[Feat] (DirectVulkan): relax fragment precision where the bound formats allow it

- WIP, parked: measures 80.9 -> 94.8 fps on Adreno 650 / MC 26.2 (same scene,
  device cooled to 38-40C), but is NOT validated. Desktop GLSL carries no
  precision qualifiers, so every fragment value reaches the driver as fp32 while
  Adreno runs fp16 at twice the rate.
- RelaxTextureDerivedPrecisionPass taints the values a fragment shader derives
  from built-in inputs and decorates everything else RelaxedPrecision. The
  taint direction matters: whitelisting outward from texture reads captures
  nothing, because MC multiplies every texel by an interpolated colour and a UBO
  value and one un-relaxed operand vetoes the expression - measured at 80.4 fps,
  i.e. no gain, both with and without varyings seeded. Precision-critical
  sources are few (gl_FragCoord cannot even hold a 3044-pixel x exactly), so
  tainting them and relaxing the rest is what actually pays.
- SPIR-V cannot see the bound formats - sampler2D yields vec4 whether the
  texture is RGBA8 or RGBA32F - so the decision is made per draw and passed in
  as a compile option, the same shape ExplicitLod0Sampling already uses.
  RelaxedFragmentPrecision is only requested when every sampled texture and
  every colour attachment is an 8-bit-or-less normalized format, where fp16's
  11-bit mantissa already carries the value exactly. Shaderpack HDR gbuffers,
  float data textures and 16-bit normalized targets therefore keep full
  precision, as do shaders that write gl_FragDepth or gl_SampleMask.
- LocalMultiStoreElim runs first: glslang emits function-local variables, and a
  load can never be relaxed, so without SSA promotion the analysis dies at the
  first temporary.
- WHY THIS IS PARKED: the retrace correctness gate never ran green. Every
  DirectVulkan retrace on Adreno 650 dies with DEVICE_LOST in
  UploadDirtyMipLevels on unmodified dev (pre-existing, device-gated), and on
  Adreno 830 - where the gate does pass on dev - minecraft-1.21.4-in-world times
  out at 900s with this change, which still needs explaining. Do not merge until
  that is understood and vanilla plus non-Photon shaderpack cases pass.
  (photon-v1.3b is broken on Adreno independently of this work.)
- The /sdcard/MG/exp_relaxed_precision_all and exp_no_relaxed_precision file
  toggles are development scaffolding for A/B measurement; they must go before
  this ships.

.github/scripts/fetch-trace-fixture-lfs.sh

@@ -9,7 +9,23 @@ fi
 case_name="$1"
 fixture_dir="${2:-tools/trace_replay/fixtures}"
 python_bin="${PYTHON:-python3}"
-mirror_base="${MOBILEGL_TRACE_FIXTURE_MIRROR_BASE:-https://repo.miawa.cn/mgl/tools/trace_replay/fixtures}"
+# Fixture mirrors, tried in order before falling back to Git LFS. Override the
+# whole list with MOBILEGL_TRACE_FIXTURE_MIRROR_BASES (whitespace separated);
+# MOBILEGL_TRACE_FIXTURE_MIRROR_BASE still works and is tried first.
+default_mirror_bases=(
+  "https://git.hit.moe/swung0x48/MobileGL/media/branch/dev/tools/trace_replay/fixtures"
+  "https://repo.miawa.cn/mgl/tools/trace_replay/fixtures"
+)
+if [ -n "${MOBILEGL_TRACE_FIXTURE_MIRROR_BASES:-}" ]; then
+  read -r -a mirror_bases <<< "${MOBILEGL_TRACE_FIXTURE_MIRROR_BASES}"
+else
+  mirror_bases=("${default_mirror_bases[@]}")
+fi
+if [ -n "${MOBILEGL_TRACE_FIXTURE_MIRROR_BASE:-}" ]; then
+  mirror_bases=("${MOBILEGL_TRACE_FIXTURE_MIRROR_BASE}" "${mirror_bases[@]}")
+fi
+# Optional bearer token for mirrors that require authentication (private Gitea).
+mirror_token="${MOBILEGL_TRACE_FIXTURE_MIRROR_TOKEN:-}"
 download_attempts="${MOBILEGL_TRACE_FIXTURE_DOWNLOAD_ATTEMPTS:-5}"
 retry_delay="${MOBILEGL_TRACE_FIXTURE_RETRY_DELAY:-2}"
 
@@ -30,7 +46,8 @@ fixture_list="$("${python_bin}" tools/trace_replay/trace_cases.py \
   --format fixture-files \
   --case "${case_name}" \
   --fixture-root "${fixture_dir}")"
-mapfile -t files <<< "${fixture_list}"
+# Strip CR so the script also works when python emits CRLF (Git Bash on Windows).
+mapfile -t files < <(printf '%s\n' "${fixture_list}" | tr -d '\r')
 
 include="$(IFS=,; echo "${files[*]}")"
 if [ "${case_name}" = "OpenRA" ]; then
@@ -106,6 +123,7 @@ fetch_file_from_mirror() {
   local attempt
   local partial_size
   local curl_status
+  local curl_auth
 
   metadata="$(get_lfs_metadata "${file}")" || return 1
   read -r expected_oid expected_size <<< "${metadata}"
@@ -136,7 +154,11 @@ fetch_file_from_mirror() {
       echo "Starting mirror download for ${file} (attempt ${attempt}/${download_attempts})"
     fi
 
-    if curl -L --fail --show-error --continue-at - --output "${tmp_file}" "${url}"; then
+    curl_auth=()
+    if [ -n "${mirror_token}" ]; then
+      curl_auth=(--header "Authorization: token ${mirror_token}")
+    fi
+    if curl -L --fail --show-error --continue-at - "${curl_auth[@]}" --output "${tmp_file}" "${url}"; then
       if verify_fixture_file "${tmp_file}" "${file}" "${expected_oid}" "${expected_size}"; then
         mv "${tmp_file}" "${file}"
         return 0
@@ -184,10 +206,19 @@ fetch_from_mirror() {
   for file in "${files[@]}"; do
     local name
     local url
+    local base
+    local fetched=0
     name="$(basename "${file}")"
-    url="${mirror_base%/}/${name}"
-    echo "Fetching trace fixture from mirror: ${url}"
-    if ! fetch_file_from_mirror "${file}" "${url}"; then
+    for base in "${mirror_bases[@]}"; do
+      url="${base%/}/${name}"
+      echo "Fetching trace fixture from mirror: ${url}"
+      if fetch_file_from_mirror "${file}" "${url}"; then
+        fetched=1
+        break
+      fi
+      echo "Mirror did not serve ${name}; trying the next mirror" >&2
+    done
+    if [ "${fetched}" -ne 1 ]; then
       return 1
     fi
   done
@@ -196,7 +227,7 @@ fetch_from_mirror() {
 if fetch_from_mirror; then
   echo "Fetched trace fixture files for ${case_name} from mirror: ${include}"
 else
-  echo "Mirror fetch failed for ${case_name}; falling back to Git LFS: ${include}"
+  echo "All mirrors failed for ${case_name}; falling back to Git LFS: ${include}"
   git lfs install --local
   git lfs pull --include="${include}" --exclude=""
 fi

.github/workflows/test.yml

@@ -455,6 +455,15 @@ jobs:
           if [ '${{ matrix.backend }}' = 'DirectVulkan' ]; then
             export MOBILEGL_MAGMA_R11G11B10F_FALLBACK=1
           fi
+          # The blended depth-write quirk auto-enables only on Qualcomm, which no CI
+          # runner has, so force it on for the OIT case it exists to fix. ForceOn
+          # bypasses only the vendor gate, so this exercises the real strip on
+          # lavapipe. The Android AVD lane deliberately leaves it off, keeping the
+          # unstripped path covered for the same trace.
+          if [ '${{ matrix.backend }}' = 'DirectVulkan' ] \
+             && [ '${{ matrix.case }}' = 'improved-transparency-minecraft-26.3' ]; then
+            export MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE=1
+          fi
           ctest -V --no-tests=error -R '^MobileGLTraceReplay\.${{ matrix.case }}\.${{ matrix.backend }}$'
 
       - name: Upload actual image

CMakeLists.txt

@@ -188,9 +188,12 @@ set(SOURCE_FILES
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/EliminateFloatEqualsZeroPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/RenameSamplerFunctionParameterPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DecomposeWorkgroupVec3Pass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DecoratePositionInvariantPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/LowerDrawParametersPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/RebaseInstanceIndexPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/StripUboMemberRelaxedPrecisionPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/StripNoPerspectivePass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/EmulateNoPerspectivePass.cpp
 
     MobileGL/MG_Util/BackendLoaders/OpenGL/Loader.cpp
     MobileGL/MG_Util/BackendLoaders/Vulkan/Loader.cpp
@@ -300,6 +303,13 @@ if (ANDROID)
     )
 endif()
 
+if (WIN32)
+    list(APPEND SOURCE_FILES
+        MobileGL/MG_Impl/WGLImpl/WGLImpl.cpp
+        MobileGL/MG_Impl/WGLImpl/Exporting/Definitions.cpp
+    )
+endif()
+
 set(MOBILEGL_LINK_LIBRARIES
         glslang::glslang
         spirv-cross-c
@@ -328,10 +338,18 @@ set(MOBILEGL_INCLUDE_DIR
         ${SPIRV-Headers_SOURCE_DIR}/include
 )
 
-add_library(${CMAKE_PROJECT_NAME} SHARED 
+add_library(${CMAKE_PROJECT_NAME} SHARED
     ${SOURCE_FILES}
 )
 
+if (WIN32)
+    # The wgl* entry points are exported via .def (see the comment in wgl.def);
+    # only the shared library links it.
+    target_sources(${CMAKE_PROJECT_NAME} PRIVATE
+        MobileGL/MG_Impl/WGLImpl/Exporting/wgl.def
+    )
+endif()
+
 if (CMAKE_BUILD_TYPE STREQUAL "Debug")
     set_target_properties(${CMAKE_PROJECT_NAME} PROPERTIES
         C_VISIBILITY_PRESET default
@@ -374,6 +392,18 @@ if(UNIX AND NOT APPLE AND NOT ANDROID)
     endforeach()
 endif()
 
+if(WIN32)
+    # Drop-in for the classic GL loader path: a copy named opengl32.dll placed
+    # next to a host executable is what LoadLibrary("opengl32.dll") and gdi32's
+    # pixel-format forwarding will resolve.
+    add_custom_command(TARGET ${CMAKE_PROJECT_NAME} POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different
+                "$<TARGET_FILE:${CMAKE_PROJECT_NAME}>"
+                "$<TARGET_FILE_DIR:${CMAKE_PROJECT_NAME}>/opengl32.dll"
+        COMMENT "Creating opengl32.dll drop-in copy"
+    )
+endif()
+
 if(NOT ANDROID)
     add_library(${CMAKE_PROJECT_NAME}_s STATIC 
         ${SOURCE_FILES}
@@ -425,8 +455,21 @@ if (ANDROID)
 endif()
 
 if (APPLE AND NOT MOBILEGL_IOS)
+    # MobileGL statically embeds glslang, SPIRV-Tools, and SPIRV-Cross. When
+    # this dylib is injected with DYLD_INSERT_LIBRARIES, exporting those C++
+    # symbols interposes incompatible copies embedded by host libraries such
+    # as shaderc. Keep only the public GL/EGL/CGL loader surface globally
+    # visible; GetProcAddress can still return pointers to hidden internals.
+    set(MOBILEGL_MACOS_EXPORTED_SYMBOLS
+        "${CMAKE_CURRENT_SOURCE_DIR}/MobileGL/MG_Impl/DyldInterpose/ExportedSymbols.txt")
+    target_link_options(${CMAKE_PROJECT_NAME} PRIVATE
+        "LINKER:-exported_symbols_list,${MOBILEGL_MACOS_EXPORTED_SYMBOLS}")
+    set_property(TARGET ${CMAKE_PROJECT_NAME} APPEND PROPERTY
+        LINK_DEPENDS "${MOBILEGL_MACOS_EXPORTED_SYMBOLS}")
+
     target_link_libraries(${CMAKE_PROJECT_NAME} PUBLIC
         "-framework Cocoa"
+        "-framework CoreVideo"
         "-framework QuartzCore"
         "-framework Foundation"
         "-framework OpenGL"
@@ -434,6 +477,7 @@ if (APPLE AND NOT MOBILEGL_IOS)
     if(TARGET ${CMAKE_PROJECT_NAME}_s)
         target_link_libraries(${CMAKE_PROJECT_NAME}_s PUBLIC
             "-framework Cocoa"
+            "-framework CoreVideo"
             "-framework QuartzCore"
             "-framework Foundation"
             "-framework OpenGL"

MobileGL/Config.h

@@ -20,6 +20,15 @@ namespace MobileGL::MG_Config {
 
     extern BackendType ActiveBackendType;
 
+    // Tri-state override for device-specific quirks: Auto lets the detected device decide,
+    // ForceOn/ForceOff bypass the detection in either direction. ForceOn only bypasses the
+    // device gate - each quirk keeps its structural safety checks.
+    enum class QuirkOverride : Uint8 {
+        Auto = 0,
+        ForceOn,
+        ForceOff,
+    };
+
     // Feature toggles parsed once from environment variables in MG_ConfigLoader::Init()
     // (ConfigLoader.cpp), before the accepted-env map is destroyed. All Bool fields share
     // one truthy rule: the variable is set, non-empty, not "0", and not "false"
@@ -67,6 +76,21 @@ namespace MobileGL::MG_Config {
         // explicitly request a core profile via EGL_CONTEXT_OPENGL_PROFILE_MASK / a >=3.1
         // version request.
         Bool RelaxedSemantics = false;
+        // MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN: overrides the shader-source quirk that
+        // rewrites the recognized workgroup prefix-scan template on Qualcomm devices with
+        // subgroups wider than 32 lanes (see ShaderSourceProcessor's quirk registry).
+        QuirkOverride SubgroupPrefixScanQuirk = QuirkOverride::Auto;
+        // MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE: overrides the DirectVulkan quirk that
+        // strips depth writes from accumulation-blended pipelines (MIN/MAX or additive
+        // ONE+ONE - the multi-pass depth-equality signature) on drivers without
+        // cross-pipeline vertex position invariance. Sorted-transparency "over" blends,
+        // gl_FragDepth writers, and fully color-masked attachments are exempt (see
+        // PipelineFactory::ShouldSuppressDepthWrite). Auto detects Qualcomm.
+        QuirkOverride MagmaDisableBlendedDepthWriteQuirk = QuirkOverride::Auto;
+        // MOBILEGL_DISABLE_ROBUST_BUFFER_ACCESS: leave the Vulkan robustBufferAccess device
+        // feature off. It is enabled by default to match GL's defined out-of-range fetch
+        // behavior; this escape hatch exists to measure or dodge its GPU cost on a device.
+        Bool DisableRobustBufferAccess = false;
     };
     extern FeaturesTable Features;
 } // namespace MobileGL::MG_Config

MobileGL/ConfigLoader.cpp

@@ -86,6 +86,17 @@ namespace MobileGL::MG_ConfigLoader {
         return it != acceptedEnvVariablesMap->end() && IsTruthyValue(it->second);
     }
 
+    // Quirk overrides are tri-state: an unset variable keeps device auto-detection, a truthy
+    // value forces the quirk on, anything else set ("0", "false", "") forces it off.
+    inline MG_Config::QuirkOverride QueryEnvQuirkOverride(const String& key) {
+        auto it = acceptedEnvVariablesMap->find(key);
+        if (it == acceptedEnvVariablesMap->end()) {
+            return MG_Config::QuirkOverride::Auto;
+        }
+        return IsTruthyValue(it->second) ? MG_Config::QuirkOverride::ForceOn
+                                         : MG_Config::QuirkOverride::ForceOff;
+    }
+
     inline Uint32 QueryEnvUint32(const String& key, Uint32 defaultValue, Uint32 minValue, Uint32 maxValue) {
         auto it = acceptedEnvVariablesMap->find(key);
         if (it == acceptedEnvVariablesMap->end()) {
@@ -123,6 +134,10 @@ namespace MobileGL::MG_ConfigLoader {
         features.TraceSkipAutodestroy = QueryEnvFlag("MOBILEGL_TRACE_SKIP_AUTODESTROY");
         features.DisableUboRing = QueryEnvFlag("MOBILEGL_DISABLE_UBO_RING");
         features.RelaxedSemantics = QueryEnvFlag("MOBILEGL_RELAXED_SEMANTICS");
+        features.SubgroupPrefixScanQuirk = QueryEnvQuirkOverride("MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN");
+        features.MagmaDisableBlendedDepthWriteQuirk =
+            QueryEnvQuirkOverride("MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE");
+        features.DisableRobustBufferAccess = QueryEnvFlag("MOBILEGL_DISABLE_ROBUST_BUFFER_ACCESS");
     }
 
     inline void InitBackendType() {

MobileGL/Defines.h

@@ -34,6 +34,7 @@
 #define MOBILEGL_EGL_API MOBILEGL_API
 #define MOBILEGL_CGL_API MOBILEGL_API
 #define MOBILEGL_NSOPENGL_API MOBILEGL_API
+#define MOBILEGL_WGL_API MOBILEGL_API
 
 // ====================== MobileGL configurations ======================= //
 #ifndef MOBILEGL_LOG_ACTIVE_LEVEL

MobileGL/GlobalObjects.cpp

@@ -14,7 +14,13 @@ namespace MobileGL {
     } // namespace MG_Config
 
     namespace MG_Backend {
-        UniquePtr<BackendObject> pActiveBackendObject;
+        // Leak-at-exit storage: the UniquePtr itself lives on the heap and is
+        // never destroyed by the runtime, so process exit runs no backend
+        // destructors (static destruction order across TUs is undefined).
+        // Deterministic teardown happens inside the EGL lifecycle instead:
+        // the last eglTerminate calls MobileGL::Destroy(), which .reset()s
+        // these singletons while the process is still healthy.
+        UniquePtr<BackendObject>& pActiveBackendObject = *new UniquePtr<BackendObject>();
         GlobalBackendFunctionsTable gBackendFunctionsTable;
     } // namespace MG_Backend
 } // namespace MobileGL

MobileGL/Init.cpp

@@ -9,14 +9,25 @@
 #include "Init.h"
 #include "Config.h"
 #include <MG_Backend/BackendObjects.h>
+#include <MG_Backend/DirectVulkan/DirectVulkan.h>
 #include <MG_State/GLState/Core.h>
 #include <MG_State/EGLState/Core.h>
 #include <MG_Impl/GLImpl/Texture/ProxyTexture.h>
 #include <MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.h>
+#include <MG_Impl/GLImpl/Sync/GL_Sync.h>
+
+#include <atomic>
+#include <mutex>
 
 namespace MobileGL {
     namespace {
-        Bool g_isInitialized = false;
+        std::atomic<Bool> g_isInitialized = false;
+        thread_local Bool tl_initializing = false;
+
+        std::mutex& InitMutex() {
+            static std::mutex mutex;
+            return mutex;
+        }
 
         void DestroyImpl(Bool logLifecycle) {
             if (!g_isInitialized) {
@@ -27,6 +38,12 @@ namespace MobileGL {
                 MGLOG_I("MobileGL closing...");
             }
             glslang::FinalizeProcess();
+            // GL syncs die with their contexts, and every context is gone by the
+            // time full teardown runs: drain the live-sync registry while the
+            // backend function table can still release the backend handles (and
+            // before a re-initialized library could pair them with the wrong
+            // backend's DeleteSync).
+            MG_Impl::GLImpl::DestroyAllSyncObjects();
             MG_Backend::pActiveBackendObject.reset();
             MG_State::pGLContext.reset();
             MG_State::pEGLContext.reset();
@@ -64,40 +81,37 @@ namespace MobileGL {
         MGLOG_I("MobileGL initialized");
     }
 
+    void EnsureInitialized() {
+        if (g_isInitialized.load(std::memory_order_acquire)) {
+            return;
+        }
+        // Re-entrant call while this thread is already inside Initialize()
+        // (e.g. an init step routing back through a public entry point).
+        if (tl_initializing) {
+            return;
+        }
+        const std::lock_guard<std::mutex> lock(InitMutex());
+        if (g_isInitialized.load(std::memory_order_acquire)) {
+            return;
+        }
+        tl_initializing = true;
+        Initialize();
+        tl_initializing = false;
+    }
+
     void Destroy() {
         DestroyImpl(true);
     }
 
-#if defined(__linux__) || defined(__APPLE__)
-    __attribute__((constructor)) static void AutoInit() {
-        Initialize();
-    }
-
-    __attribute__((destructor)) static void AutoDestroy() {
-        if (MG_Config::Features.TraceSkipAutodestroy) {
-            return;
-        }
-#if defined(__APPLE__)
-        // macOS injected dylibs can run destructors after logging/backend static state is already torn down.
-        return;
-#else
-        DestroyImpl(false);
-#endif
-    }
-#endif
-
-#ifdef _WIN32
-    BOOL WINAPI DllMain(HMODULE hModule, DWORD ul_reason_for_call, LPVOID lpReserved) {
-        switch (ul_reason_for_call) {
-        case DLL_PROCESS_ATTACH:
-            Initialize();
-            break;
-
-        case DLL_PROCESS_DETACH:
-            Destroy();
-            break;
-        }
-        return TRUE;
-    }
-#endif
+    // MobileGL's lifecycle is owned entirely by the host-API layers
+    // (EGL/WGL/CGL): initialization happens lazily on the first entry point
+    // via EnsureInitialized(), and full teardown happens deterministically
+    // when the last EGL display is terminated with nothing current (EGLImpl
+    // calls Destroy()). There is intentionally no backend-initializing static
+    // constructor, no static destructor, and no DllMain: the global singletons
+    // use leak-at-exit storage (see GlobalObjects.cpp), so a process that exits
+    // without eglTerminate simply leaks them to the OS instead of running
+    // backend destructors during static teardown. macOS has a lightweight
+    // dyld constructor that installs NSOpenGL dispatch hooks only; full backend
+    // initialization still enters here from the first hooked CGL context.
 } // namespace MobileGL

MobileGL/Init.h

@@ -11,6 +11,13 @@
 
 namespace MobileGL {
     void Initialize();
+    // Thread-safe, idempotent, and re-entrant wrapper around Initialize().
+    // Host layers (EGL/WGL/CGL entry points) call this lazily on first use so
+    // full backend initialization never depends on ELF/DLL static constructors,
+    // and so a fresh init can follow a full Destroy() (e.g. after the last
+    // eglTerminate). The macOS dyld bootstrap installs only lightweight
+    // NSOpenGL method hooks.
+    void EnsureInitialized();
     void Destroy();
 
     namespace MG_Util::Debug {

MobileGL/MG_Backend/BackendObject.h

@@ -230,6 +230,21 @@ namespace MobileGL {
             void (*SetSwapInterval)(Int interval);
         };
 
+        // Coarse GPU vendor identity for gating device-specific quirks. Detected from the
+        // Vulkan physical-device vendorID or the GLES GL_VENDOR/GL_RENDERER strings; stays
+        // Unknown when detection is inconclusive, in which case auto-gated quirks stay off.
+        enum class GpuVendorKind : Uint8 {
+            Unknown = 0,
+            Qualcomm,
+            Arm,
+            Nvidia,
+            Amd,
+            Intel,
+            ImgTec,
+            // Software rasterizers (llvmpipe/lavapipe, SwiftShader).
+            Software,
+        };
+
         struct DynamicBackendParameters {
             SizeT UniformBufferOffsetAlignment = 256;
             // GL_MAX_TEXTURE_MAX_ANISOTROPY_EXT. 1.0 means the backend cannot filter anisotropically,
@@ -291,13 +306,15 @@ namespace MobileGL {
             Uint32 SubgroupSupportedStages = 0;
             Uint32 SubgroupSupportedFeatures = 0;
             Bool SubgroupQuadOperationsInAllStages = false;
+            GpuVendorKind GpuVendor = GpuVendorKind::Unknown;
         };
 
         enum class WindowBackend {
             Android,
             X11,
             MetalLayer,
-            // TODO: Wayland, Windows, etc.
+            Win32, // Handle is an HWND
+            // TODO: Wayland, etc.
             WindowBackendCount,
             Unknown = -1
         };

MobileGL/MG_Backend/BackendObjects.h

@@ -13,6 +13,6 @@
 #include "DirectVulkan/BackendObject_DirectVulkan.h"
 
 namespace MobileGL::MG_Backend {
-    extern UniquePtr<BackendObject> pActiveBackendObject;
+    extern UniquePtr<BackendObject>& pActiveBackendObject;
     extern GlobalBackendFunctionsTable gBackendFunctionsTable;
 } // namespace MobileGL::MG_Backend

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.cpp

@@ -701,9 +701,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
         if ((handle.Backend != WindowBackend::Android &&
              handle.Backend != WindowBackend::X11 &&
-             handle.Backend != WindowBackend::MetalLayer) ||
+             handle.Backend != WindowBackend::MetalLayer &&
+             handle.Backend != WindowBackend::Win32) ||
             !handle.Handle) {
-            MGLOG_E("DirectGLES backend only supports Android, X11, and CAMetalLayer native windows");
+            MGLOG_E("DirectGLES backend only supports Android, X11, CAMetalLayer, and Win32 native windows");
             return false;
         }
 
@@ -826,7 +827,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                           E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object,
                                           E_GL_EXT_framebuffer_object, E_GL_ARB_depth_texture, E_GL_ARB_buffer_storage,
                                           E_GL_ARB_texture_storage, E_GL_ARB_texture_storage_multisample,
-                                          E_GL_ARB_direct_state_access,
+                                          E_GL_ARB_clear_texture, E_GL_ARB_direct_state_access,
                                           E_GL_ARB_multi_draw_indirect, E_GL_ARB_indirect_parameters,
                                           E_GL_ARB_shader_draw_parameters, E_GL_ARB_gpu_shader5, E_GL_ARB_multi_bind,
                                           E_GL_ARB_shading_language_420pack, E_GL_ARB_vertex_attrib_binding,
@@ -1035,6 +1036,32 @@ namespace MobileGL::MG_Backend::DirectGLES {
         m_dynamicParameters.ViewportSubpixelBits = m_GLESCapabilities.ViewportSubpixelBits;
         m_dynamicParameters.SupportsWideLines =
             m_GLESCapabilities.AliasedLineWidthRangeMax > 1.0f || m_GLESCapabilities.SmoothLineWidthRangeMax > 1.0f;
+
+        const auto containsAny = [](const String& haystack, std::initializer_list<const char*> needles) {
+            return std::any_of(needles.begin(), needles.end(), [&](const char* needle) {
+                return haystack.find(needle) != String::npos;
+            });
+        };
+        const String vendorAndRenderer =
+            m_GLESCapabilities.GLESVendorString + " " + m_GLESCapabilities.GLESRendererString;
+        if (containsAny(vendorAndRenderer, {"llvmpipe", "SwiftShader", "softpipe"})) {
+            // Check software rasterizers first: ANGLE-on-llvmpipe reports both.
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Software;
+        } else if (containsAny(vendorAndRenderer, {"Qualcomm", "Adreno"})) {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Qualcomm;
+        } else if (containsAny(vendorAndRenderer, {"Mali", "ARM"})) {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Arm;
+        } else if (containsAny(vendorAndRenderer, {"NVIDIA"})) {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Nvidia;
+        } else if (containsAny(vendorAndRenderer, {"AMD", "Radeon"})) {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Amd;
+        } else if (containsAny(vendorAndRenderer, {"Intel"})) {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Intel;
+        } else if (containsAny(vendorAndRenderer, {"Imagination", "PowerVR"})) {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::ImgTec;
+        } else {
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Unknown;
+        }
     }
 
     const MG_External::GLESFunctionsTable& BackendObject_DirectGLES::GetGLESFunctions() const {

MobileGL/MG_Backend/DirectGLES/Managers.cpp

@@ -267,16 +267,26 @@ namespace MobileGL::MG_Backend::DirectGLES {
             Uint g_boundArrayBufferId = 0;
             Bool g_boundArrayBufferKnown = false;
 
+            // Driver-level GL_PIXEL_PACK/UNPACK_BUFFER binding shadows (see
+            // Managers.h). Resting state between operations is 0; scopes in the
+            // readback/upload paths bind what they need through the cache and
+            // return to 0, so a stale user PBO can never capture a later
+            // readback that meant to target client memory.
+            Uint g_boundPixelPackBufferId = 0;
+            Bool g_boundPixelPackBufferKnown = false;
+            Uint g_boundPixelUnpackBufferId = 0;
+            Bool g_boundPixelUnpackBufferKnown = false;
+
             // Bumped whenever the backend ES context is destroyed; resources with
             // an older generation hold ids from a dead context.
             Uint g_bufferContextGeneration = 1;
 
             // Defined next to the indexed-binding shadow below; forward-declared so
             // every glDeleteBuffers site in this namespace can scrub stale shadow
-            // entries (GL resets a deleted buffer's indexed bindings to 0, and a
-            // recycled name matching a stale shadow entry would otherwise
-            // false-skip the rebind).
-            void ScrubIndexedBufferBindingShadowForId(Uint id);
+            // entries (GL resets a deleted buffer's bindings - indexed and pixel
+            // pack/unpack alike - to 0, and a recycled name matching a stale shadow
+            // entry would otherwise false-skip the rebind).
+            void ScrubBufferBindingShadowsForId(Uint id);
 
             // Resources whose owning BufferObject died; ids deleted at the next
             // sync point with a current ES context.
@@ -318,10 +328,16 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 if (g_boundArrayBufferKnown && g_boundArrayBufferId == r.id) {
                     InvalidateArrayBufferBindingCache();
                 }
+                // Pooling keeps the id alive (and thus any driver binding of it);
+                // drop to unknown rather than claiming the post-delete 0 state.
+                if ((g_boundPixelPackBufferKnown && g_boundPixelPackBufferId == r.id) ||
+                    (g_boundPixelUnpackBufferKnown && g_boundPixelUnpackBufferId == r.id)) {
+                    InvalidatePixelBufferBindingCaches();
+                }
                 const std::lock_guard<std::mutex> lock(g_poolMutex);
                 auto& bucket = g_bufferPool[r.storageSize];
                 if (bucket.size() >= kMaxEntriesPerBucket || g_pooledBytes + r.storageSize > kMaxPoolBytes) {
-                    ScrubIndexedBufferBindingShadowForId(r.id);
+                    ScrubBufferBindingShadowsForId(r.id);
                     g_GLESFuncs.glDeleteBuffers(1, &r.id); // over budget: don't pool
                     r.id = 0;
                     return;
@@ -502,7 +518,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 // Need a fresh id: glBufferStorage fails on a buffer that already has
                 // immutable storage, and any prior mutable store is replaced anyway.
                 if (resource->id != 0) {
-                    ScrubIndexedBufferBindingShadowForId(resource->id);
+                    ScrubBufferBindingShadowsForId(resource->id);
                     g_GLESFuncs.glDeleteBuffers(1, &resource->id);
                     resource->id = 0;
                 }
@@ -630,7 +646,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                         if (g_boundArrayBufferKnown && g_boundArrayBufferId == glesResource->id) {
                             InvalidateArrayBufferBindingCache();
                         }
-                        ScrubIndexedBufferBindingShadowForId(glesResource->id);
+                        ScrubBufferBindingShadowsForId(glesResource->id);
                         g_GLESFuncs.glDeleteBuffers(1, &glesResource->id);
                         glesResource->id = 0;
                     }
@@ -668,6 +684,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
         void OnBackendContextDestroyed() {
             UnregisterBufferBackendOps();
             ++g_bufferContextGeneration;
+            InvalidateArrayBufferBindingCache();
+            InvalidateIndexedBufferBindingCache();
+            InvalidatePixelBufferBindingCaches();
             // The global-UBO ring's id and persistent map died with the context;
             // drop the handles (no GL) and let the next draw recreate the ring.
             ResetUboRingForNewContext();
@@ -694,7 +713,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     if (g_boundArrayBufferKnown && g_boundArrayBufferId == glesResource->id) {
                         InvalidateArrayBufferBindingCache();
                     }
-                    ScrubIndexedBufferBindingShadowForId(glesResource->id);
+                    ScrubBufferBindingShadowsForId(glesResource->id);
                     g_GLESFuncs.glDeleteBuffers(1, &glesResource->id);
                     glesResource->id = 0;
                 }
@@ -819,6 +838,41 @@ namespace MobileGL::MG_Backend::DirectGLES {
             g_boundArrayBufferKnown = false;
         }
 
+        void BindPixelPackBufferId(Uint id) {
+            if (g_boundPixelPackBufferKnown && g_boundPixelPackBufferId == id) {
+                return;
+            }
+            g_GLESFuncs.glBindBuffer(GL_PIXEL_PACK_BUFFER, id);
+            g_boundPixelPackBufferId = id;
+            g_boundPixelPackBufferKnown = true;
+        }
+
+        void BindPixelUnpackBufferId(Uint id) {
+            if (g_boundPixelUnpackBufferKnown && g_boundPixelUnpackBufferId == id) {
+                return;
+            }
+            g_GLESFuncs.glBindBuffer(GL_PIXEL_UNPACK_BUFFER, id);
+            g_boundPixelUnpackBufferId = id;
+            g_boundPixelUnpackBufferKnown = true;
+        }
+
+        void InvalidatePixelBufferBindingCaches() {
+            g_boundPixelPackBufferId = 0;
+            g_boundPixelPackBufferKnown = false;
+            g_boundPixelUnpackBufferId = 0;
+            g_boundPixelUnpackBufferKnown = false;
+        }
+
+        void NoteBufferIdDeleted(Uint id) {
+            if (id == 0) {
+                return;
+            }
+            if (g_boundArrayBufferKnown && g_boundArrayBufferId == id) {
+                InvalidateArrayBufferBindingCache();
+            }
+            ScrubBufferBindingShadowsForId(id);
+        }
+
         namespace {
             // Shadow of the GL indexed buffer bindings so redundant glBindBufferBase/Range
             // (same index + id + range) are skipped. isBase distinguishes a whole-buffer
@@ -840,12 +894,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 return nullptr;
             }
 
-            // glDeleteBuffers resets the deleted buffer's bindings (indexed ones
-            // included) to 0 in the current context; mirror that in the shadow, or a
-            // later buffer recycling the same name with a matching range would
-            // false-skip its rebind. Default IndexedBufferBinding{} == base(0) ==
-            // the post-delete GL state.
-            void ScrubIndexedBufferBindingShadowForId(Uint id) {
+            // glDeleteBuffers resets the deleted buffer's bindings (indexed and
+            // pixel pack/unpack ones included) to 0 in the current context; mirror
+            // that in the shadows, or a later buffer recycling the same name with a
+            // matching shadow entry would false-skip its rebind. Default
+            // IndexedBufferBinding{} == base(0) == the post-delete GL state.
+            void ScrubBufferBindingShadowsForId(Uint id) {
                 if (id == 0) return;
                 for (auto& binding : g_indexedUBOBindings) {
                     if (binding.id == id) binding = {};
@@ -853,6 +907,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 for (auto& binding : g_indexedSSBOBindings) {
                     if (binding.id == id) binding = {};
                 }
+                if (g_boundPixelPackBufferKnown && g_boundPixelPackBufferId == id) {
+                    g_boundPixelPackBufferId = 0;
+                }
+                if (g_boundPixelUnpackBufferKnown && g_boundPixelUnpackBufferId == id) {
+                    g_boundPixelUnpackBufferId = 0;
+                }
             }
         } // namespace
 
@@ -897,7 +957,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 auto& bucket = g_bufferPool[oldestKey];
                 PooledBuffer& e = bucket[oldestIdx];
                 if (e.contextGeneration == g_bufferContextGeneration && e.id != 0) {
-                    ScrubIndexedBufferBindingShadowForId(e.id);
+                    ScrubBufferBindingShadowsForId(e.id);
                     g_GLESFuncs.glDeleteBuffers(1, &e.id);
                 }
                 g_pooledBytes -= e.size;
@@ -1064,7 +1124,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 const Bool staleContext = entry.contextGeneration != g_bufferContextGeneration;
                 if (!staleContext && entry.retireSerial > completed) continue;
                 if (!staleContext && entry.id != 0) {
-                    ScrubIndexedBufferBindingShadowForId(entry.id);
+                    ScrubBufferBindingShadowsForId(entry.id);
                     g_GLESFuncs.glDeleteBuffers(1, &entry.id);
                 }
                 g_retiredUboRings[i] = g_retiredUboRings.back();
@@ -1152,6 +1212,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             }
             for (auto& bufferId : m_clientAttributeBufferIds) {
                 if (bufferId != 0) {
+                    BufferImpl::NoteBufferIdDeleted(bufferId);
                     g_GLESFuncs.glDeleteBuffers(1, &bufferId);
                     bufferId = 0;
                 }
@@ -1324,6 +1385,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
             g_GLESFuncs.glGenTextures(1, &m_backendTextureId);
+            m_contextGeneration = g_textureContextGeneration;
             if (m_backendTextureId == 0) {
                 MGLOG_E("Failed to generate texture object.");
                 MGLOG_E("ES glGetError(): %s", MG_Util::ConvertGLEnumToString(g_GLESFuncs.glGetError()).c_str());
@@ -1332,6 +1394,27 @@ namespace MobileGL::MG_Backend::DirectGLES {
             }
         }
 
+        BackendTextureObject::~BackendTextureObject() {
+            if (m_backendTextureId == 0) {
+                return;
+            }
+            // Scrub every driver-state shadow that could false-skip when the name
+            // or this heap address is recycled - regardless of whether the id can
+            // still be deleted.
+            ScratchFBOImpl::NoteTextureIdDeleted(m_backendTextureId);
+            for (auto& unitCache : g_boundTexturesCache) {
+                for (auto& boundTexture : unitCache) {
+                    if (boundTexture == this) {
+                        boundTexture = nullptr;
+                    }
+                }
+            }
+            if (m_contextGeneration == g_textureContextGeneration && g_GLESFuncs.glDeleteTextures) {
+                g_GLESFuncs.glDeleteTextures(1, &m_backendTextureId);
+            }
+            m_backendTextureId = 0;
+        }
+
         void BackendTextureObject::Bind(GLenum target, Uint unit) {
 #ifdef TRACY_ENABLE
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
@@ -1364,7 +1447,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
         void BackendTextureObject::RecreateBackendTexture() {
             if (m_backendTextureId != 0) {
-                g_GLESFuncs.glDeleteTextures(1, &m_backendTextureId);
+                ScratchFBOImpl::NoteTextureIdDeleted(m_backendTextureId);
+                if (m_contextGeneration == g_textureContextGeneration) {
+                    g_GLESFuncs.glDeleteTextures(1, &m_backendTextureId);
+                }
                 for (auto& unitCache : g_boundTexturesCache) {
                     for (auto& boundTexture : unitCache) {
                         if (boundTexture == this) {
@@ -1375,6 +1461,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
             }
 
             g_GLESFuncs.glGenTextures(1, &m_backendTextureId);
+            m_contextGeneration = g_textureContextGeneration;
             if (m_backendTextureId == 0) {
                 MGLOG_E("Failed to regenerate texture object.");
                 MGLOG_E("ES glGetError(): %s", MG_Util::ConvertGLEnumToString(g_GLESFuncs.glGetError()).c_str());
@@ -1391,7 +1478,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // glGetIntegerv - that query forces a driver pipeline sync and, because texture
         // uploads run it per dirty texture per frame, it dominated the DirectGLES draw
         // path. The backend unpack state is set ONLY by MobileGL's own save/restore
-        // helpers (this class, TempPixelStoreParameterSync, the R32F copy path), all of
+        // helpers (this class and, historically, the R32F copy path), all of
         // which restore to the resting default, so the shadow stays accurate; a one-time
         // forced sync pins the backend to that known default up front. Apply() is
         // compare-and-set, so the (now redundant) glPixelStorei calls also usually no-op.
@@ -1563,6 +1650,56 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return convertedData.data();
         }
 
+        // RGB565/RGB5_A1 shadow data is stored as 8-bit unorm; uploading it as GL_UNSIGNED_BYTE
+        // leaves the 8-bit -> 5/6-bit requantization to the driver, whose rounding direction is
+        // implementation-defined: Adreno rounds to nearest (lossless round trip) but Mali floors,
+        // drifting mid-range texels one 5-bit step down and failing the KHR-GL3x
+        // pixelstoragemodes.teximage3d rgb565/rgb5a1 1/32-eps checks. Repack the shadow rows into
+        // the packed 16-bit client type with round-to-nearest instead - that recovers the original
+        // 5/6-bit values exactly (the shadow expansion round(v * 255 / max) is injective), so the
+        // driver stores them verbatim with no requantization left to its discretion. 4-bit formats
+        // (RGBA4) are exempt: their 8-bit expansion (v * 17) is exact under either rounding.
+        // Always retargets *inOutType for these formats (even for null data) so every upload of a
+        // level uses the same client type.
+        static const void* PreparePackedNormUpload(TextureInternalFormat format, const IntVec3& texelSize,
+                                                   const void* data, SizeT byteSize, GLenum* inOutType,
+                                                   Vector<Uint8>& packedData) {
+            if (format != TextureInternalFormat::RGB5 && format != TextureInternalFormat::RGB5A1) {
+                return data;
+            }
+            const Bool hasAlpha = format == TextureInternalFormat::RGB5A1;
+            const GLenum packedType = hasAlpha ? GL_UNSIGNED_SHORT_5_5_5_1 : GL_UNSIGNED_SHORT_5_6_5;
+            // Idempotent across a region's level loop: glType is shared, so later levels arrive with
+            // the already-retargeted packed type and must still be converted.
+            if (*inOutType != GL_UNSIGNED_BYTE && *inOutType != packedType) {
+                return data;
+            }
+            *inOutType = packedType;
+            if (data == nullptr || byteSize == 0) {
+                return data;
+            }
+            const SizeT srcPixelBytes = hasAlpha ? 4 : 3;
+            const SizeT texelCount = std::min(static_cast<SizeT>(std::max(texelSize.x(), 0)) *
+                                                  static_cast<SizeT>(std::max(texelSize.y(), 0)) *
+                                                  static_cast<SizeT>(std::max(texelSize.z(), 1)),
+                                              byteSize / srcPixelBytes);
+            packedData.resize(texelCount * sizeof(Uint16));
+            const Uint8* src = static_cast<const Uint8*>(data);
+            auto* dst = reinterpret_cast<Uint16*>(packedData.data());
+            for (SizeT i = 0; i < texelCount; ++i, src += srcPixelBytes) {
+                const Uint32 r = (static_cast<Uint32>(src[0]) * 31u + 127u) / 255u;
+                const Uint32 b = (static_cast<Uint32>(src[2]) * 31u + 127u) / 255u;
+                if (hasAlpha) {
+                    const Uint32 g = (static_cast<Uint32>(src[1]) * 31u + 127u) / 255u;
+                    dst[i] = static_cast<Uint16>((r << 11) | (g << 6) | (b << 1) | (src[3] >= 128 ? 1u : 0u));
+                } else {
+                    const Uint32 g = (static_cast<Uint32>(src[1]) * 63u + 127u) / 255u;
+                    dst[i] = static_cast<Uint16>((r << 11) | (g << 5) | b);
+                }
+            }
+            return packedData.data();
+        }
+
         void BackendTextureObject::SyncMipmapsToBackend(
             const SharedPtr<MG_State::GLState::ITextureObject>& stateTextureObject) {
             if (!stateTextureObject) {
@@ -1706,9 +1843,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
                             const void* uploadData = PrepareNormFloatFallbackUpload(
                                 textureMipmapObject->GetFormat(), levelTexelSize, pData, levelByteSize, glType,
                                 convertedUploadData);
+                            Vector<Uint8> packedUploadData;
+                            uploadData = PreparePackedNormUpload(textureMipmapObject->GetFormat(), levelTexelSize,
+                                                                 uploadData, levelByteSize, &glType, packedUploadData);
 
                             DebugImpl::ErrorLopper::Clear();
-                            g_GLESFuncs.glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
+                            BufferImpl::BindPixelUnpackBufferId(0); // no-op once the resting 0 state is pinned
                             const IntVec3 uploadSize =
                                 GetBackendUploadSize(stateTextureObject->GetTarget(), levelTexelSize);
                             switch (MapToBackendTextureTarget(stateTextureObject->GetTarget())) {
@@ -1760,7 +1900,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     const auto& uploadTargets = textureMipmapObject->GetUploadTargets();
                     if (TextureImpl::IsMultisampleTextureTarget(targetInternal)) {
                         DebugImpl::ErrorLopper::Clear();
-                        g_GLESFuncs.glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
+                        BufferImpl::BindPixelUnpackBufferId(0); // no-op once the resting 0 state is pinned
                         switch (targetInternal) {
                         case TextureTarget::Texture2DMultisample:
                             g_GLESFuncs.glTexStorage2DMultisample(
@@ -1787,7 +1927,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                         }
                     } else if (stateTextureObject->IsImmutable() || m_imageBindableStorageRequired) {
                         DebugImpl::ErrorLopper::Clear();
-                        g_GLESFuncs.glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
+                        BufferImpl::BindPixelUnpackBufferId(0); // no-op once the resting 0 state is pinned
                         const IntVec3 storageSize = GetBackendUploadSize(targetInternal, baseSize);
                         switch (MapToBackendTextureTarget(targetInternal)) {
                         case TextureTarget::Texture2D:
@@ -1831,9 +1971,13 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                     const void* uploadData = PrepareNormFloatFallbackUpload(
                                         textureMipmapObject->GetFormat(), levelTexelSize, pData, levelByteSize, glType,
                                         convertedUploadData);
+                                    Vector<Uint8> packedUploadData;
+                                    uploadData =
+                                        PreparePackedNormUpload(textureMipmapObject->GetFormat(), levelTexelSize,
+                                                                uploadData, levelByteSize, &glType, packedUploadData);
 
                                     DebugImpl::ErrorLopper::Clear();
-                                    g_GLESFuncs.glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
+                                    BufferImpl::BindPixelUnpackBufferId(0); // no-op once the resting 0 state is pinned
                                     const IntVec3 uploadSize =
                                         GetBackendUploadSize(targetInternal, levelTexelSize);
                                     switch (MapToBackendTextureTarget(targetInternal)) {
@@ -1885,6 +2029,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                 const void* uploadData = PrepareNormFloatFallbackUpload(
                                     textureMipmapObject->GetFormat(), levelTexelSize, pData, levelByteSize, glType,
                                     convertedUploadData);
+                                Vector<Uint8> packedUploadData;
+                                uploadData =
+                                    PreparePackedNormUpload(textureMipmapObject->GetFormat(), levelTexelSize,
+                                                            uploadData, levelByteSize, &glType, packedUploadData);
                                 MGLOG_D("%s: target: %s: syncing mip %d: %dx%dx%d, byteSize = %d, pData = %p, "
                                         "levelDirty = %s",
                                         __func__, MG_Util::ConvertTextureUploadTargetToString(uploadTarget).c_str(),
@@ -1892,7 +2040,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                         levelByteSize, pData, levelDirty ? "true" : "false");
 
                                 DebugImpl::ErrorLopper::Clear();
-                                g_GLESFuncs.glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
+                                BufferImpl::BindPixelUnpackBufferId(0); // no-op once the resting 0 state is pinned
                                 auto textureTarget = stateTextureObject->GetTarget();
                                 const IntVec3 uploadSize = GetBackendUploadSize(textureTarget, levelTexelSize);
                                 switch (MapToBackendTextureTarget(textureTarget)) {
@@ -1978,7 +2126,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                                         textureMipmapObject->GetMipmapTexelSize(uploadTarget, level).y(), byteSize);
 
                             auto glUploadTarget = ConvertTextureUploadTargetToBackendGLEnum(uploadTarget);
-                            g_GLESFuncs.glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
+                            BufferImpl::BindPixelUnpackBufferId(0); // no-op once the resting 0 state is pinned
                             DebugImpl::ErrorLopper::Loop(
                                 [file = __FILE__, line = __LINE__, func = __func__](GLenum err) {
                                     MGLOG_D("%s(%s:%d) ES error: %s", func, file, line,
@@ -1990,6 +2138,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
                             const void* uploadData = PrepareNormFloatFallbackUpload(
                                 textureMipmapObject->GetFormat(), texelSize, mipData, byteSize, glType,
                                 convertedUploadData);
+                            Vector<Uint8> packedUploadData;
+                            uploadData = PreparePackedNormUpload(textureMipmapObject->GetFormat(), texelSize,
+                                                                 uploadData, byteSize, &glType, packedUploadData);
                             const IntVec3 uploadSize =
                                 GetBackendUploadSize(stateTextureObject->GetTarget(), texelSize);
                             switch (MapToBackendTextureTarget(stateTextureObject->GetTarget())) {
@@ -2216,11 +2367,17 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 return;
             }
 
-            if (TextureImpl::IsMultisampleTextureTarget(targetInternal)) {
+            // Multisample targets reject the *sampler* parameters (LOD range, border color) but
+            // GL_TEXTURE_SWIZZLE_* is texture state, not sampler state, and ES accepts it on them.
+            // Bailing out entirely used to drop every swizzle write on the floor, which is what the
+            // frontend already assumes is legal (see GL_Texture.cpp's MS-invalid pname list, which
+            // deliberately omits the swizzle enums). Note the caches for the skipped parameters are
+            // still refreshed so they never look stale, but m_cacheSwizzleParams must NOT be, or the
+            // change detection below would swallow the very writes we came here to emit.
+            const Bool isMultisampleTarget = TextureImpl::IsMultisampleTextureTarget(targetInternal);
+            if (isMultisampleTarget) {
                 m_cacheLodRange = stateTextureObject->GetLevelRange();
-                m_cacheSwizzleParams = stateTextureObject->GetAllSwizzleParams();
                 m_cacheBorderColor = stateTextureObject->GetBorderColor();
-                return;
             }
 
             Bind(target);
@@ -2233,14 +2390,14 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
             const auto& levelRange = stateTextureObject->GetLevelRange();
 
-            if (m_cacheLodRange.x() != levelRange.x()) {
+            if (!isMultisampleTarget && m_cacheLodRange.x() != levelRange.x()) {
                 g_GLESFuncs.glTexParameteri(target, GL_TEXTURE_BASE_LEVEL, static_cast<GLint>(levelRange.x()));
                 m_cacheLodRange.x() = levelRange.x();
             }
             DebugImpl::ErrorLopper::Loop([file = __FILE__, line = __LINE__, func = __func__](GLenum err) {
                 MGLOG_D("%s(%s:%d) ES error %s", func, file, line, MG_Util::ConvertGLEnumToString(err).c_str());
             });
-            if (m_cacheLodRange.y() != levelRange.y()) {
+            if (!isMultisampleTarget && m_cacheLodRange.y() != levelRange.y()) {
                 g_GLESFuncs.glTexParameteri(target, GL_TEXTURE_MAX_LEVEL, static_cast<GLint>(levelRange.y()));
                 m_cacheLodRange.y() = levelRange.y();
             }
@@ -2266,7 +2423,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 });
             }
 
-            if (m_cacheBorderColor != stateTextureObject->GetBorderColor()) {
+            if (!isMultisampleTarget && m_cacheBorderColor != stateTextureObject->GetBorderColor()) {
                 const auto& borderColor = stateTextureObject->GetBorderColor();
                 GLfloat borderColorArray[4] = {borderColor.x(), borderColor.y(), borderColor.z(), borderColor.w()};
                 g_GLESFuncs.glTexParameterfv(target, GL_TEXTURE_BORDER_COLOR, borderColorArray);
@@ -2295,6 +2452,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
         }
 
         Uint g_activeTextureUnit = 0;
+        Uint g_textureContextGeneration = 1;
         Array<Array<BackendTextureObject*, (SizeT)TextureTarget::TextureTargetCount>,
               MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS>
             g_boundTexturesCache;
@@ -2320,9 +2478,59 @@ namespace MobileGL::MG_Backend::DirectGLES {
             ZoneScopedC(TRACY_ZONECOLOR_BACKEND);
 #endif
             if (target == FramebufferTarget::Read)
-                g_GLESFuncs.glBindFramebuffer(GL_READ_FRAMEBUFFER, m_backendFBOId);
+                BindFramebufferId(GL_READ_FRAMEBUFFER, m_backendFBOId);
             else
-                g_GLESFuncs.glBindFramebuffer(GL_DRAW_FRAMEBUFFER, m_backendFBOId);
+                BindFramebufferId(GL_DRAW_FRAMEBUFFER, m_backendFBOId);
+        }
+
+        namespace {
+            // Driver-level framebuffer-binding shadow (see Managers.h). Indexed by
+            // FramebufferTarget {Draw, Read}.
+            Array<Uint, SizeT(FramebufferTarget::FramebufferTargetCount)> g_driverFBOBindings = {0, 0};
+            Array<Bool, SizeT(FramebufferTarget::FramebufferTargetCount)> g_driverFBOBindingKnown = {false, false};
+        } // namespace
+
+        void BindFramebufferId(GLenum fbTarget, Uint id) {
+            const Bool bindsDraw = fbTarget == GL_DRAW_FRAMEBUFFER || fbTarget == GL_FRAMEBUFFER;
+            const Bool bindsRead = fbTarget == GL_READ_FRAMEBUFFER || fbTarget == GL_FRAMEBUFFER;
+            const SizeT drawIdx = SizeT(FramebufferTarget::Draw);
+            const SizeT readIdx = SizeT(FramebufferTarget::Read);
+            const Bool drawMatches =
+                !bindsDraw || (g_driverFBOBindingKnown[drawIdx] && g_driverFBOBindings[drawIdx] == id);
+            const Bool readMatches =
+                !bindsRead || (g_driverFBOBindingKnown[readIdx] && g_driverFBOBindings[readIdx] == id);
+            if (drawMatches && readMatches) {
+                return;
+            }
+            g_GLESFuncs.glBindFramebuffer(fbTarget, id);
+            if (bindsDraw) {
+                g_driverFBOBindings[drawIdx] = id;
+                g_driverFBOBindingKnown[drawIdx] = true;
+            }
+            if (bindsRead) {
+                g_driverFBOBindings[readIdx] = id;
+                g_driverFBOBindingKnown[readIdx] = true;
+            }
+        }
+
+        Uint CurrentFramebufferBinding(FramebufferTarget target) {
+            const SizeT idx = SizeT(target);
+            if (!g_driverFBOBindingKnown[idx]) {
+                // Cold path: pin the shadow from the driver once (init probes and
+                // pre-shadow code bind raw but restore what they found).
+                GLint binding = 0;
+                g_GLESFuncs.glGetIntegerv(
+                    target == FramebufferTarget::Read ? GL_READ_FRAMEBUFFER_BINDING : GL_DRAW_FRAMEBUFFER_BINDING,
+                    &binding);
+                g_driverFBOBindings[idx] = static_cast<Uint>(binding);
+                g_driverFBOBindingKnown[idx] = true;
+            }
+            return g_driverFBOBindings[idx];
+        }
+
+        void InvalidateFramebufferBindingCache() {
+            g_driverFBOBindings = {0, 0};
+            g_driverFBOBindingKnown = {false, false};
         }
 
         void BackendFramebufferObject::InvalidateSyncedState() {
@@ -2376,7 +2584,13 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     if (glTextureTarget == GL_UNKNOWN_MGL) {
                         glTextureTarget = TextureImpl::ConvertTextureTargetToBackendGLEnum(textureObject->GetTarget());
                     }
-                    backendTextureObject->Bind(glTextureTarget);
+                    // glBindTexture rejects cube-face enums (INVALID_ENUM with no
+                    // bind, while Bind() would still record the cube-map cache slot
+                    // as bound): bind via the owning cube target; the attach below
+                    // keeps the face target.
+                    const Bool isCubeFace = glTextureTarget >= GL_TEXTURE_CUBE_MAP_POSITIVE_X &&
+                                            glTextureTarget <= GL_TEXTURE_CUBE_MAP_NEGATIVE_Z;
+                    backendTextureObject->Bind(isCubeFace ? GL_TEXTURE_CUBE_MAP : glTextureTarget);
                     g_GLESFuncs.glFramebufferTexture2D(glFBOTarget, glBackendAttachment, glTextureTarget,
                                                        backendTextureObject->GetBackendTextureId(),
                                                        static_cast<GLint>(attachmentObject.GetTextureLevel()));
@@ -2465,6 +2679,28 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return false;
         }
 
+        void BackendFramebufferObject::SyncReadBufferToBackend(
+            const SharedPtr<MG_State::GLState::FramebufferObject>& stateFBOObject) {
+            if (!stateFBOObject) {
+                return;
+            }
+            auto frontendReadBuf = stateFBOObject->GetReadBuffer();
+            if (frontendReadBuf == m_frontendReadBuffer) {
+                return;
+            }
+            m_frontendReadBuffer = frontendReadBuf;
+
+            GLenum glBackendReadBuffer = GetBackendAttachmentType(frontendReadBuf);
+            if (m_backendReadBuffer != glBackendReadBuffer) {
+                m_backendReadBuffer = glBackendReadBuffer;
+                // glReadBuffer targets whatever FBO is bound to GL_READ_FRAMEBUFFER. When this is
+                // reached from SyncCurrentFBO's "same FBO as draw" skip path the backend FBO was
+                // only bound as DRAW, so bind it as READ first to route the read buffer correctly.
+                Bind(FramebufferTarget::Read);
+                g_GLESFuncs.glReadBuffer(glBackendReadBuffer);
+            }
+        }
+
         void BackendFramebufferObject::SyncToBackend(
             const SharedPtr<MG_State::GLState::FramebufferObject>& stateFBOObject, FramebufferTarget asTarget) {
 #ifdef TRACY_ENABLE
@@ -2546,16 +2782,8 @@ namespace MobileGL::MG_Backend::DirectGLES {
 
             // 2. Remap read buffer. glReadBuffer writes the READ-bound FBO's state, so
             // only apply (and stamp the memo) when this object is bound as READ.
-            auto frontendReadBuf = stateFBOObject->GetReadBuffer();
-            if (frontendReadBuf != m_frontendReadBuffer && asTarget == FramebufferTarget::Read) {
-                m_frontendReadBuffer = frontendReadBuf;
-
-                GLenum glBackendReadBuffer = GetBackendAttachmentType(frontendReadBuf);
-
-                if (m_backendReadBuffer != glBackendReadBuffer) {
-                    m_backendReadBuffer = glBackendReadBuffer;
-                    g_GLESFuncs.glReadBuffer(glBackendReadBuffer);
-                }
+            if (asTarget == FramebufferTarget::Read) {
+                SyncReadBufferToBackend(stateFBOObject);
             }
 
             // -------------------- Attach texture to backend FBO -----------------------
@@ -2662,6 +2890,295 @@ namespace MobileGL::MG_Backend::DirectGLES {
             g_fboSyncedObjects = {};
     } // namespace FramebufferImpl
 
+    namespace ScratchFBOImpl {
+        namespace {
+            ScratchFramebuffer g_tempFramebuffer;
+            ScratchFramebuffer g_blitReadFramebuffer;
+            ScratchFramebuffer g_blitDrawFramebuffer;
+            Uint g_completeTinyFBOId = 0;
+            Uint g_completeTinyRBOId = 0;
+
+            // Detach every point the shadow no longer vouches for. Used when the
+            // shadow is unknown (context reset, texture id deleted while attached).
+            void ScrubAllAttachments(ScratchFramebuffer& fb, GLenum fbTarget) {
+                g_GLESFuncs.glFramebufferTexture2D(fbTarget, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, 0, 0);
+                g_GLESFuncs.glFramebufferTexture2D(fbTarget, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0, 0);
+                fb.colorTex = 0;
+                fb.colorTarget = 0;
+                fb.colorLevel = 0;
+                fb.colorLayer = -1;
+                fb.depthTex = 0;
+                fb.depthTarget = 0;
+                fb.depthLevel = 0;
+                fb.depthHasStencil = false;
+                fb.attachmentsKnown = true;
+            }
+
+            void PrepareForUse(ScratchFramebuffer& fb, GLenum fbTarget) {
+                if (!fb.attachmentsKnown) {
+                    ScrubAllAttachments(fb, fbTarget);
+                }
+            }
+
+            // The post-attach glGetError probe below must not misread an error some
+            // earlier operation left queued; drain before attaching (rare path -
+            // only runs when the attachment actually changes).
+            void DrainPendingGLErrors() {
+                while (g_GLESFuncs.glGetError() != GL_NO_ERROR) {
+                }
+            }
+
+            // Record the color point as detached when the shadow said something was
+            // there; the actual detach call is the caller's (it may be replaced by
+            // the new attach directly when the point is being overwritten).
+            void RecordNoColor(ScratchFramebuffer& fb) {
+                fb.colorTex = 0;
+                fb.colorTarget = 0;
+                fb.colorLevel = 0;
+                fb.colorLayer = -1;
+            }
+
+            void RecordNoDepth(ScratchFramebuffer& fb) {
+                fb.depthTex = 0;
+                fb.depthTarget = 0;
+                fb.depthLevel = 0;
+                fb.depthHasStencil = false;
+            }
+        } // namespace
+
+        ScratchFramebuffer& TempFramebuffer() {
+            return g_tempFramebuffer;
+        }
+        ScratchFramebuffer& BlitReadFramebuffer() {
+            return g_blitReadFramebuffer;
+        }
+        ScratchFramebuffer& BlitDrawFramebuffer() {
+            return g_blitDrawFramebuffer;
+        }
+
+        Uint EnsureId(ScratchFramebuffer& fb) {
+            if (fb.id == 0) {
+                g_GLESFuncs.glGenFramebuffers(1, &fb.id);
+                // A fresh FBO has nothing attached and COLOR_ATTACHMENT0 read/draw
+                // buffers (the ES defaults for a non-default framebuffer).
+                fb.attachmentsKnown = true;
+                RecordNoColor(fb);
+                RecordNoDepth(fb);
+                fb.readBuffer = GL_COLOR_ATTACHMENT0;
+                fb.drawBuffer = GL_COLOR_ATTACHMENT0;
+            }
+            return fb.id;
+        }
+
+        void EnsureColorAttachment2D(ScratchFramebuffer& fb, GLenum fbTarget, Uint tex, GLenum texTarget,
+                                     GLint level) {
+            PrepareForUse(fb, fbTarget);
+            if (fb.depthTex != 0) {
+                g_GLESFuncs.glFramebufferTexture2D(fbTarget, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0, 0);
+                RecordNoDepth(fb);
+            }
+            if (fb.colorTex == tex && fb.colorTarget == texTarget && fb.colorLevel == level && fb.colorLayer < 0) {
+                return;
+            }
+            if (fb.colorTex != 0) {
+                // Detach first: if the new attach fails, the point must read as
+                // missing (incomplete FBO), not silently keep the old texture.
+                g_GLESFuncs.glFramebufferTexture2D(fbTarget, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, 0, 0);
+            }
+            DrainPendingGLErrors();
+            g_GLESFuncs.glFramebufferTexture2D(fbTarget, GL_COLOR_ATTACHMENT0, texTarget, tex, level);
+            if (g_GLESFuncs.glGetError() != GL_NO_ERROR) {
+                RecordNoColor(fb);
+                return;
+            }
+            fb.colorTex = tex;
+            fb.colorTarget = texTarget;
+            fb.colorLevel = level;
+            fb.colorLayer = -1;
+        }
+
+        void EnsureColorAttachmentLayer(ScratchFramebuffer& fb, GLenum fbTarget, Uint tex, GLint level, GLint layer) {
+            PrepareForUse(fb, fbTarget);
+            if (fb.depthTex != 0) {
+                g_GLESFuncs.glFramebufferTexture2D(fbTarget, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0, 0);
+                RecordNoDepth(fb);
+            }
+            if (fb.colorTex == tex && fb.colorTarget == 0 && fb.colorLevel == level && fb.colorLayer == layer) {
+                return;
+            }
+            if (fb.colorTex != 0) {
+                g_GLESFuncs.glFramebufferTexture2D(fbTarget, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, 0, 0);
+            }
+            DrainPendingGLErrors();
+            g_GLESFuncs.glFramebufferTextureLayer(fbTarget, GL_COLOR_ATTACHMENT0, tex, level, layer);
+            if (g_GLESFuncs.glGetError() != GL_NO_ERROR) {
+                RecordNoColor(fb);
+                return;
+            }
+            fb.colorTex = tex;
+            fb.colorTarget = 0;
+            fb.colorLevel = level;
+            fb.colorLayer = layer;
+        }
+
+        void EnsureDepthAttachment2D(ScratchFramebuffer& fb, GLenum fbTarget, Uint tex, GLenum texTarget, GLint level,
+                                     Bool withStencil) {
+            PrepareForUse(fb, fbTarget);
+            if (fb.colorTex != 0) {
+                g_GLESFuncs.glFramebufferTexture2D(fbTarget, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, 0, 0);
+                RecordNoColor(fb);
+            }
+            if (fb.depthTex == tex && fb.depthTarget == texTarget && fb.depthLevel == level &&
+                fb.depthHasStencil == withStencil) {
+                return;
+            }
+            if (fb.depthTex != 0) {
+                // One call clears both depth and stencil points regardless of how
+                // the previous attachment was made.
+                g_GLESFuncs.glFramebufferTexture2D(fbTarget, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0, 0);
+            }
+            DrainPendingGLErrors();
+            g_GLESFuncs.glFramebufferTexture2D(fbTarget,
+                                               withStencil ? GL_DEPTH_STENCIL_ATTACHMENT : GL_DEPTH_ATTACHMENT,
+                                               texTarget, tex, level);
+            if (g_GLESFuncs.glGetError() != GL_NO_ERROR) {
+                RecordNoDepth(fb);
+                return;
+            }
+            fb.depthTex = tex;
+            fb.depthTarget = texTarget;
+            fb.depthLevel = level;
+            fb.depthHasStencil = withStencil;
+        }
+
+        void EnsureNoColorAttachment(ScratchFramebuffer& fb, GLenum fbTarget) {
+            PrepareForUse(fb, fbTarget);
+            if (fb.colorTex != 0) {
+                g_GLESFuncs.glFramebufferTexture2D(fbTarget, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, 0, 0);
+                RecordNoColor(fb);
+            }
+        }
+
+        void EnsureNoDepthAttachment(ScratchFramebuffer& fb, GLenum fbTarget) {
+            PrepareForUse(fb, fbTarget);
+            if (fb.depthTex != 0) {
+                g_GLESFuncs.glFramebufferTexture2D(fbTarget, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0, 0);
+                RecordNoDepth(fb);
+            }
+        }
+
+        void EnsureReadBuffer(ScratchFramebuffer& fb, GLenum readBuffer) {
+            if (fb.readBuffer == readBuffer) {
+                return;
+            }
+            g_GLESFuncs.glReadBuffer(readBuffer);
+            fb.readBuffer = readBuffer;
+        }
+
+        void EnsureDrawBuffer(ScratchFramebuffer& fb, GLenum drawBuffer) {
+            if (fb.drawBuffer == drawBuffer) {
+                return;
+            }
+            g_GLESFuncs.glDrawBuffers(1, &drawBuffer);
+            fb.drawBuffer = drawBuffer;
+        }
+
+        Uint EnsureCompleteTinyFramebufferId() {
+            if (g_completeTinyFBOId != 0) {
+                return g_completeTinyFBOId;
+            }
+            // One-time creation: the renderbuffer binding is context state with no
+            // shadow, so save/restore it by query here (cold path only).
+            GLint prevRenderbuffer = 0;
+            g_GLESFuncs.glGetIntegerv(GL_RENDERBUFFER_BINDING, &prevRenderbuffer);
+            g_GLESFuncs.glGenFramebuffers(1, &g_completeTinyFBOId);
+            g_GLESFuncs.glGenRenderbuffers(1, &g_completeTinyRBOId);
+            FramebufferImpl::BindFramebufferId(GL_FRAMEBUFFER, g_completeTinyFBOId);
+            g_GLESFuncs.glBindRenderbuffer(GL_RENDERBUFFER, g_completeTinyRBOId);
+            g_GLESFuncs.glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, 1, 1);
+            g_GLESFuncs.glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER,
+                                                  g_completeTinyRBOId);
+            const GLenum drawBuffer = GL_COLOR_ATTACHMENT0;
+            g_GLESFuncs.glDrawBuffers(1, &drawBuffer);
+            g_GLESFuncs.glReadBuffer(GL_COLOR_ATTACHMENT0);
+            MOBILEGL_ASSERT(g_GLESFuncs.glCheckFramebufferStatus(GL_FRAMEBUFFER) == GL_FRAMEBUFFER_COMPLETE,
+                            "Scratch 1x1 framebuffer is incomplete.");
+            g_GLESFuncs.glBindRenderbuffer(GL_RENDERBUFFER, static_cast<Uint>(prevRenderbuffer));
+            return g_completeTinyFBOId;
+        }
+
+        void NoteTextureIdDeleted(Uint textureId) {
+            if (textureId == 0) {
+                return;
+            }
+            for (ScratchFramebuffer* fb : {&g_tempFramebuffer, &g_blitReadFramebuffer, &g_blitDrawFramebuffer}) {
+                if (fb->colorTex == textureId || fb->depthTex == textureId) {
+                    fb->attachmentsKnown = false;
+                }
+            }
+        }
+
+        void OnBackendContextDestroyed() {
+            g_tempFramebuffer = {};
+            g_blitReadFramebuffer = {};
+            g_blitDrawFramebuffer = {};
+            g_completeTinyFBOId = 0;
+            g_completeTinyRBOId = 0;
+        }
+    } // namespace ScratchFBOImpl
+
+    namespace PixelStoreImpl {
+        namespace {
+            PackState g_packState;
+            Bool g_packStateKnown = false;
+
+            void PinPackState(const PackState& value) {
+                g_GLESFuncs.glPixelStorei(GL_PACK_ALIGNMENT, value.Alignment);
+                g_GLESFuncs.glPixelStorei(GL_PACK_ROW_LENGTH, value.RowLength);
+                g_GLESFuncs.glPixelStorei(GL_PACK_SKIP_ROWS, value.SkipRows);
+                g_GLESFuncs.glPixelStorei(GL_PACK_SKIP_PIXELS, value.SkipPixels);
+                g_packState = value;
+                g_packStateKnown = true;
+            }
+        } // namespace
+
+        void ApplyPackState(const PackState& desired) {
+            if (!g_packStateKnown) {
+                PinPackState(desired);
+                return;
+            }
+            if (desired.Alignment != g_packState.Alignment) {
+                g_GLESFuncs.glPixelStorei(GL_PACK_ALIGNMENT, desired.Alignment);
+                g_packState.Alignment = desired.Alignment;
+            }
+            if (desired.RowLength != g_packState.RowLength) {
+                g_GLESFuncs.glPixelStorei(GL_PACK_ROW_LENGTH, desired.RowLength);
+                g_packState.RowLength = desired.RowLength;
+            }
+            if (desired.SkipRows != g_packState.SkipRows) {
+                g_GLESFuncs.glPixelStorei(GL_PACK_SKIP_ROWS, desired.SkipRows);
+                g_packState.SkipRows = desired.SkipRows;
+            }
+            if (desired.SkipPixels != g_packState.SkipPixels) {
+                g_GLESFuncs.glPixelStorei(GL_PACK_SKIP_PIXELS, desired.SkipPixels);
+                g_packState.SkipPixels = desired.SkipPixels;
+            }
+        }
+
+        PackState CurrentPackState() {
+            if (!g_packStateKnown) {
+                // Fresh/unknown context: pin to the GL defaults (what a new context
+                // starts with; writing them makes the shadow authoritative either way).
+                PinPackState(PackState{});
+            }
+            return g_packState;
+        }
+
+        void InvalidatePackStateCache() {
+            g_packStateKnown = false;
+        }
+    } // namespace PixelStoreImpl
+
     namespace PrgramImpl {
         Uint32 g_snormFallbackClampOutputMask = 0;
         Uint32 g_unormFallbackClampOutputMask = 0;
@@ -2784,6 +3301,21 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     effectiveSpirv = &uboPrecisionSpirv;
                 }
 
+                // noperspective is core desktop GLSL and reaches here as the SPIR-V NoPerspective
+                // decoration. SPIRV-Cross renders it as ESSL `noperspective` + `#extension
+                // GL_NV_shader_noperspective_interpolation : require`; a driver without that extension
+                // rejects the require. So on such devices emulate screen-linear interpolation instead
+                // (pre-multiply outputs by gl_Position.w, recover inputs via gl_FragCoord.w) and drop
+                // the decoration - exact, extension-free. Devices that have the extension keep the
+                // decoration and let the hardware do it natively.
+                Vector<unsigned int> noperspectiveSpirv;
+                if (!g_GLESCapabilities.SupportsNoperspectiveInterpolation &&
+                    MG_Util::ShaderTranspiler::ShaderCompiler::EmulateNoPerspectiveForEssl(
+                        *effectiveSpirv, noperspectiveSpirv) &&
+                    !noperspectiveSpirv.empty()) {
+                    effectiveSpirv = &noperspectiveSpirv;
+                }
+
                 MG_Util::ShaderTranspiler::SpvcSession spvcSession(*effectiveSpirv,
                     MG_Util::ShaderTranspiler::SessionUsageBit::Transpile);
 

MobileGL/MG_Backend/DirectGLES/Managers.h

@@ -178,6 +178,21 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // glBindBuffer with a redundant-bind cache for GL_ARRAY_BUFFER.
         void BindBufferId(GLenum target, Uint id);
         void InvalidateArrayBufferBindingCache();
+        // Redundant-bind caches for the driver-level GL_PIXEL_PACK/UNPACK_BUFFER
+        // bindings. Every backend readback (glReadPixels / pack-PBO map) and pixel
+        // upload site routes its binding through these so the shadow always matches
+        // the driver; the resting state between operations is 0, which keeps any
+        // path that implicitly assumes "no PBO bound" correct. Scrubbed when a
+        // buffer id is deleted/pooled (GL resets a deleted buffer's bindings to 0,
+        // and a recycled name matching the shadow would false-skip the rebind) and
+        // invalidated on MakeCurrent (context may reset).
+        void BindPixelPackBufferId(Uint id);
+        void BindPixelUnpackBufferId(Uint id);
+        void InvalidatePixelBufferBindingCaches();
+        // A GL buffer id is being deleted by code outside BufferImpl (e.g. the VAO
+        // client-attribute staging buffers): scrub every buffer-binding shadow that
+        // could false-skip when the name is recycled.
+        void NoteBufferIdDeleted(Uint id);
         // Redundant-bind cache for INDEXED buffer bindings (glBindBufferBase/Range on
         // GL_UNIFORM_BUFFER / GL_SHADER_STORAGE_BUFFER): skips the GL call when the
         // (id, range) already at that index matches, like the array-buffer/texture/
@@ -332,6 +347,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
         class BackendTextureObject {
         public:
             BackendTextureObject();
+            // Deletes the GL texture (frontend glDeleteTextures used to leak every
+            // backend id for the context lifetime) and scrubs the binding/scratch-FBO
+            // shadows so a recycled name or heap address cannot false-skip a rebind.
+            ~BackendTextureObject();
+            BackendTextureObject(const BackendTextureObject&) = delete;
+            BackendTextureObject& operator=(const BackendTextureObject&) = delete;
             void SyncMipmapsToBackend(const SharedPtr<MG_State::GLState::ITextureObject>& stateTextureObject);
             void SyncBuiltinSamplerToBackend(const SharedPtr<MG_State::GLState::ITextureObject>& stateTextureObject);
             void SyncTextureParamsToBackend(const SharedPtr<MG_State::GLState::ITextureObject>& stateTextureObject);
@@ -343,6 +364,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
             void RecreateBackendTexture();
 
             Uint m_backendTextureId = 0;
+            // ES context generation the id was created under; a dtor running after
+            // that context died must not delete a foreign (recycled) name.
+            Uint m_contextGeneration = 0;
             Bool m_isInitialized = false;
             Bool m_imageBindableStorageRequired = false;
             Bool m_backendStorageImmutable = false;
@@ -367,6 +391,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
                      MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS>
             g_boundTexturesCache;
         extern Uint g_activeTextureUnit;
+        // Bumped when the backend ES context is destroyed; texture ids stamped with
+        // an older generation belong to a dead context and must not be deleted.
+        extern Uint g_textureContextGeneration;
     } // namespace TextureImpl
 
     namespace FramebufferImpl {
@@ -375,6 +402,10 @@ namespace MobileGL::MG_Backend::DirectGLES {
             BackendFramebufferObject();
             void SyncToBackend(const SharedPtr<MG_State::GLState::FramebufferObject>& stateFBOObject,
                                FramebufferTarget asTarget);
+            // Apply only this FBO's read buffer (glReadBuffer) to the backend. Split out so it can
+            // still run when SyncCurrentFBO skips the READ-target sync because the same GL FBO is
+            // bound as both draw and read (otherwise glReadBuffer changes would be silently dropped).
+            void SyncReadBufferToBackend(const SharedPtr<MG_State::GLState::FramebufferObject>& stateFBOObject);
             void InvalidateSyncedState();
             Uint GetBackendFramebufferId() const { return m_backendFBOId; }
             void Bind(FramebufferTarget target) const;
@@ -414,8 +445,99 @@ namespace MobileGL::MG_Backend::DirectGLES {
         extern Array<Uint16, SizeT(FramebufferTarget::FramebufferTargetCount)> g_fboSyncedObjectVersions;
         extern Array<MG_State::GLState::FramebufferObject*, SizeT(FramebufferTarget::FramebufferTargetCount)>
             g_fboSyncedObjects;
+
+        // Driver-level READ/DRAW framebuffer-binding shadow. Every backend
+        // glBindFramebuffer routes through BindFramebufferId so scoped helpers can
+        // save/restore the current binding without a glGetIntegerv round-trip (that
+        // query forces a driver pipeline sync) and so redundant rebinds no-op.
+        // Starts unknown; the first CurrentFramebufferBinding() query pins it from
+        // the driver once. Invalidated on MakeCurrent (context may reset).
+        // GL_FRAMEBUFFER binds both targets.
+        void BindFramebufferId(GLenum fbTarget, Uint id);
+        Uint CurrentFramebufferBinding(FramebufferTarget target);
+        void InvalidateFramebufferBindingCache();
     } // namespace FramebufferImpl
 
+    // Shared scratch framebuffers for the readback/copy/blit emulation paths, with a
+    // driver-side attachment shadow: repeated uses skip redundant detach/attach GL
+    // calls, and an attachment left by one use (e.g. a depth copy's DEPTH_STENCIL
+    // texture) is detached exactly when a later use of another aspect would
+    // otherwise inherit it (stale cross-aspect attachments made the shared temp FBO
+    // incomplete and silently degraded later readbacks).
+    namespace ScratchFBOImpl {
+        struct ScratchFramebuffer {
+            Uint id = 0;
+            // false => attachment state unknown; scrub every point on next use.
+            // A fresh FBO starts with nothing attached, so creation sets it true.
+            Bool attachmentsKnown = false;
+            Uint colorTex = 0;
+            GLenum colorTarget = 0;
+            GLint colorLevel = 0;
+            GLint colorLayer = -1; // >= 0 => attached via glFramebufferTextureLayer
+            Uint depthTex = 0;
+            GLenum depthTarget = 0;
+            GLint depthLevel = 0;
+            Bool depthHasStencil = false;
+            // Per-FBO read/draw buffer state (0 = unknown, set on first use).
+            GLenum readBuffer = 0;
+            GLenum drawBuffer = 0;
+        };
+        ScratchFramebuffer& TempFramebuffer();     // GetTexImage READ / CopyTex*Image2D depth DRAW
+        ScratchFramebuffer& BlitReadFramebuffer(); // texture-to-texture blit source
+        ScratchFramebuffer& BlitDrawFramebuffer(); // texture-to-texture blit destination
+        // Returns the GL id, generating it if needed (requires a current ES context).
+        Uint EnsureId(ScratchFramebuffer& fb);
+        // The fb must currently be bound at fbTarget (glReadBuffer/glDrawBuffers
+        // target the READ/DRAW binding respectively). Each Ensure* performs the
+        // minimal detach/attach set and keeps the shadow in sync; a failed attach
+        // records the point as detached so the completeness check fails instead of
+        // silently reading a stale attachment.
+        void EnsureColorAttachment2D(ScratchFramebuffer& fb, GLenum fbTarget, Uint tex, GLenum texTarget, GLint level);
+        void EnsureColorAttachmentLayer(ScratchFramebuffer& fb, GLenum fbTarget, Uint tex, GLint level, GLint layer);
+        void EnsureDepthAttachment2D(ScratchFramebuffer& fb, GLenum fbTarget, Uint tex, GLenum texTarget, GLint level,
+                                     Bool withStencil);
+        void EnsureNoColorAttachment(ScratchFramebuffer& fb, GLenum fbTarget);
+        void EnsureNoDepthAttachment(ScratchFramebuffer& fb, GLenum fbTarget);
+        void EnsureReadBuffer(ScratchFramebuffer& fb, GLenum readBuffer);
+        void EnsureDrawBuffer(ScratchFramebuffer& fb, GLenum drawBuffer);
+        // A 1x1 RGBA8-renderbuffer-complete FBO (GenerateMipmap needs a complete
+        // binding while respecifying texture storage). Attachment is set once at
+        // creation and never changes.
+        Uint EnsureCompleteTinyFramebufferId();
+        // A backend texture id is being deleted or respecified: a scratch FBO still
+        // referencing it would hold a dangling attachment (ES only auto-detaches
+        // from the *bound* framebuffer), and a recycled name could false-skip a
+        // re-attach; force a full scrub on next use.
+        void NoteTextureIdDeleted(Uint textureId);
+        // The ES context (and the scratch FBO ids with it) is going away.
+        void OnBackendContextDestroyed();
+    } // namespace ScratchFBOImpl
+
+    // Driver-level GL_PACK_* pixel-store shadow, the readback-side sibling of the
+    // upload path's ScopedDefaultUnpackState (Managers.cpp): the backend PACK state
+    // is written ONLY through ApplyPackState, so scoped helpers can save/restore it
+    // from the shadow instead of glGetIntegerv (which forces a driver pipeline
+    // sync), and redundant glPixelStorei calls no-op. The first Apply/Current call
+    // pins the driver to the shadow by writing all fields once. Invalidated on
+    // MakeCurrent (context may reset). PACK_IMAGE_HEIGHT/SKIP_IMAGES/SWAP_BYTES/
+    // LSB_FIRST have no ES equivalents; readbacks honor them on the CPU from the
+    // frontend context state instead.
+    namespace PixelStoreImpl {
+        struct PackState {
+            GLint Alignment = 4;
+            GLint RowLength = 0;
+            GLint SkipRows = 0;
+            GLint SkipPixels = 0;
+            Bool operator==(const PackState& o) const {
+                return Alignment == o.Alignment && RowLength == o.RowLength && SkipRows == o.SkipRows &&
+                       SkipPixels == o.SkipPixels;
+            }
+        };
+        void ApplyPackState(const PackState& desired);
+        PackState CurrentPackState();
+        void InvalidatePackStateCache();
+    } // namespace PixelStoreImpl
+
     // Image uniforms take their unit from the layout(binding=N) qualifier baked into
     // the transpiled ESSL; unlike samplers they must not (and in ES cannot) be
     // assigned through glUniform1i.

MobileGL/MG_Backend/DirectGLES/Utils.cpp

@@ -764,5 +764,95 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 }
             }
         }
+
+        static SizeT AlignReadbackRow(SizeT rowBytes, Int alignment) {
+            const SizeT align = alignment > 0 ? static_cast<SizeT>(alignment) : 1;
+            return (rowBytes + align - 1) / align * align;
+        }
+
+    // Repacks wide RGBA(_INTEGER) rows into the client's (format, type) layout, honoring the
+        // client-side PACK parameters and the bound pixel-pack buffer. `wide` holds
+        // `sliceHeight * sliceCount` rows of `width` texels (slice-major, tightly stacked),
+        // 4 components x GetReadbackComponentSize(wideType) bytes each.
+        // applyPackImageParams: GL_PACK_IMAGE_HEIGHT / GL_PACK_SKIP_IMAGES apply only to GetTexImage
+        // of 3D/array images; ReadPixels and 2D GetTexImage ignore them (GL 3.3 sections 4.3.1, 6.1.4).
+        // Per the GL addressing rules, slice k row j lands at
+        // SKIP_IMAGES*imageStride + SKIP_ROWS*rowStride + SKIP_PIXELS*pixelBytes
+        //   + k*imageStride + j*rowStride, with imageStride = max(IMAGE_HEIGHT, sliceHeight)*rowStride.
+        Bool StoreWideRowsToClient(const Uint8* wide, GLenum wideType, GLsizei width, GLsizei sliceHeight,
+                                   GLsizei sliceCount, const ReadbackChannelMapping& mapping, GLenum type,
+                                   void* pixels, Bool applyPackImageParams) {
+        const SizeT dstPixelBytes = GetReadbackDstPixelSize(mapping, type);
+        if (dstPixelBytes == 0) {
+            return false;
+        }
+        PackedReadbackLayout packedLayout{};
+        const Bool isPackedType = GetPackedReadbackLayout(type, packedLayout);
+        const SizeT dstComponentSize = GetReadbackComponentSize(type);
+
+        const auto& pixelPackBufferObject =
+            MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::PixelPack).GetBoundObject();
+
+        // Destination layout is computed from the client-side PACK parameters; only the actual pixel
+        // rows are written so skip regions of the destination stay untouched.
+        const auto packParams = MG_State::pGLContext->GetPixelStoreParameters(false);
+        const SizeT rowPixels = static_cast<SizeT>(packParams.RowLength > 0 ? packParams.RowLength : width);
+        const SizeT dstRowStride = AlignReadbackRow(rowPixels * dstPixelBytes, packParams.Alignment);
+        const SizeT imageRows =
+            applyPackImageParams && packParams.ImageHeight > 0
+                ? static_cast<SizeT>(packParams.ImageHeight)
+                : static_cast<SizeT>(sliceHeight);
+        const SizeT dstImageStride = imageRows * dstRowStride;
+        const SizeT skipImages =
+            applyPackImageParams ? static_cast<SizeT>(std::max(packParams.SkipImages, 0)) : SizeT{0};
+        const SizeT dstSkipOffset = skipImages * dstImageStride +
+                                    static_cast<SizeT>(std::max(packParams.SkipRows, 0)) * dstRowStride +
+                                    static_cast<SizeT>(std::max(packParams.SkipPixels, 0)) * dstPixelBytes;
+        const SizeT dstRowBytes = static_cast<SizeT>(width) * dstPixelBytes;
+
+        const SizeT pboBaseOffset = reinterpret_cast<SizeT>(pixels); // with a PBO, `pixels` is an offset
+        if (pixelPackBufferObject) {
+            const SizeT requiredSize = pboBaseOffset + dstSkipOffset +
+                                    static_cast<SizeT>(sliceCount - 1) * dstImageStride +
+                                    static_cast<SizeT>(sliceHeight - 1) * dstRowStride + dstRowBytes;
+            if (requiredSize > pixelPackBufferObject->GetSize()) {
+                MGLOG_E("Readback conversion: pixel pack buffer is too small");
+                return true;
+            }
+        }
+
+        const SizeT srcComponentSize = GetReadbackComponentSize(wideType);
+        const SizeT srcPixelBytes = 4 * srcComponentSize;
+        Vector<Uint8> convertedRow(dstRowBytes);
+
+        for (GLsizei slice = 0; slice < sliceCount; ++slice) {
+            for (GLsizei row = 0; row < sliceHeight; ++row) {
+                const SizeT flatRow = static_cast<SizeT>(slice) * static_cast<SizeT>(sliceHeight) +
+                                   static_cast<SizeT>(row);
+                const Uint8* srcRow = wide + flatRow * static_cast<SizeT>(width) * srcPixelBytes;
+                ConvertWideReadbackRow(srcRow, convertedRow.data(), static_cast<SizeT>(width), wideType,
+                                                  mapping, type);
+
+                if (packParams.SwapBytes) {
+                    const SizeT groupSize = isPackedType ? packedLayout.byteSize : dstComponentSize;
+                    if (groupSize > 1) {
+                        for (SizeT offset = 0; offset + groupSize <= dstRowBytes; offset += groupSize) {
+                            std::reverse(convertedRow.data() + offset, convertedRow.data() + offset + groupSize);
+                        }
+                    }
+                }
+
+                const SizeT dstOffset = dstSkipOffset + static_cast<SizeT>(slice) * dstImageStride +
+                                     static_cast<SizeT>(row) * dstRowStride;
+                if (pixelPackBufferObject) {
+                    pixelPackBufferObject->WritebackFromBackend({convertedRow.data(), dstRowBytes},
+                                                             pboBaseOffset + dstOffset);
+                } else {
+                    Memcpy(static_cast<Uint8*>(pixels) + dstOffset, convertedRow.data(), dstRowBytes);
+                }
+            }
+        }
+            return true;
+        }
     } // namespace ReadbackImpl
 } // namespace MobileGL::MG_Backend::DirectGLES

MobileGL/MG_Backend/DirectGLES/Utils.h

@@ -88,6 +88,14 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // bytes, dst receives width * GetReadbackDstPixelSize(mapping, type) bytes.
         void ConvertWideReadbackRow(const Uint8* src, Uint8* dst, SizeT width, GLenum wideType,
                                     const ReadbackChannelMapping& mapping, GLenum type);
+
+        // Stores wide RGBA(_INTEGER) rows into the client pointer or the bound PACK pixel buffer,
+        // honoring the client-side PACK pixel-store parameters (row length, alignment, skips,
+        // swap-bytes, and - when applyPackImageParams - image height/skip images). Shared by the
+        // DirectGLES and DirectVulkan readback conversion paths.
+        Bool StoreWideRowsToClient(const Uint8* wide, GLenum wideType, GLsizei width, GLsizei sliceHeight,
+                                   GLsizei sliceCount, const ReadbackChannelMapping& mapping, GLenum type,
+                                   void* pixels, Bool applyPackImageParams);
     } // namespace ReadbackImpl
 
     namespace PrgramImpl {

MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.cpp

@@ -140,6 +140,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             case TextureInternalFormat::RGB:
             case TextureInternalFormat::RGB8:
                 return TextureInternalFormat::RGBA8;
+            // Legacy low-bit-depth formats with no (or rarely supported) native Vulkan
+            // encoding; a wider normalized fallback keeps at least the required precision.
+            case TextureInternalFormat::R3G3B2:
+            case TextureInternalFormat::RGB4:
+            case TextureInternalFormat::RGB5:
+            case TextureInternalFormat::RGBA2:
+            case TextureInternalFormat::RGBA4:
+            case TextureInternalFormat::RGB5A1:
+                return TextureInternalFormat::RGBA8;
+            case TextureInternalFormat::RGB10:
+                return TextureInternalFormat::RGB10A2;
+            case TextureInternalFormat::RGB12:
+            case TextureInternalFormat::RGBA12:
+                return TextureInternalFormat::RGBA16;
             case TextureInternalFormat::SRGB8:
                 return TextureInternalFormat::SRGB8Alpha8;
             case TextureInternalFormat::RGB8Snorm:
@@ -397,8 +411,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         if (!handle.Handle || (handle.Backend != WindowBackend::Android &&
                                handle.Backend != WindowBackend::X11 &&
-                               handle.Backend != WindowBackend::MetalLayer)) {
-            MGLOG_E("DirectVulkan backend only supports Android, X11, and CAMetalLayer native windows");
+                               handle.Backend != WindowBackend::MetalLayer &&
+                               handle.Backend != WindowBackend::Win32)) {
+            MGLOG_E("DirectVulkan backend only supports Android, X11, CAMetalLayer, and Win32 native windows");
             return false;
         }
 
@@ -454,6 +469,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // treat them as signaled/available with zero results from here on.
         BumpRendererGeneration();
         pVulkanRenderer.reset();
+        // The reflection cache is file-scope, not renderer-owned; without this the
+        // deleted programs' reflection strings survive full context teardown.
+        ClearProgramResourceCaches();
         BackendObject::ReleaseEGLResources();
     }
 
@@ -463,6 +481,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // treat them as signaled/available with zero results from here on.
         BumpRendererGeneration();
         pVulkanRenderer.reset();
+        // The reflection cache is file-scope, not renderer-owned; without this the
+        // deleted programs' reflection strings survive full context teardown.
+        ClearProgramResourceCaches();
     }
 
     const RendererInfo& BackendObject_DirectVulkan::GetRendererInfo() const {
@@ -794,5 +815,32 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     m_vulkanCaps.MaxShaderStorageBlockSize,
                     m_dynamicParameters.MaxShaderStorageBlockSize);
         }
+        switch (m_vulkanCaps.VendorId) {
+        case 0x5143u: // VK_VENDOR_ID: Qualcomm
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Qualcomm;
+            break;
+        case 0x13B5u: // ARM
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Arm;
+            break;
+        case 0x10DEu: // NVIDIA
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Nvidia;
+            break;
+        case 0x1002u: // AMD
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Amd;
+            break;
+        case 0x8086u: // Intel
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Intel;
+            break;
+        case 0x1010u: // Imagination
+            m_dynamicParameters.GpuVendor = GpuVendorKind::ImgTec;
+            break;
+        case 0x10005u: // Mesa software (lavapipe)
+        case 0x1AE0u:  // Google (SwiftShader)
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Software;
+            break;
+        default:
+            m_dynamicParameters.GpuVendor = GpuVendorKind::Unknown;
+            break;
+        }
     }
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/DirectVulkan.cpp

@@ -20,7 +20,8 @@
 #include <spirv_reflect.h>
 
 namespace MobileGL::MG_Backend::DirectVulkan {
-    UniquePtr<VulkanRenderer> pVulkanRenderer = nullptr;
+    // Leak-at-exit storage; see GlobalObjects.cpp.
+    UniquePtr<VulkanRenderer>& pVulkanRenderer = *new UniquePtr<VulkanRenderer>();
 
     namespace {
         // Generation of the live VulkanRenderer instance, mirroring
@@ -60,6 +61,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         };
 
         struct ProgramResourceCache {
+            // Lifetime id of the program the cached reflection belongs to. GL names are
+            // recycled (IndexGenerator hands freed indices straight back), and a
+            // recreated program's backendStateVersion restarts at the same small values,
+            // so the version alone can collide; the never-reused lifetime id makes the
+            // slot's ownership unambiguous.
+            Uint64 programLifetimeId = 0;
             Uint32 backendStateVersion = 0;
             Vector<StorageBlockResource> storageBlocks;
             Vector<BufferVariableResource> bufferVariables;
@@ -81,6 +88,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint32 baseInstance = 0;
         };
 
+        // Keyed by GL program name so the freed-name reuse in IndexGenerator bounds the
+        // map at the peak-simultaneous-program high-water mark; each slot's ownership is
+        // checked against the program's lifetime id before it is served (see
+        // GetProgramResourceCache). Cleared wholesale at EGL teardown via
+        // ClearProgramResourceCaches.
         UnorderedMap<GLuint, ProgramResourceCache> g_programResourceCaches;
 
         void ClearReadPixelsOutput(GLsizei width, GLsizei height, GLenum format, GLenum type, void* pixels) {
@@ -141,13 +153,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         ProgramResourceCache& GetProgramResourceCache(const MG_State::GLState::ProgramObject& program) {
             auto& cache = g_programResourceCaches[program.GetExternalIndex()];
+            const Uint64 programLifetimeId = program.GetLifetimeId();
             const Uint32 backendStateVersion = program.GetBackendStateVersion();
-            if (cache.backendStateVersion == backendStateVersion &&
+            // The lifetime id must match too: a new program that reuses a deleted
+            // program's name and happens to land on the same backendStateVersion (both
+            // count from zero) would otherwise be served the dead program's reflection.
+            if (cache.programLifetimeId == programLifetimeId &&
+                cache.backendStateVersion == backendStateVersion &&
                 (!cache.storageBlocks.empty() || !cache.bufferVariables.empty())) {
                 return cache;
             }
 
             cache = {};
+            cache.programLifetimeId = programLifetimeId;
             cache.backendStateVersion = backendStateVersion;
 
             Vector<SpvReflectShaderModule> modules;
@@ -365,6 +383,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     } // namespace
 
+    void ClearProgramResourceCaches() {
+        // Called from EGL teardown while the backend's m_eglStateMutex is held; GL
+        // calls are serialized in this codebase (contexts migrate threads but never
+        // run concurrently), so no other thread can be inside the unsynchronized map.
+        // Live programs in another context self-heal: their entry rebuilds from the
+        // retained generated SPIR-V on the next resource query.
+        g_programResourceCaches.clear();
+    }
+
     GLuint GetShaderStorageBlockIndex(const MG_State::GLState::ProgramObject& program, const String& name) {
         auto& cache = GetProgramResourceCache(program);
         const auto it = std::find_if(cache.storageBlocks.begin(), cache.storageBlocks.end(),

MobileGL/MG_Backend/DirectVulkan/DirectVulkan.h

@@ -12,7 +12,7 @@
 #include "Renderer/VulkanRenderer.h"
 
 namespace MobileGL::MG_Backend::DirectVulkan {
-    extern UniquePtr<VulkanRenderer> pVulkanRenderer;
+    extern UniquePtr<VulkanRenderer>& pVulkanRenderer;
 
     // Generation of the live VulkanRenderer instance, mirroring DirectGLES's
     // g_syncContextGeneration. BackendObject_DirectVulkan bumps it wherever
@@ -23,6 +23,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     Uint64 GetRendererGeneration();
     void BumpRendererGeneration();
 
+    // Drops every cached program-resource reflection entry (CPU-side strings/vectors
+    // only, no Vulkan handles). Called at EGL teardown next to the renderer reset;
+    // safe because GL calls are serialized in this codebase, and any still-live
+    // program rebuilds its entry from the retained generated SPIR-V on demand.
+    void ClearProgramResourceCaches();
+
     void ClearBufferfi(GLenum buffer, GLint drawbuffer, GLfloat depth, GLint stencil);
     void ClearBufferfv(GLenum buffer, GLint drawbuffer, const GLfloat* value);
     void ClearBufferuiv(GLenum buffer, GLint drawbuffer, const GLuint* value);

MobileGL/MG_Backend/DirectVulkan/Renderer/FrameContext.cpp

@@ -150,12 +150,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     Bool FrameContext::TransitionToPresent(VkImage image, VkImageLayout oldLayout, VkImageLayout presentLayout) {
         auto& frame = GetCurrent();
-        if (frame.hasCommandBufferRecorded || frame.isCommandRecording || oldLayout == presentLayout ||
-            oldLayout == VK_IMAGE_LAYOUT_SHARED_PRESENT_KHR) {
+        if (oldLayout == presentLayout || oldLayout == VK_IMAGE_LAYOUT_SHARED_PRESENT_KHR) {
             return false;
         }
 
-        auto& commandBuffer = BeginCommandRecording();
+        // The barrier belongs in the frame's own recording. Bailing out because
+        // something was already recorded (the previous behaviour) dropped the
+        // transition entirely for every frame that never ran a default-framebuffer
+        // render pass - the only other thing that carries the image to
+        // PRESENT_SRC_KHR, via that pass's finalLayout - so the swapchain image was
+        // handed to the WSI still in the layout it was acquired in.
+        // A closed-but-unsubmitted buffer can only come from a submit that already
+        // failed (SubmitPendingCommandBuffer leaves the flag set on error), and
+        // appending to it is illegal while reopening would reset the frame's own
+        // commands away. The device is gone on that path anyway - stay silent-safe
+        // rather than trade a lost device for a barrier into a closed buffer.
+        if (frame.hasCommandBufferRecorded) {
+            MGLOG_E("TransitionToPresent: command buffer already closed; skipping the present barrier");
+            return false;
+        }
+
+        // Reopening a recording here would vkResetCommandBuffer this frame's own
+        // commands away, so append to the open one and let the caller close it.
+        const Bool openedRecording = !frame.isCommandRecording;
+        VkCommandBuffer commandBuffer = openedRecording ? BeginCommandRecording() : frame.commandBuffer;
 
         VkImageMemoryBarrier presentBarrier{};
         presentBarrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
@@ -174,7 +192,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         vkCmdPipelineBarrier(commandBuffer, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, 0, 0,
                              nullptr, 0, nullptr, 1, &presentBarrier);
 
-        EndCommandRecording();
+        if (openedRecording) {
+            EndCommandRecording();
+        }
         return true;
     }
 
@@ -227,12 +247,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         result = vkAcquireNextImageKHR(device, swapchain, timeout, frame.imageAvailableSemaphore, acquireFence,
                                        &outImageIndex);
-        if (result != VK_SUCCESS) {
+        // VK_SUBOPTIMAL_KHR is a success code: an image *was* acquired and
+        // imageAvailableSemaphore *will* be signaled. Bailing out on it skipped both
+        // the consumed-flag reset (leaving a stale "already consumed", so the next
+        // submit never waited on the pending signal) and the fence reset (leaving
+        // the slot's fence signaled for the next submit to reuse). Only a genuine
+        // failure - VK_ERROR_OUT_OF_DATE_KHR and friends, where nothing is acquired
+        // and nothing is signaled - skips the bookkeeping.
+        if (result != VK_SUCCESS && result != VK_SUBOPTIMAL_KHR) {
             return result;
         }
 
         frame.imageAvailableSemaphoreConsumed = false;
-        return vkResetFences(device, 1, &frame.imageInFlightFence);
+        const VkResult resetResult = vkResetFences(device, 1, &frame.imageInFlightFence);
+        // Hand the acquire's own code back so the caller can schedule a rebuild.
+        return resetResult == VK_SUCCESS ? result : resetResult;
     }
 
     Uint32 FrameContext::GetCurrentFrameIndex() const {
@@ -264,7 +293,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (result != VK_SUCCESS) {
             return result;
         }
-        frame.retiredCommandBuffers.push_back(frame.commandBuffer);
+        // lastSubmitIndex was just written by the renderer for the submission
+        // that carried this command buffer.
+        frame.retiredCommandBuffers.push_back({frame.commandBuffer, frame.lastSubmitIndex});
         frame.commandBuffer = replacement;
         return VK_SUCCESS;
     }
@@ -274,12 +305,40 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return;
         }
         if (m_device != VK_NULL_HANDLE && m_commandPool != VK_NULL_HANDLE) {
-            vkFreeCommandBuffers(m_device, m_commandPool, static_cast<Uint32>(frame.retiredCommandBuffers.size()),
-                                 frame.retiredCommandBuffers.data());
+            for (const auto& retired : frame.retiredCommandBuffers) {
+                vkFreeCommandBuffers(m_device, m_commandPool, 1, &retired.commandBuffer);
+            }
         }
         frame.retiredCommandBuffers.clear();
     }
 
+    void FrameContext::FreeRetiredCommandBuffersCompletedUpTo(Uint64 completedSubmitIndex) {
+        if (m_device == VK_NULL_HANDLE || m_commandPool == VK_NULL_HANDLE) {
+            return;
+        }
+        for (auto& frame : m_frames) {
+            // Retired buffers are appended in submit order, so the completed
+            // ones form a prefix.
+            SizeT completedCount = 0;
+            while (completedCount < frame.retiredCommandBuffers.size() &&
+                   frame.retiredCommandBuffers[completedCount].submitIndex <= completedSubmitIndex) {
+                vkFreeCommandBuffers(m_device, m_commandPool, 1,
+                                     &frame.retiredCommandBuffers[completedCount].commandBuffer);
+                ++completedCount;
+            }
+            if (completedCount > 0) {
+                frame.retiredCommandBuffers.erase(frame.retiredCommandBuffers.begin(),
+                                                  frame.retiredCommandBuffers.begin() + completedCount);
+            }
+        }
+    }
+
+    void FrameContext::FreeAllRetiredCommandBuffers() {
+        for (auto& frame : m_frames) {
+            FreeRetiredCommandBuffers(frame);
+        }
+    }
+
     void FrameContext::AssertValidFrameIndex(Uint32 frameIndex) const {
         MOBILEGL_ASSERT(frameIndex < m_frames.size(), "FrameContext index out of range");
     }

MobileGL/MG_Backend/DirectVulkan/Renderer/FrameContext.h

@@ -40,6 +40,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkPresentInfoKHR presentInfo{VK_STRUCTURE_TYPE_PRESENT_INFO_KHR};
         };
 
+        // A command buffer submitted mid-frame (FlushPendingCommands), tagged
+        // with the submit-tracker index it was submitted under so it can be
+        // freed as soon as that submission is observed complete - without
+        // waiting for the slot's fence to be waited again (present-less flush
+        // loops never wait it).
+        struct RetiredCommandBuffer {
+            VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
+            Uint64 submitIndex = 0;
+        };
+
         struct FrameData {
             VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
             VkSemaphore imageAvailableSemaphore = VK_NULL_HANDLE;
@@ -47,10 +57,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Bool isCommandRecording = false;
             Bool hasCommandBufferRecorded = false;
             Bool imageAvailableSemaphoreConsumed = false;
-            // Command buffers submitted mid-frame (FlushPendingCommands) whose
-            // execution is only known complete once this slot's fence has been
-            // waited again; freed at that point.
-            Vector<VkCommandBuffer> retiredCommandBuffers;
+            // Command buffers submitted mid-frame (FlushPendingCommands),
+            // appended in submit order; freed once their submission is known
+            // complete (fence wait or completion poll).
+            Vector<RetiredCommandBuffer> retiredCommandBuffers;
             // Submit-tracker index of this slot's most recent queue submission
             // (written by the renderer at submit time).
             Uint64 lastSubmitIndex = 0;
@@ -79,9 +89,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Parks the current (already ended and submitted) command buffer on the
         // slot's retired list and installs a freshly allocated one, so recording
         // can restart while the submitted buffer is still executing. Retired
-        // buffers are freed after the slot's fence is next waited.
+        // buffers are freed after the slot's fence is next waited, or as soon
+        // as their submission is observed complete.
         VkResult RetireCurrentCommandBuffer();
 
+        // Frees every retired command buffer whose tagged submission index is
+        // known complete. Driven by the renderer's submit tracker on completion
+        // events (fence waits and non-blocking polls), so present-less flush
+        // loops reclaim their buffers without any extra wait.
+        void FreeRetiredCommandBuffersCompletedUpTo(Uint64 completedSubmitIndex);
+        // Frees every slot's retired command buffers. Only valid when the
+        // caller has proven every queue submission complete.
+        void FreeAllRetiredCommandBuffers();
+
         Uint32 GetCurrentFrameIndex() const;
         Uint32 GetFrameCount() const;
 

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.cpp

@@ -8,6 +8,8 @@
 
 #include "PipelineFactory.h"
 
+#include <algorithm>
+
 namespace MobileGL::MG_Backend::DirectVulkan {
     static const char* PrimitiveTopologyToString(VkPrimitiveTopology topology) {
         switch (topology) {
@@ -108,6 +110,81 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                   "vkCreatePipelineCache");
     }
 
+    // Must be called once, before any pipeline is created: the flag is not part of the
+    // pipeline hash, so flipping it mid-life would serve cached pipelines built under the
+    // old value.
+    void PipelineFactory::SetSuppressBlendedDepthWrite(Bool enabled) {
+        s_suppressBlendedDepthWrite = enabled;
+    }
+
+    Bool PipelineFactory::ShouldSuppressBlendedDepthWriteForDevice(MG_Config::QuirkOverride quirkOverride,
+                                                                   Uint32 vendorId) {
+        static constexpr Uint32 kVendorIdQualcomm = 0x5143;
+        switch (quirkOverride) {
+        case MG_Config::QuirkOverride::ForceOn:
+            return true;
+        case MG_Config::QuirkOverride::ForceOff:
+            return false;
+        case MG_Config::QuirkOverride::Auto:
+        default:
+            return vendorId == kVendorIdQualcomm;
+        }
+    }
+
+    namespace {
+        // MIN/MAX extremum blending: the signature of a depth-bounds accumulation pass
+        // (MC 26.3 OIT writes vec4(-linD, linD, deviceZ, 0) under GL_MAX while writing
+        // depth for its equality chain). MIN/MAX ignore blend factors per the Vulkan spec.
+        //
+        // Deliberately the ONLY shape stripped. A quirk should touch as little unrelated
+        // content as possible, and a trace sweep of every fixture showed the wider
+        // alternatives all cost more than they fix:
+        //   - additive ONE+ONE with a depth write matched zero draws of the 26.3 chain
+        //     (its transmittance/accumulate passes disable depth writes themselves) - the
+        //     only real content it caught was harmless additive glow effects (Create);
+        //   - sorted-transparency "over" blends (SRC_ALPHA-style) are order-dependent,
+        //     drawn once per surface, and rely on their depth writes for occlusion;
+        //   - separate-alpha accumulation over an over-blending color channel has no
+        //     known pairing with a depth-equality chain (color channel only, see tests).
+        // If a future workload pairs another blend shape with an equality chain, widen
+        // this with that evidence in hand rather than pre-emptively.
+        Bool IsAccumulationBlend(const VkPipelineColorBlendAttachmentState& attachment) {
+            return attachment.colorBlendOp == VK_BLEND_OP_MIN ||
+                   attachment.colorBlendOp == VK_BLEND_OP_MAX;
+        }
+    } // namespace
+
+    Bool PipelineFactory::ShouldSuppressDepthWrite(const PipelineCreatePayload& payload) {
+        if (!payload.depthWriteEnable) {
+            return false;
+        }
+        // A shader that assigns gl_FragDepth supplies depth itself rather than taking the
+        // pipeline's interpolated Z, so a driver that varies the vertex position math
+        // between pipelines cannot desynchronize it. (A gl_FragDepth = gl_FragCoord.z
+        // passthrough is the exception that stays exposed; no known content pairs one with
+        // an equality chain, and 26.3's composite is a genuine computed-depth writer.)
+        if (payload.fragmentReplacesDepth) {
+            return false;
+        }
+        for (Uint32 i = 0; i < payload.colorAttachmentCount; ++i) {
+            const VkPipelineColorBlendAttachmentState& attachment = payload.colorBlendAttachments[i];
+            if (attachment.blendEnable != VK_TRUE) {
+                continue;
+            }
+            // All color writes masked: blending is moot (depth-prepass pattern that left
+            // GL_BLEND enabled); stripping the depth write would delete the whole prepass.
+            if (attachment.colorWriteMask == 0) {
+                continue;
+            }
+            // Any attachment qualifies, not just attachment 0: the 26.3 transmittance pass
+            // accumulates into a 2-target MRT and must stay stripped.
+            if (IsAccumulationBlend(attachment)) {
+                return true;
+            }
+        }
+        return false;
+    }
+
     PipelineFactory::~PipelineFactory() {
         DestroyAll();
         if (m_pipelineCache != VK_NULL_HANDLE) {
@@ -152,6 +229,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             XXH64_update(m_hashState, &payload.backStencilDepthFailOp, sizeof(payload.backStencilDepthFailOp)));
         XXHASH_VERIFY(
             XXH64_update(m_hashState, &payload.backStencilCompareOp, sizeof(payload.backStencilCompareOp)));
+        XXHASH_VERIFY(
+            XXH64_update(m_hashState, &payload.fragmentReplacesDepth, sizeof(payload.fragmentReplacesDepth)));
         if (payload.colorAttachmentCount > 0) {
             XXHASH_VERIFY(XXH64_update(
                 m_hashState,
@@ -165,23 +244,108 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const HashType hash = ComputeHash(payload);
         auto it = m_cache.find(hash);
         if (it != m_cache.end()) {
-            return it->second;
+            it->second.lastUsedFrame = m_frameCounter;
+            return it->second.pipeline;
         }
 
         VkPipeline pipeline = CreatePipeline(payload);
-        m_cache.emplace(hash, pipeline);
+        m_cache.emplace(hash, PipelineCacheEntry{pipeline, payload.programHash, payload.renderPass,
+                                                 m_frameCounter});
         return pipeline;
     }
 
     void PipelineFactory::DestroyAll() {
         for (auto& pair : m_cache) {
-            if (pair.second != VK_NULL_HANDLE) {
-                vkDestroyPipeline(m_device, pair.second, nullptr);
+            if (pair.second.pipeline != VK_NULL_HANDLE) {
+                vkDestroyPipeline(m_device, pair.second.pipeline, nullptr);
             }
         }
         m_cache.clear();
     }
 
+    Uint32 PipelineFactory::OnFrameBoundary() {
+        ++m_frameCounter;
+
+        // Sweep cadence and retire age mirror VkRenderPassManager::OnPresent: an entry
+        // idle for more than kRetireAgeFrames frame boundaries cannot be referenced by
+        // any in-flight command buffer (frames-in-flight <= MOBILEGL_MAGMA_FRAMESINFLIGHT),
+        // so immediate vkDestroyPipeline is safe. The caller must drop its "last
+        // pipeline" memo when this returns non-zero: the memo can return a cached
+        // handle without touching this cache, so an evicted pipeline may still be
+        // memoized (present-less flush loops never reset the memo per frame).
+        constexpr Uint64 kSweepInterval = 256;
+        constexpr Uint64 kRetireAgeFrames = 1024;
+        if ((m_frameCounter % kSweepInterval) != 0) {
+            return 0;
+        }
+
+        Uint32 evicted = 0;
+        for (auto it = m_cache.begin(); it != m_cache.end();) {
+            if (m_frameCounter - it->second.lastUsedFrame > kRetireAgeFrames) {
+                if (it->second.pipeline != VK_NULL_HANDLE) {
+                    vkDestroyPipeline(m_device, it->second.pipeline, nullptr);
+                }
+                it = m_cache.erase(it);
+                ++evicted;
+            } else {
+                ++it;
+            }
+        }
+        if (evicted > 0) {
+            MGLOG_D("PipelineFactory::OnFrameBoundary: evicted %u idle pipelines (%zu remain)", evicted,
+                    m_cache.size());
+        }
+        return evicted;
+    }
+
+    Uint32 PipelineFactory::EvictByRenderPasses(const Vector<VkRenderPass>& renderPasses) {
+        if (renderPasses.empty() || m_cache.empty()) {
+            return 0;
+        }
+        // Sorted-batch membership test keeps a mass eviction (shader-pack switch,
+        // dimension exit) at one O(cache * log batch) scan instead of one full scan
+        // per dying pass.
+        Vector<VkRenderPass> sortedPasses = renderPasses;
+        std::sort(sortedPasses.begin(), sortedPasses.end());
+        Uint32 evicted = 0;
+        for (auto it = m_cache.begin(); it != m_cache.end();) {
+            if (std::binary_search(sortedPasses.begin(), sortedPasses.end(), it->second.renderPass)) {
+                if (it->second.pipeline != VK_NULL_HANDLE) {
+                    vkDestroyPipeline(m_device, it->second.pipeline, nullptr);
+                }
+                it = m_cache.erase(it);
+                ++evicted;
+            } else {
+                ++it;
+            }
+        }
+        if (evicted > 0) {
+            MGLOG_D("PipelineFactory::EvictByRenderPasses: evicted %u pipelines for %zu destroyed render passes",
+                    evicted, sortedPasses.size());
+        }
+        return evicted;
+    }
+
+    Uint32 PipelineFactory::EvictByProgramHash(HashType programHash) {
+        Uint32 evicted = 0;
+        for (auto it = m_cache.begin(); it != m_cache.end();) {
+            if (it->second.programHash == programHash) {
+                if (it->second.pipeline != VK_NULL_HANDLE) {
+                    vkDestroyPipeline(m_device, it->second.pipeline, nullptr);
+                }
+                it = m_cache.erase(it);
+                ++evicted;
+            } else {
+                ++it;
+            }
+        }
+        if (evicted > 0) {
+            MGLOG_D("PipelineFactory::EvictByProgramHash: evicted %u pipelines for program hash 0x%llx",
+                    evicted, static_cast<unsigned long long>(programHash));
+        }
+        return evicted;
+    }
+
     VkPipeline PipelineFactory::CreatePipeline(const PipelineCreatePayload& payload) const {
         MOBILEGL_ASSERT(payload.stages != nullptr && !payload.stages->empty(), "PipelineFactory: stages are empty");
         MOBILEGL_ASSERT(payload.vertexInputState != nullptr, "PipelineFactory: vertexInputState is null");
@@ -258,6 +422,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         for (Uint32 i = 0; i < payload.colorAttachmentCount; ++i) {
             colorAttachments[i] = payload.colorBlendAttachments[i];
         }
+        // Suppress depth writes on accumulation-blended pipelines when the active driver
+        // cannot keep vertex positions invariant across the pipelines of a multi-pass
+        // depth-equality chain (see SetSuppressBlendedDepthWrite). The decision is narrowed
+        // in ShouldSuppressDepthWrite: sorted-transparency "over" blends (vanilla MC water),
+        // gl_FragDepth writers, and masked-out attachments keep their depth writes.
+        // This bakes the decision into the pipeline, which only works because depth write is
+        // static state here - adding VK_DYNAMIC_STATE_DEPTH_WRITE_ENABLE to kDynamicStates
+        // would let the record-time value override it and silently disable the quirk.
+        if (s_suppressBlendedDepthWrite && ShouldSuppressDepthWrite(payload)) {
+            depthStencil.depthWriteEnable = VK_FALSE;
+        }
         VkPipelineColorBlendStateCreateInfo blend{VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO};
         blend.logicOpEnable = payload.logicOpEnable ? VK_TRUE : VK_FALSE;
         blend.logicOp = payload.logicOp;

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.h

@@ -49,6 +49,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkStencilOp backStencilPassOp = VK_STENCIL_OP_KEEP;
             VkStencilOp backStencilDepthFailOp = VK_STENCIL_OP_KEEP;
             VkCompareOp backStencilCompareOp = VK_COMPARE_OP_ALWAYS;
+            // The fragment module writes gl_FragDepth (SPIR-V DepthReplacing); exempts the
+            // pipeline from the blended depth-write quirk (see ShouldSuppressDepthWrite).
+            Bool fragmentReplacesDepth = false;
             Array<VkPipelineColorBlendAttachmentState, kMaxColorAttachments> colorBlendAttachments{};
             const Vector<VkPipelineShaderStageCreateInfo>* stages = nullptr;
             const VkPipelineVertexInputStateCreateInfo* vertexInputState = nullptr;
@@ -62,13 +65,69 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkPipeline GetOrCreatePipeline(const PipelineCreatePayload& payload);
         void DestroyAll();
 
+        // Frame boundary hook: ages the pipeline cache and destroys long-unused entries
+        // (their command buffers retired many frames ago), mirroring
+        // VkRenderPassManager::OnPresent's sweep. Returns the number of pipelines
+        // destroyed so the caller can drop any memoized VkPipeline handle.
+        Uint32 OnFrameBoundary();
+        // Destroys every cached pipeline hashed on one of `renderPasses`. Only safe
+        // when the caller guarantees GPU idleness for them - the render-pass manager
+        // calls this (via the renderer) for passes its own >1024-boundary-idle sweep
+        // just evicted, and a pipeline hashed on those handles is only ever bound by
+        // draws that also hit the render-pass entries. Also closes the handle-recycling
+        // hazard: a recycled VkRenderPass value must never serve a stale pipeline.
+        // Batched: one cache scan regardless of how many passes died in the sweep.
+        // Returns the number destroyed (callers invalidate memos when non-zero).
+        Uint32 EvictByRenderPasses(const Vector<VkRenderPass>& renderPasses);
+        // Destroys every cached pipeline built from the program with content hash
+        // `programHash`. Called from the ProgramFactory eviction path, which proves the
+        // same >1024-boundary idleness (the program's pipelines are only bound by draws
+        // that stamp its factory entry). Returns the number destroyed.
+        Uint32 EvictByProgramHash(HashType programHash);
+
+        // Driver quirk: suppress depth writes on accumulation-blended pipelines. Multi-pass
+        // depth-equality rendering (a blended prepass writes depth that later passes re-test
+        // with an equality-inclusive compare on the re-rasterized geometry) requires
+        // cross-pipeline position invariance that some mobile compilers do not provide, even
+        // with the SPIR-V Invariant decoration; whole primitives then drop out of the later
+        // passes. Only MIN/MAX extremum blends are stripped - the signature of such a
+        // chain's depth-bounds pass (MC 26.3 OIT), and per a fixture-wide trace sweep the
+        // only depth-writing shape the chain actually uses - so every other blend
+        // (sorted-transparency "over" like vanilla MC water, additive glows, ...) keeps
+        // its depth writes. Set at renderer initialization based on the active driver.
+        static void SetSuppressBlendedDepthWrite(Bool enabled);
+        static Bool IsSuppressBlendedDepthWriteEnabled() { return s_suppressBlendedDepthWrite; }
+        // Device gate for the quirk: ForceOn/ForceOff bypass detection, Auto enables it on
+        // the known-affected vendor (Qualcomm).
+        static Bool ShouldSuppressBlendedDepthWriteForDevice(MG_Config::QuirkOverride quirkOverride,
+                                                             Uint32 vendorId);
+        // Pure per-pipeline strip decision (exempts gl_FragDepth writers, masked-out and
+        // non-accumulation blends); combined with the device flag in CreatePipeline. Static
+        // and payload-only so tests can pin the contract without a VkDevice.
+        static Bool ShouldSuppressDepthWrite(const PipelineCreatePayload& payload);
+
     private:
+        struct PipelineCacheEntry {
+            VkPipeline pipeline = VK_NULL_HANDLE;
+            // The hashed inputs the eviction paths key on: programHash ties the entry to
+            // its ProgramFactory entry, renderPass records the exact handle the hash
+            // folded in (the hash is one-way, so targeted eviction needs them verbatim).
+            HashType programHash = 0;
+            VkRenderPass renderPass = VK_NULL_HANDLE;
+            // Frame-boundary counter value of the last GetOrCreatePipeline hit; drives
+            // cache eviction (see OnFrameBoundary).
+            Uint64 lastUsedFrame = 0;
+        };
+
         VkPipeline CreatePipeline(const PipelineCreatePayload& payload) const;
 
         VkDevice m_device = VK_NULL_HANDLE;
         const VulkanRendererConfig& m_config;
         VkPipelineCache m_pipelineCache = VK_NULL_HANDLE;
-        UnorderedMap<HashType, VkPipeline> m_cache;
+        UnorderedMap<HashType, PipelineCacheEntry> m_cache;
+        // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
+        Uint64 m_frameCounter = 0;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
+        static inline Bool s_suppressBlendedDepthWrite = false;
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.h

@@ -42,6 +42,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             SurfaceRotate90 = 1 << 2,
             SurfaceRotate180 = 1 << 3,
             SurfaceRotate270 = 1 << 4,
+            // Rewrites the fragment stage's implicit-LOD image samples to explicit LOD 0.
+            // Only ever set for a draw whose every sampler binding is clamped to a single mip
+            // level, which makes the two forms produce identical texels (the implicit lambda is
+            // clamped into [minLod, maxLod] = [0, 0] regardless of derivatives or bias).
+            ExplicitLod0Sampling = 1 << 5,
+            // Fragment arithmetic may run at relaxed (fp16) precision. Only requested for draws
+            // where every sampled texture and every colour attachment is an 8-bit-or-less
+            // normalized format, so nothing the shader reads or writes carries more precision
+            // than fp16 already represents exactly.
+            RelaxedFragmentPrecision = 1 << 6,
         };
         using CompileOptionFlags = Flags<CompileOptionBit>;
         using HashType = Uint64;
@@ -59,6 +69,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Vector<DescriptorBindingKind> bindingKinds;
             Vector<Uint32> dynamicBindings;
             Vector<Int> uniformBlockIndexByBinding;
+            // Descriptor count per binding (1 except for UBO instance arrays, which occupy one
+            // binding with descriptorCount = N).
+            Vector<Uint16> bindingDescriptorCounts;
+            // Per-element GL uniform block indices for arrayed UBO bindings (count > 1);
+            // element 0 of a non-arrayed binding stays in uniformBlockIndexByBinding.
+            UnorderedMap<Uint32, Vector<Int>> arrayedUniformBlockIndicesByBinding;
             Vector<String> samplerNameByBinding;
             Vector<Int> samplerUniformLocationByBinding;
             Vector<TextureTarget> samplerTextureTargetByBinding;
@@ -67,6 +83,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Vector<Bool> storageImageUsesBindingFormatByBinding;
             Vector<String> storageBlockNameByBinding;
             Vector<Int> storageBlockIndexByBinding;
+            // Set once during ReflectLayout so the per-draw path can skip the whole
+            // storage-image preparation for the overwhelming majority of programs.
+            Bool hasStorageImages = false;
             Int globalUboBinding = -1;
             Uint32 activeVertexInputLocationMask = 0;
             Array<GLenum, kMaxVertexInputLocations> vertexInputTypes{};
@@ -75,6 +94,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             ShaderStage rasterizationProducerStage = ShaderStage::Unknown;
             Uint32 producerOutputComponentCount = 0;
             Uint32 fragmentInputComponentCount = 0;
+            // The fragment module declares the DepthReplacing execution mode (writes
+            // gl_FragDepth); shader-computed depth is immune to the cross-pipeline
+            // position-invariance quirk (see PipelineFactory::ShouldSuppressDepthWrite).
+            Bool fragmentReplacesDepth = false;
+            // Frame-boundary counter value of the last GetOrCreateProgram hit; drives
+            // cache eviction (see OnFrameBoundary).
+            Uint64 lastUsedFrame = 0;
 
             static inline VkDevice s_device = VK_NULL_HANDLE;
 
@@ -90,6 +116,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 bindingKinds = std::move(other.bindingKinds);
                 dynamicBindings = std::move(other.dynamicBindings);
                 uniformBlockIndexByBinding = std::move(other.uniformBlockIndexByBinding);
+                bindingDescriptorCounts = std::move(other.bindingDescriptorCounts);
+                arrayedUniformBlockIndicesByBinding = std::move(other.arrayedUniformBlockIndicesByBinding);
                 samplerNameByBinding = std::move(other.samplerNameByBinding);
                 samplerUniformLocationByBinding = std::move(other.samplerUniformLocationByBinding);
                 samplerTextureTargetByBinding = std::move(other.samplerTextureTargetByBinding);
@@ -99,6 +127,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     std::move(other.storageImageUsesBindingFormatByBinding);
                 storageBlockNameByBinding = std::move(other.storageBlockNameByBinding);
                 storageBlockIndexByBinding = std::move(other.storageBlockIndexByBinding);
+                hasStorageImages = other.hasStorageImages;
                 globalUboBinding = other.globalUboBinding;
                 activeVertexInputLocationMask = other.activeVertexInputLocationMask;
                 vertexInputTypes = other.vertexInputTypes;
@@ -107,15 +136,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 rasterizationProducerStage = other.rasterizationProducerStage;
                 producerOutputComponentCount = other.producerOutputComponentCount;
                 fragmentInputComponentCount = other.fragmentInputComponentCount;
+                fragmentReplacesDepth = other.fragmentReplacesDepth;
+                lastUsedFrame = other.lastUsedFrame;
                 other.hash = 0;
                 other.descriptorSetLayout = VK_NULL_HANDLE;
                 other.pipelineLayout = VK_NULL_HANDLE;
+                other.hasStorageImages = false;
                 other.globalUboBinding = -1;
                 other.activeVertexInputLocationMask = 0;
                 other.activeFragmentOutputLocationMask = 0;
                 other.rasterizationProducerStage = ShaderStage::Unknown;
                 other.producerOutputComponentCount = 0;
                 other.fragmentInputComponentCount = 0;
+                other.fragmentReplacesDepth = false;
+                other.lastUsedFrame = 0;
             }
             VkProgramObject& operator=(VkProgramObject&& other) noexcept {
                 if (this == &other) {
@@ -130,6 +164,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 bindingKinds = std::move(other.bindingKinds);
                 dynamicBindings = std::move(other.dynamicBindings);
                 uniformBlockIndexByBinding = std::move(other.uniformBlockIndexByBinding);
+                bindingDescriptorCounts = std::move(other.bindingDescriptorCounts);
+                arrayedUniformBlockIndicesByBinding = std::move(other.arrayedUniformBlockIndicesByBinding);
                 samplerNameByBinding = std::move(other.samplerNameByBinding);
                 samplerUniformLocationByBinding = std::move(other.samplerUniformLocationByBinding);
                 samplerTextureTargetByBinding = std::move(other.samplerTextureTargetByBinding);
@@ -139,6 +175,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     std::move(other.storageImageUsesBindingFormatByBinding);
                 storageBlockNameByBinding = std::move(other.storageBlockNameByBinding);
                 storageBlockIndexByBinding = std::move(other.storageBlockIndexByBinding);
+                hasStorageImages = other.hasStorageImages;
                 globalUboBinding = other.globalUboBinding;
                 activeVertexInputLocationMask = other.activeVertexInputLocationMask;
                 vertexInputTypes = other.vertexInputTypes;
@@ -147,15 +184,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 rasterizationProducerStage = other.rasterizationProducerStage;
                 producerOutputComponentCount = other.producerOutputComponentCount;
                 fragmentInputComponentCount = other.fragmentInputComponentCount;
+                fragmentReplacesDepth = other.fragmentReplacesDepth;
+                lastUsedFrame = other.lastUsedFrame;
                 other.hash = 0;
                 other.descriptorSetLayout = VK_NULL_HANDLE;
                 other.pipelineLayout = VK_NULL_HANDLE;
+                other.hasStorageImages = false;
                 other.globalUboBinding = -1;
                 other.activeVertexInputLocationMask = 0;
                 other.activeFragmentOutputLocationMask = 0;
                 other.rasterizationProducerStage = ShaderStage::Unknown;
                 other.producerOutputComponentCount = 0;
                 other.fragmentInputComponentCount = 0;
+                other.fragmentReplacesDepth = false;
+                other.lastUsedFrame = 0;
                 return *this;
             }
 
@@ -185,6 +227,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             }
         };
 
+        // Notified when the OnFrameBoundary sweep destroys an aged-out cache entry,
+        // carrying the entry's content hash and the VkDescriptorSetLayout it owned.
+        // Dependent caches (compute pipelines, PipelineFactory entries, UniformManager's
+        // per-layout descriptor sets) must purge in the same step: after vkDestroy the
+        // layout handle value may be recycled for an unrelated layout, and the program
+        // hash may be re-inserted by a later rebuild of the same content.
+        class IEvictionObserver {
+        public:
+            virtual ~IEvictionObserver() = default;
+            virtual void OnProgramEvicted(HashType programHash, VkDescriptorSetLayout descriptorSetLayout) = 0;
+        };
+
         explicit ProgramFactory(VkDevice device, const VulkanRendererConfig& config, Uint32 maxBindings = 16,
                                 Bool shaderDrawParametersEnabled = false,
                                 Bool unformattedFloatStorageImagesEnabled = false)
@@ -200,9 +254,24 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkProgramObject& GetOrCreateProgram(
             const MG_State::GLState::ProgramObject& program, CompileOptionFlags flags);
 
+        // Observer may be null (no notifications). Not owned.
+        void SetEvictionObserver(IEvictionObserver* observer) { m_evictionObserver = observer; }
+        // Frame boundary hook: ages the program cache and evicts long-unused entries
+        // (their command buffers retired many frames ago), mirroring
+        // VkRenderPassManager::OnPresent's sweep.
+        void OnFrameBoundary();
+
         static VkShaderStageFlagBits ToVkStage(ShaderStage stage);
         static VkFormat ConvertSpirvImageFormatToVkFormat(SpvImageFormat format);
         static SamplerNumericDomain UniformTypeToSamplerNumericDomain(GLenum glType);
+        // True when any entry point declares the DepthReplacing execution mode, i.e. the
+        // shader assigns gl_FragDepth. Exposed so the blended depth-write quirk's exemption
+        // can be pinned by tests. A false negative loses the exemption, so such a shader is
+        // stripped conservatively and forfeits its depth write.
+        static Bool ReflectedFragmentReplacesDepth(const SpvReflectShaderModule& reflectModule);
+        // True when an entry point reads the InstanceIndex builtin. Only gates a diagnostic:
+        // without shaderDrawParameters such a shader cannot have gl_InstanceID rebased.
+        static Bool ReflectedReadsInstanceIndexBuiltin(const SpvReflectShaderModule& reflectModule);
 
     private:
         struct ProgramLookupCache {
@@ -233,6 +302,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // shaderStorageImageReadWithoutFormat and shaderStorageImageWriteWithoutFormat.
         Bool m_unformattedFloatStorageImagesEnabled = false;
         mutable ProgramLookupCache m_lastLookup;
+        // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
+        Uint64 m_frameCounter = 0;
+        IEvictionObserver* m_evictionObserver = nullptr;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/SwapchainObject.cpp

@@ -247,6 +247,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         m_surfaceFormat = {createInfo.imageFormat, createInfo.imageColorSpace};
         m_extent = createInfo.imageExtent;
+        // The surface-space extent this swapchain was built from, i.e. before the
+        // quarter-turn swap above. Out-of-date checks must compare in THIS space: comparing a
+        // freshly queried currentExtent against the swapped m_extent flips axes every rotation
+        // and makes the comparison alternate forever.
+        m_surfaceExtent = defaultFramebufferExtent;
         m_preTransform = createInfo.preTransform;
 
         VK_VERIFY(vkCreateSwapchainKHR(device, &createInfo, nullptr, &m_swapchain));

MobileGL/MG_Backend/DirectVulkan/Renderer/SwapchainObject.h

@@ -35,6 +35,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkSwapchainKHR GetHandle() const { return m_swapchain; }
         const VkSurfaceFormatKHR& GetSurfaceFormat() const { return m_surfaceFormat; }
         VkExtent2D GetExtent() const { return m_extent; }
+        // Surface-space extent (before the pre-rotation quarter-turn swap) this swapchain was
+        // created from - the value to compare a freshly queried currentExtent against.
+        VkExtent2D GetSurfaceExtent() const { return m_surfaceExtent; }
         VkSurfaceTransformFlagBitsKHR GetPreTransform() const { return m_preTransform; }
         const Vector<VkImage>& GetImages() const { return m_images; }
         const Vector<VkImageView>& GetImageViews() const { return m_imageViews; }
@@ -63,6 +66,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkSwapchainKHR m_swapchain = VK_NULL_HANDLE;
         VkSurfaceFormatKHR m_surfaceFormat{};
         VkExtent2D m_extent{};
+        VkExtent2D m_surfaceExtent{};
         VkSurfaceTransformFlagBitsKHR m_preTransform = VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR;
         Vector<VkImage> m_images;
         Vector<VkImageView> m_imageViews;

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.cpp

@@ -16,6 +16,7 @@
 #include "MG_Util/Converters/GLToMG/TextureEnumConverter.h"
 #include "MG_Util/Converters/MGToStr/FramebufferEnumConverter.h"
 #include "MG_Util/Converters/MGToVk/TextureEnumConverter.h"
+#include <vulkan/utility/vk_format_utils.h>
 #include "MG_Util/Metrics/TextureMetrics.h"
 #include <Config.h>
 #include <cstdio>
@@ -211,6 +212,40 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     }
 
+    void UniformManager::OnDescriptorSetLayoutDestroyed(VkDescriptorSetLayout descriptorSetLayout) {
+        SizeT purgedSets = 0;
+        for (auto& frame : m_frames) {
+            const auto it = frame.descriptorSetCacheByLayout.find(descriptorSetLayout);
+            if (it == frame.descriptorSetCacheByLayout.end()) {
+                continue;
+            }
+            // Free the sets back to their pools and credit the bucket accounting, so
+            // program churn recycles pool capacity instead of abandoning the slots.
+            // GPU-safe: the layout only dies after >1024 idle frame boundaries, so no
+            // in-flight command buffer references these sets.
+            for (const auto& cached : it->second.sets) {
+                if (cached.set == VK_NULL_HANDLE) {
+                    continue;
+                }
+                vkFreeDescriptorSets(m_device, cached.pool, 1, &cached.set);
+                const auto bucket = std::find_if(
+                    frame.descriptorPools.begin(), frame.descriptorPools.end(),
+                    [&cached](const DescriptorPoolBucket& candidate) { return candidate.handle == cached.pool; });
+                if (bucket != frame.descriptorPools.end() && bucket->allocatedSets > 0) {
+                    --bucket->allocatedSets;
+                }
+            }
+            purgedSets += it->second.sets.size();
+            frame.descriptorSetCacheByLayout.erase(it);
+        }
+        if (purgedSets > 0) {
+            // The per-draw reuse memo folds the layout handle into its signature; drop
+            // it so a recycled handle value cannot revive a purged set mid-frame.
+            m_hasLastDescriptor = false;
+            MGLOG_D("UniformDescriptorBinder: freed %zu descriptor sets for destroyed layout", purgedSets);
+        }
+    }
+
     Bool UniformManager::ResolveSamplerDescriptor(VkCommandBuffer commandBuffer,
                                                             const MG_State::GLState::ProgramObject& program,
                                                             const ProgramFactory::VkProgramObject& programObj,
@@ -298,8 +333,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // descriptor valid.
         const Bool forceNearestFiltering = numericDomain == SamplerNumericDomain::SignedInteger ||
                                            numericDomain == SamplerNumericDomain::UnsignedInteger;
-        const VkFormat sampledViewFormat =
-            VkTextureManager::ResolveSampledImageViewFormat(resource->format, numericDomain);
+        SamplerResolveMemo* viewFormatMemo =
+            binding < m_samplerResolveMemo.size() ? &m_samplerResolveMemo[binding] : nullptr;
+        VkFormat sampledViewFormat;
+        if (viewFormatMemo != nullptr && viewFormatMemo->viewFormatValid &&
+            viewFormatMemo->viewFormatSource == resource->format &&
+            viewFormatMemo->viewFormatDomain == numericDomain) {
+            sampledViewFormat = viewFormatMemo->viewFormat;
+        } else {
+            sampledViewFormat =
+                VkTextureManager::ResolveSampledImageViewFormat(resource->format, numericDomain);
+            if (viewFormatMemo != nullptr) {
+                viewFormatMemo->viewFormatSource = resource->format;
+                viewFormatMemo->viewFormatDomain = numericDomain;
+                viewFormatMemo->viewFormat = sampledViewFormat;
+                viewFormatMemo->viewFormatValid = true;
+            }
+        }
         if (sampledViewFormat == VK_FORMAT_UNDEFINED) {
             MGLOG_E("ResolveSamplerDescriptor: no compatible sampled view for binding=%u ('%s') "
                     "textureId=%d imageFormat=%d numericDomain=%d",
@@ -307,8 +357,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     static_cast<Int>(resource->format), static_cast<Int>(numericDomain));
             return false;
         }
+        // No reinterpretation requested: bind the depth-or-color aspect view the sync above
+        // already produced instead of re-entering GetOrCreateSampledImageView's sync path.
         const VkImageView sampledImageView =
-            m_textureManager->GetOrCreateSampledImageView(*texture, sampledViewFormat);
+            sampledViewFormat == resource->format
+                ? resource->sampledView
+                : m_textureManager->GetOrCreateSampledImageView(*texture, sampledViewFormat);
         if (sampledImageView == VK_NULL_HANDLE) {
             MGLOG_E("ResolveSamplerDescriptor: failed to resolve sampled view for binding=%u ('%s') "
                     "textureId=%d imageFormat=%d viewFormat=%d numericDomain=%d",
@@ -331,24 +385,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const Uint16 samplerVersion = samplerToUse->GetVersion();
             const Uint64 textureLifetimeId = texture->GetLifetimeId();
             const Uint16 textureParamsVersion = texture->GetTextureParamsVersion();
+            // The sampler's LOD clamp depends on how many levels the sampled view exposes, and that
+            // follows uploads as well as GL parameters - so it belongs in the memo key too.
+            const Uint32 viewLevelCount = resource->sampledLevelCount;
             if (memo.valid && memo.samplerLifetimeId == samplerLifetimeId && memo.samplerVersion == samplerVersion &&
                 memo.textureLifetimeId == textureLifetimeId && memo.textureParamsVersion == textureParamsVersion &&
-                memo.forceNearestFiltering == forceNearestFiltering) {
+                memo.forceNearestFiltering == forceNearestFiltering && memo.viewLevelCount == viewLevelCount) {
                 resolvedSampler = memo.sampler;
             } else {
-                resolvedSampler =
-                    m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture, forceNearestFiltering);
+                resolvedSampler = m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture,
+                                                                       forceNearestFiltering, viewLevelCount);
                 memo.samplerLifetimeId = samplerLifetimeId;
                 memo.samplerVersion = samplerVersion;
                 memo.textureLifetimeId = textureLifetimeId;
                 memo.textureParamsVersion = textureParamsVersion;
                 memo.forceNearestFiltering = forceNearestFiltering;
+                memo.viewLevelCount = viewLevelCount;
                 memo.sampler = resolvedSampler;
                 memo.valid = true;
             }
         } else {
-            resolvedSampler =
-                m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture, forceNearestFiltering);
+            resolvedSampler = m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture, forceNearestFiltering,
+                                                                   resource->sampledLevelCount);
         }
         outImageInfo = {
             .sampler = resolvedSampler,
@@ -389,6 +447,106 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return outImageInfo.sampler != VK_NULL_HANDLE;
     }
 
+    namespace {
+        // fp16 carries an 11-bit mantissa, so an 8-bit normalized channel round-trips exactly.
+        // Anything wider - 16-bit normalized, half float, full float, and every packed HDR
+        // encoding - holds precision or range that relaxing the arithmetic would throw away.
+        Bool IsLowPrecisionNormalizedFormat(VkFormat format) {
+            if (format == VK_FORMAT_UNDEFINED) return false;
+            if (!vkuFormatIsUNORM(format) && !vkuFormatIsSNORM(format) && !vkuFormatIsSRGB(format)) {
+                return false;
+            }
+            const struct VKU_FORMAT_INFO info = vkuGetFormatInfo(format);
+            for (Uint32 i = 0; i < info.component_count; ++i) {
+                if (info.components[i].size > 8) return false;
+            }
+            return info.component_count > 0;
+        }
+    } // namespace
+
+    Bool UniformManager::DrawTargetIsLowPrecision(const MG_State::GLState::FramebufferObject* drawFramebuffer) {
+        // Default framebuffer: the swapchain is an 8-bit normalized surface.
+        if (drawFramebuffer == nullptr) return true;
+
+        Bool sawColour = false;
+        for (Int i = static_cast<Int>(FramebufferAttachmentType::Color0);
+             i < static_cast<Int>(FramebufferAttachmentType::FramebufferAttachmentTypeCount);
+             ++i) {
+            const auto& attachment =
+                drawFramebuffer->GetAttachment(static_cast<FramebufferAttachmentType>(i));
+            VkFormat format = VK_FORMAT_UNDEFINED;
+            if (const auto& texture = attachment.GetTexture()) {
+                format = MG_Util::ConvertTextureInternalFormatToVkEnum(texture->GetFormat());
+            } else if (const auto& renderbuffer = attachment.GetRenderbuffer()) {
+                format = MG_Util::ConvertTextureInternalFormatToVkEnum(
+                    renderbuffer->GetInternalFormat());
+            } else {
+                continue;
+            }
+            if (!IsLowPrecisionNormalizedFormat(format)) return false;
+            sawColour = true;
+        }
+        return sawColour;
+    }
+
+    Bool UniformManager::ProgramSamplesOnlyLowPrecisionTextures(
+        const MG_State::GLState::ProgramObject& program, const ProgramFactory::VkProgramObject& programObj) {
+        for (Uint32 binding = 0; binding < programObj.bindingKinds.size(); ++binding) {
+            if (programObj.bindingKinds[binding] != ProgramFactory::DescriptorBindingKind::CombinedImageSampler) {
+                continue;
+            }
+            const auto* texture = ResolveSamplerTextureRaw(program, programObj, binding);
+            // An unresolvable binding is unknown territory, not licence to relax.
+            if (texture == nullptr) return false;
+            const VkFormat format =
+                MG_Util::ConvertTextureInternalFormatToVkEnum(texture->GetFormat());
+            if (!IsLowPrecisionNormalizedFormat(format)) return false;
+        }
+        return true;
+    }
+
+    Bool UniformManager::ProgramSamplesOnlySingleLevelTextures(
+        const MG_State::GLState::ProgramObject& program, const ProgramFactory::VkProgramObject& programObj) {
+        Bool sawSampler = false;
+        for (Uint32 binding = 0; binding < programObj.bindingKinds.size(); ++binding) {
+            if (programObj.bindingKinds[binding] != ProgramFactory::DescriptorBindingKind::CombinedImageSampler) {
+                continue;
+            }
+            const auto* texture = ResolveSamplerTextureRaw(program, programObj, binding);
+            if (texture == nullptr) return false;
+            const auto& levelRange = texture->GetLevelRange();
+            if (levelRange.x() != levelRange.y()) return false;
+
+            // An explicit-LOD sample is a single filtered tap, so it also gives up anisotropic
+            // filtering - which a single-level view can still have. Resolve the sampler exactly
+            // the way ResolveSamplerDescriptor does and bail if anisotropy would apply.
+            const Int location = programObj.samplerUniformLocationByBinding[binding];
+            const Int unit = ResolveSamplerUnitIndex(program, location, binding);
+            const auto& samplerOverride = MG_State::pGLContext->GetTextureUnitObject(unit).GetSamplerObject();
+            const auto* effectiveSampler =
+                samplerOverride ? samplerOverride.get() : texture->GetSamplerObject().get();
+            if (effectiveSampler == nullptr) return false;
+            if (effectiveSampler->GetMaxAnisotropy() > 1.0f &&
+                effectiveSampler->GetMinFilter() == SamplerFilterMode::Linear &&
+                effectiveSampler->GetMagFilter() == SamplerFilterMode::Linear) {
+                return false;
+            }
+
+            // An explicit LOD 0 makes lambda exactly 0, which is the magnification side of the
+            // min/mag decision. That only matches the implicit form when lambda could not have been
+            // positive anyway (the LOD clamp already pins it at or below 0), or when the two
+            // filters are the same and the choice cannot be observed.
+            const Float effectiveMaxLod = effectiveSampler->GetMipmapMode() == SamplerMipmapMode::None
+                                              ? 0.0f
+                                              : effectiveSampler->GetMaxLod();
+            if (effectiveMaxLod > 0.0f && effectiveSampler->GetMinFilter() != effectiveSampler->GetMagFilter()) {
+                return false;
+            }
+            sawSampler = true;
+        }
+        return sawSampler;
+    }
+
     Bool UniformManager::ResolveSamplerTexture(const MG_State::GLState::ProgramObject& program,
                                                          const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
                                                          SharedPtr<MG_State::GLState::ITextureObject>& outTexture) {
@@ -746,7 +904,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     Bool UniformManager::ResolveUniformBufferPayload(const MG_State::GLState::ProgramObject& program,
                                                      const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
-                                                     UboBindResult& out) const {
+                                                     Uint32 arrayElement, UboBindResult& out) const {
         const void* outData = nullptr;
         VkDeviceSize outSize = 0;
 
@@ -772,7 +930,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         MOBILEGL_ASSERT(binding < programObj.uniformBlockIndexByBinding.size(),
                         "ResolveUniformBufferPayload: UBO mapping binding %u out of range", binding);
-        const Int blockIndex = programObj.uniformBlockIndexByBinding[binding];
+        Int blockIndex = programObj.uniformBlockIndexByBinding[binding];
+        if (arrayElement > 0) {
+            const auto arrayIt = programObj.arrayedUniformBlockIndicesByBinding.find(binding);
+            const Bool elementValid = arrayIt != programObj.arrayedUniformBlockIndicesByBinding.end() &&
+                                      arrayElement < arrayIt->second.size();
+            MOBILEGL_ASSERT(elementValid,
+                            "ResolveUniformBufferPayload: UBO binding %u has no array element %u", binding,
+                            arrayElement);
+            if (!elementValid) {
+                return false;
+            }
+            blockIndex = arrayIt->second[arrayElement];
+        }
         MOBILEGL_ASSERT(blockIndex >= 0,
                         "ResolveUniformBufferPayload: no uniform block mapped to descriptor binding %u", binding);
 
@@ -880,6 +1050,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         VkDescriptorPoolCreateInfo poolInfo{};
         poolInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO;
+        // FREE_DESCRIPTOR_SET_BIT lets a destroyed layout's cached sets be freed back
+        // (OnDescriptorSetLayoutDestroyed) so program churn recycles pool capacity.
+        // The cost is on set allocation only, which happens when a layout's per-frame
+        // cache grows - never on the per-draw reuse path.
+        poolInfo.flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT;
         poolInfo.maxSets = maxSets;
         poolInfo.poolSizeCount = static_cast<Uint32>(std::size(poolSizes));
         poolInfo.pPoolSizes = poolSizes;
@@ -959,7 +1134,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         auto& frame = m_frames[frameIndex];
         auto& cache = frame.descriptorSetCacheByLayout[programObj.descriptorSetLayout];
         if (cache.cursor < cache.sets.size()) {
-            outDescriptorSet = cache.sets[cache.cursor++];
+            outDescriptorSet = cache.sets[cache.cursor++].set;
         } else {
             VkResult allocResult = AllocateDescriptorSetsFromActivePool(frameIndex, programObj, outDescriptorSet);
             if (allocResult == VK_ERROR_OUT_OF_POOL_MEMORY || allocResult == VK_ERROR_FRAGMENTED_POOL) {
@@ -973,7 +1148,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 return allocResult;
             }
 
-            cache.sets.push_back(outDescriptorSet);
+            // The successful allocation came from the bucket the alloc helper left
+            // active; record it so a layout-destroyed purge can free the set back.
+            cache.sets.push_back({outDescriptorSet, frame.descriptorPools[frame.activeDescriptorPoolIndex].handle});
             ++cache.cursor;
             MGLOG_D("UniformDescriptorBinder: cached descriptor set count for frame=%u grew to %zu", frameIndex,
                     cache.sets.size());
@@ -1019,11 +1196,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         imageInfos.clear();
         texelBufferViews.clear();
         dynamicOffsets.clear();
+        // Arrayed UBO bindings contribute extra buffer infos and dynamic offsets; reserve for
+        // the worst case so the pBufferInfo pointers taken below never dangle on reallocation.
+        Uint32 uboArrayExtra = 0;
+        for (const auto& arrayEntry : programObj.arrayedUniformBlockIndicesByBinding) {
+            uboArrayExtra += static_cast<Uint32>(arrayEntry.second.size()) - 1u;
+        }
         writes.reserve(m_maxBindings);
-        bufferInfos.reserve(m_maxBindings);
+        bufferInfos.reserve(m_maxBindings + uboArrayExtra);
         imageInfos.reserve(m_maxBindings);
         texelBufferViews.reserve(m_maxBindings);
-        dynamicOffsets.reserve(programObj.dynamicBindings.size());
+        dynamicOffsets.reserve(programObj.dynamicBindings.size() + uboArrayExtra);
 
         const Uint32 bindingCount =
             std::min<Uint32>(m_maxBindings, static_cast<Uint32>(programObj.bindingKinds.size()));
@@ -1041,40 +1224,51 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             write.descriptorCount = 1;
 
             if (kind == ProgramFactory::DescriptorBindingKind::UniformBufferDynamic) {
-                UboBindResult ubo{};
-                const Bool hasPayload = ResolveUniformBufferPayload(program, programObj, binding, ubo);
-                MOBILEGL_ASSERT(hasPayload && ubo.payload != nullptr && ubo.payloadSize > 0,
-                                "UniformDescriptorBinder::BindProgramUniformBuffers failed: missing UBO payload on binding %u",
-                                binding);
+                const Uint32 descriptorCount =
+                    binding < programObj.bindingDescriptorCounts.size()
+                        ? std::max<Uint32>(1, programObj.bindingDescriptorCounts[binding])
+                        : 1u;
+                const SizeT firstBufferInfoIndex = bufferInfos.size();
+                for (Uint32 element = 0; element < descriptorCount; ++element) {
+                    UboBindResult ubo{};
+                    const Bool hasPayload =
+                        ResolveUniformBufferPayload(program, programObj, binding, element, ubo);
+                    MOBILEGL_ASSERT(hasPayload && ubo.payload != nullptr && ubo.payloadSize > 0,
+                                    "UniformDescriptorBinder::BindProgramUniformBuffers failed: missing UBO payload on binding %u element %u",
+                                    binding, element);
 
-                VkDescriptorBufferInfo bufferInfo{};
-                // Keep offset 0 (sub-range selected via the dynamic offset) so the hashed bufferInfo
-                // is stable across draws and the descriptor-set reuse cache keeps hitting.
-                bufferInfo.offset = 0;
-                Uint32 dynOffset;
-                if (ubo.directBindable) {
-                    // Zero-copy: bind the app's resident VkBuffer directly, no per-draw memcpy.
-                    bufferInfo.buffer = ubo.buffer;
-                    bufferInfo.range = ubo.range;
-                    dynOffset = static_cast<Uint32>(ubo.dynamicOffset);
-                } else {
-                    BufferSlice slice{};
-                    if (!m_bufferManager->UploadTransient(BufferKind::Uniform, frameIndex, ubo.payload,
-                                                          ubo.payloadSize, m_minDynamicOffsetAlignment, slice)) {
-                        MOBILEGL_ASSERT(false, "UniformDescriptorBinder::BindProgramUniformBuffers failed: UBO upload failed on binding %u",
-                                binding);
-                        return false;
+                    VkDescriptorBufferInfo bufferInfo{};
+                    // Keep offset 0 (sub-range selected via the dynamic offset) so the hashed bufferInfo
+                    // is stable across draws and the descriptor-set reuse cache keeps hitting.
+                    bufferInfo.offset = 0;
+                    Uint32 dynOffset;
+                    if (ubo.directBindable) {
+                        // Zero-copy: bind the app's resident VkBuffer directly, no per-draw memcpy.
+                        bufferInfo.buffer = ubo.buffer;
+                        bufferInfo.range = ubo.range;
+                        dynOffset = static_cast<Uint32>(ubo.dynamicOffset);
+                    } else {
+                        BufferSlice slice{};
+                        if (!m_bufferManager->UploadTransient(BufferKind::Uniform, frameIndex, ubo.payload,
+                                                              ubo.payloadSize, m_minDynamicOffsetAlignment, slice)) {
+                            MOBILEGL_ASSERT(false, "UniformDescriptorBinder::BindProgramUniformBuffers failed: UBO upload failed on binding %u element %u",
+                                    binding, element);
+                            return false;
+                        }
+                        bufferInfo.buffer = slice.buffer;
+                        bufferInfo.range = ubo.payloadSize;
+                        dynOffset = static_cast<Uint32>(slice.offset);
                     }
-                    bufferInfo.buffer = slice.buffer;
-                    bufferInfo.range = ubo.payloadSize;
-                    dynOffset = static_cast<Uint32>(slice.offset);
+                    bufferInfos.push_back(bufferInfo);
+                    // Dynamic offsets are consumed in binding order, then array element order,
+                    // matching Vulkan's dynamic-offset consumption rules.
+                    dynamicOffsets.push_back(dynOffset);
                 }
-                bufferInfos.push_back(bufferInfo);
 
                 write.descriptorType = VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC;
-                write.pBufferInfo = &bufferInfos.back();
+                write.descriptorCount = descriptorCount;
+                write.pBufferInfo = &bufferInfos[firstBufferInfoIndex];
                 writes.push_back(write);
-                dynamicOffsets.push_back(dynOffset);
             } else if (kind == ProgramFactory::DescriptorBindingKind::UniformTexelBuffer) {
                 VkBufferView bufferView = VK_NULL_HANDLE;
                 if (!ResolveTexelBufferDescriptor(program, programObj, binding, frameIndex, bufferView) ||

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -39,6 +39,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void Shutdown();
 
         void BeginFrame(Uint32 frameIndex);
+        // A ProgramFactory eviction just destroyed this layout: purge every frame
+        // slot's cached descriptor sets for it, so a recycled handle value can never
+        // stale-hit sets written for the dead layout's bindings. The sets are
+        // vkFreeDescriptorSets'd back to their pools (created with
+        // FREE_DESCRIPTOR_SET_BIT) and the pool accounting is credited, so program
+        // churn recycles pool capacity instead of abandoning it. GPU-safe: the layout
+        // only dies after >1024 idle frame boundaries, so no in-flight command buffer
+        // references its sets. This is the only eviction path for the per-layout
+        // caches - a live layout's entry must never be purged (its sets would be
+        // unreachable pool slots), so there is deliberately no age-based sweep here.
+        void OnDescriptorSetLayoutDestroyed(VkDescriptorSetLayout descriptorSetLayout);
         Bool CollectSampledTextures(const MG_State::GLState::ProgramObject& program,
                                     const ProgramFactory::VkProgramObject& programObj,
                                     Vector<MG_State::GLState::ITextureObject*>& outTextures);
@@ -58,6 +69,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         static VkFormat ResolveStorageImageViewFormat(VkFormat reflectedFormat, GLenum bindingFormat,
                                                       VkFormat resourceFormat, Bool useBindingFormat);
 
+        // True when the program reads at least one sampler and every one of them is bound to a
+        // texture whose GL level range is a single level. Such a sampler resolves to
+        // minLod = maxLod = 0 (see VkSamplerManager::GetOrCreateSampler), so an implicit-LOD sample
+        // and an explicit LOD 0 sample must read the same texel - which is what makes the
+        // ExplicitLod0Sampling SPIR-V rewrite safe to request. Deliberately conservative: it reads
+        // only GL state, so a texture that ends up single-level for another reason (one uploaded
+        // level under a wide level range) merely misses the rewrite.
+        // True when every texture this program samples is an 8-bit-or-less normalized format, so
+        // relaxing the fragment stage to fp16 cannot lose a bit the texel ever carried. Says
+        // nothing about the render target - the caller must check that too.
+        static Bool ProgramSamplesOnlyLowPrecisionTextures(const MG_State::GLState::ProgramObject& program,
+                                                          const ProgramFactory::VkProgramObject& programObj);
+        // True when every colour attachment the draw writes is an 8-bit-or-less normalized
+        // format (nullptr = default framebuffer, which is). Blending happens at attachment
+        // precision, so a wider target must keep the fragment stage at full precision.
+        static Bool DrawTargetIsLowPrecision(const MG_State::GLState::FramebufferObject* drawFramebuffer);
+        static Bool ProgramSamplesOnlySingleLevelTextures(const MG_State::GLState::ProgramObject& program,
+                                                          const ProgramFactory::VkProgramObject& programObj);
+
     private:
         struct DescriptorPoolBucket {
             VkDescriptorPool handle = VK_NULL_HANDLE;
@@ -65,8 +95,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint32 allocatedSets = 0;
         };
 
+        // A cached descriptor set together with the pool it was allocated from, so a
+        // layout-destroyed purge can vkFreeDescriptorSets it back and credit the
+        // owning bucket's accounting.
+        struct CachedDescriptorSet {
+            VkDescriptorSet set = VK_NULL_HANDLE;
+            VkDescriptorPool pool = VK_NULL_HANDLE;
+        };
+
         struct DescriptorSetCacheEntry {
-            Vector<VkDescriptorSet> sets;
+            Vector<CachedDescriptorSet> sets;
             Uint32 cursor = 0;
         };
 
@@ -116,7 +154,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         };
         Bool ResolveUniformBufferPayload(const MG_State::GLState::ProgramObject& program,
                                          const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
-                                         UboBindResult& out) const;
+                                         Uint32 arrayElement, UboBindResult& out) const;
         Bool CreateDescriptorPool(Uint32 maxSets, VkDescriptorPool& outPool) const;
         Bool GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex);
         VkResult AllocateDescriptorSetsFromActivePool(
@@ -172,10 +210,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint64 samplerLifetimeId = 0;
             Uint64 textureLifetimeId = 0;
             VkSampler sampler = VK_NULL_HANDLE;
+            Uint32 viewLevelCount = 0;
             Uint16 samplerVersion = 0;
             Uint16 textureParamsVersion = 0;
             Bool forceNearestFiltering = false;
             Bool valid = false;
+            // ResolveSampledImageViewFormat is pure in (image format, numeric domain), but a
+            // domain mismatch walks a ~184-entry format table. Memo the resolution per binding
+            // so a reinterpreted sampler pays that scan once, not once per draw.
+            VkFormat viewFormatSource = VK_FORMAT_UNDEFINED;
+            SamplerNumericDomain viewFormatDomain = SamplerNumericDomain::Unknown;
+            VkFormat viewFormat = VK_FORMAT_UNDEFINED;
+            Bool viewFormatValid = false;
         };
         mutable Vector<SamplerResolveMemo> m_samplerResolveMemo;
     };

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.cpp

@@ -32,6 +32,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             XXHASH_VERIFY(XXH64_update(m_hashState, &attr.IsBgra, sizeof(attr.IsBgra)));
             XXHASH_VERIFY(XXH64_update(m_hashState, &attr.Divisor, sizeof(attr.Divisor)));
 
+            // The buffer's heap address is an identity component of the key: a freed
+            // buffer's reused address can alias an old cache entry, but only under a
+            // byte-identical attribute layout - and the entry payload is a pure function
+            // of the hashed inputs, with the draw path re-resolving bindingBufferKeys
+            // against the live VAO attribute pointers, so an aliased hit returns exactly
+            // what a rebuild would. Address drift only grows the map; the OnFrameBoundary
+            // aging sweep bounds that.
             const SizeT bufferKey = reinterpret_cast<SizeT>(attr.Buffer.get());
             XXHASH_VERIFY(XXH64_update(m_hashState, &bufferKey, sizeof(bufferKey)));
         }
@@ -58,6 +65,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const MG_State::GLState::VertexArrayObject& vao, HashType hash) {
         auto it = m_cache.find(hash);
         if (it != m_cache.end()) {
+            it->second.lastUsedFrameBoundary = m_frameBoundaryCounter;
             return it->second;
         }
 
@@ -126,8 +134,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const Bool packedAttribute = attr.Type == DataType::Int2101010Rev ||
                                          attr.Type == DataType::Uint2101010Rev;
             const SizeT requiredAlignment = packedAttribute ? attribByteSize : GetComponentSize(attr.Type);
+            // For a client-memory array attr.Offset holds the raw client pointer, and the
+            // draw path re-uploads the data to a 16-aligned transient slice with attribute
+            // offset 0, so only the stride can violate Vulkan's fetch alignment there.
+            const Bool clientMemoryAttribute = attr.Buffer == nullptr;
             if (conversion == VertexStreamConversion::None && requiredAlignment > 1 &&
-                ((sourceStride % requiredAlignment) != 0 || (attr.Offset % requiredAlignment) != 0)) {
+                ((sourceStride % requiredAlignment) != 0 ||
+                 (!clientMemoryAttribute && (attr.Offset % requiredAlignment) != 0))) {
                 // GL accepts arbitrary byte strides and offsets. Core Vulkan vertex fetches do not
                 // unless VK_EXT_legacy_vertex_attributes is available, so deinterleave this one
                 // attribute into a tightly packed transient stream without changing its format.
@@ -161,6 +174,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         auto& entry = m_cache[hash];
         entry.hash = hash;
+        entry.lastUsedFrameBoundary = m_frameBoundaryCounter;
         entry.bindings = builder.GetBindings();
         entry.attributes = builder.GetAttributes();
         entry.bindingBufferKeys = std::move(bindingBufferKeys);
@@ -175,6 +189,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return entry;
     }
 
+    void VertexInputStateFactory::OnFrameBoundary() {
+        ++m_frameBoundaryCounter;
+
+        // Sweep occasionally; evict entries whose last hit is far in the past.
+        // Erasure happens only here, never mid-frame: the draw path holds a
+        // reference into the current entry across its setup, and unordered_map
+        // erase would invalidate it. Entries are CPU-side only, so no GPU-idle
+        // proof is needed; an evicted entry that is used again is simply rebuilt
+        // from the VAO state (same hash, same content).
+        constexpr Uint64 kSweepInterval = 256;
+        constexpr Uint64 kRetireAgeBoundaries = 1024;
+        if ((m_frameBoundaryCounter % kSweepInterval) != 0) {
+            return;
+        }
+
+        for (auto it = m_cache.begin(); it != m_cache.end();) {
+            if (m_frameBoundaryCounter - it->second.lastUsedFrameBoundary > kRetireAgeBoundaries) {
+                it = m_cache.erase(it);
+            } else {
+                ++it;
+            }
+        }
+    }
+
     VkFormat VertexInputStateFactory::ToVkVertexFormat(DataType type, Int size, Bool normalized, Bool isInteger,
                                                        Bool isBgra) {
         if (isBgra) {

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.h

@@ -27,6 +27,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         struct BackendVertexInputState {
             HashType hash = 0;
+            // Frame boundary of the last cache hit; entries idle past the
+            // OnFrameBoundary retirement age are evicted (CPU heap only).
+            Uint64 lastUsedFrameBoundary = 0;
             Vector<VkVertexInputBindingDescription> bindings;
             Vector<VkVertexInputAttributeDescription> attributes;
             Vector<SizeT> bindingBufferKeys;
@@ -55,6 +58,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const BackendVertexInputState& GetOrCreateVertexInputState(
             const MG_State::GLState::VertexArrayObject& vao, HashType hash);
         const BackendVertexInputState& GetOrCreateVertexInputState(const MG_State::GLState::VertexArrayObject& vao);
+        // Frame boundary hook: ages the cache and evicts entries not hit for many
+        // frames. The key mixes buffer heap addresses, so buffer/VAO churn keeps
+        // minting fresh keys; without eviction the map grows for the whole session.
+        // Entries hold no Vulkan handles (pipeline creation copies the descriptions)
+        // and the draw path's entry reference never spans a frame boundary, so
+        // eviction here needs no GPU-idle proof. Self-gated: one counter bump and
+        // compare except on sweep boundaries.
+        void OnFrameBoundary();
         static SizeT GetComponentSize(DataType type);
         // Tightly-packed byte size of one vertex element for this attribute: componentSize * size for
         // normal types, and 4 (one packed word) for the 2_10_10_10 types and GL_BGRA. Returns 0 for
@@ -70,6 +81,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VulkanRendererConfig& m_config;
         VkPhysicalDevice m_physicalDevice = VK_NULL_HANDLE;
         UnorderedMap<HashType, BackendVertexInputState> m_cache;
+        // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
+        Uint64 m_frameBoundaryCounter = 0;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferManager.cpp

@@ -141,6 +141,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_transientUploadArena.BeginFrame(frameIndex);
     }
 
+    void VkBufferManager::CollectAllDeferredReleases() {
+        for (Uint32 frameIndex = 0; frameIndex < m_deferredBufferReleases.size(); ++frameIndex) {
+            CollectDeferredReleases(frameIndex);
+        }
+        for (Uint32 frameIndex = 0; frameIndex < m_transientUploadArena.GetFrameCount(); ++frameIndex) {
+            m_transientUploadArena.CollectDeferredReleases(frameIndex);
+        }
+    }
+
     void VkBufferManager::NotifyDeviceIdle() {
         // Everything submitted so far has completed. Work recorded for the
         // current frame has not been submitted yet, so the current serial

MobileGL/MG_Backend/DirectVulkan/Renderer/VkBufferManager.h

@@ -77,6 +77,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Recreate all per-frame transient arenas
         Bool RecreateTransientArenas(Uint32 frameCount);
         void BeginFrame(Uint32 frameIndex);
+        // Drains every frame slot's deferred buffer/resource releases (and the
+        // transient arena's parked superseded blocks). Only valid when the
+        // caller has proven every queue submission complete; used by the
+        // present-less frame-boundary drain.
+        void CollectAllDeferredReleases();
         // All previously submitted GPU work has completed (vkDeviceWaitIdle).
         void NotifyDeviceIdle();
         // A frame slot's submission fence has been waited: every serial up to

MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.cpp

@@ -180,6 +180,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         sampleCount = VK_SAMPLE_COUNT_1_BIT;
         internalFormat = TextureInternalFormat::Unknown;
         samples = 0;
+        deadSinceFrame = kNeverObservedDead;
     }
 
     VkRenderPassManager::VkRenderPassManager(VkDevice device,
@@ -206,6 +207,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             resource.Destroy(m_device, m_allocator);
         }
         m_renderbufferResources.clear();
+        CollectDeferredRenderbufferReleases(/*destroyAll=*/true); // caller guarantees device idle
         m_pendingRenderbufferClears.clear();
         RenderPassEntry::s_textureResourcesScratch.clear();
         s_activeRenderPass = {};
@@ -213,22 +215,75 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_rpFastValid = false;
     }
 
-    void VkRenderPassManager::CollectRenderbufferGarbage() {
-        Vector<MG_State::GLState::RenderbufferObject*> deadRenderbuffers;
-        deadRenderbuffers.reserve(m_renderbufferResources.size());
-        for (auto& [renderbuffer, resource] : m_renderbufferResources) {
-            const auto liveRenderbuffer = resource.renderbuffer.lock();
-            if (!liveRenderbuffer || liveRenderbuffer.get() != renderbuffer) {
-                deadRenderbuffers.emplace_back(renderbuffer);
-            }
+    Uint64 VkRenderPassManager::RetireAgeFrames() const {
+        // MaxFramesInFlight + 2 covers the frame ring plus one boundary for the
+        // recording-to-submit gap and one because OnPresent runs ahead of Present's
+        // fence wait; the floor of 8 keeps a margin over the default ring of 3 while
+        // still releasing multi-MB attachment memory promptly (the render-pass cache's
+        // 1024-frame retirement would pin it for no additional safety).
+        return std::max<Uint64>(8, static_cast<Uint64>(m_config.MaxFramesInFlight) + 2);
+    }
+
+    void VkRenderPassManager::DeferRenderbufferBackingRelease(RenderbufferResource& resource) {
+        // The superseded backing may still be referenced by in-flight command buffers
+        // (glRenderbufferStorage can respecify a renderbuffer drawn this very frame),
+        // so it is parked and destroyed only after RetireAgeFrames() boundaries.
+        if (resource.image == VK_NULL_HANDLE && resource.view == VK_NULL_HANDLE) {
+            return;
         }
-        for (auto* renderbuffer : deadRenderbuffers) {
-            auto resourceIt = m_renderbufferResources.find(renderbuffer);
-            if (resourceIt != m_renderbufferResources.end()) {
-                resourceIt->second.Destroy(m_device, m_allocator);
-                m_renderbufferResources.erase(resourceIt);
+        m_deferredRenderbufferReleases.push_back({resource.image, resource.allocation, resource.view, m_frameCounter});
+        resource.image = VK_NULL_HANDLE;
+        resource.allocation = nullptr;
+        resource.view = VK_NULL_HANDLE;
+    }
+
+    void VkRenderPassManager::CollectDeferredRenderbufferReleases(Bool destroyAll) {
+        if (m_deferredRenderbufferReleases.empty()) {
+            return;
+        }
+        const Uint64 retireAgeFrames = RetireAgeFrames();
+        std::erase_if(m_deferredRenderbufferReleases, [&](DeferredRenderbufferRelease& release) {
+            if (!destroyAll && m_frameCounter - release.deferredAtFrame < retireAgeFrames) {
+                return false;
             }
-            m_pendingRenderbufferClears.erase(renderbuffer);
+            if (release.view != VK_NULL_HANDLE) {
+                vkDestroyImageView(m_device, release.view, nullptr);
+            }
+            if (release.image != VK_NULL_HANDLE) {
+                vmaDestroyImage(m_allocator, release.image, release.allocation);
+            }
+            return true;
+        });
+    }
+
+    void VkRenderPassManager::CollectRenderbufferGarbage() {
+        // Two-phase reclamation: a dead renderbuffer's VkImage may still be referenced by
+        // command buffers submitted up to frames-in-flight frames ago (it was legally
+        // attached and drawn right up to its deletion), so the first observation of an
+        // expired weak reference only stamps the current frame counter; Destroy runs once
+        // enough frame boundaries have passed that the stamping frame's submission fence
+        // has provably been waited (see RetireAgeFrames).
+        const Uint64 retireAgeFrames = RetireAgeFrames();
+        for (auto it = m_renderbufferResources.begin(); it != m_renderbufferResources.end();) {
+            auto& resource = it->second;
+            const auto liveRenderbuffer = resource.renderbuffer.lock();
+            if (liveRenderbuffer && liveRenderbuffer.get() == it->first) {
+                resource.deadSinceFrame = RenderbufferResource::kNeverObservedDead;
+                ++it;
+                continue;
+            }
+            if (resource.deadSinceFrame == RenderbufferResource::kNeverObservedDead) {
+                resource.deadSinceFrame = m_frameCounter;
+                ++it;
+                continue;
+            }
+            if (m_frameCounter - resource.deadSinceFrame < retireAgeFrames) {
+                ++it;
+                continue;
+            }
+            m_pendingRenderbufferClears.erase(it->first);
+            resource.Destroy(m_device, m_allocator);
+            it = m_renderbufferResources.erase(it);
         }
     }
 
@@ -251,11 +306,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const auto internalFormat = renderbuffer->GetInternalFormat();
         const VkFormat format = MG_Util::ConvertTextureInternalFormatToVkEnum(internalFormat);
         const VkImageAspectFlags aspect = ResolveImageAspectMaskForFormat(format);
-        if ((aspect & VK_IMAGE_ASPECT_COLOR_BIT) != 0) {
-            MGLOG_E("GetOrCreateRenderbufferResource: color renderbuffer %u is not supported by DirectVulkan render passes yet",
-                    renderbuffer->GetExternalIndex());
-            return nullptr;
-        }
+        // Renderbuffers are never sampled (GL has no way to bind one to a sampler), so the
+        // usage set is attachment + transfer: transfer covers readback (vkCmdCopyImageToBuffer),
+        // BlitFramebuffer, CopyTexImage sources, and out-of-render-pass clear materialization.
+        const VkImageUsageFlags imageUsage =
+            ((aspect & VK_IMAGE_ASPECT_COLOR_BIT) != 0 ? VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT
+                                                       : VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) |
+            VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT;
 
         auto& resource = m_renderbufferResources[renderbuffer.get()];
         const Bool needsCreate =
@@ -268,9 +325,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             resource.samples != renderbuffer->GetSamples();
         if (!needsCreate) {
             resource.renderbuffer = renderbuffer;
+            // A new renderbuffer at a recycled address may adopt a compatible entry that
+            // was already stamped dead; it is alive again, so cancel the aging.
+            resource.deadSinceFrame = RenderbufferResource::kNeverObservedDead;
             return &resource;
         }
 
+        // Respecify: park the old backing for aged destruction instead of destroying
+        // inline - it may still be referenced by in-flight command buffers.
+        DeferRenderbufferBackingRelease(resource);
         resource.Destroy(m_device, m_allocator);
         resource.renderbuffer = renderbuffer;
 
@@ -285,7 +348,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         imageInfo.format = format;
         imageInfo.tiling = VK_IMAGE_TILING_OPTIMAL;
         imageInfo.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
-        imageInfo.usage = VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
+        imageInfo.usage = imageUsage;
         imageInfo.samples = sampleCount;
         imageInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
 
@@ -386,6 +449,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     void VkRenderPassManager::QueueRenderbufferClear(
         GLbitfield mask, const ClearFramebufferPayload& clearPayload,
         const MG_State::GLState::FramebufferObject& drawFbo) {
+        if ((mask & GL_COLOR_BUFFER_BIT) != 0) {
+            // Color renderbuffer draw buffers take the framebuffer-level clear too; texture
+            // attachments are skipped by the per-attachment overload's IsRenderbuffer guard.
+            for (const auto attachmentType : drawFbo.GetDrawBuffers()) {
+                if (attachmentType == FramebufferAttachmentType::None) {
+                    continue;
+                }
+                QueueRenderbufferClear(
+                    ClearAttachmentPayload{.mask = GL_COLOR_BUFFER_BIT, .color = clearPayload.color},
+                    drawFbo.GetAttachment(attachmentType));
+            }
+        }
         if ((mask & GL_DEPTH_BUFFER_BIT) != 0) {
             QueueRenderbufferClear(
                 ClearAttachmentPayload{.mask = GL_DEPTH_BUFFER_BIT, .depth = clearPayload.depth},
@@ -682,6 +757,83 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // assuming default FBO has the right param
         for (Uint32 i = 0; i < colorAttachmentSlotCount; ++i) {
             auto drawbuf = drawbufs[i];
+
+            // Renderbuffer color attachments mirror the texture path below, with the
+            // resource (image/view/format/layout) coming from the render-pass manager's
+            // renderbuffer store instead of the texture manager.
+            if (drawbuf != FramebufferAttachmentType::None && !isDefaultFbo) {
+                const auto& rbAtt = fbo.GetAttachment(drawbuf);
+                if (rbAtt.IsRenderbuffer() && rbAtt.IsComplete()) {
+                    const auto& renderbuffer = rbAtt.GetRenderbuffer();
+                    auto* rbResource = GetOrCreateRenderbufferResource(renderbuffer);
+                    if (rbResource == nullptr || (rbResource->aspect & VK_IMAGE_ASPECT_COLOR_BIT) == 0) {
+                        MGLOG_E("GetOrCreateRenderPass: draw buffer slot %u on FBO %u has an unsupported color "
+                                "renderbuffer %u; using VK_ATTACHMENT_UNUSED",
+                                i, fbo.GetExternalIndex(), renderbuffer->GetExternalIndex());
+                        continue;
+                    }
+
+                    const Uint32 rbAttachmentIndex = static_cast<Uint32>(attachmentDescriptions.size());
+                    attachmentDescriptions.emplace_back();
+                    VkAttachmentDescription& rbDesc = attachmentDescriptions.back();
+
+                    ClearAttachmentPayload rbClearPayload{};
+                    Bool rbHasClear = GetPendingRenderbufferClear(renderbuffer.get(), rbClearPayload) &&
+                                      (rbClearPayload.mask & GL_COLOR_BUFFER_BIT) != 0;
+                    if (rbHasClear &&
+                        MG_Util::GetBaseInternalFormatComponentCount(renderbuffer->GetInternalFormat()) == 3) {
+                        // RGB renderbuffers are backed by an RGBA image; the missing alpha reads as 1.
+                        rbClearPayload.color =
+                            FloatVec4(rbClearPayload.color.x(), rbClearPayload.color.y(),
+                                      rbClearPayload.color.z(), 1.0f);
+                    }
+
+                    const VkImageLayout trackedRbLayout = rbResource->layout;
+                    rbDesc.flags = 0;
+                    rbDesc.format = rbResource->format;
+                    rbDesc.samples = rbResource->sampleCount;
+                    rbDesc.loadOp = rbHasClear ? VK_ATTACHMENT_LOAD_OP_CLEAR :
+                                    (trackedRbLayout == VK_IMAGE_LAYOUT_UNDEFINED ? VK_ATTACHMENT_LOAD_OP_DONT_CARE
+                                                                                  : VK_ATTACHMENT_LOAD_OP_LOAD);
+                    rbDesc.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
+                    rbDesc.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
+                    rbDesc.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE;
+                    rbDesc.initialLayout = (rbHasClear || trackedRbLayout == VK_IMAGE_LAYOUT_UNDEFINED) ?
+                        VK_IMAGE_LAYOUT_UNDEFINED : trackedRbLayout;
+                    rbDesc.finalLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
+                    adoptRenderPassSampleCount(rbResource->sampleCount, "color",
+                                               static_cast<Int>(renderbuffer->GetExternalIndex()));
+
+                    if (rbHasClear) {
+                        pendingClearAttachments.emplace_back(PendingClearAttachmentInfo {
+                            .attachmentIndex = rbAttachmentIndex,
+                            .colorAttachmentSlot = i,
+                            .renderbuffer = renderbuffer.get(),
+                            .hasInlinePayload = true,
+                            .inlinePayload = rbClearPayload,
+                        });
+                    }
+
+                    if (width == 0)
+                        width = static_cast<Int>(rbResource->extent.width);
+                    if (height == 0)
+                        height = static_cast<Int>(rbResource->extent.height);
+
+                    trackedAttachmentLayouts.emplace_back(TrackedAttachmentLayoutInfo {
+                        .target = TrackedAttachmentTarget::Renderbuffer,
+                        .renderbuffer = renderbuffer,
+                        .finalLayout = rbDesc.finalLayout,
+                    });
+                    textureResources.emplace_back(nullptr);
+                    attachmentViews.emplace_back(rbResource->view);
+                    MOBILEGL_ASSERT(attachmentViews.back() != VK_NULL_HANDLE,
+                                    "GetOrCreateRenderPass: renderbuffer view missing at color attachment %d", i);
+
+                    colorAttachmentRefs[i].attachment = rbAttachmentIndex;
+                    continue;
+                }
+            }
+
             auto* texture = ResolveCompleteColorAttachmentTexture(fbo, drawbuf, i);
             if (texture == nullptr)
                 continue;
@@ -700,6 +852,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 case TextureTarget::Texture2D:
                 case TextureTarget::Texture2DArray:
                 case TextureTarget::Texture2DMultisample:
+                case TextureTarget::Texture2DMultisampleArray:
+                case TextureTarget::Texture3D:
+                case TextureTarget::TextureCubeMap:
+                case TextureTarget::TextureCubeMapArray:
                 case TextureTarget::TextureRectangle: {
                     desc.flags = 0;
                     desc.format = isDefaultFbo ?
@@ -1083,6 +1239,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     void VkRenderPassManager::OnPresent() {
         ++m_frameCounter;
 
+        // Runs every frame boundary, ahead of the render-pass sweep gate below: the walk
+        // is O(#renderbuffer resources) — single digits in practice — and per-frame
+        // invocation keeps dead-resource reclaim latency at the aging bound instead of
+        // coupling it to renderbuffer *use* (the GetOrCreateRenderbufferResource call
+        // site never runs again once an app stops using renderbuffers).
+        CollectRenderbufferGarbage();
+        CollectDeferredRenderbufferReleases(/*destroyAll=*/false);
+
         // Sweep occasionally; evict entries whose last use is far past every
         // in-flight frame so their VkRenderPass/VkFramebuffer can be destroyed
         // safely (RenderPassEntry's destructor releases the handles).
@@ -1092,6 +1256,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return;
         }
 
+        // Collect the dying handles and notify once after the loop: pipelines hashed
+        // on them share the entries' >kRetireAgeFrames idleness (they are only bound
+        // by draws that hit those entries), so the observer may destroy them
+        // immediately - and a single batched notification costs one pipeline-cache
+        // scan instead of one per evicted pass.
+        Vector<VkRenderPass> destroyedRenderPasses;
         const Uint64 activeHash = s_hasActiveRenderPass ? s_activeRenderPass.hash : 0;
         for (auto it = m_renderPasses.begin(); it != m_renderPasses.end();) {
             const Bool isActive = s_hasActiveRenderPass && it->first == activeHash;
@@ -1099,11 +1269,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 if (m_rpFastValid && m_rpFastRenderPassHash == it->first) {
                     m_rpFastValid = false;
                 }
+                destroyedRenderPasses.push_back(it->second.renderPass);
                 it = m_renderPasses.erase(it);
             } else {
                 ++it;
             }
         }
+        if (!destroyedRenderPasses.empty() && m_evictionObserver != nullptr) {
+            m_evictionObserver->OnRenderPassesDestroyed(destroyedRenderPasses);
+        }
     }
 
     Bool VkRenderPassManager::BeginRenderPass(VkCommandBuffer commandBuffer, RenderPassEntry& renderPassEntry) {

MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.h

@@ -157,11 +157,31 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     class VkRenderPassManager {
     public:
         using HashType = Uint64;
+
+        // Notified once per OnPresent sweep with every aged-out entry's VkRenderPass
+        // value: pipelines are hashed on the raw handle, and once destroyed the value
+        // may be recycled for an incompatible pass, so dependent caches must purge
+        // everything keyed on them before any new pass can be created (the sweep and
+        // the notification run back-to-back with no creation in between; observers
+        // compare the values, never dereference them). Batched so a mass-idle cohort
+        // (shader-pack switch, dimension exit) costs the observer one pipeline-cache
+        // scan, not one per dying pass. The wholesale paths
+        // (Shutdown/RecreateSwapchain) do not notify - their callers already drop
+        // every pipeline outright.
+        class IEvictionObserver {
+        public:
+            virtual ~IEvictionObserver() = default;
+            virtual void OnRenderPassesDestroyed(const Vector<VkRenderPass>& renderPasses) = 0;
+        };
+
         VkRenderPassManager(VkDevice device,
             VkPhysicalDevice physicalDevice, VmaAllocator allocator, const VulkanRendererConfig& config,
             VkClearManager& clearManager, VkTextureManager& textureManager, SwapchainObject& swapchainObject);
         ~VkRenderPassManager();
 
+        // Observer may be null (no notifications). Not owned.
+        void SetEvictionObserver(IEvictionObserver* observer) { m_evictionObserver = observer; }
+
         Bool Initialize();
         void Shutdown();
 
@@ -192,6 +212,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         UnorderedMap<Uint64, RenderPassEntry> m_renderPasses;
         // Monotonic frame counter (bumped in OnPresent) for render-pass cache aging.
         Uint64 m_frameCounter = 0;
+        IEvictionObserver* m_evictionObserver = nullptr;
 
         // Bumped whenever a renderbuffer VkImage is (re)created; together with the texture
         // manager's image epoch this invalidates the render-pass fast path on any attachment
@@ -211,7 +232,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Uint64 m_rpFastRbEpoch = 0;
         Uint64 m_rpFastRenderPassHash = 0;
 
+    public:
         struct RenderbufferResource {
+            // deadSinceFrame sentinel: the owning weak reference has not been observed
+            // expired. Dead resources age past every in-flight frame before Destroy
+            // (see CollectRenderbufferGarbage); the GPU may still reference the image
+            // for frames-in-flight frames after the GL object dies.
+            static constexpr Uint64 kNeverObservedDead = UINT64_MAX;
+
             WeakPtr<MG_State::GLState::RenderbufferObject> renderbuffer;
             VkImage image = VK_NULL_HANDLE;
             VmaAllocation allocation = nullptr;
@@ -223,25 +251,47 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkSampleCountFlagBits sampleCount = VK_SAMPLE_COUNT_1_BIT;
             TextureInternalFormat internalFormat = TextureInternalFormat::Unknown;
             Int samples = 0;
+            // m_frameCounter value at which the weak reference was first seen expired.
+            Uint64 deadSinceFrame = kNeverObservedDead;
 
             void Destroy(VkDevice device, VmaAllocator allocator);
         };
 
+        // Public so the renderer's blit/copy/readback bindings can source renderbuffer
+        // attachments the same way texture attachments go through the texture manager.
+        RenderbufferResource* GetOrCreateRenderbufferResource(
+            const SharedPtr<MG_State::GLState::RenderbufferObject>& renderbuffer);
+        Bool GetPendingRenderbufferClear(MG_State::GLState::RenderbufferObject* renderbuffer,
+                                         ClearAttachmentPayload& outPayload) const;
+
+    private:
         struct PendingRenderbufferClear {
             WeakPtr<MG_State::GLState::RenderbufferObject> renderbuffer;
             ClearAttachmentPayload payload{};
         };
 
+        // A superseded renderbuffer backing (glRenderbufferStorage respecify) parked
+        // until enough frame boundaries have passed that no in-flight command buffer
+        // can still reference it; destroyed in OnPresent (see RetireAgeFrames).
+        struct DeferredRenderbufferRelease {
+            VkImage image = VK_NULL_HANDLE;
+            VmaAllocation allocation = nullptr;
+            VkImageView view = VK_NULL_HANDLE;
+            Uint64 deferredAtFrame = 0;
+        };
+
         UnorderedMap<MG_State::GLState::RenderbufferObject*, RenderbufferResource> m_renderbufferResources;
         UnorderedMap<MG_State::GLState::RenderbufferObject*, PendingRenderbufferClear> m_pendingRenderbufferClears;
+        Vector<DeferredRenderbufferRelease> m_deferredRenderbufferReleases;
 
-        RenderbufferResource* GetOrCreateRenderbufferResource(
-            const SharedPtr<MG_State::GLState::RenderbufferObject>& renderbuffer);
-        Bool GetPendingRenderbufferClear(MG_State::GLState::RenderbufferObject* renderbuffer,
-                                         ClearAttachmentPayload& outPayload) const;
         Bool HasPendingRenderbufferClear(
             const MG_State::GLState::FramebufferAttachmentObject& attachment) const;
         void CollectRenderbufferGarbage();
+        // Frame-boundary margin after which a resource last referenced by a retired
+        // GL object (or superseded backing) is provably past every in-flight frame.
+        Uint64 RetireAgeFrames() const;
+        void DeferRenderbufferBackingRelease(RenderbufferResource& resource);
+        void CollectDeferredRenderbufferReleases(Bool destroyAll);
 
         static inline XXH64_state_t* m_hashState = XXH64_createState();
         static inline ActiveRenderPassInfo s_activeRenderPass{};

MobileGL/MG_Backend/DirectVulkan/Renderer/VkSamplerManager.cpp

@@ -51,6 +51,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Float ResolveEffectiveMinLod(const MG_State::GLState::SamplerObject& sampler, Float effectiveMaxLod) {
             return std::min(sampler.GetMinLod(), effectiveMaxLod);
         }
+
+        // A single-level view can only ever deliver the base level, but the LOD clamp must not be
+        // collapsed to exactly 0: both GL and Vulkan pick magFilter over minFilter from the
+        // *clamped* lambda, so maxLod = 0 would make every fragment magnify and quietly retire the
+        // min filter. 0.25 is the value VkSamplerCreateInfo's own note prescribes for emulating
+        // GL's non-mipmapped minification - large enough for lambda to stay positive, small enough
+        // that a NEAREST mip mode still rounds down to level 0. Clamped rather than assigned, so a
+        // texture whose GL_TEXTURE_MAX_LOD really is 0 keeps magnifying as GL says it must.
+        Float ResolveSingleLevelMaxLod(const MG_State::GLState::SamplerObject& sampler, Bool singleLevelView) {
+            const Float maxLod = ResolveEffectiveMaxLod(sampler);
+            return singleLevelView ? std::min(maxLod, 0.25f) : maxLod;
+        }
     } // namespace
 
     Bool VkSamplerManager::Initialize(const InitInfo& initInfo) {
@@ -89,15 +101,43 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         m_device = VK_NULL_HANDLE;
         m_config = nullptr;
+        m_frameBoundaryCounter = 0;
+    }
+
+    void VkSamplerManager::OnFrameBoundary() {
+        ++m_frameBoundaryCounter;
+
+        // Sweep occasionally; destroy samplers whose last use is far past every
+        // in-flight frame. Destroy and erase must stay atomic, or Shutdown would
+        // double-free the handle; an evicted key that recurs simply re-creates
+        // its sampler on the next miss.
+        constexpr Uint64 kSweepInterval = 256;
+        constexpr Uint64 kRetireAgeBoundaries = 1024;
+        if ((m_frameBoundaryCounter % kSweepInterval) != 0) {
+            return;
+        }
+
+        for (auto it = m_samplers.begin(); it != m_samplers.end();) {
+            auto& entry = it->second;
+            if (m_frameBoundaryCounter - entry.lastUsedFrameBoundary > kRetireAgeBoundaries) {
+                if (m_device != VK_NULL_HANDLE && entry.handle != VK_NULL_HANDLE) {
+                    vkDestroySampler(m_device, entry.handle, nullptr);
+                }
+                it = m_samplers.erase(it);
+            } else {
+                ++it;
+            }
+        }
     }
 
     Uint64 VkSamplerManager::BuildSamplerKey(const MG_State::GLState::SamplerObject& sampler,
                                              const MG_State::GLState::ITextureObject& texture,
-                                             Bool forceNearestFiltering) const {
+                                             Bool forceNearestFiltering, Bool singleLevelView) const {
         MOBILEGL_ASSERT(m_config != nullptr, "VkSamplerManager::BuildSamplerKey: m_config is null");
         XXHASH_VERIFY(XXH64_reset(m_hashState, m_config->CacheVersion));
 
         XXHASH_VERIFY(XXH64_update(m_hashState, &forceNearestFiltering, sizeof(forceNearestFiltering)));
+        XXHASH_VERIFY(XXH64_update(m_hashState, &singleLevelView, sizeof(singleLevelView)));
 
         const auto minFilter = sampler.GetMinFilter();
         XXHASH_VERIFY(XXH64_update(m_hashState, &minFilter, sizeof(minFilter)));
@@ -111,7 +151,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         XXHASH_VERIFY(XXH64_update(m_hashState, &wrapT, sizeof(wrapT)));
         const auto wrapR = sampler.GetWrapR();
         XXHASH_VERIFY(XXH64_update(m_hashState, &wrapR, sizeof(wrapR)));
-        const auto maxLod = ResolveEffectiveMaxLod(sampler);
+        const auto maxLod = ResolveSingleLevelMaxLod(sampler, singleLevelView);
         const auto minLod = ResolveEffectiveMinLod(sampler, maxLod);
         XXHASH_VERIFY(XXH64_update(m_hashState, &minLod, sizeof(minLod)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &maxLod, sizeof(maxLod)));
@@ -133,10 +173,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     VkSampler VkSamplerManager::GetOrCreateSampler(const MG_State::GLState::SamplerObject& sampler,
                                                    const MG_State::GLState::ITextureObject& texture,
-                                                   Bool forceNearestFiltering) {
-        const Uint64 key = BuildSamplerKey(sampler, texture, forceNearestFiltering);
+                                                   Bool forceNearestFiltering, Uint32 viewLevelCount) {
+        // A view that exposes a single mip level has no second level to blend with, so GL's
+        // *_MIPMAP_* minification filters degenerate to plain filtering on the base level -
+        // sampling is unchanged by pinning the Vulkan sampler to NEAREST mip mode at LOD 0.
+        // It is not cosmetic: MobileGL backs such a view with a fully allocated mip chain whose
+        // tail is never written, and a LINEAR mip mode lets the texture unit issue the level+1
+        // fetch anyway. On Adreno that fetch lands in uninitialized UBWC pages (or past the
+        // allocation for a genuinely single-level image) and faults the GPU - the same failure
+        // the default-framebuffer blit shader had to work around with an explicit-LOD sample.
+        const Bool singleLevelView = viewLevelCount == 1;
+        const Uint64 key = BuildSamplerKey(sampler, texture, forceNearestFiltering, singleLevelView);
         auto it = m_samplers.find(key);
         if (it != m_samplers.end()) {
+            it->second.lastUsedFrameBoundary = m_frameBoundaryCounter;
             return it->second.handle;
         }
 
@@ -144,8 +194,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         samplerInfo.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO;
         samplerInfo.magFilter = forceNearestFiltering ? VK_FILTER_NEAREST : ToVkFilter(sampler.GetMagFilter());
         samplerInfo.minFilter = forceNearestFiltering ? VK_FILTER_NEAREST : ToVkFilter(sampler.GetMinFilter());
-        samplerInfo.mipmapMode = forceNearestFiltering ? VK_SAMPLER_MIPMAP_MODE_NEAREST
-                                                       : ToVkMipmapMode(sampler.GetMipmapMode());
+        samplerInfo.mipmapMode = (forceNearestFiltering || singleLevelView)
+                                     ? VK_SAMPLER_MIPMAP_MODE_NEAREST
+                                     : ToVkMipmapMode(sampler.GetMipmapMode());
         samplerInfo.addressModeU = ToVkAddressMode(sampler.GetWrapS());
         samplerInfo.addressModeV = ToVkAddressMode(sampler.GetWrapT());
         samplerInfo.addressModeW = ToVkAddressMode(sampler.GetWrapR());
@@ -157,7 +208,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         samplerInfo.maxAnisotropy = maxAnisotropy;
         samplerInfo.compareEnable = sampler.GetCompareMode() == SamplerCompareMode::CompareToTexture ? VK_TRUE : VK_FALSE;
         samplerInfo.compareOp = ToVkCompareOp(ResolveCompareFunc(sampler, texture));
-        samplerInfo.maxLod = ResolveEffectiveMaxLod(sampler);
+        // Must match BuildSamplerKey's resolution exactly.
+        samplerInfo.maxLod = ResolveSingleLevelMaxLod(sampler, singleLevelView);
         samplerInfo.minLod = ResolveEffectiveMinLod(sampler, samplerInfo.maxLod);
         samplerInfo.borderColor = ResolveVkBorderColor(sampler, texture);
         samplerInfo.unnormalizedCoordinates = VK_FALSE;
@@ -169,6 +221,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         entry.handle = vkSampler;
         entry.externalIndex = sampler.GetExternalIndex();
         entry.version = sampler.GetVersion();
+        entry.lastUsedFrameBoundary = m_frameBoundaryCounter;
         m_samplers[key] = entry;
         return vkSampler;
     }

MobileGL/MG_Backend/DirectVulkan/Renderer/VkSamplerManager.h

@@ -33,20 +33,38 @@ public:
     Bool Initialize(const InitInfo& initInfo);
     void Shutdown();
 
+    // viewLevelCount is the mip-level count of the image view this sampler will be paired
+    // with; 0 means "unknown, do not narrow". See GetOrCreateSampler for why it matters.
     VkSampler GetOrCreateSampler(const MG_State::GLState::SamplerObject& sampler,
                                  const MG_State::GLState::ITextureObject& texture,
-                                 Bool forceNearestFiltering = false);
+                                 Bool forceNearestFiltering = false,
+                                 Uint32 viewLevelCount = 0);
+    // Frame boundary hook: ages the sampler cache and destroys samplers not used
+    // for many frames. The key hashes continuous float state (lodBias, LOD clamps,
+    // anisotropy), so an app animating those would otherwise mint an unbounded
+    // stream of never-destroyed VkSamplers and eventually exhaust the device's
+    // maxSamplerAllocationCount. A sampler idle for over a thousand frame
+    // boundaries cannot be referenced by any in-flight command buffer (frames in
+    // flight are single digits), and every descriptor set the GPU consumes is
+    // written that same frame with live handles (the per-binding resolve memo and
+    // descriptor-set reuse are both frame-reset), so destruction here needs no
+    // fence wait. Self-gated: one counter bump and compare except on sweep
+    // boundaries.
+    void OnFrameBoundary();
 
 private:
     struct SamplerCacheEntry {
         VkSampler handle = VK_NULL_HANDLE;
         Uint externalIndex = 0;
         Uint16 version = 0;
+        // Frame boundary of the last cache hit; entries idle past the
+        // OnFrameBoundary retirement age have their VkSampler destroyed.
+        Uint64 lastUsedFrameBoundary = 0;
     };
 
     Uint64 BuildSamplerKey(const MG_State::GLState::SamplerObject& sampler,
                            const MG_State::GLState::ITextureObject& texture,
-                           Bool forceNearestFiltering) const;
+                           Bool forceNearestFiltering, Bool singleLevelView) const;
     static VkFilter ToVkFilter(SamplerFilterMode mode);
     static VkSamplerMipmapMode ToVkMipmapMode(SamplerMipmapMode mode);
     static VkSamplerAddressMode ToVkAddressMode(SamplerWrapMode mode);
@@ -67,6 +85,8 @@ private:
     Bool m_samplerAnisotropySupported = false;
     Float m_maxSamplerAnisotropy = 1.0f;
     UnorderedMap<Uint64, SamplerCacheEntry> m_samplers;
+    // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
+    Uint64 m_frameBoundaryCounter = 0;
     static inline XXH64_state_t* m_hashState = XXH64_createState();
 };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.cpp

@@ -375,7 +375,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         switch (format) {
         case TextureInternalFormat::RGB:
         case TextureInternalFormat::RGB8:
+        // Legacy low-bit RGB formats share the UNorm8 canonical shadow layout (see
+        // TextureFormatProcessor), so they upload exactly like RGB8 with an alpha expand.
+        case TextureInternalFormat::R3G3B2:
+        case TextureInternalFormat::RGB4:
+        case TextureInternalFormat::RGB5:
             return {VK_FORMAT_R8G8B8A8_UNORM, true, 1, {0xFF, 0x00, 0x00, 0x00}};
+        // Low-bit RGBA formats: UNorm8x4 canonical shadow, no expansion needed.
+        case TextureInternalFormat::RGBA2:
+        case TextureInternalFormat::RGBA4:
+        case TextureInternalFormat::RGB5A1:
+            return {VK_FORMAT_R8G8B8A8_UNORM, false, 0, {0, 0, 0, 0}};
+        // 10/12-bit RGB(A): UNorm16 canonical shadow.
+        case TextureInternalFormat::RGB10:
+        case TextureInternalFormat::RGB12:
+            return {VK_FORMAT_R16G16B16A16_UNORM, true, 2, {0xFF, 0xFF, 0x00, 0x00}};
+        case TextureInternalFormat::RGBA12:
+            return {VK_FORMAT_R16G16B16A16_UNORM, false, 0, {0, 0, 0, 0}};
         case TextureInternalFormat::SRGB8:
             return {VK_FORMAT_R8G8B8A8_SRGB, true, 1, {0xFF, 0x00, 0x00, 0x00}};
         case TextureInternalFormat::RGB8Snorm:
@@ -571,6 +587,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_allocator = initInfo.allocator;
         m_commandPool = initInfo.commandPool;
         m_graphicsQueue = initInfo.graphicsQueue;
+        m_imageFormatListSupported = initInfo.imageFormatListSupported;
         m_currentFrameIndex = 0;
         m_deferredReleases.clear();
         m_deferredReleases.resize(initInfo.frameCount);
@@ -593,6 +610,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         DestroyDeferredReleases();
         m_textureResources.clear();
         m_aliveObjects.clear();
+        m_storageImageTextures.clear();
 
         m_device = VK_NULL_HANDLE;
         m_physicalDevice = VK_NULL_HANDLE;
@@ -611,6 +629,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                         frameIndex, m_deferredViewReleases.size());
         m_currentFrameIndex = frameIndex;
         CollectDeferredReleases(frameIndex);
+
+        // Frame-boundary GC: every 64 frame boundaries (~1 s at 60 fps) bounds the reclaim
+        // latency for dead textures regardless of draw traffic — workloads that churn
+        // textures through clears/readbacks alone never reach the draw-gated
+        // CollectGarbage. Must run after CollectDeferredReleases above: the prune defers
+        // its releases into this frame's slot, which was just drained, so they are
+        // destroyed only after the slot's fence has been waited again one full frame-ring
+        // cycle from now (never while an in-flight frame may still reference them).
+        constexpr Uint32 kGcFrameInterval = 64;
+        ++m_gcFrameCounter;
+        if (m_gcFrameCounter % kGcFrameInterval == 0) {
+            PruneDeadTextures();
+        }
+    }
+
+    void VkTextureManager::CollectAllDeferredReleases() {
+        const SizeT frameCount = std::min(m_deferredReleases.size(), m_deferredViewReleases.size());
+        for (SizeT frameIndex = 0; frameIndex < frameCount; ++frameIndex) {
+            CollectDeferredReleases(static_cast<Uint32>(frameIndex));
+        }
     }
 
     void VkTextureManager::EraseTrackedTexture(const TextureIdentity& identity) {
@@ -620,6 +658,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_textureResources.erase(resourceIt);
         }
         m_aliveObjects.erase(identity);
+        m_storageImageTextures.erase(identity);
     }
 
     void VkTextureManager::PruneStaleTextureAliases(MG_State::GLState::ITextureObject* texture) {
@@ -686,9 +725,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // construction introduces a new identity. Doing this unconditionally made every
         // sampled-texture sync scan the entire alive-texture map per draw.
         if (aliveIt == m_aliveObjects.end()) {
+            WeakPtr<MG_State::GLState::ITextureObject> aliveTexture;
             const auto& liveTexture = MG_State::pGLContext->GetTextureObject(texture.GetExternalIndex());
             if (liveTexture && liveTexture.get() == &texture) {
-                m_aliveObjects[identity] = WeakPtr<MG_State::GLState::ITextureObject>(liveTexture);
+                aliveTexture = liveTexture;
+            } else {
+                // The name lookup legally fails while the object is alive: the name was
+                // deleted with the texture still attached to an FBO (the attachment's
+                // SharedPtr keeps it alive), or the name was reused by a new texture, or
+                // this is a default texture object (name 0 lives outside the name map).
+                // Register through the object's own control block so the resource created
+                // below still participates in weak-expiry GC instead of becoming an
+                // orphan no reclamation path can reach until Shutdown.
+                aliveTexture = texture.weak_from_this();
+            }
+            if (!aliveTexture.expired()) {
+                m_aliveObjects[identity] = Move(aliveTexture);
                 PruneStaleTextureAliases(&texture);
             }
         }
@@ -1140,6 +1192,47 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return ok;
     }
 
+    void VkTextureManager::MarkStorageImageTexture(MG_State::GLState::ITextureObject& texture) {
+        m_storageImageTextures.insert(MakeTextureIdentity(&texture));
+    }
+
+    Bool VkTextureManager::NeedsStorageUsageUpgrade(MG_State::GLState::ITextureObject& texture) const {
+        const TextureIdentity identity = MakeTextureIdentity(&texture);
+        if (m_storageImageTextures.find(identity) == m_storageImageTextures.end()) {
+            return false;
+        }
+        const auto it = m_textureResources.find(identity);
+        // No image yet: the first sync creates it with STORAGE straight away, so there is nothing
+        // to preserve and nothing to order against.
+        return it != m_textureResources.end() && it->second.image != VK_NULL_HANDLE &&
+               !it->second.storageUsageResolved;
+    }
+
+    Bool VkTextureManager::NeedsStorageImagePreparation(MG_State::GLState::ITextureObject& texture) const {
+        const TextureIdentity identity = MakeTextureIdentity(&texture);
+        const auto it = m_textureResources.find(identity);
+        if (it == m_textureResources.end()) {
+            return true;
+        }
+        const TextureResource& resource = it->second;
+        if (resource.image == VK_NULL_HANDLE || resource.layout != VK_IMAGE_LAYOUT_GENERAL) {
+            return true;
+        }
+        // The image predates this texture's first image-unit binding, so it was created without
+        // STORAGE usage and has to be recreated - which is illegal inside a render pass.
+        if (!resource.storageUsageResolved &&
+            m_storageImageTextures.find(identity) != m_storageImageTextures.end()) {
+            return true;
+        }
+        // Mirror SyncTexture's cross-draw skip condition: any version drift means the sync
+        // path may upload or rebuild, both of which need the render pass ended first.
+        const auto* mipTexture = MG_State::GLState::AsMipmapTexture(&texture);
+        const Uint32 mipLevelCount = mipTexture != nullptr ? mipTexture->GetMipmapLevelCount() : 0u;
+        return resource.syncedContentVersion != texture.GetContentVersion() ||
+               resource.syncedTextureParamsVersion != texture.GetTextureParamsVersion() ||
+               resource.syncedMipLevelCount != mipLevelCount;
+    }
+
     Bool VkTextureManager::TransitionImageLayout(VkCommandBuffer commandBuffer, VkImage image,
                                                  VkImageLayout& trackedLayout, VkImageLayout newLayout,
                                                  VkPipelineStageFlags srcStageMask, VkPipelineStageFlags dstStageMask,
@@ -1178,10 +1271,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     }
 
     SizeT VkTextureManager::CollectGarbage() {
+        // Draw-gated stagger (1 in 256 calls): keeps the per-draw cost at one counter
+        // bump. The guaranteed reclaim path is the frame-boundary prune in BeginFrame;
+        // this remains as a cheap assist so draw-heavy workloads reclaim sooner.
         m_gcCounter++;
         if (m_gcCounter != 0) {
             return 0;
         }
+        return PruneDeadTextures();
+    }
+
+    SizeT VkTextureManager::PruneDeadTextures() {
+        // Erasing entries would dangle the raw TextureResource pointers memoized for the
+        // current draw; every call path (BeginFrame, and CollectGarbage at the top of a
+        // freshly opened draw-sync scope) runs before any memo entry is recorded.
+        MOBILEGL_ASSERT(m_drawSyncedThisDraw.empty(),
+                        "PruneDeadTextures: draw-sync memo holds raw resource pointers an erase would dangle");
 
         Vector<MG_State::GLState::ITextureObject*> expiredTextures;
         expiredTextures.reserve(m_aliveObjects.size());
@@ -1193,7 +1298,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         for (auto* texture : expiredTextures) {
             PruneStaleTextureAliases(texture);
         }
-        return expiredTextures.size();
+        SizeT prunedCount = expiredTextures.size();
+
+        // Orphan sweep: after the pass above, m_aliveObjects holds only live entries.
+        // Registration in SyncTextureAndGetDescriptor cannot fail for a SharedPtr-owned
+        // texture (weak_from_this fallback), so a resource whose identity has no alive
+        // entry has no trackable owner: its GL-side object is gone, or was never
+        // shared-owned, in which case recreation on a later sync is the safe fallback.
+        // Destruction goes through the per-frame deferred queues, never immediate.
+        Vector<TextureIdentity> orphanIdentities;
+        for (auto it = m_textureResources.begin(); it != m_textureResources.end(); ++it) {
+            if (m_aliveObjects.find(it->first) == m_aliveObjects.end()) {
+                orphanIdentities.emplace_back(it->first);
+            }
+        }
+        for (const auto& identity : orphanIdentities) {
+            EraseTrackedTexture(identity);
+        }
+        prunedCount += orphanIdentities.size();
+        return prunedCount;
     }
 
     Bool VkTextureManager::SyncTexture(MG_State::GLState::ITextureObject &texture,
@@ -1207,7 +1330,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const auto* syncingMipTexture = MG_State::GLState::AsMipmapTexture(&texture);
         const Uint32 syncingMipLevelCount =
             syncingMipTexture != nullptr ? syncingMipTexture->GetMipmapLevelCount() : 0u;
-        if (outResource.image != VK_NULL_HANDLE &&
+        // A pending storage-usage upgrade also has to bust the skip: nothing about the texture's
+        // content or params changed, but the image itself must be recreated with STORAGE usage
+        // before it can back an image-unit descriptor.
+        const Bool storageUpgradePending =
+            !outResource.storageUsageResolved &&
+            m_storageImageTextures.find(MakeTextureIdentity(&texture)) != m_storageImageTextures.end();
+        if (outResource.image != VK_NULL_HANDLE && !storageUpgradePending &&
             outResource.syncedContentVersion == syncingContentVersion &&
             outResource.syncedTextureParamsVersion == texture.GetTextureParamsVersion() &&
             outResource.syncedMipLevelCount == syncingMipLevelCount) {
@@ -1318,15 +1447,44 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkImageAspectFlags aspect = GetAspectMaskForFormat(format);
         VkFormatProperties formatProperties{};
         vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &formatProperties);
-        const Bool supportsStorageImage =
+        // Only textures that have actually been bound to a GL image unit get STORAGE usage (and
+        // the MUTABLE_FORMAT it drags in for format-reinterpreting image views). Requesting it
+        // for every storage-capable colour texture costs real bandwidth: Adreno cannot keep UBWC
+        // compression on an image that may be written through a storage descriptor, so the whole
+        // render target - MC's included - runs uncompressed. MarkStorageImageTexture upgrades a
+        // texture before its first image-unit draw, and the usage below feeds the compatibility
+        // check so the upgrade recreates the image.
+        const Bool markedAsStorageImage =
+            m_storageImageTextures.find(MakeTextureIdentity(
+                const_cast<MG_State::GLState::ITextureObject*>(&texture))) != m_storageImageTextures.end();
+        // Storage-image CAPABILITY (does the format allow it at all) is deliberately separate from
+        // whether this texture actually needs the usage. MUTABLE_FORMAT keys off capability, as
+        // before: format-reinterpreting views are not a storage-only concern - the SAMPLED path
+        // needs them too (GetOrCreateSampledImageView bails out without it, see ~line 892), so
+        // tying MUTABLE_FORMAT to the image-unit mark would break sampled format reinterpretation
+        // for every texture that never becomes a storage image.
+        const Bool storageImageCapable =
             !isMultisampleTexture &&
             (aspect & VK_IMAGE_ASPECT_COLOR_BIT) != 0 &&
             (formatProperties.optimalTilingFeatures & VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT) != 0;
+        const Bool supportsStorageImage = storageImageCapable && markedAsStorageImage;
         VkImageCreateFlags imageCreateFlags = shapeInfo.imageFlags;
-        if (supportsStorageImage && IsMutableStorageImageFormat(format)) {
+        if (storageImageCapable && IsMutableStorageImageFormat(format) &&
+            m_mutableFormatUnsupported.find(format) == m_mutableFormatUnsupported.end()) {
             imageCreateFlags |= VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
         }
 
+        VkImageUsageFlags desiredUsage =
+            VK_IMAGE_USAGE_SAMPLED_BIT |
+            (supportsStorageImage ? VK_IMAGE_USAGE_STORAGE_BIT : 0) |
+            ((aspect & VK_IMAGE_ASPECT_COLOR_BIT) ? VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT : 0) |
+            (((aspect & VK_IMAGE_ASPECT_DEPTH_BIT) || (aspect & VK_IMAGE_ASPECT_STENCIL_BIT)) ?
+                 VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT :
+                 0);
+        if (!isMultisampleTexture) {
+            desiredUsage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
+        }
+
         const Bool compatible = resource.image != VK_NULL_HANDLE && resource.format == format &&
                                 resource.extent.width == static_cast<Uint32>(texelSize.x()) &&
                                 resource.extent.height == static_cast<Uint32>(texelSize.y()) &&
@@ -1335,6 +1493,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                 resource.viewType == shapeInfo.viewType &&
                                 resource.sampleCount == resolvedSampleCount &&
                                 resource.imageCreateFlags == imageCreateFlags &&
+                                resource.usageFlags == desiredUsage &&
                                 resource.mipLevels == backingMipLevels;
         if (compatible) {
             if (resource.perMipViews.size() != backingMipLevels) {
@@ -1343,6 +1502,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (resource.perMipSampledViews.size() != backingMipLevels) {
                 resource.perMipSampledViews.resize(backingMipLevels, VK_NULL_HANDLE);
             }
+            // Keeping the image is itself the answer to the mark: either it already carries
+            // STORAGE, or this format can never carry it. Either way there is nothing left to
+            // recreate, so stop reporting the texture as needing preparation.
+            resource.storageUsageResolved = markedAsStorageImage;
             return true;
         }
 
@@ -1357,7 +1520,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             resource.sampleCount == resolvedSampleCount &&
             resource.imageCreateFlags == imageCreateFlags &&
             resolvedSampleCount == VK_SAMPLE_COUNT_1_BIT &&
-            resource.mipLevels < backingMipLevels &&
+            // '<=' rather than '<': a storage-usage upgrade recreates the image with an
+            // unchanged mip count, and its contents (a render target's pixels live only on the
+            // GPU) still have to survive. The vkCmdCopyImage below copies min(mipLevels).
+            resource.mipLevels <= backingMipLevels &&
             resource.layout != VK_IMAGE_LAYOUT_UNDEFINED;
 
         std::unique_ptr<TextureResource> preservedResource;
@@ -1379,21 +1545,60 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         imageInfo.format = format;
         imageInfo.tiling = VK_IMAGE_TILING_OPTIMAL;
         imageInfo.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
-        imageInfo.usage = VK_IMAGE_USAGE_SAMPLED_BIT |
-                          (supportsStorageImage ? VK_IMAGE_USAGE_STORAGE_BIT : 0) |
-                          ((aspect & VK_IMAGE_ASPECT_COLOR_BIT) ? VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT : 0) |
-                          (((aspect & VK_IMAGE_ASPECT_DEPTH_BIT) || (aspect & VK_IMAGE_ASPECT_STENCIL_BIT)) ?
-                               VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT :
-                               0);
-        if (!isMultisampleTexture) {
-            imageInfo.usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
-        }
+        imageInfo.usage = desiredUsage;
         imageInfo.samples = resolvedSampleCount;
+
+        // Bound the mutability. A blindly-mutable image has to be laid out so that ANY format in
+        // its compatibility class can be viewed, which costs bandwidth compression on tilers;
+        // naming the exact set instead lets the driver keep it. Only safe when that set really is
+        // exhaustive, so it is restricted to textures that are not image-unit bound: sampled views
+        // can only ever ask for ResolveSampledImageViewFormat's output, whereas glBindImageTexture
+        // may name any compatible format, which nothing here can enumerate ahead of time.
+        Vector<VkFormat> viewFormats;
+        VkImageFormatListCreateInfo formatListInfo{};
+        if (m_imageFormatListSupported && !supportsStorageImage &&
+            (imageInfo.flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) != 0) {
+            viewFormats.push_back(format);
+            for (const SamplerNumericDomain domain : {SamplerNumericDomain::Float,
+                                                      SamplerNumericDomain::SignedInteger,
+                                                      SamplerNumericDomain::UnsignedInteger}) {
+                const VkFormat viewFormat = ResolveSampledImageViewFormat(format, domain);
+                if (viewFormat == VK_FORMAT_UNDEFINED) {
+                    continue;
+                }
+                if (std::find(viewFormats.begin(), viewFormats.end(), viewFormat) == viewFormats.end()) {
+                    viewFormats.push_back(viewFormat);
+                }
+            }
+            formatListInfo.sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_LIST_CREATE_INFO;
+            formatListInfo.viewFormatCount = static_cast<Uint32>(viewFormats.size());
+            formatListInfo.pViewFormats = viewFormats.data();
+            imageInfo.pNext = &formatListInfo;
+        }
+
         if (isMultisampleTexture || (imageInfo.flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) != 0) {
             VkImageFormatProperties imageFormatProperties{};
-            const VkResult imageFormatResult = vkGetPhysicalDeviceImageFormatProperties(
+            VkResult imageFormatResult = vkGetPhysicalDeviceImageFormatProperties(
                 m_physicalDevice, format, imageInfo.imageType, imageInfo.tiling, imageInfo.usage,
                 imageInfo.flags, &imageFormatProperties);
+            if (imageFormatResult != VK_SUCCESS && !isMultisampleTexture &&
+                (imageInfo.flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) != 0) {
+                // Losing reinterpreted views only degrades the formatless-image feature for
+                // this texture; failing creation would lose the texture entirely, so retry
+                // as a plain immutable-format image.
+                MGLOG_W("%s: mutable image format=%d is unsupported for textureId=%d; creating "
+                        "without VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT (format reinterpretation "
+                        "will be unavailable for it)",
+                        __func__, static_cast<Int>(format), texture.GetExternalIndex());
+                // Remember the verdict so later syncs of same-format textures neither retry
+                // the probe nor flag-mismatch against this image and recreate it.
+                m_mutableFormatUnsupported.insert(format);
+                imageInfo.flags &= ~VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
+                imageCreateFlags = imageInfo.flags;
+                imageFormatResult = vkGetPhysicalDeviceImageFormatProperties(
+                    m_physicalDevice, format, imageInfo.imageType, imageInfo.tiling, imageInfo.usage,
+                    imageInfo.flags, &imageFormatProperties);
+            }
             if (imageFormatResult != VK_SUCCESS ||
                 (isMultisampleTexture && (imageFormatProperties.sampleCounts & resolvedSampleCount) == 0)) {
                 MGLOG_D("%s: image flags=0x%x sampleCount=%d are unsupported for textureId=%d target=%s "
@@ -1427,6 +1632,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         resource.viewType = shapeInfo.viewType;
         resource.sampleCount = resolvedSampleCount;
         resource.imageCreateFlags = imageCreateFlags;
+        resource.usageFlags = imageInfo.usage;
+        resource.storageUsageResolved = markedAsStorageImage;
         resource.syncedTextureParamsVersion = 0;
 
         if (preservedResource) {
@@ -1921,6 +2128,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     VkFormat VkTextureManager::ResolveSampledImageViewFormat(VkFormat imageFormat,
                                                               SamplerNumericDomain numericDomain) {
+        // Depth/stencil images always sample through the existing depth-aspect sampledView.
+        // Combined formats (D24S8, D32FS8) are multi-numeric, so vkuFormatIsSampledFloat is
+        // false for them by design, yet their depth aspect reads as float in every GL depth
+        // texture mode; Vulkan also forbids reinterpreting them through color-class views.
+        // Integer domains keep the same view (pre-reinterpretation behavior for stencil-index
+        // style access) rather than failing the draw.
+        if (vkuFormatIsDepthOrStencil(imageFormat)) {
+            return imageFormat;
+        }
         if (imageFormat == VK_FORMAT_UNDEFINED || numericDomain == SamplerNumericDomain::Unknown ||
             FormatMatchesSamplerNumericDomain(imageFormat, numericDomain)) {
             return imageFormat;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.h

@@ -13,6 +13,7 @@
 #include <MG_State/GLState/TextureState/TextureObject.h>
 #include <vk_mem_alloc.h>
 #include <unordered_map>
+#include <unordered_set>
 
 namespace MobileGL::MG_State::GLState {
 class ITextureObject;
@@ -52,6 +53,9 @@ public:
         VkCommandPool commandPool = VK_NULL_HANDLE;
         VkQueue graphicsQueue = VK_NULL_HANDLE;
         Uint32 frameCount = 0;
+        // VK_KHR_image_format_list is enabled: MUTABLE_FORMAT images can name the exact set of
+        // formats they will be viewed as, which is what lets a tiler keep them compressed.
+        Bool imageFormatListSupported = false;
     };
 
     struct TextureResource {
@@ -156,6 +160,17 @@ public:
         VkImageViewType viewType = VK_IMAGE_VIEW_TYPE_2D;
         VkSampleCountFlagBits sampleCount = VK_SAMPLE_COUNT_1_BIT;
         VkImageCreateFlags imageCreateFlags = 0;
+        // Usage the live image was created with. STORAGE is only requested for textures that
+        // have actually been bound to a GL image unit, because on Adreno a storage-capable
+        // image loses UBWC bandwidth compression; a later image binding upgrades the usage
+        // and recreates the image, so the resolved usage has to be part of the compatibility
+        // check that decides whether the existing image can be kept.
+        VkImageUsageFlags usageFlags = 0;
+        // True once this image was (re)resolved while the texture was already marked as an
+        // image-unit texture. Distinguishes "not upgraded yet" from "cannot be upgraded"
+        // (a format whose optimalTilingFeatures lack STORAGE_IMAGE never gains the bit), so
+        // NeedsStorageImagePreparation cannot ask for a recreate that will never happen.
+        Bool storageUsageResolved = false;
         Uint16 syncedTextureParamsVersion = 0;
         // Snapshot of ITextureObject::GetContentVersion() at the last successful sync;
         // lets SyncTexture skip the whole re-check/re-upload when content is unchanged.
@@ -189,6 +204,8 @@ public:
             std::swap(this->viewType, that.viewType);
             std::swap(this->sampleCount, that.sampleCount);
             std::swap(this->imageCreateFlags, that.imageCreateFlags);
+            std::swap(this->usageFlags, that.usageFlags);
+            std::swap(this->storageUsageResolved, that.storageUsageResolved);
             std::swap(this->syncedTextureParamsVersion, that.syncedTextureParamsVersion);
             std::swap(this->syncedContentVersion, that.syncedContentVersion);
             std::swap(this->syncedMipLevelCount, that.syncedMipLevelCount);
@@ -250,6 +267,8 @@ public:
             viewType = VK_IMAGE_VIEW_TYPE_2D;
             sampleCount = VK_SAMPLE_COUNT_1_BIT;
             imageCreateFlags = 0;
+            usageFlags = 0;
+            storageUsageResolved = false;
             syncedTextureParamsVersion = 0;
             syncedContentVersion = 0;
             syncedMipLevelCount = 0;
@@ -266,6 +285,10 @@ public:
     Bool Initialize(const InitInfo& initInfo);
     void Shutdown();
     void BeginFrame(Uint32 frameIndex);
+    // Drains every frame slot's deferred image/view releases. Only valid when
+    // the caller has proven every queue submission complete; used by the
+    // present-less frame-boundary drain.
+    void CollectAllDeferredReleases();
 
     TextureResource* SyncTextureAndGetDescriptor(
         MG_State::GLState::ITextureObject& texture);
@@ -284,6 +307,22 @@ public:
                                                       VkImageLayout newLayout);
     Bool TransitionTextureForSampling(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
     Bool TransitionTextureForStorageImage(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
+    // Records that this texture is bound to a GL image unit, so its image must carry
+    // VK_IMAGE_USAGE_STORAGE_BIT. Must be called before NeedsStorageImagePreparation, and
+    // therefore before the render pass is committed: an image that has to be upgraded is
+    // recreated, which is illegal inside a render pass. Sticky for the texture's lifetime -
+    // GL lets an image binding come and go, and re-creating the image every time it does
+    // would cost far more than the compression it wins back.
+    void MarkStorageImageTexture(MG_State::GLState::ITextureObject& texture);
+    // True when this texture is marked but its live image predates the mark, i.e. the next sync
+    // will recreate it with STORAGE usage and copy the old contents forward. Callers use this to
+    // submit their pending recording first, so that copy cannot read pre-flush content.
+    Bool NeedsStorageUsageUpgrade(MG_State::GLState::ITextureObject& texture) const;
+    // Non-mutating probe for the per-draw storage-image fast path: true when preparing this
+    // texture as a storage image may need work that is illegal inside a render pass (resource
+    // creation, dirty-content upload, or a layout transition to GENERAL). Unknown state reports
+    // true - a false positive merely ends the render pass, a false negative would skip a barrier.
+    Bool NeedsStorageImagePreparation(MG_State::GLState::ITextureObject& texture) const;
 
     static VkImageAspectFlags ResolveSampledImageViewAspectMask(VkImageAspectFlags imageAspect);
     static VkFormat ResolveSampledImageViewFormat(VkFormat imageFormat, SamplerNumericDomain numericDomain);
@@ -358,15 +397,20 @@ private:
     static TextureIdentity MakeTextureIdentity(MG_State::GLState::ITextureObject* texture);
     void EraseTrackedTexture(const TextureIdentity& identity);
     void PruneStaleTextureAliases(MG_State::GLState::ITextureObject* texture);
+    SizeT PruneDeadTextures();
 
     VkDevice m_device = VK_NULL_HANDLE;
     VkPhysicalDevice m_physicalDevice = VK_NULL_HANDLE;
     VmaAllocator m_allocator = nullptr;
     VkCommandPool m_commandPool = VK_NULL_HANDLE;
     VkQueue m_graphicsQueue = VK_NULL_HANDLE;
+    Bool m_imageFormatListSupported = false;
     Uint32 m_currentFrameIndex = 0;
 
     Uint8 m_gcCounter = 0;
+    // Frame-boundary GC gate: counts BeginFrame calls, not draws, so texture churn
+    // through non-draw paths (FBO clears, readbacks) still reaches the prune.
+    Uint32 m_gcFrameCounter = 0;
     // Active only between BeginDrawSyncScope/EndDrawSyncScope; identities of
     // textures already fully synced in the current draw (small N -> flat scan).
     Bool m_drawSyncScopeActive = false;
@@ -379,8 +423,13 @@ private:
         TextureResource* resource = nullptr;
     };
     Vector<DrawSyncedTexture> m_drawSyncedThisDraw;
+    // Formats whose mutable-image probe failed on this device; their images are created
+    // without MUTABLE_FORMAT_BIT so repeat syncs neither re-probe nor flag-mismatch.
+    std::unordered_set<VkFormat> m_mutableFormatUnsupported;
     std::unordered_map<TextureIdentity, WeakPtr<MG_State::GLState::ITextureObject>, TextureIdentityHash> m_aliveObjects;
     std::unordered_map<TextureIdentity, TextureResource, TextureIdentityHash> m_textureResources;
+    // Textures that have been bound to a GL image unit (see MarkStorageImageTexture).
+    std::unordered_set<TextureIdentity, TextureIdentityHash> m_storageImageTextures;
     Vector<Vector<TextureResource>> m_deferredReleases;
     Vector<Vector<VkImageView>> m_deferredViewReleases;
 };

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h

@@ -51,6 +51,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Uint32 instanceCount = 1;
         Uint32 firstVertex = 0;
         Uint32 firstInstance = 0;
+        // Indexed-draw metadata for bounding vertex-stream conversion. baseVertex is the
+        // draw's base-vertex offset; indexRangeIsExactView is true only when the draw
+        // fetches exactly the indices its IndexBufferView describes (direct DrawElements;
+        // multi/indirect forms leave it false because the CPU cannot bound their ranges).
+        Int32 baseVertex = 0;
+        Bool indexRangeIsExactView = false;
     };
 
     struct DrawIndexedCmdParam {
@@ -108,7 +114,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     };
 
-    class VulkanRenderer : public IBufferCopyCommandProvider, public FrameContext::IRecordingObserver {
+    class VulkanRenderer : public IBufferCopyCommandProvider,
+                           public FrameContext::IRecordingObserver,
+                           public VkRenderPassManager::IEvictionObserver,
+                           public ProgramFactory::IEvictionObserver {
     public:
         VulkanRenderer(NativeWindowType window, const VulkanRendererConfig& cfg = {});
         ~VulkanRenderer();
@@ -125,6 +134,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // recording, before any render pass.
         void OnFrameCommandRecordingBegan(VkCommandBuffer commandBuffer) override;
 
+        // VkRenderPassManager::IEvictionObserver: the render-pass aging sweep just
+        // destroyed these VkRenderPasses; evict every graphics pipeline hashed on a
+        // dying handle (they share its >1024-boundary idleness, so immediate
+        // destruction is safe) and drop the last-pipeline memo if any went.
+        void OnRenderPassesDestroyed(const Vector<VkRenderPass>& renderPasses) override;
+
+        // ProgramFactory::IEvictionObserver: an aged-out program entry was
+        // destroyed; evict its compute pipeline and graphics pipelines (same
+        // idleness guarantee - they are only bound through draws/dispatches that
+        // stamp the program entry) and purge the descriptor-set cache entries
+        // keyed by its now-recyclable VkDescriptorSetLayout handle.
+        void OnProgramEvicted(ProgramFactory::HashType programHash,
+                              VkDescriptorSetLayout descriptorSetLayout) override;
+
         Bool SetupDraw(FrameContext::FrameData& frame, GLenum mode, Flags<DrawSetupAspect> aspects,
                        const DrawCmdParam& drawParams,
                        const IndexBufferView* pIndexBufferView = nullptr);
@@ -251,7 +274,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Uint64 GetTimerQueryTimestampNs(const VkTimerQueryManager::TimestampRecord& record) const;
 
         void RequestSwapchainResize(Uint32 width, Uint32 height);
-        void RecreateSwapchain();
+        // Re-query the surface and report whether the live swapchain no longer matches it
+        // (size or orientation). This - not a VK_SUBOPTIMAL_KHR result - is what decides a
+        // rebuild, so a surface the driver merely considers suboptimal cannot thrash.
+        Bool SwapchainIsOutOfDate();
+        // Returns false when the surface is zero-area (minimized/hidden window):
+        // no new swapchain is installed and presentation must stay suspended.
+        Bool RecreateSwapchain();
 
     private:
         struct BlitUniformData {
@@ -337,11 +366,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkFence AcquirePooledSubmitFence();
         void DestroySubmitFencePool();
         Bool HasPendingRecordedWork() const;
+        // Frame-boundary housekeeping for paths that never reach Present's
+        // tail (present-less readback loops, suspended presentation, blocking
+        // sync waits): runs the same per-frame drains Present performs, but
+        // only when every queue submission has been observed complete AND no
+        // recorded-but-unsubmitted commands exist - i.e. when CPU-GPU overlap
+        // is provably already zero. Never blocks (non-blocking fence poll
+        // only), so the presenting path's frames-in-flight pipelining is
+        // untouched. Returns true when the drain ran.
+        Bool TryDrainFrameTransients();
 
         Vector<SubmitRecord> m_inFlightSubmits;
         Vector<VkFence> m_freeSubmitFences;
         Uint64 m_submitCounter = 0;
         Uint64 m_completedSubmitCounter = 0;
+        // Drains since the last Present, gating the drain's frame-boundary-equivalent
+        // work (arena rewind + cache aging): a presenting app's mid-frame
+        // readbacks/waits must neither churn the transient caches nor accelerate the
+        // aging clocks, while present-less loops still cross a boundary every few
+        // iterations. Reset in Present.
+        Uint32 m_drainsSinceLastPresent = 0;
 
         NativeWindowType m_window = 0;
         void* m_platformDisplay = nullptr;
@@ -349,12 +393,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void* m_platformCloseDisplay = nullptr;
         VulkanRendererConfig m_config;
         Bool m_swapchainResizeRequested = false;
+        // Presentation is suspended while the window is zero-area (minimized): the
+        // swapchain is unusable/out of date, so Present drops frames instead of
+        // submitting on a signaled fence / presenting never-acquired images.
+        Bool m_presentSuspended = false;
 
         // Vulkan objects
         Bool m_validationLayersEnabled = false;
         Vector<VkExtensionProperties> m_extensions;
         VkInstance m_instance = VK_NULL_HANDLE;
         VkDebugUtilsMessengerEXT m_debugMessenger = VK_NULL_HANDLE;
+        // Fallback reporting channel for drivers that ship the validation layers but
+        // only expose the older VK_EXT_debug_report (Adreno 650 / Vulkan 1.1.128).
+        VkDebugReportCallbackEXT m_debugReportCallback = VK_NULL_HANDLE;
         PhysicalDevice m_physicalDevice;
         VkDevice m_device = VK_NULL_HANDLE;
         VmaAllocator m_allocator = nullptr;
@@ -488,12 +539,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             }
         };
 
-        UnorderedMap<ConvertedVertexStreamKey, BufferSlice, ConvertedVertexStreamKeyHash>
+        struct ConvertedVertexStream {
+            BufferSlice slice;
+            // Number of source elements the cached slice covers. A draw needing a prefix of
+            // this range reuses the slice (converted streams are tightly packed); a draw
+            // needing more reconverts and replaces the entry, so per (buffer, layout) a
+            // frame converts at most the largest range any draw asked for.
+            SizeT elementCount = 0;
+            // Pins the source buffer for the frame so its heap address cannot be reused by
+            // a new BufferObject while this pointer-keyed entry is alive.
+            SharedPtr<const MG_State::GLState::BufferObject> sourcePin;
+        };
+        UnorderedMap<ConvertedVertexStreamKey, ConvertedVertexStream, ConvertedVertexStreamKeyHash>
             m_convertedVertexStreams;
 
         void CreateInstance();
         VkResult SetupDebugMessenger();
         VkResult DestroyDebugMessenger();
+        VkResult SetupDebugReportCallback();
+        void DestroyDebugReportCallback();
         VkDebugUtilsMessengerCreateInfoEXT PopulateDebugMessengerCreateInfo();
         void CreateSurface();
         void PickPhysicalDevice();
@@ -512,14 +576,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const RenderPassEntry& renderPassEntry);
         VkPipeline GetOrCreateComputePipeline(const ProgramFactory::VkProgramObject& programObj);
         void DestroyComputePipelines();
+        // Takes the frame rather than a command buffer: a first-time storage-usage upgrade has to
+        // flush the pending recording (see the body), which retires the current command buffer.
         Bool PrepareStorageImageTextures(
-            VkCommandBuffer commandBuffer,
+            FrameContext::FrameData& frame,
             const MG_State::GLState::ProgramObject& program,
             const ProgramFactory::VkProgramObject& programObj);
 
         Bool UploadAndBindVertexBuffers(VkCommandBuffer commandBuffer, const MG_State::GLState::VertexArrayObject& vao,
                                         const ProgramFactory::VkProgramObject& programObj,
-                                        const DrawCmdParam& drawParams, Bool indexedDraw);
+                                        const DrawCmdParam& drawParams,
+                                        const IndexBufferView* pIndexBufferView);
         Bool UploadAndBindIndexBuffer(FrameContext::FrameData& frame,
                                      const MG_State::GLState::VertexArrayObject& vao,
                                       const IndexBufferView* pIndexBufferView = nullptr);
@@ -537,6 +604,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                    GLenum filter);
         Bool MaterializePendingClearForTexture(VkCommandBuffer commandBuffer,
                                                MG_State::GLState::ITextureObject& texture);
+        Bool MaterializePendingClearForRenderbuffer(
+            VkCommandBuffer commandBuffer,
+            const SharedPtr<MG_State::GLState::RenderbufferObject>& renderbuffer);
         VkPipeline GetOrCreateBlitPipeline(const RenderPassEntry& renderPassEntry);
         Bool GenerateDepthMipmapWithShader(FrameContext::FrameData& frame,
                                            MG_State::GLState::ITextureObject& texture,
@@ -571,6 +641,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                  const PhysicalDevice& compareWithDevice,
                                                  PhysicalDevice& outBetterDevice);
         static constexpr const char* s_validationLayerNames[] = {"VK_LAYER_KHRONOS_validation"};
+        // VK_KHR_image_format_list: lets MUTABLE_FORMAT images declare their exact view-format
+        // set so the driver can keep bandwidth compression (see CreateLogicalDeviceAndQueues).
+        Bool m_imageFormatListExtensionEnabled = false;
+
         static constexpr const char* s_deviceExtensionNames[] = {VK_KHR_SWAPCHAIN_EXTENSION_NAME};
         static Bool CheckValidationLayerSupport();
 

MobileGL/MG_Impl/CGLImpl/CGLImpl.cpp

@@ -24,6 +24,7 @@ namespace MobileGL::MG_Impl::CGLImpl {
             GLint Samples = 0;
             GLint Profile = kCGLOGLPVersion_3_2_Core;
             GLint RendererId = 0x4d474c;
+            GLint DisplayMask = 0;
         };
 
         struct ContextObject {
@@ -134,6 +135,9 @@ namespace MobileGL::MG_Impl::CGLImpl {
             case kCGLPFARendererID:
                 pixelFormat.RendererId = value;
                 break;
+            case kCGLPFADisplayMask:
+                pixelFormat.DisplayMask = value;
+                break;
             default:
                 break;
             }
@@ -343,6 +347,9 @@ namespace MobileGL::MG_Impl::CGLImpl {
         case kCGLPFARendererID:
             *value = pixelFormat->RendererId;
             return kCGLNoError;
+        case kCGLPFADisplayMask:
+            *value = pixelFormat->DisplayMask;
+            return kCGLNoError;
         case kCGLPFAOpenGLProfile:
             *value = pixelFormat->Profile;
             return kCGLNoError;
@@ -481,6 +488,32 @@ namespace MobileGL::MG_Impl::CGLImpl {
         return it == currentContexts.end() ? nullptr : it->second;
     }
 
+    CGLError SetVirtualScreen(CGLContextObj ctx, GLint screen) {
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        auto* object = TryGetContext(ctx);
+        if (!object) {
+            return kCGLBadContext;
+        }
+        if (screen != 0) {
+            return kCGLBadValue;
+        }
+        object->VirtualScreen = screen;
+        return kCGLNoError;
+    }
+
+    CGLError GetVirtualScreen(CGLContextObj ctx, GLint* screen) {
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        auto* object = TryGetContext(ctx);
+        if (!object) {
+            return kCGLBadContext;
+        }
+        if (!screen) {
+            return kCGLBadAddress;
+        }
+        *screen = object->VirtualScreen;
+        return kCGLNoError;
+    }
+
     CGLError SetParameter(CGLContextObj ctx, CGLContextParameter pname, const GLint* params) {
         const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
         auto* object = TryGetContext(ctx);

MobileGL/MG_Impl/CGLImpl/CGLImpl.h

@@ -32,6 +32,8 @@ namespace MobileGL::MG_Impl::CGLImpl {
 
     CGLError SetCurrentContext(CGLContextObj ctx);
     CGLContextObj GetCurrentContext();
+    CGLError SetVirtualScreen(CGLContextObj ctx, GLint screen);
+    CGLError GetVirtualScreen(CGLContextObj ctx, GLint* screen);
     CGLError SetParameter(CGLContextObj ctx, CGLContextParameter pname, const GLint* params);
     CGLError GetParameter(CGLContextObj ctx, CGLContextParameter pname, GLint* params);
     CGLError UpdateContext(CGLContextObj ctx);

MobileGL/MG_Impl/CGLImpl/Exporting/Definitions.cpp

@@ -71,6 +71,14 @@ MOBILEGL_CGL_API CGLContextObj CGLGetCurrentContext(void) {
     return MobileGL::MG_Impl::CGLImpl::GetCurrentContext();
 }
 
+MOBILEGL_CGL_API CGLError CGLSetVirtualScreen(CGLContextObj ctx, GLint screen) {
+    return MobileGL::MG_Impl::CGLImpl::SetVirtualScreen(ctx, screen);
+}
+
+MOBILEGL_CGL_API CGLError CGLGetVirtualScreen(CGLContextObj ctx, GLint* screen) {
+    return MobileGL::MG_Impl::CGLImpl::GetVirtualScreen(ctx, screen);
+}
+
 MOBILEGL_CGL_API CGLError CGLSetParameter(CGLContextObj ctx, CGLContextParameter pname, const GLint* params) {
     return MobileGL::MG_Impl::CGLImpl::SetParameter(ctx, pname, params);
 }

MobileGL/MG_Impl/DyldInterpose/DyldInterpose.cpp

@@ -10,8 +10,12 @@
 
 #if defined(__APPLE__)
 
+#include "MG_Impl/CGLImpl/CGLImpl.h"
 #include "MG_Impl/GetProcAddress.h"
 
+#include <CoreGraphics/CoreGraphics.h>
+#include <CoreVideo/CVDisplayLink.h>
+#include <cstdint>
 #include <dlfcn.h>
 
 namespace {
@@ -47,10 +51,52 @@ namespace {
         return dlsym(handle, symbol);
     }
 
+    CGDirectDisplayID DisplayForMask(GLint displayMask) {
+        constexpr std::uint32_t MaxDisplays = sizeof(CGOpenGLDisplayMask) * 8;
+        CGDirectDisplayID displays[MaxDisplays] = {};
+        std::uint32_t displayCount = 0;
+        if (displayMask != 0 &&
+            CGGetActiveDisplayList(MaxDisplays, displays, &displayCount) == kCGErrorSuccess) {
+            const auto mask = static_cast<CGOpenGLDisplayMask>(displayMask);
+            for (std::uint32_t i = 0; i < displayCount; ++i) {
+                if ((CGDisplayIDToOpenGLDisplayMask(displays[i]) & mask) != 0) {
+                    return displays[i];
+                }
+            }
+        }
+        return CGMainDisplayID();
+    }
+
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wdeprecated-declarations"
+    CVReturn MobileGLCVDisplayLinkSetCurrentCGDisplayFromOpenGLContext(
+        CVDisplayLinkRef displayLink,
+        CGLContextObj context,
+        CGLPixelFormatObj pixelFormat) {
+        GLint virtualScreen = 0;
+        if (MobileGL::MG_Impl::CGLImpl::GetVirtualScreen(context, &virtualScreen) == kCGLNoError) {
+            GLint displayMask = 0;
+            if (!displayLink ||
+                MobileGL::MG_Impl::CGLImpl::DescribePixelFormat(
+                    pixelFormat, virtualScreen, kCGLPFADisplayMask, &displayMask) != kCGLNoError) {
+                return kCVReturnInvalidArgument;
+            }
+            return CVDisplayLinkSetCurrentCGDisplay(displayLink, DisplayForMask(displayMask));
+        }
+
+        using OriginalFunction = CVReturn (*)(CVDisplayLinkRef, CGLContextObj, CGLPixelFormatObj);
+        static const auto original = reinterpret_cast<OriginalFunction>(
+            dlsym(RTLD_NEXT, "CVDisplayLinkSetCurrentCGDisplayFromOpenGLContext"));
+        return original ? original(displayLink, context, pixelFormat) : kCVReturnError;
+    }
+
     __attribute__((used)) static const DyldInterposeEntry kMobileGLDyldInterpose[]
         __attribute__((section("__DATA,__interpose"))) = {
             {reinterpret_cast<const void*>(MobileGLDlsym), reinterpret_cast<const void*>(dlsym)},
+            {reinterpret_cast<const void*>(MobileGLCVDisplayLinkSetCurrentCGDisplayFromOpenGLContext),
+             reinterpret_cast<const void*>(CVDisplayLinkSetCurrentCGDisplayFromOpenGLContext)},
         };
+#pragma clang diagnostic pop
 } // namespace
 
 #endif

MobileGL/MG_Impl/DyldInterpose/ExportedSymbols.txt

@@ -0,0 +1,10 @@
+# Public CGL entry points.
+_CGL*
+
+# Public EGL entry points.
+_egl*
+
+# Public OpenGL and GLX entry points. OpenGL function names always use an
+# uppercase letter or digit after the "gl" prefix; excluding lowercase here
+# deliberately prevents glslang_* from matching this pattern.
+_gl[A-Z0-9]*

MobileGL/MG_Impl/EGLImpl/EGLImpl.cpp

@@ -8,6 +8,7 @@
 
 #include "EGLImpl.h"
 #include "../GetProcAddress.h"
+#include <Init.h>
 #include <MG_Backend/BackendObjects.h>
 #include <MG_State/EGLState/Core.h>
 #include <mutex>
@@ -25,6 +26,17 @@ namespace MobileGL::MG_Impl::EGLImpl {
             return MG_State::pEGLContext.get();
         }
 
+        // Entry points that can legitimately be an application's FIRST EGL
+        // call (display/proc-address/string queries) lazily bring MobileGL
+        // up here, so the library needs no static constructor and can
+        // re-initialize after the last eglTerminate tore everything down.
+        // Teardown-ish entry points keep using GetState() and fail benignly
+        // when MobileGL is not initialized.
+        EGLStateContext* GetStateEnsureInitialized() {
+            MobileGL::EnsureInitialized();
+            return GetState();
+        }
+
         MG_Backend::BackendObject* GetBackendObject(EGLStateContext* state) {
             auto* backendObject = MG_Backend::pActiveBackendObject.get();
             if (!backendObject && state) {
@@ -49,6 +61,8 @@ namespace MobileGL::MG_Impl::EGLImpl {
             return MG_Backend::WindowBackend::Android;
 #elif defined(__APPLE__)
             return MG_Backend::WindowBackend::MetalLayer;
+#elif defined(_WIN32)
+            return MG_Backend::WindowBackend::Win32;
 #elif defined(__linux__)
             return MG_Backend::WindowBackend::X11;
 #else
@@ -187,7 +201,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
     }
 
     EGLBoolean Initialize(EGLDisplay dpy, EGLint* major, EGLint* minor) {
-        auto* state = GetState();
+        auto* state = GetStateEnsureInitialized();
         if (!state) {
             return EGL_FALSE;
         }
@@ -208,7 +222,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
     }
 
     EGLDisplay GetDisplay(NativeDisplayType display) {
-        auto* state = GetState();
+        auto* state = GetStateEnsureInitialized();
         if (!state) {
             return EGL_NO_DISPLAY;
         }
@@ -313,6 +327,14 @@ namespace MobileGL::MG_Impl::EGLImpl {
         if (auto* backendObject = MG_Backend::pActiveBackendObject.get()) {
             backendObject->ReleaseEGLResources();
         }
+        // The last initialized display is gone and nothing is current on any
+        // thread: tear the whole library down deterministically inside the
+        // EGL lifecycle (backend, GL/EGL state, glslang). A later EGL call
+        // re-initializes lazily via GetStateEnsureInitialized(); process exit
+        // then has nothing left to destroy.
+        if (!state->HasAnyInitializedDisplay() && !state->HasAnyCurrentContext()) {
+            MobileGL::Destroy();
+        }
         return EGL_TRUE;
     }
 
@@ -345,7 +367,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
     }
 
     EGLBoolean BindAPI(EGLenum api) {
-        auto* state = GetState();
+        auto* state = GetStateEnsureInitialized();
         if (!state) {
             return EGL_FALSE;
         }
@@ -378,7 +400,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
     }
 
     char const* QueryString(EGLDisplay display, EGLint name) {
-        auto* state = GetState();
+        auto* state = GetStateEnsureInitialized();
         if (!state) {
             return nullptr;
         }
@@ -641,7 +663,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
     EGLDisplay GetPlatformDisplay(EGLenum platform, void* native_display, const EGLAttrib* attrib_list) {
         (void)attrib_list;
 
-        auto* state = GetState();
+        auto* state = GetStateEnsureInitialized();
         if (!state) {
             return EGL_NO_DISPLAY;
         }
@@ -737,6 +759,7 @@ namespace MobileGL::MG_Impl::EGLImpl {
         if (!name) {
             return nullptr;
         }
+        MobileGL::EnsureInitialized();
 
         MGLOG_D("eglGetProcAddress(%s)", name);
         void* proc = MG_Impl::GetProcAddress(name);

MobileGL/MG_Impl/GLImpl/Exporting/Definitions.cpp

@@ -295,7 +295,13 @@ DECLARE_GL_FUNCTION_HEAD(void, VertexAttribDivisor, GLuint index, GLuint divisor
 DECLARE_GL_FUNCTION_STUB_HEAD(void, BindTransformFeedback, GLenum target, GLuint id) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, BindTransformFeedback, target, id)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, DeleteTransformFeedbacks, GLsizei n, const GLuint* ids) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DeleteTransformFeedbacks, n, ids)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GenTransformFeedbacks, GLsizei n, GLuint* ids) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GenTransformFeedbacks, n, ids)
-DECLARE_GL_FUNCTION_STUB_HEAD(GLboolean, IsTransformFeedback, GLuint id) DECLARE_GL_FUNCTION_STUB_END(GLboolean, IsTransformFeedback, id)
+// Transform feedback objects are not implemented, so no name is ever a live object. The shared
+// stub returns (type)1, telling a probing caller that every id it invents already exists; GL_FALSE
+// is both truthful and what the spec requires for a name that was never generated.
+MOBILEGL_GL_API GLboolean glIsTransformFeedback(GLuint id) {
+    MGLOG_W("Stub function: %s(...)", __FUNCTION__);
+    return GL_FALSE;
+}
 DECLARE_GL_FUNCTION_STUB_HEAD(void, PauseTransformFeedback) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, PauseTransformFeedback)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, ResumeTransformFeedback) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ResumeTransformFeedback)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetProgramBinary, GLuint program, GLsizei bufSize, GLsizei* length, GLenum* binaryFormat, void* binary) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetProgramBinary, program, bufSize, length, binaryFormat, binary)
@@ -418,7 +424,7 @@ DECLARE_GL_FUNCTION_HEAD(void, DrawRangeElementsBaseVertex, GLenum mode, GLuint
 DECLARE_GL_FUNCTION_HEAD(void, DrawElementsInstancedBaseVertex, GLenum mode, GLsizei count, GLenum type, const void* indices, GLsizei instancecount, GLint basevertex) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawElementsInstancedBaseVertex, mode, count, type, indices, instancecount, basevertex)
 DECLARE_GL_FUNCTION_HEAD(void, FramebufferTexture, GLenum target, GLenum attachment, GLuint texture, GLint level) DECLARE_GL_FUNCTION_END_NO_RETURN(void, FramebufferTexture, target, attachment, texture, level)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, PrimitiveBoundingBox, GLfloat minX, GLfloat minY, GLfloat minZ, GLfloat minW, GLfloat maxX, GLfloat maxY, GLfloat maxZ, GLfloat maxW) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, PrimitiveBoundingBox, minX, minY, minZ, minW, maxX, maxY, maxZ, maxW)
-DECLARE_GL_FUNCTION_STUB_HEAD(GLenum, GetGraphicsResetStatus) DECLARE_GL_FUNCTION_STUB_END(GLenum, GetGraphicsResetStatus)
+DECLARE_GL_FUNCTION_HEAD(GLenum, GetGraphicsResetStatus) DECLARE_GL_FUNCTION_END(GLenum, GetGraphicsResetStatus)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, ReadnPixels, GLint x, GLint y, GLsizei width, GLsizei height, GLenum format, GLenum type, GLsizei bufSize, void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ReadnPixels, x, y, width, height, format, type, bufSize, data)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetnUniformfv, GLuint program, GLint location, GLsizei bufSize, GLfloat* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetnUniformfv, program, location, bufSize, params)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetnUniformiv, GLuint program, GLint location, GLsizei bufSize, GLint* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetnUniformiv, program, location, bufSize, params)
@@ -2583,7 +2589,10 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, TransformFeedbackStreamAttribsNV, GLsizei co
 DECLARE_GL_FUNCTION_STUB_HEAD(void, BindTransformFeedbackNV, GLenum target, GLuint id) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, BindTransformFeedbackNV, target, id)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, DeleteTransformFeedbacksNV, GLsizei n, const GLuint* ids) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DeleteTransformFeedbacksNV, n, ids)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GenTransformFeedbacksNV, GLsizei n, GLuint* ids) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GenTransformFeedbacksNV, n, ids)
-DECLARE_GL_FUNCTION_STUB_HEAD(GLboolean, IsTransformFeedbackNV, GLuint id) DECLARE_GL_FUNCTION_STUB_END(GLboolean, IsTransformFeedbackNV, id)
+MOBILEGL_GL_API GLboolean glIsTransformFeedbackNV(GLuint id) {
+    MGLOG_W("Stub function: %s(...)", __FUNCTION__);
+    return GL_FALSE;
+}
 DECLARE_GL_FUNCTION_STUB_HEAD(void, PauseTransformFeedbackNV, void) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, PauseTransformFeedbackNV, )
 DECLARE_GL_FUNCTION_STUB_HEAD(void, ResumeTransformFeedbackNV, void) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ResumeTransformFeedbackNV, )
 DECLARE_GL_FUNCTION_STUB_HEAD(void, DrawTransformFeedbackNV, GLenum mode, GLuint id) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DrawTransformFeedbackNV, mode, id)

MobileGL/MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.cpp

@@ -2144,6 +2144,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     namespace FramebufferImpl {
-        UniquePtr<DefaultFramebufferInfo> pDefaultFramebufferInfo;
+        // Leak-at-exit storage; see GlobalObjects.cpp.
+        UniquePtr<DefaultFramebufferInfo>& pDefaultFramebufferInfo = *new UniquePtr<DefaultFramebufferInfo>();
     } // namespace FramebufferImpl
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Framebuffer/GL_Framebuffer.h

@@ -78,6 +78,6 @@ namespace MobileGL::MG_Impl::GLImpl {
             SharedPtr<MG_State::GLState::ITextureObject> stencilAttachment;
         };
 
-        extern UniquePtr<DefaultFramebufferInfo> pDefaultFramebufferInfo;
+        extern UniquePtr<DefaultFramebufferInfo>& pDefaultFramebufferInfo;
     } // namespace FramebufferImpl
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Getter/GL_Getter.cpp

@@ -1996,4 +1996,12 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         return MG_Util::ConvertErrorCodeToGLEnum(error->get()->code);
     }
+
+    GLenum GetGraphicsResetStatus() {
+        // MobileGL does not implement robustness reset notification, so report GL_NO_ERROR
+        // ("no reset detected"). Returning the generic stub's (GLenum)1 makes dEQP read a lost
+        // device after every case (gl3cTestPackages.cpp:121) and, under the default
+        // --deqp-terminate-on-device-lost=enable, tear the whole CTS run down.
+        return GL_NO_ERROR;
+    }
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Getter/GL_Getter.h

@@ -21,4 +21,5 @@ namespace MobileGL::MG_Impl::GLImpl {
     void GetIntegeri_v(GLenum target, GLuint index, GLint* data);
     void GetInteger64i_v(GLenum target, GLuint index, GLint64* data);
     GLenum GetError();
+    GLenum GetGraphicsResetStatus();
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Sync/GL_Sync.cpp

@@ -133,4 +133,33 @@ namespace MobileGL::MG_Impl::GLImpl {
             values[0] = value;
         }
     }
+
+    void DestroyAllSyncObjects() {
+        // Detach the registry under the lock, release outside it. Entries the app
+        // already deleted were erased by DeleteSync, so nothing here double-frees;
+        // a DeleteSync racing this sweep finds an empty registry and returns. A
+        // thread still blocked inside ClientWaitSync/GetSynciv during teardown
+        // holds a raw SyncObject* these deletes invalidate - the same undefined
+        // race an app-driven DeleteSync already has.
+        UnorderedMap<GLsync, SyncObject*> orphans;
+        {
+            const std::lock_guard<std::mutex> lock(g_syncObjectsMutex);
+            orphans.swap(g_liveSyncObjects);
+        }
+        if (orphans.empty()) {
+            return;
+        }
+        // Both backends' DeleteSync only free the heap wrapper once their GL
+        // context/renderer is gone (generation/current-thread guards), so this is
+        // safe after the backend has released its EGL resources - but not after
+        // the function table itself is cleared.
+        const auto backendDeleteSync = MG_Backend::gBackendFunctionsTable.GL.DeleteSync;
+        for (const auto& [_, syncObject] : orphans) {
+            if (backendDeleteSync && syncObject->backendHandle) {
+                backendDeleteSync(syncObject->backendHandle);
+            }
+            delete syncObject;
+        }
+        MGLOG_D("DestroyAllSyncObjects: reclaimed %zu sync object(s) the app left undeleted", orphans.size());
+    }
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Sync/GL_Sync.h

@@ -16,4 +16,12 @@ namespace MobileGL::MG_Impl::GLImpl {
     void WaitSync(GLsync sync, GLbitfield flags, GLuint64 timeout);
     void DeleteSync(GLsync sync);
     void GetSynciv(GLsync sync, GLenum pname, GLsizei bufSize, GLsizei* length, GLint* values);
+    // Destroys every still-registered sync object exactly as DeleteSync would.
+    // GL requires syncs to die with their context; called only from full library
+    // teardown (DestroyImpl), where no context survives on any thread, so the
+    // process-global registry can be drained wholesale. Must run while the
+    // backend function table is still populated: each backend handle has to be
+    // released by the backend that created it, never by a later re-initialized
+    // one.
+    void DestroyAllSyncObjects();
 } // namespace MobileGL::MG_Impl::GLImpl

MobileGL/MG_Impl/GLImpl/Texture/GL_Texture.cpp

@@ -350,6 +350,10 @@ namespace MobileGL::MG_Impl::GLImpl {
                 texture.AllocateStorage(uploadTarget, level, {levelTexelSize, levelByteSize});
                 texture.MarkStorageDirty(uploadTarget, level, false);
             }
+            // glGenerateMipmap defines exactly levels 0..requiredLevelCount-1. AllocateStorage only
+            // grows, so a previously longer chain (a bigger base image before respecification) would
+            // otherwise keep a tail of stale levels here and read as incomplete.
+            texture.TruncateMipmapLevels(uploadTarget, requiredLevelCount);
             // Mip generation grows/regenerates the level set on the GPU without marking any CPU
             // level dirty (MarkStorageDirty(...,false) above). Bump the content version so the
             // backend re-syncs: a cached sampled VkImageView built for the pre-generate level
@@ -429,8 +433,10 @@ namespace MobileGL::MG_Impl::GLImpl {
 
             const Int maxSamples = GetMaxSupportedTextureSamples(textureInternalFormat);
             if (samples > maxSamples) {
+                // GL specifies INVALID_OPERATION - not INVALID_VALUE - when the sample count
+                // exceeds what the format supports, and the native Adreno driver agrees.
                 MG_State::pGLContext->RecordError(
-                    ErrorCode::InvalidValue,
+                    ErrorCode::InvalidOperation,
                     MakeUnique<GenericErrorInfo>(
                         "MG_Impl/GLImpl", caller,
                         std::format("Sample count {} exceeds the supported maximum {} for this texture format.",
@@ -454,8 +460,56 @@ namespace MobileGL::MG_Impl::GLImpl {
             textureObject->SetSamples(samples);
             textureObject->SetFixedSampleLocations(fixedsamplelocations == GL_TRUE);
             textureMipmapObject->AllocateStorage(textureUploadTarget, 0, {{width, height, depth}, 0});
+            // Multisample textures are single-level by definition, so a name that previously held a
+            // mip chain must not keep its tail now that AllocateStorage only grows.
+            textureMipmapObject->TruncateMipmapLevels(textureUploadTarget, 1);
             textureMipmapObject->MarkStorageDirty(textureUploadTarget, 0, false);
         }
+
+        // Redefining level 0 of a texture that already had a base image drops the rest of the chain,
+        // which is exactly what AllocateLevel used to do implicitly for every level. Keeping that
+        // behaviour for level 0 - and only for level 0 - is what makes the grow-only change safe:
+        // any level-0 respecification leaves the chain in precisely the state it would have had
+        // before, while an upload to level N no longer destroys the levels beneath it.
+        //
+        // Why it has to be *every* level-0 respecification and not just a size change: Minecraft's
+        // Mipmap Levels setting rebuilds the block atlas at the SAME dimensions with a different
+        // level count. A size-only test would leave the old tail in place, and because Mojang
+        // terminates its chains with a 0x0 level the result is the zero-then-nonzero pattern that
+        // IsComplete() rejects (TextureObject.cpp) - whereupon DirectGLES skips syncing the texture
+        // entirely (Managers.cpp) and the atlas samples black.
+        //
+        // The "already has a base image" test is what lets the fix work at all: a level that was
+        // never written reads back as {0,0,0}, so building a chain top-down - upload level N first,
+        // then level 0 - must not discard the levels just uploaded. That ordering is what
+        // KHR-GL33.texture_repeat_mode does.
+        // Scoped to the respecified upload target only, which is what AllocateLevel already did.
+        // Cube maps keep six independent chains while reporting a single level count (face +X), so
+        // respecifying a face other than +X can leave the count longer than that face - but that
+        // asymmetry predates this change and widening the truncation to all six faces would destroy
+        // mip data for faces the application never touched. Left alone deliberately.
+        void DiscardMipmapChainOnBaseRespecification(MG_State::GLState::TextureObjectMipmap* texture,
+                                                     TextureUploadTarget uploadTarget, Uint level) {
+            if (level != 0) return;
+
+            const IntVec3 existingBaseSize = texture->GetMipmapTexelSize(uploadTarget, 0);
+            const Bool hasExistingBaseImage =
+                existingBaseSize.x() > 0 && existingBaseSize.y() > 0 && existingBaseSize.z() > 0;
+            if (!hasExistingBaseImage) return;
+
+            texture->TruncateMipmapLevels(uploadTarget, 1);
+        }
+
+        // Compressed texture upload is not implemented yet. GL_NUM_COMPRESSED_TEXTURE_FORMATS
+        // reports 0, so every compressed internalformat is by definition unsupported and
+        // GL_INVALID_ENUM is the specified error - unlike THROW_UNIMPL_EXCEPTION, which unwinds
+        // a C++ exception through the C GL ABI and takes the process down.
+        void RecordUnsupportedCompressedFormat(const char* caller) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidEnum,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                             "Compressed texture formats are not supported."));
+        }
     } // namespace
 
     const SharedPtr<MG_State::GLState::ITextureObject>& GetTextureObjectByName(GLuint texture, const char* caller) {
@@ -493,8 +547,15 @@ namespace MobileGL::MG_Impl::GLImpl {
             }
 
             auto mipmapTexture = std::static_pointer_cast<MG_State::GLState::TextureObjectMipmap>(textureObject);
-            if (level < 0 || static_cast<Uint>(level) >= mipmapTexture->GetMipmapLevelCount()) {
+            if (level < 0) {
                 RecordClearTextureError(caller, ErrorCode::InvalidValue,
+                                        std::format("Texture level {} is negative.", level));
+                return nullptr;
+            }
+            // ARB_clear_texture: clearing an image that was never defined by TexImage*/
+            // TexStorage* is INVALID_OPERATION, not INVALID_VALUE.
+            if (static_cast<Uint>(level) >= mipmapTexture->GetMipmapLevelCount()) {
+                RecordClearTextureError(caller, ErrorCode::InvalidOperation,
                                         std::format("Texture level {} is not defined.", level));
                 return nullptr;
             }
@@ -536,6 +597,12 @@ namespace MobileGL::MG_Impl::GLImpl {
             return true;
         }
 
+        // Writes the clear into the CPU shadow and marks the whole level dirty, exactly like
+        // TexSubImage*_State does. Shared limitation of the level-granular shadow sync: the
+        // shadow does not reflect GPU-side writes (FBO rendering, imageStore), so a PARTIAL
+        // clear of a GPU-written level re-uploads stale shadow bytes outside the region on
+        // the next sync. Full-level clears (glClearTexImage, or a sub-clear covering the
+        // level) rewrite the entire shadow and are always correct.
         Bool ClearMipmapRegion(const SharedPtr<MG_State::GLState::TextureObjectMipmap>& textureObject,
                                TextureUploadTarget uploadTarget, GLint level,
                                GLint xoffset, GLint yoffset, GLint zoffset,
@@ -1657,6 +1724,21 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
+        // RGTC is a 2D-only compression scheme, so a 3D target rejects it. This has to be tested on
+        // the raw enum: the RGTC formats resolve to plain R8/RG8/SNORM storage on the way in (see
+        // GLToMG's TextureEnumConverter), so once the internal format is converted there is nothing
+        // left to distinguish them from an ordinary one- or two-channel upload.
+        if ((textureUploadTarget == TextureUploadTarget::Texture3D ||
+             textureUploadTarget == TextureUploadTarget::ProxyTexture3D) &&
+            (internalformat == GL_COMPRESSED_RED_RGTC1 || internalformat == GL_COMPRESSED_SIGNED_RED_RGTC1 ||
+             internalformat == GL_COMPRESSED_RG_RGTC2 || internalformat == GL_COMPRESSED_SIGNED_RG_RGTC2)) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "RGTC compressed formats are invalid for 3D texture targets"));
+            return;
+        }
+
         // TODO: GL_INVALID_OPERATION is generated if a non-zero buffer object name is bound to the
         // GL_PIXEL_UNPACK_BUFFER target and the buffer object's data store is currently mapped.
         // GL_INVALID_OPERATION is generated if a non-zero buffer object name is bound to the GL_PIXEL_UNPACK_BUFFER
@@ -1714,6 +1796,7 @@ namespace MobileGL::MG_Impl::GLImpl {
         if (isProxy) {
             MGLOG_D("%s: isProxy = true, not allocating", __func__);
         } else {
+            DiscardMipmapChainOnBaseRespecification(textureMipmapObject, textureUploadTarget, level);
             textureMipmapObject->AllocateStorage(textureUploadTarget, level, {{width, height, depth}, internalBytes});
         }
 
@@ -1841,6 +1924,7 @@ namespace MobileGL::MG_Impl::GLImpl {
             MGLOG_D("%s: isProxy = true, not allocating", __func__);
         } else {
             MGLOG_D("%s: Allocating %d bytes at mip %d", __func__, internalBytes, level);
+            DiscardMipmapChainOnBaseRespecification(textureMipmapObject, textureUploadTarget, level);
             textureMipmapObject->AllocateStorage(textureUploadTarget, level,
                                                  {{width, height, 1}, internalBytes});
         }
@@ -1929,6 +2013,7 @@ namespace MobileGL::MG_Impl::GLImpl {
                         "Texture object here should always be an object with mipmap");
         auto textureMipmapObject = static_cast<MG_State::GLState::TextureObjectMipmap*>(textureObject.get());
         if (!isProxy) {
+            DiscardMipmapChainOnBaseRespecification(textureMipmapObject, textureUploadTarget, level);
             textureMipmapObject->AllocateStorage(textureUploadTarget, level, {{width, 1, 1}, internalBytes});
         }
 
@@ -2581,7 +2666,13 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void GetCompressedTexImage_State(GLenum target, GLint level, void* img) {
-        // TODO: implement
+        // TODO: implement compressed readback. Reporting success while writing nothing hands
+        // the caller stale memory with GL_NO_ERROR; no texture can be compressed yet, and GL
+        // specifies GL_INVALID_OPERATION when the bound level is not compressed.
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidOperation,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                         "Texture level is not stored in a compressed format."));
     }
 
     void GenTextures_State(GLsizei n, GLuint* textures) {
@@ -2768,20 +2859,20 @@ namespace MobileGL::MG_Impl::GLImpl {
     void CompressedTexSubImage3D_State(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLint zoffset,
                                        GLsizei width, GLsizei height, GLsizei depth, GLenum format, GLsizei imageSize,
                                        const void* data) {
-        // TODO: implement
-        THROW_UNIMPL_EXCEPTION;
+        // TODO: implement compressed upload - see CompressedTexImage2D_State.
+        RecordUnsupportedCompressedFormat(__func__);
     }
 
     void CompressedTexSubImage2D_State(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLsizei width,
                                        GLsizei height, GLenum format, GLsizei imageSize, const void* data) {
-        // TODO: implement
-        THROW_UNIMPL_EXCEPTION;
+        // TODO: implement compressed upload - see CompressedTexImage2D_State.
+        RecordUnsupportedCompressedFormat(__func__);
     }
 
     void CompressedTexSubImage1D_State(GLenum target, GLint level, GLint xoffset, GLsizei width, GLenum format,
                                        GLsizei imageSize, const void* data) {
-        // TODO: implement
-        THROW_UNIMPL_EXCEPTION;
+        // TODO: implement compressed upload - see CompressedTexImage2D_State.
+        RecordUnsupportedCompressedFormat(__func__);
     }
 
     void CompressedTexImage3D_State(GLenum target, GLint level, GLenum internalformat, GLsizei width, GLsizei height,
@@ -2791,8 +2882,8 @@ namespace MobileGL::MG_Impl::GLImpl {
         auto& textureObject = GetTextureObjectByTarget(textureUploadTarget, textureTarget);
         if (!ValidateTextureMutable(textureObject, __func__)) return;
 
-        // TODO: implement
-        THROW_UNIMPL_EXCEPTION;
+        // TODO: implement compressed upload - see CompressedTexImage2D_State.
+        RecordUnsupportedCompressedFormat(__func__);
     }
 
     void CompressedTexImage2D_State(GLenum target, GLint level, GLenum internalformat, GLsizei width, GLsizei height,
@@ -2802,8 +2893,11 @@ namespace MobileGL::MG_Impl::GLImpl {
         auto& textureObject = GetTextureObjectByTarget(textureUploadTarget, textureTarget);
         if (!ValidateTextureMutable(textureObject, __func__)) return;
 
-        // TODO: implement
-        THROW_UNIMPL_EXCEPTION;
+        // TODO: implement compressed upload. Until then report the spec error for an
+        // unsupported compressed format rather than throwing - a C++ exception unwinding
+        // through the C GL ABI is a hard crash for the caller, while GL_INVALID_ENUM is
+        // exactly what GL_NUM_COMPRESSED_TEXTURE_FORMATS == 0 promises.
+        RecordUnsupportedCompressedFormat(__func__);
     }
 
     void CompressedTexImage1D_State(GLenum target, GLint level, GLenum internalformat, GLsizei width, GLint border,
@@ -2813,8 +2907,8 @@ namespace MobileGL::MG_Impl::GLImpl {
         auto& textureObject = GetTextureObjectByTarget(textureUploadTarget, textureTarget);
         if (!ValidateTextureMutable(textureObject, __func__)) return;
 
-        // TODO: implement
-        THROW_UNIMPL_EXCEPTION;
+        // TODO: implement compressed upload - see CompressedTexImage2D_State.
+        RecordUnsupportedCompressedFormat(__func__);
     }
 
     void BindTexture_State(GLenum target, GLuint texture) {
@@ -3160,6 +3254,9 @@ namespace MobileGL::MG_Impl::GLImpl {
             textureMipmapObject->AllocateStorage(textureUploadTarget, level, {{levelWidth, 1, 1}, byteSize});
             textureMipmapObject->MarkStorageDirty(textureUploadTarget, level, false);
         }
+        // Immutable storage defines exactly `levels` levels; AllocateStorage only grows, so a
+        // longer pre-existing chain has to be dropped explicitly.
+        textureMipmapObject->TruncateMipmapLevels(textureUploadTarget, static_cast<Uint>(levels));
         textureObject->SetImmutableLevels(static_cast<Uint>(levels));
     }
 
@@ -3212,6 +3309,8 @@ namespace MobileGL::MG_Impl::GLImpl {
             textureMipmapObject->AllocateStorage(textureUploadTarget, level, {{levelWidth, levelHeight, 1}, byteSize});
             textureMipmapObject->MarkStorageDirty(textureUploadTarget, level, false);
         }
+        // See TextureStorage1D.
+        textureMipmapObject->TruncateMipmapLevels(textureUploadTarget, static_cast<Uint>(levels));
         textureObject->SetImmutableLevels(static_cast<Uint>(levels));
     }
 
@@ -3264,6 +3363,8 @@ namespace MobileGL::MG_Impl::GLImpl {
                                                  {{levelWidth, levelHeight, levelDepth}, byteSize});
             textureMipmapObject->MarkStorageDirty(textureUploadTarget, level, false);
         }
+        // See TextureStorage1D.
+        textureMipmapObject->TruncateMipmapLevels(textureUploadTarget, static_cast<Uint>(levels));
         textureObject->SetImmutableLevels(static_cast<Uint>(levels));
     }
 
@@ -4025,8 +4126,21 @@ namespace MobileGL::MG_Impl::GLImpl {
     void CopyTextureSubImage2D(GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint x, GLint y,
                                GLsizei width, GLsizei height) {
         auto textureObject = GetTextureObjectByName(texture, __func__);
-        WithTemporarilyBoundNamedTexture(textureObject, [&](GLenum target) {
-            CopyTexSubImage2D_Backend(target, level, xoffset, yoffset, x, y, width, height);
+        if (!textureObject) return;
+        // GL 4.6 sec. 8.8: the 2D form only accepts these effective targets; cube maps must
+        // go through CopyTextureSubImage3D with the face as a layer.
+        const auto target = textureObject->GetTarget();
+        if (target != TextureTarget::Texture2D && target != TextureTarget::Texture1DArray &&
+            target != TextureTarget::TextureRectangle) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "CopyTextureSubImage2D requires a 2D, 1D-array, or "
+                                             "rectangle texture."));
+            return;
+        }
+        WithTemporarilyBoundNamedTexture(textureObject, [&](GLenum glTarget) {
+            CopyTexSubImage2D_Backend(glTarget, level, xoffset, yoffset, x, y, width, height);
         });
     }
 

MobileGL/MG_Impl/GLImpl/Texture/ProxyTexture.cpp

@@ -14,7 +14,8 @@
 #include <MG_State/GLState/TextureState/TextureObjectStubs.h>
 
 namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
-    UniquePtr<ProxyTextureManager> pProxyTextureManager;
+    // Leak-at-exit storage; see GlobalObjects.cpp.
+    UniquePtr<ProxyTextureManager>& pProxyTextureManager = *new UniquePtr<ProxyTextureManager>();
 
     Bool IsProxyTextureTarget(TextureUploadTarget target) {
         switch (target) {

MobileGL/MG_Impl/GLImpl/Texture/ProxyTexture.h

@@ -23,5 +23,5 @@ namespace MobileGL::MG_Impl::GLImpl::TextureImpl {
         UnorderedMap<TextureUploadTarget, SharedPtr<MG_State::GLState::ITextureObject>> m_proxyTexturesMap;
     };
 
-    extern UniquePtr<ProxyTextureManager> pProxyTextureManager;
+    extern UniquePtr<ProxyTextureManager>& pProxyTextureManager;
 } // namespace MobileGL::MG_Impl::GLImpl::TextureImpl

MobileGL/MG_Impl/GetProcAddress.cpp

@@ -85,6 +85,8 @@ namespace MobileGL::MG_Impl {
         GETPROC(CGLGetPixelFormat, name);
         GETPROC(CGLSetCurrentContext, name);
         GETPROC(CGLGetCurrentContext, name);
+        GETPROC(CGLSetVirtualScreen, name);
+        GETPROC(CGLGetVirtualScreen, name);
         GETPROC(CGLSetParameter, name);
         GETPROC(CGLGetParameter, name);
         GETPROC(CGLUpdateContext, name);

MobileGL/MG_Impl/NSOpenGLImpl/NSOpenGLImpl.cpp

@@ -29,10 +29,19 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
         char kContextViewKey;
         char kContextLayerKey;
 
-        std::once_flag g_installOnce;
         IMP g_pixelFormatDealloc = nullptr;
         IMP g_contextDealloc = nullptr;
 
+        std::mutex& HookInstallMutex() {
+            static auto* mutex = new std::mutex();
+            return *mutex;
+        }
+
+        Bool& HooksInstalled() {
+            static auto* installed = new Bool(false);
+            return *installed;
+        }
+
         template <typename Fn>
         Fn ObjcMsgSend() {
             return reinterpret_cast<Fn>(objc_msgSend);
@@ -431,12 +440,12 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             method_setImplementation(method, replacement);
         }
 
-        void InstallHooksOnce() {
+        Bool InstallHooksOnce() {
             Class pixelFormatClass = objc_getClass("NSOpenGLPixelFormat");
             Class contextClass = objc_getClass("NSOpenGLContext");
             if (!pixelFormatClass || !contextClass) {
                 MGLOG_W("NSOpenGLImpl: NSOpenGL classes are not loaded; hooks not installed");
-                return;
+                return false;
             }
 
             ReplaceInstanceMethod(pixelFormatClass, "initWithAttributes:",
@@ -471,11 +480,34 @@ namespace MobileGL::MG_Impl::NSOpenGLImpl {
             ReplaceInstanceMethod(contextClass, "dealloc", reinterpret_cast<IMP>(ContextDealloc), &g_contextDealloc);
 
             MGLOG_I("NSOpenGLImpl hooks installed");
+            return true;
         }
     } // namespace
 
     void InstallHooks() {
-        std::call_once(g_installOnce, InstallHooksOnce);
+        const std::lock_guard<std::mutex> lock(HookInstallMutex());
+        if (!HooksInstalled()) {
+            // Do not permanently consume the install attempt when the OpenGL
+            // framework has not registered its Objective-C classes yet. The
+            // dyld bootstrap normally runs after framework dependencies, but
+            // an explicitly loaded/static-linked MobileGL can arrive earlier.
+            HooksInstalled() = InstallHooksOnce();
+        }
     }
 } // namespace MobileGL::MG_Impl::NSOpenGLImpl
+
+namespace {
+    // SDL's Cocoa backend creates NSOpenGLPixelFormat/NSOpenGLContext before
+    // its first dlsym("glGetString") or other MobileGL host-API call. Install
+    // only the lightweight Objective-C dispatch hooks while the injected dylib
+    // is loading so those first Cocoa objects are routed through CGLImpl. The
+    // hooked context constructor reaches EGLImpl::GetDisplay(), which performs
+    // the full, thread-safe MobileGL initialization outside this bootstrap.
+    //
+    // There is intentionally no matching destructor: backend teardown remains
+    // owned by the EGL lifecycle and process-exit globals remain leak-at-exit.
+    __attribute__((constructor)) void BootstrapNSOpenGLHooks() {
+        MobileGL::MG_Impl::NSOpenGLImpl::InstallHooks();
+    }
+} // namespace
 #endif

MobileGL/MG_Impl/WGLImpl/Exporting/Definitions.cpp

@@ -0,0 +1,156 @@
+// MobileGL - MobileGL/MG_Impl/WGLImpl/Exporting/Definitions.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+// wingdi.h declares most wgl* entry points as WINGDIAPI (__declspec(dllimport)),
+// which would reject our definitions. _GDI32_ is the SDK's "I am the module that
+// implements these" switch: it turns WINGDIAPI into a plain declaration. It must
+// be defined before the first windows.h inclusion in this translation unit.
+#if defined(_WIN32) && !defined(_GDI32_)
+#define _GDI32_ 1
+#endif
+
+#include <Includes.h>
+
+#if defined(_WIN32)
+#include "../WGLImpl.h"
+
+namespace WGL = MobileGL::MG_Impl::WGLImpl;
+
+// ---- Pixel-format entry points (gdi32 forwards ChoosePixelFormat/SetPixelFormat/
+// ---- DescribePixelFormat/GetPixelFormat/SwapBuffers into these exports) ----
+
+extern "C" int WINAPI wglChoosePixelFormat(HDC hdc, CONST PIXELFORMATDESCRIPTOR* ppfd) {
+    return WGL::ChoosePixelFormat(hdc, ppfd);
+}
+
+extern "C" int WINAPI wglDescribePixelFormat(HDC hdc, int iPixelFormat, UINT nBytes,
+                                                   LPPIXELFORMATDESCRIPTOR ppfd) {
+    return WGL::DescribePixelFormat(hdc, iPixelFormat, nBytes, ppfd);
+}
+
+extern "C" int WINAPI wglGetPixelFormat(HDC hdc) {
+    return WGL::GetPixelFormat(hdc);
+}
+
+extern "C" BOOL WINAPI wglSetPixelFormat(HDC hdc, int iPixelFormat, CONST PIXELFORMATDESCRIPTOR* ppfd) {
+    return WGL::SetPixelFormat(hdc, iPixelFormat, ppfd);
+}
+
+extern "C" BOOL WINAPI wglSwapBuffers(HDC hdc) {
+    return WGL::SwapBuffers(hdc);
+}
+
+// ---- Context management ----
+
+extern "C" HGLRC WINAPI wglCreateContext(HDC hdc) {
+    return WGL::CreateContext(hdc);
+}
+
+extern "C" HGLRC WINAPI wglCreateLayerContext(HDC hdc, int iLayerPlane) {
+    return iLayerPlane == 0 ? WGL::CreateContext(hdc) : nullptr;
+}
+
+extern "C" BOOL WINAPI wglCopyContext(HGLRC, HGLRC, UINT) {
+    MGLOG_W("wglCopyContext is not supported");
+    SetLastError(ERROR_NOT_SUPPORTED);
+    return FALSE;
+}
+
+extern "C" BOOL WINAPI wglDeleteContext(HGLRC hglrc) {
+    return WGL::DeleteContext(hglrc);
+}
+
+extern "C" HGLRC WINAPI wglGetCurrentContext(VOID) {
+    return WGL::GetCurrentContext();
+}
+
+extern "C" HDC WINAPI wglGetCurrentDC(VOID) {
+    return WGL::GetCurrentDC();
+}
+
+extern "C" BOOL WINAPI wglMakeCurrent(HDC hdc, HGLRC hglrc) {
+    return WGL::MakeCurrent(hdc, hglrc);
+}
+
+extern "C" BOOL WINAPI wglShareLists(HGLRC hglrcShare, HGLRC hglrcDest) {
+    return WGL::ShareLists(hglrcShare, hglrcDest);
+}
+
+// ---- Proc address ----
+
+extern "C" PROC WINAPI wglGetProcAddress(LPCSTR lpszProc) {
+    return WGL::GetProcAddress(lpszProc);
+}
+
+extern "C" PROC WINAPI wglGetDefaultProcAddress(LPCSTR lpszProc) {
+    return WGL::GetProcAddress(lpszProc);
+}
+
+// ---- Layer planes and palettes (unsupported; overlay planes do not exist here) ----
+
+extern "C" BOOL WINAPI wglDescribeLayerPlane(HDC, int, int, UINT, LPLAYERPLANEDESCRIPTOR) {
+    return FALSE;
+}
+
+extern "C" int WINAPI wglSetLayerPaletteEntries(HDC, int, int, int, CONST COLORREF*) {
+    return 0;
+}
+
+extern "C" int WINAPI wglGetLayerPaletteEntries(HDC, int, int, int, COLORREF*) {
+    return 0;
+}
+
+extern "C" BOOL WINAPI wglRealizeLayerPalette(HDC, int, BOOL) {
+    return FALSE;
+}
+
+extern "C" BOOL WINAPI wglSwapLayerBuffers(HDC hdc, UINT fuPlanes) {
+    if (fuPlanes & WGL_SWAP_MAIN_PLANE) {
+        return WGL::SwapBuffers(hdc);
+    }
+    return FALSE;
+}
+
+extern "C" DWORD WINAPI wglSwapMultipleBuffers(UINT n, CONST WGLSWAP* ps) {
+    if (!ps) {
+        return 0;
+    }
+    DWORD swapped = 0;
+    for (UINT i = 0; i < n; ++i) {
+        if (WGL::SwapBuffers(ps[i].hdc)) {
+            ++swapped;
+        }
+    }
+    return swapped;
+}
+
+// ---- Font rendering (legacy immediate-mode feature; not supported) ----
+
+extern "C" BOOL WINAPI wglUseFontBitmapsA(HDC, DWORD, DWORD, DWORD) {
+    MGLOG_W("wglUseFontBitmapsA is not supported");
+    return FALSE;
+}
+
+extern "C" BOOL WINAPI wglUseFontBitmapsW(HDC, DWORD, DWORD, DWORD) {
+    MGLOG_W("wglUseFontBitmapsW is not supported");
+    return FALSE;
+}
+
+extern "C" BOOL WINAPI wglUseFontOutlinesA(HDC, DWORD, DWORD, DWORD, FLOAT, FLOAT, int,
+                                                 LPGLYPHMETRICSFLOAT) {
+    MGLOG_W("wglUseFontOutlinesA is not supported");
+    return FALSE;
+}
+
+extern "C" BOOL WINAPI wglUseFontOutlinesW(HDC, DWORD, DWORD, DWORD, FLOAT, FLOAT, int,
+                                                 LPGLYPHMETRICSFLOAT) {
+    MGLOG_W("wglUseFontOutlinesW is not supported");
+    return FALSE;
+}
+
+#endif // _WIN32

MobileGL/MG_Impl/WGLImpl/Exporting/wgl.def

@@ -0,0 +1,30 @@
+; MobileGL WGL exports. The wgl* entry points are defined without
+; __declspec(dllexport) because wingdi.h pre-declares them (with _GDI32_ they
+; become plain declarations, and MSVC rejects adding dllexport afterwards),
+; so this .def file is what actually exports them from the DLL.
+EXPORTS
+    wglChoosePixelFormat
+    wglCopyContext
+    wglCreateContext
+    wglCreateLayerContext
+    wglDeleteContext
+    wglDescribeLayerPlane
+    wglDescribePixelFormat
+    wglGetCurrentContext
+    wglGetCurrentDC
+    wglGetDefaultProcAddress
+    wglGetLayerPaletteEntries
+    wglGetPixelFormat
+    wglGetProcAddress
+    wglMakeCurrent
+    wglRealizeLayerPalette
+    wglSetLayerPaletteEntries
+    wglSetPixelFormat
+    wglShareLists
+    wglSwapBuffers
+    wglSwapLayerBuffers
+    wglSwapMultipleBuffers
+    wglUseFontBitmapsA
+    wglUseFontBitmapsW
+    wglUseFontOutlinesA
+    wglUseFontOutlinesW

MobileGL/MG_Impl/WGLImpl/WGLImpl.cpp

@@ -0,0 +1,732 @@
+// MobileGL - MobileGL/MG_Impl/WGLImpl/WGLImpl.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "WGLImpl.h"
+
+#if defined(_WIN32)
+#include "../EGLImpl/EGLImpl.h"
+#include "../GetProcAddress.h"
+#include <Init.h>
+
+namespace MobileGL::MG_Impl::WGLImpl {
+    namespace {
+        // WGL_ARB_pixel_format
+        constexpr int WGL_NUMBER_PIXEL_FORMATS_ARB = 0x2000;
+        constexpr int WGL_DRAW_TO_WINDOW_ARB = 0x2001;
+        constexpr int WGL_DRAW_TO_BITMAP_ARB = 0x2002;
+        constexpr int WGL_ACCELERATION_ARB = 0x2003;
+        constexpr int WGL_NEED_PALETTE_ARB = 0x2004;
+        constexpr int WGL_NEED_SYSTEM_PALETTE_ARB = 0x2005;
+        constexpr int WGL_SWAP_LAYER_BUFFERS_ARB = 0x2006;
+        constexpr int WGL_SWAP_METHOD_ARB = 0x2007;
+        constexpr int WGL_NUMBER_OVERLAYS_ARB = 0x2008;
+        constexpr int WGL_NUMBER_UNDERLAYS_ARB = 0x2009;
+        constexpr int WGL_TRANSPARENT_ARB = 0x200A;
+        constexpr int WGL_SHARE_DEPTH_ARB = 0x200C;
+        constexpr int WGL_SHARE_STENCIL_ARB = 0x200D;
+        constexpr int WGL_SHARE_ACCUM_ARB = 0x200E;
+        constexpr int WGL_SUPPORT_GDI_ARB = 0x200F;
+        constexpr int WGL_SUPPORT_OPENGL_ARB = 0x2010;
+        constexpr int WGL_DOUBLE_BUFFER_ARB = 0x2011;
+        constexpr int WGL_STEREO_ARB = 0x2012;
+        constexpr int WGL_PIXEL_TYPE_ARB = 0x2013;
+        constexpr int WGL_COLOR_BITS_ARB = 0x2014;
+        constexpr int WGL_RED_BITS_ARB = 0x2015;
+        constexpr int WGL_RED_SHIFT_ARB = 0x2016;
+        constexpr int WGL_GREEN_BITS_ARB = 0x2017;
+        constexpr int WGL_GREEN_SHIFT_ARB = 0x2018;
+        constexpr int WGL_BLUE_BITS_ARB = 0x2019;
+        constexpr int WGL_BLUE_SHIFT_ARB = 0x201A;
+        constexpr int WGL_ALPHA_BITS_ARB = 0x201B;
+        constexpr int WGL_ALPHA_SHIFT_ARB = 0x201C;
+        constexpr int WGL_ACCUM_BITS_ARB = 0x201D;
+        constexpr int WGL_ACCUM_RED_BITS_ARB = 0x201E;
+        constexpr int WGL_ACCUM_GREEN_BITS_ARB = 0x201F;
+        constexpr int WGL_ACCUM_BLUE_BITS_ARB = 0x2020;
+        constexpr int WGL_ACCUM_ALPHA_BITS_ARB = 0x2021;
+        constexpr int WGL_DEPTH_BITS_ARB = 0x2022;
+        constexpr int WGL_STENCIL_BITS_ARB = 0x2023;
+        constexpr int WGL_AUX_BUFFERS_ARB = 0x2024;
+        constexpr int WGL_NO_ACCELERATION_ARB = 0x2025;
+        constexpr int WGL_FULL_ACCELERATION_ARB = 0x2027;
+        constexpr int WGL_SWAP_EXCHANGE_ARB = 0x2028;
+        constexpr int WGL_TYPE_RGBA_ARB = 0x202B;
+        // WGL_ARB_multisample
+        constexpr int WGL_SAMPLE_BUFFERS_ARB = 0x2041;
+        constexpr int WGL_SAMPLES_ARB = 0x2042;
+        // WGL_ARB_create_context / _profile / _no_error
+        constexpr int WGL_CONTEXT_MAJOR_VERSION_ARB = 0x2091;
+        constexpr int WGL_CONTEXT_MINOR_VERSION_ARB = 0x2092;
+        constexpr int WGL_CONTEXT_LAYER_PLANE_ARB = 0x2093;
+        constexpr int WGL_CONTEXT_FLAGS_ARB = 0x2094;
+        constexpr int WGL_CONTEXT_PROFILE_MASK_ARB = 0x9126;
+        constexpr int WGL_CONTEXT_DEBUG_BIT_ARB = 0x0001;
+        constexpr int WGL_CONTEXT_FORWARD_COMPATIBLE_BIT_ARB = 0x0002;
+        constexpr int WGL_CONTEXT_CORE_PROFILE_BIT_ARB = 0x00000001;
+        constexpr int WGL_CONTEXT_COMPATIBILITY_PROFILE_BIT_ARB = 0x00000002;
+        constexpr int WGL_CONTEXT_OPENGL_NO_ERROR_ARB = 0x31B3;
+        constexpr DWORD ERROR_INVALID_VERSION_ARB = 0x2095;
+        constexpr DWORD ERROR_INVALID_PROFILE_ARB = 0x2096;
+
+        struct PixelFormatInfo {
+            GLint AlphaBits;
+            GLint DepthBits;
+            GLint StencilBits;
+        };
+
+        // Mirrors the two EGLState configs (RGBA8 + depth24, stencil 8 / stencil 0).
+        constexpr PixelFormatInfo kPixelFormats[] = {
+            {8, 24, 8},
+            {8, 24, 0},
+        };
+        constexpr int kPixelFormatCount = static_cast<int>(std::size(kPixelFormats));
+
+        struct ContextObject {
+            EGLDisplay Display = EGL_NO_DISPLAY;
+            EGLConfig Config = nullptr;
+            EGLContext Context = EGL_NO_CONTEXT;
+        };
+
+        struct WindowSurface {
+            EGLDisplay Display = EGL_NO_DISPLAY;
+            EGLSurface Surface = EGL_NO_SURFACE;
+            Uint32 Width = 0;
+            Uint32 Height = 0;
+        };
+
+        std::recursive_mutex& RegistryMutex() {
+            static auto* mutex = new std::recursive_mutex();
+            return *mutex;
+        }
+
+        UnorderedMap<HGLRC, ContextObject>& Contexts() {
+            static auto* contexts = new UnorderedMap<HGLRC, ContextObject>();
+            return *contexts;
+        }
+
+        UnorderedMap<HWND, WindowSurface>& WindowSurfaces() {
+            static auto* surfaces = new UnorderedMap<HWND, WindowSurface>();
+            return *surfaces;
+        }
+
+        UnorderedMap<HWND, int>& WindowPixelFormats() {
+            static auto* formats = new UnorderedMap<HWND, int>();
+            return *formats;
+        }
+
+        Uint64& NextContextHandle() {
+            static auto* handle = new Uint64(0x10000);
+            return *handle;
+        }
+
+        struct ThreadCurrent {
+            HDC DC = nullptr;
+            HGLRC Context = nullptr;
+        };
+        thread_local ThreadCurrent t_current;
+
+        Int& SwapIntervalShadow() {
+            static auto* interval = new Int(1);
+            return *interval;
+        }
+
+        void EnsureInitialized() {
+            // Initialize() loads backend libraries and glslang, which must not run
+            // under the loader lock; first WGL call is the earliest safe moment.
+            // MobileGL::EnsureInitialized (not a local once_flag) so a fresh init
+            // can follow a full teardown from the last eglTerminate.
+            MobileGL::EnsureInitialized();
+        }
+
+        EGLDisplay EnsureDisplay() {
+            EnsureInitialized();
+            EGLDisplay display = EGLImpl::GetDisplay(EGL_DEFAULT_DISPLAY);
+            if (display == EGL_NO_DISPLAY) {
+                return EGL_NO_DISPLAY;
+            }
+            if (!EGLImpl::Initialize(display, nullptr, nullptr)) {
+                return EGL_NO_DISPLAY;
+            }
+            return display;
+        }
+
+        HGLRC EncodeContext(Uint64 handle) {
+            return reinterpret_cast<HGLRC>(static_cast<SizeT>(handle));
+        }
+
+        ContextObject* TryGetContext(HGLRC hglrc) {
+            auto& contexts = Contexts();
+            auto it = contexts.find(hglrc);
+            return it == contexts.end() ? nullptr : &it->second;
+        }
+
+        const PixelFormatInfo& PixelFormatForWindow(HWND hwnd) {
+            auto& formats = WindowPixelFormats();
+            auto it = formats.find(hwnd);
+            int index = it == formats.end() ? 1 : it->second;
+            if (index < 1 || index > kPixelFormatCount) {
+                index = 1;
+            }
+            return kPixelFormats[index - 1];
+        }
+
+        Bool QueryClientSize(HWND hwnd, Uint32& width, Uint32& height) {
+            RECT rect{};
+            if (!GetClientRect(hwnd, &rect)) {
+                return false;
+            }
+            width = static_cast<Uint32>(std::max<LONG>(rect.right - rect.left, 1));
+            height = static_cast<Uint32>(std::max<LONG>(rect.bottom - rect.top, 1));
+            return true;
+        }
+
+        // The backends never query the HWND client size themselves; the WGL layer
+        // owns size discovery and pushes changes through the internal resize hook
+        // (same contract as the macOS CGL layer).
+        void SyncSurfaceSize(HWND hwnd, WindowSurface& surface) {
+            Uint32 width = 0;
+            Uint32 height = 0;
+            if (!QueryClientSize(hwnd, width, height)) {
+                return;
+            }
+            if (width == surface.Width && height == surface.Height) {
+                return;
+            }
+            if (EGLImpl::ResizePlatformWindowSurface(surface.Display, surface.Surface,
+                                                     static_cast<EGLint>(width), static_cast<EGLint>(height))) {
+                surface.Width = width;
+                surface.Height = height;
+            }
+        }
+
+        WindowSurface* EnsureWindowSurface(HWND hwnd, const ContextObject& context) {
+            auto& surfaces = WindowSurfaces();
+            auto it = surfaces.find(hwnd);
+            if (it != surfaces.end()) {
+                SyncSurfaceSize(hwnd, it->second);
+                return &it->second;
+            }
+
+            Uint32 width = 0;
+            Uint32 height = 0;
+            if (!QueryClientSize(hwnd, width, height)) {
+                MGLOG_E("wgl: GetClientRect failed for HWND %p", hwnd);
+                return nullptr;
+            }
+
+            const EGLAttrib attribs[] = {
+                EGL_WIDTH, static_cast<EGLAttrib>(width),
+                EGL_HEIGHT, static_cast<EGLAttrib>(height),
+                EGL_NONE,
+            };
+            EGLSurface surface =
+                EGLImpl::CreatePlatformWindowSurface(context.Display, context.Config, hwnd, attribs);
+            if (surface == EGL_NO_SURFACE) {
+                MGLOG_E("wgl: failed to create window surface for HWND %p (%ux%u)", hwnd, width, height);
+                return nullptr;
+            }
+
+            WindowSurface record;
+            record.Display = context.Display;
+            record.Surface = surface;
+            record.Width = width;
+            record.Height = height;
+            auto [inserted, _] = surfaces.emplace(hwnd, record);
+            return &inserted->second;
+        }
+
+        HGLRC CreateContextFromEGLAttribs(HDC hdc, HGLRC share, const EGLint* contextAttribs) {
+            const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+            EGLDisplay display = EnsureDisplay();
+            if (display == EGL_NO_DISPLAY) {
+                MGLOG_E("wgl: no EGL display");
+                return nullptr;
+            }
+            EGLImpl::BindAPI(EGL_OPENGL_API);
+
+            EGLContext shareContext = EGL_NO_CONTEXT;
+            if (share) {
+                auto* shareObject = TryGetContext(share);
+                if (!shareObject) {
+                    SetLastError(ERROR_INVALID_HANDLE);
+                    return nullptr;
+                }
+                shareContext = shareObject->Context;
+            }
+
+            HWND hwnd = WindowFromDC(hdc);
+            const PixelFormatInfo& pixelFormat = PixelFormatForWindow(hwnd);
+            const EGLint configAttribs[] = {
+                EGL_RED_SIZE, 8,
+                EGL_GREEN_SIZE, 8,
+                EGL_BLUE_SIZE, 8,
+                EGL_ALPHA_SIZE, pixelFormat.AlphaBits,
+                EGL_DEPTH_SIZE, pixelFormat.DepthBits,
+                EGL_STENCIL_SIZE, pixelFormat.StencilBits,
+                EGL_SURFACE_TYPE, EGL_WINDOW_BIT | EGL_PBUFFER_BIT,
+                EGL_RENDERABLE_TYPE, EGL_OPENGL_BIT,
+                EGL_NONE,
+            };
+            EGLConfig config = nullptr;
+            EGLint configCount = 0;
+            if (!EGLImpl::ChooseConfig(display, configAttribs, &config, 1, &configCount) || configCount <= 0) {
+                MGLOG_E("wgl: eglChooseConfig failed");
+                return nullptr;
+            }
+
+            EGLContext eglContext = EGLImpl::CreateContext(display, config, shareContext, contextAttribs);
+            if (eglContext == EGL_NO_CONTEXT) {
+                MGLOG_E("wgl: eglCreateContext failed");
+                return nullptr;
+            }
+
+            ContextObject object;
+            object.Display = display;
+            object.Config = config;
+            object.Context = eglContext;
+            const auto handle = EncodeContext(NextContextHandle()++);
+            Contexts()[handle] = object;
+            MGLOG_I("wgl: created context %p (EGL context %p)", handle, eglContext);
+            return handle;
+        }
+
+        // ---- WGL extension entry points (resolved via wglGetProcAddress only) ----
+
+        const char* WINAPI Ext_GetExtensionsStringARB(HDC) {
+            return "WGL_ARB_create_context WGL_ARB_create_context_no_error WGL_ARB_create_context_profile "
+                   "WGL_ARB_extensions_string WGL_ARB_pixel_format WGL_EXT_extensions_string WGL_EXT_swap_control";
+        }
+
+        const char* WINAPI Ext_GetExtensionsStringEXT() {
+            return Ext_GetExtensionsStringARB(nullptr);
+        }
+
+        HGLRC WINAPI Ext_CreateContextAttribsARB(HDC hdc, HGLRC hShareContext, const int* attribList) {
+            EnsureInitialized();
+            int major = 1;
+            int minor = 0;
+            int profileMask = 0;
+            int flags = 0;
+            if (attribList) {
+                for (SizeT i = 0; attribList[i] != 0; i += 2) {
+                    const int attrib = attribList[i];
+                    const int value = attribList[i + 1];
+                    switch (attrib) {
+                    case WGL_CONTEXT_MAJOR_VERSION_ARB:
+                        major = value;
+                        break;
+                    case WGL_CONTEXT_MINOR_VERSION_ARB:
+                        minor = value;
+                        break;
+                    case WGL_CONTEXT_PROFILE_MASK_ARB:
+                        profileMask = value;
+                        break;
+                    case WGL_CONTEXT_FLAGS_ARB:
+                        flags = value;
+                        break;
+                    case WGL_CONTEXT_LAYER_PLANE_ARB:
+                        if (value != 0) {
+                            SetLastError(ERROR_INVALID_PARAMETER);
+                            return nullptr;
+                        }
+                        break;
+                    case WGL_CONTEXT_OPENGL_NO_ERROR_ARB:
+                        // Accepted and ignored: MobileGL always validates.
+                        break;
+                    default:
+                        MGLOG_D("wglCreateContextAttribsARB: ignoring attrib 0x%04x = 0x%x", attrib, value);
+                        break;
+                    }
+                }
+            }
+
+            if (major < 1 || (profileMask & ~(WGL_CONTEXT_CORE_PROFILE_BIT_ARB |
+                                              WGL_CONTEXT_COMPATIBILITY_PROFILE_BIT_ARB))) {
+                SetLastError(profileMask ? ERROR_INVALID_PROFILE_ARB : ERROR_INVALID_VERSION_ARB);
+                return nullptr;
+            }
+
+            Vector<EGLint> attribs = {
+                EGL_CONTEXT_MAJOR_VERSION, major,
+                EGL_CONTEXT_MINOR_VERSION, minor,
+            };
+            const Bool wantsCompat = (profileMask & WGL_CONTEXT_COMPATIBILITY_PROFILE_BIT_ARB) != 0;
+            if (major > 3 || (major == 3 && minor >= 2) || profileMask != 0) {
+                attribs.push_back(EGL_CONTEXT_OPENGL_PROFILE_MASK);
+                attribs.push_back(wantsCompat ? EGL_CONTEXT_OPENGL_COMPATIBILITY_PROFILE_BIT
+                                              : EGL_CONTEXT_OPENGL_CORE_PROFILE_BIT);
+            }
+            if (flags & WGL_CONTEXT_FORWARD_COMPATIBLE_BIT_ARB) {
+                attribs.push_back(EGL_CONTEXT_OPENGL_FORWARD_COMPATIBLE);
+                attribs.push_back(EGL_TRUE);
+            }
+            if (flags & WGL_CONTEXT_DEBUG_BIT_ARB) {
+                attribs.push_back(EGL_CONTEXT_OPENGL_DEBUG);
+                attribs.push_back(EGL_TRUE);
+            }
+            attribs.push_back(EGL_NONE);
+
+            return CreateContextFromEGLAttribs(hdc, hShareContext, attribs.data());
+        }
+
+        BOOL WINAPI Ext_SwapIntervalEXT(int interval) {
+            const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+            EGLDisplay display = EnsureDisplay();
+            if (display == EGL_NO_DISPLAY) {
+                return FALSE;
+            }
+            if (interval < 0) {
+                // Adaptive vsync is not supported; clamp to regular vsync.
+                interval = 1;
+            }
+            EGLImpl::SwapInterval(display, interval);
+            SwapIntervalShadow() = interval;
+            return TRUE;
+        }
+
+        int WINAPI Ext_GetSwapIntervalEXT() {
+            const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+            return SwapIntervalShadow();
+        }
+
+        int PixelFormatAttribValue(int format, int attrib) {
+            const PixelFormatInfo& info = kPixelFormats[format - 1];
+            switch (attrib) {
+            case WGL_NUMBER_PIXEL_FORMATS_ARB:
+                return kPixelFormatCount;
+            case WGL_SUPPORT_OPENGL_ARB:
+            case WGL_DRAW_TO_WINDOW_ARB:
+            case WGL_DOUBLE_BUFFER_ARB:
+                return 1;
+            case WGL_ACCELERATION_ARB:
+                return WGL_FULL_ACCELERATION_ARB;
+            case WGL_PIXEL_TYPE_ARB:
+                return WGL_TYPE_RGBA_ARB;
+            case WGL_COLOR_BITS_ARB:
+                return 32;
+            case WGL_RED_BITS_ARB:
+            case WGL_GREEN_BITS_ARB:
+            case WGL_BLUE_BITS_ARB:
+                return 8;
+            case WGL_RED_SHIFT_ARB:
+                return 16;
+            case WGL_GREEN_SHIFT_ARB:
+                return 8;
+            case WGL_BLUE_SHIFT_ARB:
+                return 0;
+            case WGL_ALPHA_BITS_ARB:
+                return info.AlphaBits;
+            case WGL_ALPHA_SHIFT_ARB:
+                return 24;
+            case WGL_DEPTH_BITS_ARB:
+                return info.DepthBits;
+            case WGL_STENCIL_BITS_ARB:
+                return info.StencilBits;
+            case WGL_SWAP_METHOD_ARB:
+                return WGL_SWAP_EXCHANGE_ARB;
+            case WGL_DRAW_TO_BITMAP_ARB:
+            case WGL_NEED_PALETTE_ARB:
+            case WGL_NEED_SYSTEM_PALETTE_ARB:
+            case WGL_SWAP_LAYER_BUFFERS_ARB:
+            case WGL_NUMBER_OVERLAYS_ARB:
+            case WGL_NUMBER_UNDERLAYS_ARB:
+            case WGL_TRANSPARENT_ARB:
+            case WGL_SHARE_DEPTH_ARB:
+            case WGL_SHARE_STENCIL_ARB:
+            case WGL_SHARE_ACCUM_ARB:
+            case WGL_SUPPORT_GDI_ARB:
+            case WGL_STEREO_ARB:
+            case WGL_ACCUM_BITS_ARB:
+            case WGL_ACCUM_RED_BITS_ARB:
+            case WGL_ACCUM_GREEN_BITS_ARB:
+            case WGL_ACCUM_BLUE_BITS_ARB:
+            case WGL_ACCUM_ALPHA_BITS_ARB:
+            case WGL_AUX_BUFFERS_ARB:
+            case WGL_SAMPLE_BUFFERS_ARB:
+            case WGL_SAMPLES_ARB:
+            default:
+                return 0;
+            }
+        }
+
+        BOOL WINAPI Ext_GetPixelFormatAttribivARB(HDC, int iPixelFormat, int iLayerPlane, UINT nAttributes,
+                                                  const int* piAttributes, int* piValues) {
+            if (iLayerPlane != 0 || !piAttributes || !piValues) {
+                return FALSE;
+            }
+            // Format 0 is only valid for WGL_NUMBER_PIXEL_FORMATS_ARB queries.
+            if (iPixelFormat < 0 || iPixelFormat > kPixelFormatCount) {
+                return FALSE;
+            }
+            const int format = iPixelFormat == 0 ? 1 : iPixelFormat;
+            for (UINT i = 0; i < nAttributes; ++i) {
+                piValues[i] = PixelFormatAttribValue(format, piAttributes[i]);
+            }
+            return TRUE;
+        }
+
+        BOOL WINAPI Ext_GetPixelFormatAttribfvARB(HDC hdc, int iPixelFormat, int iLayerPlane, UINT nAttributes,
+                                                  const int* piAttributes, FLOAT* pfValues) {
+            if (!pfValues) {
+                return FALSE;
+            }
+            Vector<int> values(nAttributes);
+            if (!Ext_GetPixelFormatAttribivARB(hdc, iPixelFormat, iLayerPlane, nAttributes, piAttributes,
+                                               values.data())) {
+                return FALSE;
+            }
+            for (UINT i = 0; i < nAttributes; ++i) {
+                pfValues[i] = static_cast<FLOAT>(values[i]);
+            }
+            return TRUE;
+        }
+
+        BOOL WINAPI Ext_ChoosePixelFormatARB(HDC, const int* piAttribIList, const FLOAT*, UINT nMaxFormats,
+                                             int* piFormats, UINT* nNumFormats) {
+            if (!piFormats || !nNumFormats) {
+                return FALSE;
+            }
+            int wantedStencil = 0;
+            if (piAttribIList) {
+                for (SizeT i = 0; piAttribIList[i] != 0; i += 2) {
+                    if (piAttribIList[i] == WGL_STENCIL_BITS_ARB) {
+                        wantedStencil = piAttribIList[i + 1];
+                    }
+                }
+            }
+            UINT count = 0;
+            const int preferred = wantedStencil > 0 ? 1 : 2;
+            const int fallback = wantedStencil > 0 ? 2 : 1;
+            if (count < nMaxFormats) {
+                piFormats[count++] = preferred;
+            }
+            if (count < nMaxFormats) {
+                piFormats[count++] = fallback;
+            }
+            *nNumFormats = count;
+            return TRUE;
+        }
+
+        struct WGLExtensionProc {
+            const char* Name;
+            PROC Proc;
+        };
+
+        const WGLExtensionProc kWGLExtensionProcs[] = {
+            {"wglGetExtensionsStringARB", reinterpret_cast<PROC>(Ext_GetExtensionsStringARB)},
+            {"wglGetExtensionsStringEXT", reinterpret_cast<PROC>(Ext_GetExtensionsStringEXT)},
+            {"wglCreateContextAttribsARB", reinterpret_cast<PROC>(Ext_CreateContextAttribsARB)},
+            {"wglSwapIntervalEXT", reinterpret_cast<PROC>(Ext_SwapIntervalEXT)},
+            {"wglGetSwapIntervalEXT", reinterpret_cast<PROC>(Ext_GetSwapIntervalEXT)},
+            {"wglGetPixelFormatAttribivARB", reinterpret_cast<PROC>(Ext_GetPixelFormatAttribivARB)},
+            {"wglGetPixelFormatAttribfvARB", reinterpret_cast<PROC>(Ext_GetPixelFormatAttribfvARB)},
+            {"wglChoosePixelFormatARB", reinterpret_cast<PROC>(Ext_ChoosePixelFormatARB)},
+        };
+    } // namespace
+
+    int ChoosePixelFormat(HDC hdc, const PIXELFORMATDESCRIPTOR* pfd) {
+        EnsureInitialized();
+        MGLOG_D("wglChoosePixelFormat(hdc=%p)", hdc);
+        // Format 1 (RGBA8 + depth24/stencil8) satisfies every request; a format
+        // exceeding the asked-for capabilities is a legal ChoosePixelFormat answer.
+        (void)pfd;
+        return 1;
+    }
+
+    int DescribePixelFormat(HDC hdc, int format, UINT size, PIXELFORMATDESCRIPTOR* pfd) {
+        EnsureInitialized();
+        MGLOG_D("wglDescribePixelFormat(hdc=%p, format=%d)", hdc, format);
+        if (!pfd) {
+            return kPixelFormatCount;
+        }
+        if (size < sizeof(PIXELFORMATDESCRIPTOR) || format < 1 || format > kPixelFormatCount) {
+            return 0;
+        }
+
+        const PixelFormatInfo& info = kPixelFormats[format - 1];
+        std::memset(pfd, 0, sizeof(PIXELFORMATDESCRIPTOR));
+        pfd->nSize = sizeof(PIXELFORMATDESCRIPTOR);
+        pfd->nVersion = 1;
+        pfd->dwFlags = PFD_DRAW_TO_WINDOW | PFD_SUPPORT_OPENGL | PFD_DOUBLEBUFFER | PFD_SWAP_EXCHANGE
+#if defined(PFD_SUPPORT_COMPOSITION)
+                       | PFD_SUPPORT_COMPOSITION
+#endif
+            ;
+        pfd->iPixelType = PFD_TYPE_RGBA;
+        pfd->cColorBits = 32;
+        pfd->cRedBits = 8;
+        pfd->cRedShift = 16;
+        pfd->cGreenBits = 8;
+        pfd->cGreenShift = 8;
+        pfd->cBlueBits = 8;
+        pfd->cBlueShift = 0;
+        pfd->cAlphaBits = static_cast<BYTE>(info.AlphaBits);
+        pfd->cAlphaShift = 24;
+        pfd->cDepthBits = static_cast<BYTE>(info.DepthBits);
+        pfd->cStencilBits = static_cast<BYTE>(info.StencilBits);
+        pfd->iLayerType = PFD_MAIN_PLANE;
+        return kPixelFormatCount;
+    }
+
+    int GetPixelFormat(HDC hdc) {
+        EnsureInitialized();
+        HWND hwnd = WindowFromDC(hdc);
+        if (!hwnd) {
+            return 0;
+        }
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        auto& formats = WindowPixelFormats();
+        auto it = formats.find(hwnd);
+        return it == formats.end() ? 0 : it->second;
+    }
+
+    BOOL SetPixelFormat(HDC hdc, int format, const PIXELFORMATDESCRIPTOR*) {
+        EnsureInitialized();
+        MGLOG_D("wglSetPixelFormat(hdc=%p, format=%d)", hdc, format);
+        if (format < 1 || format > kPixelFormatCount) {
+            SetLastError(ERROR_INVALID_PARAMETER);
+            return FALSE;
+        }
+        HWND hwnd = WindowFromDC(hdc);
+        if (!hwnd) {
+            SetLastError(ERROR_INVALID_HANDLE);
+            return FALSE;
+        }
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        WindowPixelFormats()[hwnd] = format;
+        return TRUE;
+    }
+
+    BOOL SwapBuffers(HDC hdc) {
+        HWND hwnd = WindowFromDC(hdc);
+        if (!hwnd) {
+            SetLastError(ERROR_INVALID_HANDLE);
+            return FALSE;
+        }
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        auto& surfaces = WindowSurfaces();
+        auto it = surfaces.find(hwnd);
+        if (it == surfaces.end()) {
+            MGLOG_W("wglSwapBuffers: no surface for HWND %p", hwnd);
+            return FALSE;
+        }
+        SyncSurfaceSize(hwnd, it->second);
+        return EGLImpl::SwapBuffers(it->second.Display, it->second.Surface) == EGL_TRUE ? TRUE : FALSE;
+    }
+
+    HGLRC CreateContext(HDC hdc) {
+        EnsureInitialized();
+        MGLOG_I("wglCreateContext(hdc=%p)", hdc);
+        // A legacy WGL context is a compatibility-profile context; MobileGL keys
+        // its relaxed-semantics mode off the explicit compatibility bit.
+        const EGLint attribs[] = {
+            EGL_CONTEXT_MAJOR_VERSION, 3,
+            EGL_CONTEXT_MINOR_VERSION, 3,
+            EGL_CONTEXT_OPENGL_PROFILE_MASK, EGL_CONTEXT_OPENGL_COMPATIBILITY_PROFILE_BIT,
+            EGL_NONE,
+        };
+        return CreateContextFromEGLAttribs(hdc, nullptr, attribs);
+    }
+
+    BOOL DeleteContext(HGLRC hglrc) {
+        EnsureInitialized();
+        MGLOG_I("wglDeleteContext(%p)", hglrc);
+        if (t_current.Context == hglrc) {
+            MakeCurrent(nullptr, nullptr);
+        }
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        auto* object = TryGetContext(hglrc);
+        if (!object) {
+            SetLastError(ERROR_INVALID_HANDLE);
+            return FALSE;
+        }
+        if (object->Context != EGL_NO_CONTEXT) {
+            EGLImpl::DestroyContext(object->Display, object->Context);
+        }
+        Contexts().erase(hglrc);
+        return TRUE;
+    }
+
+    BOOL MakeCurrent(HDC hdc, HGLRC hglrc) {
+        EnsureInitialized();
+        MGLOG_D("wglMakeCurrent(hdc=%p, hglrc=%p)", hdc, hglrc);
+        if (!hglrc) {
+            if (!t_current.Context) {
+                t_current = {};
+                return TRUE;
+            }
+            const EGLBoolean released =
+                EGLImpl::MakeCurrent(EGL_NO_DISPLAY, EGL_NO_SURFACE, EGL_NO_SURFACE, EGL_NO_CONTEXT);
+            t_current = {};
+            return released == EGL_TRUE ? TRUE : FALSE;
+        }
+
+        HWND hwnd = WindowFromDC(hdc);
+        if (!hwnd) {
+            SetLastError(ERROR_INVALID_HANDLE);
+            return FALSE;
+        }
+
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        auto* object = TryGetContext(hglrc);
+        if (!object) {
+            SetLastError(ERROR_INVALID_HANDLE);
+            return FALSE;
+        }
+
+        WindowSurface* surface = EnsureWindowSurface(hwnd, *object);
+        if (!surface) {
+            return FALSE;
+        }
+
+        if (!EGLImpl::MakeCurrent(object->Display, surface->Surface, surface->Surface, object->Context)) {
+            MGLOG_E("wglMakeCurrent: eglMakeCurrent failed (hdc=%p, hglrc=%p)", hdc, hglrc);
+            return FALSE;
+        }
+        t_current = {hdc, hglrc};
+        return TRUE;
+    }
+
+    HGLRC GetCurrentContext() {
+        return t_current.Context;
+    }
+
+    HDC GetCurrentDC() {
+        return t_current.DC;
+    }
+
+    BOOL ShareLists(HGLRC hglrcShare, HGLRC hglrcDest) {
+        // All MobileGL contexts alias one global GL object namespace, so every
+        // pair of contexts already shares.
+        const std::lock_guard<std::recursive_mutex> lock(RegistryMutex());
+        if (!TryGetContext(hglrcShare) || !TryGetContext(hglrcDest)) {
+            SetLastError(ERROR_INVALID_HANDLE);
+            return FALSE;
+        }
+        return TRUE;
+    }
+
+    PROC GetProcAddress(const char* name) {
+        EnsureInitialized();
+        if (!name) {
+            return nullptr;
+        }
+        if (name[0] == 'w' && name[1] == 'g' && name[2] == 'l') {
+            for (const auto& entry : kWGLExtensionProcs) {
+                if (std::strcmp(entry.Name, name) == 0) {
+                    return entry.Proc;
+                }
+            }
+            MGLOG_D("wglGetProcAddress: unknown wgl entry point %s", name);
+            return nullptr;
+        }
+        return reinterpret_cast<PROC>(MG_Impl::GetProcAddress(name));
+    }
+} // namespace MobileGL::MG_Impl::WGLImpl
+
+#endif // _WIN32

MobileGL/MG_Impl/WGLImpl/WGLImpl.h

@@ -0,0 +1,34 @@
+// MobileGL - MobileGL/MG_Impl/WGLImpl/WGLImpl.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include <Includes.h>
+
+#if defined(_WIN32)
+
+namespace MobileGL::MG_Impl::WGLImpl {
+    // Classic opengl32.dll surface. gdi32's ChoosePixelFormat/SetPixelFormat/
+    // DescribePixelFormat/GetPixelFormat/SwapBuffers forward into the loaded
+    // opengl32.dll's wgl* exports, so these back both call paths.
+    int ChoosePixelFormat(HDC hdc, const PIXELFORMATDESCRIPTOR* pfd);
+    int DescribePixelFormat(HDC hdc, int format, UINT size, PIXELFORMATDESCRIPTOR* pfd);
+    int GetPixelFormat(HDC hdc);
+    BOOL SetPixelFormat(HDC hdc, int format, const PIXELFORMATDESCRIPTOR* pfd);
+    BOOL SwapBuffers(HDC hdc);
+
+    HGLRC CreateContext(HDC hdc);
+    BOOL DeleteContext(HGLRC hglrc);
+    BOOL MakeCurrent(HDC hdc, HGLRC hglrc);
+    HGLRC GetCurrentContext();
+    HDC GetCurrentDC();
+    BOOL ShareLists(HGLRC hglrcShare, HGLRC hglrcDest);
+
+    PROC GetProcAddress(const char* name);
+} // namespace MobileGL::MG_Impl::WGLImpl
+
+#endif // _WIN32

MobileGL/MG_State/EGLState/Core.cpp

@@ -368,6 +368,26 @@ namespace MobileGL {
                 return true;
             }
 
+            Bool EGLContext::HasAnyInitializedDisplay() const {
+                const std::lock_guard<std::recursive_mutex> lock(m_mutex);
+                for (const auto& [handle, displayObject] : m_displays) {
+                    if (displayObject.Initialized) {
+                        return true;
+                    }
+                }
+                return false;
+            }
+
+            Bool EGLContext::HasAnyCurrentContext() const {
+                const std::lock_guard<std::recursive_mutex> lock(m_mutex);
+                for (const auto& [threadId, current] : m_threadCurrents) {
+                    if (current.Context != nullptr) {
+                        return true;
+                    }
+                }
+                return false;
+            }
+
             Bool EGLContext::ChooseConfig(EGLDisplayHandle display, const EGLint* attribList, EGLConfigHandle* configs,
                                           EGLint configSize, EGLint* numConfig) {
                 const std::lock_guard<std::recursive_mutex> lock(m_mutex);
@@ -1415,6 +1435,7 @@ namespace MobileGL {
             }
         } // namespace EGLState
 
-        UniquePtr<EGLState::EGLContext> pEGLContext;
+        // Leak-at-exit storage; see GlobalObjects.cpp.
+        UniquePtr<EGLState::EGLContext>& pEGLContext = *new UniquePtr<EGLState::EGLContext>();
     } // namespace MG_State
 } // namespace MobileGL

MobileGL/MG_State/EGLState/Core.h

@@ -39,6 +39,10 @@ namespace MobileGL {
                 Bool IsDisplayInitialized(EGLDisplayHandle display) const;
                 Bool InitializeDisplay(EGLDisplayHandle display, EGLint* major, EGLint* minor);
                 Bool TerminateDisplay(EGLDisplayHandle display);
+                // Whole-library idle checks used by EGLImpl::Terminate to decide
+                // when the last eglTerminate may tear MobileGL down entirely.
+                Bool HasAnyInitializedDisplay() const;
+                Bool HasAnyCurrentContext() const;
 
                 // Config
                 Bool ChooseConfig(EGLDisplayHandle display, const EGLint* attribList, EGLConfigHandle* configs,
@@ -262,6 +266,6 @@ namespace MobileGL {
             };
         } // namespace EGLState
 
-        extern UniquePtr<EGLState::EGLContext> pEGLContext;
+        extern UniquePtr<EGLState::EGLContext>& pEGLContext;
     } // namespace MG_State
 } // namespace MobileGL

MobileGL/MG_State/GLState/Core.cpp

@@ -721,5 +721,6 @@ namespace MobileGL::MG_State {
         }
     } // namespace GLState
 
-    UniquePtr<GLState::GLContext> pGLContext;
+    // Leak-at-exit storage; see GlobalObjects.cpp.
+    UniquePtr<GLState::GLContext>& pGLContext = *new UniquePtr<GLState::GLContext>();
 } // namespace MobileGL::MG_State

MobileGL/MG_State/GLState/Core.h

@@ -252,7 +252,7 @@ namespace MobileGL {
             };
         } // namespace GLState
 
-        extern UniquePtr<GLState::GLContext> pGLContext;
+        extern UniquePtr<GLState::GLContext>& pGLContext;
 
         // True when relaxed GL semantics apply. Strict core rules are enforced only when the
         // current EGL context explicitly requested a core profile (core bit in

MobileGL/MG_State/GLState/ProgramState/ProgramObject.h

@@ -334,16 +334,32 @@ namespace MobileGL::MG_State::GLState {
         // draw. The memo is keyed by (backendStateVersion, flags); ResetLinkArtifacts and
         // the binding setters below invalidate it by bumping m_backendStateVersion.
         Bool GetBackendHashMemo(Uint flags, Uint64& outHash) const {
-            if (m_backendHashMemoVersion != m_backendStateVersion || m_backendHashMemoFlags != flags) {
-                return false;
+            if (m_backendHashMemoVersion != m_backendStateVersion) return false;
+            for (const auto& slot : m_backendHashMemoSlots) {
+                if (slot.valid && slot.flags == flags) {
+                    outHash = slot.hash;
+                    return true;
+                }
             }
-            outHash = m_backendHashMemo;
-            return true;
+            return false;
         }
         void SetBackendHashMemo(Uint flags, Uint64 hash) const {
-            m_backendHashMemo = hash;
-            m_backendHashMemoVersion = m_backendStateVersion;
-            m_backendHashMemoFlags = flags;
+            if (m_backendHashMemoVersion != m_backendStateVersion) {
+                for (auto& slot : m_backendHashMemoSlots) slot.valid = false;
+                m_backendHashMemoVersion = m_backendStateVersion;
+                m_backendHashMemoNextSlot = 0;
+            }
+            for (auto& slot : m_backendHashMemoSlots) {
+                if (slot.valid && slot.flags == flags) {
+                    slot.hash = hash;
+                    return;
+                }
+            }
+            auto& slot = m_backendHashMemoSlots[m_backendHashMemoNextSlot];
+            slot.flags = flags;
+            slot.hash = hash;
+            slot.valid = true;
+            m_backendHashMemoNextSlot = (m_backendHashMemoNextSlot + 1) % kBackendHashMemoSlotCount;
         }
 
         void SetUniformSamplerOrImageUnitIndex(Uint location, Int unit) {
@@ -527,10 +543,19 @@ namespace MobileGL::MG_State::GLState {
         Uint32 m_backendStateVersion = 0;
 
         // Backend-owned content-hash memo (see GetBackendHashMemo): valid only while
-        // m_backendStateVersion and the compile flags match the recorded values.
-        mutable Uint64 m_backendHashMemo = 0;
+        // m_backendStateVersion matches. Several slots, not one: a backend may resolve the same
+        // program under more than one compile-flag set within a frame (surface rotation, and the
+        // explicit-LOD sampling variant), and a single slot would then miss on every lookup and
+        // re-hash the program's whole SPIR-V once per draw.
+        static constexpr SizeT kBackendHashMemoSlotCount = 4;
+        struct BackendHashMemoSlot {
+            Uint64 hash = 0;
+            Uint flags = 0;
+            Bool valid = false;
+        };
+        mutable Array<BackendHashMemoSlot, kBackendHashMemoSlotCount> m_backendHashMemoSlots{};
+        mutable SizeT m_backendHashMemoNextSlot = 0;
         mutable Uint32 m_backendHashMemoVersion = ~0u;
-        mutable Uint m_backendHashMemoFlags = 0;
         Uint32 m_uboContentVersion = 0;
         Uint32 m_linkVersion = 0;
     };

MobileGL/MG_State/GLState/TextureState/MipmapStorage.cpp

@@ -16,17 +16,32 @@ namespace MobileGL {
             }
 
             void MipmapStorage::AllocateLevel(Uint level, MipmapInput input) {
-                m_data.reserve(std::bit_ceil(level + 1));
-                m_data.resize(level + 1);
-                m_texelSizes.reserve(std::bit_ceil(level + 1));
-                m_texelSizes.resize(level + 1);
-                m_texelSizes[level] = input.texelSize;
-                m_isDirty.resize(level + 1, false);
+                // Grow only. GL respecifies exactly the level it is handed, so allocating level 0
+                // must not disturb the levels above it - but resize() shrinks as readily as it
+                // grows, so this used to truncate the whole chain to a single level. Callers that
+                // genuinely redefine the complete level set say so with TruncateToLevelCount.
+                const SizeT requiredLevelCount = static_cast<SizeT>(level) + 1;
+                if (m_data.size() < requiredLevelCount) {
+                    m_data.reserve(std::bit_ceil(requiredLevelCount));
+                    m_data.resize(requiredLevelCount);
+                    m_texelSizes.reserve(std::bit_ceil(requiredLevelCount));
+                    m_texelSizes.resize(requiredLevelCount);
+                    m_isDirty.resize(requiredLevelCount, false);
+                }
 
+                m_texelSizes[level] = input.texelSize;
                 auto& data = m_data[level];
                 data.resize(input.byteSize, 0);
             }
 
+            void MipmapStorage::TruncateToLevelCount(SizeT levelCount) {
+                if (levelCount >= m_data.size()) return;
+
+                m_data.resize(levelCount);
+                m_texelSizes.resize(levelCount);
+                m_isDirty.resize(levelCount);
+            }
+
             void MipmapStorage::UpdateSubData(Uint level, DataPtr input) {
                 auto& targetData = m_data;
                 MOBILEGL_ASSERT(level < targetData.size(), "UpdateSubData: level out of range");
@@ -55,6 +70,7 @@ namespace MobileGL {
             }
 
             SizeT MipmapStorage::GetByteSize(Uint level) const {
+                if (level >= m_data.size()) return 0;
                 return m_data[level].size();
             }
 

MobileGL/MG_State/GLState/TextureState/MipmapStorage.h

@@ -19,6 +19,10 @@ namespace MobileGL {
             public:
                 SizeT GetLevelCount() const;
                 void AllocateLevel(Uint level, MipmapInput input);
+                // Discard every level at or above levelCount. AllocateLevel never shrinks, so this
+                // is the only way a chain gets shorter - use it where the caller defines the whole
+                // level set (glTexStorage*, mip regeneration, atlas respecification).
+                void TruncateToLevelCount(SizeT levelCount);
                 void UpdateSubData(Uint level, DataPtr input);
                 void* MapData(Uint level);
                 IntVec3 GetTexelSize(Uint level) const;

MobileGL/MG_State/GLState/TextureState/MipmapUploadTargetArray.h

@@ -29,6 +29,14 @@ namespace MobileGL {
                     m_storage[targetIndex].AllocateLevel(level, input);
                 }
 
+                // Per-target, like AllocateLevel: cube-map faces are respecified independently, so
+                // truncating one face must not disturb the others.
+                void TruncateToLevelCount(Uint targetIndex, SizeT levelCount) {
+                    MOBILEGL_ASSERT(targetIndex < TargetCount, "TruncateToLevelCount: target invalid");
+
+                    m_storage[targetIndex].TruncateToLevelCount(levelCount);
+                }
+
                 void UpdateSubData(Uint targetIndex, Uint level, DataPtr input) {
                     MOBILEGL_ASSERT(targetIndex < TargetCount, "UpdateSubData: target invalid");
                     m_storage[targetIndex].UpdateSubData(level, input);

MobileGL/MG_State/GLState/TextureState/TextureObject.cpp

@@ -271,6 +271,10 @@ namespace MobileGL {
                 m_textureStorage.AllocateLevel(GetIndexOfTextureUploadTarget(uploadTarget), mipmapLevel, input);
             }
 
+            void TextureObjectWithOneMipmap::TruncateMipmapLevels(TextureUploadTarget uploadTarget, Uint levelCount) {
+                m_textureStorage.TruncateToLevelCount(GetIndexOfTextureUploadTarget(uploadTarget), levelCount);
+            }
+
             void TextureObjectWithOneMipmap::UpdateMipmapSubData(TextureUploadTarget uploadTarget, Uint mipmapLevel,
                                                                  DataPtr input) {
                 m_textureStorage.UpdateSubData(GetIndexOfTextureUploadTarget(uploadTarget), mipmapLevel, input);

MobileGL/MG_State/GLState/TextureState/TextureObject.h

@@ -15,7 +15,11 @@
 #include <MG_Util/Math/VectorTypes.h>
 
 namespace MobileGL::MG_State::GLState {
-    class ITextureObject {
+    // Texture objects are always SharedPtr-owned (TextureState creates every instance via
+    // MakeShared, including the per-target default objects). enable_shared_from_this lets
+    // backends that only receive a reference (e.g. syncing a name-deleted texture kept
+    // alive by an FBO attachment) still register a weak liveness reference for GC.
+    class ITextureObject : public std::enable_shared_from_this<ITextureObject> {
     public:
         using TargetEnum = TextureTarget;
         virtual ~ITextureObject() = default;
@@ -134,6 +138,10 @@ namespace MobileGL::MG_State::GLState {
         virtual const IntVec3 GetMipmapTexelSize(TextureUploadTarget target, Uint mipmapLevel) const = 0;
         virtual const SizeT GetMipmapByteSize(TextureUploadTarget target, Uint mipmapLevel) const = 0;
         virtual void AllocateStorage(TextureUploadTarget uploadTarget, Uint mipmapLevel, MipmapInput input) = 0;
+        // AllocateStorage only ever grows the chain. Callers that define the complete level set -
+        // glTexStorage*, mip regeneration, or a level-0 respecification at a new size - drop the
+        // leftovers explicitly, so a stale tail can never make the texture silently incomplete.
+        virtual void TruncateMipmapLevels(TextureUploadTarget uploadTarget, Uint levelCount) = 0;
         virtual void UpdateMipmapSubData(TextureUploadTarget uploadTarget, Uint mipmapLevel, DataPtr input) = 0;
         virtual void* MapMipmapData(TextureUploadTarget uploadTarget, Uint mipmapLevel) = 0;
         virtual void MarkStorageDirty(TextureUploadTarget uploadTarget, Uint mipmapLevel, Bool dirty = true) = 0;
@@ -175,6 +183,7 @@ namespace MobileGL::MG_State::GLState {
         const IntVec3 GetMipmapTexelSize(TextureUploadTarget target, Uint mipmapLevel) const override;
         const SizeT GetMipmapByteSize(TextureUploadTarget target, Uint mipmapLevel) const override;
         void AllocateStorage(TextureUploadTarget uploadTarget, Uint mipmapLevel, MipmapInput input) override;
+        void TruncateMipmapLevels(TextureUploadTarget uploadTarget, Uint levelCount) override;
         void UpdateMipmapSubData(TextureUploadTarget uploadTarget, Uint mipmapLevel, DataPtr input) override;
         void* MapMipmapData(TextureUploadTarget uploadTarget, Uint mipmapLevel) override;
         void MarkStorageDirty(TextureUploadTarget uploadTarget, Uint mipmapLevel, Bool dirty) override;

MobileGL/MG_State/GLState/TextureState/TextureObject2DCube.cpp

@@ -31,6 +31,10 @@ namespace MobileGL {
                 m_textureStorage.AllocateLevel(GetIndexOfTextureUploadTarget(uploadTarget), mipmapLevel, input);
             }
 
+            void TextureObject2DCube::TruncateMipmapLevels(TextureUploadTarget uploadTarget, Uint levelCount) {
+                m_textureStorage.TruncateToLevelCount(GetIndexOfTextureUploadTarget(uploadTarget), levelCount);
+            }
+
             void TextureObject2DCube::UpdateMipmapSubData(TextureUploadTarget uploadTarget, Uint mipmapLevel,
                                                           DataPtr input) {
                 m_textureStorage.UpdateSubData(GetIndexOfTextureUploadTarget(uploadTarget), mipmapLevel, input);

MobileGL/MG_State/GLState/TextureState/TextureObject2DCube.h

@@ -22,6 +22,7 @@ namespace MobileGL {
                 const IntVec3 GetMipmapTexelSize(TextureUploadTarget target, Uint mipmapLevel) const override;
                 const SizeT GetMipmapByteSize(TextureUploadTarget target, Uint mipmapLevel) const override;
                 void AllocateStorage(TextureUploadTarget uploadTarget, Uint mipmapLevel, MipmapInput input) override;
+                void TruncateMipmapLevels(TextureUploadTarget uploadTarget, Uint levelCount) override;
                 void UpdateMipmapSubData(TextureUploadTarget uploadTarget, Uint mipmapLevel, DataPtr input) override;
                 void* MapMipmapData(TextureUploadTarget uploadTarget, Uint mipmapLevel) override;
                 void MarkStorageDirty(TextureUploadTarget uploadTarget, Uint mipmapLevel, bool dirty) override;

MobileGL/MG_Test/CMakeLists.txt

@@ -72,6 +72,8 @@ add_subdirectory(Texture)
 add_subdirectory(VertexArray)
 add_subdirectory(Program)
 add_subdirectory(Query)
+add_subdirectory(Pipeline)
+add_subdirectory(ShaderTranspiler)
 if (ENABLE_INTEGRATION_TESTS)
     add_subdirectory(Backend/DirectVulkan)
 endif()

MobileGL/MG_Test/Pipeline/CMakeLists.txt

@@ -0,0 +1,27 @@
+cmake_minimum_required(VERSION 3.14)
+
+add_executable(
+        PipelineQuirkTest
+        PipelineQuirkTest.cpp
+)
+
+target_include_directories(PipelineQuirkTest PRIVATE
+        ${MGL_ROOT}/include
+        ${MGL_ROOT}/MobileGL
+        ${MGL_ROOT}/3rdparty/xxHash
+        ${MGL_ROOT}/3rdparty/Vulkan-Headers/include
+        ${MGL_ROOT}/3rdparty/SPIRV-Reflect
+)
+
+target_link_libraries(
+        PipelineQuirkTest PRIVATE
+        GTest::gtest_main
+        ${LINK_LIBRARIES}
+)
+
+if (MSVC)
+        target_compile_options(PipelineQuirkTest PRIVATE /Zc:preprocessor)
+endif()
+
+include(GoogleTest)
+gtest_discover_tests(PipelineQuirkTest DISCOVERY_TIMEOUT 30 PROPERTIES LABELS unit)

MobileGL/MG_Test/Pipeline/PipelineQuirkTest.cpp

@@ -0,0 +1,459 @@
+// MobileGL - MobileGL/MG_Test/Pipeline/PipelineQuirkTest.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include <gtest/gtest.h>
+
+#include <Config.h>
+#include <MG_Backend/DirectVulkan/Renderer/PipelineFactory.h>
+#include <MG_Backend/DirectVulkan/Renderer/ProgramFactory.h>
+
+using namespace MobileGL;
+using MobileGL::MG_Backend::DirectVulkan::PipelineFactory;
+using MobileGL::MG_Backend::DirectVulkan::ProgramFactory;
+using MobileGL::MG_Config::QuirkOverride;
+
+namespace {
+    constexpr Uint32 kVendorIdQualcomm = 0x5143;
+    constexpr Uint32 kVendorIdArm = 0x13B5;
+
+    constexpr VkColorComponentFlags kFullColorWriteMask =
+        VK_COLOR_COMPONENT_R_BIT | VK_COLOR_COMPONENT_G_BIT |
+        VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT;
+
+    // Builds non-separate blend state: the alpha channel repeats the color factors/op, which
+    // is what glBlendFunc/glBlendEquation (as opposed to their *Separate forms) produce.
+    // ShouldSuppressDepthWrite deliberately decides on the color channel alone, so these
+    // cases cover its whole input space; SeparateAlphaAccumulationIsNotStripped below pins
+    // the separate-alpha contract.
+    VkPipelineColorBlendAttachmentState MakeBlendAttachment(Bool blendEnable,
+                                                            VkBlendFactor srcColor,
+                                                            VkBlendFactor dstColor,
+                                                            VkBlendOp colorOp,
+                                                            VkColorComponentFlags colorWriteMask) {
+        VkPipelineColorBlendAttachmentState attachment{};
+        attachment.blendEnable = blendEnable ? VK_TRUE : VK_FALSE;
+        attachment.srcColorBlendFactor = srcColor;
+        attachment.dstColorBlendFactor = dstColor;
+        attachment.colorBlendOp = colorOp;
+        attachment.srcAlphaBlendFactor = srcColor;
+        attachment.dstAlphaBlendFactor = dstColor;
+        attachment.alphaBlendOp = colorOp;
+        attachment.colorWriteMask = colorWriteMask;
+        return attachment;
+    }
+
+    // glslangValidator -V output for:
+    //   #version 450
+    //   layout(location = 0) out vec4 outColor;
+    //   void main() { outColor = vec4(1.0); gl_FragDepth = 0.5; }
+    // Assigning gl_FragDepth makes glslang emit OpExecutionMode ... DepthReplacing.
+    constexpr Uint32 kFragDepthWriterSpirv[] = {
+        0x07230203u, 0x00010000u, 0x0008000bu, 0x0000000fu, 0x00000000u, 0x00020011u,
+        0x00000001u, 0x0006000bu, 0x00000001u, 0x4c534c47u, 0x6474732eu, 0x3035342eu,
+        0x00000000u, 0x0003000eu, 0x00000000u, 0x00000001u, 0x0007000fu, 0x00000004u,
+        0x00000004u, 0x6e69616du, 0x00000000u, 0x00000009u, 0x0000000du, 0x00030010u,
+        0x00000004u, 0x00000007u, 0x00030010u, 0x00000004u, 0x0000000cu, 0x00030003u,
+        0x00000002u, 0x000001c2u, 0x00040005u, 0x00000004u, 0x6e69616du, 0x00000000u,
+        0x00050005u, 0x00000009u, 0x4374756fu, 0x726f6c6fu, 0x00000000u, 0x00060005u,
+        0x0000000du, 0x465f6c67u, 0x44676172u, 0x68747065u, 0x00000000u, 0x00040047u,
+        0x00000009u, 0x0000001eu, 0x00000000u, 0x00040047u, 0x0000000du, 0x0000000bu,
+        0x00000016u, 0x00020013u, 0x00000002u, 0x00030021u, 0x00000003u, 0x00000002u,
+        0x00030016u, 0x00000006u, 0x00000020u, 0x00040017u, 0x00000007u, 0x00000006u,
+        0x00000004u, 0x00040020u, 0x00000008u, 0x00000003u, 0x00000007u, 0x0004003bu,
+        0x00000008u, 0x00000009u, 0x00000003u, 0x0004002bu, 0x00000006u, 0x0000000au,
+        0x3f800000u, 0x0007002cu, 0x00000007u, 0x0000000bu, 0x0000000au, 0x0000000au,
+        0x0000000au, 0x0000000au, 0x00040020u, 0x0000000cu, 0x00000003u, 0x00000006u,
+        0x0004003bu, 0x0000000cu, 0x0000000du, 0x00000003u, 0x0004002bu, 0x00000006u,
+        0x0000000eu, 0x3f000000u, 0x00050036u, 0x00000002u, 0x00000004u, 0x00000000u,
+        0x00000003u, 0x000200f8u, 0x00000005u, 0x0003003eu, 0x00000009u, 0x0000000bu,
+        0x0003003eu, 0x0000000du, 0x0000000eu, 0x000100fdu, 0x00010038u,
+    };
+
+    // Same shader without the gl_FragDepth assignment.
+    constexpr Uint32 kPlainFragmentSpirv[] = {
+        0x07230203u, 0x00010000u, 0x0008000bu, 0x0000000cu, 0x00000000u, 0x00020011u,
+        0x00000001u, 0x0006000bu, 0x00000001u, 0x4c534c47u, 0x6474732eu, 0x3035342eu,
+        0x00000000u, 0x0003000eu, 0x00000000u, 0x00000001u, 0x0006000fu, 0x00000004u,
+        0x00000004u, 0x6e69616du, 0x00000000u, 0x00000009u, 0x00030010u, 0x00000004u,
+        0x00000007u, 0x00030003u, 0x00000002u, 0x000001c2u, 0x00040005u, 0x00000004u,
+        0x6e69616du, 0x00000000u, 0x00050005u, 0x00000009u, 0x4374756fu, 0x726f6c6fu,
+        0x00000000u, 0x00040047u, 0x00000009u, 0x0000001eu, 0x00000000u, 0x00020013u,
+        0x00000002u, 0x00030021u, 0x00000003u, 0x00000002u, 0x00030016u, 0x00000006u,
+        0x00000020u, 0x00040017u, 0x00000007u, 0x00000006u, 0x00000004u, 0x00040020u,
+        0x00000008u, 0x00000003u, 0x00000007u, 0x0004003bu, 0x00000008u, 0x00000009u,
+        0x00000003u, 0x0004002bu, 0x00000006u, 0x0000000au, 0x3f800000u, 0x0007002cu,
+        0x00000007u, 0x0000000bu, 0x0000000au, 0x0000000au, 0x0000000au, 0x0000000au,
+        0x00050036u, 0x00000002u, 0x00000004u, 0x00000000u, 0x00000003u, 0x000200f8u,
+        0x00000005u, 0x0003003eu, 0x00000009u, 0x0000000bu, 0x000100fdu, 0x00010038u,
+    };
+
+
+    // glslangValidator -V output for a vertex shader reading gl_InstanceIndex:
+    //   #version 450
+    //   layout(location = 0) in vec4 inPos;
+    //   void main() { gl_Position = inPos + vec4(float(gl_InstanceIndex)); }
+    constexpr Uint32 kInstanceIndexVertexSpirv[] = {
+        0x07230203u, 0x00010000u, 0x0008000bu, 0x0000001bu, 0x00000000u, 0x00020011u,
+        0x00000001u, 0x0006000bu, 0x00000001u, 0x4c534c47u, 0x6474732eu, 0x3035342eu,
+        0x00000000u, 0x0003000eu, 0x00000000u, 0x00000001u, 0x0008000fu, 0x00000000u,
+        0x00000004u, 0x6e69616du, 0x00000000u, 0x0000000du, 0x00000011u, 0x00000014u,
+        0x00030003u, 0x00000002u, 0x000001c2u, 0x00040005u, 0x00000004u, 0x6e69616du,
+        0x00000000u, 0x00060005u, 0x0000000bu, 0x505f6c67u, 0x65567265u, 0x78657472u,
+        0x00000000u, 0x00060006u, 0x0000000bu, 0x00000000u, 0x505f6c67u, 0x7469736fu,
+        0x006e6f69u, 0x00070006u, 0x0000000bu, 0x00000001u, 0x505f6c67u, 0x746e696fu,
+        0x657a6953u, 0x00000000u, 0x00070006u, 0x0000000bu, 0x00000002u, 0x435f6c67u,
+        0x4470696cu, 0x61747369u, 0x0065636eu, 0x00070006u, 0x0000000bu, 0x00000003u,
+        0x435f6c67u, 0x446c6c75u, 0x61747369u, 0x0065636eu, 0x00030005u, 0x0000000du,
+        0x00000000u, 0x00040005u, 0x00000011u, 0x6f506e69u, 0x00000073u, 0x00070005u,
+        0x00000014u, 0x495f6c67u, 0x6174736eu, 0x4965636eu, 0x7865646eu, 0x00000000u,
+        0x00030047u, 0x0000000bu, 0x00000002u, 0x00050048u, 0x0000000bu, 0x00000000u,
+        0x0000000bu, 0x00000000u, 0x00050048u, 0x0000000bu, 0x00000001u, 0x0000000bu,
+        0x00000001u, 0x00050048u, 0x0000000bu, 0x00000002u, 0x0000000bu, 0x00000003u,
+        0x00050048u, 0x0000000bu, 0x00000003u, 0x0000000bu, 0x00000004u, 0x00040047u,
+        0x00000011u, 0x0000001eu, 0x00000000u, 0x00040047u, 0x00000014u, 0x0000000bu,
+        0x0000002bu, 0x00020013u, 0x00000002u, 0x00030021u, 0x00000003u, 0x00000002u,
+        0x00030016u, 0x00000006u, 0x00000020u, 0x00040017u, 0x00000007u, 0x00000006u,
+        0x00000004u, 0x00040015u, 0x00000008u, 0x00000020u, 0x00000000u, 0x0004002bu,
+        0x00000008u, 0x00000009u, 0x00000001u, 0x0004001cu, 0x0000000au, 0x00000006u,
+        0x00000009u, 0x0006001eu, 0x0000000bu, 0x00000007u, 0x00000006u, 0x0000000au,
+        0x0000000au, 0x00040020u, 0x0000000cu, 0x00000003u, 0x0000000bu, 0x0004003bu,
+        0x0000000cu, 0x0000000du, 0x00000003u, 0x00040015u, 0x0000000eu, 0x00000020u,
+        0x00000001u, 0x0004002bu, 0x0000000eu, 0x0000000fu, 0x00000000u, 0x00040020u,
+        0x00000010u, 0x00000001u, 0x00000007u, 0x0004003bu, 0x00000010u, 0x00000011u,
+        0x00000001u, 0x00040020u, 0x00000013u, 0x00000001u, 0x0000000eu, 0x0004003bu,
+        0x00000013u, 0x00000014u, 0x00000001u, 0x00040020u, 0x00000019u, 0x00000003u,
+        0x00000007u, 0x00050036u, 0x00000002u, 0x00000004u, 0x00000000u, 0x00000003u,
+        0x000200f8u, 0x00000005u, 0x0004003du, 0x00000007u, 0x00000012u, 0x00000011u,
+        0x0004003du, 0x0000000eu, 0x00000015u, 0x00000014u, 0x0004006fu, 0x00000006u,
+        0x00000016u, 0x00000015u, 0x00070050u, 0x00000007u, 0x00000017u, 0x00000016u,
+        0x00000016u, 0x00000016u, 0x00000016u, 0x00050081u, 0x00000007u, 0x00000018u,
+        0x00000012u, 0x00000017u, 0x00050041u, 0x00000019u, 0x0000001au, 0x0000000du,
+        0x0000000fu, 0x0003003eu, 0x0000001au, 0x00000018u, 0x000100fdu, 0x00010038u,
+    };
+
+
+    // Same, but reading gl_VertexIndex instead: a DIFFERENT input builtin. glslang emits
+    // this for GL's gl_VertexID, so nearly every real vertex shader has one - it is what
+    // separates "declares some builtin" from "declares the InstanceIndex builtin".
+    constexpr Uint32 kVertexIndexVertexSpirv[] = {
+        0x07230203u, 0x00010000u, 0x0008000bu, 0x0000001bu, 0x00000000u, 0x00020011u,
+        0x00000001u, 0x0006000bu, 0x00000001u, 0x4c534c47u, 0x6474732eu, 0x3035342eu,
+        0x00000000u, 0x0003000eu, 0x00000000u, 0x00000001u, 0x0008000fu, 0x00000000u,
+        0x00000004u, 0x6e69616du, 0x00000000u, 0x0000000du, 0x00000011u, 0x00000014u,
+        0x00030003u, 0x00000002u, 0x000001c2u, 0x00040005u, 0x00000004u, 0x6e69616du,
+        0x00000000u, 0x00060005u, 0x0000000bu, 0x505f6c67u, 0x65567265u, 0x78657472u,
+        0x00000000u, 0x00060006u, 0x0000000bu, 0x00000000u, 0x505f6c67u, 0x7469736fu,
+        0x006e6f69u, 0x00070006u, 0x0000000bu, 0x00000001u, 0x505f6c67u, 0x746e696fu,
+        0x657a6953u, 0x00000000u, 0x00070006u, 0x0000000bu, 0x00000002u, 0x435f6c67u,
+        0x4470696cu, 0x61747369u, 0x0065636eu, 0x00070006u, 0x0000000bu, 0x00000003u,
+        0x435f6c67u, 0x446c6c75u, 0x61747369u, 0x0065636eu, 0x00030005u, 0x0000000du,
+        0x00000000u, 0x00040005u, 0x00000011u, 0x6f506e69u, 0x00000073u, 0x00060005u,
+        0x00000014u, 0x565f6c67u, 0x65747265u, 0x646e4978u, 0x00007865u, 0x00030047u,
+        0x0000000bu, 0x00000002u, 0x00050048u, 0x0000000bu, 0x00000000u, 0x0000000bu,
+        0x00000000u, 0x00050048u, 0x0000000bu, 0x00000001u, 0x0000000bu, 0x00000001u,
+        0x00050048u, 0x0000000bu, 0x00000002u, 0x0000000bu, 0x00000003u, 0x00050048u,
+        0x0000000bu, 0x00000003u, 0x0000000bu, 0x00000004u, 0x00040047u, 0x00000011u,
+        0x0000001eu, 0x00000000u, 0x00040047u, 0x00000014u, 0x0000000bu, 0x0000002au,
+        0x00020013u, 0x00000002u, 0x00030021u, 0x00000003u, 0x00000002u, 0x00030016u,
+        0x00000006u, 0x00000020u, 0x00040017u, 0x00000007u, 0x00000006u, 0x00000004u,
+        0x00040015u, 0x00000008u, 0x00000020u, 0x00000000u, 0x0004002bu, 0x00000008u,
+        0x00000009u, 0x00000001u, 0x0004001cu, 0x0000000au, 0x00000006u, 0x00000009u,
+        0x0006001eu, 0x0000000bu, 0x00000007u, 0x00000006u, 0x0000000au, 0x0000000au,
+        0x00040020u, 0x0000000cu, 0x00000003u, 0x0000000bu, 0x0004003bu, 0x0000000cu,
+        0x0000000du, 0x00000003u, 0x00040015u, 0x0000000eu, 0x00000020u, 0x00000001u,
+        0x0004002bu, 0x0000000eu, 0x0000000fu, 0x00000000u, 0x00040020u, 0x00000010u,
+        0x00000001u, 0x00000007u, 0x0004003bu, 0x00000010u, 0x00000011u, 0x00000001u,
+        0x00040020u, 0x00000013u, 0x00000001u, 0x0000000eu, 0x0004003bu, 0x00000013u,
+        0x00000014u, 0x00000001u, 0x00040020u, 0x00000019u, 0x00000003u, 0x00000007u,
+        0x00050036u, 0x00000002u, 0x00000004u, 0x00000000u, 0x00000003u, 0x000200f8u,
+        0x00000005u, 0x0004003du, 0x00000007u, 0x00000012u, 0x00000011u, 0x0004003du,
+        0x0000000eu, 0x00000015u, 0x00000014u, 0x0004006fu, 0x00000006u, 0x00000016u,
+        0x00000015u, 0x00070050u, 0x00000007u, 0x00000017u, 0x00000016u, 0x00000016u,
+        0x00000016u, 0x00000016u, 0x00050081u, 0x00000007u, 0x00000018u, 0x00000012u,
+        0x00000017u, 0x00050041u, 0x00000019u, 0x0000001au, 0x0000000du, 0x0000000fu,
+        0x0003003eu, 0x0000001au, 0x00000018u, 0x000100fdu, 0x00010038u,
+    };
+
+    // Owns the reflection module so each test case cleans up after itself.
+    class ReflectModule {
+    public:
+        template <SizeT WordCount>
+        explicit ReflectModule(const Uint32 (&spirv)[WordCount]) {
+            m_created = spvReflectCreateShaderModule(sizeof(spirv), spirv, &m_module) ==
+                        SPV_REFLECT_RESULT_SUCCESS;
+        }
+        ~ReflectModule() {
+            if (m_created) {
+                spvReflectDestroyShaderModule(&m_module);
+            }
+        }
+        ReflectModule(const ReflectModule&) = delete;
+        ReflectModule& operator=(const ReflectModule&) = delete;
+
+        Bool Created() const { return m_created; }
+        const SpvReflectShaderModule& Get() const { return m_module; }
+
+    private:
+        SpvReflectShaderModule m_module{};
+        Bool m_created = false;
+    };
+
+    PipelineFactory::PipelineCreatePayload MakeDepthWritingPayload(
+        const VkPipelineColorBlendAttachmentState& attachment0) {
+        PipelineFactory::PipelineCreatePayload payload{};
+        payload.colorAttachmentCount = 1;
+        payload.depthTestEnable = true;
+        payload.depthWriteEnable = true;
+        payload.colorBlendAttachments[0] = attachment0;
+        return payload;
+    }
+} // namespace
+
+// --- Device gate: MOBILEGL_MAGMA_DISABLE_BLENDED_DEPTH_WRITE tri-state ---
+
+TEST(PipelineQuirkDeviceGate, ForceOnEnablesOnAnyVendor) {
+    EXPECT_TRUE(PipelineFactory::ShouldSuppressBlendedDepthWriteForDevice(QuirkOverride::ForceOn,
+                                                                          kVendorIdArm));
+    EXPECT_TRUE(PipelineFactory::ShouldSuppressBlendedDepthWriteForDevice(QuirkOverride::ForceOn,
+                                                                          kVendorIdQualcomm));
+}
+
+TEST(PipelineQuirkDeviceGate, ForceOffDisablesEvenOnQualcomm) {
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressBlendedDepthWriteForDevice(QuirkOverride::ForceOff,
+                                                                           kVendorIdQualcomm));
+}
+
+TEST(PipelineQuirkDeviceGate, AutoDetectsQualcommOnly) {
+    EXPECT_TRUE(PipelineFactory::ShouldSuppressBlendedDepthWriteForDevice(QuirkOverride::Auto,
+                                                                          kVendorIdQualcomm));
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressBlendedDepthWriteForDevice(QuirkOverride::Auto,
+                                                                           kVendorIdArm));
+}
+
+TEST(PipelineQuirkDeviceGate, ForceOnRoundTripsThroughTheFactoryFlag) {
+    const Bool previous = PipelineFactory::IsSuppressBlendedDepthWriteEnabled();
+    PipelineFactory::SetSuppressBlendedDepthWrite(
+        PipelineFactory::ShouldSuppressBlendedDepthWriteForDevice(QuirkOverride::ForceOn, kVendorIdArm));
+    EXPECT_TRUE(PipelineFactory::IsSuppressBlendedDepthWriteEnabled());
+    PipelineFactory::SetSuppressBlendedDepthWrite(previous);
+}
+
+// --- Per-pipeline strip decision against the pipeline create-info payload ---
+
+TEST(PipelineQuirkStripDecision, MaxBlendIsStripped) {
+    // MC 26.3 OIT depth_bounds: GL_MAX accumulation writing depth - the case the quirk fixes.
+    const auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ZERO, VK_BLEND_OP_MAX, kFullColorWriteMask));
+    EXPECT_TRUE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, MinBlendIsStripped) {
+    const auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ZERO, VK_BLEND_OP_MIN, kFullColorWriteMask));
+    EXPECT_TRUE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, AdditiveOnePlusOneIsNotStripped) {
+    // ONE+ONE additive with a depth write matched zero draws of the 26.3 chain in the
+    // fixture sweep (transmittance/accumulate disable depth writes themselves); the only
+    // real content with this shape was harmless additive glow effects (Create). A quirk
+    // touches as little unrelated content as possible, so the shape stays exempt.
+    const auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ONE, VK_BLEND_OP_ADD, kFullColorWriteMask));
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, SortedTransparencyOverBlendIsNotStripped) {
+    // Vanilla MC translucent layer (water, stained glass): SRC_ALPHA "over" compositing
+    // draws each surface once and depends on its depth writes to occlude particles, rain,
+    // and clouds drawn later - it must keep them.
+    const auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_SRC_ALPHA, VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA, VK_BLEND_OP_ADD,
+        kFullColorWriteMask));
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, EffectivelyOpaqueBlendIsNotStripped) {
+    // GL_BLEND left enabled with ONE/ZERO+ADD factors is opaque in effect; stripping its
+    // depth write would break occlusion for plainly opaque geometry.
+    const auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ZERO, VK_BLEND_OP_ADD, kFullColorWriteMask));
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, FullyMaskedAccumulationBlendIsNotStripped) {
+    // Depth-prepass pattern: colorMask(0,0,0,0) with blending left enabled - blending is
+    // moot, and stripping would delete the entire prepass. MAX so the exemption, not the
+    // blend-op filter, is what keeps the depth write.
+    const auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ONE, VK_BLEND_OP_MAX, 0));
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, DisabledBlendIsNotStripped) {
+    const auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        false, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ONE, VK_BLEND_OP_MAX, kFullColorWriteMask));
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, NoDepthWriteMeansNoStrip) {
+    auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ONE, VK_BLEND_OP_MAX, kFullColorWriteMask));
+    payload.depthWriteEnable = false;
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, FragDepthWriterIsExempt) {
+    // gl_FragDepth output does not go through per-pipeline vertex position math, so the
+    // cross-pipeline invariance hazard cannot affect it (e.g. the 26.3 OIT composite).
+    auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ONE, VK_BLEND_OP_MAX, kFullColorWriteMask));
+    payload.fragmentReplacesDepth = true;
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, AccumulationOnSecondaryAttachmentIsStripped) {
+    // The scan is not limited to attachment 0: an extremum accumulation on any live
+    // attachment marks the pipeline.
+    PipelineFactory::PipelineCreatePayload payload{};
+    payload.colorAttachmentCount = 2;
+    payload.depthTestEnable = true;
+    payload.depthWriteEnable = true;
+    payload.colorBlendAttachments[0] = MakeBlendAttachment(
+        false, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ZERO, VK_BLEND_OP_ADD, kFullColorWriteMask);
+    payload.colorBlendAttachments[1] = MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ONE, VK_BLEND_OP_MAX, kFullColorWriteMask);
+    EXPECT_TRUE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, AlphaWeightedAdditiveIsNotStripped) {
+    // SRC_ALPHA,ONE additive: the classic *sorted* particle/glow blend. Kept exempt like
+    // every other ADD-op shape now that the strip is extremum-only.
+    const auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_SRC_ALPHA, VK_BLEND_FACTOR_ONE, VK_BLEND_OP_ADD, kFullColorWriteMask));
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, ReverseSubtractIsNotStripped) {
+    // Deliberate narrowing: only the MIN/MAX extremum ops carry the depth-bounds
+    // signature. SUBTRACT-class ops stay outside the quirk until content demands them.
+    const auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ONE, VK_BLEND_OP_REVERSE_SUBTRACT,
+        kFullColorWriteMask));
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, PartiallyMaskedAccumulationIsStripped) {
+    // Only a fully masked attachment is exempt; a live alpha channel still accumulates.
+    const auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ONE, VK_BLEND_OP_MAX, VK_COLOR_COMPONENT_A_BIT));
+    EXPECT_TRUE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, NoColorAttachmentsMeansNoStrip) {
+    // Depth-only FBO: the loop must not read the (stale) attachment array at all.
+    PipelineFactory::PipelineCreatePayload payload{};
+    payload.colorAttachmentCount = 0;
+    payload.depthTestEnable = true;
+    payload.depthWriteEnable = true;
+    payload.colorBlendAttachments[0] = MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ONE, VK_BLEND_OP_MAX, kFullColorWriteMask);
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+TEST(PipelineQuirkStripDecision, SeparateAlphaAccumulationIsNotStripped) {
+    // glBlendEquationSeparate(GL_FUNC_ADD, GL_MAX) over an ordinary color over-blend: the
+    // alpha channel accumulates but the color channel does not. Pins that the decision is
+    // color-channel only - widening it to alpha would re-capture sorted transparency.
+    auto attachment = MakeBlendAttachment(true, VK_BLEND_FACTOR_SRC_ALPHA,
+                                          VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA, VK_BLEND_OP_ADD,
+                                          kFullColorWriteMask);
+    attachment.srcAlphaBlendFactor = VK_BLEND_FACTOR_ONE;
+    attachment.dstAlphaBlendFactor = VK_BLEND_FACTOR_ONE;
+    attachment.alphaBlendOp = VK_BLEND_OP_MAX;
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(MakeDepthWritingPayload(attachment)));
+}
+
+TEST(PipelineQuirkStripDecision, MixedOverAndMaskedAttachmentsAreNotStripped) {
+    PipelineFactory::PipelineCreatePayload payload{};
+    payload.colorAttachmentCount = 2;
+    payload.depthTestEnable = true;
+    payload.depthWriteEnable = true;
+    payload.colorBlendAttachments[0] = MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_SRC_ALPHA, VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA, VK_BLEND_OP_ADD,
+        kFullColorWriteMask);
+    payload.colorBlendAttachments[1] = MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ONE, VK_BLEND_OP_MAX, 0);
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+// --- DepthReplacing reflection feeding the gl_FragDepth exemption ---
+
+TEST(ReflectedFragmentReplacesDepth, TrueForAShaderThatAssignsFragDepth) {
+    const ReflectModule module(kFragDepthWriterSpirv);
+    ASSERT_TRUE(module.Created());
+    EXPECT_TRUE(ProgramFactory::ReflectedFragmentReplacesDepth(module.Get()));
+}
+
+TEST(ReflectedFragmentReplacesDepth, FalseForAPlainFragmentShader) {
+    const ReflectModule module(kPlainFragmentSpirv);
+    ASSERT_TRUE(module.Created());
+    EXPECT_FALSE(ProgramFactory::ReflectedFragmentReplacesDepth(module.Get()));
+}
+
+TEST(ReflectedFragmentReplacesDepth, FalseForAnEmptyModule) {
+    // A default-constructed module has no entry points; the scan must not dereference.
+    SpvReflectShaderModule emptyModule{};
+    EXPECT_FALSE(ProgramFactory::ReflectedFragmentReplacesDepth(emptyModule));
+}
+
+TEST(ReflectedFragmentReplacesDepth, ReflectedFlagFlipsTheStripDecision) {
+    // The two fixtures differ only by the gl_FragDepth assignment, so they pin that the
+    // reflected flag is what flips the strip decision for an otherwise identical pipeline.
+    const ReflectModule depthWriter(kFragDepthWriterSpirv);
+    const ReflectModule plain(kPlainFragmentSpirv);
+    ASSERT_TRUE(depthWriter.Created());
+    ASSERT_TRUE(plain.Created());
+
+    auto payload = MakeDepthWritingPayload(MakeBlendAttachment(
+        true, VK_BLEND_FACTOR_ONE, VK_BLEND_FACTOR_ONE, VK_BLEND_OP_MAX, kFullColorWriteMask));
+
+    payload.fragmentReplacesDepth = ProgramFactory::ReflectedFragmentReplacesDepth(plain.Get());
+    EXPECT_TRUE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+
+    payload.fragmentReplacesDepth = ProgramFactory::ReflectedFragmentReplacesDepth(depthWriter.Get());
+    EXPECT_FALSE(PipelineFactory::ShouldSuppressDepthWrite(payload));
+}
+
+
+// --- InstanceIndex reflection feeding the shaderDrawParameters diagnostic ---
+
+TEST(ReflectedReadsInstanceIndexBuiltin, TrueForAShaderReadingInstanceIndex) {
+    const ReflectModule module(kInstanceIndexVertexSpirv);
+    ASSERT_TRUE(module.Created());
+    EXPECT_TRUE(ProgramFactory::ReflectedReadsInstanceIndexBuiltin(module.Get()));
+}
+
+TEST(ReflectedReadsInstanceIndexBuiltin, FalseForAShaderReadingADifferentBuiltin) {
+    // Discriminates the builtin's identity, not merely its presence: weakening the check to
+    // "has any BuiltIn decoration" would fire the diagnostic on every real vertex shader.
+    const ReflectModule module(kVertexIndexVertexSpirv);
+    ASSERT_TRUE(module.Created());
+    EXPECT_FALSE(ProgramFactory::ReflectedReadsInstanceIndexBuiltin(module.Get()));
+}
+
+TEST(ReflectedReadsInstanceIndexBuiltin, FalseForAShaderWithNoInputBuiltins) {
+    const ReflectModule module(kPlainFragmentSpirv);
+    ASSERT_TRUE(module.Created());
+    EXPECT_FALSE(ProgramFactory::ReflectedReadsInstanceIndexBuiltin(module.Get()));
+}
+
+TEST(ReflectedReadsInstanceIndexBuiltin, FalseForAnEmptyModule) {
+    SpvReflectShaderModule emptyModule{};
+    EXPECT_FALSE(ProgramFactory::ReflectedReadsInstanceIndexBuiltin(emptyModule));
+}

MobileGL/MG_Test/Program/ProgramUtilTest.cpp

@@ -238,6 +238,77 @@ void main() {
     }
 }
 
+// KHR-GL33.shaders.preprocessor.* — a block comment is one preprocessing token that the C/GLSL
+// preprocessor replaces with a single space, even when it spans newlines inside a directive. glslang
+// handles this natively, so MobileGL must not mangle it. These reproduce the CTS cases that failed
+// because comment blanking preserved the interior newline, truncating multi-line #define bodies.
+static void ExpectCompiles(MobileGL::ShaderStage stage, GLenum glStage, MobileGL::String source) {
+    using namespace MG_Util::ShaderTranspiler;
+    PreprocessShaderSource(stage, source);
+    ShaderAttrib attrib{.shaderType = glStage, .sourceStr = source};
+    auto res = ShaderCompiler::CompileShader(attrib);
+    if (!res) {
+        FAIL() << "errc: " << res.error().errc << "\nlog: " << res.error().log << "\nsource:\n" << source;
+    }
+}
+
+TEST_F(ProgramUtilTest, PreprocessMultilineCommentInDefineBodyCompiles) {
+    ExpectCompiles(ShaderStage::Fragment, GL_FRAGMENT_SHADER,
+                   R"(#version 330
+precision mediump float;
+out float out0;
+#define VALUE /* current
+            value */ 4.2
+
+void main()
+{
+    out0 = VALUE;
+})");
+}
+
+TEST_F(ProgramUtilTest, PreprocessRedefineObjectMultilineCommentCompiles) {
+    ExpectCompiles(ShaderStage::Fragment, GL_FRAGMENT_SHADER,
+                   R"(#version 330
+precision mediump float;
+out float out0;
+#    define  VAL1 1.0
+#define        VAL2 2.0
+
+#define RES2 /* fdsjklfdsjkl
+                dsfjkhfdsjkh
+                fdsjklhfdsjkh */ (RES1 * VAL2)
+#define RES1    (VAL2 / VAL1)
+#define RES2    /* ewrlkjhsadf */ (RES1 * VAL2)
+#define VALUE    (RES2 + RES1)
+
+void main()
+{
+    out0 = VALUE;
+})");
+}
+
+TEST_F(ProgramUtilTest, PreprocessFunctionMacroRedefinitionMultilineCommentCompiles) {
+    ExpectCompiles(ShaderStage::Fragment, GL_FRAGMENT_SHADER,
+                   R"(#version 330
+precision mediump float;
+out float out0;
+# define FUNC(a,b)        (a  +b)
+# define FUNC(a,b)(a    /* comment
+                         */ +b)
+
+void main()
+{
+    out0 = FUNC(1.0, 2.0);
+})");
+}
+
+// Note: KHR-GL3x.shaders.preprocessor.conditional_inclusion.basic_2 (`#define AAA defined(BBB)` used
+// in `#if !AAA`) is intentionally NOT handled here. Generating the `defined` operator via macro
+// expansion is undefined per the C/GLSL preprocessor spec, and glslang deliberately rejects it
+// ("'defined' : cannot use in preprocessor expression when expanded from macros"). Making it pass
+// would require MobileGL to run its own macro expansion ahead of glslang, which is exactly the
+// preprocessing we defer to glslang; the two cases stay failing by design.
+
 TEST_F(ProgramUtilTest, PreprocessLegacyFragmentShaderModernizesGlmarkStyleSource) {
     using namespace MG_Util::ShaderTranspiler;
 
@@ -426,6 +497,62 @@ void main() {
     verifyVersion("#version 460 core");
 }
 
+// KHR-GL33.shaders.preprocessor.directive.version_* (also re-run verbatim under GL40-GL44): the
+// compiler must REJECT a malformed #version line. MobileGL used to rewrite the whole line to
+// "#version 330 core" whenever it could scrape a leading integer - or treat an unknown profile token
+// as core - which silently legalized every form below. CTS compiles the shader's own #version
+// verbatim, so the rejection has to survive preprocessing (and the 460 retry).
+TEST_F(ProgramUtilTest, PreprocessRejectsMalformedVersionDirectives) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    const char* body = "\nout vec4 fragColor;\nvoid main() { fragColor = vec4(1.0); }\n";
+    const auto rejects = [](const String& fullSource) {
+        String src = fullSource;
+        PreprocessShaderSource(ShaderStage::Fragment, src);
+        ShaderAttrib attrib{.shaderType = GL_FRAGMENT_SHADER, .sourceStr = src};
+        auto res = ShaderCompiler::CompileShader(attrib);
+        return res ? false : true;  // "rejects" == compile failed
+    };
+
+    // Silently legalized today - the five this fix must flip to rejection:
+    EXPECT_TRUE(rejects(String("#version 329") + body))        << "329 is not a real version";
+    EXPECT_TRUE(rejects(String("#version 331") + body))        << "331 is not a real version";
+    EXPECT_TRUE(rejects(String("#version 330 foo") + body))    << "unknown profile keyword";
+    EXPECT_TRUE(rejects(String("#version 330.0") + body))      << "float literal, not an int token";
+    EXPECT_TRUE(rejects(String("#version 330 foobar") + body)) << "trailing tokens after a valid decl";
+
+    // Already rejected (no leading integer, or #version is not the first token) - pinned so a future
+    // change to the normalizer cannot start legalizing them either:
+    EXPECT_TRUE(rejects(String("#version") + body))            << "missing version number";
+    EXPECT_TRUE(rejects(String("#version foobar") + body))     << "identifier where the int belongs";
+    EXPECT_TRUE(rejects(String("#version AAA") + body))        << "identifier where the int belongs";
+    EXPECT_TRUE(rejects(String("precision mediump float;\n#version 330") + body))
+        << "#version must be the first statement";
+    EXPECT_TRUE(rejects(String("#define FOO BAR\n#version 330") + body))
+        << "#version must precede a #define";
+}
+
+// The PASS half of the same CTS group: a valid decl, and #version preceded only by whitespace or a
+// comment, must still compile. Guards the fix above from over-rejecting.
+TEST_F(ProgramUtilTest, PreprocessKeepsValidVersionDirectivesCompiling) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    const char* body = "\nout vec4 fragColor;\nvoid main() { fragColor = vec4(1.0); }\n";
+    const auto compiles = [](const String& fullSource) {
+        String src = fullSource;
+        PreprocessShaderSource(ShaderStage::Fragment, src);
+        ShaderAttrib attrib{.shaderType = GL_FRAGMENT_SHADER, .sourceStr = src};
+        auto res = ShaderCompiler::CompileShader(attrib);
+        return res ? true : false;
+    };
+
+    EXPECT_TRUE(compiles(String("#version 330 core") + body));
+    EXPECT_TRUE(compiles(String("\n#version 330 core") + body))
+        << "leading whitespace is legal before #version";
+    EXPECT_TRUE(compiles(String("// test\n#version 330 core") + body))
+        << "a leading comment is legal before #version";
+}
+
 TEST_F(ProgramUtilTest, PreprocessUsesRealSpacedVersionDirectiveForInjectedOutput) {
     using namespace MG_Util::ShaderTranspiler;
 
@@ -446,6 +573,9 @@ void main() {
     EXPECT_NE(versionPos, String::npos);
     EXPECT_EQ(outputPos, versionPos + std::strlen("#version 330 core\n"));
     EXPECT_NE(source.find("// #version 460 core"), String::npos);
+    // This #line sits ahead of the version directive, where GLSL would never have honoured it, so
+    // it is still dropped. Directives that follow the version line are kept - see
+    // PreprocessKeepsPlainLineDirectivesAndSparesLookalikeIdentifiers.
     EXPECT_EQ(source.find("#line"), String::npos);
 
     ShaderAttrib attrib{.shaderType = GL_FRAGMENT_SHADER, .sourceStr = source};
@@ -455,6 +585,105 @@ void main() {
     }
 }
 
+// A banner line like "//*** NOTE ***" contains "/*" at offset 1 and no "*/" anywhere after it. The
+// old hand-rolled comment stripper searched for "/*" with no lexical state, found that, failed to
+// find a terminator, and erased everything from there to the end of the file - deleting the entire
+// shader. Banner comments in that exact shape are common in Iris and OptiFine packs.
+TEST_F(ProgramUtilTest, PreprocessKeepsShaderBodyAfterAStarredLineComment) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    String source = R"(#version 330 core
+//*** lighting pass ***
+out vec4 fragColor;
+void main() {
+    fragColor = vec4(1.0);
+}
+)";
+    PreprocessShaderSource(ShaderStage::Fragment, source);
+
+    EXPECT_NE(source.find("void main()"), String::npos) << "shader body was truncated:\n" << source;
+    EXPECT_NE(source.find("fragColor = vec4(1.0);"), String::npos);
+
+    ShaderAttrib attrib{.shaderType = GL_FRAGMENT_SHADER, .sourceStr = source};
+    auto res = ShaderCompiler::CompileShader(attrib);
+    if (!res) {
+        FAIL() << "errc: " << res.error().errc << "\nlog: " << res.error().log << "\nsource:\n" << source;
+    }
+}
+
+// The builtin-shadowing rename only fires when the shader really defines its own round/tanh/etc.
+// Deciding that from a commented-out definition renames every genuine call to the builtin to a
+// mg_ name that nothing defines, which fails to link.
+TEST_F(ProgramUtilTest, PreprocessIgnoresCommentedOutBuiltinShadowingDefinition) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    String source = R"(#version 330 core
+// float round(float x) { return floor(x + 0.5); }
+out vec4 fragColor;
+void main() {
+    fragColor = vec4(round(1.25));
+}
+)";
+    PreprocessShaderSource(ShaderStage::Fragment, source);
+
+    EXPECT_NE(source.find("round(1.25)"), String::npos) << "call was renamed from a comment:\n" << source;
+    EXPECT_EQ(source.find("mg_round"), String::npos);
+}
+
+// A block-commented extension directive must not be treated as a real one - the int64 filter turns
+// unsupported directives into #error, so reading one out of a comment manufactures a compile
+// failure for a shader that never asked for the extension.
+TEST_F(ProgramUtilTest, PreprocessIgnoresBlockCommentedExtensionDirectives) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    String source = R"(#version 330 core
+/*
+#extension GL_ARB_gpu_shader_int64 : require
+*/
+out vec4 fragColor;
+void main() {
+    fragColor = vec4(1.0);
+}
+)";
+    PreprocessShaderSource(ShaderStage::Fragment, source);
+
+    EXPECT_EQ(source.find("#error"), String::npos) << "#error synthesized from a comment:\n" << source;
+
+    ShaderAttrib attrib{.shaderType = GL_FRAGMENT_SHADER, .sourceStr = source};
+    auto res = ShaderCompiler::CompileShader(attrib);
+    if (!res) {
+        FAIL() << "errc: " << res.error().errc << "\nlog: " << res.error().log << "\nsource:\n" << source;
+    }
+}
+
+// KHR-GL33.shaders.preprocessor.builtin.line_* checks that __LINE__ follows #line. That only works
+// if the directive reaches glslang, so a plain integer form must pass through untouched - while
+// "#linear" and friends must not be mistaken for it.
+TEST_F(ProgramUtilTest, PreprocessKeepsPlainLineDirectivesAndSparesLookalikeIdentifiers) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    String source = R"(#version 330 core
+out vec4 fragColor;
+#line 42
+float linear(float x) { return x; }
+void main() {
+#line 100
+    fragColor = vec4(linear(float(__LINE__)));
+}
+)";
+    PreprocessShaderSource(ShaderStage::Fragment, source);
+
+    EXPECT_NE(source.find("#line 42"), String::npos) << source;
+    EXPECT_NE(source.find("#line 100"), String::npos) << source;
+    EXPECT_NE(source.find("float linear(float x)"), String::npos) << "identifier lookalike was eaten:\n" << source;
+
+    ShaderAttrib attrib{.shaderType = GL_FRAGMENT_SHADER, .sourceStr = source};
+    auto res = ShaderCompiler::CompileShader(attrib);
+    if (!res) {
+        FAIL() << "errc: " << res.error().errc << "\nlog: " << res.error().log << "\nsource:\n" << source;
+    }
+}
+
 TEST_F(ProgramUtilTest, PreprocessModernSampleQualifierStaysAtVersion460) {
     using namespace MG_Util::ShaderTranspiler;
 
@@ -745,6 +974,16 @@ TEST_F(ProgramUtilTest, RetargetLegacyVersionDirectiveOnlyTouchesNormalizedDeskt
     String commented = "// #version 330 core\nvoid main() {}\n";
     EXPECT_FALSE(RetargetLegacyVersionDirectiveTo460(commented));
     EXPECT_EQ(commented.find("#version 460"), String::npos);
+
+    // A malformed directive must NOT be rescued to 460 - that is what silently legalized the CTS
+    // directive.version_* rejection cases. The bad version stays put so glslang keeps rejecting it.
+    String badNumber = "#version 331\nvoid main() {}\n";
+    EXPECT_FALSE(RetargetLegacyVersionDirectiveTo460(badNumber));
+    EXPECT_EQ(badNumber.find("#version 460"), String::npos);
+
+    String badProfile = "#version 330 foo\nvoid main() {}\n";
+    EXPECT_FALSE(RetargetLegacyVersionDirectiveTo460(badProfile));
+    EXPECT_EQ(badProfile.find("#version 460"), String::npos);
 }
 
 const char* fs = R"(#version 150
@@ -921,6 +1160,353 @@ TEST_F(ProgramUtilTest, CompileFragmentShaderWithDiscard) {
     }
 }
 
+// noperspective is core desktop GLSL (1.30+) and maps to the SPIR-V NoPerspective decoration. It must
+// reach glslang (not be stripped as text) so the SPIR-V carries the decoration; SPIRV-Cross then emits
+// ESSL `noperspective` + the GL_NV_shader_noperspective_interpolation extension. Shader packs
+// (Iris/Complementary) depend on it, and KHR-GL33.glsl_noperspective fails if the result matches
+// smooth. This is the DirectGLES path with the NV extension available (SPIRV-Cross's default).
+TEST_F(ProgramUtilTest, NoperspectiveInterpolationSurvivesToEssl) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    String fs = R"(#version 330 core
+noperspective in vec4 vColor;
+out vec4 fragColor;
+void main() { fragColor = vColor; }
+)";
+    ShaderAttrib attrib{.shaderType = GL_FRAGMENT_SHADER, .sourceStr = fs};
+    auto res = ShaderCompiler::CompileShader(attrib);
+    if (!res) FAIL() << "compile errc: " << res.error().errc << "\nlog: " << res.error().log;
+
+    ProgramAttrib programAttrib{.shaders = {res.value()}};
+    auto program_res = ShaderCompiler::LinkProgram(programAttrib);
+    if (!program_res) FAIL() << "link errc: " << program_res.error().errc << "\nlog: " << program_res.error().log;
+
+    ProgramBinaryAttrib binaryAttrib{.shaderTypes = {GL_FRAGMENT_SHADER}, .program = *program_res.value()};
+    auto bin_res = ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
+    if (!bin_res) FAIL() << "spirv errc: " << bin_res.error().errc << "\nlog: " << bin_res.error().log;
+    ASSERT_EQ(bin_res.value().size(), 1u);
+
+    SpvcSession session(bin_res.value()[0], SessionUsageBit::Transpile);
+    auto essl = ShaderCompiler::DecompileShader(session);
+    if (!essl) FAIL() << "decompile errc: " << essl.error().errc << "\nlog: " << essl.error().log;
+
+    EXPECT_NE(essl.value().find("noperspective"), String::npos)
+        << "noperspective was lost before it reached SPIR-V:\n" << essl.value();
+    EXPECT_NE(essl.value().find("GL_NV_shader_noperspective_interpolation"), String::npos)
+        << "SPIRV-Cross must require the NV extension for ES noperspective:\n" << essl.value();
+}
+
+// The old handling was a naked substring erase of "noperspective", so any identifier that merely
+// contained those characters (a uniform named noperspectiveBlend, say) got mangled. Removing the
+// strip fixes it - glslang, which is identifier-aware, is the only thing that should see the keyword.
+TEST_F(ProgramUtilTest, PreprocessDoesNotCorruptIdentifiersContainingNoperspective) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    String source = R"(#version 330 core
+uniform float noperspectiveBlend;
+out vec4 fragColor;
+void main() { fragColor = vec4(noperspectiveBlend); }
+)";
+    PreprocessShaderSource(ShaderStage::Fragment, source);
+    EXPECT_NE(source.find("noperspectiveBlend"), String::npos)
+        << "identifier was corrupted by substring stripping:\n" << source;
+}
+
+// The DirectGLES fallback for devices without GL_NV_shader_noperspective_interpolation: stripping the
+// NoPerspective decoration makes SPIRV-Cross emit a plain smooth varying with no `#extension … :
+// require`, so the shader still compiles (rendering as smooth) instead of being rejected by the driver.
+TEST_F(ProgramUtilTest, StripNoPerspectiveFallbackProducesPlainEssl) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    String fs = R"(#version 330 core
+noperspective in vec4 vColor;
+out vec4 fragColor;
+void main() { fragColor = vColor; }
+)";
+    ShaderAttrib attrib{.shaderType = GL_FRAGMENT_SHADER, .sourceStr = fs};
+    auto res = ShaderCompiler::CompileShader(attrib);
+    if (!res) FAIL() << "compile errc: " << res.error().errc << "\nlog: " << res.error().log;
+
+    ProgramAttrib programAttrib{.shaders = {res.value()}};
+    auto program_res = ShaderCompiler::LinkProgram(programAttrib);
+    if (!program_res) FAIL() << "link errc: " << program_res.error().errc << "\nlog: " << program_res.error().log;
+
+    ProgramBinaryAttrib binaryAttrib{.shaderTypes = {GL_FRAGMENT_SHADER}, .program = *program_res.value()};
+    auto bin_res = ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
+    if (!bin_res) FAIL() << "spirv errc: " << bin_res.error().errc << "\nlog: " << bin_res.error().log;
+    ASSERT_EQ(bin_res.value().size(), 1u);
+
+    // Precondition: with the decoration present the default decompile requires the NV extension.
+    {
+        SpvcSession session(bin_res.value()[0], SessionUsageBit::Transpile);
+        auto essl = ShaderCompiler::DecompileShader(session);
+        if (!essl) FAIL() << "decompile errc: " << essl.error().errc;
+        ASSERT_NE(essl.value().find("noperspective"), String::npos) << essl.value();
+    }
+
+    // The fallback strips the decoration -> plain smooth ESSL, no extension require.
+    Vector<Uint32> stripped;
+    ASSERT_TRUE(ShaderCompiler::StripNoPerspectiveForEssl(bin_res.value()[0], stripped));
+    ASSERT_FALSE(stripped.empty());
+
+    SpvcSession session(stripped, SessionUsageBit::Transpile);
+    auto essl = ShaderCompiler::DecompileShader(session);
+    if (!essl) FAIL() << "decompile errc: " << essl.error().errc << "\nlog: " << essl.error().log;
+    EXPECT_EQ(essl.value().find("noperspective"), String::npos)
+        << "the decoration should be gone:\n" << essl.value();
+    EXPECT_EQ(essl.value().find("GL_NV_shader_noperspective_interpolation"), String::npos)
+        << "no extension require without the decoration:\n" << essl.value();
+}
+
+// Directly exercises BOTH decoration forms StripNoPerspectivePass handles: a plain-variable
+// OpDecorate NoPerspective (in-operand 1) and an interface-block-member OpMemberDecorate NoPerspective
+// (in-operand 2). The ESSL round-trip tests above use only a scalar input, so they never reach the
+// member-decorate branch, which a block varying like `in Block { noperspective vec4 c; }` (common in
+// shader packs) produces. Unrelated decorations (Flat, Location) must survive untouched.
+TEST_F(ProgramUtilTest, StripNoPerspectivePassRemovesBothDecorateForms) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    const String spirvText = R"(
+               OpCapability Shader
+               OpMemoryModel Logical GLSL450
+               OpEntryPoint Fragment %main "main" %plainVar %blockVar %flatVar
+               OpExecutionMode %main OriginUpperLeft
+               OpName %main "main"
+               OpDecorate %plainVar Location 0
+               OpDecorate %plainVar NoPerspective
+               OpMemberDecorate %Block 0 NoPerspective
+               OpDecorate %blockVar Location 1
+               OpDecorate %flatVar Location 2
+               OpDecorate %flatVar Flat
+       %void = OpTypeVoid
+     %mainFn = OpTypeFunction %void
+      %float = OpTypeFloat 32
+    %v4float = OpTypeVector %float 4
+        %int = OpTypeInt 32 1
+    %inV4Ptr = OpTypePointer Input %v4float
+   %plainVar = OpVariable %inV4Ptr Input
+      %Block = OpTypeStruct %v4float
+ %inBlockPtr = OpTypePointer Input %Block
+   %blockVar = OpVariable %inBlockPtr Input
+   %inIntPtr = OpTypePointer Input %int
+    %flatVar = OpVariable %inIntPtr Input
+       %main = OpFunction %void None %mainFn
+   %mainBody = OpLabel
+               OpReturn
+               OpFunctionEnd
+)";
+
+    spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_1);
+    Vector<uint32_t> inputBinary;
+    ASSERT_TRUE(tools.Assemble(spirvText, &inputBinary));
+
+    const auto countNoPerspective = [](const String& text) {
+        SizeT count = 0, offset = 0;
+        while ((offset = text.find("NoPerspective", offset)) != String::npos) {
+            ++count;
+            offset += std::strlen("NoPerspective");
+        }
+        return count;
+    };
+
+    String inputText;
+    ASSERT_TRUE(tools.Disassemble(inputBinary, &inputText));
+    ASSERT_EQ(countNoPerspective(inputText), 2u)
+        << "fixture must carry both a plain and a member NoPerspective:\n" << inputText;
+
+    Vector<uint32_t> outputBinary;
+    ASSERT_TRUE(ShaderCompiler::StripNoPerspectiveForEssl(inputBinary, outputBinary));
+    ASSERT_FALSE(outputBinary.empty());
+
+    String outputText;
+    ASSERT_TRUE(tools.Disassemble(outputBinary, &outputText));
+    EXPECT_EQ(countNoPerspective(outputText), 0u)
+        << "both NoPerspective decorations (OpDecorate and OpMemberDecorate) must be stripped:\n" << outputText;
+    EXPECT_NE(outputText.find("Flat"), String::npos)
+        << "the unrelated Flat decoration must survive:\n" << outputText;
+    EXPECT_NE(outputText.find("Location"), String::npos)
+        << "Location decorations must survive:\n" << outputText;
+}
+
+// Phase 2 emulation - fragment side. On a device without the NV extension the NoPerspective input is
+// recovered as `load * gl_FragCoord.w` and the decoration removed; gl_FragCoord is synthesized because
+// the shader did not otherwise use it. The emulated SPIR-V must validate and decompile without the
+// extension require.
+TEST_F(ProgramUtilTest, EmulateNoperspectiveFragmentRecoversWithFragCoordW) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    String fs = R"(#version 330 core
+noperspective in vec4 vColor;
+out vec4 f;
+void main() { f = vColor; }
+)";
+    ShaderAttrib attrib{.shaderType = GL_FRAGMENT_SHADER, .sourceStr = fs};
+    auto res = ShaderCompiler::CompileShader(attrib);
+    if (!res) FAIL() << "compile: " << res.error().log;
+    ProgramAttrib pa{.shaders = {res.value()}};
+    auto pr = ShaderCompiler::LinkProgram(pa);
+    if (!pr) FAIL() << "link: " << pr.error().log;
+    ProgramBinaryAttrib ba{.shaderTypes = {GL_FRAGMENT_SHADER}, .program = *pr.value()};
+    auto br = ShaderCompiler::GetSpirvBinaryFromProgram(ba);
+    if (!br) FAIL() << "spirv: " << br.error().log;
+    ASSERT_EQ(br.value().size(), 1u);
+
+    Vector<uint32_t> emulated;
+    ASSERT_TRUE(ShaderCompiler::EmulateNoPerspectiveForEssl(br.value()[0], emulated));
+    ASSERT_FALSE(emulated.empty());
+
+    spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_1);
+    String dis;
+    ASSERT_TRUE(tools.Disassemble(emulated, &dis));
+    ASSERT_TRUE(tools.Validate(emulated)) << "emulated SPIR-V must be valid:\n" << dis;
+    EXPECT_EQ(dis.find("NoPerspective"), String::npos) << "decoration must be stripped:\n" << dis;
+    EXPECT_NE(dis.find("FragCoord"), String::npos) << "gl_FragCoord must be synthesized:\n" << dis;
+    EXPECT_NE(dis.find("OpVectorTimesScalar"), String::npos) << "the recovery multiply must be present:\n" << dis;
+
+    SpvcSession session(emulated, SessionUsageBit::Transpile);
+    auto essl = ShaderCompiler::DecompileShader(session);
+    if (!essl) FAIL() << "decompile: " << essl.error().log;
+    EXPECT_EQ(essl.value().find("noperspective"), String::npos) << essl.value();
+    EXPECT_EQ(essl.value().find("GL_NV_shader_noperspective_interpolation"), String::npos) << essl.value();
+    EXPECT_NE(essl.value().find("gl_FragCoord"), String::npos) << "recovery must reference gl_FragCoord:\n" << essl.value();
+}
+
+// Phase 2 emulation - vertex side. The NoPerspective output is pre-multiplied by gl_Position.w before
+// return and the decoration removed. Emulated SPIR-V must validate and decompile without the extension.
+TEST_F(ProgramUtilTest, EmulateNoperspectiveVertexPreMultipliesByPositionW) {
+    using namespace MG_Util::ShaderTranspiler;
+
+    String vs = R"(#version 330 core
+in vec4 pos;
+noperspective out vec4 vColor;
+void main() { gl_Position = pos; vColor = pos; }
+)";
+    ShaderAttrib attrib{.shaderType = GL_VERTEX_SHADER, .sourceStr = vs};
+    auto res = ShaderCompiler::CompileShader(attrib);
+    if (!res) FAIL() << "compile: " << res.error().log;
+    ProgramAttrib pa{.shaders = {res.value()}};
+    auto pr = ShaderCompiler::LinkProgram(pa);
+    if (!pr) FAIL() << "link: " << pr.error().log;
+    ProgramBinaryAttrib ba{.shaderTypes = {GL_VERTEX_SHADER}, .program = *pr.value()};
+    auto br = ShaderCompiler::GetSpirvBinaryFromProgram(ba);
+    if (!br) FAIL() << "spirv: " << br.error().log;
+    ASSERT_EQ(br.value().size(), 1u);
+
+    Vector<uint32_t> emulated;
+    ASSERT_TRUE(ShaderCompiler::EmulateNoPerspectiveForEssl(br.value()[0], emulated));
+    ASSERT_FALSE(emulated.empty());
+
+    spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_1);
+    String dis;
+    ASSERT_TRUE(tools.Disassemble(emulated, &dis));
+    ASSERT_TRUE(tools.Validate(emulated)) << "emulated SPIR-V must be valid:\n" << dis;
+    EXPECT_EQ(dis.find("NoPerspective"), String::npos) << "decoration must be stripped:\n" << dis;
+    EXPECT_NE(dis.find("OpVectorTimesScalar"), String::npos) << "the pre-multiply must be present:\n" << dis;
+
+    SpvcSession session(emulated, SessionUsageBit::Transpile);
+    auto essl = ShaderCompiler::DecompileShader(session);
+    if (!essl) FAIL() << "decompile: " << essl.error().log;
+    EXPECT_EQ(essl.value().find("noperspective"), String::npos) << essl.value();
+    EXPECT_NE(essl.value().find("gl_Position"), String::npos) << "pre-multiply must reference gl_Position:\n" << essl.value();
+}
+
+namespace {
+// Compiles one shader stage through the full pipeline and returns its SPIR-V, or fails the test.
+MobileGL::Vector<uint32_t> CompileStageSpirv(GLenum type, const char* src) {
+    using namespace MG_Util::ShaderTranspiler;
+    ShaderAttrib attrib{.shaderType = type, .sourceStr = src};
+    auto res = ShaderCompiler::CompileShader(attrib);
+    EXPECT_TRUE(static_cast<bool>(res)) << (res ? "" : res.error().log);
+    if (!res) return {};
+    ProgramAttrib pa{.shaders = {res.value()}};
+    auto pr = ShaderCompiler::LinkProgram(pa);
+    EXPECT_TRUE(static_cast<bool>(pr)) << (pr ? "" : pr.error().log);
+    if (!pr) return {};
+    ProgramBinaryAttrib ba{.shaderTypes = {type}, .program = *pr.value()};
+    auto br = ShaderCompiler::GetSpirvBinaryFromProgram(ba);
+    EXPECT_TRUE(static_cast<bool>(br)) << (br ? "" : br.error().log);
+    if (!br || br.value().empty()) return {};
+    return br.value()[0];
+}
+} // namespace
+
+// Regression: the vertex pre-multiply must be applied exactly once (in main), not once per function.
+// glslang does not inline, so a helper function survives as its own OpFunction; instrumenting its
+// return too would scale the varying by gl_Position.w twice (w^2).
+TEST_F(ProgramUtilTest, EmulateNoperspectiveVertexWithHelperScalesExactlyOnce) {
+    using namespace MG_Util::ShaderTranspiler;
+    // helper() returns via OpReturnValue and adds (no vector*scalar), so the ONLY OpVectorTimesScalar
+    // in the module is the emulation's pre-multiply. The old all-functions code injected it at both
+    // helper's and main's return -> count 2; restricted to the entry function it is 1.
+    auto spirv = CompileStageSpirv(GL_VERTEX_SHADER, R"(#version 330 core
+in vec4 pos;
+noperspective out vec4 vColor;
+vec4 helper(vec4 x) { return x + vec4(1.0); }
+void main() { gl_Position = pos; vColor = helper(pos); }
+)");
+    ASSERT_FALSE(spirv.empty());
+
+    Vector<uint32_t> emulated;
+    ASSERT_TRUE(ShaderCompiler::EmulateNoPerspectiveForEssl(spirv, emulated));
+    spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_1);
+    String dis;
+    ASSERT_TRUE(tools.Disassemble(emulated, &dis));
+    ASSERT_TRUE(tools.Validate(emulated)) << dis;
+
+    SizeT count = 0, off = 0;
+    while ((off = dis.find("OpVectorTimesScalar", off)) != String::npos) {
+        ++count;
+        off += std::strlen("OpVectorTimesScalar");
+    }
+    EXPECT_EQ(count, 1u) << "the gl_Position.w pre-multiply must happen exactly once, not per function:\n" << dis;
+}
+
+// Regression: a single-component read (vColor.x), which glslang lowers via OpAccessChain, must still be
+// recovered with gl_FragCoord.w - not silently left un-scaled.
+TEST_F(ProgramUtilTest, EmulateNoperspectiveFragmentComponentReadIsRecovered) {
+    using namespace MG_Util::ShaderTranspiler;
+    auto spirv = CompileStageSpirv(GL_FRAGMENT_SHADER, R"(#version 330 core
+noperspective in vec4 vColor;
+out vec4 f;
+void main() { f = vec4(vColor.x); }
+)");
+    ASSERT_FALSE(spirv.empty());
+
+    Vector<uint32_t> emulated;
+    ASSERT_TRUE(ShaderCompiler::EmulateNoPerspectiveForEssl(spirv, emulated));
+    spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_1);
+    String dis;
+    ASSERT_TRUE(tools.Disassemble(emulated, &dis));
+    ASSERT_TRUE(tools.Validate(emulated)) << dis;
+    EXPECT_EQ(dis.find("NoPerspective"), String::npos) << dis;
+    EXPECT_NE(dis.find("FragCoord"), String::npos)
+        << "the component read must still be recovered via gl_FragCoord.w:\n" << dis;
+}
+
+// Coverage: a scalar float varying exercises the OpFMul path; a vector varying the OpVectorTimesScalar
+// path; multiple noperspective varyings in one stage are all handled.
+TEST_F(ProgramUtilTest, EmulateNoperspectiveHandlesScalarAndMultipleVaryings) {
+    using namespace MG_Util::ShaderTranspiler;
+    auto spirv = CompileStageSpirv(GL_FRAGMENT_SHADER, R"(#version 330 core
+noperspective in float a;
+noperspective in vec2 b;
+out vec4 f;
+void main() { f = vec4(a, b, 1.0); }
+)");
+    ASSERT_FALSE(spirv.empty());
+
+    Vector<uint32_t> emulated;
+    ASSERT_TRUE(ShaderCompiler::EmulateNoPerspectiveForEssl(spirv, emulated));
+    spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_1);
+    String dis;
+    ASSERT_TRUE(tools.Disassemble(emulated, &dis));
+    ASSERT_TRUE(tools.Validate(emulated)) << dis;
+    EXPECT_EQ(dis.find("NoPerspective"), String::npos) << dis;
+    EXPECT_NE(dis.find("OpFMul"), String::npos) << "the scalar varying must scale with OpFMul:\n" << dis;
+    EXPECT_NE(dis.find("OpVectorTimesScalar"), String::npos)
+        << "the vector varying must scale with OpVectorTimesScalar:\n" << dis;
+}
+
 const char* vs_location = R"(#version 460
 
 in vec4 Position;
@@ -1623,4 +2209,26 @@ TEST_F(ProgramUtilTest, RewriteLinearSubgroupPrefixScanRejectsPartialOrUnsafeTem
     ASSERT_NE(consumerEnd, String::npos);
     nestedScan.insert(consumerEnd + 1, "\n    }");
     expectUnchanged(std::move(nestedScan));
+
+    // ARB/NV spellings of lane-width-sensitive builtins must block the rewrite exactly
+    // like their KHR counterparts.
+    String arbSubgroupBuiltin = MakeLinearSubgroupPrefixScanShader();
+    arbSubgroupBuiltin.insert(arbSubgroupBuiltin.find("float importance"),
+                              "uint arbLane = gl_SubGroupInvocationARB;\n    ");
+    expectUnchanged(std::move(arbSubgroupBuiltin));
+
+    String arbBallotCall = MakeLinearSubgroupPrefixScanShader();
+    arbBallotCall.insert(arbBallotCall.find("float importance"),
+                         "uint64_t arbMask = ballotARB(true);\n    ");
+    expectUnchanged(std::move(arbBallotCall));
+
+    String nvWarpBuiltin = MakeLinearSubgroupPrefixScanShader();
+    nvWarpBuiltin.insert(nvWarpBuiltin.find("float importance"),
+                         "uint warpSize = gl_WarpSizeNV;\n    ");
+    expectUnchanged(std::move(nvWarpBuiltin));
+
+    String nvShuffleCall = MakeLinearSubgroupPrefixScanShader();
+    nvShuffleCall.insert(nvShuffleCall.find("float importance"),
+                         "float other = shuffleNV(1.0f, 0u, 32u);\n    ");
+    expectUnchanged(std::move(nvShuffleCall));
 }

MobileGL/MG_Test/SanityTest.cpp

@@ -33,6 +33,7 @@
 #include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
 #include <MG_Util/ShaderTranspiler/ShaderSourceProcessor.h>
 #include <MG_Util/Debug/Log.h>
+#include <FastSTL/UnorderedMap.h>
 
 namespace {
     class DynamicParameterBackend final : public MobileGL::MG_Backend::BackendObject {
@@ -794,6 +795,33 @@ TEST(DirectVulkanSanity, ReadbackConvertsRgba8AndRgba16fPixels) {
     EXPECT_FLOAT_EQ(rgba16fFloatResult[3], 1.0f);
 }
 
+TEST(DirectVulkanSanity, ReadbackDecodesSingleChannel32BitFormats) {
+    using MobileGL::MG_Backend::DirectVulkan::VulkanRenderer;
+
+    // The reinterpretation feature makes R32F/R32UI-class images common readback sources
+    // (iterationRP custom images). Missing channels take GL defaults: 0 for GB, 1 for alpha.
+    const MobileGL::Float r32f[] = {0.75f, -2.0f};
+    MobileGL::Float r32fResult[8]{};
+    ASSERT_TRUE(VulkanRenderer::ConvertReadbackPixels(
+        reinterpret_cast<const MobileGL::Uint8*>(r32f), VK_FORMAT_R32_SFLOAT,
+        2, 1, GL_RGBA, GL_FLOAT, sizeof(MobileGL::Float) * 8,
+        reinterpret_cast<MobileGL::Uint8*>(r32fResult)));
+    EXPECT_FLOAT_EQ(r32fResult[0], 0.75f);
+    EXPECT_FLOAT_EQ(r32fResult[1], 0.0f);
+    EXPECT_FLOAT_EQ(r32fResult[2], 0.0f);
+    EXPECT_FLOAT_EQ(r32fResult[3], 1.0f);
+    EXPECT_FLOAT_EQ(r32fResult[4], -2.0f);
+
+    const MobileGL::Uint32 r32ui[] = {12345u};
+    MobileGL::Float r32uiResult[4]{};
+    ASSERT_TRUE(VulkanRenderer::ConvertReadbackPixels(
+        reinterpret_cast<const MobileGL::Uint8*>(r32ui), VK_FORMAT_R32_UINT,
+        1, 1, GL_RGBA, GL_FLOAT, sizeof(MobileGL::Float) * 4,
+        reinterpret_cast<MobileGL::Uint8*>(r32uiResult)));
+    EXPECT_FLOAT_EQ(r32uiResult[0], 12345.0f);
+    EXPECT_FLOAT_EQ(r32uiResult[3], 1.0f);
+}
+
 TEST(DirectVulkanSanity, DrawIndexedIndirectCommandMatchesGlAndVulkanLayout) {
     using namespace MobileGL::MG_Backend::DirectVulkan;
 
@@ -999,9 +1027,27 @@ TEST(DirectVulkanSanity, SampledViewFormatMatchesSamplerNumericDomainWithoutChan
     EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
                   VK_FORMAT_B10G11R11_UFLOAT_PACK32, SamplerNumericDomain::UnsignedInteger),
               VK_FORMAT_UNDEFINED);
+
+    // Depth/stencil formats never resolve through color-class reinterpretation; they pass
+    // through unchanged so the existing depth-aspect sampled view is used. Combined
+    // depth-stencil formats are multi-numeric (vkuFormatIsSampledFloat is false for them),
+    // so without the passthrough a plain sampler2D/sampler2DShadow on GL_DEPTH24_STENCIL8
+    // would resolve to UNDEFINED and the draw would be dropped.
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_D24_UNORM_S8_UINT, SamplerNumericDomain::Float),
+              VK_FORMAT_D24_UNORM_S8_UINT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_D32_SFLOAT_S8_UINT, SamplerNumericDomain::Float),
+              VK_FORMAT_D32_SFLOAT_S8_UINT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_D32_SFLOAT, SamplerNumericDomain::Float),
+              VK_FORMAT_D32_SFLOAT);
+    EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
+                  VK_FORMAT_D24_UNORM_S8_UINT, SamplerNumericDomain::UnsignedInteger),
+              VK_FORMAT_D24_UNORM_S8_UINT);
     EXPECT_EQ(VkTextureManager::ResolveSampledImageViewFormat(
                   VK_FORMAT_D32_SFLOAT, SamplerNumericDomain::UnsignedInteger),
-              VK_FORMAT_UNDEFINED);
+              VK_FORMAT_D32_SFLOAT);
 
     EXPECT_TRUE(VkTextureManager::AreSampledImageViewFormatsCompatible(
         VK_FORMAT_R32_SFLOAT, VK_FORMAT_R32_UINT));
@@ -1386,3 +1432,366 @@ TEST(RenderStateSanity, PrimitiveRestartIndexStoresAndReadsBack) {
 
     MG_State::pGLContext.reset();
 }
+
+
+// ---- DirectGLES readback driver-state shadows ----------------------------------------------------
+// Regression coverage for the readback-path state-leak overhaul: the pixel-PBO
+// binding cache, the framebuffer-binding shadow, the PACK pixel-store shadow and
+// the scratch-FBO attachment shadow must (a) leave the driver in the documented
+// resting state, (b) skip redundant GL calls, and (c) scrub correctly on
+// deletion. All drive the real Managers.cpp implementations against a recording
+// mock GLES table.
+namespace {
+    struct StateGuardCallLog {
+        MobileGL::Vector<MobileGL::String> calls;
+
+        MobileGL::SizeT Count(const MobileGL::String& prefix) const {
+            MobileGL::SizeT n = 0;
+            for (const auto& c : calls) {
+                if (c.compare(0, prefix.size(), prefix) == 0) ++n;
+            }
+            return n;
+        }
+    };
+
+    StateGuardCallLog* g_stateGuardLog = nullptr;
+    GLuint g_nextStateGuardFBOId = 201;
+
+    void SG_Log(MobileGL::String entry) {
+        if (g_stateGuardLog) g_stateGuardLog->calls.push_back(MobileGL::Move(entry));
+    }
+    void SG_BindBuffer(GLenum target, GLuint buffer) {
+        SG_Log("BindBuffer:" + std::to_string(target) + ":" + std::to_string(buffer));
+    }
+    void SG_BindFramebuffer(GLenum target, GLuint framebuffer) {
+        SG_Log("BindFramebuffer:" + std::to_string(target) + ":" + std::to_string(framebuffer));
+    }
+    void SG_GetIntegerv(GLenum pname, GLint* data) {
+        SG_Log("GetIntegerv:" + std::to_string(pname));
+        if (data) *data = 0;
+    }
+    void SG_PixelStorei(GLenum pname, GLint param) {
+        SG_Log("PixelStorei:" + std::to_string(pname) + ":" + std::to_string(param));
+    }
+    void SG_GenFramebuffers(GLsizei count, GLuint* framebuffers) {
+        for (GLsizei i = 0; i < count; ++i) framebuffers[i] = g_nextStateGuardFBOId++;
+    }
+    void SG_FramebufferTexture2D(GLenum target, GLenum attachment, GLenum textarget, GLuint texture, GLint level) {
+        SG_Log("FramebufferTexture2D:" + std::to_string(target) + ":" + std::to_string(attachment) + ":" +
+               std::to_string(textarget) + ":" + std::to_string(texture) + ":" + std::to_string(level));
+    }
+    void SG_FramebufferTextureLayer(GLenum target, GLenum attachment, GLuint texture, GLint level, GLint layer) {
+        SG_Log("FramebufferTextureLayer:" + std::to_string(target) + ":" + std::to_string(attachment) + ":" +
+               std::to_string(texture) + ":" + std::to_string(level) + ":" + std::to_string(layer));
+    }
+    void SG_ReadBuffer(GLenum src) {
+        SG_Log("ReadBuffer:" + std::to_string(src));
+    }
+    void SG_DrawBuffers(GLsizei n, const GLenum* bufs) {
+        SG_Log("DrawBuffers:" + std::to_string(n) + ":" + std::to_string(n > 0 && bufs ? bufs[0] : 0));
+    }
+    GLenum SG_NoError() {
+        return GL_NO_ERROR;
+    }
+
+    // Installs the recording table and resets every readback driver-state shadow on
+    // both ends, so these tests cannot bleed into (or inherit from) other tests.
+    struct ScopedStateGuardMocks {
+        ScopedStateGuardMocks(): previousFunctions(MobileGL::MG_Backend::DirectGLES::g_GLESFuncs) {
+            ResetShadows();
+            MobileGL::MG_External::GLESFunctionsTable functions{};
+            functions.glBindBuffer = SG_BindBuffer;
+            functions.glBindFramebuffer = SG_BindFramebuffer;
+            functions.glGetIntegerv = SG_GetIntegerv;
+            functions.glPixelStorei = SG_PixelStorei;
+            functions.glGenFramebuffers = SG_GenFramebuffers;
+            functions.glFramebufferTexture2D = SG_FramebufferTexture2D;
+            functions.glFramebufferTextureLayer = SG_FramebufferTextureLayer;
+            functions.glReadBuffer = SG_ReadBuffer;
+            functions.glDrawBuffers = SG_DrawBuffers;
+            functions.glGetError = SG_NoError;
+            MobileGL::MG_Backend::DirectGLES::SetGLESFuncsTable(functions);
+            g_stateGuardLog = &log;
+        }
+
+        ~ScopedStateGuardMocks() {
+            g_stateGuardLog = nullptr;
+            MobileGL::MG_Backend::DirectGLES::SetGLESFuncsTable(previousFunctions);
+            ResetShadows();
+        }
+
+        ScopedStateGuardMocks(const ScopedStateGuardMocks&) = delete;
+        ScopedStateGuardMocks& operator=(const ScopedStateGuardMocks&) = delete;
+
+        static void ResetShadows() {
+            MobileGL::MG_Backend::DirectGLES::BufferImpl::InvalidatePixelBufferBindingCaches();
+            MobileGL::MG_Backend::DirectGLES::FramebufferImpl::InvalidateFramebufferBindingCache();
+            MobileGL::MG_Backend::DirectGLES::PixelStoreImpl::InvalidatePackStateCache();
+            MobileGL::MG_Backend::DirectGLES::ScratchFBOImpl::OnBackendContextDestroyed();
+        }
+
+        StateGuardCallLog log;
+        MobileGL::MG_External::GLESFunctionsTable previousFunctions;
+    };
+} // namespace
+
+TEST(DirectGLESStateGuards, PixelPackBindingCacheSkipsRedundantBindsAndRestsAtZero) {
+    using namespace MobileGL::MG_Backend::DirectGLES;
+    ScopedStateGuardMocks mocks;
+
+    BufferImpl::BindPixelPackBufferId(5);
+    EXPECT_EQ(mocks.log.Count("BindBuffer:"), 1u);
+    BufferImpl::BindPixelPackBufferId(5); // redundant: must not reach the driver
+    EXPECT_EQ(mocks.log.Count("BindBuffer:"), 1u);
+    BufferImpl::BindPixelPackBufferId(0); // scope exit: resting state
+    EXPECT_EQ(mocks.log.Count("BindBuffer:"), 2u);
+    BufferImpl::BindPixelPackBufferId(0);
+    EXPECT_EQ(mocks.log.Count("BindBuffer:"), 2u);
+
+    // After invalidation (MakeCurrent / context reset) the first bind must reach
+    // the driver again even for the same value.
+    BufferImpl::InvalidatePixelBufferBindingCaches();
+    BufferImpl::BindPixelPackBufferId(0);
+    EXPECT_EQ(mocks.log.Count("BindBuffer:"), 3u);
+}
+
+TEST(DirectGLESStateGuards, FramebufferBindingShadowPinsOnceThenSkips) {
+    using namespace MobileGL::MG_Backend::DirectGLES;
+    ScopedStateGuardMocks mocks;
+
+    // Cold path: one driver query pins the shadow; further reads are free.
+    (void)FramebufferImpl::CurrentFramebufferBinding(MobileGL::FramebufferTarget::Read);
+    EXPECT_EQ(mocks.log.Count("GetIntegerv:"), 1u);
+    (void)FramebufferImpl::CurrentFramebufferBinding(MobileGL::FramebufferTarget::Read);
+    EXPECT_EQ(mocks.log.Count("GetIntegerv:"), 1u);
+
+    FramebufferImpl::BindFramebufferId(GL_READ_FRAMEBUFFER, 7);
+    EXPECT_EQ(mocks.log.Count("BindFramebuffer:"), 1u);
+    FramebufferImpl::BindFramebufferId(GL_READ_FRAMEBUFFER, 7);
+    EXPECT_EQ(mocks.log.Count("BindFramebuffer:"), 1u);
+    // GL_FRAMEBUFFER touches both targets; DRAW is still unknown so it must bind.
+    FramebufferImpl::BindFramebufferId(GL_FRAMEBUFFER, 7);
+    EXPECT_EQ(mocks.log.Count("BindFramebuffer:"), 2u);
+    // Both halves now match: no further calls for either single target.
+    FramebufferImpl::BindFramebufferId(GL_DRAW_FRAMEBUFFER, 7);
+    FramebufferImpl::BindFramebufferId(GL_READ_FRAMEBUFFER, 7);
+    FramebufferImpl::BindFramebufferId(GL_FRAMEBUFFER, 7);
+    EXPECT_EQ(mocks.log.Count("BindFramebuffer:"), 2u);
+    EXPECT_EQ(FramebufferImpl::CurrentFramebufferBinding(MobileGL::FramebufferTarget::Draw), 7u);
+    EXPECT_EQ(mocks.log.Count("GetIntegerv:"), 1u); // shadow answered, no new query
+}
+
+TEST(DirectGLESStateGuards, PackStateShadowAppliesMinimalDeltas) {
+    using namespace MobileGL::MG_Backend::DirectGLES;
+    ScopedStateGuardMocks mocks;
+
+    // First application pins all four parameters.
+    PixelStoreImpl::ApplyPackState(PixelStoreImpl::PackState{4, 0, 0, 0});
+    EXPECT_EQ(mocks.log.Count("PixelStorei:"), 4u);
+    // Identical state: zero driver calls.
+    PixelStoreImpl::ApplyPackState(PixelStoreImpl::PackState{4, 0, 0, 0});
+    EXPECT_EQ(mocks.log.Count("PixelStorei:"), 4u);
+    // One field changed: exactly one driver call.
+    PixelStoreImpl::ApplyPackState(PixelStoreImpl::PackState{1, 0, 0, 0});
+    EXPECT_EQ(mocks.log.Count("PixelStorei:"), 5u);
+
+    const auto current = PixelStoreImpl::CurrentPackState();
+    EXPECT_EQ(current.Alignment, 1);
+    EXPECT_EQ(current.RowLength, 0);
+    EXPECT_EQ(current.SkipRows, 0);
+    EXPECT_EQ(current.SkipPixels, 0);
+}
+
+TEST(DirectGLESStateGuards, ScratchFBODetachesCrossAspectResidue) {
+    using namespace MobileGL::MG_Backend::DirectGLES;
+    ScopedStateGuardMocks mocks;
+
+    auto& fb = ScratchFBOImpl::TempFramebuffer();
+    EXPECT_NE(ScratchFBOImpl::EnsureId(fb), 0u);
+
+    // A depth copy leaves a DEPTH_STENCIL attachment (the pre-fix code never
+    // detached it, wedging every later color readback through this FBO).
+    ScratchFBOImpl::EnsureDepthAttachment2D(fb, GL_DRAW_FRAMEBUFFER, 11, GL_TEXTURE_2D, 0, /*withStencil=*/true);
+    const MobileGL::String dsAttach = "FramebufferTexture2D:" + std::to_string(GL_DRAW_FRAMEBUFFER) + ":" +
+                                      std::to_string(GL_DEPTH_STENCIL_ATTACHMENT);
+    EXPECT_EQ(mocks.log.Count(dsAttach), 1u);
+
+    // The next color use must detach the stale depth-stencil attachment exactly once.
+    mocks.log.calls.clear();
+    ScratchFBOImpl::EnsureColorAttachment2D(fb, GL_READ_FRAMEBUFFER, 22, GL_TEXTURE_2D, 0);
+    const MobileGL::String dsDetach = "FramebufferTexture2D:" + std::to_string(GL_READ_FRAMEBUFFER) + ":" +
+                                      std::to_string(GL_DEPTH_STENCIL_ATTACHMENT) + ":" +
+                                      std::to_string(GL_TEXTURE_2D) + ":0:0";
+    const MobileGL::String colorAttach = "FramebufferTexture2D:" + std::to_string(GL_READ_FRAMEBUFFER) + ":" +
+                                         std::to_string(GL_COLOR_ATTACHMENT0) + ":" +
+                                         std::to_string(GL_TEXTURE_2D) + ":22:0";
+    EXPECT_EQ(mocks.log.Count(dsDetach), 1u);
+    EXPECT_EQ(mocks.log.Count(colorAttach), 1u);
+
+    // Back-to-back identical color use: no driver traffic at all.
+    mocks.log.calls.clear();
+    ScratchFBOImpl::EnsureColorAttachment2D(fb, GL_READ_FRAMEBUFFER, 22, GL_TEXTURE_2D, 0);
+    EXPECT_EQ(mocks.log.Count("FramebufferTexture2D:"), 0u);
+}
+
+TEST(DirectGLESStateGuards, ScratchFBOTextureDeletionForcesFullScrub) {
+    using namespace MobileGL::MG_Backend::DirectGLES;
+    ScopedStateGuardMocks mocks;
+
+    auto& fb = ScratchFBOImpl::TempFramebuffer();
+    ScratchFBOImpl::EnsureId(fb);
+    ScratchFBOImpl::EnsureColorAttachment2D(fb, GL_READ_FRAMEBUFFER, 22, GL_TEXTURE_2D, 0);
+
+    // The attached texture id dies: the shadow can no longer vouch for the FBO
+    // (ES does not auto-detach from unbound FBOs, and the name may be recycled),
+    // so the next use must scrub and re-attach instead of skipping.
+    ScratchFBOImpl::NoteTextureIdDeleted(22);
+    mocks.log.calls.clear();
+    ScratchFBOImpl::EnsureColorAttachment2D(fb, GL_READ_FRAMEBUFFER, 22, GL_TEXTURE_2D, 0);
+    EXPECT_GE(mocks.log.Count("FramebufferTexture2D:"), 2u); // scrub (color + depth) ...
+    const MobileGL::String colorAttach = "FramebufferTexture2D:" + std::to_string(GL_READ_FRAMEBUFFER) + ":" +
+                                         std::to_string(GL_COLOR_ATTACHMENT0) + ":" +
+                                         std::to_string(GL_TEXTURE_2D) + ":22:0";
+    EXPECT_EQ(mocks.log.Count(colorAttach), 1u); // ... then the real re-attach
+}
+
+TEST(DirectGLESStateGuards, ScratchFBOReadDrawBufferStateCached) {
+    using namespace MobileGL::MG_Backend::DirectGLES;
+    ScopedStateGuardMocks mocks;
+
+    auto& fb = ScratchFBOImpl::BlitReadFramebuffer();
+    ScratchFBOImpl::EnsureId(fb);
+
+    // Fresh FBOs default to COLOR_ATTACHMENT0 for both buffers: no call needed.
+    ScratchFBOImpl::EnsureReadBuffer(fb, GL_COLOR_ATTACHMENT0);
+    EXPECT_EQ(mocks.log.Count("ReadBuffer:"), 0u);
+    // Depth blits want GL_NONE; the transition costs one call, repeats are free.
+    ScratchFBOImpl::EnsureReadBuffer(fb, GL_NONE);
+    ScratchFBOImpl::EnsureReadBuffer(fb, GL_NONE);
+    EXPECT_EQ(mocks.log.Count("ReadBuffer:"), 1u);
+    ScratchFBOImpl::EnsureDrawBuffer(fb, GL_NONE);
+    ScratchFBOImpl::EnsureDrawBuffer(fb, GL_NONE);
+    EXPECT_EQ(mocks.log.Count("DrawBuffers:"), 1u);
+}
+
+namespace {
+    MobileGL::Vector<GLuint>* g_deletedTextureIds = nullptr;
+
+    void SG_DeleteTextures(GLsizei count, const GLuint* textures) {
+        if (!g_deletedTextureIds) return;
+        for (GLsizei i = 0; i < count; ++i) g_deletedTextureIds->push_back(textures[i]);
+    }
+
+    // Clears the recording hook even when a gtest assertion unwinds the test body
+    // (a dangling pointer to the dead stack vector would corrupt later tests).
+    struct ScopedDeletedTextureRecording {
+        explicit ScopedDeletedTextureRecording(MobileGL::Vector<GLuint>& sink) { g_deletedTextureIds = &sink; }
+        ~ScopedDeletedTextureRecording() { g_deletedTextureIds = nullptr; }
+        ScopedDeletedTextureRecording(const ScopedDeletedTextureRecording&) = delete;
+        ScopedDeletedTextureRecording& operator=(const ScopedDeletedTextureRecording&) = delete;
+    };
+} // namespace
+
+TEST(DirectGLESBackendTexture, DestructorDeletesIdAndScrubsBindingCache) {
+    using namespace MobileGL::MG_Backend::DirectGLES;
+    ScopedDirectGLESTextureBindings scoped; // installs glGenTextures/glBindTexture mocks + resets caches
+    MobileGL::Vector<GLuint> deleted;
+    ScopedDeletedTextureRecording recording(deleted);
+    auto functions = g_GLESFuncs;
+    functions.glDeleteTextures = SG_DeleteTextures;
+    SetGLESFuncsTable(functions);
+
+    const auto texture2DSlot = static_cast<MobileGL::SizeT>(MobileGL::TextureTarget::Texture2D);
+    GLuint id = 0;
+    {
+        auto backendTexture = MobileGL::MakeShared<TextureImpl::BackendTextureObject>();
+        id = backendTexture->GetBackendTextureId();
+        ASSERT_NE(id, 0u);
+        backendTexture->Bind(GL_TEXTURE_2D, 0);
+        ASSERT_EQ(TextureImpl::g_boundTexturesCache[0][texture2DSlot], backendTexture.get());
+    }
+    // Frontend glDeleteTextures used to leak the backend id forever and leave the
+    // cache pointer dangling (heap-address reuse then false-skips a later Bind).
+    ASSERT_EQ(deleted.size(), 1u);
+    EXPECT_EQ(deleted[0], id);
+    EXPECT_EQ(TextureImpl::g_boundTexturesCache[0][texture2DSlot], nullptr);
+
+    // A wrapper whose context died must NOT delete a foreign (recycled) name.
+    {
+        auto backendTexture = MobileGL::MakeShared<TextureImpl::BackendTextureObject>();
+        ++TextureImpl::g_textureContextGeneration;
+        backendTexture.reset();
+        --TextureImpl::g_textureContextGeneration; // restore for later tests
+        EXPECT_EQ(deleted.size(), 1u);
+    }
+}
+
+TEST(DirectGLESStateGuards, DefaultFramebufferBindGoesThroughShadow) {
+    using namespace MobileGL::MG_Backend::DirectGLES;
+    ScopedStateGuardMocks mocks;
+
+    // The regression this guards against: binding framebuffer 0 raw while the
+    // shadow keeps a user-FBO id makes the next re-bind of that FBO false-skip.
+    FramebufferImpl::BindFramebufferId(GL_DRAW_FRAMEBUFFER, 7);
+    FramebufferImpl::BindFramebufferId(GL_DRAW_FRAMEBUFFER, 0); // default-FBO path must use this API
+    FramebufferImpl::BindFramebufferId(GL_DRAW_FRAMEBUFFER, 7); // must reach the driver again
+    EXPECT_EQ(mocks.log.Count("BindFramebuffer:"), 3u);
+}
+
+// FastSTL::unordered_map::erase(iterator) regression coverage. The open-addressing
+// iterator constructor snaps forward from a tombstoned slot to the successor, so
+// erase must NOT advance the rebuilt iterator again: the old double-advance skipped
+// one live element per erase, and erasing the element in the highest occupied
+// bucket pushed the returned index past bucket_count where it never compared equal
+// to end() again - erase-while-iterating sweeps (pipeline/program cache eviction)
+// then ran off the bucket array and fed garbage handles to vkDestroyPipeline
+// (device crash on first mass eviction during world load).
+TEST(FastSTLSanity, EraseWhileIteratingVisitsEveryElementExactlyOnce) {
+    FastSTL::unordered_map<MobileGL::Uint64, MobileGL::Uint64> map;
+    constexpr MobileGL::Uint64 kCount = 1000;
+    for (MobileGL::Uint64 key = 0; key < kCount; ++key) {
+        map.emplace(key * 0x9e3779b97f4a7c15ull, key);
+    }
+    ASSERT_EQ(map.size(), kCount);
+
+    MobileGL::SizeT visited = 0;
+    for (auto it = map.begin(); it != map.end();) {
+        it = map.erase(it);
+        ++visited;
+        ASSERT_LE(visited, kCount); // old code: runaway past end / skipped entries
+    }
+    EXPECT_EQ(visited, kCount);
+    EXPECT_EQ(map.size(), 0u);
+}
+
+TEST(FastSTLSanity, EraseReturnsTheSuccessorElement) {
+    FastSTL::unordered_map<MobileGL::Uint32, MobileGL::Uint32> map;
+    for (MobileGL::Uint32 key = 1; key <= 64; ++key) {
+        map.emplace(key, key);
+    }
+
+    // Erasing every other visited element must still visit all 64 exactly once:
+    // the iterator returned by erase names the very next element, not one past it.
+    MobileGL::SizeT visited = 0;
+    MobileGL::SizeT erased = 0;
+    for (auto it = map.begin(); it != map.end();) {
+        ++visited;
+        if ((visited & 1) != 0) {
+            it = map.erase(it);
+            ++erased;
+        } else {
+            ++it;
+        }
+        ASSERT_LE(visited, 64u);
+    }
+    EXPECT_EQ(visited, 64u);
+    EXPECT_EQ(map.size(), 64u - erased);
+}
+
+TEST(FastSTLSanity, ErasingTheOnlyElementReturnsEnd) {
+    FastSTL::unordered_map<MobileGL::Uint32, MobileGL::Uint32> map;
+    map.emplace(42u, 1u);
+    auto next = map.erase(map.begin());
+    EXPECT_EQ(next, map.end());
+    EXPECT_TRUE(map.empty());
+}

MobileGL/MG_Test/ShaderTranspiler/CMakeLists.txt

@@ -0,0 +1,25 @@
+cmake_minimum_required(VERSION 3.14)
+
+add_executable(
+        SpirvPassTest
+        SpirvPassTest.cpp
+)
+
+target_include_directories(SpirvPassTest PRIVATE
+        ${MGL_ROOT}/include
+        ${MGL_ROOT}/MobileGL
+        ${MGL_ROOT}/3rdparty/SPIRV-Reflect
+)
+
+target_link_libraries(
+        SpirvPassTest PRIVATE
+        GTest::gtest_main
+        ${LINK_LIBRARIES}
+)
+
+if (MSVC)
+        target_compile_options(SpirvPassTest PRIVATE /Zc:preprocessor)
+endif()
+
+include(GoogleTest)
+gtest_discover_tests(SpirvPassTest DISCOVERY_TIMEOUT 30 PROPERTIES LABELS unit)

MobileGL/MG_Test/ShaderTranspiler/SpirvPassTest.cpp

@@ -0,0 +1,170 @@
+// MobileGL - MobileGL/MG_Test/ShaderTranspiler/SpirvPassTest.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include <gtest/gtest.h>
+
+#include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
+
+#include <spirv_reflect.h>
+
+using namespace MobileGL;
+using MobileGL::MG_Util::ShaderTranspiler::ShaderCompiler;
+
+namespace {
+    // glslangValidator -V output. Both are vertex shaders writing gl_Position through
+    // the gl_PerVertex block, i.e. the Position builtin arrives as OpMemberDecorate rather
+    // than a plain OpDecorate - the shape real glslang output actually takes.
+
+    //   #version 450
+    //   layout(location = 0) in vec4 inPos;
+    //   void main() { gl_Position = inPos; }
+    constexpr Uint32 kPlainVertexSpirv[] = {
+        0x07230203u, 0x00010000u, 0x0008000bu, 0x00000015u, 0x00000000u, 0x00020011u,
+        0x00000001u, 0x0006000bu, 0x00000001u, 0x4c534c47u, 0x6474732eu, 0x3035342eu,
+        0x00000000u, 0x0003000eu, 0x00000000u, 0x00000001u, 0x0007000fu, 0x00000000u,
+        0x00000004u, 0x6e69616du, 0x00000000u, 0x0000000du, 0x00000011u, 0x00030003u,
+        0x00000002u, 0x000001c2u, 0x00040005u, 0x00000004u, 0x6e69616du, 0x00000000u,
+        0x00060005u, 0x0000000bu, 0x505f6c67u, 0x65567265u, 0x78657472u, 0x00000000u,
+        0x00060006u, 0x0000000bu, 0x00000000u, 0x505f6c67u, 0x7469736fu, 0x006e6f69u,
+        0x00070006u, 0x0000000bu, 0x00000001u, 0x505f6c67u, 0x746e696fu, 0x657a6953u,
+        0x00000000u, 0x00070006u, 0x0000000bu, 0x00000002u, 0x435f6c67u, 0x4470696cu,
+        0x61747369u, 0x0065636eu, 0x00070006u, 0x0000000bu, 0x00000003u, 0x435f6c67u,
+        0x446c6c75u, 0x61747369u, 0x0065636eu, 0x00030005u, 0x0000000du, 0x00000000u,
+        0x00040005u, 0x00000011u, 0x6f506e69u, 0x00000073u, 0x00030047u, 0x0000000bu,
+        0x00000002u, 0x00050048u, 0x0000000bu, 0x00000000u, 0x0000000bu, 0x00000000u,
+        0x00050048u, 0x0000000bu, 0x00000001u, 0x0000000bu, 0x00000001u, 0x00050048u,
+        0x0000000bu, 0x00000002u, 0x0000000bu, 0x00000003u, 0x00050048u, 0x0000000bu,
+        0x00000003u, 0x0000000bu, 0x00000004u, 0x00040047u, 0x00000011u, 0x0000001eu,
+        0x00000000u, 0x00020013u, 0x00000002u, 0x00030021u, 0x00000003u, 0x00000002u,
+        0x00030016u, 0x00000006u, 0x00000020u, 0x00040017u, 0x00000007u, 0x00000006u,
+        0x00000004u, 0x00040015u, 0x00000008u, 0x00000020u, 0x00000000u, 0x0004002bu,
+        0x00000008u, 0x00000009u, 0x00000001u, 0x0004001cu, 0x0000000au, 0x00000006u,
+        0x00000009u, 0x0006001eu, 0x0000000bu, 0x00000007u, 0x00000006u, 0x0000000au,
+        0x0000000au, 0x00040020u, 0x0000000cu, 0x00000003u, 0x0000000bu, 0x0004003bu,
+        0x0000000cu, 0x0000000du, 0x00000003u, 0x00040015u, 0x0000000eu, 0x00000020u,
+        0x00000001u, 0x0004002bu, 0x0000000eu, 0x0000000fu, 0x00000000u, 0x00040020u,
+        0x00000010u, 0x00000001u, 0x00000007u, 0x0004003bu, 0x00000010u, 0x00000011u,
+        0x00000001u, 0x00040020u, 0x00000013u, 0x00000003u, 0x00000007u, 0x00050036u,
+        0x00000002u, 0x00000004u, 0x00000000u, 0x00000003u, 0x000200f8u, 0x00000005u,
+        0x0004003du, 0x00000007u, 0x00000012u, 0x00000011u, 0x00050041u, 0x00000013u,
+        0x00000014u, 0x0000000du, 0x0000000fu, 0x0003003eu, 0x00000014u, 0x00000012u,
+        0x000100fdu, 0x00010038u,
+    };
+
+    //   ... plus `invariant gl_Position;` - already carries OpMemberDecorate %gl_PerVertex 0
+    //   Invariant, so the pass must not add a duplicate.
+    constexpr Uint32 kAlreadyInvariantVertexSpirv[] = {
+        0x07230203u, 0x00010000u, 0x0008000bu, 0x00000015u, 0x00000000u, 0x00020011u,
+        0x00000001u, 0x0006000bu, 0x00000001u, 0x4c534c47u, 0x6474732eu, 0x3035342eu,
+        0x00000000u, 0x0003000eu, 0x00000000u, 0x00000001u, 0x0007000fu, 0x00000000u,
+        0x00000004u, 0x6e69616du, 0x00000000u, 0x0000000du, 0x00000011u, 0x00030003u,
+        0x00000002u, 0x000001c2u, 0x00040005u, 0x00000004u, 0x6e69616du, 0x00000000u,
+        0x00060005u, 0x0000000bu, 0x505f6c67u, 0x65567265u, 0x78657472u, 0x00000000u,
+        0x00060006u, 0x0000000bu, 0x00000000u, 0x505f6c67u, 0x7469736fu, 0x006e6f69u,
+        0x00070006u, 0x0000000bu, 0x00000001u, 0x505f6c67u, 0x746e696fu, 0x657a6953u,
+        0x00000000u, 0x00070006u, 0x0000000bu, 0x00000002u, 0x435f6c67u, 0x4470696cu,
+        0x61747369u, 0x0065636eu, 0x00070006u, 0x0000000bu, 0x00000003u, 0x435f6c67u,
+        0x446c6c75u, 0x61747369u, 0x0065636eu, 0x00030005u, 0x0000000du, 0x00000000u,
+        0x00040005u, 0x00000011u, 0x6f506e69u, 0x00000073u, 0x00030047u, 0x0000000bu,
+        0x00000002u, 0x00050048u, 0x0000000bu, 0x00000000u, 0x0000000bu, 0x00000000u,
+        0x00040048u, 0x0000000bu, 0x00000000u, 0x00000012u, 0x00050048u, 0x0000000bu,
+        0x00000001u, 0x0000000bu, 0x00000001u, 0x00050048u, 0x0000000bu, 0x00000002u,
+        0x0000000bu, 0x00000003u, 0x00050048u, 0x0000000bu, 0x00000003u, 0x0000000bu,
+        0x00000004u, 0x00040047u, 0x00000011u, 0x0000001eu, 0x00000000u, 0x00020013u,
+        0x00000002u, 0x00030021u, 0x00000003u, 0x00000002u, 0x00030016u, 0x00000006u,
+        0x00000020u, 0x00040017u, 0x00000007u, 0x00000006u, 0x00000004u, 0x00040015u,
+        0x00000008u, 0x00000020u, 0x00000000u, 0x0004002bu, 0x00000008u, 0x00000009u,
+        0x00000001u, 0x0004001cu, 0x0000000au, 0x00000006u, 0x00000009u, 0x0006001eu,
+        0x0000000bu, 0x00000007u, 0x00000006u, 0x0000000au, 0x0000000au, 0x00040020u,
+        0x0000000cu, 0x00000003u, 0x0000000bu, 0x0004003bu, 0x0000000cu, 0x0000000du,
+        0x00000003u, 0x00040015u, 0x0000000eu, 0x00000020u, 0x00000001u, 0x0004002bu,
+        0x0000000eu, 0x0000000fu, 0x00000000u, 0x00040020u, 0x00000010u, 0x00000001u,
+        0x00000007u, 0x0004003bu, 0x00000010u, 0x00000011u, 0x00000001u, 0x00040020u,
+        0x00000013u, 0x00000003u, 0x00000007u, 0x00050036u, 0x00000002u, 0x00000004u,
+        0x00000000u, 0x00000003u, 0x000200f8u, 0x00000005u, 0x0004003du, 0x00000007u,
+        0x00000012u, 0x00000011u, 0x00050041u, 0x00000013u, 0x00000014u, 0x0000000du,
+        0x0000000fu, 0x0003003eu, 0x00000014u, 0x00000012u, 0x000100fdu, 0x00010038u,
+    };
+
+    // OpMemberDecorate <struct-id> <member> <decoration>
+    constexpr Uint32 kOpMemberDecorate = 72;
+    constexpr Uint32 kDecorationInvariant = 18;
+    constexpr Uint32 kSpirvHeaderWordCount = 5;
+
+    // Test-side reference walker. Deliberately independent of the production code so a bug in
+    // the pass cannot hide behind the same helper; only used to count what the pass emitted.
+    Uint32 CountInvariantMemberDecorations(const Vector<Uint32>& spirv) {
+        Uint32 count = 0;
+        for (SizeT i = kSpirvHeaderWordCount; i < spirv.size();) {
+            const Uint32 wordCount = spirv[i] >> 16;
+            const Uint32 opcode = spirv[i] & 0xFFFFu;
+            if (wordCount == 0 || i + wordCount > spirv.size()) {
+                break;
+            }
+            if (opcode == kOpMemberDecorate && wordCount >= 4 && spirv[i + 3] == kDecorationInvariant) {
+                ++count;
+            }
+            i += wordCount;
+        }
+        return count;
+    }
+
+    template <SizeT WordCount>
+    Vector<Uint32> ToVector(const Uint32 (&words)[WordCount]) {
+        return Vector<Uint32>(words, words + WordCount);
+    }
+} // namespace
+
+// --- DecoratePositionInvariantPass ---
+
+TEST(DecoratePositionInvariant, AddsInvariantToThePositionMember) {
+    const Vector<Uint32> input = ToVector(kPlainVertexSpirv);
+    ASSERT_EQ(CountInvariantMemberDecorations(input), 0u);
+
+    Vector<Uint32> output;
+    ASSERT_TRUE(ShaderCompiler::DecoratePositionInvariantForVulkan(input, output));
+    EXPECT_EQ(CountInvariantMemberDecorations(output), 1u);
+}
+
+TEST(DecoratePositionInvariant, DoesNotDuplicateAnExistingInvariant) {
+    const Vector<Uint32> input = ToVector(kAlreadyInvariantVertexSpirv);
+    ASSERT_EQ(CountInvariantMemberDecorations(input), 1u);
+
+    Vector<Uint32> output;
+    ASSERT_TRUE(ShaderCompiler::DecoratePositionInvariantForVulkan(input, output));
+    EXPECT_EQ(CountInvariantMemberDecorations(output), 1u);
+    // The pass reports SuccessWithoutChange here, and SPIRV-Tools asserts (in assert-enabled
+    // builds) that such a run round-trips byte-identically. Pin that from the outside so an
+    // assert-enabled CI build cannot be the first thing to discover a violation.
+    EXPECT_EQ(output, input);
+}
+
+TEST(DecoratePositionInvariant, IsIdempotent) {
+    Vector<Uint32> once;
+    ASSERT_TRUE(ShaderCompiler::DecoratePositionInvariantForVulkan(ToVector(kPlainVertexSpirv), once));
+    Vector<Uint32> twice;
+    ASSERT_TRUE(ShaderCompiler::DecoratePositionInvariantForVulkan(once, twice));
+    EXPECT_EQ(CountInvariantMemberDecorations(twice), 1u);
+}
+
+TEST(DecoratePositionInvariant, OutputStaysAReflectableModule) {
+    Vector<Uint32> output;
+    ASSERT_TRUE(ShaderCompiler::DecoratePositionInvariantForVulkan(ToVector(kPlainVertexSpirv), output));
+
+    SpvReflectShaderModule module{};
+    ASSERT_EQ(spvReflectCreateShaderModule(output.size() * sizeof(Uint32), output.data(), &module),
+              SPV_REFLECT_RESULT_SUCCESS);
+    EXPECT_EQ(module.entry_point_count, 1u);
+    spvReflectDestroyShaderModule(&module);
+}
+
+TEST(DecoratePositionInvariant, RejectsGarbageInput) {
+    const Vector<Uint32> notSpirv{0xdeadbeefu, 0u, 0u, 0u, 0u};
+    Vector<Uint32> output;
+    EXPECT_FALSE(ShaderCompiler::DecoratePositionInvariantForVulkan(notSpirv, output));
+}

MobileGL/MG_Test/Texture/TextureTest.cpp

@@ -318,6 +318,49 @@ TEST_F(TextureTest, CopyTextureSubImage2DUsesNamedObjectAndRestoresBinding) {
     EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
 }
 
+TEST_F(TextureTest, CopyTextureSubImage2DRejectsCubeMapTargets) {
+    const ScopedTextureBackendFunctionsOverride backendGuard;
+    MG_Backend::gBackendFunctionsTable.GL.CopyTexSubImage2D = RecordCopyTexSubImage2D;
+    g_copyTexSubImage2DCall = {};
+
+    GLuint cubeTexture = 0;
+    MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_CUBE_MAP, 1, &cubeTexture);
+    MG_Impl::GLImpl::CopyTextureSubImage2D(cubeTexture, 0, 0, 0, 0, 0, 1, 1);
+
+    // GL 4.6 sec. 8.8: the 2D form only accepts 2D/1D-array/rectangle effective targets.
+    EXPECT_FALSE(g_copyTexSubImage2DCall.Called);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), static_cast<GLenum>(GL_INVALID_OPERATION));
+}
+
+TEST_F(TextureTest, ClearTexImageErrorContracts) {
+    GLuint texture = 0;
+    MG_Impl::GLImpl::GenTextures(1, &texture);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, texture);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 2, 2, 0,
+                                GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+
+    // Zero texture name is INVALID_OPERATION (ARB_clear_texture).
+    MG_Impl::GLImpl::ClearTexImage(0, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), static_cast<GLenum>(GL_INVALID_OPERATION));
+
+    // A negative level is INVALID_VALUE...
+    MG_Impl::GLImpl::ClearTexImage(texture, -1, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), static_cast<GLenum>(GL_INVALID_VALUE));
+
+    // ...but clearing a level that was never defined is INVALID_OPERATION.
+    MG_Impl::GLImpl::ClearTexImage(texture, 5, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), static_cast<GLenum>(GL_INVALID_OPERATION));
+
+    // A clear region outside the level is INVALID_VALUE.
+    MG_Impl::GLImpl::ClearTexSubImage(texture, 0, 1, 1, 0, 4, 4, 1,
+                                      GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), static_cast<GLenum>(GL_INVALID_VALUE));
+
+    // An invalid pixel-transfer format is INVALID_ENUM from the shared validators.
+    MG_Impl::GLImpl::ClearTexImage(texture, 0, GL_NONE, GL_UNSIGNED_BYTE, nullptr);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), static_cast<GLenum>(GL_INVALID_ENUM));
+}
+
 // GL_MAX_TEXTURE_MAX_ANISOTROPY_EXT is float state that must answer every numeric query: GetFloatv
 // is authoritative and GetIntegerv would otherwise fall through to its INVALID_ENUM default.
 TEST_F(TextureTest, MaxTextureMaxAnisotropyIsAnsweredFromTheBackendLimit) {
@@ -1384,6 +1427,158 @@ TEST_F(TextureTest, TextureStorage1DAndSubImageModifyNamedObjectOnly) {
     EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
 }
 
+// Building a mip chain top-down - upload level N, then level 0 - must not destroy the levels
+// already uploaded. AllocateLevel used to resize() the storage down to level+1 on every call, so
+// the level-0 upload truncated the chain to a single level; the higher level then read back as
+// {0,0,0}, IsComplete() rejected the zero-then-nonzero pattern, and DirectGLES answered that by
+// skipping the texture's sync entirely. This is the shape KHR-GL33.texture_repeat_mode uses, and
+// it accounted for 108 CTS failures in every GL version.
+TEST_F(TextureTest, TexImage2DOnLevelZeroKeepsAnAlreadyUploadedHigherLevel) {
+    GLuint texture = 0;
+    MG_Impl::GLImpl::GenTextures(1, &texture);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, texture);
+
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 1, GL_RGBA8, 49, 23, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 98, 46, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+
+    const auto textureObject = MG_State::pGLContext->GetTextureObject(texture);
+    auto* mipmapObject = static_cast<MG_State::GLState::TextureObjectMipmap*>(textureObject.get());
+    ASSERT_NE(mipmapObject, nullptr);
+    EXPECT_EQ(mipmapObject->GetMipmapLevelCount(), 2u);
+    EXPECT_EQ(mipmapObject->GetMipmapTexelSize(TextureUploadTarget::Texture2D, 0), IntVec3(98, 46, 1));
+    EXPECT_EQ(mipmapObject->GetMipmapTexelSize(TextureUploadTarget::Texture2D, 1), IntVec3(49, 23, 1));
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
+// The other half of the contract: respecifying a level 0 that already held an image still drops
+// the chain, exactly as before. Minecraft rebinds the block-atlas name and calls glTexImage2D on
+// level 0 before uploading the new levels; leaving the previous chain in place would strand a tail
+// at the wrong sizes and - because Mojang terminates its chains with a 0x0 level - reproduce the
+// same incomplete-texture black atlas the fix above exists to prevent.
+TEST_F(TextureTest, TexImage2DRespecifyingAnExistingLevelZeroDropsTheStaleChain) {
+    GLuint texture = 0;
+    MG_Impl::GLImpl::GenTextures(1, &texture);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, texture);
+
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 8, 8, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 1, GL_RGBA8, 4, 4, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 2, GL_RGBA8, 2, 2, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+
+    const auto textureObject = MG_State::pGLContext->GetTextureObject(texture);
+    auto* mipmapObject = static_cast<MG_State::GLState::TextureObjectMipmap*>(textureObject.get());
+    ASSERT_NE(mipmapObject, nullptr);
+    ASSERT_EQ(mipmapObject->GetMipmapLevelCount(), 3u);
+
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 16, 16, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+
+    EXPECT_EQ(mipmapObject->GetMipmapLevelCount(), 1u);
+    EXPECT_EQ(mipmapObject->GetMipmapTexelSize(TextureUploadTarget::Texture2D, 0), IntVec3(16, 16, 1));
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
+// Same-size respecification has to drop the chain too. The Mipmap Levels video setting rebuilds
+// the atlas at identical dimensions with a different level count, so a size-change-only test would
+// let the old tail survive.
+TEST_F(TextureTest, TexImage2DRespecifyingLevelZeroAtTheSameSizeStillDropsTheChain) {
+    GLuint texture = 0;
+    MG_Impl::GLImpl::GenTextures(1, &texture);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, texture);
+
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 8, 8, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 1, GL_RGBA8, 4, 4, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 8, 8, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+
+    const auto textureObject = MG_State::pGLContext->GetTextureObject(texture);
+    auto* mipmapObject = static_cast<MG_State::GLState::TextureObjectMipmap*>(textureObject.get());
+    ASSERT_NE(mipmapObject, nullptr);
+    EXPECT_EQ(mipmapObject->GetMipmapLevelCount(), 1u);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
+// glTexStorage2D defines exactly `levels` levels. AllocateStorage only grows now, so the immutable
+// path has to drop a longer pre-existing chain explicitly.
+TEST_F(TextureTest, TexStorage2DTrimsALongerPreExistingMipChain) {
+    GLuint texture = 0;
+    MG_Impl::GLImpl::GenTextures(1, &texture);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, texture);
+
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, 8, 8, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 1, GL_RGBA8, 4, 4, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 2, GL_RGBA8, 2, 2, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 3, GL_RGBA8, 1, 1, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+
+    MG_Impl::GLImpl::TexStorage2D(GL_TEXTURE_2D, 2, GL_RGBA8, 8, 8);
+
+    const auto textureObject = MG_State::pGLContext->GetTextureObject(texture);
+    auto* mipmapObject = static_cast<MG_State::GLState::TextureObjectMipmap*>(textureObject.get());
+    ASSERT_NE(mipmapObject, nullptr);
+    EXPECT_EQ(mipmapObject->GetMipmapLevelCount(), 2u);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
+// glTexImage2D used to reject every GL_COMPRESSED_* internal format with GL_INVALID_ENUM, because
+// none of them mapped to a TextureInternalFormat and the "unknown format" gate fired. They now
+// resolve to the uncompressed storage that backs them - what GL prescribes for the generic formats,
+// and a deliberate deviation for RGTC, which ES cannot compress. The (format, type) pairs below are
+// the ones KHR-GL33.packed_pixels uploads with, so this table doubles as a pin for those 480 cases.
+TEST_F(TextureTest, CompressedInternalFormatsResolveToTheirUncompressedStorage) {
+    struct Case {
+        GLenum internalFormat;
+        GLenum format;
+        GLenum type;
+        TextureInternalFormat expected;
+    };
+    const Case cases[] = {
+        {GL_COMPRESSED_RED, GL_RED, GL_UNSIGNED_BYTE, TextureInternalFormat::R8},
+        {GL_COMPRESSED_RG, GL_RG, GL_UNSIGNED_BYTE, TextureInternalFormat::RG8},
+        {GL_COMPRESSED_RGB, GL_RGB, GL_UNSIGNED_BYTE, TextureInternalFormat::RGB8},
+        {GL_COMPRESSED_RGBA, GL_RGBA, GL_UNSIGNED_BYTE, TextureInternalFormat::RGBA8},
+        {GL_COMPRESSED_SRGB, GL_RGB, GL_UNSIGNED_BYTE, TextureInternalFormat::SRGB8},
+        {GL_COMPRESSED_SRGB_ALPHA, GL_RGBA, GL_UNSIGNED_BYTE, TextureInternalFormat::SRGB8Alpha8},
+        {GL_COMPRESSED_RED_RGTC1, GL_RED, GL_UNSIGNED_BYTE, TextureInternalFormat::R8},
+        {GL_COMPRESSED_RG_RGTC2, GL_RG, GL_UNSIGNED_BYTE, TextureInternalFormat::RG8},
+        // The signed RGTC pair is uploaded as GL_BYTE and must land on SNORM storage - resolving
+        // them to plain R8/RG8 would silently reinterpret negative texels.
+        {GL_COMPRESSED_SIGNED_RED_RGTC1, GL_RED, GL_BYTE, TextureInternalFormat::R8Snorm},
+        {GL_COMPRESSED_SIGNED_RG_RGTC2, GL_RG, GL_BYTE, TextureInternalFormat::RG8Snorm},
+    };
+
+    MG_Impl::GLImpl::PixelStorei(GL_UNPACK_ALIGNMENT, 1);
+    for (const auto& c : cases) {
+        GLuint texture = 0;
+        MG_Impl::GLImpl::GenTextures(1, &texture);
+        MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, texture);
+        MG_Impl::GLImpl::TexImage2D(GL_TEXTURE_2D, 0, c.internalFormat, 4, 4, 0, c.format, c.type, nullptr);
+
+        const auto textureObject = MG_State::pGLContext->GetTextureObject(texture);
+        ASSERT_NE(textureObject, nullptr) << "internalFormat 0x" << std::hex << c.internalFormat;
+        EXPECT_EQ(textureObject->GetFormat(), c.expected) << "internalFormat 0x" << std::hex << c.internalFormat;
+        EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR) << "internalFormat 0x" << std::hex << c.internalFormat;
+    }
+}
+
+// RGTC compresses 4x4 blocks of a 2D image and has no 3D form, so glTexImage3D must reject it even
+// though the same enum is accepted on a 2D target. The generic compressed formats carry no such
+// restriction and stay legal in 3D.
+TEST_F(TextureTest, RgtcInternalFormatsAreRejectedOnThreeDimensionalTargets) {
+    const GLenum rgtc[] = {GL_COMPRESSED_RED_RGTC1, GL_COMPRESSED_SIGNED_RED_RGTC1, GL_COMPRESSED_RG_RGTC2,
+                           GL_COMPRESSED_SIGNED_RG_RGTC2};
+    for (const GLenum internalFormat : rgtc) {
+        GLuint texture = 0;
+        MG_Impl::GLImpl::GenTextures(1, &texture);
+        MG_Impl::GLImpl::BindTexture(GL_TEXTURE_3D, texture);
+        MG_Impl::GLImpl::TexImage3D(GL_TEXTURE_3D, 0, internalFormat, 4, 4, 4, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
+        EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_INVALID_OPERATION)
+            << "internalFormat 0x" << std::hex << internalFormat;
+    }
+
+    GLuint generic = 0;
+    MG_Impl::GLImpl::GenTextures(1, &generic);
+    MG_Impl::GLImpl::BindTexture(GL_TEXTURE_3D, generic);
+    MG_Impl::GLImpl::TexImage3D(GL_TEXTURE_3D, 0, GL_COMPRESSED_RGBA, 4, 4, 4, 0, GL_RGBA, GL_UNSIGNED_BYTE, nullptr);
+    EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+}
+
 TEST_F(TextureTest, TextureStorage3DAndSubImageModifyNamedObjectOnly) {
     GLuint texture = 0;
     MG_Impl::GLImpl::CreateTextures(GL_TEXTURE_3D, 1, &texture);

MobileGL/MG_Util/BackendLoaders/OpenGL/Loader.cpp

@@ -9,7 +9,12 @@
 #include "Loader.h"
 #include "MG_Util/Types.h"
 #include <Config.h>
-#if !defined(__WIN32) && !defined(_WIN32)
+#if defined(_WIN32)
+#ifndef WIN32_LEAN_AND_MEAN
+#define WIN32_LEAN_AND_MEAN 1
+#endif
+#include <windows.h>
+#else
 #include <dlfcn.h>
 #endif
 
@@ -60,7 +65,14 @@ namespace MobileGL::MG_Util::BackendLoader {
 #endif
 
     static void* OpenLib(const Vector<String>& names) {
-#if !defined(__WIN32) && !defined(_WIN32) && (!defined(__APPLE__) || defined(MOBILEGL_IOS))
+#if defined(_WIN32)
+        for (const auto& name : names) {
+            if (HMODULE lib = LoadLibraryA(name.c_str())) {
+                MGLOG_I("Loaded GL backend library: %s", name.c_str());
+                return reinterpret_cast<void*>(lib);
+            }
+        }
+#elif !defined(__APPLE__) || defined(MOBILEGL_IOS)
         static const String LibPathPrefixes[] = {
 #if defined(MOBILEGL_IOS)
             "@rpath/", "@executable_path/Frameworks/", "@loader_path/Frameworks/",
@@ -104,7 +116,9 @@ namespace MobileGL::MG_Util::BackendLoader {
     }
 
     inline void* ProcAddress(void* lib, const char* name) {
-#if !defined(__WIN32) && !defined(_WIN32) && (!defined(__APPLE__) || defined(MOBILEGL_IOS))
+#if defined(_WIN32)
+        return reinterpret_cast<void*>(::GetProcAddress(reinterpret_cast<HMODULE>(lib), name));
+#elif !defined(__APPLE__) || defined(MOBILEGL_IOS)
         return dlsym(lib, name);
 #else
         return nullptr;
@@ -520,6 +534,15 @@ namespace MobileGL::MG_Util::BackendLoader {
 #if defined(MOBILEGL_TRACE_ANGLE_VARIANTS) && defined(__ANDROID__)
         void* angleGlesLib = nullptr;
 #endif
+#if defined(_WIN32)
+        // ANGLE is the GLES provider on Windows regardless of UseAngle(). Preload
+        // libGLESv2.dll so libEGL.dll resolves its dependency from the same directory.
+        if (!OpenLib({"libGLESv2.dll"})) {
+            MGLOG_E("Failed to open ANGLE libGLESv2.dll");
+            return;
+        }
+        eglLib = OpenLib({"libEGL.dll"});
+#else
         if (UseAngle()) {
             void* glesLib = OpenLib({"libGLESv2_angle.so"});
             if (!glesLib) {
@@ -541,6 +564,7 @@ namespace MobileGL::MG_Util::BackendLoader {
             eglLib = OpenLib({"libEGL.so"});
 #endif
         }
+#endif // !_WIN32
 
         if (!eglLib) {
             MGLOG_E("Failed to open EGL library");
@@ -811,6 +835,9 @@ namespace MobileGL::MG_Util::BackendLoader {
                 if (std::strcmp(extension, "GL_EXT_blend_func_extended") == 0) {
                     caps.SupportsDualSourceBlend = true;
                 }
+                if (std::strcmp(extension, "GL_NV_shader_noperspective_interpolation") == 0) {
+                    caps.SupportsNoperspectiveInterpolation = true;
+                }
             }
         }
 

MobileGL/MG_Util/BackendLoaders/OpenGL/Loader.h

@@ -1050,6 +1050,11 @@ namespace MobileGL {
             // factors and layout(index = 1) fragment outputs. GLES core has no dual-source blending,
             // so without this a draw using a SRC1 factor cannot proceed.
             Bool SupportsDualSourceBlend = false;
+            // GL_NV_shader_noperspective_interpolation is present: the driver accepts the
+            // `noperspective` interpolation qualifier in ESSL. GLES core has none, so without this
+            // SPIRV-Cross's `#extension ... : require` would fail to compile and MobileGL falls back
+            // to stripping the NoPerspective decoration (smooth interpolation) via StripNoPerspectivePass.
+            Bool SupportsNoperspectiveInterpolation = false;
             // GL_RENDERER contains "ANGLE".
             Bool IsAngleRenderer = false;
             // GL_RENDERER contains both "ANGLE" and "llvmpipe".

MobileGL/MG_Util/BackendLoaders/Vulkan/Loader.cpp

@@ -121,6 +121,7 @@ namespace MobileGL::MG_Util::BackendLoader {
         caps.VulkanAPIVersion = DecodeApiVersion(p.apiVersion);
         caps.DeviceName = p.deviceName;
         caps.DriverVersionString = DecodeDriverVersion(p.driverVersion);
+        caps.VendorId = p.vendorID;
         caps.UniformBufferOffsetAlignment = static_cast<int>(p.limits.minUniformBufferOffsetAlignment);
         caps.AliasedLineWidthRangeMin = p.limits.lineWidthRange[0];
         caps.AliasedLineWidthRangeMax = p.limits.lineWidthRange[1];
@@ -210,6 +211,7 @@ namespace MobileGL::MG_Util::BackendLoader {
         caps.VulkanAPIVersion = DecodeApiVersion(properties.apiVersion);
         caps.DeviceName = properties.deviceName;
         caps.DriverVersionString = DecodeDriverVersion(properties.driverVersion);
+        caps.VendorId = properties.vendorID;
         caps.UniformBufferOffsetAlignment = static_cast<int>(properties.limits.minUniformBufferOffsetAlignment);
         caps.AliasedLineWidthRangeMin = properties.limits.lineWidthRange[0];
         caps.AliasedLineWidthRangeMax = properties.limits.lineWidthRange[1];

MobileGL/MG_Util/BackendLoaders/Vulkan/Loader.h

@@ -15,6 +15,8 @@ namespace MobileGL {
             Version VulkanAPIVersion{1, 0, 0};
             String DeviceName;
             String DriverVersionString;
+            // VkPhysicalDeviceProperties::vendorID, for device-quirk vendor gating.
+            Uint32 VendorId = 0;
             Int UniformBufferOffsetAlignment = 256;
             Float AliasedLineWidthRangeMin = 1.0f;
             Float AliasedLineWidthRangeMax = 1.0f;

MobileGL/MG_Util/Converters/GLToMG/TextureEnumConverter.cpp

@@ -255,6 +255,37 @@ namespace MobileGL {
                 return TextureInternalFormat::DepthComponent;
             case GL_DEPTH_STENCIL:
                 return TextureInternalFormat::DepthStencil;
+            // Compressed internal formats resolve to the uncompressed storage that backs them.
+            //
+            // For the six generic formats this is exactly what GL prescribes: the implementation
+            // picks a specific compressed format, and when none is available it falls back to the
+            // corresponding base format. Nothing downstream ever sees a compressed enum, so the
+            // metrics, pixel-store and backend tables keep their "one format, N bytes per texel"
+            // invariant instead of each needing a compressed-aware arm.
+            //
+            // The four RGTC formats are a deliberate deviation: they are specific formats that GL
+            // 3.3 requires, but ES exposes no RGTC compressor to hand the data to. Storing the
+            // texels uncompressed keeps them renderable at the cost of the memory saving, which is
+            // strictly better than the INVALID_ENUM the application used to get. Note the signed
+            // variants must land on SNORM storage - CTS uploads them as GL_BYTE.
+            case GL_COMPRESSED_RED:
+            case GL_COMPRESSED_RED_RGTC1:
+                return TextureInternalFormat::R8;
+            case GL_COMPRESSED_SIGNED_RED_RGTC1:
+                return TextureInternalFormat::R8Snorm;
+            case GL_COMPRESSED_RG:
+            case GL_COMPRESSED_RG_RGTC2:
+                return TextureInternalFormat::RG8;
+            case GL_COMPRESSED_SIGNED_RG_RGTC2:
+                return TextureInternalFormat::RG8Snorm;
+            case GL_COMPRESSED_RGB:
+                return TextureInternalFormat::RGB8;
+            case GL_COMPRESSED_RGBA:
+                return TextureInternalFormat::RGBA8;
+            case GL_COMPRESSED_SRGB:
+                return TextureInternalFormat::SRGB8;
+            case GL_COMPRESSED_SRGB_ALPHA:
+                return TextureInternalFormat::SRGB8Alpha8;
             case GL_ALPHA:
             case GL_RED:
                 return TextureInternalFormat::Red;

MobileGL/MG_Util/Converters/MGToVk/TextureEnumConverter.cpp

@@ -133,9 +133,11 @@ namespace MobileGL {
             case TextureInternalFormat::RGBA8Snorm:
                 return VK_FORMAT_R8G8B8A8_SNORM;
             case TextureInternalFormat::RGB10A2:
-                return VK_FORMAT_A2R10G10B10_UNORM_PACK32;
+                // GL_UNSIGNED_INT_2_10_10_10_REV puts R in bits 0-9, which is Vulkan's
+                // A2B10G10R10 layout - A2R10G10B10 silently swaps R and B on upload.
+                return VK_FORMAT_A2B10G10R10_UNORM_PACK32;
             case TextureInternalFormat::RGB10A2UI:
-                return VK_FORMAT_A2R10G10B10_UINT_PACK32;
+                return VK_FORMAT_A2B10G10R10_UINT_PACK32;
             case TextureInternalFormat::RGBA16:
                 return VK_FORMAT_R16G16B16A16_UNORM;
             case TextureInternalFormat::RGBA16Snorm:

MobileGL/MG_Util/SelfTest/DriverPost.cpp

@@ -401,6 +401,230 @@ namespace MobileGL::MG_Util::SelfTest {
                              disabledNote);
         }
 
+        // Compiles + links a two-stage program on the probe context. Returns 0 on failure and writes a
+        // human-readable reason into |detail|.
+        GLuint CompileLinkProgram(const MG_External::GLESFunctionsTable& g, const char* vs, const char* fs,
+                                  String& detail) {
+            const auto compile = [&](GLenum stage, const char* src, GLuint& out) -> bool {
+                out = g.glCreateShader(stage);
+                if (out == 0) {
+                    detail = "glCreateShader returned 0";
+                    return false;
+                }
+                g.glShaderSource(out, 1, &src, nullptr);
+                g.glCompileShader(out);
+                GLint ok = GL_FALSE;
+                g.glGetShaderiv(out, GL_COMPILE_STATUS, &ok);
+                if (ok != GL_TRUE) {
+                    GLchar log[512] = {};
+                    GLsizei len = 0;
+                    g.glGetShaderInfoLog(out, static_cast<GLsizei>(sizeof(log) - 1), &len, log);
+                    detail = format("{} shader compile failed: {}",
+                                    stage == GL_VERTEX_SHADER ? "vertex" : "fragment",
+                                    len > 0 ? log : "(no info log)");
+                    return false;
+                }
+                return true;
+            };
+            GLuint v = 0, f = 0;
+            const ScopeGuard delV([&]() { if (v) g.glDeleteShader(v); });
+            const ScopeGuard delF([&]() { if (f) g.glDeleteShader(f); });
+            if (!compile(GL_VERTEX_SHADER, vs, v) || !compile(GL_FRAGMENT_SHADER, fs, f)) {
+                return 0;
+            }
+            const GLuint prog = g.glCreateProgram();
+            if (prog == 0) {
+                detail = "glCreateProgram returned 0";
+                return 0;
+            }
+            g.glAttachShader(prog, v);
+            g.glAttachShader(prog, f);
+            g.glLinkProgram(prog);
+            GLint linked = GL_FALSE;
+            g.glGetProgramiv(prog, GL_LINK_STATUS, &linked);
+            if (linked != GL_TRUE) {
+                detail = "program link failed";
+                g.glDeleteProgram(prog);
+                return 0;
+            }
+            return prog;
+        }
+
+        // "noperspective interpolation" row - a real correctness render, not just a compile. A viewport-
+        // filling quad is drawn with strong perspective (left clip-w 1, right clip-w 8) and a varying that
+        // runs 0..1 across it. At the screen centre screen-linear interpolation gives 0.5 while perspective-
+        // correct gives 1/(w+1) ~= 0.11, so reading the centre texel tells the two apart. The varying is
+        // carried either through the native `noperspective` qualifier (extension present) or through the
+        // exact gl_Position.w / gl_FragCoord.w rewrite MobileGL applies when it is absent. Verdict:
+        //   PASS  - extension present and the native noperspective result is screen-linear;
+        //   WARN  - extension absent but the gl_Position.w/gl_FragCoord.w emulation renders screen-linear
+        //           (correct, just the fallback path shipping shader packs hit on such devices);
+        //   FAIL  - either path renders perspective-correct / wrong (noperspective does not actually work),
+        //           or the program will not compile/link, or the render errors.
+        // Requires the probe context to still be current.
+        void ProbeGlesNoperspective(ReportBuilder& builder, const MG_External::GLESCapabilities& caps,
+                                    const MG_External::GLESFunctionsTable& g) {
+            const Bool native = caps.SupportsNoperspectiveInterpolation;
+            const String pathNote = native ? "GL_NV_shader_noperspective_interpolation present (native path)"
+                                           : "GL_NV_shader_noperspective_interpolation absent (gl_Position.w / "
+                                             "gl_FragCoord.w emulation path)";
+            const auto fail = [&](const String& detail) {
+                builder.Fail("noperspective interpolation", pathNote + "; " + detail);
+            };
+
+            if (!g.glCreateShader || !g.glShaderSource || !g.glCompileShader || !g.glGetShaderiv ||
+                !g.glGetShaderInfoLog || !g.glDeleteShader || !g.glCreateProgram || !g.glAttachShader ||
+                !g.glLinkProgram || !g.glGetProgramiv || !g.glUseProgram || !g.glDeleteProgram ||
+                !g.glGenFramebuffers || !g.glBindFramebuffer || !g.glDeleteFramebuffers ||
+                !g.glGenRenderbuffers || !g.glBindRenderbuffer || !g.glRenderbufferStorage ||
+                !g.glFramebufferRenderbuffer || !g.glDeleteRenderbuffers || !g.glCheckFramebufferStatus ||
+                !g.glGenBuffers || !g.glBindBuffer || !g.glBufferData || !g.glDeleteBuffers ||
+                !g.glGetAttribLocation || !g.glVertexAttribPointer || !g.glEnableVertexAttribArray ||
+                !g.glViewport || !g.glClearColor || !g.glClear || !g.glDrawArrays || !g.glReadPixels ||
+                !g.glFinish || !g.glGetError) {
+                fail("the render entry points did not resolve through eglGetProcAddress");
+                return;
+            }
+
+            // Match MobileGL's own ESSL target (the device's version). At #version 300 es some drivers
+            // (Adreno) still treat `noperspective` as reserved even with the extension enabled; the ES 3.2
+            // form the backend actually emits compiles. Emulated shaders are version-agnostic but use the
+            // same header for consistency.
+            const Int esslVer = caps.GLESVersion.Major * 100 + caps.GLESVersion.Minor * 10;
+            const String header = format("#version {} es\n", esslVer >= 300 ? esslVer : 300);
+            static const char* const kVsNativeBody =
+                "#extension GL_NV_shader_noperspective_interpolation : require\n"
+                "in vec4 a_pos;\n"
+                "in float a_v;\n"
+                "noperspective out highp float v_out;\n"
+                "void main() { gl_Position = a_pos; v_out = a_v; }\n";
+            static const char* const kFsNativeBody =
+                "#extension GL_NV_shader_noperspective_interpolation : require\n"
+                "precision highp float;\n"
+                "noperspective in highp float v_out;\n"
+                "out vec4 fragColor;\n"
+                "void main() { fragColor = vec4(v_out, 0.0, 0.0, 1.0); }\n";
+            // Exactly MobileGL's emulation (verified against EmulateNoPerspectivePass output): pre-multiply
+            // the varying by clip-w in the vertex stage, recover with gl_FragCoord.w in the fragment stage,
+            // no noperspective qualifier (so the driver interpolates it perspective-correct).
+            static const char* const kVsEmuBody =
+                "in vec4 a_pos;\n"
+                "in float a_v;\n"
+                "out highp float v_out;\n"
+                "void main() { gl_Position = a_pos; v_out = a_v * gl_Position.w; }\n";
+            static const char* const kFsEmuBody =
+                "precision highp float;\n"
+                "in highp float v_out;\n"
+                "out vec4 fragColor;\n"
+                "void main() { fragColor = vec4(v_out * gl_FragCoord.w, 0.0, 0.0, 1.0); }\n";
+
+            while (g.glGetError() != GL_NO_ERROR) {
+            }
+
+            const String vsSrc = header + (native ? kVsNativeBody : kVsEmuBody);
+            const String fsSrc = header + (native ? kFsNativeBody : kFsEmuBody);
+            String linkDetail;
+            const GLuint prog = CompileLinkProgram(g, vsSrc.c_str(), fsSrc.c_str(), linkDetail);
+            if (prog == 0) {
+                fail(native ? "a noperspective program failed to build though the extension is advertised: " +
+                                  linkDetail
+                            : "the emulation program failed to build: " + linkDetail);
+                return;
+            }
+            const ScopeGuard delProg([&]() { g.glDeleteProgram(prog); });
+
+            // 9x9 so the centre texel (4,4) sits exactly at NDC (0,0).
+            constexpr GLsizei kDim = 9;
+            GLuint rbo = 0, fbo = 0, vbo = 0;
+            g.glGenRenderbuffers(1, &rbo);
+            const ScopeGuard delRbo([&]() { if (rbo) g.glDeleteRenderbuffers(1, &rbo); });
+            g.glBindRenderbuffer(GL_RENDERBUFFER, rbo);
+            g.glRenderbufferStorage(GL_RENDERBUFFER, GL_RGBA8, kDim, kDim);
+            g.glGenFramebuffers(1, &fbo);
+            const ScopeGuard delFbo([&]() {
+                if (fbo) {
+                    g.glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                    g.glDeleteFramebuffers(1, &fbo);
+                }
+            });
+            g.glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            g.glFramebufferRenderbuffer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_RENDERBUFFER, rbo);
+            if (g.glCheckFramebufferStatus(GL_FRAMEBUFFER) != GL_FRAMEBUFFER_COMPLETE) {
+                fail("the probe framebuffer is incomplete");
+                return;
+            }
+
+            // Interleaved [vec4 clip-pos, float v]. Left w=1, right w=8; x/y pre-multiplied by w so the quad
+            // still fills NDC after the perspective divide.
+            const GLfloat verts[] = {
+                -1.f, -1.f, 0.f, 1.f, 0.f, //
+                8.f,  -8.f, 0.f, 8.f, 1.f, //
+                -1.f, 1.f,  0.f, 1.f, 0.f, //
+                8.f,  8.f,  0.f, 8.f, 1.f, //
+            };
+            g.glGenBuffers(1, &vbo);
+            const ScopeGuard delVbo([&]() { if (vbo) g.glDeleteBuffers(1, &vbo); });
+            g.glBindBuffer(GL_ARRAY_BUFFER, vbo);
+            g.glBufferData(GL_ARRAY_BUFFER, sizeof(verts), verts, GL_STATIC_DRAW);
+
+            g.glUseProgram(prog);
+            const GLint posLoc = g.glGetAttribLocation(prog, "a_pos");
+            const GLint vLoc = g.glGetAttribLocation(prog, "a_v");
+            if (posLoc < 0 || vLoc < 0) {
+                fail("the probe vertex attributes did not resolve");
+                return;
+            }
+            g.glEnableVertexAttribArray(static_cast<GLuint>(posLoc));
+            g.glVertexAttribPointer(static_cast<GLuint>(posLoc), 4, GL_FLOAT, GL_FALSE, 5 * sizeof(GLfloat),
+                                    reinterpret_cast<const void*>(0));
+            g.glEnableVertexAttribArray(static_cast<GLuint>(vLoc));
+            g.glVertexAttribPointer(static_cast<GLuint>(vLoc), 1, GL_FLOAT, GL_FALSE, 5 * sizeof(GLfloat),
+                                    reinterpret_cast<const void*>(4 * sizeof(GLfloat)));
+
+            g.glViewport(0, 0, kDim, kDim);
+            g.glClearColor(0.f, 0.f, 0.f, 1.f);
+            g.glClear(GL_COLOR_BUFFER_BIT);
+            g.glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+            g.glFinish();
+
+            const GLenum drawError = g.glGetError();
+            if (drawError != GL_NO_ERROR) {
+                fail(format("GL error 0x{:x} while rendering the probe quad", drawError));
+                return;
+            }
+
+            GLubyte center[4] = {};
+            g.glReadPixels(kDim / 2, kDim / 2, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, center);
+            const GLenum readError = g.glGetError();
+            if (readError != GL_NO_ERROR) {
+                fail(format("GL error 0x{:x} while reading the probe pixel back", readError));
+                return;
+            }
+
+            // At the centre: screen-linear -> 0.5 (~128); perspective-correct -> 1/(8+1) ~= 0.111 (~28).
+            const float observed = static_cast<float>(center[0]) / 255.0f;
+            const int observedByte = center[0];
+            constexpr float kScreenLinear = 0.5f;
+            const bool screenLinear = observed > 0.5f * (kScreenLinear + 1.0f / 9.0f); // midpoint ~= 0.306
+            if (!screenLinear) {
+                fail(format("the centre texel read {} (~{:.3f}); expected the screen-linear ~0.5 - "
+                            "interpolation came out perspective-correct, so noperspective does not work here",
+                            observedByte, observed));
+                return;
+            }
+            if (native) {
+                builder.Pass("noperspective interpolation",
+                             pathNote + format("; native noperspective renders screen-linear (centre {} ~= 0.5)",
+                                               observedByte));
+            } else {
+                builder.Warn("noperspective interpolation",
+                             pathNote +
+                                 format("; the emulation renders screen-linear correctly (centre {} ~= 0.5), "
+                                        "but this is the fallback path with less driver coverage",
+                                        observedByte));
+            }
+        }
+
         // Everything the "MobileGL reported ..." rows need from the GLES device probe.
         struct GlesProbeSummary {
             Bool capsValid = false;
@@ -527,6 +751,7 @@ namespace MobileGL::MG_Util::SelfTest {
             builder.report.rendererInfo = format("{} ({})", caps.GLESRendererString, caps.GLESVersionString);
             EvaluateGlesChecklist(builder, caps, glesFuncs);
             ProbeGlesTimerQuery(builder, caps, glesFuncs);
+            ProbeGlesNoperspective(builder, caps, glesFuncs);
             builder.report.formatCapabilities.emplace();
             MG_Backend::DirectGLES::PopulateFormatCapabilities(
                 glesFuncs, caps, builder.report.formatCapabilities.value());

MobileGL/MG_Util/ShaderTranspiler/ShaderCompiler.cpp

@@ -16,9 +16,12 @@
 #include "SpirvPasses/FlattenInterfaceStructPass.h"
 #include "SpirvPasses/RenameSamplerFunctionParameterPass.h"
 #include "SpirvPasses/DecomposeWorkgroupVec3Pass.h"
+#include "SpirvPasses/DecoratePositionInvariantPass.h"
 #include "SpirvPasses/LowerDrawParametersPass.h"
 #include "SpirvPasses/RebaseInstanceIndexPass.h"
 #include "SpirvPasses/StripUboMemberRelaxedPrecisionPass.h"
+#include "SpirvPasses/StripNoPerspectivePass.h"
+#include "SpirvPasses/EmulateNoPerspectivePass.h"
 #include "spirv-tools/libspirv.h"
 #include "spirv-tools/optimizer.hpp"
 
@@ -26,6 +29,7 @@
 #include <MG_Backend/BackendObjects.h>
 #include <MG_Util/Converters/GLToStr/GLEnumConverter.h>
 #include <MG_Util/Converters/GLToGlslang/ProgramEnumConverter.h>
+#include <cstdlib>
 
 namespace MobileGL {
     namespace MG_Util {
@@ -332,6 +336,30 @@ namespace MobileGL {
                 return optimizer.Run(inputBinary.data(), inputBinary.size(), &outputBinary, options);
             }
 
+            bool ShaderCompiler::StripNoPerspectiveForEssl(const Vector<Uint32>& inputBinary,
+                                                           Vector<uint32_t>& outputBinary) {
+                using namespace spvtools;
+                OptimizerOptions options;
+                options.set_run_validator(false);
+
+                Optimizer optimizer(SPV_ENV_VULKAN_1_1);
+                optimizer.RegisterPass(StripNoPerspectivePass::CreateStripNoPerspectivePass());
+
+                return optimizer.Run(inputBinary.data(), inputBinary.size(), &outputBinary, options);
+            }
+
+            bool ShaderCompiler::EmulateNoPerspectiveForEssl(const Vector<Uint32>& inputBinary,
+                                                             Vector<uint32_t>& outputBinary) {
+                using namespace spvtools;
+                OptimizerOptions options;
+                options.set_run_validator(false);
+
+                Optimizer optimizer(SPV_ENV_VULKAN_1_1);
+                optimizer.RegisterPass(EmulateNoPerspectivePass::CreateEmulateNoPerspectivePass());
+
+                return optimizer.Run(inputBinary.data(), inputBinary.size(), &outputBinary, options);
+            }
+
             bool ShaderCompiler::RebaseInstanceIndexForVulkan(const Vector<Uint32>& inputBinary,
                                                               Vector<uint32_t>& outputBinary) {
                 using namespace spvtools;
@@ -344,6 +372,18 @@ namespace MobileGL {
                 return optimizer.Run(inputBinary.data(), inputBinary.size(), &outputBinary, options);
             }
 
+            bool ShaderCompiler::DecoratePositionInvariantForVulkan(const Vector<Uint32>& inputBinary,
+                                                                    Vector<uint32_t>& outputBinary) {
+                using namespace spvtools;
+                OptimizerOptions options;
+                options.set_run_validator(false);
+
+                Optimizer optimizer(SPV_ENV_VULKAN_1_1);
+                optimizer.RegisterPass(DecoratePositionInvariantPass::CreateDecoratePositionInvariantPass());
+
+                return optimizer.Run(inputBinary.data(), inputBinary.size(), &outputBinary, options);
+            }
+
             bool ShaderCompiler::UseUnformattedFloatStorageImagesForVulkan(
                 const Vector<Uint32>& inputBinary, Vector<uint32_t>& outputBinary) {
                 constexpr SizeT kSpirvHeaderWordCount = 5;

MobileGL/MG_Util/ShaderTranspiler/ShaderCompiler.h

@@ -33,12 +33,29 @@ namespace MobileGL {
                 // Only for the DirectGLES transpile path.
                 static bool StripUboMemberRelaxedPrecisionForEssl(const Vector<Uint32>& inputBinary,
                                                                   Vector<uint32_t>& outputBinary);
+                // Removes NoPerspective decorations so SPIRV-Cross emits plain (smooth) ESSL varyings.
+                // DirectGLES fallback only, for devices lacking GL_NV_shader_noperspective_interpolation
+                // (SPIRV-Cross would otherwise require that extension and the driver would reject it).
+                static bool StripNoPerspectiveForEssl(const Vector<Uint32>& inputBinary,
+                                                      Vector<uint32_t>& outputBinary);
+                // Emulates noperspective (screen-linear) interpolation via gl_Position.w / gl_FragCoord.w
+                // so no NV extension is needed; strips what it cannot emulate. DirectGLES fallback for
+                // devices lacking GL_NV_shader_noperspective_interpolation. See EmulateNoPerspectivePass.
+                static bool EmulateNoPerspectiveForEssl(const Vector<Uint32>& inputBinary,
+                                                        Vector<uint32_t>& outputBinary);
                 // Rebases loads of the InstanceIndex builtin to (InstanceIndex - BaseInstance) so
                 // shaders see GL's zero-based gl_InstanceID. Vertex shaders only; DirectVulkan
                 // backend only (glslang's relaxed mode aliases gl_InstanceID to gl_InstanceIndex,
                 // which wrongly includes baseInstance).
                 static bool RebaseInstanceIndexForVulkan(const Vector<Uint32>& inputBinary,
                                                          Vector<uint32_t>& outputBinary);
+                // Adds the Invariant decoration to every Position builtin output. GL apps
+                // routinely rely on cross-program position invariance for multi-pass
+                // equality depth tests (e.g. GEQUAL re-draws of the same geometry), and
+                // mobile drivers that optimize per-pipeline break that without the
+                // decoration. DirectVulkan only.
+                static bool DecoratePositionInvariantForVulkan(const Vector<Uint32>& inputBinary,
+                                                               Vector<uint32_t>& outputBinary);
                 // Replaces the declared format of float storage images with Unknown and adds the
                 // matching SPIR-V capabilities. DirectVulkan uses this only when both Vulkan
                 // shaderStorageImage*WithoutFormat features are enabled, allowing the

MobileGL/MG_Util/ShaderTranspiler/ShaderSourceProcessor.cpp

@@ -12,6 +12,7 @@
 #include <cctype>
 #include <initializer_list>
 #include <utility>
+#include <Config.h>
 #include <MG_Backend/BackendObjects.h>
 
 namespace {
@@ -95,6 +96,77 @@ namespace {
         return masked;
     }
 
+    // Blank out block comments in place, leaving line comments and every other byte where it is.
+    //
+    // The passes that follow scan the source as raw text, so block comments have to stop being
+    // visible to them - but they must not be *deleted*: replacing the bytes with spaces keeps every
+    // later offset valid and keeps newlines, so glslang's diagnostics still point at the line the
+    // application wrote. It also has to be lexically aware. A banner line such as
+    //
+    //     //*** lighting pass ***
+    //
+    // contains "/*" one byte in, and a naive search for that opener treats the rest of the file as
+    // an unterminated comment.
+    void BlankBlockComments(MobileGL::String& source) {
+        enum class Region { Code, SingleLineComment, MultiLineComment, QuotedText };
+
+        Region region = Region::Code;
+        char quote = '\0';
+        bool escaped = false;
+
+        for (SizeT pos = 0; pos < source.size(); pos++) {
+            const char ch = source[pos];
+            const char next = pos + 1 < source.size() ? source[pos + 1] : '\0';
+
+            if (region == Region::Code) {
+                if (ch == '/' && next == '/') {
+                    pos++;
+                    region = Region::SingleLineComment;
+                } else if (ch == '/' && next == '*') {
+                    source[pos] = ' ';
+                    source[pos + 1] = ' ';
+                    pos++;
+                    region = Region::MultiLineComment;
+                } else if (ch == '"' || ch == '\'') {
+                    quote = ch;
+                    escaped = false;
+                    region = Region::QuotedText;
+                }
+                continue;
+            }
+
+            if (region == Region::SingleLineComment) {
+                if (ch == '\n' || ch == '\r') region = Region::Code;
+                continue;
+            }
+
+            if (region == Region::MultiLineComment) {
+                if (ch == '*' && next == '/') {
+                    source[pos] = ' ';
+                    source[pos + 1] = ' ';
+                    pos++;
+                    region = Region::Code;
+                } else if (ch != '\n' && ch != '\r') {
+                    source[pos] = ' ';
+                }
+                continue;
+            }
+
+            // GLSL has no multi-line string literals, so a quote that reaches end of line was never
+            // a literal to begin with - most likely an apostrophe in a #error or #pragma message.
+            // Ending the region here keeps one stray apostrophe from swallowing the rest of the file.
+            if (ch == '\n' || ch == '\r') {
+                region = Region::Code;
+            } else if (escaped) {
+                escaped = false;
+            } else if (ch == '\\') {
+                escaped = true;
+            } else if (ch == quote) {
+                region = Region::Code;
+            }
+        }
+    }
+
     struct CodeToken {
         String text;
         SizeT begin = 0;
@@ -365,7 +437,21 @@ namespace {
             HasIdentifierWithPrefixOutsideAllowed(tokens, "subgroup", {"subgroupInclusiveAdd"}) ||
             HasIdentifierWithPrefixOutsideAllowed(
                 tokens, "gl_Subgroup",
-                {"gl_SubgroupInvocationID", "gl_SubgroupSize", "gl_SubgroupID", "gl_NumSubgroups"})) {
+                {"gl_SubgroupInvocationID", "gl_SubgroupSize", "gl_SubgroupID", "gl_NumSubgroups"}) ||
+            // ARB/NV spellings of lane-width-sensitive builtins and functions
+            // (gl_SubGroupSizeARB, ballotARB, gl_WarpSizeNV, shuffleNV, ...) must block the
+            // rewrite just like their KHR counterparts: they would silently keep native-width
+            // semantics in a module rewritten to the virtual 32-lane model.
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "gl_SubGroup", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "gl_Warp", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "gl_Thread", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "gl_SMID", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "ballot", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "shuffle", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "readInvocation", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "readFirstInvocation", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "anyInvocation", {}) ||
+            HasIdentifierWithPrefixOutsideAllowed(tokens, "allInvocations", {})) {
             return false;
         }
 
@@ -487,6 +573,23 @@ namespace {
                static_cast<unsigned char>(source[1]) == 0xbb && static_cast<unsigned char>(source[2]) == 0xbf;
     }
 
+    // The GLSL versions MobileGL is willing to normalize. Anything else in a #version line - a number
+    // that is not a real language version (329, 331), a bad profile keyword, a float/identifier where
+    // the integer belongs, or trailing tokens - is left untouched so glslang rejects it, matching
+    // KHR-GL33.shaders.preprocessor.directive.version_*. The set is deliberately generous (every real
+    // desktop and ES version) so the normalizer never starts rejecting a form it used to accept.
+    bool IsRecognizedGlslVersion(unsigned version) {
+        switch (version) {
+            case 100: case 110: case 120: case 130: case 140: case 150:
+            case 300: case 310: case 320:
+            case 330: case 400: case 410: case 420: case 430:
+            case 440: case 450: case 460:
+                return true;
+            default:
+                return false;
+        }
+    }
+
     struct ShaderLanguageInfo {
         unsigned version = 110;
         MobileGL::ShaderProfile profile = MobileGL::ShaderProfile::Core;
@@ -494,6 +597,9 @@ namespace {
         SizeT versionDirectiveEnd = MobileGL::String::npos;
         bool hasUtf8Bom = false;
         bool enablesGpuShader5 = false;
+        // Whether the parsed #version directive is a well-formed one MobileGL should rewrite. A
+        // malformed directive (see IsRecognizedGlslVersion) is left alone for glslang to reject.
+        bool hasValidVersionDirective = false;
 
         bool HasVersionDirective() const { return versionDirectiveStart != MobileGL::String::npos; }
     };
@@ -537,13 +643,25 @@ namespace {
                         info.versionDirectiveEnd = lineEnd + (hasLineBreak ? 1 : 0);
                         SkipDirectiveWhitespace(code, probe, lineEnd);
                         const MobileGL::String profile = ReadDirectiveIdentifier(code, probe, lineEnd);
-                        if (profile == "es" || profile == "ES") {
+                        bool profileTokenValid = true;
+                        if (profile.empty() || profile == "core") {
+                            info.profile = MobileGL::ShaderProfile::Core;
+                        } else if (profile == "es" || profile == "ES") {
                             info.profile = MobileGL::ShaderProfile::ES;
                         } else if (profile == "compatibility") {
                             info.profile = MobileGL::ShaderProfile::Compatibility;
                         } else {
+                            // "#version 330 foo": an unrecognized profile keyword. Keep Core for any
+                            // downstream routing, but mark the directive malformed.
                             info.profile = MobileGL::ShaderProfile::Core;
+                            profileTokenValid = false;
                         }
+                        // Comments are already masked to spaces, so anything non-blank left on the
+                        // line is real trailing garbage: "#version 330 foobar" / "#version 330.0".
+                        SkipDirectiveWhitespace(code, probe, lineEnd);
+                        const bool hasTrailingTokens = probe < lineEnd;
+                        info.hasValidVersionDirective =
+                            IsRecognizedGlslVersion(info.version) && profileTokenValid && !hasTrailingTokens;
                     }
                 } else if (directive == "extension") {
                     SkipDirectiveWhitespace(code, probe, lineEnd);
@@ -590,6 +708,17 @@ namespace {
     }
 
     void NormalizeVersionDirective(MobileGL::String& source, const ShaderLanguageInfo& info) {
+        // A malformed #version (329, 331, bad profile, float/trailing tokens) is left exactly as the
+        // application wrote it so glslang rejects it - rewriting it to "#version 330 core" would
+        // silently legalize the CTS directive.version_* rejection cases. Still drop a leading BOM so
+        // the reported error is the bad version rather than a stray byte-order mark.
+        if (info.HasVersionDirective() && !info.hasValidVersionDirective) {
+            if (info.hasUtf8Bom) {
+                source.erase(0, 3);
+            }
+            return;
+        }
+
         const MobileGL::String replacement = GetNormalizedVersionDirective(info);
         if (info.HasVersionDirective()) {
             source.replace(info.versionDirectiveStart, info.versionDirectiveEnd - info.versionDirectiveStart,
@@ -671,7 +800,10 @@ namespace {
 
     void RenameBuiltinShadowingFunction(MobileGL::String& source, const char* from, const char* to) {
         const MobileGL::String fromName = from;
-        if (!HasSingleLineFunctionDefinition(source, fromName)) {
+        // Decide from a comment-free view. A commented-out definition is not a definition, and
+        // acting on one renames every genuine call to the builtin to a name nothing defines - which
+        // then fails to resolve. Line comments survive BlankBlockComments, so this matters.
+        if (!HasSingleLineFunctionDefinition(MaskCommentsAndQuotedText(source), fromName)) {
             return;
         }
 
@@ -744,6 +876,58 @@ namespace {
         return info.HasVersionDirective() ? info.versionDirectiveEnd : 0;
     }
 
+    // GLSL's #line takes integer expressions only, but plenty of shader-pack preprocessors emit the
+    // C form with a quoted filename. Deleting every #line outright made those harmless - at the cost
+    // of __LINE__ reporting the position in MobileGL's rewritten text rather than the one the pack
+    // author wrote, and of every later diagnostic pointing at the wrong line. Dropping just the
+    // quoted operand keeps the directive doing its job and still hands glslang something it accepts.
+    void NormalizeLineDirectives(MobileGL::String& source) {
+        const MobileGL::String masked = MaskCommentsAndQuotedText(source);
+        const SizeT versionEnd = FindAfterVersionDirective(source);
+        MobileGL::String result;
+        result.reserve(source.size());
+
+        SizeT lineStart = 0;
+        while (lineStart <= source.size()) {
+            SizeT lineEnd = source.find('\n', lineStart);
+            const bool lastLine = lineEnd == MobileGL::String::npos;
+            if (lastLine) lineEnd = source.size();
+
+            SizeT probe = lineStart;
+            while (probe < lineEnd && (source[probe] == ' ' || source[probe] == '\t')) probe++;
+
+            const bool isLineDirective = masked.compare(probe, 5, "#line") == 0 &&
+                                         (probe + 5 >= lineEnd || !IsIdentifierChar(source[probe + 5]));
+            if (isLineDirective && lineStart < versionEnd) {
+                // #version has to be the first token in the shader, so a #line ahead of it could
+                // never have taken effect. Drop it rather than hand glslang a source it must reject
+                // - some pack preprocessors emit their directives before the version line.
+            } else if (isLineDirective) {
+                // Keep everything up to the first quote that the masker identified as string text.
+                SizeT quotePos = MobileGL::String::npos;
+                for (SizeT i = probe + 5; i < lineEnd; i++) {
+                    if (source[i] == '"' || source[i] == '\'') {
+                        quotePos = i;
+                        break;
+                    }
+                }
+                if (quotePos != MobileGL::String::npos) {
+                    result.append(source, lineStart, quotePos - lineStart);
+                } else {
+                    result.append(source, lineStart, lineEnd - lineStart);
+                }
+            } else {
+                result.append(source, lineStart, lineEnd - lineStart);
+            }
+
+            if (lastLine) break;
+            result.push_back('\n');
+            lineStart = lineEnd + 1;
+        }
+
+        source = std::move(result);
+    }
+
     bool IsExtensionAdvertised(MobileGL::GLExtension extension) {
         const auto& activeBackendObject = MobileGL::MG_Backend::pActiveBackendObject;
         if (!activeBackendObject) {
@@ -772,15 +956,28 @@ namespace {
             return;
         }
 
+        // Detect the directive on a comment/string-masked copy so a commented-out
+        // "#extension GL_ARB_gpu_shader_int64" is never turned into a synthesized #error. Comments are
+        // no longer blanked in the delivered source (glslang handles them), so this pass must mask
+        // locally like its siblings. Masking preserves offsets, so edits collected against the scan
+        // apply verbatim to `source`; they are applied back-to-front to keep earlier offsets valid.
+        const MobileGL::String scan = MaskCommentsAndQuotedText(source);
+        struct DirectiveEdit {
+            SizeT pos;
+            SizeT len;
+            MobileGL::String replacement;
+        };
+        Vector<DirectiveEdit> edits;
+
         SizeT lineStart = 0;
-        while (lineStart < source.size()) {
-            SizeT lineEnd = source.find('\n', lineStart);
+        while (lineStart < scan.size()) {
+            SizeT lineEnd = scan.find('\n', lineStart);
             const bool hasLineBreak = lineEnd != MobileGL::String::npos;
             if (!hasLineBreak) {
-                lineEnd = source.size();
+                lineEnd = scan.size();
             }
 
-            const MobileGL::String line = source.substr(lineStart, lineEnd - lineStart);
+            const MobileGL::String line = scan.substr(lineStart, lineEnd - lineStart);
             SizeT probe = 0;
             while (probe < line.size() && std::isspace(static_cast<unsigned char>(line[probe]))) {
                 probe++;
@@ -822,16 +1019,12 @@ namespace {
                             const MobileGL::String behavior = TrimDirectiveToken(line.substr(probe));
                             const SizeT replaceLen = lineEnd - lineStart + (hasLineBreak ? 1 : 0);
                             if (behavior == "require") {
-                                const MobileGL::String replacement =
-                                    "#error GL_ARB_gpu_shader_int64 is not advertised by MobileGL\n";
-                                source.replace(lineStart, replaceLen, replacement);
-                                lineStart += replacement.size();
+                                edits.push_back({lineStart, replaceLen,
+                                                 "#error GL_ARB_gpu_shader_int64 is not advertised by MobileGL\n"});
                             } else if (behavior == "enable" || behavior == "warn") {
-                                source.replace(lineStart, replaceLen, "\n");
-                                lineStart++;
-                            } else {
-                                lineStart = lineEnd + (hasLineBreak ? 1 : 0);
+                                edits.push_back({lineStart, replaceLen, "\n"});
                             }
+                            lineStart = lineEnd + (hasLineBreak ? 1 : 0);
                             continue;
                         }
                     }
@@ -841,6 +1034,10 @@ namespace {
             lineStart = lineEnd + (hasLineBreak ? 1 : 0);
         }
 
+        for (auto it = edits.rbegin(); it != edits.rend(); ++it) {
+            source.replace(it->pos, it->len, it->replacement);
+        }
+
         ReplaceIdentifier(source, "GL_ARB_gpu_shader_int64", "MG_DISABLED_GL_ARB_gpu_shader_int64");
     }
 
@@ -967,6 +1164,14 @@ namespace MobileGL {
                 const Vector<CodeToken> tokens = TokenizeCode(source);
                 LinearPrefixScanMatch match;
                 if (!ParseLinearPrefixScanTemplate(tokens, match)) {
+                    // Diagnosability: when the trigger op is present but the template no longer
+                    // matches (e.g. the pack shipped a new shader revision), the affected device
+                    // silently falls back to the driver's miscompiled path. Make that visible.
+                    if (CountToken(tokens, "subgroupInclusiveAdd") > 0) {
+                        MGLOG_W("%s: subgroupInclusiveAdd present but the linear prefix-scan template "
+                                "did not match; the wide-subgroup rewrite was NOT applied",
+                                __func__);
+                    }
                     return false;
                 }
 
@@ -979,48 +1184,98 @@ namespace MobileGL {
                 return true;
             }
 
+            namespace {
+                struct ShaderSourceQuirkContext {
+                    ShaderStage stage = ShaderStage::Unknown;
+                    BackendType backend = BackendType::Unknown;
+                    MG_Backend::GpuVendorKind vendor = MG_Backend::GpuVendorKind::Unknown;
+                    Uint32 subgroupSize = 0;
+                };
+
+                // Device-quirk registry. Every entry is a narrowly scoped source rewrite that
+                // works around a specific driver defect. A quirk runs when its env override
+                // forces it on, or when the override is Auto and DeviceApplies matches the
+                // detected device. ForceOn bypasses only the device gate - each Apply keeps
+                // its own structural safety checks. Add new per-device workarounds here
+                // instead of open-coding them in PreprocessShaderSource.
+                struct ShaderSourceQuirk {
+                    const char* name;
+                    MG_Config::QuirkOverride (*GetOverride)();
+                    Bool (*DeviceApplies)(const ShaderSourceQuirkContext&);
+                    Bool (*Apply)(const ShaderSourceQuirkContext&, String&);
+                };
+
+                constexpr ShaderSourceQuirk kShaderSourceQuirks[] = {
+                    {
+                        // MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN
+                        "subgroup-prefix-scan-rewrite",
+                        [] { return MG_Config::Features.SubgroupPrefixScanQuirk; },
+                        [](const ShaderSourceQuirkContext& ctx) {
+                            // Qualcomm's Vulkan driver miscompiles the recognized float
+                            // InclusiveScan pattern for native subgroups wider than the
+                            // captured 32 lanes; other vendors compile it correctly and
+                            // should keep their native scan.
+                            return ctx.backend == BackendType::DirectVulkan &&
+                                   ctx.vendor == MG_Backend::GpuVendorKind::Qualcomm;
+                        },
+                        [](const ShaderSourceQuirkContext& ctx, String& source) {
+                            return RewriteLinearSubgroupPrefixScanForVulkan(ctx.stage, ctx.subgroupSize,
+                                                                            source);
+                        },
+                    },
+                };
+
+                void ApplyShaderSourceQuirks(ShaderStage stage, String& source) {
+                    const auto& activeBackend = MG_Backend::pActiveBackendObject;
+                    if (!activeBackend) {
+                        return;
+                    }
+                    const auto& dynamicParameters = activeBackend->GetDynamicParameters();
+                    const ShaderSourceQuirkContext quirkContext{
+                        stage,
+                        activeBackend->GetBackendType(),
+                        dynamicParameters.GpuVendor,
+                        dynamicParameters.SubgroupSize,
+                    };
+                    for (const ShaderSourceQuirk& quirk : kShaderSourceQuirks) {
+                        const MG_Config::QuirkOverride quirkOverride = quirk.GetOverride();
+                        if (quirkOverride == MG_Config::QuirkOverride::ForceOff) {
+                            continue;
+                        }
+                        if (quirkOverride == MG_Config::QuirkOverride::Auto &&
+                            !quirk.DeviceApplies(quirkContext)) {
+                            continue;
+                        }
+                        if (quirk.Apply(quirkContext, source)) {
+                            MGLOG_I("ApplyShaderSourceQuirks: applied '%s'%s", quirk.name,
+                                    quirkOverride == MG_Config::QuirkOverride::ForceOn ? " (forced on)" : "");
+                        }
+                    }
+                }
+            } // namespace
+
             void PreprocessShaderSource(ShaderStage stage, String& source) {
                 // Normalize while the inspector's source span still refers to the untouched input. Later passes
                 // remove comments and directives, so any subsequent insertion re-inspects the current source.
                 const ShaderLanguageInfo originalLanguage = InspectShaderLanguage(source);
                 NormalizeVersionDirective(source, originalLanguage);
 
-                // remove multi-line comment
-                size_t commentStartPos = source.find("/*");
-                while (commentStartPos != String::npos) {
-                    size_t commentEndPos = source.find("*/", commentStartPos);
-                    if (commentEndPos == String::npos) {
-                        source.erase(commentStartPos);
-                        break;
-                    }
-                    // + length of "*/"
-                    source = source.replace(commentStartPos, commentEndPos - commentStartPos + 2, "");
-                    commentStartPos = source.find("/*", commentStartPos);
-                }
+                // Comments are left intact for glslang's own preprocessor: a block comment is a single
+                // preprocessing token that collapses to one space even across newlines and inside a
+                // directive, so blanking it here (which preserved the interior newlines) truncated
+                // multi-line #define bodies and broke otherwise-valid shaders (KHR-GL3x.shaders.
+                // preprocessor multiline_comment_define / redefine_object / function_redefinition).
+                // Every MobileGL pass that must ignore comment/string text already masks them locally
+                // via MaskCommentsAndQuotedText/TokenizeCode, so the source we hand glslang keeps them.
+                NormalizeLineDirectives(source);
 
-                // remove #line directives
-                SizeT linedirPos = source.find("#line");
-                while (linedirPos != String::npos) {
-                    SizeT newlinePos = source.find('\n', linedirPos);
-                    if (newlinePos == String::npos) {
-                        source.erase(linedirPos);
-                        break;
-                    }
-
-                    // Preserve a line break so adjacent preprocessor directives do not merge.
-                    source = source.replace(linedirPos, newlinePos - linedirPos + 1, "\n");
-                    linedirPos = source.find("#line", linedirPos);
-                }
-
-                // remove "noperspective"
-                const char* str_np = "noperspective";
-                const SizeT len_np = strlen(str_np);
-                SizeT noperspectivePos = source.find(str_np);
-                while (noperspectivePos != String::npos) {
-                    // + length of "\n"
-                    source = source.replace(noperspectivePos, len_np, "");
-                    noperspectivePos = source.find(str_np);
-                }
+                // noperspective is intentionally NOT touched here. It is core in desktop GLSL (1.30+)
+                // and maps to the core SPIR-V NoPerspective decoration, which DirectVulkan renders
+                // natively and SPIRV-Cross turns into ESSL `noperspective` + the
+                // GL_NV_shader_noperspective_interpolation extension. The old naked substring erase
+                // both discarded that interpolation (shader packs need it) and corrupted any
+                // identifier that merely contained the word. The GLES fallback for devices without
+                // the extension lives in the backend, where device capabilities are known.
 
                 FilterUnsupportedGpuShaderInt64(source);
                 CoerceUniformBlockPackingToStd140(source);
@@ -1035,12 +1290,7 @@ namespace MobileGL {
                 ModernizeLegacyGLSL(stage, source);
                 InjectDepthRangeBuiltinShim(stage, source);
 
-                const auto& activeBackend = MG_Backend::pActiveBackendObject;
-                if (stage == ShaderStage::Compute && activeBackend &&
-                    activeBackend->GetBackendType() == BackendType::DirectVulkan) {
-                    RewriteLinearSubgroupPrefixScanForVulkan(stage, activeBackend->GetDynamicParameters().SubgroupSize,
-                                                             source);
-                }
+                ApplyShaderSourceQuirks(stage, source);
             }
 
             Bool RetargetLegacyVersionDirectiveTo460(String& source) {
@@ -1049,6 +1299,10 @@ namespace MobileGL {
                 // must not be mistaken for the real one.
                 const ShaderLanguageInfo info = InspectShaderLanguage(source);
                 if (!info.HasVersionDirective()) return false;
+                // Never rescue a malformed directive to 460: that is precisely what re-legalized the
+                // CTS directive.version_* rejection cases after the first compile failed. The shader-
+                // pack retry this exists for only ever sees a valid low version (a real "#version 330").
+                if (!info.hasValidVersionDirective) return false;
                 // Only the set NormalizeVersionDirective downgraded: desktop core below 400. ES and
                 // compatibility shaders keep whatever they declared.
                 if (info.profile != ShaderProfile::Core || info.version >= 400) return false;

MobileGL/MG_Util/ShaderTranspiler/ShaderSourceProcessor.h

@@ -27,8 +27,10 @@ namespace MobileGL {
             // wider than the capture's 32 lanes. For the narrowly recognized, uniform-control-
             // flow template, replace the subgroup-local scan with a shared-memory, strict
             // left-fold over virtual 32-lane segments. Returns true only when the complete safe
-            // template was recognized and rewritten. DirectVulkan calls this through
-            // PreprocessShaderSource; the explicit entry point exists for deterministic tests.
+            // template was recognized and rewritten. PreprocessShaderSource reaches this through
+            // its device-quirk registry: by default only on detected Qualcomm Vulkan devices,
+            // overridable either way with MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN=1/0. The explicit
+            // entry point exists for deterministic tests.
             Bool RewriteLinearSubgroupPrefixScanForVulkan(ShaderStage stage, Uint32 nativeSubgroupSize, String& source);
 
             // Rewrites a "#version 330 core" directive that PreprocessShaderSource normalized down

MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DecoratePositionInvariantPass.cpp

@@ -0,0 +1,139 @@
+// MobileGL - MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DecoratePositionInvariantPass.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "DecoratePositionInvariantPass.h"
+
+#include "spirv.hpp"
+#include "source/opt/instruction.h"
+#include "source/opt/ir_context.h"
+#include "source/util/make_unique.h"
+
+#include <algorithm>
+#include <vector>
+
+namespace MobileGL {
+    namespace MG_Util {
+        namespace ShaderTranspiler {
+            namespace {
+                using spvtools::opt::Instruction;
+                using spvtools::opt::IRContext;
+                using spvtools::opt::Operand;
+
+                // Identifies one member of a decorated struct (gl_PerVertex's Position slot).
+                struct MemberKey {
+                    uint32_t id = 0;
+                    uint32_t member = 0;
+                    bool operator==(const MemberKey& other) const {
+                        return id == other.id && member == other.member;
+                    }
+                };
+
+                // OpDecorate      <target-id> <decoration> [literals...]
+                // OpMemberDecorate <struct-id> <member> <decoration> [literals...]
+                constexpr uint32_t kDecorateTargetOperand = 0;
+                constexpr uint32_t kDecorateDecorationOperand = 1;
+                constexpr uint32_t kDecorateBuiltInOperand = 2;
+                constexpr uint32_t kMemberDecorateStructOperand = 0;
+                constexpr uint32_t kMemberDecorateMemberOperand = 1;
+                constexpr uint32_t kMemberDecorateDecorationOperand = 2;
+                constexpr uint32_t kMemberDecorateBuiltInOperand = 3;
+            } // namespace
+
+            spvtools::opt::Pass::Status DecoratePositionInvariantPass::Process() {
+                auto* irContext = context();
+
+                // Collect first: AddAnnotationInst mutates the list being walked.
+                std::vector<uint32_t> invariantIds;
+                std::vector<MemberKey> invariantMembers;
+                std::vector<uint32_t> positionIds;
+                std::vector<MemberKey> positionMembers;
+
+                for (const Instruction& annotation : irContext->annotations()) {
+                    if (annotation.opcode() == spv::Op::OpDecorate) {
+                        if (annotation.NumInOperands() <= kDecorateDecorationOperand) {
+                            continue;
+                        }
+                        const auto decoration = static_cast<spv::Decoration>(
+                            annotation.GetSingleWordInOperand(kDecorateDecorationOperand));
+                        const uint32_t target = annotation.GetSingleWordInOperand(kDecorateTargetOperand);
+                        if (decoration == spv::Decoration::Invariant) {
+                            invariantIds.push_back(target);
+                        } else if (decoration == spv::Decoration::BuiltIn &&
+                                   annotation.NumInOperands() > kDecorateBuiltInOperand &&
+                                   static_cast<spv::BuiltIn>(annotation.GetSingleWordInOperand(
+                                       kDecorateBuiltInOperand)) == spv::BuiltIn::Position) {
+                            positionIds.push_back(target);
+                        }
+                    } else if (annotation.opcode() == spv::Op::OpMemberDecorate) {
+                        if (annotation.NumInOperands() <= kMemberDecorateDecorationOperand) {
+                            continue;
+                        }
+                        const auto decoration = static_cast<spv::Decoration>(
+                            annotation.GetSingleWordInOperand(kMemberDecorateDecorationOperand));
+                        const MemberKey key{
+                            annotation.GetSingleWordInOperand(kMemberDecorateStructOperand),
+                            annotation.GetSingleWordInOperand(kMemberDecorateMemberOperand)};
+                        if (decoration == spv::Decoration::Invariant) {
+                            invariantMembers.push_back(key);
+                        } else if (decoration == spv::Decoration::BuiltIn &&
+                                   annotation.NumInOperands() > kMemberDecorateBuiltInOperand &&
+                                   static_cast<spv::BuiltIn>(annotation.GetSingleWordInOperand(
+                                       kMemberDecorateBuiltInOperand)) == spv::BuiltIn::Position) {
+                            positionMembers.push_back(key);
+                        }
+                    }
+                }
+
+                Bool changed = false;
+
+                for (const uint32_t target : positionIds) {
+                    if (std::find(invariantIds.begin(), invariantIds.end(), target) != invariantIds.end()) {
+                        continue;
+                    }
+                    irContext->AddAnnotationInst(spvtools::MakeUnique<Instruction>(
+                        irContext, spv::Op::OpDecorate, 0, 0,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_ID, {target}},
+                            {SPV_OPERAND_TYPE_DECORATION,
+                             {static_cast<uint32_t>(spv::Decoration::Invariant)}}}));
+                    // Guard against a second Position decoration on the same target.
+                    invariantIds.push_back(target);
+                    changed = true;
+                }
+
+                for (const MemberKey& key : positionMembers) {
+                    if (std::find(invariantMembers.begin(), invariantMembers.end(), key) !=
+                        invariantMembers.end()) {
+                        continue;
+                    }
+                    irContext->AddAnnotationInst(spvtools::MakeUnique<Instruction>(
+                        irContext, spv::Op::OpMemberDecorate, 0, 0,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_ID, {key.id}},
+                            {SPV_OPERAND_TYPE_LITERAL_INTEGER, {key.member}},
+                            {SPV_OPERAND_TYPE_DECORATION,
+                             {static_cast<uint32_t>(spv::Decoration::Invariant)}}}));
+                    invariantMembers.push_back(key);
+                    changed = true;
+                }
+
+                if (!changed) {
+                    return Status::SuccessWithoutChange;
+                }
+
+                irContext->InvalidateAnalysesExceptFor(IRContext::kAnalysisNone);
+                return Status::SuccessWithChange;
+            }
+
+            spvtools::Optimizer::PassToken
+            DecoratePositionInvariantPass::CreateDecoratePositionInvariantPass() {
+                return spvtools::Optimizer::PassToken(MakeUnique<DecoratePositionInvariantPass>());
+            }
+        } // namespace ShaderTranspiler
+    } // namespace MG_Util
+} // namespace MobileGL

MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DecoratePositionInvariantPass.h

@@ -0,0 +1,35 @@
+// MobileGL - MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DecoratePositionInvariantPass.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include "source/opt/pass.h"
+#include "spirv-tools/optimizer.hpp"
+
+#include <Includes.h>
+
+namespace MobileGL {
+    namespace MG_Util {
+        namespace ShaderTranspiler {
+            // Adds the Invariant decoration to every Position builtin output. GL apps
+            // routinely rely on cross-program position invariance for multi-pass equality
+            // depth tests - MC 26.3's OIT re-draws the cloud geometry with GEQUAL against the
+            // depth its own first pass wrote - and a driver that optimizes each pipeline
+            // separately may otherwise vary the position math between passes, dropping whole
+            // primitives from the later ones. Both the plain (OpDecorate on a Position
+            // variable) and the block-member (OpMemberDecorate on gl_PerVertex) spellings are
+            // handled; targets that already carry Invariant are left alone. DirectVulkan only.
+            class DecoratePositionInvariantPass : public spvtools::opt::Pass {
+            public:
+                const char* name() const override { return "decorate-position-invariant"; }
+                Status Process() override;
+
+                static spvtools::Optimizer::PassToken CreateDecoratePositionInvariantPass();
+            };
+        } // namespace ShaderTranspiler
+    } // namespace MG_Util
+} // namespace MobileGL

MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/EmulateNoPerspectivePass.cpp

@@ -0,0 +1,407 @@
+// MobileGL - MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/EmulateNoPerspectivePass.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "EmulateNoPerspectivePass.h"
+
+#include "spirv.hpp"
+#include "source/opt/constants.h"
+#include "source/opt/def_use_manager.h"
+#include "source/opt/instruction.h"
+#include "source/opt/ir_context.h"
+#include "source/opt/module.h"
+#include "source/opt/type_manager.h"
+#include "source/opt/types.h"
+#include "source/util/make_unique.h"
+
+#include <algorithm>
+#include <vector>
+
+namespace MobileGL {
+    namespace MG_Util {
+        namespace ShaderTranspiler {
+            namespace {
+                using spvtools::opt::Instruction;
+                using spvtools::opt::IRContext;
+                using spvtools::opt::Operand;
+                namespace analysis = spvtools::opt::analysis;
+
+                spv::ExecutionModel EntryExecutionModel(IRContext* ctx) {
+                    for (Instruction& ep : ctx->module()->entry_points()) {
+                        return static_cast<spv::ExecutionModel>(ep.GetSingleWordInOperand(0));
+                    }
+                    return spv::ExecutionModel::Max;
+                }
+
+                uint32_t VariablePointeeType(IRContext* ctx, Instruction* var) {
+                    Instruction* ptrType = ctx->get_def_use_mgr()->GetDef(var->type_id());
+                    // OpTypePointer <storage-class> <pointee>
+                    return ptrType->GetSingleWordInOperand(1);
+                }
+
+                // If |typeId| is float or a vector of float, returns true and reports the scalar float
+                // type and whether it is a vector. Matrices, structs, ints etc. are not emulatable.
+                bool IsFloatScalarOrVector(IRContext* ctx, uint32_t typeId, uint32_t& floatTypeId, bool& isVector) {
+                    Instruction* t = ctx->get_def_use_mgr()->GetDef(typeId);
+                    if (t == nullptr) return false;
+                    if (t->opcode() == spv::Op::OpTypeFloat) {
+                        floatTypeId = typeId;
+                        isVector = false;
+                        return true;
+                    }
+                    if (t->opcode() == spv::Op::OpTypeVector) {
+                        const uint32_t comp = t->GetSingleWordInOperand(0);
+                        Instruction* ct = ctx->get_def_use_mgr()->GetDef(comp);
+                        if (ct != nullptr && ct->opcode() == spv::Op::OpTypeFloat) {
+                            floatTypeId = comp;
+                            isVector = true;
+                            return true;
+                        }
+                    }
+                    return false;
+                }
+
+                uint32_t PointerTypeTo(IRContext* ctx, uint32_t pointeeId, spv::StorageClass sc) {
+                    analysis::Type* pointee = ctx->get_type_mgr()->GetType(pointeeId);
+                    analysis::Pointer ptr(pointee, sc);
+                    return ctx->get_type_mgr()->GetTypeInstruction(&ptr);
+                }
+
+                uint32_t V4FloatType(IRContext* ctx) {
+                    analysis::Float f(32);
+                    analysis::Type* freg = ctx->get_type_mgr()->GetRegisteredType(&f);
+                    analysis::Vector v(freg, 4);
+                    return ctx->get_type_mgr()->GetTypeInstruction(&v);
+                }
+
+                uint32_t FloatType(IRContext* ctx) {
+                    analysis::Float f(32);
+                    return ctx->get_type_mgr()->GetTypeInstruction(&f);
+                }
+
+                uint32_t SignedIntConstant(IRContext* ctx, int32_t value) {
+                    analysis::Integer i(32, true);
+                    analysis::Type* reg = ctx->get_type_mgr()->GetRegisteredType(&i);
+                    const analysis::Constant* c =
+                        ctx->get_constant_mgr()->GetConstant(reg, {static_cast<uint32_t>(value)});
+                    return ctx->get_constant_mgr()->GetDefiningInstruction(c)->result_id();
+                }
+
+                // Multiply |valueId| (of type |valueTypeId|) by the scalar |scalarId|, inserting the op
+                // before |before|. Returns the product's id.
+                uint32_t InsertScale(IRContext* ctx, Instruction* before, uint32_t valueTypeId,
+                                     uint32_t valueId, uint32_t scalarId, bool isVector) {
+                    const uint32_t productId = ctx->TakeNextId();
+                    const spv::Op op = isVector ? spv::Op::OpVectorTimesScalar : spv::Op::OpFMul;
+                    before->InsertBefore(spvtools::MakeUnique<Instruction>(
+                        ctx, op, valueTypeId, productId,
+                        std::initializer_list<Operand>{{SPV_OPERAND_TYPE_ID, {valueId}},
+                                                       {SPV_OPERAND_TYPE_ID, {scalarId}}}));
+                    return productId;
+                }
+
+                // --- Vertex stage: gl_Position discovery ------------------------------------------
+
+                // Finds gl_Position as member |memberIndex| of a gl_PerVertex-style block whose Output
+                // variable is |blockVarId|; |v4floatTypeId| is that member's (vec4) type. Returns false
+                // if gl_Position is not a block member (older plain-variable form is left to the strip).
+                bool FindPositionBlock(IRContext* ctx, uint32_t& blockVarId, uint32_t& memberIndex,
+                                       uint32_t& v4floatTypeId) {
+                    uint32_t structId = 0;
+                    uint32_t member = 0;
+                    for (Instruction& ann : ctx->annotations()) {
+                        if (ann.opcode() == spv::Op::OpMemberDecorate && ann.NumInOperands() >= 4 &&
+                            static_cast<spv::Decoration>(ann.GetSingleWordInOperand(2)) ==
+                                spv::Decoration::BuiltIn &&
+                            static_cast<spv::BuiltIn>(ann.GetSingleWordInOperand(3)) ==
+                                spv::BuiltIn::Position) {
+                            structId = ann.GetSingleWordInOperand(0);
+                            member = ann.GetSingleWordInOperand(1);
+                            break;
+                        }
+                    }
+                    if (structId == 0) return false;
+
+                    Instruction* structType = ctx->get_def_use_mgr()->GetDef(structId);
+                    if (structType == nullptr || member >= structType->NumInOperands()) return false;
+                    v4floatTypeId = structType->GetSingleWordInOperand(member);
+
+                    for (Instruction& inst : ctx->module()->types_values()) {
+                        if (inst.opcode() == spv::Op::OpVariable &&
+                            static_cast<spv::StorageClass>(inst.GetSingleWordInOperand(0)) ==
+                                spv::StorageClass::Output &&
+                            VariablePointeeType(ctx, &inst) == structId) {
+                            blockVarId = inst.result_id();
+                            memberIndex = member;
+                            return true;
+                        }
+                    }
+                    return false;
+                }
+
+                // --- Fragment stage: gl_FragCoord discovery/synthesis -----------------------------
+
+                Instruction* FindBuiltinInput(IRContext* ctx, spv::BuiltIn builtin) {
+                    for (Instruction& ann : ctx->annotations()) {
+                        if (ann.opcode() != spv::Op::OpDecorate || ann.NumInOperands() < 3) continue;
+                        if (static_cast<spv::Decoration>(ann.GetSingleWordInOperand(1)) !=
+                            spv::Decoration::BuiltIn)
+                            continue;
+                        if (static_cast<spv::BuiltIn>(ann.GetSingleWordInOperand(2)) != builtin) continue;
+                        Instruction* var = ctx->get_def_use_mgr()->GetDef(ann.GetSingleWordInOperand(0));
+                        if (var != nullptr && var->opcode() == spv::Op::OpVariable &&
+                            static_cast<spv::StorageClass>(var->GetSingleWordInOperand(0)) ==
+                                spv::StorageClass::Input) {
+                            return var;
+                        }
+                    }
+                    return nullptr;
+                }
+
+                uint32_t SynthesizeFragCoord(IRContext* ctx, uint32_t v4floatTypeId) {
+                    const uint32_t ptrType = PointerTypeTo(ctx, v4floatTypeId, spv::StorageClass::Input);
+                    const uint32_t varId = ctx->TakeNextId();
+                    ctx->AddGlobalValue(spvtools::MakeUnique<Instruction>(
+                        ctx, spv::Op::OpVariable, ptrType, varId,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_STORAGE_CLASS,
+                             {static_cast<uint32_t>(spv::StorageClass::Input)}}}));
+                    ctx->AddAnnotationInst(spvtools::MakeUnique<Instruction>(
+                        ctx, spv::Op::OpDecorate, 0, 0,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_ID, {varId}},
+                            {SPV_OPERAND_TYPE_DECORATION,
+                             {static_cast<uint32_t>(spv::Decoration::BuiltIn)}},
+                            {SPV_OPERAND_TYPE_LITERAL_INTEGER,
+                             {static_cast<uint32_t>(spv::BuiltIn::FragCoord)}}}));
+                    for (Instruction& ep : ctx->module()->entry_points()) {
+                        ep.AddOperand({SPV_OPERAND_TYPE_ID, {varId}});
+                    }
+                    return varId;
+                }
+            } // namespace
+
+            spvtools::opt::Pass::Status EmulateNoPerspectivePass::Process() {
+                auto* ctx = context();
+                const spv::ExecutionModel model = EntryExecutionModel(ctx);
+                const bool isVertex = model == spv::ExecutionModel::Vertex;
+                const bool isFragment = model == spv::ExecutionModel::Fragment;
+
+                // Collect NoPerspective-decorated plain variables and every NoPerspective annotation.
+                std::vector<uint32_t> plainVarIds;
+                std::vector<Instruction*> decorationsToKill;
+                for (Instruction& ann : ctx->annotations()) {
+                    if (ann.opcode() == spv::Op::OpDecorate && ann.NumInOperands() >= 2 &&
+                        static_cast<spv::Decoration>(ann.GetSingleWordInOperand(1)) ==
+                            spv::Decoration::NoPerspective) {
+                        plainVarIds.push_back(ann.GetSingleWordInOperand(0));
+                        decorationsToKill.push_back(&ann);
+                    } else if (ann.opcode() == spv::Op::OpMemberDecorate && ann.NumInOperands() >= 3 &&
+                               static_cast<spv::Decoration>(ann.GetSingleWordInOperand(2)) ==
+                                   spv::Decoration::NoPerspective) {
+                        // Block-member noperspective is not emulated here; the decoration is stripped
+                        // (smooth fallback) so SPIRV-Cross does not require the NV extension.
+                        decorationsToKill.push_back(&ann);
+                    }
+                }
+
+                if (decorationsToKill.empty()) {
+                    return Status::SuccessWithoutChange;
+                }
+
+                const spv::StorageClass wantStorage =
+                    isVertex ? spv::StorageClass::Output : spv::StorageClass::Input;
+
+                // Emulatable = plain variable of the stage's interface direction, float or floatN.
+                struct Target {
+                    Instruction* var;
+                    uint32_t typeId;
+                    uint32_t floatTypeId;
+                    bool isVector;
+                };
+                std::vector<Target> targets;
+                if (isVertex || isFragment) {
+                    for (const uint32_t id : plainVarIds) {
+                        Instruction* var = ctx->get_def_use_mgr()->GetDef(id);
+                        if (var == nullptr || var->opcode() != spv::Op::OpVariable) continue;
+                        if (static_cast<spv::StorageClass>(var->GetSingleWordInOperand(0)) != wantStorage)
+                            continue;
+                        const uint32_t pointee = VariablePointeeType(ctx, var);
+                        uint32_t floatTypeId = 0;
+                        bool isVector = false;
+                        if (IsFloatScalarOrVector(ctx, pointee, floatTypeId, isVector)) {
+                            targets.push_back({var, pointee, floatTypeId, isVector});
+                        }
+                    }
+                }
+
+                // Force highp on the varyings we emulate: the a*w round-trip overflows a mediump (fp16)
+                // varying at large clip-space w. Dropping RelaxedPrecision makes SPIRV-Cross emit them
+                // highp on both stages, keeping the emulation exact. Only touches emulated variables.
+                if (!targets.empty()) {
+                    std::vector<uint32_t> targetIds;
+                    targetIds.reserve(targets.size());
+                    for (const Target& t : targets) targetIds.push_back(t.var->result_id());
+                    for (Instruction& ann : ctx->annotations()) {
+                        if (ann.opcode() == spv::Op::OpDecorate && ann.NumInOperands() >= 2 &&
+                            static_cast<spv::Decoration>(ann.GetSingleWordInOperand(1)) ==
+                                spv::Decoration::RelaxedPrecision &&
+                            std::find(targetIds.begin(), targetIds.end(),
+                                      ann.GetSingleWordInOperand(0)) != targetIds.end()) {
+                            decorationsToKill.push_back(&ann);
+                        }
+                    }
+                }
+
+                if (isVertex && !targets.empty()) {
+                    uint32_t blockVarId = 0;
+                    uint32_t memberIndex = 0;
+                    uint32_t v4floatTypeId = 0;
+                    if (FindPositionBlock(ctx, blockVarId, memberIndex, v4floatTypeId)) {
+                        const uint32_t ptrOutV4 =
+                            PointerTypeTo(ctx, v4floatTypeId, spv::StorageClass::Output);
+                        const uint32_t memberConst = SignedIntConstant(ctx, static_cast<int32_t>(memberIndex));
+                        const uint32_t floatTy = FloatType(ctx);
+
+                        uint32_t entryFuncId = 0;
+                        for (Instruction& ep : ctx->module()->entry_points()) {
+                            // OpEntryPoint <model> <function> "name" <interface...>
+                            entryFuncId = ep.GetSingleWordInOperand(1);
+                            break;
+                        }
+
+                        // Pre-multiply every target output by gl_Position.w before each return of the
+                        // ENTRY function only. glslang does not inline, so a called helper survives as
+                        // its own OpFunction; instrumenting its returns too would scale the varying
+                        // more than once (w^2), breaking the identity.
+                        for (auto funcIt = ctx->module()->begin(); funcIt != ctx->module()->end(); ++funcIt) {
+                            if (funcIt->result_id() != entryFuncId) continue;
+                            funcIt->ForEachInst([&](Instruction* inst) {
+                                if (inst->opcode() != spv::Op::OpReturn &&
+                                    inst->opcode() != spv::Op::OpReturnValue) {
+                                    return;
+                                }
+                                const uint32_t posPtrId = ctx->TakeNextId();
+                                inst->InsertBefore(spvtools::MakeUnique<Instruction>(
+                                    ctx, spv::Op::OpAccessChain, ptrOutV4, posPtrId,
+                                    std::initializer_list<Operand>{{SPV_OPERAND_TYPE_ID, {blockVarId}},
+                                                                   {SPV_OPERAND_TYPE_ID, {memberConst}}}));
+                                const uint32_t posId = ctx->TakeNextId();
+                                inst->InsertBefore(spvtools::MakeUnique<Instruction>(
+                                    ctx, spv::Op::OpLoad, v4floatTypeId, posId,
+                                    std::initializer_list<Operand>{{SPV_OPERAND_TYPE_ID, {posPtrId}}}));
+                                const uint32_t wId = ctx->TakeNextId();
+                                inst->InsertBefore(spvtools::MakeUnique<Instruction>(
+                                    ctx, spv::Op::OpCompositeExtract, floatTy, wId,
+                                    std::initializer_list<Operand>{{SPV_OPERAND_TYPE_ID, {posId}},
+                                                                   {SPV_OPERAND_TYPE_LITERAL_INTEGER, {3u}}}));
+                                for (const Target& t : targets) {
+                                    const uint32_t valId = ctx->TakeNextId();
+                                    inst->InsertBefore(spvtools::MakeUnique<Instruction>(
+                                        ctx, spv::Op::OpLoad, t.typeId, valId,
+                                        std::initializer_list<Operand>{
+                                            {SPV_OPERAND_TYPE_ID, {t.var->result_id()}}}));
+                                    const uint32_t scaledId =
+                                        InsertScale(ctx, inst, t.typeId, valId, wId, t.isVector);
+                                    inst->InsertBefore(spvtools::MakeUnique<Instruction>(
+                                        ctx, spv::Op::OpStore, 0, 0,
+                                        std::initializer_list<Operand>{
+                                            {SPV_OPERAND_TYPE_ID, {t.var->result_id()}},
+                                            {SPV_OPERAND_TYPE_ID, {scaledId}}}));
+                                }
+                            });
+                        }
+                    }
+                }
+
+                if (isFragment && !targets.empty()) {
+                    Instruction* fragCoord = FindBuiltinInput(ctx, spv::BuiltIn::FragCoord);
+                    uint32_t fragCoordId = 0;
+                    uint32_t v4floatTypeId = 0;
+                    if (fragCoord != nullptr) {
+                        fragCoordId = fragCoord->result_id();
+                        v4floatTypeId = VariablePointeeType(ctx, fragCoord);
+                    } else {
+                        v4floatTypeId = V4FloatType(ctx);
+                        fragCoordId = SynthesizeFragCoord(ctx, v4floatTypeId);
+                    }
+                    const uint32_t floatTy = FloatType(ctx);
+
+                    auto* defUse = ctx->get_def_use_mgr();
+                    for (const Target& t : targets) {
+                        // Collect every load that reads the varying. glslang lowers a whole-variable
+                        // read to OpLoad(var), but a single-component read (v.x) to
+                        // OpAccessChain(var) + OpLoad(chain). Both must be scaled; the identity is
+                        // per-component, so scaling one loaded component by gl_FragCoord.w is valid.
+                        std::vector<Instruction*> loads;
+                        defUse->ForEachUser(t.var, [&](Instruction* user) {
+                            if (user->opcode() == spv::Op::OpLoad &&
+                                user->GetSingleWordInOperand(0) == t.var->result_id()) {
+                                loads.push_back(user);
+                            } else if (user->opcode() == spv::Op::OpAccessChain &&
+                                       user->GetSingleWordInOperand(0) == t.var->result_id()) {
+                                const uint32_t chainId = user->result_id();
+                                defUse->ForEachUser(user, [&](Instruction* chainUser) {
+                                    if (chainUser->opcode() == spv::Op::OpLoad &&
+                                        chainUser->GetSingleWordInOperand(0) == chainId) {
+                                        loads.push_back(chainUser);
+                                    }
+                                });
+                            }
+                        });
+
+                        // Rewrite `%r = OpLoad %ty %ptr` into
+                        //   %orig = OpLoad %ty %ptr
+                        //   %fc   = OpLoad %v4float %fragCoord
+                        //   %w    = OpCompositeExtract %float %fc 3
+                        //   %r    = OpVectorTimesScalar/OpFMul %ty %orig %w   (reuse %r: uses stay intact)
+                        // The op is chosen from the LOAD's own result type: a whole-vector load scales
+                        // with OpVectorTimesScalar, a scalar component load with OpFMul.
+                        for (Instruction* load : loads) {
+                            const uint32_t loadType = load->type_id();
+                            uint32_t componentFloat = 0;
+                            bool loadIsVector = false;
+                            if (!IsFloatScalarOrVector(ctx, loadType, componentFloat, loadIsVector)) {
+                                continue;
+                            }
+                            const uint32_t ptrId = load->GetSingleWordInOperand(0);
+                            const uint32_t origId = ctx->TakeNextId();
+                            load->InsertBefore(spvtools::MakeUnique<Instruction>(
+                                ctx, spv::Op::OpLoad, loadType, origId,
+                                std::initializer_list<Operand>{{SPV_OPERAND_TYPE_ID, {ptrId}}}));
+                            const uint32_t fcId = ctx->TakeNextId();
+                            load->InsertBefore(spvtools::MakeUnique<Instruction>(
+                                ctx, spv::Op::OpLoad, v4floatTypeId, fcId,
+                                std::initializer_list<Operand>{{SPV_OPERAND_TYPE_ID, {fragCoordId}}}));
+                            const uint32_t wId = ctx->TakeNextId();
+                            load->InsertBefore(spvtools::MakeUnique<Instruction>(
+                                ctx, spv::Op::OpCompositeExtract, floatTy, wId,
+                                std::initializer_list<Operand>{{SPV_OPERAND_TYPE_ID, {fcId}},
+                                                               {SPV_OPERAND_TYPE_LITERAL_INTEGER, {3u}}}));
+                            load->SetOpcode(loadIsVector ? spv::Op::OpVectorTimesScalar : spv::Op::OpFMul);
+                            load->SetInOperands(Instruction::OperandList{
+                                {SPV_OPERAND_TYPE_ID, {origId}}, {SPV_OPERAND_TYPE_ID, {wId}}});
+                        }
+                    }
+                }
+
+                // Strip every NoPerspective decoration: emulated varyings now transport smooth, and
+                // non-emulatable ones fall back to smooth.
+                for (Instruction* dec : decorationsToKill) {
+                    ctx->KillInst(dec);
+                }
+
+                ctx->InvalidateAnalysesExceptFor(IRContext::kAnalysisNone);
+                return Status::SuccessWithChange;
+            }
+
+            spvtools::Optimizer::PassToken EmulateNoPerspectivePass::CreateEmulateNoPerspectivePass() {
+                return spvtools::Optimizer::PassToken(MakeUnique<EmulateNoPerspectivePass>());
+            }
+        } // namespace ShaderTranspiler
+    } // namespace MG_Util
+} // namespace MobileGL

MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/EmulateNoPerspectivePass.h

@@ -0,0 +1,41 @@
+// MobileGL - MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/EmulateNoPerspectivePass.h
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#pragma once
+#include "source/opt/pass.h"
+#include "spirv-tools/optimizer.hpp"
+
+#include <Includes.h>
+
+namespace MobileGL {
+    namespace MG_Util {
+        namespace ShaderTranspiler {
+            // Emulates 'noperspective' (screen-linear) interpolation on GLES devices that lack
+            // GL_NV_shader_noperspective_interpolation, so no NV extension is required. The hardware
+            // interpolates perspective-correct; screen-linear L(a) is recovered from the identity
+            //     L(a) = P(a * w) * gl_FragCoord.w
+            // where P is perspective-correct interpolation and w is the vertex clip-space w. So each
+            // NoPerspective-decorated output is pre-multiplied by gl_Position.w in the vertex stage
+            // and each NoPerspective-decorated input is multiplied by gl_FragCoord.w in the fragment
+            // stage; the decoration is then removed so the varying transports smooth. This is exact
+            // (modulo float precision - the emulated varyings want highp).
+            //
+            // Scope: plain interface variables of float or floatN type. Anything it cannot emulate
+            // (interface-block members, matrices, or a stage lacking the needed builtin) has its
+            // NoPerspective decoration stripped instead, degrading to smooth - the same result the
+            // extension-less fallback produced before, and never invalid SPIR-V. DirectGLES only.
+            class EmulateNoPerspectivePass : public spvtools::opt::Pass {
+            public:
+                const char* name() const override { return "emulate-noperspective"; }
+                Status Process() override;
+
+                static spvtools::Optimizer::PassToken CreateEmulateNoPerspectivePass();
+            };
+        } // namespace ShaderTranspiler
+    } // namespace MG_Util
+} // namespace MobileGL

MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/StripNoPerspectivePass.cpp

@@ -0,0 +1,72 @@
+// MobileGL - MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/StripNoPerspectivePass.cpp
+// Copyright (c) 2025-2026 MobileGL-Dev
+// Licensed under the GNU Lesser General Public License v3.0:
+//   https://www.gnu.org/licenses/gpl-3.0.txt
+//   https://www.gnu.org/licenses/lgpl-3.0.txt
+// SPDX-License-Identifier: LGPL-3.0-only
+// End of Source File Header
+
+#include "StripNoPerspectivePass.h"
+
+#include "spirv.hpp"
+#include "source/opt/instruction.h"
+#include "source/opt/ir_context.h"
+#include "source/util/make_unique.h"
+
+#include <vector>
+
+namespace MobileGL {
+    namespace MG_Util {
+        namespace ShaderTranspiler {
+            namespace {
+                using spvtools::opt::Instruction;
+                using spvtools::opt::IRContext;
+
+                // OpDecorate       <target-id> <decoration> [literals...]
+                // OpMemberDecorate <struct-id>  <member> <decoration> [literals...]
+                constexpr uint32_t kDecorateDecorationOperand = 1;
+                constexpr uint32_t kMemberDecorateDecorationOperand = 2;
+            } // namespace
+
+            spvtools::opt::Pass::Status StripNoPerspectivePass::Process() {
+                auto* irContext = context();
+
+                // Collect first: KillInst mutates the annotation list being walked.
+                std::vector<Instruction*> toKill;
+                for (Instruction& annotation : irContext->annotations()) {
+                    uint32_t decorationOperand = 0;
+                    if (annotation.opcode() == spv::Op::OpDecorate) {
+                        decorationOperand = kDecorateDecorationOperand;
+                    } else if (annotation.opcode() == spv::Op::OpMemberDecorate) {
+                        decorationOperand = kMemberDecorateDecorationOperand;
+                    } else {
+                        continue;
+                    }
+
+                    if (annotation.NumInOperands() <= decorationOperand) {
+                        continue;
+                    }
+                    if (static_cast<spv::Decoration>(annotation.GetSingleWordInOperand(decorationOperand)) ==
+                        spv::Decoration::NoPerspective) {
+                        toKill.push_back(&annotation);
+                    }
+                }
+
+                if (toKill.empty()) {
+                    return Status::SuccessWithoutChange;
+                }
+
+                for (Instruction* inst : toKill) {
+                    irContext->KillInst(inst);
+                }
+
+                irContext->InvalidateAnalysesExceptFor(IRContext::kAnalysisNone);
+                return Status::SuccessWithChange;
+            }
+
+            spvtools::Optimizer::PassToken StripNoPerspectivePass::CreateStripNoPerspectivePass() {
+                return spvtools::Optimizer::PassToken(MakeUnique<StripNoPerspectivePass>());
+            }
+        } // namespace ShaderTranspiler
+    } // namespace MG_Util
+} // namespace MobileGL