Emulate SHA-NI and fix SSE4a EXTRQ/INSERTQ register form (#126)

* Emulate SHA-NI on illegal instruction faults

* Fix SSE4a EXTRQ/INSERTQ register form

* Fix SHA-NI memory operand emulation

* Revert "Fix SSE4a EXTRQ/INSERTQ register form"

This reverts commit ea2b54a4d0.

---------

Co-authored-by: neobugs1 <neobugs1@users.noreply.github.com>
Co-authored-by: nmzik <Nmzik@mail.ru>

.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
 
@@ -69,45 +90,302 @@ jobs:
         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="$QT_ROOT_DIR"
+
+      - name: Build
+        shell: bash
+        run: |
+          cmake --build _Build/macos --target launcher --parallel
+
+      - name: Install
+        shell: bash
+        run: |
+          cmake --install _Build/macos --prefix _Build/macos/install
+
+      - name: Bundle MoltenVK
+        shell: bash
+        env:
+          MOLTENVK_VERSION: v1.4.2
+          MOLTENVK_SHA256: f95765a6229cb7b915990a2890ce12ebe36a730b021545d3d52ae69ce4c4024e
+        run: |
+          archive="$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 \
+            --parallel
+
+      - name: Test
+        shell: bash
+        run: |
+          ctest --test-dir _Build/linux --output-on-failure \
+            -R '^(page_manager|memory_tracker)$'
+
+      - name: Install
+        shell: bash
+        run: |
+          cmake --install _Build/linux --prefix _Build/linux/install
+
+      - name: Verify artifacts
+        shell: bash
+        run: |
+          file _Build/linux/install/launcher
+          file _Build/linux/install/kyty_emulator
+          file _Build/linux/install/kyty_emulator | grep -q "ELF 64-bit LSB .*x86-64"
+          ldd _Build/linux/install/kyty_emulator > /dev/null
+          readelf -d _Build/linux/install/launcher | grep -q 'RPATH.*\$ORIGIN/lib'
+          while IFS= read -r -d '' binary; do
+            while read -r dependency; do
+              test -e "_Build/linux/install/lib/$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/

3rdparty/CMakeLists.txt

@@ -44,7 +44,17 @@ set(SPDLOG_NO_EXCEPTIONS ON CACHE BOOL "" FORCE)
 add_subdirectory(spdlog EXCLUDE_FROM_ALL)
 
 if (NOT TARGET FFmpeg::ffmpeg)
-	if(CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm64|ARM64)$")
+	# On macOS the target arch is driven by CMAKE_OSX_ARCHITECTURES, not the host
+	# CMAKE_SYSTEM_PROCESSOR (which reports arm64 even for an x86_64/Rosetta build).
+	if(APPLE AND CMAKE_OSX_ARCHITECTURES)
+		if(CMAKE_OSX_ARCHITECTURES MATCHES "^(x86_64)$")
+			set(ARCHITECTURE x86_64)
+		elseif(CMAKE_OSX_ARCHITECTURES MATCHES "^(arm64)$")
+			set(ARCHITECTURE arm64)
+		else()
+			set(ARCHITECTURE x86_64 arm64) # universal
+		endif()
+	elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm64|ARM64)$")
 		set(ARCHITECTURE arm64)
 	else()
 		set(ARCHITECTURE x86_64)

3rdparty/cpuinfo/CMakeLists.txt

@@ -5,7 +5,14 @@ set(cpuinfo_src
 	deps/clog/src/clog.c
 )
 
-if(CMAKE_SYSTEM_PROCESSOR MATCHES "^(i[3-6]86|AMD64|x86(_64)?)$")
+# On Apple the target arch is driven by CMAKE_OSX_ARCHITECTURES, not the host
+# CMAKE_SYSTEM_PROCESSOR (which is arm64 even for an x86_64/Rosetta build).
+set(CPUINFO_TARGET_PROC "${CMAKE_SYSTEM_PROCESSOR}")
+if(APPLE AND CMAKE_OSX_ARCHITECTURES)
+	list(GET CMAKE_OSX_ARCHITECTURES 0 CPUINFO_TARGET_PROC)
+endif()
+
+if(CPUINFO_TARGET_PROC MATCHES "^(i[3-6]86|AMD64|x86(_64)?)$")
 	list(APPEND cpuinfo_src
 		src/x86/init.c
 		src/x86/info.c
@@ -19,11 +26,13 @@ if(CMAKE_SYSTEM_PROCESSOR MATCHES "^(i[3-6]86|AMD64|x86(_64)?)$")
 		src/x86/cache/deterministic.c)
 	if(LINUX OR ANDROID)
 		list(APPEND cpuinfo_src src/x86/linux/init.c src/x86/linux/cpuinfo.c)
+	elseif(APPLE)
+		list(APPEND cpuinfo_src src/x86/mach/init.c src/mach/topology.c)
 	else()
 		list(APPEND cpuinfo_src src/x86/windows/init.c)
 	endif()
-	
-elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(armv[5-8].*|aarch64|arm64)$")
+
+elseif(CPUINFO_TARGET_PROC MATCHES "^(armv[5-8].*|aarch64|arm64)$")
 	list(APPEND cpuinfo_src src/arm/uarch.c src/arm/cache.c)
 	
 	if(LINUX OR ANDROID)

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 +24,11 @@ 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. Compatibility with the same games on Windows and macOS has not yet
+been tested.
 
 ## Bugs and Issues
 
@@ -45,7 +50,7 @@ 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 +58,28 @@ 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 +114,11 @@ the Vulkan/SPIR-V validation rules.
 
 ### System requirements
 
-- Windows 10 version 1803
+- Windows 10 version 1803, or a current Linux distribution
 - A 64-bit x86 processor
 - A Vulkan 1.3-capable GPU with current drivers
 
-### Build requirements
+### Build requirements (Windows)
 
 - Git
 - CMake 3.12 or newer
@@ -134,11 +152,48 @@ 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.
+
 ### 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 +215,10 @@ 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 +229,11 @@ 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"
+```
+
+Run `kyty_emulator --help` to see the available graphics, logging, validation, profiling, and
 debugging options.
 
 ### AI Use

src/CMakeLists.txt

@@ -10,8 +10,8 @@ if(CMAKE_SYSTEM_NAME MATCHES ".*Linux")
 	set(LINUX TRUE)
 endif()
 
-if (NOT (WIN32 OR LINUX))
-    message(FATAL_ERROR "only Windows and Linux builds are supported")
+if (NOT (WIN32 OR LINUX OR APPLE))
+    message(FATAL_ERROR "only Windows, Linux, and macOS builds are supported")
 endif()
 
 set(CMAKE_CXX_STANDARD 20)
@@ -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)
@@ -45,8 +47,9 @@ else()
 	set(KYTY_BUILD KYTY_BUILD_RELEASE)
 endif()
 
-if(LINUX)
-	set(KYTY_PLATFORM KYTY_PLATFORM_LINUX)	
+if(LINUX OR APPLE)
+	# macOS rides the POSIX/Linux code paths until it gets a dedicated platform
+	set(KYTY_PLATFORM KYTY_PLATFORM_LINUX)
 else()
 	set(KYTY_PLATFORM KYTY_PLATFORM_WINDOWS)
 endif()
@@ -63,6 +66,8 @@ if(CMAKE_CXX_COMPILER_ID MATCHES "^(Apple)?Clang$")
 	if(WIN32 AND KYTY_CXX_COMPILER_NAME STREQUAL "clang-cl")
 		set(KYTY_CLANG_CL TRUE)
 		set(KYTY_LINKER LLD_LINK)
+	elseif(APPLE)
+		set(KYTY_LINKER LD64) # Apple's default linker; lld flags don't apply
 	else()
 		set(KYTY_LINKER LLD)
 	endif()
@@ -126,7 +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
@@ -153,16 +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
@@ -246,6 +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
@@ -286,8 +319,8 @@ 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/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})
@@ -319,9 +352,9 @@ 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})
@@ -330,20 +363,20 @@ 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})
@@ -389,6 +422,14 @@ 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 image_page_table COMMAND $<TARGET_FILE:image_page_table_tests>)
 	add_test(NAME memory_tracker COMMAND $<TARGET_FILE:memory_tracker_tests>)
@@ -405,23 +446,25 @@ 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()
 endif()
 
@@ -441,6 +484,8 @@ 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}")
@@ -461,6 +506,18 @@ endif()
 
 add_dependencies(kyty_emulator KytyGitVersion)
 
+if(APPLE)
+	# The emulator writes x86-64 trampolines/PLT stubs into memory and executes them, and
+	# runs guest code under Rosetta. That requires the JIT / unsigned-executable-memory
+	# entitlements, so re-sign the binary after every link (an unsigned/relinked binary
+	# reverts to the hardened defaults and aborts when it first executes written code).
+	add_custom_command(TARGET kyty_emulator POST_BUILD
+		COMMAND codesign -s - --force --options runtime
+			--entitlements "${CMAKE_CURRENT_SOURCE_DIR}/macos_jit.entitlements"
+			$<TARGET_FILE:kyty_emulator>
+		COMMENT "Codesign kyty_emulator with JIT entitlements (macOS)")
+endif()
+
 install(TARGETS kyty_emulator DESTINATION .)
 if(KYTY_BUILD_LAUNCHER)
 	add_subdirectory(launcher)

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 +312,48 @@ 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);
+
+	// 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/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 +8,57 @@
 
 #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 +75,33 @@ 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 +115,6 @@ 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 +188,11 @@ 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);
@@ -136,7 +173,7 @@ sys_file_t* SysFileCreate(const std::filesystem::path& file_name) {
 }
 
 sys_file_t* SysFileOpenR(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;
 
 	ret->type = SYS_FILE_FILE;
@@ -150,6 +187,8 @@ 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;
@@ -180,7 +219,7 @@ sys_file_t* SysFileCreate() {
 }
 
 sys_file_t* SysFileOpenW(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);
@@ -194,13 +233,15 @@ 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 +255,8 @@ 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 +282,17 @@ 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 +309,18 @@ 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 +441,7 @@ 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 +460,7 @@ 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 +480,36 @@ 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 +615,110 @@ 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);
 
@@ -83,14 +154,13 @@ 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;
+		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++) {
@@ -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
 		}
 	}
 
@@ -154,7 +250,7 @@ uint64_t SysVirtualAllocAligned(uint64_t address, uint64_t size, VirtualMemory::
 	}
 
 	pthread_mutex_lock(&g_virtual_mutex);
-	(*g_allocs)[ret_addr] = size;
+	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++) {
@@ -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);
@@ -218,7 +336,7 @@ 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;
+		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++) {
@@ -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

@@ -7,6 +7,7 @@
 #include <atomic>
 #include <chrono>             // IWYU pragma: keep
 #include <condition_variable> // IWYU pragma: keep
+#include <cerrno>
 #include <mutex>
 #include <vector>
 
@@ -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/emulator.cpp

@@ -159,13 +159,14 @@ 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);
 }
@@ -207,7 +208,7 @@ 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/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 |
@@ -2321,14 +2334,18 @@ 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 "

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/host_gpu/memoryTracker.h

@@ -38,10 +38,51 @@ public:
 	                                    bool downloaded) noexcept;
 	[[nodiscard]] bool InvalidateRegion(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) {
+				               const auto changed = manager->ChangeState<DirtySource::Cpu, true>(
+				                   manager->GetCpuAddr() + offset, bytes);
+				               manager->ApplyProtection(changed, false);
+			               });
+			return false;
+		};
+
+		if (!update_cpu_state()) {
+			return;
+		}
+		std::forward<Flush>(on_flush)();
+		if (update_cpu_state()) {
+			EXIT("memory invalidation retained GPU-owned pages\n");
+		}
+	}
 	[[nodiscard]] bool InvalidateVirtualGpuWrite(PageFaultAccess access, uint64_t vaddr,
 	                                             uint64_t size, PageFaultPhase phase) noexcept;
-	void ValidateGpuDirtyPages(const RangeSet& dirty, uint64_t vaddr, uint64_t size,
-	                           const char* operation) const noexcept;
+	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);
 

src/graphics/host_gpu/pageManager.cpp

@@ -2,6 +2,7 @@
 
 #include "graphics/host_gpu/regionDefinitions.h"
 
+#include <algorithm>
 #include <array>
 #include <atomic>
 #include <cstdarg>
@@ -19,6 +20,20 @@
 #include <windows.h>
 #undef min
 #undef max
+#elif defined(__APPLE__)
+#include <mach/mach.h>
+#include <mach/mach_vm.h>
+#include <pthread.h>
+#include <sys/mman.h>
+#include <unistd.h>
+#else
+#include <cerrno>
+#include <cstring>
+#include <execinfo.h>
+#include <fcntl.h>
+#include <sys/mman.h>
+#include <sys/syscall.h>
+#include <unistd.h>
 #endif
 
 namespace Libs::Graphics {
@@ -28,16 +43,58 @@ 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 (on
+// Windows they come from <windows.h> and are what VirtualQuery returns). Mirror the
+// canonical Win32 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;
+
+#if defined(__APPLE__)
+// Map the tracker's Win32-style protection tags to POSIX mprotect flags.
+static int PageProtToPosix(uint32_t protection) {
+	switch (protection) {
+		case PAGE_NOACCESS: return PROT_NONE;
+		case PAGE_READONLY: return PROT_READ;
+		case PAGE_READWRITE: return PROT_READ | PROT_WRITE;
+		default: return PROT_NONE;
+	}
+}
+
+// Query the current protection of the page containing vaddr via the Mach VM map and
+// collapse it to the tracker's read/write tags (execute is irrelevant to write tracking).
+static uint32_t MachQueryPageProt(uint64_t vaddr) {
+	auto                           region_addr = static_cast<mach_vm_address_t>(vaddr);
+	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 || region_addr > vaddr) {
+		return PAGE_NOACCESS; // no region covering vaddr
+	}
+	if ((info.protection & VM_PROT_WRITE) != 0) {
+		return PAGE_READWRITE;
+	}
+	if ((info.protection & VM_PROT_READ) != 0) {
+		return PAGE_READONLY;
+	}
+	return PAGE_NOACCESS;
+}
+#elif defined(__linux__)
+// Zero is the unknown protection sentinel.
+constexpr uint32_t UNKNOWN_PROTECTION = 0;
 #endif
 
 thread_local bool g_in_fault_resolution = false;
@@ -56,6 +113,10 @@ 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
@@ -78,11 +139,146 @@ thread_local bool g_in_fault_resolution = false;
 uint32_t CurrentThread() noexcept {
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
 	return GetCurrentThreadId();
+#elif defined(__APPLE__)
+	return static_cast<uint32_t>(pthread_mach_thread_np(pthread_self()));
+#elif defined(__linux__)
+	static thread_local const uint32_t tid = [] {
+		const auto raw = static_cast<uint32_t>(::syscall(SYS_gettid));
+		if (raw == 0) {
+			FailFast("gettid returned the reserved zero owner token");
+		}
+		return raw;
+	}();
+	return tid;
 #else
-	FailFast();
+	FailFast("page tracking thread identity is unsupported on this platform");
 #endif
 }
 
+#if defined(__linux__)
+int ToHostProtection(uint32_t protection) {
+	switch (protection) {
+		case NO_ACCESS_PROTECTION: return PROT_NONE;
+		case READ_ONLY_PROTECTION: return PROT_READ;
+		case READ_WRITE_PROTECTION: return PROT_READ | PROT_WRITE;
+		default: Fatal("unmappable protection 0x%08" PRIx32, protection);
+	}
+}
+
+struct HostMapping {
+	uint64_t end        = 0;
+	uint32_t protection = UNKNOWN_PROTECTION;
+};
+
+// Async-signal-safe lookup in the address-ordered /proc/self/maps.
+HostMapping QueryHostMapping(uint64_t vaddr) noexcept {
+	int fd = ::open("/proc/self/maps", O_RDONLY | O_CLOEXEC); // NOLINT
+	if (fd < 0) {
+		return {};
+	}
+
+	enum class Field { Start, End, Perms, Rest };
+
+	HostMapping result {};
+	auto        field      = Field::Start;
+	uint64_t    start      = 0;
+	uint64_t    end        = 0;
+	char        perms[4]   = {};
+	uint32_t    perms_len  = 0;
+	bool        line_valid = true;
+
+	char buffer[8192];
+
+	for (bool done = false; !done;) {
+		const auto got = ::read(fd, buffer, sizeof(buffer));
+		if (got < 0) {
+			if (errno == EINTR) {
+				continue;
+			}
+			break;
+		}
+		if (got == 0) {
+			break;
+		}
+
+		for (ssize_t i = 0; i < got && !done; i++) {
+			const char c = buffer[i];
+
+			if (c == '\n') {
+				field      = Field::Start;
+				start      = 0;
+				end        = 0;
+				perms_len  = 0;
+				line_valid = true;
+				continue;
+			}
+
+			if (!line_valid) {
+				continue;
+			}
+
+			switch (field) {
+				case Field::Start:
+				case Field::End: {
+					uint64_t digit = 0;
+					if (c >= '0' && c <= '9') {
+						digit = static_cast<uint64_t>(c - '0');
+					} else if (c >= 'a' && c <= 'f') {
+						digit = static_cast<uint64_t>(c - 'a') + 10;
+					} else if (c == '-' && field == Field::Start) {
+						field = Field::End;
+						break;
+					} else if (c == ' ' && field == Field::End) {
+						field     = Field::Perms;
+						perms_len = 0;
+						break;
+					} else {
+						line_valid = false;
+						break;
+					}
+
+					auto& value = (field == Field::Start ? start : end);
+					value       = (value << 4u) | digit;
+					break;
+				}
+
+				case Field::Perms: {
+					if (c != ' ') {
+						if (perms_len < sizeof(perms)) {
+							perms[perms_len] = c;
+						}
+						perms_len++;
+						break;
+					}
+
+					if (vaddr < start) {
+						done = true;
+					} else if (vaddr < end && perms_len >= 2) {
+						result.end        = end;
+						result.protection = perms[1] == 'w'   ? READ_WRITE_PROTECTION
+						                    : perms[0] == 'r' ? READ_ONLY_PROTECTION
+						                                      : NO_ACCESS_PROTECTION;
+						done              = true;
+					} else {
+						field = Field::Rest;
+					}
+					break;
+				}
+
+				case Field::Rest: break;
+			}
+		}
+	}
+
+	::close(fd);
+	return result;
+}
+
+uint32_t QueryHostProtection(uint64_t vaddr) noexcept {
+	return QueryHostMapping(vaddr).protection;
+}
+#endif
+
 class SpinGuard final {
 public:
 	explicit SpinGuard(std::atomic_flag& lock): m_lock(lock) {
@@ -116,24 +312,48 @@ 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;
+		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;
+#if defined(__linux__)
+		// Shadow the protection applied through Protect().
+		uint32_t current_protection = UNKNOWN_PROTECTION;
+#endif
+		bool resolving            = false;
+		bool resolving_read_write = false;
+		bool late_read_pending    = false;
+		bool late_write_pending   = false;
 	};
 
 	struct Region {
 		std::array<PageState, REGION_PAGES> pages;
 	};
 
+	class PageRangeGuard final {
+	public:
+		explicit PageRangeGuard(std::span<PageState*> pages): m_pages(pages) {
+			for (auto* page: m_pages) {
+				while (page->lock.test_and_set(std::memory_order_acquire)) {
+					std::atomic_signal_fence(std::memory_order_seq_cst);
+				}
+			}
+		}
+		~PageRangeGuard() {
+			for (auto it = m_pages.rbegin(); it != m_pages.rend(); ++it) {
+				(*it)->lock.clear(std::memory_order_release);
+			}
+		}
+		KYTY_CLASS_NO_COPY(PageRangeGuard);
+
+	private:
+		std::span<PageState*> m_pages;
+	};
+
 	Impl(PageFaultHandler handler, void* context): fault_handler(handler), fault_context(context) {
 		if (fault_handler == nullptr) {
 			Fatal("null fault handler");
@@ -145,8 +365,16 @@ 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++) {
@@ -207,31 +435,61 @@ struct PageManager::Impl {
 		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) &&
+		if ((old_protection == NO_ACCESS_PROTECTION || old_protection == READ_ONLY_PROTECTION) &&
 		    new_protection == READ_WRITE_PROTECTION) {
 			page.late_write_pending = true;
 		}
 	}
 
-	static uint32_t QueryProtection(uint64_t vaddr) {
+	static void ValidateInitialProtection(std::span<PageState*> pages, uint64_t vaddr) {
+		const auto end = vaddr + pages.size() * PAGE_SIZE;
 #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));
+		for (auto address = vaddr; address < end;) {
+			MEMORY_BASIC_INFORMATION info {};
+			if (VirtualQuery(reinterpret_cast<const void*>(static_cast<uintptr_t>(address)), &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 ")",
+				      address, static_cast<uint32_t>(info.State),
+				      static_cast<uint32_t>(info.Protect));
+			}
+			const auto region_end = reinterpret_cast<uint64_t>(info.BaseAddress) + info.RegionSize;
+			if (region_end <= address) {
+				Fatal("VirtualQuery returned an invalid region at 0x%016" PRIx64, address);
+			}
+			address = std::min(end, region_end);
+		}
+#elif defined(__APPLE__)
+		for (auto address = vaddr; address < end; address += PAGE_SIZE) {
+			const uint32_t protection = MachQueryPageProt(address);
+			if (protection != PAGE_READWRITE) {
+				Fatal("basic path requires PAGE_READWRITE at 0x%016" PRIx64
+				      " (protection=0x%08" PRIx32 ")",
+				      address, protection);
+			}
 		}
-		return info.Protect;
 #else
-		(void)vaddr;
-		Fatal("page query is unsupported on this platform");
+		for (auto address = vaddr; address < end;) {
+			const auto mapping = QueryHostMapping(address);
+			if (mapping.protection != READ_WRITE_PROTECTION || mapping.end <= address) {
+				Fatal("basic path requires a read/write mapping at 0x%016" PRIx64
+				      " (protection=0x%08" PRIx32 ")",
+				      address, mapping.protection);
+			}
+			address = std::min(end, mapping.end);
+		}
+		for (auto* page: pages) {
+			page->current_protection = READ_WRITE_PROTECTION;
+		}
 #endif
+		for (auto* page: pages) {
+			page->original_protection = READ_WRITE_PROTECTION;
+		}
 	}
 
-	static bool AllowsAccess(uint64_t vaddr, PageFaultAccess access) noexcept {
+	static bool AllowsAccess([[maybe_unused]] const PageState& page, 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,
@@ -245,34 +503,149 @@ struct PageManager::Impl {
 			case PageFaultAccess::Write: return info.Protect == PAGE_READWRITE;
 			default: return false;
 		}
+#elif defined(__APPLE__)
+		const uint32_t protection = MachQueryPageProt(vaddr);
+		switch (access) {
+			case PageFaultAccess::Read:
+				return protection == PAGE_READONLY || protection == PAGE_READWRITE;
+			case PageFaultAccess::Write: return protection == PAGE_READWRITE;
+			default: return false;
+		}
 #else
-		(void)vaddr;
-		return false;
+		const auto permitted = [](uint32_t protection, PageFaultAccess wanted) {
+			switch (wanted) {
+				case PageFaultAccess::Read:
+					return protection == READ_ONLY_PROTECTION ||
+					       protection == READ_WRITE_PROTECTION;
+				case PageFaultAccess::Write: return protection == READ_WRITE_PROTECTION;
+				default: return false;
+			}
+		};
+
+		if (!permitted(page.current_protection, access)) {
+			return false;
+		}
+		return permitted(QueryHostProtection(vaddr), access);
 #endif
 	}
 
-	static void Protect(uint64_t vaddr, uint32_t protection, uint32_t expected_old,
-	                    bool fault_path) noexcept {
+	static void ProtectRange(std::span<PageState*> pages, uint64_t vaddr, uint32_t protection,
+	                         std::span<const uint32_t> expected_old, bool fault_path) noexcept {
+		const auto size = pages.size() * PAGE_SIZE;
+		if (pages.size() != expected_old.size()) {
+			FailFast("protection range state size mismatch");
+		}
 #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");
+		struct HostRange {
+			uint64_t begin = 0;
+			uint64_t end   = 0;
+		};
+		std::vector<HostRange> host_ranges;
+		const auto             end = vaddr + size;
+		for (auto address = vaddr; address < end;) {
+			MEMORY_BASIC_INFORMATION info {};
+			if (VirtualQuery(reinterpret_cast<const void*>(static_cast<uintptr_t>(address)), &info,
+			                 sizeof(info)) == 0 ||
+			    info.State != MEM_COMMIT) {
+				if (fault_path) {
+					FailFast("VirtualProtect fault transition did not match expected protection");
+				}
+				Fatal("invalid protection transition at 0x%016" PRIx64 ", state=0x%08" PRIx32
+				      ", new=0x%08" PRIx32,
+				      address, static_cast<uint32_t>(info.State), protection);
 			}
-			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);
+			const auto region_end = reinterpret_cast<uint64_t>(info.BaseAddress) + info.RegionSize;
+			const auto query_end  = std::min(end, region_end);
+			if (query_end <= address) {
+				if (fault_path) {
+					FailFast("VirtualQuery returned an invalid fault transition region");
+				}
+				Fatal("VirtualQuery returned an invalid region at 0x%016" PRIx64, address);
+			}
+			const auto first_page = static_cast<size_t>((address - vaddr) / PAGE_SIZE);
+			const auto last_page =
+			    static_cast<size_t>((query_end - vaddr + PAGE_SIZE - 1) / PAGE_SIZE);
+			for (auto page = first_page; page < last_page; page++) {
+				if (info.Protect != expected_old[page]) {
+					if (fault_path) {
+						FailFast(
+						    "VirtualProtect fault transition did not match expected protection");
+					}
+					Fatal("invalid protection transition at 0x%016" PRIx64 ", actual=0x%08" PRIx32
+					      ", expected=0x%08" PRIx32 ", new=0x%08" PRIx32,
+					      vaddr + page * PAGE_SIZE, static_cast<uint32_t>(info.Protect),
+					      expected_old[page], protection);
+				}
+			}
+			const auto allocation = reinterpret_cast<uint64_t>(info.AllocationBase);
+			if (host_ranges.empty() || allocation != host_ranges.back().begin) {
+				host_ranges.push_back({allocation, query_end});
+			} else {
+				host_ranges.back().end = query_end;
+			}
+			address = query_end;
+		}
+		for (auto range: host_ranges) {
+			range.begin               = std::max(range.begin, vaddr);
+			DWORD      old_protection = 0;
+			const auto first_page     = static_cast<size_t>((range.begin - vaddr) / PAGE_SIZE);
+			if (VirtualProtect(reinterpret_cast<void*>(static_cast<uintptr_t>(range.begin)),
+			                   range.end - range.begin, protection, &old_protection) == 0 ||
+			    old_protection != expected_old[first_page]) {
+				if (fault_path) {
+					FailFast("VirtualProtect fault transition did not match expected protection");
+				}
+				Fatal("invalid protection transition at 0x%016" PRIx64 ", old=0x%08" PRIx32
+				      ", expected=0x%08" PRIx32 ", new=0x%08" PRIx32,
+				      range.begin, static_cast<uint32_t>(old_protection), expected_old[first_page],
+				      protection);
+			}
+		}
+#elif defined(__APPLE__)
+		// mprotect cannot report the previous protection, so the expected_old comparison
+		// is dropped; the tracker is the sole mutator of these pages and drives the
+		// transition from its own shadow state.
+		(void)expected_old;
+		if (mprotect(reinterpret_cast<void*>(static_cast<uintptr_t>(vaddr)), size,
+		             PageProtToPosix(protection)) != 0) {
+			if (fault_path) {
+				FailFast("mprotect fault transition failed");
+			}
+			Fatal("mprotect failed at 0x%016" PRIx64 ", new=0x%08" PRIx32, vaddr, protection);
 		}
 #else
-		(void)vaddr;
-		(void)protection;
-		(void)fault_path;
-		FailFast("page protection is unsupported on this platform");
+		for (size_t i = 0; i < pages.size(); i++) {
+			const auto actual = pages[i]->current_protection;
+			if (actual != UNKNOWN_PROTECTION && actual != expected_old[i]) {
+				if (fault_path) {
+					FailFast("mprotect fault transition did not match expected protection");
+				}
+				Fatal("invalid protection transition at 0x%016" PRIx64 ", old=0x%08" PRIx32
+				      ", expected=0x%08" PRIx32 ", new=0x%08" PRIx32,
+				      vaddr + i * PAGE_SIZE, actual, expected_old[i], protection);
+			}
+		}
+		if (::mprotect(reinterpret_cast<void*>(static_cast<uintptr_t>(vaddr)), size,
+		               ToHostProtection(protection)) != 0) {
+			if (fault_path) {
+				FailFast("mprotect failed on the fault path");
+			}
+			Fatal("mprotect failed at 0x%016" PRIx64 ", new=0x%08" PRIx32 " (%s)", vaddr,
+			      protection, std::strerror(errno));
+		}
+		for (auto* page: pages) {
+			page->current_protection = protection;
+		}
 #endif
 	}
 
+	static void Protect(PageState& page, uint64_t vaddr, uint32_t protection, uint32_t expected_old,
+	                    bool fault_path) noexcept {
+		PageState* pages[]    = {&page};
+		uint32_t   expected[] = {expected_old};
+		ProtectRange(pages, vaddr, protection, expected, fault_path);
+	}
+
 	std::unique_ptr<std::atomic<Region*>[]> regions;
 	std::vector<std::unique_ptr<Region>>    region_storage;
 	std::mutex                              region_mutex;
@@ -326,26 +699,6 @@ bool PageManager::IsMapped(uint64_t vaddr, uint64_t size) const noexcept {
 	return true;
 }
 
-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;
-}
-
 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");
@@ -376,65 +729,134 @@ void PageManager::UpdatePageWatchers(bool track, uint64_t vaddr, uint64_t size,
 	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);
+	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);
+		auto*      region =
+		    track ? m_impl->GetOrCreateRegion(chunk_begin) : m_impl->FindRegion(chunk_begin);
 		if (region == nullptr) {
-			Fatal("untracking unknown page 0x%016" PRIx64, page_vaddr);
+			Fatal("untracking unknown page 0x%016" PRIx64, chunk_begin);
 		}
-		auto&     page = m_impl->GetPage(*region, page_vaddr);
-		SpinGuard lock(page.lock);
-		if (page.resolving && track) {
-			FailFast("new page watcher raced active fault resolution");
+
+		const auto page_count = static_cast<size_t>((chunk_end - chunk_begin) / PAGE_SIZE);
+		std::vector<Impl::PageState*> pages;
+		pages.reserve(page_count);
+		for (auto address = chunk_begin; address < chunk_end; address += PAGE_SIZE) {
+			pages.push_back(&m_impl->GetPage(*region, address));
 		}
-		if (page.mappings == 0) {
-			Fatal("watching unmapped page 0x%016" PRIx64, page_vaddr);
+		Impl::PageRangeGuard lock(pages);
+
+		std::vector<uint8_t> first_watchers(page_count);
+		for (size_t i = 0; i < page_count; i++) {
+			auto&      page    = *pages[i];
+			const auto address = chunk_begin + i * PAGE_SIZE;
+			if (page.resolving && track) {
+				FailFast("new page watcher raced active fault resolution");
+			}
+			if (page.mappings == 0) {
+				Fatal("watching unmapped page 0x%016" PRIx64, address);
+			}
+			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, address);
+				}
+				first_watchers[i] = page.write_watchers == 0 && page.access_watchers == 0;
+			} else {
+				if (watchers == 0) {
+					Fatal("watcher underflow at 0x%016" PRIx64, address);
+				}
+				if (page.backing_writer != 0 && page.backing_writer != CurrentThread()) {
+					Fatal("backing write ownership changed at 0x%016" PRIx64, address);
+				}
+			}
 		}
-		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;
+			for (size_t first = 0; first < page_count;) {
+				while (first < page_count && first_watchers[first] == 0) {
+					first++;
+				}
+				auto last = first;
+				while (last < page_count && first_watchers[last] != 0) {
+					last++;
+				}
+				if (first != last) {
+					Impl::ValidateInitialProtection(std::span {pages}.subspan(first, last - first),
+					                                chunk_begin + first * PAGE_SIZE);
+				}
+				first = last;
 			}
 		}
+
+		std::vector<uint32_t> old_protections(page_count);
+		std::vector<uint32_t> new_protections(page_count);
+		std::vector<uint8_t>  transitions(page_count);
+		for (size_t i = 0; i < page_count; i++) {
+			auto& page = *pages[i];
+			auto& watchers =
+			    (mode == PageWatchMode::ReadWrite ? page.access_watchers : page.write_watchers);
+			const auto old_protection = Impl::WatcherProtection(page);
+			if (track) {
+				watchers++;
+			} else {
+				watchers--;
+			}
+			const auto new_protection = Impl::WatcherProtection(page);
+			old_protections[i]        = old_protection;
+			new_protections[i]        = new_protection;
+			if (new_protection != old_protection && (track || page.backing_writer == 0)) {
+				transitions[i] = 1;
+			}
+		}
+
+		for (size_t first = 0; first < page_count;) {
+			while (first < page_count && transitions[first] == 0) {
+				first++;
+			}
+			if (first == page_count) {
+				break;
+			}
+			const auto protection = new_protections[first];
+			auto       current    = first + 1;
+			auto       last       = current;
+			for (; current < page_count && new_protections[current] == protection; current++) {
+				if (old_protections[current] != new_protections[current] &&
+				    transitions[current] == 0) {
+					break;
+				}
+				if (transitions[current] != 0) {
+					last = current + 1;
+				}
+			}
+			Impl::ProtectRange(std::span {pages}.subspan(first, last - first),
+			                   chunk_begin + first * PAGE_SIZE, protection,
+			                   std::span {old_protections}.subspan(first, last - first), false);
+			first = current;
+		}
+
+		for (size_t i = 0; i < page_count; i++) {
+			auto&      page       = *pages[i];
+			const auto protection = new_protections[i];
+			if (track) {
+				switch (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 (page.backing_writer == 0) {
+				Impl::PublishDelayedFaults(page, old_protections[i], protection);
+				if (page.write_watchers == 0 && page.access_watchers == 0) {
+					page.original_protection = 0;
+				}
+			}
+		}
+		chunk_begin = chunk_end;
 	}
 }
 
@@ -459,6 +881,12 @@ void PageManager::OnGpuMap(uint64_t vaddr, uint64_t size, GpuAccess access) {
 		page.mappings++;
 		page.gpu_read_mappings += gpu_read ? 1u : 0u;
 		page.gpu_write_mappings += gpu_write ? 1u : 0u;
+#if defined(__linux__)
+		// New guest mappings start read/write.
+		if (page.current_protection == UNKNOWN_PROTECTION) {
+			page.current_protection = READ_WRITE_PROTECTION;
+		}
+#endif
 	}
 }
 
@@ -580,7 +1008,7 @@ void PageManager::EndBackingWrite(uint64_t vaddr, uint64_t size) noexcept {
 		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::Protect(page, address, new_protection, old_protection, false);
 		}
 		Impl::PublishDelayedFaults(page, old_protection, new_protection);
 		if (page.write_watchers == 0 && page.access_watchers == 0) {
@@ -605,12 +1033,12 @@ bool PageManager::HandleFault(PageFaultAccess access, uint64_t fault_vaddr) noex
 	while (true) {
 		SpinGuard lock(page.lock);
 		if (access == PageFaultAccess::Read && page.late_read_pending &&
-		    Impl::AllowsAccess(fault_vaddr, access)) {
+		    Impl::AllowsAccess(page, fault_vaddr, access)) {
 			page.late_read_pending = false;
 			return true;
 		}
 		if (access == PageFaultAccess::Write && page.late_write_pending &&
-		    Impl::AllowsAccess(fault_vaddr, access)) {
+		    Impl::AllowsAccess(page, fault_vaddr, access)) {
 			page.late_write_pending = false;
 			return true;
 		}
@@ -632,7 +1060,7 @@ bool PageManager::HandleFault(PageFaultAccess access, uint64_t fault_vaddr) noex
 			}
 			bool&      pending = (access == PageFaultAccess::Read ? page.late_read_pending
 			                                                      : page.late_write_pending);
-			const bool allowed = Impl::AllowsAccess(fault_vaddr, access);
+			const bool allowed = Impl::AllowsAccess(page, fault_vaddr, access);
 			pending            = false;
 			if (waited && !allowed) {
 				FailFast("page remained inaccessible after waiting for its resolver");
@@ -681,12 +1109,12 @@ bool PageManager::HandleFault(PageFaultAccess access, uint64_t fault_vaddr) noex
 				page.write_watchers = 0;
 			}
 			const auto restored_protection = Impl::WatcherProtection(page);
-			Impl::Protect(PageStart(fault_vaddr), restored_protection, old_protection, true);
+			Impl::Protect(page, 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)) {
+		} else if (!Impl::AllowsAccess(page, fault_vaddr, access)) {
 			FailFast("fault completion left the page inaccessible");
 		}
 		page.resolving            = false;
@@ -702,22 +1130,4 @@ bool PageManager::HandleFault(PageFaultAccess access, uint64_t fault_vaddr) noex
 	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;
-}
-
 } // namespace Libs::Graphics

src/graphics/host_gpu/pageManager.h

@@ -41,7 +41,6 @@ public:
 	[[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,
@@ -50,9 +49,8 @@ public:
 	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);
+	ReserveBackingWrites(std::span<const RangeSet::Range> ranges);
 
 private:
 	void BeginBackingWrite(uint64_t vaddr, uint64_t size) noexcept;

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/regionManager.h

@@ -16,6 +16,11 @@
 #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 {
@@ -49,6 +54,12 @@ 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

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>
@@ -183,9 +183,8 @@ 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");
 	}
@@ -215,40 +214,38 @@ 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();
+	auto       downloads = RecordDownloads(copies);
+	const auto tick      = m_scheduler.CurrentTick();
 	BeginBackingPublication(retire.address, retire.size, tick);
-	m_scheduler.DeferOperation(
-	    [this, downloads = std::move(downloads), retire = std::move(retire), tick]() 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);
-			    }
-			    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);
-	    });
+	m_scheduler.DeferOperation([this, downloads = std::move(downloads), retire = std::move(retire),
+	                            tick]() 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);
+			}
+			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);
+	});
 }
 
 BufferCache::BufferCache(GraphicContext& graphics, CommandScheduler& scheduler,
@@ -301,14 +298,13 @@ BufferCache::~BufferCache() {
 bool BufferCache::SynchronizeBacking(uint64_t vaddr, uint64_t size) {
 	bool waited = false;
 	for (;;) {
-		uint64_t tick = 0;
+		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);
+		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;
+			bool               changed = true;
 			while (changed) {
 				changed = false;
 				for (const auto& [address, cached]: m_buffers) {
@@ -384,6 +380,73 @@ BufferBinding BufferCache::UploadTransient(const void* data, uint64_t size, uint
 	return {owner, owner->Handle(), 0};
 }
 
+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");
+	}
+	(void)SynchronizeBacking(vaddr, size);
+	if (!HasPageOverlap(vaddr, size)) {
+		return;
+	}
+	m_memory_tracker.InvalidateRegion(vaddr, size,
+	                                  [this, vaddr, size] { ReadMemory(vaddr, size); });
+}
+
+void BufferCache::ReadMemory(uint64_t vaddr, uint64_t size) {
+	(void)SynchronizeBacking(vaddr, size);
+	std::vector<DownloadCopy> copies;
+	{
+		FaultSafeCacheLock lock(this, m_mutex);
+		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;
+				    }
+			    }
+		    });
+	}
+	if (copies.empty()) {
+		return;
+	}
+	auto downloads = RecordDownloads(copies);
+	m_scheduler.FinishCurrent();
+	PublishDownloads(downloads);
+	{
+		FaultSafeCacheLock lock(this, m_mutex);
+		m_memory_tracker.ForEachDownloadRange<true>(
+		    vaddr, size,
+		    [&](uint64_t address, uint64_t bytes) noexcept {
+			    m_memory_tracker.ValidateGpuDirtyPages(m_gpu_modified_ranges, address, bytes,
+			                                           "memory invalidation completion");
+		    },
+		    [](uint64_t, uint64_t) noexcept {});
+		for (const auto& range: downloads) {
+			m_gpu_modified_ranges.Subtract(range.address, range.size);
+		}
+	}
+}
+
 bool BufferCache::InvalidateMemory(PageFaultAccess access, uint64_t vaddr, uint64_t size,
                                    PageFaultPhase phase) noexcept {
 	const auto page = vaddr & ~(TRACKER_PAGE_SIZE - 1);
@@ -544,8 +607,8 @@ void BufferCache::UnmapMemory(uint64_t vaddr, uint64_t size) {
 			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");
+				    m_memory_tracker.ValidateGpuDirtyPages(m_gpu_modified_ranges, address,
+				                                           download_size, "unmap retirement");
 			    },
 			    [](uint64_t, uint64_t) noexcept {});
 		}
@@ -722,12 +785,11 @@ 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 invalidated = !m_image_invalidated_ranges.Intersections(vaddr, size).empty();
+		const bool requested_gpu_owned = !dirty.empty();
 		m_memory_tracker.ValidateGpuDirtyOwnership(m_gpu_modified_ranges, vaddr, size,
 		                                           "image source");
 
@@ -807,8 +869,8 @@ ImageBufferSource BufferCache::ObtainBufferForImage(uint64_t vaddr, uint64_t siz
 	}
 
 	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 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()) {
@@ -831,9 +893,8 @@ 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 - stage_address,
+			        cached.buffer->Offset(address), upload_size, vk::AccessFlagBits::eHostWrite);
 		    }
 	    });
 	DiscardGpuDirtyBytesLocked(vaddr, size, "staged image source transfer");
@@ -917,9 +978,8 @@ std::pair<std::shared_ptr<Buffer>, uint64_t> BufferCache::ObtainBufferForImageWr
 	    [&]() 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 - stage_address,
+			        cached.buffer->Offset(address), upload_size, vk::AccessFlagBits::eHostWrite);
 		    }
 	    });
 	return {cached.buffer, cached.buffer->Offset(vaddr)};
@@ -946,12 +1006,12 @@ void BufferCache::FillBuffer(uint64_t vaddr, uint64_t size, uint32_t value, bool
 		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));
@@ -992,7 +1052,7 @@ void BufferCache::CopyBuffer(uint64_t dst_vaddr, uint64_t src_vaddr, uint64_t si
 		if (src_memory) {
 			(void)SynchronizeBacking(src_vaddr, size);
 		}
-		const auto      src_region =
+		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 {};
@@ -1004,7 +1064,7 @@ void BufferCache::CopyBuffer(uint64_t dst_vaddr, uint64_t src_vaddr, uint64_t si
 		    !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;) {
@@ -1037,10 +1097,10 @@ 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");
 	}
@@ -1107,7 +1167,7 @@ void BufferCache::BeginBackingPublication(uint64_t vaddr, uint64_t size, uint64_
 
 void BufferCache::CompleteBackingPublication(uint64_t vaddr, uint64_t size, uint64_t tick) {
 	std::lock_guard lock(m_publication_mutex);
-	const auto publication =
+	const auto      publication =
 	    std::ranges::find_if(m_pending_backing_publications, [&](const auto& pending) {
 		    return pending.address == vaddr && pending.size == size && pending.tick == tick;
 	    });
@@ -1170,7 +1230,7 @@ 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 +1250,8 @@ 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 +1265,13 @@ 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});
 				    }
 			    });
 		}

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

@@ -6,7 +6,7 @@
 #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>
@@ -49,6 +49,8 @@ public:
 
 	[[nodiscard]] bool InvalidateMemory(PageFaultAccess access, uint64_t vaddr, uint64_t size,
 	                                    PageFaultPhase phase) noexcept;
+	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,
@@ -71,8 +73,8 @@ public:
 	[[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);
+	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;
@@ -91,23 +93,21 @@ private:
 	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]] 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,

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

@@ -1,4 +1,4 @@
-#include "graphics/host_gpu/renderer/gpuResourceManager.h"
+#include "graphics/host_gpu/renderer/cache/gpuResourceManager.h"
 
 #include "common/assert.h"
 #include "graphics/guest_gpu/command_processor/commandProcessor.h"
@@ -36,7 +36,8 @@ bool GpuResourceManager::InvalidateMemory(PageFaultAccess access, uint64_t vaddr
 }
 
 bool GpuResourceManager::HandleFault(PageFaultAccess access, uint64_t fault_vaddr) noexcept {
-	if (!m_page_manager.IsMapped(fault_vaddr, 1)) {
+	constexpr uint64_t fault_size = 8;
+	if (!IsMapped(fault_vaddr, fault_size)) {
 		return false;
 	}
 	if (CommandScheduler::InDeferredOperation()) {
@@ -47,10 +48,15 @@ bool GpuResourceManager::HandleFault(PageFaultAccess access, uint64_t fault_vadd
 	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);
+			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();
 	};
@@ -68,47 +74,52 @@ bool GpuResourceManager::HandleFault(PageFaultAccess access, uint64_t fault_vadd
 	return handled;
 }
 
-void GpuResourceManager::PrepareHostWrite(uint64_t vaddr, uint64_t size) {
-	if (!m_page_manager.HasAnyMapping(vaddr, size)) {
-		return;
+bool GpuResourceManager::InvalidateMemory(uint64_t vaddr, uint64_t size) {
+	if (!IsMapped(vaddr, size)) {
+		return false;
 	}
 	if (CommandScheduler::InDeferredOperation()) {
-		EXIT("unsupported host write from an asynchronous GPU completion, addr=0x%016" PRIx64
-		     " size=0x%016" PRIx64 "\n",
+		EXIT("unsupported memory invalidation 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) {
+	const auto resolve = [this, vaddr, size](CommandProcessor& cp) {
 		cp.BeginReadbackTransaction();
 		{
 			ResourceMutex::FaultScope fault(m_resource_mutex);
-			handle_range();
+			m_buffer_cache.InvalidateMemory(vaddr, size);
+			m_texture_cache.InvalidateMemory(vaddr, size);
 		}
 		cp.EndReadbackTransaction();
 	};
 	if (auto* cp = Gpu::CurrentCommandProcessor(); cp != nullptr) {
 		resolve(*cp);
-		return;
+		return true;
 	}
 	if (m_resource_mutex.IsOwnedByCurrentThread()) {
-		EXIT("unsupported host write from a pre-owned resource transaction, addr=0x%016" PRIx64
-		     " size=0x%016" PRIx64 "\n",
+		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 {
-	return m_page_manager.IsMapped(vaddr, size);
+	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, GpuAccess access) {
+	{
+		std::lock_guard lock(m_mapped_ranges_mutex);
+		m_mapped_ranges.Add(vaddr, size);
+	}
 	m_page_manager.OnGpuMap(vaddr, size, access);
 }
 
@@ -120,6 +131,8 @@ void GpuResourceManager::UnmapMemory(uint64_t vaddr, uint64_t size, GpuAccess ac
 		m_texture_cache.UnmapMemory(vaddr, size);
 		m_buffer_cache.UnmapMemory(vaddr, size);
 		m_page_manager.OnGpuUnmap(vaddr, size, access);
+		std::lock_guard lock(m_mapped_ranges_mutex);
+		m_mapped_ranges.Subtract(vaddr, size);
 	};
 	if (m_gpu == nullptr) {
 		if (m_resource_mutex.IsOwnedByCurrentThread()) {

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 {
 
@@ -26,7 +27,7 @@ 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);
@@ -38,11 +39,13 @@ private:
 	[[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;
+	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/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"

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"

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
@@ -903,39 +1028,29 @@ 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_gpu_owned = false;
+	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_gpu_owned |= source.gpu_owned;
+		data_imported = true;
+		UploadImage(image, desc, *source.buffer, source.offset);
+	}
 	if (data_imported) {
 		image.ClearBufferModified();
 	}
@@ -945,39 +1060,27 @@ void TextureCache::InitializeImage(ImageId id, const ImageDesc& desc) {
 	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);
@@ -1029,7 +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)) {
@@ -1079,20 +1182,19 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
 				}
 				replacement_buffer = resolved.IsBufferModified();
 				DeleteImage(result);
-				result          = {};
-				replacing_image = true;
+				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)) {
 				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 +1203,7 @@ 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 +1352,7 @@ 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);
 }
 
@@ -1265,13 +1366,10 @@ void TextureCache::CommitGpuWrite(Image& image) {
 	}
 	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,
@@ -1335,8 +1433,7 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
 	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 +1458,12 @@ 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) {
@@ -1567,18 +1662,15 @@ bool TextureCache::SynchronizeImageToBuffer(ImageId id) {
 	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) {
+	const auto range = image.info.data;
+	if (image.IsCpuDirty()) {
 		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)) {
+	if (image.IsDefinitelyCpuDirty() || image.IsBufferModified()) {
 		EXIT("TextureCache: image mirror source is not native-current\n");
 	}
 	auto [destination, offset] =
@@ -1591,7 +1683,7 @@ bool TextureCache::SynchronizeImageToBuffer(ImageId id) {
 	m_buffer_cache.PublishImageBuffer(range.address, range.size);
 	image.MarkBufferModified();
 	RetainImage(m_scheduler.Current(), id);
-	ReleaseGpuTracking(id);
+	ClearGpuModified(id);
 	return true;
 }
 
@@ -1633,7 +1725,7 @@ 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 +1752,40 @@ 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) {
@@ -1755,27 +1837,25 @@ bool TextureCache::InvalidateMemory(PageFaultAccess access, uint64_t address, ui
 		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);
+		CacheLock  lock(*this, m_lock);
+		const bool tracked =
+		    std::ranges::any_of(FindImagesInRegion(address, size, true), [&](ImageId id) {
+			    const auto owner = ResolveOwner(id);
+			    return owner != nullptr && !owner->depth_id && owner->IsTracked();
+		    });
+		if (tracked) {
 			InvalidateCpuAliases(address, size);
 		}
-		return action != CpuFaultAction::Untracked;
+		return tracked;
 	}
-
-	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) {
+	if (phase != PageFaultPhase::Complete && phase != PageFaultPhase::Release) {
 		return false;
 	}
-	(void)m_memory_tracker.InvalidateVirtualGpuWrite(access, address, size, phase);
-	return true;
+	CacheLock lock(*this, m_lock);
+	return std::ranges::any_of(FindImagesInRegion(address, size, true), [&](ImageId id) {
+		const auto owner = ResolveOwner(id);
+		return owner != nullptr && !owner->depth_id;
+	});
 }
 
 void TextureCache::UnmapMemory(uint64_t address, uint64_t size) {
@@ -1794,16 +1874,10 @@ 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 +1923,7 @@ 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>
@@ -64,7 +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);
+	void               InvalidateMemory(uint64_t address, uint64_t size);
 	[[nodiscard]] bool SynchronizeImageToBuffer(uint64_t address, uint64_t size);
 	[[nodiscard]] bool InvalidateMemoryFromGPU(uint64_t address, uint64_t size,
 	                                           bool formatted_buffer_write = false);
@@ -109,11 +110,17 @@ 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,
@@ -148,8 +155,7 @@ 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);
@@ -160,7 +166,7 @@ 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;

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/image/textureCommon.h"
 #include "graphics/host_gpu/renderer/debug.h"
-#include "graphics/host_gpu/renderer/descriptorCache.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"

src/graphics/host_gpu/renderer/colorRenderTarget.h

@@ -3,7 +3,7 @@
 
 #include "graphics/guest_gpu/gpu_defs.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
-#include "graphics/host_gpu/renderer/textureCache.h"
+#include "graphics/host_gpu/renderer/cache/textureCache.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 
 #include <cstdint>

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/pipeline/descriptorCache.h"
+#include "graphics/host_gpu/renderer/image/imageView.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/vma.h"

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/image/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/pipeline/descriptorCache.h"
+#include "graphics/host_gpu/renderer/image/imageView.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/vulkanCommon.h"

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/image/imageView.h"
 #include "graphics/host_gpu/renderer/renderTarget.h"
-#include "graphics/host_gpu/renderer/textureCache.h"
+#include "graphics/host_gpu/renderer/cache/textureCache.h"
 #include "graphics/host_gpu/vulkanCommon.h"
 
 #include <cstdint>

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>

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>
@@ -99,9 +99,9 @@ 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;
@@ -130,25 +130,25 @@ 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 +159,10 @@ 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 +191,7 @@ 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 +200,8 @@ 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 +217,9 @@ 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 +232,15 @@ 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 +253,7 @@ 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 +261,7 @@ 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 +272,15 @@ 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 +294,11 @@ 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 +307,10 @@ 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 +319,11 @@ 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 +333,7 @@ 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 +341,7 @@ 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 +358,30 @@ 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 +389,40 @@ 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 +436,21 @@ 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 +465,66 @@ 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 +534,9 @@ 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 +551,13 @@ 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 +569,7 @@ 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 +582,9 @@ 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 +655,11 @@ 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 +709,9 @@ 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/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>
 

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 {
@@ -103,10 +103,7 @@ 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 +119,10 @@ 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>
@@ -102,6 +102,10 @@ 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,

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"
@@ -51,4 +51,4 @@ bool TextureBuildGpuTileInfos(uint64_t size,
 
 } // 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"
@@ -15,8 +15,8 @@
 #include "gpu_tiler_shaders/gpu_tiler_swap_bgra16_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/streamBuffer.h"
+#include "graphics/host_gpu/renderer/image/image.h"
+#include "graphics/host_gpu/renderer/cache/streamBuffer.h"
 
 #include <algorithm>
 #include <array>

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>
 

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"
 

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>
@@ -239,11 +239,11 @@ bool IsSupportedDepthTextureEncoding(const ShaderTextureResource& descriptor, co
 	const uint32_t     field3_expected =
 	    (descriptor.Type() << 28u) | field3_common | descriptor.DstSelXYZW();
 	const uint32_t field4_expected = descriptor.Depth() | (descriptor.BaseArray5() << 16u);
-	const bool common = (descriptor.fields[1] & field1_reserved_mask) == 0 &&
-	                    (descriptor.fields[2] & field2_reserved_mask) == 0 &&
-	                    descriptor.fields[3] == field3_expected &&
-	                    descriptor.fields[4] == field4_expected &&
-	                    descriptor.fields[5] == field5_expected;
+	const bool     common          = (descriptor.fields[1] & field1_reserved_mask) == 0 &&
+	                                 (descriptor.fields[2] & field2_reserved_mask) == 0 &&
+	                                 descriptor.fields[3] == field3_expected &&
+	                                 descriptor.fields[4] == field4_expected &&
+	                                 descriptor.fields[5] == field5_expected;
 	if (!common || (descriptor.fields[6] == 0 && descriptor.fields[7] != 0)) {
 		return false;
 	}
@@ -318,8 +318,8 @@ 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 +333,20 @@ 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 +377,10 @@ 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 format_ok =
+	    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;
 	}
@@ -616,12 +620,12 @@ RenderExecutor::ResolveTexture(const ShaderRecompiler::IR::ImageResource&   reso
 		                                             resource.written);
 	}
 
-	const auto              pixel_format = TextureGetFormat(format);
+	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              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>

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"
 

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/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/renderer/pipeline/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>
@@ -729,7 +729,13 @@ void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descripto
 
 	vk::PipelineRasterizationStateCreateInfo rasterizer {};
 	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;
@@ -807,7 +813,13 @@ void CreatePipelineInternal(GraphicContext& graphics, DescriptorCache& descripto
 
 	vk::PipelineColorBlendStateCreateInfo color_blending {};
 	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;
@@ -873,7 +885,11 @@ 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 +909,9 @@ 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]));

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/pipeline/descriptorCache.h"
+#include "graphics/host_gpu/renderer/pipeline/descriptors.h"
+#include "graphics/host_gpu/renderer/image/imageInfo.h"
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.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/renderer/pipeline/shaderResourceBarrier.h"
+#include "graphics/host_gpu/renderer/pipeline/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.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/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/cache/gpuResourceManager.h"
+#include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"
+#include "graphics/host_gpu/renderer/cache/samplerCache.h"
+#include "graphics/host_gpu/renderer/cache/textureCache.h"
 #include "kernel/eventQueue.h"
 
 #include <memory>

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/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"

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

src/graphics/presentation/videoOut.cpp

@@ -12,7 +12,7 @@
 #include "graphics/guest_gpu/gpu_defs.h"
 #include "graphics/guest_gpu/graphicsRun.h"
 #include "graphics/guest_gpu/tile.h"
-#include "graphics/host_gpu/renderer/imageInfo.h"
+#include "graphics/host_gpu/renderer/image/imageInfo.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/presentation/presenter.h"

src/graphics/presentation/window/swapchain.cpp

@@ -586,7 +586,13 @@ void Swapchain::RefreshSurfaceSize() {
 void Swapchain::Recreate(bool surface_lost) {
 	Destroy();
 	if (surface_lost) {
+#if defined(__APPLE__)
+		// Surface recreation goes through SDL_Vulkan_CreateSurface, which touches the
+		// window's view/layer and must run on the main thread on macOS.
+		m_window.RunOnMainThread([this] { m_window.RecreateSurface(); }, true);
+#else
 		m_window.RecreateSurface();
+#endif
 	}
 	RefreshSurfaceSize();
 	Create();
@@ -797,9 +803,20 @@ void Presenter::Present(Frame& frame, bool reuse) {
 
 	auto& window = m_impl->window;
 	if (window.window_hidden) {
+#if defined(__APPLE__)
+		// AppKit traps if a window is shown off the main thread; marshal and wait so the
+		// swapchain below is recreated against a visible window.
+		window.RunOnMainThread(
+		    [&window] {
+			    window.UpdateIcon();
+			    SDL_ShowWindow(window.window);
+		    },
+		    true);
+#else
 		window.UpdateIcon();
 
 		SDL_ShowWindow(window.window);
+#endif
 
 		window.window_hidden = false;
 		m_impl->RecoverSwapchain(Swapchain::Status::Recreate);

src/graphics/presentation/window/vulkanWindow.cpp

@@ -217,7 +217,9 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 
 		if (color_write_ext.colorWriteEnable != VK_TRUE) {
 			LOGF("colorWriteEnable is not supported\n");
+#if !defined(__APPLE__)
 			skip_device = true;
+#endif
 		}
 
 		if (depth_clip_control.depthClipControl != VK_TRUE) {
@@ -226,7 +228,9 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		}
 		if (depth_clip_enable.depthClipEnable != VK_TRUE) {
 			LOGF("depthClipEnable is not supported\n");
+#if !defined(__APPLE__)
 			skip_device = true;
+#endif
 		}
 
 		if (features12.samplerMirrorClampToEdge != VK_TRUE) {
@@ -271,7 +275,9 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 		}
 		if (device_features2.features.depthBounds != VK_TRUE) {
 			LOGF("depthBounds is not supported\n");
+#if !defined(__APPLE__)
 			skip_device = true;
+#endif
 		}
 		if (device_features2.features.shaderStorageImageWriteWithoutFormat != VK_TRUE) {
 			LOGF("shaderStorageImageWriteWithoutFormat is not supported\n");
@@ -494,7 +500,14 @@ static vk::Device VulkanCreateDevice(vk::PhysicalDevice physical_device, const V
 
 	vk::PhysicalDeviceDepthClipControlFeaturesEXT depth_clip_control {};
 	depth_clip_control.sType = vk::StructureType::ePhysicalDeviceDepthClipControlFeaturesEXT;
+	// MoltenVK lacks VK_EXT_depth_clip_enable and VK_EXT_color_write_enable, so drop those
+	// feature structs from the chain on macOS (the renderer falls back to default depth
+	// clipping and static color-write masks).
+#if defined(__APPLE__)
+	depth_clip_control.pNext = nullptr;
+#else
 	depth_clip_control.pNext = &depth_clip_enable;
+#endif
 	depth_clip_control.depthClipControl = VK_TRUE;
 
 	vk::PhysicalDeviceVulkan12Features features12 {};
@@ -541,7 +554,9 @@ static vk::Device VulkanCreateDevice(vk::PhysicalDevice physical_device, const V
 	device_features.fragmentStoresAndAtomics             = VK_TRUE;
 	device_features.samplerAnisotropy                    = VK_TRUE;
 	device_features.robustBufferAccess                   = VK_TRUE;
-	device_features.depthBounds                          = VK_TRUE;
+#if !defined(__APPLE__)
+	device_features.depthBounds                          = VK_TRUE; // unsupported by MoltenVK
+#endif
 	device_features.shaderStorageImageWriteWithoutFormat = VK_TRUE;
 	device_features.shaderStorageImageReadWithoutFormat  = VK_TRUE;
 	device_features.shaderImageGatherExtended            = VK_TRUE;
@@ -894,10 +909,20 @@ void WindowContext::CreateVulkan() {
 	}
 	surface = native_surface;
 
-	std::vector<const char*> device_extensions = {
-	    VK_KHR_SWAPCHAIN_EXTENSION_NAME, VK_EXT_DEPTH_CLIP_ENABLE_EXTENSION_NAME,
-	    VK_EXT_DEPTH_CLIP_CONTROL_EXTENSION_NAME, VK_EXT_COLOR_WRITE_ENABLE_EXTENSION_NAME,
-	    VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME, "VK_KHR_maintenance1"};
+	std::vector<const char*> device_extensions = {VK_KHR_SWAPCHAIN_EXTENSION_NAME,
+	                                              VK_EXT_DEPTH_CLIP_CONTROL_EXTENSION_NAME,
+	                                              VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME,
+	                                              "VK_KHR_maintenance1"};
+
+#if defined(__APPLE__)
+	// MoltenVK lacks VK_EXT_depth_clip_enable and VK_EXT_color_write_enable; the renderer
+	// falls back to default depth clipping and static color-write masks on macOS. It also
+	// requires VK_KHR_portability_subset per the Vulkan portability spec.
+	device_extensions.push_back("VK_KHR_portability_subset");
+#else
+	device_extensions.push_back(VK_EXT_DEPTH_CLIP_ENABLE_EXTENSION_NAME);
+	device_extensions.push_back(VK_EXT_COLOR_WRITE_ENABLE_EXTENSION_NAME);
+#endif
 
 #ifdef KYTY_ENABLE_DEBUG_PRINTF
 	if (Config::SpirvDebugPrintfEnabled()) {

src/graphics/presentation/window/window.cpp

@@ -1,5 +1,7 @@
 #include "graphics/presentation/window.h"
 
+#include <cstdlib>
+
 #include "SDL.h"
 #include "SDL_error.h"
 #include "SDL_events.h"
@@ -697,6 +699,52 @@ void WindowContext::ProcessEvent(double time_s) {
 	}
 }
 
+#if defined(__APPLE__)
+void WindowContext::RunOnMainThread(std::function<void()> task, bool wait) {
+	if (Common::Thread::IsMainThread()) {
+		task();
+		return;
+	}
+
+	uint64_t ticket = 0;
+	{
+		Common::LockGuard lock(main_task_mutex);
+		main_tasks.push_back(std::move(task));
+		ticket = ++main_tasks_queued;
+	}
+
+	// Wake the main loop in case it is blocked in SDL_WaitEvent.
+	SDL_Event event {};
+	event.type = SDL_USEREVENT;
+	SDL_PushEvent(&event);
+
+	if (!wait) {
+		return;
+	}
+	Common::LockGuard lock(main_task_mutex);
+	while (main_tasks_run < ticket) {
+		main_task_done.Wait(&main_task_mutex);
+	}
+}
+
+void WindowContext::DrainMainThreadTasks() {
+	std::vector<std::function<void()>> tasks;
+	{
+		Common::LockGuard lock(main_task_mutex);
+		tasks.swap(main_tasks);
+	}
+	if (tasks.empty()) {
+		return;
+	}
+	for (auto& task: tasks) {
+		task();
+	}
+	Common::LockGuard lock(main_task_mutex);
+	main_tasks_run += tasks.size();
+	main_task_done.SignalAll();
+}
+#endif
+
 void WindowContext::Run() {
 	Common::Timer timer;
 	timer.Start();
@@ -706,6 +754,9 @@ void WindowContext::Run() {
 	loop.paused.store(false, std::memory_order_release);
 
 	while (!loop.need_exit) {
+#if defined(__APPLE__)
+		DrainMainThreadTasks();
+#endif
 		if (SDL_PollEvent(&loop.event) != 0) {
 			ProcessEvent(timer.GetTimeS());
 			continue;
@@ -747,9 +798,18 @@ static void WindowCreate(WindowContext& context) {
 
 	LOGF("WindowCreate(): width = %d, height = %d\n", width, height);
 
+	uint32_t window_flags = KYTY_SDL_WINDOW_FLAGS;
+#if defined(__APPLE__)
+	// macOS 26 window chrome (CoreUI asset decode, SwiftUI titlebar) has been observed
+	// throwing NSExceptions under Rosetta during the first CATransaction commit. A
+	// borderless window skips that machinery entirely.
+	if (std::getenv("KYTY_BORDERLESS") != nullptr) {
+		window_flags |= static_cast<uint32_t>(SDL_WINDOW_BORDERLESS);
+	}
+#endif
 	context.window =
 	    SDL_CreateWindow(KYTY_SDL_WINDOW_CAPTION, KYTY_SDL_WINDOWPOS_CENTERED,
-	                     KYTY_SDL_WINDOWPOS_CENTERED, width, height, KYTY_SDL_WINDOW_FLAGS);
+	                     KYTY_SDL_WINDOWPOS_CENTERED, width, height, window_flags);
 
 	context.window_hidden = true;
 
@@ -896,7 +956,13 @@ void WindowContext::UpdateTitle() {
 	                       (has_app_ver ? " " : ""), device_name, processor_name,
 	                       frame_num, current_fps);
 
+#if defined(__APPLE__)
+	// AppKit traps on title changes off the main thread; fire-and-forget keeps present pacing.
+	RunOnMainThread([this, fps = std::move(fps)] { SDL_SetWindowTitle(window, fps.c_str()); },
+	                false);
+#else
 	SDL_SetWindowTitle(window, fps.c_str());
+#endif
 }
 
 } // namespace Libs::Graphics

src/graphics/presentation/window/windowInternal.h

@@ -12,6 +12,10 @@
 #include <memory>
 #include <vector>
 
+#if defined(__APPLE__)
+#include <functional>
+#endif
+
 namespace Libs::Graphics {
 
 class Presenter;
@@ -47,6 +51,14 @@ struct WindowContext {
 	void ProcessEvent(double time_seconds);
 	void Run();
 
+#if defined(__APPLE__)
+	// AppKit only allows window operations (show, icon, title, view/layer changes) on the
+	// main thread; the present thread marshals them through the SDL main loop with these.
+	// wait=true blocks until the task has run on the main thread.
+	void RunOnMainThread(std::function<void()> task, bool wait);
+	void DrainMainThreadTasks();
+#endif
+
 	GraphicContext      graphic_ctx;
 	SDL_Window*         window        = nullptr;
 	bool                window_hidden = true;
@@ -60,6 +72,14 @@ struct WindowContext {
 	char processor_name[64]                            = {0};
 
 	Common::Mutex mutex;
+
+#if defined(__APPLE__)
+	Common::Mutex                      main_task_mutex;
+	Common::CondVar                    main_task_done;
+	std::vector<std::function<void()>> main_tasks;        // guarded by main_task_mutex
+	uint64_t                           main_tasks_queued = 0; // guarded by main_task_mutex
+	uint64_t                           main_tasks_run    = 0; // guarded by main_task_mutex
+#endif
 };
 
 } // namespace Libs::Graphics

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -2,17 +2,17 @@
 
 #include "common/assert.h"
 #include "common/logging/log.h"
-#include "graphics/shader/recompiler/BindingLayout.h"
-#include "graphics/shader/recompiler/ResourceMaterialization.h"
-#include "graphics/shader/recompiler/ResourceTracking.h"
-#include "graphics/shader/recompiler/ScalarProvenance.h"
-#include "graphics/shader/recompiler/ShaderCFG.h"
-#include "graphics/shader/recompiler/ShaderDecoder.h"
-#include "graphics/shader/recompiler/ShaderIR.h"
-#include "graphics/shader/recompiler/ShaderInfoCollection.h"
-#include "graphics/shader/recompiler/SpirvEmitter.h"
-#include "graphics/shader/recompiler/SrtPatcher.h"
-#include "graphics/shader/recompiler/SrtWalker.h"
+#include "graphics/shader/recompiler/ir/BindingLayout.h"
+#include "graphics/shader/recompiler/ir/ResourceMaterialization.h"
+#include "graphics/shader/recompiler/ir/ResourceTracking.h"
+#include "graphics/shader/recompiler/ir/ScalarProvenance.h"
+#include "graphics/shader/recompiler/cfg/ShaderCFG.h"
+#include "graphics/shader/recompiler/decompiler/ShaderDecoder.h"
+#include "graphics/shader/recompiler/ir/ShaderIR.h"
+#include "graphics/shader/recompiler/ir/ShaderInfoCollection.h"
+#include "graphics/shader/recompiler/emitter/SpirvEmitter.h"
+#include "graphics/shader/recompiler/ir/SrtPatcher.h"
+#include "graphics/shader/recompiler/ir/SrtWalker.h"
 
 #include <algorithm>
 #include <array>

src/graphics/shader/recompiler/ShaderRecompiler.h

@@ -3,7 +3,7 @@
 
 #include "common/common.h"
 #include "common/stringUtils.h"
-#include "graphics/shader/recompiler/ResourceMaterialization.h"
+#include "graphics/shader/recompiler/ir/ResourceMaterialization.h"
 #include "graphics/shader/shader.h"
 
 #include <optional>

src/graphics/shader/recompiler/cfg/ShaderCFG.cpp

@@ -1,4 +1,4 @@
-#include "graphics/shader/recompiler/ShaderCFG.h"
+#include "graphics/shader/recompiler/cfg/ShaderCFG.h"
 
 #include <algorithm>
 #include <fmt/format.h>

src/graphics/shader/recompiler/cfg/ShaderCFG.h

@@ -3,7 +3,7 @@
 
 #include "common/common.h"
 #include "common/stringUtils.h"
-#include "graphics/shader/recompiler/ShaderDecoder.h"
+#include "graphics/shader/recompiler/decompiler/ShaderDecoder.h"
 
 #include <vector>
 

src/graphics/shader/recompiler/decompiler/ExportOps.cpp

@@ -1,4 +1,4 @@
-#include "graphics/shader/recompiler/ExportOps.h"
+#include "graphics/shader/recompiler/decompiler/ExportOps.h"
 
 #include <fmt/format.h>
 

src/graphics/shader/recompiler/decompiler/ExportOps.h

@@ -1,7 +1,7 @@
 #ifndef EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_EXPORTOPS_H_
 #define EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_EXPORTOPS_H_
 
-#include "graphics/shader/recompiler/ShaderDecoder.h"
+#include "graphics/shader/recompiler/decompiler/ShaderDecoder.h"
 
 namespace Libs::Graphics::ShaderRecompiler::Decoder {
 

src/graphics/shader/recompiler/decompiler/ImageOps.cpp

@@ -1,4 +1,4 @@
-#include "graphics/shader/recompiler/ImageOps.h"
+#include "graphics/shader/recompiler/decompiler/ImageOps.h"
 
 #include <algorithm>
 #include <fmt/format.h>

src/graphics/shader/recompiler/decompiler/ImageOps.h

@@ -1,7 +1,7 @@
 #ifndef EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_IMAGEOPS_H_
 #define EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_IMAGEOPS_H_
 
-#include "graphics/shader/recompiler/ShaderDecoder.h"
+#include "graphics/shader/recompiler/decompiler/ShaderDecoder.h"
 
 namespace Libs::Graphics::ShaderRecompiler::Decoder {
 

src/graphics/shader/recompiler/decompiler/MemoryOps.cpp

@@ -1,4 +1,4 @@
-#include "graphics/shader/recompiler/MemoryOps.h"
+#include "graphics/shader/recompiler/decompiler/MemoryOps.h"
 
 #include <fmt/format.h>
 #include <iterator>
@@ -421,13 +421,13 @@ bool DecodeDs(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
 	const uint32_t data0   = (word1 >> 8u) & 0xffu;
 	const uint32_t addr    = word1 & 0xffu;
 
-	inst.pc         = pc;
-	inst.word       = word0;
-	inst.word_count = 2;
-	inst.offset     = offset0 | (offset1 << 8u);
-	inst.gds        = ((word0 >> 17u) & 1u) != 0u;
-	inst.family     = Family::DS;
-	inst.opcode_id  = opcode;
+	inst.pc          = pc;
+	inst.word        = word0;
+	inst.word_count  = 2;
+	inst.offset      = offset0 | (offset1 << 8u);
+	inst.gds         = ((word0 >> 17u) & 1u) != 0u;
+	inst.family      = Family::DS;
+	inst.opcode_id   = opcode;
 	const auto* info = LookupMemoryOpcode(DS_OPS, static_cast<uint32_t>(std::size(DS_OPS)), opcode);
 	ApplyMemoryInfo(inst, info);
 	SetRawWords(inst, code, word_index, 2);
@@ -442,11 +442,6 @@ bool DecodeDs(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
 	     inst.opcode == Opcode::DsReadAddtidB32)) {
 		SetUnsupported(inst, Family::DS, opcode, "DS swizzle/addtid is available only for LDS");
 	}
-	if (inst.gds && (inst.opcode == Opcode::DsAppend || inst.opcode == Opcode::DsConsume) &&
-	    inst.offset != 0u) {
-		SetUnsupported(inst, Family::DS, opcode,
-		               "GDS append/consume requires a zero instruction offset");
-	}
 	if (inst.opcode == Opcode::DsWriteAddtidB32 && data1 != 0u) {
 		SetUnsupported(inst, Family::DS, opcode,
 		               "DS write addtid data1 operand is not implemented");

src/graphics/shader/recompiler/decompiler/MemoryOps.h

@@ -1,7 +1,7 @@
 #ifndef EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_MEMORYOPS_H_
 #define EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_MEMORYOPS_H_
 
-#include "graphics/shader/recompiler/ShaderDecoder.h"
+#include "graphics/shader/recompiler/decompiler/ShaderDecoder.h"
 
 namespace Libs::Graphics::ShaderRecompiler::Decoder {
 

src/graphics/shader/recompiler/decompiler/ScalarAluOps.cpp

@@ -1,4 +1,4 @@
-#include "graphics/shader/recompiler/ScalarAluOps.h"
+#include "graphics/shader/recompiler/decompiler/ScalarAluOps.h"
 
 #include <iterator>
 

src/graphics/shader/recompiler/decompiler/ScalarAluOps.h

@@ -1,7 +1,7 @@
 #ifndef EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_SCALARALUOPS_H_
 #define EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_SCALARALUOPS_H_
 
-#include "graphics/shader/recompiler/ShaderDecoder.h"
+#include "graphics/shader/recompiler/decompiler/ShaderDecoder.h"
 
 namespace Libs::Graphics::ShaderRecompiler::Decoder {
 

src/graphics/shader/recompiler/decompiler/ShaderDecoder.cpp

@@ -1,10 +1,10 @@
-#include "graphics/shader/recompiler/ShaderDecoder.h"
+#include "graphics/shader/recompiler/decompiler/ShaderDecoder.h"
 
-#include "graphics/shader/recompiler/ExportOps.h"
-#include "graphics/shader/recompiler/ImageOps.h"
-#include "graphics/shader/recompiler/MemoryOps.h"
-#include "graphics/shader/recompiler/ScalarAluOps.h"
-#include "graphics/shader/recompiler/VectorAluOps.h"
+#include "graphics/shader/recompiler/decompiler/ExportOps.h"
+#include "graphics/shader/recompiler/decompiler/ImageOps.h"
+#include "graphics/shader/recompiler/decompiler/MemoryOps.h"
+#include "graphics/shader/recompiler/decompiler/ScalarAluOps.h"
+#include "graphics/shader/recompiler/decompiler/VectorAluOps.h"
 
 #include <algorithm>
 #include <bit>

src/graphics/shader/recompiler/decompiler/VectorAluOps.cpp

@@ -1,4 +1,4 @@
-#include "graphics/shader/recompiler/VectorAluOps.h"
+#include "graphics/shader/recompiler/decompiler/VectorAluOps.h"
 
 #include <fmt/format.h>
 #include <iterator>

src/graphics/shader/recompiler/decompiler/VectorAluOps.h

@@ -1,7 +1,7 @@
 #ifndef EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_VECTORALUOPS_H_
 #define EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_VECTORALUOPS_H_
 
-#include "graphics/shader/recompiler/ShaderDecoder.h"
+#include "graphics/shader/recompiler/decompiler/ShaderDecoder.h"
 
 namespace Libs::Graphics::ShaderRecompiler::Decoder {
 

src/graphics/shader/recompiler/emitter/SpirvBuilder.cpp

@@ -1,4 +1,4 @@
-#include "graphics/shader/recompiler/SpirvBuilder.h"
+#include "graphics/shader/recompiler/emitter/SpirvBuilder.h"
 
 #include <cstdio>
 #include <cstring>

src/graphics/shader/recompiler/emitter/SpirvEmitter.cpp

@@ -1,7 +1,7 @@
-#include "graphics/shader/recompiler/SpirvEmitter.h"
+#include "graphics/shader/recompiler/emitter/SpirvEmitter.h"
 
-#include "graphics/shader/recompiler/SrtWalker.h"
-#include "graphics/shader/recompiler/spirvEmitter/spirvEmitterInternal.h"
+#include "graphics/shader/recompiler/emitter/spirvEmitterInternal.h"
+#include "graphics/shader/recompiler/ir/SrtWalker.h"
 
 #include <algorithm>
 #include <array>
@@ -160,8 +160,7 @@ bool ValidateInstructionContract(const IR::Instruction& inst, std::string* error
 	                            inst.dst.kind != IR::OperandKind::Null)) ||
 	    ((inst.op == IR::Opcode::DsAppend || inst.op == IR::Opcode::DsConsume) &&
 	     (!ds_kind || !ds_resource || inst.src_count != 1 ||
-	      inst.dst.kind != IR::OperandKind::Register ||
-	      (kind == IR::ResourceKind::Gds && inst.memory.offset != 0))) ||
+	      inst.dst.kind != IR::OperandKind::Register)) ||
 	    ((inst.op == IR::Opcode::DsMinF32 || inst.op == IR::Opcode::DsMaxF32) &&
 	     (!ds_kind || !ds_resource || inst.src_count != 3 ||
 	      inst.dst.kind != IR::OperandKind::Null)) ||

src/graphics/shader/recompiler/emitter/SpirvEmitter.h

@@ -3,7 +3,7 @@
 
 #include "common/common.h"
 #include "common/stringUtils.h"
-#include "graphics/shader/recompiler/ResourceMaterialization.h"
+#include "graphics/shader/recompiler/ir/ResourceMaterialization.h"
 
 #include <vector>
 

src/graphics/shader/recompiler/emitter/spirvEmitterAluOps.cpp

@@ -1,4 +1,4 @@
-#include "graphics/shader/recompiler/spirvEmitter/spirvEmitterInternal.h"
+#include "graphics/shader/recompiler/emitter/spirvEmitterInternal.h"
 
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {