endif()
 if (CLANG AND NOT KYTY_CLANG_CL)
 	target_link_libraries(kyty_emulator pthread)
 endif()
+# dlopen/dlsym/dladdr for RenderDoc.
+target_link_libraries(kyty_emulator ${CMAKE_DL_LIBS})
 target_include_directories(kyty_emulator PRIVATE ${inc_headers})
 
 clang_tidy_check(kyty_emulator "" "${check_headers}" "${inc_headers}")
@@ -456,7 +550,12 @@ set(KYTY_EMULATOR_MAP_LINK_PATH "${CMAKE_CURRENT_BINARY_DIR}/${KYTY_EMULATOR_MAP
 set(KYTY_EMULATOR_PDB_LINK_PATH "${CMAKE_CURRENT_BINARY_DIR}/kyty_emulator.pdb")
 
 if(KYTY_CLANG_CL)
-	set_target_properties(kyty_emulator PROPERTIES LINK_FLAGS "/DYNAMICBASE:NO /DEBUG:FULL /PDB:${KYTY_EMULATOR_PDB_LINK_PATH} /lldmap:${KYTY_EMULATOR_MAP_LINK_PATH}")
+	target_link_options(kyty_emulator PRIVATE
+		"/DYNAMICBASE:NO"
+		"/DEBUG:FULL"
+		"/PDB:${KYTY_EMULATOR_PDB_LINK_PATH}"
+		"/lldmap:${KYTY_EMULATOR_MAP_LINK_PATH}"
+	)
 	add_custom_command(TARGET kyty_emulator POST_BUILD COMMAND ${CMAKE_COMMAND} -E copy_if_different "${KYTY_THIRD_PARTY_DIR}/winpthread/bin/libwinpthread-1.dll" $<TARGET_FILE_DIR:kyty_emulator>/libwinpthread-1.dll)
 elseif(WIN32 OR LINUX)
 	set_target_properties(kyty_emulator PROPERTIES LINK_FLAGS "${KYTY_LD_OPTIONS} -Wl,-Map=${KYTY_EMULATOR_MAP_LINK_PATH}")

src/common/bitArray.h

@@ -0,0 +1,260 @@
+#ifndef EMULATOR_SRC_COMMON_BITARRAY_H_
+#define EMULATOR_SRC_COMMON_BITARRAY_H_
+
+#include <array>
+#include <bit>
+#include <cstddef>
+#include <cstdint>
+#include <iterator>
+#include <utility>
+
+namespace Common {
+
+template <size_t N>
+class BitArray final {
+	static_assert(N != 0, "BitArray size must be nonzero");
+	static_assert(N % 64 == 0, "BitArray size must be a multiple of 64 bits");
+
+	static constexpr size_t BITS_PER_WORD = 64;
+	static constexpr size_t WORD_COUNT    = N / BITS_PER_WORD;
+
+public:
+	using Range = std::pair<size_t, size_t>;
+
+	class Iterator final {
+	public:
+		using iterator_category = std::forward_iterator_tag;
+		using value_type        = Range;
+		using difference_type   = std::ptrdiff_t;
+		using pointer           = const Range*;
+		using reference         = const Range&;
+
+		Iterator(const BitArray& bits, size_t start)
+		    : m_bits(bits), m_range(bits.FirstRangeFrom(start)) {}
+
+		Iterator& operator++() {
+			m_range = m_bits.FirstRangeFrom(m_range.second);
+			return *this;
+		}
+
+		[[nodiscard]] bool operator==(const Iterator& other) const {
+			return &m_bits == &other.m_bits && m_range == other.m_range;
+		}
+
+		[[nodiscard]] bool operator!=(const Iterator& other) const { return !(*this == other); }
+
+		[[nodiscard]] reference operator*() const { return m_range; }
+		[[nodiscard]] pointer   operator->() const { return &m_range; }
+
+	private:
+		const BitArray& m_bits;
+		Range           m_range;
+	};
+
+	using const_iterator = Iterator;
+
+	constexpr BitArray() = default;
+
+	constexpr BitArray(const BitArray& other, size_t start, size_t end) {
+		if (start >= end || end > N) {
+			return;
+		}
+
+		const auto first_word = start / BITS_PER_WORD;
+		const auto last_word  = (end - 1) / BITS_PER_WORD;
+		const auto start_bit  = start % BITS_PER_WORD;
+		const auto end_bit    = (end - 1) % BITS_PER_WORD;
+		const auto start_mask = ~uint64_t {0} << start_bit;
+		const auto end_mask =
+		    end_bit == BITS_PER_WORD - 1 ? ~uint64_t {0} : (uint64_t {1} << (end_bit + 1)) - 1;
+
+		if (first_word == last_word) {
+			m_data[first_word] = other.m_data[first_word] & start_mask & end_mask;
+			return;
+		}
+
+		m_data[first_word] = other.m_data[first_word] & start_mask;
+		for (auto word = first_word + 1; word < last_word; word++) {
+			m_data[word] = other.m_data[word];
+		}
+		m_data[last_word] = other.m_data[last_word] & end_mask;
+	}
+
+	[[nodiscard]] constexpr bool Get(size_t index) const {
+		return (m_data[index / BITS_PER_WORD] & (uint64_t {1} << (index % BITS_PER_WORD))) != 0;
+	}
+
+	constexpr void Set(size_t index) {
+		m_data[index / BITS_PER_WORD] |= uint64_t {1} << (index % BITS_PER_WORD);
+	}
+
+	constexpr void Unset(size_t index) {
+		m_data[index / BITS_PER_WORD] &= ~(uint64_t {1} << (index % BITS_PER_WORD));
+	}
+
+	constexpr void SetRange(size_t start, size_t end) {
+		if (start >= end || end > N) {
+			return;
+		}
+
+		const auto first_word = start / BITS_PER_WORD;
+		const auto last_word  = (end - 1) / BITS_PER_WORD;
+		const auto start_bit  = start % BITS_PER_WORD;
+		const auto end_bit    = (end - 1) % BITS_PER_WORD;
+		const auto start_mask = ~uint64_t {0} << start_bit;
+		const auto end_mask =
+		    end_bit == BITS_PER_WORD - 1 ? ~uint64_t {0} : (uint64_t {1} << (end_bit + 1)) - 1;
+
+		if (first_word == last_word) {
+			m_data[first_word] |= start_mask & end_mask;
+			return;
+		}
+
+		m_data[first_word] |= start_mask;
+		for (auto word = first_word + 1; word < last_word; word++) {
+			m_data[word] = ~uint64_t {0};
+		}
+		m_data[last_word] |= end_mask;
+	}
+
+	constexpr void UnsetRange(size_t start, size_t end) {
+		if (start >= end || end > N) {
+			return;
+		}
+
+		const auto first_word = start / BITS_PER_WORD;
+		const auto last_word  = (end - 1) / BITS_PER_WORD;
+		const auto start_bit  = start % BITS_PER_WORD;
+		const auto end_bit    = (end - 1) % BITS_PER_WORD;
+		const auto start_mask = (uint64_t {1} << start_bit) - 1;
+		const auto end_mask =
+		    end_bit == BITS_PER_WORD - 1 ? uint64_t {0} : ~((uint64_t {1} << (end_bit + 1)) - 1);
+
+		if (first_word == last_word) {
+			m_data[first_word] &= start_mask | end_mask;
+			return;
+		}
+
+		m_data[first_word] &= start_mask;
+		for (auto word = first_word + 1; word < last_word; word++) {
+			m_data[word] = 0;
+		}
+		m_data[last_word] &= end_mask;
+	}
+
+	constexpr void Clear() { m_data.fill(0); }
+	constexpr void Fill() { m_data.fill(~uint64_t {0}); }
+
+	[[nodiscard]] constexpr bool None() const {
+		uint64_t combined = 0;
+		for (const auto word: m_data) {
+			combined |= word;
+		}
+		return combined == 0;
+	}
+
+	[[nodiscard]] constexpr bool Any() const { return !None(); }
+
+	[[nodiscard]] constexpr Range FirstRangeFrom(size_t start) const {
+		if (start >= N) {
+			return {N, N};
+		}
+
+		auto word_index = start / BITS_PER_WORD;
+		auto word       = m_data[word_index] & (~uint64_t {0} << (start % BITS_PER_WORD));
+		while (word == 0) {
+			word_index++;
+			if (word_index == WORD_COUNT) {
+				return {N, N};
+			}
+			word = m_data[word_index];
+		}
+
+		const auto first     = word_index * BITS_PER_WORD + std::countr_zero(word);
+		const auto first_bit = first % BITS_PER_WORD;
+		const auto first_ones =
+		    static_cast<size_t>(std::countr_one(m_data[word_index] >> first_bit));
+		if (first_bit + first_ones < BITS_PER_WORD) {
+			return {first, first + first_ones};
+		}
+
+		for (word_index++; word_index < WORD_COUNT; word_index++) {
+			word = m_data[word_index];
+			if (word != ~uint64_t {0}) {
+				return {first, word_index * BITS_PER_WORD + std::countr_one(word)};
+			}
+		}
+		return {first, N};
+	}
+
+	[[nodiscard]] constexpr Range FirstRange() const { return FirstRangeFrom(0); }
+
+	[[nodiscard]] constexpr Range LastRangeFrom(size_t end) const {
+		if (end == 0) {
+			return {0, 0};
+		}
+		if (end > N) {
+			end = N;
+		}
+
+		auto       word_index = (end - 1) / BITS_PER_WORD;
+		const auto end_bit    = (end - 1) % BITS_PER_WORD;
+		const auto end_mask =
+		    end_bit == BITS_PER_WORD - 1 ? ~uint64_t {0} : (uint64_t {1} << (end_bit + 1)) - 1;
+		auto word = m_data[word_index] & end_mask;
+		while (word == 0) {
+			if (word_index == 0) {
+				return {0, 0};
+			}
+			word = m_data[--word_index];
+		}
+
+		const auto empty_bits = static_cast<size_t>(std::countl_zero(word));
+		const auto ones       = static_cast<size_t>(std::countl_one(word << empty_bits));
+		const auto last       = (word_index + 1) * BITS_PER_WORD - empty_bits;
+		if (empty_bits + ones < BITS_PER_WORD) {
+			return {last - ones, last};
+		}
+
+		while (word_index != 0) {
+			word = m_data[--word_index];
+			if (word != ~uint64_t {0}) {
+				return {(word_index + 1) * BITS_PER_WORD - std::countl_one(word), last};
+			}
+		}
+		return {0, last};
+	}
+
+	[[nodiscard]] constexpr Range LastRange() const { return LastRangeFrom(N); }
+
+	[[nodiscard]] const_iterator begin() const { return Iterator(*this, 0); }
+	[[nodiscard]] const_iterator end() const { return Iterator(*this, N); }
+
+	constexpr BitArray& operator^=(const BitArray& other) {
+		for (size_t word = 0; word < WORD_COUNT; word++) {
+			m_data[word] ^= other.m_data[word];
+		}
+		return *this;
+	}
+
+	[[nodiscard]] constexpr BitArray operator^(const BitArray& other) const {
+		auto result = *this;
+		result ^= other;
+		return result;
+	}
+
+	[[nodiscard]] constexpr BitArray operator~() const {
+		auto result = *this;
+		for (auto& word: result.m_data) {
+			word = ~word;
+		}
+		return result;
+	}
+
+private:
+	std::array<uint64_t, WORD_COUNT> m_data {};
+};
+
+} // namespace Common
+
+#endif // EMULATOR_SRC_COMMON_BITARRAY_H_

src/common/hostException.cpp

@@ -6,6 +6,14 @@
 
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 #include <windows.h> // IWYU pragma: keep
+#elif defined(__APPLE__)
+#include <csignal>
+#include <sys/ucontext.h>
+#else
+#include <csignal>
+#include <initializer_list>
+#include <ucontext.h> // IWYU pragma: keep
+#include <unistd.h>
 #endif
 
 // IWYU pragma: no_include <errhandlingapi.h>
@@ -16,7 +24,7 @@
 
 namespace Common::HostException {
 
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+#if !defined(__APPLE__)
 
 static std::atomic<Handler> g_handler {nullptr};
 static std::atomic_uint32_t g_install_state {0};
@@ -30,7 +38,9 @@ static_assert(decltype(g_install_state)::is_always_lock_free);
 	std::fputs(reason != nullptr ? reason : "unspecified", stderr);
 	std::fputc('\n', stderr);
 	std::fflush(stderr);
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	TerminateProcess(GetCurrentProcess(), static_cast<UINT>(EXCEPTION_NONCONTINUABLE_EXCEPTION));
+#endif
 	std::_Exit(321);
 }
 
@@ -48,6 +58,21 @@ public:
 	KYTY_CLASS_NO_COPY(FilterScope);
 };
 
+static Handler LoadInstalledHandler() noexcept {
+	if (g_install_state.load(std::memory_order_acquire) == 0) {
+		FailFast("host exception handler is not installed");
+	}
+
+	const auto handler = g_handler.load(std::memory_order_acquire);
+	if (handler == nullptr) {
+		FailFast("host exception callback is null");
+	}
+	return handler;
+}
+#endif
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+
 static LONG WINAPI ExceptionFilter(PEXCEPTION_POINTERS exception) {
 	FilterScope filter_scope;
 
@@ -106,22 +131,176 @@ static LONG WINAPI ExceptionFilter(PEXCEPTION_POINTERS exception) {
 	info.r14 = exception->ContextRecord->R14;
 	info.r15 = exception->ContextRecord->R15;
 
-	if (g_install_state.load(std::memory_order_acquire) == 0) {
-		FailFast("host exception handler is not installed");
+	const auto handler = LoadInstalledHandler();
+
+	return handler(info) ? EXCEPTION_CONTINUE_EXECUTION : EXCEPTION_CONTINUE_SEARCH;
+}
+
+#elif defined(__APPLE__)
+
+static std::atomic<Handler> g_handler {nullptr};
+static std::atomic_uint32_t g_install_state {0};
+static thread_local bool    g_in_exception_filter = false;
+
+static_assert(decltype(g_handler)::is_always_lock_free);
+static_assert(decltype(g_install_state)::is_always_lock_free);
+
+[[noreturn]] static void FailFast(const char* reason) noexcept {
+	std::fputs("HostException fail-fast: ", stderr);
+	std::fputs(reason != nullptr ? reason : "unspecified", stderr);
+	std::fputc('\n', stderr);
+	std::fflush(stderr);
+	std::_Exit(321);
+}
+
+// Translate the x86-64 page-fault error code (mcontext __es.__err) into an access type.
+// bit 1 (0x2) = write, bit 4 (0x10) = instruction fetch, otherwise a read.
+static AccessViolationType DecodeAccess(uint64_t err) {
+	if ((err & 0x10u) != 0) {
+		return AccessViolationType::Execute;
 	}
+	if ((err & 0x2u) != 0) {
+		return AccessViolationType::Write;
+	}
+	return AccessViolationType::Read;
+}
+
+// POSIX signal handler that mirrors the Windows vectored handler: build an ExceptionInfo
+// from the mcontext and dispatch. A resolved fault (handler returns true) simply returns,
+// re-executing the faulting instruction against the now-fixed protection. An unresolved
+// fault restores the default disposition so the retry terminates the process.
+static void SignalHandler(int sig, siginfo_t* si, void* uctx) {
+	if (g_in_exception_filter) {
+		FailFast("nested exception while resolving a host fault");
+	}
+	g_in_exception_filter = true;
+
+	auto*       uc = static_cast<ucontext_t*>(uctx);
+	const auto* mc = uc->uc_mcontext;
+	const auto& ss = mc->__ss;
+
+	ExceptionInfo info {};
+	info.exception_address = ss.__rip;
+	info.native_code       = static_cast<uint32_t>(si->si_code);
+	info.native_context    = uctx;
+
+	if (sig == SIGILL) {
+		info.type = ExceptionType::IllegalInstruction;
+	} else {
+		info.type                   = ExceptionType::AccessViolation;
+		info.access_violation_type  = DecodeAccess(mc->__es.__err);
+		info.access_violation_vaddr = reinterpret_cast<uint64_t>(si->si_addr);
+	}
+
+	info.rax = ss.__rax;
+	info.rbx = ss.__rbx;
+	info.rcx = ss.__rcx;
+	info.rdx = ss.__rdx;
+	info.rsi = ss.__rsi;
+	info.rdi = ss.__rdi;
+	info.rbp = ss.__rbp;
+	info.rsp = ss.__rsp;
+	info.r8  = ss.__r8;
+	info.r9  = ss.__r9;
+	info.r10 = ss.__r10;
+	info.r11 = ss.__r11;
+	info.r12 = ss.__r12;
+	info.r13 = ss.__r13;
+	info.r14 = ss.__r14;
+	info.r15 = ss.__r15;
 
 	const auto handler = g_handler.load(std::memory_order_acquire);
 	if (handler == nullptr) {
 		FailFast("host exception callback is null");
 	}
 
-	return handler(info) ? EXCEPTION_CONTINUE_EXECUTION : EXCEPTION_CONTINUE_SEARCH;
+	const bool resolved   = handler(info);
+	g_in_exception_filter = false;
+
+	if (resolved) {
+		return; // retry the faulting instruction against the fixed mapping
+	}
+
+	// Unresolved: restore the default action so the re-executed instruction terminates.
+	struct sigaction dfl {};
+	dfl.sa_handler = SIG_DFL;
+	sigemptyset(&dfl.sa_mask);
+	sigaction(sig, &dfl, nullptr);
+}
+
+#else
+
+// x86-64 page-fault error bits.
+constexpr uint64_t PAGE_FAULT_ERROR_WRITE       = 0x02;
+constexpr uint64_t PAGE_FAULT_ERROR_INSTRUCTION = 0x10;
+
+// Let the kernel handle an unresolved fault on retry.
+static void ChainToDefault(int signal_number) noexcept {
+	struct sigaction restore {};
+	restore.sa_handler = SIG_DFL;
+	sigemptyset(&restore.sa_mask);
+	restore.sa_flags = 0;
+	::sigaction(signal_number, &restore, nullptr);
+}
+
+static void SignalHandler(int signal_number, siginfo_t* signal_info, void* native_context) {
+	FilterScope filter_scope;
+
+	auto* context = static_cast<ucontext_t*>(native_context);
+	auto* gregs   = context->uc_mcontext.gregs;
+
+	ExceptionInfo info {};
+	info.exception_address = static_cast<uint64_t>(gregs[REG_RIP]);
+	info.native_code       = static_cast<uint32_t>(signal_number);
+	info.native_context    = context;
+
+	if (signal_number == SIGSEGV || signal_number == SIGBUS) {
+		info.type             = ExceptionType::AccessViolation;
+		const auto error_code = static_cast<uint64_t>(gregs[REG_ERR]);
+		if ((error_code & PAGE_FAULT_ERROR_INSTRUCTION) != 0) {
+			info.access_violation_type = AccessViolationType::Execute;
+		} else if ((error_code & PAGE_FAULT_ERROR_WRITE) != 0) {
+			info.access_violation_type = AccessViolationType::Write;
+		} else {
+			info.access_violation_type = AccessViolationType::Read;
+		}
+		info.access_violation_vaddr = reinterpret_cast<uint64_t>(signal_info->si_addr);
+	} else if (signal_number == SIGILL) {
+		info.type = ExceptionType::IllegalInstruction;
+	} else {
+		ChainToDefault(signal_number);
+		return;
+	}
+
+	info.rax = static_cast<uint64_t>(gregs[REG_RAX]);
+	info.rbx = static_cast<uint64_t>(gregs[REG_RBX]);
+	info.rcx = static_cast<uint64_t>(gregs[REG_RCX]);
+	info.rdx = static_cast<uint64_t>(gregs[REG_RDX]);
+	info.rsi = static_cast<uint64_t>(gregs[REG_RSI]);
+	info.rdi = static_cast<uint64_t>(gregs[REG_RDI]);
+	info.rbp = static_cast<uint64_t>(gregs[REG_RBP]);
+	info.rsp = static_cast<uint64_t>(gregs[REG_RSP]);
+	info.r8  = static_cast<uint64_t>(gregs[REG_R8]);
+	info.r9  = static_cast<uint64_t>(gregs[REG_R9]);
+	info.r10 = static_cast<uint64_t>(gregs[REG_R10]);
+	info.r11 = static_cast<uint64_t>(gregs[REG_R11]);
+	info.r12 = static_cast<uint64_t>(gregs[REG_R12]);
+	info.r13 = static_cast<uint64_t>(gregs[REG_R13]);
+	info.r14 = static_cast<uint64_t>(gregs[REG_R14]);
+	info.r15 = static_cast<uint64_t>(gregs[REG_R15]);
+
+	const auto handler = LoadInstalledHandler();
+
+	if (handler(info)) {
+		return;
+	}
+
+	ChainToDefault(signal_number);
 }
 
 #endif
 
 bool InstallHandler(Handler handler) {
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	if (handler == nullptr) {
 		return false;
 	}
@@ -133,19 +312,52 @@ bool InstallHandler(Handler handler) {
 
 	g_handler.store(handler, std::memory_order_release);
 
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	if (AddVectoredExceptionHandler(1, ExceptionFilter) == nullptr) {
 		g_handler.store(nullptr, std::memory_order_release);
 		g_install_state.store(0, std::memory_order_release);
 		printf("AddVectoredExceptionHandler() failed\n");
 		return false;
 	}
+#elif defined(__APPLE__)
+	struct sigaction sa {};
+	sa.sa_sigaction = SignalHandler;
+	sa.sa_flags     = SA_SIGINFO;
+	sigemptyset(&sa.sa_mask);
+	// The guest signal-dispatch path (KernelRaiseException) interrupts threads with
+	// SIGUSR1; block it while a fault is being resolved so a stop-the-world request
+	// cannot preempt the handler between the protection fix and the retry.
+	sigaddset(&sa.sa_mask, SIGUSR1);
+
+	// macOS raises SIGBUS for protection faults on some paths and SIGSEGV on others;
+	// SIGILL covers instructions the host cannot execute (routed to the x64 emulator).
+	bool ok = sigaction(SIGSEGV, &sa, nullptr) == 0 && sigaction(SIGBUS, &sa, nullptr) == 0 &&
+	          sigaction(SIGILL, &sa, nullptr) == 0;
+	if (!ok) {
+		g_handler.store(nullptr, std::memory_order_release);
+		g_install_state.store(0, std::memory_order_release);
+		printf("sigaction() failed to install the host fault handler\n");
+		return false;
+	}
+#else
+	struct sigaction action {};
+	action.sa_sigaction = SignalHandler;
+	sigemptyset(&action.sa_mask);
+	// Fault resolution needs the normal thread stack.
+	action.sa_flags = SA_SIGINFO | SA_RESTART;
+
+	for (const int signal_number: {SIGSEGV, SIGBUS, SIGILL}) {
+		if (::sigaction(signal_number, &action, nullptr) != 0) {
+			g_handler.store(nullptr, std::memory_order_release);
+			g_install_state.store(0, std::memory_order_release);
+			printf("sigaction(%d) failed\n", signal_number);
+			return false;
+		}
+	}
+#endif
 
 	g_install_state.store(2, std::memory_order_release);
 	return true;
-#else
-	(void)handler;
-	return false;
-#endif
 }
 
 } // namespace Common::HostException

src/common/logging/log.cpp

@@ -178,6 +178,11 @@ Direction GetDirection() {
 	return g_direction;
 }
 
+bool IsSilent() {
+	// Before init LOGF must keep writing to stdout, so report non-silent.
+	return g_initialized && g_direction == Direction::Silent;
+}
+
 void Write(std::string_view text) {
 	WriteImpl(text);
 }

src/common/logging/log.h

@@ -15,6 +15,7 @@ KYTY_SUBSYSTEM_DEFINE(Log);
 enum class Direction { Silent, Console, File };
 
 Direction GetDirection();
+bool      IsSilent();
 void      Write(std::string_view text);
 void      Write(fmt::text_style style, std::string_view text);
 void      WriteFatal(std::string_view text);
@@ -41,8 +42,18 @@ inline constexpr auto BrightWhite   = fmt::fg(fmt::terminal_color::bright_white)
 } // namespace Log
 
 // NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
-#define LOGF(...) ::Log::Write(::fmt::sprintf(__VA_ARGS__))
+#define LOGF(...)                                                                                  \
+	do {                                                                                           \
+		if (!::Log::IsSilent()) {                                                                  \
+			::Log::Write(::fmt::sprintf(__VA_ARGS__));                                             \
+		}                                                                                          \
+	} while (false)
 // NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
-#define LOGF_COLOR(style, ...) ::Log::Write((style), ::fmt::sprintf(__VA_ARGS__))
+#define LOGF_COLOR(style, ...)                                                                     \
+	do {                                                                                           \
+		if (!::Log::IsSilent()) {                                                                  \
+			::Log::Write((style), ::fmt::sprintf(__VA_ARGS__));                                    \
+		}                                                                                          \
+	} while (false)
 
 #endif /* KYTY_COMMON_LOGGING_LOG_H_ */

src/common/lruCache.h

@@ -19,10 +19,10 @@ class LeastRecentlyUsedCache {
 
 public:
 	[[nodiscard]] size_t Insert(Object object, Tick tick) {
-		const auto id = Build();
+		const auto id   = Build();
 		auto&      item = m_items[id];
-		item.object = std::move(object);
-		item.tick   = tick;
+		item.object     = std::move(object);
+		item.tick       = tick;
 		Attach(item);
 		return id;
 	}
@@ -49,8 +49,7 @@ public:
 
 	template <typename Function>
 	void ForEachItemBelow(Tick tick, Function&& function) {
-		constexpr bool ReturnsBool =
-		    std::is_same_v<std::invoke_result_t<Function, Object>, bool>;
+		constexpr bool ReturnsBool = std::is_same_v<std::invoke_result_t<Function, Object>, bool>;
 		for (auto* item = m_first; item != nullptr;) {
 			if (item->tick > tick) {
 				return;
@@ -87,10 +86,10 @@ private:
 			m_last = &item;
 			return;
 		}
-		item.prev   = m_last;
+		item.prev    = m_last;
 		m_last->next = &item;
-		item.next   = nullptr;
-		m_last      = &item;
+		item.next    = nullptr;
+		m_last       = &item;
 	}
 
 	void Detach(Item& item) {

src/common/platform/sysDbg.h

@@ -23,6 +23,10 @@ struct sys_dbg_stack_info_t {
 	size_t    commited_size;
 	size_t    total_size;
 	size_t    code_size;
+
+	// Full stack reservation reported by pthread.
+	uintptr_t reserved_addr;
+	size_t    reserved_size;
 #endif
 };
 

src/common/platform/sysLinuxDbg.cpp

@@ -8,12 +8,56 @@
 
 #include <cstdlib>
 #include <cstring>
+#include <execinfo.h>
+#include <pthread.h>
 #include <sys/param.h>
 #include <sys/types.h>
 #include <unistd.h>
+#if defined(__APPLE__)
+#include <libgen.h> // POSIX basename() lives here on macOS, not in <cstring>
+#endif
 
-void SysStackWalk(void** /*stack*/, int* depth) {
-	*depth = 0;
+// Avoid unwinding a guest-owned stack.
+static bool OnOwnStack() {
+	const char* probe = reinterpret_cast<const char*>(&probe);
+
+	pthread_attr_t attr {};
+#if defined(__APPLE__)
+	const auto* top  = static_cast<const char*>(pthread_get_stackaddr_np(pthread_self()));
+	const auto  size = pthread_get_stacksize_np(pthread_self());
+	(void)attr;
+	return top != nullptr && size != 0 && probe < top && probe >= top - size;
+#else
+	if (pthread_getattr_np(pthread_self(), &attr) != 0) {
+		return false;
+	}
+	void*      base = nullptr;
+	size_t     size = 0;
+	const bool ok = pthread_attr_getstack(&attr, &base, &size) == 0 && base != nullptr && size != 0;
+	pthread_attr_destroy(&attr);
+	if (!ok) {
+		return false;
+	}
+	const auto* low = static_cast<const char*>(base);
+	return probe >= low && probe < low + size;
+#endif
+}
+
+void SysStackWalk(void** stack, int* depth) {
+	if (stack == nullptr || depth == nullptr || *depth <= 0) {
+		if (depth != nullptr) {
+			*depth = 0;
+		}
+		return;
+	}
+
+	if (!OnOwnStack()) {
+		*depth = 0;
+		return;
+	}
+
+	const int n = ::backtrace(stack, *depth);
+	*depth      = (n < 0 ? 0 : n);
 }
 
 void SysStackUsagePrint(sys_dbg_stack_info_t& stack) {
@@ -30,6 +74,33 @@ void SysStackUsage(sys_dbg_stack_info_t& s) {
 
 	[[maybe_unused]] int result = 0;
 
+	memset(&s, 0, sizeof(sys_dbg_stack_info_t));
+
+	// Record the reservation before the Linux /proc walk.
+	{
+		pthread_attr_t self_attr {};
+#if defined(__APPLE__)
+		void*        stack_top  = pthread_get_stackaddr_np(pthread_self());
+		const size_t stack_size = pthread_get_stacksize_np(pthread_self());
+		if (stack_top != nullptr && stack_size != 0) {
+			s.reserved_addr = reinterpret_cast<uintptr_t>(stack_top) - stack_size;
+			s.reserved_size = stack_size;
+		}
+		(void)self_attr;
+#else
+		if (pthread_getattr_np(pthread_self(), &self_attr) == 0) {
+			void*  stack_base = nullptr;
+			size_t stack_size = 0;
+			if (pthread_attr_getstack(&self_attr, &stack_base, &stack_size) == 0 &&
+			    stack_base != nullptr && stack_size != 0) {
+				s.reserved_addr = reinterpret_cast<uintptr_t>(stack_base);
+				s.reserved_size = stack_size;
+			}
+			pthread_attr_destroy(&self_attr);
+		}
+#endif
+	}
+
 	char str[1024];
 	char str2[1024];
 	result = sprintf(str, "/proc/%d/exe", static_cast<int>(pid));
@@ -43,8 +114,6 @@ void SysStackUsage(sys_dbg_stack_info_t& s) {
 
 	result = sprintf(str, "/proc/%d/maps", static_cast<int>(pid));
 
-	memset(&s, 0, sizeof(sys_dbg_stack_info_t));
-
 	FILE* f = fopen(str, "r");
 
 	if (f == nullptr) {
@@ -118,6 +187,11 @@ void SysStackUsage(sys_dbg_stack_info_t& s) {
 	}
 
 	result = fclose(f);
+
+	if (s.reserved_addr == 0) {
+		s.reserved_addr = s.addr;
+		s.reserved_size = s.total_size;
+	}
 }
 
 #endif

src/common/platform/sysLinuxFileIO.cpp

@@ -11,7 +11,11 @@
 
 #include <cerrno>
 #include <cstdlib>
+#include <dirent.h>
+#include <fcntl.h>
+#include <filesystem>
 #include <sys/stat.h>
+#include <system_error>
 #include <unistd.h>
 #include <utime.h>
 
@@ -39,10 +43,43 @@ struct sys_file_t {
 	};
 };
 
+// Darwin uses BSD timestamp member names.
+#if defined(__APPLE__)
+#define KYTY_STAT_ATIME_NS(st) ((st).st_atimespec.tv_nsec)
+#define KYTY_STAT_MTIME_NS(st) ((st).st_mtimespec.tv_nsec)
+#else
+#define KYTY_STAT_ATIME_NS(st) ((st).st_atim.tv_nsec)
+#define KYTY_STAT_MTIME_NS(st) ((st).st_mtim.tv_nsec)
+#endif
+
 static std::filesystem::path get_internal_name(const std::filesystem::path& name) {
 	return name.is_absolute() ? name : (std::filesystem::path(".") / name);
 }
 
+// Pass access-pattern hints to the host.
+static void apply_cache_hint(FILE* f, sys_file_cache_type_t cache_type) {
+	if (f == nullptr) {
+		return;
+	}
+
+#if !defined(__APPLE__)
+	int advice = POSIX_FADV_NORMAL;
+	switch (cache_type) {
+		case SYS_FILE_CACHE_RANDOM_ACCESS: advice = POSIX_FADV_RANDOM; break;
+		case SYS_FILE_CACHE_SEQUENTIAL_SCAN: advice = POSIX_FADV_SEQUENTIAL; break;
+		case SYS_FILE_CACHE_AUTO:
+		default: return;
+	}
+	::posix_fadvise(fileno(f), 0, 0, advice);
+#else
+	if (cache_type == SYS_FILE_CACHE_SEQUENTIAL_SCAN) {
+		::fcntl(fileno(f), F_RDAHEAD, 1);
+	} else if (cache_type == SYS_FILE_CACHE_RANDOM_ACCESS) {
+		::fcntl(fileno(f), F_RDAHEAD, 0);
+	}
+#endif
+}
+
 void SysFileRead(void* data, uint32_t size, sys_file_t& f, uint32_t* bytes_read) {
 	if (f.type == SYS_FILE_FILE) {
 		size_t w = fread(data, 1, size, f.f);
@@ -135,8 +172,7 @@ sys_file_t* SysFileCreate(const std::filesystem::path& file_name) {
 	return ret;
 }
 
-sys_file_t* SysFileOpenR(const std::filesystem::path& file_name,
-                         sys_file_cache_type_t /*cache_type*/) {
+sys_file_t* SysFileOpenR(const std::filesystem::path& file_name, sys_file_cache_type_t cache_type) {
 	auto* ret = new sys_file_t;
 
 	ret->type = SYS_FILE_FILE;
@@ -150,6 +186,8 @@ sys_file_t* SysFileOpenR(const std::filesystem::path& file_name,
 		ret->type = SYS_FILE_ERROR;
 	}
 
+	apply_cache_hint(f, cache_type);
+
 	ret->f = f;
 
 	return ret;
@@ -179,8 +217,7 @@ sys_file_t* SysFileCreate() {
 	return ret;
 }
 
-sys_file_t* SysFileOpenW(const std::filesystem::path& file_name,
-                         sys_file_cache_type_t /*cache_type*/) {
+sys_file_t* SysFileOpenW(const std::filesystem::path& file_name, sys_file_cache_type_t cache_type) {
 	auto* ret = new sys_file_t;
 
 	auto real_name     = get_internal_name(file_name);
@@ -194,13 +231,15 @@ sys_file_t* SysFileOpenW(const std::filesystem::path& file_name,
 		ret->type = SYS_FILE_FILE;
 	}
 
+	apply_cache_hint(f, cache_type);
+
 	ret->f = f;
 
 	return ret;
 }
 
 sys_file_t* SysFileOpenRw(const std::filesystem::path& file_name,
-                          sys_file_cache_type_t /*cache_type*/) {
+                          sys_file_cache_type_t        cache_type) {
 	auto* ret = new sys_file_t;
 
 	auto real_name     = get_internal_name(file_name);
@@ -214,6 +253,8 @@ sys_file_t* SysFileOpenRw(const std::filesystem::path& file_name,
 		ret->type = SYS_FILE_FILE;
 	}
 
+	apply_cache_hint(f, cache_type);
+
 	ret->f = f;
 
 	return ret;
@@ -239,11 +280,17 @@ uint64_t SysFileSize(sys_file_t& f) {
 	[[maybe_unused]] int result = 0;
 
 	if (f.type == SYS_FILE_FILE) {
-		uint32_t pos  = ftell(f.f);
-		result        = fseek(f.f, 0, SEEK_END);
-		uint32_t size = ftell(f.f);
-		result        = fseek(f.f, pos, SEEK_SET);
-		return size;
+		// Preserve sizes above 4 GiB.
+		const off_t pos = ftello(f.f);
+		if (pos < 0) {
+			return 0;
+		}
+		if (fseeko(f.f, 0, SEEK_END) != 0) {
+			return 0;
+		}
+		const off_t size = ftello(f.f);
+		result           = fseeko(f.f, pos, SEEK_SET);
+		return (size < 0 ? 0 : static_cast<uint64_t>(size));
 	}
 
 	if (f.type == SYS_FILE_MEMORY_STAT || f.type == SYS_FILE_MEMORY_DYN) {
@@ -260,8 +307,18 @@ uint64_t SysFileSize(const std::filesystem::path& file_name) {
 	return size;
 }
 
-bool SysFileTruncate(sys_file_t& /*f*/, uint64_t /*size*/) {
-	return false;
+bool SysFileTruncate(sys_file_t& f, uint64_t size) {
+	bool ok = false;
+	if (f.type == SYS_FILE_FILE) {
+		// Flush before resizing and restore the caller's position.
+		const auto position = ftell(f.f);
+		fflush(f.f);
+		ok = (ftruncate(fileno(f.f), static_cast<off_t>(size)) == 0);
+		if (position >= 0) {
+			fseek(f.f, position, SEEK_SET);
+		}
+	}
+	return ok;
 }
 
 bool SysFileUnlink(sys_file_t& /*f*/, const std::filesystem::path& name) {
@@ -382,6 +439,7 @@ SysFileTimeStruct SysFileGetLastAccessTimeUtc(const std::filesystem::path& name)
 	} else {
 		r.is_invalid = false;
 		r.time       = s.st_atime;
+		r.nanos      = KYTY_STAT_ATIME_NS(s);
 	}
 
 	return r;
@@ -400,6 +458,7 @@ SysFileTimeStruct SysFileGetLastWriteTimeUtc(const std::filesystem::path& name)
 	} else {
 		r.is_invalid = false;
 		r.time       = s.st_mtime;
+		r.nanos      = KYTY_STAT_MTIME_NS(s);
 	}
 
 	return r;
@@ -419,13 +478,36 @@ void SysFileGetLastAccessAndWriteTimeUtc(const std::filesystem::path& name, SysF
 		a.is_invalid = false;
 		w.is_invalid = false;
 		a.time       = s.st_atime;
+		a.nanos      = KYTY_STAT_ATIME_NS(s);
 		w.time       = s.st_mtime;
+		w.nanos      = KYTY_STAT_MTIME_NS(s);
 	}
 }
 
-void SysFileGetLastAccessAndWriteTimeUtc(sys_file_t& /*f*/, SysFileTimeStruct& /*a*/,
-                                         SysFileTimeStruct& /*w*/) {
-	EXIT("not implemented\n");
+void SysFileGetLastAccessAndWriteTimeUtc(sys_file_t& f, SysFileTimeStruct& a,
+                                         SysFileTimeStruct& w) {
+	if (f.type == SYS_FILE_FILE) {
+		struct stat s {};
+
+		const bool ok = (0 == fstat(fileno(f.f), &s));
+
+		a.is_invalid = w.is_invalid = !ok;
+
+		if (ok) {
+			a.time  = s.st_atime;
+			a.nanos = KYTY_STAT_ATIME_NS(s);
+			w.time  = s.st_mtime;
+			w.nanos = KYTY_STAT_MTIME_NS(s);
+		}
+	} else if (f.type == SYS_FILE_MEMORY_STAT || f.type == SYS_FILE_MEMORY_DYN) {
+		// Memory-backed files use the current time.
+		SysTimeStruct t {};
+		SysGetSystemTimeUtc(t);
+		SysSystemToFileTimeUtc(t, a);
+		SysSystemToFileTimeUtc(t, w);
+	} else {
+		a.is_invalid = w.is_invalid = true;
+	}
 }
 
 bool SysFileSetLastAccessTimeUtc(const std::filesystem::path& name, SysFileTimeStruct& access) {
@@ -531,27 +613,110 @@ bool SysFileSetLastAccessAndWriteTimeUtc(const std::filesystem::path& name,
 	//	}
 }
 
-void SysFileFindFiles(const std::filesystem::path& /*path*/,
-                      std::vector<sys_file_find_t>& /*out*/) {
-	EXIT("not implemented\n");
+// Recursively collect regular files.
+void SysFileFindFiles(const std::filesystem::path& path, std::vector<sys_file_find_t>& out) {
+	auto real_path = get_internal_name(path);
+
+	DIR* dir = opendir(real_path.string().c_str());
+	if (dir == nullptr) {
+		return;
+	}
+
+	for (const dirent* entry = readdir(dir); entry != nullptr; entry = readdir(dir)) {
+		const std::string file_name(entry->d_name);
+
+		if (file_name == "." || file_name == "..") {
+			continue;
+		}
+
+		auto        child = real_path / file_name;
+		struct stat s {};
+
+		// lstat, so a symlink is never followed into a cycle during the recursive walk.
+		if (0 != lstat(child.string().c_str(), &s)) {
+			continue;
+		}
+
+		if (S_ISDIR(s.st_mode)) {
+			SysFileFindFiles(child, out);
+		} else if (S_ISREG(s.st_mode)) {
+			sys_file_find_t r {};
+
+			r.path_with_name              = child;
+			r.size                        = static_cast<uint64_t>(s.st_size);
+			r.last_access_time.is_invalid = false;
+			r.last_access_time.time       = s.st_atime;
+			r.last_access_time.nanos      = KYTY_STAT_ATIME_NS(s);
+			r.last_write_time.is_invalid  = false;
+			r.last_write_time.time        = s.st_mtime;
+			r.last_write_time.nanos       = KYTY_STAT_MTIME_NS(s);
+
+			out.push_back(r);
+		}
+	}
+
+	closedir(dir);
 }
 
-void SysFileGetDents(const std::filesystem::path& /*path*/, std::vector<sys_dir_entry_t>& /*out*/) {
-	EXIT("not implemented\n");
+// Keep "." and ".." to match FindFirstFileW.
+void SysFileGetDents(const std::filesystem::path& path, std::vector<sys_dir_entry_t>& out) {
+	auto real_path = get_internal_name(path);
+
+	DIR* dir = opendir(real_path.string().c_str());
+	if (dir == nullptr) {
+		return;
+	}
+
+	for (const dirent* entry = readdir(dir); entry != nullptr; entry = readdir(dir)) {
+		sys_dir_entry_t r {};
+
+		r.name = entry->d_name;
+
+		if (entry->d_type == DT_UNKNOWN) {
+			// Some filesystems do not populate d_type.
+			struct stat s {};
+			r.is_file = 0 == lstat((real_path / r.name).string().c_str(), &s) && S_ISREG(s.st_mode);
+		} else {
+			r.is_file = entry->d_type != DT_DIR;
+		}
+
+		out.push_back(r);
+	}
+
+	closedir(dir);
 }
 
-bool SysFileCopyFile(const std::filesystem::path& /*src*/, const std::filesystem::path& /*dst*/) {
-	EXIT("not implemented\n");
-	return false;
+bool SysFileCopyFile(const std::filesystem::path& src, const std::filesystem::path& dst) {
+	std::error_code error;
+	return std::filesystem::copy_file(get_internal_name(src), get_internal_name(dst),
+	                                  std::filesystem::copy_options::overwrite_existing, error) &&
+	       !error;
 }
 
-bool SysFileMoveFile(const std::filesystem::path& /*src*/, const std::filesystem::path& /*dst*/) {
-	EXIT("not implemented\n");
-	return false;
+bool SysFileMoveFile(const std::filesystem::path& src, const std::filesystem::path& dst) {
+	auto real_src = get_internal_name(src);
+	auto real_dst = get_internal_name(dst);
+
+	// Match MoveFileW: fail when the destination exists.
+	std::error_code error;
+	if (std::filesystem::exists(real_dst, error)) {
+		return false;
+	}
+
+	return 0 == rename(real_src.string().c_str(), real_dst.string().c_str());
 }
 
-void SysFileRemoveReadonly(const std::filesystem::path& /*name*/) {
-	EXIT("not implemented\n");
+void SysFileRemoveReadonly(const std::filesystem::path& name) {
+	auto real_name     = get_internal_name(name);
+	auto real_name_str = real_name.string();
+
+	struct stat s {};
+
+	if (0 != stat(real_name_str.c_str(), &s)) {
+		return;
+	}
+
+	chmod(real_name_str.c_str(), s.st_mode | S_IWUSR);
 }
 
 #endif

src/common/platform/sysLinuxVirtual.cpp

@@ -8,9 +8,16 @@
 #include "common/platform/sysVirtual.h"
 #include "common/virtualMemory.h"
 
+#include <atomic>
 #include <map>
 #include <pthread.h>
 #include <sys/mman.h>
+#include <unistd.h>
+
+#if defined(__APPLE__)
+#include <mach/mach.h>
+#include <mach/mach_vm.h>
+#endif
 
 // IWYU pragma: no_include <asm/mman-common.h>
 // IWYU pragma: no_include <asm/mman.h>
@@ -31,8 +38,8 @@ void SysVirtualInit() {
 	pthread_mutexattr_t attr {};
 
 	pthread_mutexattr_init(&attr);
-#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX
-	pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_FAST_NP);
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX && !defined(__APPLE__)
+	pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_FAST_NP); // glibc-only fast mutex
 #else
 	pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_NORMAL);
 #endif
@@ -76,6 +83,70 @@ static VirtualMemory::Mode get_protection_flag(int mode) {
 	}
 }
 
+// Keep automatic mappings inside the guest and GPU-addressable low window.
+#ifdef KYTY_FIXED_NOREPLACE
+static constexpr uintptr_t LOW_ARENA_LIMIT = 0x000000FC00000000ULL; // libc mspace window ceiling
+static constexpr uintptr_t LOW_ARENA_FLOOR = 0x000000A000000000ULL; // 640 GiB
+static constexpr uintptr_t LOW_ARENA_GRAIN = 0x0000000000010000ULL; // 64 KiB
+
+static_assert(LOW_ARENA_LIMIT <= 0x0000010000000000ULL,
+              "arena must stay inside the GPU page tracker's 1<<40 window");
+static_assert(LOW_ARENA_FLOOR < LOW_ARENA_LIMIT, "arena floor must sit below its ceiling");
+
+static std::atomic<uintptr_t> g_low_arena_next {LOW_ARENA_LIMIT};
+#endif
+
+// Caller holds g_virtual_mutex.
+static void record_alloc(uintptr_t addr, size_t size) {
+	auto next = g_allocs->upper_bound(addr);
+	if (next != g_allocs->begin()) {
+		auto       it         = std::prev(next);
+		const auto alloc_addr = it->first;
+		const auto alloc_end  = alloc_addr + it->second;
+		if (alloc_addr <= addr && addr + size <= alloc_end) {
+			g_allocs->erase(it);
+			if (alloc_addr < addr) {
+				(*g_allocs)[alloc_addr] = addr - alloc_addr;
+			}
+			if (addr + size < alloc_end) {
+				(*g_allocs)[addr + size] = alloc_end - (addr + size);
+			}
+		}
+	}
+	(*g_allocs)[addr] = size;
+}
+
+#ifdef KYTY_FIXED_NOREPLACE
+static uintptr_t align_up_to(uintptr_t addr, uint64_t alignment) {
+	return (addr + alignment - 1) & ~(alignment - 1);
+}
+#endif
+
+// Freed arena addresses are not reused while GPU caches remain keyed by address.
+static void* map_anonymous(uintptr_t addr, size_t size, int protect, int flags) {
+	if (addr != 0) {
+		return mmap(reinterpret_cast<void*>(addr), size, protect, flags, -1, 0); // NOLINT
+	}
+
+#ifdef KYTY_FIXED_NOREPLACE
+	const auto step = align_up_to(size, LOW_ARENA_GRAIN);
+	for (int attempt = 0; attempt < 256; attempt++) {
+		const auto top = g_low_arena_next.fetch_sub(step, std::memory_order_relaxed);
+		if (top < step || top - step < LOW_ARENA_FLOOR) {
+			break;
+		}
+		const auto hint = (top - step) & ~(LOW_ARENA_GRAIN - 1);
+		void* ptr = mmap(reinterpret_cast<void*>(hint), size, protect, flags | MAP_FIXED_NOREPLACE,
+		                 -1, 0); // NOLINT
+		if (ptr != MAP_FAILED) {
+			return ptr;
+		}
+	}
+#endif
+
+	return mmap(nullptr, size, protect, flags, -1, 0); // NOLINT
+}
+
 uint64_t SysVirtualAlloc(uint64_t address, uint64_t size, VirtualMemory::Mode mode) {
 	EXIT_IF(g_allocs == nullptr);
 
@@ -83,16 +154,15 @@ uint64_t SysVirtualAlloc(uint64_t address, uint64_t size, VirtualMemory::Mode mo
 
 	int protect = get_protection_flag(mode);
 
-	void* ptr =
-	    mmap(reinterpret_cast<void*>(addr), size, protect, MAP_PRIVATE | MAP_ANON, -1, 0); // NOLINT
+	void* ptr = map_anonymous(addr, size, protect, MAP_PRIVATE | MAP_ANON);
 
 	auto ret_addr = reinterpret_cast<uintptr_t>(ptr);
 
 	if (ptr != MAP_FAILED) {
 		pthread_mutex_lock(&g_virtual_mutex);
-		(*g_allocs)[ret_addr] = size;
-		uintptr_t page_start  = ret_addr >> 12u;
-		uintptr_t page_end    = (ret_addr + size - 1) >> 12u;
+		record_alloc(ret_addr, size);
+		uintptr_t page_start = ret_addr >> 12u;
+		uintptr_t page_end   = (ret_addr + size - 1) >> 12u;
 		for (uintptr_t page = page_start; page <= page_end; page++) {
 			(*g_protects)[page] = protect;
 		}
@@ -117,18 +187,43 @@ uint64_t SysVirtualAllocAligned(uint64_t address, uint64_t size, VirtualMemory::
 	auto addr    = static_cast<uintptr_t>(address);
 	int  protect = get_protection_flag(mode);
 
-	void* ptr =
-	    mmap(reinterpret_cast<void*>(addr), size, protect, MAP_PRIVATE | MAP_ANON, -1, 0); // NOLINT
+	void* ptr = map_anonymous(addr, size, protect, MAP_PRIVATE | MAP_ANON);
 
 	auto ret_addr = reinterpret_cast<uintptr_t>(ptr);
 
 	if (ptr != MAP_FAILED && ((ret_addr & (alignment - 1)) != 0)) {
 		munmap(ptr, size);
 
-		ptr      = mmap(reinterpret_cast<void*>(addr), size + alignment, protect,
-		                MAP_PRIVATE | MAP_ANON | MAP_NORESERVE, -1, 0); // NOLINT
+		ptr =
+		    map_anonymous(addr, size + alignment, protect, MAP_PRIVATE | MAP_ANON | MAP_NORESERVE);
 		ret_addr = reinterpret_cast<uintptr_t>(ptr);
 		if (ptr != MAP_FAILED) {
+#if defined(__APPLE__)
+			// Carve the aligned subrange out of the live mapping with MAP_FIXED (in-place
+			// replacement) and trim the slack; never munmap the whole range first, or a
+			// concurrent host mapping (dyld, Rosetta, Metal) could claim the hole and be
+			// destroyed by the MAP_FIXED. Other platforms keep the original path below.
+			auto aligned_addr = align_up(ret_addr, alignment);
+			// NOLINTNEXTLINE
+			void* fixed = mmap(reinterpret_cast<void*>(aligned_addr), size, protect,
+			                   MAP_FIXED | MAP_PRIVATE | MAP_ANON, -1, 0);
+			if (fixed == MAP_FAILED) {
+				munmap(ptr, size + alignment);
+				ret_addr = 0;
+				ptr      = MAP_FAILED;
+			} else {
+				if (aligned_addr > ret_addr) {
+					munmap(reinterpret_cast<void*>(ret_addr), aligned_addr - ret_addr);
+				}
+				const uintptr_t tail_start = aligned_addr + size;
+				const uintptr_t resv_end   = ret_addr + size + alignment;
+				if (resv_end > tail_start) {
+					munmap(reinterpret_cast<void*>(tail_start), resv_end - tail_start);
+				}
+				ptr      = fixed;
+				ret_addr = aligned_addr;
+			}
+#else
 			munmap(ptr, size + alignment);
 			auto aligned_addr = align_up(ret_addr, alignment);
 #ifdef KYTY_FIXED_NOREPLACE
@@ -146,6 +241,7 @@ uint64_t SysVirtualAllocAligned(uint64_t address, uint64_t size, VirtualMemory::
 				ret_addr = 0;
 				ptr      = MAP_FAILED;
 			}
+#endif
 		}
 	}
 
@@ -154,9 +250,9 @@ uint64_t SysVirtualAllocAligned(uint64_t address, uint64_t size, VirtualMemory::
 	}
 
 	pthread_mutex_lock(&g_virtual_mutex);
-	(*g_allocs)[ret_addr] = size;
-	uintptr_t page_start  = ret_addr >> 12u;
-	uintptr_t page_end    = (ret_addr + size - 1) >> 12u;
+	record_alloc(ret_addr, size);
+	uintptr_t page_start = ret_addr >> 12u;
+	uintptr_t page_end   = (ret_addr + size - 1) >> 12u;
 	for (uintptr_t page = page_start; page <= page_end; page++) {
 		(*g_protects)[page] = protect;
 	}
@@ -165,6 +261,27 @@ uint64_t SysVirtualAllocAligned(uint64_t address, uint64_t size, VirtualMemory::
 	return ret_addr;
 }
 
+#if defined(__APPLE__)
+// macOS has no /proc/self/maps; query the Mach VM map directly. mach_vm_region returns
+// the first mapped region at or above `region_addr`; if it begins before the end of the
+// requested range, the range overlaps an existing mapping.
+static bool is_mapped(void* ptr, size_t length) {
+	auto                           query_addr  = reinterpret_cast<mach_vm_address_t>(ptr);
+	mach_vm_address_t              region_addr = query_addr;
+	mach_vm_size_t                 region_size = 0;
+	vm_region_basic_info_data_64_t info {};
+	mach_msg_type_number_t         count       = VM_REGION_BASIC_INFO_COUNT_64;
+	mach_port_t                    object_name = MACH_PORT_NULL;
+
+	kern_return_t kr =
+	    mach_vm_region(mach_task_self(), &region_addr, &region_size, VM_REGION_BASIC_INFO_64,
+	                   reinterpret_cast<vm_region_info_t>(&info), &count, &object_name);
+	if (kr != KERN_SUCCESS) {
+		return false; // no region at or above the address → unmapped
+	}
+	return region_addr < (query_addr + length);
+}
+#else
 static bool is_mapped(void* ptr, size_t length) {
 	FILE* file = fopen("/proc/self/maps", "r");
 	char  line[1024];
@@ -189,6 +306,7 @@ static bool is_mapped(void* ptr, size_t length) {
 	fclose(file);
 	return ret;
 }
+#endif
 
 bool SysVirtualAllocFixed(uint64_t address, uint64_t size, VirtualMemory::Mode mode) {
 	EXIT_IF(g_allocs == nullptr);
@@ -218,9 +336,9 @@ bool SysVirtualAllocFixed(uint64_t address, uint64_t size, VirtualMemory::Mode m
 
 	if (ptr != MAP_FAILED) {
 		pthread_mutex_lock(&g_virtual_mutex);
-		(*g_allocs)[ret_addr] = size;
-		uintptr_t page_start  = ret_addr >> 12u;
-		uintptr_t page_end    = (ret_addr + size - 1) >> 12u;
+		record_alloc(ret_addr, size);
+		uintptr_t page_start = ret_addr >> 12u;
+		uintptr_t page_end   = (ret_addr + size - 1) >> 12u;
 		for (uintptr_t page = page_start; page <= page_end; page++) {
 			(*g_protects)[page] = protect;
 		}
@@ -249,18 +367,44 @@ uint64_t SysVirtualReserveAligned(uint64_t address, uint64_t size, uint64_t alig
 
 	auto addr = static_cast<uintptr_t>(address);
 
-	void* ptr = mmap(reinterpret_cast<void*>(addr), size, PROT_NONE,
-	                 MAP_PRIVATE | MAP_ANON | MAP_NORESERVE, -1, 0); // NOLINT
+	void* ptr = map_anonymous(addr, size, PROT_NONE, MAP_PRIVATE | MAP_ANON | MAP_NORESERVE);
 
 	auto ret_addr = reinterpret_cast<uintptr_t>(ptr);
 
 	if (ptr != MAP_FAILED && ((ret_addr & (alignment - 1)) != 0)) {
 		munmap(ptr, size);
 
-		ptr      = mmap(reinterpret_cast<void*>(addr), size + alignment, PROT_NONE,
-		                MAP_PRIVATE | MAP_ANON | MAP_NORESERVE, -1, 0); // NOLINT
+		ptr      = map_anonymous(addr, size + alignment, PROT_NONE,
+		                         MAP_PRIVATE | MAP_ANON | MAP_NORESERVE);
 		ret_addr = reinterpret_cast<uintptr_t>(ptr);
 		if (ptr != MAP_FAILED) {
+#if defined(__APPLE__)
+			// Carve the aligned subrange out of the live reservation with MAP_FIXED (an
+			// in-place replacement), then trim the slack. The range must never be
+			// returned to the OS in between: another thread (dyld, Rosetta, Metal,
+			// malloc) could claim the hole, and the subsequent MAP_FIXED would silently
+			// destroy its mapping. Other platforms keep the original path below.
+			auto aligned_addr = align_up(ret_addr, alignment);
+			// NOLINTNEXTLINE
+			void* fixed = mmap(reinterpret_cast<void*>(aligned_addr), size, PROT_NONE,
+			                   MAP_FIXED | MAP_PRIVATE | MAP_ANON | MAP_NORESERVE, -1, 0);
+			if (fixed == MAP_FAILED) {
+				munmap(ptr, size + alignment);
+				ret_addr = 0;
+				ptr      = MAP_FAILED;
+			} else {
+				if (aligned_addr > ret_addr) {
+					munmap(reinterpret_cast<void*>(ret_addr), aligned_addr - ret_addr);
+				}
+				const uintptr_t tail_start = aligned_addr + size;
+				const uintptr_t resv_end   = ret_addr + size + alignment;
+				if (resv_end > tail_start) {
+					munmap(reinterpret_cast<void*>(tail_start), resv_end - tail_start);
+				}
+				ptr      = fixed;
+				ret_addr = aligned_addr;
+			}
+#else
 			munmap(ptr, size + alignment);
 			auto aligned_addr = align_up(ret_addr, alignment);
 #ifdef KYTY_FIXED_NOREPLACE
@@ -278,6 +422,7 @@ uint64_t SysVirtualReserveAligned(uint64_t address, uint64_t size, uint64_t alig
 				ret_addr = 0;
 				ptr      = MAP_FAILED;
 			}
+#endif
 		}
 	}
 
@@ -286,12 +431,7 @@ uint64_t SysVirtualReserveAligned(uint64_t address, uint64_t size, uint64_t alig
 	}
 
 	pthread_mutex_lock(&g_virtual_mutex);
-	(*g_allocs)[ret_addr] = size;
-	uintptr_t page_start  = ret_addr >> 12u;
-	uintptr_t page_end    = (ret_addr + size - 1) >> 12u;
-	for (uintptr_t page = page_start; page <= page_end; page++) {
-		(*g_protects)[page] = PROT_NONE;
-	}
+	record_alloc(ret_addr, size);
 	pthread_mutex_unlock(&g_virtual_mutex);
 
 	return ret_addr;
@@ -324,12 +464,7 @@ bool SysVirtualReserveFixed(uint64_t address, uint64_t size) {
 
 	if (ptr != MAP_FAILED) {
 		pthread_mutex_lock(&g_virtual_mutex);
-		(*g_allocs)[ret_addr] = size;
-		uintptr_t page_start  = ret_addr >> 12u;
-		uintptr_t page_end    = (ret_addr + size - 1) >> 12u;
-		for (uintptr_t page = page_start; page <= page_end; page++) {
-			(*g_protects)[page] = PROT_NONE;
-		}
+		record_alloc(ret_addr, size);
 		pthread_mutex_unlock(&g_virtual_mutex);
 
 		return true;
@@ -339,7 +474,29 @@ bool SysVirtualReserveFixed(uint64_t address, uint64_t size) {
 }
 
 bool SysVirtualDecommit(uint64_t address, uint64_t size) {
-	return SysVirtualProtect(address, size, VirtualMemory::Mode::NoAccess);
+	// Drop physical pages while preserving the reservation.
+	if (!SysVirtualProtect(address, size, VirtualMemory::Mode::NoAccess)) {
+		return false;
+	}
+
+	if (size != 0) {
+#if defined(__APPLE__)
+		constexpr int RECLAIM_ADVICE = MADV_FREE;
+#else
+		constexpr int RECLAIM_ADVICE = MADV_DONTNEED;
+#endif
+		const auto page_size = static_cast<uintptr_t>(sysconf(_SC_PAGESIZE));
+		if (page_size != 0) {
+			// Do not discard pages outside the requested range.
+			const auto begin = (static_cast<uintptr_t>(address) + page_size - 1) & ~(page_size - 1);
+			const auto end   = (static_cast<uintptr_t>(address) + size) & ~(page_size - 1);
+			if (end > begin) {
+				::madvise(reinterpret_cast<void*>(begin), end - begin, RECLAIM_ADVICE);
+			}
+		}
+	}
+
+	return true;
 }
 
 bool SysVirtualFree(uint64_t address) {
@@ -391,15 +548,34 @@ bool SysVirtualFreeRange(uint64_t address, uint64_t size) {
 		pthread_mutex_unlock(&g_virtual_mutex);
 		return false;
 	}
-	auto       it         = std::prev(next);
-	const auto alloc_addr = it->first;
-	const auto alloc_end  = alloc_addr + it->second;
-	if (addr < alloc_addr || end > alloc_end || munmap(reinterpret_cast<void*>(addr), size) != 0) {
+
+	// A reservation may have been split into several adjacent records.
+	auto       first      = std::prev(next);
+	const auto alloc_addr = first->first;
+	if (addr < alloc_addr || alloc_addr + first->second <= addr) {
 		pthread_mutex_unlock(&g_virtual_mutex);
 		return false;
 	}
 
-	g_allocs->erase(it);
+	auto      last   = first;
+	uintptr_t cursor = alloc_addr + first->second;
+	while (cursor < end) {
+		auto following = std::next(last);
+		if (following == g_allocs->end() || following->first != cursor) {
+			pthread_mutex_unlock(&g_virtual_mutex);
+			return false;
+		}
+		last   = following;
+		cursor = following->first + following->second;
+	}
+	const auto alloc_end = cursor;
+
+	if (munmap(reinterpret_cast<void*>(addr), size) != 0) {
+		pthread_mutex_unlock(&g_virtual_mutex);
+		return false;
+	}
+
+	g_allocs->erase(first, std::next(last));
 	if (alloc_addr < addr) {
 		(*g_allocs)[alloc_addr] = addr - alloc_addr;
 	}

src/common/platform/sysTimer.h

@@ -27,7 +27,9 @@ struct SysFileTimeStruct {
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	FILETIME time;
 #elif KYTY_PLATFORM == KYTY_PLATFORM_LINUX
+	// Nanoseconds preserve sub-second file timestamps.
 	time_t time;
+	long   nanos;
 #endif
 	bool is_invalid;
 };
@@ -139,7 +141,7 @@ inline void SysFileToSystemTimeUtc(const SysFileTimeStruct& f, SysTimeStruct& t)
 	t.Hour         = i.tm_hour;
 	t.Minute       = i.tm_min;
 	t.Second       = (i.tm_sec == 60 ? 59 : i.tm_sec);
-	t.Milliseconds = 0;
+	t.Milliseconds = static_cast<uint16_t>((f.nanos / 1000000) % 1000);
 }
 
 inline void SysTimeTToSystem(time_t t, SysTimeStruct& s) {
@@ -168,10 +170,11 @@ inline void SysSystemToFileTimeUtc(const SysTimeStruct& f, SysFileTimeStruct& t)
 
 // Retrieves the current local date and time.
 inline void SysGetSystemTime(SysTimeStruct& t) {
-	time_t    st {};
+	// Preserve millisecond precision.
+	timespec  now {};
 	struct tm i {};
 
-	if (time(&st) == static_cast<time_t>(-1) || localtime_r(&st, &i) == nullptr) {
+	if (clock_gettime(CLOCK_REALTIME, &now) != 0 || localtime_r(&now.tv_sec, &i) == nullptr) {
 		t.is_invalid = true;
 		return;
 	}
@@ -183,15 +186,16 @@ inline void SysGetSystemTime(SysTimeStruct& t) {
 	t.Hour         = i.tm_hour;
 	t.Minute       = i.tm_min;
 	t.Second       = (i.tm_sec == 60 ? 59 : i.tm_sec);
-	t.Milliseconds = 0;
+	t.Milliseconds = static_cast<uint16_t>((now.tv_nsec / 1000000) % 1000);
 }
 
 // Retrieves the current system date and time in Coordinated Universal Time (UTC).
 inline void SysGetSystemTimeUtc(SysTimeStruct& t) {
-	time_t    st {};
+	// Preserve millisecond precision.
+	timespec  now {};
 	struct tm i {};
 
-	if (time(&st) == static_cast<time_t>(-1) || gmtime_r(&st, &i) == nullptr) {
+	if (clock_gettime(CLOCK_REALTIME, &now) != 0 || gmtime_r(&now.tv_sec, &i) == nullptr) {
 		t.is_invalid = true;
 		return;
 	}
@@ -203,7 +207,7 @@ inline void SysGetSystemTimeUtc(SysTimeStruct& t) {
 	t.Hour         = i.tm_hour;
 	t.Minute       = i.tm_min;
 	t.Second       = (i.tm_sec == 60 ? 59 : i.tm_sec);
-	t.Milliseconds = 0;
+	t.Milliseconds = static_cast<uint16_t>((now.tv_nsec / 1000000) % 1000);
 }
 
 inline void SysQueryPerformanceFrequency(uint64_t* freq) {

src/common/threads.cpp

@@ -5,6 +5,7 @@
 
 #include <algorithm>
 #include <atomic>
+#include <cerrno>
 #include <chrono>             // IWYU pragma: keep
 #include <condition_variable> // IWYU pragma: keep
 #include <mutex>
@@ -14,6 +15,12 @@
 #define KYTY_WIN_CS
 #endif
 
+// macOS has no clock_nanosleep.
+#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS && !defined(__APPLE__)
+#define KYTY_POSIX_HIGH_RES_SLEEP
+#include <ctime>
+#endif
+
 #include <sstream>
 #include <string>
 #include <thread>
@@ -121,6 +128,42 @@ static SleepConditionVariableCS_func_t ResolveSleepConditionVariableCS() {
 
 #endif
 
+#ifdef KYTY_POSIX_HIGH_RES_SLEEP
+// Spin for very short waits; use an absolute deadline for longer waits.
+static void SleepHighResolutionNanos(uint64_t nanos) {
+	if (nanos == 0) {
+		return;
+	}
+
+	constexpr uint64_t NANOS_PER_SEC = 1000000000;
+	constexpr uint64_t SPIN_LIMIT_NS = 50000; // below this a context switch dominates
+
+	timespec deadline {};
+	if (clock_gettime(CLOCK_MONOTONIC, &deadline) != 0) {
+		std::this_thread::sleep_for(std::chrono::nanoseconds(nanos));
+		return;
+	}
+
+	auto target_nsec = static_cast<uint64_t>(deadline.tv_nsec) + nanos;
+	deadline.tv_sec += static_cast<time_t>(target_nsec / NANOS_PER_SEC);
+	deadline.tv_nsec = static_cast<long>(target_nsec % NANOS_PER_SEC);
+
+	if (nanos <= SPIN_LIMIT_NS) {
+		timespec now {};
+		do {
+			if (clock_gettime(CLOCK_MONOTONIC, &now) != 0) {
+				return;
+			}
+		} while (now.tv_sec < deadline.tv_sec ||
+		         (now.tv_sec == deadline.tv_sec && now.tv_nsec < deadline.tv_nsec));
+		return;
+	}
+
+	while (clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME, &deadline, nullptr) == EINTR) {
+	}
+}
+#endif
+
 namespace Common {
 
 using thread_id_t = std::thread::id;
@@ -249,6 +292,8 @@ void Thread::Sleep(uint32_t millis) {
 void Thread::SleepMicro(uint32_t micros) {
 #ifdef KYTY_WIN_CS
 	SleepHighResolution100ns(static_cast<uint64_t>(micros) * 10);
+#elif defined(KYTY_POSIX_HIGH_RES_SLEEP)
+	SleepHighResolutionNanos(static_cast<uint64_t>(micros) * 1000);
 #else
 	std::this_thread::sleep_for(std::chrono::microseconds(micros));
 #endif
@@ -257,6 +302,8 @@ void Thread::SleepMicro(uint32_t micros) {
 void Thread::SleepNano(uint64_t nanos) {
 #ifdef KYTY_WIN_CS
 	SleepHighResolution100ns((nanos + 99) / 100);
+#elif defined(KYTY_POSIX_HIGH_RES_SLEEP)
+	SleepHighResolutionNanos(nanos);
 #else
 	std::this_thread::sleep_for(std::chrono::nanoseconds(nanos));
 #endif

src/common/uniqueFunction.h

@@ -11,7 +11,7 @@ template <typename Result, typename... Args>
 class UniqueFunction {
 	class CallableBase {
 	public:
-		virtual ~CallableBase() = default;
+		virtual ~CallableBase()               = default;
 		virtual Result Invoke(Args&&... args) = 0;
 	};
 
@@ -20,9 +20,7 @@ class UniqueFunction {
 	public:
 		explicit Callable(Function function): m_function(std::move(function)) {}
 
-		Result Invoke(Args&&... args) override {
-			return m_function(std::forward<Args>(args)...);
-		}
+		Result Invoke(Args&&... args) override { return m_function(std::forward<Args>(args)...); }
 
 	private:
 		Function m_function;

src/common/virtualMemory.cpp

@@ -58,25 +58,6 @@ bool FlushInstructionCache(uint64_t address, uint64_t size) {
 	return SysVirtualFlushInstructionCache(address, size);
 }
 
-bool PatchReplace(uint64_t vaddr, uint64_t value) {
-	Mode old_mode {};
-	Protect(vaddr, 8, Mode::ReadWrite, &old_mode);
-
-	auto* ptr = reinterpret_cast<uint64_t*>(vaddr);
-
-	bool ret = (*ptr != value);
-
-	*ptr = value;
-
-	Protect(vaddr, 8, old_mode);
-
-	if (IsExecute(old_mode)) {
-		FlushInstructionCache(vaddr, 8);
-	}
-
-	return ret;
-}
-
 } // namespace VirtualMemory
 
 } // namespace Common

src/common/virtualMemory.h

@@ -37,7 +37,6 @@ bool     Free(uint64_t address);
 bool     FreeRange(uint64_t address, uint64_t size);
 bool     Protect(uint64_t address, uint64_t size, Mode mode, Mode* old_mode = nullptr);
 bool     FlushInstructionCache(uint64_t address, uint64_t size);
-bool     PatchReplace(uint64_t vaddr, uint64_t value);
 
 } // namespace VirtualMemory
 

src/emulator.cpp

@@ -105,7 +105,7 @@ static void ClearDebugTextureFolder() {
 	}
 }
 
-static void Init(const Config::ConfigOptions& cfg) {
+static void Init(const Config::ConfigOptions& cfg, const std::filesystem::path& param_json) {
 	EXIT_IF(!Common::Thread::IsMainThread());
 
 	auto* slist = Common::SubsystemsList::Instance();
@@ -127,12 +127,21 @@ static void Init(const Config::ConfigOptions& cfg) {
 	slist->InitAll(true);
 
 	Config::Load(cfg);
+	slist->Add(log, {core, config});
+	slist->InitAll(true);
+
+	if (Common::File::IsFileExisting(param_json)) {
+		Loader::SystemContentLoadParamSfo(param_json);
+		if (const auto flexible_memory_size = Loader::SystemContentGetFlexibleMemorySize();
+		    flexible_memory_size != 0) {
+			Libs::LibKernel::Memory::SetFlexibleMemorySize(flexible_memory_size);
+		}
+	}
 
 	slist->Add(audio, {core, log, pthread, memory});
 	slist->Add(controller, {core, log, config});
 	slist->Add(file_system, {core, log, pthread});
 	slist->Add(graphics, {core, log, pthread, memory, config, profiler, controller});
-	slist->Add(log, {core, config});
 	slist->Add(memory, {core, log});
 	slist->Add(network, {core, log, pthread});
 	slist->Add(profiler, {core, config});
@@ -159,13 +168,14 @@ static void LoadElf(const std::filesystem::path& elf, bool dbg_print_reloc = fal
 	}
 }
 
-static void Execute() {
+static void Execute(const std::filesystem::path& game_patch) {
+	auto           patch_path = game_patch;
 	Common::Thread guest_thread(
-	    [](void* /*unused*/) {
+	    [](void* param) {
 		    auto* rt = Common::Singleton<Loader::RuntimeLinker>::Instance();
-		    rt->Execute();
+		    rt->Execute(*static_cast<const std::filesystem::path*>(param));
 	    },
-	    nullptr);
+	    &patch_path);
 	Libs::Graphics::WindowRun();
 	std::quick_exit(0);
 }
@@ -179,7 +189,8 @@ void Run(const RunOptions& options) {
 		EXIT("ELF is required\n");
 	}
 
-	Init(options.config);
+	const auto param_json = options.app0_dir / "sce_sys" / "param.json";
+	Init(options.config, param_json);
 
 	ClearDebugTextureFolder();
 
@@ -191,15 +202,6 @@ void Run(const RunOptions& options) {
 	Libs::LibKernel::FileSystem::Mount(options.app0_dir, "/app0");
 	Libs::LibKernel::FileSystem::Mount(options.app0_dir, "/hostapp");
 
-	auto param_json = options.app0_dir / "sce_sys" / "param.json";
-	if (Common::File::IsFileExisting(param_json)) {
-		Loader::SystemContentLoadParamSfo(param_json);
-		if (auto flexible_memory_size = Loader::SystemContentGetFlexibleMemorySize();
-		    flexible_memory_size != 0) {
-			Libs::LibKernel::Memory::SetFlexibleMemorySize(flexible_memory_size);
-		}
-	}
-
 	MountSandboxDirs();
 
 	auto* rt = Common::Singleton<Loader::RuntimeLinker>::Instance();
@@ -207,7 +209,7 @@ void Run(const RunOptions& options) {
 
 	LoadElf(options.elf);
 
-	Execute();
+	Execute(options.game_patch);
 }
 
 } // namespace Emulator

src/emulator.h

@@ -12,6 +12,7 @@ struct RunOptions {
 	Config::ConfigOptions config;
 	std::filesystem::path app0_dir;
 	std::filesystem::path elf;
+	std::filesystem::path game_patch;
 };
 
 void Run(const RunOptions& options);

src/graphics/guest_gpu/command_processor/commandProcessor.h

@@ -158,7 +158,7 @@ private:
 	void                 CheckBuffer() const { GetScheduler().CheckActive(); }
 	GpuResourceManager&  GetGpuResources() const { return m_renderer.GetGpuResources(); }
 
-	RenderContext&    m_renderer;
+	RenderContext&   m_renderer;
 	HW::Context      m_ctx;
 	HW::UserConfig   m_ucfg;
 	HW::Shader       m_sh_ctx;
@@ -170,9 +170,9 @@ private:
 	uint64_t         m_dispatch_indirect_args_base_addr = 0;
 	uint32_t         m_num_instances                    = 1;
 
-	uint32_t m_de_count          = 0;
-	uint32_t m_ce_count          = 0;
-	bool     m_ce_complete       = false;
+	uint32_t m_de_count        = 0;
+	uint32_t m_ce_count        = 0;
+	bool     m_ce_complete     = false;
 	bool     m_readback_active = false;
 
 	uint32_t m_const_ram[0x3000] = {0};

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -65,10 +65,23 @@ constexpr uint32_t GcrKnownMask             = GcrGl2MetadataInvalidate | GcrGl0V
                                               GcrGl2Writeback | GcrOrder012 | GcrOrder210;
 constexpr uint32_t RegisterSelectorMask     = 0x70000000u;
 
-uint32_t NormalizeRegisterOffset(uint32_t raw_offset) {
-	return (raw_offset & ~RegisterSelectorMask);
+constexpr uint32_t NormalizeRegisterOffset(uint32_t raw_offset) {
+	return raw_offset & ~RegisterSelectorMask;
 }
 
+// Indirect Cx descriptors retain their selector. Selector 1 offsets 0..31 address the
+// SPI_PS_INPUT_CNTL register bank; ordinary context-register offsets remain unchanged.
+constexpr uint32_t DecodeIndirectCxRegisterOffset(uint32_t raw_offset) {
+	const auto offset = NormalizeRegisterOffset(raw_offset);
+	return (raw_offset & RegisterSelectorMask) == Pm4::CX_PS_SHADER_USAGE_BASE && offset < 32u
+	           ? Pm4::SPI_PS_INPUT_CNTL_0 + offset
+	           : offset;
+}
+
+static_assert(DecodeIndirectCxRegisterOffset(Pm4::CX_PS_SHADER_USAGE_BASE + 2u) ==
+              Pm4::SPI_PS_INPUT_CNTL_0 + 2u);
+static_assert(DecodeIndirectCxRegisterOffset(Pm4::DB_Z_INFO) == Pm4::DB_Z_INFO);
+
 bool ReleaseMemGcrNeedsBarrier(uint32_t eop_event_type, uint32_t gcr_cntl) {
 	return eop_event_type != 0x28u ||
 	       (gcr_cntl & (GcrGl2MetadataInvalidate | GcrGl0VectorInvalidate | GcrGl1Invalidate |
@@ -1904,17 +1917,23 @@ KYTY_CP_OP_PARSER(CpOpCopyData) {
 
 	EXIT_NOT_IMPLEMENTED(cmd_id != KYTY_PM4(6, Pm4::IT_COPY_DATA, 0u));
 
-	const uint32_t control       = buffer[0];
-	const uint32_t src_sel       = ((control & 0xfu) << 1u) | ((control >> 30u) & 0x1u);
-	const uint32_t dst_sel       = ((control >> 8u) & 0xfu) << 1u;
-	const uint8_t  src_cache     = static_cast<uint8_t>((control >> 13u) & 0x3u);
-	const uint8_t  dst_cache     = static_cast<uint8_t>((control >> 25u) & 0x3u);
-	const uint8_t  write_confirm = static_cast<uint8_t>((control >> 20u) & 0x1u);
-	const uint32_t num_bytes     = ((control >> 16u) & 0x1u) != 0 ? 8u : 4u;
-	const uint64_t src           = buffer[1] | (static_cast<uint64_t>(buffer[2]) << 32u);
-	const uint64_t dst           = buffer[3] | (static_cast<uint64_t>(buffer[4]) << 32u);
-	if (src_sel == (9u << 1u)) {
-		if (dst_sel != (2u << 1u) || dst == 0 || (dst & (num_bytes - 1u)) != 0) {
+	const uint32_t control             = buffer[0];
+	const uint32_t src_sel             = ((control & 0xfu) << 1u) | ((control >> 30u) & 0x1u);
+	const uint32_t dst_sel             = ((control >> 8u) & 0xfu) << 1u;
+	const uint8_t  src_cache           = static_cast<uint8_t>((control >> 13u) & 0x3u);
+	const uint8_t  dst_cache           = static_cast<uint8_t>((control >> 25u) & 0x3u);
+	const uint8_t  write_confirm       = static_cast<uint8_t>((control >> 20u) & 0x1u);
+	const uint32_t num_bytes           = ((control >> 16u) & 0x1u) != 0 ? 8u : 4u;
+	const uint64_t src                 = buffer[1] | (static_cast<uint64_t>(buffer[2]) << 32u);
+	const uint64_t dst                 = buffer[3] | (static_cast<uint64_t>(buffer[4]) << 32u);
+	uint32_t       reference_clock_dst = 0;
+	switch (src_sel) {
+		case 9u: reference_clock_dst = 2u; break;
+		case 18u: reference_clock_dst = 4u; break;
+		default: break;
+	}
+	if (reference_clock_dst != 0) {
+		if (dst_sel != reference_clock_dst || dst == 0 || (dst & (num_bytes - 1u)) != 0) {
 			EXIT("unsupported reference-clock copyData, src_sel=0x%02" PRIx32
 			     " dst_sel=0x%02" PRIx32 " dst=0x%016" PRIx64 " size=%u\n",
 			     src_sel, dst_sel, dst, num_bytes);
@@ -2321,14 +2340,18 @@ KYTY_CP_OP_PARSER(CpOpIndirectCxRegs) {
 		EXIT("indirect CX registers have null address, num_regs = %" PRIu32 "\n", indirect_num_dw);
 	}
 	for (uint32_t i = 0; i < indirect_num_dw; i++, indirect_buffer += 2) {
-		auto cmd_offset = indirect_buffer[0];
-		auto value      = indirect_buffer[1];
+		// Keep the encoded offset for packet control values, and use the decoded offset only
+		// for register dispatch.
+		auto raw_cmd_offset = indirect_buffer[0];
+		auto cmd_offset     = DecodeIndirectCxRegisterOffset(raw_cmd_offset);
+		auto value          = indirect_buffer[1];
 
 		if (HwCtxTrySetFakeRegister(cmd_offset, value)) {
 			continue;
 		}
 
-		if (cmd_offset == 0xffffffffu) {
+		// The sentinel is an encoded descriptor value and must be checked before normalization.
+		if (raw_cmd_offset == 0xffffffffu) {
 			static bool logged = false;
 			if (!logged) {
 				LOGF("\t temporary: skipping indirect CX sentinel pair offset = 0xffffffff, value "
@@ -3373,6 +3396,12 @@ void GraphicsInitJmpTablesCxIndirect() {
 	g_hw_ctx_indirect_func[Pm4::DB_COUNT_CONTROL] = [](KYTY_HW_CTX_INDIRECT_ARGS) {
 		HwCtxIgnoreDepthMetadataRegister(cmd_offset, value);
 	};
+	for (auto cmd_offset = Pm4::DB_SRESULTS_COMPARE_STATE0;
+	     cmd_offset <= Pm4::DB_SRESULTS_COMPARE_STATE1; cmd_offset++) {
+		g_hw_ctx_indirect_func[cmd_offset] = [](KYTY_HW_CTX_INDIRECT_ARGS) {
+			HwCtxIgnoreDepthMetadataRegister(cmd_offset, value);
+		};
+	}
 	g_hw_ctx_indirect_func[Pm4::DB_RENDER_OVERRIDE] = [](KYTY_HW_CTX_INDIRECT_ARGS) {
 		HwCtxIgnoreDepthMetadataRegister(cmd_offset, value);
 	};

src/graphics/guest_gpu/gpu_defs.h

@@ -72,6 +72,7 @@ enum class ChannelLayout : uint32_t {
 	k32_32       = 11,
 	k16_16_16_16 = 12,
 	k32_32_32_32 = 14,
+	k5_6_5       = 16,
 	k5_5_5_1     = 17,
 	k4_4_4_4     = 19,
 	kBc1         = 35,
@@ -374,6 +375,8 @@ enum class BufferFormat : uint32_t {
 	k32_32_32_32UInt    = 75,
 	k32_32_32_32SInt    = 76,
 	k32_32_32_32Float   = 77,
+	k8Srgb              = 128,
+	k8_8Srgb            = 129,
 	k8_8_8_8Srgb        = 130,
 	k9_9_9_5Float       = 132,
 	k5_6_5UNorm         = 133,

src/graphics/guest_gpu/gpu_format.cpp

@@ -39,6 +39,7 @@ constexpr FormatInfo kFormatInfo[] = {
     {GpuEnumValue(BufferFormat::k16_16Float), 4, 0, 4, true, false},
     {GpuEnumValue(BufferFormat::k11_11_10Float), 4, 0, 4, true, false},
     {GpuEnumValue(BufferFormat::k10_10_10_2UNorm), 4, 0, 4, true, false},
+    {GpuEnumValue(BufferFormat::k10_10_10_2UInt), 4, 0, 4, true, true},
     {GpuEnumValue(BufferFormat::k8_8_8_8UNorm), 4, 0, 4, true, false},
     {GpuEnumValue(BufferFormat::k8_8_8_8SNorm), 4, 0, 4, true, false},
     {GpuEnumValue(BufferFormat::k8_8_8_8UInt), 4, 0, 4, true, true},
@@ -57,6 +58,8 @@ constexpr FormatInfo kFormatInfo[] = {
     {GpuEnumValue(BufferFormat::k32_32_32_32UInt), 16, 0, 16, true, true},
     {GpuEnumValue(BufferFormat::k32_32_32_32SInt), 16, 0, 16, false, false},
     {GpuEnumValue(BufferFormat::k32_32_32_32Float), 16, 0, 16, true, false},
+    {GpuEnumValue(BufferFormat::k8Srgb), 1, 0, 0, true, false},
+    {GpuEnumValue(BufferFormat::k8_8Srgb), 2, 0, 0, true, false},
     {GpuEnumValue(BufferFormat::k8_8_8_8Srgb), 4, 0, 4, true, false},
     {GpuEnumValue(BufferFormat::k9_9_9_5Float), 4, 0, 0, true, false},
     {GpuEnumValue(BufferFormat::k5_6_5UNorm), 2, 0, 2, true, false},

src/graphics/guest_gpu/graphicsRun.cpp

@@ -32,11 +32,10 @@
 
 namespace Libs::Graphics {
 
-static thread_local CommandProcessor* g_current_processor      = nullptr;
-static thread_local Pm4Execution*     g_current_execution      = nullptr;
-static thread_local uint32_t          g_submission_pause_depth = 0;
-static thread_local bool              g_gpu_mutex_owned        = false;
-static thread_local bool              g_gpu_thread             = false;
+static thread_local CommandProcessor* g_current_processor = nullptr;
+static thread_local Pm4Execution*     g_current_execution = nullptr;
+static thread_local bool              g_gpu_mutex_owned   = false;
+static thread_local bool              g_gpu_thread        = false;
 
 class GpuMutexLock final {
 public:
@@ -98,8 +97,6 @@ public:
 	                   bool trigger_agc_interrupt_on_done);
 	void SubmitFlipPreparation(uint64_t request_id);
 	void Done();
-	void PauseSubmissions();
-	void ResumeSubmissions();
 	void Shutdown();
 	[[nodiscard]] bool IsStopping();
 	void               SendCommand(Common::UniqueFunction<void>&& command);
@@ -410,8 +407,13 @@ void CommandProcessor::WriteData(uint32_t* dst, const uint32_t* src, uint32_t dw
 	const uint32_t increment    = (write_control >> 16u) & 0x1u;
 	const uint32_t write_confirm = (write_control >> 20u) & 0x1u;
 
-	if (dst_sel != 0 && dst_sel != 2 && dst_sel != 4 && dst_sel != 5) {
-		EXIT("unsupported writeData destination selector 0x%02" PRIx32 "\n", dst_sel);
+	switch (dst_sel) {
+		case 0:
+		case 2:
+		case 4:
+		case 5:
+		case 6: break;
+		default: EXIT("unsupported writeData destination selector 0x%02" PRIx32 "\n", dst_sel);
 	}
 	EXIT_NOT_IMPLEMENTED(increment != 0);
 
@@ -691,26 +693,6 @@ bool GpuState::Process(Submission& submission) {
 	return complete;
 }
 
-void GpuState::PauseSubmissions() {
-	if (g_gpu_mutex_owned) {
-		EXIT("GPU submissions are already paused by this thread\n");
-	}
-	g_gpu_mutex_owned = true;
-	m_submission_mutex.Lock();
-	if (!IsGpuThread()) {
-		WaitLocked();
-	}
-	m_renderer.GetCommandScheduler().DrainPriorityOperations();
-}
-
-void GpuState::ResumeSubmissions() {
-	if (!g_gpu_mutex_owned) {
-		EXIT("GPU submissions resumed without an active pause\n");
-	}
-	m_submission_mutex.Unlock();
-	g_gpu_mutex_owned = false;
-}
-
 Pm4ProcessResult CommandProcessor::Process(Pm4Execution& execution, uint32_t* buffer,
                                            uint32_t size_dw) {
 	KYTY_PROFILER_BLOCK("CommandProcessor::Process");
@@ -962,9 +944,8 @@ void CommandProcessor::DrawIndexOffset(uint32_t index_offset, uint32_t index_cou
 	auto* index_addr = reinterpret_cast<const void*>(
 	    m_index_base_addr + static_cast<uint64_t>(index_offset) * index_size);
 
-	m_renderer.GetRenderExecutor().DrawIndex(m_submit_id, CurrentBuffer(),
-	                                             m_index_type_and_size, index_count, index_addr,
-	                                             flags, 1, m_num_instances);
+	m_renderer.GetRenderExecutor().DrawIndex(m_submit_id, CurrentBuffer(), m_index_type_and_size,
+	                                         index_count, index_addr, flags, 1, m_num_instances);
 }
 
 void CommandProcessor::DrawIndirect(uint32_t data_offset, uint32_t draw_initiator, bool indexed) {
@@ -1190,8 +1171,8 @@ void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_g
 			}
 		}
 
-		m_renderer.GetRenderExecutor().DispatchDirect(
-		    m_submit_id, CurrentBuffer(), thread_group_x, thread_group_y, thread_group_z, mode);
+		m_renderer.GetRenderExecutor().DispatchDirect(m_submit_id, CurrentBuffer(), thread_group_x,
+		                                              thread_group_y, thread_group_z, mode);
 	}
 
 	constexpr uint32_t DispatchInitiatorUseThreadDimensions = 1u << 5u;
@@ -1237,16 +1218,16 @@ void CommandProcessor::DrawIndexAuto(uint32_t index_count, uint32_t flags,
                                      uint32_t first_vertex, uint32_t first_instance) {
 	CheckBuffer();
 
-	m_renderer.GetRenderExecutor().DrawAuto(
-	    m_submit_id, CurrentBuffer(), index_count, flags, render_target_slice_offset,
-	    instance_count, first_vertex, first_instance);
+	m_renderer.GetRenderExecutor().DrawAuto(m_submit_id, CurrentBuffer(), index_count, flags,
+	                                        render_target_slice_offset, instance_count,
+	                                        first_vertex, first_instance);
 }
 
 void CommandProcessor::WaitFlipDone(uint32_t video_out_handle, uint32_t display_buffer_index) {
 	BufferFlush();
 
 	m_renderer.GetVideoOut().WaitFlipDone(static_cast<int>(video_out_handle),
-	                                              static_cast<int>(display_buffer_index));
+	                                      static_cast<int>(display_buffer_index));
 }
 
 template <typename T>
@@ -1317,8 +1298,8 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
 				if (eop_event_type == 0x2f && cache_action == 0x00 && event_index == 0x06) {
 					auto* dst = static_cast<uint32_t*>(dst_gpu_addr);
 					SynchronizeGpu();
-					Sync::ReadGds(m_renderer.GetBufferCache().GetGdsBuffer(), dst,
-					              value & 0xffffu, value >> 16u);
+					Sync::ReadGds(m_renderer.GetBufferCache().GetGdsBuffer(), dst, value & 0xffffu,
+					              value >> 16u);
 					Sync::WriteAtEndOfPipeGds32(m_submit_id, CurrentBuffer(), dst, value & 0xffffu,
 					                            value >> 16u);
 					return;
@@ -1486,8 +1467,7 @@ void CommandProcessor::EmitGlobalBarrier() {
 	barrier.srcStageMask  = vk::PipelineStageFlagBits2::eAllCommands;
 	barrier.srcAccessMask = vk::AccessFlagBits2::eMemoryWrite;
 	barrier.dstStageMask  = vk::PipelineStageFlagBits2::eAllCommands;
-	barrier.dstAccessMask =
-	    vk::AccessFlagBits2::eMemoryRead | vk::AccessFlagBits2::eMemoryWrite;
+	barrier.dstAccessMask = vk::AccessFlagBits2::eMemoryRead | vk::AccessFlagBits2::eMemoryWrite;
 
 	vk::DependencyInfo dependency {};
 	dependency.memoryBarrierCount = 1;
@@ -1690,32 +1670,6 @@ int Gpu::GetFrameNum() const {
 	return m_state->GetFrameNum();
 }
 
-void Gpu::PauseSubmissions() {
-	m_state->PauseSubmissions();
-}
-
-void Gpu::ResumeSubmissions() {
-	m_state->ResumeSubmissions();
-}
-
-Gpu::SubmissionLock::SubmissionLock(Gpu& gpu): m_gpu(gpu) {
-	if (g_current_processor != nullptr || g_submission_pause_depth == UINT32_MAX) {
-		EXIT("cannot acquire GPU submission lock in the current state\n");
-	}
-	if (g_submission_pause_depth++ == 0) {
-		m_gpu.PauseSubmissions();
-	}
-}
-
-Gpu::SubmissionLock::~SubmissionLock() {
-	if (g_submission_pause_depth == 0) {
-		EXIT("GPU submission lock released without ownership\n");
-	}
-	if (--g_submission_pause_depth == 0) {
-		m_gpu.ResumeSubmissions();
-	}
-}
-
 bool Gpu::IsCommandProcessorThread() noexcept {
 	return g_current_processor != nullptr;
 }
@@ -1724,12 +1678,4 @@ CommandProcessor* Gpu::CurrentCommandProcessor() noexcept {
 	return g_current_processor;
 }
 
-bool Gpu::SubmissionLockHeld() noexcept {
-	return g_submission_pause_depth != 0;
-}
-
-bool Gpu::MutexHeld() noexcept {
-	return g_gpu_mutex_owned;
-}
-
 } // namespace Libs::Graphics

src/graphics/guest_gpu/graphicsRun.h

@@ -35,25 +35,8 @@ public:
 
 	[[nodiscard]] static bool              IsCommandProcessorThread() noexcept;
 	[[nodiscard]] static CommandProcessor* CurrentCommandProcessor() noexcept;
-	[[nodiscard]] static bool              SubmissionLockHeld() noexcept;
-	[[nodiscard]] static bool              MutexHeld() noexcept;
-
-	class SubmissionLock final {
-	public:
-		explicit SubmissionLock(Gpu& gpu);
-		~SubmissionLock();
-		KYTY_CLASS_NO_COPY(SubmissionLock);
-
-	private:
-		Gpu& m_gpu;
-	};
 
 private:
-	friend class SubmissionLock;
-
-	void PauseSubmissions();
-	void ResumeSubmissions();
-
 	std::unique_ptr<GpuState> m_state;
 };
 } // namespace Libs::Graphics

src/graphics/guest_gpu/pm4.cpp

@@ -110,7 +110,6 @@ void DumpPm4PacketStream(Common::File* file, uint32_t* cmd_buffer, uint32_t star
 	auto* cmd = cmd_buffer + start_dw;
 	auto  dw  = num_dw;
 	while (dw != 0) {
-		EXIT_NOT_IMPLEMENTED(dw < 2);
 		EXIT_NOT_IMPLEMENTED(dw > num_dw);
 
 		auto cmd_id = *cmd++;
@@ -120,6 +119,9 @@ void DumpPm4PacketStream(Common::File* file, uint32_t* cmd_buffer, uint32_t star
 		uint32_t len = 0;
 
 		const auto packet_type = static_cast<PacketType>(cmd_id >> 30u);
+		// Type-2 packets are header-only padding; every other packet type requires a body.
+		EXIT_NOT_IMPLEMENTED(dw < 2 && packet_type != PacketType::Type2);
+
 		switch (packet_type) {
 			case PacketType::Type3: {
 				const bool    sh_gx = (cmd_id & 0x2u) == 0;

src/graphics/guest_gpu/pm4.h

@@ -385,6 +385,9 @@ constexpr uint32_t SPI_SHADER_POS_FORMAT  = 0x1C3;
 constexpr uint32_t SPI_SHADER_Z_FORMAT    = 0x1C4;
 constexpr uint32_t SPI_SHADER_COL_FORMAT  = 0x1C5;
 
+// Indirect Cx descriptor selector for the 32-entry PS input-control register bank.
+constexpr uint32_t CX_PS_SHADER_USAGE_BASE = 0x10000000u;
+
 constexpr uint32_t CB_BLEND0_CONTROL                            = 0x1E0;
 constexpr uint32_t CB_BLEND0_CONTROL_COLOR_SRCBLEND_SHIFT       = 0;
 constexpr uint32_t CB_BLEND0_CONTROL_COLOR_SRCBLEND_MASK        = 0x1F;

src/graphics/guest_gpu/tile.h

@@ -65,41 +65,41 @@ struct TileVolumeLayout {
 };
 
 bool TileGetBlockLayout(TileBlockFamily family, uint32_t bytes_per_element,
-	                    TileBlockLayout& layout);
+                        TileBlockLayout& layout);
 bool TileGetBlockOffset(const TileBlockLayout& layout, uint32_t x, uint32_t y, uint32_t z,
-	                    uint32_t& byte_offset);
+                        uint32_t& byte_offset);
 bool TileGetBlockXor(const TileBlockLayout& layout, uint32_t block_x, uint32_t block_y,
-	                 uint32_t& byte_offset);
+                     uint32_t& byte_offset);
 bool TileGetBlockXor(const TileBlockLayout& layout, uint32_t block_x, uint32_t block_y,
-	                 uint32_t block_z, uint32_t& byte_offset);
+                     uint32_t block_z, uint32_t& byte_offset);
 bool TileIsStandard256BTextureSupported(uint32_t format);
 bool TileIsStandard4KBTextureSupported(uint32_t format);
 bool TileIsStandard64KBTextureSupported(uint32_t format);
 bool TileGetTextureVolumeLayout(uint32_t format, uint32_t width, uint32_t height, uint32_t depth,
-	                            uint32_t levels, uint32_t tile, TileVolumeLayout& layout);
+                                uint32_t levels, uint32_t tile, TileVolumeLayout& layout);
 
 bool     TileGetHtileSize(uint32_t width, uint32_t height, TileSizeAlign& htile_size);
 bool     TileGetDepthSize(uint32_t width, uint32_t height, uint32_t pitch, uint32_t z_format,
-	                      uint32_t stencil_format, bool htile, TileSizeAlign& stencil_size,
-	                      TileSizeAlign& htile_size, TileSizeAlign& depth_size,
+                          uint32_t stencil_format, bool htile, TileSizeAlign& stencil_size,
+                          TileSizeAlign& htile_size, TileSizeAlign& depth_size,
                           uint32_t num_fragments_log2 = 0);
 uint32_t TileGetRenderTargetPitch(uint32_t width, uint32_t bytes_per_element,
                                   uint32_t num_fragments_log2 = 0);
 uint32_t TileGetDepthPitch(uint32_t width, uint32_t bytes_per_element,
                            uint32_t num_fragments_log2 = 0);
 bool     TileGetRenderTargetSize(uint32_t width, uint32_t height, uint32_t pitch,
-	                             uint32_t bytes_per_element, TileSizeAlign& total_size,
+                                 uint32_t bytes_per_element, TileSizeAlign& total_size,
                                  uint32_t num_fragments_log2 = 0);
 bool     TileGetRenderTargetMipLayout(uint32_t width, uint32_t height, uint32_t pitch,
-	                                  uint32_t bytes_per_element, uint32_t levels,
-	                                  TileSizeAlign& total_size, TileSizeOffset* level_sizes,
+                                      uint32_t bytes_per_element, uint32_t levels,
+                                      TileSizeAlign& total_size, TileSizeOffset* level_sizes,
                                       TilePaddedSize* padded_size);
 void     TileGetTextureSize(uint32_t format, uint32_t width, uint32_t height, uint32_t pitch,
                             uint32_t levels, uint32_t tile, TileSizeAlign* total_size,
                             TileSizeOffset* level_sizes, TilePaddedSize* padded_size);
 void TileGetTextureTotalSize(uint32_t format, uint32_t width, uint32_t height, uint32_t depth,
                              uint32_t pitch, uint32_t levels, uint32_t tile, bool volume_texture,
-	                         TileSizeAlign& total_size);
+                             TileSizeAlign& total_size);
 uint32_t TileGetTexturePitch(uint32_t format, uint32_t width, uint32_t levels, uint32_t tile);
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/graphicContext.h

@@ -60,19 +60,19 @@ struct VulkanImage {
 	VulkanImage() = default;
 	KYTY_CLASS_NO_COPY(VulkanImage);
 
-	vk::Format             format      = vk::Format::eUndefined;
-	vk::ImageType          image_type  = vk::ImageType::e2D;
-	vk::Extent3D           extent      = {1, 1, 1};
-	uint32_t               guest_pitch = 0;
-	uint32_t               layers      = 1;
-	uint32_t               mip_levels  = 1;
-	uint32_t               samples     = 1;
-	vk::ImageUsageFlags    usage       = {};
-	vk::ImageCreateFlags   flags       = {};
-	vk::Image              image       = nullptr;
-	VulkanImageState       state;
+	vk::Format                    format      = vk::Format::eUndefined;
+	vk::ImageType                 image_type  = vk::ImageType::e2D;
+	vk::Extent3D                  extent      = {1, 1, 1};
+	uint32_t                      guest_pitch = 0;
+	uint32_t                      layers      = 1;
+	uint32_t                      mip_levels  = 1;
+	uint32_t                      samples     = 1;
+	vk::ImageUsageFlags           usage       = {};
+	vk::ImageCreateFlags          flags       = {};
+	vk::Image                     image       = nullptr;
+	VulkanImageState              state;
 	std::vector<VulkanImageState> subresource_states;
-	Graphics::VulkanMemory memory;
+	Graphics::VulkanMemory        memory;
 };
 
 struct VulkanBuffer {

src/graphics/host_gpu/hostMemory.cpp

@@ -30,7 +30,7 @@ bool IsAccessible(DWORD protect, HostMemoryAccess access) {
 } // namespace
 
 bool HostMemoryQueryRange(uint64_t addr, uint64_t requested_size, HostMemoryAccess access,
-	                      uint64_t& accessible_size) {
+                          uint64_t& accessible_size) {
 	accessible_size = 0;
 	if (addr == 0 || requested_size == 0) {
 		return false;

src/graphics/host_gpu/hostMemory.h

@@ -8,7 +8,7 @@ namespace Libs::Graphics {
 enum class HostMemoryAccess { Read, Mapped };
 
 bool HostMemoryQueryRange(uint64_t addr, uint64_t requested_size, HostMemoryAccess access,
-	                      uint64_t& accessible_size);
+                          uint64_t& accessible_size);
 bool HostMemoryQueryReadable(uint64_t addr, uint64_t requested_size, uint64_t& readable_size);
 bool HostMemoryIsReadable(uint64_t addr);
 bool HostMemoryRangeIsReadable(uint64_t addr, uint64_t size);

src/graphics/host_gpu/memoryTracker.cpp

@@ -4,25 +4,9 @@
 
 namespace Libs::Graphics {
 
-#if defined(KYTY_MEMORY_TRACKER_TESTS)
-namespace {
-std::atomic<MemoryTracker::UnmapContentionHook> g_unmap_contention_hook {nullptr};
-}
-
-void MemoryTracker::SetUnmapContentionHook(UnmapContentionHook hook) noexcept {
-	g_unmap_contention_hook.store(hook, std::memory_order_release);
-}
-#endif
-
 static_assert(std::atomic<void*>::is_always_lock_free);
 
-MemoryTracker::MemoryTracker(PageManager& page_manager, PageWatchMode gpu_watch_mode)
-    : m_page_manager(page_manager), m_gpu_watch_mode(gpu_watch_mode) {
-	switch (m_gpu_watch_mode) {
-		case PageWatchMode::Write:
-		case PageWatchMode::ReadWrite: break;
-		default: EXIT("unsupported memory tracker GPU page-watch mode\n");
-	}
+MemoryTracker::MemoryTracker(PageManager& page_manager): m_page_manager(page_manager) {
 	m_regions = std::make_unique<std::atomic<RegionManager*>[]>(REGION_COUNT);
 	for (size_t i = 0; i < REGION_COUNT; i++) {
 		m_regions[i].store(nullptr, std::memory_order_relaxed);
@@ -31,6 +15,7 @@ MemoryTracker::MemoryTracker(PageManager& page_manager, PageWatchMode gpu_watch_
 
 MemoryTracker::~MemoryTracker() = default;
 
+#if KYTY_BUILD == KYTY_BUILD_DEBUG
 void MemoryTracker::ValidateGpuDirtyPages(const RangeSet& dirty, uint64_t vaddr, uint64_t size,
                                           const char* operation) const noexcept {
 	if (vaddr == 0 || size == 0 || size > UINT64_MAX - vaddr ||
@@ -68,6 +53,7 @@ void MemoryTracker::ValidateGpuDirtyOwnership(const RangeSet& dirty, uint64_t va
 		}
 	}
 }
+#endif
 
 void MemoryTracker::ValidateRange(uint64_t vaddr, uint64_t size) {
 	if (vaddr == 0 || size == 0 || vaddr >= TRACKER_ADDRESS_SIZE ||
@@ -94,7 +80,6 @@ RegionManager* MemoryTracker::GetOrCreateRegion(uint64_t index) {
 bool MemoryTracker::IsRegionCpuModified(uint64_t vaddr, uint64_t size) {
 	CheckNotInUploadCallback();
 	std::lock_guard access(m_access_mutex);
-	RequireMapped(vaddr, size);
 	return Iterate<true>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 		std::scoped_lock lock(manager->lock);
 		return manager->IsModified<DirtySource::Cpu>(offset, bytes);
@@ -104,7 +89,6 @@ bool MemoryTracker::IsRegionCpuModified(uint64_t vaddr, uint64_t size) {
 bool MemoryTracker::IsRegionGpuModified(uint64_t vaddr, uint64_t size) {
 	CheckNotInUploadCallback();
 	std::lock_guard access(m_access_mutex);
-	RequireMapped(vaddr, size);
 	return Iterate<false>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 		std::scoped_lock lock(manager->lock);
 		return manager->IsModified<DirtySource::Gpu>(offset, bytes);
@@ -114,45 +98,31 @@ bool MemoryTracker::IsRegionGpuModified(uint64_t vaddr, uint64_t size) {
 void MemoryTracker::MarkRegionAsCpuModified(uint64_t vaddr, uint64_t size) {
 	CheckNotInUploadCallback();
 	std::lock_guard access(m_access_mutex);
-	RequireMapped(vaddr, size);
 	Iterate<true>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 		std::scoped_lock lock(manager->lock);
-		const auto       changed =
-		    manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
-		manager->ApplyProtection(changed, false);
+		manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
 }
 
 void MemoryTracker::MarkRegionAsGpuModified(uint64_t vaddr, uint64_t size) {
 	CheckNotInUploadCallback();
 	std::lock_guard access(m_access_mutex);
-	RequireMapped(vaddr, size);
-	Iterate<true>(vaddr, size, [this](RegionManager* manager, uint64_t offset, uint64_t bytes) {
+	Iterate<true>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 		std::scoped_lock lock(manager->lock);
-		const auto       changed =
-		    manager->ChangeState<DirtySource::Gpu, true>(manager->GetCpuAddr() + offset, bytes);
-		manager->ApplyGpuProtection(changed, true, m_gpu_watch_mode);
+		manager->ChangeState<DirtySource::Gpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
 }
 
 void MemoryTracker::UnmarkRegionAsGpuModified(uint64_t vaddr, uint64_t size) {
 	CheckNotInUploadCallback();
 	std::lock_guard access(m_access_mutex);
-	RequireMapped(vaddr, size);
-	Iterate<true>(vaddr, size, [this](RegionManager* manager, uint64_t offset, uint64_t bytes) {
+	Iterate<false>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 		std::scoped_lock lock(manager->lock);
-		if (!manager->IsFullyModified<DirtySource::Gpu>(offset, bytes)) {
-			EXIT("cannot clear partially GPU-dirty tracking range\n");
-		}
-		const auto changed =
-		    manager->ChangeState<DirtySource::Gpu, false>(manager->GetCpuAddr() + offset, bytes);
-		manager->ApplyGpuProtection(changed, false, m_gpu_watch_mode);
+		manager->ChangeState<DirtySource::Gpu, false>(manager->GetCpuAddr() + offset, bytes);
 	});
 }
 
 void MemoryTracker::UntrackMemoryLocked(uint64_t vaddr, uint64_t size) {
-	RequireMapped(vaddr, size);
-
 	std::vector<RegionManager*> managers;
 	managers.reserve((vaddr % TRACKER_REGION_SIZE + size + TRACKER_REGION_SIZE - 1) /
 	                 TRACKER_REGION_SIZE);
@@ -171,10 +141,7 @@ void MemoryTracker::UntrackMemoryLocked(uint64_t vaddr, uint64_t size) {
 		EXIT("cannot untrack GPU-dirty memory\n");
 	}
 	Iterate<false>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
-		const auto changed =
-		    manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
-		manager->ApplyProtection(changed, false);
-		manager->Untrack(manager->GetCpuAddr() + offset, bytes);
+		manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
 	locks.clear();
 }
@@ -185,109 +152,4 @@ void MemoryTracker::UntrackMemory(uint64_t vaddr, uint64_t size) {
 	UntrackMemoryLocked(vaddr, size);
 }
 
-void MemoryTracker::UnmapMemory(uint64_t vaddr, uint64_t size) {
-	CheckNotInUploadCallback();
-	std::unique_lock access(m_access_mutex, std::try_to_lock);
-	if (!access.owns_lock()) {
-#if defined(KYTY_MEMORY_TRACKER_TESTS)
-		if (const auto hook = g_unmap_contention_hook.load(std::memory_order_acquire);
-		    hook != nullptr) {
-			hook();
-		}
-#endif
-		access.lock();
-	}
-	UntrackMemoryLocked(vaddr, size);
-	m_page_manager.OnGpuUnmap(vaddr, size);
-}
-
-bool MemoryTracker::InvalidateRegion(uint64_t vaddr, uint64_t size, PageFaultPhase phase) noexcept {
-	switch (phase) {
-		case PageFaultPhase::Release: return true;
-		case PageFaultPhase::Invalidate: {
-			const auto action = BeginCpuFault(vaddr, size);
-			switch (action) {
-				case CpuFaultAction::Untracked: return false;
-				case CpuFaultAction::Continue: return true;
-				case CpuFaultAction::Download:
-					EXIT("generic region invalidation cannot download GPU-dirty memory\n");
-			}
-		}
-		case PageFaultPhase::Complete:
-			return CompleteCpuFault(vaddr, size, PageFaultAccess::Write, false);
-	}
-	EXIT("unsupported region invalidation phase\n");
-}
-
-bool MemoryTracker::InvalidateVirtualGpuWrite(PageFaultAccess access, uint64_t vaddr, uint64_t size,
-                                              PageFaultPhase phase) noexcept {
-	switch (phase) {
-		case PageFaultPhase::Release: return true;
-		case PageFaultPhase::Invalidate: {
-			const bool gpu_modified = Iterate<false>(
-			    vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
-				    std::scoped_lock lock(manager->lock);
-				    return manager->IsModified<DirtySource::Gpu>(offset, bytes);
-			    });
-			if (!gpu_modified) {
-				return false;
-			}
-			const auto action = BeginCpuFault(vaddr, size);
-			if (access != PageFaultAccess::Write || action != CpuFaultAction::Download) {
-				EXIT("virtual GPU write fault requires write access to GPU-dirty memory\n");
-			}
-			return true;
-		}
-		case PageFaultPhase::Complete: {
-			if (access != PageFaultAccess::Write) {
-				EXIT("virtual GPU write completion requires write access\n");
-			}
-			bool completed = false;
-			Iterate<false>(
-			    vaddr, size, [&completed](RegionManager* manager, uint64_t offset, uint64_t bytes) {
-				    std::scoped_lock lock(manager->lock);
-				    if (completed) {
-					    EXIT("virtual GPU write fault spans multiple tracked regions\n");
-				    }
-				    completed =
-				        manager->CompleteVirtualGpuWrite(manager->GetCpuAddr() + offset, bytes);
-			    });
-			return completed;
-		}
-	}
-	EXIT("unsupported virtual GPU write invalidation phase\n");
-}
-
-CpuFaultAction MemoryTracker::BeginCpuFault(uint64_t vaddr, uint64_t size,
-                                            PageFaultAccess access) noexcept {
-	CheckNotInUploadCallback();
-	CpuFaultAction action = CpuFaultAction::Untracked;
-	Iterate<false>(
-	    vaddr, size, [&action, access](RegionManager* manager, uint64_t offset, uint64_t bytes) {
-		    std::scoped_lock lock(manager->lock);
-		    if (action != CpuFaultAction::Untracked) {
-			    EXIT("CPU fault spans multiple tracked regions\n");
-		    }
-		    action = manager->BeginCpuFault(manager->GetCpuAddr() + offset, bytes, access);
-	    });
-	return action;
-}
-
-bool MemoryTracker::CompleteCpuFault(uint64_t vaddr, uint64_t size, PageFaultAccess access,
-                                     bool downloaded) noexcept {
-	CheckNotInUploadCallback();
-	bool found = false;
-	Iterate<false>(
-	    vaddr, size,
-	    [&found, access, downloaded](RegionManager* manager, uint64_t offset, uint64_t bytes) {
-		    std::scoped_lock lock(manager->lock);
-		    if (found) {
-			    EXIT("CPU fault completion spans multiple tracked regions\n");
-		    }
-		    found = manager->CompleteCpuFault(manager->GetCpuAddr() + offset, bytes, access,
-		                                      downloaded);
-	    });
-	return found;
-}
-
 } // namespace Libs::Graphics

src/graphics/host_gpu/memoryTracker.h

@@ -18,8 +18,7 @@ namespace Libs::Graphics {
 
 class MemoryTracker final {
 public:
-	explicit MemoryTracker(PageManager&  page_manager,
-	                       PageWatchMode gpu_watch_mode = PageWatchMode::ReadWrite);
+	explicit MemoryTracker(PageManager& page_manager);
 	~MemoryTracker();
 
 	KYTY_CLASS_NO_COPY(MemoryTracker);
@@ -30,28 +29,62 @@ public:
 	void               MarkRegionAsGpuModified(uint64_t vaddr, uint64_t size);
 	void               UnmarkRegionAsGpuModified(uint64_t vaddr, uint64_t size);
 	void               UntrackMemory(uint64_t vaddr, uint64_t size);
-	void               UnmapMemory(uint64_t vaddr, uint64_t size);
-	[[nodiscard]] CpuFaultAction
-	                   BeginCpuFault(uint64_t vaddr, uint64_t size,
-	                                 PageFaultAccess access = PageFaultAccess::Write) noexcept;
-	[[nodiscard]] bool CompleteCpuFault(uint64_t vaddr, uint64_t size, PageFaultAccess access,
-	                                    bool downloaded) noexcept;
-	[[nodiscard]] bool InvalidateRegion(uint64_t vaddr, uint64_t size,
-	                                    PageFaultPhase phase) noexcept;
-	[[nodiscard]] bool InvalidateVirtualGpuWrite(PageFaultAccess access, uint64_t vaddr,
-	                                             uint64_t size, PageFaultPhase phase) noexcept;
+	template <typename Flush>
+	void InvalidateRegion(uint64_t vaddr, uint64_t size, Flush&& on_flush) {
+		static_assert(std::is_invocable_v<Flush&>);
+		CheckNotInUploadCallback();
+		ValidateRange(vaddr, size);
+
+		const auto update_cpu_state = [this, vaddr, size] {
+			std::lock_guard             access(m_access_mutex);
+			std::vector<RegionManager*> managers;
+			Iterate<false>(vaddr, size, [&](RegionManager* manager, uint64_t, uint64_t) {
+				managers.push_back(manager);
+			});
+			std::vector<std::unique_lock<TrackingSpinLock>> locks;
+			locks.reserve(managers.size());
+			for (auto* manager: managers) {
+				locks.emplace_back(manager->lock);
+			}
+			const bool gpu_modified = Iterate<false>(
+			    vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
+				    return manager->IsModified<DirtySource::Gpu>(offset, bytes);
+			    });
+			if (gpu_modified) {
+				return true;
+			}
+			Iterate<false>(vaddr, size,
+			               [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
+				               manager->ChangeState<DirtySource::Cpu, true>(
+				                   manager->GetCpuAddr() + offset, bytes);
+			               });
+			return false;
+		};
+
+		if (!update_cpu_state()) {
+			return;
+		}
+		std::forward<Flush>(on_flush)();
+		if (update_cpu_state()) {
+			EXIT("memory invalidation retained GPU-owned pages\n");
+		}
+	}
+#if KYTY_BUILD == KYTY_BUILD_DEBUG
 	void ValidateGpuDirtyPages(const RangeSet& dirty, uint64_t vaddr, uint64_t size,
 	                           const char* operation) const noexcept;
 	void ValidateGpuDirtyOwnership(const RangeSet& dirty, uint64_t vaddr, uint64_t size,
 	                               const char* operation);
+#else
+	void ValidateGpuDirtyPages(const RangeSet&, uint64_t, uint64_t, const char*) const noexcept {}
+	void ValidateGpuDirtyOwnership(const RangeSet&, uint64_t, uint64_t, const char*) {}
+#endif
 
 	template <bool clear, typename Preflight, typename Func>
 	void ForEachDownloadRange(uint64_t vaddr, uint64_t size, Preflight&& preflight, Func&& func) {
 		static_assert(std::is_nothrow_invocable_v<Preflight&, uint64_t, uint64_t>);
 		static_assert(std::is_nothrow_invocable_v<Func&, uint64_t, uint64_t>);
 		CheckNotInUploadCallback();
-		std::lock_guard access(m_access_mutex);
-		RequireMapped(vaddr, size);
+		std::lock_guard             access(m_access_mutex);
 		std::vector<RegionManager*> managers;
 		Iterate<false>(vaddr, size, [&](RegionManager* manager, uint64_t, uint64_t) {
 			managers.push_back(manager);
@@ -63,9 +96,6 @@ public:
 		}
 		Iterate<false>(vaddr, size, [&](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 			const auto address = manager->GetCpuAddr() + offset;
-			if (manager->HasPendingFault(address, bytes)) {
-				EXIT("GPU download synchronization raced a pending CPU fault\n");
-			}
 			manager->template ForEachModifiedRange<DirtySource::Gpu, false>(address, bytes,
 			                                                                preflight);
 		});
@@ -77,10 +107,8 @@ public:
 			Iterate<false>(vaddr, size,
 			               [&](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 				               const auto address = manager->GetCpuAddr() + offset;
-				               const auto changed =
-				                   manager->template ForEachModifiedRange<DirtySource::Gpu, true>(
-				                       address, bytes, [](uint64_t, uint64_t) noexcept {});
-				               manager->ApplyGpuProtection(changed, false, m_gpu_watch_mode);
+				               manager->template ForEachModifiedRange<DirtySource::Gpu, true>(
+				                   address, bytes, [](uint64_t, uint64_t) noexcept {});
 			               });
 		}
 	}
@@ -91,11 +119,6 @@ public:
 		    vaddr, size, [](uint64_t, uint64_t) noexcept {}, std::forward<Func>(func));
 	}
 
-#if defined(KYTY_MEMORY_TRACKER_TESTS)
-	using UnmapContentionHook = void (*)() noexcept;
-	static void SetUnmapContentionHook(UnmapContentionHook hook) noexcept;
-#endif
-
 	template <typename RangeFunc, typename UploadFunc>
 	void ForEachUploadRange(uint64_t vaddr, uint64_t size, bool is_written, RangeFunc&& range_func,
 	                        UploadFunc&& upload_func) {
@@ -103,12 +126,10 @@ public:
 		static_assert(std::is_nothrow_invocable_v<UploadFunc&>);
 		CheckNotInUploadCallback();
 		std::unique_lock access(m_access_mutex);
-		RequireMapped(vaddr, size);
 		Iterate<true>(vaddr, size, [](RegionManager*, uint64_t, uint64_t) {});
 		const auto* previous_upload_owner = std::exchange(s_upload_owner, this);
 		Iterate<false>(vaddr, size, [&](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 			manager->lock.lock();
-			manager->Track(manager->GetCpuAddr() + offset, bytes);
 			manager->ForEachModifiedRange<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset,
 			                                                      bytes, range_func);
 			if (!is_written) {
@@ -117,13 +138,12 @@ public:
 		});
 		upload_func();
 		if (is_written) {
-			Iterate<false>(
-			    vaddr, size, [this](RegionManager* manager, uint64_t offset, uint64_t bytes) {
-				    const auto changed = manager->template ChangeState<DirtySource::Gpu, true>(
-				        manager->GetCpuAddr() + offset, bytes);
-				    manager->ApplyGpuProtection(changed, true, m_gpu_watch_mode);
-				    manager->lock.unlock();
-			    });
+			Iterate<false>(vaddr, size,
+			               [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
+				               manager->template ChangeState<DirtySource::Gpu, true>(
+				                   manager->GetCpuAddr() + offset, bytes);
+				               manager->lock.unlock();
+			               });
 		}
 		s_upload_owner = previous_upload_owner;
 	}
@@ -168,16 +188,8 @@ private:
 		return false;
 	}
 
-	static void ValidateRange(uint64_t vaddr, uint64_t size);
-	void        UntrackMemoryLocked(uint64_t vaddr, uint64_t size);
-	void        RequireMapped(uint64_t vaddr, uint64_t size) const {
-		ValidateRange(vaddr, size);
-		if (!m_page_manager.IsMapped(vaddr, size)) {
-			EXIT("memory tracker range [0x%llx, 0x%llx) is not mapped\n",
-			     static_cast<unsigned long long>(vaddr),
-			     static_cast<unsigned long long>(vaddr + size));
-		}
-	}
+	static void    ValidateRange(uint64_t vaddr, uint64_t size);
+	void           UntrackMemoryLocked(uint64_t vaddr, uint64_t size);
 	RegionManager* GetOrCreateRegion(uint64_t index);
 
 	std::unique_ptr<std::atomic<RegionManager*>[]> m_regions;
@@ -185,7 +197,6 @@ private:
 	std::mutex                                     m_region_mutex;
 	std::mutex                                     m_access_mutex;
 	PageManager&                                   m_page_manager;
-	PageWatchMode                                  m_gpu_watch_mode = PageWatchMode::ReadWrite;
 };
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/pageManager.cpp

@@ -1,13 +1,14 @@
 #include "graphics/host_gpu/pageManager.h"
 
 #include "graphics/host_gpu/regionDefinitions.h"
+#include "kernel/memory.h"
 
+#include <algorithm>
 #include <array>
 #include <atomic>
 #include <cstdarg>
 #include <cstdio>
 #include <cstdlib>
-#include <limits>
 #include <memory>
 #include <mutex>
 #include <vector>
@@ -19,6 +20,11 @@
 #include <windows.h>
 #undef min
 #undef max
+#elif defined(__APPLE__)
+#include <unistd.h>
+#else
+#include <execinfo.h>
+#include <unistd.h>
 #endif
 
 namespace Libs::Graphics {
@@ -28,19 +34,19 @@ constexpr uint64_t PAGE_SIZE    = TRACKER_PAGE_SIZE;
 constexpr uint64_t REGION_SIZE  = TRACKER_REGION_SIZE;
 constexpr uint64_t ADDRESS_SIZE = TRACKER_ADDRESS_SIZE;
 constexpr uint64_t REGION_COUNT = ADDRESS_SIZE / REGION_SIZE;
+
+#if KYTY_PLATFORM != KYTY_PLATFORM_WINDOWS
+// The tracker reuses Win32 memory-protection tags as internal page-state values.
+// Mirror their canonical numeric values so the shared state-machine logic is identical.
+constexpr uint32_t PAGE_NOACCESS  = 0x01;
+constexpr uint32_t PAGE_READONLY  = 0x02;
+constexpr uint32_t PAGE_READWRITE = 0x04;
+#endif
 constexpr uint64_t REGION_PAGES = REGION_SIZE / PAGE_SIZE;
 
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-constexpr uint32_t NO_ACCESS_PROTECTION = PAGE_NOACCESS;
-constexpr uint32_t READ_ONLY_PROTECTION = PAGE_READONLY;
+constexpr uint32_t NO_ACCESS_PROTECTION  = PAGE_NOACCESS;
+constexpr uint32_t READ_ONLY_PROTECTION  = PAGE_READONLY;
 constexpr uint32_t READ_WRITE_PROTECTION = PAGE_READWRITE;
-#else
-constexpr uint32_t NO_ACCESS_PROTECTION = 0;
-constexpr uint32_t READ_ONLY_PROTECTION = 1;
-constexpr uint32_t READ_WRITE_PROTECTION = 2;
-#endif
-
-thread_local bool g_in_fault_resolution = false;
 
 [[noreturn]] void FailFast(const char* reason = nullptr) noexcept {
 	std::fputs("PageManager fail-fast: ", stderr);
@@ -56,6 +62,10 @@ thread_local bool g_in_fault_resolution = false;
 		std::fprintf(stderr, "  frame[%u]=0x%016" PRIxPTR " image_rva=0x%016" PRIxPTR "\n", i,
 		             address, address >= image_base ? address - image_base : 0);
 	}
+#elif !defined(__APPLE__)
+	void*     frames[16] {};
+	const int frame_count = ::backtrace(frames, static_cast<int>(std::size(frames)));
+	::backtrace_symbols_fd(frames, frame_count, STDERR_FILENO);
 #endif
 	std::fflush(stderr);
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
@@ -75,12 +85,13 @@ thread_local bool g_in_fault_resolution = false;
 	std::_Exit(322);
 }
 
-uint32_t CurrentThread() noexcept {
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-	return GetCurrentThreadId();
-#else
-	FailFast();
-#endif
+Common::VirtualMemory::Mode ToMemoryMode(uint32_t protection) {
+	switch (protection) {
+		case NO_ACCESS_PROTECTION: return Common::VirtualMemory::Mode::NoAccess;
+		case READ_ONLY_PROTECTION: return Common::VirtualMemory::Mode::Read;
+		case READ_WRITE_PROTECTION: return Common::VirtualMemory::Mode::ReadWrite;
+		default: Fatal("unmappable protection 0x%08" PRIx32, protection);
+	}
 }
 
 class SpinGuard final {
@@ -116,28 +127,61 @@ uint64_t PageEnd(uint64_t vaddr, uint64_t size) {
 
 struct PageManager::Impl {
 	struct PageState {
-		std::atomic_flag lock                 = ATOMIC_FLAG_INIT;
-		uint32_t         mappings             = 0;
-		uint32_t         gpu_read_mappings    = 0;
-		uint32_t         gpu_write_mappings   = 0;
-		uint32_t         write_watchers       = 0;
-		uint32_t         access_watchers      = 0;
-		uint32_t         original_protection  = 0;
-		uint32_t         backing_writer       = 0;
-		bool             resolving            = false;
-		bool             resolving_read_write = false;
-		bool             late_read_pending    = false;
-		bool             late_write_pending   = false;
+		uint8_t write_watchers  : 7 = 0;
+		uint8_t access_watchers : 1 = 0;
+
+		[[nodiscard]] uint32_t Perms() const noexcept {
+			if (access_watchers != 0) {
+				return NO_ACCESS_PROTECTION;
+			}
+			if (write_watchers != 0) {
+				return READ_ONLY_PROTECTION;
+			}
+			return READ_WRITE_PROTECTION;
+		}
+
+		template <int delta, bool is_read>
+		uint32_t AddDelta(uint64_t address) {
+			static_assert(delta >= -1 && delta <= 1);
+			if constexpr (is_read) {
+				if constexpr (delta == 1) {
+					if (access_watchers != 0) {
+						Fatal("read-watcher overflow at 0x%016" PRIx64, address);
+					}
+					return ++access_watchers;
+				} else if constexpr (delta == -1) {
+					if (access_watchers == 0) {
+						Fatal("read-watcher underflow at 0x%016" PRIx64, address);
+					}
+					return --access_watchers;
+				} else {
+					return access_watchers;
+				}
+			} else {
+				if constexpr (delta == 1) {
+					if (write_watchers == 0x7f) {
+						Fatal("write-watcher overflow at 0x%016" PRIx64, address);
+					}
+					return ++write_watchers;
+				} else if constexpr (delta == -1) {
+					if (write_watchers == 0) {
+						Fatal("write-watcher underflow at 0x%016" PRIx64, address);
+					}
+					return --write_watchers;
+				} else {
+					return write_watchers;
+				}
+			}
+		}
 	};
+	static_assert(sizeof(PageState) == 1);
 
 	struct Region {
+		std::atomic_flag                    lock = ATOMIC_FLAG_INIT;
 		std::array<PageState, REGION_PAGES> pages;
 	};
 
-	Impl(PageFaultHandler handler, void* context): fault_handler(handler), fault_context(context) {
-		if (fault_handler == nullptr) {
-			Fatal("null fault handler");
-		}
+	Impl() {
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 		SYSTEM_INFO info {};
 		GetSystemInfo(&info);
@@ -145,8 +189,16 @@ struct PageManager::Impl {
 			Fatal("unsupported host page size 0x%08" PRIx32,
 			      static_cast<uint32_t>(info.dwPageSize));
 		}
+#elif defined(__APPLE__)
+		// Under Rosetta the host page size is 4 KB, matching TRACKER_PAGE_SIZE.
+		if (static_cast<uint64_t>(getpagesize()) != PAGE_SIZE) {
+			Fatal("unsupported host page size 0x%08" PRIx32, static_cast<uint32_t>(getpagesize()));
+		}
 #else
-		Fatal("page-fault invalidation is not implemented on this platform");
+		const auto host_page_size = ::sysconf(_SC_PAGESIZE);
+		if (host_page_size < 0 || static_cast<uint64_t>(host_page_size) != PAGE_SIZE) {
+			Fatal("unsupported host page size %ld", static_cast<long>(host_page_size));
+		}
 #endif
 		regions = std::make_unique<std::atomic<Region*>[]>(REGION_COUNT);
 		for (uint64_t i = 0; i < REGION_COUNT; i++) {
@@ -156,11 +208,9 @@ struct PageManager::Impl {
 
 	~Impl() {
 		for (const auto& region: region_storage) {
+			SpinGuard lock(region->lock);
 			for (auto& page: region->pages) {
-				SpinGuard lock(page.lock);
-				if (page.mappings != 0 || page.gpu_read_mappings != 0 ||
-				    page.gpu_write_mappings != 0 || page.write_watchers != 0 ||
-				    page.access_watchers != 0 || page.backing_writer != 0 || page.resolving) {
+				if (page.write_watchers != 0 || page.access_watchers != 0) {
 					FailFast("PageManager destroyed with live page state");
 				}
 			}
@@ -188,536 +238,140 @@ struct PageManager::Impl {
 		return ptr;
 	}
 
-	PageState& GetPage(Region& region, uint64_t vaddr) const {
-		return region.pages[(vaddr % REGION_SIZE) / PAGE_SIZE];
-	}
-
-	static uint32_t WatcherProtection(const PageState& page) {
-		if (page.access_watchers != 0) {
-			return NO_ACCESS_PROTECTION;
-		}
-		if (page.write_watchers != 0) {
-			return READ_ONLY_PROTECTION;
-		}
-		return page.original_protection;
-	}
-
-	static void PublishDelayedFaults(PageState& page, uint32_t old_protection,
-	                                 uint32_t new_protection) {
-		if (old_protection == NO_ACCESS_PROTECTION && new_protection != NO_ACCESS_PROTECTION) {
-			page.late_read_pending = true;
-		}
-		if ((old_protection == NO_ACCESS_PROTECTION ||
-		     old_protection == READ_ONLY_PROTECTION) &&
-		    new_protection == READ_WRITE_PROTECTION) {
-			page.late_write_pending = true;
+	void Protect(uint64_t vaddr, uint64_t size, uint32_t protection) noexcept {
+		if (!Libs::LibKernel::Memory::ProtectGuestHostMemory(vaddr, size,
+		                                                     ToMemoryMode(protection))) {
+			Fatal("address-space protection failed at 0x%016" PRIx64 ", new=0x%08" PRIx32, vaddr,
+			      protection);
 		}
 	}
 
-	static uint32_t QueryProtection(uint64_t vaddr) {
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-		MEMORY_BASIC_INFORMATION info {};
-		if (VirtualQuery(reinterpret_cast<const void*>(static_cast<uintptr_t>(vaddr)), &info,
-		                 sizeof(info)) == 0 ||
-		    info.State != MEM_COMMIT || info.Protect != PAGE_READWRITE) {
-			Fatal("basic path requires PAGE_READWRITE at 0x%016" PRIx64 " (state=0x%08" PRIx32
-			      ", protection=0x%08" PRIx32 ")",
-			      vaddr, static_cast<uint32_t>(info.State), static_cast<uint32_t>(info.Protect));
-		}
-		return info.Protect;
-#else
-		(void)vaddr;
-		Fatal("page query is unsupported on this platform");
-#endif
-	}
+	template <bool track, bool is_read, bool masked>
+	void UpdateRegionWatchers(Region& region, uint64_t base_addr, size_t first, size_t last,
+	                          const RegionBits* mask = nullptr) {
+		SpinGuard lock(region.lock);
+		auto      perms                 = region.pages[first].Perms();
+		uint64_t  range_begin           = 0;
+		uint64_t  range_bytes           = 0;
+		uint64_t  potential_range_bytes = 0;
 
-	static bool AllowsAccess(uint64_t vaddr, PageFaultAccess access) noexcept {
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-		MEMORY_BASIC_INFORMATION info {};
-		if (VirtualQuery(reinterpret_cast<const void*>(static_cast<uintptr_t>(vaddr)), &info,
-		                 sizeof(info)) == 0 ||
-		    info.State != MEM_COMMIT) {
-			return false;
-		}
-		switch (access) {
-			case PageFaultAccess::Read:
-				return info.Protect == PAGE_READONLY || info.Protect == PAGE_READWRITE;
-			case PageFaultAccess::Write: return info.Protect == PAGE_READWRITE;
-			default: return false;
-		}
-#else
-		(void)vaddr;
-		return false;
-#endif
-	}
+		const auto release_pending = [&] {
+			if (range_bytes != 0) {
+				Protect(base_addr + range_begin * PAGE_SIZE, range_bytes, perms);
+				range_bytes           = 0;
+				potential_range_bytes = 0;
+			}
+		};
 
-	static void Protect(uint64_t vaddr, uint32_t protection, uint32_t expected_old,
-	                    bool fault_path) noexcept {
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-		DWORD old_protection = 0;
-		if (VirtualProtect(reinterpret_cast<void*>(static_cast<uintptr_t>(vaddr)), PAGE_SIZE,
-		                   protection, &old_protection) == 0 ||
-		    old_protection != expected_old) {
-			if (fault_path) {
-				FailFast("VirtualProtect fault transition did not match expected protection");
+		for (size_t page_index = first; page_index < last; page_index++) {
+			auto&      page    = region.pages[page_index];
+			const auto address = base_addr + page_index * PAGE_SIZE;
+			const bool update  = !masked || mask->Get(page_index);
+
+			const auto old_perms = page.Perms();
+			const auto new_count = update ? page.AddDelta<track ? 1 : -1, is_read>(address)
+			                              : page.AddDelta<0, is_read>(address);
+			const auto new_perms = page.Perms();
+
+			if (new_perms != perms) [[unlikely]] {
+				release_pending();
+				perms = new_perms;
+			} else if (range_bytes != 0) {
+				potential_range_bytes += PAGE_SIZE;
+			}
+
+			if (!update) {
+				continue;
+			}
+
+			const bool watcher_edge = (track && new_count == 1) || (!track && new_count == 0);
+			if (watcher_edge && old_perms != new_perms) {
+				if (range_bytes == 0) {
+					range_begin           = page_index;
+					potential_range_bytes = PAGE_SIZE;
+				}
+				range_bytes = potential_range_bytes;
 			}
-			Fatal("invalid protection transition at 0x%016" PRIx64 ", old=0x%08" PRIx32
-			      ", expected=0x%08" PRIx32 ", new=0x%08" PRIx32,
-			      vaddr, static_cast<uint32_t>(old_protection), expected_old, protection);
 		}
-#else
-		(void)vaddr;
-		(void)protection;
-		(void)fault_path;
-		FailFast("page protection is unsupported on this platform");
-#endif
+
+		release_pending();
+	}
+
+	template <bool track, bool is_read>
+	void UpdatePageWatchers(uint64_t vaddr, uint64_t size) {
+		const auto begin = PageStart(vaddr);
+		const auto end   = PageEnd(vaddr, size);
+		for (auto chunk_begin = begin; chunk_begin < end;) {
+			const auto chunk_end   = std::min(end, (chunk_begin / REGION_SIZE + 1) * REGION_SIZE);
+			const auto region_base = chunk_begin / REGION_SIZE * REGION_SIZE;
+			auto*      region = track ? GetOrCreateRegion(chunk_begin) : FindRegion(chunk_begin);
+			if (region == nullptr) {
+				Fatal("untracking unknown page 0x%016" PRIx64, chunk_begin);
+			}
+			const auto first = static_cast<size_t>((chunk_begin - region_base) / PAGE_SIZE);
+			const auto last  = static_cast<size_t>((chunk_end - region_base) / PAGE_SIZE);
+			UpdateRegionWatchers<track, is_read, false>(*region, region_base, first, last);
+			chunk_begin = chunk_end;
+		}
 	}
 
 	std::unique_ptr<std::atomic<Region*>[]> regions;
 	std::vector<std::unique_ptr<Region>>    region_storage;
 	std::mutex                              region_mutex;
-	PageFaultHandler                        fault_handler = nullptr;
-	void*                                   fault_context = nullptr;
 };
 
 static_assert(std::atomic<void*>::is_always_lock_free);
 
-PageManager::PageManager(PageFaultHandler fault_handler, void* fault_context)
-    : m_impl(std::make_unique<Impl>(fault_handler, fault_context)) {}
+PageManager::PageManager(): m_impl(std::make_unique<Impl>()) {}
 
 PageManager::~PageManager() = default;
 
 uint64_t PageManager::GetPageSize() const {
-	if (g_in_fault_resolution) {
-		FailFast("nested page fault while resolving a watched page");
-	}
 	return PAGE_SIZE;
 }
 
-bool PageManager::IsTracked(uint64_t vaddr) const noexcept {
-	if (g_in_fault_resolution) {
-		FailFast("IsTracked called during fault resolution");
+template <bool track>
+void PageManager::UpdatePageWatchers(uint64_t vaddr, uint64_t size) {
+	m_impl->UpdatePageWatchers<track, false>(vaddr, size);
+}
+
+template void PageManager::UpdatePageWatchers<true>(uint64_t, uint64_t);
+template void PageManager::UpdatePageWatchers<false>(uint64_t, uint64_t);
+
+template <bool track, bool is_read>
+void PageManager::UpdatePageWatchersForRegion(uint64_t base_addr, RegionBits& mask) {
+	if (base_addr % REGION_SIZE != 0 || base_addr >= ADDRESS_SIZE ||
+	    REGION_SIZE > ADDRESS_SIZE - base_addr) {
+		Fatal("invalid tracking region base 0x%016" PRIx64, base_addr);
 	}
-	auto* region = m_impl->FindRegion(vaddr);
+
+	const auto start_range = mask.FirstRange();
+	const auto end_range   = mask.LastRange();
+	if (start_range.first == REGION_PAGES) {
+		FailFast("empty region watcher mask");
+	}
+	const auto first = start_range.first;
+	const auto last  = end_range.second;
+	if (start_range.second == end_range.second) {
+		m_impl->UpdatePageWatchers<track, is_read>(base_addr + first * PAGE_SIZE,
+		                                           (last - first) * PAGE_SIZE);
+		return;
+	}
+
+	auto* region = track ? m_impl->GetOrCreateRegion(base_addr) : m_impl->FindRegion(base_addr);
 	if (region == nullptr) {
-		return false;
+		Fatal("untracking unknown region 0x%016" PRIx64, base_addr);
 	}
-	auto&     page = m_impl->GetPage(*region, vaddr);
-	SpinGuard lock(page.lock);
-	return page.write_watchers != 0 || page.access_watchers != 0;
+	m_impl->UpdateRegionWatchers<track, is_read, true>(*region, base_addr, first, last, &mask);
 }
 
-bool PageManager::IsMapped(uint64_t vaddr, uint64_t size) const noexcept {
-	if (vaddr == 0 || size == 0 || vaddr >= ADDRESS_SIZE || size > ADDRESS_SIZE - vaddr) {
-		return false;
-	}
-	const auto end = PageStart(vaddr + size - 1) + PAGE_SIZE;
-	for (auto page_vaddr = PageStart(vaddr); page_vaddr < end; page_vaddr += PAGE_SIZE) {
-		auto* region = m_impl->FindRegion(page_vaddr);
-		if (region == nullptr) {
-			return false;
-		}
-		auto&     page = m_impl->GetPage(*region, page_vaddr);
-		SpinGuard lock(page.lock);
-		if (page.mappings == 0) {
-			return false;
-		}
-	}
-	return true;
-}
+template void PageManager::UpdatePageWatchersForRegion<true, true>(uint64_t, RegionBits&);
+template void PageManager::UpdatePageWatchersForRegion<true, false>(uint64_t, RegionBits&);
+template void PageManager::UpdatePageWatchersForRegion<false, true>(uint64_t, RegionBits&);
+template void PageManager::UpdatePageWatchersForRegion<false, false>(uint64_t, RegionBits&);
 
-bool PageManager::HasAnyMapping(uint64_t vaddr, uint64_t size) const noexcept {
-	if (g_in_fault_resolution || vaddr == 0 || size == 0 || vaddr >= ADDRESS_SIZE ||
-	    size > ADDRESS_SIZE - vaddr) {
-		return false;
-	}
-	const auto end = PageEnd(vaddr, size);
-	for (auto page_vaddr = PageStart(vaddr); page_vaddr < end; page_vaddr += PAGE_SIZE) {
-		auto* region = m_impl->FindRegion(page_vaddr);
-		if (region == nullptr) {
-			continue;
-		}
-		auto&     page = m_impl->GetPage(*region, page_vaddr);
-		SpinGuard lock(page.lock);
-		if (page.mappings != 0) {
-			return true;
-		}
-	}
-	return false;
-}
+void PageManager::OnGpuMap(uint64_t, uint64_t) {}
 
-bool PageManager::HasGpuAccess(uint64_t vaddr, uint64_t size, GpuAccess access) const noexcept {
-	if (access != GpuAccess::Read && access != GpuAccess::Write && access != GpuAccess::ReadWrite) {
-		FailFast("HasGpuAccess received an invalid GPU access mode");
-	}
-	const bool need_read  = access == GpuAccess::Read || access == GpuAccess::ReadWrite;
-	const bool need_write = access == GpuAccess::Write || access == GpuAccess::ReadWrite;
-	if (vaddr == 0 || size == 0 || vaddr >= ADDRESS_SIZE || size > ADDRESS_SIZE - vaddr) {
-		return false;
-	}
-	const auto end = PageEnd(vaddr, size);
-	for (auto addr = PageStart(vaddr); addr < end; addr += PAGE_SIZE) {
-		auto* region = m_impl->FindRegion(addr);
-		if (region == nullptr) {
-			return false;
-		}
-		auto&     page = m_impl->GetPage(*region, addr);
-		SpinGuard lock(page.lock);
-		if ((need_read && page.gpu_read_mappings == 0) ||
-		    (need_write && page.gpu_write_mappings == 0)) {
-			return false;
-		}
-	}
-	return true;
-}
-
-void PageManager::UpdatePageWatchers(bool track, uint64_t vaddr, uint64_t size,
-                                     PageWatchMode mode) {
-	if (mode != PageWatchMode::Write && mode != PageWatchMode::ReadWrite) {
-		Fatal("invalid watcher mode");
-	}
-	const auto end = PageEnd(vaddr, size);
-	for (auto page_vaddr = PageStart(vaddr); page_vaddr < end; page_vaddr += PAGE_SIZE) {
-		auto* region =
-		    track ? m_impl->GetOrCreateRegion(page_vaddr) : m_impl->FindRegion(page_vaddr);
-		if (region == nullptr) {
-			Fatal("untracking unknown page 0x%016" PRIx64, page_vaddr);
-		}
-		auto&     page = m_impl->GetPage(*region, page_vaddr);
-		SpinGuard lock(page.lock);
-		if (page.resolving && track) {
-			FailFast("new page watcher raced active fault resolution");
-		}
-		if (page.mappings == 0) {
-			Fatal("watching unmapped page 0x%016" PRIx64, page_vaddr);
-		}
-		auto& watchers =
-		    (mode == PageWatchMode::ReadWrite ? page.access_watchers : page.write_watchers);
-		if (track) {
-			if (watchers == std::numeric_limits<uint32_t>::max()) {
-				Fatal("watcher overflow at 0x%016" PRIx64, page_vaddr);
-			}
-			const bool first_watcher = page.write_watchers == 0 && page.access_watchers == 0;
-			if (first_watcher) {
-				page.original_protection = Impl::QueryProtection(page_vaddr);
-			}
-			const auto old_protection = Impl::WatcherProtection(page);
-			watchers++;
-			const auto new_protection = Impl::WatcherProtection(page);
-			if (new_protection != old_protection) {
-				Impl::Protect(page_vaddr, new_protection, old_protection, false);
-			}
-			switch (new_protection) {
-				case NO_ACCESS_PROTECTION:
-					page.late_read_pending  = false;
-					page.late_write_pending = false;
-					break;
-				case READ_ONLY_PROTECTION: page.late_write_pending = false; break;
-				default: break;
-			}
-		} else {
-			if (watchers == 0) {
-				Fatal("watcher underflow at 0x%016" PRIx64, page_vaddr);
-			}
-			if (page.backing_writer != 0 && page.backing_writer != CurrentThread()) {
-				Fatal("backing write ownership changed at 0x%016" PRIx64, page_vaddr);
-			}
-			const auto old_protection = Impl::WatcherProtection(page);
-			watchers--;
-			const auto new_protection = Impl::WatcherProtection(page);
-			if (page.backing_writer == 0 && new_protection != old_protection) {
-				Impl::Protect(page_vaddr, new_protection, old_protection, false);
-			}
-			if (page.backing_writer == 0) {
-				Impl::PublishDelayedFaults(page, old_protection, new_protection);
-			}
-			if (page.backing_writer == 0 && page.write_watchers == 0 && page.access_watchers == 0) {
-				page.original_protection = 0;
-			}
-		}
-	}
-}
-
-void PageManager::OnGpuMap(uint64_t vaddr, uint64_t size, GpuAccess access) {
-	if (g_in_fault_resolution) {
-		FailFast("GPU mapping changed during fault resolution");
-	}
-	if (access != GpuAccess::Read && access != GpuAccess::Write && access != GpuAccess::ReadWrite) {
-		FailFast("GPU map received an invalid access mode");
-	}
-	const bool gpu_read  = access == GpuAccess::Read || access == GpuAccess::ReadWrite;
-	const bool gpu_write = access == GpuAccess::Write || access == GpuAccess::ReadWrite;
-	const auto end       = PageEnd(vaddr, size);
-	for (auto addr = PageStart(vaddr); addr < end; addr += PAGE_SIZE) {
-		auto&     page = m_impl->GetPage(*m_impl->GetOrCreateRegion(addr), addr);
-		SpinGuard lock(page.lock);
-		if (page.resolving || page.mappings == std::numeric_limits<uint32_t>::max() ||
-		    (gpu_read && page.gpu_read_mappings == std::numeric_limits<uint32_t>::max()) ||
-		    (gpu_write && page.gpu_write_mappings == std::numeric_limits<uint32_t>::max())) {
-			Fatal("invalid map state at 0x%016" PRIx64, addr);
-		}
-		page.mappings++;
-		page.gpu_read_mappings += gpu_read ? 1u : 0u;
-		page.gpu_write_mappings += gpu_write ? 1u : 0u;
-	}
-}
-
-void PageManager::OnGpuUnmap(uint64_t vaddr, uint64_t size, GpuAccess access) {
-	if (g_in_fault_resolution) {
-		FailFast("GPU unmapping changed during fault resolution");
-	}
-	if (access != GpuAccess::Read && access != GpuAccess::Write && access != GpuAccess::ReadWrite) {
-		FailFast("GPU unmap received an invalid access mode");
-	}
-	const bool gpu_read  = access == GpuAccess::Read || access == GpuAccess::ReadWrite;
-	const bool gpu_write = access == GpuAccess::Write || access == GpuAccess::ReadWrite;
-	const auto end       = PageEnd(vaddr, size);
-	for (auto page_vaddr = PageStart(vaddr); page_vaddr < end; page_vaddr += PAGE_SIZE) {
-		auto* region = m_impl->FindRegion(page_vaddr);
-		if (region == nullptr) {
-			Fatal("unmapping unknown page 0x%016" PRIx64, page_vaddr);
-		}
-		auto&     page = m_impl->GetPage(*region, page_vaddr);
-		SpinGuard lock(page.lock);
-		if (page.resolving || page.mappings == 0 || (gpu_read && page.gpu_read_mappings == 0) ||
-		    (gpu_write && page.gpu_write_mappings == 0) ||
-		    (page.mappings == 1 && (page.write_watchers != 0 || page.access_watchers != 0))) {
-			Fatal("invalid unmap state at 0x%016" PRIx64, page_vaddr);
-		}
-		page.mappings--;
-		page.gpu_read_mappings -= gpu_read ? 1u : 0u;
-		page.gpu_write_mappings -= gpu_write ? 1u : 0u;
-		if (page.mappings == 0) {
-			if (page.gpu_read_mappings != 0 || page.gpu_write_mappings != 0) {
-				FailFast("GPU unmap left nonzero GPU mapping counts");
-			}
-			page.late_read_pending  = false;
-			page.late_write_pending = false;
-		}
-	}
-}
-
-PageManager::BackingWrite::BackingWrite(PageManager& manager, uint64_t vaddr,
-                                        uint64_t size) noexcept
-    : m_manager(manager), m_vaddr(vaddr), m_size(size) {
-	m_manager.BeginBackingWrite(vaddr, size);
-}
-
-PageManager::BackingWrite::~BackingWrite() {
-	m_manager.EndBackingWrite(m_vaddr, m_size);
-}
-
-std::vector<std::unique_ptr<PageManager::BackingWrite>>
-PageManager::ReserveBackingWrites(std::span<const RangeSet::Range> ranges) {
-	if (ranges.empty()) {
-		Fatal("cannot reserve empty backing-write ranges");
-	}
-	std::vector<std::unique_ptr<BackingWrite>> writes;
-	writes.reserve(ranges.size());
-	uint64_t begin = 0;
-	uint64_t end   = 0;
-	for (const auto& range: ranges) {
-		if (range.address == 0 || range.size == 0 || range.size > UINT64_MAX - range.address ||
-		    range.address + range.size > UINT64_MAX - (PAGE_SIZE - 1)) {
-			Fatal("invalid backing-write range");
-		}
-		const auto page_begin = PageStart(range.address);
-		const auto page_end   = PageStart(range.address + range.size + PAGE_SIZE - 1);
-		if (begin != 0 && page_begin > end) {
-			writes.push_back(std::make_unique<BackingWrite>(*this, begin, end - begin));
-			begin = 0;
-		}
-		if (begin == 0) {
-			begin = page_begin;
-			end   = page_end;
-		} else {
-			end = std::max(end, page_end);
-		}
-	}
-	writes.push_back(std::make_unique<BackingWrite>(*this, begin, end - begin));
-	return writes;
-}
-
-void PageManager::BeginBackingWrite(uint64_t vaddr, uint64_t size) noexcept {
-	if (g_in_fault_resolution) {
-		FailFast("backing write began during fault resolution");
-	}
-	const auto end    = PageEnd(vaddr, size);
-	const auto writer = CurrentThread();
-	for (auto address = PageStart(vaddr); address < end; address += PAGE_SIZE) {
-		auto* region = m_impl->FindRegion(address);
-		if (region == nullptr) {
-			Fatal("backing write reserves an unknown page at 0x%016" PRIx64, address);
-		}
-		auto&     page = m_impl->GetPage(*region, address);
-		SpinGuard lock(page.lock);
-		if (page.mappings == 0 || page.resolving || page.backing_writer != 0 ||
-		    page.access_watchers == 0) {
-			Fatal("backing write races page resolution at 0x%016" PRIx64, address);
-		}
-		page.resolving            = true;
-		page.resolving_read_write = true;
-		page.backing_writer       = writer;
-	}
-}
-
-void PageManager::EndBackingWrite(uint64_t vaddr, uint64_t size) noexcept {
-	if (g_in_fault_resolution) {
-		FailFast("backing write ended during fault resolution");
-	}
-	const auto end    = PageEnd(vaddr, size);
-	const auto writer = CurrentThread();
-	for (auto address = PageStart(vaddr); address < end; address += PAGE_SIZE) {
-		auto* region = m_impl->FindRegion(address);
-		if (region == nullptr) {
-			FailFast("backing write ended for an unknown page");
-		}
-		auto&     page = m_impl->GetPage(*region, address);
-		SpinGuard lock(page.lock);
-		if (!page.resolving || page.backing_writer != writer) {
-			FailFast("backing write ended without matching owner and resolving state");
-		}
-		const auto old_protection = NO_ACCESS_PROTECTION;
-		const auto new_protection = Impl::WatcherProtection(page);
-		if (new_protection != old_protection) {
-			Impl::Protect(address, new_protection, old_protection, false);
-		}
-		Impl::PublishDelayedFaults(page, old_protection, new_protection);
-		if (page.write_watchers == 0 && page.access_watchers == 0) {
-			page.original_protection = 0;
-		}
-		page.backing_writer       = 0;
-		page.resolving            = false;
-		page.resolving_read_write = false;
-	}
-}
-
-bool PageManager::HandleFault(PageFaultAccess access, uint64_t fault_vaddr) noexcept {
-	if (g_in_fault_resolution) {
-		FailFast("nested HandleFault call");
-	}
-	auto* region = m_impl->FindRegion(fault_vaddr);
-	if (region == nullptr) {
-		return false;
-	}
-	auto& page   = m_impl->GetPage(*region, fault_vaddr);
-	bool  waited = false;
-	while (true) {
-		SpinGuard lock(page.lock);
-		if (access == PageFaultAccess::Read && page.late_read_pending &&
-		    Impl::AllowsAccess(fault_vaddr, access)) {
-			page.late_read_pending = false;
-			return true;
-		}
-		if (access == PageFaultAccess::Write && page.late_write_pending &&
-		    Impl::AllowsAccess(fault_vaddr, access)) {
-			page.late_write_pending = false;
-			return true;
-		}
-		if (page.resolving) {
-			if (page.backing_writer == CurrentThread()) {
-				FailFast("backing writer faulted on its own reserved page");
-			}
-			if ((!page.resolving_read_write && access != PageFaultAccess::Write) ||
-			    (page.resolving_read_write && access != PageFaultAccess::Read &&
-			     access != PageFaultAccess::Write)) {
-				FailFast("fault access is incompatible with the active resolver");
-			}
-			waited = true;
-			continue;
-		}
-		if (page.write_watchers == 0 && page.access_watchers == 0) {
-			if (access != PageFaultAccess::Read && access != PageFaultAccess::Write) {
-				return false;
-			}
-			bool&      pending = (access == PageFaultAccess::Read ? page.late_read_pending
-			                                                      : page.late_write_pending);
-			const bool allowed = Impl::AllowsAccess(fault_vaddr, access);
-			pending            = false;
-			if (waited && !allowed) {
-				FailFast("page remained inaccessible after waiting for its resolver");
-			}
-			// More than one CPU can fault before a protection transition becomes visible. The first
-			// delayed fault consumes the hint bit; later faults must also resume once the mapped
-			// page already permits the requested access. A genuinely read-only/no-access page still
-			// falls through to the guest exception path.
-			return allowed;
-		}
-		if ((access != PageFaultAccess::Read && access != PageFaultAccess::Write) ||
-		    (access == PageFaultAccess::Read && page.access_watchers == 0)) {
-			FailFast("fault access is incompatible with active page watchers");
-		}
-		page.resolving            = true;
-		page.resolving_read_write = page.access_watchers != 0;
-		break;
-	}
-	g_in_fault_resolution = true;
-	const bool handled    = m_impl->fault_handler(m_impl->fault_context, access, fault_vaddr, 1,
-	                                              PageFaultPhase::Invalidate);
-	g_in_fault_resolution = false;
-	{
-		SpinGuard lock(page.lock);
-		if (!handled || !page.resolving) {
-			FailFast("fault invalidation did not preserve the resolving state");
-		}
-	}
-	g_in_fault_resolution = true;
-	const bool completed  = m_impl->fault_handler(m_impl->fault_context, access, fault_vaddr, 1,
-	                                              PageFaultPhase::Complete);
-	g_in_fault_resolution = false;
-	{
-		SpinGuard lock(page.lock);
-		if (!completed || !page.resolving) {
-			FailFast("fault completion did not preserve the resolving state");
-		}
-		if (page.write_watchers != 0 || page.access_watchers != 0) {
-			const auto old_protection  = Impl::WatcherProtection(page);
-			const bool read_only_fault = access == PageFaultAccess::Read;
-			if (read_only_fault && page.access_watchers == 0) {
-				FailFast("read fault completed without a read/write watcher");
-			}
-			page.access_watchers = 0;
-			if (!read_only_fault) {
-				page.write_watchers = 0;
-			}
-			const auto restored_protection = Impl::WatcherProtection(page);
-			Impl::Protect(PageStart(fault_vaddr), restored_protection, old_protection, true);
-			if (page.write_watchers == 0) {
-				page.original_protection = 0;
-			}
-			Impl::PublishDelayedFaults(page, old_protection, restored_protection);
-		} else if (!Impl::AllowsAccess(fault_vaddr, access)) {
-			FailFast("fault completion left the page inaccessible");
-		}
-		page.resolving            = false;
-		page.resolving_read_write = false;
-	}
-	g_in_fault_resolution = true;
-	const bool released   = m_impl->fault_handler(m_impl->fault_context, access, fault_vaddr, 1,
-	                                              PageFaultPhase::Release);
-	g_in_fault_resolution = false;
-	if (!released) {
-		FailFast("fault release callback failed");
-	}
-	return true;
-}
-
-bool PageManager::HandleWriteRange(uint64_t vaddr, uint64_t size) noexcept {
-	if (g_in_fault_resolution || vaddr == 0 || size == 0 || vaddr >= ADDRESS_SIZE ||
-	    size > ADDRESS_SIZE - vaddr) {
-		return false;
-	}
-	const auto end = PageEnd(vaddr, size);
-	for (auto page_vaddr = PageStart(vaddr); page_vaddr < end; page_vaddr += PAGE_SIZE) {
-		if (!IsMapped(page_vaddr, 1)) {
-			continue;
-		}
-		const auto fault_vaddr = std::max(page_vaddr, vaddr);
-		if (!HandleFault(PageFaultAccess::Write, fault_vaddr)) {
-			return false;
-		}
-	}
-	return true;
-}
+void PageManager::OnGpuUnmap(uint64_t, uint64_t) {}
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/pageManager.h

@@ -2,62 +2,32 @@
 #define EMULATOR_SRC_GRAPHICS_HOST_GPU_PAGEMANAGER_H_
 
 #include "common/common.h"
-#include "graphics/host_gpu/rangeSet.h"
+#include "graphics/host_gpu/regionDefinitions.h"
 
 #include <memory>
-#include <span>
-#include <vector>
 
 namespace Libs::Graphics {
 
 enum class PageFaultAccess { Read, Write, Execute, Unknown };
-enum class PageFaultPhase { Invalidate, Complete, Release };
-enum class PageWatchMode { Write, ReadWrite };
-enum class GpuAccess { Read, Write, ReadWrite };
-
-using PageFaultHandler = bool (*)(void* context, PageFaultAccess access, uint64_t vaddr,
-                                  uint64_t size, PageFaultPhase phase) noexcept;
 
 class PageManager final {
 public:
-	class BackingWrite final {
-	public:
-		BackingWrite(PageManager& manager, uint64_t vaddr, uint64_t size) noexcept;
-		~BackingWrite();
-		KYTY_CLASS_NO_COPY(BackingWrite);
-
-	private:
-		PageManager& m_manager;
-		uint64_t     m_vaddr = 0;
-		uint64_t     m_size  = 0;
-	};
-
-	PageManager(PageFaultHandler fault_handler, void* fault_context);
+	PageManager();
 	// The owner must stop all PageManager callers before destruction.
 	~PageManager();
 
 	KYTY_CLASS_NO_COPY(PageManager);
 
 	[[nodiscard]] uint64_t GetPageSize() const;
-	[[nodiscard]] bool     IsTracked(uint64_t vaddr) const noexcept;
-	[[nodiscard]] bool     IsMapped(uint64_t vaddr, uint64_t size) const noexcept;
-	[[nodiscard]] bool     HasAnyMapping(uint64_t vaddr, uint64_t size) const noexcept;
-	[[nodiscard]] bool HasGpuAccess(uint64_t vaddr, uint64_t size, GpuAccess access) const noexcept;
 
-	void UpdatePageWatchers(bool track, uint64_t vaddr, uint64_t size,
-	                        PageWatchMode mode = PageWatchMode::Write);
-	void OnGpuMap(uint64_t vaddr, uint64_t size, GpuAccess access = GpuAccess::ReadWrite);
-	void OnGpuUnmap(uint64_t vaddr, uint64_t size, GpuAccess access = GpuAccess::ReadWrite);
-
-	[[nodiscard]] bool HandleFault(PageFaultAccess access, uint64_t fault_vaddr) noexcept;
-	[[nodiscard]] bool HandleWriteRange(uint64_t vaddr, uint64_t size) noexcept;
-	[[nodiscard]] std::vector<std::unique_ptr<BackingWrite>>
-	     ReserveBackingWrites(std::span<const RangeSet::Range> ranges);
+	template <bool track>
+	void UpdatePageWatchers(uint64_t vaddr, uint64_t size);
+	template <bool track, bool is_read = false>
+	void UpdatePageWatchersForRegion(uint64_t base_addr, RegionBits& mask);
+	void OnGpuMap(uint64_t vaddr, uint64_t size);
+	void OnGpuUnmap(uint64_t vaddr, uint64_t size);
 
 private:
-	void BeginBackingWrite(uint64_t vaddr, uint64_t size) noexcept;
-	void EndBackingWrite(uint64_t vaddr, uint64_t size) noexcept;
-
 	struct Impl;
 	std::unique_ptr<Impl> m_impl;
 };

src/graphics/host_gpu/rangeSet.h

@@ -59,6 +59,16 @@ public:
 		return result;
 	}
 
+	[[nodiscard]] bool Contains(uint64_t address, uint64_t size) const {
+		const auto end = End(address, size);
+		auto       it  = m_ranges.upper_bound(address);
+		if (it == m_ranges.begin()) {
+			return false;
+		}
+		--it;
+		return it->first <= address && it->second >= end;
+	}
+
 	template <typename Func>
 	void ForEachIntersection(uint64_t address, uint64_t size, Func&& func) const {
 		const auto end = End(address, size);

src/graphics/host_gpu/regionDefinitions.h

@@ -1,10 +1,9 @@
 #ifndef EMULATOR_SRC_GRAPHICS_HOST_GPU_REGIONDEFINITIONS_H_
 #define EMULATOR_SRC_GRAPHICS_HOST_GPU_REGIONDEFINITIONS_H_
 
+#include "common/bitArray.h"
 #include "common/common.h"
 
-#include <bitset>
-
 namespace Libs::Graphics {
 
 constexpr uint64_t TRACKER_PAGE_SIZE    = 4ull * 1024ull;
@@ -13,7 +12,8 @@ constexpr uint64_t TRACKER_ADDRESS_SIZE = 1ull << 40u;
 constexpr size_t   TRACKER_REGION_PAGES = TRACKER_REGION_SIZE / TRACKER_PAGE_SIZE;
 
 enum class DirtySource { Cpu, Gpu };
-using RegionBits = std::bitset<TRACKER_REGION_PAGES>;
+using RegionBits = Common::BitArray<TRACKER_REGION_PAGES>;
+static_assert(sizeof(RegionBits) == TRACKER_REGION_PAGES / 8);
 
 } // namespace Libs::Graphics
 

src/graphics/host_gpu/regionManager.h

@@ -16,12 +16,15 @@
 #include <windows.h>
 #undef min
 #undef max
+#elif defined(__APPLE__)
+#include <pthread.h>
+#elif defined(__linux__)
+#include <sys/syscall.h>
+#include <unistd.h>
 #endif
 
 namespace Libs::Graphics {
 
-enum class CpuFaultAction { Untracked, Continue, Download };
-
 class TrackingSpinLock final {
 public:
 	void lock() noexcept {
@@ -49,6 +52,12 @@ private:
 	static uint32_t CurrentThread() noexcept {
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 		return GetCurrentThreadId();
+#elif defined(__APPLE__)
+		// mach thread port is a nonzero per-thread id (0 is the "no owner" sentinel).
+		return static_cast<uint32_t>(pthread_mach_thread_np(pthread_self()));
+#elif defined(__linux__)
+		static thread_local const uint32_t tid = static_cast<uint32_t>(::syscall(SYS_gettid));
+		return tid;
 #else
 		EXIT("region tracking thread identity is unsupported on this platform\n");
 #endif
@@ -67,229 +76,93 @@ public:
 		if (m_cpu_addr % TRACKER_REGION_SIZE != 0) {
 			EXIT("invalid region tracking manager construction\n");
 		}
-		m_cpu_dirty.set();
-		m_writable.set();
+		m_cpu_dirty.Fill();
+		m_writable.Fill();
+		m_readable.Fill();
 	}
 
 	KYTY_CLASS_NO_COPY(RegionManager);
 
 	[[nodiscard]] uint64_t GetCpuAddr() const { return m_cpu_addr; }
-	void                   Track(uint64_t vaddr, uint64_t size) {
-		const auto [start, end] = GetPageRange(vaddr, size);
-		for (auto page = start; page < end; page++) {
-			m_tracked.set(page);
-		}
-	}
-	void Untrack(uint64_t vaddr, uint64_t size) {
-		const auto [start, end] = GetPageRange(vaddr, size);
-		for (auto page = start; page < end; page++) {
-			m_tracked.reset(page);
-		}
-	}
 	template <DirtySource source>
 	[[nodiscard]] bool IsModified(uint64_t offset, uint64_t size) const {
 		const auto [start, end] = GetPageRange(m_cpu_addr + offset, size);
 		const auto& bits        = GetBits<source>();
-		for (auto page = start; page < end; page++) {
-			if (bits.test(page)) {
-				return true;
-			}
-		}
-		return false;
-	}
-
-	template <DirtySource source>
-	[[nodiscard]] bool IsFullyModified(uint64_t offset, uint64_t size) const {
-		const auto [start, end] = GetPageRange(m_cpu_addr + offset, size);
-		const auto& bits        = GetBits<source>();
-		for (auto page = start; page < end; page++) {
-			if (!bits.test(page)) {
-				return false;
-			}
-		}
-		return true;
+		return RegionBits(bits, start, end).Any();
 	}
 
 	template <DirtySource source, bool enable>
-	RegionBits ChangeState(uint64_t vaddr, uint64_t size) {
+	void ChangeState(uint64_t vaddr, uint64_t size) {
 		const auto [start, end] = GetPageRange(vaddr, size);
 		if constexpr (source == DirtySource::Cpu && enable) {
-			for (auto page = start; page < end; page++) {
-				if (m_gpu_dirty.test(page) || m_fault_pending.test(page)) {
-					EXIT("CPU dirty state conflicts with GPU dirty or pending fault state\n");
-				}
+			if (RegionBits(m_gpu_dirty, start, end).Any()) {
+				EXIT("CPU dirty state conflicts with GPU dirty state\n");
 			}
 		}
 		if constexpr (source == DirtySource::Gpu && enable) {
-			for (auto page = start; page < end; page++) {
-				if (m_cpu_dirty.test(page) || m_fault_pending.test(page)) {
-					EXIT("GPU dirty state conflicts with CPU dirty or pending fault state\n");
-				}
+			if (RegionBits(m_cpu_dirty, start, end).Any()) {
+				EXIT("GPU dirty state conflicts with CPU dirty state\n");
 			}
 		}
-		auto& bits    = GetBits<source>();
-		auto  changed = bits;
-		for (auto page = start; page < end; page++) {
-			bits.set(page, enable);
+		auto& bits = GetBits<source>();
+		if constexpr (enable) {
+			bits.SetRange(start, end);
+		} else {
+			bits.UnsetRange(start, end);
 		}
-		changed ^= bits;
 		if constexpr (source == DirtySource::Cpu) {
-			changed    = m_cpu_dirty ^ m_writable;
-			m_writable = m_cpu_dirty;
+			UpdateCpuProtection<!enable>();
+		} else {
+			UpdateGpuProtection<enable>();
 		}
-		return changed;
-	}
-
-	[[nodiscard]] CpuFaultAction BeginCpuFault(uint64_t vaddr, uint64_t size,
-	                                           PageFaultAccess access = PageFaultAccess::Write) {
-		if (access != PageFaultAccess::Read && access != PageFaultAccess::Write) {
-			EXIT("unsupported CPU fault access while beginning ownership transfer\n");
-		}
-		const auto [start, end] = GetPageRange(vaddr, size);
-		const bool tracked      = m_tracked.test(start);
-		for (auto page = start; page < end; page++) {
-			if (m_tracked.test(page) != tracked) {
-				EXIT("CPU fault spans mixed tracked and untracked pages\n");
-			}
-			if (m_fault_pending.test(page)) {
-				return CpuFaultAction::Untracked;
-			}
-			if (m_cpu_dirty.test(page) != m_writable.test(page) ||
-			    (m_gpu_dirty.test(page) && (m_cpu_dirty.test(page) || m_writable.test(page)))) {
-				EXIT("inconsistent CPU fault page state\n");
-			}
-		}
-		if (!tracked) {
-			return CpuFaultAction::Untracked;
-		}
-		bool gpu_dirty = m_gpu_dirty.test(start);
-		bool writable  = m_writable.test(start);
-		for (auto page = start + 1; page < end; page++) {
-			if (m_gpu_dirty.test(page) != gpu_dirty || m_writable.test(page) != writable) {
-				EXIT("CPU fault spans pages with incompatible dirty or writable state\n");
-			}
-		}
-		for (auto page = start; page < end; page++) {
-			if (!gpu_dirty && access == PageFaultAccess::Write) {
-				m_cpu_dirty.set(page);
-				m_writable.set(page);
-			}
-			m_fault_pending.set(page);
-		}
-		return gpu_dirty ? CpuFaultAction::Download : CpuFaultAction::Continue;
-	}
-
-	[[nodiscard]] bool CompleteCpuFault(uint64_t vaddr, uint64_t size, PageFaultAccess access,
-	                                    bool downloaded) {
-		const auto [start, end] = GetPageRange(vaddr, size);
-		for (auto page = start; page < end; page++) {
-			if (!m_fault_pending.test(page)) {
-				return false;
-			}
-		}
-		for (auto page = start; page < end; page++) {
-			const bool gpu_dirty = m_gpu_dirty.test(page);
-			if (gpu_dirty != downloaded) {
-				EXIT("CPU fault download result disagrees with GPU dirty state\n");
-			}
-			if (gpu_dirty) {
-				m_gpu_dirty.reset(page);
-				switch (access) {
-					case PageFaultAccess::Read: break;
-					case PageFaultAccess::Write:
-						m_cpu_dirty.set(page);
-						m_writable.set(page);
-						break;
-					default: EXIT("unsupported CPU fault access after GPU download\n");
-				}
-			}
-			m_fault_pending.reset(page);
-		}
-		return true;
-	}
-
-	[[nodiscard]] bool HasPendingFault(uint64_t vaddr, uint64_t size) const {
-		const auto [start, end] = GetPageRange(vaddr, size);
-		for (auto page = start; page < end; page++) {
-			if (m_fault_pending.test(page)) {
-				return true;
-			}
-		}
-		return false;
-	}
-
-	[[nodiscard]] bool CompleteVirtualGpuWrite(uint64_t vaddr, uint64_t size) {
-		const auto [start, end] = GetPageRange(vaddr, size);
-		for (auto page = start; page < end; page++) {
-			if (!m_fault_pending.test(page)) {
-				return false;
-			}
-			if (!m_gpu_dirty.test(page)) {
-				EXIT("virtual GPU write completion found a non-GPU-dirty page\n");
-			}
-		}
-		for (auto page = start; page < end; page++) {
-			m_gpu_dirty.reset(page);
-			m_cpu_dirty.set(page);
-			m_writable.set(page);
-			m_fault_pending.reset(page);
-		}
-		return true;
 	}
 
 	template <DirtySource source, bool clear, typename Func>
-	RegionBits ForEachModifiedRange(uint64_t vaddr, uint64_t size, Func&& func) {
+	void ForEachModifiedRange(uint64_t vaddr, uint64_t size, Func&& func) {
 		const auto [start, end] = GetPageRange(vaddr, size);
-		auto mask               = GetBits<source>();
-		if constexpr (source == DirtySource::Cpu) {
-			mask &= ~m_fault_pending;
-		}
-		for (auto page = 0u; page < start; page++) {
-			mask.reset(page);
-		}
-		for (auto page = end; page < TRACKER_REGION_PAGES; page++) {
-			mask.reset(page);
-		}
+		RegionBits mask(GetBits<source>(), start, end);
 		if constexpr (clear) {
-			auto& bits = GetBits<source>();
-			for (auto page = start; page < end; page++) {
-				if (mask.test(page)) {
-					bits.reset(page);
-				}
-			}
+			GetBits<source>().UnsetRange(start, end);
 		}
 		if constexpr (source == DirtySource::Cpu && clear) {
-			auto changed = m_cpu_dirty ^ m_writable;
-			m_writable   = m_cpu_dirty;
-			ApplyProtection(changed, true);
+			UpdateCpuProtection<true>();
 			ForEachRange(mask, std::forward<Func>(func));
-			return changed;
+			return;
+		}
+		if constexpr (source == DirtySource::Gpu && clear) {
+			UpdateGpuProtection<false>();
 		}
 		ForEachRange(mask, std::forward<Func>(func));
-		if constexpr (clear) {
-			return mask;
-		}
-		return {};
-	}
-
-	void ApplyProtection(const RegionBits& changed, bool track) {
-		ForEachRange(changed, [this, track](uint64_t vaddr, uint64_t size) {
-			m_page_manager.UpdatePageWatchers(track, vaddr, size);
-		});
-	}
-
-	void ApplyGpuProtection(const RegionBits& changed, bool track, PageWatchMode mode) {
-		if (mode != PageWatchMode::Write && mode != PageWatchMode::ReadWrite) {
-			EXIT("unsupported GPU page-watch mode\n");
-		}
-		ForEachRange(changed, [this, track, mode](uint64_t vaddr, uint64_t size) {
-			m_page_manager.UpdatePageWatchers(track, vaddr, size, mode);
-		});
 	}
 
 	TrackingSpinLock lock;
 
 private:
+	template <bool track>
+	void UpdateCpuProtection() {
+		auto mask  = m_cpu_dirty ^ m_writable;
+		m_writable = m_cpu_dirty;
+		if (mask.None()) {
+			return;
+		}
+		m_page_manager.UpdatePageWatchersForRegion<track>(m_cpu_addr, mask);
+	}
+
+	template <bool track>
+	void UpdateGpuProtection() {
+		auto readable = ~m_gpu_dirty;
+		auto mask     = readable ^ m_readable;
+		m_readable    = readable;
+		if (mask.None()) {
+			return;
+		}
+		if constexpr (track) {
+			m_page_manager.UpdatePageWatchersForRegion<true, true>(m_cpu_addr, mask);
+		} else {
+			m_page_manager.UpdatePageWatchersForRegion<false, true>(m_cpu_addr, mask);
+		}
+	}
+
 	template <DirtySource source>
 	RegionBits& GetBits() {
 		if constexpr (source == DirtySource::Cpu) {
@@ -320,18 +193,8 @@ private:
 
 	template <typename Func>
 	void ForEachRange(const RegionBits& bits, Func&& func) const {
-		size_t page = 0;
-		while (page < TRACKER_REGION_PAGES) {
-			while (page < TRACKER_REGION_PAGES && !bits.test(page)) {
-				page++;
-			}
-			const auto start = page;
-			while (page < TRACKER_REGION_PAGES && bits.test(page)) {
-				page++;
-			}
-			if (start != page) {
-				func(m_cpu_addr + start * TRACKER_PAGE_SIZE, (page - start) * TRACKER_PAGE_SIZE);
-			}
+		for (const auto [start, end]: bits) {
+			func(m_cpu_addr + start * TRACKER_PAGE_SIZE, (end - start) * TRACKER_PAGE_SIZE);
 		}
 	}
 
@@ -340,8 +203,7 @@ private:
 	RegionBits   m_cpu_dirty;
 	RegionBits   m_gpu_dirty;
 	RegionBits   m_writable;
-	RegionBits   m_fault_pending;
-	RegionBits   m_tracked;
+	RegionBits   m_readable;
 };
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/renderer/cache/bufferCache.cpp

@@ -1,13 +1,13 @@
-#include "graphics/host_gpu/renderer/bufferCache.h"
+#include "graphics/host_gpu/renderer/cache/bufferCache.h"
 
 #include "common/assert.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
 #include "graphics/host_gpu/graphicContext.h"
+#include "graphics/host_gpu/renderer/cache/resourceMutex.h"
+#include "graphics/host_gpu/renderer/cache/textureCache.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
 #include "graphics/host_gpu/renderer/render.h"
-#include "graphics/host_gpu/renderer/resourceMutex.h"
-#include "graphics/host_gpu/renderer/textureCache.h"
 #include "kernel/memory.h"
 
 #include <algorithm>
@@ -123,30 +123,6 @@ struct BufferCache::RetiredBuffer {
 	std::shared_ptr<Buffer> owner;
 };
 
-struct BufferCache::FaultReadback {
-	PageFaultAccess            access = PageFaultAccess::Unknown;
-	uint64_t                   vaddr  = 0;
-	uint64_t                   size   = 0;
-	std::vector<DownloadRange> ranges;
-	bool                       installed = false;
-
-	[[nodiscard]] bool Active() const noexcept { return !ranges.empty(); }
-
-	void Reset() {
-		access    = PageFaultAccess::Unknown;
-		vaddr     = 0;
-		size      = 0;
-		installed = false;
-		ranges.clear();
-	}
-};
-
-struct BufferCache::PendingBackingPublication {
-	uint64_t address = 0;
-	uint64_t size    = 0;
-	uint64_t tick    = 0;
-};
-
 std::pair<uint64_t, uint64_t> BufferCache::DownloadEnvelope(const DownloadCopy& copy) {
 	if (copy.owner == nullptr || copy.size == 0 || copy.source_offset > copy.owner->Size() ||
 	    copy.size > copy.owner->Size() - copy.source_offset) {
@@ -183,9 +159,8 @@ BufferCache::RecordDownloads(std::span<const DownloadCopy> copies) {
 		return {};
 	}
 
-	auto& download = m_download_buffer;
-	const auto [mapped, base_offset] =
-	    download.Map(reservation_size, DOWNLOAD_ALIGNMENT);
+	auto& download                   = m_download_buffer;
+	const auto [mapped, base_offset] = download.Map(reservation_size, DOWNLOAD_ALIGNMENT);
 	if (mapped == nullptr) {
 		EXIT("BufferCache: download batch could not reserve the shared stream\n");
 	}
@@ -215,39 +190,29 @@ void BufferCache::PublishDownloads(std::span<const DownloadRange> downloads) {
 	}
 }
 
-void BufferCache::QueueGarbageDownload(std::span<const DownloadCopy> copies,
-                                       RetiredBuffer retire) {
+void BufferCache::QueueGarbageDownload(std::span<const DownloadCopy> copies, RetiredBuffer retire) {
 	if (copies.empty()) {
 		return;
 	}
-	auto downloads  = RecordDownloads(copies);
-	const auto tick = m_scheduler.CurrentTick();
-	BeginBackingPublication(retire.address, retire.size, tick);
+	auto downloads = RecordDownloads(copies);
 	m_scheduler.DeferOperation(
-	    [this, downloads = std::move(downloads), retire = std::move(retire), tick]() mutable {
+	    [this, downloads = std::move(downloads), retire = std::move(retire)]() mutable {
 		    PublishDownloads(downloads);
 		    {
 			    FaultSafeCacheLock lock(this, m_mutex);
-			    if (m_memory_tracker.IsRegionGpuModified(retire.address, retire.size)) {
-				    m_memory_tracker.ForEachDownloadRange<true>(
-				        retire.address, retire.size,
-				        [&](uint64_t address, uint64_t size) noexcept {
-					        m_memory_tracker.ValidateGpuDirtyPages(
-					            m_gpu_modified_ranges, address, size,
-					            "asynchronous garbage retirement");
-				        },
-				        [](uint64_t, uint64_t) noexcept {});
-			    }
 			    for (const auto& range: downloads) {
 				    m_gpu_modified_ranges.Subtract(range.address, range.size);
 			    }
+			    // ForEachDownloadRange reports full tracker pages, and every exact GPU-owned
+			    // interval on those pages was downloaded and removed. Clearing the original
+			    // query therefore cannot orphan a dirty sibling on an edge page.
+			    m_memory_tracker.UnmarkRegionAsGpuModified(retire.address, retire.size);
 			    if (m_memory_tracker.IsRegionGpuModified(retire.address, retire.size) ||
 			        !m_gpu_modified_ranges.Intersections(retire.address, retire.size).empty()) {
 				    EXIT("BufferCache: asynchronous garbage collection retained GPU ownership\n");
 			    }
 			    m_memory_tracker.UntrackMemory(retire.address, retire.size);
 		    }
-		    CompleteBackingPublication(retire.address, retire.size, tick);
 	    });
 }
 
@@ -256,13 +221,12 @@ BufferCache::BufferCache(GraphicContext& graphics, CommandScheduler& scheduler,
                          ResourceMutex& resource_mutex)
     : m_graphics(graphics), m_scheduler(scheduler),
       m_gds_buffer(graphics, scheduler, MemoryUsage::Stream, 0, AllFlags, GdsBufferSize),
-      m_fault_readback(std::make_unique<FaultReadback>()), m_memory_tracker(page_manager),
+      m_memory_tracker(page_manager),
       m_staging_buffer(graphics, scheduler, MemoryUsage::Upload, 512 * MiB),
       m_stream_buffer(graphics, scheduler, MemoryUsage::Stream, 64 * MiB),
       m_download_buffer(graphics, scheduler, MemoryUsage::Download, 32 * MiB),
       m_device_buffer(graphics, scheduler, MemoryUsage::DeviceLocal, 128 * MiB),
-      m_page_manager(page_manager), m_texture_cache(texture_cache),
-      m_resource_mutex(resource_mutex) {
+      m_texture_cache(texture_cache), m_resource_mutex(resource_mutex) {
 	std::memset(m_gds_buffer.Mapped().data(), 0, static_cast<size_t>(m_gds_buffer.Size()));
 	m_gds_buffer.Flush(0, m_gds_buffer.Size());
 	if (!m_graphics.CanReportMemoryUsage()) {
@@ -280,15 +244,9 @@ BufferCache::BufferCache(GraphicContext& graphics, CommandScheduler& scheduler,
 }
 
 BufferCache::~BufferCache() {
-	if (m_fault_readback->Active()) {
-		EXIT("BufferCache: destroyed with an active fault readback\n");
-	}
 	if (!m_gpu_modified_ranges.Empty()) {
 		EXIT("BufferCache: destroyed with pending GPU-modified ranges\n");
 	}
-	if (!m_pending_backing_publications.empty()) {
-		EXIT("BufferCache: destroyed with pending backing publications\n");
-	}
 	for (const auto& [vaddr, cached]: m_buffers) {
 		(void)vaddr;
 		if (m_memory_tracker.IsRegionGpuModified(cached->vaddr, cached->size)) {
@@ -298,69 +256,6 @@ BufferCache::~BufferCache() {
 	m_buffers.clear();
 }
 
-bool BufferCache::SynchronizeBacking(uint64_t vaddr, uint64_t size) {
-	bool waited = false;
-	for (;;) {
-		uint64_t tick = 0;
-		const auto page_begin = vaddr & ~(TRACKER_PAGE_SIZE - 1);
-		const auto page_end =
-		    (vaddr + size + TRACKER_PAGE_SIZE - 1) & ~(TRACKER_PAGE_SIZE - 1);
-		CacheRange affected {.address = page_begin, .size = page_end - page_begin};
-		{
-			FaultSafeCacheLock lock(this, m_mutex);
-			bool changed = true;
-			while (changed) {
-				changed = false;
-				for (const auto& [address, cached]: m_buffers) {
-					const CacheRange previous = affected;
-					if (ResolveOverlap(affected, {address, cached->size}) &&
-					    (previous.address != affected.address || previous.size != affected.size)) {
-						changed = true;
-					}
-				}
-			}
-		}
-		{
-			std::lock_guard lock(m_publication_mutex);
-			for (const auto& publication: m_pending_backing_publications) {
-				if (publication.address < affected.address + affected.size &&
-				    affected.address < publication.address + publication.size) {
-					tick = std::max(tick, publication.tick);
-				}
-			}
-		}
-		if (tick == 0) {
-			return waited;
-		}
-		waited = true;
-		m_scheduler.Wait(tick);
-		m_scheduler.WaitPriorityOperations(tick);
-	}
-}
-
-void BufferCache::RefreshInvalidatedRanges(CommandBuffer& command, CachedBuffer& cached,
-                                           uint64_t vaddr, uint64_t size, bool upload) {
-	const auto invalidated = m_image_invalidated_ranges.Intersections(vaddr, size);
-	if (upload) {
-		std::array<uint8_t, 64 * 1024> bytes;
-		for (const auto& range: invalidated) {
-			for (uint64_t copied = 0; copied < range.size;) {
-				const auto chunk = std::min<uint64_t>(range.size - copied, bytes.size());
-				if (!Libs::LibKernel::Memory::TryReadBacking(range.address + copied, bytes.data(),
-				                                             chunk)) {
-					EXIT("BufferCache: failed to refresh an invalidated image alias\n");
-				}
-				Upload(command, *cached.buffer, cached.buffer->Offset(range.address + copied),
-				       bytes.data(), chunk);
-				copied += chunk;
-			}
-		}
-	}
-	if (!invalidated.empty()) {
-		m_image_invalidated_ranges.Subtract(vaddr, size);
-	}
-}
-
 StreamBuffer& BufferCache::GetUtilityBuffer(MemoryUsage usage) noexcept {
 	switch (usage) {
 		case MemoryUsage::Upload: return m_staging_buffer;
@@ -384,104 +279,73 @@ BufferBinding BufferCache::UploadTransient(const void* data, uint64_t size, uint
 	return {owner, owner->Handle(), 0};
 }
 
-bool BufferCache::InvalidateMemory(PageFaultAccess access, uint64_t vaddr, uint64_t size,
-                                   PageFaultPhase phase) noexcept {
-	const auto page = vaddr & ~(TRACKER_PAGE_SIZE - 1);
-	if (size == 0 || size > page + TRACKER_PAGE_SIZE - vaddr) {
-		EXIT("BufferCache: invalid page-fault range\n");
+void BufferCache::InvalidateMemory(uint64_t vaddr, uint64_t size) {
+	if (vaddr == 0 || size == 0 || vaddr >= TRACKER_ADDRESS_SIZE ||
+	    size > TRACKER_ADDRESS_SIZE - vaddr) {
+		EXIT("BufferCache: invalid memory-invalidation range\n");
 	}
-
-	if (phase == PageFaultPhase::Complete) {
-		FaultSafeCacheLock lock(this, m_mutex);
-		auto&              fault = *m_fault_readback;
-		if (!fault.Active()) {
-			return m_memory_tracker.CompleteCpuFault(vaddr, size, access, false);
-		}
-		if (fault.access != access || fault.vaddr != vaddr || fault.size != size ||
-		    fault.installed) {
-			EXIT("BufferCache: mismatched fault readback completion\n");
-		}
-		PublishDownloads(fault.ranges);
-		if (!m_memory_tracker.CompleteCpuFault(vaddr, size, access, true)) {
-			EXIT("BufferCache: failed to complete downloaded CPU fault\n");
-		}
-		fault.installed = true;
-		return true;
+	if (!HasPageOverlap(vaddr, size)) {
+		return;
 	}
+	m_memory_tracker.InvalidateRegion(vaddr, size,
+	                                  [this, vaddr, size] { ReadMemory(vaddr, size); });
+}
 
-	if (phase == PageFaultPhase::Release) {
-		FaultSafeCacheLock lock(this, m_mutex);
-		auto&              fault = *m_fault_readback;
-		if (fault.Active()) {
-			if (fault.access != access || fault.vaddr != vaddr || fault.size != size ||
-			    !fault.installed) {
-				EXIT("BufferCache: mismatched fault readback release\n");
-			}
-			for (const auto& range: fault.ranges) {
-				m_gpu_modified_ranges.Subtract(range.address, range.size);
-			}
-			fault.Reset();
-		}
-		return true;
-	}
-
-	if (phase != PageFaultPhase::Invalidate) {
-		EXIT("BufferCache: unsupported page-fault phase\n");
-	}
-
-	const auto action = m_memory_tracker.BeginCpuFault(vaddr, size, access);
-	if (action != CpuFaultAction::Download) {
-		return action == CpuFaultAction::Continue;
-	}
-
-	auto&                     fault = *m_fault_readback;
+void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size) {
 	std::vector<DownloadCopy> copies;
 	{
 		FaultSafeCacheLock lock(this, m_mutex);
-		if (fault.Active()) {
-			EXIT("BufferCache: nested fault readback\n");
-		}
-		fault.access = access;
-		fault.vaddr  = vaddr;
-		fault.size   = size;
-
-		m_gpu_modified_ranges.ForEachIntersection(
-		    page, TRACKER_PAGE_SIZE, [&](RangeSet::Range range) {
-			    auto owner = m_buffers.upper_bound(range.address);
-			    if (owner == m_buffers.begin()) {
-				    EXIT("BufferCache: fault readback has no buffer owner\n");
+		m_memory_tracker.ForEachDownloadRange<false>(
+		    vaddr, size,
+		    [&](uint64_t address, uint64_t bytes) noexcept {
+			    m_memory_tracker.ValidateGpuDirtyPages(m_gpu_modified_ranges, address, bytes,
+			                                           "memory invalidation");
+		    },
+		    [&](uint64_t address, uint64_t bytes) noexcept {
+			    for (const auto range: m_gpu_modified_ranges.Intersections(address, bytes)) {
+				    for (uint64_t copied = 0; copied < range.size;) {
+					    const auto copy_address = range.address + copied;
+					    auto       owner        = m_buffers.upper_bound(copy_address);
+					    if (owner == m_buffers.begin()) {
+						    EXIT("BufferCache: invalidation readback has no buffer owner\n");
+					    }
+					    auto& cached = *std::prev(owner)->second;
+					    if (!cached.buffer->IsInBounds(copy_address, 1)) {
+						    EXIT(
+						        "BufferCache: invalidation readback is outside its buffer owner\n");
+					    }
+					    const auto copy_size = std::min(range.size - copied,
+					                                    cached.vaddr + cached.size - copy_address);
+					    copies.push_back({cached.buffer, cached.buffer->Offset(copy_address),
+					                      copy_address, copy_size});
+					    copied += copy_size;
+				    }
 			    }
-			    --owner;
-			    auto& cached = *owner->second;
-			    if (!cached.buffer->IsInBounds(range.address, range.size)) {
-				    EXIT("BufferCache: fault readback is outside its buffer owner\n");
-			    }
-			    copies.push_back({cached.buffer, cached.buffer->Offset(range.address),
-			                      range.address, range.size});
 		    });
-		if (copies.empty()) {
-			EXIT("BufferCache: GPU-dirty fault page has no dirty byte ranges\n");
-		}
 	}
-	fault.ranges = RecordDownloads(copies);
-	if (!fault.Active()) {
-		EXIT("BufferCache: GPU-dirty fault page has no dirty byte ranges\n");
+	if (copies.empty()) {
+		return;
 	}
+	auto downloads = RecordDownloads(copies);
 	m_scheduler.FinishCurrent();
-	return true;
+	PublishDownloads(downloads);
+	{
+		FaultSafeCacheLock lock(this, m_mutex);
+		for (const auto& range: downloads) {
+			m_gpu_modified_ranges.Subtract(range.address, range.size);
+		}
+		// The enumeration above covered whole dirty pages and every exact interval on them.
+		m_memory_tracker.UnmarkRegionAsGpuModified(vaddr, size);
+	}
 }
 
 void BufferCache::UnmapMemory(uint64_t vaddr, uint64_t size) {
 	if (vaddr == 0 || size == 0 || size > UINT64_MAX - vaddr) {
 		EXIT("BufferCache: invalid unmap range\n");
 	}
-	(void)SynchronizeBacking(vaddr, size);
-
-	std::vector<DownloadCopy>                               copies;
-	std::vector<RangeSet::Range>                            dirty_ranges;
-	std::vector<std::pair<uint64_t, uint64_t>>              modified_buffers;
-	std::vector<std::unique_ptr<PageManager::BackingWrite>> backing_writes;
-	std::vector<std::pair<uint64_t, uint64_t>>              retired_buffers;
+	std::vector<DownloadCopy>                  copies;
+	std::vector<std::pair<uint64_t, uint64_t>> modified_buffers;
+	std::vector<std::pair<uint64_t, uint64_t>> retired_buffers;
 	{
 		FaultSafeCacheLock lock(this, m_mutex);
 		for (const auto& [begin, cached]: m_buffers) {
@@ -498,12 +362,8 @@ void BufferCache::UnmapMemory(uint64_t vaddr, uint64_t size) {
 			if (dirty.empty()) {
 				EXIT("BufferCache: GPU-modified buffer has no dirty ranges\n");
 			}
-			dirty_ranges.insert(dirty_ranges.end(), dirty.begin(), dirty.end());
 			modified_buffers.emplace_back(begin, cached->size);
 		}
-		if (!dirty_ranges.empty()) {
-			backing_writes = m_page_manager.ReserveBackingWrites(dirty_ranges);
-		}
 		for (const auto& [begin, bytes]: modified_buffers) {
 			auto owner = m_buffers.find(begin);
 			if (owner == m_buffers.end() || owner->second->size != bytes) {
@@ -533,24 +393,11 @@ void BufferCache::UnmapMemory(uint64_t vaddr, uint64_t size) {
 		// command stream before removing such backing.
 		m_scheduler.FinishCurrent();
 	}
-	backing_writes.clear();
-
 	{
 		FaultSafeCacheLock lock(this, m_mutex);
-		for (const auto& [begin, bytes]: modified_buffers) {
-			if (!m_memory_tracker.IsRegionGpuModified(begin, bytes)) {
-				continue;
-			}
-			m_memory_tracker.ForEachDownloadRange<true>(
-			    begin, bytes,
-			    [&](uint64_t address, uint64_t download_size) noexcept {
-				    m_memory_tracker.ValidateGpuDirtyPages(
-				        m_gpu_modified_ranges, address, download_size, "unmap retirement");
-			    },
-			    [](uint64_t, uint64_t) noexcept {});
-		}
 		for (const auto& [begin, bytes]: modified_buffers) {
 			m_gpu_modified_ranges.Subtract(begin, bytes);
+			m_memory_tracker.UnmarkRegionAsGpuModified(begin, bytes);
 		}
 		for (const auto& [begin, bytes]: retired_buffers) {
 			m_memory_tracker.MarkRegionAsCpuModified(begin, bytes);
@@ -558,7 +405,6 @@ void BufferCache::UnmapMemory(uint64_t vaddr, uint64_t size) {
 		if (!m_gpu_modified_ranges.Intersections(vaddr, size).empty()) {
 			EXIT("BufferCache: unmap retained dirty byte ranges\n");
 		}
-		m_image_invalidated_ranges.Subtract(vaddr, size);
 		m_memory_tracker.UntrackMemory(vaddr, size);
 		for (auto it = m_buffers.begin(); it != m_buffers.end();) {
 			if (vaddr < it->first + it->second->size && it->first < vaddr + size) {
@@ -651,22 +497,30 @@ BufferBinding BufferCache::ObtainBuffer(CommandBuffer& command, uint64_t vaddr,
 	if (command.IsInvalid() || command.IsExecute()) {
 		EXIT("BufferCache: buffer request requires a recording command buffer\n");
 	}
-	ValidateGpuAccess(vaddr, size, is_read, is_written);
 	std::lock_guard transaction(m_resource_mutex);
-	(void)SynchronizeBacking(vaddr, size);
 
 	if (is_read && !is_written && size <= CACHING_PAGE_SIZE &&
 	    !m_memory_tracker.IsRegionGpuModified(vaddr, size) &&
 	    m_memory_tracker.IsRegionCpuModified(vaddr, size)) {
-		std::vector<uint8_t> data(size);
-		if (Libs::LibKernel::Memory::TryReadBacking(vaddr, data.data(), size)) {
-			return UploadTransient(data.data(), size, 16);
+		const auto alignment = std::max<uint64_t>(
+		    m_graphics.physical_device_properties.limits.minUniformBufferOffsetAlignment, 1);
+		if (auto [mapped, offset] = m_stream_buffer.Map(size, alignment, false);
+		    mapped != nullptr) {
+			if (Libs::LibKernel::Memory::TryReadBacking(vaddr, mapped, size)) {
+				m_stream_buffer.Commit();
+				return {{}, m_stream_buffer.Handle(), offset};
+			}
+		} else {
+			auto owner = std::make_shared<Buffer>(m_graphics, m_scheduler, MemoryUsage::Upload, 0,
+			                                      AllFlags, size);
+			if (Libs::LibKernel::Memory::TryReadBacking(vaddr, owner->Mapped().data(), size)) {
+				owner->Flush(0, size);
+				return {owner, owner->Handle(), 0};
+			}
 		}
 	}
 
-	if (is_formatted && is_read && !is_written) {
-		(void)m_texture_cache.SynchronizeImageToBuffer(vaddr, size);
-	} else if (is_formatted && is_written) {
+	if (is_formatted && is_written) {
 		(void)m_texture_cache.InvalidateMemoryFromGPU(vaddr, size, true);
 	}
 
@@ -682,10 +536,12 @@ BufferBinding BufferCache::ObtainBuffer(CommandBuffer& command, uint64_t vaddr,
 			           reinterpret_cast<const void*>(address), bytes);
 		    }
 	    });
-	RefreshInvalidatedRanges(command, cached, vaddr, size, is_read);
 	if (is_written) {
 		m_gpu_modified_ranges.Add(vaddr, size);
 	}
+	if (is_formatted && is_read && !is_written) {
+		(void)SynchronizeBufferFromImage(*cached.buffer, vaddr, size);
+	}
 	return {cached.buffer, cached.buffer->Handle(), cached.buffer->Offset(vaddr)};
 }
 
@@ -710,7 +566,6 @@ ImageBufferSource BufferCache::ObtainBufferForImage(uint64_t vaddr, uint64_t siz
 	    size > TRACKER_ADDRESS_SIZE - vaddr) {
 		EXIT("BufferCache: invalid image source\n");
 	}
-	(void)SynchronizeBacking(vaddr, size);
 	auto find_owner = [&]() {
 		auto owner = m_buffers.upper_bound(vaddr);
 		if (owner == m_buffers.begin()) {
@@ -722,17 +577,15 @@ ImageBufferSource BufferCache::ObtainBufferForImage(uint64_t vaddr, uint64_t siz
 
 	{
 		FaultSafeCacheLock lock(this, m_mutex);
-		const bool         cpu_modified        = m_memory_tracker.IsRegionCpuModified(vaddr, size);
-		const bool         gpu_modified        = m_memory_tracker.IsRegionGpuModified(vaddr, size);
-		const auto         dirty               = m_gpu_modified_ranges.Intersections(vaddr, size);
-		const bool invalidated =
-		    !m_image_invalidated_ranges.Intersections(vaddr, size).empty();
-		const bool         requested_gpu_owned = !dirty.empty();
+		const bool         cpu_modified = m_memory_tracker.IsRegionCpuModified(vaddr, size);
+		const bool         gpu_modified = m_memory_tracker.IsRegionGpuModified(vaddr, size);
+		const auto         dirty        = m_gpu_modified_ranges.Intersections(vaddr, size);
+		const bool         has_dirty_buffer_source = !dirty.empty();
 		m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size,
 		                                           "image source");
 
 		auto owner = find_owner();
-		if (requested_gpu_owned && owner == m_buffers.end()) {
+		if (has_dirty_buffer_source && owner == m_buffers.end()) {
 			CacheRange merged {.address = AlignDown(vaddr),
 			                   .size    = AlignUp(vaddr + size) - AlignDown(vaddr)};
 			using Iterator = decltype(m_buffers.begin());
@@ -782,42 +635,32 @@ ImageBufferSource BufferCache::ObtainBufferForImage(uint64_t vaddr, uint64_t siz
 				EXIT("BufferCache: merged image source does not contain the requested range\n");
 			}
 		}
-		if (owner != m_buffers.end() && !cpu_modified && !invalidated &&
-		    (!gpu_modified || requested_gpu_owned)) {
-			DiscardGpuDirtyBytesLocked(vaddr, size, "image source transfer");
+		if (owner != m_buffers.end() && !cpu_modified &&
+		    (!gpu_modified || has_dirty_buffer_source)) {
 			owner->second->tick_accessed_last = m_gc_tick;
-			return {owner->second->buffer.get(), owner->second->buffer->Offset(vaddr),
-			        requested_gpu_owned};
+			return {owner->second->buffer.get(), owner->second->buffer->Offset(vaddr)};
 		}
-		if (requested_gpu_owned && owner == m_buffers.end()) {
+		if (has_dirty_buffer_source && owner == m_buffers.end()) {
 			EXIT("BufferCache: GPU-dirty image source could not resolve its native owner\n");
 		}
 	}
 
-	// Direct-memory backing remains readable while PageManager protects the guest mapping. The
-	// fallback exists for plain host mappings used by standalone renderer tests and is deliberately
-	// performed outside the cache lock so a page fault cannot recurse into BufferCache.
-	const auto stage_address = vaddr & ~(TRACKER_PAGE_SIZE - 1);
-	const auto stage_end     = (vaddr + size + TRACKER_PAGE_SIZE - 1) & ~(TRACKER_PAGE_SIZE - 1);
-	const auto stage_size    = stage_end - stage_address;
-	(void)SynchronizeBacking(stage_address, stage_size);
-	std::vector<uint8_t> bytes(stage_size);
-	if (!Libs::LibKernel::Memory::TryReadBacking(stage_address, bytes.data(), stage_size)) {
+	auto [staging, stage_offset] = m_staging_buffer.Map(size, 16);
+	if (staging == nullptr || !Libs::LibKernel::Memory::TryReadBacking(vaddr, staging, size)) {
 		EXIT("BufferCache: failed to read mapped guest image backing\n");
 	}
+	m_staging_buffer.Commit();
 
 	FaultSafeCacheLock lock(this, m_mutex);
-	const auto         dirty               = m_gpu_modified_ranges.Intersections(vaddr, size);
-	const bool invalidated = !m_image_invalidated_ranges.Intersections(vaddr, size).empty();
-	const bool         requested_gpu_owned = !dirty.empty();
-	auto               owner               = find_owner();
-	if (requested_gpu_owned && owner == m_buffers.end()) {
+	const auto         dirty                   = m_gpu_modified_ranges.Intersections(vaddr, size);
+	const bool         has_dirty_buffer_source = !dirty.empty();
+	auto               owner                   = find_owner();
+	if (has_dirty_buffer_source && owner == m_buffers.end()) {
 		EXIT("BufferCache: GPU-dirty image source lost its native owner\n");
 	}
-	const auto stage_offset = m_staging_buffer.Copy(bytes.data(), stage_size, 16);
-	if (owner == m_buffers.end() || invalidated ||
-	    (m_memory_tracker.IsRegionGpuModified(vaddr, size) && !requested_gpu_owned)) {
-		return {&m_staging_buffer, stage_offset + vaddr - stage_address, false};
+	if (owner == m_buffers.end() ||
+	    (m_memory_tracker.IsRegionGpuModified(vaddr, size) && !has_dirty_buffer_source)) {
+		return {&m_staging_buffer, stage_offset};
 	}
 
 	auto& cached              = *owner->second;
@@ -831,43 +674,17 @@ ImageBufferSource BufferCache::ObtainBufferForImage(uint64_t vaddr, uint64_t siz
 	    [&]() noexcept {
 		    for (const auto& [address, upload_size]: uploads) {
 			    cached.buffer->CopyFrom(
-			        m_scheduler.Current(), m_staging_buffer,
-			        stage_offset + address - stage_address, cached.buffer->Offset(address),
-			        upload_size, vk::AccessFlagBits::eHostWrite);
+			        m_scheduler.Current(), m_staging_buffer, stage_offset + address - vaddr,
+			        cached.buffer->Offset(address), upload_size, vk::AccessFlagBits::eHostWrite);
 		    }
 	    });
-	DiscardGpuDirtyBytesLocked(vaddr, size, "staged image source transfer");
-	return {cached.buffer.get(), cached.buffer->Offset(vaddr), requested_gpu_owned};
-}
-
-void BufferCache::DiscardGpuDirtyBytesLocked(uint64_t vaddr, uint64_t size, const char* operation) {
-	m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size, operation);
-	m_gpu_modified_ranges.Subtract(vaddr, size);
-	const auto page_begin = vaddr & ~(TRACKER_PAGE_SIZE - 1);
-	const auto page_end   = (vaddr + size + TRACKER_PAGE_SIZE - 1) & ~(TRACKER_PAGE_SIZE - 1);
-	for (auto page = page_begin; page < page_end; page += TRACKER_PAGE_SIZE) {
-		if (m_gpu_modified_ranges.Intersections(page, TRACKER_PAGE_SIZE).empty() &&
-		    m_memory_tracker.IsRegionGpuModified(page, TRACKER_PAGE_SIZE)) {
-			m_memory_tracker.UnmarkRegionAsGpuModified(page, TRACKER_PAGE_SIZE);
-		}
-	}
-	m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size, operation);
-}
-
-void BufferCache::DiscardGpuDirtyBytes(uint64_t vaddr, uint64_t size) {
-	if (vaddr == 0 || size == 0 || vaddr >= TRACKER_ADDRESS_SIZE ||
-	    size > TRACKER_ADDRESS_SIZE - vaddr) {
-		EXIT("BufferCache: invalid dirty-byte discard range\n");
-	}
-	FaultSafeCacheLock lock(this, m_mutex);
-	DiscardGpuDirtyBytesLocked(vaddr, size, "image output supersession");
+	return {cached.buffer.get(), cached.buffer->Offset(vaddr)};
 }
 
 void BufferCache::WriteHostMemory(uint64_t vaddr, std::span<const uint8_t> data) {
 	if (vaddr == 0 || data.empty() || data.size() > UINT64_MAX - vaddr) {
 		EXIT("BufferCache: invalid host DMA write\n");
 	}
-	(void)SynchronizeBacking(vaddr, data.size());
 	Libs::LibKernel::Memory::WriteBacking(vaddr, data.data(), data.size());
 
 	FaultSafeCacheLock lock(this, m_mutex);
@@ -883,46 +700,6 @@ void BufferCache::WriteHostMemory(uint64_t vaddr, std::span<const uint8_t> data)
 		       data.data() + begin - vaddr, range_end - begin);
 		cached->tick_accessed_last = m_gc_tick;
 	}
-	m_image_invalidated_ranges.Subtract(vaddr, data.size());
-}
-
-std::pair<std::shared_ptr<Buffer>, uint64_t> BufferCache::ObtainBufferForImageWrite(uint64_t vaddr,
-                                                                                    uint64_t size) {
-	if (vaddr == 0 || size == 0 || vaddr >= TRACKER_ADDRESS_SIZE ||
-	    size > TRACKER_ADDRESS_SIZE - vaddr) {
-		EXIT("BufferCache: invalid image destination\n");
-	}
-	const auto stage_address = vaddr & ~(TRACKER_PAGE_SIZE - 1);
-	const auto stage_end     = (vaddr + size + TRACKER_PAGE_SIZE - 1) & ~(TRACKER_PAGE_SIZE - 1);
-	const auto stage_size    = stage_end - stage_address;
-	(void)SynchronizeBacking(stage_address, stage_size);
-	std::vector<uint8_t> bytes(stage_size);
-	if (!Libs::LibKernel::Memory::TryReadBacking(stage_address, bytes.data(), stage_size)) {
-		EXIT("BufferCache: failed to preserve guest bytes around an image mirror\n");
-	}
-	FaultSafeCacheLock lock(this, m_mutex);
-	auto&              cached = GetOrCreateBuffer(m_scheduler.Current(), vaddr, size);
-	m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size,
-	                                           "image destination");
-	if (!m_gpu_modified_ranges.Intersections(vaddr, size).empty()) {
-		EXIT("BufferCache: image destination aliases GPU-owned buffer bytes\n");
-	}
-	const auto stage_offset = m_staging_buffer.Copy(bytes.data(), stage_size, 16);
-	std::vector<std::pair<uint64_t, uint64_t>> uploads;
-	m_memory_tracker.ForEachUploadRange(
-	    vaddr, size, false,
-	    [&](uint64_t address, uint64_t upload_size) noexcept {
-		    uploads.emplace_back(address, upload_size);
-	    },
-	    [&]() noexcept {
-		    for (const auto& [address, upload_size]: uploads) {
-			    cached.buffer->CopyFrom(
-			        m_scheduler.Current(), m_staging_buffer,
-			        stage_offset + address - stage_address, cached.buffer->Offset(address),
-			        upload_size, vk::AccessFlagBits::eHostWrite);
-		    }
-	    });
-	return {cached.buffer, cached.buffer->Offset(vaddr)};
 }
 
 void BufferCache::FillBuffer(uint64_t vaddr, uint64_t size, uint32_t value, bool is_gds) {
@@ -939,19 +716,18 @@ void BufferCache::FillBuffer(uint64_t vaddr, uint64_t size, uint32_t value, bool
 	if (vaddr == 0) {
 		EXIT("BufferCache: invalid fill memory address\n");
 	}
-	ValidateGpuAccess(vaddr, size, false, true);
 	(void)m_texture_cache.ClearMeta(vaddr);
 	{
 		std::lock_guard transaction(m_resource_mutex);
 		const auto      region = m_texture_cache.QueryRegion(vaddr, size);
 		if (!HasGpuDirtyBytes(vaddr, size) && !region.gpu_image_bytes) {
 			if (region.image_bytes) {
-				m_texture_cache.PrepareHostWrite(vaddr, size);
+				m_texture_cache.InvalidateMemory(vaddr, size);
 			}
 			std::array<uint32_t, 4096> values;
 			values.fill(value);
-			const std::span<const uint8_t> bytes {
-			    reinterpret_cast<const uint8_t*>(values.data()), sizeof(values)};
+			const std::span<const uint8_t> bytes {reinterpret_cast<const uint8_t*>(values.data()),
+			                                      sizeof(values)};
 			for (uint64_t offset = 0; offset < size;) {
 				const auto chunk = std::min<uint64_t>(size - offset, bytes.size());
 				WriteHostMemory(vaddr + offset, bytes.first(chunk));
@@ -981,30 +757,17 @@ void BufferCache::CopyBuffer(uint64_t dst_vaddr, uint64_t src_vaddr, uint64_t si
 	    (src_gds && (src_vaddr > m_gds_buffer.Size() || size > m_gds_buffer.Size() - src_vaddr))) {
 		EXIT("BufferCache: invalid or overlapping copy range\n");
 	}
-	if (src_memory) {
-		ValidateGpuAccess(src_vaddr, size, true, false);
-	}
-	if (dst_memory) {
-		ValidateGpuAccess(dst_vaddr, size, false, true);
-	}
 	if (src_memory || dst_memory) {
 		std::lock_guard transaction(m_resource_mutex);
-		if (src_memory) {
-			(void)SynchronizeBacking(src_vaddr, size);
-		}
 		const auto      src_region =
 		    src_memory ? m_texture_cache.QueryRegion(src_vaddr, size) : TextureCache::RegionInfo {};
 		const auto dst_region =
 		    dst_memory ? m_texture_cache.QueryRegion(dst_vaddr, size) : TextureCache::RegionInfo {};
-		if (src_memory && src_region.gpu_image_bytes &&
-		    !m_texture_cache.SynchronizeImageToBuffer(src_vaddr, size)) {
-			EXIT("BufferCache: GPU copy source image could not be synchronized\n");
-		}
 		if (src_memory && dst_memory && !HasGpuDirtyBytes(src_vaddr, size) &&
 		    !HasGpuDirtyBytes(dst_vaddr, size) && !src_region.gpu_image_bytes &&
 		    !dst_region.gpu_image_bytes) {
 			if (dst_region.image_bytes) {
-				m_texture_cache.PrepareHostWrite(dst_vaddr, size);
+				m_texture_cache.InvalidateMemory(dst_vaddr, size);
 			}
 			std::array<uint8_t, 64 * 1024> bytes;
 			for (uint64_t offset = 0; offset < size;) {
@@ -1021,7 +784,7 @@ void BufferCache::CopyBuffer(uint64_t dst_vaddr, uint64_t src_vaddr, uint64_t si
 	}
 
 	auto& command = m_scheduler.Current();
-	auto  src = src_memory ? ObtainBuffer(command, src_vaddr, size, false, true)
+	auto  src = src_memory ? ObtainBuffer(command, src_vaddr, size, false, true, true)
 	                       : BufferBinding {.buffer = m_gds_buffer.Handle(), .offset = src_vaddr};
 	auto  dst = dst_memory ? ObtainBuffer(command, dst_vaddr, size, true, false, true)
 	                       : BufferBinding {.buffer = m_gds_buffer.Handle(), .offset = dst_vaddr};
@@ -1037,10 +800,10 @@ void BufferCache::CopyBuffer(uint64_t dst_vaddr, uint64_t src_vaddr, uint64_t si
 		EXIT("BufferCache: resolved Vulkan copy ranges overlap\n");
 	}
 	auto& source = src.owner != nullptr ? *std::static_pointer_cast<Buffer>(src.owner)
-	               : src_gds           ? m_gds_buffer
-	                                   : m_stream_buffer;
-	auto& destination = dst.owner != nullptr ? *std::static_pointer_cast<Buffer>(dst.owner)
-	                                         : m_gds_buffer;
+	               : src_gds            ? m_gds_buffer
+	                                    : m_stream_buffer;
+	auto& destination =
+	    dst.owner != nullptr ? *std::static_pointer_cast<Buffer>(dst.owner) : m_gds_buffer;
 	if (source.Handle() != src.buffer || destination.Handle() != dst.buffer) {
 		EXIT("BufferCache: resolved copy owner does not match its Vulkan handle\n");
 	}
@@ -1074,95 +837,13 @@ bool BufferCache::IsRegionCpuModified(uint64_t vaddr, uint64_t size) {
 	return m_memory_tracker.IsRegionCpuModified(vaddr, size);
 }
 
-void BufferCache::InvalidateImageAliases(uint64_t vaddr, uint64_t size) {
-	if (vaddr == 0 || size == 0 || vaddr >= TRACKER_ADDRESS_SIZE ||
-	    size > TRACKER_ADDRESS_SIZE - vaddr) {
-		EXIT("BufferCache: invalid image-alias invalidation\n");
-	}
-	FaultSafeCacheLock lock(this, m_mutex);
-	const auto         end = vaddr + size;
-	for (const auto& [address, cached]: m_buffers) {
-		const auto cached_end = address + cached->size;
-		const auto begin      = std::max(vaddr, address);
-		const auto range_end  = std::min(end, cached_end);
-		if (begin >= range_end) {
-			continue;
-		}
-		const auto bytes = range_end - begin;
-		if (!m_gpu_modified_ranges.Intersections(begin, bytes).empty()) {
-			EXIT("BufferCache: image ownership overlaps exact dirty buffer bytes\n");
-		}
-		m_image_invalidated_ranges.Add(begin, bytes);
-	}
-}
-
-void BufferCache::BeginBackingPublication(uint64_t vaddr, uint64_t size, uint64_t tick) {
-	if (vaddr == 0 || size == 0 || tick == 0 || vaddr >= TRACKER_ADDRESS_SIZE ||
-	    size > TRACKER_ADDRESS_SIZE - vaddr) {
-		EXIT("BufferCache: invalid pending backing publication\n");
-	}
-	std::lock_guard lock(m_publication_mutex);
-	m_pending_backing_publications.push_back({vaddr, size, tick});
-}
-
-void BufferCache::CompleteBackingPublication(uint64_t vaddr, uint64_t size, uint64_t tick) {
-	std::lock_guard lock(m_publication_mutex);
-	const auto publication =
-	    std::ranges::find_if(m_pending_backing_publications, [&](const auto& pending) {
-		    return pending.address == vaddr && pending.size == size && pending.tick == tick;
-	    });
-	if (publication == m_pending_backing_publications.end()) {
-		EXIT("BufferCache: completed an unknown backing publication\n");
-	}
-	m_pending_backing_publications.erase(publication);
-}
-
-void BufferCache::PublishImageBuffer(uint64_t vaddr, uint64_t size) {
-	FaultSafeCacheLock lock(this, m_mutex);
-	auto               owner = m_buffers.end();
-	for (auto it = m_buffers.begin(); it != m_buffers.end(); ++it) {
-		if (!PageOverlaps(vaddr, size, it->second->vaddr, it->second->size)) {
-			continue;
-		}
-		if (owner != m_buffers.end() || !it->second->buffer->IsInBounds(vaddr, size)) {
-			EXIT("BufferCache: image destination aliases a non-containing cached buffer\n");
-		}
-		owner = it;
-	}
-	m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size,
-	                                           "image destination publication");
-	if (owner == m_buffers.end() || m_memory_tracker.IsRegionCpuModified(vaddr, size) ||
-	    !m_gpu_modified_ranges.Intersections(vaddr, size).empty()) {
-		EXIT("BufferCache: image destination requires clean buffer ownership\n");
-	}
-	m_memory_tracker.MarkRegionAsGpuModified(vaddr, size);
-	m_gpu_modified_ranges.Add(vaddr, size);
-	m_image_invalidated_ranges.Subtract(vaddr, size);
-	m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size,
-	                                           "published image destination");
-	owner->second->tick_accessed_last = m_gc_tick;
-}
-
-void BufferCache::ValidateGpuAccess(uint64_t vaddr, uint64_t size, bool is_read,
-                                    bool is_written) const {
-	if ((!is_read && !is_written) || vaddr == 0 || size == 0 || size > UINT64_MAX - vaddr) {
-		EXIT("BufferCache: invalid GPU access request\n");
-	}
-	if (is_read && !m_page_manager.HasGpuAccess(vaddr, size, GpuAccess::Read)) {
-		EXIT("BufferCache: GPU-read access denied\n");
-	}
-	if (is_written && !m_page_manager.HasGpuAccess(vaddr, size, GpuAccess::Write)) {
-		EXIT("BufferCache: GPU-write access denied\n");
-	}
-}
-
 void BufferCache::RunGarbageCollector() {
 	std::lock_guard transaction(m_resource_mutex);
 	const auto      tick = m_gc_tick++;
 	if (m_graphics.CanReportMemoryUsage()) {
 		m_total_used_memory = m_graphics.GetDeviceMemoryUsage();
 	}
-	if (m_total_used_memory < m_trigger_gc_memory || m_fault_readback->Active()) {
+	if (m_total_used_memory < m_trigger_gc_memory) {
 		return;
 	}
 
@@ -1170,7 +851,7 @@ void BufferCache::RunGarbageCollector() {
 	const uint64_t age        = std::min<uint64_t>(aggressive ? 80 : 160, tick);
 	const size_t   limit      = aggressive ? 64 : 32;
 
-	std::vector<RetiredBuffer> retires;
+	std::vector<RetiredBuffer>                                       retires;
 	std::vector<std::pair<RetiredBuffer, std::vector<DownloadCopy>>> dirty_retires;
 	{
 		FaultSafeCacheLock    lock(this, m_mutex);
@@ -1190,8 +871,8 @@ void BufferCache::RunGarbageCollector() {
 		}
 		for (const auto address: candidates) {
 			auto& cached = *m_buffers.at(address);
-			m_memory_tracker.ValidateGpuDirtyOwnership(
-			    m_gpu_modified_ranges, cached.vaddr, cached.size, "garbage collection");
+			m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, cached.vaddr,
+			                                           cached.size, "garbage collection");
 			retires.push_back({address, cached.size, cached.buffer});
 			// GC runs immediately before submission. Preserve every source referenced by commands
 			// already recorded in the active batch.
@@ -1205,13 +886,13 @@ void BufferCache::RunGarbageCollector() {
 			m_memory_tracker.ForEachDownloadRange<false>(
 			    retire.address, retire.size,
 			    [&](uint64_t address, uint64_t size) noexcept {
-				    m_memory_tracker.ValidateGpuDirtyPages(
-				        m_gpu_modified_ranges, address, size, "garbage collection");
+				    m_memory_tracker.ValidateGpuDirtyPages(m_gpu_modified_ranges, address, size,
+				                                           "garbage collection");
 			    },
 			    [&](uint64_t address, uint64_t size) noexcept {
 				    for (const auto range: m_gpu_modified_ranges.Intersections(address, size)) {
-					    copies.push_back({retire.owner, range.address - retire.address, range.address,
-					                      range.size});
+					    copies.push_back({retire.owner, range.address - retire.address,
+					                      range.address, range.size});
 				    }
 			    });
 		}
@@ -1231,7 +912,6 @@ void BufferCache::RunGarbageCollector() {
 		if (!m_memory_tracker.IsRegionGpuModified(retire.address, retire.size)) {
 			m_memory_tracker.UntrackMemory(retire.address, retire.size);
 		}
-		m_image_invalidated_ranges.Subtract(retire.address, retire.size);
 		if (retire.size > m_total_used_memory) {
 			EXIT("BufferCache: allocation accounting underflow\n");
 		}

src/graphics/host_gpu/renderer/cache/bufferCache.h

@@ -6,11 +6,10 @@
 #include "common/threads.h"
 #include "graphics/host_gpu/memoryTracker.h"
 #include "graphics/host_gpu/rangeSet.h"
-#include "graphics/host_gpu/renderer/streamBuffer.h"
+#include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 
 #include <map>
 #include <memory>
-#include <mutex>
 #include <span>
 #include <utility>
 #include <vector>
@@ -30,9 +29,8 @@ struct BufferBinding {
 };
 
 struct ImageBufferSource {
-	Buffer*  buffer    = nullptr;
-	uint64_t offset    = 0;
-	bool     gpu_owned = false;
+	Buffer*  buffer = nullptr;
+	uint64_t offset = 0;
 };
 
 class BufferCache {
@@ -47,9 +45,9 @@ public:
 	~BufferCache();
 	KYTY_CLASS_NO_COPY(BufferCache);
 
-	[[nodiscard]] bool InvalidateMemory(PageFaultAccess access, uint64_t vaddr, uint64_t size,
-	                                    PageFaultPhase phase) noexcept;
-	void               UnmapMemory(uint64_t vaddr, uint64_t size);
+	void                        InvalidateMemory(uint64_t vaddr, uint64_t size);
+	void                        ReadMemory(uint64_t vaddr, uint64_t size);
+	void                        UnmapMemory(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] BufferBinding ObtainBuffer(CommandBuffer& command, uint64_t vaddr, uint64_t size,
 	                                         bool is_written = false, bool is_read = true,
 	                                         bool is_formatted = false);
@@ -60,9 +58,6 @@ public:
 	                                            uint64_t alignment);
 	[[nodiscard]] std::shared_ptr<Buffer> ObtainNullBuffer();
 	[[nodiscard]] ImageBufferSource       ObtainBufferForImage(uint64_t vaddr, uint64_t size);
-	[[nodiscard]] std::pair<std::shared_ptr<Buffer>, uint64_t>
-	     ObtainBufferForImageWrite(uint64_t vaddr, uint64_t size);
-	void DiscardGpuDirtyBytes(uint64_t vaddr, uint64_t size);
 	void FillBuffer(uint64_t vaddr, uint64_t size, uint32_t value, bool is_gds = false);
 	void CopyBuffer(uint64_t dst_vaddr, uint64_t src_vaddr, uint64_t size, bool dst_gds = false,
 	                bool src_gds = false);
@@ -70,13 +65,7 @@ public:
 	[[nodiscard]] bool HasGpuDirtyBytes(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] bool IsRegionCpuModified(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] bool IsRegionGpuModified(uint64_t vaddr, uint64_t size);
-	void               InvalidateImageAliases(uint64_t vaddr, uint64_t size);
-	void BeginBackingPublication(uint64_t vaddr, uint64_t size, uint64_t tick);
-	void CompleteBackingPublication(uint64_t vaddr, uint64_t size, uint64_t tick);
-	[[nodiscard]] bool SynchronizeBacking(uint64_t vaddr, uint64_t size);
-	void               PublishImageBuffer(uint64_t vaddr, uint64_t size);
-	void ValidateGpuAccess(uint64_t vaddr, uint64_t size, bool is_read, bool is_written) const;
-	void RunGarbageCollector();
+	void               RunGarbageCollector();
 
 private:
 	friend struct BufferCacheTestAccess;
@@ -89,30 +78,24 @@ private:
 	struct DownloadCopy;
 	struct DownloadRange;
 	struct RetiredBuffer;
-	struct FaultReadback;
-	struct PendingBackingPublication;
-	static constexpr uint64_t DOWNLOAD_ALIGNMENT = 64;
-	[[nodiscard]] static uint64_t AlignDown(uint64_t value) noexcept;
-	[[nodiscard]] static uint64_t AlignUp(uint64_t value);
+	static constexpr uint64_t               DOWNLOAD_ALIGNMENT = 64;
+	[[nodiscard]] static uint64_t           AlignDown(uint64_t value) noexcept;
+	[[nodiscard]] static uint64_t           AlignUp(uint64_t value);
 	[[nodiscard]] static constexpr uint64_t AlignDownload(uint64_t size) noexcept {
 		return (size + DOWNLOAD_ALIGNMENT - 1) & ~(DOWNLOAD_ALIGNMENT - 1);
 	}
 	[[nodiscard]] static bool PageOverlaps(uint64_t left, uint64_t left_size, uint64_t right,
-	                                      uint64_t right_size) noexcept;
-	[[nodiscard]] static std::pair<uint64_t, uint64_t>
-	     DownloadEnvelope(const DownloadCopy& copy);
-	[[nodiscard]] static bool   ResolveOverlap(CacheRange& merged, CacheRange candidate) noexcept;
+	                                       uint64_t right_size) noexcept;
+	[[nodiscard]] static std::pair<uint64_t, uint64_t> DownloadEnvelope(const DownloadCopy& copy);
+	[[nodiscard]] static bool ResolveOverlap(CacheRange& merged, CacheRange candidate) noexcept;
 	void Upload(CommandBuffer& command, Buffer& destination, uint64_t destination_offset,
 	            const void* source, uint64_t size);
 	[[nodiscard]] CachedBuffer& GetOrCreateBuffer(CommandBuffer& command, uint64_t vaddr,
 	                                              uint64_t size);
-	[[nodiscard]] std::vector<DownloadRange>
-	     RecordDownloads(std::span<const DownloadCopy> copies);
+	[[nodiscard]] bool SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uint64_t size);
+	[[nodiscard]] std::vector<DownloadRange> RecordDownloads(std::span<const DownloadCopy> copies);
 	void PublishDownloads(std::span<const DownloadRange> downloads);
 	void QueueGarbageDownload(std::span<const DownloadCopy> copies, RetiredBuffer retire);
-	void RefreshInvalidatedRanges(CommandBuffer& command, CachedBuffer& cached, uint64_t vaddr,
-	                              uint64_t size, bool upload);
-	void DiscardGpuDirtyBytesLocked(uint64_t vaddr, uint64_t size, const char* operation);
 	void WriteHostMemory(uint64_t vaddr, std::span<const uint8_t> data);
 
 	GraphicContext&                                   m_graphics;
@@ -121,17 +104,12 @@ private:
 	Common::Mutex                                     m_mutex;
 	std::shared_ptr<Buffer>                           m_null_buffer;
 	std::map<uint64_t, std::unique_ptr<CachedBuffer>> m_buffers;
-	std::unique_ptr<FaultReadback>                    m_fault_readback;
 	RangeSet                                          m_gpu_modified_ranges;
-	RangeSet                                          m_image_invalidated_ranges;
-	std::mutex                                        m_publication_mutex;
-	std::vector<PendingBackingPublication>            m_pending_backing_publications;
 	MemoryTracker                                     m_memory_tracker;
 	StreamBuffer                                      m_staging_buffer;
 	StreamBuffer                                      m_stream_buffer;
 	StreamBuffer                                      m_download_buffer;
 	StreamBuffer                                      m_device_buffer;
-	PageManager&                                      m_page_manager;
 	TextureCache&                                     m_texture_cache;
 	ResourceMutex&                                    m_resource_mutex;
 	uint64_t                                          m_total_used_memory = 0;

src/graphics/host_gpu/renderer/cache/gpuResourceManager.cpp

@@ -0,0 +1,140 @@
+#include "graphics/host_gpu/renderer/cache/gpuResourceManager.h"
+
+#include "common/assert.h"
+#include "graphics/guest_gpu/command_processor/commandProcessor.h"
+#include "graphics/guest_gpu/graphicsRun.h"
+#include "graphics/host_gpu/renderer/commandScheduler.h"
+namespace Libs::Graphics {
+
+GpuResourceManager::GpuResourceManager(GraphicContext& graphics, CommandScheduler& scheduler)
+    : m_scheduler(scheduler),
+      m_buffer_cache(graphics, scheduler, m_page_manager, m_texture_cache, m_resource_mutex),
+      m_texture_cache(graphics, scheduler, m_page_manager, m_buffer_cache, m_resource_mutex) {}
+
+GpuResourceManager::~GpuResourceManager() = default;
+
+bool GpuResourceManager::HandleFault(PageFaultAccess access, uint64_t fault_vaddr) noexcept {
+	constexpr uint64_t fault_size = 8;
+	if (!IsMapped(fault_vaddr, fault_size)) {
+		return false;
+	}
+	if (CommandScheduler::InDeferredOperation()) {
+		EXIT("unsupported guest-memory fault from an asynchronous GPU completion, "
+		     "addr=0x%016" PRIx64 " access=%u\n",
+		     fault_vaddr, static_cast<uint32_t>(access));
+	}
+	bool       handled = false;
+	const auto resolve = [this, access, fault_vaddr, &handled](CommandProcessor& cp) {
+		cp.BeginReadbackTransaction();
+		{
+			ResourceMutex::FaultScope fault(m_resource_mutex);
+			if (access == PageFaultAccess::Write) {
+				m_buffer_cache.InvalidateMemory(fault_vaddr, fault_size);
+				m_texture_cache.InvalidateMemory(fault_vaddr, fault_size);
+			} else {
+				m_buffer_cache.ReadMemory(fault_vaddr, fault_size);
+			}
+			handled = true;
+		}
+		cp.EndReadbackTransaction();
+	};
+	if (auto* cp = Gpu::CurrentCommandProcessor(); cp != nullptr) {
+		resolve(*cp);
+		return handled;
+	}
+	if (m_resource_mutex.IsOwnedByCurrentThread()) {
+		EXIT("unsupported page fault from a pre-owned resource transaction, addr=0x%016" PRIx64
+		     " access=%u\n",
+		     fault_vaddr, static_cast<uint32_t>(access));
+	}
+	EXIT_IF(m_gpu == nullptr);
+	m_gpu->SendCommandSyncWithProcessor(resolve);
+	return handled;
+}
+
+bool GpuResourceManager::InvalidateMemory(uint64_t vaddr, uint64_t size) {
+	if (!IsMapped(vaddr, size)) {
+		return false;
+	}
+	if (CommandScheduler::InDeferredOperation()) {
+		EXIT("unsupported memory invalidation from an asynchronous GPU completion, "
+		     "addr=0x%016" PRIx64 " size=0x%016" PRIx64 "\n",
+		     vaddr, size);
+	}
+	const auto resolve = [this, vaddr, size](CommandProcessor& cp) {
+		cp.BeginReadbackTransaction();
+		{
+			ResourceMutex::FaultScope fault(m_resource_mutex);
+			m_buffer_cache.InvalidateMemory(vaddr, size);
+			m_texture_cache.InvalidateMemory(vaddr, size);
+		}
+		cp.EndReadbackTransaction();
+	};
+	if (auto* cp = Gpu::CurrentCommandProcessor(); cp != nullptr) {
+		resolve(*cp);
+		return true;
+	}
+	if (m_resource_mutex.IsOwnedByCurrentThread()) {
+		EXIT("unsupported memory invalidation from a pre-owned resource transaction, "
+		     "addr=0x%016" PRIx64 " size=0x%016" PRIx64 "\n",
+		     vaddr, size);
+	}
+	EXIT_IF(m_gpu == nullptr);
+	m_gpu->SendCommandSyncWithProcessor(resolve);
+	return true;
+}
+
+bool GpuResourceManager::IsMapped(uint64_t vaddr, uint64_t size) const noexcept {
+	if (vaddr == 0 || size == 0 || vaddr >= TRACKER_ADDRESS_SIZE ||
+	    size > TRACKER_ADDRESS_SIZE - vaddr) {
+		return false;
+	}
+	std::shared_lock lock(m_mapped_ranges_mutex);
+	return m_mapped_ranges.Contains(vaddr, size);
+}
+
+void GpuResourceManager::MapMemory(uint64_t vaddr, uint64_t size) {
+	{
+		std::lock_guard lock(m_mapped_ranges_mutex);
+		m_mapped_ranges.Add(vaddr, size);
+	}
+	m_page_manager.OnGpuMap(vaddr, size);
+}
+
+void GpuResourceManager::UnmapMemory(uint64_t vaddr, uint64_t size) {
+	if (CommandScheduler::InDeferredOperation()) {
+		EXIT("unsupported memory unmap from an asynchronous GPU completion, "
+		     "addr=0x%016" PRIx64 " size=0x%016" PRIx64 "\n",
+		     vaddr, size);
+	}
+	if (m_resource_mutex.IsOwnedByCurrentThread()) {
+		EXIT("unsupported memory unmap from a pre-owned resource transaction, "
+		     "addr=0x%016" PRIx64 " size=0x%016" PRIx64 "\n",
+		     vaddr, size);
+	}
+	const auto unmap = [this, vaddr, size] {
+		if (m_scheduler.Active()) {
+			const auto tick = m_scheduler.CurrentTick();
+			m_scheduler.FinishCurrent();
+			m_scheduler.WaitPriorityOperations(tick);
+		}
+		m_buffer_cache.UnmapMemory(vaddr, size);
+		m_texture_cache.UnmapMemory(vaddr, size);
+		m_page_manager.OnGpuUnmap(vaddr, size);
+		std::lock_guard lock(m_mapped_ranges_mutex);
+		m_mapped_ranges.Subtract(vaddr, size);
+	};
+	if (m_gpu == nullptr) {
+		unmap();
+		return;
+	}
+	m_gpu->SendCommandSync(unmap);
+}
+
+void GpuResourceManager::RunGarbageCollector() {
+	m_texture_cache.ProcessDownloadImages();
+	m_texture_cache.RunGarbageCollector();
+	m_buffer_cache.RunGarbageCollector();
+}
+
+} // namespace Libs::Graphics

src/graphics/host_gpu/renderer/cache/gpuResourceManager.h

@@ -4,11 +4,12 @@
 #include "common/abi.h"
 #include "common/common.h"
 #include "graphics/host_gpu/pageManager.h"
-#include "graphics/host_gpu/renderer/bufferCache.h"
-#include "graphics/host_gpu/renderer/resourceMutex.h"
-#include "graphics/host_gpu/renderer/textureCache.h"
+#include "graphics/host_gpu/renderer/cache/bufferCache.h"
+#include "graphics/host_gpu/renderer/cache/resourceMutex.h"
+#include "graphics/host_gpu/renderer/cache/textureCache.h"
 
 #include <cstdint>
+#include <shared_mutex>
 
 namespace Libs::Graphics {
 
@@ -26,23 +27,21 @@ public:
 	void                        SetGpu(Gpu* gpu) noexcept { m_gpu = gpu; }
 
 	[[nodiscard]] bool HandleFault(PageFaultAccess access, uint64_t fault_vaddr) noexcept;
-	void               PrepareHostWrite(uint64_t vaddr, uint64_t size);
+	[[nodiscard]] bool InvalidateMemory(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] bool IsMapped(uint64_t vaddr, uint64_t size) const noexcept;
-	void               MapMemory(uint64_t vaddr, uint64_t size, GpuAccess access);
-	void               UnmapMemory(uint64_t vaddr, uint64_t size, GpuAccess access);
+	void               MapMemory(uint64_t vaddr, uint64_t size);
+	void               UnmapMemory(uint64_t vaddr, uint64_t size);
 	void               RunGarbageCollector();
 
 private:
-	static bool FaultThunk(void* context, PageFaultAccess access, uint64_t vaddr, uint64_t size,
-	                       PageFaultPhase phase) noexcept;
-	[[nodiscard]] bool InvalidateMemory(PageFaultAccess access, uint64_t vaddr, uint64_t size,
-	                                    PageFaultPhase phase) noexcept;
-
-	PageManager   m_page_manager;
-	ResourceMutex m_resource_mutex;
-	BufferCache   m_buffer_cache;
-	TextureCache  m_texture_cache;
-	Gpu*          m_gpu = nullptr;
+	PageManager               m_page_manager;
+	ResourceMutex             m_resource_mutex;
+	CommandScheduler&         m_scheduler;
+	BufferCache               m_buffer_cache;
+	TextureCache              m_texture_cache;
+	mutable std::shared_mutex m_mapped_ranges_mutex;
+	RangeSet                  m_mapped_ranges;
+	Gpu*                      m_gpu = nullptr;
 };
 
 } // namespace Libs::Graphics

src/graphics/host_gpu/renderer/cache/multiLevelPageTable.h

@@ -219,7 +219,7 @@ private:
 		typename CoarseTable::PageRange   coarse_range {};
 		typename TrackingTable::PageRange tracking_range {};
 		if (!CoarseTable::TryGetPageRange(address, size, coarse_range) ||
-		    !TrackingTable::TryGetPageRange(address, size, tracking_range)) {
+		    (!strict_bytes && !TrackingTable::TryGetPageRange(address, size, tracking_range))) {
 			return {};
 		}
 		MembershipList candidates;
@@ -230,8 +230,8 @@ private:
 		}
 		std::vector<OwnerT> result;
 		for (const Registration* registration: candidates) {
-			if ((!strict_bytes || Overlaps(registration->ranges, address, size)) &&
-			    HasTrackingMembership(registration, tracking_range) &&
+			if ((strict_bytes ? Overlaps(registration->ranges, address, size)
+			                  : HasTrackingMembership(registration, tracking_range)) &&
 			    predicate(registration->owner)) {
 				result.push_back(registration->owner);
 			}

src/graphics/host_gpu/renderer/cache/resourceMutex.cpp

@@ -1,4 +1,4 @@
-#include "graphics/host_gpu/renderer/resourceMutex.h"
+#include "graphics/host_gpu/renderer/cache/resourceMutex.h"
 
 #include "common/assert.h"
 

src/graphics/host_gpu/renderer/cache/samplerCache.cpp

@@ -1,4 +1,4 @@
-#include "graphics/host_gpu/renderer/samplerCache.h"
+#include "graphics/host_gpu/renderer/cache/samplerCache.h"
 
 #include "common/assert.h"
 #include "common/logging/log.h"
@@ -56,7 +56,10 @@ vk::Sampler SamplerCache::GetSampler(const ShaderSamplerResource& r) {
 			case Prospero::SamplerAnisoRatio::kFour: aniso_ratio = 4.0f; break;
 			case Prospero::SamplerAnisoRatio::kEight: aniso_ratio = 8.0f; break;
 			case Prospero::SamplerAnisoRatio::kSixteen: aniso_ratio = 16.0f; break;
-			default: EXIT("unknown ratio: %d\n", static_cast<int>(r.MaxAnisoRatio()));
+			default:
+				EXIT("unknown ratio: %d dwords=%08x,%08x,%08x,%08x\n",
+				     static_cast<int>(r.MaxAnisoRatio()), r.fields[0], r.fields[1], r.fields[2],
+				     r.fields[3]);
 		}
 	}
 

src/graphics/host_gpu/renderer/cache/streamBuffer.cpp

@@ -1,4 +1,4 @@
-#include "graphics/host_gpu/renderer/streamBuffer.h"
+#include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 
 #include "common/assert.h"
 #include "common/profiler.h"
@@ -135,8 +135,8 @@ void Buffer::Write(uint64_t offset, const void* source, uint64_t size) {
 void Buffer::Flush(uint64_t offset, uint64_t size) {
 	EXIT_IF(m_mapped.empty() || offset > m_size || size > m_size - offset);
 	if (!m_is_coherent && size != 0) {
-		const auto result = vmaFlushAllocation(m_graphics->allocator, m_buffer->memory.allocation,
-		                                      offset, size);
+		const auto result =
+		    vmaFlushAllocation(m_graphics->allocator, m_buffer->memory.allocation, offset, size);
 		EXIT_NOT_IMPLEMENTED(static_cast<vk::Result>(result) != vk::Result::eSuccess);
 	}
 }
@@ -144,8 +144,8 @@ void Buffer::Flush(uint64_t offset, uint64_t size) {
 vk::BufferMemoryBarrier Buffer::Barrier(uint64_t offset, uint64_t size, vk::AccessFlags source,
                                         vk::AccessFlags destination) const {
 	if (Handle() == nullptr || size == 0 || offset > m_size || size > m_size - offset) {
-		EXIT("Buffer: invalid DMA barrier, handle=%p offset=0x%016" PRIx64
-		     " size=0x%016" PRIx64 " capacity=0x%016" PRIx64 "\n",
+		EXIT("Buffer: invalid DMA barrier, handle=%p offset=0x%016" PRIx64 " size=0x%016" PRIx64
+		     " capacity=0x%016" PRIx64 "\n",
 		     static_cast<const void*>(Handle()), offset, size, m_size);
 	}
 	vk::BufferMemoryBarrier barrier {};
@@ -175,10 +175,9 @@ void Buffer::CopyFrom(CommandBuffer& command, const Buffer& source, uint64_t sou
 	command.EndRendering();
 	const vk::BufferMemoryBarrier before[] = {
 	    source.Barrier(source_offset, size, source_before, vk::AccessFlagBits::eTransferRead),
-	    Barrier(destination_offset, size, destination_before,
-	            vk::AccessFlagBits::eTransferWrite),
+	    Barrier(destination_offset, size, destination_before, vk::AccessFlagBits::eTransferWrite),
 	};
-	const auto host_access = vk::AccessFlagBits::eHostRead | vk::AccessFlagBits::eHostWrite;
+	const auto host_access  = vk::AccessFlagBits::eHostRead | vk::AccessFlagBits::eHostWrite;
 	auto       before_stage = vk::PipelineStageFlags {vk::PipelineStageFlagBits::eAllCommands};
 	if (static_cast<bool>((source_before | destination_before) & host_access)) {
 		before_stage |= vk::PipelineStageFlagBits::eHost;
@@ -214,9 +213,8 @@ void Buffer::Fill(uint64_t offset, uint64_t size, uint32_t value) {
 	                       vk::PipelineStageFlagBits::eTransfer, vk::DependencyFlagBits::eByRegion,
 	                       0, nullptr, 1, &before, 0, nullptr);
 	native.fillBuffer(Handle(), offset, size, value);
-	const auto after =
-	    Barrier(offset, size, vk::AccessFlagBits::eTransferWrite,
-	            vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite);
+	const auto after = Barrier(offset, size, vk::AccessFlagBits::eTransferWrite,
+	                           vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite);
 	native.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer,
 	                       vk::PipelineStageFlagBits::eAllCommands,
 	                       vk::DependencyFlagBits::eByRegion, 0, nullptr, 1, &after, 0, nullptr);
@@ -250,8 +248,8 @@ std::pair<uint8_t*, uint64_t> StreamBuffer::Map(uint64_t size, uint64_t alignmen
 	if (Mapped().empty()) {
 		return {nullptr, 0};
 	}
-	uint64_t mapped_size = size;
-	const auto atom = Graphics().physical_device_properties.limits.nonCoherentAtomSize;
+	uint64_t   mapped_size = size;
+	const auto atom        = Graphics().physical_device_properties.limits.nonCoherentAtomSize;
 	if (!NormalizeReservation(IsCoherent(), atom, mapped_size, alignment)) {
 		return {nullptr, 0};
 	}

src/graphics/host_gpu/renderer/cache/streamBuffer.h

@@ -54,16 +54,15 @@ public:
 	[[nodiscard]] bool IsInBounds(uint64_t address, uint64_t size) const noexcept;
 	void               Write(uint64_t offset, const void* source, uint64_t size);
 	void               Flush(uint64_t offset, uint64_t size);
-	void               CopyFrom(
-	                  CommandBuffer& command, const Buffer& source, uint64_t source_offset,
-	                  uint64_t destination_offset, uint64_t size,
-	                  vk::AccessFlags source_before = vk::AccessFlagBits::eMemoryWrite,
-	                  vk::AccessFlags destination_before =
-	                      vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite,
-	                  vk::AccessFlags source_after =
-	                      vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite,
-	                  vk::AccessFlags destination_after =
-	                      vk::AccessFlagBits::eMemoryRead | vk::AccessFlagBits::eMemoryWrite);
+	void CopyFrom(CommandBuffer& command, const Buffer& source, uint64_t source_offset,
+	              uint64_t destination_offset, uint64_t size,
+	              vk::AccessFlags source_before      = vk::AccessFlagBits::eMemoryWrite,
+	              vk::AccessFlags destination_before = vk::AccessFlagBits::eMemoryRead |
+	                                                   vk::AccessFlagBits::eMemoryWrite,
+	              vk::AccessFlags source_after       = vk::AccessFlagBits::eMemoryRead |
+	                                                   vk::AccessFlagBits::eMemoryWrite,
+	              vk::AccessFlags destination_after  = vk::AccessFlagBits::eMemoryRead |
+	                                                   vk::AccessFlagBits::eMemoryWrite);
 	void Fill(uint64_t offset, uint64_t size, uint32_t value);
 
 protected:
@@ -107,13 +106,13 @@ private:
 		uint64_t upper_bound = 0;
 	};
 
-	void               ReserveWatches(std::vector<Watch>& watches, size_t grow_size);
+	void                      ReserveWatches(std::vector<Watch>& watches, size_t grow_size);
 	[[nodiscard]] static bool NormalizeReservation(bool coherent, uint64_t atom, uint64_t& size,
 	                                               uint64_t& alignment);
-	[[nodiscard]] bool WaitPendingOperations(const std::vector<Watch>& watches,
-	                                         std::optional<size_t>     invalidation_mark,
-	                                         uint64_t requested_upper_bound, bool allow_wait,
-	                                         size_t& wait_cursor, uint64_t& wait_bound);
+	[[nodiscard]] bool        WaitPendingOperations(const std::vector<Watch>& watches,
+	                                                std::optional<size_t>     invalidation_mark,
+	                                                uint64_t requested_upper_bound, bool allow_wait,
+	                                                size_t& wait_cursor, uint64_t& wait_bound);
 
 	uint64_t              m_offset      = 0;
 	uint64_t              m_mapped_size = 0;

src/graphics/host_gpu/renderer/cache/textureCache.cpp

@@ -1,4 +1,4 @@
-#include "graphics/host_gpu/renderer/textureCache.h"
+#include "graphics/host_gpu/renderer/cache/textureCache.h"
 
 #include "common/assert.h"
 #include "common/emulatorConfig.h"
@@ -7,13 +7,13 @@
 #include "graphics/guest_gpu/gpu_format.h"
 #include "graphics/guest_gpu/tile.h"
 #include "graphics/host_gpu/graphicContext.h"
-#include "graphics/host_gpu/objects/textureCommon.h"
-#include "graphics/host_gpu/renderer/bufferCache.h"
+#include "graphics/host_gpu/renderer/cache/bufferCache.h"
+#include "graphics/host_gpu/renderer/cache/resourceMutex.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
-#include "graphics/host_gpu/renderer/imageView.h"
+#include "graphics/host_gpu/renderer/image/imageView.h"
+#include "graphics/host_gpu/renderer/image/textureCommon.h"
+#include "graphics/host_gpu/renderer/image/tiler.h"
 #include "graphics/host_gpu/renderer/render.h"
-#include "graphics/host_gpu/renderer/resourceMutex.h"
-#include "graphics/host_gpu/renderer/tiler.h"
 #include "kernel/memory.h"
 
 #include <algorithm>
@@ -58,8 +58,8 @@ private:
 TextureCache::TextureCache(GraphicContext& graphics, CommandScheduler& scheduler,
                            PageManager& page_manager, BufferCache& buffer_cache,
                            ResourceMutex& resource_mutex)
-    : m_graphics(graphics), m_scheduler(scheduler),
-      m_memory_tracker(page_manager, PageWatchMode::Write), m_blit_helper(graphics, scheduler),
+    : m_graphics(graphics), m_scheduler(scheduler), m_page_manager(page_manager),
+      m_blit_helper(graphics, scheduler),
       m_tiler(std::make_unique<TileManager>(graphics, scheduler,
                                             buffer_cache.GetUtilityBuffer(MemoryUsage::Stream))),
       m_buffer_cache(buffer_cache), m_resource_mutex(resource_mutex),
@@ -80,7 +80,7 @@ TextureCache::TextureCache(GraphicContext& graphics, CommandScheduler& scheduler
 TextureCache::~TextureCache() {
 	for (uint32_t index = 0; index < m_slots.size(); index++) {
 		if (m_slots[index].image != nullptr && m_slots[index].image->registered) {
-			UnregisterImage({index, m_slots[index].generation}, false);
+			UnregisterImage({index, m_slots[index].generation});
 		}
 		m_slots[index].image.reset();
 	}
@@ -88,15 +88,34 @@ TextureCache::~TextureCache() {
 
 bool TextureCache::SameBacking(const ImageInfo& cached, const ImageInfo& requested,
                                bool exact_format) {
-	const bool unit_extent =
-	    requested.extent.width == 1 && requested.extent.height == 1 && requested.extent.depth == 1;
-	return cached.data == requested.data && cached.extent == requested.extent &&
-	       cached.samples == requested.samples &&
-	       cached.bytes_per_block == requested.bytes_per_block &&
-	       (cached.type == requested.type || unit_extent) &&
-	       (exact_format
-	            ? cached.pixel_format == requested.pixel_format
-	            : ImageViewOps::FormatsCompatible(cached.pixel_format, requested.pixel_format));
+	if (cached.data.address != requested.data.address) {
+		return false;
+	}
+	if (cached.data.size != requested.data.size) {
+		return false;
+	}
+	if (cached.extent != requested.extent) {
+		return false;
+	}
+	if (cached.samples != requested.samples) {
+		return false;
+	}
+	if (cached.bytes_per_block != requested.bytes_per_block) {
+		return false;
+	}
+	if (cached.tile_mode != requested.tile_mode) {
+		return false;
+	}
+	if (!ImageViewOps::FormatsCompatible(cached.pixel_format, requested.pixel_format)) {
+		return false;
+	}
+	if (cached.type != requested.type && requested.extent != vk::Extent3D {1, 1, 1}) {
+		return false;
+	}
+	if (exact_format && cached.pixel_format != requested.pixel_format) {
+		return false;
+	}
+	return true;
 }
 
 TextureCache::BindingType TextureCache::UploadBinding(const Image& image) {
@@ -117,8 +136,7 @@ bool TextureCache::SafeToDownload(const Image& image) {
 		return false;
 	}
 	const auto range = image.info.data;
-	return !m_buffer_cache.HasGpuDirtyBytes(range.address, range.size) &&
-	       !m_memory_tracker.IsRegionCpuModified(range.address, range.size);
+	return !m_buffer_cache.HasGpuDirtyBytes(range.address, range.size);
 }
 
 Image& TextureCache::ResolveImage(ImageId id) {
@@ -187,21 +205,17 @@ void TextureCache::RegisterImage(ImageId id) {
 	m_total_used_memory += image.AccountedSize();
 }
 
-void TextureCache::UnregisterImage(ImageId id, bool release_tracking) {
+void TextureCache::UnregisterImage(ImageId id) {
 	auto& image = ResolveImage(id);
 	if (!image.registered) {
 		return;
 	}
+	UntrackImage(id);
 	std::vector<ImageOwnerIndex::ByteRange> releases;
 	if (!m_image_owner_index.Unregister(id, releases)) {
 		EXIT("TextureCache: image missing from owner index\n");
 	}
 	m_lru_cache.Free(image.lru_id);
-	if (release_tracking) {
-		for (const auto& range: releases) {
-			m_memory_tracker.UntrackMemory(range.address, range.size);
-		}
-	}
 	const auto accounted = image.AccountedSize();
 	if (accounted > m_total_used_memory) {
 		EXIT("TextureCache: image accounting underflow\n");
@@ -210,7 +224,7 @@ void TextureCache::UnregisterImage(ImageId id, bool release_tracking) {
 	image.registered = false;
 }
 
-void TextureCache::DeleteImage(ImageId id, bool release_tracking) {
+void TextureCache::DeleteImage(ImageId id) {
 	auto owner = ResolveOwner(id);
 	if (owner == nullptr || !owner->registered) {
 		return;
@@ -224,7 +238,7 @@ void TextureCache::DeleteImage(ImageId id, bool release_tracking) {
 			}
 		}
 		for (const auto association: associations) {
-			ReleaseGpuTracking(association);
+			ClearGpuModified(association);
 			DeleteImage(association);
 		}
 	}
@@ -235,7 +249,7 @@ void TextureCache::DeleteImage(ImageId id, bool release_tracking) {
 	if (owner->info.metadata.kind == ImageMetadataKind::Htile) {
 		m_surface_metas.erase(owner->info.metadata.range.address);
 	}
-	UnregisterImage(id, release_tracking);
+	UnregisterImage(id);
 	const auto erase_slot = [this, id, retained = owner] {
 		auto& slot = m_slots[id.index];
 		if (slot.generation != id.generation || slot.image != retained) {
@@ -266,10 +280,10 @@ void TextureCache::DeleteImages(std::span<const ImageId> ids,
 			continue;
 		}
 		if (native_source == id) {
-			ReleaseGpuTracking(id);
+			ClearGpuModified(id);
 		} else if (owner->IsGpuModified()) {
 			DownloadImage(id);
-			ReleaseGpuTracking(id);
+			ClearGpuModified(id);
 		}
 		DeleteImage(id);
 	}
@@ -296,6 +310,121 @@ void TextureCache::TouchImage(Image& image) {
 	}
 }
 
+void TextureCache::TrackImage(ImageId id) {
+	auto& image = ResolveImage(id);
+	if (!image.registered) {
+		return;
+	}
+	const auto image_begin = image.info.data.address;
+	const auto image_end   = image.info.data.End();
+	if (image_begin == image.track_addr && image_end == image.track_addr_end) {
+		return;
+	}
+	if (!image.IsTracked()) {
+		image.track_addr     = image_begin;
+		image.track_addr_end = image_end;
+		m_page_manager.UpdatePageWatchers<true>(image_begin, image.info.data.size);
+		return;
+	}
+	if (image_begin < image.track_addr) {
+		TrackImageHead(id);
+	}
+	if (image.track_addr_end < image_end) {
+		TrackImageTail(id);
+	}
+}
+
+void TextureCache::TrackImageHead(ImageId id) {
+	auto& image = ResolveImage(id);
+	if (!image.registered) {
+		return;
+	}
+	const auto image_begin = image.info.data.address;
+	if (image_begin == image.track_addr) {
+		return;
+	}
+	if (!image.IsTracked() || image_begin > image.track_addr) {
+		EXIT("TextureCache: invalid image head tracking range\n");
+	}
+	const auto size  = image.track_addr - image_begin;
+	image.track_addr = image_begin;
+	m_page_manager.UpdatePageWatchers<true>(image_begin, size);
+}
+
+void TextureCache::TrackImageTail(ImageId id) {
+	auto& image = ResolveImage(id);
+	if (!image.registered) {
+		return;
+	}
+	const auto image_end = image.info.data.End();
+	if (image_end == image.track_addr_end) {
+		return;
+	}
+	if (!image.IsTracked() || image.track_addr_end > image_end) {
+		EXIT("TextureCache: invalid image tail tracking range\n");
+	}
+	const auto address   = image.track_addr_end;
+	const auto size      = image_end - address;
+	image.track_addr_end = image_end;
+	m_page_manager.UpdatePageWatchers<true>(address, size);
+}
+
+void TextureCache::UntrackImage(ImageId id) {
+	auto& image = ResolveImage(id);
+	if (!image.IsTracked()) {
+		return;
+	}
+	const auto address   = image.track_addr;
+	const auto size      = image.track_addr_end - image.track_addr;
+	image.track_addr     = 0;
+	image.track_addr_end = 0;
+	if (size != 0) {
+		m_page_manager.UpdatePageWatchers<false>(address, size);
+	}
+}
+
+void TextureCache::UntrackImageHead(ImageId id) {
+	auto&      image = ResolveImage(id);
+	const auto begin = image.info.data.address;
+	if (!image.IsTracked() || begin < image.track_addr) {
+		return;
+	}
+	const auto address = (begin + TRACKER_PAGE_SIZE) & ~(TRACKER_PAGE_SIZE - 1);
+	const auto size    = address - begin;
+	image.track_addr   = address;
+	if (image.track_addr == image.track_addr_end) {
+		image.MarkMaybeCpuDirty();
+		if (image.NeedsMaybeCpuHash()) {
+			image.SetMaybeCpuHash(image.HashGuestEdges());
+		}
+		UntrackImage(id);
+	}
+	if (size != 0) {
+		m_page_manager.UpdatePageWatchers<false>(begin, size);
+	}
+}
+
+void TextureCache::UntrackImageTail(ImageId id) {
+	auto&      image = ResolveImage(id);
+	const auto end   = image.info.data.End();
+	if (!image.IsTracked() || image.track_addr_end < end) {
+		return;
+	}
+	const auto address   = end & ~(TRACKER_PAGE_SIZE - 1);
+	const auto size      = end - address;
+	image.track_addr_end = address;
+	if (image.track_addr == image.track_addr_end) {
+		image.MarkMaybeCpuDirty();
+		if (image.NeedsMaybeCpuHash()) {
+			image.SetMaybeCpuHash(image.HashGuestEdges());
+		}
+		UntrackImage(id);
+	}
+	if (size != 0) {
+		m_page_manager.UpdatePageWatchers<false>(address, size);
+	}
+}
+
 void TextureCache::TrackImageDownload(ImageId id) {
 	std::lock_guard transaction(m_resource_mutex);
 	CacheLock       lock(*this, m_lock);
@@ -376,9 +505,6 @@ void TextureCache::ValidateImageDesc(const ImageDesc& desc) const {
 }
 
 void TextureCache::PrepareImageCopy(Image& image) {
-	const auto range = image.info.data;
-	m_memory_tracker.ForEachUploadRange(
-	    range.address, range.size, false, [](uint64_t, uint64_t) noexcept {}, []() noexcept {});
 	if (image.IsCpuDirty()) {
 		image.RefreshComplete();
 	}
@@ -471,6 +597,7 @@ void TextureCache::CopyImage(ImageId destination_id, ImageId source_id) {
 	RefreshCopySource(source_id);
 	auto& destination = ResolveImage(destination_id);
 	auto& source      = ResolveImage(source_id);
+	TrackImage(destination_id);
 	if (source.backing.samples != destination.backing.samples) {
 		EXIT("TextureCache: cannot issue an unequal-sample image copy\n");
 	}
@@ -479,7 +606,6 @@ void TextureCache::CopyImage(ImageId destination_id, ImageId source_id) {
 		if (source.info.data == destination.info.data) {
 			destination.MarkBufferModified();
 		}
-		RestoreGpuTracking(destination);
 		return;
 	}
 	const bool source_depth = source.info.IsDepth();
@@ -503,7 +629,6 @@ void TextureCache::CopyImage(ImageId destination_id, ImageId source_id) {
 		destination.MarkGpuModified();
 	}
 	destination.ClearBufferModified();
-	RestoreGpuTracking(destination);
 }
 
 void TextureCache::CopyImageMip(ImageId destination_id, ImageId source_id, uint32_t mip,
@@ -511,6 +636,7 @@ void TextureCache::CopyImageMip(ImageId destination_id, ImageId source_id, uint3
 	RefreshCopySource(source_id);
 	auto& destination = ResolveImage(destination_id);
 	auto& source      = ResolveImage(source_id);
+	TrackImage(destination_id);
 	if (source.IsBufferModified() || source.backing.samples != destination.backing.samples) {
 		EXIT("TextureCache: invalid mip-copy ownership or sample count\n");
 	}
@@ -520,7 +646,6 @@ void TextureCache::CopyImageMip(ImageId destination_id, ImageId source_id, uint3
 	if (source.IsGpuModified()) {
 		destination.MarkGpuModified();
 	}
-	RestoreGpuTracking(destination);
 }
 
 ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingType binding,
@@ -590,7 +715,7 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingTyp
 	if (copied) {
 		DeleteImages(std::array {cached_id}, cached_id);
 	} else {
-		ReleaseGpuTracking(cached_id);
+		ClearGpuModified(cached_id);
 		DeleteImage(cached_id);
 	}
 	return replacement_id;
@@ -629,6 +754,12 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 		    (requested.IsVolume() || cached.info.IsVolume())) {
 			return {ExpandImage(requested, cached_id)};
 		}
+		if (requested.tile_mode != cached.info.tile_mode) {
+			if (safe_to_delete) {
+				DeleteImages(std::array {cached_id}, cached_id);
+			}
+			return {merged_id};
+		}
 		if (requested.pixel_format != cached.info.pixel_format ||
 		    requested.data.size <= cached.info.data.size) {
 			const auto result_id = merged_id ? merged_id : cached_id;
@@ -641,12 +772,6 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
 		if (requested.type == cached.info.type && requested.resources > cached.info.resources) {
 			return {ExpandImage(requested, cached_id)};
 		}
-		if (requested.tile_mode != cached.info.tile_mode) {
-			if (safe_to_delete) {
-				DeleteImages(std::array {cached_id}, cached_id);
-			}
-			return {merged_id};
-		}
 		EXIT("TextureCache: unresolvable equal-address image overlap, address=0x%016" PRIx64
 		     " requested=%ux%u "
 		     "cached=%ux%u requested_size=0x%016" PRIx64 " cached_size=0x%016" PRIx64
@@ -742,22 +867,28 @@ TextureCache::BuildColorTransfer(const Image& image, BindingType binding,
 			case BindingType::Texture: break;
 			case BindingType::Storage: owner = "StorageTextureCache"; break;
 			case BindingType::RenderTarget:
+				if (info.resources.layers == 0 || info.data.size % info.resources.layers != 0 ||
+				    info.samples != 1 || image.backing.samples != 1) {
+					EXIT("TextureCache: invalid color-attachment upload\n");
+				}
+				format           = ImageOps::RenderTargetTransferFormat(info.bytes_per_block);
+				allow_depth_tile = false;
+				plan.swap_bgra16 = info.bgra16;
+				owner            = "RenderTarget";
+				break;
 			case BindingType::VideoOut:
 				if (info.resources.layers == 0 || info.data.size % info.resources.layers != 0 ||
 				    info.samples != 1 || image.backing.samples != 1 ||
-				    (binding == BindingType::VideoOut &&
-				     info.metadata.compression != VideoOutCompression::Uncompressed)) {
+				    info.metadata.compression != VideoOutCompression::Uncompressed) {
 					EXIT("TextureCache: invalid color-attachment upload\n");
 				}
-				format           = binding == BindingType::RenderTarget
-				                       ? ImageOps::RenderTargetTransferFormat(info.bytes_per_block)
-				                       : info.guest_format;
+				format           = info.guest_format;
 				layers           = info.resources.layers;
 				volume           = false;
 				layered          = layers > 1;
 				allow_depth_tile = false;
 				plan.swap_bgra16 = info.bgra16;
-				owner = binding == BindingType::RenderTarget ? "RenderTarget" : "VideoOut";
+				owner            = "VideoOut";
 				break;
 			case BindingType::DepthTarget: return plan;
 		}
@@ -903,81 +1034,54 @@ void TextureCache::InitializeImage(ImageId id, const ImageDesc& desc) {
 	if (image.info.data.Empty()) {
 		return;
 	}
+	TrackImage(id);
 	if (image.info.metadata.compression != VideoOutCompression::Uncompressed) {
-		m_memory_tracker.ForEachUploadRange(
-		    image.info.data.address, image.info.data.size, false,
-		    [](uint64_t, uint64_t) noexcept {}, []() noexcept {});
 		if (image.IsCpuDirty()) {
 			image.RefreshComplete();
 		}
 		return;
 	}
 	if (image.info.samples > 1) {
-		RestoreGpuTracking(image);
 		return;
 	}
-	bool data_gpu_owned = false;
-	bool data_imported  = false;
-	bool uploaded       = false;
-	m_memory_tracker.ForEachUploadRange(
-	    image.info.data.address, image.info.data.size, false,
-	    [&](uint64_t, uint64_t) noexcept { uploaded = true; },
-	    [&]() noexcept {
-		    uploaded |= image.IsBufferModified() || image.IsDefinitelyCpuDirty();
-		    if (!uploaded) {
-			    return;
-		    }
-		    const auto source =
-		        m_buffer_cache.ObtainBufferForImage(image.info.data.address, image.info.data.size);
-		    if (source.buffer == nullptr) {
-			    EXIT("TextureCache: failed to obtain image upload source\n");
-		    }
-		    data_gpu_owned |= source.gpu_owned;
-		    data_imported = true;
-		    UploadImage(image, desc, *source.buffer, source.offset);
-	    });
+	bool       data_imported = false;
+	const bool upload        = image.IsBufferModified() || image.IsCpuDirty();
+	if (upload) {
+		const auto source =
+		    m_buffer_cache.ObtainBufferForImage(image.info.data.address, image.info.data.size);
+		if (source.buffer == nullptr) {
+			EXIT("TextureCache: failed to obtain image upload source\n");
+		}
+		data_imported = true;
+		UploadImage(image, desc, *source.buffer, source.offset);
+	}
 	if (data_imported) {
 		image.ClearBufferModified();
 	}
-	if (data_gpu_owned) {
-		image.MarkGpuModified();
-	}
 	if (image.IsCpuDirty()) {
 		image.RefreshComplete();
 	}
-	RestoreGpuTracking(image);
 }
 
 void TextureCache::RefreshImage(ImageId id, const ImageDesc& desc) {
-	auto& image           = ResolveImage(id);
-	bool  unchanged_maybe = false;
+	TrackImage(id);
+	auto& image = ResolveImage(id);
 	if (image.IsMaybeCpuDirty()) {
 		const auto hash = image.HashGuestEdges();
 		if (image.NeedsMaybeCpuHash()) {
 			image.SetMaybeCpuHash(hash);
 			return;
 		}
-		unchanged_maybe = !image.ResolveMaybeCpuHash(hash);
-		if (unchanged_maybe) {
-			m_memory_tracker.ForEachUploadRange(
-			    image.info.data.address, image.info.data.size, false,
-			    [](uint64_t, uint64_t) noexcept {}, []() noexcept {});
-		}
+		(void)image.ResolveMaybeCpuHash(hash);
 	}
 	bool cpu_dirty = image.IsBufferModified() || image.IsDefinitelyCpuDirty();
-	if (!unchanged_maybe) {
-		cpu_dirty |=
-		    m_memory_tracker.IsRegionCpuModified(image.info.data.address, image.info.data.size);
-	}
 	if (image.info.metadata.compression != VideoOutCompression::Uncompressed) {
 		if (cpu_dirty) {
 			EXIT("TextureCache: compressed guest image refresh is unsupported\n");
 		}
-		RestoreGpuTracking(image);
 		return;
 	}
 	if (!cpu_dirty) {
-		RestoreGpuTracking(image);
 		return;
 	}
 	InitializeImage(id, desc);
@@ -1028,8 +1132,7 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 	}
 
 	ImageId result {};
-	bool    replacement_buffer = false;
-	bool    replacing_image    = false;
+	bool    inserted_new = false;
 	{
 		std::lock_guard            transaction(m_resource_mutex);
 		CacheLock                  lock(*this, m_lock);
@@ -1038,7 +1141,7 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 
 		for (const auto id: candidates) {
 			const auto owner = ResolveOwner(id);
-			if (owner == nullptr || owner->info.data != desc.info.data) {
+			if (owner == nullptr) {
 				continue;
 			}
 			if (SameBacking(owner->info, desc.info, exact_format)) {
@@ -1056,8 +1159,8 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 				if (owner == nullptr) {
 					continue;
 				}
-				const auto merged_info = result ? ResolveImage(result).info : desc.info;
-				const auto overlap     = ResolveOverlap(merged_info, desc.type, candidate, result);
+				const auto& merged_info = result ? ResolveImage(result).info : desc.info;
+				const auto  overlap     = ResolveOverlap(merged_info, desc.type, candidate, result);
 				if (overlap.image) {
 					result     = overlap.image;
 					view_mip   = overlap.mip;
@@ -1071,28 +1174,19 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 			if (exact_format && resolved.info.pixel_format != desc.info.pixel_format) {
 				result = {};
 			} else if (resolved.info.resources < desc.info.resources) {
-				ImageDesc refresh {
-				    .info = resolved.info, .view_info = {}, .type = UploadBinding(resolved)};
-				RefreshImage(result, refresh);
-				if (resolved.IsGpuModified() && !SynchronizeImageToBuffer(result)) {
-					EXIT("TextureCache: cannot preserve an unsupported replacement image\n");
-				}
-				replacement_buffer = resolved.IsBufferModified();
-				DeleteImage(result);
-				result          = {};
-				replacing_image = true;
+				result = ExpandImage(desc.info, result);
 			}
 		}
 		if (!result) {
 			result         = InsertImage(desc.info);
+			inserted_new   = true;
 			auto& inserted = ResolveImage(result);
-			if (replacement_buffer || m_buffer_cache.HasGpuDirtyBytes(inserted.info.data.address,
-			                                                          inserted.info.data.size)) {
+			if (m_buffer_cache.HasGpuDirtyBytes(inserted.info.data.address,
+			                                    inserted.info.data.size)) {
 				inserted.MarkBufferModified();
-			} else if (replacing_image) {
-				m_memory_tracker.MarkRegionAsCpuModified(inserted.info.data.address,
-				                                         inserted.info.data.size);
 			}
+		}
+		if (inserted_new) {
 			InitializeImage(result, desc);
 		} else {
 			RefreshImage(result, desc);
@@ -1101,9 +1195,7 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 		auto& image = ResolveImage(result);
 		if (desc.type == BindingType::VideoOut &&
 		    desc.info.metadata.compression != VideoOutCompression::Uncompressed) {
-			const bool guest_dirty =
-			    image.IsBufferModified() || image.IsCpuDirty() ||
-			    m_memory_tracker.IsRegionCpuModified(image.info.data.address, image.info.data.size);
+			const bool guest_dirty = image.IsBufferModified() || image.IsCpuDirty();
 			const bool native_current =
 			    (image.usage.render_target || image.IsGpuModified()) && !guest_dirty;
 			if (!native_current) {
@@ -1252,6 +1344,7 @@ void TextureCache::MarkGpuWritten(ImageId id) {
 	if (!image.registered || image.depth_id) {
 		EXIT("TextureCache: cannot mark an unavailable image GPU-written\n");
 	}
+	TrackImage(id);
 	CommitGpuWrite(image);
 }
 
@@ -1259,19 +1352,11 @@ void TextureCache::CommitGpuWrite(Image& image) {
 	if (image.depth_id || image.backing.image == nullptr) {
 		EXIT("TextureCache: stencil association cannot own image contents\n");
 	}
-	const auto range = image.info.data;
-	if (m_buffer_cache.HasGpuDirtyBytes(range.address, range.size)) {
-		m_buffer_cache.DiscardGpuDirtyBytes(range.address, range.size);
-	}
-	m_buffer_cache.InvalidateImageAliases(range.address, range.size);
 	image.ClearBufferModified();
-	m_memory_tracker.ForEachUploadRange(
-	    range.address, range.size, true, [](uint64_t, uint64_t) noexcept {}, []() noexcept {});
 	if (image.IsCpuDirty()) {
 		image.RefreshComplete();
 	}
 	image.MarkGpuModified();
-	RestoreGpuTracking(image);
 }
 
 bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address, uint64_t size,
@@ -1279,7 +1364,6 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 	if (command.IsInvalid() || !GuestRange {address, size}.Valid()) {
 		EXIT("TextureCache: invalid image clear\n");
 	}
-	m_buffer_cache.ValidateGpuAccess(address, size, false, true);
 	std::lock_guard      transaction(m_resource_mutex);
 	CacheLock            lock(*this, m_lock);
 	ImageId              selected {};
@@ -1332,11 +1416,7 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 			return false;
 		}
 	}
-	if (m_buffer_cache.HasGpuDirtyBytes(address, size)) {
-		m_buffer_cache.DiscardGpuDirtyBytes(address, size);
-	}
-	if (image.IsBufferModified() || image.IsCpuDirty() ||
-	    m_memory_tracker.IsRegionCpuModified(image.info.data.address, image.info.data.size)) {
+	if (image.IsBufferModified() || image.IsCpuDirty()) {
 		ImageDesc refresh {.info = image.info, .view_info = {}, .type = UploadBinding(image)};
 		InitializeImage(selected, refresh);
 		if (image.info.samples == 1 && (image.IsBufferModified() || image.IsCpuDirty())) {
@@ -1361,14 +1441,12 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 	return true;
 }
 
-void TextureCache::PrepareHostWrite(uint64_t address, uint64_t size) {
+void TextureCache::InvalidateMemory(uint64_t address, uint64_t size) {
 	if (!GuestRange {address, size}.Valid()) {
-		EXIT("TextureCache: invalid host-write range\n");
+		EXIT("TextureCache: invalid memory-invalidation range\n");
 	}
 	CacheLock lock(*this, m_lock);
 	InvalidateCpuAliases(address, size);
-	m_memory_tracker.ForEachDownloadRange<true>(address, size, [](uint64_t, uint64_t) noexcept {});
-	m_memory_tracker.MarkRegionAsCpuModified(address, size);
 }
 
 void TextureCache::DownloadDepth(Image& image, Buffer& destination, uint64_t destination_offset) {
@@ -1457,11 +1535,14 @@ void TextureCache::DownloadDepth(Image& image, Buffer& destination, uint64_t des
 }
 
 void TextureCache::DownloadImageData(Image& image, Buffer& destination, uint64_t destination_offset,
-                                     DownloadPlan plan) {
+                                     uint64_t destination_size, DownloadPlan plan) {
 	if (!plan.valid) {
 		EXIT("TextureCache: invalid image download plan\n");
 	}
 	if (plan.depth) {
+		if (destination_size != image.info.data.size) {
+			EXIT("TextureCache: partial depth image download is unsupported\n");
+		}
 		DownloadDepth(image, destination, destination_offset);
 		return;
 	}
@@ -1471,22 +1552,118 @@ void TextureCache::DownloadImageData(Image& image, Buffer& destination, uint64_t
 	                                         : TileManager::ColorTransform::None;
 	if (!color.tiled) {
 		if (transform == TileManager::ColorTransform::SwapBgra16) {
-			auto linear = m_tiler->GetScratchBuffer(image.info.data.size);
+			auto linear = m_tiler->GetScratchBuffer(destination_size);
 			image.Download(color.regions, linear.buffer, 0, linear.size);
 			m_tiler->SwapBgra16(linear,
-			                    {destination.Handle(), destination_offset, image.info.data.size});
+			                    {destination.Handle(), destination_offset, destination_size});
 			return;
 		}
 		for (auto& copy: color.regions) {
 			copy.bufferOffset += destination_offset;
 		}
-		image.Download(color.regions, destination.Handle(), destination_offset,
-		               image.info.data.size);
+		image.Download(color.regions, destination.Handle(), destination_offset, destination_size);
 		return;
 	}
 
 	m_tiler->TileImage(image, color.regions, destination.Handle(), destination_offset,
-	                   image.info.data.size, image.info.data.size, color.tiles, transform);
+	                   destination_size, destination_size, color.tiles, transform);
+}
+
+bool BufferCache::SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uint64_t size) {
+	CacheLock            lock(m_texture_cache, m_texture_cache.m_lock);
+	std::vector<ImageId> matches;
+	for (const auto id: m_texture_cache.FindImagesInRegion(vaddr, size, false)) {
+		auto owner = m_texture_cache.ResolveOwner(id);
+		if (owner == nullptr || owner->info.data.address != vaddr) {
+			continue;
+		}
+		if (owner->depth_id) {
+			owner = m_texture_cache.ResolveOwner(owner->depth_id);
+		}
+		if (owner != nullptr && owner->SafeToDownload()) {
+			matches.push_back(id);
+		}
+	}
+
+	ImageId selected {};
+	if (matches.size() == 1) {
+		selected = matches.front();
+	} else {
+		for (const auto id: matches) {
+			const auto& image = m_texture_cache.ResolveImage(id);
+			if (image.info.data.size == size) {
+				selected = id;
+				break;
+			}
+		}
+	}
+	if (!selected) {
+		return false;
+	}
+	if (const auto owner = m_texture_cache.ResolveOwner(selected);
+	    owner != nullptr && owner->depth_id) {
+		selected = owner->depth_id;
+	}
+
+	auto& image = m_texture_cache.ResolveImage(selected);
+	if (!buffer.IsInBounds(image.info.data.address, 1)) {
+		return false;
+	}
+	const auto buf_offset = buffer.Offset(image.info.data.address);
+	const auto available  = buffer.Size() - buf_offset;
+	uint32_t   levels     = 0;
+	uint64_t   copy_size  = 0;
+	if (image.info.IsVolume()) {
+		// Volume mips contain strided block slices, so a mip's linear span cannot prove that
+		// every retained slice fits. Keep volume synchronization whole-image only.
+		if (!buffer.IsInBounds(image.info.data.address, image.info.data.size)) {
+			return false;
+		}
+		levels    = image.info.resources.levels;
+		copy_size = image.info.data.size;
+	} else {
+		for (; levels < image.info.resources.levels; ++levels) {
+			const auto& mip = image.info.mip_layout[levels];
+			if (mip.size == 0 || mip.offset > available || mip.size > available - mip.offset) {
+				break;
+			}
+			copy_size = std::max(copy_size, mip.offset + mip.size);
+		}
+	}
+	if (copy_size == 0) {
+		return false;
+	}
+	auto plan = m_texture_cache.BuildDownload(image);
+	if (!plan.valid) {
+		return false;
+	}
+	if (plan.depth && copy_size != image.info.data.size) {
+		return false;
+	}
+	if (!plan.depth && levels < image.info.resources.levels) {
+		auto& color = plan.color;
+		std::erase_if(color.regions, [levels](const vk::BufferImageCopy& region) {
+			return region.imageSubresource.mipLevel >= levels;
+		});
+		if (color.regions.empty()) {
+			return false;
+		}
+		if (color.tiled) {
+			const auto binding = m_texture_cache.UploadBinding(image);
+			const auto format =
+			    binding == TextureCache::BindingType::RenderTarget
+			        ? ImageOps::RenderTargetTransferFormat(image.info.bytes_per_block)
+			        : image.info.guest_format;
+			color.tiles.clear();
+			if (!TextureBuildGpuTileInfos(copy_size, color.regions, color.layout, format,
+			                              image.info.TransferLayers(), levels, color.tiles)) {
+				return false;
+			}
+		}
+	}
+	m_texture_cache.DownloadImageData(image, buffer, buf_offset, copy_size, std::move(plan));
+	m_texture_cache.RetainImage(m_scheduler.Current(), selected);
+	return true;
 }
 
 std::pair<uint8_t*, uint64_t> TextureCache::MapDownload(uint64_t size, uint64_t alignment) {
@@ -1518,12 +1695,9 @@ void TextureCache::QueueDownload(GuestRange range, StreamBuffer& download, uint8
 	m_scheduler.Current().Handle().pipelineBarrier(vk::PipelineStageFlagBits::eAllCommands,
 	                                               vk::PipelineStageFlagBits::eHost, {}, 0, nullptr,
 	                                               1, &barrier, 0, nullptr);
-	const auto tick = m_scheduler.CurrentTick();
-	m_buffer_cache.BeginBackingPublication(range.address, range.size, tick);
-	m_scheduler.DeferPriorityOperation([this, &download, range, mapped, offset, tick] {
+	m_scheduler.DeferPriorityOperation([&download, range, mapped, offset] {
 		download.Invalidate(offset, range.size);
 		LibKernel::Memory::WriteBacking(range.address, mapped, range.size);
-		m_buffer_cache.CompleteBackingPublication(range.address, range.size, tick);
 	});
 }
 
@@ -1544,7 +1718,7 @@ bool TextureCache::TryDownloadImage(ImageId id) {
 	}
 	download.Flush(offset, range.size);
 
-	DownloadImageData(image, download, offset, std::move(plan));
+	DownloadImageData(image, download, offset, range.size, std::move(plan));
 
 	QueueDownload(range, download, mapped, offset);
 	return true;
@@ -1558,65 +1732,6 @@ void TextureCache::DownloadImage(ImageId id) {
 	m_scheduler.DrainPriorityOperations();
 }
 
-bool TextureCache::SynchronizeImageToBuffer(ImageId id) {
-	auto& image = ResolveImage(id);
-	if (image.depth_id) {
-		return true;
-	}
-	auto plan = BuildDownload(image);
-	if (!plan.valid) {
-		return false;
-	}
-	const auto range   = image.info.data;
-	const bool refresh = image.IsDefinitelyCpuDirty() ||
-	                     m_memory_tracker.IsRegionCpuModified(range.address, range.size);
-	if (refresh) {
-		RefreshImage(id,
-		             ImageDesc {.info = image.info, .view_info = {}, .type = UploadBinding(image)});
-	}
-	if (!image.IsGpuModified()) {
-		return true;
-	}
-	if (image.IsDefinitelyCpuDirty() || image.IsBufferModified() ||
-	    m_memory_tracker.IsRegionCpuModified(range.address, range.size)) {
-		EXIT("TextureCache: image mirror source is not native-current\n");
-	}
-	auto [destination, offset] =
-	    m_buffer_cache.ObtainBufferForImageWrite(range.address, range.size);
-	if (destination == nullptr) {
-		EXIT("TextureCache: failed to allocate image mirror\n");
-	}
-	DownloadImageData(image, *destination, offset, std::move(plan));
-	m_scheduler.Current().RetainResourceUntilFence(destination);
-	m_buffer_cache.PublishImageBuffer(range.address, range.size);
-	image.MarkBufferModified();
-	RetainImage(m_scheduler.Current(), id);
-	ReleaseGpuTracking(id);
-	return true;
-}
-
-bool TextureCache::SynchronizeImageToBuffer(uint64_t address, uint64_t size) {
-	if (!GuestRange {address, size}.Valid()) {
-		return false;
-	}
-	CacheLock lock(*this, m_lock);
-	ImageId   selected {};
-	for (const auto id: FindImagesInRegion(address, size, true)) {
-		auto owner = ResolveOwner(id);
-		if (owner == nullptr || !owner->GpuOverlaps(address, size)) {
-			continue;
-		}
-		if (selected) {
-			EXIT("TextureCache: ambiguous image-to-buffer synchronization\n");
-		}
-		selected = id;
-	}
-	if (!selected) {
-		return false;
-	}
-	return SynchronizeImageToBuffer(selected);
-}
-
 bool TextureCache::InvalidateMemoryFromGPU(uint64_t address, uint64_t size,
                                            bool formatted_buffer_write) {
 	if (!GuestRange {address, size}.Valid()) {
@@ -1633,7 +1748,7 @@ bool TextureCache::InvalidateMemoryFromGPU(uint64_t address, uint64_t size,
 			if (!formatted_buffer_write) {
 				EXIT("TextureCache: buffer write aliases GPU-modified image\n");
 			}
-			ReleaseGpuTracking(id);
+			ClearGpuModified(id);
 		}
 		owner->MarkBufferModified();
 		found = true;
@@ -1660,50 +1775,40 @@ TextureCache::RegionInfo TextureCache::QueryRegion(uint64_t address, uint64_t si
 }
 
 void TextureCache::InvalidateCpuAliases(uint64_t address, uint64_t size) {
+	const auto page_begin = address & ~(TRACKER_PAGE_SIZE - 1);
+	const auto page_end   = (address + size + TRACKER_PAGE_SIZE - 1) & ~(TRACKER_PAGE_SIZE - 1);
 	for (const auto id: FindImagesInRegion(address, size, true)) {
 		auto owner = ResolveOwner(id);
 		if (owner == nullptr || owner->depth_id) {
 			continue;
 		}
-		owner->InvalidateCpuWrite(address, size);
-		if (owner->NeedsMaybeCpuHash()) {
-			owner->SetMaybeCpuHash(owner->HashGuestEdges());
+		if (owner->Overlaps(address, size)) {
+			owner->InvalidateCpuWrite(address, size);
+			UntrackImage(id);
+			continue;
+		}
+		const auto image_begin = owner->info.data.address;
+		const auto image_end   = owner->info.data.End();
+		if (page_end < image_end) {
+			UntrackImageHead(id);
+		} else if (image_begin < page_begin) {
+			UntrackImageTail(id);
+		} else {
+			owner->MarkMaybeCpuDirty();
+			if (owner->NeedsMaybeCpuHash()) {
+				owner->SetMaybeCpuHash(owner->HashGuestEdges());
+			}
+			UntrackImage(id);
 		}
 	}
 }
 
-void TextureCache::RestoreGpuTracking(const Image& image) {
-	if (!image.IsGpuModified()) {
-		return;
-	}
-	constexpr uint64_t page_mask = TRACKER_PAGE_SIZE - 1;
-	const auto         range     = image.info.data;
-	const auto         begin     = range.address & ~page_mask;
-	const auto         end       = (range.End() + page_mask) & ~page_mask;
-	for (auto page = begin; page < end; page += TRACKER_PAGE_SIZE) {
-		if (!m_memory_tracker.IsRegionGpuModified(page, TRACKER_PAGE_SIZE) &&
-		    !m_memory_tracker.IsRegionCpuModified(page, TRACKER_PAGE_SIZE)) {
-			m_memory_tracker.MarkRegionAsGpuModified(page, TRACKER_PAGE_SIZE);
-		}
-	}
-}
-
-void TextureCache::ReleaseGpuTracking(ImageId id) {
+void TextureCache::ClearGpuModified(ImageId id) {
 	auto owner = ResolveOwner(id);
 	if (owner == nullptr || !owner->IsGpuModified()) {
 		return;
 	}
-	const auto released = owner->info.data;
 	owner->ClearGpuModified();
-	m_memory_tracker.ForEachDownloadRange<true>(released.address, released.size,
-	                                            [](uint64_t, uint64_t) noexcept {});
-	for (const auto candidate: FindImagesInRegion(released.address, released.size, true)) {
-		const auto survivor = ResolveOwner(candidate);
-		if (survivor != nullptr && survivor.get() != owner.get() && !survivor->depth_id) {
-			RestoreGpuTracking(*survivor);
-		}
-	}
-	RestoreGpuTracking(*owner);
 }
 
 bool TextureCache::IsMeta(uint64_t address) {
@@ -1745,39 +1850,6 @@ bool TextureCache::TouchMeta(uint64_t address, uint32_t slice, bool is_clear) {
 	return true;
 }
 
-bool TextureCache::InvalidateMemory(PageFaultAccess access, uint64_t address, uint64_t size,
-                                    PageFaultPhase phase) noexcept {
-	if ((access != PageFaultAccess::Read && access != PageFaultAccess::Write) ||
-	    !GuestRange {address, size}.Valid()) {
-		return false;
-	}
-	if (access == PageFaultAccess::Read) {
-		return false;
-	}
-	if (phase == PageFaultPhase::Invalidate) {
-		const bool gpu_image =
-		    m_memory_tracker.InvalidateVirtualGpuWrite(access, address, size, phase);
-		CpuFaultAction action = gpu_image ? CpuFaultAction::Download
-		                                  : m_memory_tracker.BeginCpuFault(address, size, access);
-		{
-			CacheLock lock(*this, m_lock);
-			InvalidateCpuAliases(address, size);
-		}
-		return action != CpuFaultAction::Untracked;
-	}
-
-	if (phase == PageFaultPhase::Complete) {
-		const bool gpu_image = m_memory_tracker.IsRegionGpuModified(address, size);
-		return gpu_image ? m_memory_tracker.InvalidateVirtualGpuWrite(access, address, size, phase)
-		                 : m_memory_tracker.CompleteCpuFault(address, size, access, false);
-	}
-	if (phase != PageFaultPhase::Release) {
-		return false;
-	}
-	(void)m_memory_tracker.InvalidateVirtualGpuWrite(access, address, size, phase);
-	return true;
-}
-
 void TextureCache::UnmapMemory(uint64_t address, uint64_t size) {
 	if (!GuestRange {address, size}.Valid()) {
 		EXIT("TextureCache: invalid unmap range\n");
@@ -1794,16 +1866,10 @@ void TextureCache::UnmapMemory(uint64_t address, uint64_t size) {
 			continue;
 		}
 		if (owner->IsGpuModified()) {
-			ReleaseGpuTracking(id);
+			ClearGpuModified(id);
 		}
 		DeleteImage(id);
 	}
-	m_memory_tracker.UntrackMemory(address, size);
-	for (const auto id: FindImagesInRegion(address, size, true)) {
-		if (const auto survivor = ResolveOwner(id); survivor != nullptr) {
-			RestoreGpuTracking(*survivor);
-		}
-	}
 }
 
 void TextureCache::RunGarbageCollector() {
@@ -1849,7 +1915,7 @@ void TextureCache::RunGarbageCollector() {
 				if (safe && !TryDownloadImage(id)) {
 					continue;
 				}
-				ReleaseGpuTracking(id);
+				ClearGpuModified(id);
 			}
 			DeleteImage(id);
 			if (m_total_used_memory < m_critical_gc_memory && aggressive) {

src/graphics/host_gpu/renderer/cache/textureCache.h

@@ -4,10 +4,11 @@
 #include "common/abi.h"
 #include "common/common.h"
 #include "common/lruCache.h"
-#include "graphics/host_gpu/memoryTracker.h"
-#include "graphics/host_gpu/renderer/blitHelper.h"
-#include "graphics/host_gpu/renderer/image.h"
-#include "graphics/host_gpu/renderer/multiLevelPageTable.h"
+#include "graphics/host_gpu/pageManager.h"
+#include "graphics/host_gpu/regionManager.h"
+#include "graphics/host_gpu/renderer/cache/multiLevelPageTable.h"
+#include "graphics/host_gpu/renderer/image/blitHelper.h"
+#include "graphics/host_gpu/renderer/image/image.h"
 
 #include <compare>
 #include <map>
@@ -64,8 +65,7 @@ public:
 
 	[[nodiscard]] bool ClearImageFromBuffer(CommandBuffer& command, uint64_t address, uint64_t size,
 	                                        uint32_t packed_clear);
-	void               PrepareHostWrite(uint64_t address, uint64_t size);
-	[[nodiscard]] bool SynchronizeImageToBuffer(uint64_t address, uint64_t size);
+	void               InvalidateMemory(uint64_t address, uint64_t size);
 	[[nodiscard]] bool InvalidateMemoryFromGPU(uint64_t address, uint64_t size,
 	                                           bool formatted_buffer_write = false);
 	[[nodiscard]] RegionInfo QueryRegion(uint64_t address, uint64_t size);
@@ -75,11 +75,9 @@ public:
 	[[nodiscard]] bool ClearMeta(uint64_t address);
 	[[nodiscard]] bool TouchMeta(uint64_t address, uint32_t slice, bool is_clear);
 
-	[[nodiscard]] bool InvalidateMemory(PageFaultAccess access, uint64_t address, uint64_t size,
-	                                    PageFaultPhase phase) noexcept;
-	void               UnmapMemory(uint64_t address, uint64_t size);
-	void               ProcessDownloadImages();
-	void               RunGarbageCollector();
+	void UnmapMemory(uint64_t address, uint64_t size);
+	void ProcessDownloadImages();
+	void RunGarbageCollector();
 
 private:
 	enum class TransferDirection { Upload, Download };
@@ -109,11 +107,17 @@ private:
 	[[nodiscard]] ImageId                InsertImage(const ImageInfo& info);
 	[[nodiscard]] ImageId                GetNullImage(const ImageDesc& desc);
 	void                                 RegisterImage(ImageId id);
-	void                                 UnregisterImage(ImageId id, bool release_tracking);
-	void                                 DeleteImage(ImageId id, bool release_tracking = true);
+	void                                 UnregisterImage(ImageId id);
+	void                                 DeleteImage(ImageId id);
 	void DeleteImages(std::span<const ImageId> ids, std::optional<ImageId> native_source = {});
 	void RetainImage(CommandBuffer& command, ImageId id);
 	void TouchImage(Image& image);
+	void TrackImage(ImageId id);
+	void TrackImageHead(ImageId id);
+	void TrackImageTail(ImageId id);
+	void UntrackImage(ImageId id);
+	void UntrackImageHead(ImageId id);
+	void UntrackImageTail(ImageId id);
 	void TrackImageDownload(ImageId id);
 	void TrackImageDownloadLocked(ImageId id, Image& image);
 	[[nodiscard]] static bool SameBacking(const ImageInfo& cached, const ImageInfo& requested,
@@ -135,7 +139,7 @@ private:
 	[[nodiscard]] DownloadPlan      BuildDownload(const Image& image) const;
 	void UploadImage(Image& image, const ImageDesc& desc, Buffer& source, uint64_t source_offset);
 	void DownloadImageData(Image& image, Buffer& destination, uint64_t destination_offset,
-	                       DownloadPlan plan);
+	                       uint64_t destination_size, DownloadPlan plan);
 	void DownloadDepth(Image& image, Buffer& destination, uint64_t destination_offset);
 	void CommitGpuWrite(Image& image);
 	void PrepareImageCopy(Image& image);
@@ -148,10 +152,8 @@ private:
 	void ValidateImageDesc(const ImageDesc& desc) const;
 
 	void InvalidateCpuAliases(uint64_t address, uint64_t size);
-	void RestoreGpuTracking(const Image& image);
-	void ReleaseGpuTracking(ImageId id);
+	void ClearGpuModified(ImageId id);
 
-	[[nodiscard]] bool                          SynchronizeImageToBuffer(ImageId id);
 	void                                        DownloadImage(ImageId id);
 	[[nodiscard]] bool                          TryDownloadImage(ImageId id);
 	[[nodiscard]] std::pair<uint8_t*, uint64_t> MapDownload(uint64_t size, uint64_t alignment);
@@ -160,7 +162,7 @@ private:
 	GraphicContext&                                   m_graphics;
 	CommandScheduler&                                 m_scheduler;
 	TrackingSpinLock                                  m_lock;
-	MemoryTracker                                     m_memory_tracker;
+	PageManager&                                      m_page_manager;
 	BlitHelper                                        m_blit_helper;
 	std::unique_ptr<TileManager>                      m_tiler;
 	BufferCache&                                      m_buffer_cache;
@@ -180,6 +182,7 @@ private:
 	bool     m_readback_linear_images = false;
 
 	friend struct TextureCacheTestAccess;
+	friend class BufferCache;
 	friend class RenderExecutor;
 };
 

src/graphics/host_gpu/renderer/colorRenderTarget.cpp

@@ -7,9 +7,9 @@
 #include "graphics/guest_gpu/hardwareContext.h"
 #include "graphics/guest_gpu/tile.h"
 #include "graphics/host_gpu/graphicContext.h"
-#include "graphics/host_gpu/objects/textureCommon.h"
 #include "graphics/host_gpu/renderer/debug.h"
-#include "graphics/host_gpu/renderer/descriptorCache.h"
+#include "graphics/host_gpu/renderer/image/textureCommon.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptorCache.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/vulkanCommon.h"
@@ -23,10 +23,10 @@ static std::atomic<uint32_t> g_render_color_log_count = 0;
 
 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
 void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandBuffer& buffer,
-                                          RenderColorInfo& r,
-                                          uint32_t render_target_slice_offset,
-                                          uint32_t render_target_slot, bool ignore_target_mask,
-                                          bool exact_format) {
+                                              RenderColorInfo& r,
+                                              uint32_t         render_target_slice_offset,
+                                              uint32_t render_target_slot, bool ignore_target_mask,
+                                              bool exact_format) {
 	KYTY_PROFILER_FUNCTION();
 	const auto& hw = buffer.GetRegisters();
 
@@ -79,10 +79,8 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 	const auto view = ResolveTargetViewInfo(
 	    rt.view.base_array_slice_index, rt.view.last_array_slice_index, render_target_slice_offset);
 	switch (view.type) {
-		case TargetViewType::Image2D: break;
-		case TargetViewType::Image2DArray:
-			EXIT("layered render-target views are unsupported: base=%u count=%u\n", view.base_layer,
-			     view.layer_count);
+		case TargetViewType::Image2D:
+		case TargetViewType::Image2DArray: break;
 		case TargetViewType::Unsupported:
 			EXIT("invalid render-target view: base=%u last=%u draw_offset=%u\n",
 			     rt.view.base_array_slice_index, rt.view.last_array_slice_index,
@@ -121,7 +119,18 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 	uint32_t   pitch  = 0;
 	uint64_t   size   = 0;
 	bool       tile   = false;
-	const bool standard64 =
+	const bool volume = rt.attrib3.dimension == 2;
+	if (rt.attrib3.dimension != 1 && !volume) {
+		EXIT("unsupported render-target dimension: %u\n", rt.attrib3.dimension);
+	}
+	if (!volume && rt.attrib3.depth != 0) {
+		EXIT("2D render target has nonzero depth: %u\n", rt.attrib3.depth);
+	}
+	if (volume && samples != 1) {
+		EXIT("multisampled 3D render targets are unsupported\n");
+	}
+	const uint32_t depth = volume ? rt.attrib3.depth + 1u : 1u;
+	const bool     standard64 =
 	    rt.attrib3.tile_mode == Prospero::GpuEnumValue(Prospero::TileMode::kStandard64KB);
 
 	switch (rt.attrib3.tile_mode) {
@@ -147,6 +156,7 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 	if (bytes_per_element == 0) {
 		EXIT("render-target format has no valid element size\n");
 	}
+	const auto transfer_format = ImageOps::RenderTargetTransferFormat(bytes_per_element);
 	if (standard64 &&
 	    (rt.attrib3.dimension != 1 || rt.attrib3.depth != 0 || levels != 1 ||
 	     rt.view.current_mip_level != 0 || view.base_layer != 0 || view.image_layers != 1 ||
@@ -167,10 +177,14 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 	if (rt.pitch.pitch_div8_minus1 != 0) {
 		pitch = (rt.pitch.pitch_div8_minus1 + 1u) << 3u;
 	} else if (tile) {
-		pitch = standard64
-		            ? TileGetTexturePitch(Prospero::GpuEnumValue(Prospero::BufferFormat::k32Float),
-		                                  width, levels, rt.attrib3.tile_mode)
-		            : TileGetRenderTargetPitch(width, bytes_per_element, rt.attrib.num_fragments);
+		if (volume) {
+			pitch = TileGetTexturePitch(transfer_format, width, levels, rt.attrib3.tile_mode);
+		} else if (standard64) {
+			pitch = TileGetTexturePitch(Prospero::GpuEnumValue(Prospero::BufferFormat::k32Float),
+			                            width, levels, rt.attrib3.tile_mode);
+		} else {
+			pitch = TileGetRenderTargetPitch(width, bytes_per_element, rt.attrib.num_fragments);
+		}
 		if (pitch == 0) {
 			EXIT("unsupported render-target pitch: width=%u bytes=%u\n", width, bytes_per_element);
 		}
@@ -178,9 +192,19 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 		pitch = width;
 	}
 
-	TileSizeOffset mip_sizes[16] {};
-	TilePaddedSize mip_padded[16] {};
-	if (tile) {
+	TileSizeOffset   mip_sizes[16] {};
+	TilePaddedSize   mip_padded[16] {};
+	TileVolumeLayout volume_layout {};
+	uint64_t         backing_size = 0;
+	if (volume) {
+		if (!tile || !TileGetTextureVolumeLayout(transfer_format, width, height, depth, levels,
+		                                         rt.attrib3.tile_mode, volume_layout)) {
+			EXIT("unsupported 3D render-target layout: %ux%ux%u levels=%u tile=%u\n", width, height,
+			     depth, levels, rt.attrib3.tile_mode);
+		}
+		size         = volume_layout.block_slice_size;
+		backing_size = volume_layout.total_size;
+	} else if (tile) {
 		TileSizeAlign layout {};
 		bool          valid_layout = false;
 		if (standard64) {
@@ -205,12 +229,6 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 			mip_sizes[0]  = {static_cast<uint32_t>(size), 0, 0, 0, 0, 0};
 			mip_padded[0] = {pitch, height};
 		}
-		if (rt.slice.slice_div64_minus1 != 0 &&
-		    (static_cast<uint64_t>(rt.slice.slice_div64_minus1) + 1u) * 64u != size) {
-			EXIT("render-target slice span mismatch: encoded=0x%016" PRIx64 " derived=0x%016" PRIx64
-			     "\n",
-			     (static_cast<uint64_t>(rt.slice.slice_div64_minus1) + 1u) * 64u, size);
-		}
 	} else {
 		size = static_cast<uint64_t>(pitch) * height * bytes_per_element * samples;
 		if (size > UINT32_MAX) {
@@ -219,23 +237,40 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 		mip_sizes[0]  = {static_cast<uint32_t>(size), 0, 0, 0, 0, 0};
 		mip_padded[0] = {pitch, height};
 	}
-	if (size == 0 || size > UINT64_MAX / view.image_layers) {
+	if (rt.slice.slice_div64_minus1 != 0 &&
+	    (static_cast<uint64_t>(rt.slice.slice_div64_minus1) + 1u) * 64u != size) {
+		EXIT("render-target slice span mismatch: encoded=0x%016" PRIx64 " derived=0x%016" PRIx64
+		     "\n",
+		     (static_cast<uint64_t>(rt.slice.slice_div64_minus1) + 1u) * 64u, size);
+	}
+	if (size == 0 || (!volume && size > UINT64_MAX / view.image_layers)) {
 		EXIT("render-target memory footprint is invalid\n");
 	}
-	const auto backing_size = size * view.image_layers;
+	if (!volume) {
+		backing_size = size * view.image_layers;
+	}
+	if (backing_size == 0) {
+		EXIT("render-target backing is empty\n");
+	}
 	if (backing_size > TRACKER_ADDRESS_SIZE - rt.base.addr) {
 		EXIT("render-target backing range is invalid\n");
 	}
 
 	const vk::Extent2D view_extent = {std::max(width >> rt.view.current_mip_level, 1u),
 	                                  std::max(height >> rt.view.current_mip_level, 1u)};
+	const uint32_t     view_depth  = std::max(depth >> rt.view.current_mip_level, 1u);
+	if (volume &&
+	    (view.base_layer >= view_depth || view.layer_count > view_depth - view.base_layer)) {
+		EXIT("3D render-target view exceeds mip depth: base=%u count=%u depth=%u mip=%u\n",
+		     view.base_layer, view.layer_count, view_depth, rt.view.current_mip_level);
+	}
 
 	auto decision_log_id = g_render_color_log_count.fetch_add(1);
 	if (decision_log_id < 128) {
 		LOGF("RenderColorTarget: slot=%" PRIu32 " addr=0x%010" PRIx64 " size=0x%016" PRIx64
-		     " extent=%ux%u view_mip=%u view_extent=%ux%u levels=%u pitch=%u"
+		     " extent=%ux%ux%u view_mip=%u view_extent=%ux%u levels=%u pitch=%u"
 		     " fmt=0x%08" PRIx32 " nfmt=0x%08" PRIx32 " order=0x%08" PRIx32 " samples=%u tile=%s\n",
-		     rt_slot, rt.base.addr, backing_size, width, height, rt.view.current_mip_level,
+		     rt_slot, rt.base.addr, backing_size, width, height, depth, rt.view.current_mip_level,
 		     view_extent.width, view_extent.height, levels, pitch, rt.info.format,
 		     rt.info.channel_type, rt.info.channel_order, samples, tile ? "tiled" : "linear");
 	}
@@ -244,15 +279,24 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 	desc.type              = TextureCache::BindingType::RenderTarget;
 	desc.info.data         = {rt.base.addr, backing_size};
 	desc.info.pixel_format = target_format.format;
-	desc.info.guest_format = ImageOps::RenderTargetTransferFormat(bytes_per_element);
-	desc.info.type   = Prospero::ImageType::kColor2D;
-	desc.info.extent = {width, height, 1};
-	desc.info.resources       = {levels, view.image_layers};
-	desc.info.pitch           = pitch;
+	desc.info.guest_format = transfer_format;
+	desc.info.type         = volume ? Prospero::ImageType::kColor3D : Prospero::ImageType::kColor2D;
+	desc.info.extent       = {width, height, depth};
+	desc.info.resources    = {levels, volume ? 1u : view.image_layers};
+	desc.info.pitch        = pitch;
 	desc.info.bytes_per_block = bytes_per_element;
 	desc.info.samples         = samples;
 	desc.info.tile_mode       = rt.attrib3.tile_mode;
 	for (uint32_t level = 0; level < levels; level++) {
+		if (volume) {
+			desc.info.mip_layout[level] = {
+			    volume_layout.level_offsets[level],
+			    volume_layout.level_sizes[level],
+			    volume_layout.level_widths[level],
+			    volume_layout.level_heights[level],
+			};
+			continue;
+		}
 		const auto level_offset =
 		    mip_sizes[level].src_size != 0 ? mip_sizes[level].src_offset : mip_sizes[level].offset;
 		const auto level_size =
@@ -275,20 +319,20 @@ void RenderExecutor::ResolveRenderColorTarget(uint64_t submit_id, RenderCommandB
 	desc.view_info.base_layer  = view.base_layer;
 	desc.view_info.layer_count = view.layer_count;
 	desc.view_info.usage       = vk::ImageUsageFlagBits::eColorAttachment;
-	auto& texture_cache = m_context.GetTextureCache();
-	r.desc              = std::move(desc);
-	r.image_id          = texture_cache.FindImage(r.desc, exact_format);
-	r.type              = RenderColorType::RenderTexture;
-	r.base_addr         = rt.base.addr;
-	r.image_view        = nullptr;
-	r.format            = r.desc.view_info.format;
-	r.extent            = view_extent;
-	r.base_mip_level    = rt.view.current_mip_level;
-	r.buffer_size       = backing_size;
-	r.samples           = samples;
-	r.export_mapping    = target_format.export_mapping;
-	r.color_clear_enable = false;
-	r.color_clear_value = {};
+	auto& texture_cache        = m_context.GetTextureCache();
+	r.desc                     = std::move(desc);
+	r.image_id                 = texture_cache.FindImage(r.desc, exact_format);
+	r.type                     = RenderColorType::RenderTexture;
+	r.base_addr                = rt.base.addr;
+	r.image_view               = nullptr;
+	r.format                   = r.desc.view_info.format;
+	r.extent                   = view_extent;
+	r.base_mip_level           = rt.view.current_mip_level;
+	r.buffer_size              = backing_size;
+	r.samples                  = samples;
+	r.export_mapping           = target_format.export_mapping;
+	r.color_clear_enable       = false;
+	r.color_clear_value        = {};
 	BindRenderTarget(r.image_id);
 }
 

src/graphics/host_gpu/renderer/colorRenderTarget.h

@@ -2,8 +2,8 @@
 #define EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_COLORRENDERTARGET_H_
 
 #include "graphics/guest_gpu/gpu_defs.h"
+#include "graphics/host_gpu/renderer/cache/textureCache.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
-#include "graphics/host_gpu/renderer/textureCache.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 
 #include <cstdint>

src/graphics/host_gpu/renderer/commandScheduler.cpp

@@ -44,13 +44,13 @@ CommandSlot* CommandScheduler::CommandPool::CreateSlot() {
 	allocate.commandPool        = m_pool;
 	allocate.level              = vk::CommandBufferLevel::ePrimary;
 	allocate.commandBufferCount = 1;
-	vk::CommandBuffer buffer = nullptr;
+	vk::CommandBuffer buffer    = nullptr;
 	EXIT_IF(graphics.device.allocateCommandBuffers(&allocate, &buffer) != vk::Result::eSuccess);
 
 	vk::FenceCreateInfo fence_create {};
 	fence_create.sType = vk::StructureType::eFenceCreateInfo;
 	fence_create.flags = vk::FenceCreateFlagBits::eSignaled;
-	vk::Fence fence = nullptr;
+	vk::Fence fence    = nullptr;
 	if (graphics.device.createFence(&fence_create, nullptr, &fence) != vk::Result::eSuccess) {
 		graphics.device.freeCommandBuffers(m_pool, 1, &buffer);
 		EXIT("failed to create command-buffer fence\n");
@@ -70,9 +70,9 @@ CommandSlot* CommandScheduler::CommandPool::Allocate(GraphicContext& graphics) {
 		Create(graphics);
 	}
 	EXIT_IF(m_graphics != &graphics);
-	auto found = std::ranges::find_if(m_slots, [](const auto& slot) { return !slot.busy; });
-	auto* slot = found != m_slots.end() ? &*found : CreateSlot();
-	slot->busy = true;
+	auto  found = std::ranges::find_if(m_slots, [](const auto& slot) { return !slot.busy; });
+	auto* slot  = found != m_slots.end() ? &*found : CreateSlot();
+	slot->busy  = true;
 	slot->Reset();
 	return slot;
 }
@@ -331,8 +331,7 @@ void CommandScheduler::WaitPriorityOperations(uint64_t tick) {
 	EXIT_IF(g_deferred_callback_scheduler == this);
 	std::unique_lock lock(m_operation_mutex);
 	m_operation_available.wait(lock, [this, tick] {
-		const bool active_before_or_at =
-		    m_priority_active && m_priority_active_tick <= tick;
+		const bool active_before_or_at = m_priority_active && m_priority_active_tick <= tick;
 		const bool queued_before_or_at =
 		    !m_priority_operations.empty() && m_priority_operations.front().tick <= tick;
 		return !active_before_or_at && !queued_before_or_at;

src/graphics/host_gpu/renderer/commandScheduler.h

@@ -47,21 +47,21 @@ public:
 	void           FinishCurrent();
 	// Deferred callbacks can observe an externally owned drain, but cannot initiate shutdown:
 	// the priority runner cannot join itself.
-	void Shutdown();
-	void Wait(uint64_t tick);
-	void PopPendingOperations();
-	void DrainPriorityOperations();
-	void WaitPriorityOperations(uint64_t tick);
-	void DeferOperation(Common::UniqueFunction<void>&& operation);
-	void DeferPriorityOperation(Common::UniqueFunction<void>&& operation);
+	void                      Shutdown();
+	void                      Wait(uint64_t tick);
+	void                      PopPendingOperations();
+	void                      DrainPriorityOperations();
+	void                      WaitPriorityOperations(uint64_t tick);
+	void                      DeferOperation(Common::UniqueFunction<void>&& operation);
+	void                      DeferPriorityOperation(Common::UniqueFunction<void>&& operation);
 	[[nodiscard]] static bool InDeferredOperation() noexcept;
 
-	[[nodiscard]] bool     Active() const noexcept { return m_current >= 0; }
-	void                   CheckActive() const;
-	RenderCommandBuffer&   Current() const;
-	[[nodiscard]] uint64_t CurrentTick() const noexcept { return m_master.CurrentTick(); }
-	[[nodiscard]] bool     IsFree(uint64_t tick);
-	[[nodiscard]] RenderContext& Context() const noexcept { return m_context; }
+	[[nodiscard]] bool            Active() const noexcept { return m_current >= 0; }
+	void                          CheckActive() const;
+	RenderCommandBuffer&          Current() const;
+	[[nodiscard]] uint64_t        CurrentTick() const noexcept { return m_master.CurrentTick(); }
+	[[nodiscard]] bool            IsFree(uint64_t tick);
+	[[nodiscard]] RenderContext&  Context() const noexcept { return m_context; }
 	[[nodiscard]] GraphicContext& Graphics() const noexcept { return m_graphics; }
 
 private:
@@ -91,11 +91,11 @@ private:
 		uint64_t                     tick = 0;
 	};
 
-	void           BindCurrent() const;
-	CommandBuffer& SubmitCurrent(SubmitInfo& submit);
-	void           BeginNext();
-	void           PriorityOperationsThread(std::stop_token stop);
-	void           RunOperation(Common::UniqueFunction<void>&& operation);
+	void                       BindCurrent() const;
+	CommandBuffer&             SubmitCurrent(SubmitInfo& submit);
+	void                       BeginNext();
+	void                       PriorityOperationsThread(std::stop_token stop);
+	void                       RunOperation(Common::UniqueFunction<void>&& operation);
 	[[nodiscard]] CommandSlot* AllocateCommandBuffer();
 	[[nodiscard]] uint64_t     NextSubmitSequence() noexcept;
 
@@ -109,14 +109,14 @@ private:
 	std::mutex                                                    m_operation_mutex;
 	std::condition_variable                                       m_operation_available;
 	std::jthread                                                  m_priority_thread;
-	bool                                                          m_priority_active = false;
+	bool                                                          m_priority_active      = false;
 	uint64_t                                                      m_priority_active_tick = 0;
-	OperationState  m_operation_state = OperationState::Open;
-	int             m_current         = -1;
-	bool            m_recording       = false;
-	HW::Context*    m_registers       = nullptr;
-	HW::UserConfig* m_user_config     = nullptr;
-	HW::Shader*     m_shaders         = nullptr;
+	OperationState        m_operation_state = OperationState::Open;
+	int                   m_current         = -1;
+	bool                  m_recording       = false;
+	HW::Context*          m_registers       = nullptr;
+	HW::UserConfig*       m_user_config     = nullptr;
+	HW::Shader*           m_shaders         = nullptr;
 	std::atomic<uint64_t> m_submit_sequence = 0;
 
 	friend class CommandBuffer;

src/graphics/host_gpu/renderer/context.cpp

@@ -8,8 +8,8 @@
 #include "graphics/host_gpu/renderer/colorRenderTarget.h"
 #include "graphics/host_gpu/renderer/debug.h"
 #include "graphics/host_gpu/renderer/depthRenderTarget.h"
-#include "graphics/host_gpu/renderer/descriptorCache.h"
-#include "graphics/host_gpu/renderer/imageView.h"
+#include "graphics/host_gpu/renderer/image/imageView.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptorCache.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/vma.h"
@@ -270,30 +270,30 @@ void CommandBuffer::BeginRendering(const RenderState& state) const {
 		colors[i].sType        = vk::StructureType::eRenderingAttachmentInfo;
 		colors[i].imageView    = attachment.image_view;
 		colors[i].imageLayout  = attachment.image_layout;
-		colors[i].loadOp = attachment.is_clear ? vk::AttachmentLoadOp::eClear
-		                                      : vk::AttachmentLoadOp::eLoad;
-		colors[i].storeOp                  = vk::AttachmentStoreOp::eStore;
-		colors[i].clearValue.color.uint32  = attachment.clear_value;
+		colors[i].loadOp =
+		    attachment.is_clear ? vk::AttachmentLoadOp::eClear : vk::AttachmentLoadOp::eLoad;
+		colors[i].storeOp                 = vk::AttachmentStoreOp::eStore;
+		colors[i].clearValue.color.uint32 = attachment.clear_value;
 	}
 
-	const auto& depth_stencil = state.depth_stencil_attachment;
+	const auto&                 depth_stencil = state.depth_stencil_attachment;
 	vk::RenderingAttachmentInfo depth {};
 	depth.sType       = vk::StructureType::eRenderingAttachmentInfo;
 	depth.imageView   = depth_stencil.image_view;
 	depth.imageLayout = depth_stencil.image_layout;
-	depth.loadOp = depth_stencil.depth_clear ? vk::AttachmentLoadOp::eClear
-	                                        : vk::AttachmentLoadOp::eLoad;
-	depth.storeOp                        = vk::AttachmentStoreOp::eStore;
+	depth.loadOp =
+	    depth_stencil.depth_clear ? vk::AttachmentLoadOp::eClear : vk::AttachmentLoadOp::eLoad;
+	depth.storeOp                       = vk::AttachmentStoreOp::eStore;
 	depth.clearValue.depthStencil.depth = std::bit_cast<float>(depth_stencil.clear_value[0]);
 
 	vk::RenderingAttachmentInfo stencil {};
 	stencil.sType       = vk::StructureType::eRenderingAttachmentInfo;
 	stencil.imageView   = depth_stencil.image_view;
 	stencil.imageLayout = depth_stencil.image_layout;
-	stencil.loadOp = depth_stencil.stencil_clear ? vk::AttachmentLoadOp::eClear
-	                                            : vk::AttachmentLoadOp::eLoad;
-	stencil.storeOp                            = vk::AttachmentStoreOp::eStore;
-	stencil.clearValue.depthStencil.stencil   = depth_stencil.clear_value[1];
+	stencil.loadOp =
+	    depth_stencil.stencil_clear ? vk::AttachmentLoadOp::eClear : vk::AttachmentLoadOp::eLoad;
+	stencil.storeOp                         = vk::AttachmentStoreOp::eStore;
+	stencil.clearValue.depthStencil.stencil = depth_stencil.clear_value[1];
 
 	vk::RenderingInfo rendering {};
 	rendering.sType                = vk::StructureType::eRenderingInfo;

src/graphics/host_gpu/renderer/debug.cpp

@@ -359,30 +359,12 @@ static void RtCheck(const HW::RenderTarget& rt) {
 				logged = true;
 			}
 		}
-		if (rt.attrib3.depth != 0x00000000) {
-			static bool logged = false;
-			if (!logged) {
-				LOGF("RenderTarget: temporary: ignoring PS5 color target depth_minus1=0x%08" PRIx32
-				     "\n",
-				     rt.attrib3.depth);
-				logged = true;
-			}
-		}
 		if (!RenderIsColorTileMode(rt.attrib3.tile_mode)) {
 			EXIT("unknown PS5 render-target tile mode: 0x%08" PRIx32 "\n", rt.attrib3.tile_mode);
 		}
 		if (!RenderIsColorDimension(rt.attrib3.dimension)) {
 			EXIT("unknown PS5 render-target dimension: 0x%08" PRIx32 "\n", rt.attrib3.dimension);
 		}
-		if (rt.attrib3.dimension != 0x00000001) {
-			static bool logged = false;
-			if (!logged) {
-				LOGF("RenderTarget: temporary: using 2D fallback for PS5 color "
-				     "dimension=0x%08" PRIx32 "\n",
-				     rt.attrib3.dimension);
-				logged = true;
-			}
-		}
 		if (!rt.attrib3.cmask_pipe_aligned) {
 			static bool logged = false;
 			if (!logged) {
@@ -497,7 +479,15 @@ static void ZPrint(const char* func, const HW::DepthRenderTarget& z) {
 }
 
 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
-static void ZCheck(const HW::DepthRenderTarget& z) {
+static void ZCheck(const HW::DepthRenderTarget& z, const HW::DepthControl& dc,
+                   const HW::RenderControl& rc) {
+	const bool depth_active =
+	    dc.z_enable || dc.z_write_enable || dc.depth_bounds_enable || rc.depth_clear_enable;
+	const bool stencil_active = dc.stencil_enable || rc.stencil_clear_enable;
+	if (!depth_active && !stencil_active) {
+		return;
+	}
+
 	EXIT_NOT_IMPLEMENTED(!z.z_info.HasValidTextureCompatibility());
 	EXIT_NOT_IMPLEMENTED(!z.stencil_info.HasValidTextureCompatibility());
 	if (z.z_info.format == 0) {
@@ -548,14 +538,6 @@ static void ZCheck(const HW::DepthRenderTarget& z) {
 		EXIT_NOT_IMPLEMENTED(z.htile_surface.prefetch_height != 0x00000000);
 		EXIT_NOT_IMPLEMENTED(z.htile_surface.dst_outside_zero_to_one != 0x00000000);
 
-		if (z.depth_view.slice_start != 0x00000000 || z.depth_view.slice_max != 0x00000000) {
-			static std::atomic<uint32_t> log_count {0};
-			if (log_count.fetch_add(1, std::memory_order_relaxed) < 16) {
-				LOGF("DepthTarget: temporary: ignoring PS5 array slice view start=0x%08" PRIx32
-				     ", max=0x%08" PRIx32 "\n",
-				     z.depth_view.slice_start, z.depth_view.slice_max);
-			}
-		}
 		if (z.depth_view.current_mip_level != 0x00000000) {
 			static std::atomic<uint32_t> log_count {0};
 			if (log_count.fetch_add(1, std::memory_order_relaxed) < 16) {
@@ -1214,7 +1196,7 @@ void hw_check(const RenderCommandBuffer& buffer) {
 	log_phase("vp");
 	VpCheck(vp, smc);
 	log_phase("z");
-	ZCheck(z);
+	ZCheck(z, d, rc);
 	log_phase("clip");
 	ClipCheck(c);
 	log_phase("rc");

src/graphics/host_gpu/renderer/depthRenderTarget.cpp

@@ -10,10 +10,10 @@
 #include "graphics/guest_gpu/hardwareContext.h"
 #include "graphics/guest_gpu/tile.h"
 #include "graphics/host_gpu/graphicContext.h"
-#include "graphics/host_gpu/objects/textureCommon.h"
 #include "graphics/host_gpu/renderer/debug.h"
-#include "graphics/host_gpu/renderer/descriptorCache.h"
-#include "graphics/host_gpu/renderer/imageView.h"
+#include "graphics/host_gpu/renderer/image/imageView.h"
+#include "graphics/host_gpu/renderer/image/textureCommon.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptorCache.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/vulkanCommon.h"
@@ -150,10 +150,8 @@ void RenderExecutor::ResolveRenderDepthTarget(uint64_t submit_id, RenderCommandB
 	    has_stencil, has_htile, z.stencil_info.htile_stencil_disabled);
 	const auto view = ResolveTargetViewInfo(z.depth_view.slice_start, z.depth_view.slice_max);
 	switch (view.type) {
-		case TargetViewType::Image2D: break;
-		case TargetViewType::Image2DArray:
-			DepthFatal("layered depth views are unsupported: base=%u count=%u", view.base_layer,
-			           view.layer_count);
+		case TargetViewType::Image2D:
+		case TargetViewType::Image2DArray: break;
 		case TargetViewType::Unsupported:
 			DepthFatal("invalid depth view: base=%u last=%u", z.depth_view.slice_start,
 			           z.depth_view.slice_max);

src/graphics/host_gpu/renderer/depthRenderTarget.h

@@ -2,9 +2,9 @@
 #define EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_DEPTHRENDERTARGET_H_
 
 #include "common/assert.h"
-#include "graphics/host_gpu/renderer/imageView.h"
+#include "graphics/host_gpu/renderer/cache/textureCache.h"
+#include "graphics/host_gpu/renderer/image/imageView.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
-#include "graphics/host_gpu/renderer/textureCache.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 
 #include <cstdint>

src/graphics/host_gpu/renderer/gpuResourceManager.cpp

@@ -1,141 +0,0 @@
-#include "graphics/host_gpu/renderer/gpuResourceManager.h"
-
-#include "common/assert.h"
-#include "graphics/guest_gpu/command_processor/commandProcessor.h"
-#include "graphics/guest_gpu/graphicsRun.h"
-#include "graphics/host_gpu/renderer/commandScheduler.h"
-
-namespace Libs::Graphics {
-
-GpuResourceManager::GpuResourceManager(GraphicContext& graphics, CommandScheduler& scheduler)
-    : m_page_manager(FaultThunk, this),
-      m_buffer_cache(graphics, scheduler, m_page_manager, m_texture_cache, m_resource_mutex),
-      m_texture_cache(graphics, scheduler, m_page_manager, m_buffer_cache, m_resource_mutex) {}
-
-GpuResourceManager::~GpuResourceManager() = default;
-
-bool GpuResourceManager::FaultThunk(void* context, PageFaultAccess access, uint64_t vaddr,
-                                    uint64_t size, PageFaultPhase phase) noexcept {
-	return static_cast<GpuResourceManager*>(context)->InvalidateMemory(access, vaddr, size, phase);
-}
-
-bool GpuResourceManager::InvalidateMemory(PageFaultAccess access, uint64_t vaddr, uint64_t size,
-                                          PageFaultPhase phase) noexcept {
-	// Let the authoritative image materialize first. A clean overlapping buffer marks a write
-	// fault CPU-dirty when it begins ownership transfer; doing that before image preflight would
-	// make the image appear to race a real CPU write. Completion and release retain buffer-first
-	// ordering so its pending fault is gone before TextureCache publishes the downloaded backing.
-	if (phase == PageFaultPhase::Invalidate) {
-		const bool image_handled  = m_texture_cache.InvalidateMemory(access, vaddr, size, phase);
-		const bool buffer_handled = m_buffer_cache.InvalidateMemory(access, vaddr, size, phase);
-		return buffer_handled || image_handled;
-	}
-	const bool buffer_handled = m_buffer_cache.InvalidateMemory(access, vaddr, size, phase);
-	const bool image_handled  = m_texture_cache.InvalidateMemory(access, vaddr, size, phase);
-	return buffer_handled || image_handled;
-}
-
-bool GpuResourceManager::HandleFault(PageFaultAccess access, uint64_t fault_vaddr) noexcept {
-	if (!m_page_manager.IsMapped(fault_vaddr, 1)) {
-		return false;
-	}
-	if (CommandScheduler::InDeferredOperation()) {
-		EXIT("unsupported guest-memory fault from an asynchronous GPU completion, "
-		     "addr=0x%016" PRIx64 " access=%u\n",
-		     fault_vaddr, static_cast<uint32_t>(access));
-	}
-	bool       handled = false;
-	const auto resolve = [this, access, fault_vaddr, &handled](CommandProcessor& cp) {
-		cp.BeginReadbackTransaction();
-		(void)m_buffer_cache.SynchronizeBacking(fault_vaddr, 1);
-		{
-			ResourceMutex::FaultScope fault(m_resource_mutex);
-			handled = m_page_manager.HandleFault(access, fault_vaddr);
-		}
-		cp.EndReadbackTransaction();
-	};
-	if (auto* cp = Gpu::CurrentCommandProcessor(); cp != nullptr) {
-		resolve(*cp);
-		return handled;
-	}
-	if (m_resource_mutex.IsOwnedByCurrentThread()) {
-		EXIT("unsupported page fault from a pre-owned resource transaction, addr=0x%016" PRIx64
-		     " access=%u\n",
-		     fault_vaddr, static_cast<uint32_t>(access));
-	}
-	EXIT_IF(m_gpu == nullptr);
-	m_gpu->SendCommandSyncWithProcessor(resolve);
-	return handled;
-}
-
-void GpuResourceManager::PrepareHostWrite(uint64_t vaddr, uint64_t size) {
-	if (!m_page_manager.HasAnyMapping(vaddr, size)) {
-		return;
-	}
-	if (CommandScheduler::InDeferredOperation()) {
-		EXIT("unsupported host write from an asynchronous GPU completion, addr=0x%016" PRIx64
-		     " size=0x%016" PRIx64 "\n",
-		     vaddr, size);
-	}
-	const auto handle_range = [this, vaddr, size] {
-		if (!m_page_manager.HandleWriteRange(vaddr, size)) {
-			EXIT("failed to prepare host write, addr=0x%016" PRIx64 " size=0x%016" PRIx64 "\n",
-			     vaddr, size);
-		}
-	};
-	const auto resolve = [this, &handle_range](CommandProcessor& cp) {
-		cp.BeginReadbackTransaction();
-		{
-			ResourceMutex::FaultScope fault(m_resource_mutex);
-			handle_range();
-		}
-		cp.EndReadbackTransaction();
-	};
-	if (auto* cp = Gpu::CurrentCommandProcessor(); cp != nullptr) {
-		resolve(*cp);
-		return;
-	}
-	if (m_resource_mutex.IsOwnedByCurrentThread()) {
-		EXIT("unsupported host write from a pre-owned resource transaction, addr=0x%016" PRIx64
-		     " size=0x%016" PRIx64 "\n",
-		     vaddr, size);
-	}
-	EXIT_IF(m_gpu == nullptr);
-	m_gpu->SendCommandSyncWithProcessor(resolve);
-}
-
-bool GpuResourceManager::IsMapped(uint64_t vaddr, uint64_t size) const noexcept {
-	return m_page_manager.IsMapped(vaddr, size);
-}
-
-void GpuResourceManager::MapMemory(uint64_t vaddr, uint64_t size, GpuAccess access) {
-	m_page_manager.OnGpuMap(vaddr, size, access);
-}
-
-void GpuResourceManager::UnmapMemory(uint64_t vaddr, uint64_t size, GpuAccess access) {
-	if (!IsMapped(vaddr, size)) {
-		EXIT("cannot unmap an unmapped GPU resource range\n");
-	}
-	const auto unmap = [this, vaddr, size, access] {
-		m_texture_cache.UnmapMemory(vaddr, size);
-		m_buffer_cache.UnmapMemory(vaddr, size);
-		m_page_manager.OnGpuUnmap(vaddr, size, access);
-	};
-	if (m_gpu == nullptr) {
-		if (m_resource_mutex.IsOwnedByCurrentThread()) {
-			EXIT("cannot synchronously unmap from a resource transaction\n");
-		}
-		unmap();
-		return;
-	}
-	Gpu::SubmissionLock submissions(*m_gpu);
-	m_gpu->SendCommandSync(unmap);
-}
-
-void GpuResourceManager::RunGarbageCollector() {
-	m_texture_cache.ProcessDownloadImages();
-	m_texture_cache.RunGarbageCollector();
-	m_buffer_cache.RunGarbageCollector();
-}
-
-} // namespace Libs::Graphics

src/graphics/host_gpu/renderer/image/blitHelper.cpp

@@ -1,11 +1,11 @@
-#include "graphics/host_gpu/renderer/blitHelper.h"
+#include "graphics/host_gpu/renderer/image/blitHelper.h"
 
 #include "common/assert.h"
 #include "gpu_blit_shaders/gpu_blit_color_to_ms_depth_spv.h"
 #include "gpu_blit_shaders/gpu_blit_fs_triangle_spv.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
-#include "graphics/host_gpu/renderer/image.h"
+#include "graphics/host_gpu/renderer/image/image.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
 
 #include <algorithm>
@@ -182,8 +182,8 @@ void BlitHelper::ReinterpretColorAsMsDepth(Image& source, Image& destination) {
 	auto  command        = command_buffer.Handle();
 	source.Transit(vk::ImageLayout::eShaderReadOnlyOptimal, vk::AccessFlagBits2::eShaderRead, {},
 	               command);
-	destination.Transit(ColorToMsDepthLayout,
-	                    vk::AccessFlagBits2::eDepthStencilAttachmentWrite, {}, command);
+	destination.Transit(ColorToMsDepthLayout, vk::AccessFlagBits2::eDepthStencilAttachmentWrite, {},
+	                    command);
 
 	vk::RenderingAttachmentInfo depth_attachment {};
 	depth_attachment.sType                   = vk::StructureType::eRenderingAttachmentInfo;

src/graphics/host_gpu/renderer/image/image.cpp

@@ -1,11 +1,11 @@
-#include "graphics/host_gpu/renderer/image.h"
+#include "graphics/host_gpu/renderer/image/image.h"
 
 #include "common/assert.h"
 #include "common/profiler.h"
+#include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
-#include "graphics/host_gpu/renderer/imageView.h"
+#include "graphics/host_gpu/renderer/image/imageView.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
-#include "graphics/host_gpu/renderer/streamBuffer.h"
 #include "kernel/memory.h"
 
 #include <algorithm>
@@ -99,12 +99,16 @@ vk::ImageAspectFlags Image::FullAspectMask(vk::Format format) noexcept {
 	}
 }
 
-Image::Barriers Image::GetBarriers(vk::ImageLayout destination_layout,
-                                   vk::AccessFlags2 destination_access,
-                                   vk::PipelineStageFlags2 destination_stage,
+Image::Barriers Image::GetBarriers(vk::ImageLayout                      destination_layout,
+                                   vk::AccessFlags2                     destination_access,
+                                   vk::PipelineStageFlags2              destination_stage,
                                    std::optional<ImageSubresourceRange> range) {
 	auto& state              = backing.state;
 	auto& subresource_states = backing.subresource_states;
+	if (range && info.IsVolume()) {
+		range->base_layer  = 0;
+		range->layer_count = 1;
+	}
 
 	const bool partial =
 	    range && (range->base_level != 0 || range->level_count != info.resources.levels ||
@@ -130,25 +134,25 @@ Image::Barriers Image::GetBarriers(vk::ImageLayout destination_layout,
 				constexpr auto write_access = vk::AccessFlagBits2::eTransferWrite |
 				                              vk::AccessFlagBits2::eShaderWrite |
 				                              vk::AccessFlagBits2::eMemoryWrite;
-				const bool repeated_write =
+				const bool     repeated_write =
 				    static_cast<bool>(subresource_state.access_mask & write_access);
 				if (subresource_state.layout != destination_layout ||
 				    subresource_state.access_mask != destination_access || repeated_write) {
 					vk::ImageMemoryBarrier2 barrier {};
-					barrier.srcStageMask                          = subresource_state.pl_stage;
-					barrier.srcAccessMask                         = subresource_state.access_mask;
-					barrier.dstStageMask                          = destination_stage;
-					barrier.dstAccessMask                         = destination_access;
-					barrier.oldLayout                             = subresource_state.layout;
-					barrier.newLayout                             = destination_layout;
-					barrier.srcQueueFamilyIndex                   = VK_QUEUE_FAMILY_IGNORED;
-					barrier.dstQueueFamilyIndex                   = VK_QUEUE_FAMILY_IGNORED;
-					barrier.image                                 = backing.image;
-					barrier.subresourceRange.aspectMask           = FullAspectMask(backing.format);
-					barrier.subresourceRange.baseMipLevel         = level;
-					barrier.subresourceRange.levelCount           = 1;
-					barrier.subresourceRange.baseArrayLayer       = layer;
-					barrier.subresourceRange.layerCount           = 1;
+					barrier.srcStageMask                    = subresource_state.pl_stage;
+					barrier.srcAccessMask                   = subresource_state.access_mask;
+					barrier.dstStageMask                    = destination_stage;
+					barrier.dstAccessMask                   = destination_access;
+					barrier.oldLayout                       = subresource_state.layout;
+					barrier.newLayout                       = destination_layout;
+					barrier.srcQueueFamilyIndex             = VK_QUEUE_FAMILY_IGNORED;
+					barrier.dstQueueFamilyIndex             = VK_QUEUE_FAMILY_IGNORED;
+					barrier.image                           = backing.image;
+					barrier.subresourceRange.aspectMask     = FullAspectMask(backing.format);
+					barrier.subresourceRange.baseMipLevel   = level;
+					barrier.subresourceRange.levelCount     = 1;
+					barrier.subresourceRange.baseArrayLayer = layer;
+					barrier.subresourceRange.layerCount     = 1;
 					barriers.push_back(barrier);
 					subresource_state = {destination_stage, destination_access, destination_layout};
 				}
@@ -159,10 +163,10 @@ Image::Barriers Image::GetBarriers(vk::ImageLayout destination_layout,
 			subresource_states.clear();
 		}
 	} else {
-		constexpr auto write_access = vk::AccessFlagBits2::eTransferWrite |
-		                              vk::AccessFlagBits2::eShaderWrite |
-		                              vk::AccessFlagBits2::eMemoryWrite;
-		const bool repeated_write = static_cast<bool>(state.access_mask & write_access);
+		constexpr auto write_access   = vk::AccessFlagBits2::eTransferWrite |
+		                                vk::AccessFlagBits2::eShaderWrite |
+		                                vk::AccessFlagBits2::eMemoryWrite;
+		const bool     repeated_write = static_cast<bool>(state.access_mask & write_access);
 		if (state.layout == destination_layout && state.access_mask == destination_access &&
 		    !repeated_write) {
 			return {};
@@ -191,8 +195,7 @@ Image::Barriers Image::GetBarriers(vk::ImageLayout destination_layout,
 }
 
 void Image::Transit(vk::ImageLayout destination_layout, vk::AccessFlags2 destination_access,
-                    std::optional<ImageSubresourceRange> range,
-                    vk::CommandBuffer command_buffer) {
+                    std::optional<ImageSubresourceRange> range, vk::CommandBuffer command_buffer) {
 	const auto transfer_access =
 	    vk::AccessFlagBits2::eTransferRead | vk::AccessFlagBits2::eTransferWrite;
 	vk::PipelineStageFlags2 destination_stage {};
@@ -201,8 +204,8 @@ void Image::Transit(vk::ImageLayout destination_layout, vk::AccessFlags2 destina
 	}
 	if (!destination_access ||
 	    static_cast<bool>(destination_access & ~vk::AccessFlags2 {transfer_access})) {
-		destination_stage |= vk::PipelineStageFlagBits2::eAllGraphics |
-		                     vk::PipelineStageFlagBits2::eComputeShader;
+		destination_stage |=
+		    vk::PipelineStageFlagBits2::eAllGraphics | vk::PipelineStageFlagBits2::eComputeShader;
 	}
 	const auto barriers =
 	    GetBarriers(destination_layout, destination_access, destination_stage, range);
@@ -218,10 +221,9 @@ void Image::Transit(vk::ImageLayout destination_layout, vk::AccessFlags2 destina
 	command_buffer.pipelineBarrier2(dependency);
 }
 
-void Image::Upload(std::span<const vk::BufferImageCopy> copies, vk::Buffer buffer,
-                   uint64_t offset, uint64_t size) {
-	EXIT_IF(m_scheduler == nullptr || copies.empty() || buffer == nullptr ||
-	        size == 0);
+void Image::Upload(std::span<const vk::BufferImageCopy> copies, vk::Buffer buffer, uint64_t offset,
+                   uint64_t size) {
+	EXIT_IF(m_scheduler == nullptr || copies.empty() || buffer == nullptr || size == 0);
 	m_scheduler->EndRendering();
 	vk::BufferMemoryBarrier2 buffer_barrier {};
 	buffer_barrier.srcStageMask        = vk::PipelineStageFlagBits2::eAllCommands;
@@ -234,16 +236,15 @@ void Image::Upload(std::span<const vk::BufferImageCopy> copies, vk::Buffer buffe
 	buffer_barrier.offset              = offset;
 	buffer_barrier.size                = size;
 	const auto image_barriers =
-	    GetBarriers(vk::ImageLayout::eTransferDstOptimal,
-	                vk::AccessFlagBits2::eTransferWrite,
+	    GetBarriers(vk::ImageLayout::eTransferDstOptimal, vk::AccessFlagBits2::eTransferWrite,
 	                vk::PipelineStageFlagBits2::eCopy, {});
 	vk::DependencyInfo dependency {};
-	dependency.dependencyFlags           = vk::DependencyFlagBits::eByRegion;
+	dependency.dependencyFlags          = vk::DependencyFlagBits::eByRegion;
 	dependency.bufferMemoryBarrierCount = 1;
 	dependency.pBufferMemoryBarriers    = &buffer_barrier;
-	dependency.imageMemoryBarrierCount = static_cast<uint32_t>(image_barriers.size());
-	dependency.pImageMemoryBarriers    = image_barriers.data();
-	auto command = m_scheduler->Current().Handle();
+	dependency.imageMemoryBarrierCount  = static_cast<uint32_t>(image_barriers.size());
+	dependency.pImageMemoryBarriers     = image_barriers.data();
+	auto command                        = m_scheduler->Current().Handle();
 	command.pipelineBarrier2(dependency);
 	command.copyBufferToImage(buffer, backing.image, vk::ImageLayout::eTransferDstOptimal,
 	                          static_cast<uint32_t>(copies.size()), copies.data());
@@ -256,8 +257,7 @@ void Image::Upload(std::span<const vk::BufferImageCopy> copies, vk::Buffer buffe
 	dependency.pImageMemoryBarriers    = nullptr;
 	command.pipelineBarrier2(dependency);
 	Transit(vk::ImageLayout::eGeneral,
-	        vk::AccessFlagBits2::eShaderRead | vk::AccessFlagBits2::eTransferRead, {},
-	        command);
+	        vk::AccessFlagBits2::eShaderRead | vk::AccessFlagBits2::eTransferRead, {}, command);
 }
 
 void Image::Download(std::span<const vk::BufferImageCopy> copies, vk::Buffer buffer,
@@ -265,7 +265,7 @@ void Image::Download(std::span<const vk::BufferImageCopy> copies, vk::Buffer buf
 	EXIT_IF(m_scheduler == nullptr || copies.empty() || buffer == nullptr || size == 0);
 	m_scheduler->EndRendering();
 	vk::BufferMemoryBarrier2 buffer_barrier {};
-	buffer_barrier.srcStageMask        = vk::PipelineStageFlagBits2::eAllCommands;
+	buffer_barrier.srcStageMask = vk::PipelineStageFlagBits2::eAllCommands;
 	buffer_barrier.srcAccessMask =
 	    vk::AccessFlagBits2::eMemoryRead | vk::AccessFlagBits2::eMemoryWrite;
 	buffer_barrier.dstStageMask        = vk::PipelineStageFlagBits2::eCopy;
@@ -276,16 +276,15 @@ void Image::Download(std::span<const vk::BufferImageCopy> copies, vk::Buffer buf
 	buffer_barrier.offset              = offset;
 	buffer_barrier.size                = size;
 	const auto image_barriers =
-	    GetBarriers(vk::ImageLayout::eTransferSrcOptimal,
-	                vk::AccessFlagBits2::eTransferRead,
+	    GetBarriers(vk::ImageLayout::eTransferSrcOptimal, vk::AccessFlagBits2::eTransferRead,
 	                vk::PipelineStageFlagBits2::eCopy, {});
 	vk::DependencyInfo dependency {};
-	dependency.dependencyFlags           = vk::DependencyFlagBits::eByRegion;
+	dependency.dependencyFlags          = vk::DependencyFlagBits::eByRegion;
 	dependency.bufferMemoryBarrierCount = 1;
 	dependency.pBufferMemoryBarriers    = &buffer_barrier;
-	dependency.imageMemoryBarrierCount = static_cast<uint32_t>(image_barriers.size());
-	dependency.pImageMemoryBarriers    = image_barriers.data();
-	auto command = m_scheduler->Current().Handle();
+	dependency.imageMemoryBarrierCount  = static_cast<uint32_t>(image_barriers.size());
+	dependency.pImageMemoryBarriers     = image_barriers.data();
+	auto command                        = m_scheduler->Current().Handle();
 	command.pipelineBarrier2(dependency);
 	command.copyImageToBuffer(backing.image, vk::ImageLayout::eTransferSrcOptimal, buffer,
 	                          static_cast<uint32_t>(copies.size()), copies.data());
@@ -299,11 +298,11 @@ void Image::Download(std::span<const vk::BufferImageCopy> copies, vk::Buffer buf
 	command.pipelineBarrier2(dependency);
 }
 
-std::pair<uint32_t, uint32_t>
-Image::SanitizeCopyLayers(const Image& source, const Image& destination, uint32_t depth) {
-	const auto source_type      = source.backing.image_type;
-	const auto destination_type = destination.backing.image_type;
-	uint32_t   source_layers    = source.backing.layers;
+std::pair<uint32_t, uint32_t> Image::SanitizeCopyLayers(const Image& source,
+                                                        const Image& destination, uint32_t depth) {
+	const auto source_type        = source.backing.image_type;
+	const auto destination_type   = destination.backing.image_type;
+	uint32_t   source_layers      = source.backing.layers;
 	uint32_t   destination_layers = destination.backing.layers;
 	if (source_type == vk::ImageType::e3D) {
 		source_layers = 1;
@@ -312,13 +311,10 @@ Image::SanitizeCopyLayers(const Image& source, const Image& destination, uint32_
 		destination_layers = 1;
 	}
 	if (source_type == destination_type) {
-		source_layers = destination_layers =
-		    std::min(source_layers, destination_layers);
-	} else if (source_type == vk::ImageType::e2D &&
-	           destination_type == vk::ImageType::e3D) {
+		source_layers = destination_layers = std::min(source_layers, destination_layers);
+	} else if (source_type == vk::ImageType::e2D && destination_type == vk::ImageType::e3D) {
 		source_layers = depth;
-	} else if (source_type == vk::ImageType::e3D &&
-	           destination_type == vk::ImageType::e2D) {
+	} else if (source_type == vk::ImageType::e3D && destination_type == vk::ImageType::e2D) {
 		destination_layers = depth;
 	}
 	return {source_layers, destination_layers};
@@ -327,12 +323,11 @@ Image::SanitizeCopyLayers(const Image& source, const Image& destination, uint32_
 void Image::CopyImage(Image& source) {
 	EXIT_IF(m_scheduler == nullptr || source.backing.samples != backing.samples);
 	m_scheduler->EndRendering();
-	const uint32_t levels =
-	    std::min(source.backing.mip_levels, backing.mip_levels);
+	const uint32_t levels     = std::min(source.backing.mip_levels, backing.mip_levels);
 	const uint32_t base_depth = backing.image_type == vk::ImageType::e3D
 	                                ? backing.extent.depth
 	                                : source.backing.extent.depth;
-	const auto source_aspect =
+	const auto     source_aspect =
 	    FullAspectMask(source.backing.format) & ~vk::ImageAspectFlagBits::eStencil;
 	const auto destination_aspect =
 	    FullAspectMask(backing.format) & ~vk::ImageAspectFlagBits::eStencil;
@@ -342,8 +337,7 @@ void Image::CopyImage(Image& source) {
 		const auto width  = std::max(source.backing.extent.width >> level, 1u);
 		const auto height = std::max(source.backing.extent.height >> level, 1u);
 		const auto depth  = std::max(base_depth >> level, 1u);
-		const auto [source_layers, destination_layers] =
-		    SanitizeCopyLayers(source, *this, depth);
+		const auto [source_layers, destination_layers] = SanitizeCopyLayers(source, *this, depth);
 		vk::ImageCopy copy {};
 		copy.srcSubresource = {source_aspect, level, 0, 1};
 		copy.dstSubresource = {destination_aspect, level, 0, 1};
@@ -351,8 +345,7 @@ void Image::CopyImage(Image& source) {
 			if (source.backing.image_type == vk::ImageType::e3D) {
 				copy.extent = {width, height, depth};
 			} else {
-				copy.srcSubresource.layerCount =
-				    std::min(source_layers, destination_layers);
+				copy.srcSubresource.layerCount = std::min(source_layers, destination_layers);
 				copy.dstSubresource.layerCount = copy.srcSubresource.layerCount;
 				copy.extent                    = {width, height, 1};
 			}
@@ -369,34 +362,30 @@ void Image::CopyImage(Image& source) {
 		return;
 	}
 	auto command = m_scheduler->Current().Handle();
-	source.Transit(vk::ImageLayout::eTransferSrcOptimal,
-	               vk::AccessFlagBits2::eTransferRead, {}, command);
-	Transit(vk::ImageLayout::eTransferDstOptimal,
-	        vk::AccessFlagBits2::eTransferWrite, {}, command);
-	command.copyImage(source.backing.image, vk::ImageLayout::eTransferSrcOptimal,
-	                  backing.image, vk::ImageLayout::eTransferDstOptimal,
-	                  static_cast<uint32_t>(copies.size()), copies.data());
+	source.Transit(vk::ImageLayout::eTransferSrcOptimal, vk::AccessFlagBits2::eTransferRead, {},
+	               command);
+	Transit(vk::ImageLayout::eTransferDstOptimal, vk::AccessFlagBits2::eTransferWrite, {}, command);
+	command.copyImage(source.backing.image, vk::ImageLayout::eTransferSrcOptimal, backing.image,
+	                  vk::ImageLayout::eTransferDstOptimal, static_cast<uint32_t>(copies.size()),
+	                  copies.data());
 	Transit(vk::ImageLayout::eGeneral,
-	        vk::AccessFlagBits2::eShaderRead | vk::AccessFlagBits2::eTransferRead, {},
-	        command);
+	        vk::AccessFlagBits2::eShaderRead | vk::AccessFlagBits2::eTransferRead, {}, command);
 }
 
 void Image::Resolve(Image& source, const ImageSubresourceRange& source_range,
                     const ImageSubresourceRange& destination_range) {
 	EXIT_IF(m_scheduler == nullptr || backing.samples != 1 ||
 	        source.backing.image_type != vk::ImageType::e2D ||
-	        backing.image_type != vk::ImageType::e2D ||
-	        source_range.level_count != 1 || destination_range.level_count != 1 ||
+	        backing.image_type != vk::ImageType::e2D || source_range.level_count != 1 ||
+	        destination_range.level_count != 1 ||
 	        source_range.base_level >= source.backing.mip_levels ||
 	        destination_range.base_level >= backing.mip_levels ||
 	        source_range.base_layer >= source.backing.layers ||
 	        destination_range.base_layer >= backing.layers);
-	const auto layers = std::min(
-	    {source_range.layer_count, destination_range.layer_count,
-	     source.backing.layers - source_range.base_layer,
-	     backing.layers - destination_range.base_layer});
-	const auto source_width =
-	    std::max(source.backing.extent.width >> source_range.base_level, 1u);
+	const auto layers       = std::min({source_range.layer_count, destination_range.layer_count,
+	                                    source.backing.layers - source_range.base_layer,
+	                                    backing.layers - destination_range.base_layer});
+	const auto source_width = std::max(source.backing.extent.width >> source_range.base_level, 1u);
 	const auto source_height =
 	    std::max(source.backing.extent.height >> source_range.base_level, 1u);
 	const auto destination_width =
@@ -404,43 +393,40 @@ void Image::Resolve(Image& source, const ImageSubresourceRange& source_range,
 	const auto destination_height =
 	    std::max(backing.extent.height >> destination_range.base_level, 1u);
 	const bool copy = source.backing.samples == 1;
-	EXIT_IF(layers == 0 || info.extent.width > source_width ||
-	        info.extent.height > source_height || info.extent.width > destination_width ||
-	        info.extent.height > destination_height ||
+	EXIT_IF(layers == 0 || info.extent.width > source_width || info.extent.height > source_height ||
+	        info.extent.width > destination_width || info.extent.height > destination_height ||
 	        (copy ? !ImageViewOps::FormatsCompatible(source.backing.format, backing.format)
 	              : source.backing.format != backing.format));
-	auto resolved_source_range      = source_range;
-	auto resolved_destination_range = destination_range;
+	auto resolved_source_range             = source_range;
+	auto resolved_destination_range        = destination_range;
 	resolved_source_range.layer_count      = layers;
 	resolved_destination_range.layer_count = layers;
 	const vk::Extent3D resolve_extent {info.extent.width, info.extent.height, 1};
 
 	m_scheduler->EndRendering();
 	auto command = m_scheduler->Current().Handle();
-	source.Transit(vk::ImageLayout::eTransferSrcOptimal,
-	               vk::AccessFlagBits2::eTransferRead, resolved_source_range, command);
-	Transit(vk::ImageLayout::eTransferDstOptimal,
-	        vk::AccessFlagBits2::eTransferWrite, resolved_destination_range, command);
+	source.Transit(vk::ImageLayout::eTransferSrcOptimal, vk::AccessFlagBits2::eTransferRead,
+	               resolved_source_range, command);
+	Transit(vk::ImageLayout::eTransferDstOptimal, vk::AccessFlagBits2::eTransferWrite,
+	        resolved_destination_range, command);
 	if (copy) {
 		vk::ImageCopy region {};
-		region.srcSubresource = {vk::ImageAspectFlagBits::eColor,
-		                         resolved_source_range.base_level,
+		region.srcSubresource = {vk::ImageAspectFlagBits::eColor, resolved_source_range.base_level,
 		                         resolved_source_range.base_layer, layers};
 		region.dstSubresource = {vk::ImageAspectFlagBits::eColor,
 		                         resolved_destination_range.base_level,
 		                         resolved_destination_range.base_layer, layers};
-		region.extent = resolve_extent;
-		command.copyImage(source.backing.image, vk::ImageLayout::eTransferSrcOptimal,
-		                  backing.image, vk::ImageLayout::eTransferDstOptimal, region);
+		region.extent         = resolve_extent;
+		command.copyImage(source.backing.image, vk::ImageLayout::eTransferSrcOptimal, backing.image,
+		                  vk::ImageLayout::eTransferDstOptimal, region);
 	} else {
 		vk::ImageResolve region {};
-		region.srcSubresource = {vk::ImageAspectFlagBits::eColor,
-		                         resolved_source_range.base_level,
+		region.srcSubresource = {vk::ImageAspectFlagBits::eColor, resolved_source_range.base_level,
 		                         resolved_source_range.base_layer, layers};
 		region.dstSubresource = {vk::ImageAspectFlagBits::eColor,
 		                         resolved_destination_range.base_level,
 		                         resolved_destination_range.base_layer, layers};
-		region.extent = resolve_extent;
+		region.extent         = resolve_extent;
 		command.resolveImage(source.backing.image, vk::ImageLayout::eTransferSrcOptimal,
 		                     backing.image, vk::ImageLayout::eTransferDstOptimal, region);
 	}
@@ -454,22 +440,21 @@ uint32_t Image::CopyRows(uint64_t row_size, uint32_t rows, uint64_t capacity) no
 }
 
 void Image::CopyImageWithBuffer(Image& source, Buffer& buffer) {
-	EXIT_IF(m_scheduler == nullptr || buffer.Handle() == nullptr ||
-	        source.backing.samples != 1 || backing.samples != 1);
+	EXIT_IF(m_scheduler == nullptr || buffer.Handle() == nullptr || source.backing.samples != 1 ||
+	        backing.samples != 1);
 	m_scheduler->EndRendering();
-	const uint32_t levels =
-	    std::min(source.backing.mip_levels, backing.mip_levels);
-	const auto source_aspect =
+	const uint32_t levels = std::min(source.backing.mip_levels, backing.mip_levels);
+	const auto     source_aspect =
 	    FullAspectMask(source.backing.format) & ~vk::ImageAspectFlagBits::eStencil;
 	const auto destination_aspect =
 	    FullAspectMask(backing.format) & ~vk::ImageAspectFlagBits::eStencil;
-	const auto source_bytes = DepthAspectTransferBytes(source.backing.format) != 0
-	                              ? DepthAspectTransferBytes(source.backing.format)
-	                              : source.info.bytes_per_block;
-	const auto destination_bytes = DepthAspectTransferBytes(backing.format) != 0
-	                                   ? DepthAspectTransferBytes(backing.format)
-	                                   : info.bytes_per_block;
-	const uint32_t source_block = source.info.IsBlock() ? 4u : 1u;
+	const auto     source_bytes      = DepthAspectTransferBytes(source.backing.format) != 0
+	                                       ? DepthAspectTransferBytes(source.backing.format)
+	                                       : source.info.bytes_per_block;
+	const auto     destination_bytes = DepthAspectTransferBytes(backing.format) != 0
+	                                       ? DepthAspectTransferBytes(backing.format)
+	                                       : info.bytes_per_block;
+	const uint32_t source_block      = source.info.IsBlock() ? 4u : 1u;
 	const uint32_t destination_block = info.IsBlock() ? 4u : 1u;
 	EXIT_IF(levels == 0 || source_bytes == 0 || source_bytes != destination_bytes ||
 	        source_block != destination_block);
@@ -484,74 +469,66 @@ void Image::CopyImageWithBuffer(Image& source, Buffer& buffer) {
 	barrier.buffer              = buffer.Handle();
 	barrier.offset              = 0;
 	vk::DependencyInfo dependency {};
-	dependency.dependencyFlags           = vk::DependencyFlagBits::eByRegion;
+	dependency.dependencyFlags          = vk::DependencyFlagBits::eByRegion;
 	dependency.bufferMemoryBarrierCount = 1;
 	dependency.pBufferMemoryBarriers    = &barrier;
-	auto command = m_scheduler->Current().Handle();
-	source.Transit(vk::ImageLayout::eTransferSrcOptimal,
-	               vk::AccessFlagBits2::eTransferRead, {}, command);
-	Transit(vk::ImageLayout::eTransferDstOptimal,
-	        vk::AccessFlagBits2::eTransferWrite, {}, command);
+	auto command                        = m_scheduler->Current().Handle();
+	source.Transit(vk::ImageLayout::eTransferSrcOptimal, vk::AccessFlagBits2::eTransferRead, {},
+	               command);
+	Transit(vk::ImageLayout::eTransferDstOptimal, vk::AccessFlagBits2::eTransferWrite, {}, command);
 	for (uint32_t level = 0; level < levels; level++) {
-		const auto width  = std::max(source.backing.extent.width >> level, 1u);
-		const auto height = std::max(source.backing.extent.height >> level, 1u);
-		const auto source_depth = source.backing.image_type == vk::ImageType::e3D
-		                              ? std::max(source.backing.extent.depth >> level, 1u)
-		                              : source.backing.layers;
+		const auto width             = std::max(source.backing.extent.width >> level, 1u);
+		const auto height            = std::max(source.backing.extent.height >> level, 1u);
+		const auto source_depth      = source.backing.image_type == vk::ImageType::e3D
+		                                   ? std::max(source.backing.extent.depth >> level, 1u)
+		                                   : source.backing.layers;
 		const auto destination_depth = backing.image_type == vk::ImageType::e3D
 		                                   ? std::max(backing.extent.depth >> level, 1u)
 		                                   : backing.layers;
-		const auto slices    = std::min(source_depth, destination_depth);
-		const auto block_rows = (height + source_block - 1) / source_block;
+		const auto slices            = std::min(source_depth, destination_depth);
+		const auto block_rows        = (height + source_block - 1) / source_block;
 		const auto row_size =
 		    static_cast<uint64_t>((width + source_block - 1) / source_block) * source_bytes;
 		const auto rows_per_copy = CopyRows(row_size, block_rows, buffer.Size());
 		EXIT_IF(slices == 0 || rows_per_copy == 0);
 		for (uint32_t slice = 0; slice < slices; slice++) {
-			for (uint32_t block_row = 0; block_row < block_rows;
-			     block_row += rows_per_copy) {
-				const auto copy_rows = std::min(rows_per_copy, block_rows - block_row);
-				const auto y         = block_row * source_block;
-				const auto copy_height =
-				    std::min(copy_rows * source_block, height - y);
-				const auto copy_size = row_size * copy_rows;
+			for (uint32_t block_row = 0; block_row < block_rows; block_row += rows_per_copy) {
+				const auto          copy_rows   = std::min(rows_per_copy, block_rows - block_row);
+				const auto          y           = block_row * source_block;
+				const auto          copy_height = std::min(copy_rows * source_block, height - y);
+				const auto          copy_size   = row_size * copy_rows;
 				vk::BufferImageCopy source_copy {};
 				source_copy.imageSubresource = {
 				    source_aspect, level,
 				    source.backing.image_type == vk::ImageType::e3D ? 0u : slice, 1};
-				source_copy.imageOffset = {
-				    0, static_cast<int32_t>(y),
-				    source.backing.image_type == vk::ImageType::e3D
-				        ? static_cast<int32_t>(slice)
-				        : 0};
-				source_copy.imageExtent = {width, copy_height, 1};
-				auto destination_copy = source_copy;
+				source_copy.imageOffset           = {0, static_cast<int32_t>(y),
+				                                     source.backing.image_type == vk::ImageType::e3D
+				                                         ? static_cast<int32_t>(slice)
+				                                         : 0};
+				source_copy.imageExtent           = {width, copy_height, 1};
+				auto destination_copy             = source_copy;
 				destination_copy.imageSubresource = {
 				    destination_aspect, level,
 				    backing.image_type == vk::ImageType::e3D ? 0u : slice, 1};
 				destination_copy.imageOffset.z =
-				    backing.image_type == vk::ImageType::e3D
-				        ? static_cast<int32_t>(slice)
-				        : 0;
+				    backing.image_type == vk::ImageType::e3D ? static_cast<int32_t>(slice) : 0;
 				barrier.size          = copy_size;
 				barrier.srcAccessMask = vk::AccessFlagBits2::eTransferRead;
 				barrier.dstAccessMask = vk::AccessFlagBits2::eTransferWrite;
 				command.pipelineBarrier2(dependency);
 				command.copyImageToBuffer(source.backing.image,
-				                          vk::ImageLayout::eTransferSrcOptimal,
-				                          buffer.Handle(), source_copy);
+				                          vk::ImageLayout::eTransferSrcOptimal, buffer.Handle(),
+				                          source_copy);
 				barrier.srcAccessMask = vk::AccessFlagBits2::eTransferWrite;
 				barrier.dstAccessMask = vk::AccessFlagBits2::eTransferRead;
 				command.pipelineBarrier2(dependency);
 				command.copyBufferToImage(buffer.Handle(), backing.image,
-				                          vk::ImageLayout::eTransferDstOptimal,
-				                          destination_copy);
+				                          vk::ImageLayout::eTransferDstOptimal, destination_copy);
 			}
 		}
 	}
 	Transit(vk::ImageLayout::eGeneral,
-	        vk::AccessFlagBits2::eShaderRead | vk::AccessFlagBits2::eTransferRead, {},
-	        command);
+	        vk::AccessFlagBits2::eShaderRead | vk::AccessFlagBits2::eTransferRead, {}, command);
 }
 
 void Image::CopyMip(Image& source, uint32_t mip, uint32_t layer) {
@@ -561,11 +538,9 @@ void Image::CopyMip(Image& source, uint32_t mip, uint32_t layer) {
 	const auto width  = std::max(backing.extent.width >> mip, 1u);
 	const auto height = std::max(backing.extent.height >> mip, 1u);
 	const auto depth  = std::max(backing.extent.depth >> mip, 1u);
-	EXIT_IF(width != source.backing.extent.width ||
-	        height != source.backing.extent.height);
-	const auto [source_layers, destination_layers] =
-	    SanitizeCopyLayers(source, *this, depth);
-	const auto aspects = FullAspectMask(source.backing.format);
+	EXIT_IF(width != source.backing.extent.width || height != source.backing.extent.height);
+	const auto [source_layers, destination_layers] = SanitizeCopyLayers(source, *this, depth);
+	const auto aspects                             = FullAspectMask(source.backing.format);
 	EXIT_IF(aspects != FullAspectMask(backing.format));
 	std::array<vk::ImageCopy, 2> copies {};
 	uint32_t                     copy_count = 0;
@@ -580,16 +555,13 @@ void Image::CopyMip(Image& source, uint32_t mip, uint32_t layer) {
 		copy.extent         = {width, height, depth};
 	}
 	auto command = m_scheduler->Current().Handle();
-	Transit(vk::ImageLayout::eTransferDstOptimal,
-	        vk::AccessFlagBits2::eTransferWrite, {}, command);
-	source.Transit(vk::ImageLayout::eTransferSrcOptimal,
-	               vk::AccessFlagBits2::eTransferRead, {}, command);
-	command.copyImage(source.backing.image, vk::ImageLayout::eTransferSrcOptimal,
-	                  backing.image, vk::ImageLayout::eTransferDstOptimal, copy_count,
-	                  copies.data());
+	Transit(vk::ImageLayout::eTransferDstOptimal, vk::AccessFlagBits2::eTransferWrite, {}, command);
+	source.Transit(vk::ImageLayout::eTransferSrcOptimal, vk::AccessFlagBits2::eTransferRead, {},
+	               command);
+	command.copyImage(source.backing.image, vk::ImageLayout::eTransferSrcOptimal, backing.image,
+	                  vk::ImageLayout::eTransferDstOptimal, copy_count, copies.data());
 	Transit(vk::ImageLayout::eGeneral,
-	        vk::AccessFlagBits2::eShaderRead | vk::AccessFlagBits2::eTransferRead, {},
-	        command);
+	        vk::AccessFlagBits2::eShaderRead | vk::AccessFlagBits2::eTransferRead, {}, command);
 }
 
 namespace ImageOps {
@@ -601,8 +573,7 @@ void Validate(const ImageInfo& info) {
 	if (info.pixel_format == vk::Format::eUndefined) {
 		const bool metadata_empty =
 		    info.metadata.range.address == 0 && info.metadata.range.size == 0 &&
-		    info.metadata.kind == ImageMetadataKind::None &&
-		    info.metadata.control == 0 &&
+		    info.metadata.kind == ImageMetadataKind::None && info.metadata.control == 0 &&
 		    info.metadata.compression == VideoOutCompression::Uncompressed &&
 		    !info.metadata.stencil_compressed;
 		if (info.data.Empty() || info.HasStencil() || !metadata_empty || info.extent.width == 0 ||
@@ -615,9 +586,9 @@ void Validate(const ImageInfo& info) {
 	}
 
 	if (info.extent.width == 0 || info.extent.height == 0 || info.extent.depth == 0 ||
-	    info.resources.levels == 0 ||
-	    info.resources.levels > info.mip_layout.size() || info.resources.layers == 0 ||
-	    info.samples == 0 || vulkan_sample_count(info.samples) == vk::SampleCountFlagBits {} ||
+	    info.resources.levels == 0 || info.resources.levels > info.mip_layout.size() ||
+	    info.resources.layers == 0 || info.samples == 0 ||
+	    vulkan_sample_count(info.samples) == vk::SampleCountFlagBits {} ||
 	    info.bytes_per_block == 0 || (info.data.address != 0 && info.pitch == 0)) {
 		EXIT("invalid image geometry or format\n");
 	}
@@ -688,11 +659,11 @@ uint32_t RenderTargetTransferFormat(uint32_t bytes_per_element) {
 
 } // namespace ImageOps
 
-Image::Image(GraphicContext& graphics, CommandScheduler& scheduler,
-             const ImageInfo& image_info)
+Image::Image(GraphicContext& graphics, CommandScheduler& scheduler, const ImageInfo& image_info)
     : info(image_info), m_graphics(&graphics), m_scheduler(&scheduler) {
 	KYTY_PROFILER_FUNCTION();
 	ImageOps::Validate(info);
+	m_cpu_dirty = !info.data.Empty();
 	if (info.pixel_format == vk::Format::eUndefined) {
 		return;
 	}
@@ -742,9 +713,9 @@ Image::Image(GraphicContext& graphics, CommandScheduler& scheduler,
 }
 
 uint64_t Image::HashGuestEdges() const {
-	constexpr uint64_t page_mask = TRACKER_PAGE_SIZE - 1;
+	constexpr uint64_t                         page_mask = TRACKER_PAGE_SIZE - 1;
 	std::array<uint8_t, TRACKER_PAGE_SIZE * 2> bytes {};
-	const auto     range        = info.data;
+	const auto                                 range = info.data;
 	const uint64_t head_end     = std::min(range.End(), (range.address + page_mask) & ~page_mask);
 	const uint64_t tail_begin   = std::max(range.address, range.End() & ~page_mask);
 	const uint64_t head_size    = head_end - range.address;

src/graphics/host_gpu/renderer/image/image.h

@@ -3,7 +3,7 @@
 
 #include "common/assert.h"
 #include "graphics/host_gpu/graphicContext.h"
-#include "graphics/host_gpu/renderer/imageInfo.h"
+#include "graphics/host_gpu/renderer/image/imageInfo.h"
 
 #include <compare>
 #include <limits>
@@ -66,16 +66,16 @@ public:
 	[[nodiscard]] vk::ImageView FindView(const ImageViewInfo& view_info);
 	void                        AssociateDepth(ImageId image_id) { depth_id = image_id; }
 	using Barriers = std::vector<vk::ImageMemoryBarrier2>;
-	[[nodiscard]] Barriers
-	GetBarriers(vk::ImageLayout destination_layout, vk::AccessFlags2 destination_access,
-	            vk::PipelineStageFlags2 destination_stage,
-	            std::optional<ImageSubresourceRange> range);
+	[[nodiscard]] Barriers GetBarriers(vk::ImageLayout                      destination_layout,
+	                                   vk::AccessFlags2                     destination_access,
+	                                   vk::PipelineStageFlags2              destination_stage,
+	                                   std::optional<ImageSubresourceRange> range);
 	void Transit(vk::ImageLayout destination_layout, vk::AccessFlags2 destination_access,
 	             std::optional<ImageSubresourceRange> range, vk::CommandBuffer command_buffer);
-	void Upload(std::span<const vk::BufferImageCopy> copies, vk::Buffer buffer,
-	            uint64_t offset, uint64_t size);
-	void Download(std::span<const vk::BufferImageCopy> copies, vk::Buffer buffer,
-	              uint64_t offset, uint64_t size);
+	void Upload(std::span<const vk::BufferImageCopy> copies, vk::Buffer buffer, uint64_t offset,
+	            uint64_t size);
+	void Download(std::span<const vk::BufferImageCopy> copies, vk::Buffer buffer, uint64_t offset,
+	              uint64_t size);
 	void CopyImage(Image& source);
 	void Resolve(Image& source, const ImageSubresourceRange& source_range,
 	             const ImageSubresourceRange& destination_range);
@@ -84,8 +84,8 @@ public:
 
 	void InvalidateCpuWrite(uint64_t vaddr, uint64_t size) {
 		if (ImageRangeOverlaps(info.data.address, info.data.size, vaddr, size)) {
-			m_cpu_dirty       = true;
-			m_maybe_cpu_dirty = false;
+			m_cpu_dirty        = true;
+			m_maybe_cpu_dirty  = false;
 			m_maybe_hash_valid = false;
 		} else if (ImagePageRangesOverlap(info.data.address, info.data.size, vaddr, size)) {
 			m_maybe_cpu_dirty = true;
@@ -95,6 +95,11 @@ public:
 	[[nodiscard]] bool IsCpuDirty() const { return m_cpu_dirty || m_maybe_cpu_dirty; }
 	[[nodiscard]] bool IsDefinitelyCpuDirty() const { return m_cpu_dirty; }
 	[[nodiscard]] bool IsMaybeCpuDirty() const { return m_maybe_cpu_dirty; }
+	void               MarkMaybeCpuDirty() {
+		if (!m_cpu_dirty) {
+			m_maybe_cpu_dirty = true;
+		}
+	}
 	[[nodiscard]] bool NeedsMaybeCpuHash() const {
 		return m_maybe_cpu_dirty && !m_maybe_hash_valid;
 	}
@@ -102,14 +107,14 @@ public:
 		if (!NeedsMaybeCpuHash()) {
 			EXIT("image cannot initialize maybe-dirty hash\n");
 		}
-		m_maybe_cpu_hash = hash;
+		m_maybe_cpu_hash   = hash;
 		m_maybe_hash_valid = true;
 	}
 	[[nodiscard]] bool ResolveMaybeCpuHash(uint64_t hash) {
 		if (!m_maybe_cpu_dirty || !m_maybe_hash_valid || m_cpu_dirty) {
 			EXIT("image cannot resolve maybe-dirty hash\n");
 		}
-		m_maybe_cpu_dirty = false;
+		m_maybe_cpu_dirty  = false;
 		m_maybe_hash_valid = false;
 		m_cpu_dirty |= hash != m_maybe_cpu_hash;
 		return m_cpu_dirty;
@@ -125,14 +130,15 @@ public:
 	}
 
 	[[nodiscard]] bool IsGpuModified() const noexcept { return m_gpu_modified; }
-	void MarkGpuModified() noexcept { m_gpu_modified = true; }
-	void ClearGpuModified() noexcept { m_gpu_modified = false; }
+	void               MarkGpuModified() noexcept { m_gpu_modified = true; }
+	void               ClearGpuModified() noexcept { m_gpu_modified = false; }
 
 	[[nodiscard]] bool IsBufferModified() const noexcept { return m_buffer_modified; }
-	void MarkBufferModified() noexcept { m_buffer_modified = true; }
-	void ClearBufferModified() noexcept { m_buffer_modified = false; }
+	void               MarkBufferModified() noexcept { m_buffer_modified = true; }
+	void               ClearBufferModified() noexcept { m_buffer_modified = false; }
 
-	[[nodiscard]] bool Overlaps(uint64_t address, uint64_t size, bool pages = false) const noexcept {
+	[[nodiscard]] bool Overlaps(uint64_t address, uint64_t size,
+	                            bool pages = false) const noexcept {
 		return pages ? ImagePageRangesOverlap(info.data.address, info.data.size, address, size)
 		             : ImageRangeOverlaps(info.data.address, info.data.size, address, size);
 	}
@@ -142,38 +148,41 @@ public:
 	[[nodiscard]] bool SafeToDownload() const noexcept {
 		return IsGpuModified() && !IsBufferModified() && !IsCpuDirty();
 	}
+	[[nodiscard]] bool IsTracked() const noexcept { return track_addr != 0 && track_addr_end != 0; }
 	[[nodiscard]] uint64_t AccountedSize() const noexcept {
 		return backing.image == nullptr ? 0 : (info.data.size + 1023) & ~uint64_t {1023};
 	}
 	[[nodiscard]] uint64_t HashGuestEdges() const;
 
-	ImageInfo            info;
-	VulkanImage          backing;
-	ImageViewCache       views;
-	ImageUsage           usage;
-	ImageBinding         binding;
-	bool                 registered          = false;
-	ImageId              depth_id {};
-	uint64_t             tick_accessed_last = 0;
-	size_t               lru_id = 0;
+	ImageInfo      info;
+	VulkanImage    backing;
+	ImageViewCache views;
+	ImageUsage     usage;
+	ImageBinding   binding;
+	bool           registered     = false;
+	uint64_t       track_addr     = 0;
+	uint64_t       track_addr_end = 0;
+	ImageId        depth_id {};
+	uint64_t       tick_accessed_last = 0;
+	size_t         lru_id             = 0;
 
 private:
 	friend struct ImageTestAccess;
 
 	[[nodiscard]] static vk::ImageAspectFlags FullAspectMask(vk::Format format) noexcept;
-	[[nodiscard]] static uint32_t CopyRows(uint64_t row_size, uint32_t rows,
-	                                       uint64_t capacity) noexcept;
+	[[nodiscard]] static uint32_t             CopyRows(uint64_t row_size, uint32_t rows,
+	                                                   uint64_t capacity) noexcept;
 	[[nodiscard]] static std::pair<uint32_t, uint32_t>
 	SanitizeCopyLayers(const Image& source, const Image& destination, uint32_t depth);
 
-	GraphicContext*   m_graphics  = nullptr;
-	CommandScheduler* m_scheduler = nullptr;
-	uint64_t        m_maybe_cpu_hash = 0;
-	bool            m_cpu_dirty = false;
-	bool            m_maybe_cpu_dirty = false;
-	bool            m_maybe_hash_valid = false;
-	bool            m_gpu_modified = false;
-	bool            m_buffer_modified = false;
+	GraphicContext*   m_graphics         = nullptr;
+	CommandScheduler* m_scheduler        = nullptr;
+	uint64_t          m_maybe_cpu_hash   = 0;
+	bool              m_cpu_dirty        = false;
+	bool              m_maybe_cpu_dirty  = false;
+	bool              m_maybe_hash_valid = false;
+	bool              m_gpu_modified     = false;
+	bool              m_buffer_modified  = false;
 };
 
 namespace ImageOps {

src/graphics/host_gpu/renderer/image/imageInfo.h

@@ -19,10 +19,9 @@ struct GuestRange {
 	uint64_t address = 0;
 	uint64_t size    = 0;
 
-	[[nodiscard]] constexpr bool     Empty() const noexcept { return address == 0 || size == 0; }
-	[[nodiscard]] constexpr bool     Valid() const noexcept {
-		return !Empty() && address < TRACKER_ADDRESS_SIZE &&
-		       size <= TRACKER_ADDRESS_SIZE - address;
+	[[nodiscard]] constexpr bool Empty() const noexcept { return address == 0 || size == 0; }
+	[[nodiscard]] constexpr bool Valid() const noexcept {
+		return !Empty() && address < TRACKER_ADDRESS_SIZE && size <= TRACKER_ADDRESS_SIZE - address;
 	}
 	[[nodiscard]] constexpr uint64_t End() const noexcept { return address + size; }
 	auto                             operator<=>(const GuestRange&) const = default;
@@ -47,10 +46,10 @@ struct ImageSubresources {
 };
 
 struct ImageSubresourceRange {
-	uint32_t base_level  = 0;
-	uint32_t level_count = 1;
-	uint32_t base_layer  = 0;
-	uint32_t layer_count = 1;
+	uint32_t base_level                                      = 0;
+	uint32_t level_count                                     = 1;
+	uint32_t base_layer                                      = 0;
+	uint32_t layer_count                                     = 1;
 	auto     operator<=>(const ImageSubresourceRange&) const = default;
 };
 
@@ -67,10 +66,10 @@ struct ImageInfo {
 	GuestRange                   stencil;
 	ImageMetadataInfo            metadata;
 	uint32_t                     htile_clear_mask = UINT32_MAX;
-	vk::Format                   pixel_format = vk::Format::eUndefined;
-	uint32_t                     guest_format = 0;
-	Prospero::ImageType          type         = Prospero::ImageType::kColor2D;
-	vk::Extent3D                 extent       = {1, 1, 1};
+	vk::Format                   pixel_format     = vk::Format::eUndefined;
+	uint32_t                     guest_format     = 0;
+	Prospero::ImageType          type             = Prospero::ImageType::kColor2D;
+	vk::Extent3D                 extent           = {1, 1, 1};
 	ImageSubresources            resources;
 	uint32_t                     pitch           = 0;
 	uint32_t                     bytes_per_block = 0;
@@ -352,8 +351,7 @@ inline bool ImageInfo::IsDepth() const noexcept {
 	}
 	const auto transfer_bytes = DepthAspectTransferBytes(info.pixel_format);
 	return transfer_bytes == info.bytes_per_block ||
-	       (info.bytes_per_block == sizeof(uint16_t) &&
-	        transfer_bytes == sizeof(uint32_t));
+	       (info.bytes_per_block == sizeof(uint16_t) && transfer_bytes == sizeof(uint32_t));
 }
 
 [[nodiscard]] inline VideoOutCompression
@@ -470,18 +468,13 @@ IsSupportedDisplayRenderTargetTileMode(uint32_t tile_mode) noexcept {
                                                  vk::ClearColorValue& clear) {
 	vk::ClearColorValue next {};
 	const auto unorm8 = [](uint32_t value) { return static_cast<float>(value & 0xffu) / 255.0f; };
-	const auto srgb8 = [](uint32_t value) {
+	const auto srgb8  = [](uint32_t value) {
 		const auto encoded = static_cast<float>(value & 0xffu) / 255.0f;
-		return encoded <= 0.04045f ? encoded / 12.92f
-		                           : std::pow((encoded + 0.055f) / 1.055f, 2.4f);
+		return encoded <= 0.04045f ? encoded / 12.92f : std::pow((encoded + 0.055f) / 1.055f, 2.4f);
 	};
 	switch (format) {
-		case vk::Format::eR32Uint:
-			next.uint32[0] = packed;
-			break;
-		case vk::Format::eR32Sint:
-			next.int32[0] = static_cast<int32_t>(packed);
-			break;
+		case vk::Format::eR32Uint: next.uint32[0] = packed; break;
+		case vk::Format::eR32Sint: next.int32[0] = static_cast<int32_t>(packed); break;
 		case vk::Format::eR8G8B8A8Srgb:
 			next.float32[0] = srgb8(packed);
 			next.float32[1] = srgb8(packed >> 8u);

src/graphics/host_gpu/renderer/image/imageView.cpp

@@ -1,8 +1,8 @@
-#include "graphics/host_gpu/renderer/imageView.h"
+#include "graphics/host_gpu/renderer/image/imageView.h"
 
 #include "common/assert.h"
 #include "graphics/host_gpu/graphicContext.h"
-#include "graphics/host_gpu/renderer/image.h"
+#include "graphics/host_gpu/renderer/image/image.h"
 
 #include <mutex>
 
@@ -70,15 +70,14 @@ namespace {
 			}
 		case vk::ImageType::e3D:
 			switch (info.type) {
-				case vk::ImageViewType::e3D:
-					return info.base_layer == 0 && info.layer_count == 1;
+				case vk::ImageViewType::e3D: return info.base_layer == 0 && info.layer_count == 1;
 				case vk::ImageViewType::e2D:
-					return static_cast<bool>(
-					           image.flags & vk::ImageCreateFlagBits::e2DArrayCompatible) &&
+					return static_cast<bool>(image.flags &
+					                         vk::ImageCreateFlagBits::e2DArrayCompatible) &&
 					       info.level_count == 1 && info.layer_count == 1;
 				case vk::ImageViewType::e2DArray:
-					return static_cast<bool>(
-					           image.flags & vk::ImageCreateFlagBits::e2DArrayCompatible) &&
+					return static_cast<bool>(image.flags &
+					                         vk::ImageCreateFlagBits::e2DArrayCompatible) &&
 					       info.level_count == 1;
 				default: return false;
 			}
@@ -325,11 +324,10 @@ bool FormatsCompatible(vk::Format base, vk::Format view) noexcept {
 } // namespace ImageViewOps
 
 vk::ImageView Image::FindView(const ImageViewInfo& view_info) {
-	const auto& image = backing;
+	const auto& image      = backing;
 	auto        normalized = view_info;
-	const bool  is_storage =
-	    static_cast<bool>(normalized.usage & vk::ImageUsageFlagBits::eStorage);
-	normalized.aspect = FullAspectMask(image.format);
+	const bool  is_storage = static_cast<bool>(normalized.usage & vk::ImageUsageFlagBits::eStorage);
+	normalized.aspect      = FullAspectMask(image.format);
 	if (normalized.aspect & vk::ImageAspectFlagBits::eDepth &&
 	    IsDepthViewFormat(normalized.format)) {
 		normalized.format = image.format;
@@ -340,28 +338,26 @@ vk::ImageView Image::FindView(const ImageViewInfo& view_info) {
 		normalized.format = image.format;
 		normalized.aspect = vk::ImageAspectFlagBits::eStencil;
 	}
-	normalized.usage =
-	    is_storage ? vk::ImageUsageFlagBits::eStorage : vk::ImageUsageFlags {};
+	normalized.usage = is_storage ? vk::ImageUsageFlagBits::eStorage : vk::ImageUsageFlags {};
 	const bool format_compatible = normalized.format != vk::Format::eUndefined &&
 	                               IsCompatibleViewFormat(image.format, normalized.format);
-	const bool slice_view = image.image_type == vk::ImageType::e3D &&
-	                        (normalized.type == vk::ImageViewType::e2D ||
-	                         normalized.type == vk::ImageViewType::e2DArray);
+	const bool slice_view =
+	    image.image_type == vk::ImageType::e3D && (normalized.type == vk::ImageViewType::e2D ||
+	                                               normalized.type == vk::ImageViewType::e2DArray);
 	const bool levels_valid = normalized.level_count != 0 &&
 	                          normalized.base_level < image.mip_levels &&
 	                          normalized.level_count <= image.mip_levels - normalized.base_level;
-	const auto view_layers = slice_view && levels_valid
-	                             ? std::max(image.extent.depth >> normalized.base_level, 1u)
-	                             : image.layers;
-	const bool ranges_valid = levels_valid &&
-	                          normalized.layer_count != 0 && normalized.base_layer < view_layers &&
+	const auto view_layers  = slice_view && levels_valid
+	                              ? std::max(image.extent.depth >> normalized.base_level, 1u)
+	                              : image.layers;
+	const bool ranges_valid = levels_valid && normalized.layer_count != 0 &&
+	                          normalized.base_layer < view_layers &&
 	                          normalized.layer_count <= view_layers - normalized.base_layer;
 	const bool mapping_valid =
 	    IsComponentSwizzle(normalized.mapping.r) && IsComponentSwizzle(normalized.mapping.g) &&
 	    IsComponentSwizzle(normalized.mapping.b) && IsComponentSwizzle(normalized.mapping.a);
 	if (image.image == nullptr || !format_compatible || !ranges_valid || !mapping_valid ||
-	    !IsValidViewType(image, normalized) ||
-	    !IsValidAspect(image, normalized.aspect)) {
+	    !IsValidViewType(image, normalized) || !IsValidAspect(image, normalized.aspect)) {
 		EXIT("invalid image view: image_format=%d view_format=%d type=%d aspect=0x%x "
 		     "mip=%u+%u layer=%u+%u usage=0x%x image_levels=%u image_layers=%u\n",
 		     static_cast<int>(image.format), static_cast<int>(normalized.format),

src/graphics/host_gpu/renderer/image/imageView.h

@@ -2,8 +2,8 @@
 #define EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_IMAGEVIEW_H_
 
 #include "common/assert.h"
-#include "graphics/host_gpu/renderer/imageInfo.h"
-#include "graphics/shader/recompiler/ShaderIR.h"
+#include "graphics/host_gpu/renderer/image/imageInfo.h"
+#include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/shader.h"
 
 namespace Libs::Graphics {
@@ -88,7 +88,9 @@ SelectSampledDepthView(vk::Format image_format, vk::Format view_format, uint32_t
 IsSupportedSampledDepthResource(const ShaderRecompiler::IR::ImageResource& resource) noexcept {
 	return resource.kind == ShaderRecompiler::IR::ResourceKind::Image &&
 	       (resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2D ||
-	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DArray) &&
+	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DArray ||
+	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DMsaa ||
+	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DMsaaArray) &&
 	       resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::None && resource.read &&
 	       !resource.written && !resource.atomic;
 }
@@ -103,10 +105,7 @@ IsSupportedSampledDepthUintResource(const ShaderRecompiler::IR::ImageResource& r
 
 inline void ValidateStorageColorView(vk::Format image_format, vk::Format view_format,
                                      uint32_t swizzle) noexcept {
-	const auto srgb_view = SrgbStorageViewFormat(image_format);
-	const bool srgb_storage_view =
-	    srgb_view != vk::Format::eUndefined && view_format == srgb_view;
-	if ((image_format != view_format && !srgb_storage_view) ||
+	if (!ImageViewOps::FormatsCompatible(image_format, view_format) ||
 	    !IsValidImageSwizzle(swizzle)) {
 		UnsupportedColorView("storage", image_format, view_format, swizzle);
 	}
@@ -122,7 +121,10 @@ IsSupportedStorageImageResource(const ShaderRecompiler::IR::ImageResource& resou
 	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim3D ||
 	        resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DArray) &&
 	       resource.mip_mode == ShaderRecompiler::IR::ImageMipMode::None && resource.written &&
-	       !resource.atomic && !resource.depth_compare;
+	       (!resource.atomic ||
+	        (resource.kind == ShaderRecompiler::IR::ResourceKind::StorageImageUint &&
+	         resource.read)) &&
+	       !resource.depth_compare;
 }
 
 inline void

src/graphics/host_gpu/renderer/image/textureCommon.cpp

@@ -1,9 +1,9 @@
-#include "graphics/host_gpu/objects/textureCommon.h"
+#include "graphics/host_gpu/renderer/image/textureCommon.h"
 
 #include "common/assert.h"
 #include "graphics/guest_gpu/gpu_defs.h"
 #include "graphics/guest_gpu/gpu_format.h"
-#include "graphics/host_gpu/renderer/tiler.h"
+#include "graphics/host_gpu/renderer/image/tiler.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 
 #include <algorithm>
@@ -70,6 +70,10 @@ constexpr RenderTargetFormatMapping kRenderTargetFormats[] = {
      Prospero::ChannelType::kFloat,
      Prospero::ChannelOrder::kStandard,
      {vk::Format::eB10G11R11UfloatPack32, 4}},
+    {Prospero::ChannelLayout::k5_6_5,
+     Prospero::ChannelType::kUNorm,
+     Prospero::ChannelOrder::kStandard,
+     {vk::Format::eB5G6R5UnormPack16, 2}},
     {Prospero::ChannelLayout::k16,
      Prospero::ChannelType::kUNorm,
      Prospero::ChannelOrder::kStandard,
@@ -102,6 +106,10 @@ constexpr RenderTargetFormatMapping kRenderTargetFormats[] = {
      Prospero::ChannelType::kUNorm,
      Prospero::ChannelOrder::kStandard,
      {vk::Format::eR16G16B16A16Unorm, 8}},
+    {Prospero::ChannelLayout::k16_16_16_16,
+     Prospero::ChannelType::kUInt,
+     Prospero::ChannelOrder::kStandard,
+     {vk::Format::eR16G16B16A16Uint, 8}},
     {Prospero::ChannelLayout::k16_16_16_16,
      Prospero::ChannelType::kFloat,
      Prospero::ChannelOrder::kStandard,
@@ -393,10 +401,10 @@ TextureUploadLayout TextureCalcUploadLayout(uint32_t fmt, uint64_t width, uint64
 	return layout;
 }
 
-std::vector<vk::BufferImageCopy>
-TextureBuildImageCopies(const TextureUploadLayout& layout, uint32_t width, uint32_t height,
-                        uint32_t depth, uint64_t levels, bool array_texture,
-                        bool volume_texture) {
+std::vector<vk::BufferImageCopy> TextureBuildImageCopies(const TextureUploadLayout& layout,
+                                                         uint32_t width, uint32_t height,
+                                                         uint32_t depth, uint64_t levels,
+                                                         bool array_texture, bool volume_texture) {
 	uint32_t mip_width  = width;
 	uint32_t mip_height = height;
 	uint32_t mip_pitch  = volume_texture && static_cast<Prospero::TileMode>(layout.tile) !=
@@ -412,14 +420,13 @@ TextureBuildImageCopies(const TextureUploadLayout& layout, uint32_t width, uint3
 		const auto mip_depth = GetTextureLevelDepth(depth, i, volume_texture);
 
 		for (uint32_t z = 0; z < mip_depth; z++) {
-			const auto slice_offset = z * layout.slice_stride;
+			const auto          slice_offset = z * layout.slice_stride;
 			vk::BufferImageCopy region {};
-			region.bufferOffset =
-			    layout.level_sizes[i].offset + slice_offset;
-			region.imageSubresource = {vk::ImageAspectFlagBits::eColor, i,
-			                           array_texture ? z : 0, 1};
-			region.imageOffset.z = volume_texture ? static_cast<int>(z) : 0;
-			region.imageExtent   = {mip_width, mip_height, 1};
+			region.bufferOffset     = layout.level_sizes[i].offset + slice_offset;
+			region.imageSubresource = {vk::ImageAspectFlagBits::eColor, i, array_texture ? z : 0,
+			                           1};
+			region.imageOffset.z    = volume_texture ? static_cast<int>(z) : 0;
+			region.imageExtent      = {mip_width, mip_height, 1};
 			const bool linear =
 			    static_cast<Prospero::TileMode>(layout.tile) == Prospero::TileMode::kLinear;
 			if (linear) {
@@ -429,9 +436,8 @@ TextureBuildImageCopies(const TextureUploadLayout& layout, uint32_t width, uint3
 				const auto align = [](uint32_t value, uint32_t block) {
 					return ((value + block - 1u) / block) * block;
 				};
-				const auto pitch = align(mip_pitch, layout.texel_block);
-				region.bufferRowLength =
-				    pitch > align(mip_width, layout.texel_block) ? pitch : 0;
+				const auto pitch       = align(mip_pitch, layout.texel_block);
+				region.bufferRowLength = pitch > align(mip_width, layout.texel_block) ? pitch : 0;
 			}
 			regions.push_back(region);
 		}
@@ -476,8 +482,7 @@ static bool SetGpuTileSize(uint64_t offset, uint64_t length, uint64_t capacity,
 	return true;
 }
 
-bool TextureBuildGpuTileInfos(uint64_t size,
-                              const std::vector<vk::BufferImageCopy>& regions,
+bool TextureBuildGpuTileInfos(uint64_t size, const std::vector<vk::BufferImageCopy>& regions,
                               const TextureUploadLayout& layout, uint32_t fmt, uint32_t depth,
                               uint64_t levels, std::vector<GpuTileInfo>& out_infos) {
 	if (size == 0 || levels == 0 || levels > 16 || depth == 0 ||
@@ -518,13 +523,12 @@ bool TextureBuildGpuTileInfos(uint64_t size,
 			for (uint32_t z = 0; z < mip_depth; z += block.block_depth) {
 				const uint32_t copy_depth = std::min(block.block_depth, mip_depth - z);
 				const auto&    region     = regions[region_base + z];
-				const auto     pitch = region.bufferRowLength != 0
-				                           ? region.bufferRowLength
-				                           : region.imageExtent.width;
-				const auto logical_height = region.bufferImageHeight != 0
-				                                ? region.bufferImageHeight
-				                                : region.imageExtent.height;
-				GpuTileInfo    info {};
+				const auto     pitch =
+				    region.bufferRowLength != 0 ? region.bufferRowLength : region.imageExtent.width;
+				const auto  logical_height = region.bufferImageHeight != 0
+				                                 ? region.bufferImageHeight
+				                                 : region.imageExtent.height;
+				GpuTileInfo info {};
 				info.family            = block.family;
 				info.bytes_per_element = block.bytes_per_element;
 				info.linear_offset     = region.bufferOffset;
@@ -540,20 +544,17 @@ bool TextureBuildGpuTileInfos(uint64_t size,
 					return false;
 				}
 				info.linear_slice_stride = linear_stride;
-				info.width = std::max(
-				    (region.imageExtent.width + element.wide - 1u) / element.wide, 1u);
-				info.height = std::max(
-				    (logical_height + element.tall - 1u) / element.tall, 1u);
-				info.depth       = copy_depth;
-				info.surface_z = block.block_depth == 1
-				                     ? static_cast<uint32_t>(region.imageOffset.z)
-				                     : 0;
-				info.pitch =
-				    std::max((pitch + element.wide - 1u) / element.wide, 1u);
-				info.tail_x      = tail ? volume.tail_x[level] : 0;
-				info.tail_y      = tail ? volume.tail_y[level] : 0;
-				info.tail        = tail;
-				info.tiled_width = volume.level_widths[level];
+				info.width =
+				    std::max((region.imageExtent.width + element.wide - 1u) / element.wide, 1u);
+				info.height = std::max((logical_height + element.tall - 1u) / element.tall, 1u);
+				info.depth  = copy_depth;
+				info.surface_z =
+				    block.block_depth == 1 ? static_cast<uint32_t>(region.imageOffset.z) : 0;
+				info.pitch        = std::max((pitch + element.wide - 1u) / element.wide, 1u);
+				info.tail_x       = tail ? volume.tail_x[level] : 0;
+				info.tail_y       = tail ? volume.tail_y[level] : 0;
+				info.tail         = tail;
+				info.tiled_width  = volume.level_widths[level];
 				info.tiled_height = volume.level_heights[level];
 				infos.push_back(info);
 			}
@@ -577,12 +578,11 @@ bool TextureBuildGpuTileInfos(uint64_t size,
 			const auto level_depth = GetTextureLevelDepth(depth, level, layout.volume_texture);
 			for (uint32_t z = 0; z < level_depth; z++) {
 				const auto& region = regions[region_index++];
-				const auto pitch = region.bufferRowLength != 0
-				                       ? region.bufferRowLength
-				                       : region.imageExtent.width;
-				const auto logical_height = region.bufferImageHeight != 0
-				                                ? region.bufferImageHeight
-				                                : region.imageExtent.height;
+				const auto  pitch =
+				    region.bufferRowLength != 0 ? region.bufferRowLength : region.imageExtent.width;
+				const auto  logical_height = region.bufferImageHeight != 0
+				                                 ? region.bufferImageHeight
+				                                 : region.imageExtent.height;
 				GpuTileInfo info {};
 				info.family            = block.family;
 				info.bytes_per_element = block.bytes_per_element;
@@ -593,16 +593,14 @@ bool TextureBuildGpuTileInfos(uint64_t size,
 				                    info.tiled_size)) {
 					return false;
 				}
-				info.width = std::max(
-				    (region.imageExtent.width + element.wide - 1u) / element.wide, 1u);
-				info.height = std::max(
-				    (logical_height + element.tall - 1u) / element.tall, 1u);
+				info.width =
+				    std::max((region.imageExtent.width + element.wide - 1u) / element.wide, 1u);
+				info.height    = std::max((logical_height + element.tall - 1u) / element.tall, 1u);
 				info.surface_z = base_family == TileBlockFamily::RenderTarget64KB ||
 				                         base_family == TileBlockFamily::Depth64KB
 				                     ? region.imageSubresource.baseArrayLayer
 				                     : 0;
-				info.pitch =
-				    std::max((pitch + element.wide - 1u) / element.wide, 1u);
+				info.pitch     = std::max((pitch + element.wide - 1u) / element.wide, 1u);
 				info.tail      = tail;
 				info.tail_x    = tail ? level_size.x : 0;
 				info.tail_y    = tail ? level_size.y : 0;

src/graphics/host_gpu/renderer/image/textureCommon.h

@@ -1,5 +1,5 @@
-#ifndef EMULATOR_INCLUDE_EMULATOR_GRAPHICS_OBJECTS_TEXTURECOMMON_H_
-#define EMULATOR_INCLUDE_EMULATOR_GRAPHICS_OBJECTS_TEXTURECOMMON_H_
+#ifndef EMULATOR_INCLUDE_EMULATOR_GRAPHICS_HOST_GPU_RENDERER_IMAGE_TEXTURECOMMON_H_
+#define EMULATOR_INCLUDE_EMULATOR_GRAPHICS_HOST_GPU_RENDERER_IMAGE_TEXTURECOMMON_H_
 
 #include "common/abi.h"
 #include "common/common.h"
@@ -32,23 +32,22 @@ struct TextureUploadLayout {
 	TilePaddedSize  padded_sizes[16]    = {};
 };
 
-vk::ComponentMapping TextureGetComponentMapping(uint32_t swizzle);
+vk::ComponentMapping   TextureGetComponentMapping(uint32_t swizzle);
 vk::Format             TextureGetFormat(uint32_t fmt);
 RenderTargetFormatInfo TextureGetRenderTargetFormat(uint32_t layout, uint32_t type, uint32_t order);
-TextureUploadLayout TextureCalcUploadLayout(uint32_t fmt, uint64_t width, uint64_t height,
-                                            uint64_t levels, uint32_t depth, uint64_t pitch,
-                                            uint64_t tile, uint64_t upload_size,
-                                            bool allow_depth_tile, bool volume_texture,
-                                            const char* owner);
-std::vector<vk::BufferImageCopy>
-TextureBuildImageCopies(const TextureUploadLayout& layout, uint32_t width, uint32_t height,
-                        uint32_t depth, uint64_t levels, bool array_texture,
-                        bool volume_texture);
-bool TextureBuildGpuTileInfos(uint64_t size,
-                              const std::vector<vk::BufferImageCopy>& regions,
+TextureUploadLayout    TextureCalcUploadLayout(uint32_t fmt, uint64_t width, uint64_t height,
+                                               uint64_t levels, uint32_t depth, uint64_t pitch,
+                                               uint64_t tile, uint64_t upload_size,
+                                               bool allow_depth_tile, bool volume_texture,
+                                               const char* owner);
+std::vector<vk::BufferImageCopy> TextureBuildImageCopies(const TextureUploadLayout& layout,
+                                                         uint32_t width, uint32_t height,
+                                                         uint32_t depth, uint64_t levels,
+                                                         bool array_texture, bool volume_texture);
+bool TextureBuildGpuTileInfos(uint64_t size, const std::vector<vk::BufferImageCopy>& regions,
                               const TextureUploadLayout& layout, uint32_t fmt, uint32_t depth,
                               uint64_t levels, std::vector<GpuTileInfo>& infos);
 
 } // namespace Libs::Graphics
 
-#endif /* EMULATOR_INCLUDE_EMULATOR_GRAPHICS_OBJECTS_TEXTURECOMMON_H_ */
+#endif /* EMULATOR_INCLUDE_EMULATOR_GRAPHICS_HOST_GPU_RENDERER_IMAGE_TEXTURECOMMON_H_ */

src/graphics/host_gpu/renderer/image/tiler.cpp

@@ -1,4 +1,4 @@
-#include "graphics/host_gpu/renderer/tiler.h"
+#include "graphics/host_gpu/renderer/image/tiler.h"
 
 #include "common/assert.h"
 #include "gpu_tiler_shaders/gpu_tiler_demote_d16_spv.h"
@@ -14,9 +14,9 @@
 #include "gpu_tiler_shaders/gpu_tiler_standard64_spv.h"
 #include "gpu_tiler_shaders/gpu_tiler_swap_bgra16_spv.h"
 #include "graphics/host_gpu/graphicContext.h"
+#include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
-#include "graphics/host_gpu/renderer/image.h"
-#include "graphics/host_gpu/renderer/streamBuffer.h"
+#include "graphics/host_gpu/renderer/image/image.h"
 
 #include <algorithm>
 #include <array>

src/graphics/host_gpu/renderer/masterSemaphore.cpp

@@ -26,8 +26,8 @@ MasterSemaphore::~MasterSemaphore() {
 }
 
 void MasterSemaphore::Refresh() {
-	uint64_t counter = 0;
-	const auto result = m_graphics.device.getSemaphoreCounterValue(m_semaphore, &counter);
+	uint64_t   counter = 0;
+	const auto result  = m_graphics.device.getSemaphoreCounterValue(m_semaphore, &counter);
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 
 	auto known = m_gpu_tick.load(std::memory_order_acquire);

src/graphics/host_gpu/renderer/masterSemaphore.h

@@ -22,7 +22,7 @@ public:
 	[[nodiscard]] uint64_t KnownGpuTick() const noexcept {
 		return m_gpu_tick.load(std::memory_order_acquire);
 	}
-	[[nodiscard]] bool IsFree(uint64_t tick) const noexcept { return KnownGpuTick() >= tick; }
+	[[nodiscard]] bool     IsFree(uint64_t tick) const noexcept { return KnownGpuTick() >= tick; }
 	[[nodiscard]] uint64_t NextTick() noexcept {
 		return m_current_tick.fetch_add(1, std::memory_order_release);
 	}

src/graphics/host_gpu/renderer/pipeline/descriptorCache.cpp

@@ -1,8 +1,8 @@
-#include "graphics/host_gpu/renderer/descriptorCache.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptorCache.h"
 
 #include "common/assert.h"
 #include "common/profiler.h"
-#include "graphics/shader/recompiler/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/ShaderIR.h"
 
 #include <array>
 
@@ -26,11 +26,15 @@ bool IsSampledImage(BindingKind kind) {
 		case BindingKind::Sampled1DArray:
 		case BindingKind::Sampled2D:
 		case BindingKind::Sampled2DArray:
+		case BindingKind::Sampled2DMsaa:
+		case BindingKind::Sampled2DMsaaArray:
 		case BindingKind::Sampled3D:
 		case BindingKind::SampledUint1D:
 		case BindingKind::SampledUint1DArray:
 		case BindingKind::SampledUint2D:
 		case BindingKind::SampledUint2DArray:
+		case BindingKind::SampledUint2DMsaa:
+		case BindingKind::SampledUint2DMsaaArray:
 		case BindingKind::SampledUint3D: return true;
 		default: return false;
 	}

src/graphics/host_gpu/renderer/pipeline/descriptorCache.h

@@ -6,7 +6,7 @@
 #include "common/common.h"
 #include "common/threads.h"
 #include "graphics/host_gpu/graphicContext.h"
-#include "graphics/host_gpu/renderer/textureCache.h"
+#include "graphics/host_gpu/renderer/cache/textureCache.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 #include "graphics/shader/shaderBindings.h"
 
@@ -95,7 +95,7 @@ private:
 	};
 
 	static vk::DescriptorImageInfo MakeImageInfo(const TextureBinding& texture);
-	void                 CreatePool();
+	void                           CreatePool();
 	VulkanDescriptorSet* Allocate(Stage stage, const ShaderRecompiler::IR::Program& program);
 	vk::DescriptorSetLayout
 	GetDescriptorSetLayoutInternal(Stage stage, const ShaderRecompiler::IR::Program& program);

src/graphics/host_gpu/renderer/pipeline/descriptors.cpp

@@ -1,4 +1,4 @@
-#include "graphics/host_gpu/renderer/descriptors.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptors.h"
 
 #include "common/assert.h"
 #include "common/common.h"
@@ -14,18 +14,18 @@
 #include "graphics/guest_gpu/tile.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/hostMemory.h"
-#include "graphics/host_gpu/objects/textureCommon.h"
 #include "graphics/host_gpu/renderer/debug.h"
-#include "graphics/host_gpu/renderer/descriptorCache.h"
-#include "graphics/host_gpu/renderer/imageView.h"
+#include "graphics/host_gpu/renderer/image/imageView.h"
+#include "graphics/host_gpu/renderer/image/textureCommon.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptorCache.h"
+#include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
-#include "graphics/host_gpu/renderer/shaderResourceBarrier.h"
 #include "graphics/host_gpu/vma.h"
 #include "graphics/host_gpu/vulkanCommon.h"
-#include "graphics/shader/recompiler/BindingLayout.h"
-#include "graphics/shader/recompiler/ResourceMaterialization.h"
-#include "graphics/shader/recompiler/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/BindingLayout.h"
+#include "graphics/shader/recompiler/ir/ResourceMaterialization.h"
+#include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/shader.h"
 
 #include <algorithm>
@@ -73,6 +73,11 @@ static Prospero::ImageType TextureBaseType(Prospero::ImageType type) {
 	}
 }
 
+static bool IsMultisampledTexture(Prospero::ImageType type) {
+	return type == Prospero::ImageType::kColor2DMsaa ||
+	       type == Prospero::ImageType::kColor2DMsaaArray;
+}
+
 static BufferView NativeStorageBuffer(RenderContext& context, CommandBuffer& command_buffer,
                                       const ShaderBufferResource&                 descriptor,
                                       const ShaderRecompiler::IR::BufferResource& resource,
@@ -159,6 +164,8 @@ static bool IsSupportedSampledColorResource(const ShaderRecompiler::IR::ImageRes
 		case ShaderRecompiler::Decoder::ImageDimension::Dim1DArray:
 		case ShaderRecompiler::Decoder::ImageDimension::Dim2D:
 		case ShaderRecompiler::Decoder::ImageDimension::Dim2DArray:
+		case ShaderRecompiler::Decoder::ImageDimension::Dim2DMsaa:
+		case ShaderRecompiler::Decoder::ImageDimension::Dim2DMsaaArray:
 			supported_dimension = true;
 			break;
 		default: break;
@@ -195,6 +202,22 @@ TargetTextureViewInfo ResolveTargetTextureView(const ShaderRecompiler::IR::Image
 			           ? TargetTextureViewInfo {vk::ImageViewType::e2DArray, base_layer,
 			                                    image_layers - base_layer}
 			           : TargetTextureViewInfo {};
+		case Prospero::ImageType::kColor2DMsaa:
+			return resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DMsaa &&
+			               base_layer == 0 && image_layers == 1
+			           ? TargetTextureViewInfo {vk::ImageViewType::e2D, 0, 1}
+			           : TargetTextureViewInfo {};
+		case Prospero::ImageType::kColor2DMsaaArray:
+			if (resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DMsaa &&
+			    base_layer == 0 && image_layers == 1) {
+				return {vk::ImageViewType::e2D, 0, 1};
+			}
+			return resource.dimension ==
+			                   ShaderRecompiler::Decoder::ImageDimension::Dim2DMsaaArray &&
+			               base_layer < image_layers
+			           ? TargetTextureViewInfo {vk::ImageViewType::e2DArray, base_layer,
+			                                    image_layers - base_layer}
+			           : TargetTextureViewInfo {};
 		default: return {};
 	}
 }
@@ -209,36 +232,59 @@ bool IsSupportedSampledVideoOutView(const ShaderRecompiler::IR::ImageResource& r
 }
 
 bool IsSupportedDepthTargetDescriptor(const ShaderTextureResource& descriptor, const Image& image) {
-	const auto width  = static_cast<uint32_t>(descriptor.Width5()) + 1u;
-	const auto height = static_cast<uint32_t>(descriptor.Height5()) + 1u;
-	const auto pitch  = TileGetTexturePitch(descriptor.Format(), width, 1, descriptor.TileMode());
-	const auto type   = static_cast<Prospero::ImageType>(descriptor.Type());
-	const bool supported_single_layer =
-	    image.info.resources.layers == 1 && descriptor.Depth() == 0 &&
-	    descriptor.BaseArray5() == 0 &&
-	    (type == Prospero::ImageType::kColor2D || type == Prospero::ImageType::kColor2DArray);
+	const auto width        = static_cast<uint32_t>(descriptor.Width5()) + 1u;
+	const auto height       = static_cast<uint32_t>(descriptor.Height5()) + 1u;
+	const auto type         = static_cast<Prospero::ImageType>(descriptor.Type());
+	const bool multisampled = IsMultisampledTexture(type);
+	const auto samples      = multisampled ? 1u << descriptor.LastLevel() : 1u;
+	const auto pitch =
+	    multisampled ? TileGetDepthPitch(width, image.info.bytes_per_block, descriptor.LastLevel())
+	                 : TileGetTexturePitch(descriptor.Format(), width, 1, descriptor.TileMode());
+	const bool supported_2d    = type == Prospero::ImageType::kColor2D &&
+	                             image.info.resources.layers == 1 && descriptor.Depth() == 0 &&
+	                             descriptor.BaseArray5() == 0;
+	const bool supported_array = type == Prospero::ImageType::kColor2DArray &&
+	                             descriptor.BaseArray5() <= descriptor.Depth() &&
+	                             descriptor.Depth() < image.info.resources.layers;
 	const bool supported_cube =
 	    type == Prospero::ImageType::kCube && width == height && image.info.resources.layers >= 6 &&
 	    image.info.resources.layers % 6u == 0 &&
 	    static_cast<uint32_t>(descriptor.Depth()) + 1u == image.info.resources.layers &&
 	    descriptor.BaseArray5() == 0;
+	const bool supported_msaa_2d    = type == Prospero::ImageType::kColor2DMsaa &&
+	                                  image.info.resources.layers == 1 && descriptor.Depth() == 0 &&
+	                                  descriptor.BaseArray5() == 0;
+	const bool supported_msaa_array = type == Prospero::ImageType::kColor2DMsaaArray &&
+	                                  descriptor.BaseArray5() <= descriptor.Depth() &&
+	                                  descriptor.Depth() < image.info.resources.layers;
+	const bool levels_ok =
+	    multisampled
+	        ? descriptor.BaseLevel() == 0 && descriptor.LastLevel() >= 1 &&
+	              descriptor.LastLevel() <= 3 && descriptor.MaxMip() == descriptor.LastLevel() &&
+	              image.info.resources.levels == 1 && image.info.samples == samples
+	        : descriptor.BaseLevel() == 0 && descriptor.LastLevel() == 0 &&
+	              descriptor.MaxMip() == 0 && image.info.samples == 1;
 	return image.info.IsDepth() && width == image.info.extent.width &&
-	       height == image.info.extent.height && (supported_single_layer || supported_cube) &&
-	       descriptor.BaseLevel() == 0 && descriptor.LastLevel() == 0 && descriptor.MaxMip() == 0 &&
-	       descriptor.MinLod() == 0 && descriptor.BaseArray5() == 0 &&
+	       height == image.info.extent.height &&
+	       (supported_2d || supported_array || supported_cube || supported_msaa_2d ||
+	        supported_msaa_array) &&
+	       levels_ok && descriptor.MinLod() == 0 &&
 	       descriptor.TileMode() == Prospero::GpuEnumValue(Prospero::TileMode::kDepth) &&
-	       descriptor.BCSwizzle() == 0 && !descriptor.MsaaDepth() && pitch >= width &&
-	       pitch == image.info.pitch;
+	       descriptor.BCSwizzle() == 0 && (!descriptor.MsaaDepth() || multisampled) &&
+	       pitch >= width && pitch == image.info.pitch;
 }
 
 bool IsSupportedDepthTextureEncoding(const ShaderTextureResource& descriptor, const Image& image) {
 	constexpr uint32_t field1_reserved_mask = 0x200fff00u;
 	constexpr uint32_t field2_reserved_mask = 0xf0003000u;
-	constexpr uint32_t field3_common        = 0x01800000u;
-	constexpr uint32_t field5_expected      = 0x00700000u;
-	const uint32_t     field3_expected =
-	    (descriptor.Type() << 28u) | field3_common | descriptor.DstSelXYZW();
-	const uint32_t field4_expected = descriptor.Depth() | (descriptor.BaseArray5() << 16u);
+	const uint32_t     field3_expected = descriptor.DstSelXYZW() |
+	                                     (static_cast<uint32_t>(descriptor.BaseLevel()) << 12u) |
+	                                     (static_cast<uint32_t>(descriptor.LastLevel()) << 16u) |
+	                                     (static_cast<uint32_t>(descriptor.TileMode()) << 20u) |
+	                                     (static_cast<uint32_t>(descriptor.Type()) << 28u);
+	const uint32_t     field4_expected = descriptor.Depth() | (descriptor.BaseArray5() << 16u);
+	const uint32_t     field5_expected =
+	    0x00700000u | (static_cast<uint32_t>(descriptor.MaxMip()) << 4u);
 	const bool common = (descriptor.fields[1] & field1_reserved_mask) == 0 &&
 	                    (descriptor.fields[2] & field2_reserved_mask) == 0 &&
 	                    descriptor.fields[3] == field3_expected &&
@@ -251,8 +297,9 @@ bool IsSupportedDepthTextureEncoding(const ShaderTextureResource& descriptor, co
 		return true;
 	}
 	constexpr uint32_t htile_control = 0x00280000u;
-	const auto         metadata_addr = descriptor.MetaAddr() << 8u;
-	return (descriptor.fields[6] & 0x00ffffffu) == htile_control && metadata_addr != 0 &&
+	const uint32_t expected_control  = htile_control | (descriptor.MsaaDepth() ? (1u << 10u) : 0u);
+	const auto     metadata_addr     = descriptor.MetaAddr() << 8u;
+	return (descriptor.fields[6] & 0x00ffffffu) == expected_control && metadata_addr != 0 &&
 	       descriptor.TileMode() == Prospero::GpuEnumValue(Prospero::TileMode::kDepth) &&
 	       image.info.tile_mode == Prospero::GpuEnumValue(Prospero::TileMode::kDepth) &&
 	       image.info.metadata.kind == ImageMetadataKind::Htile &&
@@ -318,8 +365,8 @@ static bool IsSupportedStorageTextureDescriptor(const ShaderRecompiler::IR::Imag
 	const bool valid_2d_slice =
 	    (is_color_2d && descriptor.Depth() == 0 && descriptor.BaseArray5() == 0) ||
 	    (is_color_2d_array && descriptor.BaseArray5() <= descriptor.Depth());
-	const bool is_2d = resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2D &&
-	                   valid_2d_slice;
+	const bool is_2d =
+	    resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2D && valid_2d_slice;
 	const bool is_2d_array =
 	    resource.dimension == ShaderRecompiler::Decoder::ImageDimension::Dim2DArray &&
 	    is_color_2d_array && descriptor.BaseArray5() <= descriptor.Depth();
@@ -333,18 +380,20 @@ static bool IsSupportedStorageTextureDescriptor(const ShaderRecompiler::IR::Imag
 	    !Prospero::IsFmaskTextureFormat(descriptor.Format()) && (is_2d || is_2d_array) &&
 	    TileGetBlockLayout(TileBlockFamily::Depth64KB, depth_bpe, depth_block);
 	const bool supported_standard_tile =
-	    tile == Prospero::GpuEnumValue(Prospero::TileMode::kStandard4KB) &&
-	    TileIsStandard4KBTextureSupported(descriptor.Format());
+	    (tile == Prospero::GpuEnumValue(Prospero::TileMode::kStandard4KB) &&
+	     TileIsStandard4KBTextureSupported(descriptor.Format())) ||
+	    (tile == Prospero::GpuEnumValue(Prospero::TileMode::kStandard64KB) &&
+	     TileIsStandard64KBTextureSupported(descriptor.Format()));
 	const bool supported_tile = tile == Prospero::GpuEnumValue(Prospero::TileMode::kLinear) ||
 	                            tile == Prospero::GpuEnumValue(Prospero::TileMode::kRenderTarget) ||
 	                            supported_depth_tile || supported_standard_tile;
+	const auto swizzle        = descriptor.DstSelXYZW();
 	const bool supported_swizzle =
-	    IsValidImageSwizzle(descriptor.DstSelXYZW()) &&
-	    (descriptor.DstSelXYZW() == DstSel(4, 5, 6, 7) || !resource.read);
+	    IsValidImageSwizzle(swizzle) &&
+	    (swizzle == DstSel(4, 5, 6, 7) || !resource.read || resource.atomic);
 	const bool supported_mip_view = descriptor.BaseLevel() == 0 || is_1d || is_2d;
 	return (is_1d || is_1d_array || is_2d || is_2d_array || is_3d) && supported_tile &&
-	       supported_mip_view &&
-	       descriptor.BaseLevel() == descriptor.LastLevel() &&
+	       supported_mip_view && descriptor.BaseLevel() == descriptor.LastLevel() &&
 	       descriptor.LastLevel() <= descriptor.MaxMip() && descriptor.MinLod() == 0 &&
 	       supported_swizzle && descriptor.BCSwizzle() == 0 && !descriptor.MsaaDepth();
 }
@@ -375,8 +424,14 @@ void ValidateStorageTexture(const ShaderRecompiler::IR::ImageResource& resource,
 	const bool encoding_ok   = IsSupportedStorageTextureEncoding(descriptor);
 	const bool uint_resource =
 	    resource.kind == ShaderRecompiler::IR::ResourceKind::StorageImageUint;
-	const bool format_ok = Prospero::IsSupportedTextureFormat(format) &&
-	                       uint_resource == Prospero::IsUintTextureFormat(format);
+	const bool raw_sint_storage =
+	    format == Prospero::GpuEnumValue(Prospero::BufferFormat::k32SInt) && uint_resource &&
+	    resource.written && !resource.read && !resource.atomic;
+	const bool format_ok =
+	    raw_sint_storage ||
+	    (Prospero::IsSupportedTextureFormat(format) &&
+	     uint_resource == Prospero::IsUintTextureFormat(format) &&
+	     (!resource.atomic || format == Prospero::GpuEnumValue(Prospero::BufferFormat::k32UInt)));
 	if (resource_ok && descriptor_ok && encoding_ok && format_ok && size != 0) {
 		return;
 	}
@@ -514,6 +569,7 @@ static ImageViewInfo TextureViewInfo(const ShaderRecompiler::IR::ImageResource&
 			view.layer_count = 1;
 			break;
 		case ShaderRecompiler::Decoder::ImageDimension::Dim2DArray:
+		case ShaderRecompiler::Decoder::ImageDimension::Dim2DMsaaArray:
 			view.type       = vk::ImageViewType::e2DArray;
 			view.base_layer = descriptor.BaseArray5();
 			if (view.base_layer >= image_layers) {
@@ -522,6 +578,7 @@ static ImageViewInfo TextureViewInfo(const ShaderRecompiler::IR::ImageResource&
 			view.layer_count = image_layers - view.base_layer;
 			break;
 		case ShaderRecompiler::Decoder::ImageDimension::Dim2D:
+		case ShaderRecompiler::Decoder::ImageDimension::Dim2DMsaa:
 			view.type       = vk::ImageViewType::e2D;
 			view.base_layer = descriptor.BaseArray5();
 			if (view.base_layer >= image_layers) {
@@ -552,25 +609,32 @@ RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageResource&   reso
 		return {id, nullptr, std::move(desc)};
 	}
 
-	const auto address    = descriptor.Base40();
-	const auto width      = static_cast<uint32_t>(descriptor.Width5()) + 1u;
-	const auto height     = static_cast<uint32_t>(descriptor.Height5()) + 1u;
-	const auto base_level = descriptor.BaseLevel();
-	const auto last_level = descriptor.LastLevel();
-	const auto type       = TextureType(descriptor);
-	const bool multisampled =
-	    type == Prospero::ImageType::kColor2DMsaa || type == Prospero::ImageType::kColor2DMsaaArray;
-	const auto levels     = multisampled ? 1u : static_cast<uint32_t>(descriptor.MaxMip()) + 1u;
-	const auto tile       = descriptor.TileMode();
-	const bool msaa_tile  = tile == Prospero::GpuEnumValue(Prospero::TileMode::kRenderTarget);
+	const auto address      = descriptor.Base40();
+	const auto width        = static_cast<uint32_t>(descriptor.Width5()) + 1u;
+	const auto height       = static_cast<uint32_t>(descriptor.Height5()) + 1u;
+	const auto base_level   = descriptor.BaseLevel();
+	const auto last_level   = descriptor.LastLevel();
+	const auto type         = TextureType(descriptor);
+	const bool multisampled = IsMultisampledTexture(type);
+	const auto levels       = multisampled ? 1u : static_cast<uint32_t>(descriptor.MaxMip()) + 1u;
+	const auto tile         = descriptor.TileMode();
+	const bool depth_tile = tile == Prospero::GpuEnumValue(Prospero::TileMode::kDepth);
+	const bool msaa_tile =
+	    depth_tile || tile == Prospero::GpuEnumValue(Prospero::TileMode::kRenderTarget);
 	const bool msaa_array = type == Prospero::ImageType::kColor2DMsaaArray;
 	if ((!multisampled && (base_level > last_level || last_level >= levels)) ||
 	    (multisampled &&
 	     (base_level != 0 || last_level == 0 || last_level > 3 ||
-	      descriptor.MaxMip() != last_level || !msaa_tile || descriptor.MsaaDepth() ||
+	      descriptor.MaxMip() != last_level || !msaa_tile || (descriptor.MsaaDepth() && !depth_tile) ||
 	      (!msaa_array && (descriptor.Depth() != 0 || descriptor.BaseArray5() != 0))))) {
-		EXIT("unsupported texture mip view: base=%u last=%u levels=%u\n", base_level, last_level,
-		     levels);
+		EXIT("unsupported texture mip view: base=%u last=%u levels=%u max=%u type=%u tile=%u "
+		     "kind=%u dimension=%u mip_mode=%u read=%d written=%d "
+		     "dwords=%08x,%08x,%08x,%08x,%08x,%08x,%08x,%08x\n",
+		     base_level, last_level, levels, descriptor.MaxMip(), descriptor.Type(), tile,
+		     static_cast<uint32_t>(resource.kind), static_cast<uint32_t>(resource.dimension),
+		     static_cast<uint32_t>(resource.mip_mode), resource.read, resource.written,
+		     descriptor.fields[0], descriptor.fields[1], descriptor.fields[2], descriptor.fields[3],
+		     descriptor.fields[4], descriptor.fields[5], descriptor.fields[6], descriptor.fields[7]);
 	}
 	const auto samples = multisampled ? 1u << last_level : 1u;
 	const auto view_levels =
@@ -597,7 +661,8 @@ RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageResource&   reso
 	TileSizeAlign size {};
 	if (multisampled) {
 		const auto bytes = Prospero::NumBytesPerElement(format);
-		pitch            = TileGetRenderTargetPitch(width, bytes, last_level);
+		pitch = depth_tile ? TileGetDepthPitch(width, bytes, last_level)
+		                   : TileGetRenderTargetPitch(width, bytes, last_level);
 		if (pitch == 0 || !TileGetRenderTargetSize(width, height, pitch, bytes, size, last_level) ||
 		    size.size > UINT32_MAX / image_layers) {
 			EXIT("unsupported multisample texture layout\n");
@@ -612,16 +677,17 @@ RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageResource&   reso
 	                     (address & (static_cast<uint64_t>(size.align) - 1u)) != 0);
 	if (storage) {
 		ValidateStorageTexture(resource, descriptor, size.size);
-		m_context.GetBufferCache().ValidateGpuAccess(address, size.size, resource.read,
-		                                             resource.written);
 	}
 
 	const auto              pixel_format = TextureGetFormat(format);
-	const auto              storage_view_format = SrgbStorageViewFormat(pixel_format);
-	const auto              view_format =
-	    storage && storage_view_format != vk::Format::eUndefined ? storage_view_format
-	                                                            : pixel_format;
-	const auto              block_bytes  = Prospero::BlockCompressedBytesPerBlock(format);
+	const auto              storage_view_format =
+	    storage && format == Prospero::GpuEnumValue(Prospero::BufferFormat::k32SInt)
+	        ? vk::Format::eR32Uint
+	        : SrgbStorageViewFormat(pixel_format);
+	const auto              view_format = storage && storage_view_format != vk::Format::eUndefined
+	                                          ? storage_view_format
+	                                          : pixel_format;
+	const auto              block_bytes = Prospero::BlockCompressedBytesPerBlock(format);
 	TextureCache::ImageDesc desc {};
 	desc.info.data         = {address, size.size};
 	desc.info.pixel_format = pixel_format;

src/graphics/host_gpu/renderer/pipeline/descriptors.h

@@ -2,9 +2,9 @@
 #define EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_DESCRIPTORS_H_
 
 #include "common/assert.h"
-#include "graphics/host_gpu/renderer/image.h"
+#include "graphics/host_gpu/renderer/image/image.h"
 #include "graphics/host_gpu/vulkanCommon.h"
-#include "graphics/shader/recompiler/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/shaderBindings.h"
 
 #include <cstdint>
@@ -36,7 +36,7 @@ ResolveTargetTextureView(const ShaderRecompiler::IR::ImageResource& resource,
 [[nodiscard]] bool IsSupportedDepthTargetDescriptor(const ShaderTextureResource& descriptor,
                                                     const Image&                 image);
 [[nodiscard]] bool IsSupportedDepthTextureEncoding(const ShaderTextureResource& descriptor,
-                                                    const Image&                 image);
+                                                   const Image&                 image);
 [[nodiscard]] bool
 IsSupportedSampledVideoOutView(const ShaderRecompiler::IR::ImageResource& resource,
                                const ShaderTextureResource& descriptor, const Image& image);

src/graphics/host_gpu/renderer/pipeline/pipelineCache.cpp

@@ -1,4 +1,4 @@
-#include "graphics/host_gpu/renderer/pipelineCache.h"
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
 
 #include "common/assert.h"
 #include "common/logging/log.h"
@@ -7,7 +7,7 @@
 #include "graphics/host_gpu/renderer/colorRenderTarget.h"
 #include "graphics/host_gpu/renderer/debug.h"
 #include "graphics/host_gpu/renderer/depthRenderTarget.h"
-#include "graphics/host_gpu/renderer/imageView.h"
+#include "graphics/host_gpu/renderer/image/imageView.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 
@@ -88,12 +88,12 @@ PipelineCache::GraphicsPipeline& PipelineCache::CreateGraphicsPipeline(
 
 	PipelineStaticParameters static_params {};
 	GraphicsPipeline         p {};
-	p.ps_shader_id   = ps_id;
-	p.vs_shader_id   = vs_id;
+	p.ps_shader_id = ps_id;
+	p.vs_shader_id = vs_id;
 
 	static_params.color_count = color_count;
 	PipelineRenderingState rendering {};
-	rendering.color_count = color_count;
+	rendering.color_count       = color_count;
 	uint32_t attachment_samples = 0;
 	for (uint32_t i = 0; i < color_count; i++) {
 		EXIT_IF(!colors[i].image_id || colors[i].format == vk::Format::eUndefined);
@@ -116,8 +116,8 @@ PipelineCache::GraphicsPipeline& PipelineCache::CreateGraphicsPipeline(
 		if (attachment_samples == 0) {
 			attachment_samples = depth.samples;
 		} else if (attachment_samples != depth.samples) {
-			EXIT("mixed color/depth sample counts are unsupported: %u and %u\n",
-			     attachment_samples, depth.samples);
+			EXIT("mixed color/depth sample counts are unsupported: %u and %u\n", attachment_samples,
+			     depth.samples);
 		}
 	}
 	EXIT_IF(attachment_samples == 0 ||
@@ -179,10 +179,10 @@ PipelineCache::GraphicsPipeline& PipelineCache::CreateGraphicsPipeline(
 	NormalizeStaticParamsForDynamicState(static_params);
 
 	GraphicsPipelineKey key {};
-	key.rendering      = rendering;
-	key.vs_shader_id   = p.vs_shader_id;
-	key.ps_shader_id   = p.ps_shader_id;
-	key.static_params  = static_params;
+	key.rendering     = rendering;
+	key.vs_shader_id  = p.vs_shader_id;
+	key.ps_shader_id  = p.ps_shader_id;
+	key.static_params = static_params;
 
 	if (auto iter = m_graphics_pipelines.find(key); iter != m_graphics_pipelines.end()) {
 		return *iter->second;
@@ -203,9 +203,8 @@ PipelineCache::GraphicsPipeline& PipelineCache::CreateGraphicsPipeline(
 	LogPipelineTrace("CreatePipelineInternal begin", vs_id.hash0, vs_id.crc32, ps_id.hash0,
 	                 ps_id.crc32);
 	CreatePipelineInternal(m_graphics, m_descriptor_cache, *cached, rendering, vs_input_info,
-	                       vs_spirv, ps_input_info,
-	                       ps_spirv, static_params, vs_id.hash0, vs_id.crc32, ps_id.hash0,
-	                       ps_id.crc32, ps_active);
+	                       vs_spirv, ps_input_info, ps_spirv, static_params, vs_id.hash0,
+	                       vs_id.crc32, ps_id.hash0, ps_id.crc32, ps_active);
 	LogPipelineTrace("CreatePipelineInternal done", vs_id.hash0, vs_id.crc32, ps_id.hash0,
 	                 ps_id.crc32);
 

src/graphics/host_gpu/renderer/pipeline/pipelineCache.h

@@ -88,9 +88,9 @@ static_assert(sizeof(PipelineStaticParameters) ==
 
 struct PipelineRenderingState {
 	std::array<vk::Format, RENDER_COLOR_ATTACHMENTS_MAX> color_formats {};
-	vk::Format depth_format   = vk::Format::eUndefined;
-	vk::Format stencil_format = vk::Format::eUndefined;
-	uint32_t   color_count    = 0;
+	vk::Format                                           depth_format   = vk::Format::eUndefined;
+	vk::Format                                           stencil_format = vk::Format::eUndefined;
+	uint32_t                                             color_count    = 0;
 
 	bool operator==(const PipelineRenderingState&) const = default;
 };
@@ -118,11 +118,12 @@ public:
 		ShaderId cs_shader_id;
 	};
 
-	GraphicsPipeline& CreateGraphicsPipeline(
-	    RenderColorInfo* colors, uint32_t color_count, RenderDepthInfo& depth,
-	    ShaderVertexInputInfo& vs_input_info, RenderCommandBuffer& command,
-	    ShaderPixelInputInfo* ps_input_info, vk::PrimitiveTopology topology, bool ps_active,
-	    std::span<const uint32_t> vs_spirv, std::span<const uint32_t> ps_spirv);
+	GraphicsPipeline&
+	CreateGraphicsPipeline(RenderColorInfo* colors, uint32_t color_count, RenderDepthInfo& depth,
+	                       ShaderVertexInputInfo& vs_input_info, RenderCommandBuffer& command,
+	                       ShaderPixelInputInfo* ps_input_info, vk::PrimitiveTopology topology,
+	                       bool ps_active, std::span<const uint32_t> vs_spirv,
+	                       std::span<const uint32_t> ps_spirv);
 	ComputePipeline& CreateComputePipeline(ShaderComputeInputInfo&      input_info,
 	                                       const HW::ComputeShaderInfo& cs_regs,
 	                                       std::span<const uint32_t>    cs_spirv);
@@ -199,7 +200,7 @@ private:
 		}
 	};
 
-	GraphicContext& m_graphics;
+	GraphicContext&  m_graphics;
 	DescriptorCache& m_descriptor_cache;
 	std::unordered_map<GraphicsPipelineKey, std::unique_ptr<GraphicsPipeline>,
 	                   GraphicsPipelineKeyHash>
@@ -211,16 +212,13 @@ private:
 
 void LogPipelineTrace(const char* phase, uint32_t vs_hash0, uint32_t vs_crc32, uint32_t ps_hash0,
                       uint32_t ps_crc32);
-void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descriptor_cache,
-                            PipelineCache::GraphicsPipeline& pipeline,
-                            const PipelineRenderingState&   rendering,
-                            const ShaderVertexInputInfo&    vs_input_info,
-                            std::span<const uint32_t>       vs_shader,
-                            const ShaderPixelInputInfo*     ps_input_info,
-                            std::span<const uint32_t>       ps_shader,
-                            const PipelineStaticParameters& static_params, uint32_t vs_hash0,
-                            uint32_t vs_crc32, uint32_t ps_hash0, uint32_t ps_crc32,
-                            bool ps_active);
+void CreatePipelineInternal(
+    GraphicContext& graphics, DescriptorCache& descriptor_cache,
+    PipelineCache::GraphicsPipeline& pipeline, const PipelineRenderingState& rendering,
+    const ShaderVertexInputInfo& vs_input_info, std::span<const uint32_t> vs_shader,
+    const ShaderPixelInputInfo* ps_input_info, std::span<const uint32_t> ps_shader,
+    const PipelineStaticParameters& static_params, uint32_t vs_hash0, uint32_t vs_crc32,
+    uint32_t ps_hash0, uint32_t ps_crc32, bool ps_active);
 void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descriptor_cache,
                             PipelineCache::ComputePipeline& pipeline,
                             const ShaderComputeInputInfo&   input_info,

src/graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.cpp

@@ -1,4 +1,4 @@
-#include "graphics/host_gpu/renderer/shaderResourceBarrier.h"
+#include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
 
 #include "common/assert.h"
 #include "graphics/shader/shader.h"

src/graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h

@@ -2,7 +2,7 @@
 #define EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_SHADERRESOURCEBARRIER_H_
 
 #include "graphics/host_gpu/graphicContext.h"
-#include "graphics/shader/recompiler/ResourceMaterialization.h"
+#include "graphics/shader/recompiler/ir/ResourceMaterialization.h"
 
 #include <vector>
 

src/graphics/host_gpu/renderer/pipeline/shaderSubgroup.cpp

@@ -1,6 +1,6 @@
-#include "graphics/host_gpu/renderer/shaderSubgroup.h"
+#include "graphics/host_gpu/renderer/pipeline/shaderSubgroup.h"
 
-#include "graphics/shader/recompiler/SpirvEmitter.h"
+#include "graphics/shader/recompiler/emitter/SpirvEmitter.h"
 
 namespace Libs::Graphics {
 

src/graphics/host_gpu/renderer/pipeline/shaderSubgroup.h

@@ -2,7 +2,7 @@
 #define EMULATOR_SRC_GRAPHICS_HOST_GPU_RENDERER_SHADERSUBGROUP_H_
 
 #include "graphics/host_gpu/graphicContext.h"
-#include "graphics/shader/recompiler/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/ShaderIR.h"
 
 namespace Libs::Graphics {
 

src/graphics/host_gpu/renderer/pipeline/shaders.cpp

@@ -6,14 +6,14 @@
 #include "graphics/guest_gpu/gpu_defs.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/renderer/debug.h"
-#include "graphics/host_gpu/renderer/descriptorCache.h"
-#include "graphics/host_gpu/renderer/pipelineCache.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptorCache.h"
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
+#include "graphics/host_gpu/renderer/pipeline/shaderSubgroup.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
-#include "graphics/host_gpu/renderer/shaderSubgroup.h"
 #include "graphics/host_gpu/vulkanCommon.h"
-#include "graphics/shader/recompiler/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/shader.h"
 
 #include <algorithm>
@@ -385,9 +385,8 @@ static vk::BlendOp GetBlendOp(uint32_t op) {
 	return vk::BlendOp::eAdd;
 }
 
-static void CreateLayout(DescriptorCache& descriptor_cache,
-                         std::span<vk::DescriptorSetLayout> set_layouts,
-                         uint32_t& set_layouts_num,
+static void CreateLayout(DescriptorCache&                   descriptor_cache,
+                         std::span<vk::DescriptorSetLayout> set_layouts, uint32_t& set_layouts_num,
                          std::span<vk::PushConstantRange>     push_constant_info,
                          uint32_t&                            push_constant_info_num,
                          const ShaderRecompiler::IR::Program& program,
@@ -412,12 +411,11 @@ static void CreateLayout(DescriptorCache& descriptor_cache,
 	}
 }
 
-static void ConfigureSubgroupSize(const GraphicContext&                                  graphics,
-                                  vk::ShaderStageFlagBits                                vk_stage,
+static void ConfigureSubgroupSize(const GraphicContext& graphics, vk::ShaderStageFlagBits vk_stage,
                                   const ShaderRecompiler::IR::Program&                   program,
                                   vk::PipelineShaderStageRequiredSubgroupSizeCreateInfo& required,
                                   vk::PipelineShaderStageCreateInfo&                     stage) {
-	const auto  config =
+	const auto config =
 	    ConfigureShaderSubgroup(ShaderSubgroupCapabilities {graphics}, vk_stage, program);
 	switch (config.mode) {
 		case ShaderSubgroupMode::Natural: return;
@@ -456,16 +454,13 @@ static void ConfigureSubgroupSize(const GraphicContext&
 }
 
 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
-void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descriptor_cache,
-                            PipelineCache::GraphicsPipeline& pipeline,
-                            const PipelineRenderingState&   rendering,
-                            const ShaderVertexInputInfo&    vs_input_info,
-                            std::span<const uint32_t>       vs_shader,
-                            const ShaderPixelInputInfo*     ps_input_info,
-                            std::span<const uint32_t>       ps_shader,
-                            const PipelineStaticParameters& static_params, uint32_t vs_hash0,
-                            uint32_t vs_crc32, uint32_t ps_hash0, uint32_t ps_crc32,
-                            bool ps_active) {
+void CreatePipelineInternal(
+    GraphicContext& graphics, DescriptorCache& descriptor_cache,
+    PipelineCache::GraphicsPipeline& pipeline, const PipelineRenderingState& rendering,
+    const ShaderVertexInputInfo& vs_input_info, std::span<const uint32_t> vs_shader,
+    const ShaderPixelInputInfo* ps_input_info, std::span<const uint32_t> ps_shader,
+    const PipelineStaticParameters& static_params, uint32_t vs_hash0, uint32_t vs_crc32,
+    uint32_t ps_hash0, uint32_t ps_crc32, bool ps_active) {
 	EXIT_IF(ps_active && ps_input_info == nullptr);
 
 	vk::ShaderModule vert_shader_module = nullptr;
@@ -511,8 +506,7 @@ void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descripto
 	vert_shader_stage_info.pName               = "main";
 	vert_shader_stage_info.pSpecializationInfo = nullptr;
 	EXIT_IF(!vs_input_info.stage);
-	ConfigureSubgroupSize(graphics, vk::ShaderStageFlagBits::eVertex,
-	                      *vs_input_info.stage.program,
+	ConfigureSubgroupSize(graphics, vk::ShaderStageFlagBits::eVertex, *vs_input_info.stage.program,
 	                      vert_subgroup_size, vert_shader_stage_info);
 
 	vk::PipelineShaderStageCreateInfo                     frag_shader_stage_info {};
@@ -527,8 +521,8 @@ void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descripto
 	if (ps_active) {
 		EXIT_IF(!ps_input_info->stage);
 		ConfigureSubgroupSize(graphics, vk::ShaderStageFlagBits::eFragment,
-		                      *ps_input_info->stage.program,
-		                      frag_subgroup_size, frag_shader_stage_info);
+		                      *ps_input_info->stage.program, frag_subgroup_size,
+		                      frag_shader_stage_info);
 	}
 
 	vk::PipelineShaderStageCreateInfo shader_stages[]    = {vert_shader_stage_info,
@@ -728,8 +722,14 @@ void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descripto
 	clip_ext.depthClipEnable = static_params.depth_clip_enable ? VK_TRUE : VK_FALSE;
 
 	vk::PipelineRasterizationStateCreateInfo rasterizer {};
-	rasterizer.sType                   = vk::StructureType::ePipelineRasterizationStateCreateInfo;
-	rasterizer.pNext                   = &clip_ext;
+	rasterizer.sType = vk::StructureType::ePipelineRasterizationStateCreateInfo;
+	// MoltenVK lacks VK_EXT_depth_clip_enable; omit the depth-clip struct on macOS and accept
+	// Vulkan's default depth clipping (enabled) instead of the PS5's clamp behavior.
+#if defined(__APPLE__)
+	rasterizer.pNext = nullptr;
+#else
+	rasterizer.pNext = &clip_ext;
+#endif
 	rasterizer.flags                   = {};
 	rasterizer.depthClampEnable        = VK_FALSE;
 	rasterizer.rasterizerDiscardEnable = VK_FALSE;
@@ -806,8 +806,14 @@ void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descripto
 	color_write.pColorWriteEnables = color_write_enable;
 
 	vk::PipelineColorBlendStateCreateInfo color_blending {};
-	color_blending.sType             = vk::StructureType::ePipelineColorBlendStateCreateInfo;
-	color_blending.pNext             = &color_write;
+	color_blending.sType = vk::StructureType::ePipelineColorBlendStateCreateInfo;
+	// MoltenVK lacks VK_EXT_color_write_enable; drop the dynamic color-write struct on macOS
+	// and rely on each attachment's static colorWriteMask (all channels enabled by default).
+#if defined(__APPLE__)
+	color_blending.pNext = nullptr;
+#else
+	color_blending.pNext = &color_write;
+#endif
 	color_blending.flags             = {};
 	color_blending.logicOpEnable     = VK_FALSE;
 	color_blending.logicOp           = vk::LogicOp::eCopy;
@@ -826,15 +832,13 @@ void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descripto
 
 	EXIT_IF(!vs_input_info.stage);
 	CreateLayout(descriptor_cache, set_layouts, set_layouts_num, push_constant_info,
-	             push_constant_info_num,
-	             *vs_input_info.stage.program, vk::ShaderStageFlagBits::eVertex,
-	             DescriptorCache::Stage::Vertex);
+	             push_constant_info_num, *vs_input_info.stage.program,
+	             vk::ShaderStageFlagBits::eVertex, DescriptorCache::Stage::Vertex);
 	if (ps_active) {
 		EXIT_IF(!ps_input_info->stage);
 		CreateLayout(descriptor_cache, set_layouts, set_layouts_num, push_constant_info,
-		             push_constant_info_num,
-		             *ps_input_info->stage.program, vk::ShaderStageFlagBits::eFragment,
-		             DescriptorCache::Stage::Pixel);
+		             push_constant_info_num, *ps_input_info->stage.program,
+		             vk::ShaderStageFlagBits::eFragment, DescriptorCache::Stage::Pixel);
 	}
 
 	vk::PipelineLayoutCreateInfo pipeline_layout_info {};
@@ -873,7 +877,11 @@ void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descripto
 	depth_stencil_info.depthWriteEnable = (static_params.depth_write_enable ? VK_TRUE : VK_FALSE);
 	depth_stencil_info.depthCompareOp   = static_params.depth_compare_op;
 	depth_stencil_info.depthBoundsTestEnable =
+#if defined(__APPLE__)
+	    VK_FALSE; // MoltenVK lacks the depthBounds feature; depth-bounds testing is disabled
+#else
 	    (static_params.depth_bounds_test_enable ? VK_TRUE : VK_FALSE);
+#endif
 	depth_stencil_info.stencilTestEnable = (static_params.stencil_test_enable ? VK_TRUE : VK_FALSE);
 	depth_stencil_info.front.failOp      = static_params.stencil_front.failOp;
 	depth_stencil_info.front.passOp      = static_params.stencil_front.passOp;
@@ -893,7 +901,9 @@ void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descripto
 	    vk::DynamicState::eStencilCompareMask,
 	    vk::DynamicState::eStencilReference,
 	    vk::DynamicState::eStencilWriteMask,
-	    vk::DynamicState::eColorWriteEnableEXT,
+#if !defined(__APPLE__)
+	    vk::DynamicState::eColorWriteEnableEXT, // unsupported by MoltenVK; static mask instead
+#endif
 	};
 	const auto dynamic_states_count =
 	    static_cast<uint32_t>(sizeof(dynamic_states) / sizeof(dynamic_states[0]));
@@ -905,32 +915,32 @@ void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descripto
 	dynamic_state.dynamicStateCount = dynamic_states_count;
 	dynamic_state.pDynamicStates    = dynamic_states;
 
-	vk::GraphicsPipelineCreateInfo pipeline_info {};
+	vk::GraphicsPipelineCreateInfo  pipeline_info {};
 	vk::PipelineRenderingCreateInfo rendering_info {};
 	rendering_info.sType                   = vk::StructureType::ePipelineRenderingCreateInfo;
 	rendering_info.colorAttachmentCount    = rendering.color_count;
 	rendering_info.pColorAttachmentFormats = rendering.color_formats.data();
 	rendering_info.depthAttachmentFormat   = rendering.depth_format;
 	rendering_info.stencilAttachmentFormat = rendering.stencil_format;
-	pipeline_info.sType               = vk::StructureType::eGraphicsPipelineCreateInfo;
-	pipeline_info.pNext               = &rendering_info;
-	pipeline_info.flags               = {};
-	pipeline_info.stageCount          = shader_stage_count;
-	pipeline_info.pStages             = shader_stages;
-	pipeline_info.pVertexInputState   = &vertex_input_info;
-	pipeline_info.pInputAssemblyState = &input_assembly;
-	pipeline_info.pTessellationState  = nullptr;
-	pipeline_info.pViewportState      = &viewport_state;
-	pipeline_info.pRasterizationState = &rasterizer;
-	pipeline_info.pMultisampleState   = &multisampling;
-	pipeline_info.pDepthStencilState  = (static_params.with_depth ? &depth_stencil_info : nullptr);
-	pipeline_info.pColorBlendState    = &color_blending;
-	pipeline_info.pDynamicState       = &dynamic_state;
-	pipeline_info.layout              = pipeline.pipeline_layout;
-	pipeline_info.renderPass          = nullptr;
-	pipeline_info.subpass             = 0;
-	pipeline_info.basePipelineHandle  = nullptr;
-	pipeline_info.basePipelineIndex   = -1;
+	pipeline_info.sType                    = vk::StructureType::eGraphicsPipelineCreateInfo;
+	pipeline_info.pNext                    = &rendering_info;
+	pipeline_info.flags                    = {};
+	pipeline_info.stageCount               = shader_stage_count;
+	pipeline_info.pStages                  = shader_stages;
+	pipeline_info.pVertexInputState        = &vertex_input_info;
+	pipeline_info.pInputAssemblyState      = &input_assembly;
+	pipeline_info.pTessellationState       = nullptr;
+	pipeline_info.pViewportState           = &viewport_state;
+	pipeline_info.pRasterizationState      = &rasterizer;
+	pipeline_info.pMultisampleState        = &multisampling;
+	pipeline_info.pDepthStencilState = (static_params.with_depth ? &depth_stencil_info : nullptr);
+	pipeline_info.pColorBlendState   = &color_blending;
+	pipeline_info.pDynamicState      = &dynamic_state;
+	pipeline_info.layout             = pipeline.pipeline_layout;
+	pipeline_info.renderPass         = nullptr;
+	pipeline_info.subpass            = 0;
+	pipeline_info.basePipelineHandle = nullptr;
+	pipeline_info.basePipelineIndex  = -1;
 
 	EXIT_IF(pipeline.pipeline != nullptr);
 
@@ -994,8 +1004,7 @@ void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descripto
 	comp_shader_stage_info.pName               = "main";
 	comp_shader_stage_info.pSpecializationInfo = nullptr;
 	EXIT_IF(!input_info.stage);
-	ConfigureSubgroupSize(graphics, vk::ShaderStageFlagBits::eCompute,
-	                      *input_info.stage.program,
+	ConfigureSubgroupSize(graphics, vk::ShaderStageFlagBits::eCompute, *input_info.stage.program,
 	                      comp_subgroup_size, comp_shader_stage_info);
 
 	vk::DescriptorSetLayout set_layouts[1]  = {};
@@ -1006,9 +1015,8 @@ void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descripto
 
 	EXIT_IF(!input_info.stage);
 	CreateLayout(descriptor_cache, set_layouts, set_layouts_num, push_constant_info,
-	             push_constant_info_num,
-	             *input_info.stage.program, vk::ShaderStageFlagBits::eCompute,
-	             DescriptorCache::Stage::Compute);
+	             push_constant_info_num, *input_info.stage.program,
+	             vk::ShaderStageFlagBits::eCompute, DescriptorCache::Stage::Compute);
 
 	vk::PipelineLayoutCreateInfo pipeline_layout_info {};
 	pipeline_layout_info.sType                  = vk::StructureType::ePipelineLayoutCreateInfo;

src/graphics/host_gpu/renderer/render.h

@@ -4,7 +4,7 @@
 #include "common/abi.h"
 #include "common/assert.h"
 #include "common/common.h"
-#include "graphics/host_gpu/renderer/descriptorCache.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptorCache.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -10,17 +10,17 @@
 #include "graphics/guest_gpu/graphicsRun.h"
 #include "graphics/guest_gpu/hardwareContext.h"
 #include "graphics/host_gpu/graphicContext.h"
-#include "graphics/host_gpu/renderer/descriptorCache.h"
-#include "graphics/host_gpu/renderer/descriptors.h"
-#include "graphics/host_gpu/renderer/imageInfo.h"
-#include "graphics/host_gpu/renderer/pipelineCache.h"
+#include "graphics/host_gpu/renderer/image/imageInfo.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptorCache.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptors.h"
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
+#include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
+#include "graphics/host_gpu/renderer/pipeline/shaderSubgroup.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
-#include "graphics/host_gpu/renderer/shaderResourceBarrier.h"
-#include "graphics/host_gpu/renderer/shaderSubgroup.h"
 #include "graphics/host_gpu/vulkanCommon.h"
-#include "graphics/shader/recompiler/ResourceMaterialization.h"
-#include "graphics/shader/recompiler/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/ResourceMaterialization.h"
+#include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/shader.h"
 #include "kernel/eventQueue.h"
 #include "kernel/pthread.h"

src/graphics/host_gpu/renderer/renderContext.cpp

@@ -14,8 +14,7 @@ namespace Libs::Graphics {
 RenderContext::RenderContext(GraphicContext& graphics)
     : m_graphics(graphics), m_render_executor(*this), m_command_scheduler(*this, graphics),
       m_descriptor_cache(graphics), m_pipeline_cache(graphics, m_descriptor_cache),
-      m_sampler_cache(graphics),
-      m_gpu_resources(graphics, m_command_scheduler) {
+      m_sampler_cache(graphics), m_gpu_resources(graphics, m_command_scheduler) {
 	EXIT_NOT_IMPLEMENTED(!Common::Thread::IsMainThread());
 }
 
@@ -27,7 +26,7 @@ RenderContext::~RenderContext() {
 void RenderContext::InitializeGpu(VideoOut::VideoOutDriver* video_out) {
 	EXIT_IF(m_gpu != nullptr);
 	m_video_out = video_out;
-	m_gpu = std::make_unique<Gpu>(*this);
+	m_gpu       = std::make_unique<Gpu>(*this);
 	m_gpu_resources.SetGpu(m_gpu.get());
 }
 
@@ -99,8 +98,7 @@ void RenderContext::TriggerEopEvent(uint32_t context_id) {
 		    registration.eq, static_cast<uintptr_t>(registration.id),
 		    LibKernel::EventQueue::KERNEL_EVFILT_GRAPHICS,
 		    reinterpret_cast<void*>(static_cast<uintptr_t>(context_id)));
-		if (result == LibKernel::KERNEL_ERROR_EBADF ||
-		    result == LibKernel::KERNEL_ERROR_ENOENT) {
+		if (result == LibKernel::KERNEL_ERROR_EBADF || result == LibKernel::KERNEL_ERROR_ENOENT) {
 			DeleteEopEq(registration.eq, registration.id);
 			continue;
 		}

src/graphics/host_gpu/renderer/renderContext.h

@@ -5,13 +5,13 @@
 #include "common/assert.h"
 #include "common/common.h"
 #include "common/threads.h"
-#include "graphics/host_gpu/renderer/bufferCache.h"
+#include "graphics/host_gpu/renderer/cache/bufferCache.h"
+#include "graphics/host_gpu/renderer/cache/gpuResourceManager.h"
+#include "graphics/host_gpu/renderer/cache/samplerCache.h"
+#include "graphics/host_gpu/renderer/cache/textureCache.h"
 #include "graphics/host_gpu/renderer/commandScheduler.h"
-#include "graphics/host_gpu/renderer/descriptorCache.h"
-#include "graphics/host_gpu/renderer/gpuResourceManager.h"
-#include "graphics/host_gpu/renderer/pipelineCache.h"
-#include "graphics/host_gpu/renderer/samplerCache.h"
-#include "graphics/host_gpu/renderer/textureCache.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptorCache.h"
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
 #include "kernel/eventQueue.h"
 
 #include <memory>
@@ -32,10 +32,10 @@ public:
 	~RenderContext();
 	KYTY_CLASS_NO_COPY(RenderContext);
 
-	[[nodiscard]] GraphicContext& GetGraphics() const noexcept { return m_graphics; }
-	void                          InitializeGpu(VideoOut::VideoOutDriver* video_out);
-	void                          ShutdownGpu();
-	[[nodiscard]] Gpu&            GetGpu() const;
+	[[nodiscard]] GraphicContext&           GetGraphics() const noexcept { return m_graphics; }
+	void                                    InitializeGpu(VideoOut::VideoOutDriver* video_out);
+	void                                    ShutdownGpu();
+	[[nodiscard]] Gpu&                      GetGpu() const;
 	[[nodiscard]] VideoOut::VideoOutDriver& GetVideoOut() const;
 
 	Common::Mutex&      GetMutex() { return m_mutex; }
@@ -56,18 +56,18 @@ private:
 	struct EopEqRegistration {
 		LibKernel::EventQueue::KernelEqueue    eq = LibKernel::EventQueue::KERNEL_EQUEUE_INVALID;
 		LibKernel::EventQueue::KernelEqueueRef queue;
-		int                                   id = 0;
+		int                                    id = 0;
 	};
 
-	GraphicContext&      m_graphics;
-	Common::Mutex        m_mutex;
-	RenderExecutor       m_render_executor;
-	CommandScheduler     m_command_scheduler;
-	DescriptorCache      m_descriptor_cache;
-	PipelineCache        m_pipeline_cache;
-	SamplerCache         m_sampler_cache;
-	GpuResourceManager   m_gpu_resources;
-	std::unique_ptr<Gpu> m_gpu;
+	GraphicContext&           m_graphics;
+	Common::Mutex             m_mutex;
+	RenderExecutor            m_render_executor;
+	CommandScheduler          m_command_scheduler;
+	DescriptorCache           m_descriptor_cache;
+	PipelineCache             m_pipeline_cache;
+	SamplerCache              m_sampler_cache;
+	GpuResourceManager        m_gpu_resources;
+	std::unique_ptr<Gpu>      m_gpu;
 	VideoOut::VideoOutDriver* m_video_out = nullptr;
 
 	Common::Mutex                  m_eop_mutex;

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -16,17 +16,18 @@
 #include "graphics/host_gpu/renderer/colorRenderTarget.h"
 #include "graphics/host_gpu/renderer/debug.h"
 #include "graphics/host_gpu/renderer/depthRenderTarget.h"
-#include "graphics/host_gpu/renderer/descriptorCache.h"
-#include "graphics/host_gpu/renderer/pipelineCache.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptorCache.h"
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
+#include "graphics/host_gpu/renderer/pipeline/shaderResourceBarrier.h"
+#include "graphics/host_gpu/renderer/pipeline/shaderSubgroup.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
-#include "graphics/host_gpu/renderer/shaderResourceBarrier.h"
-#include "graphics/host_gpu/renderer/shaderSubgroup.h"
 #include "graphics/host_gpu/vulkanCommon.h"
-#include "graphics/shader/recompiler/ResourceMaterialization.h"
-#include "graphics/shader/recompiler/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/ResourceMaterialization.h"
+#include "graphics/shader/recompiler/ir/ShaderIR.h"
 #include "graphics/shader/shader.h"
 #include "kernel/eventQueue.h"
+#include "kernel/memory.h"
 #include "kernel/pthread.h"
 #include "libs/errno.h"
 
@@ -221,8 +222,7 @@ static void LogDrawTargetState(const char* draw_name, const RenderColorInfo& col
 	LogMrtState(draw_name, buffer, ps_input_info);
 }
 
-static void LogDrawInputState(const RenderCommandBuffer&   buffer,
-                               const RenderColorInfo&       color,
+static void LogDrawInputState(const RenderCommandBuffer& buffer, const RenderColorInfo& color,
                               const ShaderVertexInputInfo& vs_input_info,
                               uint32_t index_type_and_size, uint32_t index_count,
                               const void* index_addr) {
@@ -395,6 +395,10 @@ static void SetDynamicParams(const RenderCommandBuffer& buffer, vk::CommandBuffe
 		                              dynamic_params.stencil_back.reference);
 	}
 
+#if defined(__APPLE__)
+	// MoltenVK has no VK_EXT_color_write_enable; the pipeline is created without the
+	// eColorWriteEnableEXT dynamic state and relies on the static colorWriteMask instead.
+#else
 	vk::Bool32 enable[RENDER_COLOR_ATTACHMENTS_MAX] = {};
 	for (uint32_t i = 0; i < dynamic_params.color_write_count; i++) {
 		enable[i] = (dynamic_params.color_write_enable[i] ? VK_TRUE : VK_FALSE);
@@ -402,6 +406,7 @@ static void SetDynamicParams(const RenderCommandBuffer& buffer, vk::CommandBuffe
 	if (dynamic_params.color_write_count != 0) {
 		vk_buffer.setColorWriteEnableEXT(dynamic_params.color_write_count, enable);
 	}
+#endif
 }
 
 static bool DrawHasValidVertexShader(const HW::Shader& sh_ctx) {
@@ -494,9 +499,9 @@ struct DrawCallInfo {
 };
 
 RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderColorInfo* colors,
-                                             uint32_t color_count, RenderDepthInfo& depth) {
+                                                 uint32_t color_count, RenderDepthInfo& depth) {
 	EXIT_IF(colors == nullptr || color_count > RENDER_COLOR_ATTACHMENTS_MAX);
-	auto& cache = m_context.GetTextureCache();
+	auto&       cache = m_context.GetTextureCache();
 	RenderState state {};
 	state.width                 = std::numeric_limits<uint32_t>::max();
 	state.height                = std::numeric_limits<uint32_t>::max();
@@ -507,8 +512,7 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 		auto& target = colors[i];
 		EXIT_IF(!target.image_id);
 		const auto old_image = cache.ResolveOwner(target.image_id);
-		if (old_image == nullptr ||
-		    (!old_image->registered && !old_image->info.data.Empty()) ||
+		if (old_image == nullptr || (!old_image->registered && !old_image->info.data.Empty()) ||
 		    old_image->binding.needs_rebind) {
 			if (old_image != nullptr) {
 				old_image->binding = {};
@@ -517,7 +521,7 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 			BindRenderTarget(target.image_id);
 		}
 		target.image_view = cache.FindRenderTarget(target.image_id, target.desc);
-		auto& image = cache.GetImage(target.image_id);
+		auto& image       = cache.GetImage(target.image_id);
 		EXIT_IF(image.backing.samples != target.samples || target.image_view == nullptr);
 		if (attachment_samples == 0) {
 			attachment_samples = target.samples;
@@ -525,20 +529,19 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 			EXIT("mixed color attachment sample counts are unsupported: %u and %u\n",
 			     attachment_samples, target.samples);
 		}
-		const auto& view = target.desc.view_info;
-		const auto layout =
-		    image.binding.is_bound ? vk::ImageLayout::eGeneral
-		                           : vk::ImageLayout::eColorAttachmentOptimal;
+		const auto& view   = target.desc.view_info;
+		const auto  layout = image.binding.is_bound ? vk::ImageLayout::eGeneral
+		                                            : vk::ImageLayout::eColorAttachmentOptimal;
 		image.Transit(layout,
 		              vk::AccessFlagBits2::eColorAttachmentRead |
 		                  vk::AccessFlagBits2::eColorAttachmentWrite,
 		              ImageSubresourceRange {view.base_level, view.level_count, view.base_layer,
 		                                     view.layer_count},
 		              buffer.Handle());
-		state.width      = std::min(state.width, target.extent.width);
-		state.height     = std::min(state.height, target.extent.height);
-		state.num_layers = std::min(state.num_layers, view.layer_count);
-		auto& attachment = state.color_attachments[i];
+		state.width             = std::min(state.width, target.extent.width);
+		state.height            = std::min(state.height, target.extent.height);
+		state.num_layers        = std::min(state.num_layers, view.layer_count);
+		auto& attachment        = state.color_attachments[i];
 		attachment.image_view   = target.image_view;
 		attachment.image_layout = layout;
 		attachment.clear_value  = target.color_clear_value.uint32;
@@ -556,8 +559,7 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 		depth.depth_meta_clear_enable =
 		    depth.htile &&
 		    cache.IsMetaCleared(depth.htile_buffer_vaddr, depth.desc.view_info.base_layer);
-		depth.depth_load_clear_enable =
-		    depth.depth_clear_enable || depth.depth_meta_clear_enable;
+		depth.depth_load_clear_enable = depth.depth_clear_enable || depth.depth_meta_clear_enable;
 		if (depth.depth_meta_clear_enable &&
 		    !cache.TouchMeta(depth.htile_buffer_vaddr, depth.desc.view_info.base_layer, false)) {
 			EXIT("failed to consume HTile clear state\n");
@@ -567,12 +569,12 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 		if (attachment_samples == 0) {
 			attachment_samples = depth.samples;
 		} else if (attachment_samples != depth.samples) {
-			EXIT("mixed color/depth sample counts are unsupported: %u and %u\n",
-			     attachment_samples, depth.samples);
+			EXIT("mixed color/depth sample counts are unsupported: %u and %u\n", attachment_samples,
+			     depth.samples);
 		}
 		const auto layout = depth_attachment_layout(depth);
 		const auto writes = depth.AttachmentWriteAspects();
-		auto access = vk::AccessFlags2 {vk::AccessFlagBits2::eDepthStencilAttachmentRead};
+		auto       access = vk::AccessFlags2 {vk::AccessFlagBits2::eDepthStencilAttachmentRead};
 		if (writes) {
 			access |= vk::AccessFlagBits2::eDepthStencilAttachmentWrite;
 		}
@@ -581,21 +583,19 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 		              ImageSubresourceRange {view.base_level, view.level_count, view.base_layer,
 		                                     view.layer_count},
 		              buffer.Handle());
-		state.width      = std::min(state.width, depth.width);
-		state.height     = std::min(state.height, depth.height);
-		state.num_layers = std::min(state.num_layers, view.layer_count);
-		const auto aspects = ImageViewOps::DepthAspectMask(depth.format);
-		auto& attachment = state.depth_stencil_attachment;
+		state.width               = std::min(state.width, depth.width);
+		state.height              = std::min(state.height, depth.height);
+		state.num_layers          = std::min(state.num_layers, view.layer_count);
+		const auto aspects        = ImageViewOps::DepthAspectMask(depth.format);
+		auto&      attachment     = state.depth_stencil_attachment;
 		attachment.image_view     = depth.image_view;
 		attachment.image_layout   = layout;
 		attachment.clear_value[0] = std::bit_cast<uint32_t>(depth.depth_clear_value);
 		attachment.clear_value[1] = depth.stencil_clear_value;
-		attachment.has_depth =
-		    static_cast<bool>(aspects & vk::ImageAspectFlagBits::eDepth);
-		attachment.depth_clear = depth.depth_load_clear_enable;
-		attachment.has_stencil =
-		    static_cast<bool>(aspects & vk::ImageAspectFlagBits::eStencil);
-		attachment.stencil_clear = depth.stencil_clear_enable;
+		attachment.has_depth      = static_cast<bool>(aspects & vk::ImageAspectFlagBits::eDepth);
+		attachment.depth_clear    = depth.depth_load_clear_enable;
+		attachment.has_stencil    = static_cast<bool>(aspects & vk::ImageAspectFlagBits::eStencil);
+		attachment.stencil_clear  = depth.stencil_clear_enable;
 	}
 	if (attachment_samples == 0 ||
 	    vulkan_sample_count(attachment_samples) == vk::SampleCountFlagBits {}) {
@@ -680,6 +680,85 @@ static uint64_t VertexBufferDescriptorSize(const ShaderVertexInputBuffer& buffer
 	                           : buffer.num_records);
 }
 
+struct VertexBufferRange {
+	uint64_t      base_address  = 0;
+	uint64_t      requested_end = 0;
+	uint64_t      acquired_end  = 0;
+	BufferBinding binding;
+
+	[[nodiscard]] uint64_t RequestedSize() const { return requested_end - base_address; }
+};
+
+static std::vector<BufferBinding> AcquireVertexBuffers(RenderCommandBuffer&         buffer,
+                                                       const ShaderVertexInputInfo& vs_input_info) {
+	// Collect the non-empty guest vertex ranges.
+	std::vector<VertexBufferRange> ranges;
+	ranges.reserve(vs_input_info.buffers_num);
+	for (int i = 0; i < vs_input_info.buffers_num; i++) {
+		const auto& vertex = vs_input_info.buffers[i];
+		const auto  size   = VertexBufferDescriptorSize(vertex);
+		if (size == 0) {
+			continue;
+		}
+		if (vertex.addr == 0 || size > UINT64_MAX - vertex.addr) {
+			EXIT("invalid vertex buffer range: addr=0x%016" PRIx64 " size=0x%016" PRIx64 "\n",
+			     vertex.addr, size);
+		}
+		ranges.push_back({vertex.addr, vertex.addr + size});
+	}
+
+	std::ranges::sort(ranges, [](const VertexBufferRange& left, const VertexBufferRange& right) {
+		return left.base_address < right.base_address;
+	});
+
+	// Merge overlapping or touching ranges before acquiring host buffers.
+	std::vector<VertexBufferRange> merged_ranges;
+	merged_ranges.reserve(ranges.size());
+	for (const auto& range: ranges) {
+		if (!merged_ranges.empty() && merged_ranges.back().requested_end >= range.base_address) {
+			merged_ranges.back().requested_end =
+			    std::max(merged_ranges.back().requested_end, range.requested_end);
+			continue;
+		}
+		merged_ranges.push_back(range);
+	}
+
+	auto& cache = buffer.GetContext().GetBufferCache();
+	for (auto& range: merged_ranges) {
+		// PPSA20298
+		const auto size =
+		    Libs::LibKernel::Memory::ClampRangeSize(range.base_address, range.RequestedSize());
+		range.acquired_end = range.base_address + size;
+		range.binding      = cache.ObtainBuffer(buffer, range.base_address, size);
+	}
+
+	// Rebuild slot bindings, offsetting non-empty slots into their acquired merged range.
+	std::vector<BufferBinding> bindings;
+	bindings.reserve(vs_input_info.buffers_num);
+	for (int i = 0; i < vs_input_info.buffers_num; i++) {
+		const auto& vertex = vs_input_info.buffers[i];
+		const auto  size   = VertexBufferDescriptorSize(vertex);
+		if (size == 0) {
+			auto owner = cache.ObtainNullBuffer();
+			bindings.push_back({owner, owner->Handle(), 0});
+			continue;
+		}
+
+		const auto range = std::ranges::find_if(merged_ranges, [&](const VertexBufferRange& value) {
+			return vertex.addr >= value.base_address && vertex.addr < value.acquired_end;
+		});
+		if (range == merged_ranges.end()) {
+			EXIT("vertex buffer address is outside the acquired range: addr=0x%016" PRIx64 "\n",
+			     vertex.addr);
+		}
+
+		auto binding = range->binding;
+		binding.offset += vertex.addr - range->base_address;
+		bindings.push_back(std::move(binding));
+	}
+	return bindings;
+}
+
 static void SetDrawDebugPhase(RenderCommandBuffer& buffer, uint64_t submit_id,
                               const DrawCallInfo& draw, uint32_t phase) {
 	EXIT_IF(draw.name == nullptr);
@@ -731,9 +810,9 @@ static bool GetDrawTopology(const HW::UserConfig& ucfg, bool auto_draw, bool use
 }
 
 bool RenderExecutor::PrepareDrawRenderState(uint64_t submit_id, RenderCommandBuffer& buffer,
-                                        const DrawCallInfo& draw,
-                                        uint32_t render_target_slice_offset,
-                                        bool log_setup_phases, DrawRenderState& state) {
+                                            const DrawCallInfo& draw,
+                                            uint32_t            render_target_slice_offset,
+                                            bool log_setup_phases, DrawRenderState& state) {
 	EXIT_IF(draw.name == nullptr);
 	auto& ctx = buffer.GetRegisters();
 
@@ -818,37 +897,13 @@ static std::vector<BufferBinding> PrepareVertexBuffers(uint64_t
 	(void)submit_id;
 
 	LogDrawPhase(draw.name, "PrepareVertexBuffers");
-	std::vector<BufferBinding> bindings;
-	bindings.reserve(vs_input_info.buffers_num);
-	for (int i = 0; i < vs_input_info.buffers_num; i++) {
-		const auto& b    = vs_input_info.buffers[i];
-		const auto  size = VertexBufferDescriptorSize(b);
-		if (size == 0) {
-			auto owner = buffer.GetContext().GetBufferCache().ObtainNullBuffer();
-			bindings.push_back({owner, owner->Handle(), 0});
-		} else {
-			bindings.push_back(
-			    buffer.GetContext().GetBufferCache().ObtainBuffer(buffer, b.addr, size));
-		}
-	}
-	return bindings;
+	return AcquireVertexBuffers(buffer, vs_input_info);
 }
 
 static void RebindVertexBuffers(RenderCommandBuffer&         buffer,
                                 const ShaderVertexInputInfo& vs_input_info,
                                 std::vector<BufferBinding>&  bindings) {
-	EXIT_IF(bindings.size() != static_cast<size_t>(vs_input_info.buffers_num));
-	for (int i = 0; i < vs_input_info.buffers_num; i++) {
-		const auto& vertex = vs_input_info.buffers[i];
-		const auto  size   = VertexBufferDescriptorSize(vertex);
-		if (size == 0) {
-			auto owner  = buffer.GetContext().GetBufferCache().ObtainNullBuffer();
-			bindings[i] = {owner, owner->Handle(), 0};
-		} else {
-			bindings[i] =
-			    buffer.GetContext().GetBufferCache().ObtainBuffer(buffer, vertex.addr, size);
-		}
-	}
+	bindings = AcquireVertexBuffers(buffer, vs_input_info);
 }
 
 static PreparedIndexBuffer PrepareIndexBuffer(RenderCommandBuffer&         buffer,
@@ -1006,17 +1061,17 @@ static void EmitDrawPrimitives(const HW::UserConfig& ucfg, vk::CommandBuffer vk_
 }
 
 void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, RenderCommandBuffer& buffer,
-                                     const DrawCallInfo& draw, DrawRenderState& state,
-                                     vk::PrimitiveTopology topology, const DrawEmitInfo& emit,
-                                     const DrawIndexBufferSource& index_source,
-                                     bool log_pipeline_phase, bool set_bind_debug,
-                                     bool set_auto_debug) {
+                                         const DrawCallInfo& draw, DrawRenderState& state,
+                                         vk::PrimitiveTopology topology, const DrawEmitInfo& emit,
+                                         const DrawIndexBufferSource& index_source,
+                                         bool log_pipeline_phase, bool set_bind_debug,
+                                         bool set_auto_debug) {
 	EXIT_IF(draw.name == nullptr);
 	auto& ucfg = buffer.GetUserConfig();
 
 	LogDrawPhase(draw.name, "PrepareBindings");
-	auto bindings = PrepareGraphicsBindings(buffer, state.vs_input_info.stage,
-	                                        state.ps_input_info.stage, state.ps_active);
+	auto bindings        = PrepareGraphicsBindings(buffer, state.vs_input_info.stage,
+	                                               state.ps_input_info.stage, state.ps_active);
 	auto vertex_bindings = PrepareVertexBuffers(submit_id, buffer, draw, state.vs_input_info);
 	auto index_binding   = PrepareIndexBuffer(buffer, index_source);
 	RebindVertexBuffers(buffer, state.vs_input_info, vertex_bindings);
@@ -1089,10 +1144,10 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, RenderCommandBuffer
 }
 
 void RenderExecutor::DrawIndex(uint64_t submit_id, RenderCommandBuffer& buffer,
-                           uint32_t index_type_and_size, uint32_t index_count,
-                           const void* index_addr, uint32_t flags, uint32_t type,
-                           uint32_t instance_count, uint32_t render_target_slice_offset,
-                           int32_t vertex_offset_add, uint32_t first_instance) {
+                               uint32_t index_type_and_size, uint32_t index_count,
+                               const void* index_addr, uint32_t flags, uint32_t type,
+                               uint32_t instance_count, uint32_t render_target_slice_offset,
+                               int32_t vertex_offset_add, uint32_t first_instance) {
 	KYTY_PROFILER_FUNCTION();
 
 	EXIT_IF(buffer.IsInvalid());
@@ -1223,11 +1278,10 @@ void RenderExecutor::DrawIndex(uint64_t submit_id, RenderCommandBuffer& buffer,
 }
 
 // NOLINTNEXTLINE(readability-function-cognitive-complexity)
-void RenderExecutor::DrawAuto(uint64_t submit_id, RenderCommandBuffer& buffer,
-                              uint32_t index_count,
-                          uint32_t flags, uint32_t render_target_slice_offset,
-                          uint32_t instance_count, uint32_t first_vertex,
-                          uint32_t first_instance) {
+void RenderExecutor::DrawAuto(uint64_t submit_id, RenderCommandBuffer& buffer, uint32_t index_count,
+                              uint32_t flags, uint32_t render_target_slice_offset,
+                              uint32_t instance_count, uint32_t first_vertex,
+                              uint32_t first_instance) {
 	KYTY_PROFILER_FUNCTION();
 
 	EXIT_IF(buffer.IsInvalid());
@@ -1285,7 +1339,8 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, RenderCommandBuffer& buffer,
 	                         instance_count,  first_instance};
 
 	DrawRenderState state {};
-	if (!PrepareDrawRenderState(submit_id, buffer, draw, render_target_slice_offset, false, state)) {
+	if (!PrepareDrawRenderState(submit_id, buffer, draw, render_target_slice_offset, false,
+	                            state)) {
 		ResetBindings();
 		return;
 	}
@@ -1335,7 +1390,7 @@ void RenderExecutor::DrawAuto(uint64_t submit_id, RenderCommandBuffer& buffer,
 }
 
 bool RenderExecutor::ResolveColorTargets(uint64_t submit_id, RenderCommandBuffer& buffer,
-                                     uint32_t render_target_slice_offset) {
+                                         uint32_t render_target_slice_offset) {
 	const auto& hw = buffer.GetRegisters();
 	if (hw.GetColorControl().mode != 3) {
 		return false;
@@ -1364,8 +1419,7 @@ bool RenderExecutor::ResolveColorTargets(uint64_t submit_id, RenderCommandBuffer
 	cache.MarkGpuWritten(dst.image_id);
 	auto& source      = cache.GetImage(src.image_id);
 	auto& destination = cache.GetImage(dst.image_id);
-	destination.Resolve(source,
-	                    {src.base_mip_level, 1, src.base_array_layer, 1},
+	destination.Resolve(source, {src.base_mip_level, 1, src.base_array_layer, 1},
 	                    {dst.base_mip_level, 1, dst.base_array_layer, 1});
 	return true;
 }

src/graphics/host_gpu/renderer/renderTarget.h

@@ -12,13 +12,13 @@ namespace Libs::Graphics {
 static constexpr uint32_t RENDER_COLOR_ATTACHMENTS_MAX = 8;
 
 struct RenderAttachment {
-	vk::ImageView           image_view   = nullptr;
-	vk::ImageLayout         image_layout = vk::ImageLayout::eUndefined;
-	std::array<uint32_t, 4> clear_value  = {};
+	vk::ImageView           image_view    = nullptr;
+	vk::ImageLayout         image_layout  = vk::ImageLayout::eUndefined;
+	std::array<uint32_t, 4> clear_value   = {};
 	bool                    is_clear      = false;
-	bool                    has_depth    = false;
-	bool                    depth_clear  = false;
-	bool                    has_stencil  = false;
+	bool                    has_depth     = false;
+	bool                    depth_clear   = false;
+	bool                    has_stencil   = false;
 	bool                    stencil_clear = false;
 
 	bool operator==(const RenderAttachment&) const = default;

src/graphics/host_gpu/renderer/sync.cpp

@@ -5,7 +5,7 @@
 #include "common/logging/log.h"
 #include "common/threads.h"
 #include "graphics/host_gpu/graphicContext.h"
-#include "graphics/host_gpu/renderer/bufferCache.h"
+#include "graphics/host_gpu/renderer/cache/bufferCache.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/presentation/videoOut.h"
@@ -251,9 +251,9 @@ uint64_t PrepareVideoOutFlip(CommandBuffer& buffer, int handle, int index, int f
                              int64_t flip_arg) {
 	for (;;) {
 		uint64_t   request_id = 0;
-		auto&      video_out = buffer.GetContext().GetVideoOut();
-		const auto result = video_out.SubmitFlipFromGpu(
-		    buffer, handle, index, flip_mode, flip_arg, request_id);
+		auto&      video_out  = buffer.GetContext().GetVideoOut();
+		const auto result =
+		    video_out.SubmitFlipFromGpu(buffer, handle, index, flip_mode, flip_arg, request_id);
 		if (result == OK) {
 			EXIT_IF(request_id == 0);
 			return request_id;

src/graphics/host_gpu/vma.cpp

@@ -122,9 +122,9 @@ uint64_t GraphicContext::GetDeviceMemoryUsage() const {
 	    physical_device_properties.deviceType == vk::PhysicalDeviceType::eDiscreteGpu;
 	uint64_t usage = 0;
 	for (uint32_t heap = 0; heap < physical_device_memory_properties.memoryHeapCount; heap++) {
-		const bool device_local = static_cast<bool>(
-		    physical_device_memory_properties.memoryHeaps[heap].flags &
-		    vk::MemoryHeapFlagBits::eDeviceLocal);
+		const bool device_local =
+		    static_cast<bool>(physical_device_memory_properties.memoryHeaps[heap].flags &
+		                      vk::MemoryHeapFlagBits::eDeviceLocal);
 		if (!discrete || device_local) {
 			usage += budgets[heap].usage;
 		}
@@ -144,7 +144,7 @@ uint64_t GraphicContext::GetTotalMemoryBudget() const {
 	uint64_t local  = 0;
 	uint64_t usage  = 0;
 	for (uint32_t heap = 0; heap < physical_device_memory_properties.memoryHeapCount; heap++) {
-		const auto& properties   = physical_device_memory_properties.memoryHeaps[heap];
+		const auto& properties = physical_device_memory_properties.memoryHeaps[heap];
 		const bool  device_local =
 		    static_cast<bool>(properties.flags & vk::MemoryHeapFlagBits::eDeviceLocal);
 		if (device_local) {
@@ -159,9 +159,8 @@ uint64_t GraphicContext::GetTotalMemoryBudget() const {
 		return budget - std::min<uint64_t>(budget / 8, 1024ull * 1024 * 1024);
 	}
 	constexpr uint64_t system_reserve = 8ull * 1024 * 1024 * 1024;
-	const auto available = budget > usage ? budget - usage : uint64_t {0};
-	return std::max(local,
-	                available > system_reserve ? available - system_reserve : uint64_t {0});
+	const auto         available      = budget > usage ? budget - usage : uint64_t {0};
+	return std::max(local, available > system_reserve ? available - system_reserve : uint64_t {0});
 }
 
 void GraphicContext::CreateBuffer(uint64_t size, VulkanBuffer& buffer) {

src/graphics/host_gpu/vulkanCommon.cpp

@@ -37,6 +37,7 @@ constexpr FormatMapping kFormatMappings[] = {
     {Prospero::BufferFormat::k16_16Float, vk::Format::eR16G16Sfloat},
     {Prospero::BufferFormat::k11_11_10Float, vk::Format::eB10G11R11UfloatPack32},
     {Prospero::BufferFormat::k10_10_10_2UNorm, vk::Format::eA2B10G10R10UnormPack32},
+    {Prospero::BufferFormat::k10_10_10_2UInt, vk::Format::eA2B10G10R10UintPack32},
     {Prospero::BufferFormat::k8_8_8_8UNorm, vk::Format::eR8G8B8A8Unorm},
     {Prospero::BufferFormat::k8_8_8_8SNorm, vk::Format::eR8G8B8A8Snorm},
     {Prospero::BufferFormat::k8_8_8_8UInt, vk::Format::eR8G8B8A8Uint},
@@ -55,6 +56,10 @@ constexpr FormatMapping kFormatMappings[] = {
     {Prospero::BufferFormat::k32_32_32_32UInt, vk::Format::eR32G32B32A32Uint},
     {Prospero::BufferFormat::k32_32_32_32SInt, vk::Format::eR32G32B32A32Sint},
     {Prospero::BufferFormat::k32_32_32_32Float, vk::Format::eR32G32B32A32Sfloat},
+    // Narrow-channel sRGB formats are optional in Vulkan. Keep a same-width fallback until
+    // sampler-aware sRGB emulation is available.
+    {Prospero::BufferFormat::k8Srgb, vk::Format::eR8Unorm},
+    {Prospero::BufferFormat::k8_8Srgb, vk::Format::eR8G8Unorm},
     {Prospero::BufferFormat::k8_8_8_8Srgb, vk::Format::eR8G8B8A8Srgb},
     {Prospero::BufferFormat::k9_9_9_5Float, vk::Format::eE5B9G9R9UfloatPack32},
     {Prospero::BufferFormat::k5_6_5UNorm, vk::Format::eB5G6R5UnormPack16},

src/graphics/presentation/presenter.h

@@ -20,14 +20,14 @@ public:
 	~Presenter();
 	KYTY_CLASS_NO_COPY(Presenter);
 
-	[[nodiscard]] Frame& PrepareFrame(CommandBuffer& command, const ImageInfo& info);
-	[[nodiscard]] Frame& PrepareBlankFrame(uint32_t width, uint32_t height, bool opaque,
-	                                       CommandBuffer* producer = nullptr);
-	[[nodiscard]] Frame* PrepareLastFrame();
-	[[nodiscard]] bool   IsGuestPaused() const noexcept;
+	[[nodiscard]] Frame&         PrepareFrame(CommandBuffer& command, const ImageInfo& info);
+	[[nodiscard]] Frame&         PrepareBlankFrame(uint32_t width, uint32_t height, bool opaque,
+	                                               CommandBuffer* producer = nullptr);
+	[[nodiscard]] Frame*         PrepareLastFrame();
+	[[nodiscard]] bool           IsGuestPaused() const noexcept;
 	[[nodiscard]] RenderContext& Renderer() const noexcept;
-	void                 Present(Frame& frame, bool reuse = false);
-	void                 Discard(Frame& frame);
+	void                         Present(Frame& frame, bool reuse = false);
+	void                         Discard(Frame& frame);
 
 private:
 	struct Impl;

src/graphics/presentation/renderDoc.cpp

@@ -19,12 +19,16 @@
 #include <windows.h>
 #undef min
 #undef max
+#else
+#include <dlfcn.h>
+
+// RenderDoc uses Windows-style names in its cross-platform API.
+#define __cdecl
+using HMODULE = void*;
 #endif
 
 namespace Libs::Graphics {
 
-#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
-
 using RenderDocDevicePointer = void*;
 using RenderDocWindowHandle  = void*;
 
@@ -107,6 +111,8 @@ static RenderDocDevicePointer GetRenderDocDevicePointer(vk::Instance instance) {
 	return VulkanHandleToPointer(instance);
 }
 
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+
 static bool BindRenderDocApi(HMODULE module) {
 	auto* get_api = reinterpret_cast<pRENDERDOC_GetAPI>(GetProcAddress(module, "RENDERDOC_GetAPI"));
 	if (get_api == nullptr) {
@@ -147,10 +153,84 @@ static RenderDocWindowHandle GetRenderDocWindowHandle(SDL_Window* window) {
 	return info.info.win.window;
 }
 
+#else
+
+static bool BindRenderDocApi(HMODULE module) {
+	auto* get_api = reinterpret_cast<pRENDERDOC_GetAPI>(::dlsym(module, "RENDERDOC_GetAPI"));
+	if (get_api == nullptr) {
+		return false;
+	}
+
+	void* api = nullptr;
+	if (get_api(eRENDERDOC_API_Version_1_4_2, &api) == 0 || api == nullptr) {
+		return false;
+	}
+
+	g_module = module;
+	g_api    = static_cast<RenderDocApi*>(api);
+
+	g_api->SetCaptureFilePathTemplate("_RenderDoc/kyty");
+	static RenderDocInputButton capture_keys[] = {eRENDERDOC_Key_F1};
+	g_api->SetCaptureKeys(capture_keys, 1);
+	g_api->UnloadCrashHandler();
+
+	Dl_info info {};
+	if (::dladdr(reinterpret_cast<void*>(get_api), &info) != 0 && info.dli_fname != nullptr) {
+		LOGF("RenderDoc: bound API from %s\n", info.dli_fname);
+	} else {
+		LOGF("RenderDoc: bound API\n");
+	}
+	return true;
+}
+
+static RenderDocWindowHandle GetRenderDocWindowHandle(SDL_Window* window) {
+	if (window == nullptr) {
+		return nullptr;
+	}
+
+	SDL_SysWMinfo info {};
+	SDL_VERSION(&info.version);
+
+	if (SDL_GetWindowWMInfo(window, &info) != SDL_TRUE) {
+		return nullptr;
+	}
+
+#if defined(SDL_VIDEO_DRIVER_X11)
+	if (info.subsystem == SDL_SYSWM_X11) {
+		// RenderDoc takes the raw xlib Window id in the pointer slot, not a Display*.
+		return reinterpret_cast<RenderDocWindowHandle>(
+		    static_cast<uintptr_t>(info.info.x11.window));
+	}
+#endif
+
+	// Wayland capture works without an active-window handle.
+	static std::atomic_bool logged = false;
+	if (!logged.exchange(true)) {
+		LOGF("RenderDoc: no native window handle for SDL subsystem %d (Wayland?); the in-app "
+		     "overlay is unavailable, but --rd captures still work\n",
+		     static_cast<int>(info.subsystem));
+	}
+	return nullptr;
+}
+
+#endif
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+
 static bool IsAvailable() {
 	return g_api != nullptr && g_device != nullptr && g_window != nullptr;
 }
 
+#else
+
+static bool IsAvailable() {
+	return g_api != nullptr && g_device != nullptr;
+}
+
+#endif
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+
 void RenderDocInit() {
 	bool expected = false;
 	if (!g_init_done.compare_exchange_strong(expected, true)) {
@@ -187,6 +267,33 @@ void RenderDocInit() {
 	}
 }
 
+#else
+
+void RenderDocInit() {
+	bool expected = false;
+	if (!g_init_done.compare_exchange_strong(expected, true)) {
+		return;
+	}
+
+	// Prefer an injected RenderDoc instance.
+	auto* module = ::dlopen("librenderdoc.so", RTLD_NOW | RTLD_NOLOAD);
+	if (module == nullptr) {
+		module = ::dlopen("librenderdoc.so", RTLD_NOW);
+	}
+	if (module == nullptr) {
+		LOGF("RenderDoc: librenderdoc.so was not found; in-app capture disabled\n");
+		return;
+	}
+
+	if (!BindRenderDocApi(module)) {
+		LOGF("RenderDoc: API 1.4.2 is not available; in-app capture disabled\n");
+		::dlclose(module);
+		return;
+	}
+}
+
+#endif
+
 void RenderDocSetActiveWindow(vk::Instance instance, SDL_Window* window) {
 	if (g_api == nullptr) {
 		return;
@@ -274,13 +381,4 @@ void RenderDocOnPresent() {
 	}
 }
 
-#else
-
-void RenderDocInit() {}
-void RenderDocSetActiveWindow(vk::Instance /*instance*/, SDL_Window* /*window*/) {}
-void RenderDocRequestCapture() {}
-void RenderDocOnPresent() {}
-
-#endif
-
 } // namespace Libs::Graphics