[Fix] (Diligent): fix y inversion and front-face winding in viewport mapping

- Translate OpenGL bottom-left viewport/scissor rects to Diligent top-left origin
- Invert FrontCounterClockwise to compensate for the Y conversion
- Add DrawsTopHalfTriangleFromMobileGLState y-orientation test

.github/workflows/apk.yml

@@ -209,7 +209,7 @@ jobs:
       - name: Load trace cases
         id: trace-cases
         run: |
-          echo "android=$(python3 tools/trace_replay/trace_cases.py --ci --format github-apk)" >> "$GITHUB_OUTPUT"
+          echo "android=$(python3 tools/trace_replay/trace_cases.py --ci --format github-apk-matrix)" >> "$GITHUB_OUTPUT"
           echo "names=$(python3 tools/trace_replay/trace_cases.py --ci --format names)" >> "$GITHUB_OUTPUT"
 
   trace-fixtures:
@@ -337,13 +337,7 @@ jobs:
     strategy:
       fail-fast: false
       max-parallel: 4
-      matrix:
-        backend:
-          - name: DirectGLES
-            gpu: software
-          - name: DirectVulkan
-            gpu: lavapipe
-        case: ${{ fromJSON(needs.trace-cases.outputs.android) }}
+      matrix: ${{ fromJSON(needs.trace-cases.outputs.android) }}
     steps:
       - name: Set Swap Space
         uses: pierotofy/set-swap-space@v1.0

.github/workflows/test.yml

@@ -491,6 +491,7 @@ jobs:
       - benchmark
       - integration
     outputs:
+      matrix: ${{ steps.trace-cases.outputs.matrix }}
       names: ${{ steps.trace-cases.outputs.names }}
     steps:
       - name: Checkout repo
@@ -498,7 +499,9 @@ jobs:
 
       - name: Load trace cases
         id: trace-cases
-        run: echo "names=$(python3 tools/trace_replay/trace_cases.py --ci --format names)" >> "$GITHUB_OUTPUT"
+        run: |
+          echo "matrix=$(python3 tools/trace_replay/trace_cases.py --ci --format github-test-matrix)" >> "$GITHUB_OUTPUT"
+          echo "names=$(python3 tools/trace_replay/trace_cases.py --ci --format names)" >> "$GITHUB_OUTPUT"
 
   trace-fixtures:
     name: trace fixture (${{ matrix.case }})
@@ -577,11 +580,7 @@ jobs:
     strategy:
       fail-fast: false
       max-parallel: 4
-      matrix:
-        backend:
-          - DirectGLES
-          - DirectVulkan
-        case: ${{ fromJSON(needs.trace-cases.outputs.names) }}
+      matrix: ${{ fromJSON(needs.trace-cases.outputs.matrix) }}
 
     steps:
       - name: Set Swap Space

.gitignore

@@ -1,4 +1,4 @@
-################################################################################
+################################################################################
 # 此 .gitignore 文件已由 Microsoft(R) Visual Studio 自动创建。
 ################################################################################
 
@@ -16,6 +16,9 @@ MobileGLCodeManager
 MobileGL/MG_Test/build
 /build_*
 /cmake-build*
+/build-*/
+/local.properties
+/.jspace/
 .idea
 MobileGL/MG*/build*
 MobileGL/MG*/cmake-build*

CMakeLists.txt

@@ -182,6 +182,7 @@ set(ENABLE_SPVREMAPPER         OFF CACHE BOOL "Enable SPVRemapper" FORCE)
 set(ENABLE_OPT                 ON  CACHE BOOL "Enable SPIRV-Tools opt usage in glslang" FORCE)
 set(BUILD_EXTERNAL             ON  CACHE BOOL "Build external deps in External/" FORCE)
 set(ENABLE_GLSLANG_INSTALL     OFF CACHE BOOL "Install glslang targets" FORCE)
+set(SPIRV_SKIP_EXECUTABLES     ON  CACHE BOOL "Skip building SPIRV-Tools executables" FORCE)
 
 set(SPIRV_CROSS_C_API          ON  CACHE BOOL "Enable C API" FORCE)
 set(SPIRV_CROSS_ENABLE_GLSL    ON  CACHE BOOL "Enable GLSL backend" FORCE)
@@ -198,7 +199,6 @@ set(SPIRV_REFLECT_ENABLE_ASSERTS OFF CACHE BOOL  "Enable asserts for debugging"
 set(SPIRV_REFLECT_ENABLE_ASAN    OFF CACHE BOOL "Use address sanitization" FORCE)
 set(SPIRV_REFLECT_INSTALL        OFF CACHE BOOL "Whether to install" FORCE)
 
-# add_subdirectory(3rdparty/DiligentCore)
 add_subdirectory(3rdparty/glslang)
 add_subdirectory(3rdparty/SPIRV-Cross)
 add_subdirectory(3rdparty/VulkanMemoryAllocator)
@@ -210,6 +210,23 @@ set(XXHASH_BUILD_XXHSUM OFF)
 option(BUILD_SHARED_LIBS OFF)
 add_subdirectory(3rdparty/xxHash/build/cmake xxhash_build EXCLUDE_FROM_ALL)
 
+# Diligent-based backend. Enabled by default on local builds; only the Vulkan
+# engine from DiligentCore is built. Added after the other 3rdparty projects so
+# DiligentCore reuses the glslang / SPIRV-Cross / SPIRV-Tools / xxHash targets
+# already defined by MobileGL instead of building its bundled copies.
+option(MOBILEGL_ENABLE_DILIGENT "Enable the Diligent/Vulkan backend" ON)
+if(MOBILEGL_ENABLE_DILIGENT)
+    set(DILIGENT_NO_DIRECT3D11 ON CACHE BOOL "Disable Direct3D11 backend" FORCE)
+    set(DILIGENT_NO_DIRECT3D12 ON CACHE BOOL "Disable Direct3D12 backend" FORCE)
+    set(DILIGENT_NO_OPENGL     ON CACHE BOOL "Disable OpenGL backend" FORCE)
+    set(DILIGENT_NO_METAL      ON CACHE BOOL "Disable Metal backend" FORCE)
+    set(DILIGENT_NO_WEBGPU     ON CACHE BOOL "Disable WebGPU backend" FORCE)
+    set(DILIGENT_NO_ARCHIVER   ON CACHE BOOL "Disable Archiver" FORCE)
+    set(DILIGENT_BUILD_TESTS   OFF CACHE BOOL "Build Diligent tests" FORCE)
+    set(DILIGENT_INSTALL_CORE  OFF CACHE BOOL "Install DiligentCore" FORCE)
+    add_subdirectory(3rdparty/DiligentCore)
+endif()
+
 set(TRACY_ENABLE ${MOBILEGL_ENABLE_TRACY} CACHE BOOL "Enable Tracy, this is an internal variable" FORCE)
 
 if (TRACY_ENABLE)
@@ -395,6 +412,14 @@ set(SOURCE_FILES
     MobileGL/MG_State/GLState/RenderbufferState/RenderbufferState.cpp
 )
 
+if(MOBILEGL_ENABLE_DILIGENT)
+    list(APPEND SOURCE_FILES
+        MobileGL/MG_Backend/Diligent/BackendObject_Diligent.cpp
+        MobileGL/MG_Backend/Diligent/DiligentVulkan.cpp
+        MobileGL/MG_Backend/Diligent/Renderer/DiligentRenderer.cpp
+    )
+endif()
+
 if (APPLE AND NOT MOBILEGL_IOS)
     list(APPEND SOURCE_FILES
         MobileGL/MG_Impl/CGLImpl/CGLImpl.cpp
@@ -435,6 +460,21 @@ set(MOBILEGL_LINK_LIBRARIES
         Threads::Threads
 )
 
+if(MOBILEGL_ENABLE_DILIGENT)
+    list(APPEND MOBILEGL_LINK_LIBRARIES
+        Diligent-GraphicsEngineVk-static
+        Diligent-GraphicsEngine
+        Diligent-GraphicsEngineNextGenBase
+        Diligent-GraphicsAccessories
+        Diligent-ShaderTools
+        Diligent-GraphicsTools
+        Diligent-Common
+        Diligent-Primitives
+        Diligent-TargetPlatform
+        Vulkan::Headers
+    )
+endif()
+
 set(MOBILEGL_COMPILE_DEF
         -DVMA_STATIC_VULKAN_FUNCTIONS=0
         -DVMA_DYNAMIC_VULKAN_FUNCTIONS=1
@@ -500,6 +540,7 @@ target_compile_definitions(${CMAKE_PROJECT_NAME}
         ${MOBILEGL_COMPILE_DEF}
         MOBILEGL_LOG_ACTIVE_LEVEL=${MOBILEGL_LOG_ACTIVE_LEVEL}
         $<$<BOOL:${MOBILEGL_TRACE_ANGLE_VARIANTS}>:MOBILEGL_TRACE_ANGLE_VARIANTS=1>
+        $<$<BOOL:${MOBILEGL_ENABLE_DILIGENT}>:MOBILEGL_ENABLE_DILIGENT=1>
 )
 
 if(UNIX AND NOT APPLE AND NOT ANDROID)
@@ -558,6 +599,7 @@ if(NOT ANDROID)
             PUBLIC
             ${MOBILEGL_COMPILE_DEF}
             MOBILEGL_LOG_ACTIVE_LEVEL=${MOBILEGL_LOG_ACTIVE_LEVEL}
+            $<$<BOOL:${MOBILEGL_ENABLE_DILIGENT}>:MOBILEGL_ENABLE_DILIGENT=1>
     )
 endif()
 

HANDOFF_DILIGENT.md

@@ -0,0 +1,278 @@
+# Handoff: Diligent/Vulkan GL3.2 Backend for MobileGL
+
+Date: 2026-08-18  
+Branch: `feat/diligent-vulkan-backend`  
+Repo: `~/MobileGL-dev`  
+Status: **Active work-in-progress. Do not mark complete yet.**
+
+---
+
+## 1. Goal
+
+Implement a complete OpenGL 3.2 front-end emulation on a new Diligent/Vulkan backend inside MobileGL, instead of the DirectVulkan / DirectGLES backends.
+
+Target state:
+- Fully wire MobileGL front-end `MG_State` (buffers, VAO, program, texture, sampler, framebuffer, render-state) into Diligent.
+- Implement all GL 3.2 core entry points through the Diligent backend.
+- Pass local non-Android GL3.2 tests on the Turnip Adreno 750 GPU.
+
+---
+
+## 2. Current Branch / Commits
+
+Latest 12 commits on `feat/diligent-vulkan-backend`:
+
+```
+2f5abf83 test(diligent): verify indexed DrawElements path from real frontend state
+c31b7381 feat(diligent): add basic texture binding and textured state-draw test
+8945c507 feat(diligent): clear depth in GL Clear when GL_DEPTH_BUFFER_BIT set
+558d3aea feat(diligent): add offscreen depth target and depth clear
+7e2f0bc8 feat(diligent): wire stencil and color-mask state into state PSO
+f99786f6 feat(diligent): wire viewport/scissor state into state draws
+be4cc3ce feat(diligent): wire blend/depth/cull render state into state PSO
+c855e6cf feat(diligent): verify state-driven draw with real MobileGL frontend state
+02e60bfa feat(diligent): add state-driven draw path (VAO/buffer/program to Diligent)
+a9515c92 feat(diligent): add dynamic vertex buffer upload path
+2f57582a feat(diligent): wire Clear/Draw/Present into GLFunctionsTable
+beb21123 feat(diligent): add real offscreen renderer with clear and triangle draw
+```
+
+Working tree is clean.
+
+---
+
+## 3. Key Files
+
+### Backend core
+
+- `MobileGL/MG_Backend/Diligent/BackendObject_Diligent.h/.cpp`
+  - `BackendObject_Diligent`
+  - Creates Diligent Vulkan device/context
+  - Owns `DiligentRenderer`
+  - Wires `GLFunctionsTable`:
+    - `Clear` (color + depth)
+    - `DrawArrays`
+    - `DrawElements`
+    - `Present`
+- `MobileGL/MG_Backend/Diligent/DiligentVulkan.h/.cpp`
+  - Backend identity helper / translation unit
+- `MobileGL/MG_Backend/Diligent/Renderer/DiligentRenderer.h/.cpp`
+  - Offscreen RGBA8 + D32F targets
+  - Clear / ClearDepth / DrawTriangle / DrawVertices
+  - `CreateTestTexture` (RGBA8 texture + SRV + sampler)
+  - `DrawFromState` (main front-end emulation draw path)
+  - `CreatePipelineFromState`:
+    - SPIR-V → Diligent shaders via SPIRV-Reflect
+    - VAO attributes → input layout
+    - primitive topology from GL mode
+    - blend / depth / cull / stencil / color-mask state
+  - `UploadVertexDataFromState`:
+    - packs enabled VAO attributes from `BufferObject` into interleaved vertex buffer
+    - supports `DrawArrays`, `DrawElements`, triangle-fan and line-loop expansion
+  - Static texture binding to `g_Texture` through PSO static variables + SRB
+
+### Integration changes
+
+- `CMakeLists.txt`
+  - New option `MOBILEGL_ENABLE_DILIGENT` (default ON for local)
+  - DiligentCore added **after** glslang/SPIRV-Cross/xxHash/Vulkan-Headers so it reuses existing CMake targets
+  - Diligent static libraries linked into `MobileGL` / `MobileGL_s`
+  - New Diligent backend sources added
+- `MobileGL/MG_Backend/BackendObject.h`
+  - New `BackendType::DiligentVulkan`
+- `MobileGL/MG_Backend/Init.cpp`
+  - New backend switch case
+- `MobileGL/ConfigLoader.cpp`
+  - `MOBILEGL_BACKEND_TYPE=DiligentVulkan` accepted
+- `MobileGL/MG_Test/CMakeLists.txt`
+  - New `MobileGL/MG_Test/Backend/Diligent` subdirectory
+- `MobileGL/MG_Test/Backend/Diligent/`
+  - `CMakeLists.txt`
+  - `SanityTest.cpp`
+
+### Local test files
+
+- `MobileGL/MG_Test/Backend/Diligent/SanityTest.cpp`
+  - `CreatesDiligentDeviceAndAdvertisesGL32`
+  - `ClearsAndDrawsTriangleOffscreen`
+  - `DrawsFromMobileGLState`
+  - `DrawsTexturedFromMobileGLState`
+  - `DrawsIndexedFromMobileGLState`
+  - `DrawsRealTexturedFromMobileGLState`
+  - `DrawsUniformFromMobileGLState`
+  - `DrawsToOffscreenFramebufferFromMobileGLState`
+  - `DrawsWithScissorFromMobileGLState`
+  - `DrawsWithBlendFromMobileGLState`
+  - `DrawsWithDepthTestFromMobileGLState`
+  - `DrawsNamedUniformBlockFromMobileGLState`
+  - `DrawsWithStencilTestFromMobileGLState`
+  - `DrawsToRenderbufferFramebufferFromMobileGLState`
+  - `DrawsToMultipleColorAttachmentsFromMobileGLState`
+  - `DrawsIndexedBaseVertexFromMobileGLState`
+
+---
+
+## 4. What Works Today
+
+Verified locally on Turnip Adreno 750:
+
+- Diligent device/context creation
+- EGL window-surface swapchain creation path through Diligent `ISwapChain` (offscreen tests still use the offscreen target)
+- GL 3.2 / GLSL 1.50 capability advertisement
+- Offscreen color + depth rendering
+- Clear color and depth
+- Real mobilegl front-end state-driven drawing:
+  - Program SPIR-V → Diligent shaders
+  - VAO attributes + bound GL buffer → interleaved vertex buffer
+  - `DrawArrays` path
+  - `DrawElements` path (index buffer)
+- Texture basics:
+  - Offscreen texture creation
+  - CPU → Diligent texture (`CreateTestTexture`)
+  - Static sampler2D binding to `g_Texture`
+  - Textured draw test passes
+- Render state:
+  - Blend enable/factors/equations
+  - Stencil clear + test enabled on a D24S8 default depth/stencil target
+  - Depth test enable/func/write mask
+  - Cull face enable/mode/front-face winding
+  - Stencil test enable/masks/ops/func/ref
+  - Color write mask
+  - Viewport
+  - Scissor rect
+- Texture/sampler full integration:
+  - `ITextureObject` → Diligent `ITexture` + SRV with automatic dirty upload
+  - `SamplerObject` / texture-object sampler → Diligent `ISampler`
+  - Real front-end `glTexImage2D` path (not only `CreateTestTexture`) verified
+- Global UBO upload:
+  - Front-end `glUniform*` shadow → Diligent uniform buffer bound as `MGL_GLOBAL_UBO`
+- User framebuffer mapping:
+  - Current draw/read FBO resolves texture attachments to Diligent RTV/DSV
+  - `ReadPixels` can read back from a user FBO color attachment
+- More GL entry points wired:
+  - `DrawRangeElements` / `DrawRangeElementsBaseVertex`
+  - `DrawElementsBaseVertex` with real baseVertex selection
+  - `MultiDrawArrays` / `MultiDrawElements` / `MultiDrawElementsBaseVertex`
+  - `DrawArraysInstanced` / `DrawElementsInstanced` family
+  - Indirect draw CPU fallback: `DrawArraysIndirect`, `DrawElementsIndirect`, `MultiDraw*Indirect`, `*IndirectCount`
+  - `ClearBufferfv` / `ClearBufferfi` / `ClearBufferiv` / `ClearBufferuiv` (incl. stencil clear)
+  - `BlitFramebuffer` / `BlitNamedFramebuffer` (same-size color copy between read/draw FBOs)
+  - `CopyTexImage2D` / `CopyTexSubImage2D` (whole-color copy fallback)
+  - `CopyImageSubData` (whole-texture copy between two texture objects)
+  - `GenerateMipmap` (Diligent GPU mip generation on state textures)
+  - `GetTexImage` / `GetTextureImage` (RGBA8 readback)
+  - Fence sync entries (`FenceSync` / `ClientWaitSync` / `WaitSync` / `DeleteSync` / `GetSyncStatus`) as CPU always-signaled fallback
+  - Timer query entries (`BeginTimeElapsedQuery` / `EndTimeElapsedQuery` / `QueryCounterTimestamp` / `GetQueryResult64` etc.) as CPU `steady_clock` fallback
+  - `ReadPixels` from default and user color attachments
+- Primitive expansion:
+  - `GL_TRIANGLE_FAN` expanded to triangle list
+  - `GL_LINE_LOOP` expanded to line strip
+- Local test result:
+
+```
+[  PASSED  ] 16 tests
+```
+
+---
+
+## 5. How to Build and Run Locally
+
+From repo root `~/MobileGL-dev`:
+
+```bash
+cmake -S . -B build-diligent -G Ninja \
+  -DCMAKE_BUILD_TYPE=Debug \
+  -DMOBILEGL_ENABLE_DILIGENT=ON \
+  -DMOBILEGL_BUILD_TEST=ON \
+  -DMOBILEGL_BUILD_BENCHMARK=OFF \
+  -DFETCHCONTENT_SOURCE_DIR_GOOGLETEST="$PWD/3rdparty/DiligentCore/ThirdParty/googletest"
+
+cmake --build build-diligent --target DiligentVulkanSanityTest -j 4
+
+./build-diligent/MobileGL/MG_Test/Backend/Diligent/DiligentVulkanSanityTest --gtest_color=no
+```
+
+Notes:
+- `MOBILEGL_BUILD_BENCHMARK=OFF` avoids network fetch of google/benchmark in this environment.
+- `FETCHCONTENT_SOURCE_DIR_GOOGLETEST` pins googletest to DiligentCore's bundled copy, avoiding flaky network clone.
+- Max 4 cores is intentional: use `-j 4`.
+
+---
+
+## 6. Environment Notes
+
+- Host: Linux `aarch64`, glibc 2.43 (Fedora container on Android/Droidspaces)
+- GPU: Turnip Adreno 750, Vulkan API 1.4.354
+- GPU nodes available:
+  - `/dev/dri/renderD128`
+  - `/dev/kgsl-3d0`
+- Android SDK/NDK: `~/android-sdk` (aarch64 glibc)
+  - NDK `27.3.13750724`
+  - CMake `3.22.1`
+- JDK/Gradle for APK builds:
+  - `~/android-build-tools/jdk17`
+  - `~/android-build-tools/gradle/gradle-8.10.2`
+
+---
+
+## 7. Known Limitations / Not Yet Implemented
+
+- User framebuffers now support texture color attachments, renderbuffer color readback, multiple simultaneous color targets, and depth/stencil texture or renderbuffer attachments.
+- Textures auto-sync `ITextureObject` → Diligent resources, including mip levels and sampler state; compressed textures and integer/3-channel formats that Diligent lacks are still skipped.
+- Global UBO (default-block `glUniform*`) and named application UBO blocks (through `glBindBufferBase`/`glUniformBlockBinding`) now upload and bind; SSBOs are still not fed from frontend buffer bindings.
+- Swapchain creation and resize are wired for native EGL window surfaces via `Diligent::ISwapChain`; `Present()` presents the active swap chain when present and otherwise flushes the offscreen target. Actual on-screen EGL presentation is still untested in this headless environment, and the X11 display/connection fields are not yet plumbed through `WindowHandle`. `SetSwapInterval` now forwards the requested sync interval to `ISwapChain::Present()`.
+- No transform feedback / GPU-accelerated queries / non-color readback; fence sync and timer queries use CPU fallbacks.
+- Draw range, multi-draw, instanced-draw wrappers, clear-buffer, blit, read-pixels, CopyTexImage*, CopyImageSubData, GenerateMipmap, GetTexImage/GetTextureImage and indirect draws are now wired; buffer subdata paths still remain.
+- A last-PSO cache now avoids recreating the pipeline when program/render-state/topology/VAO layout is unchanged; texture/UBO resources are still rebound dynamically per draw.
+- The `GLFunctionsTable` is only partially populated.
+
+---
+
+## 8. Recommended Next Steps
+
+1. **Framebuffer / Renderbuffer mapping**
+   - [x] Map `MG_State::GLState::FramebufferObject` attachments to Diligent `ITextureView` / `ITexture`.
+   - [x] Support default framebuffer as current offscreen target.
+   - [x] Support `glBindFramebuffer`, `glFramebufferTexture2D`, renderbuffer color/depth attachments and renderbuffer color readback.
+   - [x] Multiple simultaneous color attachments.
+
+2. **Texture / Sampler full integration**
+   - [x] Translate MobileGL `ITextureObject` to Diligent `ITexture` and cache by `GetLifetimeId()`.
+   - [x] Propagate texture unit bindings into the PSO SRB.
+   - [x] Translate `SamplerObject` state into Diligent `SamplerDesc`.
+
+3. **Uniform / UBO support**
+   - [x] Create Diligent buffer for `ProgramObject::GetUBOData()` / `GetUBOSize()`.
+   - [x] Bind the global UBO as a dynamic shader resource.
+   - [x] Handle per-program uniform block bindings / named UBO blocks.
+
+4. **PSO / resource caching**
+   - [~] Cache PSOs by program + VAO config + render state + topology (single last-PSO fast path).
+   - [~] Cache textures and samplers; buffers/SRBs can still be re-bound per draw.
+
+5. **More GL 3.2 entry points**
+   - [x] `DrawRangeElements`
+   - [x] `MultiDraw*`
+   - [x] `BlitFramebuffer` (same-size color copy)
+   - [x] `ReadPixels` from non-default framebuffer
+   - [x] `CopyTexImage*` / `CopyImageSubData` wired as whole-resource copies
+   - [x] `GetTexImage` / `GetTextureImage` (RGBA8)
+   - [x] Indirect draws (CPU fallback)
+
+6. **Expand local test suite**
+   - [x] Scissor test
+   - [x] Blend test
+   - [x] Texture filtering / sampler state test
+   - [x] framebuffer offscreen render-to-texture test
+   - [x] Depth test visual test
+   - [x] Stencil test
+
+---
+
+## 9. Handoff Notes for Next Agent
+
+- Do **not** reference `origin/Deprecated/Feat/Diligent`; that old implementation is intentionally ignored.
+- Work from this branch, keep tests green.
+- The command `./build-diligent/.../DiligentVulkanSanityTest` runs all 5 Diligent tests.
+- If a new test crashes during shader resource binding, remember Diligent texture SRVs need a sampler attached via `ITextureView::SetSampler()` before `InitializeStaticSRBResources()`.
+- When re-creating a PSO or buffer, call `Release()` (or assign `nullptr`) before the create call to avoid Diligent debug “Overwriting reference” assertions.

MobileGL/Config.h

@@ -66,14 +66,12 @@ namespace MobileGL::MG_Config {
     //   - DISPLAY: X11 session variable, not MobileGL configuration.
     //   - MOBILEGL_LOG_FILE_PATH: log-file init runs before MG_ConfigLoader::Init
     //     (see MG_Util/Debug/Log.cpp).
-    //   - MOBILEGL_VALIDATE_SPIRV: test suites like SpirvPassTest exercise
-    //     ShaderCompiler without ever running MobileGL::Initialize(), and every
-    //     Initialize() re-runs MG_ConfigLoader::Init, which would clobber a
-    //     programmatic override stored here (see ShaderCompiler.cpp,
-    //     SpirvValidationEnabled).
     struct FeaturesTable {
         // MOBILEGL_DISABLE_TIMERQUERY: do not advertise or use GPU timer queries.
         Bool DisableTimerQuery = false;
+        // MOBILEGL_ENABLE_SPIRV_VALIDATION: validate generated and transformed SPIR-V.
+        // Disabled by default because validation is a diagnostics-only cost.
+        Bool EnableSpirvValidation = false;
         // MOBILEGL_USE_ANGLE: load ANGLE EGL/GLES libraries.
         Bool UseAngle = false;
 #if defined(MOBILEGL_TRACE_ANGLE_VARIANTS)
@@ -130,10 +128,6 @@ 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

MobileGL/ConfigLoader.cpp

@@ -162,6 +162,7 @@ namespace MobileGL::MG_ConfigLoader {
     inline void InitFeatures() {
         auto& features = MG_Config::Features;
         features.DisableTimerQuery = QueryEnvFlag("MOBILEGL_DISABLE_TIMERQUERY");
+        features.EnableSpirvValidation = QueryEnvFlag("MOBILEGL_ENABLE_SPIRV_VALIDATION");
         features.UseAngle = QueryEnvFlag("MOBILEGL_USE_ANGLE");
 #if defined(MOBILEGL_TRACE_ANGLE_VARIANTS)
         QueryEnvVariable("MOBILEGL_TRACE_ANGLE_VARIANT", features.TraceAngleVariant, "");
@@ -179,7 +180,6 @@ namespace MobileGL::MG_ConfigLoader {
         features.EsprytForceDepthStencilReadbackEmulation =
             QueryEnvFlag("MOBILEGL_ESPRYT_FORCE_DS_READBACK_EMULATION");
         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");
@@ -202,6 +202,7 @@ namespace MobileGL::MG_ConfigLoader {
     }
         ENTRY(DirectGLES)
         ENTRY(DirectVulkan)
+        ENTRY(DiligentVulkan)
         ENTRY(Unknown)
         MG_Config::ActiveBackendType = BackendType::Unknown;
 #undef ENTRY

MobileGL/MG_Backend/BackendObject.h

@@ -19,6 +19,7 @@ namespace MobileGL {
     enum class BackendType {
         DirectGLES,
         DirectVulkan,
+        DiligentVulkan,
         BackendTypeCount,
         Unknown = -1
     };

MobileGL/MG_Backend/Diligent/BackendObject_Diligent.cpp

@@ -0,0 +1,906 @@
+// MobileGL - MobileGL/MG_Backend/Diligent/BackendObject_Diligent.cpp
+// Copyright (c) 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
+
+#include "BackendObject_Diligent.h"
+#include "DiligentVulkan.h"
+#include "Renderer/DiligentRenderer.h"
+#include <MG_Backend/BackendObject.h>
+#include <MG_Backend/BackendObjects.h>
+#include <MG_State/GLState/Core.h>
+
+#include <EngineFactoryVk.h>
+#include <RenderDevice.h>
+#include <DeviceContext.h>
+
+#include <exception>
+#include <chrono>
+
+namespace MobileGL::MG_Backend::DiligentBackend {
+    namespace {
+        const RendererInfo BuildInitialRendererInfo() {
+            RendererInfo info;
+            info.RendererName = "MobileGL (Diligent/Vulkan)";
+            info.BackendName = "Diligent Vulkan";
+            info.RendererGLInfo.TargetGLVersion = {3, 2, 0};
+            info.RendererGLInfo.TargetGLSLVersion = {1, 50, 0};
+            info.RendererGLInfo.IsCompatibilityProfile = false;
+            return info;
+        }
+
+        DiligentRenderer* GetActiveRenderer() {
+            auto* backend = dynamic_cast<BackendObject_Diligent*>(pActiveBackendObject.get());
+            return backend != nullptr ? backend->GetRenderer() : nullptr;
+        }
+
+        struct DrawArraysIndirectCommand {
+            Uint32 Count = 0;
+            Uint32 InstanceCount = 0;
+            Uint32 First = 0;
+            Uint32 BaseInstance = 0;
+        };
+
+        struct DrawElementsIndirectCommand {
+            Uint32 Count = 0;
+            Uint32 InstanceCount = 0;
+            Uint32 FirstIndex = 0;
+            Int32 BaseVertex = 0;
+            Uint32 BaseInstance = 0;
+        };
+
+        struct CpuTimerQuery {
+            std::chrono::steady_clock::time_point Start;
+            Uint64 TimestampNs = 0;
+            Bool Available = false;
+        };
+
+        const Uint8* ResolveIndirectCommandBytes(const void* indirect, SizeT requiredBytes, const char* label) {
+            auto drawBuffer = MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::DrawIndirect).GetBoundObject();
+            if (drawBuffer) {
+                drawBuffer->SyncPersistentMappedRange();
+                const SizeT commandOffset = reinterpret_cast<SizeT>(indirect);
+                if (drawBuffer->MappedData() == nullptr || commandOffset + requiredBytes > drawBuffer->GetSize()) {
+                    MGLOG_E_ONCE("%s skipped: invalid GL_DRAW_INDIRECT_BUFFER binding or range", label);
+                    return nullptr;
+                }
+                return drawBuffer->MappedData() + commandOffset;
+            }
+            if (indirect == nullptr) {
+                MGLOG_E_ONCE("%s skipped: indirect pointer is null", label);
+                return nullptr;
+            }
+            return reinterpret_cast<const Uint8*>(indirect);
+        }
+
+        void Clear(GLbitfield mask) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || MG_State::pGLContext == nullptr) {
+                return;
+            }
+            if ((mask & GL_COLOR_BUFFER_BIT) != 0) {
+                const auto& color = MG_State::pGLContext->GetClearColor();
+                renderer->Clear(color.x(), color.y(), color.z(), color.w());
+            }
+            if ((mask & GL_DEPTH_BUFFER_BIT) != 0) {
+                renderer->ClearDepth(MG_State::pGLContext->GetClearDepth());
+            }
+            if ((mask & GL_STENCIL_BUFFER_BIT) != 0) {
+                renderer->ClearStencil(MG_State::pGLContext->GetClearStencil());
+            }
+        }
+
+        void DrawArrays(GLenum mode, GLint first, GLsizei count) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer != nullptr) {
+                renderer->DrawFromState(mode, first, count, 0, nullptr);
+            }
+        }
+
+        void DrawElements(GLenum mode, GLsizei count, GLenum type, const void* indices) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer != nullptr) {
+                renderer->DrawFromState(mode, 0, count, type, indices);
+            }
+        }
+
+        void DrawRangeElements(GLenum mode, GLuint start, GLuint end, GLsizei count, GLenum type,
+                               const void* indices) {
+            // The CPU-side UploadVertexDataFromState path already honors the selected index
+            // range. start/end only restrict which indices may be referenced; they do not
+            // change the vertex buffer layout for this backend.
+            (void)start;
+            (void)end;
+            DrawElements(mode, count, type, indices);
+        }
+
+        void DrawRangeElementsBaseVertex(GLenum mode, GLuint start, GLuint end, GLsizei count, GLenum type,
+                                         const void* indices, GLint basevertex) {
+            (void)start;
+            (void)end;
+            auto* renderer = GetActiveRenderer();
+            if (renderer != nullptr) {
+                renderer->DrawFromState(mode, 0, count, type, indices, basevertex);
+            }
+        }
+
+        void MultiDrawArrays(GLenum mode, const GLint* first, const GLsizei* count, GLsizei drawcount) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr) {
+                return;
+            }
+            for (GLsizei i = 0; i < drawcount; ++i) {
+                if (count[i] > 0) {
+                    renderer->DrawFromState(mode, first[i], count[i], 0, nullptr);
+                }
+            }
+        }
+
+        void MultiDrawElements(GLenum mode, const GLsizei* count, GLenum type, const GLvoid* const* indices,
+                               GLsizei drawcount) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr) {
+                return;
+            }
+            for (GLsizei i = 0; i < drawcount; ++i) {
+                if (count[i] > 0) {
+                    renderer->DrawFromState(mode, 0, count[i], type, indices[i]);
+                }
+            }
+        }
+
+        void DrawElementsBaseVertex(GLenum mode, GLsizei count, GLenum type, const void* indices,
+                                    GLint basevertex) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer != nullptr) {
+                renderer->DrawFromState(mode, 0, count, type, indices, basevertex);
+            }
+        }
+
+        void MultiDrawElementsBaseVertex(GLenum mode, const GLsizei* count, GLenum type,
+                                         const GLvoid* const* indices, GLsizei drawcount,
+                                         const GLint* basevertex) {
+            for (GLsizei i = 0; i < drawcount; ++i) {
+                if (count[i] > 0) {
+                    DrawElementsBaseVertex(mode, count[i], type, indices[i],
+                                           basevertex != nullptr ? basevertex[i] : 0);
+                }
+            }
+        }
+
+        void DrawArraysIndirect(GLenum mode, const void* indirect) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || MG_State::pGLContext == nullptr) {
+                return;
+            }
+            const auto* bytes = ResolveIndirectCommandBytes(indirect, sizeof(DrawArraysIndirectCommand),
+                                                            "DrawArraysIndirect");
+            if (bytes == nullptr) {
+                return;
+            }
+            DrawArraysIndirectCommand cmd{};
+            std::memcpy(&cmd, bytes, sizeof(cmd));
+            if (cmd.Count == 0 || cmd.InstanceCount == 0) {
+                return;
+            }
+            for (Uint32 i = 0; i < cmd.InstanceCount; ++i) {
+                renderer->DrawFromState(mode, static_cast<GLint>(cmd.First), static_cast<GLsizei>(cmd.Count),
+                                        0, nullptr);
+            }
+        }
+
+        void DrawElementsIndirect(GLenum mode, GLenum type, const void* indirect) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || MG_State::pGLContext == nullptr) {
+                return;
+            }
+            const SizeT indexSize = MG_Util::GetGLTypeSize(type);
+            if (indexSize == 0) {
+                return;
+            }
+            const auto* bytes = ResolveIndirectCommandBytes(indirect, sizeof(DrawElementsIndirectCommand),
+                                                            "DrawElementsIndirect");
+            if (bytes == nullptr) {
+                return;
+            }
+            DrawElementsIndirectCommand cmd{};
+            std::memcpy(&cmd, bytes, sizeof(cmd));
+            if (cmd.Count == 0 || cmd.InstanceCount == 0) {
+                return;
+            }
+            const void* indices = reinterpret_cast<const void*>(static_cast<SizeT>(cmd.FirstIndex) * indexSize);
+            for (Uint32 i = 0; i < cmd.InstanceCount; ++i) {
+                renderer->DrawFromState(mode, 0, static_cast<GLsizei>(cmd.Count), type, indices,
+                                        static_cast<GLint>(cmd.BaseVertex));
+            }
+        }
+
+        void MultiDrawArraysIndirect(GLenum mode, const void* indirect, GLsizei drawcount, GLsizei stride) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || MG_State::pGLContext == nullptr || drawcount <= 0) {
+                return;
+            }
+            const GLsizei realStride = stride == 0 ? static_cast<GLsizei>(sizeof(DrawArraysIndirectCommand)) : stride;
+            for (GLsizei i = 0; i < drawcount; ++i) {
+                const auto* bytes = ResolveIndirectCommandBytes(
+                    static_cast<const Uint8*>(indirect) + static_cast<SizeT>(i) * static_cast<SizeT>(realStride),
+                    sizeof(DrawArraysIndirectCommand), "MultiDrawArraysIndirect");
+                if (bytes == nullptr) {
+                    continue;
+                }
+                DrawArraysIndirectCommand cmd{};
+                std::memcpy(&cmd, bytes, sizeof(cmd));
+                if (cmd.Count == 0 || cmd.InstanceCount == 0) {
+                    continue;
+                }
+                for (Uint32 instance = 0; instance < cmd.InstanceCount; ++instance) {
+                    renderer->DrawFromState(mode, static_cast<GLint>(cmd.First),
+                                            static_cast<GLsizei>(cmd.Count), 0, nullptr);
+                }
+            }
+        }
+
+        void MultiDrawElementsIndirect(GLenum mode, GLenum type, const void* indirect, GLsizei drawcount,
+                                       GLsizei stride) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || MG_State::pGLContext == nullptr || drawcount <= 0) {
+                return;
+            }
+            const SizeT indexSize = MG_Util::GetGLTypeSize(type);
+            if (indexSize == 0) {
+                return;
+            }
+            const GLsizei realStride = stride == 0 ? static_cast<GLsizei>(sizeof(DrawElementsIndirectCommand)) : stride;
+            for (GLsizei i = 0; i < drawcount; ++i) {
+                const auto* bytes = ResolveIndirectCommandBytes(
+                    static_cast<const Uint8*>(indirect) + static_cast<SizeT>(i) * static_cast<SizeT>(realStride),
+                    sizeof(DrawElementsIndirectCommand), "MultiDrawElementsIndirect");
+                if (bytes == nullptr) {
+                    continue;
+                }
+                DrawElementsIndirectCommand cmd{};
+                std::memcpy(&cmd, bytes, sizeof(cmd));
+                if (cmd.Count == 0 || cmd.InstanceCount == 0) {
+                    continue;
+                }
+                const void* indices = reinterpret_cast<const void*>(static_cast<SizeT>(cmd.FirstIndex) * indexSize);
+                for (Uint32 instance = 0; instance < cmd.InstanceCount; ++instance) {
+                    renderer->DrawFromState(mode, 0, static_cast<GLsizei>(cmd.Count), type, indices,
+                                            static_cast<GLint>(cmd.BaseVertex));
+                }
+            }
+        }
+
+        void MultiDrawArraysIndirectCount(GLenum mode, const void* indirect, GLintptr drawcount,
+                                          GLsizei maxdrawcount, GLsizei stride) {
+            if (MG_State::pGLContext == nullptr) {
+                return;
+            }
+            auto paramBuffer = MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::Parameter).GetBoundObject();
+            if (!paramBuffer) {
+                return;
+            }
+            paramBuffer->SyncPersistentMappedRange();
+            const Uint8* paramData = paramBuffer->MappedData();
+            if (paramData == nullptr) {
+                return;
+            }
+            Uint32 actualDrawCount = 0;
+            std::memcpy(&actualDrawCount, paramData + static_cast<SizeT>(drawcount), sizeof(actualDrawCount));
+            actualDrawCount = std::min<Uint32>(actualDrawCount, static_cast<Uint32>(maxdrawcount));
+            MultiDrawArraysIndirect(mode, indirect, static_cast<GLsizei>(actualDrawCount), stride);
+        }
+
+        void MultiDrawElementsIndirectCount(GLenum mode, GLenum type, const void* indirect,
+                                            GLintptr drawcount, GLsizei maxdrawcount, GLsizei stride) {
+            if (MG_State::pGLContext == nullptr) {
+                return;
+            }
+            auto paramBuffer = MG_State::pGLContext->GetBufferBindingSlot(BufferTarget::Parameter).GetBoundObject();
+            if (!paramBuffer) {
+                return;
+            }
+            paramBuffer->SyncPersistentMappedRange();
+            const Uint8* paramData = paramBuffer->MappedData();
+            if (paramData == nullptr) {
+                return;
+            }
+            Uint32 actualDrawCount = 0;
+            std::memcpy(&actualDrawCount, paramData + static_cast<SizeT>(drawcount), sizeof(actualDrawCount));
+            actualDrawCount = std::min<Uint32>(actualDrawCount, static_cast<Uint32>(maxdrawcount));
+            MultiDrawElementsIndirect(mode, type, indirect, static_cast<GLsizei>(actualDrawCount), stride);
+        }
+
+        void DrawArraysInstanced(GLenum mode, GLint first, GLsizei count, GLsizei instancecount) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || instancecount <= 0) {
+                return;
+            }
+            for (GLsizei i = 0; i < instancecount; ++i) {
+                renderer->DrawFromState(mode, first, count, 0, nullptr);
+            }
+        }
+
+        void DrawArraysInstancedBaseInstance(GLenum mode, GLint first, GLsizei count, GLsizei instancecount,
+                                             GLuint baseinstance) {
+            (void)baseinstance;
+            DrawArraysInstanced(mode, first, count, instancecount);
+        }
+
+        void DrawElementsInstanced(GLenum mode, GLsizei count, GLenum type, const void* indices,
+                                   GLsizei instancecount) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || instancecount <= 0) {
+                return;
+            }
+            for (GLsizei i = 0; i < instancecount; ++i) {
+                renderer->DrawFromState(mode, 0, count, type, indices);
+            }
+        }
+
+        void DrawElementsInstancedBaseVertex(GLenum mode, GLsizei count, GLenum type, const void* indices,
+                                             GLsizei instancecount, GLint basevertex) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || instancecount <= 0) {
+                return;
+            }
+            for (GLsizei i = 0; i < instancecount; ++i) {
+                renderer->DrawFromState(mode, 0, count, type, indices, basevertex);
+            }
+        }
+
+        void DrawElementsInstancedBaseInstance(GLenum mode, GLsizei count, GLenum type, const void* indices,
+                                               GLsizei instancecount, GLuint baseinstance) {
+            (void)baseinstance;
+            DrawElementsInstanced(mode, count, type, indices, instancecount);
+        }
+
+        void DrawElementsInstancedBaseVertexBaseInstance(GLenum mode, GLsizei count, GLenum type,
+                                                         const void* indices, GLsizei instancecount,
+                                                         GLint basevertex, GLuint baseinstance) {
+            (void)baseinstance;
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || instancecount <= 0) {
+                return;
+            }
+            for (GLsizei i = 0; i < instancecount; ++i) {
+                renderer->DrawFromState(mode, 0, count, type, indices, basevertex);
+            }
+        }
+
+        void ClearBufferfv(GLenum buffer, GLint drawbuffer, const GLfloat* value) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || value == nullptr) {
+                return;
+            }
+            if (buffer == GL_COLOR && drawbuffer == 0) {
+                renderer->Clear(value[0], value[1], value[2], value[3]);
+            } else if (buffer == GL_DEPTH && drawbuffer == 0) {
+                renderer->ClearDepth(value[0]);
+            }
+        }
+
+        void ClearBufferiv(GLenum buffer, GLint drawbuffer, const GLint* value) {
+            if (value == nullptr) {
+                return;
+            }
+            if (buffer == GL_STENCIL) {
+                auto* renderer = GetActiveRenderer();
+                if (renderer != nullptr) {
+                    renderer->ClearStencil(static_cast<Uint32>(value[0]));
+                }
+                return;
+            }
+            Float color[4] = {
+                static_cast<Float>(value[0]) / 255.0f,
+                static_cast<Float>(value[1]) / 255.0f,
+                static_cast<Float>(value[2]) / 255.0f,
+                static_cast<Float>(value[3]) / 255.0f,
+            };
+            ClearBufferfv(buffer, drawbuffer, color);
+        }
+
+        void ClearBufferuiv(GLenum buffer, GLint drawbuffer, const GLuint* value) {
+            if (value == nullptr) {
+                return;
+            }
+            if (buffer == GL_STENCIL) {
+                auto* renderer = GetActiveRenderer();
+                if (renderer != nullptr) {
+                    renderer->ClearStencil(value[0]);
+                }
+                return;
+            }
+            Float color[4] = {
+                static_cast<Float>(value[0]) / 255.0f,
+                static_cast<Float>(value[1]) / 255.0f,
+                static_cast<Float>(value[2]) / 255.0f,
+                static_cast<Float>(value[3]) / 255.0f,
+            };
+            ClearBufferfv(buffer, drawbuffer, color);
+        }
+
+        void ClearBufferfi(GLenum buffer, GLint drawbuffer, GLfloat depth, GLint stencil) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || buffer != GL_DEPTH_STENCIL) {
+                return;
+            }
+            (void)drawbuffer;
+            renderer->ClearDepth(depth);
+            renderer->ClearStencil(static_cast<Uint32>(stencil));
+        }
+
+        void ReadPixels(GLint x, GLint y, GLsizei width, GLsizei height, GLenum format, GLenum type, void* pixels) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || pixels == nullptr) {
+                return;
+            }
+            // The Diligent backend's offscreen targets are RGBA8; the frontend currently
+            // uses this entry for the common GL_RGBA/GL_UNSIGNED_BYTE readback path.
+            if (format != GL_RGBA || type != GL_UNSIGNED_BYTE) {
+                return;
+            }
+            renderer->ReadPixels(static_cast<Uint32>(x), static_cast<Uint32>(y),
+                                 static_cast<Uint32>(width), static_cast<Uint32>(height), pixels);
+        }
+
+        void BlitFramebuffer(GLint srcX0, GLint srcY0, GLint srcX1, GLint srcY1,
+                             GLint dstX0, GLint dstY0, GLint dstX1, GLint dstY1,
+                             GLbitfield mask, GLenum filter) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer != nullptr) {
+                renderer->BlitFramebuffer(srcX0, srcY0, srcX1, srcY1, dstX0, dstY0, dstX1, dstY1,
+                                          mask, filter);
+            }
+        }
+
+        void BlitNamedFramebuffer(const SharedPtr<MG_State::GLState::FramebufferObject>& readFramebuffer,
+                                  const SharedPtr<MG_State::GLState::FramebufferObject>& drawFramebuffer,
+                                  GLint srcX0, GLint srcY0, GLint srcX1, GLint srcY1,
+                                  GLint dstX0, GLint dstY0, GLint dstX1, GLint dstY1,
+                                  GLbitfield mask, GLenum filter) {
+            (void)srcX0;
+            (void)srcY0;
+            (void)srcX1;
+            (void)srcY1;
+            (void)dstX0;
+            (void)dstY0;
+            (void)dstX1;
+            (void)dstY1;
+            (void)filter;
+            auto* renderer = GetActiveRenderer();
+            if (renderer != nullptr) {
+                renderer->BlitNamedFramebuffer(readFramebuffer, drawFramebuffer, mask);
+            }
+        }
+
+        void CopyTexImage2D(GLenum target, GLint level, GLenum internalformat, GLint x, GLint y,
+                            GLsizei width, GLsizei height, GLint border) {
+            (void)level;
+            (void)internalformat;
+            (void)x;
+            (void)y;
+            (void)width;
+            (void)height;
+            (void)border;
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || MG_State::pGLContext == nullptr || target != GL_TEXTURE_2D) {
+                return;
+            }
+            auto& unit = MG_State::pGLContext->GetTextureUnitObject(MG_State::pGLContext->GetActiveTextureUnit());
+            auto texture = unit.GetBindingSlot(TextureTarget::Texture2D).GetBoundObject();
+            if (texture) {
+                renderer->CopyReadFramebufferToTexture(*texture);
+            }
+        }
+
+        void CopyTexSubImage2D(GLenum target, GLint level, GLint xoffset, GLint yoffset, GLint x, GLint y,
+                               GLsizei width, GLsizei height) {
+            (void)level;
+            (void)xoffset;
+            (void)yoffset;
+            (void)x;
+            (void)y;
+            (void)width;
+            (void)height;
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || MG_State::pGLContext == nullptr || target != GL_TEXTURE_2D) {
+                return;
+            }
+            auto& unit = MG_State::pGLContext->GetTextureUnitObject(MG_State::pGLContext->GetActiveTextureUnit());
+            auto texture = unit.GetBindingSlot(TextureTarget::Texture2D).GetBoundObject();
+            if (texture) {
+                renderer->CopyReadFramebufferToTexture(*texture);
+            }
+        }
+
+        void GetTexImage(GLenum target, GLint level, GLenum format, GLenum type, GLvoid* pixels) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || MG_State::pGLContext == nullptr || target != GL_TEXTURE_2D ||
+                format != GL_RGBA || type != GL_UNSIGNED_BYTE || pixels == nullptr) {
+                return;
+            }
+            auto& unit = MG_State::pGLContext->GetTextureUnitObject(MG_State::pGLContext->GetActiveTextureUnit());
+            auto texture = unit.GetBindingSlot(TextureTarget::Texture2D).GetBoundObject();
+            if (texture) {
+                renderer->ReadTextureImage(*texture, static_cast<Uint32>(level), pixels);
+            }
+        }
+
+        void GetTextureImage(const SharedPtr<MG_State::GLState::ITextureObject>& texture,
+                             TextureUploadTarget uploadTarget, GLint level, GLenum format, GLenum type,
+                             GLsizei bufSize, GLvoid* pixels) {
+            (void)uploadTarget;
+            (void)bufSize;
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || !texture || format != GL_RGBA || type != GL_UNSIGNED_BYTE ||
+                pixels == nullptr) {
+                return;
+            }
+            renderer->ReadTextureImage(*texture, static_cast<Uint32>(level), pixels);
+        }
+
+        void CopyImageSubData(const SharedPtr<MG_State::GLState::ITextureObject>& srcTexture,
+                              GLenum srcTarget, GLint srcLevel, GLint srcX, GLint srcY, GLint srcZ,
+                              const SharedPtr<MG_State::GLState::ITextureObject>& dstTexture,
+                              GLenum dstTarget, GLint dstLevel, GLint dstX, GLint dstY, GLint dstZ,
+                              GLsizei srcWidth, GLsizei srcHeight, GLsizei srcDepth) {
+            (void)srcTarget;
+            (void)srcLevel;
+            (void)srcX;
+            (void)srcY;
+            (void)srcZ;
+            (void)dstTarget;
+            (void)dstLevel;
+            (void)dstX;
+            (void)dstY;
+            (void)dstZ;
+            (void)srcWidth;
+            (void)srcHeight;
+            (void)srcDepth;
+            auto* renderer = GetActiveRenderer();
+            if (renderer != nullptr && srcTexture && dstTexture) {
+                renderer->CopyTextureSubData(*srcTexture, *dstTexture);
+            }
+        }
+
+        void GenerateMipmap(GLenum target) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer == nullptr || MG_State::pGLContext == nullptr || target != GL_TEXTURE_2D) {
+                return;
+            }
+            auto& unit = MG_State::pGLContext->GetTextureUnitObject(MG_State::pGLContext->GetActiveTextureUnit());
+            auto texture = unit.GetBindingSlot(TextureTarget::Texture2D).GetBoundObject();
+            if (texture) {
+                renderer->GenerateMipmap(*texture);
+            }
+        }
+
+        Bool IsTimerQuerySupported() {
+            return true;
+        }
+
+        BackendQueryHandle BeginTimeElapsedQuery() {
+            auto* query = new CpuTimerQuery;
+            query->Start = std::chrono::steady_clock::now();
+            query->Available = false;
+            return query;
+        }
+
+        void EndTimeElapsedQuery(BackendQueryHandle query) {
+            if (query == nullptr) {
+                return;
+            }
+            auto* cpuQuery = static_cast<CpuTimerQuery*>(query);
+            const auto now = std::chrono::steady_clock::now();
+            cpuQuery->TimestampNs = static_cast<Uint64>(
+                std::chrono::duration_cast<std::chrono::nanoseconds>(now - cpuQuery->Start).count());
+            cpuQuery->Available = true;
+        }
+
+        BackendQueryHandle QueryCounterTimestamp() {
+            auto* query = new CpuTimerQuery;
+            query->TimestampNs = static_cast<Uint64>(
+                std::chrono::duration_cast<std::chrono::nanoseconds>(
+                    std::chrono::steady_clock::now().time_since_epoch()).count());
+            query->Available = true;
+            return query;
+        }
+
+        Bool IsQueryResultAvailable(BackendQueryHandle query) {
+            return query != nullptr && static_cast<CpuTimerQuery*>(query)->Available;
+        }
+
+        Bool GetQueryResult64(BackendQueryHandle query, Bool wait, Uint64* outNanoseconds) {
+            if (query == nullptr || outNanoseconds == nullptr) {
+                return false;
+            }
+            auto* cpuQuery = static_cast<CpuTimerQuery*>(query);
+            if (!cpuQuery->Available && !wait) {
+                return false;
+            }
+            *outNanoseconds = cpuQuery->TimestampNs;
+            return true;
+        }
+
+        void DeleteBackendQuery(BackendQueryHandle query) {
+            delete static_cast<CpuTimerQuery*>(query);
+        }
+
+        BackendSyncHandle FenceSync() {
+            // CPU fallback fence: always signaled is a valid implementation for a
+            // backend without native sync primitives. The handle still round-trips
+            // through ClientWaitSync/DeleteSync so frontend state stays balanced.
+            return new int(0);
+        }
+
+        GLenum ClientWaitSync(BackendSyncHandle sync, GLbitfield flags, GLuint64 timeout) {
+            (void)sync;
+            (void)flags;
+            (void)timeout;
+            return GL_ALREADY_SIGNALED;
+        }
+
+        void WaitSync(BackendSyncHandle sync, GLbitfield flags, GLuint64 timeout) {
+            (void)sync;
+            (void)flags;
+            (void)timeout;
+        }
+
+        void DeleteSync(BackendSyncHandle sync) {
+            delete static_cast<int*>(sync);
+        }
+
+        Bool GetSyncStatus(BackendSyncHandle sync) {
+            (void)sync;
+            return true;
+        }
+
+        void SetSwapInterval(Int interval) {
+            auto* renderer = GetActiveRenderer();
+            if (renderer != nullptr) {
+                renderer->SetSwapInterval(interval > 0 ? static_cast<Uint32>(interval) : 0);
+            }
+        }
+
+        void Present() {
+            auto* renderer = GetActiveRenderer();
+            if (renderer != nullptr) {
+                renderer->Present();
+            }
+        }
+    } // namespace
+
+    BackendObject_Diligent::BackendObject_Diligent()
+        : m_rendererInfo(BuildInitialRendererInfo()) {}
+
+    BackendObject_Diligent::~BackendObject_Diligent() {
+        m_pRenderer.reset();
+        m_pContext.Release();
+        m_pDevice.Release();
+        m_pFactoryVk = nullptr;
+    }
+
+    Bool BackendObject_Diligent::CreateDiligentDevice() {
+        if (m_pDevice && m_pContext) {
+            return true;
+        }
+
+        try {
+            if (m_pFactoryVk == nullptr) {
+                m_pFactoryVk = ::Diligent::GetEngineFactoryVk();
+                if (m_pFactoryVk == nullptr) {
+                    MGLOG_E("Diligent: failed to load Vulkan engine factory");
+                    return false;
+                }
+                m_pFactoryVk->SetBreakOnError(false);
+            }
+
+            ::Diligent::Uint32 numAdapters = 0;
+            m_pFactoryVk->EnumerateAdapters(::Diligent::Version{}, numAdapters, nullptr);
+            if (numAdapters == 0) {
+                MGLOG_W("Diligent: no Vulkan adapters available; skipping device creation");
+                return false;
+            }
+
+            ::Diligent::EngineVkCreateInfo engineCI;
+            ::Diligent::ImmediateContextCreateInfo ctxCI;
+            ctxCI.Name = "MobileGL Diligent Main Context";
+            ctxCI.QueueId = 0;
+            ctxCI.Priority = ::Diligent::QUEUE_PRIORITY_MEDIUM;
+            engineCI.NumImmediateContexts = 1;
+            engineCI.pImmediateContextInfo = &ctxCI;
+
+            ::Diligent::IRenderDevice* pDevice = nullptr;
+            ::Diligent::IDeviceContext* pContext = nullptr;
+            m_pFactoryVk->CreateDeviceAndContextsVk(engineCI, &pDevice, &pContext);
+            if (pDevice == nullptr || pContext == nullptr) {
+                MGLOG_E("Diligent: failed to create Vulkan device/context");
+                return false;
+            }
+
+            m_pDevice.Attach(pDevice);
+            m_pContext.Attach(pContext);
+            MGLOG_I("Diligent: Vulkan device created");
+            return true;
+        } catch (const std::exception& e) {
+            MGLOG_W("Diligent: Vulkan device creation failed: %s", e.what());
+            return false;
+        } catch (...) {
+            MGLOG_W("Diligent: Vulkan device creation failed");
+            return false;
+        }
+    }
+
+    void BackendObject_Diligent::Initialize() {
+        if (m_initialized) {
+            return;
+        }
+        if (!CreateDiligentDevice()) {
+            MGLOG_W("Diligent: backend initialization failed");
+            return;
+        }
+
+        m_pRenderer = std::make_unique<DiligentRenderer>(m_pDevice, m_pContext);
+        if (!m_pRenderer->Initialize(256, 256)) {
+            MGLOG_W("Diligent: renderer initialization failed");
+            m_pRenderer.reset();
+            return;
+        }
+
+        m_functions.GL.Clear = Clear;
+        m_functions.GL.DrawArrays = DrawArrays;
+        m_functions.GL.DrawElements = DrawElements;
+        m_functions.GL.DrawElementsBaseVertex = DrawElementsBaseVertex;
+        m_functions.GL.DrawRangeElements = DrawRangeElements;
+        m_functions.GL.DrawRangeElementsBaseVertex = DrawRangeElementsBaseVertex;
+        m_functions.GL.MultiDrawArrays = MultiDrawArrays;
+        m_functions.GL.MultiDrawElements = MultiDrawElements;
+        m_functions.GL.MultiDrawElementsBaseVertex = MultiDrawElementsBaseVertex;
+        m_functions.GL.DrawArraysInstanced = DrawArraysInstanced;
+        m_functions.GL.DrawArraysInstancedBaseInstance = DrawArraysInstancedBaseInstance;
+        m_functions.GL.DrawElementsInstanced = DrawElementsInstanced;
+        m_functions.GL.DrawElementsInstancedBaseVertex = DrawElementsInstancedBaseVertex;
+        m_functions.GL.DrawElementsInstancedBaseInstance = DrawElementsInstancedBaseInstance;
+        m_functions.GL.DrawElementsInstancedBaseVertexBaseInstance = DrawElementsInstancedBaseVertexBaseInstance;
+        m_functions.GL.DrawArraysIndirect = DrawArraysIndirect;
+        m_functions.GL.DrawElementsIndirect = DrawElementsIndirect;
+        m_functions.GL.MultiDrawArraysIndirect = MultiDrawArraysIndirect;
+        m_functions.GL.MultiDrawElementsIndirect = MultiDrawElementsIndirect;
+        m_functions.GL.MultiDrawArraysIndirectCount = MultiDrawArraysIndirectCount;
+        m_functions.GL.MultiDrawElementsIndirectCount = MultiDrawElementsIndirectCount;
+        m_functions.GL.ClearBufferfv = ClearBufferfv;
+        m_functions.GL.ClearBufferfi = ClearBufferfi;
+        m_functions.GL.ClearBufferiv = ClearBufferiv;
+        m_functions.GL.ClearBufferuiv = ClearBufferuiv;
+        m_functions.GL.BlitFramebuffer = BlitFramebuffer;
+        m_functions.GL.BlitNamedFramebuffer = BlitNamedFramebuffer;
+        m_functions.GL.CopyTexImage2D = CopyTexImage2D;
+        m_functions.GL.CopyTexSubImage2D = CopyTexSubImage2D;
+        m_functions.GL.CopyImageSubData = CopyImageSubData;
+        m_functions.GL.GenerateMipmap = GenerateMipmap;
+        m_functions.GL.GetTexImage = GetTexImage;
+        m_functions.GL.GetTextureImage = GetTextureImage;
+        m_functions.GL.ReadPixels = ReadPixels;
+        m_functions.GL.FenceSync = FenceSync;
+        m_functions.GL.ClientWaitSync = ClientWaitSync;
+        m_functions.GL.WaitSync = WaitSync;
+        m_functions.GL.DeleteSync = DeleteSync;
+        m_functions.GL.GetSyncStatus = GetSyncStatus;
+        m_functions.GL.IsTimerQuerySupported = IsTimerQuerySupported;
+        m_functions.GL.BeginTimeElapsedQuery = BeginTimeElapsedQuery;
+        m_functions.GL.EndTimeElapsedQuery = EndTimeElapsedQuery;
+        m_functions.GL.QueryCounterTimestamp = QueryCounterTimestamp;
+        m_functions.GL.IsQueryResultAvailable = IsQueryResultAvailable;
+        m_functions.GL.GetQueryResult64 = GetQueryResult64;
+        m_functions.GL.DeleteBackendQuery = DeleteBackendQuery;
+        m_functions.Present = Present;
+        m_functions.SetSwapInterval = SetSwapInterval;
+
+        m_initialized = true;
+    }
+
+    DiligentRenderer* BackendObject_Diligent::GetRenderer() {
+        return m_pRenderer.get();
+    }
+
+    Bool BackendObject_Diligent::InitCapabilities() {
+        // Skeleton: no format probing yet. The backend advertises GL 3.2 core
+        // capability, and the capability tables will be filled as resource
+        // creation paths are ported.
+        m_backendCapabilitiesInitialized = true;
+        return true;
+    }
+
+    Bool BackendObject_Diligent::InitWindowSurface() {
+        if (!m_windowHandle.Handle) {
+            MGLOG_E("BackendObject_Diligent::InitWindowSurface failed: native window handle is null");
+            return false;
+        }
+        if (m_pRenderer == nullptr || m_pFactoryVk == nullptr) {
+            MGLOG_E("BackendObject_Diligent::InitWindowSurface failed: renderer/factory is not ready");
+            return false;
+        }
+        return m_pRenderer->CreateSwapChain(m_pFactoryVk, m_windowHandle,
+                                            m_windowHandle.Width, m_windowHandle.Height);
+    }
+
+    Bool BackendObject_Diligent::InitPbufferSurface(EGLint width, EGLint height) {
+        // The Diligent backend keeps its offscreen target for pbuffer EGL surfaces.
+        // A future enhancement can resize/recreate the offscreen target to match the
+        // pbuffer dimensions.
+        (void)width;
+        (void)height;
+        return m_pRenderer != nullptr;
+    }
+
+    void BackendObject_Diligent::ReleaseEGLResources() {
+        if (m_pRenderer != nullptr) {
+            m_pRenderer->ReleaseSwapChain();
+        }
+        BackendObject::ReleaseEGLResources();
+    }
+
+    void BackendObject_Diligent::OnEGLSurfaceReleased(EGLSurface surface) {
+        (void)surface;
+        if (m_pRenderer != nullptr) {
+            m_pRenderer->ReleaseSwapChain();
+        }
+    }
+
+    Bool BackendObject_Diligent::CreateEGLWindowSurface(EGLSurface surface, const WindowHandle& handle) {
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        if (!m_initialized) {
+            MGLOG_E("BackendObject_Diligent::CreateEGLWindowSurface failed: backend not initialized");
+            return false;
+        }
+        if (!handle.Handle || (handle.Backend != WindowBackend::Android && handle.Backend != WindowBackend::X11 &&
+                               handle.Backend != WindowBackend::MetalLayer && handle.Backend != WindowBackend::Win32)) {
+            MGLOG_E("BackendObject_Diligent::CreateEGLWindowSurface failed: unsupported native window backend");
+            return false;
+        }
+        return RegisterEGLWindowSurface(surface, handle);
+    }
+
+    Bool BackendObject_Diligent::CreateEGLPbufferSurface(EGLSurface surface, EGLint width, EGLint height) {
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        if (!m_initialized) {
+            MGLOG_E("BackendObject_Diligent::CreateEGLPbufferSurface failed: backend not initialized");
+            return false;
+        }
+        return RegisterEGLPbufferSurface(surface, width, height);
+    }
+
+    Bool BackendObject_Diligent::ResizeEGLWindowSurface(EGLSurface surface, Uint32 width, Uint32 height) {
+        const std::lock_guard<std::recursive_mutex> lock(m_eglStateMutex);
+        if (!BackendObject::ResizeEGLWindowSurface(surface, width, height)) {
+            return false;
+        }
+        if (m_eglSurface == surface && m_pRenderer != nullptr) {
+            return m_pRenderer->ResizeSwapChain(width, height);
+        }
+        return true;
+    }
+
+    const RendererInfo& BackendObject_Diligent::GetRendererInfo() const {
+        return m_rendererInfo;
+    }
+
+    String BackendObject_Diligent::GetBackendAPIVersionString() const {
+        return "Diligent Vulkan 0.1 (GL 3.2 skeleton)";
+    }
+
+    const GlobalBackendFunctionsTable& BackendObject_Diligent::GetBackendFunctions() const {
+        return m_functions;
+    }
+
+    const DynamicBackendParameters& BackendObject_Diligent::GetDynamicParameters() const {
+        return m_dynamicParameters;
+    }
+
+    BackendType BackendObject_Diligent::GetBackendType() const {
+        return BackendType::DiligentVulkan;
+    }
+} // namespace MobileGL::MG_Backend::DiligentBackend

MobileGL/MG_Backend/Diligent/BackendObject_Diligent.h

@@ -0,0 +1,72 @@
+// MobileGL - MobileGL/MG_Backend/Diligent/BackendObject_Diligent.h
+// Copyright (c) 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
+
+#pragma once
+
+#include <Includes.h>
+#include "../BackendObject.h"
+
+// X11 (pulled in by Includes.h through Vulkan-Headers) defines True/False as
+// macros, which collide with Diligent's Bool constants in BasicTypes.h.
+#if defined(True)
+#undef True
+#endif
+#if defined(False)
+#undef False
+#endif
+
+#include <RefCntAutoPtr.hpp>
+
+namespace Diligent {
+    struct IEngineFactoryVk;
+    struct IRenderDevice;
+    struct IDeviceContext;
+}
+
+namespace MobileGL::MG_Backend::DiligentBackend {
+    class DiligentRenderer;
+
+    // New Diligent/Vulkan backend, implemented from scratch on top of
+    // DiligentCore. The backend object owns the Diligent device/context and
+    // currently advertises OpenGL 3.2 core capability; the GL function table
+    // is intentionally empty until drawing/resource paths are ported.
+    class BackendObject_Diligent : public BackendObject {
+    public:
+        BackendObject_Diligent();
+        ~BackendObject_Diligent() override;
+
+        void Initialize() override;
+        Bool InitCapabilities() override;
+        Bool InitWindowSurface() override;
+        Bool InitPbufferSurface(EGLint width, EGLint height) override;
+        Bool CreateEGLWindowSurface(EGLSurface surface, const WindowHandle& handle) override;
+        Bool CreateEGLPbufferSurface(EGLSurface surface, EGLint width, EGLint height) override;
+        Bool ResizeEGLWindowSurface(EGLSurface surface, Uint32 width, Uint32 height) override;
+        void OnEGLSurfaceReleased(EGLSurface surface) override;
+
+        const RendererInfo& GetRendererInfo() const override;
+        String GetBackendAPIVersionString() const override;
+        const GlobalBackendFunctionsTable& GetBackendFunctions() const override;
+        const DynamicBackendParameters& GetDynamicParameters() const override;
+        BackendType GetBackendType() const override;
+        void ReleaseEGLResources() override;
+
+        DiligentRenderer* GetRenderer();
+
+    private:
+        Bool CreateDiligentDevice();
+
+        RendererInfo m_rendererInfo;
+        DynamicBackendParameters m_dynamicParameters;
+        GlobalBackendFunctionsTable m_functions{};
+        ::Diligent::IEngineFactoryVk* m_pFactoryVk = nullptr;
+        ::Diligent::RefCntAutoPtr<::Diligent::IRenderDevice> m_pDevice;
+        ::Diligent::RefCntAutoPtr<::Diligent::IDeviceContext> m_pContext;
+        std::unique_ptr<DiligentRenderer> m_pRenderer;
+        Bool m_initialized = false;
+    };
+} // namespace MobileGL::MG_Backend::DiligentBackend

MobileGL/MG_Backend/Diligent/DiligentVulkan.cpp

@@ -0,0 +1,8 @@
+// MobileGL - MobileGL/MG_Backend/Diligent/DiligentVulkan.cpp
+// Copyright (c) 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
+
+#include "DiligentVulkan.h"

MobileGL/MG_Backend/Diligent/DiligentVulkan.h

@@ -0,0 +1,17 @@
+// MobileGL - MobileGL/MG_Backend/Diligent/DiligentVulkan.h
+// Copyright (c) 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
+
+#pragma once
+
+#include <Includes.h>
+
+namespace MobileGL::MG_Backend::DiligentBackend {
+    // Backend identity string used by the backend object and local smoke tests.
+    inline String GetDiligentVulkanBackendName() {
+        return "DiligentVulkan";
+    }
+} // namespace MobileGL::MG_Backend::DiligentBackend

MobileGL/MG_Backend/Diligent/Renderer/DiligentRenderer.h

@@ -0,0 +1,153 @@
+// MobileGL - MobileGL/MG_Backend/Diligent/Renderer/DiligentRenderer.h
+// Copyright (c) 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
+
+#pragma once
+
+#include <Includes.h>
+
+// X11 (pulled in by Includes.h through Vulkan-Headers) defines True/False as
+// macros, which collide with Diligent's Bool constants in BasicTypes.h.
+#if defined(True)
+#undef True
+#endif
+#if defined(False)
+#undef False
+#endif
+
+#include <RefCntAutoPtr.hpp>
+
+namespace MobileGL::MG_Backend {
+    struct WindowHandle;
+}
+
+namespace Diligent {
+    struct IRenderDevice;
+    struct IDeviceContext;
+    struct ITexture;
+    struct ITextureView;
+    struct IPipelineState;
+    struct IBuffer;
+    struct ISampler;
+    struct IShaderResourceBinding;
+    struct ISwapChain;
+    struct IEngineFactoryVk;
+}
+
+namespace MobileGL::MG_State::GLState {
+    class ITextureObject;
+    class SamplerObject;
+    class ProgramObject;
+    class RenderbufferObject;
+    class FramebufferObject;
+}
+
+namespace MobileGL::MG_Backend::DiligentBackend {
+    // Minimal real Diligent renderer used to prove the GL 3.2 basic path:
+    // clear an offscreen color target, draw a hardcoded triangle, and read
+    // pixels back. This is the first concrete rendering layer on top of the
+    // Diligent device; it will be expanded into the full MobileGL backend.
+    class DiligentRenderer {
+    public:
+        DiligentRenderer(::Diligent::IRenderDevice* device, ::Diligent::IDeviceContext* context);
+        ~DiligentRenderer();
+
+        Bool Initialize(Uint32 width, Uint32 height);
+        void Clear(Float r, Float g, Float b, Float a);
+        void ClearDepth(Float depth);
+        void ClearStencil(Uint32 stencil);
+        void DrawTriangle();
+        void DrawVertices(const Float* vertices, Uint32 vertexCount);
+        // Creates a real Diligent swap chain for a native EGL window surface.
+        Bool CreateSwapChain(::Diligent::IEngineFactoryVk* factory, const WindowHandle& handle,
+                             Uint32 width, Uint32 height);
+        Bool ResizeSwapChain(Uint32 width, Uint32 height);
+        void SetSwapInterval(Uint32 interval);
+        // Creates a simple 2D RGBA8 texture from CPU data and makes it available
+        // to state PSOs under the shader variable name "g_Texture".
+        Bool CreateTestTexture(const void* data, Uint32 width, Uint32 height);
+        // Draws using the live MG_State GL context: current program, VAO and
+        // bound buffers. This is the front-end emulation entry point.
+        void DrawFromState(GLenum mode, GLint first, GLsizei count, GLenum type, const void* indices,
+                           GLint baseVertex = 0);
+        void ReadPixels(Uint32 x, Uint32 y, Uint32 width, Uint32 height, void* pixels);
+        void BlitFramebuffer(GLint srcX0, GLint srcY0, GLint srcX1, GLint srcY1,
+                             GLint dstX0, GLint dstY0, GLint dstX1, GLint dstY1,
+                             GLbitfield mask, GLenum filter);
+        void BlitNamedFramebuffer(const SharedPtr<MG_State::GLState::FramebufferObject>& readFbo,
+                                  const SharedPtr<MG_State::GLState::FramebufferObject>& drawFbo,
+                                  GLbitfield mask);
+        void CopyReadFramebufferToTexture(MG_State::GLState::ITextureObject& dst);
+        void CopyTextureSubData(MG_State::GLState::ITextureObject& src, MG_State::GLState::ITextureObject& dst);
+        void GenerateMipmap(MG_State::GLState::ITextureObject& texture);
+        Bool ReadTextureImage(MG_State::GLState::ITextureObject& texture, Uint32 level, void* pixels);
+        void ReleaseSwapChain();
+        void Present();
+
+        ::Diligent::IRenderDevice* GetDevice() const { return m_pDevice; }
+        ::Diligent::IDeviceContext* GetContext() const { return m_pContext; }
+
+    private:
+        struct TextureResource {
+            ::Diligent::RefCntAutoPtr<::Diligent::ITexture> Texture;
+            ::Diligent::RefCntAutoPtr<::Diligent::ITextureView> SRV;
+            ::Diligent::RefCntAutoPtr<::Diligent::ITextureView> RTV;
+            ::Diligent::RefCntAutoPtr<::Diligent::ITextureView> DSV;
+            Uint64 ContentVersion = 0;
+            Uint16 ParamsVersion = 0;
+            Bool IsDepth = false;
+        };
+
+        struct SamplerResource {
+            ::Diligent::RefCntAutoPtr<::Diligent::ISampler> Sampler;
+            Uint16 Version = 0;
+        };
+
+        Bool CreateOffscreenTargets();
+        Bool CreatePipeline();
+        Bool CreateVertexBuffer();
+        Bool CreatePipelineFromState(GLenum mode);
+        Bool UploadVertexDataFromState(GLenum mode, GLint first, GLsizei count, GLenum type, const void* indices,
+                                        GLint baseVertex = 0);
+        ::Diligent::ITextureView* SyncTexture(MG_State::GLState::ITextureObject& texture);
+        ::Diligent::ITextureView* SyncTextureForAttachment(MG_State::GLState::ITextureObject& texture, Bool depth);
+        ::Diligent::ITextureView* SyncRenderbuffer(MG_State::GLState::RenderbufferObject& renderbuffer);
+        ::Diligent::ISampler* SyncSampler(const MG_State::GLState::SamplerObject& sampler);
+        Bool BindShaderResourcesFromState(const MG_State::GLState::ProgramObject& program);
+        Bool UploadUBOFromState(const MG_State::GLState::ProgramObject& program);
+        Bool ResolveCurrentRenderTargets(Vector<::Diligent::ITextureView*>& rtvs,
+                                         ::Diligent::ITextureView*& dsv);
+
+        ::Diligent::IRenderDevice* m_pDevice = nullptr;
+        ::Diligent::IDeviceContext* m_pContext = nullptr;
+        ::Diligent::RefCntAutoPtr<::Diligent::ITexture> m_pColorTarget;
+        ::Diligent::RefCntAutoPtr<::Diligent::ITextureView> m_pColorRTV;
+        ::Diligent::RefCntAutoPtr<::Diligent::ITexture> m_pDepthTarget;
+        ::Diligent::RefCntAutoPtr<::Diligent::ITextureView> m_pDepthDSV;
+        ::Diligent::RefCntAutoPtr<::Diligent::ISwapChain> m_pSwapChain;
+        ::Diligent::RefCntAutoPtr<::Diligent::ITexture> m_pTestTexture;
+        ::Diligent::RefCntAutoPtr<::Diligent::ITextureView> m_pTestSRV;
+        ::Diligent::RefCntAutoPtr<::Diligent::ISampler> m_pTestSampler;
+        ::Diligent::RefCntAutoPtr<::Diligent::IShaderResourceBinding> m_pStateSRB;
+        ::Diligent::RefCntAutoPtr<::Diligent::IPipelineState> m_pPSO;
+        ::Diligent::RefCntAutoPtr<::Diligent::IBuffer> m_pVertexBuffer;
+        ::Diligent::RefCntAutoPtr<::Diligent::IBuffer> m_pUBO;
+        Uint32 m_uboSize = 0;
+        Uint32 m_uboContentVersion = 0;
+        Uint64 m_uboProgramLifetimeId = 0;
+        UnorderedMap<Uint64, TextureResource> m_textureCache;
+        UnorderedMap<Uint64, SamplerResource> m_samplerCache;
+        UnorderedMap<Uint32, TextureResource> m_renderbufferCache;
+        UnorderedMap<Uint64, ::Diligent::RefCntAutoPtr<::Diligent::IBuffer>> m_namedUboCache;
+        Uint32 m_width = 256;
+        Uint32 m_height = 256;
+        Uint32 m_swapInterval = 0;
+        Uint32 m_lastDrawVertexCount = 0;
+        Uint64 m_lastPSOKey = 0;
+        Bool m_hasCachedPSO = false;
+        Bool m_initialized = false;
+    };
+} // namespace MobileGL::MG_Backend::DiligentBackend

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.cpp

@@ -712,9 +712,9 @@ namespace MobileGL::MG_Backend::DirectGLES {
                     {
                         .TargetGLVersion = {4, 0, 0},   // GL target version
                         .TargetGLSLVersion = {4, 6, 0}, // Target Shading Language Version
-                        // Baseline advertisement (no timer queries / anisotropy yet); reconciled
-                        // once the ES capabilities exist, see UpdateAdvertisedCapabilityExtensions.
-                        .Extensions = BuildAdvertisedExtensions(false, false),
+                        // Baseline advertisement (no runtime capabilities yet); reconciled once
+                        // the ES capabilities exist, see UpdateAdvertisedCapabilityExtensions.
+                        .Extensions = BuildAdvertisedExtensions(false, false, false, false),
                         .IsCompatibilityProfile = false // Is Compatibility Profile
                     },
                 .StaticBackendCapability = {.AllowVSOnlyPrograms = false} // Backend Capability
@@ -734,9 +734,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
         // thread can only observe the extension string after the
         // advertisement for its context has settled; rebuilding the whole
         // list keeps the re-run after a context recreation idempotent.
-        void UpdateAdvertisedCapabilityExtensions(Bool anisotropicFilteringSupported) {
-            MutableRendererInfo().RendererGLInfo.Extensions =
-                BuildAdvertisedExtensions(AreTimerQueriesSupported(), anisotropicFilteringSupported);
+        void UpdateAdvertisedCapabilityExtensions(const MG_External::GLESCapabilities& capabilities) {
+            MutableRendererInfo().RendererGLInfo.Extensions = BuildAdvertisedExtensions(
+                AreTimerQueriesSupported(), capabilities.SupportsTextureFilterAnisotropy,
+                capabilities.SupportsDrawIndirect,
+                capabilities.SupportsDrawIndirect && capabilities.SupportsBaseInstance);
         }
     } // namespace
 
@@ -779,11 +781,11 @@ namespace MobileGL::MG_Backend::DirectGLES {
             return false;
         }
         DirectGLES::SetGLESCapabilities(m_GLESCapabilities);
-        // Now that g_GLESCapabilities knows about GL_EXT_disjoint_timer_query and
-        // GL_EXT_texture_filter_anisotropic, reconcile the advertisement (see the comment on
-        // UpdateAdvertisedCapabilityExtensions for why it cannot happen when the extension
-        // list is first built).
-        UpdateAdvertisedCapabilityExtensions(m_GLESCapabilities.SupportsTextureFilterAnisotropy);
+        // Now that g_GLESCapabilities knows the host extensions, entry points, and ES version,
+        // reconcile every runtime-gated advertisement (see the comment on
+        // UpdateAdvertisedCapabilityExtensions for why this cannot happen when the list is first
+        // built).
+        UpdateAdvertisedCapabilityExtensions(m_GLESCapabilities);
         UpdateDynamicBackendParameters();
         PopulateFormatCapabilities(m_GLESFunctions, m_GLESCapabilities, MutableFormatCapabilities());
         PrintFormatCapabilities(GetFormatCapabilities());
@@ -924,11 +926,13 @@ namespace MobileGL::MG_Backend::DirectGLES {
         return MutableRendererInfo();
     }
 
-    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported) {
+    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported,
+                                                  Bool drawIndirectSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported) {
         Vector<GLExtension> extensions = {
             V_OpenGL30, V_OpenGL31, V_OpenGL32, V_OpenGL33, V_OpenGL40, E_GL_ARB_draw_buffers_blend,
             E_GL_ARB_compute_shader, E_GL_ARB_shader_storage_buffer_object, E_GL_ARB_shader_image_load_store,
-            E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_EXT_framebuffer_object,
+            E_GL_ARB_clear_buffer_object, 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_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,
@@ -955,6 +959,19 @@ namespace MobileGL::MG_Backend::DirectGLES {
             // extension explicitly permits. It is also the only thing that
             // exposes glProgramParameteri before GL 4.1.
             E_GL_ARB_get_program_binary};
+        // Minecraft 26.3 checks this prerequisite before it even considers
+        // GL_ARB_multi_draw_indirect. ES 3.1 supplies both single-draw entry points; the loader
+        // folds the version and pointer checks into SupportsDrawIndirect.
+        if (drawIndirectSupported) {
+            extensions.push_back(E_GL_ARB_draw_indirect);
+        }
+        // ARB_base_instance also defines the last word of an indirect command. Direct calls are
+        // emulated on every Espryt device, but without host GL_EXT_base_instance a native indirect
+        // draw cannot shift divisor attributes by a GPU-authored non-zero value, so do not promise
+        // that incomplete case.
+        if (drawIndirectSupported && nonZeroIndirectBaseInstanceSupported) {
+            extensions.push_back(E_GL_ARB_base_instance);
+        }
         // GL_KHR_parallel_shader_compile is MobileGL's own capability, not the host ES
         // driver's: the compiler threads are MobileGL's, and glCompileShader/glLinkProgram
         // are serviced entirely inside the frontend. Whether the device driver advertises

MobileGL/MG_Backend/DirectGLES/BackendObject_DirectGLES.h

@@ -67,9 +67,12 @@ namespace MobileGL::MG_Backend::DirectGLES {
     const RendererInfo& GetRendererIdentity();
 
     // The full OpenGL extension list Espryt advertises (glGetString(GL_EXTENSIONS))
-    // for a device whose timer queries / anisotropic filtering are (or are not) usable.
+    // for a device whose timer queries / anisotropic filtering / native indirect draws /
+    // non-zero indirect baseInstance semantics are (or are not) usable.
     // The MOBILEGL_DISABLE_TIMERQUERY escape hatch is applied inside.
-    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported);
+    Vector<GLExtension> BuildAdvertisedExtensions(Bool timerQueriesSupported, Bool anisotropicFilteringSupported,
+                                                  Bool drawIndirectSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported);
 
     // Format: <OpenGL ES Renderer>, OpenGL ES <Major>.<Minor> — the exact string an
     // initialized backend returns from GetBackendAPIVersionString (and that ends up

MobileGL/MG_Backend/DirectGLES/DirectGLES.cpp

@@ -1600,7 +1600,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 !g_hasSyncedRenderState || std::memcmp(currentBytes + kBlendSpanEnd, syncedBytes + kBlendSpanEnd,
                                                        sizeof(RenderStateParameters) - kBlendSpanEnd) != 0;
 
-            IntVec4 backendViewport = parameters.Viewport;
+            IntVec4 backendViewport = MG_State::pGLContext->GetViewport();
             if (backendViewport.z() <= 0 || backendViewport.w() <= 0) {
                 Int surfaceWidth = 0;
                 Int surfaceHeight = 0;
@@ -1614,7 +1614,8 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 g_syncedBackendViewport = backendViewport;
             }
 
-            // All 12 capability bools live after LogicOp in the struct, i.e. in the tail span.
+            // Every capability bool (and the scissor-test mask below) lives after LogicOp in the
+            // struct, i.e. in the tail span.
             if (tailSpanDirty) {
 #define SYNC_CAPABILITY(cap_mg, cap_gl)                                                                                \
     if (forceFullPush || parameters.cap_mg##Enabled != g_syncedRenderStateParameters.cap_mg##Enabled) {                                 \
@@ -1633,11 +1634,26 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 SYNC_CAPABILITY(SampleMask, GL_SAMPLE_MASK);
                 SYNC_CAPABILITY(PolygonOffsetFill, GL_POLYGON_OFFSET_FILL);
                 SYNC_CAPABILITY(RasterizerDiscard, GL_RASTERIZER_DISCARD);
-                SYNC_CAPABILITY(ScissorTest, GL_SCISSOR_TEST);
                 SYNC_CAPABILITY(StencilTest, GL_STENCIL_TEST);
                 SYNC_CAPABILITY(CullFace, GL_CULL_FACE);
 
 #undef SYNC_CAPABILITY
+
+                // GL_SCISSOR_TEST is per-viewport enable state (ARB_viewport_array), so it is a
+                // 16-bit mask and not a "<Name>Enabled" bool the macro above could key off. ES
+                // has exactly one scissor rectangle and one scissor enable, so only bit 0 - the
+                // index every ES draw rasterizes against - can be forwarded; a program that
+                // enables the test for viewport 3 alone gets viewport 0's answer here. That is
+                // the same limitation as the unemulated gl_ViewportIndex on this backend and is
+                // why the multi-viewport half of KHR-GL43.viewport_array stays red on Espryt.
+                {
+                    const Bool scissorTest = (parameters.ScissorTestEnabledMask & 1u) != 0;
+                    const Bool syncedScissorTest =
+                        (g_syncedRenderStateParameters.ScissorTestEnabledMask & 1u) != 0;
+                    if (forceFullPush || scissorTest != syncedScissorTest) {
+                        scissorTest ? g_GLESFuncs.glEnable(GL_SCISSOR_TEST) : g_GLESFuncs.glDisable(GL_SCISSOR_TEST);
+                    }
+                }
             }
 
             if (tailSpanDirty && g_GLESCapabilities.SupportsClipDistance) {
@@ -1864,8 +1880,8 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 if (forceFullPush || parameters.DepthMask != g_syncedRenderStateParameters.DepthMask) {
                     g_GLESFuncs.glDepthMask(parameters.DepthMask ? GL_TRUE : GL_FALSE);
                 }
-                if (forceFullPush || parameters.DepthRange != g_syncedRenderStateParameters.DepthRange) {
-                    g_GLESFuncs.glDepthRangef(parameters.DepthRange.x(), parameters.DepthRange.y());
+                if (forceFullPush || parameters.DepthRanges[0] != g_syncedRenderStateParameters.DepthRanges[0]) {
+                    g_GLESFuncs.glDepthRangef(parameters.DepthRanges[0].x(), parameters.DepthRanges[0].y());
                 }
             }
 
@@ -2003,7 +2019,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
               // everything drawn with GL_SCISSOR_TEST enabled before the app's first glScissor
               // is clipped away - Minecraft 26.2 keeps only its unscissored sky and hand and
               // loses the terrain and the whole GUI.
-                IntVec4 backendScissorBox = parameters.ScissorBox;
+                IntVec4 backendScissorBox = parameters.ScissorBoxes[0];
                 if (backendScissorBox.z() <= 0 || backendScissorBox.w() <= 0) {
                     Int surfaceWidth = 0;
                     Int surfaceHeight = 0;
@@ -3045,10 +3061,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
     }
 
     static Bool SupportsNativeIndirectDraws() {
-        const auto& version = g_GLESCapabilities.GLESVersion;
-        const Bool esVersionOk = version.Major > 3 || (version.Major == 3 && version.Minor >= 1);
-        return esVersionOk && g_GLESFuncs.glDrawElementsIndirect != nullptr &&
-               g_GLESFuncs.glDrawArraysIndirect != nullptr;
+        return g_GLESCapabilities.SupportsDrawIndirect;
     }
 
     // Runs an (indexed) indirect multi-draw. When a GL_DRAW_INDIRECT_BUFFER is bound the draws
@@ -4762,7 +4775,8 @@ namespace MobileGL::MG_Backend::DirectGLES {
     // restores the app state on exit, tracked via the render-state shadow.
     class ScopedScissorDisable {
     public:
-        ScopedScissorDisable() : m_wasEnabled(RenderStateImpl::g_syncedRenderStateParameters.ScissorTestEnabled) {
+        ScopedScissorDisable()
+            : m_wasEnabled((RenderStateImpl::g_syncedRenderStateParameters.ScissorTestEnabledMask & 1u) != 0) {
             if (m_wasEnabled) g_GLESFuncs.glDisable(GL_SCISSOR_TEST);
         }
         ~ScopedScissorDisable() {

MobileGL/MG_Backend/DirectGLES/Managers.cpp

@@ -729,8 +729,13 @@ namespace MobileGL::MG_Backend::DirectGLES {
             void Ops_ReadbackFromGpu(BufferObject& bufferObject) {
                 auto* resource = ResourceOf(bufferObject);
                 if (!resource || resource->id == 0 || !resource->storageInitialized) return;
-                if (resource->persistentMapped) return; // shadow already IS the GPU storage
                 if (!CanTouchGLNow() || resource->contextGeneration != g_bufferContextGeneration) return;
+                if (resource->persistentMapped) {
+                    // Host writes to a persistent map must not race shader writes already queued
+                    // on this context. There is no backend copy to read back in this case.
+                    if (g_GLESFuncs.glFinish) g_GLESFuncs.glFinish();
+                    return;
+                }
                 if (!g_GLESFuncs.glMapBufferRange || !g_GLESFuncs.glUnmapBuffer) return;
                 const SizeT size = std::min<SizeT>(bufferObject.GetSize(), resource->storageSize);
                 if (size == 0) return;
@@ -4741,6 +4746,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 MGLOG_D("%s:", src.empty() ? "" : src.c_str());
             }
             auto& shaderSpirvs = stateProgramObject->GetGeneratedSpirv();
+            const Bool enableSpirvValidation = stateProgramObject->GetSpirvValidationEnabled();
 
             // Blocks a transform-feedback capture request names a member of ("StageData" of
             // "StageData.attrib[0]"). The Adreno ES driver accepts such a request, links, and
@@ -4794,7 +4800,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 Vector<unsigned int> loweredSpirv;
                 const Vector<unsigned int>* effectiveSpirv = &spirvCode;
                 if (glShaderType == GL_VERTEX_SHADER &&
-                    MG_Util::ShaderTranspiler::ShaderCompiler::LowerDrawParametersForEssl(spirvCode, loweredSpirv) &&
+                    MG_Util::ShaderTranspiler::ShaderCompiler::LowerDrawParametersForEssl(spirvCode, loweredSpirv, enableSpirvValidation) &&
                     !loweredSpirv.empty()) {
                     effectiveSpirv = &loweredSpirv;
                 }
@@ -4804,7 +4810,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 Vector<unsigned int> splitArrayInputSpirv;
                 if (glShaderType == GL_VERTEX_SHADER &&
                     MG_Util::ShaderTranspiler::ShaderCompiler::SplitArrayVertexInputsForEssl(
-                        *effectiveSpirv, splitArrayInputSpirv) &&
+                        *effectiveSpirv, splitArrayInputSpirv, enableSpirvValidation) &&
                     !splitArrayInputSpirv.empty() && splitArrayInputSpirv != *effectiveSpirv) {
                     // Only when the pass ACTUALLY split something. The optimizer hands back a
                     // re-serialised copy either way, and adopting that copy for every vertex
@@ -4826,7 +4832,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 if (!xfbCaptureBlockNames.empty() &&
                     MG_Util::ShaderTranspiler::ShaderCompiler::FlattenXfbInterfaceBlocksForEssl(
                         *effectiveSpirv, xfbCaptureBlockNames, stageFlattenedXfbBlockNames,
-                        flattenedXfbSpirv) &&
+                        flattenedXfbSpirv, enableSpirvValidation) &&
                     !flattenedXfbSpirv.empty() && !stageFlattenedXfbBlockNames.empty()) {
                     effectiveSpirv = &flattenedXfbSpirv;
                     flattenedXfbBlockNames.insert(stageFlattenedXfbBlockNames.begin(),
@@ -4842,7 +4848,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 // declare the member highp; nothing else about emission changes.
                 Vector<unsigned int> uboPrecisionSpirv;
                 if (MG_Util::ShaderTranspiler::ShaderCompiler::StripUboMemberRelaxedPrecisionForEssl(
-                        *effectiveSpirv, uboPrecisionSpirv) &&
+                        *effectiveSpirv, uboPrecisionSpirv, enableSpirvValidation) &&
                     !uboPrecisionSpirv.empty()) {
                     effectiveSpirv = &uboPrecisionSpirv;
                 }
@@ -4857,7 +4863,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 Vector<unsigned int> noperspectiveSpirv;
                 if (!g_GLESCapabilities.SupportsNoperspectiveInterpolation &&
                     MG_Util::ShaderTranspiler::ShaderCompiler::EmulateNoPerspectiveForEssl(
-                        *effectiveSpirv, noperspectiveSpirv) &&
+                        *effectiveSpirv, noperspectiveSpirv, enableSpirvValidation) &&
                     !noperspectiveSpirv.empty()) {
                     effectiveSpirv = &noperspectiveSpirv;
                 }
@@ -4867,7 +4873,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 // divides the coordinate of every normalized-coordinate lookup by the texture
                 // size, which is the whole of the difference between the two.
                 Vector<unsigned int> rectLoweredSpirv;
-                if (MG_Util::ShaderTranspiler::ShaderCompiler::LowerRectImages(*effectiveSpirv, rectLoweredSpirv) &&
+                if (MG_Util::ShaderTranspiler::ShaderCompiler::LowerRectImages(*effectiveSpirv, rectLoweredSpirv, enableSpirvValidation) &&
                     !rectLoweredSpirv.empty()) {
                     effectiveSpirv = &rectLoweredSpirv;
                 }
@@ -4881,7 +4887,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 // coordinate to (u, 0, layer) - before SPIRV-Cross can apply its own.
                 Vector<unsigned int> arrayImageSpirv;
                 if (MG_Util::ShaderTranspiler::ShaderCompiler::Lower1DArrayImagesForEssl(*effectiveSpirv,
-                                                                                          arrayImageSpirv) &&
+                                                                                          arrayImageSpirv, enableSpirvValidation) &&
                     !arrayImageSpirv.empty()) {
                     effectiveSpirv = &arrayImageSpirv;
                 }
@@ -4900,7 +4906,8 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 if (!imageFormatBake.glFormatByUniformName.empty() &&
                     MG_Util::ShaderTranspiler::ShaderCompiler::DeclaresFormatlessStorageImage(*effectiveSpirv) &&
                     MG_Util::ShaderTranspiler::ShaderCompiler::BakeImageFormatsForEssl(
-                        *effectiveSpirv, imageFormatBake.glFormatByUniformName, imageFormatSpirv) &&
+                        *effectiveSpirv, imageFormatBake.glFormatByUniformName, imageFormatSpirv,
+                        enableSpirvValidation) &&
                     !imageFormatSpirv.empty()) {
                     effectiveSpirv = &imageFormatSpirv;
                 }
@@ -4916,7 +4923,7 @@ namespace MobileGL::MG_Backend::DirectGLES {
                 Vector<unsigned int> outputIndexSpirv;
                 if (glShaderType == GL_FRAGMENT_SHADER &&
                     MG_Util::ShaderTranspiler::ShaderCompiler::LegalizeFragmentOutputIndexingForEssl(
-                        *effectiveSpirv, outputIndexSpirv) &&
+                        *effectiveSpirv, outputIndexSpirv, enableSpirvValidation) &&
                     !outputIndexSpirv.empty()) {
                     effectiveSpirv = &outputIndexSpirv;
                 }

MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.cpp

@@ -10,6 +10,7 @@
 #include "MG_Backend/BackendObject.h"
 #include "DirectVulkan.h"
 #include "MG_State/GLState/FramebufferState/FramebufferObject.h"
+#include "MG_State/GLState/Core.h"
 #include "MG_State/GLState/TextureState/TextureState.h"
 #include "MG_Util/Classifiers/TextureEnumClassifier.h"
 #include "MG_Util/Converters/MGToGL/TextureEnumConverter.h"
@@ -383,6 +384,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         UpdateDynamicBackendParameters();
         UpdateAdvertisedExtensions();
+        if (MG_State::pGLContext) {
+            MG_State::pGLContext->InvalidateCompileEnv();
+        }
         PopulateFormatCapabilities(physicalDevice.handle, vkGetPhysicalDeviceFormatProperties, m_vulkanCaps,
                                    MutableFormatCapabilities());
         PrintFormatCapabilities(GetFormatCapabilities());
@@ -497,20 +501,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             .ExtraVendor = Nullopt,
             .RendererGLInfo = {.TargetGLVersion = {4, 0, 0},
                                .TargetGLSLVersion = {4, 6, 0},
-                               // Baseline advertisement (no shader subgroup, no timer queries); a
-                               // live backend reconciles its copy in UpdateAdvertisedExtensions.
-                               .Extensions = BuildAdvertisedExtensions(false, false, false),
+                               // Baseline advertisement (no runtime-gated capabilities); a live
+                               // backend reconciles its copy in UpdateAdvertisedExtensions.
+                               .Extensions = BuildAdvertisedExtensions(false, false, false, false),
                                .IsCompatibilityProfile = false},
             .StaticBackendCapability = {.AllowVSOnlyPrograms = false}};
         return rendererInfo;
     }
 
     Vector<GLExtension> BuildAdvertisedExtensions(Bool shaderSubgroupSupported, Bool timerQueriesSupported,
-                                                  Bool anisotropicFilteringSupported) {
+                                                  Bool anisotropicFilteringSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported) {
         Vector<GLExtension> extensions = {
             V_OpenGL30, V_OpenGL31, V_OpenGL32, V_OpenGL33, V_OpenGL40, E_GL_ARB_draw_buffers_blend,
             E_GL_ARB_compute_shader, E_GL_ARB_shader_storage_buffer_object, E_GL_ARB_shader_image_load_store,
-            E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_ARB_multi_draw_indirect,
+            E_GL_ARB_clear_buffer_object, E_GL_ARB_program_interface_query, E_GL_ARB_framebuffer_object, E_GL_ARB_draw_indirect,
+            E_GL_ARB_multi_draw_indirect,
             E_GL_ARB_indirect_parameters, 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_texture_multisample,
             E_GL_ARB_clear_texture, E_GL_ARB_direct_state_access, E_GL_ARB_shader_draw_parameters,
@@ -530,6 +536,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // extension explicitly permits. It is also the only thing that
             // exposes glProgramParameteri before GL 4.1.
             E_GL_ARB_get_program_binary};
+        // Vulkan's drawIndirectFirstInstance feature is optional. Direct base-instance calls work
+        // without it, but ARB_base_instance also promises non-zero firstInstance in GPU indirect
+        // commands; the renderer supplies true only when that word is legal and gl_InstanceID can
+        // be rebased to OpenGL's zero-based semantics.
+        if (nonZeroIndirectBaseInstanceSupported) {
+            extensions.push_back(E_GL_ARB_base_instance);
+        }
         if (shaderSubgroupSupported && !MG_Config::Features.DisableSubgroup) {
             extensions.push_back(E_GL_KHR_shader_subgroup);
         }
@@ -678,6 +691,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_vulkanCaps = capabilities;
         UpdateDynamicBackendParameters();
         UpdateAdvertisedExtensions();
+        if (MG_State::pGLContext) {
+            MG_State::pGLContext->InvalidateCompileEnv();
+        }
         MutableFormatCapabilities().Clear();
     }
 
@@ -690,7 +706,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // the whole list keeps re-runs idempotent.
         m_rendererInfo.RendererGLInfo.Extensions = BuildAdvertisedExtensions(
             m_vulkanCaps.SupportsShaderSubgroup, pVulkanRenderer && pVulkanRenderer->IsTimerQuerySupported(),
-            pVulkanRenderer && pVulkanRenderer->IsSamplerAnisotropySupported());
+            pVulkanRenderer && pVulkanRenderer->IsSamplerAnisotropySupported(),
+            pVulkanRenderer && pVulkanRenderer->IsNonZeroIndirectBaseInstanceSupported());
     }
 
     void BackendObject_DirectVulkan::UpdateDynamicBackendParameters() {

MobileGL/MG_Backend/DirectVulkan/BackendObject_DirectVulkan.h

@@ -62,8 +62,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     // POST screen shows.
 
     // Static identity of the Magma renderer (renderer/backend names, target GL/GLSL
-    // versions, ExtraVendor) with the baseline extension advertisement (no shader
-    // subgroup, no timer queries). A live backend copies this in its constructor and
+    // versions, ExtraVendor) with the baseline extension advertisement (no runtime-gated
+    // capabilities). A live backend copies this in its constructor and
     // reconciles the Extensions in UpdateAdvertisedExtensions once real capabilities
     // exist; callers that need the advertised list for a known capability set must
     // use BuildAdvertisedExtensions instead.
@@ -74,7 +74,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     // MOBILEGL_DISABLE_TIMERQUERY escape hatches are applied inside, so callers pass
     // the detected device support (passing an already-gated value is harmless).
     Vector<GLExtension> BuildAdvertisedExtensions(Bool shaderSubgroupSupported, Bool timerQueriesSupported,
-                                                  Bool anisotropicFilteringSupported);
+                                                  Bool anisotropicFilteringSupported,
+                                                  Bool nonZeroIndirectBaseInstanceSupported);
 
     // Format: <GPU Name>, Vulkan <Vulkan Version>, Driver <Driver Version> — the exact
     // string an initialized backend returns from GetBackendAPIVersionString (and that

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.cpp

@@ -206,6 +206,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         XXHASH_VERIFY(
             XXH64_update(m_hashState, &payload.primitiveRestartEnable, sizeof(payload.primitiveRestartEnable)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &payload.patchControlPoints, sizeof(payload.patchControlPoints)));
+        XXHASH_VERIFY(XXH64_update(m_hashState, &payload.viewportCount, sizeof(payload.viewportCount)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &payload.polygonMode, sizeof(payload.polygonMode)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &payload.cullMode, sizeof(payload.cullMode)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &payload.frontFace, sizeof(payload.frontFace)));
@@ -406,8 +407,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         tessellation.patchControlPoints = payload.patchControlPoints;
 
         VkPipelineViewportStateCreateInfo vpci{VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO};
-        vpci.viewportCount = 1;
-        vpci.scissorCount = 1;
+        // Both counts move together: GL has one scissor rectangle per viewport, and Vulkan
+        // requires viewportCount == scissorCount whenever both are dynamic
+        // (VUID-VkPipelineViewportStateCreateInfo-scissorCount-04136). The caller has already
+        // clamped this to the device's multiViewport capability.
+        vpci.viewportCount = std::max<Uint32>(payload.viewportCount, 1u);
+        vpci.scissorCount = vpci.viewportCount;
 
         VkPipelineRasterizationStateCreateInfo raster{VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO};
         raster.polygonMode = payload.polygonMode;

MobileGL/MG_Backend/DirectVulkan/Renderer/PipelineFactory.h

@@ -42,6 +42,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Bool primitiveRestartEnable = false;
             // GL_PATCH_VERTICES; only read for a PATCH_LIST topology.
             Uint32 patchControlPoints = 3;
+            // How many of ARB_viewport_array's viewports this pipeline rasterizes into. 1 for
+            // every program that never assigns gl_ViewportIndex, which is all of them outside the
+            // conformance suite - the wide shape costs a longer vkCmdSetViewport/Scissor per state
+            // change and can cost hardware fast paths, so it is opt-in per program. Baked into the
+            // pipeline (viewportCount is not dynamic without VK_EXT_extended_dynamic_state) and
+            // therefore hashed; the DYNAMIC viewport/scissor arrays the draw pushes must have
+            // exactly this many elements (VUID-vkCmdDraw-viewportCount-03417/-03418).
+            Uint32 viewportCount = 1;
             VkPolygonMode polygonMode = VK_POLYGON_MODE_FILL;
             VkCullModeFlags cullMode = VK_CULL_MODE_BACK_BIT;
             VkFrontFace frontFace = VK_FRONT_FACE_CLOCKWISE;

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.cpp

@@ -1997,6 +1997,29 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return ReflectedDeclaresInputBuiltin(reflectModule, SpvBuiltInBaseVertex);
     }
 
+    // gl_ViewportIndex on the last pre-rasterization stage. glslang emits it natively for Vulkan
+    // (BuiltIn ViewportIndex plus OpCapability MultiViewport), and nothing in the SpirvPasses
+    // chain touches it, so a plain reflection of the declared output builtins is the whole test.
+    Bool ProgramFactory::ReflectedWritesViewportIndexBuiltin(const SpvReflectShaderModule& reflectModule) {
+        return ReflectedDeclaresOutputBuiltin(reflectModule, SpvBuiltInViewportIndex);
+    }
+
+    Bool ProgramFactory::ReflectedDeclaresOutputBuiltin(const SpvReflectShaderModule& reflectModule,
+                                                       SpvBuiltIn builtin) {
+        for (Uint32 entryIndex = 0; entryIndex < reflectModule.entry_point_count; ++entryIndex) {
+            const SpvReflectEntryPoint& entryPoint = reflectModule.entry_points[entryIndex];
+            for (Uint32 variableIndex = 0; variableIndex < entryPoint.output_variable_count; ++variableIndex) {
+                const SpvReflectInterfaceVariable* variable = entryPoint.output_variables[variableIndex];
+                if (variable != nullptr &&
+                    (variable->decoration_flags & SPV_REFLECT_DECORATION_BUILT_IN) != 0 &&
+                    variable->built_in == builtin) {
+                    return true;
+                }
+            }
+        }
+        return false;
+    }
+
     Bool ProgramFactory::ReflectedDeclaresInputBuiltin(const SpvReflectShaderModule& reflectModule,
                                                        SpvBuiltIn builtin) {
         for (Uint32 entryIndex = 0; entryIndex < reflectModule.entry_point_count; ++entryIndex) {
@@ -2339,6 +2362,46 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
     }
 
+    // Which pre-rasterization stage assigns gl_ViewportIndex is not fixed: GL 4.1 allows only the
+    // geometry stage, ARB_shader_viewport_layer_array/GL 4.6 also the vertex and tessellation
+    // evaluation stages. Rather than guess which one is last, every non-fragment, non-compute
+    // module is asked - one writer anywhere means this program's draws need a multi-viewport
+    // pipeline, and a false positive costs only a wider viewportCount.
+    void ProgramFactory::ReflectViewportIndexUsage(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
+                                                   const Vector<Vector<Uint>>& spirv,
+                                                   VkProgramObject& entry) const {
+        entry.writesViewportIndexBuiltin = false;
+
+        for (SizeT moduleIndex = 0; moduleIndex < shaders.size() && moduleIndex < spirv.size(); ++moduleIndex) {
+            if (!shaders[moduleIndex]) continue;
+            const ShaderStage stage = shaders[moduleIndex]->GetShaderStage();
+            if (stage == ShaderStage::Fragment || stage == ShaderStage::Compute) continue;
+
+            const auto& module = spirv[moduleIndex];
+            if (module.empty()) continue;
+
+            SpvReflectShaderModule reflectModule{};
+            const SpvReflectResult createResult =
+                spvReflectCreateShaderModule(module.size() * sizeof(Uint), module.data(), &reflectModule);
+            if (createResult != SPV_REFLECT_RESULT_SUCCESS) {
+                // Fail toward the wide pipeline. Missing a real gl_ViewportIndex writer would
+                // silently collapse every viewport onto 0 (the exact bug this reflection exists
+                // to fix); over-declaring costs one extra viewport slot on a program that never
+                // uses it.
+                MGLOG_E_ONCE("ProgramFactory::ReflectViewportIndexUsage: reflection failed (result=%d); assuming the "
+                             "program writes gl_ViewportIndex",
+                             static_cast<Int>(createResult));
+                entry.writesViewportIndexBuiltin = true;
+                continue;
+            }
+
+            if (ReflectedWritesViewportIndexBuiltin(reflectModule)) {
+                entry.writesViewportIndexBuiltin = true;
+            }
+            spvReflectDestroyShaderModule(&reflectModule);
+        }
+    }
+
     void ProgramFactory::ReflectFragmentOutputs(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
                                                 const Vector<Vector<Uint>>& spirv,
                                                 VkProgramObject& entry) const {
@@ -2910,8 +2973,73 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             bindings.push_back(layoutBinding);
         }
 
+        // UPDATE_AFTER_BIND is strictly an optional per-layout acceleration. The GL
+        // descriptor model still resolves every sampler uniform element independently
+        // (including its texture-unit sampler-object override); selecting this path
+        // changes neither that resolution nor the set versioning in UniformManager.
+        // A conservative count keeps a layout on ordinary descriptors whenever any
+        // relevant update-after-bind limit is not large enough, rather than asking a
+        // driver to reject it during vkCreateDescriptorSetLayout.
+        Uint32 updateAfterBindSamplers = 0;
+        Uint32 updateAfterBindUniformBuffers = 0;
+        Uint32 updateAfterBindStorageBuffers = 0;
+        Uint32 updateAfterBindSampledImages = 0;
+        Uint32 updateAfterBindStorageImages = 0;
+        for (Uint32 binding = 0; binding < m_maxBindings; ++binding) {
+            const Uint32 count = entry.bindingDescriptorCounts[binding];
+            switch (entry.bindingKinds[binding]) {
+                case DescriptorBindingKind::UniformBufferDynamic:
+                    updateAfterBindUniformBuffers += count;
+                    break;
+                case DescriptorBindingKind::CombinedImageSampler:
+                    updateAfterBindSamplers += count;
+                    updateAfterBindSampledImages += count;
+                    break;
+                case DescriptorBindingKind::UniformTexelBuffer:
+                    updateAfterBindSampledImages += count;
+                    break;
+                case DescriptorBindingKind::StorageBuffer:
+                case DescriptorBindingKind::StorageTexelBuffer:
+                    updateAfterBindStorageBuffers += count;
+                    break;
+                case DescriptorBindingKind::StorageImage:
+                    updateAfterBindStorageImages += count;
+                    break;
+                case DescriptorBindingKind::None:
+                    break;
+            }
+        }
+        const Uint32 updateAfterBindResources = updateAfterBindUniformBuffers + updateAfterBindStorageBuffers +
+                                                updateAfterBindSampledImages + updateAfterBindStorageImages;
+        const auto& uab = m_updateAfterBindLimits;
+        entry.usesUpdateAfterBind =
+            uab.enabled && updateAfterBindSamplers <= uab.maxPerStageSamplers &&
+            updateAfterBindUniformBuffers <= uab.maxPerStageUniformBuffers &&
+            updateAfterBindStorageBuffers <= uab.maxPerStageStorageBuffers &&
+            updateAfterBindSampledImages <= uab.maxPerStageSampledImages &&
+            updateAfterBindStorageImages <= uab.maxPerStageStorageImages &&
+            updateAfterBindResources <= uab.maxPerStageResources &&
+            updateAfterBindSamplers <= uab.maxSetSamplers &&
+            updateAfterBindUniformBuffers <= uab.maxSetUniformBuffers &&
+            updateAfterBindUniformBuffers <= uab.maxSetUniformBuffersDynamic &&
+            updateAfterBindStorageBuffers <= uab.maxSetStorageBuffers &&
+            updateAfterBindStorageBuffers <= uab.maxSetStorageBuffersDynamic &&
+            updateAfterBindSampledImages <= uab.maxSetSampledImages &&
+            updateAfterBindStorageImages <= uab.maxSetStorageImages;
+
+        Vector<VkDescriptorBindingFlags> bindingFlags;
+        VkDescriptorSetLayoutBindingFlagsCreateInfo bindingFlagsInfo{};
+        if (entry.usesUpdateAfterBind) {
+            bindingFlags.assign(bindings.size(), VK_DESCRIPTOR_BINDING_UPDATE_AFTER_BIND_BIT);
+            bindingFlagsInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_BINDING_FLAGS_CREATE_INFO;
+            bindingFlagsInfo.bindingCount = static_cast<Uint32>(bindingFlags.size());
+            bindingFlagsInfo.pBindingFlags = bindingFlags.data();
+        }
+
         VkDescriptorSetLayoutCreateInfo setLayoutInfo{};
         setLayoutInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO;
+        setLayoutInfo.flags = entry.usesUpdateAfterBind ? VK_DESCRIPTOR_SET_LAYOUT_CREATE_UPDATE_AFTER_BIND_POOL_BIT : 0;
+        setLayoutInfo.pNext = entry.usesUpdateAfterBind ? &bindingFlagsInfo : nullptr;
         setLayoutInfo.bindingCount = static_cast<Uint32>(bindings.size());
         setLayoutInfo.pBindings = bindings.data();
         VK_VERIFY(vkCreateDescriptorSetLayout(m_device, &setLayoutInfo, nullptr, &entry.descriptorSetLayout),
@@ -2991,6 +3119,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         auto& shaders = program.GetAttachedShaders();
         auto& spirv = program.GetGeneratedSpirv();
         Vector<Vector<Uint>> moduleSpirvs(spirv.size());
+        const Bool enableSpirvValidation = program.GetSpirvValidationEnabled();
+        if (enableSpirvValidation) {
+            MG_Util::ShaderTranspiler::ShaderCompiler::PrepareSpirvValidation();
+        }
 
         const ShaderStage fixupStage = PickClipFixupStage(shaders);
 
@@ -3036,7 +3168,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // stored as - which addresses [0,1] where the application addressed texels.
             {
                 Vector<Uint> rectLoweredSpirv;
-                if (MG_Util::ShaderTranspiler::ShaderCompiler::LowerRectImages(moduleSpirvs[i], rectLoweredSpirv) &&
+                if (MG_Util::ShaderTranspiler::ShaderCompiler::LowerRectImages(moduleSpirvs[i], rectLoweredSpirv, enableSpirvValidation) &&
                     !rectLoweredSpirv.empty()) {
                     moduleSpirvs[i] = Move(rectLoweredSpirv);
                 }
@@ -3049,7 +3181,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             {
                 Vector<Uint> invariantSpirv;
                 if (MG_Util::ShaderTranspiler::ShaderCompiler::DecoratePositionInvariantForVulkan(
-                        moduleSpirvs[i], invariantSpirv)) {
+                        moduleSpirvs[i], invariantSpirv, enableSpirvValidation)) {
                     moduleSpirvs[i] = std::move(invariantSpirv);
                 } else {
                     // The pass round-trips through SPIRV-Tools IR, so an unparseable module
@@ -3073,7 +3205,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 m_shaderDrawParametersEnabled) {
                 Vector<Uint> rebasedSpirv;
                 if (MG_Util::ShaderTranspiler::ShaderCompiler::RebaseInstanceIndexForVulkan(moduleSpirvs[i],
-                                                                                            rebasedSpirv)) {
+                                                                                            rebasedSpirv, enableSpirvValidation)) {
                     moduleSpirvs[i] = std::move(rebasedSpirv);
                 } else {
                     MGLOG_E("ProgramFactory: failed to rebase gl_InstanceID for program %u; "
@@ -3091,7 +3223,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 (flags & CompileOptionBit::ZeroBaseVertex)) {
                 Vector<Uint> zeroedSpirv;
                 if (MG_Util::ShaderTranspiler::ShaderCompiler::ZeroBaseVertexForVulkan(moduleSpirvs[i],
-                                                                                       zeroedSpirv)) {
+                                                                                       zeroedSpirv, enableSpirvValidation)) {
                     moduleSpirvs[i] = std::move(zeroedSpirv);
                 } else {
                     // Failing open keeps the native builtin, which is the pre-fix behavior:
@@ -3114,7 +3246,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (shaders[i] && shaders[i]->GetShaderStage() == ShaderStage::Vertex) {
                 Vector<Uint> packedSpirv;
                 const Bool packOk = MG_Util::ShaderTranspiler::ShaderCompiler::PackDoubleVertexInputsForVulkan(
-                    moduleSpirvs[i], packedSpirv);
+                    moduleSpirvs[i], packedSpirv, enableSpirvValidation);
                 MOBILEGL_ASSERT(packOk,
                                 "ProgramFactory: 64-bit vertex input packing failed for program %u; the "
                                 "vertex-input format and the shader input type now disagree",
@@ -3138,7 +3270,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (m_unformattedFloatStorageImagesEnabled) {
                 Vector<Uint> unformattedSpirv;
                 if (MG_Util::ShaderTranspiler::ShaderCompiler::UseUnformattedFloatStorageImagesForVulkan(
-                        moduleSpirvs[i], unformattedSpirv)) {
+                        moduleSpirvs[i], unformattedSpirv, enableSpirvValidation)) {
                     moduleSpirvs[i] = std::move(unformattedSpirv);
                 } else {
                     MGLOG_E("ProgramFactory: failed to make float storage images unformatted for program %u",
@@ -3159,7 +3291,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 #else
             // Final module the driver receives; also checked in the INFO-level CI/test
             // lanes, where the DEBUG gate above is compiled out.
-            if (MG_Util::ShaderTranspiler::ShaderCompiler::SpirvValidationEnabled()) {
+            if (enableSpirvValidation) {
                 ValidateTransformedSpirv(moduleSpv, shaders[i]->GetShaderStage(), program.GetExternalIndex());
             }
 #endif
@@ -3189,6 +3321,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ValidateRasterizationStageInterface(shaders, moduleSpirvs, entry, program.GetExternalIndex());
 #endif
         ReflectVertexInputs(shaders, moduleSpirvs, entry);
+        ReflectViewportIndexUsage(shaders, moduleSpirvs, entry);
         ReflectFragmentOutputs(shaders, moduleSpirvs, entry);
         ReflectPassthroughTessControlNeed(shaders, moduleSpirvs, entry);
         ReflectLayout(program, moduleSpirvs, entry);
@@ -3235,14 +3368,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const VkDescriptorSetLayout descriptorSetLayout = it->second.descriptorSetLayout;
                 MGLOG_D("ProgramFactory::OnFrameBoundary: evicting idle program entry hash=0x%llx",
                         static_cast<unsigned long long>(hash));
-                // erase runs ~VkProgramObject (modules/layouts destroyed); notify after
-                // so an observer never observes a half-destroyed entry through a lookup.
-                // Observers only need the handle values to purge their keyed caches.
-                ++m_cacheStructureEpoch; // erase moves/kills entries: memoised pointers die
-                it = m_cache.erase(it);
+                // The observer destroys dependent pipelines and frees descriptor sets while
+                // this entry still owns its layout. Vulkan requires every descriptor set to be
+                // freed before its VkDescriptorSetLayout is destroyed.
                 if (m_evictionObserver != nullptr) {
                     m_evictionObserver->OnProgramEvicted(hash, descriptorSetLayout);
                 }
+                ++m_cacheStructureEpoch; // erase moves/kills entries: memoised pointers die
+                it = m_cache.erase(it);
             } else {
                 ++it;
             }
@@ -3376,7 +3509,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 #if MOBILEGL_LOG_ACTIVE_LEVEL <= MOBILEGL_LOG_LEVEL_DEBUG
         ValidateTransformedSpirv(spirv, ShaderStage::TessControl, 0);
 #else
-        if (MG_Util::ShaderTranspiler::ShaderCompiler::SpirvValidationEnabled()) {
+        if (m_enableSpirvValidation) {
+            MG_Util::ShaderTranspiler::ShaderCompiler::PrepareSpirvValidation();
             ValidateTransformedSpirv(spirv, ShaderStage::TessControl, 0);
         }
 #endif

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.h

@@ -76,6 +76,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         using CompileOptionFlags = Flags<CompileOptionBit>;
         using HashType = Uint64;
 
+        struct UpdateAfterBindLimits {
+            Bool enabled = false;
+            Uint32 maxPerStageSamplers = 0;
+            Uint32 maxPerStageUniformBuffers = 0;
+            Uint32 maxPerStageStorageBuffers = 0;
+            Uint32 maxPerStageSampledImages = 0;
+            Uint32 maxPerStageStorageImages = 0;
+            Uint32 maxPerStageResources = 0;
+            Uint32 maxSetSamplers = 0;
+            Uint32 maxSetUniformBuffers = 0;
+            Uint32 maxSetUniformBuffersDynamic = 0;
+            Uint32 maxSetStorageBuffers = 0;
+            Uint32 maxSetStorageBuffersDynamic = 0;
+            Uint32 maxSetSampledImages = 0;
+            Uint32 maxSetStorageImages = 0;
+        };
+
         struct VkProgramObject {
             static constexpr Uint32 kMaxVertexInputLocations = 32;
 
@@ -88,6 +105,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
             // Layout data (previously in separate VkProgramLayout)
             VkDescriptorSetLayout descriptorSetLayout = VK_NULL_HANDLE;
+            // True only when this layout passed every descriptor-indexing feature and
+            // update-after-bind limit gate at reflection time. It controls both the
+            // layout/binding flags and the pool class used by UniformManager.
+            Bool usesUpdateAfterBind = false;
             VkPipelineLayout pipelineLayout = VK_NULL_HANDLE;
             Vector<DescriptorBindingKind> bindingKinds;
             // The bindings this program actually declares, ascending. bindingKinds is sized to the
@@ -151,6 +172,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // PROGRAM rather than of the variant: the zeroed variant leaves the variable
             // declared, so both variants answer the same and the draw path can ask either.
             Bool readsBaseVertexBuiltin = false;
+            // Some pre-rasterization stage assigns gl_ViewportIndex. Its pipeline declares
+            // viewportCount = the renderer's rasterizable viewport count instead of 1, and its
+            // draws push the whole viewport/scissor array; every other program keeps the
+            // single-viewport fast path untouched. Part of the program's identity (folded into
+            // the pipeline hash through programHash), so no memo can serve the wrong shape.
+            Bool writesViewportIndexBuiltin = false;
             // This program has a tessellation EVALUATION stage and no tessellation CONTROL
             // stage. GL allows that (4.6 core 11.2.2: with no control shader the input patch
             // is passed through unmodified, the output patch size is PATCH_VERTICES, and the
@@ -190,6 +217,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 // a pipeline failure would be reported against the wrong SPIR-V.
                 stageSpirvDigests = std::move(other.stageSpirvDigests);
                 descriptorSetLayout = other.descriptorSetLayout;
+                usesUpdateAfterBind = other.usesUpdateAfterBind;
                 pipelineLayout = other.pipelineLayout;
                 bindingKinds = std::move(other.bindingKinds);
                 activeBindings = std::move(other.activeBindings);
@@ -218,11 +246,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 fragmentInputComponentCount = other.fragmentInputComponentCount;
                 fragmentReplacesDepth = other.fragmentReplacesDepth;
                 readsBaseVertexBuiltin = other.readsBaseVertexBuiltin;
+                writesViewportIndexBuiltin = other.writesViewportIndexBuiltin;
                 needsPassthroughTessControl = other.needsPassthroughTessControl;
                 passthroughTessControlEmulatable = other.passthroughTessControlEmulatable;
                 lastUsedFrame = other.lastUsedFrame;
                 other.hash = 0;
                 other.descriptorSetLayout = VK_NULL_HANDLE;
+                other.usesUpdateAfterBind = false;
                 other.pipelineLayout = VK_NULL_HANDLE;
                 other.hasStorageImages = false;
                 other.declinedDescriptors = false;
@@ -234,6 +264,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.fragmentInputComponentCount = 0;
                 other.fragmentReplacesDepth = false;
                 other.readsBaseVertexBuiltin = false;
+                other.writesViewportIndexBuiltin = false;
                 other.needsPassthroughTessControl = false;
                 other.passthroughTessControlEmulatable = false;
                 other.lastUsedFrame = 0;
@@ -248,6 +279,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 modules = std::move(other.modules);
                 stageSpirvDigests = std::move(other.stageSpirvDigests); // travels with `modules` - see the move ctor
                 descriptorSetLayout = other.descriptorSetLayout;
+                usesUpdateAfterBind = other.usesUpdateAfterBind;
                 pipelineLayout = other.pipelineLayout;
                 bindingKinds = std::move(other.bindingKinds);
                 activeBindings = std::move(other.activeBindings);
@@ -276,11 +308,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 fragmentInputComponentCount = other.fragmentInputComponentCount;
                 fragmentReplacesDepth = other.fragmentReplacesDepth;
                 readsBaseVertexBuiltin = other.readsBaseVertexBuiltin;
+                writesViewportIndexBuiltin = other.writesViewportIndexBuiltin;
                 needsPassthroughTessControl = other.needsPassthroughTessControl;
                 passthroughTessControlEmulatable = other.passthroughTessControlEmulatable;
                 lastUsedFrame = other.lastUsedFrame;
                 other.hash = 0;
                 other.descriptorSetLayout = VK_NULL_HANDLE;
+                other.usesUpdateAfterBind = false;
                 other.pipelineLayout = VK_NULL_HANDLE;
                 other.hasStorageImages = false;
                 other.declinedDescriptors = false;
@@ -292,6 +326,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 other.fragmentInputComponentCount = 0;
                 other.fragmentReplacesDepth = false;
                 other.readsBaseVertexBuiltin = false;
+                other.writesViewportIndexBuiltin = false;
                 other.needsPassthroughTessControl = false;
                 other.passthroughTessControlEmulatable = false;
                 other.lastUsedFrame = 0;
@@ -337,12 +372,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             virtual void OnProgramEvicted(HashType programHash, VkDescriptorSetLayout descriptorSetLayout) = 0;
         };
 
-        explicit ProgramFactory(VkDevice device, const VulkanRendererConfig& config, Uint32 maxBindings = 16,
-                                Bool shaderDrawParametersEnabled = false,
-                                Bool unformattedFloatStorageImagesEnabled = false)
+        explicit ProgramFactory(VkDevice device, const VulkanRendererConfig& config, Uint32 maxBindings,
+                                Bool shaderDrawParametersEnabled,
+                                Bool unformattedFloatStorageImagesEnabled,
+                                Bool enableSpirvValidation,
+                                UpdateAfterBindLimits updateAfterBindLimits)
             : m_device(device), m_maxBindings(maxBindings), m_config(config),
               m_shaderDrawParametersEnabled(shaderDrawParametersEnabled),
-              m_unformattedFloatStorageImagesEnabled(unformattedFloatStorageImagesEnabled) {
+              m_unformattedFloatStorageImagesEnabled(unformattedFloatStorageImagesEnabled),
+              m_enableSpirvValidation(enableSpirvValidation),
+              m_updateAfterBindLimits(updateAfterBindLimits) {
             VkProgramObject::s_device = device;
         }
         // Destroys the pass-through tessellation control modules. Runs while the device is
@@ -400,6 +439,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Shared by the two above: does any entry point list an input variable decorated with
         // this builtin?
         static Bool ReflectedDeclaresInputBuiltin(const SpvReflectShaderModule& reflectModule, SpvBuiltIn builtin);
+        // True when an entry point writes the ViewportIndex builtin (gl_ViewportIndex), i.e. when
+        // the program can route primitives to a viewport other than 0 and its pipeline therefore
+        // has to declare more than one. Asks about OUTPUT variables because that is the direction
+        // a pre-rasterization stage declares it in.
+        static Bool ReflectedWritesViewportIndexBuiltin(const SpvReflectShaderModule& reflectModule);
+        static Bool ReflectedDeclaresOutputBuiltin(const SpvReflectShaderModule& reflectModule, SpvBuiltIn builtin);
 
         // The pass-through tessellation control stage GL 4.6 core 11.2.2 describes for a
         // program that has an evaluation stage and no control stage, for an input patch of
@@ -434,6 +479,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void ReflectVertexInputs(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
                      const Vector<Vector<Uint>>& spirv,
                      VkProgramObject& entry) const;
+        void ReflectViewportIndexUsage(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
+                                       const Vector<Vector<Uint>>& spirv,
+                                       VkProgramObject& entry) const;
         void ReflectFragmentOutputs(const Vector<SharedPtr<MG_State::GLState::ShaderObject>>& shaders,
                         const Vector<Vector<Uint>>& spirv,
                         VkProgramObject& entry) const;
@@ -456,6 +504,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // True only when the logical device enabled both
         // shaderStorageImageReadWithoutFormat and shaderStorageImageWriteWithoutFormat.
         Bool m_unformattedFloatStorageImagesEnabled = false;
+        // Startup snapshot used only by internally synthesized shader modules, which do not
+        // originate from a ProgramLinkTask.
+        Bool m_enableSpirvValidation = false;
+        // Device feature and limit gate resolved before vkCreateDevice. Keeping it in
+        // the factory lets each reflected layout choose ordinary descriptors when its
+        // own counts would exceed the update-after-bind budget.
+        UpdateAfterBindLimits m_updateAfterBindLimits{};
         // See SetDefaultFramebufferHeight. 0 means "not known yet"; the FragCoordYFlip bit is
         // never set before the swapchain exists, so no variant can be compiled against it.
         Uint32 m_defaultFramebufferHeight = 0;

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.cpp

@@ -156,13 +156,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             frame.descriptorPools.clear();
 
             VkDescriptorPool initialPool = VK_NULL_HANDLE;
-            if (!CreateDescriptorPool(m_setsPerFrame, initialPool)) {
+            if (!CreateDescriptorPool(m_setsPerFrame, false, initialPool)) {
                 MGLOG_E_ONCE("UniformDescriptorBinder::Initialize failed: cannot create frame descriptor pool %u",
                         frameIndex);
                 Shutdown();
                 return false;
             }
-            frame.descriptorPools.push_back({initialPool, m_setsPerFrame, 0});
+            frame.descriptorPools.push_back({initialPool, m_setsPerFrame, 0, false});
             MGLOG_D("UniformDescriptorBinder: frame %u descriptor pool created (maxSets=%u)", frameIndex,
                     m_setsPerFrame);
         }
@@ -305,7 +305,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // texture/sampler resolution, completeness probe, sync, layout handling, sampler
         // and view lookups - would recompute the identical descriptor.
         if (trustUnchangedHint && descriptorMemoUsable && binding < m_samplerResolveMemo.size() &&
-            m_samplerResolveMemo[binding].infoValid) {
+            m_samplerResolveMemo[binding].infoValid &&
+            m_samplerResolveMemo[binding].infoProgramLifetimeId == program.GetLifetimeId()) {
             outImageInfo = m_samplerResolveMemo[binding].info;
             return true;
         }
@@ -504,6 +505,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (binding < m_samplerResolveMemo.size()) {
             if (descriptorMemoUsable) {
                 m_samplerResolveMemo[binding].info = outImageInfo;
+                m_samplerResolveMemo[binding].infoProgramLifetimeId = program.GetLifetimeId();
                 m_samplerResolveMemo[binding].infoValid = true;
             } else {
                 // An arrayed binding publishes nothing here, and clears what a previous program
@@ -1388,7 +1390,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return true;
     }
 
-    Bool UniformManager::CreateDescriptorPool(Uint32 maxSets, VkDescriptorPool& outPool) const {
+    Bool UniformManager::CreateDescriptorPool(Uint32 maxSets, Bool updateAfterBind, VkDescriptorPool& outPool) const {
         outPool = VK_NULL_HANDLE;
         if (m_device == VK_NULL_HANDLE || maxSets == 0 || m_maxBindings == 0) {
             return false;
@@ -1431,7 +1433,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // (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.flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT |
+                         (updateAfterBind ? VK_DESCRIPTOR_POOL_CREATE_UPDATE_AFTER_BIND_BIT : 0);
         poolInfo.maxSets = maxSets;
         poolInfo.poolSizeCount = static_cast<Uint32>(std::size(poolSizes));
         poolInfo.pPoolSizes = poolSizes;
@@ -1445,24 +1448,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return true;
     }
 
-    Bool UniformManager::GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex) {
+    Bool UniformManager::GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex, Bool updateAfterBind) {
         if (frame.descriptorPools.empty()) {
             return false;
         }
 
-        const auto& currentBucket = frame.descriptorPools[frame.activeDescriptorPoolIndex];
-        const Uint32 currentMaxSets = std::max<Uint32>(1, currentBucket.maxSets);
+        const auto matchingBucket = std::find_if(
+            frame.descriptorPools.begin(), frame.descriptorPools.end(),
+            [updateAfterBind](const DescriptorPoolBucket& candidate) { return candidate.updateAfterBind == updateAfterBind; });
+        const Uint32 currentMaxSets = matchingBucket != frame.descriptorPools.end()
+                                          ? std::max<Uint32>(1, matchingBucket->maxSets)
+                                          : m_setsPerFrame;
         const Uint32 grownMaxSets = currentMaxSets <= (std::numeric_limits<Uint32>::max() / 2) ? (currentMaxSets * 2)
                                                                                                  : currentMaxSets;
 
         VkDescriptorPool grownPool = VK_NULL_HANDLE;
-        if (!CreateDescriptorPool(grownMaxSets, grownPool)) {
+        if (!CreateDescriptorPool(grownMaxSets, updateAfterBind, grownPool)) {
             MGLOG_E_ONCE("UniformDescriptorBinder::GrowFrameDescriptorPool failed: cannot create grown pool (%u -> %u sets)",
                     currentMaxSets, grownMaxSets);
             return false;
         }
 
-        frame.descriptorPools.push_back({grownPool, grownMaxSets, 0});
+        frame.descriptorPools.push_back({grownPool, grownMaxSets, 0, updateAfterBind});
         frame.activeDescriptorPoolIndex = static_cast<Uint32>(frame.descriptorPools.size() - 1);
         MGLOG_D(
             "UniformDescriptorBinder: frame %u descriptor pool exhausted, grew pool (%u -> %u sets), poolCount=%zu",
@@ -1472,14 +1479,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     VkResult UniformManager::AllocateDescriptorSetsFromActivePool(Uint32 frameIndex, const ProgramFactory::VkProgramObject& programObj, VkDescriptorSet& outDescriptorSet) {
         auto& frame = m_frames[frameIndex];
-        if (frame.activeDescriptorPoolIndex >= frame.descriptorPools.size()) {
-            frame.activeDescriptorPoolIndex = 0;
-        }
-        if (frame.descriptorPools[frame.activeDescriptorPoolIndex].allocatedSets >=
-            frame.descriptorPools[frame.activeDescriptorPoolIndex].maxSets) {
+        const Bool updateAfterBind = programObj.usesUpdateAfterBind;
+        if (frame.activeDescriptorPoolIndex >= frame.descriptorPools.size() ||
+            frame.descriptorPools[frame.activeDescriptorPoolIndex].updateAfterBind != updateAfterBind ||
+            frame.descriptorPools[frame.activeDescriptorPoolIndex].allocatedSets >=
+                frame.descriptorPools[frame.activeDescriptorPoolIndex].maxSets) {
             const auto availableBucket = std::find_if(
                 frame.descriptorPools.begin(), frame.descriptorPools.end(),
-                [](const DescriptorPoolBucket& candidate) { return candidate.allocatedSets < candidate.maxSets; });
+                [updateAfterBind](const DescriptorPoolBucket& candidate) {
+                    return candidate.updateAfterBind == updateAfterBind && candidate.allocatedSets < candidate.maxSets;
+                });
             if (availableBucket == frame.descriptorPools.end()) {
                 outDescriptorSet = VK_NULL_HANDLE;
                 return VK_ERROR_OUT_OF_POOL_MEMORY;
@@ -1515,7 +1524,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         } else {
             VkResult allocResult = AllocateDescriptorSetsFromActivePool(frameIndex, programObj, outDescriptorSet);
             if (allocResult == VK_ERROR_OUT_OF_POOL_MEMORY || allocResult == VK_ERROR_FRAGMENTED_POOL) {
-                if (!GrowFrameDescriptorPool(frame, frameIndex)) {
+                if (!GrowFrameDescriptorPool(frame, frameIndex, programObj.usesUpdateAfterBind)) {
                     MGLOG_E_ONCE("UniformDescriptorBinder::AcquireDescriptorSet failed: descriptor pool growth failed");
                     return allocResult;
                 }

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -114,6 +114,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkDescriptorPool handle = VK_NULL_HANDLE;
             Uint32 maxSets = 0;
             Uint32 allocatedSets = 0;
+            Bool updateAfterBind = false;
         };
 
         // A cached descriptor set together with the pool it was allocated from, so a
@@ -223,8 +224,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void BindDescriptorSetDeduped(VkCommandBuffer commandBuffer, VkPipelineBindPoint bindPoint,
                                       VkPipelineLayout pipelineLayout, VkDescriptorSet descriptorSet,
                                       const Vector<Uint32>& dynamicOffsets);
-        Bool CreateDescriptorPool(Uint32 maxSets, VkDescriptorPool& outPool) const;
-        Bool GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex);
+        Bool CreateDescriptorPool(Uint32 maxSets, Bool updateAfterBind, VkDescriptorPool& outPool) const;
+        Bool GrowFrameDescriptorPool(FrameResources& frame, Uint32 frameIndex, Bool updateAfterBind);
         VkResult AllocateDescriptorSetsFromActivePool(
             Uint32 frameIndex, const ProgramFactory::VkProgramObject& programObj, VkDescriptorSet& outDescriptorSet);
         VkResult AcquireDescriptorSet(Uint32 frameIndex,
@@ -341,8 +342,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // lifetime id, so a freed-and-reallocated sampler or texture at the same heap address
         // always gets a fresh id and misses (a raw pointer would false-hit that ABA) - so a
         // stale guess can only miss and fall through to the hash, never resolve wrong. Still
-        // reset each frame alongside the descriptor-set cache. Indexed by binding.
+        // reset each frame alongside the descriptor-set cache. Indexed by binding, but the
+        // whole-descriptor entry is additionally keyed by program lifetime: Vulkan binding
+        // numbers are layout-local and unrelated programs routinely reuse binding 0/1.
         struct SamplerResolveMemo {
+            Uint64 infoProgramLifetimeId = 0;
             Uint64 samplerLifetimeId = 0;
             Uint64 textureLifetimeId = 0;
             VkSampler sampler = VK_NULL_HANDLE;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.cpp

@@ -300,8 +300,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, newResource.image, newResource.layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
             VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
-            0, VK_ACCESS_TRANSFER_WRITE_BIT, newResource.aspect, 0, newResource.mipLevels,
-            newResource.arrayLayers);
+            0, VK_ACCESS_TRANSFER_WRITE_BIT, newResource.aspect, 0, newResource.mipLevels);
         MOBILEGL_ASSERT(ok, "PreserveTextureContentsOnRecreate: failed to prepare destination image");
 
         VkImageLayout srcTrackedLayout = oldResource.layout;
@@ -311,8 +310,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, oldResource.image, srcTrackedLayout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
             srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
-            srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, oldResource.aspect, 0, preservedMipLevels,
-            oldResource.arrayLayers);
+            srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, oldResource.aspect, 0, preservedMipLevels);
         MOBILEGL_ASSERT(ok, "PreserveTextureContentsOnRecreate: failed to prepare source image");
 
         Vector<VkImageCopy> copyRegions;
@@ -344,8 +342,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, newResource.image, newResource.layout, oldResource.layout,
             VK_PIPELINE_STAGE_TRANSFER_BIT, dstStageMask,
-            VK_ACCESS_TRANSFER_WRITE_BIT, dstAccessMask, newResource.aspect, 0, newResource.mipLevels,
-            newResource.arrayLayers);
+            VK_ACCESS_TRANSFER_WRITE_BIT, dstAccessMask, newResource.aspect, 0, newResource.mipLevels);
         MOBILEGL_ASSERT(ok, "PreserveTextureContentsOnRecreate: failed to restore destination layout");
 
         VK_VERIFY(vkEndCommandBuffer(commandBuffer), "vkEndCommandBuffer(texture preserve)");
@@ -1191,7 +1188,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const Bool lowerTransitioned = TransitionImageLayout(
                     commandBuffer, resource.image, lowerMipLayout, newLayout,
                     srcStageMask, dstStageMask, srcAccessMask, dstAccessMask,
-                    resource.aspect, 0, writtenMipLevel, resource.arrayLayers);
+                    resource.aspect, 0, writtenMipLevel);
                 MOBILEGL_ASSERT(lowerTransitioned,
                                 "UpdateTrackedImageLayoutAfterAttachmentWrite: failed to transition lower mip levels for textureId=%d",
                                 texture->GetExternalIndex());
@@ -1203,8 +1200,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const Bool upperTransitioned = TransitionImageLayout(
                     commandBuffer, resource.image, upperMipLayout, newLayout,
                     srcStageMask, dstStageMask, srcAccessMask, dstAccessMask,
-                    resource.aspect, upperBaseMipLevel, resource.mipLevels - upperBaseMipLevel,
-                    resource.arrayLayers);
+                    resource.aspect, upperBaseMipLevel, resource.mipLevels - upperBaseMipLevel);
                 MOBILEGL_ASSERT(upperTransitioned,
                                 "UpdateTrackedImageLayoutAfterAttachmentWrite: failed to transition upper mip levels for textureId=%d",
                                 texture->GetExternalIndex());
@@ -1257,8 +1253,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const Bool ok = TransitionImageLayout(commandBuffer, resource->image, resource->layout, targetLayout, srcStageMask,
                                               s_sampledReadStages, srcAccessMask,
-                                              VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels,
-                                              resource->arrayLayers);
+                                              VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionTextureForSampling: transition failed for textureId=%d", texture.GetExternalIndex());
         // Pre-pass stream bookkeeping: a command referencing the image was recorded.
         StampResourceRecordingUse(*resource);
@@ -1288,7 +1283,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                               VK_IMAGE_LAYOUT_GENERAL, srcStageMask,
                                               VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, srcAccessMask,
                                               VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
-                                              resource->aspect, 0, resource->mipLevels, resource->arrayLayers);
+                                              resource->aspect, 0, resource->mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionTextureForStorageImage: transition failed for textureId=%d",
                         texture.GetExternalIndex());
         // Pre-pass stream bookkeeping: a command referencing the image was recorded.
@@ -1355,8 +1350,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                  VkImageLayout& trackedLayout, VkImageLayout newLayout,
                                                  VkPipelineStageFlags srcStageMask, VkPipelineStageFlags dstStageMask,
                                                  VkAccessFlags srcAccessMask, VkAccessFlags dstAccessMask,
-                                                 VkImageAspectFlags aspectMask, Uint32 baseMipLevel, Uint32 levelCount,
-                                                 Uint32 layerCount) {
+                                                 VkImageAspectFlags aspectMask, Uint32 baseMipLevel,
+                                                 Uint32 levelCount) {
         MOBILEGL_ASSERT(image != VK_NULL_HANDLE, "TransitionImageLayout: m_image == VK_NULL_HANDLE");
         MOBILEGL_ASSERT(!((dstAccessMask & VK_ACCESS_TRANSFER_READ_BIT) != 0 &&
                           (dstStageMask & VK_PIPELINE_STAGE_TRANSFER_BIT) == 0),
@@ -1381,7 +1376,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         barrier.subresourceRange.baseMipLevel = baseMipLevel;
         barrier.subresourceRange.levelCount = levelCount;
         barrier.subresourceRange.baseArrayLayer = 0;
-        barrier.subresourceRange.layerCount = layerCount;
+        // Every layer, always - see the declaration for why layout tracking leaves no other
+        // correct answer. VK_REMAINING_ARRAY_LAYERS rather than the image's own `arrayLayers`
+        // because those are not the same number for a 3D image: MobileGL creates 3D images
+        // 2D_ARRAY_COMPATIBLE and their arrayLayers is 1, which today Vulkan reads as "all depth
+        // slices" but will read as "depth slice 0" once VK_KHR_maintenance9 is enabled. The
+        // validation layer warns about that literal 1 by name.
+        barrier.subresourceRange.layerCount = VK_REMAINING_ARRAY_LAYERS;
         vkCmdPipelineBarrier(commandBuffer, srcStageMask, dstStageMask, 0, 0, nullptr, 0, nullptr, 1, &barrier);
 
         trackedLayout = newLayout;
@@ -2605,7 +2606,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                         VK_PIPELINE_STAGE_TRANSFER_BIT,
                                         uploadSrcAccessMask,
                                         VK_ACCESS_TRANSFER_WRITE_BIT,
-                                        aspectMask, 0, outResource.mipLevels, outResource.arrayLayers);
+                                        aspectMask, 0, outResource.mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionImageLayout to VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL failed");
 
         // Array textures keep their GL "depth" in VkImage array layers, so the
@@ -2709,7 +2710,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                    s_sampledReadStages,
                                    VK_ACCESS_TRANSFER_WRITE_BIT,
                                    VK_ACCESS_SHADER_READ_BIT,
-                                   aspectMask, 0, outResource.mipLevels, outResource.arrayLayers);
+                                   aspectMask, 0, outResource.mipLevels);
         MOBILEGL_ASSERT(ok, "TransitionImageLayout to sampled read-only layout failed");
         outResource.layout = finalLayout;
 

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.h

@@ -388,12 +388,24 @@ public:
     static Bool AreSampledImageViewFormatsCompatible(VkFormat imageFormat, VkFormat viewFormat);
     static Bool AreStorageImageViewFormatsCompatible(VkFormat imageFormat, VkFormat viewFormat);
 
+    // Moves `image` to `newLayout` and writes the new layout back through `trackedLayout`.
+    //
+    // The barrier covers EVERY array layer of the image, and there is deliberately no layer
+    // parameter to say otherwise: layout here is tracked per IMAGE (one `TextureResource::layout`,
+    // or one caller-owned variable), so a barrier narrower than the image would leave the layers it
+    // skipped in the old layout while the tracker claims they moved. Every transfer against a
+    // framebuffer attachment above layer 0 - glReadPixels, glBlitFramebuffer, glCopyTexSubImage,
+    // glCopyImageSubData - then ran its copy on a layer no barrier had transitioned.
+    //
+    // The mip range IS a parameter, because mip levels really are transitioned piecewise (see
+    // UpdateTrackedImageLayoutAfterAttachmentWrite and the mipmap generation loops): those callers
+    // move the complement of the level they wrote so the whole image converges on one layout again.
+    // Nothing does, or can, do that per layer.
     static Bool TransitionImageLayout(VkCommandBuffer commandBuffer, VkImage image, VkImageLayout& trackedLayout,
                                VkImageLayout newLayout, VkPipelineStageFlags srcStageMask,
                                VkPipelineStageFlags dstStageMask, VkAccessFlags srcAccessMask,
                                VkAccessFlags dstAccessMask, VkImageAspectFlags aspectMask,
-                               Uint32 baseMipLevel = 0, Uint32 levelCount = 1,
-                               Uint32 layerCount = 1);
+                               Uint32 baseMipLevel = 0, Uint32 levelCount = 1);
 
     SizeT CollectGarbage();
 

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.cpp

@@ -238,11 +238,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     // nothing else across all of gl33.
     //
     // The mapping below is derived from - and at full extent exactly reproduces - the pixel
-    // mapping RemapDefaultFboReadbackToGLOrientation has always used:
+    // mapping VulkanRenderer::RemapDefaultFramebufferReadback uses:
     //     identity : image(x, H-1-y)      -> flip Y
     //     180      : image(W-1-x, y)      -> mirror X (the rotation already flips the rows)
-    // Quarter turns swap the axes; nothing in this renderer models that (the readback declines to
-    // remap them and the viewport path only rescales), so they are left exactly as they were.
+    // Quarter turns swap the axes and are handled by MapDefaultFramebufferReadbackRect rather than
+    // this same-axis helper.
     struct DefaultFramebufferRectMapping {
         Bool flipY = false;
         Bool mirrorX = false;
@@ -335,19 +335,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // Complete input inventory of ApplyDynamicDrawStateTail, one line per reader
         // (each accessor it replaces is a verified plain field read of the same
         // RenderStateParameters field - RenderState.cpp):
-        //   ApplyGLViewportState    : Viewport, DepthRange, + extent/isDefaultFbo/preTransform
+        //   ApplyGLViewportState    : Viewports[0], DepthRanges[0], + extent/isDefaultFbo/preTransform
         //   ApplyBlendConstants     : BlendColor
         //   ApplyPolygonOffsetState : PolygonOffsetUnits, PolygonOffsetFactor
         //   ApplyLineWidthState     : LineWidth (see the caveat below)
         //   ApplyStencilState       : StencilStates[0..1].{ValueMask, WriteMask, Ref}
-        //   scissor rect            : ScissorTestEnabled, ScissorBox,
+        //   scissor rect            : ScissorTestEnabledMask bit 0, ScissorBoxes[0],
         //                             + extent/isDefaultFbo/preTransform
         // Caveat, unchanged from the version-only gate: ApplyLineWidthState also clamps
         // to the ACTIVE BACKEND OBJECT's aliased line-width range. Those are device
         // limits queried once at backend init and constant for the renderer's lifetime,
         // so they are not part of the key (the version gate never covered them either).
         struct DynamicTailKey {
-            Int viewport[4] = {0, 0, 0, 0};
+            Float viewport[4] = {0.0f, 0.0f, 0.0f, 0.0f};
             Float depthRange[2] = {0.0f, 0.0f};
             Float blendColor[4] = {0.0f, 0.0f, 0.0f, 0.0f};
             Float polygonOffsetFactor = 0.0f;
@@ -437,12 +437,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         vkCmdSetScissor(commandBuffer, 0, 1, &scissor);
     }
 
-    static void ApplyGLViewportState(VkCommandBuffer commandBuffer,
-                                     const IntVec2& framebufferExtent,
-                                     VkSurfaceTransformFlagBitsKHR preTransform,
-                                     Bool isDefaultFramebuffer) {
-        const IntVec4& viewportState = MG_State::pGLContext->GetViewport();
-        const FloatVec2& depthRange = MG_State::pGLContext->GetDepthRange();
+    // One viewport of the ARB_viewport_array state, mapped into Vulkan's frame. Split out of
+    // ApplyGLViewportState so the multi-viewport path derives index i through EXACTLY the same
+    // arithmetic as index 0 - the default-framebuffer Y-flip and pre-transform rotation
+    // especially, which is the classic way a multi-viewport port comes out upside down for every
+    // index but the one that was tested.
+    static VkViewport ComputeGLViewport(Uint32 index,
+                                        const IntVec2& framebufferExtent,
+                                        VkSurfaceTransformFlagBitsKHR preTransform,
+                                        Bool isDefaultFramebuffer) {
+        // Snapped to integers. The viewport is float STATE (glViewportIndexedf may set a
+        // fractional origin, and GetFloati_v hands it back verbatim), but what rasterizes here is
+        // the rounded rectangle - a deliberate, documented infidelity rather than a spec claim:
+        // MobileGL passes the driver's VIEWPORT_SUBPIXEL_BITS through, so it does advertise
+        // subpixel viewport precision it does not deliver. Nothing in KHR-GL43.viewport_array or
+        // in Minecraft sets a fractional viewport (the conformance checks are all on the state
+        // round trip), which is why the honest-but-lossy path was kept over widening every
+        // default-framebuffer Y-flip/pre-transform helper to floats. See the KNOWN INFIDELITY
+        // note in MG_IntegrationTest/Scenarios/AdvertisedLimitsScenario.cpp.
+        const FloatVec4& stored = MG_State::pGLContext->GetViewportIndexed(index);
+        const IntVec4 viewportState(static_cast<Int>(std::lround(stored.x())),
+                                    static_cast<Int>(std::lround(stored.y())),
+                                    static_cast<Int>(std::lround(stored.z())),
+                                    static_cast<Int>(std::lround(stored.w())));
+        const FloatVec2& depthRange = MG_State::pGLContext->GetDepthRangeIndexed(index);
         const IntVec2 logicalExtent = isDefaultFramebuffer
             ? ResolveDefaultFramebufferLogicalExtent(preTransform, framebufferExtent)
             : framebufferExtent;
@@ -477,6 +495,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         viewport.height = static_cast<float>(viewportHeight);
         viewport.minDepth = depthRange.x();
         viewport.maxDepth = depthRange.y();
+        return viewport;
+    }
+
+    static void ApplyGLViewportState(VkCommandBuffer commandBuffer,
+                                     const IntVec2& framebufferExtent,
+                                     VkSurfaceTransformFlagBitsKHR preTransform,
+                                     Bool isDefaultFramebuffer) {
+        const VkViewport viewport = ComputeGLViewport(0, framebufferExtent, preTransform, isDefaultFramebuffer);
         auto& shadow = g_dynamicStateShadow;
         if (shadow.viewportValid && shadow.viewport.x == viewport.x && shadow.viewport.y == viewport.y &&
             shadow.viewport.width == viewport.width && shadow.viewport.height == viewport.height &&
@@ -1454,7 +1480,8 @@ void main() {
             const char* label = nullptr;
         };
 
-        static Uint32 ComputeMaxProgramBindings(const VkPhysicalDeviceProperties& properties) {
+        static Uint32 ComputeMaxProgramBindings(const VkPhysicalDeviceProperties& properties,
+                                                const ProgramFactory::UpdateAfterBindLimits& updateAfterBindLimits) {
             const auto& limits = properties.limits;
             static constexpr Uint32 kMinProgramBindings = 16;
             static constexpr Uint32 kMaxProgramBindingsCap = 256;
@@ -1469,6 +1496,21 @@ void main() {
             maxBindings = std::min(maxBindings, maxCombinedImageSamplers);
             maxBindings = std::min(maxBindings, maxSampledImages + maxDynamicUniformBuffers);
 
+            if (updateAfterBindLimits.enabled) {
+                const Uint32 updateAfterBindSamplers = std::min(updateAfterBindLimits.maxPerStageSamplers,
+                                                                 updateAfterBindLimits.maxSetSamplers);
+                const Uint32 updateAfterBindSampledImages = std::min(updateAfterBindLimits.maxPerStageSampledImages,
+                                                                      updateAfterBindLimits.maxSetSampledImages);
+                const Uint32 updateAfterBindDynamicUniformBuffers =
+                    std::min(updateAfterBindLimits.maxPerStageUniformBuffers,
+                             updateAfterBindLimits.maxSetUniformBuffersDynamic);
+                Uint32 updateAfterBindBindings = updateAfterBindLimits.maxPerStageResources;
+                updateAfterBindBindings = std::min(updateAfterBindBindings, updateAfterBindSamplers);
+                updateAfterBindBindings =
+                    std::min(updateAfterBindBindings, updateAfterBindSampledImages + updateAfterBindDynamicUniformBuffers);
+                maxBindings = std::max(maxBindings, updateAfterBindBindings);
+            }
+
             maxBindings = std::max(kMinProgramBindings, maxBindings);
             maxBindings = std::min(kMaxProgramBindingsCap, maxBindings);
             return maxBindings;
@@ -2125,47 +2167,6 @@ void main() {
             return static_cast<Uint8>(value * 255.0f + 0.5f);
         }
 
-        // Re-order the copied BLOCK - not the whole image - from the default framebuffer's stored
-        // orientation into GL's. The caller has already aimed the copy at the right place with
-        // MapDefaultFramebufferRectAxis, so what arrives here is exactly the requested
-        // rectWidth x rectHeight rect, and all that is left is the order of rows (identity) or of
-        // columns (180) WITHIN it.
-        //
-        // This used to iterate the full swapchain extent and index both sides with that stride,
-        // which is why its caller could only use it on an exact full-extent read - and why every
-        // partial glReadPixels of the default framebuffer came back in Vulkan row order. Only
-        // identity/180 share the swapchain extent with the default framebuffer; 90/270 swap
-        // extents and are still declined.
-        static Bool RemapDefaultFboReadbackToGLOrientation(const Uint8* rawPixels,
-                                                            Uint32 rectWidth,
-                                                            Uint32 rectHeight,
-                                                            VkSurfaceTransformFlagBitsKHR preTransform,
-                                                            SizeT texelSize,
-                                                            Uint8* outPixels) {
-            if (IsQuarterTurnPreTransform(preTransform)) {
-                return false;
-            }
-            if (rectWidth == 0 || rectHeight == 0 || texelSize == 0) {
-                return false;
-            }
-            const DefaultFramebufferRectMapping mapping = GetDefaultFramebufferRectMapping(preTransform);
-            const SizeT rowBytes = static_cast<SizeT>(rectWidth) * texelSize;
-            for (Uint32 outY = 0; outY < rectHeight; ++outY) {
-                const Uint32 srcY = mapping.flipY ? (rectHeight - 1 - outY) : outY;
-                const Uint8* srcRow = rawPixels + static_cast<SizeT>(srcY) * rowBytes;
-                Uint8* dstRow = outPixels + static_cast<SizeT>(outY) * rowBytes;
-                if (!mapping.mirrorX) {
-                    Memcpy(dstRow, srcRow, rowBytes);
-                    continue;
-                }
-                for (Uint32 outX = 0; outX < rectWidth; ++outX) {
-                    Memcpy(dstRow + static_cast<SizeT>(outX) * texelSize,
-                           srcRow + static_cast<SizeT>(rectWidth - 1 - outX) * texelSize, texelSize);
-                }
-            }
-            return true;
-        }
-
         static SizeT AlignPixelRow(SizeT rowBytes, Int alignment) {
             const SizeT resolvedAlignment = static_cast<SizeT>(std::max(alignment, 1));
             return (rowBytes + resolvedAlignment - 1) & ~(resolvedAlignment - 1);
@@ -2712,6 +2713,95 @@ void main() {
         return formatInfo.texel_block_size;
     }
 
+    Bool VulkanRenderer::MapDefaultFramebufferReadbackRect(
+            GLint x, GLint y, GLsizei width, GLsizei height, VkExtent2D imageExtent,
+            VkSurfaceTransformFlagBitsKHR preTransform, VkOffset2D* imageOffset,
+            VkExtent2D* imageCopyExtent) {
+        if (width <= 0 || height <= 0 || imageOffset == nullptr || imageCopyExtent == nullptr) {
+            return false;
+        }
+
+        const Int imageWidth = static_cast<Int>(imageExtent.width);
+        const Int imageHeight = static_cast<Int>(imageExtent.height);
+        Int mappedX = x;
+        Int mappedY = y;
+        Uint32 mappedWidth = static_cast<Uint32>(width);
+        Uint32 mappedHeight = static_cast<Uint32>(height);
+
+        // InsertPositionFixup first flips GL Y and then applies the surface transform. In pixel
+        // coordinates that gives these half-open rectangle mappings into the stored image:
+        //   identity: (x, H-y-h), 90: (y, x), 180: (W-x-w, y), 270: (H-y-h, W-x-w).
+        // Quarter turns also transpose the copied block's extent.
+        switch (preTransform) {
+        case VK_SURFACE_TRANSFORM_ROTATE_90_BIT_KHR:
+            mappedX = y;
+            mappedY = x;
+            mappedWidth = static_cast<Uint32>(height);
+            mappedHeight = static_cast<Uint32>(width);
+            break;
+        case VK_SURFACE_TRANSFORM_ROTATE_180_BIT_KHR:
+            mappedX = imageWidth - x - width;
+            mappedY = y;
+            break;
+        case VK_SURFACE_TRANSFORM_ROTATE_270_BIT_KHR:
+            mappedX = imageWidth - y - height;
+            mappedY = imageHeight - x - width;
+            mappedWidth = static_cast<Uint32>(height);
+            mappedHeight = static_cast<Uint32>(width);
+            break;
+        default:
+            mappedY = imageHeight - y - height;
+            break;
+        }
+
+        if (mappedX < 0 || mappedY < 0 || mappedWidth > imageExtent.width ||
+            mappedHeight > imageExtent.height ||
+            static_cast<Uint64>(mappedX) + mappedWidth > imageExtent.width ||
+            static_cast<Uint64>(mappedY) + mappedHeight > imageExtent.height) {
+            return false;
+        }
+        *imageOffset = {mappedX, mappedY};
+        *imageCopyExtent = {mappedWidth, mappedHeight};
+        return true;
+    }
+
+    Bool VulkanRenderer::RemapDefaultFramebufferReadback(
+            const Uint8* rawPixels, Uint32 logicalWidth, Uint32 logicalHeight,
+            VkSurfaceTransformFlagBitsKHR preTransform, SizeT texelSize, Uint8* outPixels) {
+        if (rawPixels == nullptr || outPixels == nullptr || logicalWidth == 0 || logicalHeight == 0 ||
+            texelSize == 0) {
+            return false;
+        }
+
+        const Uint32 rawWidth = IsQuarterTurnPreTransform(preTransform) ? logicalHeight : logicalWidth;
+        for (Uint32 outY = 0; outY < logicalHeight; ++outY) {
+            for (Uint32 outX = 0; outX < logicalWidth; ++outX) {
+                Uint32 srcX = outX;
+                Uint32 srcY = outY;
+                switch (preTransform) {
+                case VK_SURFACE_TRANSFORM_ROTATE_90_BIT_KHR:
+                    srcX = outY;
+                    srcY = outX;
+                    break;
+                case VK_SURFACE_TRANSFORM_ROTATE_180_BIT_KHR:
+                    srcX = logicalWidth - 1 - outX;
+                    break;
+                case VK_SURFACE_TRANSFORM_ROTATE_270_BIT_KHR:
+                    srcX = logicalHeight - 1 - outY;
+                    srcY = logicalWidth - 1 - outX;
+                    break;
+                default:
+                    srcY = logicalHeight - 1 - outY;
+                    break;
+                }
+                Memcpy(outPixels + (static_cast<SizeT>(outY) * logicalWidth + outX) * texelSize,
+                       rawPixels + (static_cast<SizeT>(srcY) * rawWidth + srcX) * texelSize,
+                       texelSize);
+            }
+        }
+        return true;
+    }
+
     Bool VulkanRenderer::ConvertReadbackPixels(const Uint8* sourcePixels, VkFormat sourceFormat,
                                                GLsizei width, GLsizei height, GLenum destinationFormat,
                                                GLenum destinationType, SizeT destinationRowStride,
@@ -2936,7 +3026,7 @@ void main() {
         succeeded = m_renderPassManager->Initialize();
         MOBILEGL_ASSERT(succeeded, "VkRenderPassManager initialization failed.");
 
-        const Uint32 maxProgramBindings = ComputeMaxProgramBindings(m_physicalDevice.properties);
+        const Uint32 maxProgramBindings = ComputeMaxProgramBindings(m_physicalDevice.properties, m_updateAfterBindLimits);
         MGLOG_I("DirectVulkan: using %u program descriptor bindings", maxProgramBindings);
         if (IsPowerVRDevice(m_physicalDevice.properties)) {
             m_config.DisablePipelineCache = true;
@@ -2970,7 +3060,9 @@ void main() {
         }
         m_programFactory = MakeUnique<ProgramFactory>(m_device, m_config, maxProgramBindings,
                                                       m_shaderDrawParametersFeatureEnabled,
-                                                      m_unformattedFloatStorageImagesEnabled);
+                                                      m_unformattedFloatStorageImagesEnabled,
+                                                      MG_Config::Features.EnableSpirvValidation,
+                                                      m_updateAfterBindLimits);
         MOBILEGL_ASSERT(m_programFactory != nullptr, "ProgramFactory creation failed.");
         // The swapchain already exists at this point (Initialize creates it first), so seed the
         // height the factory could not be told about from CreateSwapchain.
@@ -4879,6 +4971,7 @@ void main() {
             .topology = vkTopology,
             .primitiveRestartEnable = primitiveRestartEnabled,
             .patchControlPoints = static_cast<Uint32>(MG_State::pGLContext->GetPatchVertices()),
+            .viewportCount = ResolveDrawViewportCount(programObj.writesViewportIndexBuiltin),
             .polygonMode = effectivePolygonMode,
             .cullMode = cullFaceEnabled
                 ? MG_Util::ConvertCullFaceModeToVkEnum(MG_State::pGLContext->GetCullFaceMode(), invertClockwise)
@@ -5321,9 +5414,71 @@ void main() {
     }
 
 
-    void VulkanRenderer::ApplyDynamicDrawStateTail(FrameContext::FrameData& frame, const IntVec2& extent,
-                                                   Bool isDefaultFbo) {
+    // The scissor rectangle Vulkan needs for ARB_viewport_array index `index`. Vulkan has no
+    // per-viewport scissor-test TOGGLE - a scissor rectangle always applies - so an index whose
+    // GL scissor test is disabled gets the whole framebuffer, which is exactly "the test always
+    // passes" (GL 4.6 core 17.3.2).
+    VkRect2D VulkanRenderer::ComputeGLScissorRect(Uint32 index, const IntVec2& extent,
+                                                  VkSurfaceTransformFlagBitsKHR preTransform,
+                                                  Bool isDefaultFbo) const {
+        const auto& parameters = MG_State::pGLContext->GetRenderStateParameters();
+        if ((parameters.ScissorTestEnabledMask & (1u << index)) == 0) {
+            VkRect2D full{};
+            full.offset = {0, 0};
+            full.extent = {static_cast<Uint32>(extent.x()), static_cast<Uint32>(extent.y())};
+            return full;
+        }
+        const IntVec4& scissorBox = parameters.ScissorBoxes[index];
+        return isDefaultFbo ? MakeDefaultFramebufferScissorRect(scissorBox, extent, preTransform)
+                            : MakeClampedScissorRect(scissorBox, extent);
+    }
+
+    // The wide half of ApplyDynamicDrawStateTail: a pipeline built for a gl_ViewportIndex-writing
+    // program declares viewportCount > 1, and Vulkan then requires that many viewports AND that
+    // many scissors to have been set before the draw
+    // (VUID-vkCmdDraw-viewportCount-03417/-03418). Deliberately unmemoized: only conformance
+    // shaders reach it, the single-element dynamic-state shadow cannot describe an array, and
+    // leaving that shadow invalidated is what makes the next ordinary draw re-push its own
+    // single viewport instead of believing the array's element 0 is already bound.
+    void VulkanRenderer::ApplyMultiViewportDynamicState(VkCommandBuffer commandBuffer, Uint32 viewportCount,
+                                                        const IntVec2& extent,
+                                                        VkSurfaceTransformFlagBitsKHR preTransform,
+                                                        Bool isDefaultFbo) {
+        MOBILEGL_ASSERT(viewportCount <= RenderStateParameters::MAX_VIEWPORTS,
+                        "ApplyMultiViewportDynamicState: viewportCount=%u exceeds the indexed state width",
+                        viewportCount);
+        const Uint32 count = std::min<Uint32>(viewportCount, RenderStateParameters::MAX_VIEWPORTS);
+
+        Array<VkViewport, RenderStateParameters::MAX_VIEWPORTS> viewports{};
+        Array<VkRect2D, RenderStateParameters::MAX_VIEWPORTS> scissors{};
+        for (Uint32 i = 0; i < count; ++i) {
+            viewports[i] = ComputeGLViewport(i, extent, preTransform, isDefaultFbo);
+            scissors[i] = ComputeGLScissorRect(i, extent, preTransform, isDefaultFbo);
+        }
+        vkCmdSetViewport(commandBuffer, 0, count, viewports.data());
+        vkCmdSetScissor(commandBuffer, 0, count, scissors.data());
+
         auto& shadow = g_dynamicStateShadow;
+        shadow.viewportValid = false;
+        shadow.scissorValid = false;
+        shadow.dynamicTailValid = false;
+    }
+
+    void VulkanRenderer::ApplyDynamicDrawStateTail(FrameContext::FrameData& frame, const IntVec2& extent,
+                                                   Bool isDefaultFbo, Uint32 viewportCount) {
+        auto& shadow = g_dynamicStateShadow;
+        if (viewportCount > 1) {
+            // The other five Apply* still run: blend constants, depth bias, line width and the
+            // stencil masks are not per-viewport and a multi-viewport draw needs them just as
+            // much. Only the viewport/scissor pair takes the array shape.
+            ApplyBlendConstants(frame.commandBuffer);
+            ApplyPolygonOffsetState(frame.commandBuffer);
+            ApplyLineWidthState(frame.commandBuffer);
+            ApplyStencilState(frame.commandBuffer);
+            ApplyMultiViewportDynamicState(frame.commandBuffer, viewportCount, extent,
+                                           m_swapchainObject.GetPreTransform(), isDefaultFbo);
+            return;
+        }
         // One compare for the whole tail: see the gate's declaration in
         // DynamicStateShadow for why (version, extent, default-FBO flag) pins every
         // input the six Apply* below read.
@@ -5342,12 +5497,14 @@ void main() {
         DynamicStateShadow::DynamicTailKey key;
         {
             const RenderStateParameters& p = MG_State::pGLContext->GetRenderStateParameters();
-            key.viewport[0] = p.Viewport.x();
-            key.viewport[1] = p.Viewport.y();
-            key.viewport[2] = p.Viewport.z();
-            key.viewport[3] = p.Viewport.w();
-            key.depthRange[0] = p.DepthRange.x();
-            key.depthRange[1] = p.DepthRange.y();
+            // Viewport 0 and its depth range: ApplyGLViewportState reads exactly those two
+            // (per-index state for indices > 0 is keyed separately, see multiViewportKey below).
+            key.viewport[0] = p.Viewports[0].x();
+            key.viewport[1] = p.Viewports[0].y();
+            key.viewport[2] = p.Viewports[0].z();
+            key.viewport[3] = p.Viewports[0].w();
+            key.depthRange[0] = p.DepthRanges[0].x();
+            key.depthRange[1] = p.DepthRanges[0].y();
             key.blendColor[0] = p.BlendColor.x();
             key.blendColor[1] = p.BlendColor.y();
             key.blendColor[2] = p.BlendColor.z();
@@ -5362,11 +5519,11 @@ void main() {
                 key.stencilWriteMask[face] = p.StencilStates[face].WriteMask;
                 key.stencilRef[face] = p.StencilStates[face].Ref;
             }
-            key.scissorEnabled = p.ScissorTestEnabled;
-            key.scissorBox[0] = p.ScissorBox.x();
-            key.scissorBox[1] = p.ScissorBox.y();
-            key.scissorBox[2] = p.ScissorBox.z();
-            key.scissorBox[3] = p.ScissorBox.w();
+            key.scissorEnabled = (p.ScissorTestEnabledMask & 1u) != 0;
+            key.scissorBox[0] = p.ScissorBoxes[0].x();
+            key.scissorBox[1] = p.ScissorBoxes[0].y();
+            key.scissorBox[2] = p.ScissorBoxes[0].z();
+            key.scissorBox[3] = p.ScissorBoxes[0].w();
             key.extentX = extent.x();
             key.extentY = extent.y();
             key.preTransform = static_cast<Uint32>(preTransform);
@@ -5759,7 +5916,7 @@ void main() {
             const Bool idxUploadOk = UploadAndBindIndexBuffer(frame, vao, pIndexBufferView);
             MOBILEGL_ASSERT(idxUploadOk, "SetupDraw fast path: failed to upload index buffer");
         }
-        ApplyDynamicDrawStateTail(frame, snap.renderPassExtent, snap.drawFboIsDefault);
+        ApplyDynamicDrawStateTail(frame, snap.renderPassExtent, snap.drawFboIsDefault, snap.viewportCount);
         return true;
     }
 
@@ -6199,7 +6356,8 @@ void main() {
             MOBILEGL_ASSERT(idxUploadOk, "SetupDraw skipped: failed to upload index buffer");
         }
 
-        ApplyDynamicDrawStateTail(frame, renderPassEntry->extent, drawFbo->IsDefaultFramebuffer());
+        ApplyDynamicDrawStateTail(frame, renderPassEntry->extent, drawFbo->IsDefaultFramebuffer(),
+                                  ResolveDrawViewportCount(programObj.writesViewportIndexBuiltin));
 
         // Snapshot the fully resolved configuration for the consecutive-draw
         // fast path (see TrySetupDrawFastPath).
@@ -6218,6 +6376,7 @@ void main() {
                 snap.drawFbo = drawFbo.get();
                 snap.fboVersion = drawFbo->GetObjectVersion();
                 snap.drawFboIsDefault = drawFboIsDefault;
+                snap.viewportCount = ResolveDrawViewportCount(programObj.writesViewportIndexBuiltin);
                 snap.renderStateVersion = MG_State::pGLContext->GetPipelineStateVersion();
                 snap.bindGeneration = MG_State::pGLContext->GetTextureBindGeneration();
                 snap.baseTransformFlags = GetBaseTransformFlagsRaw(drawFboIsDefault);
@@ -7113,7 +7272,7 @@ void main() {
         Bool ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, resource->image, resource->layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
             srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT, srcAccessMask, VK_ACCESS_TRANSFER_WRITE_BIT,
-            resource->aspect, 0, resource->mipLevels, resource->arrayLayers);
+            resource->aspect, 0, resource->mipLevels);
         MOBILEGL_ASSERT(ok,
                         "MaterializePendingClearForTexture: failed to transition textureId=%d to TRANSFER_DST",
                         texture.GetExternalIndex());
@@ -7231,8 +7390,7 @@ void main() {
         ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, resource->image, clearLayout, sampledLayout,
             VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT,
-            VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels,
-            resource->arrayLayers);
+            VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels);
         MOBILEGL_ASSERT(ok,
                         "MaterializePendingClearForTexture: failed to transition textureId=%d to sampled layout",
                         texture.GetExternalIndex());
@@ -7270,7 +7428,7 @@ void main() {
         Bool ok = VkTextureManager::TransitionImageLayout(
             commandBuffer, resource->image, resource->layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
             srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT, srcAccessMask, VK_ACCESS_TRANSFER_WRITE_BIT,
-            resource->aspect, 0, 1, 1);
+            resource->aspect, 0, 1);
         MOBILEGL_ASSERT(ok,
                         "MaterializePendingClearForRenderbuffer: failed to transition renderbuffer %u to TRANSFER_DST",
                         renderbuffer->GetExternalIndex());
@@ -7321,7 +7479,7 @@ void main() {
             VK_ACCESS_COLOR_ATTACHMENT_READ_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
                 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
                 VK_ACCESS_TRANSFER_READ_BIT,
-            resource->aspect, 0, 1, 1);
+            resource->aspect, 0, 1);
         MOBILEGL_ASSERT(ok,
                         "MaterializePendingClearForRenderbuffer: failed to transition renderbuffer %u to steady layout",
                         renderbuffer->GetExternalIndex());
@@ -7928,6 +8086,9 @@ void main() {
             VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
             VkAccessFlags srcAccessMask = 0;
             GetImageTransitionSourceState(srcOriginalLayout, srcStageMask, srcAccessMask);
+            // Both blit regions below name `baseArrayLayer` from their binding, and a layered depth
+            // attachment puts that above 0. These barriers carry a mip range only - their layer
+            // range is every layer (see VkTextureManager::TransitionImageLayout).
             if (readIsDefaultFbo) {
                 VkImageLayout srcTrackedLayout = srcOriginalLayout;
                 Bool ok = VkTextureManager::TransitionImageLayout(
@@ -8755,30 +8916,22 @@ void main() {
         VkAccessFlags srcAccessMask = 0;
         GetImageTransitionSourceState(srcOriginalLayout, srcStageMask, srcAccessMask);
         VkImageLayout srcCopyLayout = srcOriginalLayout;
-        // The barrier has to name every layer the copy touches, not just layer 0 - otherwise the
-        // slice fix above lands the copy on layers the barrier never transitioned, which is the
-        // same defect one level down. TransitionImageLayout always starts its range at
-        // baseArrayLayer 0, so VK_REMAINING_ARRAY_LAYERS is the whole range and a superset of
-        // [baseSlice, baseSlice + depth).
-        //
-        // Not `arrayLayers`, which is 1 for a 3D image: MobileGL creates 3D images
-        // 2D_ARRAY_COMPATIBLE, and a literal 1 on one of those means "every depth slice" today but
-        // "depth slice 0" once VK_KHR_maintenance9 is enabled - i.e. it would silently become a
-        // single-slice barrier again on a newer driver. The validation layer says so by name.
-        static constexpr Uint32 kAllLayers = VK_REMAINING_ARRAY_LAYERS;
+        // The barriers below name a MIP range only. Their layer range is not a parameter:
+        // TransitionImageLayout always covers every layer of the image, which is a superset of the
+        // [baseSlice, baseSlice + depth) the slice mapping above hands the copy.
         if (srcOriginalLayout == VK_IMAGE_LAYOUT_UNDEFINED) {
             Bool srcReady = VkTextureManager::TransitionImageLayout(
                 frame.commandBuffer, srcResource->image, srcResource->layout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
                 srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
                 srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT,
-                srcResource->aspect, 0, srcResource->mipLevels, kAllLayers);
+                srcResource->aspect, 0, srcResource->mipLevels);
             MOBILEGL_ASSERT(srcReady, "%s: failed to transition undefined source image", __func__);
             srcCopyLayout = srcResource->layout;
         } else {
             Bool srcReady = VkTextureManager::TransitionImageLayout(
                 frame.commandBuffer, srcResource->image, srcCopyLayout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
                 srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
-                srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, copyAspectMask, srcMipLevel, 1, kAllLayers);
+                srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, copyAspectMask, srcMipLevel, 1);
             MOBILEGL_ASSERT(srcReady, "%s: failed to transition source image", __func__);
         }
 
@@ -8791,14 +8944,14 @@ void main() {
                 frame.commandBuffer, dstResource->image, dstResource->layout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
                 dstStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
                 dstAccessMask, VK_ACCESS_TRANSFER_WRITE_BIT,
-                dstResource->aspect, 0, dstResource->mipLevels, kAllLayers);
+                dstResource->aspect, 0, dstResource->mipLevels);
             MOBILEGL_ASSERT(dstReady, "%s: failed to transition undefined destination image", __func__);
             dstCopyLayout = dstResource->layout;
         } else {
             Bool dstReady = VkTextureManager::TransitionImageLayout(
                 frame.commandBuffer, dstResource->image, dstCopyLayout, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
                 dstStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
-                dstAccessMask, VK_ACCESS_TRANSFER_WRITE_BIT, copyAspectMask, dstMipLevel, 1, kAllLayers);
+                dstAccessMask, VK_ACCESS_TRANSFER_WRITE_BIT, copyAspectMask, dstMipLevel, 1);
             MOBILEGL_ASSERT(dstReady, "%s: failed to transition destination image", __func__);
         }
 
@@ -8840,13 +8993,13 @@ void main() {
                 frame.commandBuffer, srcResource->image, srcResource->layout, srcRestoreLayout,
                 VK_PIPELINE_STAGE_TRANSFER_BIT, srcRestoreStageMask,
                 VK_ACCESS_TRANSFER_READ_BIT, srcRestoreAccessMask,
-                srcResource->aspect, 0, srcResource->mipLevels, kAllLayers);
+                srcResource->aspect, 0, srcResource->mipLevels);
             MOBILEGL_ASSERT(srcRestored, "%s: failed to restore undefined source image layout", __func__);
         } else {
             Bool srcRestored = VkTextureManager::TransitionImageLayout(
                 frame.commandBuffer, srcResource->image, srcCopyLayout, srcRestoreLayout,
                 VK_PIPELINE_STAGE_TRANSFER_BIT, srcRestoreStageMask,
-                VK_ACCESS_TRANSFER_READ_BIT, srcRestoreAccessMask, copyAspectMask, srcMipLevel, 1, kAllLayers);
+                VK_ACCESS_TRANSFER_READ_BIT, srcRestoreAccessMask, copyAspectMask, srcMipLevel, 1);
             MOBILEGL_ASSERT(srcRestored, "%s: failed to restore source image layout", __func__);
         }
 
@@ -8858,13 +9011,13 @@ void main() {
                 frame.commandBuffer, dstResource->image, dstResource->layout, dstRestoreLayout,
                 VK_PIPELINE_STAGE_TRANSFER_BIT, dstRestoreStageMask,
                 VK_ACCESS_TRANSFER_WRITE_BIT, dstRestoreAccessMask,
-                dstResource->aspect, 0, dstResource->mipLevels, kAllLayers);
+                dstResource->aspect, 0, dstResource->mipLevels);
             MOBILEGL_ASSERT(dstRestored, "%s: failed to restore undefined destination image layout", __func__);
         } else {
             Bool dstRestored = VkTextureManager::TransitionImageLayout(
                 frame.commandBuffer, dstResource->image, dstCopyLayout, dstRestoreLayout,
                 VK_PIPELINE_STAGE_TRANSFER_BIT, dstRestoreStageMask,
-                VK_ACCESS_TRANSFER_WRITE_BIT, dstRestoreAccessMask, copyAspectMask, dstMipLevel, 1, kAllLayers);
+                VK_ACCESS_TRANSFER_WRITE_BIT, dstRestoreAccessMask, copyAspectMask, dstMipLevel, 1);
             MOBILEGL_ASSERT(dstRestored, "%s: failed to restore destination image layout", __func__);
         }
 
@@ -9023,6 +9176,9 @@ void main() {
         VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
         VkAccessFlags srcAccessMask = 0;
         GetImageTransitionSourceState(srcOriginalLayout, srcStageMask, srcAccessMask);
+        // The copy below reads `srcBinding.baseArrayLayer`, which for a glFramebufferTextureLayer
+        // attachment is any layer of the array - the barrier covers all of them (see
+        // VkTextureManager::TransitionImageLayout), so the layer being read is one it moved.
         if (readIsDefaultFbo) {
             VkImageLayout trackedLayout = srcOriginalLayout;
             Bool ok = VkTextureManager::TransitionImageLayout(
@@ -9048,19 +9204,18 @@ void main() {
         // The GL rect, aimed at the default framebuffer's stored orientation. Using the GL y
         // verbatim copied rows [y, y+h) counted from the TOP of the image, i.e. the wrong band for
         // every read that was not full-height.
-        Int32 copyOffsetX = x;
-        Int32 copyOffsetY = y;
+        VkOffset2D copyOffset{x, y};
+        VkExtent2D copyExtent{static_cast<Uint32>(width), static_cast<Uint32>(height)};
         if (readIsDefaultFbo) {
             const VkExtent2D defaultFboExtent = m_swapchainObject.GetExtent();
-            const DefaultFramebufferRectMapping mapping =
-                GetDefaultFramebufferRectMapping(m_swapchainObject.GetPreTransform());
-            copyOffsetX = MapDefaultFramebufferRectAxis(x, width, static_cast<Int>(defaultFboExtent.width),
-                                                        mapping.mirrorX);
-            copyOffsetY = MapDefaultFramebufferRectAxis(y, height, static_cast<Int>(defaultFboExtent.height),
-                                                        mapping.flipY);
+            const Bool mapped = MapDefaultFramebufferReadbackRect(
+                x, y, width, height, defaultFboExtent, m_swapchainObject.GetPreTransform(), &copyOffset,
+                &copyExtent);
+            MOBILEGL_ASSERT(mapped, "ReadPixels: default framebuffer read rectangle is out of bounds");
+            if (!mapped) return;
         }
-        copyRegion.imageOffset = {copyOffsetX, copyOffsetY, static_cast<Int32>(srcBinding.depthOffset)};
-        copyRegion.imageExtent = {static_cast<Uint32>(width), static_cast<Uint32>(height), 1};
+        copyRegion.imageOffset = {copyOffset.x, copyOffset.y, static_cast<Int32>(srcBinding.depthOffset)};
+        copyRegion.imageExtent = {copyExtent.width, copyExtent.height, 1};
         vkCmdCopyImageToBuffer(frame.commandBuffer, srcBinding.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
                                readback.GetHandle(), 1, &copyRegion);
 
@@ -9102,16 +9257,14 @@ void main() {
             // already aimed with the same mapping. The gate is exactly what made every partial
             // read of the default framebuffer come back in Vulkan row order.
             Vector<Uint8> remapped(static_cast<SizeT>(width) * static_cast<SizeT>(height) * sourceTexelSize);
-            if (RemapDefaultFboReadbackToGLOrientation(mapped, static_cast<Uint32>(width),
-                                                       static_cast<Uint32>(height), preTransform, sourceTexelSize,
-                                                       remapped.data())) {
+            if (RemapDefaultFramebufferReadback(mapped, static_cast<Uint32>(width),
+                                                static_cast<Uint32>(height), preTransform, sourceTexelSize,
+                                                remapped.data())) {
                 PackReadbackToClientOrPbo(remapped.data(), srcFormat, width, height, 1, format, type, pixels,
                                           /*applyPackImageParams=*/false, /*applyReadColorClamp=*/true);
                 return;
             }
-            // Only a quarter-turn pre-transform reaches this, and nothing in this renderer models
-            // one. MGLOG_I because the INFO builds are the ones that run conformance.
-            MGLOG_D("DirectVulkan::ReadPixels: default-FBO remap declined (w=%d h=%d preTransform=%d); falling back "
+            MGLOG_D("DirectVulkan::ReadPixels: default-FBO remap failed (w=%d h=%d preTransform=%d); falling back "
                     "to raw readback",
                     width, height, static_cast<Int>(preTransform));
         }
@@ -9492,16 +9645,15 @@ void main() {
         // The swapchain's depth/stencil image is stored display-side-up like its colour twin, so
         // the GL rect has to be mapped into that space before the copy and the copied rows
         // re-oriented afterwards - the same two halves the colour ReadPixels path applies.
-        Int32 copyOffsetX = x;
-        Int32 copyOffsetY = y;
+        VkOffset2D copyOffset{x, y};
+        VkExtent2D copyExtent{static_cast<Uint32>(width), static_cast<Uint32>(height)};
         if (defaultFramebufferOrientation) {
             const VkExtent2D defaultFboExtent = m_swapchainObject.GetExtent();
-            const DefaultFramebufferRectMapping mapping =
-                GetDefaultFramebufferRectMapping(m_swapchainObject.GetPreTransform());
-            copyOffsetX = MapDefaultFramebufferRectAxis(x, width, static_cast<Int>(defaultFboExtent.width),
-                                                        mapping.mirrorX);
-            copyOffsetY = MapDefaultFramebufferRectAxis(y, height, static_cast<Int>(defaultFboExtent.height),
-                                                        mapping.flipY);
+            const Bool mapped = MapDefaultFramebufferReadbackRect(
+                x, y, width, height, defaultFboExtent, m_swapchainObject.GetPreTransform(), &copyOffset,
+                &copyExtent);
+            MOBILEGL_ASSERT(mapped, "ReadDepthStencilPixels: default framebuffer read rectangle is out of bounds");
+            if (!mapped) return;
         }
 
         VkBufferImageCopy regions[2]{};
@@ -9513,8 +9665,8 @@ void main() {
             region.imageSubresource.mipLevel = mipLevel;
             region.imageSubresource.baseArrayLayer = baseArrayLayer;
             region.imageSubresource.layerCount = 1;
-            region.imageOffset = {copyOffsetX, copyOffsetY, 0};
-            region.imageExtent = {static_cast<Uint32>(width), static_cast<Uint32>(height), 1};
+            region.imageOffset = {copyOffset.x, copyOffset.y, 0};
+            region.imageExtent = {copyExtent.width, copyExtent.height, 1};
         }
         if (wantStencil) {
             auto& region = regions[regionCount++];
@@ -9523,8 +9675,8 @@ void main() {
             region.imageSubresource.mipLevel = mipLevel;
             region.imageSubresource.baseArrayLayer = baseArrayLayer;
             region.imageSubresource.layerCount = 1;
-            region.imageOffset = {copyOffsetX, copyOffsetY, 0};
-            region.imageExtent = {static_cast<Uint32>(width), static_cast<Uint32>(height), 1};
+            region.imageOffset = {copyOffset.x, copyOffset.y, 0};
+            region.imageExtent = {copyExtent.width, copyExtent.height, 1};
         }
         vkCmdCopyImageToBuffer(frame.commandBuffer, image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, readback.GetHandle(),
                                regionCount, regions);
@@ -9558,23 +9710,21 @@ void main() {
             Bool remapped = true;
             if (wantDepth && depthCopyBytes > 0) {
                 remappedDepth.resize(pixelCount * depthCopyBytes);
-                remapped = RemapDefaultFboReadbackToGLOrientation(depthSrc, static_cast<Uint32>(width),
-                                                                  static_cast<Uint32>(height), preTransform,
-                                                                  depthCopyBytes, remappedDepth.data());
+                remapped = RemapDefaultFramebufferReadback(depthSrc, static_cast<Uint32>(width),
+                                                           static_cast<Uint32>(height), preTransform,
+                                                           depthCopyBytes, remappedDepth.data());
             }
             if (remapped && wantStencil) {
                 remappedStencil.resize(pixelCount);
-                remapped = RemapDefaultFboReadbackToGLOrientation(stencilSrc, static_cast<Uint32>(width),
-                                                                  static_cast<Uint32>(height), preTransform, 1,
-                                                                  remappedStencil.data());
+                remapped = RemapDefaultFramebufferReadback(stencilSrc, static_cast<Uint32>(width),
+                                                           static_cast<Uint32>(height), preTransform, 1,
+                                                           remappedStencil.data());
             }
             if (remapped) {
                 if (!remappedDepth.empty()) depthSrc = remappedDepth.data();
                 if (!remappedStencil.empty()) stencilSrc = remappedStencil.data();
             } else {
-                // Only a quarter-turn pre-transform reaches this, and nothing in this renderer
-                // models one. MGLOG_I because the INFO builds are the ones that run conformance.
-                MGLOG_D("DirectVulkan::ReadDepthStencilPixels: default-FBO remap declined (w=%d h=%d "
+                MGLOG_D("DirectVulkan::ReadDepthStencilPixels: default-FBO remap failed (w=%d h=%d "
                         "preTransform=%d); falling back to raw readback",
                         width, height, static_cast<Int>(preTransform));
             }
@@ -9825,14 +9975,13 @@ void main() {
         VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
         VkAccessFlags srcAccessMask = 0;
         GetImageTransitionSourceState(originalLayout, srcStageMask, srcAccessMask);
-        // The copy below reads EVERY layer of the level, so the barrier has to name every layer
-        // too; a layerCount of 1 left an array texture's layers 1.. in whatever layout they were
-        // last left in while the transfer read them.
+        // The copy below reads EVERY layer of the level, which is exactly the range
+        // TransitionImageLayout barriers cover.
         Bool ok = VkTextureManager::TransitionImageLayout(
             frame.commandBuffer, resource->image, resource->layout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
             srcStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT,
             srcAccessMask, VK_ACCESS_TRANSFER_READ_BIT, resource->aspect,
-            static_cast<Uint32>(level), 1, VK_REMAINING_ARRAY_LAYERS);
+            static_cast<Uint32>(level), 1);
         MOBILEGL_ASSERT(ok, "%s: failed to transition texture image", __func__);
 
         VkBufferImageCopy copyRegion{};
@@ -9852,7 +10001,7 @@ void main() {
             frame.commandBuffer, resource->image, resource->layout, originalLayout,
             VK_PIPELINE_STAGE_TRANSFER_BIT, restoreStageMask,
             VK_ACCESS_TRANSFER_READ_BIT, restoreAccessMask, resource->aspect,
-            static_cast<Uint32>(level), 1, VK_REMAINING_ARRAY_LAYERS);
+            static_cast<Uint32>(level), 1);
         MOBILEGL_ASSERT(ok, "%s: failed to restore texture image layout", __func__);
 
         if (!SubmitReadbackCommandsAndWait(frame)) {
@@ -9960,7 +10109,7 @@ void main() {
             Bool transitioned = VkTextureManager::TransitionImageLayout(
                 frame.commandBuffer, resource->image, resource->layout, finalLayout,
                 VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
-                0, VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels, resource->arrayLayers);
+                0, VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels);
             MOBILEGL_ASSERT(transitioned, "GenerateMipmap: failed to transition uninitialized mip chain");
             return;
         }
@@ -11768,6 +11917,17 @@ void main() {
     }
 
     void VulkanRenderer::CreateInstance() {
+#if defined(VK_USE_PLATFORM_METAL_EXT)
+        // MoltenVK snapshots its configuration when the loader first discovers the ICD. Set
+        // this before instance-extension enumeration, while preserving an explicit user value.
+        if (std::getenv("MVK_CONFIG_USE_METAL_ARGUMENT_BUFFERS") == nullptr) {
+            if (::setenv("MVK_CONFIG_USE_METAL_ARGUMENT_BUFFERS", "1", 0) == 0) {
+                MGLOG_I("MoltenVK: enabling Metal argument buffers");
+            } else {
+                MGLOG_W("MoltenVK: could not enable Metal argument buffers before ICD discovery");
+            }
+        }
+#endif
         m_extensions = EnumerateInstanceExtensions();
         MGLOG_I("Got %d Vulkan instance extensions: ", m_extensions.size());
         for (auto& extension : m_extensions) {
@@ -11909,17 +12069,19 @@ void main() {
 
         auto debugMessengerCreateInfo = PopulateDebugMessengerCreateInfo();
         // Layers
+        const void* instanceCreatePNext = nullptr;
         if (m_validationLayersEnabled) {
             MGLOG_I("Enabling validation layer...");
             instanceInfo.enabledLayerCount = static_cast<uint32_t>(std::size(s_validationLayerNames));
             instanceInfo.ppEnabledLayerNames = s_validationLayerNames;
             // Chaining the messenger create-info is only legal with the extension on.
-            instanceInfo.pNext = debugUtilsAvailable ? &debugMessengerCreateInfo : nullptr;
+            instanceCreatePNext = debugUtilsAvailable ? &debugMessengerCreateInfo : nullptr;
         } else {
             instanceInfo.enabledLayerCount = 0;
-            instanceInfo.pNext = nullptr;
         }
 
+        instanceInfo.pNext = instanceCreatePNext;
+
         VK_VERIFY(vkCreateInstance(&instanceInfo, nullptr, &m_instance), "vkCreateInstance failed");
 
         if (debugUtilsAvailable) {
@@ -12218,6 +12380,28 @@ void main() {
         m_fillModeNonSolidFeatureEnabled = deviceFeatures.fillModeNonSolid == VK_TRUE;
         deviceFeatures.dualSrcBlend = supportedDeviceFeatures.dualSrcBlend;
         m_dualSrcBlendFeatureEnabled = deviceFeatures.dualSrcBlend == VK_TRUE;
+        // ARB_viewport_array rasterization. Without multiViewport a pipeline may declare exactly
+        // one viewport (VUID-VkPipelineViewportStateCreateInfo-viewportCount-01216), so a shader's
+        // gl_ViewportIndex can only ever select viewport 0 and the other fifteen rectangles are
+        // state with nowhere to go. The GL state stays 16 wide either way - GL 4.3 core requires
+        // MAX_VIEWPORTS >= 16 and that is a frontend promise, not a device one; this gate decides
+        // only whether a DRAW can rasterize into more than one of them.
+        deviceFeatures.multiViewport = supportedDeviceFeatures.multiViewport;
+        m_multiViewportFeatureEnabled = deviceFeatures.multiViewport == VK_TRUE;
+        m_maxRasterizableViewports =
+            m_multiViewportFeatureEnabled
+                ? std::min<Uint32>(RenderStateParameters::MAX_VIEWPORTS,
+                                   std::max<Uint32>(m_physicalDevice.properties.limits.maxViewports, 1u))
+                : 1u;
+        MGLOG_I("Vulkan: multiViewport %s; rasterizable viewports=%u (device limit %u, GL state width %u)",
+                m_multiViewportFeatureEnabled ? "enabled" : "UNAVAILABLE", m_maxRasterizableViewports,
+                m_physicalDevice.properties.limits.maxViewports,
+                static_cast<Uint32>(RenderStateParameters::MAX_VIEWPORTS));
+        if (!m_multiViewportFeatureEnabled) {
+            MGLOG_W("Vulkan: the device does not support the multiViewport feature; gl_ViewportIndex will always "
+                    "select viewport 0 and per-viewport scissor/depth-range state past index 0 cannot be "
+                    "rasterized (the state itself is still stored and queryable)");
+        }
         deviceFeatures.logicOp = supportedDeviceFeatures.logicOp;
         deviceFeatures.shaderClipDistance = supportedDeviceFeatures.shaderClipDistance;
         deviceFeatures.shaderCullDistance = supportedDeviceFeatures.shaderCullDistance;
@@ -12323,6 +12507,75 @@ void main() {
                 vkGetInstanceProcAddr(m_instance, "vkGetPhysicalDeviceFeatures2KHR"));
         }
 
+        m_updateAfterBindLimits = {};
+        VkPhysicalDeviceDescriptorIndexingFeatures descriptorIndexingFeatures{};
+        descriptorIndexingFeatures.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_FEATURES;
+        VkPhysicalDeviceDescriptorIndexingProperties descriptorIndexingProperties{};
+        descriptorIndexingProperties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_DESCRIPTOR_INDEXING_PROPERTIES;
+        const Bool descriptorIndexingCore = m_physicalDevice.properties.apiVersion >= VK_API_VERSION_1_2;
+        const Bool descriptorIndexingExtension =
+            IsExtensionSupported(availableExtensions, VK_EXT_DESCRIPTOR_INDEXING_EXTENSION_NAME);
+        auto getPhysicalDeviceProperties2 = reinterpret_cast<PFN_vkGetPhysicalDeviceProperties2>(
+            vkGetInstanceProcAddr(m_instance, "vkGetPhysicalDeviceProperties2"));
+        if (getPhysicalDeviceProperties2 == nullptr) {
+            getPhysicalDeviceProperties2 = reinterpret_cast<PFN_vkGetPhysicalDeviceProperties2>(
+                vkGetInstanceProcAddr(m_instance, "vkGetPhysicalDeviceProperties2KHR"));
+        }
+        if ((descriptorIndexingCore || descriptorIndexingExtension) && getPhysicalDeviceFeatures2 != nullptr &&
+            getPhysicalDeviceProperties2 != nullptr) {
+            VkPhysicalDeviceFeatures2 featureQuery{};
+            featureQuery.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2;
+            featureQuery.pNext = &descriptorIndexingFeatures;
+            getPhysicalDeviceFeatures2(m_physicalDevice.handle, &featureQuery);
+            VkPhysicalDeviceProperties2 propertyQuery{};
+            propertyQuery.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2;
+            propertyQuery.pNext = &descriptorIndexingProperties;
+            getPhysicalDeviceProperties2(m_physicalDevice.handle, &propertyQuery);
+
+            // This renderer emits every descriptor category listed below, including
+            // dynamic UBOs and combined image samplers. Do not enable a partial
+            // descriptor-indexing contract: it would make a later reflected program
+            // fail in the driver instead of choosing its ordinary descriptor layout.
+            const Bool allUpdateAfterBindFeatures =
+                descriptorIndexingFeatures.descriptorBindingUniformBufferUpdateAfterBind == VK_TRUE &&
+                descriptorIndexingFeatures.descriptorBindingSampledImageUpdateAfterBind == VK_TRUE &&
+                descriptorIndexingFeatures.descriptorBindingStorageImageUpdateAfterBind == VK_TRUE &&
+                descriptorIndexingFeatures.descriptorBindingStorageBufferUpdateAfterBind == VK_TRUE &&
+                descriptorIndexingFeatures.descriptorBindingUniformTexelBufferUpdateAfterBind == VK_TRUE &&
+                descriptorIndexingFeatures.descriptorBindingStorageTexelBufferUpdateAfterBind == VK_TRUE &&
+                (!deviceFeatures.robustBufferAccess || descriptorIndexingProperties.robustBufferAccessUpdateAfterBind);
+            if (allUpdateAfterBindFeatures) {
+                if (!descriptorIndexingCore && !IsExtensionAlreadyEnabled(
+                                                   enabledDeviceExtensions,
+                                                   VK_EXT_DESCRIPTOR_INDEXING_EXTENSION_NAME)) {
+                    enabledDeviceExtensions.push_back(VK_EXT_DESCRIPTOR_INDEXING_EXTENSION_NAME);
+                }
+                descriptorIndexingFeatures.pNext = const_cast<void*>(deviceCreateInfo.pNext);
+                deviceCreateInfo.pNext = &descriptorIndexingFeatures;
+                m_updateAfterBindLimits = {
+                    true,
+                    descriptorIndexingProperties.maxPerStageDescriptorUpdateAfterBindSamplers,
+                    descriptorIndexingProperties.maxPerStageDescriptorUpdateAfterBindUniformBuffers,
+                    descriptorIndexingProperties.maxPerStageDescriptorUpdateAfterBindStorageBuffers,
+                    descriptorIndexingProperties.maxPerStageDescriptorUpdateAfterBindSampledImages,
+                    descriptorIndexingProperties.maxPerStageDescriptorUpdateAfterBindStorageImages,
+                    descriptorIndexingProperties.maxPerStageUpdateAfterBindResources,
+                    descriptorIndexingProperties.maxDescriptorSetUpdateAfterBindSamplers,
+                    descriptorIndexingProperties.maxDescriptorSetUpdateAfterBindUniformBuffers,
+                    descriptorIndexingProperties.maxDescriptorSetUpdateAfterBindUniformBuffersDynamic,
+                    descriptorIndexingProperties.maxDescriptorSetUpdateAfterBindStorageBuffers,
+                    descriptorIndexingProperties.maxDescriptorSetUpdateAfterBindStorageBuffersDynamic,
+                    descriptorIndexingProperties.maxDescriptorSetUpdateAfterBindSampledImages,
+                    descriptorIndexingProperties.maxDescriptorSetUpdateAfterBindStorageImages};
+                MGLOG_I("Vulkan: update-after-bind descriptor layouts enabled");
+            } else {
+                MGLOG_I("Vulkan: descriptor indexing is present but lacks the complete update-after-bind feature set; "
+                        "using ordinary descriptor layouts");
+            }
+        } else {
+            MGLOG_I("Vulkan: descriptor indexing unavailable; using ordinary descriptor layouts");
+        }
+
         VkPhysicalDeviceIndexTypeUint8Features indexTypeUint8Features{};
         indexTypeUint8Features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INDEX_TYPE_UINT8_FEATURES;
         if (indexTypeUint8ExtensionName != nullptr) {

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h

@@ -229,6 +229,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                     GLint dstY, GLint width, GLint height, VkImageLayout srcRestoreLayout,
                                     VkImageLayout dstRestoreLayout, Bool stencilAspect);
         static SizeT GetReadbackTexelSize(VkFormat sourceFormat);
+        // Map a GL bottom-left-origin rectangle into the display-oriented swapchain image.
+        // Quarter-turn surface transforms swap the copy extent's axes.
+        static Bool MapDefaultFramebufferReadbackRect(GLint x, GLint y, GLsizei width, GLsizei height,
+                                                      VkExtent2D imageExtent,
+                                                      VkSurfaceTransformFlagBitsKHR preTransform,
+                                                      VkOffset2D* imageOffset, VkExtent2D* imageCopyExtent);
+        // Reorder a tightly packed block copied with MapDefaultFramebufferReadbackRect back into
+        // GL row order. The input block has swapped dimensions for 90/270 degree transforms.
+        static Bool RemapDefaultFramebufferReadback(const Uint8* rawPixels, Uint32 logicalWidth,
+                                                    Uint32 logicalHeight,
+                                                    VkSurfaceTransformFlagBitsKHR preTransform,
+                                                    SizeT texelSize, Uint8* outPixels);
         static Bool ConvertReadbackPixels(const Uint8* sourcePixels, VkFormat sourceFormat,
                                           GLsizei width, GLsizei height, GLenum destinationFormat,
                                           GLenum destinationType, SizeT destinationRowStride,
@@ -298,6 +310,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // The samplerAnisotropy device feature was granted, so GL_TEXTURE_MAX_ANISOTROPY_EXT is
         // honored rather than accepted-and-ignored.
         Bool IsSamplerAnisotropySupported() const { return m_samplerAnisotropyFeatureEnabled; }
+        // ARB_base_instance extends indirect command records with a non-zero firstInstance and
+        // requires gl_InstanceID to remain zero-based. Vulkan needs both features to honor that
+        // complete contract: one legalizes the command word, the other enables the shader rebase.
+        Bool IsNonZeroIndirectBaseInstanceSupported() const {
+            return m_drawIndirectFirstInstanceFeatureEnabled && m_shaderDrawParametersFeatureEnabled;
+        }
         // Ensures the frame command buffer is recording (same lazy pattern as
         // SetupDraw) and writes a bottom-of-pipe timestamp into the current
         // frame's pool. Null when unsupported or the pool is exhausted.
@@ -537,6 +555,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool m_shaderDrawParametersExtensionEnabled = false;
         Bool m_shaderDrawParametersFeatureEnabled = false;
         Bool m_unformattedFloatStorageImagesEnabled = false;
+        // Set only after descriptor-indexing feature AND property queries prove that
+        // update-after-bind is legal for every descriptor category this renderer emits.
+        ProgramFactory::UpdateAfterBindLimits m_updateAfterBindLimits{};
         // fillModeNonSolid gates VK_POLYGON_MODE_LINE/_POINT (glPolygonMode); independentBlend gates
         // per-draw-buffer color write masks (glColorMaski). Both are cached at device creation and
         // drive a runtime fallback when the device lacks them.
@@ -547,6 +568,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // needs no feature). Both cached at device creation and drive a hard-fail-at-draw when absent.
         Bool m_dualSrcBlendFeatureEnabled = false;
         Bool m_primitiveTopologyListRestartFeatureEnabled = false;
+        // multiViewport gates rasterizing into more than one of ARB_viewport_array's 16 viewports
+        // (gl_ViewportIndex). m_maxRasterizableViewports is min(MAX_VIEWPORTS, device limit), or 1
+        // when the feature is off, and is the viewportCount a gl_ViewportIndex-writing pipeline
+        // declares - it is NOT what GL_MAX_VIEWPORTS reports, which is the frontend state width.
+        Bool m_multiViewportFeatureEnabled = false;
+        Uint32 m_maxRasterizableViewports = 1;
         // Union of shader stages sampled-read barriers may name; built at device creation
         // because geometry/tessellation stage bits are invalid in a barrier when their
         // feature is off (VUID-vkCmdPipelineBarrier-srcStageMask-04090/-04091), and
@@ -830,6 +857,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // re-resolve just the pipeline against the active pass; a change that
             // flips it must fall back to the full path's pass selection.
             Bool drawUsesDepthStencil = false;
+            // The snapshotting draw's pipeline viewportCount. A pure function of the PROGRAM
+            // (writesViewportIndexBuiltin) and of a device feature fixed at renderer init, both
+            // of which the programLifetimeId/programVersion guards above already pin - carried
+            // here so the fast path does not re-fetch the program object to re-derive it.
+            Uint32 viewportCount = 1;
             IntVec2 renderPassExtent = {0, 0};
             // colorAttachmentCount of the snapshotting draw's render pass: the
             // pipeline-state hash input, so the fast path can refresh that hash and
@@ -1120,7 +1152,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // The per-draw dynamic-state tail (viewport, scissor, blend constants, depth
         // bias, line width, stencil), gated behind one render-state-parameters-version
         // compare per command buffer - see the gate fields in DynamicStateShadow.
-        void ApplyDynamicDrawStateTail(FrameContext::FrameData& frame, const IntVec2& extent, Bool isDefaultFbo);
+        // viewportCount is the bound pipeline's declared viewport count: 1 for every program that
+        // does not write gl_ViewportIndex (the memoized fast path), otherwise the renderer's
+        // rasterizable viewport count, which takes the unmemoized array path.
+        void ApplyDynamicDrawStateTail(FrameContext::FrameData& frame, const IntVec2& extent, Bool isDefaultFbo,
+                                       Uint32 viewportCount = 1);
+        void ApplyMultiViewportDynamicState(VkCommandBuffer commandBuffer, Uint32 viewportCount, const IntVec2& extent,
+                                            VkSurfaceTransformFlagBitsKHR preTransform, Bool isDefaultFbo);
+        VkRect2D ComputeGLScissorRect(Uint32 index, const IntVec2& extent,
+                                      VkSurfaceTransformFlagBitsKHR preTransform, Bool isDefaultFbo) const;
+        // How many viewports a draw with this program rasterizes into: 1 unless the program
+        // assigns gl_ViewportIndex AND the device enabled multiViewport. Both the pipeline's
+        // baked viewportCount and the dynamic arrays come from this one answer, so they cannot
+        // disagree.
+        Uint32 ResolveDrawViewportCount(Bool programWritesViewportIndex) const {
+            return programWritesViewportIndex && m_multiViewportFeatureEnabled ? m_maxRasterizableViewports : 1u;
+        }
 
         Bool UploadAndBindVertexBuffers(VkCommandBuffer commandBuffer, const MG_State::GLState::VertexArrayObject& vao,
                                         const ProgramFactory::VkProgramObject& programObj,

MobileGL/MG_Backend/Init.cpp

@@ -10,6 +10,9 @@
 #include <Config.h>
 #include <MG_Util/BackendLoaders/OpenGL/Loader.h>
 #include <MG_Util/Converters/MGToStr/GLExtensionConverter.h>
+#if defined(MOBILEGL_ENABLE_DILIGENT)
+#include <MG_Backend/Diligent/BackendObject_Diligent.h>
+#endif
 
 namespace MobileGL::MG_Backend {
     void LogBackendInfo() {
@@ -55,6 +58,11 @@ namespace MobileGL::MG_Backend {
         case BackendType::DirectVulkan:
             pActiveBackendObject = MakeUnique<DirectVulkan::BackendObject_DirectVulkan>();
             break;
+#if defined(MOBILEGL_ENABLE_DILIGENT)
+        case BackendType::DiligentVulkan:
+            pActiveBackendObject = MakeUnique<DiligentBackend::BackendObject_Diligent>();
+            break;
+#endif
         case BackendType::Unknown:
         default:
             MGLOG_W("Unknown backend type, defaulting to unknown backend");

MobileGL/MG_Impl/GLImpl/Buffer/GL_Buffer.cpp

@@ -18,6 +18,7 @@
 #include <MG_Util/Converters/GLToStr/GLEnumConverter.h>
 #include <MG_Util/Converters/GLToMG/BufferEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/BufferEnumConverter.h>
+#include <MG_Util/Texture/PixelStoreProcessor.h>
 
 namespace MobileGL::MG_Impl::GLImpl {
     namespace {
@@ -31,6 +32,8 @@ namespace MobileGL::MG_Impl::GLImpl {
             NamedBufferData,
             NamedBufferSubData,
             CopyNamedBufferSubData,
+            ClearBufferData,
+            ClearBufferSubData,
             ClearNamedBufferData,
             ClearNamedBufferSubData,
             MapBufferRange,
@@ -65,6 +68,10 @@ namespace MobileGL::MG_Impl::GLImpl {
                 return "NamedBufferSubData";
             case BufferOp::CopyNamedBufferSubData:
                 return "CopyNamedBufferSubData";
+            case BufferOp::ClearBufferData:
+                return "ClearBufferData";
+            case BufferOp::ClearBufferSubData:
+                return "ClearBufferSubData";
             case BufferOp::ClearNamedBufferData:
                 return "ClearNamedBufferData";
             case BufferOp::ClearNamedBufferSubData:
@@ -143,16 +150,6 @@ namespace MobileGL::MG_Impl::GLImpl {
                 return 0;
             }
 
-            // The pattern is replicated verbatim, which is only the whole story while the client
-            // layout already matches the internal format - the case every entry point in practice
-            // uses, and the only one the conversion machinery here can express. Say so rather than
-            // quietly writing a differently-sized pattern.
-            const SizeT sourceSize = MG_Util::GetInputBytesPerPixel(inputFormat, pixelType);
-            if (sourceSize != elementSize) {
-                MGLOG_W_ONCE("%s: clear pattern is %zu bytes but internalformat 0x%X stores %zu; "
-                        "converting between them is not implemented",
-                        GetBufferOpName(op), sourceSize, internalformat, elementSize);
-            }
             return elementSize;
         }
 
@@ -194,27 +191,59 @@ namespace MobileGL::MG_Impl::GLImpl {
             return true;
         }
 
-        void ClearNamedBufferRange_State(GLuint buffer, GLenum internalformat, GLintptr offset, GLsizeiptr size,
-                                         GLenum format, GLenum type, const void* data, BufferOp op) {
+        Bool BuildClearPattern(GLenum internalformat, GLenum format, GLenum type, const void* data,
+                               SizeT patternSize, BufferOp op, Vector<Uint8>& pattern) {
+            const TextureInternalFormat internal = MG_Util::ConvertGLEnumToTextureInternalFormat(internalformat);
+            const TextureInputFormat inputFormat = MG_Util::ConvertGLEnumToTextureInputFormat(format);
+            const TexturePixelDataType inputType = MG_Util::ConvertGLEnumToTexturePixelDataType(type);
+
+            Vector<Uint8> zeroInput;
+            const void* inputPixel = data;
+            if (inputPixel == nullptr) {
+                const SizeT inputSize = MG_Util::GetInputBytesPerPixel(inputFormat, inputType);
+                if (inputSize == 0) {
+                    MG_State::pGLContext->RecordError(
+                        ErrorCode::InvalidValue,
+                        MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", GetBufferOpName(op),
+                                                     "format and type do not describe a source pixel."));
+                    return false;
+                }
+                zeroInput.resize(inputSize);
+                inputPixel = zeroInput.data();
+            }
+
+            if (!MG_Util::PixelStoreProcessor::ConvertOnePixelToInternal(
+                    internal, inputFormat, inputType, inputPixel, pattern)) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidValue,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", GetBufferOpName(op),
+                        std::format("Cannot convert one ({}, {}) pixel into internalformat 0x{:X}.",
+                                    MG_Util::ConvertGLEnumToString(format), MG_Util::ConvertGLEnumToString(type),
+                                    internalformat)));
+                return false;
+            }
+
+            if (data == nullptr) {
+                // GL defines a null clear value as all zero bits in the destination store, while
+                // retaining the format/type validation above.
+                pattern.assign(patternSize, 0);
+            }
+            return true;
+        }
+
+        void ClearBufferRange_State(const SharedPtr<MG_State::GLState::BufferObject>& bufferObject,
+                                    GLenum internalformat, GLintptr offset, GLsizeiptr size,
+                                    GLenum format, GLenum type, const void* data, BufferOp op) {
             const SizeT patternSize = GetClearPatternSize(internalformat, format, type, op);
             if (patternSize == 0) return;
-
-            auto bufferObject = GetNamedBufferObject(buffer, op);
-            if (!bufferObject) return;
             if (!ValidateBufferClearRange(bufferObject, offset, size, patternSize, op)) return;
             if (size == 0) return;
 
-            Vector<Uint8> clearData(static_cast<SizeT>(size));
-            if (data) {
-                const auto* pattern = static_cast<const Uint8*>(data);
-                for (SizeT at = 0; at < clearData.size(); at += patternSize) {
-                    Memcpy(clearData.data() + at, pattern, patternSize);
-                }
-            } else {
-                Memset(clearData.data(), 0, clearData.size());
-            }
-
-            bufferObject->UploadSubData({clearData.data(), clearData.size()}, static_cast<SizeT>(offset));
+            Vector<Uint8> pattern;
+            if (!BuildClearPattern(internalformat, format, type, data, patternSize, op, pattern)) return;
+            bufferObject->FillSubData({pattern.data(), pattern.size()}, static_cast<SizeT>(offset),
+                                      static_cast<SizeT>(size));
         }
 
         auto& GetBufferBindingSlot(BufferTarget target) {
@@ -1197,17 +1226,34 @@ namespace MobileGL::MG_Impl::GLImpl {
                                         static_cast<SizeT>(writeOffset), static_cast<SizeT>(size));
     }
 
+    void ClearBufferData_State(GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data) {
+        auto bufferObject = GetBoundBufferObject(target, BufferOp::ClearBufferData);
+        if (!bufferObject) return;
+        ClearBufferRange_State(bufferObject, internalformat, 0, static_cast<GLsizeiptr>(bufferObject->GetSize()), format,
+                               type, data, BufferOp::ClearBufferData);
+    }
+
+    void ClearBufferSubData_State(GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size,
+                                  GLenum format, GLenum type, const void* data) {
+        auto bufferObject = GetBoundBufferObject(target, BufferOp::ClearBufferSubData);
+        if (!bufferObject) return;
+        ClearBufferRange_State(bufferObject, internalformat, offset, size, format, type, data,
+                               BufferOp::ClearBufferSubData);
+    }
+
     void ClearNamedBufferData_State(GLuint buffer, GLenum internalformat, GLenum format, GLenum type, const void* data) {
         auto bufferObject = GetNamedBufferObject(buffer, BufferOp::ClearNamedBufferData);
         if (!bufferObject) return;
-        ClearNamedBufferRange_State(buffer, internalformat, 0, static_cast<GLsizeiptr>(bufferObject->GetSize()), format,
-                                    type, data, BufferOp::ClearNamedBufferData);
+        ClearBufferRange_State(bufferObject, internalformat, 0, static_cast<GLsizeiptr>(bufferObject->GetSize()), format,
+                               type, data, BufferOp::ClearNamedBufferData);
     }
 
     void ClearNamedBufferSubData_State(GLuint buffer, GLenum internalformat, GLintptr offset, GLsizeiptr size,
                                        GLenum format, GLenum type, const void* data) {
-        ClearNamedBufferRange_State(buffer, internalformat, offset, size, format, type, data,
-                                    BufferOp::ClearNamedBufferSubData);
+        auto bufferObject = GetNamedBufferObject(buffer, BufferOp::ClearNamedBufferSubData);
+        if (!bufferObject) return;
+        ClearBufferRange_State(bufferObject, internalformat, offset, size, format, type, data,
+                               BufferOp::ClearNamedBufferSubData);
     }
 
     void* MapNamedBuffer_State(GLuint buffer, GLenum access) {
@@ -1662,6 +1708,15 @@ namespace MobileGL::MG_Impl::GLImpl {
         CopyNamedBufferSubData_State(readBuffer, writeBuffer, readOffset, writeOffset, size);
     }
 
+    void ClearBufferData(GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data) {
+        ClearBufferData_State(target, internalformat, format, type, data);
+    }
+
+    void ClearBufferSubData(GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format,
+                            GLenum type, const void* data) {
+        ClearBufferSubData_State(target, internalformat, offset, size, format, type, data);
+    }
+
     void ClearNamedBufferData(GLuint buffer, GLenum internalformat, GLenum format, GLenum type, const void* data) {
         ClearNamedBufferData_State(buffer, internalformat, format, type, data);
     }

MobileGL/MG_Impl/GLImpl/Buffer/GL_Buffer.h

@@ -27,6 +27,9 @@ namespace MobileGL::MG_Impl::GLImpl {
     void NamedBufferSubData(GLuint buffer, GLintptr offset, GLsizeiptr size, const void* data);
     void CopyNamedBufferSubData(GLuint readBuffer, GLuint writeBuffer, GLintptr readOffset, GLintptr writeOffset,
                                 GLsizeiptr size);
+    void ClearBufferData(GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data);
+    void ClearBufferSubData(GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format,
+                            GLenum type, const void* data);
     void ClearNamedBufferData(GLuint buffer, GLenum internalformat, GLenum format, GLenum type, const void* data);
     void ClearNamedBufferSubData(GLuint buffer, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format,
                                  GLenum type, const void* data);

MobileGL/MG_Impl/GLImpl/Exporting/Definitions.cpp

@@ -969,14 +969,14 @@ DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribL3dv, GLuint index, const GLdoub
 DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribL4dv, GLuint index, const GLdouble* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, VertexAttribL4dv, index, v)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, VertexAttribLPointer, GLuint index, GLint size, GLenum type, GLsizei stride, const void* pointer) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, VertexAttribLPointer, index, size, type, stride, pointer)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetVertexAttribLdv, GLuint index, GLenum pname, GLdouble* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetVertexAttribLdv, index, pname, params)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ViewportArrayv, GLuint first, GLsizei count, const GLfloat* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ViewportArrayv, first, count, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ViewportIndexedf, GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ViewportIndexedf, index, x, y, w, h)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ViewportIndexedfv, GLuint index, const GLfloat* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ViewportIndexedfv, index, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorArrayv, GLuint first, GLsizei count, const GLint* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ScissorArrayv, first, count, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorIndexed, GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ScissorIndexed, index, left, bottom, width, height)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ScissorIndexedv, GLuint index, const GLint* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ScissorIndexedv, index, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, DepthRangeArrayv, GLuint first, GLsizei count, const GLdouble* v) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DepthRangeArrayv, first, count, v)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, DepthRangeIndexed, GLuint index, GLdouble n, GLdouble f) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, DepthRangeIndexed, index, n, f)
+DECLARE_GL_FUNCTION_HEAD(void, ViewportArrayv, GLuint first, GLsizei count, const GLfloat* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ViewportArrayv, first, count, v)
+DECLARE_GL_FUNCTION_HEAD(void, ViewportIndexedf, GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ViewportIndexedf, index, x, y, w, h)
+DECLARE_GL_FUNCTION_HEAD(void, ViewportIndexedfv, GLuint index, const GLfloat* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ViewportIndexedfv, index, v)
+DECLARE_GL_FUNCTION_HEAD(void, ScissorArrayv, GLuint first, GLsizei count, const GLint* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ScissorArrayv, first, count, v)
+DECLARE_GL_FUNCTION_HEAD(void, ScissorIndexed, GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ScissorIndexed, index, left, bottom, width, height)
+DECLARE_GL_FUNCTION_HEAD(void, ScissorIndexedv, GLuint index, const GLint* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ScissorIndexedv, index, v)
+DECLARE_GL_FUNCTION_HEAD(void, DepthRangeArrayv, GLuint first, GLsizei count, const GLdouble* v) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DepthRangeArrayv, first, count, v)
+DECLARE_GL_FUNCTION_HEAD(void, DepthRangeIndexed, GLuint index, GLdouble n, GLdouble f) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DepthRangeIndexed, index, n, f)
 DECLARE_GL_FUNCTION_HEAD(void, GetFloati_v, GLenum target, GLuint index, GLfloat* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GetFloati_v, target, index, data)
 DECLARE_GL_FUNCTION_HEAD(void, GetDoublei_v, GLenum target, GLuint index, GLdouble* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GetDoublei_v, target, index, data)
 DECLARE_GL_FUNCTION_HEAD(void, DrawArraysInstancedBaseInstance, GLenum mode, GLint first, GLsizei count, GLsizei instancecount, GLuint baseinstance) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawArraysInstancedBaseInstance, mode, first, count, instancecount, baseinstance)
@@ -985,8 +985,8 @@ DECLARE_GL_FUNCTION_HEAD(void, DrawElementsInstancedBaseVertexBaseInstance, GLen
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetActiveAtomicCounterBufferiv, GLuint program, GLuint bufferIndex, GLenum pname, GLint* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetActiveAtomicCounterBufferiv, program, bufferIndex, pname, params)
 DECLARE_GL_FUNCTION_HEAD(void, DrawTransformFeedbackInstanced, GLenum mode, GLuint id, GLsizei instancecount) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawTransformFeedbackInstanced, mode, id, instancecount)
 DECLARE_GL_FUNCTION_HEAD(void, DrawTransformFeedbackStreamInstanced, GLenum mode, GLuint id, GLuint stream, GLsizei instancecount) DECLARE_GL_FUNCTION_END_NO_RETURN(void, DrawTransformFeedbackStreamInstanced, mode, id, stream, instancecount)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ClearBufferData, GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ClearBufferData, target, internalformat, format, type, data)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, ClearBufferSubData, GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, ClearBufferSubData, target, internalformat, offset, size, format, type, data)
+DECLARE_GL_FUNCTION_HEAD(void, ClearBufferData, GLenum target, GLenum internalformat, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ClearBufferData, target, internalformat, format, type, data)
+DECLARE_GL_FUNCTION_HEAD(void, ClearBufferSubData, GLenum target, GLenum internalformat, GLintptr offset, GLsizeiptr size, GLenum format, GLenum type, const void* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, ClearBufferSubData, target, internalformat, offset, size, format, type, data)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, GetInternalformati64v, GLenum target, GLenum internalformat, GLenum pname, GLsizei count, GLint64* params) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetInternalformati64v, target, internalformat, pname, count, params)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, InvalidateTexSubImage, GLuint texture, GLint level, GLint xoffset, GLint yoffset, GLint zoffset, GLsizei width, GLsizei height, GLsizei depth) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, InvalidateTexSubImage, texture, level, xoffset, yoffset, zoffset, width, height, depth)
 DECLARE_GL_FUNCTION_STUB_HEAD(void, InvalidateTexImage, GLuint texture, GLint level) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, InvalidateTexImage, texture, level)

MobileGL/MG_Impl/GLImpl/Getter/GL_Getter.cpp

@@ -16,6 +16,7 @@
 #include <MG_State/GLState/ErrorState/ErrorInfo.h>
 #include <MG_Util/Converters/GLToStr/GLEnumConverter.h>
 #include <MG_Util/Converters/GLToMG/BufferEnumConverter.h>
+#include <MG_Util/Converters/GLToMG/RenderStateEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/FramebufferEnumConverter.h>
 #include <MG_Util/Converters/MGToGL/ErrorCodeConverter.h>
 #include <MG_Util/Converters/MGToGL/TextureEnumConverter.h>
@@ -27,6 +28,11 @@
 #include <MG_Backend/BackendObjects.h>
 
 namespace MobileGL::MG_Impl::GLImpl {
+    // Declared rather than #included from GL_RenderState.h on purpose: that header also declares
+    // a free function named BlendEquation, which would hide the ::MobileGL::BlendEquation enum
+    // this file's blend-state queries name unqualified.
+    GLboolean IsEnabledi(GLenum target, GLuint index);
+
     namespace {
         enum class IndexedBufferQueryKind {
             Binding,
@@ -339,26 +345,70 @@ namespace MobileGL::MG_Impl::GLImpl {
             return sampler ? static_cast<GLint>(sampler->GetExternalIndex()) : 0;
         }
 
-        // The ARB_viewport_array indexed rectangles. MobileGL keeps exactly one viewport, one
-        // scissor box and one depth range, so every in-range index answers with that single
-        // value - but it has to come from the frontend state the non-indexed getters read.
-        // The generic path at the bottom of GetIntegeri_v is a raw backend passthrough that
-        // has no case for these, so routing them through it returned zeros.
+        // The ARB_viewport_array indexed rectangles. Each of these is genuinely per-viewport
+        // frontend state (RenderStateParameters::Viewports / ScissorBoxes / DepthRanges), so the
+        // indexed getters must read the indexed storage - the generic path at the bottom of
+        // GetIntegeri_v is a raw backend passthrough that has no case for them and returned
+        // zeros, and routing them to the NON-indexed getter (what this used to do) answered every
+        // index with viewport 0's value, which is what
+        // KHR-GL43.viewport_array.{viewport,scissor,depth_range}_api caught.
         Bool IsIndexedViewportQuery(GLenum target) {
             return target == GL_VIEWPORT || target == GL_SCISSOR_BOX || target == GL_DEPTH_RANGE;
         }
 
-        // ARB_viewport_array: `index` selects a viewport and MAX_VIEWPORTS bounds it.
+        // Component count of an indexed viewport-array query, so every width of getter writes the
+        // caller's whole buffer instead of just element 0 (GL 4.6 core 22.1).
+        GLsizei IndexedViewportQueryComponents(GLenum target) {
+            return target == GL_DEPTH_RANGE ? 2 : 4;
+        }
+
+        // ARB_viewport_array: `index` selects a viewport and MAX_VIEWPORTS bounds it. The bound is
+        // the frontend's own state width, which is also exactly what GL_MAX_VIEWPORTS reports -
+        // taking it from the backend caps instead would let a device limit of 1 (a Vulkan device
+        // without the multiViewport feature) make index 1 illegal even though the state exists.
         Bool ValidateViewportQueryIndex(GLuint index, const char* caller) {
-            GLint maxViewports = 0;
-            GetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
-            if (index < static_cast<GLuint>(std::max(maxViewports, 1))) return true;
+            if (index < RenderStateParameters::MAX_VIEWPORTS) return true;
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller, "Viewport index is out of range."));
             return false;
         }
 
+        // The indexed viewport/scissor/depth-range state as floats, which is the widest lossless
+        // shape MobileGL stores (the viewport really is float state; the scissor box is integral
+        // and well inside float's exact range, and every depth range is in [0, 1]). Every indexed
+        // getter width funnels through this so they can never disagree with each other.
+        void ReadIndexedViewportStateFloat(GLenum target, GLuint index, GLfloat* out) {
+            switch (target) {
+            case GL_VIEWPORT: {
+                const FloatVec4& viewport = MG_State::pGLContext->GetViewportIndexed(index);
+                out[0] = viewport.x();
+                out[1] = viewport.y();
+                out[2] = viewport.z();
+                out[3] = viewport.w();
+                return;
+            }
+            case GL_SCISSOR_BOX: {
+                const IntVec4& box = MG_State::pGLContext->GetScissorBoxIndexed(index);
+                out[0] = static_cast<GLfloat>(box.x());
+                out[1] = static_cast<GLfloat>(box.y());
+                out[2] = static_cast<GLfloat>(box.z());
+                out[3] = static_cast<GLfloat>(box.w());
+                return;
+            }
+            case GL_DEPTH_RANGE: {
+                const FloatVec2& range = MG_State::pGLContext->GetDepthRangeIndexed(index);
+                out[0] = range.x();
+                out[1] = range.y();
+                return;
+            }
+            default:
+                MOBILEGL_ASSERT(false, "ReadIndexedViewportStateFloat: unexpected target 0x%x",
+                                static_cast<Uint32>(target));
+                return;
+            }
+        }
+
         void CopyIntsToBooleans(const GLint* src, SizeT count, GLboolean* dst) {
             for (SizeT i = 0; i < count; ++i) {
                 dst[i] = src[i] ? GL_TRUE : GL_FALSE;
@@ -629,6 +679,17 @@ namespace MobileGL::MG_Impl::GLImpl {
             params[1] = dynamicParameters.ViewportBoundsRangeMax;
             return;
         }
+        // Viewport 0's rectangle, verbatim. Falling through to the integer width below would
+        // round the fractional rectangle a glViewportIndexedf(0, ...) is allowed to set, and
+        // glGetFloatv(GL_VIEWPORT) is a lossless query of float state.
+        case GL_VIEWPORT: {
+            const FloatVec4& viewport = MG_State::pGLContext->GetViewportIndexed(0);
+            params[0] = viewport.x();
+            params[1] = viewport.y();
+            params[2] = viewport.z();
+            params[3] = viewport.w();
+            return;
+        }
         case GL_MIN_FRAGMENT_INTERPOLATION_OFFSET:
         case GL_MAX_FRAGMENT_INTERPOLATION_OFFSET:
         case GL_FRAGMENT_INTERPOLATION_OFFSET_BITS: {
@@ -792,15 +853,32 @@ namespace MobileGL::MG_Impl::GLImpl {
             return;
         }
 
+        // GL 4.6 core 22.1: an indexed query answers EVERY indexed state, and GL_SCISSOR_TEST is
+        // indexed by viewport just like GL_BLEND is by draw buffer. Without this the integer
+        // width fell through to the backend passthrough and answered GL_INVALID_ENUM, which is
+        // the sticky error KHR-GL43.viewport_array.queries trips over at its next error check.
+        if (MG_Util::ConvertGLEnumToCapabilityInput(target) != CapabilityInput::Unknown) {
+            *data = IsEnabledi(target, index);
+            return;
+        }
+
         switch (target) {
         // ARB_viewport_array queries the indexed rectangles through glGetIntegeri_v as well
-        // (gl4cMultiBindTests and the viewport_array group both do). The frontend keeps one
-        // viewport and one scissor box, so every in-range index reports that one.
+        // (gl4cMultiBindTests and the viewport_array group both do).
         case GL_VIEWPORT:
         case GL_SCISSOR_BOX:
+        case GL_DEPTH_RANGE: {
             if (!ValidateViewportQueryIndex(index, __func__)) return;
-            GetIntegerv(target, data);
+            GLfloat values[4] = {};
+            ReadIndexedViewportStateFloat(target, index, values);
+            const GLsizei components = IndexedViewportQueryComponents(target);
+            for (GLsizei i = 0; i < components; ++i) {
+                // Round, not truncate: glGetIntegerv on floating-point state rounds to nearest
+                // (GL 4.6 core 22.2), so a 255.875-wide viewport reads back as 256 and not 255.
+                data[i] = static_cast<GLint>(std::lround(values[i]));
+            }
             return;
+        }
         // The vertex buffer binding points of the vertex array object that is bound. Indexed by
         // binding point, not by attribute (GL 4.6 core 10.3.1).
         case GL_VERTEX_BINDING_BUFFER:
@@ -927,7 +1005,10 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         if (IsIndexedViewportQuery(target)) {
             if (!ValidateViewportQueryIndex(index, __func__)) return;
-            GetFloatv(target, data);
+            // Verbatim, NOT via the integer width: the viewport is float state and
+            // KHR-GL43.viewport_array.viewport_api compares the read-back with ==, so a
+            // glViewportIndexedf(i, 0.125f, ...) has to come back as 0.125f exactly.
+            ReadIndexedViewportStateFloat(target, index, data);
             return;
         }
         GLint ints[4] = {};
@@ -944,7 +1025,12 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         if (IsIndexedViewportQuery(target)) {
             if (!ValidateViewportQueryIndex(index, __func__)) return;
-            GetDoublev(target, data);
+            GLfloat values[4] = {};
+            ReadIndexedViewportStateFloat(target, index, values);
+            const GLsizei components = IndexedViewportQueryComponents(target);
+            for (GLsizei i = 0; i < components; ++i) {
+                data[i] = static_cast<GLdouble>(values[i]);
+            }
             return;
         }
         GLint ints[4] = {};
@@ -1020,7 +1106,12 @@ namespace MobileGL::MG_Impl::GLImpl {
         // frontend-only value simply is not in the driver's table.
         GLint values[4] = {};
         GetIntegeri_v(target, index, values);
-        *data = static_cast<GLint64>(values[0]);
+        // The viewport-array rectangles are the only multi-component indexed state here; every
+        // other pname is scalar, so widening element 0 alone would silently truncate them.
+        const GLsizei components = IsIndexedViewportQuery(target) ? IndexedViewportQueryComponents(target) : 1;
+        for (GLsizei i = 0; i < components; ++i) {
+            data[i] = static_cast<GLint64>(values[i]);
+        }
     }
 
     void GetInteger64v(GLenum pname, GLint64* params) {
@@ -2192,7 +2283,15 @@ namespace MobileGL::MG_Impl::GLImpl {
             params[1] = dynamicParameters.MaxViewportHeight;
             break;
         case GL_MAX_VIEWPORTS:
-            *params = dynamicParameters.MaxViewports;
+            // The frontend's own state width, not the backend's device limit. GL 4.3 core
+            // requires MAX_VIEWPORTS >= 16 and every indexed viewport entry point validates
+            // against RenderStateParameters::MAX_VIEWPORTS, so reporting anything else would
+            // either advertise viewports the state cannot hold or reject indices it can. A
+            // Vulkan device without the multiViewport feature reports maxViewports == 1, which
+            // limits what can be RASTERIZED to more than one rectangle (see the multiViewport
+            // gate in VulkanRenderer), not what the GL state can hold; caps.MaxViewports keeps
+            // carrying that device number for exactly that decision.
+            *params = static_cast<GLint>(RenderStateParameters::MAX_VIEWPORTS);
             break;
         case GL_MINOR_VERSION:
             *params = rendererInfo.RendererGLInfo.TargetGLVersion.Minor;

MobileGL/MG_Impl/GLImpl/RenderState/GL_RenderState.cpp

@@ -20,28 +20,118 @@ namespace MobileGL::MG_Impl::GLImpl {
         return std::clamp(static_cast<Float>(value), 0.0f, 1.0f);
     }
 
-    static Bool ValidateIndexedBlendCapability(GLenum target, GLuint index, const char* functionName) {
-        if (target != GL_BLEND) {
+    // GL 4.6 core 17.3.2 and 22.1 give exactly two indexed capabilities: GL_BLEND, indexed by
+    // draw buffer, and GL_SCISSOR_TEST, indexed by viewport. They have DIFFERENT bounds
+    // (MAX_DRAW_BUFFERS vs MAX_VIEWPORTS), so the limit is picked per target rather than shared.
+    static Bool ValidateIndexedCapability(GLenum target, GLuint index, const char* functionName) {
+        GLuint limit = 0;
+        const char* indexName = nullptr;
+        switch (target) {
+        case GL_BLEND:
+            limit = MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS;
+            indexName = "Buffer";
+            break;
+        case GL_SCISSOR_TEST:
+            limit = RenderStateParameters::MAX_VIEWPORTS;
+            indexName = "Viewport";
+            break;
+        default:
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidEnum,
                 MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
-                                             "Only GL_BLEND is supported for indexed capability state."));
+                                             "Only GL_BLEND and GL_SCISSOR_TEST are supported for indexed "
+                                             "capability state."));
             return false;
         }
 
-        if (index >= MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS) {
+        if (index >= limit) {
             MG_State::pGLContext->RecordError(
                 ErrorCode::InvalidValue,
-                MakeUnique<GenericErrorInfo>(
-                    "MG_Impl/GLImpl", functionName,
-                    "Buffer index " + std::to_string(index) + " is out of range. Max supported is " +
-                        std::to_string(MG_State::GLState::FramebufferObject::MAX_DRAW_BUFFERS - 1) + "."));
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                             String(indexName) + " index " + std::to_string(index) +
+                                                 " is out of range. Max supported is " + std::to_string(limit - 1) +
+                                                 "."));
             return false;
         }
 
         return true;
     }
 
+    // ------------------ ARB_viewport_array parameter validation ------------------
+    // All three families share the same two shapes, so they share the two checkers. GL 4.6 core
+    // 13.6.1/17.3.2: an out-of-range index is GL_INVALID_VALUE, and so is a negative width or
+    // height. `first + count == MAX_VIEWPORTS` is LEGAL - only strictly greater is an error,
+    // which KHR-GL43.viewport_array.api_errors checks explicitly in both directions.
+    static Bool ValidateViewportIndex(GLuint index, const char* functionName) {
+        if (index < RenderStateParameters::MAX_VIEWPORTS) return true;
+
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                         "Viewport index " + std::to_string(index) +
+                                             " is out of range. Max supported is " +
+                                             std::to_string(RenderStateParameters::MAX_VIEWPORTS - 1) + "."));
+        return false;
+    }
+
+    static Bool ValidateViewportRange(GLuint first, GLsizei count, const char* functionName) {
+        if (count < 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "count must not be negative."));
+            return false;
+        }
+        // Widened before adding: first is a GLuint and count a GLsizei, so `first + count` in
+        // 32 bits can wrap past MAX_VIEWPORTS and let an out-of-range range through.
+        const Uint64 last = static_cast<Uint64>(first) + static_cast<Uint64>(count);
+        if (last > RenderStateParameters::MAX_VIEWPORTS) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                             "first (" + std::to_string(first) + ") + count (" +
+                                                 std::to_string(count) + ") exceeds GL_MAX_VIEWPORTS (" +
+                                                 std::to_string(RenderStateParameters::MAX_VIEWPORTS) + ")."));
+            return false;
+        }
+        return true;
+    }
+
+    template <typename T>
+    static Bool ValidateNonNegativeExtent(T width, T height, const char* functionName) {
+        if (width >= T(0) && height >= T(0)) return true;
+
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "Width and height must be non-negative."));
+        return false;
+    }
+
+    // The array forms are all-or-nothing: one bad element rejects the whole call with a SINGLE
+    // GL_INVALID_VALUE and leaves every rectangle untouched. api_errors relies on both halves -
+    // it passes a full 16-element array with exactly one negative extent and then asserts the
+    // error queue holds exactly one entry.
+    template <typename T>
+    static Bool ValidateArrayExtents(GLsizei count, const T* v, const char* functionName) {
+        for (GLsizei i = 0; i < count; ++i) {
+            if (v[i * 4 + 2] >= T(0) && v[i * 4 + 3] >= T(0)) continue;
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName,
+                                             "Width and height must be non-negative (element " + std::to_string(i) +
+                                                 ")."));
+            return false;
+        }
+        return true;
+    }
+
+    static Bool ValidateNonNullArray(const void* v, const char* functionName) {
+        if (v != nullptr) return true;
+        MG_State::pGLContext->RecordError(
+            ErrorCode::InvalidValue,
+            MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, "value pointer cannot be null."));
+        return false;
+    }
+
     static Bool TryConvertBlendEquation(GLenum mode, const char* functionName,
                                         ::MobileGL::BlendEquation& outEquation) {
         outEquation = MG_Util::ConvertGLEnumToBlendEquation(mode);
@@ -93,16 +183,70 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Viewport_State(GLint x, GLint y, GLsizei width, GLsizei height) {
-        if (width < 0 || height < 0) {
-            MG_State::pGLContext->RecordError(ErrorCode::InvalidValue,
-                                              MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "Viewport_State",
-                                                                           "Width abd height must be non-negative."));
-            return;
-        }
+        if (!ValidateNonNegativeExtent(width, height, "Viewport_State")) return;
 
         MG_State::pGLContext->SetViewport(IntVec4(x, y, width, height));
     }
 
+    // ------------------ ARB_viewport_array setters ------------------
+    void ViewportArrayv_State(GLuint first, GLsizei count, const GLfloat* v) {
+        if (!ValidateViewportRange(first, count, "ViewportArrayv_State")) return;
+        if (count == 0) return;
+        if (!ValidateNonNullArray(v, "ViewportArrayv_State")) return;
+        if (!ValidateArrayExtents(count, v, "ViewportArrayv_State")) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->SetViewportIndexed(first + static_cast<GLuint>(i),
+                                                     FloatVec4(v[i * 4 + 0], v[i * 4 + 1], v[i * 4 + 2], v[i * 4 + 3]));
+        }
+    }
+
+    void ViewportIndexedf_State(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) {
+        if (!ValidateViewportIndex(index, "ViewportIndexedf_State")) return;
+        if (!ValidateNonNegativeExtent(w, h, "ViewportIndexedf_State")) return;
+
+        MG_State::pGLContext->SetViewportIndexed(index, FloatVec4(x, y, w, h));
+    }
+
+    void ScissorArrayv_State(GLuint first, GLsizei count, const GLint* v) {
+        if (!ValidateViewportRange(first, count, "ScissorArrayv_State")) return;
+        if (count == 0) return;
+        if (!ValidateNonNullArray(v, "ScissorArrayv_State")) return;
+        if (!ValidateArrayExtents(count, v, "ScissorArrayv_State")) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->SetScissorBoxIndexed(first + static_cast<GLuint>(i),
+                                                       IntVec4(v[i * 4 + 0], v[i * 4 + 1], v[i * 4 + 2], v[i * 4 + 3]));
+        }
+    }
+
+    void ScissorIndexed_State(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) {
+        if (!ValidateViewportIndex(index, "ScissorIndexed_State")) return;
+        if (!ValidateNonNegativeExtent(width, height, "ScissorIndexed_State")) return;
+
+        MG_State::pGLContext->SetScissorBoxIndexed(index, IntVec4(left, bottom, width, height));
+    }
+
+    void DepthRangeArrayv_State(GLuint first, GLsizei count, const GLdouble* v) {
+        if (!ValidateViewportRange(first, count, "DepthRangeArrayv_State")) return;
+        if (count == 0) return;
+        if (!ValidateNonNullArray(v, "DepthRangeArrayv_State")) return;
+
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->SetDepthRangeIndexed(
+                first + static_cast<GLuint>(i),
+                FloatVec2(ClampUnitFloat(static_cast<GLfloat>(v[i * 2 + 0])),
+                          ClampUnitFloat(static_cast<GLfloat>(v[i * 2 + 1]))));
+        }
+    }
+
+    void DepthRangeIndexed_State(GLuint index, GLdouble n, GLdouble f) {
+        if (!ValidateViewportIndex(index, "DepthRangeIndexed_State")) return;
+
+        MG_State::pGLContext->SetDepthRangeIndexed(
+            index, FloatVec2(ClampUnitFloat(static_cast<GLfloat>(n)), ClampUnitFloat(static_cast<GLfloat>(f))));
+    }
+
     void StencilOpSeparate_State(GLenum face, GLenum sfail, GLenum dpfail, GLenum dppass) {
         Bool applyFront = false;
         Bool applyBack = false;
@@ -175,12 +319,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Scissor_State(GLint x, GLint y, GLsizei width, GLsizei height) {
-        if (width < 0 || height < 0) {
-            MG_State::pGLContext->RecordError(ErrorCode::InvalidValue,
-                                              MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "Scissor_State",
-                                                                           "Width abd height must be non-negative."));
-            return;
-        }
+        if (!ValidateNonNegativeExtent(width, height, "Scissor_State")) return;
 
         MG_State::pGLContext->SetScissorBox(IntVec4(x, y, width, height));
     }
@@ -336,7 +475,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     GLboolean IsEnabledi_State(GLenum target, GLuint index) {
-        if (!ValidateIndexedBlendCapability(target, index, "IsEnabledi_State")) {
+        if (!ValidateIndexedCapability(target, index, "IsEnabledi_State")) {
             return GL_FALSE;
         }
 
@@ -392,7 +531,14 @@ namespace MobileGL::MG_Impl::GLImpl {
         }
         GLint values[4] = {};
         GetIntegeri_v(target, index, values);
-        *data = values[0] != 0 ? GL_TRUE : GL_FALSE;
+        // The ARB_viewport_array rectangles are the only multi-component indexed state that
+        // reaches here; writing element 0 alone would leave the caller's other three untouched.
+        const GLsizei components = target == GL_VIEWPORT || target == GL_SCISSOR_BOX
+            ? 4
+            : (target == GL_DEPTH_RANGE ? 2 : 1);
+        for (GLsizei i = 0; i < components; ++i) {
+            data[i] = values[i] != 0 ? GL_TRUE : GL_FALSE;
+        }
     }
 
     GLboolean IsEnabled_State(GLenum cap) {
@@ -725,7 +871,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Disablei_State(GLenum target, GLuint index) {
-        if (!ValidateIndexedBlendCapability(target, index, "Disablei_State")) {
+        if (!ValidateIndexedCapability(target, index, "Disablei_State")) {
             return;
         }
 
@@ -743,7 +889,7 @@ namespace MobileGL::MG_Impl::GLImpl {
     }
 
     void Enablei_State(GLenum target, GLuint index) {
-        if (!ValidateIndexedBlendCapability(target, index, "Enablei_State")) {
+        if (!ValidateIndexedCapability(target, index, "Enablei_State")) {
             return;
         }
 
@@ -797,6 +943,44 @@ namespace MobileGL::MG_Impl::GLImpl {
         Viewport_State(x, y, width, height);
     }
 
+    void ViewportArrayv(GLuint first, GLsizei count, const GLfloat* v) {
+        ViewportArrayv_State(first, count, v);
+    }
+
+    void ViewportIndexedf(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h) {
+        ViewportIndexedf_State(index, x, y, w, h);
+    }
+
+    void ViewportIndexedfv(GLuint index, const GLfloat* v) {
+        // The index is validated before the pointer is touched: glViewportIndexedfv(MAX, nullptr)
+        // must be one GL_INVALID_VALUE, not a null dereference.
+        if (!ValidateViewportIndex(index, "ViewportIndexedfv")) return;
+        if (!ValidateNonNullArray(v, "ViewportIndexedfv")) return;
+        ViewportIndexedf_State(index, v[0], v[1], v[2], v[3]);
+    }
+
+    void ScissorArrayv(GLuint first, GLsizei count, const GLint* v) {
+        ScissorArrayv_State(first, count, v);
+    }
+
+    void ScissorIndexed(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height) {
+        ScissorIndexed_State(index, left, bottom, width, height);
+    }
+
+    void ScissorIndexedv(GLuint index, const GLint* v) {
+        if (!ValidateViewportIndex(index, "ScissorIndexedv")) return;
+        if (!ValidateNonNullArray(v, "ScissorIndexedv")) return;
+        ScissorIndexed_State(index, v[0], v[1], v[2], v[3]);
+    }
+
+    void DepthRangeArrayv(GLuint first, GLsizei count, const GLdouble* v) {
+        DepthRangeArrayv_State(first, count, v);
+    }
+
+    void DepthRangeIndexed(GLuint index, GLdouble n, GLdouble f) {
+        DepthRangeIndexed_State(index, n, f);
+    }
+
     void StencilOpSeparate(GLenum face, GLenum sfail, GLenum dpfail, GLenum dppass) {
         StencilOpSeparate_State(face, sfail, dpfail, dppass);
     }

MobileGL/MG_Impl/GLImpl/RenderState/GL_RenderState.h

@@ -20,6 +20,16 @@ namespace MobileGL::MG_Impl::GLImpl {
     void Enablei(GLenum target, GLuint index);
     void BlendFunc(GLenum sfactor, GLenum dfactor);
     void Viewport(GLint x, GLint y, GLsizei width, GLsizei height);
+    // ARB_viewport_array (core since GL 4.1). Every one of these addresses the same 16-element
+    // indexed state the classic glViewport/glScissor/glDepthRange trio broadcasts to.
+    void ViewportArrayv(GLuint first, GLsizei count, const GLfloat* v);
+    void ViewportIndexedf(GLuint index, GLfloat x, GLfloat y, GLfloat w, GLfloat h);
+    void ViewportIndexedfv(GLuint index, const GLfloat* v);
+    void ScissorArrayv(GLuint first, GLsizei count, const GLint* v);
+    void ScissorIndexed(GLuint index, GLint left, GLint bottom, GLsizei width, GLsizei height);
+    void ScissorIndexedv(GLuint index, const GLint* v);
+    void DepthRangeArrayv(GLuint first, GLsizei count, const GLdouble* v);
+    void DepthRangeIndexed(GLuint index, GLdouble n, GLdouble f);
     void StencilOpSeparate(GLenum face, GLenum sfail, GLenum dpfail, GLenum dppass);
     void StencilOp(GLenum fail, GLenum zfail, GLenum zpass);
     void StencilMaskSeparate(GLenum face, GLuint mask);

MobileGL/MG_Impl/GLXImpl/GLXImpl.cpp

@@ -175,14 +175,14 @@ namespace MobileGL::MG_Impl::GLXImpl {
 
         struct ContextObject {
             Display* XDisplay = nullptr;
-            EGLDisplay Display = EGL_NO_DISPLAY;
+            EGLDisplay Dpy = EGL_NO_DISPLAY;
             EGLConfig Config = nullptr;
             EGLContext Context = EGL_NO_CONTEXT;
             const FBConfigInfo* FBConfig = nullptr;
         };
 
         struct DrawableSurface {
-            EGLDisplay Display = EGL_NO_DISPLAY;
+            EGLDisplay Dpy = EGL_NO_DISPLAY;
             EGLSurface Surface = EGL_NO_SURFACE;
             Uint32 Width = 0;
             Uint32 Height = 0;
@@ -294,7 +294,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
             if (width == surface.Width && height == surface.Height) {
                 return;
             }
-            if (EGLImpl::ResizePlatformWindowSurface(surface.Display, surface.Surface,
+            if (EGLImpl::ResizePlatformWindowSurface(surface.Dpy, surface.Surface,
                                                      static_cast<EGLint>(width),
                                                      static_cast<EGLint>(height))) {
                 surface.Width = width;
@@ -324,7 +324,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
                 EGL_NONE,
             };
             EGLSurface surface = EGLImpl::CreatePlatformWindowSurface(
-                context.Display, context.Config, reinterpret_cast<void*>(drawable), attribs);
+                context.Dpy, context.Config, reinterpret_cast<void*>(drawable), attribs);
             if (surface == EGL_NO_SURFACE) {
                 MGLOG_E_ONCE("glx: failed to create window surface for drawable 0x%lx (%ux%u)", drawable,
                         width, height);
@@ -332,7 +332,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
             }
 
             DrawableSurface record;
-            record.Display = context.Display;
+            record.Dpy = context.Dpy;
             record.Surface = surface;
             record.Width = width;
             record.Height = height;
@@ -388,7 +388,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
 
             ContextObject object;
             object.XDisplay = dpy;
-            object.Display = display;
+            object.Dpy = display;
             object.Config = config;
             object.Context = eglContext;
             object.FBConfig = fbconfig;
@@ -895,7 +895,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
             return;
         }
         if (object->Context != EGL_NO_CONTEXT) {
-            EGLImpl::DestroyContext(object->Display, object->Context);
+            EGLImpl::DestroyContext(object->Dpy, object->Context);
         }
         Contexts().erase(context);
     }
@@ -929,7 +929,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
             return 0;
         }
 
-        if (!EGLImpl::MakeCurrent(object->Display, surface->Surface, surface->Surface,
+        if (!EGLImpl::MakeCurrent(object->Dpy, surface->Surface, surface->Surface,
                                   object->Context)) {
             MGLOG_E_ONCE("glx: eglMakeCurrent failed (drawable=0x%lx, ctx=%p)", drawable, context);
             return 0;
@@ -962,7 +962,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
             return;
         }
         SyncSurfaceSize(dpy, drawable, it->second);
-        EGLImpl::SwapBuffers(it->second.Display, it->second.Surface);
+        EGLImpl::SwapBuffers(it->second.Dpy, it->second.Surface);
     }
 
     GLXDrawableHandle CreateWindow(Display*, GLXFBConfigHandle config, GLXDrawableHandle window,
@@ -988,7 +988,7 @@ namespace MobileGL::MG_Impl::GLXImpl {
         if (it == surfaces.end()) {
             return;
         }
-        EGLImpl::DestroySurface(it->second.Display, it->second.Surface);
+        EGLImpl::DestroySurface(it->second.Dpy, it->second.Surface);
         surfaces.erase(it);
     }
 

MobileGL/MG_IntegrationTest/CMakeLists.txt

@@ -63,6 +63,7 @@ add_executable(MobileGLIntegrationTest
     Scenarios/DepthStencilReadbackMatrixScenario.cpp
     Scenarios/DepthStencilReadbackAttachmentShapeScenario.cpp
     Scenarios/ClipDistanceScenario.cpp
+    Scenarios/ViewportArrayScenario.cpp
     Scenarios/SsboArrayLengthScenario.cpp
     Scenarios/DoublePrecisionScenario.cpp
     Scenarios/UniformInitializerScenario.cpp
@@ -80,6 +81,7 @@ add_executable(MobileGLIntegrationTest
     Scenarios/VertexArrayEnableDisableScenario.cpp
     Scenarios/CopyImageLevelRangeScenario.cpp
     Scenarios/CopyImageLayeredScenario.cpp
+    Scenarios/LayeredAttachmentBarrierScenario.cpp
 )
 
 target_include_directories(MobileGLIntegrationTest PRIVATE

MobileGL/MG_IntegrationTest/Scenarios/AdvertisedLimitsScenario.cpp

@@ -199,5 +199,61 @@ namespace MGITest {
                    "derived component limits are computed in";
         }
 
+        // ARB_viewport_array's own limits. They are advertised from three different places -
+        // GL_MAX_VIEWPORTS from the frontend's indexed state width, the bounds range and the
+        // subpixel bits from the backend caps table - and each backend fills that table from a
+        // different source, so all three are checked on both lanes.
+        //
+        // GL_VIEWPORT_BOUNDS_RANGE is the one that shipped wrong: GLES has no such query, the
+        // DirectGLES loader's glGetFloatv(GL_VIEWPORT_BOUNDS_RANGE) therefore raised
+        // GL_INVALID_ENUM and left the probe's zero-initialized array in place, and MobileGL
+        // advertised [0, 0] - a range that admits no viewport origin at all, and the check that
+        // kept KHR-GL43.viewport_array.queries red on Espryt after the indexed-state work.
+        TEST_F(AdvertisedLimitsScenario, ViewportArrayLimitsMeetTheirGL43Floors) {
+            GLint maxViewports = -1;
+            glGetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_GE(maxViewports, 16) << "GL 4.3 core table 23.53 sets the MAX_VIEWPORTS minimum at 16";
+            EXPECT_LE(maxViewports, 256) << "one viewport rectangle of indexed state is allocated per advertised "
+                                            "viewport, and the CTS sizes its arrays off this number";
+
+            GLfloat boundsRange[2] = {1.0f, -1.0f};
+            glGetFloatv(GL_VIEWPORT_BOUNDS_RANGE, boundsRange);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_LE(boundsRange[0], -32768.0f)
+                << "GL 4.6 core table 23.60 sets the VIEWPORT_BOUNDS_RANGE minimum at [-32768, 32767]; got ["
+                << boundsRange[0] << ", " << boundsRange[1] << "]";
+            EXPECT_GE(boundsRange[1], 32767.0f)
+                << "GL 4.6 core table 23.60 sets the VIEWPORT_BOUNDS_RANGE minimum at [-32768, 32767]; got ["
+                << boundsRange[0] << ", " << boundsRange[1] << "]";
+
+            // KNOWN INFIDELITY, pinned here rather than hidden. MobileGL reports the driver's own
+            // VIEWPORT_SUBPIXEL_BITS (4 on llvmpipe, i.e. 1/16-pixel viewport precision), but the
+            // float viewport rectangle glViewportIndexedf stores is snapped to integers on its
+            // way to both backends (ComputeGLViewport, DirectGLES SyncRenderState). The STATE
+            // round trip is exact - which is all KHR-GL43.viewport_array.viewport_api checks, and
+            // all this cluster set out to fix - so the gap is in rasterization only: a fractional
+            // viewport origin rasterizes as if it had been rounded. Nothing in the suite or in
+            // Minecraft sets one. Only the spec floor is asserted; tightening this to EQ(0) would
+            // mean advertising no subpixel precision at all, which is a separate decision about a
+            // limit MobileGL currently passes through from the driver.
+            GLint subpixelBits = -1;
+            glGetIntegerv(GL_VIEWPORT_SUBPIXEL_BITS, &subpixelBits);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            EXPECT_GE(subpixelBits, 0) << "GL 4.6 core table 23.60: VIEWPORT_SUBPIXEL_BITS has a minimum of 0, and "
+                                          "a negative value is what a sign-flipped uint32 looks like";
+
+            GLint viewportDims[2] = {-1, -1};
+            glGetIntegerv(GL_MAX_VIEWPORT_DIMS, viewportDims);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            GLint maxRenderbufferSize = -1;
+            glGetIntegerv(GL_MAX_RENDERBUFFER_SIZE, &maxRenderbufferSize);
+            ASSERT_EQ(FirstGLError(), GLenum(GL_NO_ERROR));
+            // GL 4.6 core 13.6.1: MAX_VIEWPORT_DIMS must be at least as large as the largest
+            // renderable surface, or a full-size framebuffer could not be fully viewported.
+            EXPECT_GE(viewportDims[0], maxRenderbufferSize);
+            EXPECT_GE(viewportDims[1], maxRenderbufferSize);
+        }
+
     } // namespace
 } // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/LayeredAttachmentBarrierScenario.cpp

@@ -0,0 +1,378 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/LayeredAttachmentBarrierScenario.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
+//
+// Scenario - A TRANSFER OFF A NON-ZERO ATTACHMENT LAYER READS THE LAYER THE BARRIER MOVED.
+//
+// Every transfer DirectVulkan performs against a framebuffer attachment is three commands: a
+// barrier that puts the image in TRANSFER_SRC/DST, the copy or blit itself, and a barrier that
+// puts it back. The copy names the attachment's layer - glFramebufferTextureLayer(.., layer) ends
+// up in `srcSubresource.baseArrayLayer` - but TransitionImageLayout used to emit `layerCount = 1`
+// from `baseArrayLayer 0`, so for every attachment on a layer above zero the barrier moved layer 0
+// and the copy read layer N. The layer the transfer touched was never transitioned: it sat in
+// COLOR_ATTACHMENT_OPTIMAL (or DEPTH_STENCIL_ATTACHMENT_OPTIMAL) while being read as TRANSFER_SRC.
+//
+// That is undefined behaviour, not a guaranteed wrong pixel: a layout is a compression/tiling
+// promise, so a driver that stores both layouts identically returns the right bytes anyway. The
+// software lanes (lavapipe) are exactly such a driver, which is why this scenario is paired with a
+// validation-layer run - the layer names the mismatch outright
+// (VUID-vkCmdCopyImageToBuffer-srcImageLayout-00189, "srcImageLayout ... doesn't match the actual
+// current layout") where the pixels here cannot. On a tiler that really does re-tile per layout,
+// these are the reads that come back as garbage.
+//
+// The four cases below are the four transfer paths that take an attachment layer from GL:
+//
+//   glReadPixels (colour)                    -> VulkanRenderer::ReadPixels
+//   glBlitFramebuffer (colour)               -> VulkanRenderer::BlitNamedFramebuffer
+//   glReadPixels (GL_DEPTH_COMPONENT)        -> VulkanRenderer::ReadDepthStencilImageToClient
+//   glBlitFramebuffer (GL_DEPTH_BUFFER_BIT)  -> VulkanRenderer::BlitNamedFramebuffer, depth leg
+//
+// Each one renders or clears INTO the non-zero layer first, so the image is genuinely sitting in
+// its attachment layout when the transfer starts - a scenario that only uploaded texels would
+// leave it in a transfer layout already and the mismatched barrier would be a no-op.
+//
+// Every case also asserts the layers it did not name still hold their own fill, so a backend that
+// "fixed" the miss by transferring the whole image passes neither half.
+//
+// DirectGLES is the control: it hands the same calls to the driver, so a failure on both backends
+// means the scenario is wrong and a failure on DirectVulkan alone means Magma is.
+
+#include <cmath>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr int kWidth = 8;
+        constexpr int kHeight = 8;
+        // Four layers with the subject at index 2: layers on both sides of it stay untouched, so
+        // "moved the whole image" and "moved layer 0" are both distinguishable from correct.
+        constexpr int kLayers = 4;
+        constexpr int kSubjectLayer = 2;
+
+        // A value no correct read can produce, so "the backend wrote nothing" fails loudly.
+        constexpr float kDepthPoison = 0.2f;
+
+        std::string Describe(const Rgba8& color) {
+            return "(" + std::to_string(color.r) + ", " + std::to_string(color.g) + ", " + std::to_string(color.b) +
+                   ", " + std::to_string(color.a) + ")";
+        }
+
+        // Per-layer fill, uniform within a layer: the defect is about WHICH layer is addressed, and
+        // a value that also varied inside the layer would make the assertions depend on row order.
+        Rgba8 LayerFill(int layer) {
+            return {static_cast<GLubyte>(17 + layer * 30), static_cast<GLubyte>(200 - layer * 25),
+                    static_cast<GLubyte>(60 + layer * 40), 255};
+        }
+
+        // What the draw paints - matches kFS below, and is deliberately none of the LayerFill
+        // values so "the draw never landed" cannot read as a pass.
+        constexpr Rgba8 kPaintedColor{26, 51, 204, 255};
+
+        constexpr const char* kVS = R"(#version 330 core
+in vec2 aPos;
+void main() { gl_Position = vec4(aPos, 0.0, 1.0); }
+)";
+
+        constexpr const char* kFS = R"(#version 330 core
+out vec4 o_color;
+void main() { o_color = vec4(0.1, 0.2, 0.8, 1.0); }
+)";
+
+        void DrawFullViewportQuad(unsigned int program) {
+            static const float kQuad[] = {-1.0f, -1.0f, 1.0f, -1.0f, -1.0f, 1.0f, 1.0f, 1.0f};
+            GLuint vao = 0, vbo = 0;
+            glGenVertexArrays(1, &vao);
+            glBindVertexArray(vao);
+            glGenBuffers(1, &vbo);
+            glBindBuffer(GL_ARRAY_BUFFER, vbo);
+            glBufferData(GL_ARRAY_BUFFER, sizeof(kQuad), kQuad, GL_STATIC_DRAW);
+            glEnableVertexAttribArray(0);
+            glVertexAttribPointer(0, 2, GL_FLOAT, GL_FALSE, 2 * sizeof(float), nullptr);
+            glUseProgram(program);
+            glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+            glBindVertexArray(0);
+            glDeleteBuffers(1, &vbo);
+            glDeleteVertexArrays(1, &vao);
+        }
+
+        class LayeredAttachmentBarrierScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+                std::string error;
+                m_program = CompileProgram(kVS, kFS, &error);
+                ASSERT_NE(m_program, 0u) << error;
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                for (const GLuint fbo : m_fbos) {
+                    glDeleteFramebuffers(1, &fbo);
+                }
+                m_fbos.clear();
+                for (const GLuint texture : m_textures) {
+                    glDeleteTextures(1, &texture);
+                }
+                m_textures.clear();
+                if (m_program != 0) {
+                    glUseProgram(0);
+                    glDeleteProgram(m_program);
+                    m_program = 0;
+                }
+            }
+
+            // An RGBA8 2D array with a different uniform colour per layer.
+            GLuint MakeColorArray() {
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                m_textures.push_back(texture);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+                glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_RGBA8, kWidth, kHeight, kLayers);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                for (int layer = 0; layer < kLayers; ++layer) {
+                    const std::vector<Rgba8> texels(static_cast<std::size_t>(kWidth) * kHeight, LayerFill(layer));
+                    glTexSubImage3D(GL_TEXTURE_2D_ARRAY, 0, 0, 0, layer, kWidth, kHeight, 1, GL_RGBA,
+                                    GL_UNSIGNED_BYTE, texels.data());
+                }
+                glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+                return texture;
+            }
+
+            // A depth 2D array. No initial upload: depth arrays are filled by clearing through an
+            // attachment, which is also the state the transfer paths have to cope with.
+            GLuint MakeDepthArray() {
+                GLuint texture = 0;
+                glGenTextures(1, &texture);
+                m_textures.push_back(texture);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, texture);
+                glTexStorage3D(GL_TEXTURE_2D_ARRAY, 1, GL_DEPTH_COMPONENT24, kWidth, kHeight, kLayers);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D_ARRAY, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glBindTexture(GL_TEXTURE_2D_ARRAY, 0);
+                return texture;
+            }
+
+            // One FBO naming `layer` of the given arrays. Depth is optional (0 = colour only).
+            GLuint MakeLayerFbo(GLuint colorArray, GLuint depthArray, int layer) {
+                GLuint fbo = 0;
+                glGenFramebuffers(1, &fbo);
+                m_fbos.push_back(fbo);
+                glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+                glFramebufferTextureLayer(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, colorArray, 0, layer);
+                if (depthArray != 0) {
+                    glFramebufferTextureLayer(GL_FRAMEBUFFER, GL_DEPTH_ATTACHMENT, depthArray, 0, layer);
+                }
+                EXPECT_EQ(glCheckFramebufferStatus(GL_FRAMEBUFFER), static_cast<GLenum>(GL_FRAMEBUFFER_COMPLETE))
+                    << "layer " << layer << " is not attachable";
+                return fbo;
+            }
+
+            // glReadPixels of one whole layer, through an FBO that names it.
+            Rgba8 ReadLayer(GLuint colorArray, int layer) {
+                const GLuint fbo = MakeLayerFbo(colorArray, 0, layer);
+                glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+                glReadBuffer(GL_COLOR_ATTACHMENT0);
+                glPixelStorei(GL_PACK_ALIGNMENT, 1);
+                std::vector<Rgba8> pixels(static_cast<std::size_t>(kWidth) * kHeight, Rgba8{});
+                glReadPixels(0, 0, kWidth, kHeight, GL_RGBA, GL_UNSIGNED_BYTE, pixels.data());
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                // The fill is uniform within a layer, so any disagreement between texels is itself
+                // a failure - reported here rather than silently reduced to pixels[0].
+                for (std::size_t i = 1; i < pixels.size(); ++i) {
+                    EXPECT_TRUE(pixels[i] == pixels[0])
+                        << "layer " << layer << " is not uniform: texel 0 is " << Describe(pixels[0]) << ", texel "
+                        << i << " is " << Describe(pixels[i]);
+                }
+                return pixels[0];
+            }
+
+            // Every layer but `changed` still holds its own fill.
+            void ExpectOtherLayersUntouched(GLuint colorArray, int changed, const char* what) {
+                for (int layer = 0; layer < kLayers; ++layer) {
+                    if (layer == changed) continue;
+                    const Rgba8 actual = ReadLayer(colorArray, layer);
+                    EXPECT_TRUE(actual == LayerFill(layer))
+                        << what << ": layer " << layer << " should still hold its fill but is " << Describe(actual)
+                        << ", expected " << Describe(LayerFill(layer));
+                }
+            }
+
+            float ReadDepthAt(int x, int y) const {
+                float depth = kDepthPoison;
+                glReadPixels(x, y, 1, 1, GL_DEPTH_COMPONENT, GL_FLOAT, &depth);
+                return depth;
+            }
+
+            std::vector<GLuint> m_textures;
+            std::vector<GLuint> m_fbos;
+            unsigned int m_program = 0;
+        };
+
+        // glReadPixels straight off a layer that was just rendered to. The image is in
+        // COLOR_ATTACHMENT_OPTIMAL when the readback barrier runs, so the barrier and the copy
+        // disagreeing about the layer is a live layout mismatch, not a bookkeeping detail.
+        TEST_F(LayeredAttachmentBarrierScenario, ReadPixelsOffRenderedNonZeroLayer) {
+            if (!Ready()) return;
+
+            const GLuint colorArray = MakeColorArray();
+            ASSERT_EQ(FirstGLError(), 0u) << "texture setup failed";
+
+            const GLuint fbo = MakeLayerFbo(colorArray, 0, kSubjectLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDisable(GL_DEPTH_TEST);
+            glDrawBuffer(GL_COLOR_ATTACHMENT0);
+            DrawFullViewportQuad(m_program);
+
+            glReadBuffer(GL_COLOR_ATTACHMENT0);
+            glPixelStorei(GL_PACK_ALIGNMENT, 1);
+            std::vector<Rgba8> pixels(static_cast<std::size_t>(kWidth) * kHeight, Rgba8{});
+            glReadPixels(0, 0, kWidth, kHeight, GL_RGBA, GL_UNSIGNED_BYTE, pixels.data());
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            EXPECT_EQ(FirstGLError(), 0u);
+
+            for (std::size_t i = 0; i < pixels.size(); ++i) {
+                ASSERT_NEAR(pixels[i].r, kPaintedColor.r, 2)
+                    << "texel " << i << " of the rendered layer is " << Describe(pixels[i]);
+                ASSERT_NEAR(pixels[i].g, kPaintedColor.g, 2) << "texel " << i;
+                ASSERT_NEAR(pixels[i].b, kPaintedColor.b, 2) << "texel " << i;
+            }
+
+            ExpectOtherLayersUntouched(colorArray, kSubjectLayer, "readback off a rendered layer");
+        }
+
+        // glBlitFramebuffer between two non-zero layers of two different arrays. Both endpoints are
+        // above layer 0, so the source and destination barriers are each wrong on their own side.
+        TEST_F(LayeredAttachmentBarrierScenario, BlitBetweenNonZeroColorLayers) {
+            if (!Ready()) return;
+
+            const GLuint sourceArray = MakeColorArray();
+            const GLuint destinationArray = MakeColorArray();
+            ASSERT_EQ(FirstGLError(), 0u) << "texture setup failed";
+
+            constexpr int kSourceLayer = 3;
+            constexpr int kDestinationLayer = 1;
+
+            const GLuint sourceFbo = MakeLayerFbo(sourceArray, 0, kSourceLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, sourceFbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDisable(GL_DEPTH_TEST);
+            glDrawBuffer(GL_COLOR_ATTACHMENT0);
+            DrawFullViewportQuad(m_program);
+
+            const GLuint destinationFbo = MakeLayerFbo(destinationArray, 0, kDestinationLayer);
+            glBindFramebuffer(GL_READ_FRAMEBUFFER, sourceFbo);
+            glReadBuffer(GL_COLOR_ATTACHMENT0);
+            glBindFramebuffer(GL_DRAW_FRAMEBUFFER, destinationFbo);
+            glDrawBuffer(GL_COLOR_ATTACHMENT0);
+            glBlitFramebuffer(0, 0, kWidth, kHeight, 0, 0, kWidth, kHeight, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            EXPECT_EQ(FirstGLError(), 0u);
+
+            const Rgba8 blitted = ReadLayer(destinationArray, kDestinationLayer);
+            EXPECT_NEAR(blitted.r, kPaintedColor.r, 2) << "blit destination layer is " << Describe(blitted);
+            EXPECT_NEAR(blitted.g, kPaintedColor.g, 2);
+            EXPECT_NEAR(blitted.b, kPaintedColor.b, 2);
+
+            ExpectOtherLayersUntouched(destinationArray, kDestinationLayer, "colour blit destination");
+            // The source layer was rendered, not blitted into, so it is checked separately.
+            const Rgba8 source = ReadLayer(sourceArray, kSourceLayer);
+            EXPECT_NEAR(source.r, kPaintedColor.r, 2) << "blit source layer is " << Describe(source);
+            ExpectOtherLayersUntouched(sourceArray, kSourceLayer, "colour blit source");
+        }
+
+        // The depth aspect of the same readback path: the depth image sits in
+        // DEPTH_STENCIL_ATTACHMENT_OPTIMAL after the clear, and the copy names the attached layer.
+        TEST_F(LayeredAttachmentBarrierScenario, ReadDepthOffClearedNonZeroLayer) {
+            if (!Ready()) return;
+
+            const GLuint colorArray = MakeColorArray();
+            const GLuint depthArray = MakeDepthArray();
+            ASSERT_EQ(FirstGLError(), 0u) << "texture setup failed";
+
+            const GLuint fbo = MakeLayerFbo(colorArray, depthArray, kSubjectLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDepthMask(GL_TRUE);
+            glClearDepth(0.375);
+            glClear(GL_DEPTH_BUFFER_BIT);
+
+            const float centre = ReadDepthAt(kWidth / 2, kHeight / 2);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            EXPECT_EQ(FirstGLError(), 0u);
+            EXPECT_NEAR(centre, 0.375f, 1.0f / 4096.0f)
+                << "glReadPixels(GL_DEPTH_COMPONENT) off layer " << kSubjectLayer << " returned " << centre
+                << (std::fabs(centre - kDepthPoison) < 1e-6f ? " - the destination was never written at all" : "");
+        }
+
+        // The depth leg of the blit path, both endpoints above layer 0. Verified by reading the
+        // destination's depth back, which is the same readback the case above pins - so a failure
+        // here with that one passing is the blit, not the readback.
+        TEST_F(LayeredAttachmentBarrierScenario, BlitDepthBetweenNonZeroLayers) {
+            if (!Ready()) return;
+
+            const GLuint sourceColor = MakeColorArray();
+            const GLuint sourceDepth = MakeDepthArray();
+            const GLuint destinationColor = MakeColorArray();
+            const GLuint destinationDepth = MakeDepthArray();
+            ASSERT_EQ(FirstGLError(), 0u) << "texture setup failed";
+
+            constexpr int kSourceLayer = 3;
+            constexpr int kDestinationLayer = 1;
+
+            const GLuint sourceFbo = MakeLayerFbo(sourceColor, sourceDepth, kSourceLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, sourceFbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDisable(GL_SCISSOR_TEST);
+            glDepthMask(GL_TRUE);
+            glClearDepth(0.625);
+            glClear(GL_DEPTH_BUFFER_BIT);
+
+            // A destination pre-cleared to something the blit must overwrite, so "the blit did
+            // nothing" and "the blit landed" are different answers.
+            const GLuint destinationFbo = MakeLayerFbo(destinationColor, destinationDepth, kDestinationLayer);
+            glBindFramebuffer(GL_FRAMEBUFFER, destinationFbo);
+            glViewport(0, 0, kWidth, kHeight);
+            glDepthMask(GL_TRUE);
+            glClearDepth(0.125);
+            glClear(GL_DEPTH_BUFFER_BIT);
+
+            glBindFramebuffer(GL_READ_FRAMEBUFFER, sourceFbo);
+            glBindFramebuffer(GL_DRAW_FRAMEBUFFER, destinationFbo);
+            glBlitFramebuffer(0, 0, kWidth, kHeight, 0, 0, kWidth, kHeight, GL_DEPTH_BUFFER_BIT, GL_NEAREST);
+            EXPECT_EQ(FirstGLError(), 0u);
+
+            glBindFramebuffer(GL_FRAMEBUFFER, destinationFbo);
+            const float blitted = ReadDepthAt(kWidth / 2, kHeight / 2);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            EXPECT_EQ(FirstGLError(), 0u);
+            EXPECT_NEAR(blitted, 0.625f, 1.0f / 4096.0f)
+                << "depth blitted onto layer " << kDestinationLayer << " reads back as " << blitted
+                << (std::fabs(blitted - 0.125f) < 1e-3f ? " - the destination kept its own clear" : "");
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/ViewportArrayScenario.cpp

@@ -0,0 +1,524 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/ViewportArrayScenario.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
+//
+// Scenario - gl_ViewportIndex ACTUALLY ROUTES, AND THE PER-INDEX STATE IT SELECTS IS REAL.
+//
+// The state half of ARB_viewport_array is asserted in MG_Test/State/RenderStateTest.cpp, which
+// is a pure set/get exercise and would pass just as green against a backend that stores all 16
+// rectangles and rasterizes only the first. This file is the other half: every case here routes
+// primitives to a viewport OTHER than 0 and then looks at where the pixels landed.
+//
+// Three claims, one per case:
+//   1. gl_ViewportIndex selects the viewport RECTANGLE - a 4x4 grid of 32x32 viewports, one
+//      geometry-shader invocation per cell, and every cell must hold its own index.
+//   2. gl_ViewportIndex selects the DEPTH RANGE - 16 one-pixel-wide viewports whose ranges are
+//      (i/16, 1 - i/16), a quad at each end of clip space, and gl_FragCoord.z read back.
+//      This is the claim that fails loudest against a single-viewport backend, because the
+//      geometry is still in the right place while every depth comes back as viewport 0's.
+//   3. The per-index SCISSOR TEST ENABLE is honoured. Vulkan has no per-viewport scissor-test
+//      toggle, so a disabled index has to be given the whole framebuffer as its rectangle; the
+//      case draws the same primitive into the same index twice, once with the test off and once
+//      with it on, and requires the two results to differ in the documented direction.
+//
+// Case 1 runs a second time against the DEFAULT framebuffer. MobileGL Y-flips (and pre-transform
+// rotates) the default framebuffer's rectangles and does not touch an FBO's, so a port that
+// applies the flip to viewport 0 and forgets the other fifteen renders a correct-looking FBO and
+// an upside-down window - the classic multi-viewport bug, and invisible to every FBO-only case.
+//
+// HONEST LIMIT OF THIS FILE. DirectGLES SKIPS every case: GLES has one viewport, one scissor
+// rectangle and no gl_ViewportIndex, so routing to index > 0 is an emulation feature that has
+// not been built (the Espryt half of KHR-GL43.viewport_array's rendering group is deliberately
+// still red). The skip is explicit rather than silent so a future emulation lands here as a
+// failing test and not as a test that was quietly never running. DirectVulkan additionally
+// skips when the device lacks the multiViewport feature - Vulkan then forbids a pipeline from
+// declaring more than one viewport at all, which is a device limit and not a MobileGL bug;
+// lavapipe (every CI lane) and both Mali/Adreno devices support it, so the cases do run where
+// it matters.
+
+#include <cmath>
+#include <string>
+#include <vector>
+
+#include "../Harness/HeadlessGL.h"
+#include "../Harness/ScenarioFixture.h"
+
+#ifdef GLAPI
+#undef GLAPI
+#endif
+#define GL_GLEXT_PROTOTYPES
+#include <GL/gl.h>
+#include <GL/glcorearb.h>
+#undef GL_GLEXT_PROTOTYPES
+
+namespace MGITest {
+    namespace {
+
+        constexpr int kViewportCount = 16;
+        constexpr int kGridSide = 4;   // 4x4 grid of viewports
+        constexpr int kCellSize = 32;  // ... each 32x32
+        constexpr int kSurfaceSide = kGridSide * kCellSize;
+        constexpr GLint kUnwritten = -1;
+
+        // A geometry shader is the only stage GL 4.1 lets write gl_ViewportIndex, and
+        // `invocations` runs it once per viewport off a single input point - the same shape
+        // KHR-GL43.viewport_array.draw_to_single_layer_with_multiple_viewports uses.
+        const char* const kVertexSource = R"(#version 410 core
+void main() { gl_Position = vec4(0.0, 0.0, 0.0, 1.0); }
+)";
+
+        const char* const kGridGeometrySource = R"(#version 410 core
+layout(points, invocations = 16) in;
+layout(triangle_strip, max_vertices = 4) out;
+flat out int gsIndex;
+void main() {
+    gsIndex = gl_InvocationID;
+    gl_ViewportIndex = gl_InvocationID;
+    gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+        // One invocation, viewport chosen by a uniform: lets a case draw the SAME primitive into
+        // the SAME index twice under two different scissor-enable states.
+        const char* const kSingleGeometrySource = R"(#version 410 core
+layout(points, invocations = 1) in;
+layout(triangle_strip, max_vertices = 4) out;
+uniform int uViewport;
+flat out int gsIndex;
+void main() {
+    gsIndex = uViewport;
+    gl_ViewportIndex = uViewport;
+    gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+        const char* const kIntFragmentSource = R"(#version 410 core
+flat in int gsIndex;
+layout(location = 0) out int fragColor;
+void main() { fragColor = gsIndex; }
+)";
+
+        // Two quads, one at each end of clip space, so the fragment stage can report the depth
+        // the viewport's range mapped them to. gl_FragCoord.z IS the post-range window depth, so
+        // it reads back the per-viewport minDepth/maxDepth directly.
+        const char* const kDepthGeometrySource = R"(#version 410 core
+layout(points, invocations = 16) in;
+layout(triangle_strip, max_vertices = 8) out;
+void main() {
+    gl_ViewportIndex = gl_InvocationID;
+    gl_Position = vec4(-1.0, -1.0, -1.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, -1.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  0.0, -1.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  0.0, -1.0, 1.0); EmitVertex();
+    EndPrimitive();
+    gl_Position = vec4(-1.0,  0.0, 1.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  0.0, 1.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 1.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 1.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+        const char* const kDepthFragmentSource = R"(#version 410 core
+layout(location = 0) out float fragColor;
+void main() { fragColor = gl_FragCoord.z; }
+)";
+
+        class ViewportArrayScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                if (Gl().BackendName() == "DirectGLES") {
+                    GTEST_SKIP() << "gl_ViewportIndex routing is not emulated on DirectGLES: GLES has one viewport "
+                                    "and one scissor rectangle, so every index rasterizes as index 0. The indexed "
+                                    "STATE is still asserted (MG_Test RenderStateTest); this is the deferred "
+                                    "rendering half of KHR-GL43.viewport_array.";
+                }
+
+                GLint maxViewports = 0;
+                glGetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
+                ASSERT_GE(maxViewports, kViewportCount) << "GL 4.3 core requires GL_MAX_VIEWPORTS >= 16";
+
+                m_program = BuildProgram(kGridGeometrySource, kIntFragmentSource);
+                ASSERT_NE(m_program, 0u) << "grid program failed to build: " << m_buildLog;
+                glGenVertexArrays(1, &m_vao);
+                glBindVertexArray(m_vao);
+                ResetViewportArrayState();
+                ASSERT_EQ(glGetError(), GL_NO_ERROR) << "setup left a GL error behind";
+            }
+
+            void TearDown() override {
+                if (!Ready() || IsSkipped()) return;
+                ResetViewportArrayState();
+                if (m_vao != 0) glDeleteVertexArrays(1, &m_vao);
+                if (m_program != 0) glDeleteProgram(m_program);
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                while (glGetError() != GL_NO_ERROR) {
+                }
+            }
+
+            // Every case starts from the same slate: this fixture shares its context with every
+            // other scenario in the process, and a leftover per-index scissor enable is exactly
+            // the kind of state that would make a later case pass or fail for the wrong reason.
+            static void ResetViewportArrayState() {
+                for (int i = 0; i < kViewportCount; ++i) {
+                    glDisablei(GL_SCISSOR_TEST, static_cast<GLuint>(i));
+                }
+                glDisable(GL_SCISSOR_TEST);
+                glViewport(0, 0, kSurfaceSide, kSurfaceSide);
+                glScissor(0, 0, kSurfaceSide, kSurfaceSide);
+                glDepthRange(0.0, 1.0);
+                glDisable(GL_DEPTH_TEST);
+            }
+
+            // The 4x4 grid: viewport y*4+x covers the cell whose lower-left corner is
+            // (x*cellW, y*cellH), in GL's bottom-left-origin window coordinates. Parameterized on
+            // the cell size because the default framebuffer this scenario also renders into is
+            // deliberately non-square (HeadlessGL is 128x96, so a transposing bug cannot hide).
+            static void SetupGridViewports(int cellW, int cellH) {
+                std::vector<GLfloat> data(static_cast<size_t>(kViewportCount) * 4);
+                for (int y = 0; y < kGridSide; ++y) {
+                    for (int x = 0; x < kGridSide; ++x) {
+                        const size_t base = static_cast<size_t>(y * kGridSide + x) * 4;
+                        data[base + 0] = static_cast<GLfloat>(x * cellW);
+                        data[base + 1] = static_cast<GLfloat>(y * cellH);
+                        data[base + 2] = static_cast<GLfloat>(cellW);
+                        data[base + 3] = static_cast<GLfloat>(cellH);
+                    }
+                }
+                glViewportArrayv(0, kViewportCount, data.data());
+            }
+
+            GLuint BuildProgram(const char* geometrySource, const char* fragmentSource) {
+                const GLuint vs = CompileStage(GL_VERTEX_SHADER, kVertexSource);
+                if (vs == 0) return 0;
+                const GLuint gs = CompileStage(GL_GEOMETRY_SHADER, geometrySource);
+                if (gs == 0) {
+                    glDeleteShader(vs);
+                    return 0;
+                }
+                const GLuint fs = CompileStage(GL_FRAGMENT_SHADER, fragmentSource);
+                if (fs == 0) {
+                    glDeleteShader(vs);
+                    glDeleteShader(gs);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, vs);
+                glAttachShader(program, gs);
+                glAttachShader(program, fs);
+                glLinkProgram(program);
+                GLint linked = 0;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                glDeleteShader(vs);
+                glDeleteShader(gs);
+                glDeleteShader(fs);
+                if (!linked) {
+                    GLint length = 0;
+                    glGetProgramiv(program, GL_INFO_LOG_LENGTH, &length);
+                    std::vector<char> log(static_cast<size_t>(length > 1 ? length : 1), '\0');
+                    glGetProgramInfoLog(program, static_cast<GLsizei>(log.size()), nullptr, log.data());
+                    m_buildLog = log.data();
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+
+            GLuint CompileStage(GLenum stage, const char* source) {
+                const GLuint shader = glCreateShader(stage);
+                glShaderSource(shader, 1, &source, nullptr);
+                glCompileShader(shader);
+                GLint compiled = 0;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled) return shader;
+                GLint length = 0;
+                glGetShaderiv(shader, GL_INFO_LOG_LENGTH, &length);
+                std::vector<char> log(static_cast<size_t>(length > 1 ? length : 1), '\0');
+                glGetShaderInfoLog(shader, static_cast<GLsizei>(log.size()), nullptr, log.data());
+                m_buildLog = log.data();
+                glDeleteShader(shader);
+                return 0;
+            }
+
+            // An R32I colour target, pre-filled with kUnwritten so "nothing was drawn here" is
+            // distinguishable from "index 0 was drawn here".
+            struct IntTarget {
+                GLuint fbo = 0;
+                GLuint texture = 0;
+            };
+
+            // The "nothing drawn here" value is UPLOADED, not cleared: the CTS fills its R32I
+            // targets the same way (fillTexture), and an upload cannot be confused with a clear
+            // that a backend defers, reorders or drops - which is exactly the ambiguity a case
+            // asserting "this cell must be untouched" cannot afford.
+            static void FillIntTarget(const IntTarget& target, int width, int height) {
+                const std::vector<GLint> unwritten(static_cast<size_t>(width) * height, kUnwritten);
+                glBindTexture(GL_TEXTURE_2D, target.texture);
+                glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, width, height, GL_RED_INTEGER, GL_INT, unwritten.data());
+            }
+
+            static IntTarget MakeIntTarget(int width, int height) {
+                IntTarget target;
+                glGenTextures(1, &target.texture);
+                glBindTexture(GL_TEXTURE_2D, target.texture);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+                glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+                glTexImage2D(GL_TEXTURE_2D, 0, GL_R32I, width, height, 0, GL_RED_INTEGER, GL_INT, nullptr);
+                glGenFramebuffers(1, &target.fbo);
+                glBindFramebuffer(GL_FRAMEBUFFER, target.fbo);
+                glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, target.texture, 0);
+                FillIntTarget(target, width, height);
+                return target;
+            }
+
+            static void DestroyIntTarget(IntTarget& target) {
+                glBindFramebuffer(GL_FRAMEBUFFER, 0);
+                if (target.fbo != 0) glDeleteFramebuffers(1, &target.fbo);
+                if (target.texture != 0) glDeleteTextures(1, &target.texture);
+            }
+
+            static std::vector<GLint> ReadInts(int width, int height) {
+                std::vector<GLint> pixels(static_cast<size_t>(width) * height, 0);
+                glReadPixels(0, 0, width, height, GL_RED_INTEGER, GL_INT, pixels.data());
+                return pixels;
+            }
+
+            // The centre of grid cell (x, y), in the bottom-left-origin coordinates glReadPixels
+            // returns. Sampling the centre rather than a corner keeps the assertion about WHICH
+            // viewport was selected rather than about edge rounding.
+            static GLint CellCentre(const std::vector<GLint>& pixels, int stride, int x, int y) {
+                const int px = x * kCellSize + kCellSize / 2;
+                const int py = y * kCellSize + kCellSize / 2;
+                return pixels[static_cast<size_t>(py) * stride + px];
+            }
+
+            std::string m_buildLog;
+            GLuint m_program = 0;
+            GLuint m_vao = 0;
+        };
+
+        // --- 1. the viewport rectangle -------------------------------------------------------
+
+        TEST_F(ViewportArrayScenario, EachViewportIndexRasterizesIntoItsOwnRectangle) {
+            IntTarget target = MakeIntTarget(kSurfaceSide, kSurfaceSide);
+            SetupGridViewports(kCellSize, kCellSize);
+            glUseProgram(m_program);
+            glBindVertexArray(m_vao);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+
+            const std::vector<GLint> pixels = ReadInts(kSurfaceSide, kSurfaceSide);
+            for (int y = 0; y < kGridSide; ++y) {
+                for (int x = 0; x < kGridSide; ++x) {
+                    const GLint expected = y * kGridSide + x;
+                    EXPECT_EQ(CellCentre(pixels, kSurfaceSide, x, y), expected)
+                        << "cell (" << x << ", " << y << ") should hold viewport index " << expected
+                        << "; a single-viewport backend paints the whole image with 15 (the last invocation)";
+                }
+            }
+            DestroyIntTarget(target);
+        }
+
+        // The same claim against the DEFAULT framebuffer, where MobileGL applies its Y-flip and
+        // pre-transform rotation. Index 0 alone getting the mapping is the classic bug.
+        TEST_F(ViewportArrayScenario, TheDefaultFramebufferAppliesTheSameFlipToEveryViewport) {
+            const int surfaceW = Gl().Width();
+            const int surfaceH = Gl().Height();
+            ASSERT_GE(surfaceW, kGridSide);
+            ASSERT_GE(surfaceH, kGridSide);
+            const int cellW = surfaceW / kGridSide;
+            const int cellH = surfaceH / kGridSide;
+
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            // Paint a value no viewport index can produce, so an unwritten cell is obvious.
+            glClearColor(0.0f, 0.0f, 0.0f, 1.0f);
+            glClear(GL_COLOR_BUFFER_BIT);
+
+            // The default framebuffer is 8-bit RGBA, so the index travels as a colour: cell i is
+            // painted with red = i * 16, which is exact in 8 bits for i in [0, 16).
+            const char* const kColorFragmentSource = R"(#version 410 core
+flat in int gsIndex;
+layout(location = 0) out vec4 fragColor;
+void main() { fragColor = vec4(float(gsIndex) * 16.0 / 255.0, 0.0, 0.0, 1.0); }
+)";
+            const GLuint colorProgram = BuildProgram(kGridGeometrySource, kColorFragmentSource);
+            ASSERT_NE(colorProgram, 0u) << "colour program failed to build: " << m_buildLog;
+
+            SetupGridViewports(cellW, cellH);
+            glUseProgram(colorProgram);
+            glBindVertexArray(m_vao);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+
+            std::vector<unsigned char> pixels(static_cast<size_t>(surfaceW) * surfaceH * 4, 0);
+            glReadPixels(0, 0, surfaceW, surfaceH, GL_RGBA, GL_UNSIGNED_BYTE, pixels.data());
+            for (int y = 0; y < kGridSide; ++y) {
+                for (int x = 0; x < kGridSide; ++x) {
+                    const int px = x * cellW + cellW / 2;
+                    const int py = y * cellH + cellH / 2;
+                    const int red = pixels[(static_cast<size_t>(py) * surfaceW + px) * 4];
+                    const int expected = (y * kGridSide + x) * 16;
+                    // One LSB of slack for an 8-bit round trip; the values are 16 apart, so this
+                    // cannot confuse two neighbouring indices.
+                    EXPECT_LE(std::abs(red - expected), 1)
+                        << "default-framebuffer cell (" << x << ", " << y << ") holds red=" << red << ", expected "
+                        << expected << ". A vertically mirrored grid means the Y-flip was applied to viewport 0 "
+                        << "only";
+                }
+            }
+            glDeleteProgram(colorProgram);
+        }
+
+        // --- 2. the depth range --------------------------------------------------------------
+
+        TEST_F(ViewportArrayScenario, EachViewportIndexUsesItsOwnDepthRange) {
+            // 16 columns one pixel wide and two rows tall: row 0 gets the near-plane quad, row 1
+            // the far-plane one, so both ends of viewport i's range land in the same column.
+            constexpr int kWidth = kViewportCount;
+            constexpr int kHeight = 2;
+
+            GLuint texture = 0;
+            GLuint fbo = 0;
+            glGenTextures(1, &texture);
+            glBindTexture(GL_TEXTURE_2D, texture);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+            glTexImage2D(GL_TEXTURE_2D, 0, GL_R32F, kWidth, kHeight, 0, GL_RED, GL_FLOAT, nullptr);
+            glGenFramebuffers(1, &fbo);
+            glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+            glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, texture, 0);
+            const GLfloat clearValue[4] = {-1.0f, 0.0f, 0.0f, 0.0f};
+            glClearBufferfv(GL_COLOR, 0, clearValue);
+
+            std::vector<GLfloat> viewports(static_cast<size_t>(kViewportCount) * 4);
+            std::vector<GLdouble> ranges(static_cast<size_t>(kViewportCount) * 2);
+            for (int i = 0; i < kViewportCount; ++i) {
+                viewports[static_cast<size_t>(i) * 4 + 0] = static_cast<GLfloat>(i);
+                viewports[static_cast<size_t>(i) * 4 + 1] = 0.0f;
+                viewports[static_cast<size_t>(i) * 4 + 2] = 1.0f;
+                viewports[static_cast<size_t>(i) * 4 + 3] = 2.0f;
+                ranges[static_cast<size_t>(i) * 2 + 0] = static_cast<GLdouble>(i) / 16.0;
+                ranges[static_cast<size_t>(i) * 2 + 1] = 1.0 - static_cast<GLdouble>(i) / 16.0;
+            }
+            glViewportArrayv(0, kViewportCount, viewports.data());
+            glDepthRangeArrayv(0, kViewportCount, ranges.data());
+
+            const GLuint depthProgram = BuildProgram(kDepthGeometrySource, kDepthFragmentSource);
+            ASSERT_NE(depthProgram, 0u) << "depth program failed to build: " << m_buildLog;
+            glUseProgram(depthProgram);
+            glBindVertexArray(m_vao);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+
+            std::vector<GLfloat> pixels(static_cast<size_t>(kWidth) * kHeight, 0.0f);
+            glReadPixels(0, 0, kWidth, kHeight, GL_RED, GL_FLOAT, pixels.data());
+            for (int i = 0; i < kViewportCount; ++i) {
+                const float near = static_cast<float>(i) / 16.0f;
+                const float far = 1.0f - static_cast<float>(i) / 16.0f;
+                // The tolerance covers depth-buffer-free rasterization of gl_FragCoord.z on a
+                // software rasterizer; the per-index values are 1/16 apart, so it cannot let a
+                // neighbouring viewport's range through, and viewport 0's range (0, 1) differs
+                // from every other index by at least 1/16.
+                EXPECT_NEAR(pixels[i], near, 1.0e-3f)
+                    << "viewport " << i << " near-plane depth; got viewport 0's range if this is 0";
+                EXPECT_NEAR(pixels[static_cast<size_t>(kWidth) + i], far, 1.0e-3f)
+                    << "viewport " << i << " far-plane depth; got viewport 0's range if this is 1";
+            }
+
+            glDeleteProgram(depthProgram);
+            glBindFramebuffer(GL_FRAMEBUFFER, 0);
+            glDeleteFramebuffers(1, &fbo);
+            glDeleteTextures(1, &texture);
+        }
+
+        // --- 3. the per-index scissor-test enable --------------------------------------------
+
+        TEST_F(ViewportArrayScenario, AnIndexedScissorEnableClipsOnlyThatIndex) {
+            IntTarget target = MakeIntTarget(kSurfaceSide, kSurfaceSide);
+
+            // One full-size viewport per index so the scissor rectangle is the ONLY thing that
+            // can shrink the quad - the same separation KHR-GL43.viewport_array.scissor uses.
+            glViewport(0, 0, kSurfaceSide, kSurfaceSide);
+            std::vector<GLint> boxes(static_cast<size_t>(kViewportCount) * 4);
+            for (int y = 0; y < kGridSide; ++y) {
+                for (int x = 0; x < kGridSide; ++x) {
+                    const size_t base = static_cast<size_t>(y * kGridSide + x) * 4;
+                    boxes[base + 0] = x * kCellSize;
+                    boxes[base + 1] = y * kCellSize;
+                    boxes[base + 2] = kCellSize;
+                    boxes[base + 3] = kCellSize;
+                }
+            }
+            glScissorArrayv(0, kViewportCount, boxes.data());
+
+            const GLuint singleProgram = BuildProgram(kSingleGeometrySource, kIntFragmentSource);
+            ASSERT_NE(singleProgram, 0u) << "single-viewport program failed to build: " << m_buildLog;
+            glUseProgram(singleProgram);
+            glBindVertexArray(m_vao);
+            const GLint uViewport = glGetUniformLocation(singleProgram, "uViewport");
+            ASSERT_NE(uViewport, -1);
+
+            constexpr GLint kProbeIndex = 6;  // grid cell (2, 1)
+            constexpr int kProbeX = kProbeIndex % kGridSide;
+            constexpr int kProbeY = kProbeIndex / kGridSide;
+
+            // (a) scissor test ENABLED for this index: the quad is clipped to its 32x32 box.
+            glUniform1i(uViewport, kProbeIndex);
+            glEnablei(GL_SCISSOR_TEST, kProbeIndex);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+            {
+                const std::vector<GLint> pixels = ReadInts(kSurfaceSide, kSurfaceSide);
+                EXPECT_EQ(CellCentre(pixels, kSurfaceSide, kProbeX, kProbeY), kProbeIndex)
+                    << "the scissored index must still paint inside its own box";
+                for (int y = 0; y < kGridSide; ++y) {
+                    for (int x = 0; x < kGridSide; ++x) {
+                        if (x == kProbeX && y == kProbeY) continue;
+                        EXPECT_EQ(CellCentre(pixels, kSurfaceSide, x, y), kUnwritten)
+                            << "cell (" << x << ", " << y << ") is outside scissor rectangle " << kProbeIndex
+                            << " and must be untouched";
+                    }
+                }
+            }
+
+            // (b) scissor test DISABLED for the same index, everything else identical: with no
+            // per-viewport toggle in Vulkan this is the case that needs the disabled index to be
+            // given the full framebuffer rectangle, and it is exactly where "leave the last
+            // rectangle bound" would show up as a still-clipped quad.
+            FillIntTarget(target, kSurfaceSide, kSurfaceSide);
+            glBindFramebuffer(GL_FRAMEBUFFER, target.fbo);
+            glDisablei(GL_SCISSOR_TEST, kProbeIndex);
+            glDrawArrays(GL_POINTS, 0, 1);
+            ASSERT_EQ(glGetError(), GL_NO_ERROR);
+            {
+                const std::vector<GLint> pixels = ReadInts(kSurfaceSide, kSurfaceSide);
+                for (int y = 0; y < kGridSide; ++y) {
+                    for (int x = 0; x < kGridSide; ++x) {
+                        EXPECT_EQ(CellCentre(pixels, kSurfaceSide, x, y), kProbeIndex)
+                            << "with the scissor test off for index " << kProbeIndex
+                            << ", its full-viewport quad must cover cell (" << x << ", " << y << ")";
+                    }
+                }
+            }
+
+            glDeleteProgram(singleProgram);
+            DestroyIntTarget(target);
+        }
+
+    } // namespace
+} // namespace MGITest

MobileGL/MG_State/GLState/BufferState/BufferObject.cpp

@@ -250,6 +250,34 @@ namespace MobileGL::MG_State::GLState {
         NotifyContentWrite(atOffset, data.size);
     }
 
+    void BufferObject::FillSubData(DataPtr pattern, SizeT atOffset, SizeT size) {
+        MOBILEGL_ASSERT(pattern.data != nullptr && pattern.size > 0,
+                        "FillSubData requires a non-empty pattern.");
+        MOBILEGL_ASSERT(size % pattern.size == 0,
+                        "FillSubData size (%zu) must be a multiple of pattern size (%zu).", size, pattern.size);
+        MOBILEGL_ASSERT(atOffset <= m_size && size <= m_size - atOffset,
+                        "FillSubData out of bounds: atOffset (%zu) + size (%zu) > m_size (%zu)", atOffset, size,
+                        m_size);
+        MOBILEGL_ASSERT(!m_isMapped || (m_mappingAccess & BufferMappingAccessBit::Persistent),
+                        "Cannot fill data while buffer is non-persistently mapped.");
+        if (size == 0) return;
+
+        // A clear is ordered after all earlier GPU writes. Partial clears additionally need the
+        // retained shadow bytes; whole-store clears need the same synchronization before writing
+        // an adopted persistent mapping that the GPU may still be accessing.
+        SyncGpuWrites();
+
+        Uint8* dst = m_resource.Bytes() + atOffset;
+        if (pattern.size == 1) {
+            Memset(dst, *static_cast<const Uint8*>(pattern.data), size);
+        } else {
+            for (SizeT at = 0; at < size; at += pattern.size) {
+                Memcpy(dst + at, pattern.data, pattern.size);
+            }
+        }
+        NotifyContentWrite(atOffset, size);
+    }
+
     void BufferObject::DownloadSubData(void* dst, SizeT atOffset, SizeT size) const {
         MOBILEGL_ASSERT(atOffset + size <= m_size,
                         "DownloadSubData out of bounds: atOffset (%zu) + size (%zu) > m_size (%zu)", atOffset, size,

MobileGL/MG_State/GLState/BufferState/BufferObject.h

@@ -132,6 +132,9 @@ namespace MobileGL {
 
             void UploadData(DataPtr data, SizeT atOffset);
             void UploadSubData(DataPtr data, SizeT atOffset);
+            // Repeats one already-converted element through [atOffset, atOffset + size) and
+            // publishes the range as one content mutation.
+            void FillSubData(DataPtr pattern, SizeT atOffset, SizeT size);
             // Reads `size` bytes from the CPU shadow at `atOffset` into `dst` (glGetBufferSubData).
             // The shadow reflects CPU writes (BufferData/SubData/maps) and backend write-backs, but not
             // arbitrary GPU-side writes.

MobileGL/MG_State/GLState/Core.cpp

@@ -39,6 +39,11 @@ namespace MobileGL::MG_State {
             return m_compileEnv;
         }
 
+        void GLContext::InvalidateCompileEnv() {
+            m_compileEnv.reset();
+            m_compileEnvBackend = nullptr;
+        }
+
         // Error
         void GLContext::RecordError(ErrorCode code, UniquePtr<ErrorInfo> info) {
             // Invariant I1, mechanically enforced: the GL error state is GL-thread-owned.
@@ -712,10 +717,18 @@ namespace MobileGL::MG_State {
             m_renderState.SetViewport(viewport);
         }
 
-        const IntVec4& GLContext::GetViewport() const {
+        IntVec4 GLContext::GetViewport() const {
             return m_renderState.GetViewport();
         }
 
+        void GLContext::SetViewportIndexed(Uint index, FloatVec4 viewport) {
+            m_renderState.SetViewportIndexed(index, viewport);
+        }
+
+        const FloatVec4& GLContext::GetViewportIndexed(Uint index) const {
+            return m_renderState.GetViewportIndexed(index);
+        }
+
         void GLContext::SetLineWidth(Float width) {
             m_renderState.SetLineWidth(width);
         }
@@ -953,6 +966,14 @@ namespace MobileGL::MG_State {
             return m_renderState.GetDepthRange();
         }
 
+        void GLContext::SetDepthRangeIndexed(Uint index, FloatVec2 range) {
+            m_renderState.SetDepthRangeIndexed(index, range);
+        }
+
+        const FloatVec2& GLContext::GetDepthRangeIndexed(Uint index) const {
+            return m_renderState.GetDepthRangeIndexed(index);
+        }
+
         void GLContext::SetSampleCoverage(Float value, Bool invert) {
             m_renderState.SetSampleCoverage(value, invert);
         }
@@ -1017,6 +1038,14 @@ namespace MobileGL::MG_State {
             return m_renderState.GetScissorBox();
         }
 
+        void GLContext::SetScissorBoxIndexed(Uint index, IntVec4 box) {
+            m_renderState.SetScissorBoxIndexed(index, box);
+        }
+
+        const IntVec4& GLContext::GetScissorBoxIndexed(Uint index) const {
+            return m_renderState.GetScissorBoxIndexed(index);
+        }
+
         // Framebuffer
         void GLContext::GenFramebufferNames(Uint number, Vector<Uint>& framebuffers) {
             m_framebufferState.GenerateNames(number, framebuffers);

MobileGL/MG_State/GLState/Core.h

@@ -198,8 +198,10 @@ namespace MobileGL {
                 // Only the pipeline-relevant subset - see RenderState::m_pipelineStateVersion.
                 Uint GetPipelineStateVersion() const;
                 const RenderStateParameters& GetRenderStateParameters() const;
-                void SetViewport(IntVec4 viewport); // x, y, width, height
-                const IntVec4& GetViewport() const; // x, y, width, height
+                void SetViewport(IntVec4 viewport); // x, y, width, height; writes ALL viewports
+                IntVec4 GetViewport() const;        // x, y, width, height; viewport 0, rounded
+                void SetViewportIndexed(Uint index, FloatVec4 viewport);
+                const FloatVec4& GetViewportIndexed(Uint index) const;
                 void SetLineWidth(Float width);
                 Float GetLineWidth() const;
                 void SetPointSize(Float size);
@@ -260,8 +262,10 @@ namespace MobileGL {
                 Uint32 GetClearStencil() const;
                 void SetBlendColor(FloatVec4 color);
                 const FloatVec4& GetBlendColor() const;
-                void SetDepthRange(FloatVec2 range);
+                void SetDepthRange(FloatVec2 range); // writes ALL viewports' depth ranges
                 const FloatVec2& GetDepthRange() const;
+                void SetDepthRangeIndexed(Uint index, FloatVec2 range);
+                const FloatVec2& GetDepthRangeIndexed(Uint index) const;
                 void SetSampleCoverage(Float value, Bool invert);
                 Float GetSampleCoverageValue() const;
                 Bool GetSampleCoverageInvert() const;
@@ -276,8 +280,10 @@ namespace MobileGL {
                 FrontFaceMode GetFrontFaceMode() const;
                 void SetProvokingVertexMode(ProvokingVertexMode mode);
                 ProvokingVertexMode GetProvokingVertexMode() const;
-                void SetScissorBox(IntVec4 box);      // x, y, width, height
-                const IntVec4& GetScissorBox() const; // x, y, width, height
+                void SetScissorBox(IntVec4 box);      // x, y, width, height; writes ALL rectangles
+                const IntVec4& GetScissorBox() const; // x, y, width, height; rectangle 0
+                void SetScissorBoxIndexed(Uint index, IntVec4 box);
+                const IntVec4& GetScissorBoxIndexed(Uint index) const;
 
                 // Transform feedback. The fields below are the state of the transform
                 // feedback object currently bound to GL_TRANSFORM_FEEDBACK; see the object
@@ -407,9 +413,12 @@ namespace MobileGL {
                 // cannot be captured in MG_State::Init() - that runs BEFORE MG_Backend::Init(),
                 // so there is no backend to query yet. Re-captured whenever the active backend
                 // object changes, which also rolls the fingerprint and therefore invalidates
-                // every P0b preprocess memo keyed against the old one.
+                // every P0b preprocess memo keyed against the old one. A backend whose dynamic
+                // capabilities become available without changing object identity must call
+                // InvalidateCompileEnv() after publishing them.
                 // GL thread only.
                 const SharedPtr<const MG_Util::ShaderTranspiler::CompileEnv>& GetCompileEnv();
+                void InvalidateCompileEnv();
 
             private:
                 // State Components

MobileGL/MG_State/GLState/ProgramState/ProgramLinkTask.h

@@ -60,6 +60,8 @@ namespace MobileGL::MG_State::GLState {
             Uint externalIndex = 0; // logs only
             Vector<LinkShaderInput> shaders; // already stage-sorted
             SharedPtr<const MG_Util::ShaderTranspiler::CompileEnv> env;
+            // Startup configuration copied with the task, never read from worker code.
+            Bool enableSpirvValidation = false;
             // The four "takes effect at the next link" request maps. Snapshotted rather than
             // referenced, which is precisely what makes glBindAttribLocation and friends
             // legal to call over a pending link without cancelling it: the pending link keeps

MobileGL/MG_State/GLState/ProgramState/ProgramObject.cpp

@@ -494,6 +494,7 @@ namespace MobileGL::MG_State::GLState {
         auto task = MakeShared<ProgramLinkTask>();
         task->in.externalIndex = m_externalIndex;
         task->in.env = MG_Util::ShaderTranspiler::GetCurrentCompileEnv();
+        task->in.enableSpirvValidation = MG_Config::Features.EnableSpirvValidation;
         task->in.explicitAttribLocations = m_explicitAttribLocations;
         task->in.explicitFragDataLocation = m_explicitFragDataLocation;
         task->in.explicitFragDataIndex = m_explicitFragDataIndex;

MobileGL/MG_State/GLState/ProgramState/ProgramObject.h

@@ -819,6 +819,9 @@ namespace MobileGL::MG_State::GLState {
         // backend asks this exactly where it used to ask GetLinkStatus(), i.e. right before
         // it builds or draws with the program.
         Bool GetSpirvStatus() const { return Spirv().spirvStatus; }
+        // Copied from the link task that generated this program's SPIR-V. Backends use it for
+        // their final transforms, which must honor the same diagnostic setting as phase B.
+        Bool GetSpirvValidationEnabled() const { return Spirv().enableSpirvValidation; }
 
         // The linked glslang reflection itself, for the ONE consumer that needs resource
         // lists no typed getter above exposes: the GL program-interface query layer
@@ -985,6 +988,7 @@ namespace MobileGL::MG_State::GLState {
         // cannot be lifted out of glslang's reflection instead.
         struct SpirvArtifacts {
             Vector<Vector<unsigned>> generatedSpirv;
+            Bool enableSpirvValidation = false;
             // Byte offset of each uniform location inside globalUboScratch, or
             // kInvalidUniformOffset. Sized maxUniformLocation + 1 by the routing pass.
             Vector<Uint> uniformOffsets;

MobileGL/MG_State/GLState/ProgramState/ProgramSpirvTask.cpp

@@ -102,7 +102,11 @@ namespace MobileGL::MG_State::GLState {
         }
 
         MGLOG_D("ProgramObject %u: Starting SPIR-V generation", externalIndex);
-        GenerateSpirv(handoff, externalIndex);
+        const Bool deferOutputValidationForDirectVulkan =
+            m_phaseA->in.env != nullptr && m_phaseA->in.env->backend == BackendType::DirectVulkan;
+        const Bool enableSpirvValidation = m_phaseA->in.enableSpirvValidation;
+        artifacts.enableSpirvValidation = enableSpirvValidation;
+        GenerateSpirv(handoff, externalIndex, deferOutputValidationForDirectVulkan, enableSpirvValidation);
         // GlslangToSpv was the only consumer of the parsed ASTs; everything after this point
         // works on the SPIR-V and on the TProgram's own self-contained reflection pool. Drop
         // them here rather than at the end of the body, which is ~87% of this node's runtime
@@ -137,7 +141,9 @@ namespace MobileGL::MG_State::GLState {
                 artifacts.generatedSpirv.size());
     }
 
-    void ProgramSpirvTask::GenerateSpirv(const ProgramLinkTask::SpirvHandoff& handoff, const Uint externalIndex) {
+    void ProgramSpirvTask::GenerateSpirv(const ProgramLinkTask::SpirvHandoff& handoff, const Uint externalIndex,
+                                         const Bool deferOutputValidationForDirectVulkan,
+                                         const Bool enableSpirvValidation) {
         /* As we passed first stage compilation/linking,
          * we'll assume all the operations here should
          * pass. We may be able to employ some optimizations
@@ -169,7 +175,8 @@ namespace MobileGL::MG_State::GLState {
         Bool allOptimized = true;
         {
             for (auto& spv : artifacts.generatedSpirv) {
-                auto success = ShaderCompiler::SanitizeAndOptimizeBinary(spv, spv);
+                auto success = ShaderCompiler::SanitizeAndOptimizeBinary(
+                    spv, spv, !deferOutputValidationForDirectVulkan, enableSpirvValidation);
                 if (!success) {
                     // The one genuine phase-B failure mode: one of the seven optimizer passes
                     // reported failure, so `spv` is whatever the run left behind. A fordebug

MobileGL/MG_State/GLState/ProgramState/ProgramSpirvTask.h

@@ -65,7 +65,8 @@ namespace MobileGL::MG_State::GLState {
     private:
         void RunBody() override;
 
-        void GenerateSpirv(const ProgramLinkTask::SpirvHandoff& handoff, Uint externalIndex);
+        void GenerateSpirv(const ProgramLinkTask::SpirvHandoff& handoff, Uint externalIndex,
+                           Bool deferOutputValidationForDirectVulkan, Bool enableSpirvValidation);
         void BuildGlobalUboRouting(const ProgramLinkTask::SpirvHandoff& handoff, Uint externalIndex);
 
         // Worker-side MGLOG replacement, replayed by the join on the GL thread. Same reason as

MobileGL/MG_State/GLState/RenderState/RenderState.cpp

@@ -25,6 +25,12 @@ namespace MobileGL {
                         return 0;
                     }
                 }
+
+                // Every viewport's scissor-test bit set, i.e. what glEnable(GL_SCISSOR_TEST) writes.
+                constexpr Uint32 kAllViewportsMask =
+                    RenderStateParameters::MAX_VIEWPORTS >= 32
+                        ? ~0u
+                        : (1u << RenderStateParameters::MAX_VIEWPORTS) - 1u;
             } // namespace
 
             RenderState::RenderState() {
@@ -32,6 +38,15 @@ namespace MobileGL {
                 for (auto& mask : m_parameters.ColorMasks) {
                     mask = BoolVec4(true, true, true, true);
                 }
+                // Every viewport's depth range starts at (0, 1) - GL 4.6 core table 23.4. The
+                // viewport and scissor rectangles legitimately start all-zero here: their spec
+                // initial value is the size of the window the context is first made current to,
+                // which the frontend does not know yet, so an all-zero rectangle means "never
+                // written" and the backends resolve it against the live surface (see
+                // DirectGLES' SyncRenderState and VulkanRenderer's ApplyGLViewportState).
+                for (auto& range : m_parameters.DepthRanges) {
+                    range = FloatVec2(0.0f, 1.0f);
+                }
             }
 
             Uint RenderState::GetVersion() const {
@@ -47,15 +62,47 @@ namespace MobileGL {
             }
 
             // -------------------- Rasterization --------------------
+            // ARB_viewport_array, "Additions to Chapter 2": Viewport(x, y, w, h) is equivalent to
+            // ViewportIndexedf(i, x, y, w, h) for every i in [0, MAX_VIEWPORTS) - it is not a
+            // synonym for "viewport 0".
             void RenderState::SetViewport(IntVec4 viewport) {
-                if (m_parameters.Viewport == viewport) return;
+                const FloatVec4 asFloat(static_cast<Float>(viewport.x()), static_cast<Float>(viewport.y()),
+                                        static_cast<Float>(viewport.z()), static_cast<Float>(viewport.w()));
+                Bool stateChanged = false;
+                for (auto& stored : m_parameters.Viewports) {
+                    if (stored == asFloat) continue;
+                    stored = asFloat;
+                    stateChanged = true;
+                }
+                if (stateChanged) ++m_version;
+            }
 
-                m_parameters.Viewport = viewport;
+            IntVec4 RenderState::GetViewport() const {
+                const FloatVec4& viewport = m_parameters.Viewports[0];
+                // Round rather than truncate: glGetIntegerv on floating-point state rounds to
+                // nearest (GL 4.6 core 22.2), and truncating a 63.5-wide viewport to 63 would
+                // also hand the backends a rectangle one pixel short of what was asked for.
+                return IntVec4(static_cast<Int>(std::lround(viewport.x())), static_cast<Int>(std::lround(viewport.y())),
+                               static_cast<Int>(std::lround(viewport.z())), static_cast<Int>(std::lround(viewport.w())));
+            }
+
+            void RenderState::SetViewportIndexed(Uint index, FloatVec4 viewport) {
+                if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                    MOBILEGL_ASSERT(false, "Viewport index out of range: %u", index);
+                    return;
+                }
+                if (m_parameters.Viewports[index] == viewport) return;
+
+                m_parameters.Viewports[index] = viewport;
                 ++m_version;
             }
 
-            const IntVec4& RenderState::GetViewport() const {
-                return m_parameters.Viewport;
+            const FloatVec4& RenderState::GetViewportIndexed(Uint index) const {
+                if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                    MOBILEGL_ASSERT(false, "Viewport index out of range: %u", index);
+                    return m_parameters.Viewports[0];
+                }
+                return m_parameters.Viewports[index];
             }
 
             void RenderState::SetLineWidth(Float width) {
@@ -223,7 +270,6 @@ namespace MobileGL {
                     SET_CAPABILITY(SampleAlphaToOne, enabled);
                     SET_CAPABILITY(SampleCoverage, enabled);
                     SET_CAPABILITY(SampleMask, enabled);
-                    SET_CAPABILITY(ScissorTest, enabled);
                     SET_CAPABILITY(StencilTest, enabled);
                     SET_CAPABILITY(ProgramPointSize, enabled);
                 case CapabilityInput::Blend: {
@@ -236,6 +282,17 @@ namespace MobileGL {
                     if (stateChanged) BumpVersions();
                     break;
                 }
+                // GL 4.6 core 17.3.2: the non-indexed Enable/Disable(SCISSOR_TEST) enables or
+                // disables the test for ALL viewports, exactly like glViewport writes all
+                // viewports. Anything narrower fails KHR-GL43.viewport_array.scissor_test_state_api,
+                // whose "enable all" phase reads every index back through glIsEnabledi.
+                case CapabilityInput::ScissorTest: {
+                    const Uint32 updated = enabled ? kAllViewportsMask : 0u;
+                    if (m_parameters.ScissorTestEnabledMask == updated) break;
+                    m_parameters.ScissorTestEnabledMask = updated;
+                    BumpVersions();
+                    break;
+                }
                 case CapabilityInput::ClipDistance0:
                 case CapabilityInput::ClipDistance1:
                 case CapabilityInput::ClipDistance2:
@@ -287,11 +344,14 @@ namespace MobileGL {
                     RETURN_CAPABILITY(SampleAlphaToOne);
                     RETURN_CAPABILITY(SampleCoverage);
                     RETURN_CAPABILITY(SampleMask);
-                    RETURN_CAPABILITY(ScissorTest);
                     RETURN_CAPABILITY(StencilTest);
                     RETURN_CAPABILITY(ProgramPointSize);
                 case CapabilityInput::Blend:
                     return m_parameters.BlendStates[0].Enabled;
+                // The non-indexed query of an indexed capability answers for index 0
+                // (GL 4.6 core 22.1), which is also the only bit either backend consumes today.
+                case CapabilityInput::ScissorTest:
+                    return (m_parameters.ScissorTestEnabledMask & 1u) != 0;
                 case CapabilityInput::ClipDistance0:
                 case CapabilityInput::ClipDistance1:
                 case CapabilityInput::ClipDistance2:
@@ -307,13 +367,29 @@ namespace MobileGL {
             }
 
             void RenderState::SetCapabilityIndexed(CapabilityInput cap, Uint index, Bool enabled) {
-                // Only for BlendState currently. The GL entry points (glEnablei/glDisablei) already
-                // reject every non-GL_BLEND target with GL_INVALID_ENUM before reaching here, so this
-                // is a backstop - but it must stay a backstop: THROW_UNIMPL_EXCEPTION unwinds a C++
-                // exception through the C GL ABI and terminates the process.
+                // GL_BLEND (indexed by draw buffer) and GL_SCISSOR_TEST (indexed by viewport) are
+                // the only indexed capabilities in GL 4.6 core. The GL entry points
+                // (glEnablei/glDisablei) already reject every other target with GL_INVALID_ENUM
+                // and every out-of-range index with GL_INVALID_VALUE before reaching here, so the
+                // guards below are backstops - but they must stay backstops:
+                // THROW_UNIMPL_EXCEPTION unwinds a C++ exception through the C GL ABI and
+                // terminates the process.
+                if (cap == CapabilityInput::ScissorTest) {
+                    if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                        MOBILEGL_ASSERT(false, "Scissor test capability index out of range: %u", index);
+                        return;
+                    }
+                    const Uint32 bit = 1u << index;
+                    const Uint32 updated = enabled ? (m_parameters.ScissorTestEnabledMask | bit)
+                                                   : (m_parameters.ScissorTestEnabledMask & ~bit);
+                    if (updated == m_parameters.ScissorTestEnabledMask) return;
+                    m_parameters.ScissorTestEnabledMask = updated;
+                    BumpVersions();
+                    return;
+                }
                 if (cap != CapabilityInput::Blend) {
                     MGLOG_I("RenderState::SetCapabilityIndexed: indexed capability state exists only for "
-                            "GL_BLEND (cap=%d, index=%u); ignoring",
+                            "GL_BLEND and GL_SCISSOR_TEST (cap=%d, index=%u); ignoring",
                             static_cast<int>(cap), index);
                     return;
                 }
@@ -328,9 +404,17 @@ namespace MobileGL {
             }
 
             Bool RenderState::IsCapabilityEnabledIndexed(CapabilityInput cap, Uint index) const {
-                // Only for BlendState currently - same backstop reasoning as SetCapabilityIndexed:
-                // glIsEnabledi has already answered GL_INVALID_ENUM/GL_FALSE for anything else, and a
-                // query must never be able to terminate the process.
+                // GL_BLEND and GL_SCISSOR_TEST only - same backstop reasoning as
+                // SetCapabilityIndexed: glIsEnabledi has already answered
+                // GL_INVALID_ENUM/GL_INVALID_VALUE for anything else, and a query must never be
+                // able to terminate the process.
+                if (cap == CapabilityInput::ScissorTest) {
+                    if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                        MOBILEGL_ASSERT(false, "Scissor test capability index out of range: %u", index);
+                        return false;
+                    }
+                    return (m_parameters.ScissorTestEnabledMask & (1u << index)) != 0;
+                }
                 if (cap != CapabilityInput::Blend) {
                     MGLOG_I("RenderState::IsCapabilityEnabledIndexed: indexed capability state exists only "
                             "for GL_BLEND (cap=%d, index=%u); reporting disabled",
@@ -591,15 +675,39 @@ namespace MobileGL {
                 return m_parameters.BlendColor;
             }
 
+            // Like Viewport: ARB_viewport_array makes DepthRange(n, f) the same as
+            // DepthRangeIndexed(i, n, f) for every i.
             void RenderState::SetDepthRange(FloatVec2 range) {
-                if (m_parameters.DepthRange == range) return;
-
-                m_parameters.DepthRange = range;
-                ++m_version;
+                Bool stateChanged = false;
+                for (auto& stored : m_parameters.DepthRanges) {
+                    if (stored == range) continue;
+                    stored = range;
+                    stateChanged = true;
+                }
+                if (stateChanged) ++m_version;
             }
 
             const FloatVec2& RenderState::GetDepthRange() const {
-                return m_parameters.DepthRange;
+                return m_parameters.DepthRanges[0];
+            }
+
+            void RenderState::SetDepthRangeIndexed(Uint index, FloatVec2 range) {
+                if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                    MOBILEGL_ASSERT(false, "Depth range index out of range: %u", index);
+                    return;
+                }
+                if (m_parameters.DepthRanges[index] == range) return;
+
+                m_parameters.DepthRanges[index] = range;
+                ++m_version;
+            }
+
+            const FloatVec2& RenderState::GetDepthRangeIndexed(Uint index) const {
+                if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                    MOBILEGL_ASSERT(false, "Depth range index out of range: %u", index);
+                    return m_parameters.DepthRanges[0];
+                }
+                return m_parameters.DepthRanges[index];
             }
 
             void RenderState::SetSampleCoverage(Float value, Bool invert) {
@@ -726,15 +834,39 @@ namespace MobileGL {
             }
 
             // --------------------- Scissor ---------------------
+            // Like Viewport: ARB_viewport_array makes Scissor(x, y, w, h) the same as
+            // ScissorIndexed(i, x, y, w, h) for every i.
             void RenderState::SetScissorBox(IntVec4 box) {
-                if (m_parameters.ScissorBox == box) return;
-
-                m_parameters.ScissorBox = box;
-                ++m_version;
+                Bool stateChanged = false;
+                for (auto& stored : m_parameters.ScissorBoxes) {
+                    if (stored == box) continue;
+                    stored = box;
+                    stateChanged = true;
+                }
+                if (stateChanged) ++m_version;
             }
 
             const IntVec4& RenderState::GetScissorBox() const {
-                return m_parameters.ScissorBox;
+                return m_parameters.ScissorBoxes[0];
+            }
+
+            void RenderState::SetScissorBoxIndexed(Uint index, IntVec4 box) {
+                if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                    MOBILEGL_ASSERT(false, "Scissor box index out of range: %u", index);
+                    return;
+                }
+                if (m_parameters.ScissorBoxes[index] == box) return;
+
+                m_parameters.ScissorBoxes[index] = box;
+                ++m_version;
+            }
+
+            const IntVec4& RenderState::GetScissorBoxIndexed(Uint index) const {
+                if (index >= RenderStateParameters::MAX_VIEWPORTS) {
+                    MOBILEGL_ASSERT(false, "Scissor box index out of range: %u", index);
+                    return m_parameters.ScissorBoxes[0];
+                }
+                return m_parameters.ScissorBoxes[index];
             }
         } // namespace GLState
     } // namespace MG_State

MobileGL/MG_State/GLState/RenderState/RenderState.h

@@ -220,8 +220,22 @@ namespace MobileGL {
     };
 
     struct RenderStateParameters {
+        // ARB_viewport_array / GL 4.6 core 13.6.1: the viewport, the scissor rectangle, the depth
+        // range and the scissor-test enable are all arrays indexed by gl_ViewportIndex, and the
+        // spec floor for MAX_VIEWPORTS is 16. MobileGL advertises exactly 16 on both backends, so
+        // this is also what GL_MAX_VIEWPORTS reports (see the backend loaders' caps.MaxViewports).
+        static constexpr Uint MAX_VIEWPORTS = 16;
+
         // Rasterization
-        IntVec4 Viewport = IntVec4(0, 0, 0, 0); // x, y, width, height
+        // The viewport rectangle is FLOAT state as of GL 4.1 - ViewportIndexedf writes fractional
+        // values and GetFloati_v(GL_VIEWPORT) must hand them back bit-exact
+        // (KHR-GL43.viewport_array.viewport_api compares with ==, no tolerance). glViewport's
+        // integers are simply one way to write it. Index 0 is what a program that never assigns
+        // gl_ViewportIndex rasterizes against, and what the classic glViewport /
+        // glGetIntegerv(GL_VIEWPORT) pair addresses. Both backends rasterize the rectangle
+        // rounded back to integers; the STATE stays exact, which is the half the conformance
+        // suite checks (see the KNOWN INFIDELITY note in AdvertisedLimitsScenario.cpp).
+        Array<FloatVec4, MAX_VIEWPORTS> Viewports{}; // x, y, width, height
         Float LineWidth = 1.0f;
         Float PointSize = 1.0f;
         // GL_PATCH_VERTICES: how many vertices one tessellation patch consumes.
@@ -247,7 +261,13 @@ namespace MobileGL {
         Float ClearDepth = 1.0f;
         Uint32 ClearStencil = 0;
         FloatVec4 BlendColor = FloatVec4(0.0f, 0.0f, 0.0f, 0.0f);
-        FloatVec2 DepthRange = FloatVec2(0.0f, 1.0f);
+        // Per-viewport depth range (glDepthRangeIndexed / glDepthRangeArrayv). Every entry is
+        // initialized to (0, 1) in RenderState's constructor - a default member initializer would
+        // not survive the Array<> aggregate. Kept float rather than double: DepthRangeArrayv takes
+        // GLdouble, but the value reaches the hardware as VkViewport::minDepth/maxDepth (float) on
+        // Magma and glDepthRangef on Espryt, so a double store would only widen the readback and
+        // then lose it again at the same place.
+        Array<FloatVec2, MAX_VIEWPORTS> DepthRanges{};
         Float SampleCoverageValue = 1.0f;
         Bool SampleCoverageInvert = false;
         Uint32 SampleMaskValue = 0xffffffffu;
@@ -299,10 +319,15 @@ namespace MobileGL {
         Bool SampleAlphaToOneEnabled = false;
         Bool SampleCoverageEnabled = false;
         Bool SampleMaskEnabled = false;
-        Bool ScissorTestEnabled = false;
         Bool StencilTestEnabled = false;
         Bool ProgramPointSizeEnabled = false;
-        IntVec4 ScissorBox = IntVec4(0, 0, 0, 0); // x, y, width, height
+        // glEnable(GL_SCISSOR_TEST) enables the test for EVERY viewport, glEnablei for one
+        // (GL 4.6 core 17.3.2), so this is 16 bits and not a bool. Bit 0 is what the classic
+        // glIsEnabled(GL_SCISSOR_TEST) reports and what both backends currently consume. Unlike
+        // ClipDistanceEnabledMask below it DOES bump the pipeline version, because DirectGLES
+        // turns it into a real glEnable/glDisable.
+        Uint32 ScissorTestEnabledMask = 0;
+        Array<IntVec4, MAX_VIEWPORTS> ScissorBoxes{}; // x, y, width, height
         // glEnable(GL_CLIP_DISTANCE0 + i) for i in [0, 8), one bit each. A bitmask rather than
         // eight bools because every consumer wants the set, not an individual flag, and because
         // the SYNC_CAPABILITY/SET_CAPABILITY macros key off a "<Name>Enabled" field name that
@@ -323,8 +348,14 @@ namespace MobileGL {
                 const RenderStateParameters& GetAllParameters() const;
 
                 // Rasterization
+                // ARB_viewport_array defines glViewport as ViewportIndexedf on EVERY index, so the
+                // classic setter broadcasts; GetViewport answers for index 0 (rounded to the
+                // integers glGetIntegerv(GL_VIEWPORT) and both backends want) and is BY VALUE for
+                // that reason. The indexed pair is the verbatim float state.
                 void SetViewport(IntVec4 viewport); // x, y, width, height
-                const IntVec4& GetViewport() const; // x, y, width, height
+                IntVec4 GetViewport() const;        // x, y, width, height, viewport 0, rounded
+                void SetViewportIndexed(Uint index, FloatVec4 viewport);
+                const FloatVec4& GetViewportIndexed(Uint index) const;
                 void SetLineWidth(Float width);
                 Float GetLineWidth() const;
                 void SetPointSize(Float size);
@@ -400,8 +431,12 @@ namespace MobileGL {
                 Uint32 GetClearStencil() const;
                 void SetBlendColor(FloatVec4 color);
                 const FloatVec4& GetBlendColor() const;
+                // glDepthRange(f) writes every viewport's range (ARB_viewport_array); the indexed
+                // pair is glDepthRangeIndexed / glDepthRangeArrayv. GetDepthRange answers index 0.
                 void SetDepthRange(FloatVec2 range);
                 const FloatVec2& GetDepthRange() const;
+                void SetDepthRangeIndexed(Uint index, FloatVec2 range);
+                const FloatVec2& GetDepthRangeIndexed(Uint index) const;
                 void SetSampleCoverage(Float value, Bool invert);
                 Float GetSampleCoverageValue() const;
                 Bool GetSampleCoverageInvert() const;
@@ -421,9 +456,12 @@ namespace MobileGL {
                 void SetProvokingVertexMode(ProvokingVertexMode mode);
                 ProvokingVertexMode GetProvokingVertexMode() const;
 
-                // Scissor
+                // Scissor. glScissor writes every rectangle (ARB_viewport_array); GetScissorBox
+                // answers for index 0.
                 void SetScissorBox(IntVec4 box);      // x, y, width, height
                 const IntVec4& GetScissorBox() const; // x, y, width, height
+                void SetScissorBoxIndexed(Uint index, IntVec4 box);
+                const IntVec4& GetScissorBoxIndexed(Uint index) const;
 
             private:
                 // Bump both: any state change invalidates the draw snapshot, and this one also

MobileGL/MG_Test/Backend/Diligent/CMakeLists.txt

@@ -0,0 +1,22 @@
+cmake_minimum_required(VERSION 3.14)
+
+message(STATUS "Generating build files for MobileGL Diligent Backend Test...")
+
+add_executable(
+        DiligentVulkanSanityTest
+        SanityTest.cpp
+)
+
+target_include_directories(DiligentVulkanSanityTest PRIVATE
+        ${MGL_ROOT}/include
+        ${MGL_ROOT}/MobileGL
+)
+
+target_link_libraries(
+        DiligentVulkanSanityTest PRIVATE
+        GTest::gtest_main
+        ${LINK_LIBRARIES}
+)
+
+include(GoogleTest)
+gtest_discover_tests(DiligentVulkanSanityTest DISCOVERY_TIMEOUT 30 PROPERTIES LABELS integration)

MobileGL/MG_Test/Backend/DirectVulkan/SanityTest.cpp

@@ -8,9 +8,28 @@
 
 #include <gtest/gtest.h>
 #include <iostream>
+#include <utility>
 #include <vector>
 
-#include <vulkan/vulkan.h>
+#include <MG_Backend/DirectVulkan/Renderer/ProgramFactory.h>
+
+TEST(DirectVulkanSanity, ProgramMovePreservesViewportIndexUsage) {
+    using VkProgramObject = MobileGL::MG_Backend::DirectVulkan::ProgramFactory::VkProgramObject;
+
+    VkProgramObject moveConstructedSource;
+    moveConstructedSource.writesViewportIndexBuiltin = true;
+    VkProgramObject moveConstructed(std::move(moveConstructedSource));
+    EXPECT_TRUE(moveConstructed.writesViewportIndexBuiltin);
+    EXPECT_FALSE(moveConstructedSource.writesViewportIndexBuiltin);
+
+    VkProgramObject moveAssignedSource;
+    moveAssignedSource.writesViewportIndexBuiltin = true;
+    VkProgramObject moveAssigned;
+    moveAssigned = std::move(moveAssignedSource);
+    EXPECT_TRUE(moveAssigned.writesViewportIndexBuiltin);
+    EXPECT_FALSE(moveAssignedSource.writesViewportIndexBuiltin);
+}
+
 TEST(DirectVulkanSanity, ExtensionEnumeration) {
     uint32_t extensionCount = 0;
     vkEnumerateInstanceExtensionProperties(nullptr, &extensionCount, nullptr);

MobileGL/MG_Test/BackendLoader/BackendLoaderTest.cpp

@@ -725,22 +725,57 @@ TEST(TextureAnisotropyCapabilities, ExtensionIsAdvertisedOnlyWhenTheHostDriverSu
         return std::find(extensions.begin(), extensions.end(), wanted) != extensions.end();
     };
 
-    const auto without = MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false);
+    const auto without = MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false, false, false);
     EXPECT_FALSE(contains(without, MobileGL::E_GL_EXT_texture_filter_anisotropic));
     EXPECT_FALSE(contains(without, MobileGL::E_GL_ARB_texture_filter_anisotropic));
 
-    const auto with = MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, true);
+    const auto with = MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, true, false, false);
     EXPECT_TRUE(contains(with, MobileGL::E_GL_EXT_texture_filter_anisotropic));
     EXPECT_TRUE(contains(with, MobileGL::E_GL_ARB_texture_filter_anisotropic));
 
     // Same rule on the Vulkan backend, where the gate is the samplerAnisotropy device feature.
-    const auto vkWithout = MobileGL::MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false);
+    const auto vkWithout = MobileGL::MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false, false);
     EXPECT_FALSE(contains(vkWithout, MobileGL::E_GL_EXT_texture_filter_anisotropic));
-    const auto vkWith = MobileGL::MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, true);
+    const auto vkWith = MobileGL::MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, true, false);
     EXPECT_TRUE(contains(vkWith, MobileGL::E_GL_EXT_texture_filter_anisotropic));
     EXPECT_TRUE(contains(vkWith, MobileGL::E_GL_ARB_texture_filter_anisotropic));
 }
 
+// Minecraft 26.3 checks ARB_draw_indirect before it considers the already-advertised
+// ARB_multi_draw_indirect, then separately requires ARB_base_instance before enabling its terrain
+// indirect path. Pin both strings and, just as importantly, the non-zero firstInstance gate.
+TEST(IndirectDrawAdvertisement, MatchesEachBackendsUsableCommandSemantics) {
+    const auto contains = [](const MobileGL::Vector<MobileGL::GLExtension>& extensions,
+                             MobileGL::GLExtension wanted) {
+        return std::find(extensions.begin(), extensions.end(), wanted) != extensions.end();
+    };
+
+    const auto esWithoutIndirect =
+        MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false, false, false);
+    EXPECT_FALSE(contains(esWithoutIndirect, MobileGL::E_GL_ARB_draw_indirect));
+    EXPECT_FALSE(contains(esWithoutIndirect, MobileGL::E_GL_ARB_base_instance));
+
+    const auto esWithoutBaseInstance =
+        MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false, true, false);
+    EXPECT_TRUE(contains(esWithoutBaseInstance, MobileGL::E_GL_ARB_draw_indirect));
+    EXPECT_FALSE(contains(esWithoutBaseInstance, MobileGL::E_GL_ARB_base_instance));
+
+    const auto esWithBoth =
+        MobileGL::MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false, true, true);
+    EXPECT_TRUE(contains(esWithBoth, MobileGL::E_GL_ARB_draw_indirect));
+    EXPECT_TRUE(contains(esWithBoth, MobileGL::E_GL_ARB_base_instance));
+
+    const auto vkWithoutBaseInstance =
+        MobileGL::MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false, false);
+    EXPECT_TRUE(contains(vkWithoutBaseInstance, MobileGL::E_GL_ARB_draw_indirect));
+    EXPECT_FALSE(contains(vkWithoutBaseInstance, MobileGL::E_GL_ARB_base_instance));
+
+    const auto vkWithBoth =
+        MobileGL::MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false, true);
+    EXPECT_TRUE(contains(vkWithBoth, MobileGL::E_GL_ARB_draw_indirect));
+    EXPECT_TRUE(contains(vkWithBoth, MobileGL::E_GL_ARB_base_instance));
+}
+
 TEST(TextureAnisotropyCapabilities, MaxAnisotropyIsQueriedOnlyWhenTheExtensionIsPresent) {
     ResetFakeDriver();
     g_fake.maxVertexSsboBlocks = 0;
@@ -836,3 +871,63 @@ TEST(MultiDrawCapabilities, ExtensionWithoutResolvedPointerIsNotSupport) {
     EXPECT_FALSE(caps.SupportsMultiDrawIndirect);
     EXPECT_FALSE(caps.SupportsMultiDrawElementsBaseVertex);
 }
+
+TEST(DrawIndirectCapabilities, RequiresEs31AndBothCoreEntryPoints) {
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    auto funcs = MakeFakeGLESFunctions();
+    funcs.glDrawElementsIndirect = [](GLenum, GLenum, const void*) {};
+
+    MobileGL::MG_External::GLESCapabilities supportedCaps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(supportedCaps, funcs));
+    EXPECT_TRUE(supportedCaps.SupportsDrawIndirect);
+
+    // The same pointers on an ES 3.0 context are not core entry points and cannot back the
+    // desktop extension contract.
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.glesMinorVersion = 0;
+    MobileGL::MG_External::GLESCapabilities es30Caps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(es30Caps, funcs));
+    EXPECT_FALSE(es30Caps.SupportsDrawIndirect);
+
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    const auto missingElements = MakeFakeGLESFunctions();
+    MobileGL::MG_External::GLESCapabilities missingEntryPointCaps;
+    ASSERT_TRUE(
+        MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(missingEntryPointCaps, missingElements));
+    EXPECT_FALSE(missingEntryPointCaps.SupportsDrawIndirect);
+}
+
+TEST(BaseInstanceCapabilities, RequiresTheExtensionAndAllThreeEntryPoints) {
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    auto funcs = MakeFakeGLESFunctions();
+    funcs.glDrawArraysInstancedBaseInstanceEXT = [](GLenum, GLint, GLsizei, GLsizei, GLuint) {};
+    funcs.glDrawElementsInstancedBaseInstanceEXT =
+        [](GLenum, GLsizei, GLenum, const void*, GLsizei, GLuint) {};
+    funcs.glDrawElementsInstancedBaseVertexBaseInstanceEXT =
+        [](GLenum, GLsizei, GLenum, const void*, GLsizei, GLint, GLuint) {};
+
+    // Resolved stubs alone must never make the capability true.
+    MobileGL::MG_External::GLESCapabilities pointersOnlyCaps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(pointersOnlyCaps, funcs));
+    EXPECT_FALSE(pointersOnlyCaps.SupportsBaseInstance);
+
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.extensions.emplace_back("GL_EXT_base_instance");
+    MobileGL::MG_External::GLESCapabilities supportedCaps;
+    ASSERT_TRUE(MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(supportedCaps, funcs));
+    EXPECT_TRUE(supportedCaps.SupportsBaseInstance);
+
+    ResetFakeDriver();
+    g_fake.maxVertexSsboBlocks = 0;
+    g_fake.extensions.emplace_back("GL_EXT_base_instance");
+    funcs.glDrawElementsInstancedBaseInstanceEXT = nullptr;
+    MobileGL::MG_External::GLESCapabilities missingEntryPointCaps;
+    ASSERT_TRUE(
+        MobileGL::MG_Util::BackendLoader::FillInGLESCapabilities(missingEntryPointCaps, funcs));
+    EXPECT_FALSE(missingEntryPointCaps.SupportsBaseInstance);
+}

MobileGL/MG_Test/Buffer/BufferTest.cpp

@@ -16,6 +16,7 @@
 #include <MG_State/GLState/Core.h>
 
 #include <MG_Impl/GLImpl/Buffer/GL_Buffer.h>
+#include <MG_Impl/GetProcAddress.h>
 #include <MG_Impl/GLImpl/Getter/GL_Getter.h>
 
 using namespace MobileGL;
@@ -599,6 +600,117 @@ TEST_F(BufferTest, ClearNamedBufferSubDataRepeatsPattern) {
     EXPECT_EQ(actual, (Vector<Uint32>{0, pattern, pattern, pattern, 0}));
     EXPECT_EQ(MobileGL::MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
 }
+TEST_F(BufferTest, ClearBufferSubDataInitializesIrisStaticSsboRange) {
+    GLuint buffer = 0;
+    MobileGL::MG_Impl::GLImpl::GenBuffers(1, &buffer);
+    MobileGL::MG_Impl::GLImpl::BindBuffer(GL_SHADER_STORAGE_BUFFER, buffer);
+
+    Vector<Uint8> initial(32, 0x7F);
+    MobileGL::MG_Impl::GLImpl::BufferData(
+        GL_SHADER_STORAGE_BUFFER, initial.size(), initial.data(), GL_STATIC_DRAW);
+    const GLbyte zero = 0;
+    const auto clear = reinterpret_cast<PFNGLCLEARBUFFERSUBDATAPROC>(
+        MobileGL::MG_Impl::GetProcAddress("glClearBufferSubData"));
+    ASSERT_NE(clear, nullptr);
+    clear(GL_SHADER_STORAGE_BUFFER, GL_R8, 4, 24, GL_RED, GL_BYTE, &zero);
+
+    Vector<Uint8> actual(initial.size());
+    auto bufferObject = MobileGL::MG_State::pGLContext->GetBufferObject(buffer);
+    ASSERT_NE(bufferObject, nullptr);
+    Memcpy(actual.data(), bufferObject->AcquireMemory(false, true, false), actual.size());
+    EXPECT_EQ(actual, (Vector<Uint8>{0x7F, 0x7F, 0x7F, 0x7F,
+                                    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                                    0x7F, 0x7F, 0x7F, 0x7F}));
+    EXPECT_EQ(MobileGL::MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MobileGL::MG_Impl::GLImpl::BindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+    MobileGL::MG_Impl::GLImpl::DeleteBuffers(1, &buffer);
+    DrainPendingGlErrors();
+}
+
+TEST_F(BufferTest, ClearBufferSubDataInitializesCompleteIrisStaticSsbo) {
+    constexpr SizeT irisStaticSsboSize = 5'000'192;
+    GLuint buffer = 0;
+    MobileGL::MG_Impl::GLImpl::GenBuffers(1, &buffer);
+    MobileGL::MG_Impl::GLImpl::BindBuffer(GL_SHADER_STORAGE_BUFFER, buffer);
+
+    Vector<Uint8> initial(irisStaticSsboSize, 0x7F);
+    MobileGL::MG_Impl::GLImpl::BufferData(
+        GL_SHADER_STORAGE_BUFFER, initial.size(), initial.data(), GL_STATIC_DRAW);
+    const GLbyte zero = 0;
+    MobileGL::MG_Impl::GLImpl::ClearBufferSubData(
+        GL_SHADER_STORAGE_BUFFER, GL_R8, 0, irisStaticSsboSize, GL_RED, GL_BYTE, &zero);
+
+    Vector<Uint8> actual(irisStaticSsboSize);
+    auto bufferObject = MobileGL::MG_State::pGLContext->GetBufferObject(buffer);
+    ASSERT_NE(bufferObject, nullptr);
+    Memcpy(actual.data(), bufferObject->AcquireMemory(false, true, false), actual.size());
+    EXPECT_EQ(actual, Vector<Uint8>(irisStaticSsboSize, 0));
+    EXPECT_EQ(MobileGL::MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MobileGL::MG_Impl::GLImpl::BindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+    MobileGL::MG_Impl::GLImpl::DeleteBuffers(1, &buffer);
+    DrainPendingGlErrors();
+}
+
+TEST_F(BufferTest, ClearBufferDataConvertsOneClientPixelBeforeRepeatingIt) {
+    GLuint buffer = 0;
+    MobileGL::MG_Impl::GLImpl::GenBuffers(1, &buffer);
+    MobileGL::MG_Impl::GLImpl::BindBuffer(GL_ARRAY_BUFFER, buffer);
+
+    Vector<Uint32> initial(4, 0u);
+    MobileGL::MG_Impl::GLImpl::BufferData(GL_ARRAY_BUFFER, initial.size() * sizeof(Uint32), initial.data(),
+                                           GL_STATIC_DRAW);
+    const Uint8 value = 0xAB;
+    MobileGL::MG_Impl::GLImpl::ClearBufferData(
+        GL_ARRAY_BUFFER, GL_R32UI, GL_RED_INTEGER, GL_UNSIGNED_BYTE, &value);
+
+    Vector<Uint32> actual(initial.size());
+    auto bufferObject = MobileGL::MG_State::pGLContext->GetBufferObject(buffer);
+    ASSERT_NE(bufferObject, nullptr);
+    Memcpy(actual.data(), bufferObject->AcquireMemory(false, true, false), actual.size() * sizeof(Uint32));
+    EXPECT_EQ(actual, Vector<Uint32>(initial.size(), value));
+    EXPECT_EQ(MobileGL::MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
+
+    MobileGL::MG_Impl::GLImpl::BindBuffer(GL_ARRAY_BUFFER, 0);
+    MobileGL::MG_Impl::GLImpl::DeleteBuffers(1, &buffer);
+    DrainPendingGlErrors();
+}
+
+TEST_F(BufferTest, ClearBufferSubDataRejectsUnboundTarget) {
+    MobileGL::MG_Impl::GLImpl::BindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+    const GLbyte zero = 0;
+    MobileGL::MG_Impl::GLImpl::ClearBufferSubData(
+        GL_SHADER_STORAGE_BUFFER, GL_R8, 0, 1, GL_RED, GL_BYTE, &zero);
+    ExpectSingleGlError(GL_INVALID_OPERATION);
+}
+
+TEST_F(BufferTest, ClearBufferDataRejectsInvalidPixelFormatTypePairs) {
+    GLuint buffer = 0;
+    MobileGL::MG_Impl::GLImpl::GenBuffers(1, &buffer);
+    MobileGL::MG_Impl::GLImpl::BindBuffer(GL_ARRAY_BUFFER, buffer);
+
+    const Vector<Uint8> initial{0x7F, 0x7F};
+    MobileGL::MG_Impl::GLImpl::BufferData(GL_ARRAY_BUFFER, initial.size(), initial.data(), GL_STATIC_DRAW);
+    const Uint16 packed = 0;
+    MobileGL::MG_Impl::GLImpl::ClearBufferData(
+        GL_ARRAY_BUFFER, GL_R16, GL_RED, GL_UNSIGNED_SHORT_5_6_5, &packed);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MobileGL::MG_Impl::GLImpl::ClearBufferData(
+        GL_ARRAY_BUFFER, GL_R16, GL_RED, GL_UNSIGNED_SHORT_5_6_5, nullptr);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    Vector<Uint8> actual(initial.size());
+    auto bufferObject = MobileGL::MG_State::pGLContext->GetBufferObject(buffer);
+    ASSERT_NE(bufferObject, nullptr);
+    Memcpy(actual.data(), bufferObject->AcquireMemory(false, true, false), actual.size());
+    EXPECT_EQ(actual, initial);
+
+    MobileGL::MG_Impl::GLImpl::BindBuffer(GL_ARRAY_BUFFER, 0);
+    MobileGL::MG_Impl::GLImpl::DeleteBuffers(1, &buffer);
+    DrainPendingGlErrors();
+}
 
 
 // GL 4.6 core 6.5: glBufferSubData fails only when the written range OVERLAPS the mapped range.

MobileGL/MG_Test/CMakeLists.txt

@@ -84,3 +84,6 @@ add_subdirectory(Backend/DirectGLES)
 if (ENABLE_INTEGRATION_TESTS)
     add_subdirectory(Backend/DirectVulkan)
 endif()
+if (MOBILEGL_ENABLE_DILIGENT)
+    add_subdirectory(Backend/Diligent)
+endif()

MobileGL/MG_Test/Pipeline/CMakeLists.txt

@@ -4,6 +4,7 @@ add_executable(
         PipelineQuirkTest
         PipelineQuirkTest.cpp
         PassthroughTessControlTest.cpp
+        ViewportIndexReflectionTest.cpp
 )
 
 target_include_directories(PipelineQuirkTest PRIVATE

MobileGL/MG_Test/Pipeline/ViewportIndexReflectionTest.cpp

@@ -0,0 +1,183 @@
+// MobileGL - MobileGL/MG_Test/Pipeline/ViewportIndexReflectionTest.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
+//
+// ProgramFactory::ReflectedWritesViewportIndexBuiltin is the switch that decides whether a
+// DirectVulkan pipeline declares one viewport or all sixteen. Getting it wrong is silent in both
+// directions and neither direction is caught by a state test:
+//
+//   - a false NEGATIVE collapses every gl_ViewportIndex onto viewport 0, which is precisely the
+//     bug the multi-viewport work exists to fix and which a set/get round trip cannot see;
+//   - a false POSITIVE widens viewportCount for an ordinary Minecraft shader, costing a longer
+//     vkCmdSetViewport per state change and, on a tiler, possibly a hardware fast path.
+//
+// So this compiles REAL GLSL through the same glslang path the renderer uses and reflects the
+// SPIR-V that comes out, rather than asserting against hand-assembled words: what has to hold is
+// that the detector agrees with what glslang actually emits for a shader that writes the builtin,
+// including the stage-by-stage question of WHERE it may be written (GL 4.1 allows the geometry
+// stage; ARB_shader_viewport_layer_array adds vertex and tessellation evaluation).
+//
+// The end-to-end claim - that a detected writer really does route pixels to its own viewport -
+// lives in MG_IntegrationTest/Scenarios/ViewportArrayScenario.cpp.
+
+#include <gtest/gtest.h>
+
+#include <string>
+#include <vector>
+
+#include "Includes.h"
+#include "Init.h"
+
+#include <MG_Backend/DirectVulkan/Renderer/ProgramFactory.h>
+#include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
+#include <MG_Util/ShaderTranspiler/Types.h>
+
+#include <spirv_reflect.h>
+
+using namespace MobileGL;
+using MobileGL::MG_Backend::DirectVulkan::ProgramFactory;
+using MobileGL::MG_Util::ShaderTranspiler::ShaderCompiler;
+
+namespace {
+
+    Vector<Uint32> CompileToSpirv(GLenum stage, const String& source) {
+        using namespace MG_Util::ShaderTranspiler;
+        ShaderAttrib shaderAttrib{.shaderType = stage, .sourceStr = source};
+        auto shaderResult = ShaderCompiler::CompileShader(shaderAttrib);
+        EXPECT_TRUE(shaderResult) << (shaderResult ? String{} : shaderResult.error().log);
+        if (!shaderResult) return {};
+
+        ProgramAttrib programAttrib{.shaders = {shaderResult.value()}};
+        auto programResult = ShaderCompiler::LinkProgram(programAttrib);
+        EXPECT_TRUE(programResult) << (programResult ? String{} : programResult.error().log);
+        if (!programResult) return {};
+
+        ProgramBinaryAttrib binaryAttrib{.shaderTypes = {stage}, .program = *programResult.value()};
+        auto binaryResult = ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
+        EXPECT_TRUE(binaryResult) << (binaryResult ? String{} : binaryResult.error().log);
+        if (!binaryResult || binaryResult->empty()) return {};
+        return binaryResult->front();
+    }
+
+    // Owns the reflection module so a failing EXPECT cannot leak it.
+    class ReflectModule {
+    public:
+        explicit ReflectModule(const Vector<Uint32>& spirv) {
+            if (spirv.empty()) return;
+            m_created = spvReflectCreateShaderModule(spirv.size() * sizeof(Uint32), spirv.data(), &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;
+    };
+
+    class ViewportIndexReflectionTest : public ::testing::Test {
+    protected:
+        void SetUp() override { MobileGL::Initialize(); }
+    };
+
+    const char* const kGeometryWritesViewportIndex = R"(#version 410 core
+layout(points, invocations = 16) in;
+layout(triangle_strip, max_vertices = 4) out;
+void main() {
+    gl_ViewportIndex = gl_InvocationID;
+    gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+    // Same stage, same shape, writing gl_Layer INSTEAD. Layered rendering and viewport routing
+    // are different features and the detector must not confuse them: a Minecraft-style cubemap
+    // pass writes gl_Layer and must keep the one-viewport pipeline.
+    const char* const kGeometryWritesLayerOnly = R"(#version 410 core
+layout(points, invocations = 6) in;
+layout(triangle_strip, max_vertices = 4) out;
+void main() {
+    gl_Layer = gl_InvocationID;
+    gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+    const char* const kPlainGeometry = R"(#version 410 core
+layout(points, invocations = 1) in;
+layout(triangle_strip, max_vertices = 4) out;
+void main() {
+    gl_Position = vec4(-1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0, -1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4(-1.0,  1.0, 0.0, 1.0); EmitVertex();
+    gl_Position = vec4( 1.0,  1.0, 0.0, 1.0); EmitVertex();
+    EndPrimitive();
+}
+)";
+
+    const char* const kPlainVertex = R"(#version 410 core
+void main() { gl_Position = vec4(0.0, 0.0, 0.0, 1.0); }
+)";
+
+    const char* const kPlainFragment = R"(#version 410 core
+layout(location = 0) out vec4 fragColor;
+void main() { fragColor = vec4(1.0); }
+)";
+
+    TEST_F(ViewportIndexReflectionTest, TrueForAGeometryShaderThatAssignsViewportIndex) {
+        const ReflectModule module(CompileToSpirv(GL_GEOMETRY_SHADER, kGeometryWritesViewportIndex));
+        ASSERT_TRUE(module.Created());
+        EXPECT_TRUE(ProgramFactory::ReflectedWritesViewportIndexBuiltin(module.Get()))
+            << "a shader that assigns gl_ViewportIndex must get a multi-viewport pipeline; missing it is what "
+               "collapses every index onto viewport 0";
+    }
+
+    TEST_F(ViewportIndexReflectionTest, FalseForAGeometryShaderThatOnlyAssignsLayer) {
+        const ReflectModule module(CompileToSpirv(GL_GEOMETRY_SHADER, kGeometryWritesLayerOnly));
+        ASSERT_TRUE(module.Created());
+        EXPECT_FALSE(ProgramFactory::ReflectedWritesViewportIndexBuiltin(module.Get()))
+            << "gl_Layer is layered rendering, not viewport routing; widening viewportCount for it costs the "
+               "single-viewport fast path for nothing";
+    }
+
+    TEST_F(ViewportIndexReflectionTest, FalseForAPlainGeometryShader) {
+        const ReflectModule module(CompileToSpirv(GL_GEOMETRY_SHADER, kPlainGeometry));
+        ASSERT_TRUE(module.Created());
+        EXPECT_FALSE(ProgramFactory::ReflectedWritesViewportIndexBuiltin(module.Get()));
+    }
+
+    TEST_F(ViewportIndexReflectionTest, FalseForTheOrdinaryVertexAndFragmentStages) {
+        // The shape every real application ships: neither stage may widen the pipeline.
+        const ReflectModule vertexModule(CompileToSpirv(GL_VERTEX_SHADER, kPlainVertex));
+        ASSERT_TRUE(vertexModule.Created());
+        EXPECT_FALSE(ProgramFactory::ReflectedWritesViewportIndexBuiltin(vertexModule.Get()));
+
+        const ReflectModule fragmentModule(CompileToSpirv(GL_FRAGMENT_SHADER, kPlainFragment));
+        ASSERT_TRUE(fragmentModule.Created());
+        EXPECT_FALSE(ProgramFactory::ReflectedWritesViewportIndexBuiltin(fragmentModule.Get()));
+    }
+
+    TEST_F(ViewportIndexReflectionTest, FalseForAnEmptyModuleWithoutDereferencing) {
+        // A default-constructed module has no entry points. The scan runs on every link, so it
+        // must survive a reflection that never got built rather than walk a null array.
+        SpvReflectShaderModule emptyModule{};
+        EXPECT_FALSE(ProgramFactory::ReflectedWritesViewportIndexBuiltin(emptyModule));
+    }
+
+} // namespace

MobileGL/MG_Test/Program/ParallelShaderCompileTest.cpp

@@ -516,17 +516,17 @@ TEST_F(ParallelShaderCompileTest, MaxShaderCompilerThreadsIgnoresTheCurrentBudge
 TEST_F(ParallelShaderCompileTest, BothBackendsAdvertiseTheExtensionIffAsyncIsEnabled) {
     {
         const AsyncModeScope async(true);
-        EXPECT_TRUE(Advertises(MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false),
+        EXPECT_TRUE(Advertises(MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false, false, false),
                                E_GL_KHR_parallel_shader_compile));
-        EXPECT_TRUE(Advertises(MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false),
+        EXPECT_TRUE(Advertises(MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false, false),
                                E_GL_KHR_parallel_shader_compile));
     }
     {
         const AsyncModeScope async(false);
-        EXPECT_FALSE(Advertises(MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false),
+        EXPECT_FALSE(Advertises(MG_Backend::DirectGLES::BuildAdvertisedExtensions(false, false, false, false),
                                 E_GL_KHR_parallel_shader_compile))
             << "MOBILEGL_ASYNC_SHADER_COMPILE=0 must withdraw the extension, not only the threading";
-        EXPECT_FALSE(Advertises(MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false),
+        EXPECT_FALSE(Advertises(MG_Backend::DirectVulkan::BuildAdvertisedExtensions(false, false, false, false),
                                 E_GL_KHR_parallel_shader_compile))
             << "MOBILEGL_ASYNC_SHADER_COMPILE=0 must withdraw the extension, not only the threading";
     }

MobileGL/MG_Test/Program/ProgramTest.cpp

@@ -2692,11 +2692,13 @@ out vec4 fragColor;
 float fma
     (float a, float b, float c) { return a * b + c; }
 float sinh(float x, float y) { return x * y; }
+float length_squared(vec3 value) { return dot(value, value); }
 float round(float x) { return floor(x + 0.5); }
 float min3(float a, float b, float c) { return min(min(a, b), c); }
 
 void main() {
-    fragColor = vec4(fma(0.1, 0.2, 0.3), sinh(0.4, 2.0), round(1.25), min3(0.1, 0.2, 0.3));
+    fragColor = vec4(fma(0.1, 0.2, 0.3), sinh(0.4, 2.0), round(1.25),
+                     min3(0.1, 0.2, 0.3) + length_squared(vec3(0.1, 0.2, 0.3)));
 }
 )";
     GLuint vs = CompileShaderChecked(GL_VERTEX_SHADER, vsSource);
@@ -2707,6 +2709,7 @@ void main() {
         if (essl.find("fragColor") == String::npos) continue; // fragment module only
         EXPECT_NE(essl.find("mg_fma("), String::npos) << essl;
         EXPECT_NE(essl.find("mg_sinh("), String::npos) << essl;
+        EXPECT_NE(essl.find("mg_length_squared("), String::npos) << essl;
         EXPECT_NE(essl.find("mg_round("), String::npos) << essl;
         EXPECT_NE(essl.find("mg_min3("), String::npos) << essl;
         EXPECT_EQ(essl.find("float fma("), String::npos) << essl;

MobileGL/MG_Test/Program/ProgramUtilTest.cpp

@@ -52,20 +52,24 @@ TEST_F(ProgramUtilTest, RenameSamplerFunctionParameterInSpirvPass) {
                OpEntryPoint Fragment %main "main" %outColor
                OpExecutionMode %main OriginUpperLeft
                OpName %globalSampler "sampler"
+               OpName %globalNew "new"
                OpName %paramSampler "sampler"
+               OpName %paramNew "new"
                OpName %main "main"
                OpDecorate %outColor Location 0
        %void = OpTypeVoid
       %float = OpTypeFloat 32
     %v4float = OpTypeVector %float 4
      %mainFn = OpTypeFunction %void
-    %paramFn = OpTypeFunction %void %float
+    %paramFn = OpTypeFunction %void %float %float
   %outV4Ptr = OpTypePointer Output %v4float
  %privatePtr = OpTypePointer Private %float
    %outColor = OpVariable %outV4Ptr Output
 %globalSampler = OpVariable %privatePtr Private
+    %globalNew = OpVariable %privatePtr Private
      %helper = OpFunction %void None %paramFn
 %paramSampler = OpFunctionParameter %float
+    %paramNew = OpFunctionParameter %float
  %helperBody = OpLabel
                OpReturn
                OpFunctionEnd
@@ -91,6 +95,7 @@ TEST_F(ProgramUtilTest, RenameSamplerFunctionParameterInSpirvPass) {
     ASSERT_TRUE(tools.Disassemble(outputBinary, &outputText));
 
     EXPECT_NE(outputText.find("\"MGL_COMPAT_sampler\""), String::npos);
+    EXPECT_NE(outputText.find("\"MGL_COMPAT_new\""), String::npos);
 
     SizeT exactSamplerNameCount = 0;
     SizeT searchOffset = 0;
@@ -99,6 +104,14 @@ TEST_F(ProgramUtilTest, RenameSamplerFunctionParameterInSpirvPass) {
         searchOffset += std::strlen("\"sampler\"");
     }
     EXPECT_EQ(exactSamplerNameCount, 1u);
+
+    SizeT exactNewNameCount = 0;
+    searchOffset = 0;
+    while ((searchOffset = outputText.find("\"new\"", searchOffset)) != String::npos) {
+        ++exactNewNameCount;
+        searchOffset += std::strlen("\"new\"");
+    }
+    EXPECT_EQ(exactNewNameCount, 1u);
 }
 
 TEST_F(ProgramUtilTest, UnformattedFloatStorageImagesKeepIntegerAtomicImagesTyped) {
@@ -2060,153 +2073,6 @@ void main() {
     EXPECT_NE(source.find("layout(std140) uniform Blk"), String::npos);
 }
 
-namespace {
-    String MakeLinearSubgroupPrefixScanShader() {
-        return R"(#version 460 core
-#extension GL_KHR_shader_subgroup_arithmetic : enable
-layout(local_size_x = 1024) in;
-shared float prefixSumCache[64];
-
-layout(std430, binding = 0) writeonly buffer OutputBuffer {
-    float outputValues[];
-};
-
-void main() {
-    float importance = 1.0f;
-    float prefixSum = subgroupInclusiveAdd(importance);
-    if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u) prefixSumCache[gl_SubgroupID] = prefixSum;
-    barrier();
-    uint loopLength = uint(findMSB(gl_NumSubgroups));
-    loopLength += uint(gl_NumSubgroups - (1u << (loopLength - 1u)) > 0u);
-    for (uint i = 0; i < loopLength; i++) {
-        if ((gl_SubgroupID & (1u << i)) > 0u) {
-            prefixSum += prefixSumCache[(gl_SubgroupID >> i << i) - 1u];
-            if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u) prefixSumCache[gl_SubgroupID] = prefixSum;
-        }
-        barrier();
-    }
-    if (gl_LocalInvocationID.x == uint(1024 - 1)) prefixSumCache[0] = prefixSum;
-    barrier();
-    float sum = prefixSumCache[0];
-    float warp = (prefixSum - importance) / sum - float(gl_LocalInvocationID.x + 1u) / float(1024);
-    outputValues[gl_GlobalInvocationID.x] = warp;
-}
-)";
-    }
-} // namespace
-
-TEST_F(ProgramUtilTest, RewriteLinearSubgroupPrefixScanUsesSharedMemoryAndProducesValidSpirv) {
-    using namespace MG_Util::ShaderTranspiler;
-
-    String source = MakeLinearSubgroupPrefixScanShader();
-    ASSERT_TRUE(RewriteLinearSubgroupPrefixScanForVulkan(ShaderStage::Compute, 64, source));
-
-    EXPECT_NE(source.find("shared float prefixSumCache[1024]"), String::npos) << source;
-    EXPECT_NE(source.find("mglVirtualSubgroupInvocation"), String::npos) << source;
-    EXPECT_NE(source.find("for (uint mglPrefixLane"), String::npos) << source;
-    EXPECT_EQ(source.find("subgroupInclusiveAdd"), String::npos) << source;
-    EXPECT_EQ(source.find("gl_Subgroup"), String::npos) << source;
-
-    const String onceRewritten = source;
-    EXPECT_FALSE(RewriteLinearSubgroupPrefixScanForVulkan(ShaderStage::Compute, 64, source));
-    EXPECT_EQ(source, onceRewritten);
-
-    ShaderAttrib shaderAttrib{.shaderType = GL_COMPUTE_SHADER, .sourceStr = source};
-    auto shaderResult = ShaderCompiler::CompileShader(shaderAttrib);
-    ASSERT_TRUE(shaderResult) << shaderResult.error().log << "\nsource:\n" << source;
-
-    ProgramAttrib programAttrib{.shaders = {shaderResult.value()}};
-    auto programResult = ShaderCompiler::LinkProgram(programAttrib);
-    ASSERT_TRUE(programResult) << programResult.error().log;
-
-    ProgramBinaryAttrib binaryAttrib{.shaderTypes = {GL_COMPUTE_SHADER}, .program = *programResult.value()};
-    auto binaryResult = ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
-    ASSERT_TRUE(binaryResult) << binaryResult.error().log;
-    ASSERT_EQ(binaryResult->size(), 1u);
-
-    String validationDiagnostics;
-    spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_1);
-    tools.SetMessageConsumer([&](spv_message_level_t, const char*, const spv_position_t&, const char* message) {
-        validationDiagnostics += message;
-        validationDiagnostics += '\n';
-    });
-    EXPECT_TRUE(tools.Validate(binaryResult->front())) << validationDiagnostics;
-
-    String spirvText;
-    ASSERT_TRUE(tools.Disassemble(binaryResult->front(), &spirvText));
-    EXPECT_EQ(spirvText.find("OpGroupNonUniform"), String::npos) << spirvText;
-}
-
-TEST_F(ProgramUtilTest, RewriteLinearSubgroupPrefixScanRejectsOtherStagesAndSubgroupWidths) {
-    using namespace MG_Util::ShaderTranspiler;
-
-    const String original = MakeLinearSubgroupPrefixScanShader();
-    for (const auto& [stage, subgroupSize] :
-         {std::pair{ShaderStage::Compute, Uint32{32}}, std::pair{ShaderStage::Fragment, Uint32{64}},
-          std::pair{ShaderStage::Compute, Uint32{96}}}) {
-        String source = original;
-        EXPECT_FALSE(RewriteLinearSubgroupPrefixScanForVulkan(stage, subgroupSize, source));
-        EXPECT_EQ(source, original);
-    }
-}
-
-TEST_F(ProgramUtilTest, RewriteLinearSubgroupPrefixScanRejectsPartialOrUnsafeTemplateMatches) {
-    using namespace MG_Util::ShaderTranspiler;
-
-    const auto expectUnchanged = [](String source) {
-        const String original = source;
-        EXPECT_FALSE(RewriteLinearSubgroupPrefixScanForVulkan(ShaderStage::Compute, 64, source));
-        EXPECT_EQ(source, original);
-    };
-
-    String wrongLocalSize = MakeLinearSubgroupPrefixScanShader();
-    wrongLocalSize.replace(wrongLocalSize.find("local_size_x = 1024"), std::strlen("local_size_x = 1024"),
-                           "local_size_x = 512");
-    expectUnchanged(std::move(wrongLocalSize));
-
-    String cacheHasAnotherUse = MakeLinearSubgroupPrefixScanShader();
-    cacheHasAnotherUse.insert(cacheHasAnotherUse.find("float importance"), "prefixSumCache[0] = 0.0f;\n    ");
-    expectUnchanged(std::move(cacheHasAnotherUse));
-
-    String extraSubgroupBuiltin = MakeLinearSubgroupPrefixScanShader();
-    extraSubgroupBuiltin.insert(extraSubgroupBuiltin.find("float importance"),
-                                "uvec4 extraMask = gl_SubgroupEqMask;\n    ");
-    expectUnchanged(std::move(extraSubgroupBuiltin));
-
-    String alteredBarrier = MakeLinearSubgroupPrefixScanShader();
-    alteredBarrier.replace(alteredBarrier.find("barrier();"), std::strlen("barrier();"), "memoryBarrierShared();");
-    expectUnchanged(std::move(alteredBarrier));
-
-    String nestedScan = MakeLinearSubgroupPrefixScanShader();
-    nestedScan.insert(nestedScan.find("float prefixSum ="), "if (importance > 0.0f) {\n    ");
-    const SizeT consumerEnd = nestedScan.find(';', nestedScan.find("float warp ="));
-    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));
-}
-
 // The LEXICAL half must fire at the source level (before the parse) for the
 // preempt-list names - the end-to-end ESSL tests cannot tell which half did the
 // rename, and for these names the parse would fail without the source rewrite.
@@ -2435,9 +2301,6 @@ TEST_F(ProgramUtilTest, CompileEnvFingerprintTracksEveryInput) {
     otherExtensions.advertisedExtensions.push_back(MobileGL::E_GL_ARB_gpu_shader_int64);
     EXPECT_NE(ComputeCompileEnvFingerprint(otherExtensions), baseline);
 
-    CompileEnv otherQuirk = base;
-    otherQuirk.subgroupPrefixScanQuirk = MobileGL::MG_Config::QuirkOverride::ForceOn;
-    EXPECT_NE(ComputeCompileEnvFingerprint(otherQuirk), baseline);
 }
 
 // The no-backend fallback must stay exactly what the pipeline used to do inline:
@@ -2836,16 +2699,6 @@ vec4 helperTint() { return vec4(1.0); }
         return binaryResult->front();
     }
 
-    struct SpirvValidationScope {
-        bool previous;
-        explicit SpirvValidationScope(bool enabled)
-            : previous(MG_Util::ShaderTranspiler::ShaderCompiler::SpirvValidationEnabled()) {
-            MG_Util::ShaderTranspiler::ShaderCompiler::SetSpirvValidationEnabled(enabled);
-        }
-        ~SpirvValidationScope() {
-            MG_Util::ShaderTranspiler::ShaderCompiler::SetSpirvValidationEnabled(previous);
-        }
-    };
 } // namespace
 
 TEST_F(ProgramUtilTest, DeadPrivateChainVertexInputIsEliminatedFromOptimizedBinary) {
@@ -2867,7 +2720,7 @@ TEST_F(ProgramUtilTest, DeadPrivateChainVertexInputIsEliminatedFromOptimizedBina
         << "entry-point-with-calls shape it exists for";
 
     Vector<Uint32> optimized;
-    ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, optimized));
+    ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, optimized, true, true));
 
     const SpirvVariableCensus after = TakeVariableCensus(optimized);
     EXPECT_EQ(after.inputCount, 1u)
@@ -2905,7 +2758,7 @@ void main() {
     ASSERT_GE(before.outputCount, 3u);
 
     Vector<Uint32> optimized;
-    ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, optimized));
+    ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, optimized, true, true));
     EXPECT_EQ(TakeVariableCensus(optimized).outputCount, before.outputCount)
         << "a declared-but-unwritten output was deleted; a fragment stage reading it now "
         << "fails to link (ES) or breaks the Vulkan stage interface";
@@ -2964,17 +2817,15 @@ void main() {
         // succeeds - fail-open call sites downstream must not see a different world),
         // and the failure latch is the signal. This is the catch that took a device
         // bisect to find when the validator was off everywhere.
-        SpirvValidationScope validationOn(true);
         const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
-        EXPECT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, optimized));
+        EXPECT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, optimized, true, true));
         EXPECT_GT(ShaderCompiler::SpirvValidationFailureCount(), failuresBefore)
             << "an invalid optimized module must bump the validation-failure latch";
     }
     {
         // The shipping configuration: same result, no validation, latch untouched.
-        SpirvValidationScope validationOff(false);
         const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
-        EXPECT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, optimized));
+        EXPECT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, optimized, true, false));
         EXPECT_EQ(ShaderCompiler::SpirvValidationFailureCount(), failuresBefore);
     }
 }
@@ -3044,10 +2895,9 @@ void main() {
     ASSERT_FALSE(raw.empty());
     ASSERT_GE(CountRectImageTypes(raw), 1u) << "glslang no longer emits Dim::Rect for sampler2DRect";
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
     Vector<Uint32> optimized;
-    ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, optimized));
+    ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, optimized, true, true));
     EXPECT_EQ(CountRectImageTypes(optimized), 0u);
     EXPECT_EQ(ShaderCompiler::SpirvValidationFailureCount(), failuresBefore)
         << "a rectangle module must leave the chain valid, not latched as a failure";
@@ -3072,10 +2922,9 @@ void main() {
     ASSERT_TRUE(AnyLocationOnUniformStorage(raw))
         << "glslang no longer keeps the explicit uniform location; the strip pass may be obsolete";
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
     Vector<Uint32> optimized;
-    ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, optimized));
+    ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, optimized, true, true));
     EXPECT_FALSE(AnyLocationOnUniformStorage(optimized));
     EXPECT_EQ(ShaderCompiler::SpirvValidationFailureCount(), failuresBefore)
         << "the stripped module must validate clean";
@@ -3172,11 +3021,10 @@ void main() {
         << "the fixture must reproduce the defect before the fix is asked to remove it:\n"
         << DisassembleSpirv(raw);
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
 
     Vector<Uint32> legalized;
-    ASSERT_TRUE(ShaderCompiler::LegalizeFragmentOutputIndexingForEssl(raw, legalized));
+    ASSERT_TRUE(ShaderCompiler::LegalizeFragmentOutputIndexingForEssl(raw, legalized, true));
     ASSERT_FALSE(legalized.empty());
 
     const String disassembly = DisassembleSpirv(legalized);
@@ -3213,11 +3061,10 @@ void main() {
     ASSERT_TRUE(LegalizeFragmentOutputIndexPass::BinaryHasDynamicOutputIndexing(raw))
         << DisassembleSpirv(raw);
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
 
     Vector<Uint32> legalized;
-    ASSERT_TRUE(ShaderCompiler::LegalizeFragmentOutputIndexingForEssl(raw, legalized));
+    ASSERT_TRUE(ShaderCompiler::LegalizeFragmentOutputIndexingForEssl(raw, legalized, true));
     ASSERT_FALSE(legalized.empty());
 
     const String disassembly = DisassembleSpirv(legalized);
@@ -3257,11 +3104,10 @@ void main() {
     ASSERT_FALSE(raw.empty());
     ASSERT_TRUE(LegalizeFragmentOutputIndexPass::BinaryHasDynamicOutputIndexing(raw));
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
 
     Vector<Uint32> legalized;
-    ASSERT_TRUE(ShaderCompiler::LegalizeFragmentOutputIndexingForEssl(raw, legalized));
+    ASSERT_TRUE(ShaderCompiler::LegalizeFragmentOutputIndexingForEssl(raw, legalized, true));
     ASSERT_FALSE(legalized.empty());
 
     const String disassembly = DisassembleSpirv(legalized);
@@ -3300,7 +3146,7 @@ void main() {
     ASSERT_FALSE(LegalizeFragmentOutputIndexPass::BinaryHasDynamicOutputIndexing(raw));
 
     Vector<Uint32> legalized;
-    ASSERT_TRUE(ShaderCompiler::LegalizeFragmentOutputIndexingForEssl(raw, legalized));
+    ASSERT_TRUE(ShaderCompiler::LegalizeFragmentOutputIndexingForEssl(raw, legalized, true));
     EXPECT_EQ(legalized, raw) << "the module must not be rewritten - not even re-serialized - when "
                                  "nothing indexes a fragment output dynamically";
 }
@@ -3550,11 +3396,10 @@ TEST_F(ProgramUtilTest, Lower1DArrayImagesRewritesTheTypeAndWidensTheCoordinate)
     ASSERT_EQ(Count1DArrayStorageImageTypes(spirv), 1u)
         << "the shared chain must leave the 1D-array image for this pass to handle";
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
 
     Vector<Uint32> lowered;
-    ASSERT_TRUE(ShaderCompiler::Lower1DArrayImagesForEssl(spirv, lowered));
+    ASSERT_TRUE(ShaderCompiler::Lower1DArrayImagesForEssl(spirv, lowered, true));
     ASSERT_FALSE(lowered.empty());
 
     EXPECT_EQ(Count1DArrayStorageImageTypes(lowered), 0u)
@@ -3602,11 +3447,10 @@ void main() { ssb.sum = imageLoad(i0, ivec2(2, 3)).r + imageLoad(i1, ivec3(1, 1,
     ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(raw, spirv));
     ASSERT_EQ(Count1DArrayStorageImageTypes(spirv), 1u);
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
 
     Vector<Uint32> lowered;
-    ASSERT_TRUE(ShaderCompiler::Lower1DArrayImagesForEssl(spirv, lowered));
+    ASSERT_TRUE(ShaderCompiler::Lower1DArrayImagesForEssl(spirv, lowered, true));
     ASSERT_FALSE(lowered.empty());
 
     EXPECT_EQ(Count1DArrayStorageImageTypes(lowered), 0u) << DisassembleSpirv(lowered);
@@ -3636,7 +3480,7 @@ void main() { ssb.sum = imageLoad(i0, 2).r; }
     ASSERT_FALSE(spirv.empty());
 
     Vector<Uint32> lowered;
-    ASSERT_TRUE(ShaderCompiler::Lower1DArrayImagesForEssl(spirv, lowered));
+    ASSERT_TRUE(ShaderCompiler::Lower1DArrayImagesForEssl(spirv, lowered, true));
     EXPECT_EQ(lowered, spirv) << "a non-arrayed 1D storage image must pass through byte for byte";
 
     const String essl = DecompileToEssl(lowered);
@@ -3660,7 +3504,7 @@ void main() { fragColor = texture(uTex, vUv); }
     ASSERT_FALSE(spirv.empty());
 
     Vector<Uint32> lowered;
-    ASSERT_TRUE(ShaderCompiler::Lower1DArrayImagesForEssl(spirv, lowered));
+    ASSERT_TRUE(ShaderCompiler::Lower1DArrayImagesForEssl(spirv, lowered, true));
     EXPECT_EQ(lowered, spirv) << "a sampled 1D-array image must pass through byte for byte";
 }
 
@@ -3684,7 +3528,7 @@ void main() { ssb.sum = uint(imageSize(i0).x) + imageLoad(i0, ivec2(0, 0)).r; }
         << "the fixture must contain the shape the pass declines";
 
     Vector<Uint32> lowered;
-    ASSERT_TRUE(ShaderCompiler::Lower1DArrayImagesForEssl(spirv, lowered));
+    ASSERT_TRUE(ShaderCompiler::Lower1DArrayImagesForEssl(spirv, lowered, true));
     EXPECT_EQ(lowered, spirv) << "a declined module must be handed back untouched, not partly rewritten";
     EXPECT_EQ(Count1DArrayStorageImageTypes(lowered), 1u)
         << "declining means the 1D-array type is still there for the driver to reject";
@@ -3735,11 +3579,10 @@ void main() { imageStore(uni_image, ivec2(gl_GlobalInvocationID.xy), uvec4(15u,
     // Precondition: SPIRV-Cross prints no format for it, which is the ESSL the driver refuses.
     EXPECT_EQ(DecompileToEssl(spirv).find("r32ui"), String::npos);
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
 
     Vector<Uint32> baked;
-    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"uni_image", kGlR32ui}}, baked));
+    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"uni_image", kGlR32ui}}, baked, true));
     ASSERT_FALSE(baked.empty());
     EXPECT_FALSE(ShaderCompiler::DeclaresFormatlessStorageImage(baked)) << DisassembleSpirv(baked);
     EXPECT_EQ(ShaderCompiler::SpirvValidationFailureCount(), failuresBefore)
@@ -3800,7 +3643,7 @@ void main() { imageStore(uni_image, ivec2(0), uvec4(1u)); }
 
     Vector<Uint32> baked;
     // Even asked to, with a format of the right component class.
-    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"uni_image", kGlR32ui}}, baked));
+    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"uni_image", kGlR32ui}}, baked, true));
     EXPECT_EQ(baked, spirv) << "a module with nothing format-less must pass through byte for byte";
     EXPECT_NE(DecompileToEssl(baked).find("rgba32ui"), String::npos);
 }
@@ -3822,11 +3665,10 @@ void main() { writeIt(uni_image); }
     ASSERT_FALSE(spirv.empty());
     ASSERT_TRUE(ShaderCompiler::DeclaresFormatlessStorageImage(spirv));
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
 
     Vector<Uint32> baked;
-    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"uni_image", kGlR32ui}}, baked));
+    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"uni_image", kGlR32ui}}, baked, true));
     EXPECT_EQ(baked, spirv) << "a shape the retype cannot follow must leave the module untouched, "
                                "not partly rewritten:\n"
                             << DisassembleSpirv(baked);
@@ -3848,18 +3690,17 @@ void main() { imageStore(uni_image, ivec2(0), vec4(1.0)); }
                                                    GL_COMPUTE_SHADER);
     ASSERT_FALSE(spirv.empty());
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
 
     Vector<Uint32> baked;
-    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"uni_image", kGlR32ui}}, baked));
+    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"uni_image", kGlR32ui}}, baked, true));
     EXPECT_EQ(baked, spirv) << "a declined module must be handed back untouched, not partly rewritten";
     EXPECT_EQ(ShaderCompiler::SpirvValidationFailureCount(), failuresBefore);
 
     // ...and the same image with a float bind format is baked, so the decline above is about the
     // class and not about the pass refusing float images.
     Vector<Uint32> bakedFloat;
-    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"uni_image", kGlR32f}}, bakedFloat));
+    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"uni_image", kGlR32f}}, bakedFloat, true));
     EXPECT_NE(DecompileToEssl(bakedFloat).find("r32f"), String::npos) << DisassembleSpirv(bakedFloat);
 }
 
@@ -3883,12 +3724,11 @@ void main() {
     ASSERT_EQ(CountSpirvOpcode(DisassembleSpirv(spirv), "OpTypeImage"), 1u)
         << "the fixture must have the two images sharing one type:\n" << DisassembleSpirv(spirv);
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
 
     Vector<Uint32> baked;
     ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(
-        spirv, {{"imgA", kGlR32ui}, {"imgB", kGlRgba32ui}}, baked));
+        spirv, {{"imgA", kGlR32ui}, {"imgB", kGlRgba32ui}}, baked, true));
     ASSERT_FALSE(baked.empty());
     EXPECT_EQ(ShaderCompiler::SpirvValidationFailureCount(), failuresBefore)
         << "splitting the shared type must not leave a dangling or duplicate declaration:\n"
@@ -3921,11 +3761,10 @@ void main() {
     ASSERT_EQ(CountSpirvOpcode(DisassembleSpirv(spirv), "OpTypeImage"), 2u)
         << "the fixture needs one Unknown-format and one r32ui image type:\n" << DisassembleSpirv(spirv);
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
 
     Vector<Uint32> baked;
-    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"formatless", kGlR32ui}}, baked));
+    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"formatless", kGlR32ui}}, baked, true));
     ASSERT_FALSE(baked.empty());
     EXPECT_EQ(ShaderCompiler::SpirvValidationFailureCount(), failuresBefore)
         << "the baked image collided with the module's own r32ui image and left a duplicate type:\n"
@@ -3950,11 +3789,10 @@ void main() {
     ASSERT_FALSE(spirv.empty());
     ASSERT_TRUE(ShaderCompiler::DeclaresFormatlessStorageImage(spirv));
 
-    SpirvValidationScope validationOn(true);
     const Uint64 failuresBefore = ShaderCompiler::SpirvValidationFailureCount();
 
     Vector<Uint32> baked;
-    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"imgs", kGlR32ui}}, baked));
+    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"imgs", kGlR32ui}}, baked, true));
     ASSERT_FALSE(baked.empty());
     EXPECT_FALSE(ShaderCompiler::DeclaresFormatlessStorageImage(baked)) << DisassembleSpirv(baked);
     EXPECT_EQ(ShaderCompiler::SpirvValidationFailureCount(), failuresBefore)
@@ -3980,7 +3818,7 @@ void main() { fragColor = texture(uni_sampler, vUv); }
         << "a sampled image must not read as a format-less STORAGE image:\n" << DisassembleSpirv(spirv);
 
     Vector<Uint32> baked;
-    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"uni_sampler", kGlR32ui}}, baked));
+    ASSERT_TRUE(ShaderCompiler::BakeImageFormatsForEssl(spirv, {{"uni_sampler", kGlR32ui}}, baked, true));
     EXPECT_EQ(baked, spirv) << "a sampled image must pass through byte for byte";
 }
 

MobileGL/MG_Test/SanityTest.cpp

@@ -31,6 +31,7 @@
 #include <MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h>
 #include <MG_Util/Math/HalfFloat.h>
 #include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
+#include <MG_Util/ShaderTranspiler/CompileEnv.h>
 #include <MG_Util/ShaderTranspiler/ShaderSourceProcessor.h>
 #include <MG_Util/Debug/Log.h>
 #include <MG_Util/Types.h>
@@ -710,6 +711,37 @@ TEST(DirectVulkanSanity, AdvertisesSubgroupOnlyWhenVulkanReportsUsableSupport) {
     EXPECT_TRUE(backend.GetDynamicParameters().SubgroupQuadOperationsInAllStages);
 }
 
+TEST(DirectVulkanSanity, CapabilityRefreshInvalidatesTheCachedCompileEnvironment) {
+    using namespace MobileGL;
+
+    auto previousContext = Move(MG_State::pGLContext);
+    auto previousBackend = Move(MG_Backend::pActiveBackendObject);
+    MG_State::pGLContext = MakeUnique<MG_State::GLState::GLContext>();
+
+    auto backend = MakeUnique<MG_Backend::DirectVulkan::BackendObject_DirectVulkan>();
+    auto* backendPtr = backend.get();
+    MG_Backend::pActiveBackendObject = Move(backend);
+
+    const auto before = MG_State::pGLContext->GetCompileEnv();
+    EXPECT_EQ(before->params.SubgroupSize, 0u);
+
+    MG_External::VulkanCapabilities caps;
+    caps.SupportsShaderSubgroup = true;
+    caps.SubgroupSize = 8;
+    caps.SubgroupSupportedStages = VK_SHADER_STAGE_COMPUTE_BIT;
+    caps.SubgroupSupportedOperations = VK_SUBGROUP_FEATURE_BASIC_BIT | VK_SUBGROUP_FEATURE_ARITHMETIC_BIT;
+    backendPtr->ApplyVulkanCapabilitiesForTesting(caps);
+
+    const auto after = MG_State::pGLContext->GetCompileEnv();
+    EXPECT_NE(after.get(), before.get());
+    EXPECT_NE(after->fingerprint, before->fingerprint);
+    EXPECT_EQ(after->backend, BackendType::DirectVulkan);
+    EXPECT_EQ(after->params.SubgroupSize, 8u);
+
+    MG_Backend::pActiveBackendObject = Move(previousBackend);
+    MG_State::pGLContext = Move(previousContext);
+}
+
 TEST(DirectVulkanSanity, KeepsOptionalGpuShaderInt64BranchForVoxyQuadDecode) {
     using namespace MobileGL;
 
@@ -936,6 +968,45 @@ TEST(DirectVulkanSanity, ReadbackUsesTheSourceFormatTexelSize) {
     EXPECT_EQ(VulkanRenderer::GetReadbackTexelSize(VK_FORMAT_R32G32B32A32_SFLOAT), 16u);
 }
 
+TEST(DirectVulkanSanity, DefaultFramebufferQuarterTurnReadbackMapsRectAndPixels) {
+    using MobileGL::MG_Backend::DirectVulkan::VulkanRenderer;
+    using MobileGL::Uint8;
+
+    VkOffset2D offset{};
+    VkExtent2D copyExtent{};
+    ASSERT_TRUE(VulkanRenderer::MapDefaultFramebufferReadbackRect(
+        1, 0, 2, 1, VkExtent2D{2, 3}, VK_SURFACE_TRANSFORM_ROTATE_90_BIT_KHR,
+        &offset, &copyExtent));
+    EXPECT_EQ(offset.x, 0);
+    EXPECT_EQ(offset.y, 1);
+    EXPECT_EQ(copyExtent.width, 1u);
+    EXPECT_EQ(copyExtent.height, 2u);
+
+    ASSERT_TRUE(VulkanRenderer::MapDefaultFramebufferReadbackRect(
+        1, 0, 2, 1, VkExtent2D{2, 3}, VK_SURFACE_TRANSFORM_ROTATE_270_BIT_KHR,
+        &offset, &copyExtent));
+    EXPECT_EQ(offset.x, 1);
+    EXPECT_EQ(offset.y, 0);
+    EXPECT_EQ(copyExtent.width, 1u);
+    EXPECT_EQ(copyExtent.height, 2u);
+
+    // Logical GL rows, bottom to top, are abc / def. The display-oriented swapchain blocks are
+    // transposed in opposite directions for 90 and 270 degrees.
+    const Uint8 raw90[] = {'a', 'd', 'b', 'e', 'c', 'f'};
+    const Uint8 raw270[] = {'f', 'c', 'e', 'b', 'd', 'a'};
+    const Uint8 expected[] = {'a', 'b', 'c', 'd', 'e', 'f'};
+    Uint8 result[sizeof(expected)]{};
+
+    ASSERT_TRUE(VulkanRenderer::RemapDefaultFramebufferReadback(
+        raw90, 3, 2, VK_SURFACE_TRANSFORM_ROTATE_90_BIT_KHR, 1, result));
+    EXPECT_TRUE(std::equal(std::begin(expected), std::end(expected), std::begin(result)));
+
+    std::fill(std::begin(result), std::end(result), 0);
+    ASSERT_TRUE(VulkanRenderer::RemapDefaultFramebufferReadback(
+        raw270, 3, 2, VK_SURFACE_TRANSFORM_ROTATE_270_BIT_KHR, 1, result));
+    EXPECT_TRUE(std::equal(std::begin(expected), std::end(expected), std::begin(result)));
+}
+
 TEST(DirectVulkanSanity, ReadbackConvertsRgba8AndRgba16fPixels) {
     using MobileGL::MG_Backend::DirectVulkan::VulkanRenderer;
     using MobileGL::MG_Util::EncodeFloatToHalfBits;

MobileGL/MG_Test/ShaderTranspiler/DemoteFloat64Test.cpp

@@ -154,7 +154,6 @@ class DemoteFloat64Test : public ::testing::Test {
 protected:
     void SetUp() override {
         MobileGL::Initialize();
-        ShaderCompiler::SetSpirvValidationEnabled(true);
         m_validationFailuresAtStart = ShaderCompiler::SpirvValidationFailureCount();
     }
 
@@ -175,7 +174,7 @@ TEST_F(DemoteFloat64Test, DemotesEveryWidthAndDropsTheCapability) {
     ASSERT_TRUE(DeclaresFloat64Capability(input));
 
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output));
+    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output, true));
 
     EXPECT_EQ(CountFloatTypesOfWidth(output, 64), 0u) << Disassemble(output);
     // And exactly one 32-bit float type survives: the merge has to happen, or spirv-val rejects
@@ -210,7 +209,7 @@ void main() {
         << Disassemble(input);
 
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output));
+    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output, true));
 
     // std140 for the demoted members: float at 4, vec2 at 8, vec3 at 16 (aligned like a vec4),
     // vec4 at 32, mat4 at 48 with a 16-byte column stride, the array at 112 with the std140
@@ -241,7 +240,7 @@ void main() {
     EXPECT_EQ(CollectOffsetsOf(input, "Ssbo"), (Vector<Uint32>{0, 32, 64})) << Disassemble(input);
 
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output));
+    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output, true));
 
     // std430, so the array packs at its element size rather than being rounded to 16: float at 0,
     // vec4 at 16, float[4] at 32 with a 4-byte stride. A storage block must NOT come out std140,
@@ -273,7 +272,7 @@ void main() {
     ASSERT_FALSE(before.empty());
 
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output));
+    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output, true));
 
     // Only the block that actually narrowed is re-laid-out. Touching the other one would be
     // churn at best, and a disagreement with glslang's own layout at worst.
@@ -287,7 +286,7 @@ TEST_F(DemoteFloat64Test, FoldsTheConversionsThatBecameIdentities) {
     ASSERT_GT(CountFConverts(input), 0u) << "the fixture no longer converts between the two widths";
 
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output));
+    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output, true));
 
     // SPIR-V requires the two component widths of an OpFConvert to differ, so every one of them
     // has to be gone: both sides are 32 bits now.
@@ -307,7 +306,7 @@ void main() {
     ASSERT_FALSE(input.empty());
 
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output));
+    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output, true));
 
     // A 64-bit literal is two words wide and a 32-bit one is a single word, so a constant left
     // unconverted is not merely imprecise - it is an unparseable instruction. Disassembling both
@@ -326,7 +325,7 @@ void main() { gl_Position = inPos; }
     ASSERT_FALSE(input.empty());
 
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output));
+    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output, true));
     // The pass reports SuccessWithoutChange here, and SPIRV-Tools asserts (in assert-enabled
     // builds) that such a run round-trips byte-identically.
     EXPECT_EQ(output, input);
@@ -351,7 +350,7 @@ void main() {
     ASSERT_EQ(CountFloatTypesOfWidth(input, 64), 1u) << Disassemble(input);
 
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output));
+    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(input, output, true));
     EXPECT_EQ(output, input) << Disassemble(output);
     EXPECT_TRUE(ShaderCompiler::ModuleDeclaresFloat64(output));
 }
@@ -362,7 +361,7 @@ TEST_F(DemoteFloat64Test, ModuleDeclaresFloat64AnswersBothWays) {
     EXPECT_TRUE(ShaderCompiler::ModuleDeclaresFloat64(wide));
 
     Vector<Uint32> demoted;
-    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(wide, demoted));
+    ASSERT_TRUE(ShaderCompiler::DemoteFloat64ToFloat32(wide, demoted, true));
     EXPECT_FALSE(ShaderCompiler::ModuleDeclaresFloat64(demoted));
 
     EXPECT_FALSE(ShaderCompiler::ModuleDeclaresFloat64({}));
@@ -375,7 +374,7 @@ TEST_F(DemoteFloat64Test, TheSharedChainDemotesToo) {
     ASSERT_FALSE(input.empty());
 
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(input, output));
+    ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(input, output, true, true));
     EXPECT_FALSE(ShaderCompiler::ModuleDeclaresFloat64(output)) << Disassemble(output);
 }
 
@@ -459,7 +458,7 @@ TEST_P(DemoteFloat64EsslTest, TheDemotedModuleCanBeEmittedAsEssl) {
     ASSERT_FALSE(input.empty());
 
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(input, output));
+    ASSERT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(input, output, true, true));
 
     SpvcSession session(output, SessionUsageBit::Transpile);
     spvc_compiler_options options;
@@ -482,7 +481,7 @@ TEST_P(DemoteFloat64EsslTest, TheDemotedModuleCanBeEmittedAsEssl) {
 TEST_F(DemoteFloat64Test, RejectsGarbageInput) {
     const Vector<Uint32> notSpirv{0xdeadbeefu, 0u, 0u, 0u, 0u};
     Vector<Uint32> output;
-    EXPECT_FALSE(ShaderCompiler::DemoteFloat64ToFloat32(notSpirv, output));
+    EXPECT_FALSE(ShaderCompiler::DemoteFloat64ToFloat32(notSpirv, output, true));
 }
 
 // EliminateFloatEqualsZeroPass turns a comparison against 0.0 into an epsilon test, a
@@ -502,7 +501,7 @@ namespace {
         EXPECT_FALSE(input.empty());
         if (input.empty()) return false;
         Vector<Uint32> output;
-        EXPECT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(input, output));
+        EXPECT_TRUE(ShaderCompiler::SanitizeAndOptimizeBinary(input, output, true, true));
         return Disassemble(output).find("FAbs") != String::npos;
     }
 

MobileGL/MG_Test/ShaderTranspiler/FlattenXfbInterfaceBlocksTest.cpp

@@ -98,7 +98,6 @@ class FlattenXfbInterfaceBlocksTest : public ::testing::Test {
 protected:
     void SetUp() override {
         MobileGL::Initialize();
-        ShaderCompiler::SetSpirvValidationEnabled(true);
         m_validationFailuresAtStart = ShaderCompiler::SpirvValidationFailureCount();
     }
 
@@ -116,7 +115,7 @@ TEST_F(FlattenXfbInterfaceBlocksTest, FlattensACapturedBlockIntoOneVariablePerMe
 
     std::set<String> flattened;
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::FlattenXfbInterfaceBlocksForEssl(input, {"StageData"}, flattened, output));
+    ASSERT_TRUE(ShaderCompiler::FlattenXfbInterfaceBlocksForEssl(input, {"StageData"}, flattened, output, true));
     ASSERT_FALSE(output.empty());
     EXPECT_EQ(flattened, (std::set<String>{"StageData"}));
 
@@ -145,7 +144,7 @@ TEST_F(FlattenXfbInterfaceBlocksTest, TheEmittedDeclarationIsAPlainArrayNotABloc
 
     std::set<String> flattened;
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::FlattenXfbInterfaceBlocksForEssl(input, {"StageData"}, flattened, output));
+    ASSERT_TRUE(ShaderCompiler::FlattenXfbInterfaceBlocksForEssl(input, {"StageData"}, flattened, output, true));
 
     const String after = Transpile(output);
     EXPECT_NE(after.find("StageData_attrib[16]"), String::npos) << after;
@@ -162,7 +161,7 @@ TEST_F(FlattenXfbInterfaceBlocksTest, GivesEachMemberItsOwnConsecutiveLocations)
 
     std::set<String> flattened;
     Vector<Uint32> output;
-    ASSERT_TRUE(ShaderCompiler::FlattenXfbInterfaceBlocksForEssl(input, {"StageData"}, flattened, output));
+    ASSERT_TRUE(ShaderCompiler::FlattenXfbInterfaceBlocksForEssl(input, {"StageData"}, flattened, output, true));
     ASSERT_FALSE(output.empty());
 
     const String dis = Disassemble(output);
@@ -184,7 +183,7 @@ TEST_F(FlattenXfbInterfaceBlocksTest, LeavesABlockNoCaptureNamesAlone) {
     std::set<String> flattened;
     Vector<Uint32> output;
     ASSERT_TRUE(
-        ShaderCompiler::FlattenXfbInterfaceBlocksForEssl(input, {"SomeOtherBlock"}, flattened, output));
+        ShaderCompiler::FlattenXfbInterfaceBlocksForEssl(input, {"SomeOtherBlock"}, flattened, output, true));
     EXPECT_TRUE(flattened.empty());
 
     const String after = Transpile(output);
@@ -200,7 +199,7 @@ TEST_F(FlattenXfbInterfaceBlocksTest, DeclinesAnEmptyRequestWithoutRewriting) {
 
     std::set<String> flattened;
     Vector<Uint32> output;
-    EXPECT_FALSE(ShaderCompiler::FlattenXfbInterfaceBlocksForEssl(input, {}, flattened, output));
+    EXPECT_FALSE(ShaderCompiler::FlattenXfbInterfaceBlocksForEssl(input, {}, flattened, output, true));
     EXPECT_TRUE(flattened.empty());
     EXPECT_TRUE(output.empty());
 }

MobileGL/MG_Test/State/RenderStateTest.cpp

@@ -6,11 +6,20 @@
 // SPDX-License-Identifier: LGPL-3.0-only
 // End of Source File Header
 //
-// Indexed capability state (glEnablei/glDisablei/glIsEnabledi) exists only for GL_BLEND in this
-// stack. Every other capability must come back as GL_INVALID_ENUM per GL 4.6 sec. 17.3.3 - and,
-// far more importantly, must come back at all: RenderState::SetCapabilityIndexed and
-// IsCapabilityEnabledIndexed used to answer a non-blend capability with THROW_UNIMPL_EXCEPTION,
+// Indexed capability state (glEnablei/glDisablei/glIsEnabledi) exists for exactly two
+// capabilities: GL_BLEND, indexed by draw buffer, and GL_SCISSOR_TEST, indexed by viewport
+// (ARB_viewport_array). Every other capability must come back as GL_INVALID_ENUM per GL 4.6
+// sec. 17.3.3 - and, far more importantly, must come back at all: RenderState::SetCapabilityIndexed
+// and IsCapabilityEnabledIndexed used to answer a non-blend capability with THROW_UNIMPL_EXCEPTION,
 // which unwinds a C++ exception through the C GL ABI and terminates the process.
+//
+// The second half of this file is the ARB_viewport_array indexed rectangle state. Every one of
+// glViewportArrayv/glViewportIndexedf(v)/glScissorArrayv/glScissorIndexed(v)/glDepthRangeArrayv/
+// glDepthRangeIndexed was a MGLOG_W_ONCE stub that raised no error and stored nothing, and the
+// indexed getters answered EVERY index with viewport 0's value, so a set/get round trip silently
+// reported the initial state. The assertions below are deliberately state-shaped rather than
+// render-shaped: this IS the state machine, and the rendering half (gl_ViewportIndex routing) is
+// asserted separately in MG_IntegrationTest/Scenarios/ViewportArrayScenario.cpp.
 
 #include <gtest/gtest.h>
 
@@ -21,6 +30,7 @@
 #include <MG_Impl/GLImpl/RenderState/GL_RenderState.h>
 #include <MG_State/GLState/Core.h>
 #include <MG_State/GLState/FramebufferState/FramebufferObject.h>
+#include <MG_State/GLState/RenderState/RenderState.h>
 
 using namespace MobileGL;
 
@@ -50,10 +60,11 @@ namespace {
     };
 } // namespace
 
-TEST_F(RenderStateTest, IndexedCapabilityTogglesRejectNonBlendCapabilities) {
+TEST_F(RenderStateTest, IndexedCapabilityTogglesRejectNonIndexedCapabilities) {
     // GL_CLIP_DISTANCE0 is a real capability, just not an indexed one - the shape an application or
-    // a CTS negative test would hit.
-    for (const GLenum cap : {GL_CLIP_DISTANCE0, GL_DEPTH_TEST, GL_SCISSOR_TEST}) {
+    // a CTS negative test would hit. GL_SCISSOR_TEST used to be in this list and is not any more:
+    // ARB_viewport_array makes it the second indexed capability (see the tests below).
+    for (const GLenum cap : {GL_CLIP_DISTANCE0, GL_DEPTH_TEST, GL_STENCIL_TEST}) {
         MG_Impl::GLImpl::Enablei(cap, 0);
         ExpectSingleGlError(GL_INVALID_ENUM);
 
@@ -89,3 +100,462 @@ TEST_F(RenderStateTest, IndexedBlendTogglesStillWork) {
     EXPECT_EQ(MG_Impl::GLImpl::IsEnabledi(GL_BLEND, 1), GL_FALSE);
     EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR);
 }
+
+// ---------------------------------------------------------------------------------------------
+// ARB_viewport_array: indexed viewport / scissor / depth-range state
+// ---------------------------------------------------------------------------------------------
+
+namespace {
+    constexpr GLuint kMaxViewports = RenderStateParameters::MAX_VIEWPORTS;
+
+    Array<Array<GLfloat, 4>, kMaxViewports> ReadAllViewports() {
+        Array<Array<GLfloat, 4>, kMaxViewports> out{};
+        for (GLuint i = 0; i < kMaxViewports; ++i) {
+            MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, i, out[i].data());
+        }
+        return out;
+    }
+
+    Array<Array<GLdouble, 2>, kMaxViewports> ReadAllDepthRanges() {
+        Array<Array<GLdouble, 2>, kMaxViewports> out{};
+        for (GLuint i = 0; i < kMaxViewports; ++i) {
+            MG_Impl::GLImpl::GetDoublei_v(GL_DEPTH_RANGE, i, out[i].data());
+        }
+        return out;
+    }
+} // namespace
+
+TEST_F(RenderStateTest, ScissorTestIsIndexedByViewport) {
+    // The exact shape of KHR-GL43.viewport_array.scissor_test_state_api's toggle loop: one index
+    // is flipped and EVERY index is read back, so a broadcast masquerading as an indexed write
+    // cannot pass.
+    MG_Impl::GLImpl::Disable(GL_SCISSOR_TEST);
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    for (GLuint toggled = 0; toggled < kMaxViewports; ++toggled) {
+        MG_Impl::GLImpl::Enablei(GL_SCISSOR_TEST, toggled);
+        EXPECT_EQ(MG_Impl::GLImpl::GetError(), GL_NO_ERROR) << "index " << toggled;
+        for (GLuint i = 0; i < kMaxViewports; ++i) {
+            EXPECT_EQ(MG_Impl::GLImpl::IsEnabledi(GL_SCISSOR_TEST, i), i == toggled ? GL_TRUE : GL_FALSE)
+                << "enabled index " << toggled << ", read index " << i;
+        }
+        MG_Impl::GLImpl::Disablei(GL_SCISSOR_TEST, toggled);
+        EXPECT_EQ(MG_Impl::GLImpl::IsEnabledi(GL_SCISSOR_TEST, toggled), GL_FALSE);
+    }
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, NonIndexedScissorTestEnableWritesEveryViewport) {
+    // GL 4.6 core 17.3.2: Enable/Disable(SCISSOR_TEST) is "for all viewports". Reading only
+    // index 0 back would let a broadcast-less implementation through, so every index is checked.
+    MG_Impl::GLImpl::Enable(GL_SCISSOR_TEST);
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        EXPECT_EQ(MG_Impl::GLImpl::IsEnabledi(GL_SCISSOR_TEST, i), GL_TRUE) << "index " << i;
+    }
+    // ... and the non-indexed query answers for viewport 0 (GL 4.6 core 22.1).
+    EXPECT_EQ(MG_Impl::GLImpl::IsEnabled(GL_SCISSOR_TEST), GL_TRUE);
+
+    MG_Impl::GLImpl::Disable(GL_SCISSOR_TEST);
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        EXPECT_EQ(MG_Impl::GLImpl::IsEnabledi(GL_SCISSOR_TEST, i), GL_FALSE) << "index " << i;
+    }
+    EXPECT_EQ(MG_Impl::GLImpl::IsEnabled(GL_SCISSOR_TEST), GL_FALSE);
+
+    // An indexed enable on a NON-zero index must not move the non-indexed answer.
+    MG_Impl::GLImpl::Enablei(GL_SCISSOR_TEST, 3);
+    EXPECT_EQ(MG_Impl::GLImpl::IsEnabled(GL_SCISSOR_TEST), GL_FALSE);
+    MG_Impl::GLImpl::Enablei(GL_SCISSOR_TEST, 0);
+    EXPECT_EQ(MG_Impl::GLImpl::IsEnabled(GL_SCISSOR_TEST), GL_TRUE);
+
+    MG_Impl::GLImpl::Disable(GL_SCISSOR_TEST);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, ScissorTestEnableRejectsAnOutOfRangeViewportIndex) {
+    MG_Impl::GLImpl::Enablei(GL_SCISSOR_TEST, kMaxViewports);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    MG_Impl::GLImpl::Disablei(GL_SCISSOR_TEST, kMaxViewports);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    EXPECT_EQ(MG_Impl::GLImpl::IsEnabledi(GL_SCISSOR_TEST, kMaxViewports), GL_FALSE);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    // MAX_VIEWPORTS - 1 is the last LEGAL index and must stay silent.
+    MG_Impl::GLImpl::Enablei(GL_SCISSOR_TEST, kMaxViewports - 1);
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::Disablei(GL_SCISSOR_TEST, kMaxViewports - 1);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, MaxViewportsMatchesTheIndexedStateWidth) {
+    // The advertised limit and the width of the state arrays are the same number by
+    // construction; a divergence would make some index simultaneously legal to the CTS and
+    // out of range to the setters.
+    GLint maxViewports = 0;
+    MG_Impl::GLImpl::GetIntegerv(GL_MAX_VIEWPORTS, &maxViewports);
+    ExpectSingleGlError(GL_NO_ERROR);
+    EXPECT_EQ(maxViewports, static_cast<GLint>(kMaxViewports));
+    EXPECT_GE(maxViewports, 16) << "GL 4.3 core requires MAX_VIEWPORTS >= 16";
+}
+
+TEST_F(RenderStateTest, ViewportArrayvRoundTripsThroughEveryGetterWidth) {
+    Array<GLfloat, kMaxViewports * 4> written{};
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        written[i * 4 + 0] = static_cast<GLfloat>(i) + 0.125f;
+        written[i * 4 + 1] = static_cast<GLfloat>(i) + 0.25f;
+        written[i * 4 + 2] = static_cast<GLfloat>(64 + i);
+        written[i * 4 + 3] = static_cast<GLfloat>(32 + i);
+    }
+    MG_Impl::GLImpl::ViewportArrayv(0, kMaxViewports, written.data());
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        GLfloat asFloat[4] = {};
+        MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, i, asFloat);
+        // Bit-exact: the fractional origin is the whole point of float viewport state, and the
+        // CTS compares with == (0.125 and 0.25 are exact binary fractions, so this is fair).
+        EXPECT_EQ(asFloat[0], written[i * 4 + 0]) << "index " << i << " must round-trip verbatim";
+        EXPECT_EQ(asFloat[1], written[i * 4 + 1]) << "index " << i;
+        EXPECT_EQ(asFloat[2], written[i * 4 + 2]) << "index " << i;
+        EXPECT_EQ(asFloat[3], written[i * 4 + 3]) << "index " << i;
+
+        GLdouble asDouble[4] = {};
+        MG_Impl::GLImpl::GetDoublei_v(GL_VIEWPORT, i, asDouble);
+        for (int c = 0; c < 4; ++c) {
+            EXPECT_EQ(asDouble[c], static_cast<GLdouble>(written[i * 4 + c])) << "index " << i << " component " << c;
+        }
+
+        // The integer widths round to nearest rather than truncate; the .5+ case is pinned by
+        // ViewportRoundsRatherThanTruncatesForIntegerQueries below.
+        GLint asInt[4] = {};
+        MG_Impl::GLImpl::GetIntegeri_v(GL_VIEWPORT, i, asInt);
+        EXPECT_EQ(asInt[2], static_cast<GLint>(64 + i)) << "index " << i;
+        EXPECT_EQ(asInt[3], static_cast<GLint>(32 + i)) << "index " << i;
+
+        GLint64 asInt64[4] = {};
+        MG_Impl::GLImpl::GetInteger64i_v(GL_VIEWPORT, i, asInt64);
+        for (int c = 0; c < 4; ++c) {
+            EXPECT_EQ(asInt64[c], static_cast<GLint64>(asInt[c])) << "index " << i << " component " << c;
+        }
+
+        GLboolean asBool[4] = {};
+        MG_Impl::GLImpl::GetBooleani_v(GL_VIEWPORT, i, asBool);
+        EXPECT_EQ(asBool[2], GL_TRUE) << "index " << i << ": a non-zero width is GL_TRUE";
+    }
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, ViewportRoundsRatherThanTruncatesForIntegerQueries) {
+    MG_Impl::GLImpl::ViewportIndexedf(2, 0.0f, 0.0f, 255.875f, 63.5f);
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    GLint asInt[4] = {};
+    MG_Impl::GLImpl::GetIntegeri_v(GL_VIEWPORT, 2, asInt);
+    EXPECT_EQ(asInt[2], 256);
+    EXPECT_EQ(asInt[3], 64);
+
+    GLfloat asFloat[4] = {};
+    MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, 2, asFloat);
+    EXPECT_EQ(asFloat[2], 255.875f) << "the integer query must not disturb the stored float";
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, ViewportIndexedWritesTouchExactlyOneIndex) {
+    MG_Impl::GLImpl::Viewport(0, 0, 8, 8);
+    const auto before = ReadAllViewports();
+
+    for (GLuint target = 0; target < kMaxViewports; ++target) {
+        const GLfloat value[4] = {0.375f, 0.375f, 0.625f, 0.625f};
+        // Alternate the two indexed entry points so both are covered by the isolation claim.
+        if (target % 2 == 0) {
+            MG_Impl::GLImpl::ViewportIndexedf(target, value[0], value[1], value[2], value[3]);
+        } else {
+            MG_Impl::GLImpl::ViewportIndexedfv(target, value);
+        }
+        ExpectSingleGlError(GL_NO_ERROR);
+
+        const auto after = ReadAllViewports();
+        for (GLuint i = 0; i < kMaxViewports; ++i) {
+            if (i == target) {
+                EXPECT_EQ(after[i][0], value[0]) << "index " << i;
+                EXPECT_EQ(after[i][2], value[2]) << "index " << i;
+            } else {
+                EXPECT_EQ(after[i], before[i]) << "write to " << target << " disturbed index " << i;
+            }
+        }
+        MG_Impl::GLImpl::ViewportIndexedf(target, before[target][0], before[target][1], before[target][2],
+                                          before[target][3]);
+    }
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, ClassicViewportWritesEveryIndexAndIsVisibleThroughIndexZero) {
+    // Both directions of the aliasing. ARB_viewport_array defines glViewport as ViewportIndexedf
+    // on every index, and glGetIntegerv(GL_VIEWPORT) as viewport 0.
+    MG_Impl::GLImpl::ViewportIndexedf(5, 1.0f, 2.0f, 3.0f, 4.0f);
+    MG_Impl::GLImpl::Viewport(0, 0, 1, 1);
+    ExpectSingleGlError(GL_NO_ERROR);
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        GLfloat data[4] = {};
+        MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, i, data);
+        EXPECT_EQ(data[0], 0.0f) << "index " << i;
+        EXPECT_EQ(data[2], 1.0f) << "glViewport must overwrite index " << i;
+    }
+
+    MG_Impl::GLImpl::ViewportIndexedf(0, 4.0f, 5.0f, 6.0f, 7.0f);
+    GLint classic[4] = {};
+    MG_Impl::GLImpl::GetIntegerv(GL_VIEWPORT, classic);
+    EXPECT_EQ(classic[0], 4);
+    EXPECT_EQ(classic[2], 6);
+    GLfloat classicFloat[4] = {};
+    MG_Impl::GLImpl::GetFloatv(GL_VIEWPORT, classicFloat);
+    EXPECT_EQ(classicFloat[2], 6.0f);
+    // Index 5 keeps its own value: writing index 0 is not a broadcast.
+    GLfloat other[4] = {};
+    MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, 5, other);
+    EXPECT_EQ(other[2], 1.0f);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, ScissorBoxRoundTripsPerIndexAndAliasesIndexZero) {
+    Array<GLint, kMaxViewports * 4> written{};
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        written[i * 4 + 0] = static_cast<GLint>(i);
+        written[i * 4 + 1] = static_cast<GLint>(i * 2);
+        written[i * 4 + 2] = static_cast<GLint>(16 + i);
+        written[i * 4 + 3] = static_cast<GLint>(8 + i);
+    }
+    MG_Impl::GLImpl::ScissorArrayv(0, kMaxViewports, written.data());
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        GLint readBack[4] = {};
+        MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, i, readBack);
+        for (int c = 0; c < 4; ++c) {
+            EXPECT_EQ(readBack[c], written[i * 4 + c]) << "index " << i << " component " << c;
+        }
+    }
+
+    // Indexed writes stay indexed; both spellings.
+    MG_Impl::GLImpl::ScissorIndexed(4, 4, 4, 8, 8);
+    const GLint indexedV[4] = {9, 9, 12, 12};
+    MG_Impl::GLImpl::ScissorIndexedv(7, indexedV);
+    ExpectSingleGlError(GL_NO_ERROR);
+    GLint probe[4] = {};
+    MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, 4, probe);
+    EXPECT_EQ(probe[2], 8);
+    MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, 7, probe);
+    EXPECT_EQ(probe[2], 12);
+    MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, 5, probe);
+    EXPECT_EQ(probe[2], static_cast<GLint>(16 + 5)) << "index 5 must be untouched";
+
+    // glScissor writes every rectangle, and glGetIntegerv(GL_SCISSOR_BOX) reports rectangle 0.
+    MG_Impl::GLImpl::Scissor(2, 3, 5, 6);
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, i, probe);
+        EXPECT_EQ(probe[0], 2) << "index " << i;
+        EXPECT_EQ(probe[2], 5) << "index " << i;
+    }
+    GLint classic[4] = {};
+    MG_Impl::GLImpl::GetIntegerv(GL_SCISSOR_BOX, classic);
+    EXPECT_EQ(classic[2], 5);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, DepthRangeRoundTripsPerIndexAndAliasesIndexZero) {
+    Array<GLdouble, kMaxViewports * 2> written{};
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        // Exact binary fractions, like the CTS uses: a float-backed store round-trips them.
+        written[i * 2 + 0] = static_cast<GLdouble>(i) / 16.0;
+        written[i * 2 + 1] = 1.0 - static_cast<GLdouble>(i) / 16.0;
+    }
+    MG_Impl::GLImpl::DepthRangeArrayv(0, kMaxViewports, written.data());
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    const auto readBack = ReadAllDepthRanges();
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        EXPECT_EQ(readBack[i][0], written[i * 2 + 0]) << "index " << i;
+        EXPECT_EQ(readBack[i][1], written[i * 2 + 1]) << "index " << i;
+    }
+
+    MG_Impl::GLImpl::DepthRangeIndexed(9, 0.25, 0.75);
+    ExpectSingleGlError(GL_NO_ERROR);
+    GLdouble probe[2] = {};
+    MG_Impl::GLImpl::GetDoublei_v(GL_DEPTH_RANGE, 9, probe);
+    EXPECT_EQ(probe[0], 0.25);
+    EXPECT_EQ(probe[1], 0.75);
+    MG_Impl::GLImpl::GetDoublei_v(GL_DEPTH_RANGE, 8, probe);
+    EXPECT_EQ(probe[0], 8.0 / 16.0) << "index 8 must be untouched";
+
+    GLfloat asFloat[2] = {};
+    MG_Impl::GLImpl::GetFloati_v(GL_DEPTH_RANGE, 9, asFloat);
+    EXPECT_EQ(asFloat[0], 0.25f);
+    EXPECT_EQ(asFloat[1], 0.75f);
+
+    // glDepthRange writes every range; glGetDoublev(GL_DEPTH_RANGE) reports range 0.
+    MG_Impl::GLImpl::DepthRange(0.0, 1.0);
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        MG_Impl::GLImpl::GetDoublei_v(GL_DEPTH_RANGE, i, probe);
+        EXPECT_EQ(probe[0], 0.0) << "index " << i;
+        EXPECT_EQ(probe[1], 1.0) << "index " << i;
+    }
+    MG_Impl::GLImpl::DepthRangeIndexed(0, 0.125, 0.875);
+    GLdouble classic[2] = {};
+    MG_Impl::GLImpl::GetDoublev(GL_DEPTH_RANGE, classic);
+    EXPECT_EQ(classic[0], 0.125);
+    EXPECT_EQ(classic[1], 0.875);
+    MG_Impl::GLImpl::DepthRange(0.0, 1.0);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, IndexedRectangleSettersRejectAnOutOfRangeIndex) {
+    const GLfloat viewport[4] = {0.0f, 0.0f, 1.0f, 1.0f};
+    const GLint scissor[4] = {0, 0, 1, 1};
+
+    for (const GLuint index : {kMaxViewports, kMaxViewports + 1}) {
+        MG_Impl::GLImpl::ViewportIndexedf(index, 0.0f, 0.0f, 1.0f, 1.0f);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        MG_Impl::GLImpl::ViewportIndexedfv(index, viewport);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        MG_Impl::GLImpl::ScissorIndexed(index, 0, 0, 1, 1);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        MG_Impl::GLImpl::ScissorIndexedv(index, scissor);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        MG_Impl::GLImpl::DepthRangeIndexed(index, 0.0, 1.0);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+    }
+
+    // The last legal index must stay silent - api_errors checks both sides of the boundary.
+    MG_Impl::GLImpl::ViewportIndexedf(kMaxViewports - 1, 0.0f, 0.0f, 1.0f, 1.0f);
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::ScissorIndexed(kMaxViewports - 1, 0, 0, 1, 1);
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::DepthRangeIndexed(kMaxViewports - 1, 0.0, 1.0);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, ArraySettersRejectAnOutOfRangeRangeButAcceptAnExactlyFullOne) {
+    Array<GLfloat, kMaxViewports * 4> viewports{};
+    Array<GLint, kMaxViewports * 4> scissors{};
+    Array<GLdouble, kMaxViewports * 2> depths{};
+    for (GLuint i = 0; i < kMaxViewports; ++i) {
+        viewports[i * 4 + 2] = 1.0f;
+        viewports[i * 4 + 3] = 1.0f;
+        scissors[i * 4 + 2] = 1;
+        scissors[i * 4 + 3] = 1;
+        depths[i * 2 + 1] = 1.0;
+    }
+
+    // first == MAX_VIEWPORTS, and first + count > MAX_VIEWPORTS.
+    MG_Impl::GLImpl::ViewportArrayv(kMaxViewports, 1, viewports.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::ViewportArrayv(1, kMaxViewports, viewports.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::ScissorArrayv(kMaxViewports, 1, scissors.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::ScissorArrayv(1, kMaxViewports, scissors.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::DepthRangeArrayv(kMaxViewports, 1, depths.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::DepthRangeArrayv(1, kMaxViewports, depths.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    // first + count == MAX_VIEWPORTS is LEGAL - the off-by-one an ">=" bound would get wrong,
+    // and one KHR-GL43.viewport_array.api_errors asserts explicitly.
+    MG_Impl::GLImpl::ViewportArrayv(1, kMaxViewports - 1, viewports.data());
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::ScissorArrayv(1, kMaxViewports - 1, scissors.data());
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::DepthRangeArrayv(1, kMaxViewports - 1, depths.data());
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    // A negative count is GL_INVALID_VALUE and must not be read as a huge unsigned length.
+    MG_Impl::GLImpl::ViewportArrayv(0, -1, viewports.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::ScissorArrayv(0, -1, scissors.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::DepthRangeArrayv(0, -1, depths.data());
+    ExpectSingleGlError(GL_INVALID_VALUE);
+}
+
+TEST_F(RenderStateTest, NegativeExtentsAreRejectedWithoutDisturbingState) {
+    MG_Impl::GLImpl::Viewport(0, 0, 4, 4);
+    MG_Impl::GLImpl::Scissor(0, 0, 4, 4);
+    ExpectSingleGlError(GL_NO_ERROR);
+
+    MG_Impl::GLImpl::Viewport(0, 0, -1, 1);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::Viewport(0, 0, 1, -1);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::Scissor(0, 0, -1, 1);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::Scissor(0, 0, 1, -1);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    for (GLuint index = 0; index < kMaxViewports; ++index) {
+        MG_Impl::GLImpl::ViewportIndexedf(index, 0.0f, 0.0f, -1.0f, 1.0f);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        MG_Impl::GLImpl::ViewportIndexedf(index, 0.0f, 0.0f, 1.0f, -1.0f);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+
+        const GLfloat badW[4] = {0.0f, 0.0f, -1.0f, 1.0f};
+        MG_Impl::GLImpl::ViewportIndexedfv(index, badW);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+
+        MG_Impl::GLImpl::ScissorIndexed(index, 0, 0, -1, 1);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        const GLint badH[4] = {0, 0, 1, -1};
+        MG_Impl::GLImpl::ScissorIndexedv(index, badH);
+        ExpectSingleGlError(GL_INVALID_VALUE);
+
+        // The array form must reject the WHOLE call for one bad element, exactly once, and
+        // leave every rectangle alone - api_errors submits a full 16-element array with a
+        // single negative extent and then requires the error queue to hold one entry.
+        Array<GLfloat, kMaxViewports * 4> viewports{};
+        Array<GLint, kMaxViewports * 4> scissors{};
+        for (GLuint i = 0; i < kMaxViewports; ++i) {
+            viewports[i * 4 + 2] = 1.0f;
+            viewports[i * 4 + 3] = 1.0f;
+            scissors[i * 4 + 2] = 1;
+            scissors[i * 4 + 3] = 1;
+        }
+        viewports[index * 4 + 2] = -1.0f;
+        scissors[index * 4 + 3] = -1;
+        MG_Impl::GLImpl::ViewportArrayv(0, kMaxViewports, viewports.data());
+        ExpectSingleGlError(GL_INVALID_VALUE);
+        MG_Impl::GLImpl::ScissorArrayv(0, kMaxViewports, scissors.data());
+        ExpectSingleGlError(GL_INVALID_VALUE);
+    }
+
+    // Nothing above may have landed.
+    GLint viewport[4] = {};
+    MG_Impl::GLImpl::GetIntegeri_v(GL_VIEWPORT, 0, viewport);
+    EXPECT_EQ(viewport[2], 4);
+    EXPECT_EQ(viewport[3], 4);
+    GLint scissor[4] = {};
+    MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, 0, scissor);
+    EXPECT_EQ(scissor[2], 4);
+    EXPECT_EQ(scissor[3], 4);
+    ExpectSingleGlError(GL_NO_ERROR);
+}
+
+TEST_F(RenderStateTest, IndexedRectangleQueriesRejectAnOutOfRangeIndex) {
+    GLint ints[4] = {};
+    GLfloat floats[4] = {};
+    GLdouble doubles[4] = {};
+
+    MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, kMaxViewports, ints);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, kMaxViewports, floats);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+    MG_Impl::GLImpl::GetDoublei_v(GL_DEPTH_RANGE, kMaxViewports, doubles);
+    ExpectSingleGlError(GL_INVALID_VALUE);
+
+    MG_Impl::GLImpl::GetIntegeri_v(GL_SCISSOR_BOX, kMaxViewports - 1, ints);
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::GetFloati_v(GL_VIEWPORT, kMaxViewports - 1, floats);
+    ExpectSingleGlError(GL_NO_ERROR);
+    MG_Impl::GLImpl::GetDoublei_v(GL_DEPTH_RANGE, kMaxViewports - 1, doubles);
+    ExpectSingleGlError(GL_NO_ERROR);
+}

MobileGL/MG_Util/BackendLoaders/OpenGL/Loader.cpp

@@ -955,6 +955,8 @@ namespace MobileGL::MG_Util::BackendLoader {
                                  (caps.GLESVersion.Major == 3 && caps.GLESVersion.Minor >= 2);
         const Bool esAtLeast31 = caps.GLESVersion.Major > 3 ||
                                  (caps.GLESVersion.Major == 3 && caps.GLESVersion.Minor >= 1);
+        caps.SupportsDrawIndirect = esAtLeast31 && glesFuncs.glDrawArraysIndirect != nullptr &&
+                                    glesFuncs.glDrawElementsIndirect != nullptr;
         caps.SupportsDrawElementsBaseVertex = (esAtLeast32 || hasDrawElementsBaseVertexExtension) &&
                                               glesFuncs.glDrawElementsBaseVertex != nullptr;
         caps.SupportsComputeShader = esAtLeast31 && glesFuncs.glDispatchCompute != nullptr &&
@@ -976,6 +978,7 @@ namespace MobileGL::MG_Util::BackendLoader {
         MGLOG_I("    indexed glColorMaski: %s", caps.SupportsIndexedColorMask ? "yes" : "no");
         MGLOG_I("    dual-source blend (EXT_blend_func_extended): %s",
                 caps.SupportsDualSourceBlend ? "yes" : "no");
+        MGLOG_I("    draw indirect (ES 3.1 core): %s", caps.SupportsDrawIndirect ? "yes" : "no");
         MGLOG_I("    multi-draw indirect (EXT_multi_draw_indirect): %s",
                 caps.SupportsMultiDrawIndirect ? "yes" : "no");
         MGLOG_I("    multi-draw base vertex (EXT/OES_draw_elements_base_vertex + EXT_multi_draw_arrays): %s",
@@ -1000,7 +1003,11 @@ namespace MobileGL::MG_Util::BackendLoader {
         GLfloat smoothLineWidthRange[2] = {1.0f, 1.0f};
         GLfloat smoothLineWidthGranularity = 1.0f;
         GLfloat aliasedPointSizeRange[2] = {1.0f, 1.0f};
-        GLfloat viewportBoundsRange[2] = {0.0f, 0.0f};
+        // GL 4.6 core table 23.60 sets the MINIMUM VIEWPORT_BOUNDS_RANGE at [-32768, 32767], and
+        // KHR-GL43.viewport_array.queries asserts exactly that floor. GLES has no such query, so
+        // the glGetFloatv below raises GL_INVALID_ENUM and leaves this untouched - starting it at
+        // {0, 0} advertised a range that admits no viewport origin at all.
+        GLfloat viewportBoundsRange[2] = {-32768.0f, 32767.0f};
         GLint maxViewportDims[2] = {16384, 16384};
         GLint viewportSubpixelBits = 0;
         GLint max3DTextureSize = 16384;
@@ -1289,8 +1296,12 @@ namespace MobileGL::MG_Util::BackendLoader {
         caps.MaxViewports = maxViewports;
         caps.MaxViewportWidth = maxViewportDims[0];
         caps.MaxViewportHeight = maxViewportDims[1];
-        caps.ViewportBoundsRangeMin = viewportBoundsRange[0];
-        caps.ViewportBoundsRangeMax = viewportBoundsRange[1];
+        // Only ever WIDER than the core minimum: a driver that answered the query is allowed to
+        // exceed the floor but never to sit inside it, and a driver that rejected the query left
+        // the floor in place. Written as a clamp rather than a plain assignment so a partial
+        // write (one component answered, the other not) cannot narrow the range either.
+        caps.ViewportBoundsRangeMin = std::min(viewportBoundsRange[0], -32768.0f);
+        caps.ViewportBoundsRangeMax = std::max(viewportBoundsRange[1], 32767.0f);
         caps.ViewportSubpixelBits = viewportSubpixelBits;
         caps.MinFragmentInterpolationOffset =
             std::isfinite(minFragmentInterpolationOffset) && minFragmentInterpolationOffset <= -0.5f

MobileGL/MG_Util/BackendLoaders/OpenGL/Loader.h

@@ -1149,6 +1149,10 @@ namespace MobileGL {
             // GLES 3.2 core or GL_OES_shader_multisample_interpolation exposes
             // interpolateAtOffset and the three fragment-offset limit queries.
             Bool SupportsShaderMultisampleInterpolation = false;
+            // ES 3.1+ exposes glDrawArraysIndirect / glDrawElementsIndirect in core. Keep the
+            // version and both entry-point checks together so extension advertisement and the
+            // DirectGLES dispatch path cannot disagree on whether native indirect draws exist.
+            Bool SupportsDrawIndirect = false;
             // GL_EXT_multi_draw_indirect is present AND glMultiDrawArraysIndirectEXT /
             // glMultiDrawElementsIndirectEXT both resolved. Multi-draw is not core in any ES
             // version, and eglGetProcAddress may return a live-looking stub on drivers without

MobileGL/MG_Util/Debug/Log.cpp

@@ -33,13 +33,13 @@ namespace MobileGL {
 
         std::string GetThreadName() {
             char buffer[64] = {0};
-#if defined(_WIN32) && !defined(__MINGW32__)
+#if defined(_WIN32)
             PWSTR desc = nullptr;
             if (SUCCEEDED(GetThreadDescription(GetCurrentThread(), &desc))) {
                 WideCharToMultiByte(CP_UTF8, 0, desc, -1, buffer, sizeof(buffer), nullptr, nullptr);
                 LocalFree(desc);
             }
-#elif defined(__ANDROID__) || defined(__linux__) || defined(__APPLE__) || defined(__MINGW32__)
+#elif defined(__ANDROID__) || defined(__linux__) || defined(__APPLE__)
             pthread_getname_np(pthread_self(), buffer, sizeof(buffer));
 #endif
             return buffer[0] ? buffer : "UnknownThread";

MobileGL/MG_Util/SelfTest/DriverPost.cpp

@@ -1168,7 +1168,9 @@ namespace MobileGL::MG_Util::SelfTest {
             backendApiVersionString = MG_Backend::DirectGLES::FormatBackendAPIVersionString(
                 summary.caps.GLESRendererString, summary.caps.GLESVersion.Major, summary.caps.GLESVersion.Minor);
             advertisedExtensions = JoinAdvertisedExtensions(MG_Backend::DirectGLES::BuildAdvertisedExtensions(
-                summary.caps.SupportsDisjointTimerQuery, summary.caps.SupportsTextureFilterAnisotropy));
+                summary.caps.SupportsDisjointTimerQuery, summary.caps.SupportsTextureFilterAnisotropy,
+                summary.caps.SupportsDrawIndirect,
+                summary.caps.SupportsDrawIndirect && summary.caps.SupportsBaseInstance));
         }
         AppendMobileGLReportedRows(builder, MG_Backend::DirectGLES::GetRendererIdentity(), backendApiVersionString,
                                    advertisedExtensions);
@@ -1461,6 +1463,8 @@ namespace MobileGL::MG_Util::SelfTest {
             Bool shaderSubgroupUsable = false;
             Bool timerQueriesSupported = false;
             Bool samplerAnisotropySupported = false;
+            Bool drawIndirectFirstInstanceSupported = false;
+            Bool shaderDrawParametersSupported = false;
         };
     } // namespace
 
@@ -1744,6 +1748,7 @@ namespace MobileGL::MG_Util::SelfTest {
         VkPhysicalDeviceFeatures features{};
         vkGetPhysicalDeviceFeaturesFn(physicalDevice, &features);
         summary.samplerAnisotropySupported = features.samplerAnisotropy == VK_TRUE;
+        summary.drawIndirectFirstInstanceSupported = features.drawIndirectFirstInstance == VK_TRUE;
         if (features.multiDrawIndirect == VK_TRUE) {
             builder.Pass("multiDrawIndirect", "indirect multi-draw batches run as single native commands");
         } else {
@@ -1910,6 +1915,7 @@ namespace MobileGL::MG_Util::SelfTest {
             builder.Warn("shaderDrawParameters",
                          "unavailable; shaders using gl_DrawID/gl_BaseInstance will not work");
         }
+        summary.shaderDrawParametersSupported = shaderDrawParameters;
 
         Bool provokingVertexLast = false;
         Bool transformFeedbackPreservesProvokingVertex = false;
@@ -2108,7 +2114,8 @@ namespace MobileGL::MG_Util::SelfTest {
             backendApiVersionString = MG_Backend::DirectVulkan::FormatBackendAPIVersionString(
                 summary.deviceName, summary.apiVersionString, summary.driverVersionString);
             advertisedExtensions = JoinAdvertisedExtensions(MG_Backend::DirectVulkan::BuildAdvertisedExtensions(
-                summary.shaderSubgroupUsable, summary.timerQueriesSupported, summary.samplerAnisotropySupported));
+                summary.shaderSubgroupUsable, summary.timerQueriesSupported, summary.samplerAnisotropySupported,
+                summary.drawIndirectFirstInstanceSupported && summary.shaderDrawParametersSupported));
         }
         AppendMobileGLReportedRows(builder, MG_Backend::DirectVulkan::GetRendererIdentity(), backendApiVersionString,
                                    advertisedExtensions);

MobileGL/MG_Util/ShaderTranspiler/CompileEnv.cpp

@@ -38,7 +38,6 @@ namespace MobileGL::MG_Util::ShaderTranspiler {
             HashBytes(state, env.advertisedExtensions.data(),
                       env.advertisedExtensions.size() * sizeof(GLExtension));
         }
-        HashValue(state, env.subgroupPrefixScanQuirk);
         return state;
     }
 
@@ -73,8 +72,6 @@ namespace MobileGL::MG_Util::ShaderTranspiler {
                                      kFrontendMaxComputeWorkGroupInvocations)
                           : kFrontendMaxComputeWorkGroupInvocations;
 
-        env->subgroupPrefixScanQuirk = MG_Config::Features.SubgroupPrefixScanQuirk;
-
         env->fingerprint = ComputeCompileEnvFingerprint(*env);
         return env;
     }
@@ -84,7 +81,6 @@ namespace MobileGL::MG_Util::ShaderTranspiler {
         // computed, and this must not run before MG_Config is loaded.
         static const SharedPtr<const CompileEnv> kDefault = [] {
             auto env = MakeShared<CompileEnv>();
-            env->subgroupPrefixScanQuirk = MG_Config::Features.SubgroupPrefixScanQuirk;
             env->fingerprint = ComputeCompileEnvFingerprint(*env);
             return SharedPtr<const CompileEnv>(Move(env));
         }();

MobileGL/MG_Util/ShaderTranspiler/CompileEnv.h

@@ -12,9 +12,9 @@
 #include <MG_Backend/BackendObject.h>
 
 namespace MobileGL::MG_Util::ShaderTranspiler {
-    // Everything the shader compile/link pipeline reads from OUTSIDE its own (stage, source)
-    // inputs: backend identity, backend limits, the advertised extension list, and the one
-    // config quirk the source rewriter branches on.
+    // everything outside (stage, source) this reads - advertised extensions and backend limits -
+    // so the transformation is a pure function of its three arguments and can run on a worker
+    // thread.
     //
     // Why it exists (P1): every one of those reads is a reach-back into
     // MG_Backend::pActiveBackendObject / gBackendFunctionsTable, and one of them
@@ -46,9 +46,6 @@ namespace MobileGL::MG_Util::ShaderTranspiler {
         MG_Backend::DynamicBackendParameters params{};   // by value, never by reference
         Vector<GLExtension> advertisedExtensions;
 
-        // --- config the source rewriter branches on ---
-        MG_Config::QuirkOverride subgroupPrefixScanQuirk = MG_Config::QuirkOverride::Auto;
-
         Uint64 fingerprint = 0;   // set by CaptureCompileEnv()
 
         Bool HasBackend() const { return backend != BackendType::Unknown; }

MobileGL/MG_Util/ShaderTranspiler/EsslBuiltinFunctionNames.h

@@ -61,7 +61,7 @@ namespace MobileGL {
                 "imageAtomicXor", "imageLoad", "imageSize", "imageStore", "imulExtended",
                 "intBitsToFloat", "interpolateAtCentroid", "interpolateAtOffset",
                 "interpolateAtSample", "inverse", "inversesqrt", "isinf", "isnan",
-                "ldexp", "length", "lessThan", "lessThanEqual", "log", "log2",
+                "ldexp", "length", "length_squared", "lessThan", "lessThanEqual", "log", "log2",
                 "matrixCompMult", "max", "max3", "memoryBarrier",
                 "memoryBarrierAtomicCounter", "memoryBarrierBuffer", "memoryBarrierImage",
                 "memoryBarrierShared", "mid3", "min", "min3", "mix", "mod", "modf",

MobileGL/MG_Util/ShaderTranspiler/ShaderCompiler.cpp

@@ -369,12 +369,6 @@ namespace MobileGL {
                 return allSpirv;
             }
 
-            // -1 unresolved, 0 off, 1 on. Resolved once from MOBILEGL_VALIDATE_SPIRV on first
-            // use. A live getenv rather than an MG_Config::Features field, for the same reason
-            // Config.h already exempts MOBILEGL_LOG_FILE_PATH: suites like SpirvPassTest never
-            // run MobileGL::Initialize(), and every Initialize() re-runs MG_ConfigLoader::Init,
-            // which would clobber a programmatic override stored in the feature table.
-            static std::atomic<int> g_validateSpirv{-1};
             // Total validation failures observed this process. This latch - not the wrappers'
             // return values - is the test-lane signal: validation must never change what a
             // wrapper returns, or the validating lanes would render differently from the
@@ -383,28 +377,6 @@ namespace MobileGL {
             static std::atomic<Uint64> g_spirvValidationFailures{0};
 
             namespace {
-                // Test lanes (desktop/CI/WSL) validate by default; device builds do not -
-                // validation costs real time per module, and on device the driver is the
-                // final validator anyway. MOBILEGL_VALIDATE_SPIRV overrides in either
-                // direction, using the ConfigLoader truthy rule.
-                constexpr bool kValidateSpirvDefault =
-#if defined(__ANDROID__)
-                    false;
-#else
-                    true;
-#endif
-
-                bool IsTruthySpirvEnvValue(const char* value) {
-                    if (value == nullptr || value[0] == '\0') {
-                        return false;
-                    }
-                    String lowered(value);
-                    for (auto& c : lowered) {
-                        c = static_cast<char>(std::tolower(static_cast<unsigned char>(c)));
-                    }
-                    return lowered != "0" && lowered != "false";
-                }
-
                 // spirv-tools' validator lazily constructs function-local static tables on
                 // its first run, which on this codebase happens on a ShaderCompilePool
                 // worker. Function-local statics are destroyed in reverse construction
@@ -445,11 +417,6 @@ namespace MobileGL {
                             tools.Validate(warmup);
                         }
                         std::atexit(+[] {
-                            // Flip validation off first: a validator table this warmup does
-                            // not know about (a future spirv-tools bump) would still be
-                            // destroyed before this handler, and workers must stop entering
-                            // Validate before the drain waits for them.
-                            g_validateSpirv.store(0, std::memory_order_release);
                             Async::ShaderCompilePool::StopAndDrainProcessPoolAtExit();
                         });
                     });
@@ -478,10 +445,10 @@ namespace MobileGL {
                 // Validation is decoupled from control flow on purpose: a failure logs and
                 // bumps the latch, and the caller proceeds exactly as the shipping (non-
                 // validating) configuration would. Tests assert on the latch delta.
-                void ValidateOrLatch(const char* site, const Vector<Uint32>& binary) {
-                    if (!ShaderCompiler::SpirvValidationEnabled()) {
-                        return;
-                    }
+                void ValidateOrLatch(const char* site, const Vector<Uint32>& binary,
+                                     const bool enableSpirvValidation) {
+                    if (!enableSpirvValidation) return;
+                    PinValidatorTablesForProcessExit();
                     spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_1);
                     tools.SetMessageConsumer(MakeSpirvMessageConsumer(site));
                     if (!tools.Validate(binary)) {
@@ -502,39 +469,23 @@ namespace MobileGL {
                 // spirv-tools drops pass diagnostics on the floor.
                 bool RunOptimizerChecked(const char* site, spvtools::Optimizer& optimizer,
                                          const Vector<Uint32>& inputBinary,
-                                         Vector<uint32_t>& outputBinary) {
+                                         Vector<uint32_t>& outputBinary, const bool validateOutput,
+                                         const bool enableSpirvValidation) {
                     spvtools::OptimizerOptions options;
                     options.set_run_validator(false);
                     optimizer.SetMessageConsumer(MakeSpirvMessageConsumer(site));
                     if (!optimizer.Run(inputBinary.data(), inputBinary.size(), &outputBinary, options)) {
                         return false;
                     }
-                    ValidateOrLatch(site, outputBinary);
+                    if (validateOutput) {
+                        ValidateOrLatch(site, outputBinary, enableSpirvValidation);
+                    }
                     return true;
                 }
             } // namespace
 
-            bool ShaderCompiler::SpirvValidationEnabled() {
-                int state = g_validateSpirv.load(std::memory_order_acquire);
-                if (state < 0) {
-                    const char* env = std::getenv("MOBILEGL_VALIDATE_SPIRV");
-                    const bool resolved = env != nullptr ? IsTruthySpirvEnvValue(env) : kValidateSpirvDefault;
-                    int expected = -1;
-                    g_validateSpirv.compare_exchange_strong(expected, resolved ? 1 : 0,
-                                                            std::memory_order_acq_rel);
-                    state = g_validateSpirv.load(std::memory_order_acquire);
-                    if (state == 1) {
-                        PinValidatorTablesForProcessExit();
-                    }
-                }
-                return state == 1;
-            }
-
-            void ShaderCompiler::SetSpirvValidationEnabled(bool enabled) {
-                g_validateSpirv.store(enabled ? 1 : 0, std::memory_order_release);
-                if (enabled) {
-                    PinValidatorTablesForProcessExit();
-                }
+            void ShaderCompiler::PrepareSpirvValidation() {
+                PinValidatorTablesForProcessExit();
             }
 
             Uint64 ShaderCompiler::NoteSpirvValidationFailure() {
@@ -604,16 +555,19 @@ namespace MobileGL {
             }
 
             bool ShaderCompiler::DemoteFloat64ToFloat32(const Vector<Uint32>& inputBinary,
-                                                        Vector<uint32_t>& outputBinary) {
+                                                        Vector<uint32_t>& outputBinary,
+                                                        const bool enableSpirvValidation) {
                 using namespace spvtools;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
                 optimizer.RegisterPass(DemoteFloat64Pass::CreateDemoteFloat64Pass());
 
-                return RunOptimizerChecked("DemoteFloat64ToFloat32", optimizer, inputBinary, outputBinary);
+                return RunOptimizerChecked("DemoteFloat64ToFloat32", optimizer, inputBinary, outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::SanitizeAndOptimizeBinary(const Vector<Uint32>& inputBinary,
-                                                           Vector<uint32_t>& outputBinary) {
+                                                           Vector<uint32_t>& outputBinary,
+                                                           const bool validateOutput,
+                                                           const bool enableSpirvValidation) {
                 using namespace spvtools;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
 
@@ -663,38 +617,41 @@ namespace MobileGL {
                 optimizer.RegisterPass(DemoteFloat64Pass::CreateDemoteFloat64Pass());
 
                 return RunOptimizerChecked("SanitizeAndOptimizeBinary", optimizer, inputBinary,
-                                           outputBinary);
+                                           outputBinary, validateOutput, enableSpirvValidation);
             }
 
             bool ShaderCompiler::LowerDrawParametersForEssl(const Vector<Uint32>& inputBinary,
-                                                            Vector<uint32_t>& outputBinary) {
+                                                            Vector<uint32_t>& outputBinary,
+                                                            const bool enableSpirvValidation) {
                 using namespace spvtools;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
                 optimizer.RegisterPass(LowerDrawParametersPass::CreateLowerDrawParametersPass());
 
                 return RunOptimizerChecked("LowerDrawParametersForEssl", optimizer, inputBinary,
-                                           outputBinary);
+                                           outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::SplitArrayVertexInputsForEssl(const Vector<Uint32>& inputBinary,
-                                                               Vector<uint32_t>& outputBinary) {
+                                                               Vector<uint32_t>& outputBinary,
+                                                               const bool enableSpirvValidation) {
                 using namespace spvtools;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
                 optimizer.RegisterPass(SplitArrayVertexInputsPass::CreateSplitArrayVertexInputsPass());
 
                 return RunOptimizerChecked("SplitArrayVertexInputsForEssl", optimizer, inputBinary,
-                                           outputBinary);
+                                           outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::BakeImageFormatsForEssl(const Vector<Uint32>& inputBinary,
                                                          const UnorderedMap<String, Uint>& glFormatByName,
-                                                         Vector<uint32_t>& outputBinary) {
+                                                         Vector<uint32_t>& outputBinary,
+                                                         const bool enableSpirvValidation) {
                 using namespace spvtools;
                 if (glFormatByName.empty()) return false;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
                 optimizer.RegisterPass(BakeImageFormatsPass::CreateBakeImageFormatsPass(glFormatByName));
 
-                return RunOptimizerChecked("BakeImageFormatsForEssl", optimizer, inputBinary, outputBinary);
+                return RunOptimizerChecked("BakeImageFormatsForEssl", optimizer, inputBinary, outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::DeclaresFormatlessStorageImage(const Vector<Uint32>& binary) {
@@ -718,7 +675,8 @@ namespace MobileGL {
             bool ShaderCompiler::FlattenXfbInterfaceBlocksForEssl(const Vector<Uint32>& inputBinary,
                                                                   const std::set<String>& blockNames,
                                                                   std::set<String>& flattenedBlockNames,
-                                                                  Vector<uint32_t>& outputBinary) {
+                                                                  Vector<uint32_t>& outputBinary,
+                                                                  const bool enableSpirvValidation) {
                 using namespace spvtools;
                 if (blockNames.empty()) return false;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
@@ -726,7 +684,7 @@ namespace MobileGL {
                     blockNames, &flattenedBlockNames));
 
                 return RunOptimizerChecked("FlattenXfbInterfaceBlocksForEssl", optimizer, inputBinary,
-                                           outputBinary);
+                                           outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::RewriteXfbCaptureNameForFlattenedBlock(
@@ -736,48 +694,53 @@ namespace MobileGL {
             }
 
             bool ShaderCompiler::PackDoubleVertexInputsForVulkan(const Vector<Uint32>& inputBinary,
-                                                                 Vector<uint32_t>& outputBinary) {
+                                                                 Vector<uint32_t>& outputBinary,
+                                                                 const bool enableSpirvValidation) {
                 using namespace spvtools;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
                 optimizer.RegisterPass(PackDoubleVertexInputsPass::CreatePackDoubleVertexInputsPass());
 
                 return RunOptimizerChecked("PackDoubleVertexInputsForVulkan", optimizer, inputBinary,
-                                           outputBinary);
+                                           outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::StripUboMemberRelaxedPrecisionForEssl(const Vector<Uint32>& inputBinary,
-                                                                       Vector<uint32_t>& outputBinary) {
+                                                                       Vector<uint32_t>& outputBinary,
+                                                                       const bool enableSpirvValidation) {
                 using namespace spvtools;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
                 optimizer.RegisterPass(
                     StripUboMemberRelaxedPrecisionPass::CreateStripUboMemberRelaxedPrecisionPass());
 
                 return RunOptimizerChecked("StripUboMemberRelaxedPrecisionForEssl", optimizer,
-                                           inputBinary, outputBinary);
+                                           inputBinary, outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::StripNoPerspectiveForEssl(const Vector<Uint32>& inputBinary,
-                                                           Vector<uint32_t>& outputBinary) {
+                                                           Vector<uint32_t>& outputBinary,
+                                                           const bool enableSpirvValidation) {
                 using namespace spvtools;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
                 optimizer.RegisterPass(StripNoPerspectivePass::CreateStripNoPerspectivePass());
 
                 return RunOptimizerChecked("StripNoPerspectiveForEssl", optimizer, inputBinary,
-                                           outputBinary);
+                                           outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::EmulateNoPerspectiveForEssl(const Vector<Uint32>& inputBinary,
-                                                             Vector<uint32_t>& outputBinary) {
+                                                             Vector<uint32_t>& outputBinary,
+                                                             const bool enableSpirvValidation) {
                 using namespace spvtools;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
                 optimizer.RegisterPass(EmulateNoPerspectivePass::CreateEmulateNoPerspectivePass());
 
                 return RunOptimizerChecked("EmulateNoPerspectiveForEssl", optimizer, inputBinary,
-                                           outputBinary);
+                                           outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::LegalizeFragmentOutputIndexingForEssl(const Vector<Uint32>& inputBinary,
-                                                                       Vector<uint32_t>& outputBinary) {
+                                                                       Vector<uint32_t>& outputBinary,
+                                                                       const bool enableSpirvValidation) {
                 using namespace spvtools;
 
                 // Detection gates everything: a module with no dynamically indexed fragment
@@ -810,7 +773,7 @@ namespace MobileGL {
 
                 Vector<uint32_t> folded;
                 if (!RunOptimizerChecked("LegalizeFragmentOutputIndexingForEssl.fold", folder, inputBinary,
-                                         folded) ||
+                                         folded, true, enableSpirvValidation) ||
                     folded.empty()) {
                     // Fail open onto the fallback rather than onto the illegal module.
                     folded = inputBinary;
@@ -829,7 +792,7 @@ namespace MobileGL {
                 lowerer.RegisterPass(CreateAggressiveDCEPass(false));
 
                 if (!RunOptimizerChecked("LegalizeFragmentOutputIndexingForEssl.lower", lowerer, folded,
-                                         outputBinary) ||
+                                         outputBinary, true, enableSpirvValidation) ||
                     outputBinary.empty()) {
                     outputBinary = folded;
                     return true;
@@ -846,16 +809,17 @@ namespace MobileGL {
             }
 
             bool ShaderCompiler::LowerRectImages(const Vector<Uint32>& inputBinary,
-                                                 Vector<uint32_t>& outputBinary) {
+                                                 Vector<uint32_t>& outputBinary,
+                                                 const bool enableSpirvValidation) {
                 using namespace spvtools;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
                 optimizer.RegisterPass(NormalizeRectCoordinatesPass::CreateNormalizeRectCoordinatesPass());
 
-                return RunOptimizerChecked("LowerRectImages", optimizer, inputBinary, outputBinary);
+                return RunOptimizerChecked("LowerRectImages", optimizer, inputBinary, outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::Lower1DArrayImagesForEssl(const Vector<Uint32>& inputBinary,
-                                                            Vector<uint32_t>& outputBinary) {
+                                                            Vector<uint32_t>& outputBinary, const bool enableSpirvValidation) {
                 using namespace spvtools;
 
                 // Declined rather than half-translated: after the rewrite the image is a 2D
@@ -897,40 +861,41 @@ namespace MobileGL {
                 // second Shader. Deduplicating afterwards collapses all three at once.
                 optimizer.RegisterPass(CreateRemoveDuplicatesPass());
 
-                return RunOptimizerChecked("Lower1DArrayImagesForEssl", optimizer, inputBinary, outputBinary);
+                return RunOptimizerChecked("Lower1DArrayImagesForEssl", optimizer, inputBinary, outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::RebaseInstanceIndexForVulkan(const Vector<Uint32>& inputBinary,
-                                                              Vector<uint32_t>& outputBinary) {
+                                                              Vector<uint32_t>& outputBinary, const bool enableSpirvValidation) {
                 using namespace spvtools;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
                 optimizer.RegisterPass(RebaseInstanceIndexPass::CreateRebaseInstanceIndexPass());
 
                 return RunOptimizerChecked("RebaseInstanceIndexForVulkan", optimizer, inputBinary,
-                                           outputBinary);
+                                           outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::ZeroBaseVertexForVulkan(const Vector<Uint32>& inputBinary,
-                                                         Vector<uint32_t>& outputBinary) {
+                                                         Vector<uint32_t>& outputBinary, const bool enableSpirvValidation) {
                 using namespace spvtools;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
                 optimizer.RegisterPass(ZeroBaseVertexPass::CreateZeroBaseVertexPass());
 
-                return RunOptimizerChecked("ZeroBaseVertexForVulkan", optimizer, inputBinary, outputBinary);
+                return RunOptimizerChecked("ZeroBaseVertexForVulkan", optimizer, inputBinary, outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::DecoratePositionInvariantForVulkan(const Vector<Uint32>& inputBinary,
-                                                                    Vector<uint32_t>& outputBinary) {
+                                                                    Vector<uint32_t>& outputBinary, const bool enableSpirvValidation) {
                 using namespace spvtools;
                 Optimizer optimizer(SPV_ENV_VULKAN_1_1);
                 optimizer.RegisterPass(DecoratePositionInvariantPass::CreateDecoratePositionInvariantPass());
 
                 return RunOptimizerChecked("DecoratePositionInvariantForVulkan", optimizer, inputBinary,
-                                           outputBinary);
+                                           outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::UseUnformattedFloatStorageImagesForVulkan(
-                const Vector<Uint32>& inputBinary, Vector<uint32_t>& outputBinary) {
+                const Vector<Uint32>& inputBinary, Vector<uint32_t>& outputBinary,
+                const bool enableSpirvValidation) {
                 constexpr SizeT kSpirvHeaderWordCount = 5;
                 outputBinary.clear();
                 if (inputBinary.size() < kSpirvHeaderWordCount || inputBinary[0] != spv::MagicNumber) {
@@ -1058,7 +1023,8 @@ namespace MobileGL {
                                     addedCapabilities.begin(), addedCapabilities.end());
                 // Hand-rolled word walk, so no Optimizer wrapper ever sees this rewrite;
                 // check the modified module explicitly in validating lanes.
-                ValidateOrLatch("UseUnformattedFloatStorageImagesForVulkan", outputBinary);
+                ValidateOrLatch("UseUnformattedFloatStorageImagesForVulkan", outputBinary,
+                                enableSpirvValidation);
                 return true;
             }
 

MobileGL/MG_Util/ShaderTranspiler/ShaderCompiler.h

@@ -23,19 +23,23 @@ namespace MobileGL {
                 static Result<SharedPtr<glslang::TProgram>> LinkProgram(const ProgramAttrib& attrib);
                 static Result<Vector<Vector<unsigned>>> GetSpirvBinaryFromProgram(const ProgramBinaryAttrib& attrib);
                 static bool SanitizeAndOptimizeBinary(const Vector<Uint32>& inputBinary,
-                                                      Vector<uint32_t>& outputBinary);
+                                                      Vector<uint32_t>& outputBinary,
+                                                      bool validateOutput = true,
+                                                      bool enableSpirvValidation = false);
                 // Demotes DrawIndex/BaseInstance/BaseVertex builtins to plain Private globals
                 // (mg_DrawID/mg_BaseInstance/mg_BaseVertex) so SPIRV-Cross can emit ESSL.
                 // Only for backends without native draw-parameter support (DirectGLES).
                 static bool LowerDrawParametersForEssl(const Vector<Uint32>& inputBinary,
-                                                       Vector<uint32_t>& outputBinary);
+                                                       Vector<uint32_t>& outputBinary,
+                                                       bool enableSpirvValidation = false);
                 // Replaces an ARRAY vertex input with one input per element at consecutive
                 // locations, seeding a Private copy of the array so indexed reads still work.
                 // GLSL ES has no array vertex inputs and SPIRV-Cross refuses the whole module
                 // rather than emulating them, so without this the stage never reaches the
                 // driver. Only for the DirectGLES transpile path.
                 static bool SplitArrayVertexInputsForEssl(const Vector<Uint32>& inputBinary,
-                                                          Vector<uint32_t>& outputBinary);
+                                                          Vector<uint32_t>& outputBinary,
+                                                          bool enableSpirvValidation = false);
                 // Replaces the named interface BLOCKS with one variable per member, named
                 // "<Block>_<member>", shadowing the block itself so the body is untouched. The
                 // Adreno ES driver silently captures NOTHING for a transform-feedback varying
@@ -46,7 +50,8 @@ namespace MobileGL {
                 static bool FlattenXfbInterfaceBlocksForEssl(const Vector<Uint32>& inputBinary,
                                                              const std::set<String>& blockNames,
                                                              std::set<String>& flattenedBlockNames,
-                                                             Vector<uint32_t>& outputBinary);
+                                                             Vector<uint32_t>& outputBinary,
+                                                             bool enableSpirvValidation = false);
                 // The capture request "StageData.attrib[0]" as the pass above renamed it,
                 // "StageData_attrib[0]", or false when it does not name a member of a block
                 // that was flattened.
@@ -58,17 +63,20 @@ namespace MobileGL {
                 // drivers reject cross-stage uniform blocks whose member precisions differ.
                 // Only for the DirectGLES transpile path.
                 static bool StripUboMemberRelaxedPrecisionForEssl(const Vector<Uint32>& inputBinary,
-                                                                  Vector<uint32_t>& outputBinary);
+                                                                  Vector<uint32_t>& outputBinary,
+                                                             bool enableSpirvValidation = false);
                 // 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);
+                                                      Vector<uint32_t>& outputBinary,
+                                                      bool enableSpirvValidation = false);
                 // 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);
+                                                        Vector<uint32_t>& outputBinary,
+                                                      bool enableSpirvValidation = false);
                 // Makes every index into a fragment-output array a constant integral
                 // expression, which is what GLSL ES requires and SPIR-V does not. Runs the
                 // stock folding chain first (loop unrolling folds the loop-derived indices
@@ -79,7 +87,8 @@ namespace MobileGL {
                 // dynamically, which is every shader but a handful.
                 // See LegalizeFragmentOutputIndexPass.
                 static bool LegalizeFragmentOutputIndexingForEssl(const Vector<Uint32>& inputBinary,
-                                                                  Vector<uint32_t>& outputBinary);
+                                                                  Vector<uint32_t>& outputBinary,
+                                                             bool enableSpirvValidation = false);
                 // 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,
@@ -88,7 +97,8 @@ namespace MobileGL {
                 // divides the coordinate of each normalized-coordinate lookup by the texture
                 // size and rewrites the image type to 2D. See NormalizeRectCoordinatesPass for
                 // what it declines and why.
-                static bool LowerRectImages(const Vector<Uint32>& inputBinary, Vector<uint32_t>& outputBinary);
+                static bool LowerRectImages(const Vector<Uint32>& inputBinary, Vector<uint32_t>& outputBinary,
+                                            bool enableSpirvValidation = false);
                 // GL_TEXTURE_1D_ARRAY storage images rewritten to the 2D-array shape the texture
                 // is actually stored in on ES, with the layer moved from the coordinate's second
                 // component to its third. DirectGLES transpile path only - Vulkan binds a real
@@ -96,7 +106,8 @@ namespace MobileGL {
                 // through untouched when the module declares no such image, which is every shader
                 // but a handful. See Lower1DArrayImagesPass for what it declines and why.
                 static bool Lower1DArrayImagesForEssl(const Vector<Uint32>& inputBinary,
-                                                      Vector<uint32_t>& outputBinary);
+                                                      Vector<uint32_t>& outputBinary,
+                                                      bool enableSpirvValidation = false);
                 // Gives each format-less storage image the format bound to its image unit, so
                 // the emitted ESSL can carry the format layout qualifier GLSL ES requires of
                 // every image and desktop GLSL lets a writeonly declaration omit. `glFormatByName`
@@ -105,7 +116,8 @@ namespace MobileGL {
                 // natively. See BakeImageFormatsPass for what it declines and why.
                 static bool BakeImageFormatsForEssl(const Vector<Uint32>& inputBinary,
                                                     const UnorderedMap<String, Uint>& glFormatByName,
-                                                    Vector<uint32_t>& outputBinary);
+                                                    Vector<uint32_t>& outputBinary,
+                                                    bool enableSpirvValidation = false);
                 // Whether the module declares a storage image with no format qualifier at all,
                 // i.e. whether BakeImageFormatsForEssl could change anything. One module parse,
                 // so the ~every shader that declares none pays no optimizer run.
@@ -124,27 +136,31 @@ namespace MobileGL {
                 // emitted text instead.
                 static bool SpirvCrossCanPrintEsslImageFormat(Uint glInternalFormat);
                 static bool RebaseInstanceIndexForVulkan(const Vector<Uint32>& inputBinary,
-                                                         Vector<uint32_t>& outputBinary);
+                                                         Vector<uint32_t>& outputBinary,
+                                                      bool enableSpirvValidation = false);
                 // Builds the non-indexed-draw variant of a vertex shader: every gl_BaseVertex
                 // read becomes zero, which is what GL defines for a command carrying no
                 // baseVertex parameter while Vulkan's builtin would report firstVertex.
                 // See ZeroBaseVertexPass.
                 static bool ZeroBaseVertexForVulkan(const Vector<Uint32>& inputBinary,
-                                                    Vector<uint32_t>& outputBinary);
+                                                    Vector<uint32_t>& outputBinary,
+                                                    bool enableSpirvValidation = false);
                 // Re-declares 64-bit float vertex inputs as their 32-bit unsigned word pair
                 // (double -> uvec2, dvec2 -> uvec4) and bitcasts them back to double at entry, so no
                 // VK_FORMAT_R64*_SFLOAT is needed - lavapipe advertises none of them for vertex
                 // buffers. Vertex stage, DirectVulkan only; pairs with the Float64 case in
                 // VertexInputStateFactory::ToVkVertexFormat.
                 static bool PackDoubleVertexInputsForVulkan(const Vector<Uint32>& inputBinary,
-                                                            Vector<uint32_t>& outputBinary);
+                                                            Vector<uint32_t>& outputBinary,
+                                                      bool enableSpirvValidation = false);
                 // 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);
+                                                               Vector<uint32_t>& outputBinary,
+                                                             bool enableSpirvValidation = false);
                 // 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
@@ -152,13 +168,15 @@ namespace MobileGL {
                 // storage images deliberately keep their declared format for GL-compatible bit
                 // reinterpretation paths (for example, R32F storage accessed as r32ui).
                 static bool UseUnformattedFloatStorageImagesForVulkan(
-                    const Vector<Uint32>& inputBinary, Vector<uint32_t>& outputBinary);
+                    const Vector<Uint32>& inputBinary, Vector<uint32_t>& outputBinary,
+                    bool enableSpirvValidation = false);
                 // Rewrites every 64-bit float in the module to a 32-bit one, preserving every
                 // block offset and stride exactly (see DemoteFloat64Pass). Already part of
                 // SanitizeAndOptimizeBinary, which is where production reaches it; exposed
                 // separately so a test can drive the demotion on its own.
                 static bool DemoteFloat64ToFloat32(const Vector<Uint32>& inputBinary,
-                                                   Vector<uint32_t>& outputBinary);
+                                                   Vector<uint32_t>& outputBinary,
+                                                   bool enableSpirvValidation = false);
                 static Result<String> DecompileShader(SpvcSession& session);
 
                 // Parses one trivial shader in each configuration the production path can
@@ -185,18 +203,16 @@ namespace MobileGL {
                 // no way left to warm it.
                 static void ResetPrewarmLatch();
 
-                // Test-environment SPIR-V validation. When enabled, every Optimizer wrapper
-                // in this file validates its OUTPUT binary - the bytes a driver can actually
-                // receive - and a failure logs the VUID (via MGLOG_I; see the consumer for
-                // why not MGLOG_E) and bumps the failure latch below WITHOUT changing the
-                // wrapper's return value: control flow must stay identical between the
-                // validating and shipping configurations, or fail-open call sites would make
-                // the two render differently. Resolved lazily from MOBILEGL_VALIDATE_SPIRV;
-                // defaults on for desktop/CI/WSL builds and off for device (__ANDROID__)
-                // builds. The setter wins over the environment and is safe to call from test
-                // fixtures at any time.
-                static bool SpirvValidationEnabled();
-                static void SetSpirvValidationEnabled(bool enabled);
+                // Validation is an explicit immutable option of each compiler operation. The
+                // program-link task snapshots MOBILEGL_ENABLE_SPIRV_VALIDATION before it can run
+                // on a worker; standalone callers pass true directly. A failure logs the VUID and
+                // bumps the latch below WITHOUT changing a wrapper's return value, so validating
+                // and shipping configurations preserve identical rendering control flow.
+
+                // Makes validator table lifetime safe before an external final-module validator
+                // runs. This has no configuration state; callers invoke it only for an enabled
+                // task-local validation option.
+                static void PrepareSpirvValidation();
 
                 // The test-lane enforcement signal: total validation failures observed this
                 // process. Tests snapshot it, run the operation under scrutiny, and assert

MobileGL/MG_Util/ShaderTranspiler/ShaderSourceProcessor.cpp

@@ -23,6 +23,7 @@
 namespace {
     using MobileGL::SizeT;
     using MobileGL::String;
+    using MobileGL::Uint32;
     using MobileGL::Vector;
 
     bool IsIdentifierChar(char ch) {
@@ -181,331 +182,6 @@ namespace {
         return std::all_of(token.text.begin() + 1, token.text.end(), IsIdentifierChar);
     }
 
-    class TokenCursor {
-    public:
-        TokenCursor(const Vector<CodeToken>& tokens, SizeT position) : m_tokens(tokens), m_position(position) {}
-
-        bool Consume(const char* expected) {
-            if (m_position >= m_tokens.size() || m_tokens[m_position].text != expected) {
-                return false;
-            }
-            ++m_position;
-            return true;
-        }
-
-        bool ConsumeAnyIdentifier(String& identifier) {
-            if (m_position >= m_tokens.size() || !IsIdentifierToken(m_tokens[m_position])) {
-                return false;
-            }
-            identifier = m_tokens[m_position++].text;
-            return true;
-        }
-
-        bool ConsumeAnyIdentifier() {
-            if (m_position >= m_tokens.size() || !IsIdentifierToken(m_tokens[m_position])) {
-                return false;
-            }
-            ++m_position;
-            return true;
-        }
-
-        bool ConsumeIdentifier(const String& expected) {
-            if (m_position >= m_tokens.size() || !IsIdentifierToken(m_tokens[m_position]) ||
-                m_tokens[m_position].text != expected) {
-                return false;
-            }
-            ++m_position;
-            return true;
-        }
-
-        SizeT Position() const { return m_position; }
-
-    private:
-        const Vector<CodeToken>& m_tokens;
-        SizeT m_position;
-    };
-
-    SizeT CountToken(const Vector<CodeToken>& tokens, const String& tokenText) {
-        return static_cast<SizeT>(std::count_if(tokens.begin(), tokens.end(),
-                                                [&](const CodeToken& token) { return token.text == tokenText; }));
-    }
-
-    bool HasIdentifierWithPrefixOutsideAllowed(const Vector<CodeToken>& tokens, const String& prefix,
-                                               std::initializer_list<const char*> allowedIdentifiers) {
-        return std::any_of(tokens.begin(), tokens.end(), [&](const CodeToken& token) {
-            if (!IsIdentifierToken(token) || !token.text.starts_with(prefix)) {
-                return false;
-            }
-            return std::none_of(allowedIdentifiers.begin(), allowedIdentifiers.end(),
-                                [&](const char* allowed) { return token.text == allowed; });
-        });
-    }
-
-    bool MatchTokenSequence(const Vector<CodeToken>& tokens, SizeT position,
-                            std::initializer_list<const char*> expected) {
-        if (position + expected.size() > tokens.size()) {
-            return false;
-        }
-        for (const char* token : expected) {
-            if (tokens[position++].text != token) {
-                return false;
-            }
-        }
-        return true;
-    }
-
-    struct LinearPrefixScanMatch {
-        SizeT sharedArraySizeBegin = 0;
-        SizeT sharedArraySizeEnd = 0;
-        SizeT scanBegin = 0;
-        SizeT scanEnd = 0;
-        String cache;
-        String importance;
-        String prefixSum;
-        String loopLength;
-        String loopIndex;
-        String sum;
-    };
-
-    bool ParseLinearPrefixScanTemplate(const Vector<CodeToken>& tokens, LinearPrefixScanMatch& match) {
-        // The workaround deliberately recognizes one complete algorithm, not merely the
-        // subgroupInclusiveAdd token. Changing scratch storage is only safe when that storage is
-        // private to this scan and the workgroup has exactly 1024 X invocations.
-        SizeT localSizeDeclarationCount = 0;
-        for (SizeT i = 0; i < tokens.size(); ++i) {
-            if (MatchTokenSequence(tokens, i, {"layout", "(", "local_size_x", "=", "1024", ")", "in", ";"})) {
-                ++localSizeDeclarationCount;
-            }
-        }
-        if (localSizeDeclarationCount != 1) {
-            return false;
-        }
-
-        SizeT sharedDeclarationIndex = String::npos;
-        SizeT sharedDeclarationCount = 0;
-        String cacheName;
-        for (SizeT i = 0; i + 6 < tokens.size(); ++i) {
-            if (tokens[i].text != "shared" || tokens[i + 1].text != "float" || !IsIdentifierToken(tokens[i + 2]) ||
-                tokens[i + 3].text != "[" || tokens[i + 4].text != "64" || tokens[i + 5].text != "]" ||
-                tokens[i + 6].text != ";") {
-                continue;
-            }
-            ++sharedDeclarationCount;
-            sharedDeclarationIndex = i;
-            cacheName = tokens[i + 2].text;
-        }
-        if (sharedDeclarationCount != 1) {
-            return false;
-        }
-
-        SizeT scanTokenIndex = String::npos;
-        SizeT scanCount = 0;
-        for (SizeT i = 0; i + 7 < tokens.size(); ++i) {
-            if (tokens[i].text == "float" && IsIdentifierToken(tokens[i + 1]) && tokens[i + 2].text == "=" &&
-                tokens[i + 3].text == "subgroupInclusiveAdd" && tokens[i + 4].text == "(" &&
-                IsIdentifierToken(tokens[i + 5]) && tokens[i + 6].text == ")" && tokens[i + 7].text == ";") {
-                ++scanCount;
-                scanTokenIndex = i;
-            }
-        }
-        if (scanCount != 1 || sharedDeclarationIndex >= scanTokenIndex) {
-            return false;
-        }
-
-        TokenCursor cursor(tokens, scanTokenIndex);
-        String prefixSum;
-        String importance;
-        String loopLength;
-        String loopIndex;
-        String sum;
-        if (!cursor.Consume("float") || !cursor.ConsumeAnyIdentifier(prefixSum) || !cursor.Consume("=") ||
-            !cursor.Consume("subgroupInclusiveAdd") || !cursor.Consume("(") ||
-            !cursor.ConsumeAnyIdentifier(importance) || !cursor.Consume(")") || !cursor.Consume(";") ||
-            !cursor.Consume("if") || !cursor.Consume("(") || !cursor.Consume("gl_SubgroupInvocationID") ||
-            !cursor.Consume("==") || !cursor.Consume("gl_SubgroupSize") || !cursor.Consume("-") ||
-            !cursor.Consume("1u") || !cursor.Consume(")") || !cursor.ConsumeIdentifier(cacheName) ||
-            !cursor.Consume("[") || !cursor.Consume("gl_SubgroupID") || !cursor.Consume("]") || !cursor.Consume("=") ||
-            !cursor.ConsumeIdentifier(prefixSum) || !cursor.Consume(";") || !cursor.Consume("barrier") ||
-            !cursor.Consume("(") || !cursor.Consume(")") || !cursor.Consume(";") || !cursor.Consume("uint") ||
-            !cursor.ConsumeAnyIdentifier(loopLength) || !cursor.Consume("=") || !cursor.Consume("uint") ||
-            !cursor.Consume("(") || !cursor.Consume("findMSB") || !cursor.Consume("(") ||
-            !cursor.Consume("gl_NumSubgroups") || !cursor.Consume(")") || !cursor.Consume(")") ||
-            !cursor.Consume(";") || !cursor.ConsumeIdentifier(loopLength) || !cursor.Consume("+=") ||
-            !cursor.Consume("uint") || !cursor.Consume("(") || !cursor.Consume("gl_NumSubgroups") ||
-            !cursor.Consume("-") || !cursor.Consume("(") || !cursor.Consume("1u") || !cursor.Consume("<<") ||
-            !cursor.Consume("(") || !cursor.ConsumeIdentifier(loopLength) || !cursor.Consume("-") ||
-            !cursor.Consume("1u") || !cursor.Consume(")") || !cursor.Consume(")") || !cursor.Consume(">") ||
-            !cursor.Consume("0u") || !cursor.Consume(")") || !cursor.Consume(";") || !cursor.Consume("for") ||
-            !cursor.Consume("(") || !cursor.Consume("uint") || !cursor.ConsumeAnyIdentifier(loopIndex) ||
-            !cursor.Consume("=") || !cursor.Consume("0") || !cursor.Consume(";") ||
-            !cursor.ConsumeIdentifier(loopIndex) || !cursor.Consume("<") || !cursor.ConsumeIdentifier(loopLength) ||
-            !cursor.Consume(";") || !cursor.ConsumeIdentifier(loopIndex) || !cursor.Consume("++") ||
-            !cursor.Consume(")") || !cursor.Consume("{") || !cursor.Consume("if") || !cursor.Consume("(") ||
-            !cursor.Consume("(") || !cursor.Consume("gl_SubgroupID") || !cursor.Consume("&") || !cursor.Consume("(") ||
-            !cursor.Consume("1u") || !cursor.Consume("<<") || !cursor.ConsumeIdentifier(loopIndex) ||
-            !cursor.Consume(")") || !cursor.Consume(")") || !cursor.Consume(">") || !cursor.Consume("0u") ||
-            !cursor.Consume(")") || !cursor.Consume("{") || !cursor.ConsumeIdentifier(prefixSum) ||
-            !cursor.Consume("+=") || !cursor.ConsumeIdentifier(cacheName) || !cursor.Consume("[") ||
-            !cursor.Consume("(") || !cursor.Consume("gl_SubgroupID") || !cursor.Consume(">>") ||
-            !cursor.ConsumeIdentifier(loopIndex) || !cursor.Consume("<<") || !cursor.ConsumeIdentifier(loopIndex) ||
-            !cursor.Consume(")") || !cursor.Consume("-") || !cursor.Consume("1u") || !cursor.Consume("]") ||
-            !cursor.Consume(";") || !cursor.Consume("if") || !cursor.Consume("(") ||
-            !cursor.Consume("gl_SubgroupInvocationID") || !cursor.Consume("==") || !cursor.Consume("gl_SubgroupSize") ||
-            !cursor.Consume("-") || !cursor.Consume("1u") || !cursor.Consume(")") ||
-            !cursor.ConsumeIdentifier(cacheName) || !cursor.Consume("[") || !cursor.Consume("gl_SubgroupID") ||
-            !cursor.Consume("]") || !cursor.Consume("=") || !cursor.ConsumeIdentifier(prefixSum) ||
-            !cursor.Consume(";") || !cursor.Consume("}") || !cursor.Consume("barrier") || !cursor.Consume("(") ||
-            !cursor.Consume(")") || !cursor.Consume(";") || !cursor.Consume("}") || !cursor.Consume("if") ||
-            !cursor.Consume("(") || !cursor.Consume("gl_LocalInvocationID") || !cursor.Consume(".") ||
-            !cursor.Consume("x") || !cursor.Consume("==") || !cursor.Consume("uint") || !cursor.Consume("(") ||
-            !cursor.Consume("1024") || !cursor.Consume("-") || !cursor.Consume("1") || !cursor.Consume(")") ||
-            !cursor.Consume(")") || !cursor.ConsumeIdentifier(cacheName) || !cursor.Consume("[") ||
-            !cursor.Consume("0") || !cursor.Consume("]") || !cursor.Consume("=") ||
-            !cursor.ConsumeIdentifier(prefixSum) || !cursor.Consume(";") || !cursor.Consume("barrier") ||
-            !cursor.Consume("(") || !cursor.Consume(")") || !cursor.Consume(";") || !cursor.Consume("float") ||
-            !cursor.ConsumeAnyIdentifier(sum) || !cursor.Consume("=") || !cursor.ConsumeIdentifier(cacheName) ||
-            !cursor.Consume("[") || !cursor.Consume("0") || !cursor.Consume("]") || !cursor.Consume(";")) {
-            return false;
-        }
-        const SizeT scanEndToken = cursor.Position() - 1;
-
-        // Require the scan's immediate consumer as well. This makes the match specific to a
-        // linear distribution warp, and avoids changing unrelated prefix scans which may rely on
-        // the implementation's native subgroup partitioning.
-        if (!cursor.Consume("float") || !cursor.ConsumeAnyIdentifier() || !cursor.Consume("=") ||
-            !cursor.Consume("(") || !cursor.ConsumeIdentifier(prefixSum) || !cursor.Consume("-") ||
-            !cursor.ConsumeIdentifier(importance) || !cursor.Consume(")") || !cursor.Consume("/") ||
-            !cursor.ConsumeIdentifier(sum) || !cursor.Consume("-") || !cursor.Consume("float") ||
-            !cursor.Consume("(") || !cursor.Consume("gl_LocalInvocationID") || !cursor.Consume(".") ||
-            !cursor.Consume("x") || !cursor.Consume("+") || !cursor.Consume("1u") || !cursor.Consume(")") ||
-            !cursor.Consume("/") || !cursor.Consume("float") || !cursor.Consume("(") || !cursor.Consume("1024") ||
-            !cursor.Consume(")") || !cursor.Consume(";")) {
-            return false;
-        }
-
-        // No other use may share the scratch array, and no additional subgroup operation or
-        // builtin may silently retain native-64 semantics after this module becomes virtual-32.
-        if (CountToken(tokens, cacheName) != 6 || CountToken(tokens, "subgroupInclusiveAdd") != 1 ||
-            CountToken(tokens, "gl_SubgroupInvocationID") != 2 || CountToken(tokens, "gl_SubgroupSize") != 2 ||
-            CountToken(tokens, "gl_SubgroupID") != 4 || CountToken(tokens, "gl_NumSubgroups") != 2 ||
-            CountToken(tokens, "gl_LocalInvocationID") != 2 || CountToken(tokens, "barrier") != 3 ||
-            CountToken(tokens, "findMSB") != 1 ||
-            HasIdentifierWithPrefixOutsideAllowed(tokens, "subgroup", {"subgroupInclusiveAdd"}) ||
-            HasIdentifierWithPrefixOutsideAllowed(
-                tokens, "gl_Subgroup",
-                {"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;
-        }
-
-        // The scan must be at the top level of the sole main() body. Its existing barriers already
-        // require uniform control flow; this check prevents us from introducing extra barriers in
-        // a nested branch or loop.
-        SizeT mainOpenBrace = String::npos;
-        SizeT mainCloseBrace = String::npos;
-        SizeT mainCount = 0;
-        for (SizeT i = 0; i + 4 < tokens.size(); ++i) {
-            if (!MatchTokenSequence(tokens, i, {"void", "main", "(", ")", "{"})) {
-                continue;
-            }
-            ++mainCount;
-            mainOpenBrace = i + 4;
-            int depth = 1;
-            for (SizeT j = mainOpenBrace + 1; j < tokens.size(); ++j) {
-                if (tokens[j].text == "{")
-                    ++depth;
-                else if (tokens[j].text == "}" && --depth == 0) {
-                    mainCloseBrace = j;
-                    break;
-                }
-            }
-        }
-        if (mainCount != 1 || mainCloseBrace == String::npos || scanTokenIndex <= mainOpenBrace ||
-            scanEndToken >= mainCloseBrace) {
-            return false;
-        }
-        int depthAtScan = 1;
-        for (SizeT i = mainOpenBrace + 1; i < scanTokenIndex; ++i) {
-            if (tokens[i].text == "{")
-                ++depthAtScan;
-            else if (tokens[i].text == "}")
-                --depthAtScan;
-        }
-        if (depthAtScan != 1) {
-            return false;
-        }
-
-        constexpr const char* injectedNames[] = {"mglPrefixScanLane",  "mglVirtualSubgroupInvocation",
-                                                 "mglVirtualSubgroup", "mglVirtualSubgroupBase",
-                                                 "mglPrefixLane",      "mglVirtualSubgroupCount"};
-        for (const char* injectedName : injectedNames) {
-            if (CountToken(tokens, injectedName) != 0) {
-                return false;
-            }
-        }
-
-        match.sharedArraySizeBegin = tokens[sharedDeclarationIndex + 4].begin;
-        match.sharedArraySizeEnd = tokens[sharedDeclarationIndex + 4].end;
-        match.scanBegin = tokens[scanTokenIndex].begin;
-        match.scanEnd = tokens[scanEndToken].end;
-        match.cache = std::move(cacheName);
-        match.importance = std::move(importance);
-        match.prefixSum = std::move(prefixSum);
-        match.loopLength = std::move(loopLength);
-        match.loopIndex = std::move(loopIndex);
-        match.sum = std::move(sum);
-        return true;
-    }
-
-    String BuildLinearPrefixScanReplacement(const LinearPrefixScanMatch& match) {
-        String replacement;
-        replacement.reserve(1800);
-        replacement += "uint mglPrefixScanLane = gl_LocalInvocationID.x;\n";
-        replacement += "uint mglVirtualSubgroupInvocation = mglPrefixScanLane & 31u;\n";
-        replacement += "uint mglVirtualSubgroup = mglPrefixScanLane >> 5u;\n";
-        replacement += "const uint mglVirtualSubgroupCount = 32u;\n";
-        replacement += match.cache + "[mglPrefixScanLane] = " + match.importance + ";\n";
-        replacement += "barrier();\n";
-        replacement += "float " + match.prefixSum + " = 0.0f;\n";
-        replacement += "uint mglVirtualSubgroupBase = mglVirtualSubgroup << 5u;\n";
-        replacement += "for (uint mglPrefixLane = mglVirtualSubgroupBase; "
-                       "mglPrefixLane <= mglPrefixScanLane; ++mglPrefixLane) {\n";
-        replacement += match.prefixSum + " += " + match.cache + "[mglPrefixLane];\n";
-        replacement += "}\n";
-        replacement += "barrier();\n";
-        replacement += "if (mglVirtualSubgroupInvocation == 31u) " + match.cache +
-                       "[mglVirtualSubgroup] = " + match.prefixSum + ";\n";
-        replacement += "barrier();\n";
-        replacement += "uint " + match.loopLength + " = uint(findMSB(mglVirtualSubgroupCount));\n";
-        replacement +=
-            match.loopLength + " += uint(mglVirtualSubgroupCount - (1u << (" + match.loopLength + " - 1u)) > 0u);\n";
-        replacement += "for (uint " + match.loopIndex + " = 0u; " + match.loopIndex + " < " + match.loopLength +
-                       "; ++" + match.loopIndex + ") {\n";
-        replacement += "if ((mglVirtualSubgroup & (1u << " + match.loopIndex + ")) > 0u) {\n";
-        replacement += match.prefixSum + " += " + match.cache + "[(mglVirtualSubgroup >> " + match.loopIndex + " << " +
-                       match.loopIndex + ") - 1u];\n";
-        replacement += "if (mglVirtualSubgroupInvocation == 31u) " + match.cache +
-                       "[mglVirtualSubgroup] = " + match.prefixSum + ";\n";
-        replacement += "}\nbarrier();\n}\n";
-        replacement += "if (mglPrefixScanLane == 1023u) " + match.cache + "[0] = " + match.prefixSum + ";\n";
-        replacement += "barrier();\n";
-        replacement += "float " + match.sum + " = " + match.cache + "[0];";
-        return replacement;
-    }
-
     void SkipDirectiveWhitespace(const MobileGL::String& source, SizeT& pos, SizeT lineEnd) {
         while (pos < lineEnd && std::isspace(static_cast<unsigned char>(source[pos]))) {
             pos++;
@@ -1253,117 +929,6 @@ namespace {
 namespace MobileGL {
     namespace MG_Util {
         namespace ShaderTranspiler {
-            Bool RewriteLinearSubgroupPrefixScanForVulkan(ShaderStage stage, Uint32 nativeSubgroupSize,
-                                                          String& source) {
-                constexpr Uint32 capturedSubgroupSize = 32;
-                if (stage != ShaderStage::Compute || nativeSubgroupSize <= capturedSubgroupSize ||
-                    nativeSubgroupSize % capturedSubgroupSize != 0) {
-                    return false;
-                }
-
-                // Vulkan subgroup widths are powers of two. Keep the workaround restricted to
-                // wider widths which are a power-of-two multiple of the captured 32-lane model.
-                const Uint32 subgroupScale = nativeSubgroupSize / capturedSubgroupSize;
-                if ((subgroupScale & (subgroupScale - 1u)) != 0u) {
-                    return false;
-                }
-
-                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_ONCE("%s: subgroupInclusiveAdd present but the linear prefix-scan template "
-                                "did not match; the wide-subgroup rewrite was NOT applied",
-                                __func__);
-                    }
-                    return false;
-                }
-
-                const String replacement = BuildLinearPrefixScanReplacement(match);
-                source.replace(match.scanBegin, match.scanEnd - match.scanBegin, replacement);
-                // The declaration occurs before the replaced scan, so its original offsets remain
-                // valid after the first replacement.
-                source.replace(match.sharedArraySizeBegin, match.sharedArraySizeEnd - match.sharedArraySizeBegin,
-                               "1024");
-                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;
-                    // Reads the override out of the captured env, never out of the live
-                    // MG_Config table: a worker must see the same config the GL thread saw.
-                    MG_Config::QuirkOverride (*GetOverride)(const CompileEnv&);
-                    Bool (*DeviceApplies)(const ShaderSourceQuirkContext&);
-                    Bool (*Apply)(const ShaderSourceQuirkContext&, String&);
-                };
-
-                constexpr ShaderSourceQuirk kShaderSourceQuirks[] = {
-                    {
-                        // MOBILEGL_QUIRK_SUBGROUP_PREFIX_SCAN
-                        "subgroup-prefix-scan-rewrite",
-                        [](const CompileEnv& env) { return env.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(const CompileEnv& env, ShaderStage stage, String& source) {
-                    // No backend at capture time means no device to match a quirk against,
-                    // and (as before) no quirk can fire - not even a forced one, because
-                    // every Apply reads device parameters that do not exist yet.
-                    if (!env.HasBackend()) {
-                        return;
-                    }
-                    const ShaderSourceQuirkContext quirkContext{
-                        stage,
-                        env.backend,
-                        env.params.GpuVendor,
-                        env.params.SubgroupSize,
-                    };
-                    for (const ShaderSourceQuirk& quirk : kShaderSourceQuirks) {
-                        const MG_Config::QuirkOverride quirkOverride = quirk.GetOverride(env);
-                        if (quirkOverride == MG_Config::QuirkOverride::ForceOff) {
-                            continue;
-                        }
-                        if (quirkOverride == MG_Config::QuirkOverride::Auto &&
-                            !quirk.DeviceApplies(quirkContext)) {
-                            continue;
-                        }
-                        if (quirk.Apply(quirkContext, source)) {
-                            MGLOG_D("ApplyShaderSourceQuirks: applied '%s'%s", quirk.name,
-                                    quirkOverride == MG_Config::QuirkOverride::ForceOn ? " (forced on)" : "");
-                        }
-                    }
-                }
-            } // namespace
-
             void PreprocessShaderSource(ShaderStage stage, String& source) {
                 PreprocessShaderSource(stage, source, *GetCurrentCompileEnv());
             }
@@ -1403,7 +968,6 @@ namespace MobileGL {
                 ModernizeLegacyGLSL(stage, source, afterVersion);
                 InjectDepthRangeBuiltinShim(stage, source, afterVersion);
 
-                ApplyShaderSourceQuirks(env, stage, source);
             }
 
             Bool RetargetLegacyVersionDirectiveTo460(String& source) {

MobileGL/MG_Util/ShaderTranspiler/ShaderSourceProcessor.h

@@ -30,18 +30,6 @@ namespace MobileGL {
             // tests and diagnostics that drive the preprocessor standalone.
             void PreprocessShaderSource(ShaderStage stage, String& source);
 
-            // Some desktop-captured compute shaders build a workgroup-wide linear prefix scan
-            // from subgroupInclusiveAdd plus a shared array of subgroup totals. Qualcomm's
-            // Vulkan driver miscompiles that exact float InclusiveScan path for native subgroups
-            // 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. 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
             // from a legacy desktop version back up to "#version 460 core". Returns false (leaving
             // the source untouched) for anything else: ES, compatibility, or an already-modern

MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/RenameSamplerFunctionParameterPass.cpp

@@ -19,23 +19,27 @@ namespace MobileGL {
     namespace MG_Util {
         namespace ShaderTranspiler {
             namespace {
-                constexpr const char* kConflictingName = "sampler";
-                constexpr const char* kCompatName = "MGL_COMPAT_sampler";
+                const char* GetCompatName(StringView name) {
+                    if (name == "sampler") return "MGL_COMPAT_sampler";
+                    if (name == "new") return "MGL_COMPAT_new";
+                    return nullptr;
+                }
 
-                Bool IsNamedSamplerFunctionParameter(spvtools::opt::IRContext* context,
-                                                     spvtools::opt::Instruction& nameInst) {
+                const char* GetConflictingFunctionParameterCompatName(spvtools::opt::IRContext* context,
+                                                                       spvtools::opt::Instruction& nameInst) {
                     if (nameInst.opcode() != spv::Op::OpName || nameInst.NumInOperands() < 2) {
-                        return false;
+                        return nullptr;
                     }
 
-                    if (nameInst.GetInOperand(1).AsString() != kConflictingName) {
-                        return false;
+                    const char* compatName = GetCompatName(nameInst.GetInOperand(1).AsString());
+                    if (compatName == nullptr) {
+                        return nullptr;
                     }
 
                     auto* defUseMgr = context->get_def_use_mgr();
                     const Uint32 targetId = nameInst.GetSingleWordInOperand(0);
                     const auto* target = defUseMgr->GetDef(targetId);
-                    return target != nullptr && target->opcode() == spv::Op::OpFunctionParameter;
+                    return target != nullptr && target->opcode() == spv::Op::OpFunctionParameter ? compatName : nullptr;
                 }
             } // namespace
 
@@ -44,12 +48,13 @@ namespace MobileGL {
                 auto* irContext = context();
 
                 for (auto& debugInst : irContext->debugs2()) {
-                    if (!IsNamedSamplerFunctionParameter(irContext, debugInst)) {
+                    const char* compatName = GetConflictingFunctionParameterCompatName(irContext, debugInst);
+                    if (compatName == nullptr) {
                         continue;
                     }
 
                     debugInst.SetInOperand(
-                        1, spvtools::utils::MakeVector<spvtools::opt::Operand::OperandData>(kCompatName));
+                        1, spvtools::utils::MakeVector<spvtools::opt::Operand::OperandData>(compatName));
                     modified = true;
                 }
 

MobileGL/MG_Util/Texture/PixelStoreProcessor.cpp

@@ -423,6 +423,26 @@ namespace MobileGL::MG_Util::PixelStoreProcessor {
             InternalPackedLayout internalPacked;
         };
 
+        Bool IsValidUnpackPixelPair(TextureInputFormat format, TexturePixelDataType type) {
+            UnpackChannelMapping mapping{};
+            if (!GetUnpackChannelMapping(format, mapping)) return false;
+
+            PackedTypeLayout packed{};
+            if (GetPackedTypeLayout(type, packed)) {
+                return packed.fieldCount == mapping.channelCount;
+            }
+
+            switch (type) {
+            case TexturePixelDataType::UnsignedInt5999Rev:
+            case TexturePixelDataType::UnsignedInt101111Rev:
+                return !mapping.isInteger && mapping.channelCount == 3;
+            default: {
+                ShadowComponent component{};
+                return GetDirectShadowComponentForType(type, mapping.isInteger, component);
+            }
+            }
+        }
+
         // Returns true when the (format, type) -> internal-format upload needs a per-texel conversion;
         // returns false both for layouts that already match the shadow bytes (memcpy fast path) and for
         // combinations the converter does not support (legacy copy behavior).
@@ -964,6 +984,32 @@ namespace MobileGL::MG_Util::PixelStoreProcessor {
         return outputPixels;
     }
 
+    Bool ConvertOnePixelToInternal(TextureInternalFormat targetInternalFormat,
+                                   TextureInputFormat textureInputFormat,
+                                   TexturePixelDataType inputDataType,
+                                   const void* inputPixel,
+                                   Vector<Uint8>& outputPixel) {
+        outputPixel.clear();
+        if (inputPixel == nullptr || !IsValidUnpackPixelPair(textureInputFormat, inputDataType)) return false;
+
+        PixelStoreParameters params{};
+        params.Alignment = 1;
+        SizeT convertedSize = 0;
+        void* converted = ProcessTexturePixelsDataUnpack(
+            inputPixel, params, targetInternalFormat, textureInputFormat, inputDataType, {1, 1, 1}, false,
+            convertedSize);
+        const SizeT expectedSize = MG_Util::GetSizedInternalFormatSizeInBytes(targetInternalFormat);
+        if (converted == nullptr || convertedSize != expectedSize || expectedSize == 0) {
+            if (converted != nullptr) free(converted);
+            return false;
+        }
+
+        outputPixel.resize(convertedSize);
+        Memcpy(outputPixel.data(), converted, convertedSize);
+        free(converted);
+        return true;
+    }
+
     void* ProcessTexturePixelsDataPack(const void* inputPixels, const PixelStoreParameters& params,
                                        TextureInternalFormat srcInternalFormat, TexturePixelDataType srcDataType,
                                        TextureInputFormat dstInputFormat, TexturePixelDataType dstDataType,

MobileGL/MG_Util/Texture/PixelStoreProcessor.h

@@ -21,6 +21,12 @@ namespace MobileGL::MG_Util::PixelStoreProcessor {
                                        TextureInternalFormat srcInternalFormat, TexturePixelDataType srcDataType,
                                        TextureInputFormat dstInputFormat, TexturePixelDataType dstDataType,
                                        IntVec3 dimension, Bool isBitmap, SizeT& outSize);
+    Bool ConvertOnePixelToInternal(TextureInternalFormat targetInternalFormat,
+                                   TextureInputFormat textureInputFormat,
+                                   TexturePixelDataType inputDataType,
+                                   const void* inputPixel,
+                                   Vector<Uint8>& outputPixel);
+
     void ProcessColorSwizzle(void* data, SizeT pixelCount, const Vector<TextureSwizzleParam>& swizzle);
 
     // True when a packed internal format's 32-bit storage word IS the client (format, type) word,

android-plugin/app/build.gradle.kts

@@ -86,7 +86,7 @@ val pluginRendererConfig = buildJsonValue {
             selectable(
                 key = "MOBILEGL_BACKEND_TYPE",
                 title = RendererConfig.MetaString("mobilegl_backend_type_title"),
-                items = RendererConfig.EnvItems("DirectGLES", listOf("DirectVulkan")),
+                items = RendererConfig.EnvItems("DirectGLES", listOf("DirectVulkan", "DiligentVulkan")),
             )
             toggleable("MOBILEGL_DISABLE_TIMERQUERY", "1", false, RendererConfig.MetaString("mobilegl_disable_timerquery_title"))
             toggleable("MOBILEGL_DISABLE_SUBGROUP", "1", false, RendererConfig.MetaString("mobilegl_disable_subgroup_title"))

include/KHR/khrplatform.h

@@ -221,10 +221,13 @@ typedef signed   char          khronos_int8_t;
 typedef unsigned char          khronos_uint8_t;
 typedef signed   short int     khronos_int16_t;
 typedef unsigned short int     khronos_uint16_t;
-typedef signed   long  int     khronos_intptr_t;
-typedef unsigned long  int     khronos_uintptr_t;
-typedef signed   long  int     khronos_ssize_t;
-typedef unsigned long  int     khronos_usize_t;
+/* `long` is 32-bit on LLP64 Windows, including 64-bit MinGW.  Use the
+ * standard pointer-sized integer types so these remain pointer-width there. */
+#include <stdint.h>
+typedef intptr_t                   khronos_intptr_t;
+typedef uintptr_t                  khronos_uintptr_t;
+typedef intptr_t                   khronos_ssize_t;
+typedef uintptr_t                  khronos_usize_t;
 
 #if KHRONOS_SUPPORT_FLOAT
 /*

tools/trace_replay/README.md

@@ -11,6 +11,8 @@ The bundled fixtures cover:
   ![Minecraft 1.21.4 startup golden](fixtures/minecraft-1.21.4-startup.0000092195.png)
 - minecraft-1.21.4-main-menu: captured from Minecraft 1.21.4's main menu.
   ![Minecraft 1.21.4 main menu golden](fixtures/minecraft-1.21.4-main-menu.0000481787.png)
+- minecraft-1.21.11-main-menu: captured from Minecraft 1.21.11's main menu on a Pixel 8 Pro through FCL MobileGL.
+  ![Minecraft 1.21.11 main menu golden](fixtures/minecraft-1.21.11-main-menu.0000205347.png)
 - minecraft-1.17-main-menu-854: captured from Minecraft 1.17's 854x480 main menu through FCL MobileGL capture.
   ![Minecraft 1.17 854x480 main menu golden](fixtures/minecraft-1.17-main-menu-854.0000117757.png)
 - minecraft-1.21.4-in-world: captured from Minecraft 1.21.4 after entering a singleplayer world.

tools/trace_replay/trace_cases.json

@@ -40,6 +40,13 @@
       "target_call": 481787,
       "timeout_seconds": 180
     },
+    {
+      "name": "minecraft-1.21.11-main-menu",
+      "trace_archive": "minecraft-1.21.11-main-menu.tgz",
+      "golden": "minecraft-1.21.11-main-menu.0000205347.png",
+      "target_call": 205347,
+      "timeout_seconds": 180
+    },
     {
       "name": "minecraft-1.17-main-menu-854",
       "trace_archive": "minecraft-1.17-main-menu-854.tgz",
@@ -270,11 +277,14 @@
     },
     {
       "name": "minecraft-1.21.4-fabric-iris-iterationrp-in-world",
-      "ci": false,
+      "ci_backends": [
+        "DirectVulkan"
+      ],
       "trace_archive": "minecraft-1.21.4-fabric-iris-iterationrp-in-world.tgz",
       "golden": "minecraft-1.21.4-fabric-iris-iterationrp-in-world.0000202020.png",
       "target_call": 202020,
-      "timeout_seconds": 1800
+      "timeout_seconds": 1800,
+      "ssim_threshold": 0.98
     },
     {
       "name": "minecraft-1.21.4-fabric-iris-bsl-esc-menu-854",

tools/trace_replay/trace_cases.py

@@ -6,6 +6,7 @@ from pathlib import Path
 
 
 TRACE_CASES_JSON = Path(__file__).with_name("trace_cases.json")
+CI_BACKENDS = ("DirectGLES", "DirectVulkan")
 
 
 def load_trace_case_manifest(path=TRACE_CASES_JSON):
@@ -71,6 +72,46 @@ def ci_trace_cases(cases):
     return [case for case in cases if case.get("ci", True)]
 
 
+def ci_backends(case):
+    backends = case.get("ci_backends")
+    if backends is None:
+        return CI_BACKENDS
+    if not isinstance(backends, list) or not backends:
+        raise ValueError(f"ci_backends must be a non-empty list for {case['name']}")
+    unknown = [backend for backend in backends if backend not in CI_BACKENDS]
+    if unknown:
+        raise ValueError(
+            f"unknown ci_backends for {case['name']}: {', '.join(unknown)}"
+        )
+    if len(set(backends)) != len(backends):
+        raise ValueError(f"ci_backends contains duplicates for {case['name']}")
+    return backends
+
+
+def github_test_matrix(cases):
+    return {
+        "include": [
+            {"backend": backend, "case": case["name"]}
+            for case in cases
+            for backend in ci_backends(case)
+        ]
+    }
+
+
+def github_apk_matrix(cases):
+    backends = {
+        "DirectGLES": {"name": "DirectGLES", "gpu": "software"},
+        "DirectVulkan": {"name": "DirectVulkan", "gpu": "lavapipe"},
+    }
+    return {
+        "include": [
+            {"backend": backends[backend], "case": github_apk_case(case)}
+            for case in cases
+            for backend in ci_backends(case)
+        ]
+    }
+
+
 def cmake_quote(value):
     return '"' + str(value).replace("\\", "/").replace('"', '\\"') + '"'
 
@@ -114,7 +155,14 @@ def parse_args():
     parser.add_argument("--fixture-root", default="tools/trace_replay/fixtures")
     parser.add_argument(
         "--format",
-        choices=("names", "github-apk", "fixture-files", "cmake"),
+        choices=(
+            "names",
+            "github-test-matrix",
+            "github-apk",
+            "github-apk-matrix",
+            "fixture-files",
+            "cmake",
+        ),
         default="names",
     )
     return parser.parse_args()
@@ -127,8 +175,12 @@ def main():
         cases = ci_trace_cases(cases)
     if args.format == "names":
         print(json.dumps([case["name"] for case in cases], separators=(",", ":")))
+    elif args.format == "github-test-matrix":
+        print(json.dumps(github_test_matrix(cases), separators=(",", ":")))
     elif args.format == "github-apk":
         print(json.dumps([github_apk_case(case) for case in cases], separators=(",", ":")))
+    elif args.format == "github-apk-matrix":
+        print(json.dumps(github_apk_matrix(cases), separators=(",", ":")))
     elif args.format == "fixture-files":
         if not args.case_name:
             print("--case is required for --format fixture-files", file=sys.stderr)