[Fix, Test] (DirectVulkan, ShaderTranspiler, TraceReplay): derive NumSubgroups behind opt-in quirk

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

.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: ${{ fromJSON(needs.trace-cases.outputs.android) }}
     steps:
       - name: Set Swap Space
         uses: pierotofy/set-swap-space@v1.0
@@ -441,6 +435,9 @@ jobs:
           if [ "${{ matrix.case.coherent_as_flush || false }}" = "true" ]; then
             extra_retrace_args+=(--coherent-as-flush)
           fi
+          if [ "${{ matrix.case.num_subgroups_quirk || false }}" = "true" ]; then
+            extra_retrace_args+=(--num-subgroups-quirk)
+          fi
 
           run_retrace() {
             timeout "$(( ${{ matrix.case.timeout_seconds }} + 300 ))" sh android-plugin/trace-replay-ci.sh \

.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

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)
@@ -283,6 +284,7 @@ set(SOURCE_FILES
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/SplitArrayVertexInputsPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/RebaseInstanceIndexPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/ZeroBaseVertexPass.cpp
+    MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DeriveNumSubgroupsPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/NormalizeRectCoordinatesPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/Lower1DArrayImagesPass.cpp
     MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/BakeImageFormatsPass.cpp
@@ -297,6 +299,7 @@ set(SOURCE_FILES
     MobileGL/MG_Util/BackendLoaders/Vulkan/Loader.cpp
 
     MobileGL/MG_Util/SelfTest/DriverPost.cpp
+    MobileGL/MG_Util/SelfTest/DriverPostProgram203Witness.cpp
 
     MobileGL/MG_Util/Texture/PixelStoreProcessor.cpp
     MobileGL/MG_Util/Texture/TextureFormatProcessor.cpp

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)
@@ -82,6 +80,11 @@ namespace MobileGL::MG_Config {
 #endif
         // MOBILEGL_DISABLE_SUBGROUP: force-disable Vulkan shader subgroup support.
         Bool DisableSubgroup = false;
+        // MOBILEGL_NUM_SUBGROUPS_QUIRK: derive compute gl_NumSubgroups from the local
+        // workgroup dimensions and gl_SubgroupSize instead of reading Vulkan's
+        // NumSubgroups builtin. Off by default; enable only for drivers whose builtin
+        // disagrees with the SubgroupId topology emitted by the same dispatch.
+        Bool NumSubgroupsQuirk = false;
         // MOBILEGL_ADVERTISE_FP64: add GL_ARB_gpu_shader_fp64 to the advertised extension
         // string. `double` in a shader always WORKS - it is narrowed to 32 bits before any
         // module reaches a backend (ShaderTranspiler::DemoteFloat64Pass) - but the extension
@@ -130,10 +133,6 @@ 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

@@ -162,11 +162,13 @@ 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, "");
 #endif
         features.DisableSubgroup = QueryEnvFlag("MOBILEGL_DISABLE_SUBGROUP");
+        features.NumSubgroupsQuirk = QueryEnvFlag("MOBILEGL_NUM_SUBGROUPS_QUIRK");
         features.AdvertiseFp64 = QueryEnvFlag("MOBILEGL_ADVERTISE_FP64");
         features.MagmaR11G11B10FFallback = QueryEnvFlag("MOBILEGL_MAGMA_R11G11B10F_FALLBACK");
         features.MagmaFramesInFlight = QueryEnvUint32("MOBILEGL_MAGMA_FRAMESINFLIGHT", 3, 1, 64);
@@ -179,7 +181,6 @@ 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");

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

@@ -3061,10 +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

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/ProgramFactory.cpp

@@ -8,6 +8,7 @@
 
 #include "ProgramFactory.h"
 
+#include "Config.h"
 #include "MG_Backend/DirectVulkan/DirectVulkanResourceState.h"
 #include "MG_Util/ShaderTranspiler/ShaderCompiler.h"
 #include "MG_Util/ShaderTranspiler/SpvcSession.h"
@@ -2973,8 +2974,73 @@ 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),
@@ -3054,6 +3120,10 @@ 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);
 
@@ -3094,12 +3164,29 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 }
             }
 
+            // NumSubgroups is defined by the local workgroup dimensions and SubgroupSize. Derive
+            // it in SPIR-V instead of trusting a driver builtin that can disagree with the
+            // SubgroupId topology produced by the same compute dispatch (Adreno reports 1 while
+            // emitting IDs 0..7 for a 512-invocation, 64-wide workgroup).
+            if (MG_Config::Features.NumSubgroupsQuirk && shaders[i] &&
+                shaders[i]->GetShaderStage() == ShaderStage::Compute) {
+                Vector<Uint> derivedNumSubgroupsSpirv;
+                if (MG_Util::ShaderTranspiler::ShaderCompiler::DeriveNumSubgroupsForVulkan(
+                        moduleSpirvs[i], derivedNumSubgroupsSpirv, enableSpirvValidation)) {
+                    moduleSpirvs[i] = std::move(derivedNumSubgroupsSpirv);
+                } else {
+                    MGLOG_E("ProgramFactory: failed to derive gl_NumSubgroups for program %u; "
+                            "compute shaders may observe a driver-inconsistent subgroup count",
+                            program.GetExternalIndex());
+                }
+            }
+
             // Vulkan's SPIR-V environment has no rectangle image dimension, so a
             // GL_TEXTURE_RECTANGLE lookup has to become the 2D one the texture is really
             // 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);
                 }
@@ -3112,7 +3199,7 @@ 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
@@ -3136,7 +3223,7 @@ 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; "
@@ -3154,7 +3241,7 @@ 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:
@@ -3177,7 +3264,7 @@ 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",
@@ -3201,7 +3288,7 @@ 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",
@@ -3222,7 +3309,7 @@ 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
@@ -3299,14 +3386,14 @@ 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;
             }
@@ -3440,7 +3527,8 @@ 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
@@ -196,6 +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);
@@ -224,11 +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;
@@ -240,6 +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;
@@ -254,6 +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);
@@ -282,11 +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;
@@ -298,6 +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;
@@ -343,12 +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
@@ -471,6 +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
@@ -540,11 +542,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         outImageInfo = {
             .sampler = m_samplerManager->GetOrCreateSampler(*samplerBindingOverride.sampler,
-                                                            *samplerBindingOverride.texture),
+                                                            *samplerBindingOverride.texture,
+                                                            samplerBindingOverride.forceNearestFiltering,
+                                                            resource->sampledLevelCount),
             .imageView = samplerBindingOverride.imageView != VK_NULL_HANDLE ?
                 samplerBindingOverride.imageView :
                 (resource->sampledView != VK_NULL_HANDLE ? resource->sampledView : resource->fullView),
-            .imageLayout = resource->layout,
+            .imageLayout = samplerBindingOverride.imageLayout != VK_IMAGE_LAYOUT_UNDEFINED ?
+                samplerBindingOverride.imageLayout : resource->layout,
         };
         return outImageInfo.sampler != VK_NULL_HANDLE;
     }
@@ -1267,6 +1272,87 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return true;
     }
 
+    Bool UniformManager::SamplerOverlapsWritableImageSubresource(Int samplerBaseLevel, Int samplerMaxLevel,
+                                                                 GLint imageLevel, GLenum imageAccess) {
+        return imageAccess != GL_READ_ONLY && imageLevel >= samplerBaseLevel && imageLevel <= samplerMaxLevel;
+    }
+
+    Bool UniformManager::CollectSamplerImageFeedback(
+        const MG_State::GLState::ProgramObject& program,
+        const ProgramFactory::VkProgramObject& programObj,
+        Vector<SamplerImageFeedbackBinding>& outBindings) const {
+        outBindings.clear();
+        MOBILEGL_ASSERT(MG_State::pGLContext != nullptr,
+                        "CollectSamplerImageFeedback: GL context is null");
+        if (programObj.declinedDescriptors) return true;
+
+        for (const Uint32 samplerBinding : programObj.activeBindings) {
+            if (samplerBinding >= m_maxBindings ||
+                programObj.bindingKinds[samplerBinding] != ProgramFactory::DescriptorBindingKind::CombinedImageSampler) {
+                continue;
+            }
+            const Uint32 samplerCount = BindingDescriptorCount(programObj, samplerBinding);
+            for (Uint32 samplerElement = 0; samplerElement < samplerCount; ++samplerElement) {
+                MG_State::GLState::ITextureObject* sampledTexture = nullptr;
+                const MG_State::GLState::SamplerObject* sampledSampler = nullptr;
+                if (!ResolveSampledBinding(program, programObj, samplerBinding, samplerElement,
+                                           sampledTexture, sampledSampler) ||
+                    sampledTexture == nullptr || sampledSampler == nullptr ||
+                    MG_State::GLState::SamplesAsIncompleteTexture(sampledTexture, sampledSampler)) {
+                    // ResolveSamplerDescriptor uses a fallback in these cases, which cannot
+                    // alias the image-unit binding of the original texture.
+                    continue;
+                }
+                // Multisample source images intentionally omit TRANSFER_SRC usage. Keep their existing
+                // direct binding instead of turning otherwise valid sampler2DMS/image2DMS dispatches
+                // into failed dispatches; a correct snapshot for them needs a same-sample-count path.
+                const TextureTarget sampledTarget = sampledTexture->GetTarget();
+                if (sampledTarget == TextureTarget::Texture2DMultisample ||
+                    sampledTarget == TextureTarget::Texture2DMultisampleArray) {
+                    continue;
+                }
+                const auto& levelRange = sampledTexture->GetLevelRange();
+                Bool aliasesWritableImage = false;
+                for (const Uint32 imageBinding : programObj.activeBindings) {
+                    if (imageBinding >= m_maxBindings ||
+                        programObj.bindingKinds[imageBinding] != ProgramFactory::DescriptorBindingKind::StorageImage) {
+                        continue;
+                    }
+                    if (imageBinding >= programObj.samplerUniformLocationByBinding.size()) return false;
+                    const Int baseLocation = programObj.samplerUniformLocationByBinding[imageBinding];
+                    if (baseLocation < 0) return false;
+                    const Uint32 imageCount = BindingDescriptorCount(programObj, imageBinding);
+                    for (Uint32 imageElement = 0; imageElement < imageCount; ++imageElement) {
+                        const Int location = ResolveDescriptorElementLocation(program, baseLocation, imageElement);
+                        if (location < 0) return false;
+                        const Int imageUnit = program.GetUniformSamplerOrImageUnitIndex(static_cast<Uint>(location));
+                        if (imageUnit < 0 || imageUnit >= MG_State::GLState::TextureState::MAX_TEXTURE_IMAGE_UNITS) {
+                            return false;
+                        }
+                        const auto& image = MG_State::pGLContext->GetImageTextureBinding(imageUnit);
+                        // A sampler view exposes all layers of its target; equal texture plus an
+                        // overlapping mip therefore aliases the writable image subresource.
+                        if (image.Texture.get() == sampledTexture &&
+                            SamplerOverlapsWritableImageSubresource(levelRange.x(), levelRange.y(),
+                                                                     image.Level, image.Access)) {
+                            aliasesWritableImage = true;
+                            break;
+                        }
+                    }
+                    if (aliasesWritableImage) break;
+                }
+                if (aliasesWritableImage) {
+                    outBindings.push_back({.samplerBinding = samplerBinding,
+                                           .samplerElement = samplerElement,
+                                           .texture = sampledTexture,
+                                           .sampler = sampledSampler,
+                                           .numericDomain = programObj.samplerNumericDomainByBinding[samplerBinding]});
+                }
+            }
+        }
+        return true;
+    }
+
     Bool UniformManager::ResolveUniformBufferPayload(const MG_State::GLState::ProgramObject& program,
                                                      const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
                                                      Uint32 arrayElement, UboBindResult& out) const {
@@ -1388,7 +1474,7 @@ 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 +1517,8 @@ 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 +1532,28 @@ 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 +1563,16 @@ 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 >=
+        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 +1608,7 @@ 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;
                 }
@@ -1633,7 +1726,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                              Uint32 frameIndex,
                                                              VkPipelineBindPoint bindPoint,
                                                              const SamplerBindingOverride* samplerBindingOverride,
-                                                             Bool samplerDescriptorsUnchangedHint) {
+                                                             Bool samplerDescriptorsUnchangedHint,
+                                                             const Vector<SamplerBindingOverride>* samplerBindingOverrides) {
         // This program has a descriptor MobileGL could not resolve (see
         // VkProgramObject::declinedDescriptors). Refusing here is the whole of the decline: the
         // binding is still declared in the layout, so the pipeline is consistent with the shader
@@ -1660,7 +1754,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // sampler binding, and an unchanged (buffer, range) for the single
         // dynamic UBO covers the rest - except the dynamic offset, which rebinding
         // the SAME set delivers without any descriptor write.
-        const Bool cacheable = (samplerBindingOverride == nullptr);
+        const Bool cacheable = samplerBindingOverride == nullptr &&
+                               (samplerBindingOverrides == nullptr || samplerBindingOverrides->empty());
         if (cacheable && samplerDescriptorsUnchangedHint && m_fastRebindMemo.valid &&
             m_fastRebindMemo.frameIndex == frameIndex &&
             m_fastRebindMemo.programLifetimeId == program.GetLifetimeId() &&
@@ -1884,13 +1979,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 const SizeT firstImageInfoIndex = imageInfos.size();
                 for (Uint32 element = 0; element < descriptorCount; ++element) {
                     VkDescriptorImageInfo imageInfo{};
-                    Bool hasImage = false;
-                    if (overrideThisBinding && element == 0) {
-                        hasImage = ResolveSamplerDescriptorOverride(*samplerBindingOverride, imageInfo);
-                    } else {
-                        hasImage = ResolveSamplerDescriptor(commandBuffer, program, programObj, binding, element,
-                                                            imageInfo, samplerDescriptorsUnchangedHint);
+                    const SamplerBindingOverride* overrideForElement =
+                        overrideThisBinding && element == 0 ? samplerBindingOverride : nullptr;
+                    if (overrideForElement == nullptr && samplerBindingOverrides != nullptr) {
+                        const auto overrideIt = std::find_if(
+                            samplerBindingOverrides->begin(), samplerBindingOverrides->end(),
+                            [binding, element](const SamplerBindingOverride& candidate) {
+                                return candidate.binding == binding && candidate.element == element;
+                            });
+                        if (overrideIt != samplerBindingOverrides->end()) {
+                            overrideForElement = &*overrideIt;
                         }
+                    }
+                    const Bool hasImage = overrideForElement != nullptr
+                                              ? ResolveSamplerDescriptorOverride(*overrideForElement, imageInfo)
+                                              : ResolveSamplerDescriptor(commandBuffer, program, programObj, binding,
+                                                                         element, imageInfo,
+                                                                         samplerDescriptorsUnchangedHint);
                     if (!hasImage) {
                         MGLOG_E_ONCE(
                             "UniformDescriptorBinder::BindProgramUniformBuffers failed: sampler binding %u element %u "

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -26,9 +26,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     public:
         struct SamplerBindingOverride {
             Uint32 binding = 0;
+            Uint32 element = 0;
             MG_State::GLState::ITextureObject* texture = nullptr;
             const MG_State::GLState::SamplerObject* sampler = nullptr;
             VkImageView imageView = VK_NULL_HANDLE;
+            VkImageLayout imageLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+            Bool forceNearestFiltering = false;
+        };
+
+        struct SamplerImageFeedbackBinding {
+            Uint32 samplerBinding = 0;
+            Uint32 samplerElement = 0;
+            MG_State::GLState::ITextureObject* texture = nullptr;
+            const MG_State::GLState::SamplerObject* sampler = nullptr;
+            SamplerNumericDomain numericDomain = SamplerNumericDomain::Unknown;
         };
 
         Bool Initialize(VkDevice device, VkBufferManager* bufferManager,
@@ -79,6 +90,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool CollectStorageImageTextures(const MG_State::GLState::ProgramObject& program,
                                          const ProgramFactory::VkProgramObject& programObj,
                                          Vector<MG_State::GLState::ITextureObject*>& outTextures) const;
+        Bool CollectSamplerImageFeedback(
+            const MG_State::GLState::ProgramObject& program,
+            const ProgramFactory::VkProgramObject& programObj,
+            Vector<SamplerImageFeedbackBinding>& outBindings) const;
+        static Bool SamplerOverlapsWritableImageSubresource(Int samplerBaseLevel, Int samplerMaxLevel,
+                                                             GLint imageLevel, GLenum imageAccess);
         // samplerDescriptorsUnchangedHint: the caller (SetupDraw fast path) proved that
         // every input of every combined-image-sampler resolution is unchanged since the
         // previous draw's resolve - same (texture, sampler) per binding, texture params
@@ -91,7 +108,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                        Uint32 frameIndex,
                                        VkPipelineBindPoint bindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
                                        const SamplerBindingOverride* samplerBindingOverride = nullptr,
-                                       Bool samplerDescriptorsUnchangedHint = false);
+                                       Bool samplerDescriptorsUnchangedHint = false,
+                                       const Vector<SamplerBindingOverride>* samplerBindingOverrides = nullptr);
 
         // Pure format-policy helper kept public for host regression tests. Formatted storage
         // images use their shader qualifier; transformed float images use glBindImageTexture's
@@ -114,6 +132,7 @@ 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 +242,8 @@ 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 +360,11 @@ 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

@@ -1291,6 +1291,158 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return ok;
     }
 
+    Bool VkTextureManager::SnapshotTextureForSampling(VkCommandBuffer commandBuffer,
+                                                      MG_State::GLState::ITextureObject& texture,
+                                                      SamplerNumericDomain numericDomain,
+                                                      VkPipelineStageFlags consumerShaderStageMask,
+                                                      SampledTextureSnapshot& outSnapshot) {
+        outSnapshot = {};
+        TextureResource* source = SyncTextureAndGetDescriptor(texture);
+        if (source == nullptr || source->image == VK_NULL_HANDLE || source->sampleCount != VK_SAMPLE_COUNT_1_BIT ||
+            source->sampledLevelCount == 0) {
+            return false;
+        }
+
+        const VkFormat sampledFormat = ResolveSampledImageViewFormat(source->format, numericDomain);
+        if (sampledFormat == VK_FORMAT_UNDEFINED ||
+            !AreSampledImageViewFormatsCompatible(source->format, sampledFormat)) {
+            MGLOG_E_ONCE("SnapshotTextureForSampling: textureId=%d cannot create sampled view format=%d from image format=%d",
+                         texture.GetExternalIndex(), static_cast<Int>(sampledFormat), static_cast<Int>(source->format));
+            return false;
+        }
+        if (sampledFormat != source->format &&
+            (source->imageCreateFlags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) == 0) {
+            MGLOG_E_ONCE("SnapshotTextureForSampling: textureId=%d needs unavailable mutable image format=%d for sampled view=%d",
+                         texture.GetExternalIndex(), static_cast<Int>(source->format), static_cast<Int>(sampledFormat));
+            return false;
+        }
+
+        VkImageType imageType = VK_IMAGE_TYPE_2D;
+        switch (source->viewType) {
+        case VK_IMAGE_VIEW_TYPE_1D:
+        case VK_IMAGE_VIEW_TYPE_1D_ARRAY:
+            imageType = VK_IMAGE_TYPE_1D;
+            break;
+        case VK_IMAGE_VIEW_TYPE_3D:
+            imageType = VK_IMAGE_TYPE_3D;
+            break;
+        default:
+            break;
+        }
+
+        TextureResource snapshot{};
+        VkImageCreateInfo imageInfo{};
+        imageInfo.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
+        imageInfo.flags = source->imageCreateFlags;
+        imageInfo.imageType = imageType;
+        imageInfo.extent = {source->extent.width, source->extent.height, source->depth};
+        imageInfo.mipLevels = source->mipLevels;
+        imageInfo.arrayLayers = source->arrayLayers;
+        imageInfo.format = source->format;
+        imageInfo.tiling = VK_IMAGE_TILING_OPTIMAL;
+        imageInfo.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+        imageInfo.usage = VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_SAMPLED_BIT;
+        imageInfo.samples = VK_SAMPLE_COUNT_1_BIT;
+        imageInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+
+        // Keep the temporary's view-format list just as narrow as the source's sampler use. This
+        // has no storage-image usage, so unlike an app image binding the exact list is knowable.
+        Vector<VkFormat> viewFormats;
+        VkImageFormatListCreateInfo formatListInfo{};
+        if (m_imageFormatListSupported && (imageInfo.flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) != 0) {
+            viewFormats.push_back(source->format);
+            if (sampledFormat != source->format) {
+                viewFormats.push_back(sampledFormat);
+            }
+            formatListInfo.sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_LIST_CREATE_INFO;
+            formatListInfo.viewFormatCount = static_cast<Uint32>(viewFormats.size());
+            formatListInfo.pViewFormats = viewFormats.data();
+            imageInfo.pNext = &formatListInfo;
+        }
+
+        VmaAllocationCreateInfo allocationInfo{};
+        allocationInfo.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
+        allocationInfo.requiredFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
+        const VkResult createResult =
+            vmaCreateImage(m_allocator, &imageInfo, &allocationInfo, &snapshot.image, &snapshot.allocation, nullptr);
+        if (createResult != VK_SUCCESS) {
+            MGLOG_E_ONCE("SnapshotTextureForSampling: vmaCreateImage failed result=%d textureId=%d", createResult,
+                         texture.GetExternalIndex());
+            return false;
+        }
+
+        snapshot.extent = source->extent;
+        snapshot.depth = source->depth;
+        snapshot.arrayLayers = source->arrayLayers;
+        snapshot.mipLevels = source->mipLevels;
+        snapshot.sampledBaseMipLevel = source->sampledBaseMipLevel;
+        snapshot.sampledLevelCount = source->sampledLevelCount;
+        snapshot.format = source->format;
+        snapshot.aspect = source->aspect;
+        snapshot.viewType = source->viewType;
+        snapshot.sampleCount = VK_SAMPLE_COUNT_1_BIT;
+        snapshot.imageCreateFlags = imageInfo.flags;
+        snapshot.usageFlags = imageInfo.usage;
+
+        const TextureFormatInfo formatInfo = ResolveTextureFormatInfo(texture.GetFormat());
+        const VkComponentMapping sampledComponents = ResolveSampledViewComponents(texture, formatInfo);
+        const VkImageAspectFlags sampledAspect =
+            ResolveSampledImageViewAspectMask(snapshot.aspect, texture.GetDepthStencilTextureMode());
+        snapshot.sampledView = CreateImageView(snapshot.image, sampledFormat, sampledAspect, snapshot.viewType,
+                                               snapshot.sampledBaseMipLevel, snapshot.sampledLevelCount, 0,
+                                               snapshot.arrayLayers, &sampledComponents);
+        if (snapshot.sampledView == VK_NULL_HANDLE) {
+            MGLOG_E_ONCE("SnapshotTextureForSampling: failed to create sampled view textureId=%d", texture.GetExternalIndex());
+            return false;
+        }
+
+        VkPipelineStageFlags sourceStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
+        VkAccessFlags sourceAccessMask = 0;
+        const VkImageLayout sourceLayout = source->layout;
+        GetImageTransitionSourceState(sourceLayout, sourceStageMask, sourceAccessMask);
+        if (!TransitionImageLayout(commandBuffer, source->image, source->layout, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+                                   sourceStageMask, VK_PIPELINE_STAGE_TRANSFER_BIT, sourceAccessMask,
+                                   VK_ACCESS_TRANSFER_READ_BIT, source->aspect, 0, source->mipLevels) ||
+            !TransitionImageLayout(commandBuffer, snapshot.image, snapshot.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, snapshot.aspect, snapshot.sampledBaseMipLevel,
+                                   snapshot.sampledLevelCount)) {
+            return false;
+        }
+
+        Vector<VkImageCopy> copyRegions;
+        copyRegions.reserve(snapshot.sampledLevelCount);
+        for (Uint32 level = snapshot.sampledBaseMipLevel;
+             level < snapshot.sampledBaseMipLevel + snapshot.sampledLevelCount; ++level) {
+            VkImageCopy copy{};
+            copy.srcSubresource = {source->aspect, level, 0, source->arrayLayers};
+            copy.dstSubresource = {snapshot.aspect, level, 0, snapshot.arrayLayers};
+            copy.extent = {std::max(source->extent.width >> level, 1u),
+                           std::max(source->extent.height >> level, 1u),
+                           std::max(source->depth >> level, 1u)};
+            copyRegions.push_back(copy);
+        }
+        vkCmdCopyImage(commandBuffer, source->image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, snapshot.image,
+                       VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, static_cast<Uint32>(copyRegions.size()), copyRegions.data());
+
+        if (!TransitionImageLayout(commandBuffer, snapshot.image, snapshot.layout,
+                                   ResolveSampledReadOnlyLayout(snapshot.aspect), VK_PIPELINE_STAGE_TRANSFER_BIT,
+                                   consumerShaderStageMask, VK_ACCESS_TRANSFER_WRITE_BIT,
+                                   VK_ACCESS_SHADER_READ_BIT, snapshot.aspect, snapshot.sampledBaseMipLevel,
+                                   snapshot.sampledLevelCount) ||
+            !TransitionImageLayout(commandBuffer, source->image, source->layout, sourceLayout,
+                                   VK_PIPELINE_STAGE_TRANSFER_BIT, consumerShaderStageMask,
+                                   VK_ACCESS_TRANSFER_READ_BIT, VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+                                   source->aspect, 0, source->mipLevels)) {
+            return false;
+        }
+
+        StampResourceRecordingUse(*source);
+        outSnapshot = {.imageView = snapshot.sampledView, .layout = snapshot.layout};
+        DeferResourceRelease(Move(snapshot));
+        return true;
+    }
+
     void VkTextureManager::MarkStorageImageTexture(MG_State::GLState::ITextureObject& texture) {
         m_storageImageTextures.insert(MakeTextureIdentity(&texture));
     }

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.h

@@ -310,6 +310,11 @@ public:
         static inline VmaAllocator s_allocator = VK_NULL_HANDLE;
     };
 
+    struct SampledTextureSnapshot {
+        VkImageView imageView = VK_NULL_HANDLE;
+        VkImageLayout layout = VK_IMAGE_LAYOUT_UNDEFINED;
+    };
+
     Bool Initialize(const InitInfo& initInfo);
     void Shutdown();
     void BeginFrame(Uint32 frameIndex);
@@ -343,6 +348,13 @@ public:
                                                       VkImageLayout newLayout);
     Bool TransitionTextureForSampling(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
     Bool TransitionTextureForStorageImage(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
+    // Copies the complete sampler-visible mip range into a transient sampled image. The source is
+    // restored to its prior layout, so image-store descriptors continue to name the original image.
+    // The transient ownership is tied to the current frame slot and is safe through its submission.
+    Bool SnapshotTextureForSampling(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture,
+                                    SamplerNumericDomain numericDomain,
+                                    VkPipelineStageFlags consumerShaderStageMask,
+                                    SampledTextureSnapshot& outSnapshot);
 
     // Recording-generation bookkeeping for the pre-pass command stream. The
     // generation advances every time the frame command buffer (re)begins

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;
@@ -1480,7 +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;
@@ -1495,6 +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;
@@ -2151,47 +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);
@@ -2738,6 +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,
@@ -2962,7 +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;
@@ -2996,7 +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.
@@ -5346,7 +5412,81 @@ void main() {
         }
         return true;
     }
+    Bool VulkanRenderer::PrepareSamplerImageFeedbackSnapshots(
+        FrameContext::FrameData& frame,
+        const MG_State::GLState::ProgramObject& program,
+        const ProgramFactory::VkProgramObject& programObj,
+        VkPipelineStageFlags consumerShaderStageMask) {
+        auto& feedbackBindings = m_samplerImageFeedbackScratch;
+        auto& overrides = m_samplerImageBindingOverridesScratch;
+        overrides.clear();
+        if (!programObj.hasStorageImages) {
+            feedbackBindings.clear();
+            return true;
+        }
+        if (!m_uniformManager->CollectSamplerImageFeedback(program, programObj, feedbackBindings)) {
+            MGLOG_E_ONCE("%s: failed to collect sampler/image feedback for program=%u", __func__,
+                         program.GetExternalIndex());
+            return false;
+        }
 
+        if (feedbackBindings.empty()) {
+            return true;
+        }
+        // Copy and layout barriers cannot be recorded inside a render pass. A graphics draw only
+        // gets here after an actual sampled/writable-image mip overlap was found, so ordinary
+        // graphics draws retain the active pass.
+        if (VkRenderPassManager::GetActiveRenderPass() != nullptr) {
+            VkRenderPassManager::EndRenderPass(frame.commandBuffer);
+        }
+
+        struct SnapshotCacheEntry {
+            MG_State::GLState::ITextureObject* texture = nullptr;
+            SamplerNumericDomain numericDomain = SamplerNumericDomain::Unknown;
+            VkTextureManager::SampledTextureSnapshot snapshot{};
+        };
+        Vector<SnapshotCacheEntry> snapshotCache;
+        snapshotCache.reserve(feedbackBindings.size());
+        overrides.reserve(feedbackBindings.size());
+        for (const auto& feedback : feedbackBindings) {
+            VkTextureManager::SampledTextureSnapshot snapshot{};
+            const auto existing = std::find_if(
+                snapshotCache.begin(), snapshotCache.end(), [&feedback](const SnapshotCacheEntry& candidate) {
+                    return candidate.texture == feedback.texture && candidate.numericDomain == feedback.numericDomain;
+                });
+            if (existing != snapshotCache.end()) {
+                snapshot = existing->snapshot;
+            } else {
+                if (!m_textureManager->SnapshotTextureForSampling(frame.commandBuffer, *feedback.texture,
+                                                                  feedback.numericDomain, consumerShaderStageMask,
+                                                                  snapshot) ||
+                    snapshot.imageView == VK_NULL_HANDLE) {
+                    MGLOG_E_ONCE("%s: failed to snapshot textureId=%d for sampler binding=%u element=%u", __func__,
+                                 feedback.texture != nullptr ? feedback.texture->GetExternalIndex() : 0,
+                                 feedback.samplerBinding, feedback.samplerElement);
+                    return false;
+                }
+                snapshotCache.push_back({.texture = feedback.texture,
+                                         .numericDomain = feedback.numericDomain,
+                                         .snapshot = snapshot});
+            }
+            overrides.push_back({
+                .binding = feedback.samplerBinding,
+                .element = feedback.samplerElement,
+                .texture = feedback.texture,
+                .sampler = feedback.sampler,
+                .imageView = snapshot.imageView,
+                .imageLayout = snapshot.layout,
+                .forceNearestFiltering = feedback.numericDomain == SamplerNumericDomain::SignedInteger ||
+                                         feedback.numericDomain == SamplerNumericDomain::UnsignedInteger,
+            });
+            if (program.GetExternalIndex() == 194 && feedback.texture->GetExternalIndex() == 75) {
+                MGLOG_D_ONCE("sampler/image feedback snapshot: program=194 texture=75 binding=%u element=%u view=%p",
+                             feedback.samplerBinding, feedback.samplerElement, snapshot.imageView);
+            }
+        }
+        return true;
+    }
 
     // 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
@@ -6028,6 +6168,11 @@ void main() {
             MGLOG_E_ONCE("SetupDraw skipped: storage image preparation failed");
             return false;
         }
+        if (!PrepareSamplerImageFeedbackSnapshots(frame, program, programObj,
+                                                  VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT)) {
+            MGLOG_E_ONCE("SetupDraw skipped: sampler/image feedback snapshot failed");
+            return false;
+        }
 
         auto* activeRenderPass = VkRenderPassManager::GetActiveRenderPass();
 
@@ -6272,7 +6417,9 @@ void main() {
         }
 
         const Bool boundUniforms = m_uniformManager->BindProgramUniformBuffers(
-            frame.commandBuffer, program, programObj, m_frameContext.GetCurrentFrameIndex());
+            frame.commandBuffer, program, programObj, m_frameContext.GetCurrentFrameIndex(),
+            VK_PIPELINE_BIND_POINT_GRAPHICS, nullptr, false,
+            m_samplerImageBindingOverridesScratch.empty() ? nullptr : &m_samplerImageBindingOverridesScratch);
         if (!boundUniforms) {
             MGLOG_E_ONCE("SetupDraw skipped: BindProgramUniformBuffers failed");
             return false;
@@ -6395,6 +6542,11 @@ void main() {
             MGLOG_E_ONCE("DispatchCompute skipped: storage image preparation failed");
             return;
         }
+        if (!PrepareSamplerImageFeedbackSnapshots(frame, program, programObj,
+                                                  VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT)) {
+            MGLOG_E_ONCE("DispatchCompute skipped: sampler/image feedback snapshot failed");
+            return;
+        }
 
         const VkPipeline pipeline = GetOrCreateComputePipeline(programObj);
         if (pipeline == VK_NULL_HANDLE) {
@@ -6406,7 +6558,8 @@ void main() {
         vkCmdBindPipeline(frame.commandBuffer, VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
         const Bool boundUniforms = m_uniformManager->BindProgramUniformBuffers(
             frame.commandBuffer, program, programObj, m_frameContext.GetCurrentFrameIndex(),
-            VK_PIPELINE_BIND_POINT_COMPUTE);
+            VK_PIPELINE_BIND_POINT_COMPUTE, nullptr, false,
+            m_samplerImageBindingOverridesScratch.empty() ? nullptr : &m_samplerImageBindingOverridesScratch);
         if (!boundUniforms) {
             MGLOG_E_ONCE("DispatchCompute skipped: BindProgramUniformBuffers failed");
             return;
@@ -6441,6 +6594,11 @@ void main() {
             MGLOG_E_ONCE("DispatchComputeIndirect skipped: storage image preparation failed");
             return;
         }
+        if (!PrepareSamplerImageFeedbackSnapshots(frame, program, programObj,
+                                                  VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT)) {
+            MGLOG_E_ONCE("DispatchComputeIndirect skipped: sampler/image feedback snapshot failed");
+            return;
+        }
 
         const VkPipeline pipeline = GetOrCreateComputePipeline(programObj);
         if (pipeline == VK_NULL_HANDLE) {
@@ -6452,7 +6610,8 @@ void main() {
         vkCmdBindPipeline(frame.commandBuffer, VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
         const Bool boundUniforms = m_uniformManager->BindProgramUniformBuffers(
             frame.commandBuffer, program, programObj, m_frameContext.GetCurrentFrameIndex(),
-            VK_PIPELINE_BIND_POINT_COMPUTE);
+            VK_PIPELINE_BIND_POINT_COMPUTE, nullptr, false,
+            m_samplerImageBindingOverridesScratch.empty() ? nullptr : &m_samplerImageBindingOverridesScratch);
         if (!boundUniforms) {
             MGLOG_E_ONCE("DispatchComputeIndirect skipped: BindProgramUniformBuffers failed");
             return;
@@ -9138,19 +9297,18 @@ 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);
 
@@ -9192,16 +9350,14 @@ 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),
+            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));
         }
@@ -9582,16 +9738,15 @@ 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]{};
@@ -9603,8 +9758,8 @@ 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++];
@@ -9613,8 +9768,8 @@ 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);
@@ -9648,13 +9803,13 @@ void main() {
             Bool remapped = true;
             if (wantDepth && depthCopyBytes > 0) {
                 remappedDepth.resize(pixelCount * depthCopyBytes);
-                remapped = RemapDefaultFboReadbackToGLOrientation(depthSrc, static_cast<Uint32>(width),
+                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),
+                remapped = RemapDefaultFramebufferReadback(stencilSrc, static_cast<Uint32>(width),
                                                            static_cast<Uint32>(height), preTransform, 1,
                                                            remappedStencil.data());
             }
@@ -9662,9 +9817,7 @@ void main() {
                 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));
             }
@@ -11857,6 +12010,17 @@ 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) {
@@ -11998,17 +12162,19 @@ 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) {
@@ -12434,6 +12600,75 @@ 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.
@@ -908,6 +929,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // already sampleable.
         Vector<VkTextureManager::TextureResource*> m_sampledResourcesScratch;
         Vector<MG_State::GLState::ITextureObject*> m_storageImageTexturesScratch;
+        Vector<UniformManager::SamplerImageFeedbackBinding> m_samplerImageFeedbackScratch;
+        Vector<UniformManager::SamplerBindingOverride> m_samplerImageBindingOverridesScratch;
         Vector<VkBuffer> m_vertexBuffersScratch;
         Vector<VkDeviceSize> m_vertexOffsetsScratch;
         Vector<VkVertexInputAttributeDescription> m_patchedAttributesScratch;
@@ -1127,6 +1150,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             FrameContext::FrameData& frame,
             const MG_State::GLState::ProgramObject& program,
             const ProgramFactory::VkProgramObject& programObj);
+        // Vulkan forbids a sampled descriptor and writable storage descriptor from naming the
+        // same image subresource in one shader operation. Snapshot only the sampler side; the
+        // storage descriptor continues to name the application texture.
+        Bool PrepareSamplerImageFeedbackSnapshots(
+            FrameContext::FrameData& frame,
+            const MG_State::GLState::ProgramObject& program,
+            const ProgramFactory::VkProgramObject& programObj,
+            VkPipelineStageFlags consumerShaderStageMask);
 
         // The per-draw dynamic-state tail (viewport, scissor, blend constants, depth
         // bias, line width, stencil), gated behind one render-state-parameters-version

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,
+        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) {
@@ -1197,16 +1226,33 @@ 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,
+        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,
+        auto bufferObject = GetNamedBufferObject(buffer, BufferOp::ClearNamedBufferSubData);
+        if (!bufferObject) return;
+        ClearBufferRange_State(bufferObject, internalformat, offset, size, format, type, data,
                                BufferOp::ClearNamedBufferSubData);
     }
 
@@ -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

@@ -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_IntegrationTest/CMakeLists.txt

@@ -68,6 +68,7 @@ add_executable(MobileGLIntegrationTest
     Scenarios/DoublePrecisionScenario.cpp
     Scenarios/UniformInitializerScenario.cpp
     Scenarios/SwizzleAccessRoutineScenario.cpp
+    Scenarios/Program203FirstReductionScenario.cpp
     Scenarios/ProgramPipelineScenario.cpp
     Scenarios/ImageLoadStoreSsoScenario.cpp
     Scenarios/ImageTargetKindScenario.cpp

MobileGL/MG_IntegrationTest/Harness/HeadlessGL.cpp

@@ -15,6 +15,11 @@
 #include <ostream>
 #include <sstream>
 
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#endif
+
 // MobileGL's own headers, in the order MobileGL/Includes.h uses them: GL/gl.h
 // first, then glcorearb.h for the 3.x+ entry points. This binary links
 // MobileGL_s, so every gl*/egl* below binds to MobileGL's implementation, not
@@ -53,6 +58,37 @@ namespace MGITest {
         constexpr int kSurfaceWidth = 128;
         constexpr int kSurfaceHeight = 96;
 
+#if defined(_WIN32)
+        HWND g_testWindow = nullptr;
+
+        HWND CreateTestWindow() {
+            static const wchar_t* const kClassName = L"MobileGLIntegrationTestWindow";
+            static bool registered = false;
+            if (!registered) {
+                WNDCLASSW windowClass{};
+                windowClass.lpfnWndProc = DefWindowProcW;
+                windowClass.hInstance = GetModuleHandleW(nullptr);
+                windowClass.lpszClassName = kClassName;
+                if (RegisterClassW(&windowClass) == 0 && GetLastError() != ERROR_CLASS_ALREADY_EXISTS) {
+                    return nullptr;
+                }
+                registered = true;
+            }
+            return CreateWindowExW(0, kClassName, L"MobileGL Integration Test", WS_OVERLAPPEDWINDOW,
+                                   CW_USEDEFAULT, CW_USEDEFAULT, kSurfaceWidth, kSurfaceHeight, nullptr, nullptr,
+                                   GetModuleHandleW(nullptr), nullptr);
+        }
+#endif
+
+        bool UseWindowSurface() {
+#if defined(_WIN32)
+            const char* value = std::getenv("MOBILEGL_ITEST_WINDOW_SURFACE");
+            return value != nullptr && value[0] != '\0' && std::strcmp(value, "0") != 0;
+#else
+            return false;
+#endif
+        }
+
         std::string EnvOr(const char* name, const char* fallback) {
             const char* value = std::getenv(name);
             return (value != nullptr && value[0] != '\0') ? std::string(value) : std::string(fallback);
@@ -134,8 +170,9 @@ namespace MGITest {
                 return 3;
             }
 
+            const bool useWindowSurface = UseWindowSurface();
             const EGLint configAttribs[] = {EGL_SURFACE_TYPE,
-                                            EGL_PBUFFER_BIT,
+                                            useWindowSurface ? EGL_WINDOW_BIT : EGL_PBUFFER_BIT,
                                             EGL_RED_SIZE,
                                             8,
                                             EGL_GREEN_SIZE,
@@ -152,7 +189,9 @@ namespace MGITest {
             EGLConfig config = nullptr;
             EGLint configCount = 0;
             if (eglChooseConfig(display, configAttribs, &config, 1, &configCount) != EGL_TRUE || configCount < 1) {
-                outReason = WithEglError("eglChooseConfig found no pbuffer-capable RGBA8/D24 config");
+                outReason = WithEglError(useWindowSurface
+                                             ? "eglChooseConfig found no window-capable RGBA8/D24 config"
+                                             : "eglChooseConfig found no pbuffer-capable RGBA8/D24 config");
                 return 4;
             }
 
@@ -166,10 +205,23 @@ namespace MGITest {
                 return 5;
             }
 
+            EGLSurface surface = EGL_NO_SURFACE;
+            if (useWindowSurface) {
+#if defined(_WIN32)
+                if (g_testWindow == nullptr) g_testWindow = CreateTestWindow();
+                if (g_testWindow == nullptr) {
+                    outReason = "failed to create the Windows integration-test window";
+                    return 6;
+                }
+                surface = eglCreateWindowSurface(display, config, g_testWindow, nullptr);
+#endif
+            } else {
                 const EGLint pbufferAttribs[] = {EGL_WIDTH, kSurfaceWidth, EGL_HEIGHT, kSurfaceHeight, EGL_NONE};
-            EGLSurface surface = eglCreatePbufferSurface(display, config, pbufferAttribs);
+                surface = eglCreatePbufferSurface(display, config, pbufferAttribs);
+            }
             if (surface == EGL_NO_SURFACE) {
-                outReason = WithEglError("eglCreatePbufferSurface failed");
+                outReason = WithEglError(useWindowSurface ? "eglCreateWindowSurface failed"
+                                                         : "eglCreatePbufferSurface failed");
                 return 6;
             }
             // The step that brings the whole backend up (DirectVulkan creates its
@@ -491,6 +543,12 @@ namespace MGITest {
         if (m_context != nullptr) eglDestroyContext(display, static_cast<EGLContext>(m_context));
         if (m_surface != nullptr) eglDestroySurface(display, static_cast<EGLSurface>(m_surface));
         eglTerminate(display);
+#if defined(_WIN32)
+        if (g_testWindow != nullptr) {
+            DestroyWindow(g_testWindow);
+            g_testWindow = nullptr;
+        }
+#endif
         m_context = nullptr;
         m_surface = nullptr;
         m_display = nullptr;

MobileGL/MG_IntegrationTest/Scenarios/Program203FirstReductionScenario.cpp

@@ -0,0 +1,878 @@
+// MobileGL - MobileGL/MG_IntegrationTest/Scenarios/Program203FirstReductionScenario.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
+// End of Source File Header
+//
+// Scenario - PROGRAM 203'S FIRST SUBGROUP REDUCTION.
+//
+// Program 203 reduces a 32 x 16 exposure tile with a vector subgroup inclusive add,
+// then a shared-memory scan of subgroup totals.  The source assumes that every
+// subgroup has a last lane, that there are 2..32 subgroups, and that local index
+// 511 belongs to the last subgroup and its last lane.  Those are source assumptions,
+// not API contracts.  This probe intentionally does not repair them: it records the
+// observed topology and makes each handoff independently observable.
+
+#include <algorithm>
+#include <array>
+#include <bit>
+#include <cstddef>
+#include <cstdint>
+#include <cstdlib>
+#include <cstring>
+#include <iomanip>
+#include <iostream>
+#include <limits>
+#include <sstream>
+#include <string>
+#include <type_traits>
+#include <utility>
+#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 std::size_t kInvocationCount = 512;
+        constexpr std::size_t kScanStageCount = 6;
+        constexpr std::uint32_t kQuietNanBits = 0x7fc00000u;
+        constexpr std::size_t kNoSlot = std::numeric_limits<std::size_t>::max();
+
+        struct UVec4 {
+            std::uint32_t x;
+            std::uint32_t y;
+            std::uint32_t z;
+            std::uint32_t w;
+        };
+
+        struct Vec4 {
+            float x;
+            float y;
+            float z;
+            float w;
+        };
+
+        // Matches the std430 block exactly. uvec4/vec4 arrays have a 16-byte
+        // stride, floats are a dense scalar array, and the outer scan array is
+        // stage-major in both GLSL and C++.
+        struct ProbeOutput {
+            std::array<UVec4, kInvocationCount> invocation;
+            std::array<UVec4, kInvocationCount> subgroup;
+            std::array<Vec4, kInvocationCount> reduction;
+            std::array<float, kInvocationCount> finalAverage;
+            std::array<std::array<float, kInvocationCount>, kScanStageCount> scanAfter;
+        };
+
+        static_assert(sizeof(UVec4) == 16);
+        static_assert(sizeof(Vec4) == 16);
+        static_assert(std::is_standard_layout_v<ProbeOutput>);
+        static_assert(offsetof(ProbeOutput, invocation) == 0);
+        static_assert(offsetof(ProbeOutput, subgroup) == 8192);
+        static_assert(offsetof(ProbeOutput, reduction) == 16384);
+        static_assert(offsetof(ProbeOutput, finalAverage) == 24576);
+        static_assert(offsetof(ProbeOutput, scanAfter) == 26624);
+        static_assert(sizeof(ProbeOutput) == 38912);
+
+        enum class InputMode {
+            SampledRgba32f,
+            IndexedSsbo,
+        };
+
+        const char* InputModeName(InputMode mode) {
+            return mode == InputMode::SampledRgba32f ? "sampled RGBA32F" : "indexed SSBO";
+        }
+
+        std::uint32_t FloatBits(float value) {
+            return std::bit_cast<std::uint32_t>(value);
+        }
+
+        bool SameBits(float lhs, float rhs) {
+            return FloatBits(lhs) == FloatBits(rhs);
+        }
+
+        bool IsQuietNanSentinel(float value) {
+            return FloatBits(value) == kQuietNanBits;
+        }
+
+        bool DrainGlErrors() {
+            bool hadError = false;
+            while (glGetError() != GL_NO_ERROR) hadError = true;
+            return hadError;
+        }
+
+        bool HasExtension(const char* wanted) {
+            GLint extensionCount = 0;
+            glGetIntegerv(GL_NUM_EXTENSIONS, &extensionCount);
+            for (GLint i = 0; i < extensionCount; ++i) {
+                const auto* extension = reinterpret_cast<const char*>(glGetStringi(GL_EXTENSIONS, static_cast<GLuint>(i)));
+                if (extension != nullptr && std::string(extension) == wanted) return true;
+            }
+            return false;
+        }
+
+        struct CapabilityInfo {
+            bool subgroupExtension = false;
+            GLint subgroupSize = 0;
+            GLint supportedStages = 0;
+            GLint supportedFeatures = 0;
+            GLint maxComputeStorageBlocks = 0;
+            GLint maxStorageBindings = 0;
+            GLint maxWorkGroupInvocations = 0;
+            std::array<GLint, 3> maxWorkGroupSize{};
+            bool queryHadError = false;
+
+            bool SupportsProbe() const {
+                const auto stages = static_cast<GLbitfield>(supportedStages);
+                const auto features = static_cast<GLbitfield>(supportedFeatures);
+                return !queryHadError && subgroupExtension &&
+                       (stages & GL_COMPUTE_SHADER_BIT) != 0 &&
+                       (features & (GL_SUBGROUP_FEATURE_BASIC_BIT_KHR | GL_SUBGROUP_FEATURE_ARITHMETIC_BIT_KHR)) ==
+                           (GL_SUBGROUP_FEATURE_BASIC_BIT_KHR | GL_SUBGROUP_FEATURE_ARITHMETIC_BIT_KHR) &&
+                       maxComputeStorageBlocks >= 2 && maxStorageBindings >= 2 &&
+                       maxWorkGroupInvocations >= static_cast<GLint>(kInvocationCount) && maxWorkGroupSize[0] >= 32 &&
+                       maxWorkGroupSize[1] >= 16 && maxWorkGroupSize[2] >= 1;
+            }
+
+            std::string MissingRequirements() const {
+                std::vector<std::string> missing;
+                const auto stages = static_cast<GLbitfield>(supportedStages);
+                const auto features = static_cast<GLbitfield>(supportedFeatures);
+                if (queryHadError) missing.emplace_back("a subgroup/compute capability query generated GL error");
+                if (!subgroupExtension) missing.emplace_back("GL_KHR_shader_subgroup");
+                if ((stages & GL_COMPUTE_SHADER_BIT) == 0) {
+                    missing.emplace_back("GL_COMPUTE_SHADER_BIT in GL_SUBGROUP_SUPPORTED_STAGES_KHR");
+                }
+                const auto requiredFeatures =
+                    GL_SUBGROUP_FEATURE_BASIC_BIT_KHR | GL_SUBGROUP_FEATURE_ARITHMETIC_BIT_KHR;
+                if ((features & requiredFeatures) != requiredFeatures) {
+                    missing.emplace_back("basic|arithmetic in GL_SUBGROUP_SUPPORTED_FEATURES_KHR");
+                }
+                if (maxComputeStorageBlocks < 2 || maxStorageBindings < 2) {
+                    missing.emplace_back("two compute SSBO bindings");
+                }
+                if (maxWorkGroupInvocations < static_cast<GLint>(kInvocationCount) || maxWorkGroupSize[0] < 32 ||
+                    maxWorkGroupSize[1] < 16 || maxWorkGroupSize[2] < 1) {
+                    missing.emplace_back("a 32x16x1 / 512-invocation compute workgroup");
+                }
+
+                std::ostringstream message;
+                for (std::size_t i = 0; i < missing.size(); ++i) {
+                    if (i != 0) message << ", ";
+                    message << missing[i];
+                }
+                return message.str();
+            }
+        };
+
+        CapabilityInfo QueryCapabilities() {
+            CapabilityInfo info;
+            DrainGlErrors();
+            info.subgroupExtension = HasExtension("GL_KHR_shader_subgroup");
+            glGetIntegerv(GL_SUBGROUP_SIZE_KHR, &info.subgroupSize);
+            glGetIntegerv(GL_SUBGROUP_SUPPORTED_STAGES_KHR, &info.supportedStages);
+            glGetIntegerv(GL_SUBGROUP_SUPPORTED_FEATURES_KHR, &info.supportedFeatures);
+            glGetIntegerv(GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS, &info.maxComputeStorageBlocks);
+            glGetIntegerv(GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS, &info.maxStorageBindings);
+            glGetIntegerv(GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS, &info.maxWorkGroupInvocations);
+            for (GLuint axis = 0; axis < info.maxWorkGroupSize.size(); ++axis) {
+                glGetIntegeri_v(GL_MAX_COMPUTE_WORK_GROUP_SIZE, axis, &info.maxWorkGroupSize[axis]);
+            }
+            info.queryHadError = DrainGlErrors();
+            return info;
+        }
+
+        void PrintMetadata(const CapabilityInfo& info, std::ostream& output) {
+            output << "Program203FirstReductionScenario metadata: "
+                   << "GL_SUBGROUP_SIZE_KHR=" << info.subgroupSize
+                   << ", GL_SUBGROUP_SUPPORTED_STAGES_KHR=0x" << std::hex
+                   << static_cast<GLbitfield>(info.supportedStages)
+                   << ", GL_SUBGROUP_SUPPORTED_FEATURES_KHR=0x"
+                   << static_cast<GLbitfield>(info.supportedFeatures) << std::dec
+                   << ", subgroupExtension=" << info.subgroupExtension
+                   << ", GL_MAX_COMPUTE_SHADER_STORAGE_BLOCKS=" << info.maxComputeStorageBlocks
+                   << ", GL_MAX_SHADER_STORAGE_BUFFER_BINDINGS=" << info.maxStorageBindings
+                   << ", GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS=" << info.maxWorkGroupInvocations
+                   << ", GL_MAX_COMPUTE_WORK_GROUP_SIZE=" << info.maxWorkGroupSize[0] << 'x'
+                   << info.maxWorkGroupSize[1] << 'x' << info.maxWorkGroupSize[2]
+                   << ", queryHadError=" << info.queryHadError << '\n';
+        }
+
+        bool DumpRequested() {
+            const char* value = std::getenv("MOBILEGL_ITEST_SUBGROUP_PROBE_DUMP");
+            return value != nullptr && std::string(value) == "1";
+        }
+
+        constexpr const char* kShaderPreamble = R"(#version 430 core
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_arithmetic : require
+
+layout(local_size_x = 32, local_size_y = 16, local_size_z = 1) in;
+
+layout(std430, binding = 1) buffer SubgroupProbeOutput {
+    uvec4 invocation[512];
+    uvec4 subgroup[512];
+    vec4 reduction[512];
+    float finalAverage[512];
+    float scanAfter[6][512];
+} outProbe;
+
+shared vec2 prefixSumCache[32];
+)";
+
+        constexpr const char* kSampledInput = R"(
+uniform sampler2D colortex2;
+uniform vec2 pixelSize;
+)";
+
+        constexpr const char* kIndexedInput = R"(
+layout(std430, binding = 0) readonly buffer Input {
+    float value[512];
+} inputData;
+)";
+
+        // Only the expression producing tileExposure differs between the two
+        // tests. The remainder is the program-203 first reduction, with stores
+        // placed after its existing barriers to expose each handoff.
+        constexpr const char* kSampledTileExposure = R"(
+    vec2 texCoord = (vec2(gl_GlobalInvocationID.xy) + 0.5) *
+                    vec2(1.0 / 32.0, 1.0 / 16.0);
+    vec2 sampleCoord = texCoord * (1.0 / 64.0);
+    sampleCoord.x += (15.0 / 32.0) + pixelSize.x * 12.0;
+
+    float tileExposure = dot(
+        textureLod(colortex2, sampleCoord, 0.0).rgb,
+        vec3(0.2125, 0.7154, 0.0721));
+)";
+
+        constexpr const char* kIndexedTileExposure = R"(
+    float tileExposure = inputData.value[gl_LocalInvocationIndex];
+)";
+
+        constexpr const char* kReductionBody = R"(
+    vec2 sampleLuminance = vec2(tileExposure, 0.0);
+    sampleLuminance = subgroupInclusiveAdd(sampleLuminance);
+    float nativeInclusive = sampleLuminance.x;
+
+    // This is a uniform, safety-only branch: it leaves an invalid source
+    // contract visible without indexing past the 32-entry cache or underflowing
+    // loopLength - 1. It is deliberately a failure on the CPU, not a skip.
+    bool sourceDomain = gl_NumSubgroups >= 2u && gl_NumSubgroups <= 32u;
+    if (!sourceDomain) {
+        float qNaN = uintBitsToFloat(0x7fc00000u);
+        uint localIndex = gl_LocalInvocationIndex;
+        outProbe.invocation[localIndex] = uvec4(localIndex, gl_LocalInvocationID);
+        outProbe.subgroup[localIndex] = uvec4(gl_SubgroupSize, gl_NumSubgroups, gl_SubgroupID,
+                                               gl_SubgroupInvocationID);
+        outProbe.reduction[localIndex] = vec4(tileExposure, nativeInclusive, qNaN, qNaN);
+        outProbe.finalAverage[localIndex] = qNaN;
+        for (uint stage = 0u; stage < 6u; ++stage)
+            outProbe.scanAfter[stage][localIndex] = qNaN;
+        return;
+    }
+
+    if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+        prefixSumCache[gl_SubgroupID] = sampleLuminance;
+    barrier();
+
+    float sourceRawSubtotal = prefixSumCache[gl_SubgroupID].x;
+
+    uint loopLength = uint(findMSB(gl_NumSubgroups));
+    loopLength += uint(gl_NumSubgroups - (1u << (loopLength - 1u)) > 0u);
+
+    for (uint scanStage = 0u; scanStage < loopLength; ++scanStage) {
+        if ((gl_SubgroupID & (1u << scanStage)) > 0u) {
+            sampleLuminance += prefixSumCache[(gl_SubgroupID >> scanStage << scanStage) - 1u];
+            if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+                prefixSumCache[gl_SubgroupID] = sampleLuminance;
+        }
+        barrier();
+        outProbe.scanAfter[scanStage][gl_LocalInvocationIndex] = sampleLuminance.x;
+    }
+
+    float sourceMergedPrefix = sampleLuminance.x;
+
+    if (gl_LocalInvocationIndex == 511u)
+        prefixSumCache[0] = sampleLuminance / 512.0;
+    barrier();
+
+    float avg = prefixSumCache[0].x;
+
+    uint localIndex = gl_LocalInvocationIndex;
+    outProbe.invocation[localIndex] = uvec4(localIndex, gl_LocalInvocationID);
+    outProbe.subgroup[localIndex] = uvec4(gl_SubgroupSize, gl_NumSubgroups, gl_SubgroupID,
+                                           gl_SubgroupInvocationID);
+    outProbe.reduction[localIndex] = vec4(tileExposure, nativeInclusive, sourceRawSubtotal, sourceMergedPrefix);
+    outProbe.finalAverage[localIndex] = avg;
+}
+)";
+
+        std::string BuildProbeShader(InputMode mode) {
+            std::string source = kShaderPreamble;
+            source += mode == InputMode::SampledRgba32f ? kSampledInput : kIndexedInput;
+            source += "\nvoid main() {\n";
+            source += mode == InputMode::SampledRgba32f ? kSampledTileExposure : kIndexedTileExposure;
+            source += kReductionBody;
+            return source;
+        }
+
+        std::string FormatFloat(float value) {
+            std::ostringstream text;
+            text << std::hexfloat << value;
+            return text.str();
+        }
+
+        struct ValidationResult {
+            bool ok = true;
+            std::string phase;
+            std::string message;
+            bool scanStageMismatch = false;
+            int scanStage = -1;
+            bool ownerEvaluated = false;
+            bool index511IsSourceLastLaneWriter = false;
+            bool index511IsHighestSubgroupMember = false;
+            std::uint32_t highestObservedSubgroup = 0;
+        };
+
+        ValidationResult Failure(std::string phase, std::string message) {
+            ValidationResult result;
+            result.ok = false;
+            result.phase = std::move(phase);
+            result.message = std::move(message);
+            return result;
+        }
+
+        constexpr float kSampledLuminance = 0.2125f + 0.7154f + 0.0721f;
+
+        float ExpectedInput(InputMode mode, std::uint32_t localIndex) {
+            return mode == InputMode::SampledRgba32f ? kSampledLuminance : static_cast<float>(localIndex + 1u);
+        }
+
+        ValidationResult ValidateProbe(const ProbeOutput& output, InputMode mode) {
+            std::array<std::size_t, kInvocationCount> slotForLocal{};
+            slotForLocal.fill(kNoSlot);
+
+            // 1. Record identity. Slots are only used to locate each reported
+            // local index; all subgroup behavior below groups recorded IDs/lanes.
+            for (std::size_t slot = 0; slot < kInvocationCount; ++slot) {
+                const std::uint32_t localIndex = output.invocation[slot].x;
+                if (localIndex >= kInvocationCount) {
+                    std::ostringstream message;
+                    message << "output slot " << slot << " reports localIndex " << localIndex << " outside [0, 511]";
+                    return Failure("record identity", message.str());
+                }
+                if (slotForLocal[localIndex] != kNoSlot) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " appears in output slots " << slotForLocal[localIndex]
+                            << " and " << slot;
+                    return Failure("record identity", message.str());
+                }
+                slotForLocal[localIndex] = slot;
+            }
+            for (std::size_t localIndex = 0; localIndex < kInvocationCount; ++localIndex) {
+                if (slotForLocal[localIndex] == kNoSlot) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " is missing from all 512 records";
+                    return Failure("record identity", message.str());
+                }
+            }
+            for (std::size_t localIndex = 0; localIndex < kInvocationCount; ++localIndex) {
+                const std::size_t slot = slotForLocal[localIndex];
+                const UVec4& invocation = output.invocation[slot];
+                const std::uint32_t expectedX = static_cast<std::uint32_t>(localIndex % 32u);
+                const std::uint32_t expectedY = static_cast<std::uint32_t>(localIndex / 32u);
+                if (invocation.y != expectedX || invocation.z != expectedY || invocation.w != 0u) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " reports local invocation (" << invocation.y << ','
+                            << invocation.z << ',' << invocation.w << "), expected (" << expectedX << ',' << expectedY
+                            << ",0)";
+                    return Failure("record identity", message.str());
+                }
+                const float expectedInput = ExpectedInput(mode, static_cast<std::uint32_t>(localIndex));
+                const float actualInput = output.reduction[slot].x;
+                if (!SameBits(actualInput, expectedInput)) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " input was " << FormatFloat(actualInput) << ", expected "
+                            << FormatFloat(expectedInput);
+                    return Failure("input", message.str());
+                }
+            }
+
+            // 2. Observed topology. Do not derive lanes or subgroup membership
+            // from local invocation indices: only the values the shader recorded
+            // participate in grouping.
+            const std::uint32_t reportedNumSubgroups = output.subgroup[slotForLocal[0]].y;
+            if (reportedNumSubgroups == 0u) {
+                return Failure("observed topology", "localIndex 0 reported gl_NumSubgroups == 0");
+            }
+            if (reportedNumSubgroups > kInvocationCount) {
+                std::ostringstream message;
+                message << "reported gl_NumSubgroups=" << reportedNumSubgroups
+                        << " exceeds the 512 recorded invocations, so at least one subgroup ID is missing";
+                return Failure("observed topology", message.str());
+            }
+            std::vector<std::vector<std::size_t>> subgroupSlots(reportedNumSubgroups);
+            for (std::size_t localIndex = 0; localIndex < kInvocationCount; ++localIndex) {
+                const std::size_t slot = slotForLocal[localIndex];
+                const UVec4& subgroup = output.subgroup[slot];
+                if (subgroup.x == 0u || subgroup.y == 0u) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " reported subgroupSize=" << subgroup.x
+                            << ", numSubgroups=" << subgroup.y;
+                    return Failure("observed topology", message.str());
+                }
+                if (subgroup.y != reportedNumSubgroups) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " reported numSubgroups=" << subgroup.y
+                            << ", while localIndex 0 reported " << reportedNumSubgroups;
+                    return Failure("observed topology", message.str());
+                }
+                if (subgroup.z >= reportedNumSubgroups) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " reported subgroupID=" << subgroup.z
+                            << " outside [0, " << (reportedNumSubgroups - 1u) << ']';
+                    return Failure("observed topology", message.str());
+                }
+                if (subgroup.w >= subgroup.x) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " reported laneID=" << subgroup.w
+                            << " outside its subgroupSize=" << subgroup.x;
+                    return Failure("observed topology", message.str());
+                }
+                subgroupSlots[subgroup.z].push_back(slot);
+            }
+            for (std::uint32_t subgroupID = 0; subgroupID < reportedNumSubgroups; ++subgroupID) {
+                if (subgroupSlots[subgroupID].empty()) {
+                    std::ostringstream message;
+                    message << "reported gl_NumSubgroups=" << reportedNumSubgroups
+                            << " but subgroupID " << subgroupID << " has no recorded members";
+                    return Failure("observed topology", message.str());
+                }
+                auto& members = subgroupSlots[subgroupID];
+                std::sort(members.begin(), members.end(), [&output](std::size_t lhs, std::size_t rhs) {
+                    return output.subgroup[lhs].w < output.subgroup[rhs].w;
+                });
+                for (std::size_t i = 1; i < members.size(); ++i) {
+                    if (output.subgroup[members[i - 1]].w == output.subgroup[members[i]].w) {
+                        std::ostringstream message;
+                        message << "subgroupID " << subgroupID << " contains duplicate laneID "
+                                << output.subgroup[members[i]].w;
+                        return Failure("observed topology", message.str());
+                    }
+                }
+            }
+
+            // 3. Native subgroup arithmetic, in the actual lane ordering emitted
+            // by the driver. The fixture values and all partial sums are exactly
+            // representable binary32 values, so compare representation, not epsilon.
+            std::array<float, kInvocationCount> nativePrefix{};
+            std::vector<float> nativeSubtotal(reportedNumSubgroups, 0.0f);
+            for (std::uint32_t subgroupID = 0; subgroupID < reportedNumSubgroups; ++subgroupID) {
+                float inclusive = 0.0f;
+                for (const std::size_t slot : subgroupSlots[subgroupID]) {
+                    const std::uint32_t localIndex = output.invocation[slot].x;
+                    inclusive += ExpectedInput(mode, localIndex);
+                    nativePrefix[slot] = inclusive;
+                    const float actualNative = output.reduction[slot].y;
+                    if (!SameBits(actualNative, inclusive)) {
+                        std::ostringstream message;
+                        message << "subgroupID " << subgroupID << ", laneID " << output.subgroup[slot].w
+                                << ", localIndex " << localIndex << " nativeInclusive was " << FormatFloat(actualNative)
+                                << ", expected " << FormatFloat(inclusive);
+                        return Failure("native subgroup arithmetic", message.str());
+                    }
+                }
+                nativeSubtotal[subgroupID] = inclusive;
+            }
+
+            // sourceDomain is the narrow source-side safety branch. It is checked
+            // after native arithmetic so an unsupported source topology still
+            // reports native subgroup behavior before failing explicitly.
+            if (reportedNumSubgroups < 2u || reportedNumSubgroups > 32u) {
+                for (std::size_t localIndex = 0; localIndex < kInvocationCount; ++localIndex) {
+                    const std::size_t slot = slotForLocal[localIndex];
+                    const Vec4& reduction = output.reduction[slot];
+                    if (!IsQuietNanSentinel(reduction.z) || !IsQuietNanSentinel(reduction.w) ||
+                        !IsQuietNanSentinel(output.finalAverage[slot])) {
+                        std::ostringstream message;
+                        message << "program 203 source reduction has no valid contract for gl_NumSubgroups="
+                                << reportedNumSubgroups << "; localIndex " << localIndex
+                                << " did not preserve its qNaN source-reduction sentinel";
+                        return Failure("source domain", message.str());
+                    }
+                    for (std::size_t stage = 0; stage < kScanStageCount; ++stage) {
+                        if (!IsQuietNanSentinel(output.scanAfter[stage][slot])) {
+                            std::ostringstream message;
+                            message << "program 203 source reduction has no valid contract for gl_NumSubgroups="
+                                    << reportedNumSubgroups << "; localIndex " << localIndex << ", scan stage " << stage
+                                    << " did not preserve its qNaN source-reduction sentinel";
+                            return Failure("source domain", message.str());
+                        }
+                    }
+                }
+                std::ostringstream message;
+                message << "program 203 source reduction has no valid contract for observed gl_NumSubgroups="
+                        << reportedNumSubgroups << " (requires 2..32); native subgroup results were recorded";
+                return Failure("source domain", message.str());
+            }
+
+            // 4. Program-203 source writer and first shared-memory handoff.
+            std::vector<std::size_t> sourceWriter(reportedNumSubgroups, kNoSlot);
+            for (std::uint32_t subgroupID = 0; subgroupID < reportedNumSubgroups; ++subgroupID) {
+                std::size_t writerCount = 0;
+                for (const std::size_t slot : subgroupSlots[subgroupID]) {
+                    const UVec4& subgroup = output.subgroup[slot];
+                    if (subgroup.w == subgroup.x - 1u) {
+                        sourceWriter[subgroupID] = slot;
+                        ++writerCount;
+                    }
+                }
+                if (writerCount != 1u) {
+                    std::ostringstream message;
+                    message << "subgroupID " << subgroupID << " has " << writerCount
+                            << " recorded lane(s) where laneID == subgroupSize - 1; program 203 leaves that "
+                               "shared-cache entry unwritten";
+                    return Failure("source writer", message.str());
+                }
+                for (const std::size_t slot : subgroupSlots[subgroupID]) {
+                    const float actualRawSubtotal = output.reduction[slot].z;
+                    if (!SameBits(actualRawSubtotal, nativeSubtotal[subgroupID])) {
+                        std::ostringstream message;
+                        message << "subgroupID " << subgroupID << ", localIndex " << output.invocation[slot].x
+                                << " sourceRawSubtotal was " << FormatFloat(actualRawSubtotal) << ", expected "
+                                << FormatFloat(nativeSubtotal[subgroupID]);
+                        return Failure("source raw subtotal", message.str());
+                    }
+                }
+            }
+
+            // 5. Reproduce the source loop exactly, including the redundant final
+            // scan iteration on power-of-two subgroup counts. Reads and writes in
+            // one iteration target disjoint cache entries, so update the cache at
+            // the CPU equivalent of the source barrier.
+            std::array<float, kInvocationCount> mergedPrefix = nativePrefix;
+            std::vector<float> cache = nativeSubtotal;
+            std::uint32_t loopLength = std::bit_width(reportedNumSubgroups) - 1u;
+            loopLength +=
+                static_cast<std::uint32_t>(reportedNumSubgroups - (1u << (loopLength - 1u)) > 0u);
+            for (std::uint32_t scanStage = 0u; scanStage < loopLength; ++scanStage) {
+                std::vector<float> cacheAfterStage = cache;
+                for (std::uint32_t subgroupID = 0; subgroupID < reportedNumSubgroups; ++subgroupID) {
+                    if ((subgroupID & (1u << scanStage)) == 0u) continue;
+                    const std::uint32_t sourceCacheIndex = (subgroupID >> scanStage << scanStage) - 1u;
+                    const float sourcePrefix = cache[sourceCacheIndex];
+                    for (const std::size_t slot : subgroupSlots[subgroupID]) {
+                        mergedPrefix[slot] += sourcePrefix;
+                    }
+                    cacheAfterStage[subgroupID] = mergedPrefix[sourceWriter[subgroupID]];
+                }
+                cache.swap(cacheAfterStage);
+                for (std::size_t localIndex = 0; localIndex < kInvocationCount; ++localIndex) {
+                    const std::size_t slot = slotForLocal[localIndex];
+                    const float actualAfterStage = output.scanAfter[scanStage][slot];
+                    if (!SameBits(actualAfterStage, mergedPrefix[slot])) {
+                        std::ostringstream message;
+                        message << "scanStage " << scanStage << ", subgroupID " << output.subgroup[slot].z
+                                << ", laneID " << output.subgroup[slot].w << ", localIndex " << localIndex
+                                << " scanAfter was " << FormatFloat(actualAfterStage) << ", expected "
+                                << FormatFloat(mergedPrefix[slot]);
+                        ValidationResult result = Failure("source scan", message.str());
+                        result.scanStageMismatch = true;
+                        result.scanStage = static_cast<int>(scanStage);
+                        return result;
+                    }
+                }
+            }
+            for (std::size_t localIndex = 0; localIndex < kInvocationCount; ++localIndex) {
+                const std::size_t slot = slotForLocal[localIndex];
+                const float actualMergedPrefix = output.reduction[slot].w;
+                if (!SameBits(actualMergedPrefix, mergedPrefix[slot])) {
+                    std::ostringstream message;
+                    message << "localIndex " << localIndex << " sourceMergedPrefix was "
+                            << FormatFloat(actualMergedPrefix) << ", expected " << FormatFloat(mergedPrefix[slot]);
+                    return Failure("source scan", message.str());
+                }
+            }
+
+            // 6. Final owner and average. The uniformity check is intentionally
+            // separate from the source's topology contract at local index 511.
+            const float firstAverage = output.finalAverage[slotForLocal[0]];
+            for (std::size_t localIndex = 1; localIndex < kInvocationCount; ++localIndex) {
+                const float actualAverage = output.finalAverage[slotForLocal[localIndex]];
+                if (!SameBits(actualAverage, firstAverage)) {
+                    std::ostringstream message;
+                    message << "finalAverage differs: localIndex 0 has " << FormatFloat(firstAverage)
+                            << ", localIndex " << localIndex << " has " << FormatFloat(actualAverage);
+                    return Failure("final average", message.str());
+                }
+            }
+
+            ValidationResult ownerResult;
+            ownerResult.ownerEvaluated = true;
+            for (std::uint32_t subgroupID = 0; subgroupID < reportedNumSubgroups; ++subgroupID) {
+                if (!subgroupSlots[subgroupID].empty()) {
+                    ownerResult.highestObservedSubgroup = std::max(ownerResult.highestObservedSubgroup, subgroupID);
+                }
+            }
+            const std::size_t index511Slot = slotForLocal[kInvocationCount - 1u];
+            const UVec4& index511Subgroup = output.subgroup[index511Slot];
+            ownerResult.index511IsSourceLastLaneWriter =
+                index511Subgroup.w == index511Subgroup.x - 1u;
+            ownerResult.index511IsHighestSubgroupMember =
+                index511Subgroup.z == ownerResult.highestObservedSubgroup;
+            if (!ownerResult.index511IsSourceLastLaneWriter || !ownerResult.index511IsHighestSubgroupMember) {
+                std::ostringstream message;
+                message << "program 203 topology incompatibility: localIndex 511 is sourceLastLaneWriter="
+                        << ownerResult.index511IsSourceLastLaneWriter << ", highestSubgroupMember="
+                        << ownerResult.index511IsHighestSubgroupMember << " (subgroupID=" << index511Subgroup.z
+                        << ", highest observed subgroupID=" << ownerResult.highestObservedSubgroup << ')';
+                ownerResult.ok = false;
+                ownerResult.phase = "final average";
+                ownerResult.message = message.str();
+                return ownerResult;
+            }
+
+            float total = 0.0f;
+            for (const float subtotal : nativeSubtotal) total += subtotal;
+            float sampledExpectedTotal = 0.0f;
+            for (std::size_t i = 0; i < kInvocationCount; ++i) sampledExpectedTotal += kSampledLuminance;
+            const float expectedTotal = mode == InputMode::IndexedSsbo ? 131328.0f : sampledExpectedTotal;
+            if (!SameBits(total, expectedTotal) || !SameBits(mergedPrefix[index511Slot], expectedTotal)) {
+                std::ostringstream message;
+                message << "program 203 source total was " << FormatFloat(mergedPrefix[index511Slot])
+                        << " (native total " << FormatFloat(total) << "), expected " << FormatFloat(expectedTotal);
+                ownerResult.ok = false;
+                ownerResult.phase = "final average";
+                ownerResult.message = message.str();
+                return ownerResult;
+            }
+
+            const float expectedAverage = mode == InputMode::IndexedSsbo ? 256.5f : sampledExpectedTotal / 512.0f;
+            if (!SameBits(firstAverage, expectedAverage)) {
+                std::ostringstream message;
+                message << "finalAverage was " << FormatFloat(firstAverage) << ", expected "
+                        << FormatFloat(expectedAverage);
+                ownerResult.ok = false;
+                ownerResult.phase = "final average";
+                ownerResult.message = message.str();
+                return ownerResult;
+            }
+            return ownerResult;
+        }
+
+        void DumpProbe(const ProbeOutput& output, const CapabilityInfo& capabilities, const ValidationResult& validation,
+                       bool includeScanStages) {
+            PrintMetadata(capabilities, std::cout);
+            if (validation.ok) {
+                std::cout << "Program203FirstReductionScenario firstFailure=none\n";
+            } else {
+                std::cout << "Program203FirstReductionScenario firstFailure=" << validation.phase << ": "
+                          << validation.message << '\n';
+            }
+            std::cout << "localIndex,localX,localY,localZ,subgroupSize,numSubgroups,subgroupID,laneID,input,"
+                         "nativeInclusive,subgroupSubtotal,mergedPrefix,finalAverage\n";
+            for (std::size_t slot = 0; slot < kInvocationCount; ++slot) {
+                const UVec4& invocation = output.invocation[slot];
+                const UVec4& subgroup = output.subgroup[slot];
+                const Vec4& reduction = output.reduction[slot];
+                std::cout << invocation.x << ',' << invocation.y << ',' << invocation.z << ',' << invocation.w << ','
+                          << subgroup.x << ',' << subgroup.y << ',' << subgroup.z << ',' << subgroup.w << ','
+                          << std::hexfloat << reduction.x << ',' << reduction.y << ',' << reduction.z << ','
+                          << reduction.w << ',' << output.finalAverage[slot] << std::defaultfloat << '\n';
+            }
+            if (includeScanStages) {
+                std::cout << "scanStage,localIndex,scanAfter\n";
+                for (std::size_t scanStage = 0; scanStage < kScanStageCount; ++scanStage) {
+                    for (std::size_t slot = 0; slot < kInvocationCount; ++slot) {
+                        std::cout << scanStage << ',' << output.invocation[slot].x << ',' << std::hexfloat
+                                  << output.scanAfter[scanStage][slot] << std::defaultfloat << '\n';
+                    }
+                }
+            }
+        }
+
+        class Program203FirstReductionScenario : public ScenarioTest {
+        protected:
+            void SetUp() override {
+                ScenarioTest::SetUp();
+                if (!Ready()) return;
+
+                m_capabilities = QueryCapabilities();
+                // GL_SUBGROUP_SIZE_KHR is diagnostic only. It is deliberately
+                // never used to infer lane placement or an expected group count.
+                PrintMetadata(m_capabilities, std::cout);
+                RecordProperty("program203_gl_subgroup_size_khr", std::to_string(m_capabilities.subgroupSize));
+                if (!m_capabilities.SupportsProbe()) {
+                    GTEST_SKIP() << "subgroup probe requires " << m_capabilities.MissingRequirements();
+                }
+            }
+
+            void TearDown() override {
+                if (!Ready()) return;
+                glUseProgram(0);
+                glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, 0);
+                glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, 0);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, 0);
+                glActiveTexture(GL_TEXTURE3);
+                glBindTexture(GL_TEXTURE_2D, 0);
+                glActiveTexture(GL_TEXTURE0);
+                if (m_texture != 0) glDeleteTextures(1, &m_texture);
+                if (m_inputBuffer != 0) glDeleteBuffers(1, &m_inputBuffer);
+                if (m_outputBuffer != 0) glDeleteBuffers(1, &m_outputBuffer);
+                if (m_program != 0) glDeleteProgram(m_program);
+                m_texture = 0;
+                m_inputBuffer = 0;
+                m_outputBuffer = 0;
+                m_program = 0;
+            }
+
+            GLuint CompileComputeProgram(const std::string& source, std::string* outError) {
+                const char* text = source.c_str();
+                const GLuint shader = glCreateShader(GL_COMPUTE_SHADER);
+                if (shader == 0) {
+                    *outError = "glCreateShader(GL_COMPUTE_SHADER) returned 0";
+                    return 0;
+                }
+                glShaderSource(shader, 1, &text, nullptr);
+                glCompileShader(shader);
+                GLint compiled = GL_FALSE;
+                glGetShaderiv(shader, GL_COMPILE_STATUS, &compiled);
+                if (compiled == GL_FALSE) {
+                    char log[8192] = {};
+                    glGetShaderInfoLog(shader, sizeof(log) - 1, nullptr, log);
+                    *outError = std::string("the subgroup probe compute shader did not compile: ") + log;
+                    glDeleteShader(shader);
+                    return 0;
+                }
+                const GLuint program = glCreateProgram();
+                glAttachShader(program, shader);
+                glLinkProgram(program);
+                glDeleteShader(shader);
+                GLint linked = GL_FALSE;
+                glGetProgramiv(program, GL_LINK_STATUS, &linked);
+                if (linked == GL_FALSE) {
+                    char log[8192] = {};
+                    glGetProgramInfoLog(program, sizeof(log) - 1, nullptr, log);
+                    *outError = std::string("the subgroup probe compute program did not link: ") + log;
+                    glDeleteProgram(program);
+                    return 0;
+                }
+                return program;
+            }
+
+            bool RunProbe(InputMode mode, ProbeOutput* output, std::string* outError) {
+                m_program = CompileComputeProgram(BuildProbeShader(mode), outError);
+                if (m_program == 0) return false;
+
+                ProbeOutput poison{};
+                std::memset(&poison, 0xa5, sizeof(poison));
+                glGenBuffers(1, &m_outputBuffer);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_outputBuffer);
+                glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(ProbeOutput), &poison, GL_DYNAMIC_COPY);
+                glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, m_outputBuffer);
+
+                if (mode == InputMode::IndexedSsbo) {
+                    std::array<float, kInvocationCount> values{};
+                    for (std::size_t i = 0; i < values.size(); ++i) values[i] = static_cast<float>(i + 1u);
+                    glGenBuffers(1, &m_inputBuffer);
+                    glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_inputBuffer);
+                    glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(values), values.data(), GL_STATIC_DRAW);
+                    glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, m_inputBuffer);
+                } else {
+                    constexpr std::array<float, 4> kOneTexel = {1.0f, 1.0f, 1.0f, 1.0f};
+                    glGenTextures(1, &m_texture);
+                    glActiveTexture(GL_TEXTURE3);
+                    glBindTexture(GL_TEXTURE_2D, m_texture);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+                    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+                    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA32F, 1, 1, 0, GL_RGBA, GL_FLOAT, kOneTexel.data());
+                }
+
+                if (const GLenum error = FirstGLError(); error != GL_NO_ERROR) {
+                    std::ostringstream message;
+                    message << "subgroup probe resource setup left " << GLErrorName(error);
+                    *outError = message.str();
+                    return false;
+                }
+
+                glUseProgram(m_program);
+                if (mode == InputMode::SampledRgba32f) {
+                    const GLint sampler = glGetUniformLocation(m_program, "colortex2");
+                    const GLint pixelSize = glGetUniformLocation(m_program, "pixelSize");
+                    if (sampler == -1 || pixelSize == -1) {
+                        *outError = "the sampled probe uniforms were optimized away or not reflected";
+                        return false;
+                    }
+                    glUniform1i(sampler, 3);
+                    glUniform2f(pixelSize, 1.0f / 854.0f, 1.0f / 480.0f);
+                }
+                glDispatchCompute(1, 1, 1);
+                glMemoryBarrier(GL_ALL_BARRIER_BITS);
+                glBindBuffer(GL_SHADER_STORAGE_BUFFER, m_outputBuffer);
+                glGetBufferSubData(GL_SHADER_STORAGE_BUFFER, 0, sizeof(ProbeOutput), output);
+                if (const GLenum error = FirstGLError(); error != GL_NO_ERROR) {
+                    std::ostringstream message;
+                    message << "subgroup probe dispatch/readback left " << GLErrorName(error);
+                    *outError = message.str();
+                    return false;
+                }
+                return true;
+            }
+
+            void RunAndValidate(InputMode mode) {
+                ProbeOutput output{};
+                std::string error;
+                ASSERT_TRUE(RunProbe(mode, &output, &error)) << InputModeName(mode) << ": " << error;
+
+                const ValidationResult validation = ValidateProbe(output, mode);
+                if (validation.ownerEvaluated) {
+                    RecordProperty("program203_index511_source_last_lane_writer",
+                                   validation.index511IsSourceLastLaneWriter ? "true" : "false");
+                    RecordProperty("program203_index511_highest_subgroup_member",
+                                   validation.index511IsHighestSubgroupMember ? "true" : "false");
+                    RecordProperty("program203_highest_observed_subgroup",
+                                   std::to_string(validation.highestObservedSubgroup));
+                    std::cout << "Program203FirstReductionScenario owner: localIndex511 sourceLastLaneWriter="
+                              << validation.index511IsSourceLastLaneWriter << ", highestSubgroupMember="
+                              << validation.index511IsHighestSubgroupMember << ", highestObservedSubgroup="
+                              << validation.highestObservedSubgroup << '\n';
+                }
+                if (!validation.ok || DumpRequested()) {
+                    DumpProbe(output, m_capabilities, validation, validation.scanStageMismatch || DumpRequested());
+                }
+                EXPECT_TRUE(validation.ok) << validation.phase << ": " << validation.message;
+            }
+
+            CapabilityInfo m_capabilities;
+            GLuint m_program = 0;
+            GLuint m_inputBuffer = 0;
+            GLuint m_outputBuffer = 0;
+            GLuint m_texture = 0;
+        };
+
+    } // namespace
+
+    TEST_F(Program203FirstReductionScenario, SampledRgba32fFirstAverage) {
+        if (!Ready() || IsSkipped()) return;
+        RunAndValidate(InputMode::SampledRgba32f);
+    }
+
+    TEST_F(Program203FirstReductionScenario, IndexedInputTopologyAndReduction) {
+        if (!Ready() || IsSkipped()) return;
+        RunAndValidate(InputMode::IndexedSsbo);
+    }
+
+} // namespace MGITest

MobileGL/MG_IntegrationTest/Scenarios/ViewportArrayScenario.cpp

@@ -428,15 +428,15 @@ void main() { fragColor = vec4(float(gsIndex) * 16.0 / 255.0, 0.0, 0.0, 1.0); }
             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;
+                const float nearDepth = static_cast<float>(i) / 16.0f;
+                const float farDepth = 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)
+                EXPECT_NEAR(pixels[i], nearDepth, 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)
+                EXPECT_NEAR(pixels[static_cast<size_t>(kWidth) + i], farDepth, 1.0e-3f)
                     << "viewport " << i << " far-plane depth; got viewport 0's range if this is 1";
             }
 

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.

MobileGL/MG_State/GLState/Core.h

@@ -413,9 +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_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

@@ -78,6 +78,7 @@ add_subdirectory(Query)
 add_subdirectory(Pipeline)
 add_subdirectory(ShaderTranspiler)
 add_subdirectory(Util)
+add_subdirectory(SelfTest)
 # The DirectGLES post-transpile ESSL passes are pure String -> String, so unlike the
 # DirectVulkan suite below this one needs no device and always builds.
 add_subdirectory(Backend/DirectGLES)

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;
@@ -2316,3 +2387,13 @@ TEST(DirectGLESTextureSync, UnitMemoRefusesToDriveATwinFromAnotherTexture) {
 
     MG_Impl::GLImpl::BindTexture(GL_TEXTURE_2D, 0);
 }
+
+TEST(DirectVulkanSanity, GraphicsSamplerFeedbackOnlyAliasesWritableOverlappingMip) {
+    using MobileGL::MG_Backend::DirectVulkan::UniformManager;
+
+    EXPECT_TRUE(UniformManager::SamplerOverlapsWritableImageSubresource(1, 3, 2, GL_WRITE_ONLY));
+    EXPECT_TRUE(UniformManager::SamplerOverlapsWritableImageSubresource(1, 3, 3, GL_READ_WRITE));
+    EXPECT_FALSE(UniformManager::SamplerOverlapsWritableImageSubresource(1, 3, 2, GL_READ_ONLY));
+    EXPECT_FALSE(UniformManager::SamplerOverlapsWritableImageSubresource(1, 3, 0, GL_WRITE_ONLY));
+    EXPECT_FALSE(UniformManager::SamplerOverlapsWritableImageSubresource(1, 3, 4, GL_WRITE_ONLY));
+}

MobileGL/MG_Test/SelfTest/CMakeLists.txt

@@ -0,0 +1,19 @@
+# MobileGL - MobileGL/MG_Test/SelfTest/CMakeLists.txt
+
+add_executable(
+        DriverPostProgram203WitnessTest
+        DriverPostProgram203WitnessTest.cpp
+)
+
+target_include_directories(DriverPostProgram203WitnessTest PRIVATE
+        ${MGL_ROOT}/include
+        ${MGL_ROOT}/MobileGL
+)
+
+target_link_libraries(DriverPostProgram203WitnessTest PRIVATE
+        GTest::gtest_main
+        ${LINK_LIBRARIES}
+)
+
+include(GoogleTest)
+gtest_discover_tests(DriverPostProgram203WitnessTest DISCOVERY_TIMEOUT 30 PROPERTIES LABELS unit)

MobileGL/MG_Test/SelfTest/DriverPostProgram203WitnessTest.cpp

@@ -0,0 +1,188 @@
+// MobileGL - MobileGL/MG_Test/SelfTest/DriverPostProgram203WitnessTest.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
+// End of Source File Header
+
+#include <gtest/gtest.h>
+
+#include <string>
+
+#include "MG_Util/SelfTest/DriverPostProgram203Witness.h"
+
+namespace MobileGL::MG_Util::SelfTest {
+    namespace {
+        Program203WitnessOutput MakeValidWitness(std::uint32_t numSubgroups) {
+            Program203WitnessOutput output{};
+            output.magic = kProgram203WitnessMagic;
+            output.numSubgroups = numSubgroups;
+            output.loopLength = ComputeProgram203WitnessLoopLength(numSubgroups);
+            output.seenSubgroupMask =
+                numSubgroups == kProgram203WitnessMaxSubgroups ? 0xffffffffu : (1u << numSubgroups) - 1u;
+
+            // Valid test layouts use equal contiguous groups of the indexed
+            // 1..512 input. The compact witness only needs their independent sums.
+            const std::uint32_t subgroupSize = kProgram203WitnessInvocationCount / numSubgroups;
+            for (std::uint32_t subgroup = 0u; subgroup < numSubgroups; ++subgroup) {
+                const std::uint32_t first = subgroup * subgroupSize + 1u;
+                const std::uint32_t last = first + subgroupSize - 1u;
+                output.lastLaneWriterCount[subgroup] = 1u;
+                output.indexedInputTotal[subgroup] = subgroupSize * (first + last) / 2u;
+                output.rawPrefix[subgroup] = {static_cast<float>(output.indexedInputTotal[subgroup]), 0.0f};
+            }
+            output.owner511 = {subgroupSize, numSubgroups, numSubgroups - 1u, subgroupSize - 1u};
+
+            auto cache = output.rawPrefix;
+            for (std::uint32_t scanStage = 0u; scanStage < output.loopLength; ++scanStage) {
+                auto cacheAfterStage = cache;
+                for (std::uint32_t subgroup = 0u; subgroup < numSubgroups; ++subgroup) {
+                    if ((subgroup & (1u << scanStage)) == 0u) continue;
+                    const std::uint32_t sourceCacheIndex = (subgroup >> scanStage << scanStage) - 1u;
+                    cacheAfterStage[subgroup].x += cache[sourceCacheIndex].x;
+                    cacheAfterStage[subgroup].y += cache[sourceCacheIndex].y;
+                }
+                cache = cacheAfterStage;
+                output.scanCache[scanStage] = cache;
+            }
+            output.finalAverage = {256.5f, 0.0f};
+            return output;
+        }
+
+        Program203WitnessLimits MakeSufficientLimits() {
+            Program203WitnessLimits limits;
+            limits.computeStageSupported = true;
+            limits.basicSubgroupSupported = true;
+            limits.arithmeticSubgroupSupported = true;
+            limits.subgroupSize = 32u;
+            limits.maxComputeWorkGroupInvocations = kProgram203WitnessInvocationCount;
+            limits.maxComputeWorkGroupSize = {32u, 16u, 1u};
+            limits.maxComputeSharedMemorySize = kProgram203WitnessSharedMemoryBytes;
+            limits.maxPerStageDescriptorStorageBuffers = 1u;
+            limits.maxDescriptorSetStorageBuffers = 1u;
+            limits.maxBoundDescriptorSets = 1u;
+            limits.maxStorageBufferRange = sizeof(Program203WitnessOutput);
+            return limits;
+        }
+    } // namespace
+
+    TEST(DriverPostProgram203WitnessTest, ValidTwoSubgroupWitness) {
+        const Program203WitnessValidationResult validation = ValidateProgram203Witness(MakeValidWitness(2u));
+        ASSERT_TRUE(validation.ok) << validation.detail;
+        EXPECT_EQ(validation.detail, "N=2, owner511=id1/lane255, 2 scan stages, average=(256.5,0)");
+    }
+
+    TEST(DriverPostProgram203WitnessTest, ValidThirtyTwoSubgroupWitness) {
+        const Program203WitnessValidationResult validation = ValidateProgram203Witness(MakeValidWitness(32u));
+        ASSERT_TRUE(validation.ok) << validation.detail;
+        EXPECT_EQ(validation.detail, "N=32, owner511=id31/lane15, 6 scan stages, average=(256.5,0)");
+    }
+
+    TEST(DriverPostProgram203WitnessTest, RejectsNonuniformNumSubgroups) {
+        Program203WitnessOutput output = MakeValidWitness(16u);
+        output.topologyFlags |= Program203WitnessNonuniformNumSubgroups;
+        const Program203WitnessValidationResult validation = ValidateProgram203Witness(output);
+        EXPECT_FALSE(validation.ok);
+        EXPECT_EQ(validation.failure, Program203WitnessValidationFailure::Topology);
+        EXPECT_NE(validation.detail.find("gl_NumSubgroups differed"), std::string::npos);
+    }
+
+    TEST(DriverPostProgram203WitnessTest, RejectsMissingAndOutOfRangeSubgroupIds) {
+        Program203WitnessOutput missing = MakeValidWitness(16u);
+        missing.seenSubgroupMask &= ~(1u << 7u);
+        Program203WitnessValidationResult validation = ValidateProgram203Witness(missing);
+        EXPECT_FALSE(validation.ok);
+        EXPECT_NE(validation.detail.find("seen subgroup-ID mask"), std::string::npos);
+
+        Program203WitnessOutput outOfRange = MakeValidWitness(16u);
+        outOfRange.topologyFlags |= Program203WitnessInvalidSubgroupId;
+        validation = ValidateProgram203Witness(outOfRange);
+        EXPECT_FALSE(validation.ok);
+        EXPECT_NE(validation.detail.find("invalid gl_SubgroupID"), std::string::npos);
+    }
+
+    TEST(DriverPostProgram203WitnessTest, RejectsInvalidMultipleAndMissingLastLaneWriters) {
+        Program203WitnessOutput invalidLane = MakeValidWitness(16u);
+        invalidLane.topologyFlags |= Program203WitnessInvalidSubgroupLane;
+        Program203WitnessValidationResult validation = ValidateProgram203Witness(invalidLane);
+        EXPECT_FALSE(validation.ok);
+        EXPECT_NE(validation.detail.find("invalid subgroup lane"), std::string::npos);
+
+        Program203WitnessOutput multiple = MakeValidWitness(16u);
+        multiple.lastLaneWriterCount[4] = 2u;
+        validation = ValidateProgram203Witness(multiple);
+        EXPECT_FALSE(validation.ok);
+        EXPECT_NE(validation.detail.find("subgroup 4 has 2 source last-lane writers"), std::string::npos);
+
+        Program203WitnessOutput missing = MakeValidWitness(16u);
+        missing.lastLaneWriterCount[6] = 0u;
+        validation = ValidateProgram203Witness(missing);
+        EXPECT_FALSE(validation.ok);
+        EXPECT_NE(validation.detail.find("subgroup 6 has 0 source last-lane writers"), std::string::npos);
+    }
+
+    TEST(DriverPostProgram203WitnessTest, ReportsEarliestCorruptSourceScanStage) {
+        Program203WitnessOutput output = MakeValidWitness(32u);
+        output.scanCache[0][1].x += 1.0f;
+        output.scanCache[3][5].x += 1.0f;
+        Program203WitnessValidationResult validation = ValidateProgram203Witness(output);
+        EXPECT_FALSE(validation.ok);
+        EXPECT_EQ(validation.failure, Program203WitnessValidationFailure::SourceScan);
+        EXPECT_EQ(validation.scanStage, 0u);
+        EXPECT_NE(validation.detail.find("source scan stage 0, subgroup 1"), std::string::npos);
+
+        output = MakeValidWitness(32u);
+        output.scanCache[3][5].x += 1.0f;
+        validation = ValidateProgram203Witness(output);
+        EXPECT_FALSE(validation.ok);
+        EXPECT_EQ(validation.failure, Program203WitnessValidationFailure::SourceScan);
+        EXPECT_EQ(validation.scanStage, 3u);
+        EXPECT_NE(validation.detail.find("source scan stage 3, subgroup 5"), std::string::npos);
+    }
+
+    TEST(DriverPostProgram203WitnessTest, RejectsOwner511OutsideHighestFinalLane) {
+        Program203WitnessOutput output = MakeValidWitness(16u);
+        output.owner511.z = 14u;
+        const Program203WitnessValidationResult validation = ValidateProgram203Witness(output);
+        EXPECT_FALSE(validation.ok);
+        EXPECT_EQ(validation.failure, Program203WitnessValidationFailure::FinalOwner);
+        EXPECT_NE(validation.detail.find("not in the highest subgroup"), std::string::npos);
+    }
+
+    TEST(DriverPostProgram203WitnessTest, RejectsIncorrectVectorFinalAverage) {
+        Program203WitnessOutput output = MakeValidWitness(16u);
+        output.finalAverage.y = 1.0f;
+        const Program203WitnessValidationResult validation = ValidateProgram203Witness(output);
+        EXPECT_FALSE(validation.ok);
+        EXPECT_EQ(validation.failure, Program203WitnessValidationFailure::FinalAverage);
+        EXPECT_NE(validation.detail.find("final average"), std::string::npos);
+    }
+
+    TEST(DriverPostProgram203WitnessTest, MissingNativeFeatureIsTheOnlySkipCondition) {
+        for (const auto toggleMissingFeature : {0u, 1u, 2u}) {
+            Program203WitnessLimits limits = MakeSufficientLimits();
+            if (toggleMissingFeature == 0u) limits.computeStageSupported = false;
+            if (toggleMissingFeature == 1u) limits.basicSubgroupSupported = false;
+            if (toggleMissingFeature == 2u) limits.arithmeticSubgroupSupported = false;
+            const Program203WitnessEligibilityResult eligibility = EvaluateProgram203WitnessEligibility(limits);
+            EXPECT_EQ(eligibility.eligibility, Program203WitnessEligibility::SkipUnsupportedNativeFeatureSet)
+                << eligibility.detail;
+        }
+
+        Program203WitnessLimits zeroSubgroupSize = MakeSufficientLimits();
+        zeroSubgroupSize.subgroupSize = 0u;
+        Program203WitnessEligibilityResult eligibility = EvaluateProgram203WitnessEligibility(zeroSubgroupSize);
+        EXPECT_EQ(eligibility.eligibility, Program203WitnessEligibility::FailInadequateLimits) << eligibility.detail;
+
+        Program203WitnessLimits limits = MakeSufficientLimits();
+        limits.maxComputeWorkGroupInvocations = 511u;
+        eligibility = EvaluateProgram203WitnessEligibility(limits);
+        EXPECT_EQ(eligibility.eligibility, Program203WitnessEligibility::FailInadequateLimits) << eligibility.detail;
+
+        limits = MakeSufficientLimits();
+        limits.maxStorageBufferRange = sizeof(Program203WitnessOutput) - 1u;
+        eligibility = EvaluateProgram203WitnessEligibility(limits);
+        EXPECT_EQ(eligibility.eligibility, Program203WitnessEligibility::FailInadequateLimits) << eligibility.detail;
+    }
+} // namespace MobileGL::MG_Util::SelfTest

MobileGL/MG_Test/ShaderTranspiler/CMakeLists.txt

@@ -3,6 +3,7 @@ cmake_minimum_required(VERSION 3.14)
 add_executable(
         SpirvPassTest
         SpirvPassTest.cpp
+        DeriveNumSubgroupsTest.cpp
         DemoteFloat64Test.cpp
         FlattenXfbInterfaceBlocksTest.cpp
 )

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/DeriveNumSubgroupsTest.cpp

@@ -0,0 +1,147 @@
+// MobileGL - MobileGL/MG_Test/ShaderTranspiler/DeriveNumSubgroupsTest.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
+// End of Source File Header
+
+#include <gtest/gtest.h>
+
+#define SPV_ENABLE_UTILITY_CODE
+#include "glslang/SPIRV/spirv.hpp11"
+#undef SPV_ENABLE_UTILITY_CODE
+
+#include "Includes.h"
+#include <MG_Util/ShaderTranspiler/ShaderCompiler.h>
+#include <MG_Util/ShaderTranspiler/Types.h>
+
+#include <spirv-tools/libspirv.hpp>
+
+using namespace MobileGL;
+using MobileGL::MG_Util::ShaderTranspiler::ShaderCompiler;
+
+namespace {
+    constexpr SizeT kSpirvHeaderWordCount = 5u;
+
+    template <typename Visitor>
+    void ForEachInstruction(const Vector<Uint32>& spirv, Visitor&& visit) {
+        for (SizeT offset = kSpirvHeaderWordCount; offset < spirv.size();) {
+            const Uint32 wordCount = spirv[offset] >> 16u;
+            if (wordCount == 0u || offset + wordCount > spirv.size()) break;
+            visit(static_cast<spv::Op>(spirv[offset] & 0xffffu), &spirv[offset], wordCount);
+            offset += wordCount;
+        }
+    }
+
+    Vector<Uint32> CompileCompute(const String& source) {
+        using namespace MobileGL::MG_Util::ShaderTranspiler;
+        ShaderAttrib shaderAttrib{.shaderType = GL_COMPUTE_SHADER, .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 = {GL_COMPUTE_SHADER}, .program = *programResult.value()};
+        auto binaryResult = ShaderCompiler::GetSpirvBinaryFromProgram(binaryAttrib);
+        EXPECT_TRUE(binaryResult) << (binaryResult ? String{} : binaryResult.error().log);
+        if (!binaryResult || binaryResult->empty()) return {};
+        return binaryResult->front();
+    }
+
+    Uint32 FindBuiltinTarget(const Vector<Uint32>& spirv, spv::BuiltIn builtin) {
+        Uint32 target = 0u;
+        ForEachInstruction(spirv, [&](spv::Op opcode, const Uint32* words, Uint32 wordCount) {
+            if (opcode == spv::Op::OpDecorate && wordCount >= 4u &&
+                static_cast<spv::Decoration>(words[2]) == spv::Decoration::BuiltIn &&
+                static_cast<spv::BuiltIn>(words[3]) == builtin) {
+                target = words[1];
+            }
+        });
+        return target;
+    }
+
+    Uint32 CountLoadsFrom(const Vector<Uint32>& spirv, Uint32 pointerId) {
+        Uint32 count = 0u;
+        ForEachInstruction(spirv, [&](spv::Op opcode, const Uint32* words, Uint32 wordCount) {
+            if (opcode == spv::Op::OpLoad && wordCount >= 4u && words[3] == pointerId) ++count;
+        });
+        return count;
+    }
+
+    Uint32 CountOpcode(const Vector<Uint32>& spirv, spv::Op wanted) {
+        Uint32 count = 0u;
+        ForEachInstruction(spirv, [&](spv::Op opcode, const Uint32*, Uint32) {
+            if (opcode == wanted) ++count;
+        });
+        return count;
+    }
+
+    bool Validates(const Vector<Uint32>& spirv) {
+        spvtools::SpirvTools tools(SPV_ENV_VULKAN_1_1);
+        return tools.Validate(spirv);
+    }
+
+    constexpr const char* kNumSubgroupsOnlySource = R"(#version 450 core
+#extension GL_KHR_shader_subgroup_basic : require
+layout(local_size_x = 32, local_size_y = 16, local_size_z = 1) in;
+layout(std430, binding = 0) buffer Output { uint value; } outputData;
+void main() {
+    if (gl_LocalInvocationIndex == 0u)
+        outputData.value = gl_NumSubgroups;
+}
+)";
+
+    constexpr const char* kNoNumSubgroupsSource = R"(#version 450 core
+layout(local_size_x = 32, local_size_y = 16, local_size_z = 1) in;
+layout(std430, binding = 0) buffer Output { uint value; } outputData;
+void main() {
+    if (gl_LocalInvocationIndex == 0u)
+        outputData.value = gl_WorkGroupSize.x;
+}
+)";
+} // namespace
+
+TEST(DeriveNumSubgroupsPass, ReplacesBuiltinLoadAndSynthesizesSubgroupSize) {
+    const Vector<Uint32> input = CompileCompute(kNumSubgroupsOnlySource);
+    ASSERT_FALSE(input.empty());
+    const Uint32 inputNumSubgroups = FindBuiltinTarget(input, spv::BuiltIn::NumSubgroups);
+    ASSERT_NE(inputNumSubgroups, 0u);
+    EXPECT_EQ(CountLoadsFrom(input, inputNumSubgroups), 1u);
+    EXPECT_EQ(FindBuiltinTarget(input, spv::BuiltIn::SubgroupSize), 0u);
+
+    Vector<Uint32> output;
+    ASSERT_TRUE(ShaderCompiler::DeriveNumSubgroupsForVulkan(input, output, true));
+    ASSERT_TRUE(Validates(output));
+
+    const Uint32 outputNumSubgroups = FindBuiltinTarget(output, spv::BuiltIn::NumSubgroups);
+    const Uint32 outputSubgroupSize = FindBuiltinTarget(output, spv::BuiltIn::SubgroupSize);
+    ASSERT_NE(outputNumSubgroups, 0u);
+    ASSERT_NE(outputSubgroupSize, 0u);
+    EXPECT_EQ(CountLoadsFrom(output, outputNumSubgroups), 0u);
+    EXPECT_EQ(CountLoadsFrom(output, outputSubgroupSize), 1u);
+    EXPECT_EQ(CountOpcode(output, spv::Op::OpCompositeExtract), 3u);
+    EXPECT_EQ(CountOpcode(output, spv::Op::OpIMul), 2u);
+    EXPECT_EQ(CountOpcode(output, spv::Op::OpUDiv), 1u);
+}
+
+TEST(DeriveNumSubgroupsPass, IsIdempotent) {
+    Vector<Uint32> once;
+    ASSERT_TRUE(ShaderCompiler::DeriveNumSubgroupsForVulkan(CompileCompute(kNumSubgroupsOnlySource), once, true));
+    Vector<Uint32> twice;
+    ASSERT_TRUE(ShaderCompiler::DeriveNumSubgroupsForVulkan(once, twice, true));
+    EXPECT_EQ(twice, once);
+}
+
+TEST(DeriveNumSubgroupsPass, LeavesUnrelatedComputeShaderUntouched) {
+    const Vector<Uint32> input = CompileCompute(kNoNumSubgroupsSource);
+    ASSERT_FALSE(input.empty());
+    Vector<Uint32> output;
+    ASSERT_TRUE(ShaderCompiler::DeriveNumSubgroupsForVulkan(input, output, true));
+    EXPECT_EQ(output, input);
+    EXPECT_EQ(FindBuiltinTarget(output, spv::BuiltIn::SubgroupSize), 0u);
+}

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_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",

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

@@ -7,6 +7,8 @@
 // End of Source File Header
 
 #include "DriverPost.h"
+#include "DriverPostProgram203Witness.h"
+#include "DriverPostProgram203WitnessSpv.h"
 #include "MG_Util/BackendLoaders/OpenGL/Loader.h"
 #include <Config.h>
 #include <MGGitHash.h>
@@ -24,6 +26,8 @@
 #include <MG_Util/Texture/TextureFormatProcessor.h>
 #include <MG_Util/Async/ShaderCompilePool.h>
 #include <chrono>
+#include <cstring>
+#include <limits>
 #include <thread>
 
 #if !defined(_WIN32)
@@ -1168,7 +1172,9 @@ 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);
@@ -1452,6 +1458,437 @@ namespace MobileGL::MG_Util::SelfTest {
                              disabledNote);
         }
 
+        // Native Program-203 compute witness. This deliberately uses a separate
+        // throwaway Vulkan device rather than the real renderer's queues, and it
+        // treats MOBILEGL_DISABLE_SUBGROUP as irrelevant: the row reports what the
+        // driver does, not what MobileGL elects to advertise to applications.
+        void ProbeVulkanProgram203Witness(ReportBuilder& builder, PFN_vkGetInstanceProcAddr getInstanceProcAddr,
+                                           VkInstance instance, VkPhysicalDevice physicalDevice,
+                                           Uint32 computeQueueFamilyIndex,
+                                           const VkPhysicalDeviceProperties& properties,
+                                           Bool subgroupPropertiesAvailable,
+                                           const VkPhysicalDeviceSubgroupProperties& subgroupProperties) {
+            constexpr const char* RowName = "Subgroup first-reduction witness";
+            const auto fail = [&](String detail) { builder.Fail(RowName, Move(detail)); };
+
+            if (!subgroupPropertiesAvailable) {
+                fail("vkGetPhysicalDeviceProperties2 could not provide raw Vulkan subgroup properties");
+                return;
+            }
+
+            Program203WitnessLimits limits{};
+            limits.computeStageSupported =
+                (subgroupProperties.supportedStages & VK_SHADER_STAGE_COMPUTE_BIT) != 0;
+            limits.basicSubgroupSupported =
+                (subgroupProperties.supportedOperations & VK_SUBGROUP_FEATURE_BASIC_BIT) != 0;
+            limits.arithmeticSubgroupSupported =
+                (subgroupProperties.supportedOperations & VK_SUBGROUP_FEATURE_ARITHMETIC_BIT) != 0;
+            limits.subgroupSize = subgroupProperties.subgroupSize;
+            limits.maxComputeWorkGroupInvocations = properties.limits.maxComputeWorkGroupInvocations;
+            limits.maxComputeWorkGroupSize = {properties.limits.maxComputeWorkGroupSize[0],
+                                               properties.limits.maxComputeWorkGroupSize[1],
+                                               properties.limits.maxComputeWorkGroupSize[2]};
+            limits.maxComputeSharedMemorySize = properties.limits.maxComputeSharedMemorySize;
+            limits.maxPerStageDescriptorStorageBuffers = properties.limits.maxPerStageDescriptorStorageBuffers;
+            limits.maxDescriptorSetStorageBuffers = properties.limits.maxDescriptorSetStorageBuffers;
+            limits.maxBoundDescriptorSets = properties.limits.maxBoundDescriptorSets;
+            limits.maxStorageBufferRange = properties.limits.maxStorageBufferRange;
+
+            const Program203WitnessEligibilityResult eligibility = EvaluateProgram203WitnessEligibility(limits);
+            if (eligibility.eligibility == Program203WitnessEligibility::SkipUnsupportedNativeFeatureSet) {
+                builder.Info(RowName, eligibility.detail);
+                return;
+            }
+            if (eligibility.eligibility == Program203WitnessEligibility::FailInadequateLimits) {
+                fail(eligibility.detail);
+                return;
+            }
+            if (computeQueueFamilyIndex == std::numeric_limits<Uint32>::max()) {
+                fail("no compute queue family is available for the native Vulkan witness");
+                return;
+            }
+
+            const auto vkGetPhysicalDeviceMemoryPropertiesFn =
+                reinterpret_cast<PFN_vkGetPhysicalDeviceMemoryProperties>(
+                    getInstanceProcAddr(instance, "vkGetPhysicalDeviceMemoryProperties"));
+            const auto vkCreateDeviceFn =
+                reinterpret_cast<PFN_vkCreateDevice>(getInstanceProcAddr(instance, "vkCreateDevice"));
+            const auto vkDestroyDeviceFn =
+                reinterpret_cast<PFN_vkDestroyDevice>(getInstanceProcAddr(instance, "vkDestroyDevice"));
+            const auto vkGetDeviceQueueFn =
+                reinterpret_cast<PFN_vkGetDeviceQueue>(getInstanceProcAddr(instance, "vkGetDeviceQueue"));
+            const auto vkCreateBufferFn =
+                reinterpret_cast<PFN_vkCreateBuffer>(getInstanceProcAddr(instance, "vkCreateBuffer"));
+            const auto vkDestroyBufferFn =
+                reinterpret_cast<PFN_vkDestroyBuffer>(getInstanceProcAddr(instance, "vkDestroyBuffer"));
+            const auto vkGetBufferMemoryRequirementsFn = reinterpret_cast<PFN_vkGetBufferMemoryRequirements>(
+                getInstanceProcAddr(instance, "vkGetBufferMemoryRequirements"));
+            const auto vkAllocateMemoryFn =
+                reinterpret_cast<PFN_vkAllocateMemory>(getInstanceProcAddr(instance, "vkAllocateMemory"));
+            const auto vkFreeMemoryFn =
+                reinterpret_cast<PFN_vkFreeMemory>(getInstanceProcAddr(instance, "vkFreeMemory"));
+            const auto vkBindBufferMemoryFn =
+                reinterpret_cast<PFN_vkBindBufferMemory>(getInstanceProcAddr(instance, "vkBindBufferMemory"));
+            const auto vkMapMemoryFn =
+                reinterpret_cast<PFN_vkMapMemory>(getInstanceProcAddr(instance, "vkMapMemory"));
+            const auto vkUnmapMemoryFn =
+                reinterpret_cast<PFN_vkUnmapMemory>(getInstanceProcAddr(instance, "vkUnmapMemory"));
+            const auto vkCreateDescriptorSetLayoutFn = reinterpret_cast<PFN_vkCreateDescriptorSetLayout>(
+                getInstanceProcAddr(instance, "vkCreateDescriptorSetLayout"));
+            const auto vkDestroyDescriptorSetLayoutFn = reinterpret_cast<PFN_vkDestroyDescriptorSetLayout>(
+                getInstanceProcAddr(instance, "vkDestroyDescriptorSetLayout"));
+            const auto vkCreateDescriptorPoolFn =
+                reinterpret_cast<PFN_vkCreateDescriptorPool>(getInstanceProcAddr(instance, "vkCreateDescriptorPool"));
+            const auto vkDestroyDescriptorPoolFn = reinterpret_cast<PFN_vkDestroyDescriptorPool>(
+                getInstanceProcAddr(instance, "vkDestroyDescriptorPool"));
+            const auto vkAllocateDescriptorSetsFn = reinterpret_cast<PFN_vkAllocateDescriptorSets>(
+                getInstanceProcAddr(instance, "vkAllocateDescriptorSets"));
+            const auto vkUpdateDescriptorSetsFn =
+                reinterpret_cast<PFN_vkUpdateDescriptorSets>(getInstanceProcAddr(instance, "vkUpdateDescriptorSets"));
+            const auto vkCreateShaderModuleFn =
+                reinterpret_cast<PFN_vkCreateShaderModule>(getInstanceProcAddr(instance, "vkCreateShaderModule"));
+            const auto vkDestroyShaderModuleFn =
+                reinterpret_cast<PFN_vkDestroyShaderModule>(getInstanceProcAddr(instance, "vkDestroyShaderModule"));
+            const auto vkCreatePipelineLayoutFn =
+                reinterpret_cast<PFN_vkCreatePipelineLayout>(getInstanceProcAddr(instance, "vkCreatePipelineLayout"));
+            const auto vkDestroyPipelineLayoutFn = reinterpret_cast<PFN_vkDestroyPipelineLayout>(
+                getInstanceProcAddr(instance, "vkDestroyPipelineLayout"));
+            const auto vkCreateComputePipelinesFn = reinterpret_cast<PFN_vkCreateComputePipelines>(
+                getInstanceProcAddr(instance, "vkCreateComputePipelines"));
+            const auto vkDestroyPipelineFn =
+                reinterpret_cast<PFN_vkDestroyPipeline>(getInstanceProcAddr(instance, "vkDestroyPipeline"));
+            const auto vkCreateCommandPoolFn =
+                reinterpret_cast<PFN_vkCreateCommandPool>(getInstanceProcAddr(instance, "vkCreateCommandPool"));
+            const auto vkDestroyCommandPoolFn =
+                reinterpret_cast<PFN_vkDestroyCommandPool>(getInstanceProcAddr(instance, "vkDestroyCommandPool"));
+            const auto vkAllocateCommandBuffersFn = reinterpret_cast<PFN_vkAllocateCommandBuffers>(
+                getInstanceProcAddr(instance, "vkAllocateCommandBuffers"));
+            const auto vkBeginCommandBufferFn =
+                reinterpret_cast<PFN_vkBeginCommandBuffer>(getInstanceProcAddr(instance, "vkBeginCommandBuffer"));
+            const auto vkEndCommandBufferFn =
+                reinterpret_cast<PFN_vkEndCommandBuffer>(getInstanceProcAddr(instance, "vkEndCommandBuffer"));
+            const auto vkCmdBindPipelineFn =
+                reinterpret_cast<PFN_vkCmdBindPipeline>(getInstanceProcAddr(instance, "vkCmdBindPipeline"));
+            const auto vkCmdBindDescriptorSetsFn = reinterpret_cast<PFN_vkCmdBindDescriptorSets>(
+                getInstanceProcAddr(instance, "vkCmdBindDescriptorSets"));
+            const auto vkCmdDispatchFn =
+                reinterpret_cast<PFN_vkCmdDispatch>(getInstanceProcAddr(instance, "vkCmdDispatch"));
+            const auto vkCmdPipelineBarrierFn =
+                reinterpret_cast<PFN_vkCmdPipelineBarrier>(getInstanceProcAddr(instance, "vkCmdPipelineBarrier"));
+            const auto vkCreateFenceFn =
+                reinterpret_cast<PFN_vkCreateFence>(getInstanceProcAddr(instance, "vkCreateFence"));
+            const auto vkDestroyFenceFn =
+                reinterpret_cast<PFN_vkDestroyFence>(getInstanceProcAddr(instance, "vkDestroyFence"));
+            const auto vkQueueSubmitFn =
+                reinterpret_cast<PFN_vkQueueSubmit>(getInstanceProcAddr(instance, "vkQueueSubmit"));
+            const auto vkWaitForFencesFn =
+                reinterpret_cast<PFN_vkWaitForFences>(getInstanceProcAddr(instance, "vkWaitForFences"));
+            const auto vkDeviceWaitIdleFn =
+                reinterpret_cast<PFN_vkDeviceWaitIdle>(getInstanceProcAddr(instance, "vkDeviceWaitIdle"));
+
+            if (vkGetPhysicalDeviceMemoryPropertiesFn == nullptr || vkCreateDeviceFn == nullptr ||
+                vkDestroyDeviceFn == nullptr || vkGetDeviceQueueFn == nullptr || vkCreateBufferFn == nullptr ||
+                vkDestroyBufferFn == nullptr || vkGetBufferMemoryRequirementsFn == nullptr ||
+                vkAllocateMemoryFn == nullptr || vkFreeMemoryFn == nullptr || vkBindBufferMemoryFn == nullptr ||
+                vkMapMemoryFn == nullptr || vkUnmapMemoryFn == nullptr || vkCreateDescriptorSetLayoutFn == nullptr ||
+                vkDestroyDescriptorSetLayoutFn == nullptr || vkCreateDescriptorPoolFn == nullptr ||
+                vkDestroyDescriptorPoolFn == nullptr || vkAllocateDescriptorSetsFn == nullptr ||
+                vkUpdateDescriptorSetsFn == nullptr || vkCreateShaderModuleFn == nullptr ||
+                vkDestroyShaderModuleFn == nullptr || vkCreatePipelineLayoutFn == nullptr ||
+                vkDestroyPipelineLayoutFn == nullptr || vkCreateComputePipelinesFn == nullptr ||
+                vkDestroyPipelineFn == nullptr || vkCreateCommandPoolFn == nullptr || vkDestroyCommandPoolFn == nullptr ||
+                vkAllocateCommandBuffersFn == nullptr || vkBeginCommandBufferFn == nullptr ||
+                vkEndCommandBufferFn == nullptr || vkCmdBindPipelineFn == nullptr ||
+                vkCmdBindDescriptorSetsFn == nullptr || vkCmdDispatchFn == nullptr ||
+                vkCmdPipelineBarrierFn == nullptr || vkCreateFenceFn == nullptr || vkDestroyFenceFn == nullptr ||
+                vkQueueSubmitFn == nullptr || vkWaitForFencesFn == nullptr || vkDeviceWaitIdleFn == nullptr) {
+                fail("vkGetInstanceProcAddr could not resolve the Vulkan entry points required for the witness");
+                return;
+            }
+
+            const Float queuePriority = 1.0f;
+            VkDeviceQueueCreateInfo queueInfo{};
+            queueInfo.sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO;
+            queueInfo.queueFamilyIndex = computeQueueFamilyIndex;
+            queueInfo.queueCount = 1;
+            queueInfo.pQueuePriorities = &queuePriority;
+
+            VkDeviceCreateInfo deviceInfo{};
+            deviceInfo.sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO;
+            deviceInfo.queueCreateInfoCount = 1;
+            deviceInfo.pQueueCreateInfos = &queueInfo;
+
+            VkDevice device = VK_NULL_HANDLE;
+            VkResult result = vkCreateDeviceFn(physicalDevice, &deviceInfo, nullptr, &device);
+            if (result != VK_SUCCESS || device == VK_NULL_HANDLE) {
+                fail(format("vkCreateDevice failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+
+            VkBuffer outputBuffer = VK_NULL_HANDLE;
+            VkDeviceMemory outputMemory = VK_NULL_HANDLE;
+            VkDescriptorSetLayout descriptorSetLayout = VK_NULL_HANDLE;
+            VkDescriptorPool descriptorPool = VK_NULL_HANDLE;
+            VkShaderModule shaderModule = VK_NULL_HANDLE;
+            VkPipelineLayout pipelineLayout = VK_NULL_HANDLE;
+            VkPipeline pipeline = VK_NULL_HANDLE;
+            VkCommandPool commandPool = VK_NULL_HANDLE;
+            VkFence fence = VK_NULL_HANDLE;
+            void* mappedOutput = nullptr;
+            Bool fenceWaitTimedOut = false;
+            const ScopeGuard destroyDeviceObjects([&]() {
+                if (fenceWaitTimedOut) {
+                    // Match ProbeVulkanTimerQuery: the command may still execute
+                    // after a timeout, so intentionally retain every device-owned
+                    // resource rather than risking a forever wait or UAF in the ICD.
+                    return;
+                }
+                vkDeviceWaitIdleFn(device);
+                if (fence != VK_NULL_HANDLE) vkDestroyFenceFn(device, fence, nullptr);
+                if (commandPool != VK_NULL_HANDLE) vkDestroyCommandPoolFn(device, commandPool, nullptr);
+                if (pipeline != VK_NULL_HANDLE) vkDestroyPipelineFn(device, pipeline, nullptr);
+                if (pipelineLayout != VK_NULL_HANDLE) vkDestroyPipelineLayoutFn(device, pipelineLayout, nullptr);
+                if (shaderModule != VK_NULL_HANDLE) vkDestroyShaderModuleFn(device, shaderModule, nullptr);
+                if (descriptorPool != VK_NULL_HANDLE) vkDestroyDescriptorPoolFn(device, descriptorPool, nullptr);
+                if (descriptorSetLayout != VK_NULL_HANDLE) {
+                    vkDestroyDescriptorSetLayoutFn(device, descriptorSetLayout, nullptr);
+                }
+                if (mappedOutput != nullptr) vkUnmapMemoryFn(device, outputMemory);
+                if (outputBuffer != VK_NULL_HANDLE) vkDestroyBufferFn(device, outputBuffer, nullptr);
+                if (outputMemory != VK_NULL_HANDLE) vkFreeMemoryFn(device, outputMemory, nullptr);
+                vkDestroyDeviceFn(device, nullptr);
+            });
+
+            VkQueue queue = VK_NULL_HANDLE;
+            vkGetDeviceQueueFn(device, computeQueueFamilyIndex, 0, &queue);
+            if (queue == VK_NULL_HANDLE) {
+                fail("vkGetDeviceQueue returned a null compute queue");
+                return;
+            }
+
+            VkBufferCreateInfo bufferInfo{};
+            bufferInfo.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
+            bufferInfo.size = sizeof(Program203WitnessOutput);
+            bufferInfo.usage = VK_BUFFER_USAGE_STORAGE_BUFFER_BIT;
+            bufferInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+            result = vkCreateBufferFn(device, &bufferInfo, nullptr, &outputBuffer);
+            if (result != VK_SUCCESS) {
+                fail(format("vkCreateBuffer(output SSBO) failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+
+            VkMemoryRequirements memoryRequirements{};
+            vkGetBufferMemoryRequirementsFn(device, outputBuffer, &memoryRequirements);
+            VkPhysicalDeviceMemoryProperties memoryProperties{};
+            vkGetPhysicalDeviceMemoryPropertiesFn(physicalDevice, &memoryProperties);
+            Uint32 memoryTypeIndex = std::numeric_limits<Uint32>::max();
+            for (Uint32 index = 0; index < memoryProperties.memoryTypeCount; ++index) {
+                const Bool compatible = (memoryRequirements.memoryTypeBits & (1u << index)) != 0u;
+                const VkMemoryPropertyFlags required = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
+                                                       VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
+                if (compatible && (memoryProperties.memoryTypes[index].propertyFlags & required) == required) {
+                    memoryTypeIndex = index;
+                    break;
+                }
+            }
+            if (memoryTypeIndex == std::numeric_limits<Uint32>::max()) {
+                fail("no host-visible/coherent memory type is compatible with the output SSBO");
+                return;
+            }
+
+            VkMemoryAllocateInfo memoryInfo{};
+            memoryInfo.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO;
+            memoryInfo.allocationSize = memoryRequirements.size;
+            memoryInfo.memoryTypeIndex = memoryTypeIndex;
+            result = vkAllocateMemoryFn(device, &memoryInfo, nullptr, &outputMemory);
+            if (result != VK_SUCCESS) {
+                fail(format("vkAllocateMemory(output SSBO) failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+            result = vkBindBufferMemoryFn(device, outputBuffer, outputMemory, 0);
+            if (result != VK_SUCCESS) {
+                fail(format("vkBindBufferMemory(output SSBO) failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+            result = vkMapMemoryFn(device, outputMemory, 0, sizeof(Program203WitnessOutput), 0, &mappedOutput);
+            if (result != VK_SUCCESS || mappedOutput == nullptr) {
+                fail(format("vkMapMemory(output SSBO) failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+            std::memset(mappedOutput, 0xa5, sizeof(Program203WitnessOutput));
+
+            VkDescriptorSetLayoutBinding outputBinding{};
+            outputBinding.binding = 0;
+            outputBinding.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
+            outputBinding.descriptorCount = 1;
+            outputBinding.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT;
+            VkDescriptorSetLayoutCreateInfo descriptorSetLayoutInfo{};
+            descriptorSetLayoutInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO;
+            descriptorSetLayoutInfo.bindingCount = 1;
+            descriptorSetLayoutInfo.pBindings = &outputBinding;
+            result = vkCreateDescriptorSetLayoutFn(device, &descriptorSetLayoutInfo, nullptr, &descriptorSetLayout);
+            if (result != VK_SUCCESS) {
+                fail(format("vkCreateDescriptorSetLayout failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+
+            VkDescriptorPoolSize poolSize{};
+            poolSize.type = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
+            poolSize.descriptorCount = 1;
+            VkDescriptorPoolCreateInfo descriptorPoolInfo{};
+            descriptorPoolInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO;
+            descriptorPoolInfo.maxSets = 1;
+            descriptorPoolInfo.poolSizeCount = 1;
+            descriptorPoolInfo.pPoolSizes = &poolSize;
+            result = vkCreateDescriptorPoolFn(device, &descriptorPoolInfo, nullptr, &descriptorPool);
+            if (result != VK_SUCCESS) {
+                fail(format("vkCreateDescriptorPool failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+
+            VkDescriptorSet descriptorSet = VK_NULL_HANDLE;
+            VkDescriptorSetAllocateInfo descriptorSetInfo{};
+            descriptorSetInfo.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO;
+            descriptorSetInfo.descriptorPool = descriptorPool;
+            descriptorSetInfo.descriptorSetCount = 1;
+            descriptorSetInfo.pSetLayouts = &descriptorSetLayout;
+            result = vkAllocateDescriptorSetsFn(device, &descriptorSetInfo, &descriptorSet);
+            if (result != VK_SUCCESS) {
+                fail(format("vkAllocateDescriptorSets failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+            VkDescriptorBufferInfo outputDescriptor{};
+            outputDescriptor.buffer = outputBuffer;
+            outputDescriptor.offset = 0;
+            outputDescriptor.range = sizeof(Program203WitnessOutput);
+            VkWriteDescriptorSet descriptorWrite{};
+            descriptorWrite.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
+            descriptorWrite.dstSet = descriptorSet;
+            descriptorWrite.dstBinding = 0;
+            descriptorWrite.descriptorCount = 1;
+            descriptorWrite.descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
+            descriptorWrite.pBufferInfo = &outputDescriptor;
+            vkUpdateDescriptorSetsFn(device, 1, &descriptorWrite, 0, nullptr);
+
+            VkShaderModuleCreateInfo shaderModuleInfo{};
+            shaderModuleInfo.sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO;
+            shaderModuleInfo.codeSize = sizeof(kDriverPostProgram203WitnessSpv);
+            shaderModuleInfo.pCode = kDriverPostProgram203WitnessSpv;
+            result = vkCreateShaderModuleFn(device, &shaderModuleInfo, nullptr, &shaderModule);
+            if (result != VK_SUCCESS) {
+                fail(format("vkCreateShaderModule failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+
+            VkPipelineLayoutCreateInfo pipelineLayoutInfo{};
+            pipelineLayoutInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
+            pipelineLayoutInfo.setLayoutCount = 1;
+            pipelineLayoutInfo.pSetLayouts = &descriptorSetLayout;
+            result = vkCreatePipelineLayoutFn(device, &pipelineLayoutInfo, nullptr, &pipelineLayout);
+            if (result != VK_SUCCESS) {
+                fail(format("vkCreatePipelineLayout failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+
+            VkPipelineShaderStageCreateInfo shaderStage{};
+            shaderStage.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
+            shaderStage.stage = VK_SHADER_STAGE_COMPUTE_BIT;
+            shaderStage.module = shaderModule;
+            shaderStage.pName = "main";
+            VkComputePipelineCreateInfo pipelineInfo{};
+            pipelineInfo.sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO;
+            pipelineInfo.stage = shaderStage;
+            pipelineInfo.layout = pipelineLayout;
+            result = vkCreateComputePipelinesFn(device, VK_NULL_HANDLE, 1, &pipelineInfo, nullptr, &pipeline);
+            if (result != VK_SUCCESS) {
+                fail(format("vkCreateComputePipelines failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+
+            VkCommandPoolCreateInfo commandPoolInfo{};
+            commandPoolInfo.sType = VK_STRUCTURE_TYPE_COMMAND_POOL_CREATE_INFO;
+            commandPoolInfo.queueFamilyIndex = computeQueueFamilyIndex;
+            result = vkCreateCommandPoolFn(device, &commandPoolInfo, nullptr, &commandPool);
+            if (result != VK_SUCCESS) {
+                fail(format("vkCreateCommandPool failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+            VkCommandBufferAllocateInfo commandBufferInfo{};
+            commandBufferInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;
+            commandBufferInfo.commandPool = commandPool;
+            commandBufferInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
+            commandBufferInfo.commandBufferCount = 1;
+            VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
+            result = vkAllocateCommandBuffersFn(device, &commandBufferInfo, &commandBuffer);
+            if (result != VK_SUCCESS) {
+                fail(format("vkAllocateCommandBuffers failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+
+            VkCommandBufferBeginInfo commandBufferBeginInfo{};
+            commandBufferBeginInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
+            commandBufferBeginInfo.flags = VK_COMMAND_BUFFER_USAGE_ONE_TIME_SUBMIT_BIT;
+            result = vkBeginCommandBufferFn(commandBuffer, &commandBufferBeginInfo);
+            if (result != VK_SUCCESS) {
+                fail(format("vkBeginCommandBuffer failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+            vkCmdBindPipelineFn(commandBuffer, VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
+            vkCmdBindDescriptorSetsFn(commandBuffer, VK_PIPELINE_BIND_POINT_COMPUTE, pipelineLayout, 0, 1,
+                                      &descriptorSet, 0, nullptr);
+            vkCmdDispatchFn(commandBuffer, 1, 1, 1);
+            VkBufferMemoryBarrier hostReadBarrier{};
+            hostReadBarrier.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
+            hostReadBarrier.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT;
+            hostReadBarrier.dstAccessMask = VK_ACCESS_HOST_READ_BIT;
+            hostReadBarrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+            hostReadBarrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+            hostReadBarrier.buffer = outputBuffer;
+            hostReadBarrier.offset = 0;
+            hostReadBarrier.size = sizeof(Program203WitnessOutput);
+            vkCmdPipelineBarrierFn(commandBuffer, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_HOST_BIT, 0,
+                                   0, nullptr, 1, &hostReadBarrier, 0, nullptr);
+            result = vkEndCommandBufferFn(commandBuffer);
+            if (result != VK_SUCCESS) {
+                fail(format("vkEndCommandBuffer failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+
+            VkFenceCreateInfo fenceInfo{};
+            fenceInfo.sType = VK_STRUCTURE_TYPE_FENCE_CREATE_INFO;
+            result = vkCreateFenceFn(device, &fenceInfo, nullptr, &fence);
+            if (result != VK_SUCCESS) {
+                fail(format("vkCreateFence failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+            VkSubmitInfo submitInfo{};
+            submitInfo.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
+            submitInfo.commandBufferCount = 1;
+            submitInfo.pCommandBuffers = &commandBuffer;
+            result = vkQueueSubmitFn(queue, 1, &submitInfo, fence);
+            if (result != VK_SUCCESS) {
+                fail(format("vkQueueSubmit failed (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+
+            constexpr Uint64 FenceTimeoutNs = 5'000'000'000ull;
+            result = vkWaitForFencesFn(device, 1, &fence, VK_TRUE, FenceTimeoutNs);
+            if (result != VK_SUCCESS) {
+                fenceWaitTimedOut = true;
+                fail(format("vkWaitForFences did not signal within 5 s (VkResult = {})", static_cast<Int>(result)));
+                return;
+            }
+
+            Program203WitnessOutput output{};
+            std::memcpy(&output, mappedOutput, sizeof(output));
+            const Program203WitnessValidationResult validation = ValidateProgram203Witness(output);
+            if (!validation.ok) {
+                fail(validation.detail);
+                return;
+            }
+            builder.Pass(RowName, validation.detail);
+        }
+
         // Everything the "MobileGL reported ..." rows need from the Vulkan device probe.
         struct VulkanProbeSummary {
             Bool devicePropsValid = false;
@@ -1461,6 +1898,8 @@ namespace MobileGL::MG_Util::SelfTest {
             Bool shaderSubgroupUsable = false;
             Bool timerQueriesSupported = false;
             Bool samplerAnisotropySupported = false;
+            Bool drawIndirectFirstInstanceSupported = false;
+            Bool shaderDrawParametersSupported = false;
         };
     } // namespace
 
@@ -1665,6 +2104,7 @@ namespace MobileGL::MG_Util::SelfTest {
         VkPhysicalDevice physicalDevice = VK_NULL_HANDLE;
         Uint32 graphicsQueueFamilyIndex = 0;
         Uint32 graphicsQueueTimestampValidBits = 0;
+        Uint32 computeQueueFamilyIndex = std::numeric_limits<Uint32>::max();
         for (VkPhysicalDevice candidate : devices) {
             Uint32 queueFamilyCount = 0;
             vkGetPhysicalDeviceQueueFamilyPropertiesFn(candidate, &queueFamilyCount, nullptr);
@@ -1680,6 +2120,13 @@ namespace MobileGL::MG_Util::SelfTest {
                 }
             }
             if (physicalDevice != VK_NULL_HANDLE) {
+                for (Uint32 familyIndex = 0; familyIndex < queueFamilyCount; ++familyIndex) {
+                    const VkQueueFamilyProperties& family = queueFamilies[familyIndex];
+                    if (family.queueCount > 0 && (family.queueFlags & VK_QUEUE_COMPUTE_BIT) != 0) {
+                        computeQueueFamilyIndex = familyIndex;
+                        break;
+                    }
+                }
                 break;
             }
         }
@@ -1744,6 +2191,7 @@ 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 +2358,7 @@ 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;
@@ -2014,13 +2463,15 @@ namespace MobileGL::MG_Util::SelfTest {
                          "change every N instances change every one");
         }
 
-        if (vkGetPhysicalDeviceProperties2Fn != nullptr && properties.apiVersion >= VK_API_VERSION_1_1) {
         VkPhysicalDeviceSubgroupProperties subgroupProperties{};
+        Bool subgroupPropertiesAvailable = false;
+        if (vkGetPhysicalDeviceProperties2Fn != nullptr && properties.apiVersion >= VK_API_VERSION_1_1) {
             subgroupProperties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_PROPERTIES;
             VkPhysicalDeviceProperties2 properties2{};
             properties2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2;
             properties2.pNext = &subgroupProperties;
             vkGetPhysicalDeviceProperties2Fn(physicalDevice, &properties2);
+            subgroupPropertiesAvailable = true;
             const Bool subgroupUsable = subgroupProperties.subgroupSize > 0 &&
                                         (subgroupProperties.supportedStages & VK_SHADER_STAGE_COMPUTE_BIT) != 0 &&
                                         (subgroupProperties.supportedOperations & VK_SUBGROUP_FEATURE_BASIC_BIT) != 0;
@@ -2040,6 +2491,9 @@ namespace MobileGL::MG_Util::SelfTest {
             builder.Warn("Compute shader subgroup", "subgroup properties could not be queried");
         }
 
+        ProbeVulkanProgram203Witness(builder, getInstanceProcAddr, instance, physicalDevice, computeQueueFamilyIndex,
+                                     properties, subgroupPropertiesAvailable, subgroupProperties);
+
         if (HasVkExtension(deviceExtensions, VK_KHR_DRAW_INDIRECT_COUNT_EXTENSION_NAME)) {
             builder.Pass("VK_KHR_draw_indirect_count",
                          "supported (count-buffer indirect draws run as single native "
@@ -2108,7 +2562,8 @@ 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/SelfTest/DriverPostProgram203Witness.comp

@@ -0,0 +1,164 @@
+// MobileGL - MobileGL/MG_Util/SelfTest/DriverPostProgram203Witness.comp
+// 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
+// End of Source File Header
+//
+// Native Vulkan GLSL 450 witness for Program 203's first subgroup reduction.
+// It is intentionally independent of the GL 430 integration scenario. The body
+// below preserves Program 203's source reduction; the surrounding diagnostics
+// only observe its topology and cache handoffs.
+
+#version 450
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_arithmetic : require
+
+layout(local_size_x = 32, local_size_y = 16, local_size_z = 1) in;
+
+const uint kTopologyNonuniformNumSubgroups = 1u << 0u;
+const uint kTopologyInvalidNumSubgroups = 1u << 1u;
+const uint kTopologyInvalidSubgroupId = 1u << 2u;
+const uint kTopologyInvalidSubgroupLane = 1u << 3u;
+const uint kWitnessMagic = 0x50323033u;
+
+layout(std430, set = 0, binding = 0) buffer Program203WitnessOutput {
+    uint magic;
+    uint topologyFlags;
+    uint numSubgroups;
+    uint loopLength;
+    uint seenSubgroupMask;
+
+    uvec4 owner511;
+
+    uint lastLaneWriterCount[32];
+    uint indexedInputTotal[32];
+
+    vec2 rawPrefix[32];
+    vec2 scanCache[6][32];
+    vec2 finalAverage;
+} outWitness;
+
+// Program 203's cache stays separate from all diagnostic shared state. In
+// particular, no instrumentation stores through prefixSumCache except source
+// writes retained below.
+shared vec2 prefixSumCache[32];
+shared uint canonicalNumSubgroups;
+shared uint topologyFlagsShared;
+shared uint seenSubgroupMaskShared;
+shared uint lastLaneWriterCountShared[32];
+shared uint indexedInputTotalShared[32];
+
+void main() {
+    const uint localInvocationIndex = gl_LocalInvocationIndex;
+
+    // Host memory is deliberately poisoned before dispatch. Initialize only
+    // shared atomic diagnostic state; owner, average, and magic remain poisoned
+    // until their required post-source-barrier writes below.
+    if (localInvocationIndex == 0u) {
+        canonicalNumSubgroups = 0u;
+        topologyFlagsShared = 0u;
+        seenSubgroupMaskShared = 0u;
+    }
+    if (localInvocationIndex < 32u) {
+        lastLaneWriterCountShared[localInvocationIndex] = 0u;
+        indexedInputTotalShared[localInvocationIndex] = 0u;
+    }
+    memoryBarrierShared();
+    barrier();
+
+    // Invocation zero defines the canonical domain. It is broadcast through
+    // shared memory before every invocation records its own raw observation.
+    if (localInvocationIndex == 0u) {
+        canonicalNumSubgroups = gl_NumSubgroups;
+        outWitness.numSubgroups = gl_NumSubgroups;
+    }
+    barrier();
+
+    const uint canonicalN = canonicalNumSubgroups;
+    if (gl_NumSubgroups != canonicalN)
+        atomicOr(topologyFlagsShared, kTopologyNonuniformNumSubgroups);
+    if (gl_NumSubgroups < 2u || gl_NumSubgroups > 32u)
+        atomicOr(topologyFlagsShared, kTopologyInvalidNumSubgroups);
+    if (gl_SubgroupID >= canonicalN || gl_SubgroupID >= 32u)
+        atomicOr(topologyFlagsShared, kTopologyInvalidSubgroupId);
+    if (gl_SubgroupSize == 0u || gl_SubgroupInvocationID >= gl_SubgroupSize)
+        atomicOr(topologyFlagsShared, kTopologyInvalidSubgroupLane);
+
+    // Keep all atomic collection bounded by the canonical valid domain. A
+    // nonuniform/broken report reaches the uniform safety branch below instead
+    // of making some lanes return before a barrier.
+    const bool canonicalDomain = canonicalN >= 2u && canonicalN <= 32u;
+    const bool idInCanonicalDomain = canonicalDomain && gl_SubgroupID < canonicalN;
+    if (idInCanonicalDomain) {
+        atomicOr(seenSubgroupMaskShared, 1u << gl_SubgroupID);
+        atomicAdd(indexedInputTotalShared[gl_SubgroupID], localInvocationIndex + 1u);
+        if (gl_SubgroupSize != 0u && gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+            atomicAdd(lastLaneWriterCountShared[gl_SubgroupID], 1u);
+    }
+    memoryBarrierShared();
+    barrier();
+
+    if (localInvocationIndex == 0u)
+        outWitness.seenSubgroupMask = seenSubgroupMaskShared;
+    if (localInvocationIndex < 32u) {
+        outWitness.lastLaneWriterCount[localInvocationIndex] = lastLaneWriterCountShared[localInvocationIndex];
+        outWitness.indexedInputTotal[localInvocationIndex] = indexedInputTotalShared[localInvocationIndex];
+    }
+
+    // This branch is uniform after collection and is solely a safety guard for
+    // broken topology reports. The valid side retains Program 203 verbatim.
+    const bool sourceDomain = canonicalDomain && topologyFlagsShared == 0u;
+    if (sourceDomain) {
+        vec2 sampleLuminance = vec2(float(gl_LocalInvocationIndex + 1u), 0.0);
+        sampleLuminance = subgroupInclusiveAdd(sampleLuminance);
+
+        if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+            prefixSumCache[gl_SubgroupID] = sampleLuminance;
+        barrier();
+
+        if (gl_LocalInvocationIndex < gl_NumSubgroups)
+            outWitness.rawPrefix[gl_LocalInvocationIndex] = prefixSumCache[gl_LocalInvocationIndex];
+        barrier();
+
+        uint loopLength = uint(findMSB(gl_NumSubgroups));
+        loopLength += uint(gl_NumSubgroups - (1u << (loopLength - 1u)) > 0u);
+        if (gl_LocalInvocationIndex == 0u)
+            outWitness.loopLength = loopLength;
+
+        for (uint scanStage = 0u; scanStage < loopLength; ++scanStage) {
+            if ((gl_SubgroupID & (1u << scanStage)) > 0u) {
+                sampleLuminance += prefixSumCache[(gl_SubgroupID >> scanStage << scanStage) - 1u];
+                if (gl_SubgroupInvocationID == gl_SubgroupSize - 1u)
+                    prefixSumCache[gl_SubgroupID] = sampleLuminance;
+            }
+            barrier();
+
+            if (gl_LocalInvocationIndex < gl_NumSubgroups)
+                outWitness.scanCache[scanStage][gl_LocalInvocationIndex] =
+                    prefixSumCache[gl_LocalInvocationIndex];
+            // A second, diagnostic-only barrier prevents a faster invocation
+            // from entering the next source stage while another reads this cache.
+            barrier();
+        }
+
+        if (gl_LocalInvocationIndex == 511u)
+            prefixSumCache[0] = sampleLuminance / 512.0;
+        barrier();
+
+        if (gl_LocalInvocationIndex == 511u) {
+            outWitness.owner511 = uvec4(gl_SubgroupSize, gl_NumSubgroups, gl_SubgroupID,
+                                        gl_SubgroupInvocationID);
+            outWitness.finalAverage = prefixSumCache[0];
+        }
+    }
+
+    // Both sides of the uniform branch reach this barrier. The magic is the
+    // completion latch and therefore cannot be written before the final barrier.
+    barrier();
+    if (localInvocationIndex == 0u) {
+        outWitness.topologyFlags = topologyFlagsShared;
+        outWitness.magic = kWitnessMagic;
+    }
+}

MobileGL/MG_Util/SelfTest/DriverPostProgram203Witness.cpp

@@ -0,0 +1,277 @@
+// MobileGL - MobileGL/MG_Util/SelfTest/DriverPostProgram203Witness.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
+// End of Source File Header
+
+#include "DriverPostProgram203Witness.h"
+
+#include <bit>
+#include <sstream>
+#include <utility>
+#include <vector>
+
+namespace MobileGL::MG_Util::SelfTest {
+    namespace {
+        [[nodiscard]] Program203WitnessValidationResult Failure(Program203WitnessValidationFailure failure,
+                                                                 std::string detail,
+                                                                 std::uint32_t scanStage = 0u,
+                                                                 std::uint32_t subgroup = 0u) {
+            Program203WitnessValidationResult result;
+            result.ok = false;
+            result.failure = failure;
+            result.scanStage = scanStage;
+            result.subgroup = subgroup;
+            result.detail = std::move(detail);
+            return result;
+        }
+
+        [[nodiscard]] std::uint32_t FloatBits(float value) {
+            return std::bit_cast<std::uint32_t>(value);
+        }
+
+        [[nodiscard]] bool SameBits(float lhs, float rhs) {
+            return FloatBits(lhs) == FloatBits(rhs);
+        }
+
+        [[nodiscard]] bool SameBits(const Program203WitnessVec2& lhs, const Program203WitnessVec2& rhs) {
+            return SameBits(lhs.x, rhs.x) && SameBits(lhs.y, rhs.y);
+        }
+
+        [[nodiscard]] std::string Vec2String(const Program203WitnessVec2& value) {
+            std::ostringstream output;
+            output << '(' << value.x << ',' << value.y << ')';
+            return output.str();
+        }
+
+        [[nodiscard]] std::uint32_t ExpectedSeenSubgroupMask(std::uint32_t numSubgroups) {
+            return numSubgroups == kProgram203WitnessMaxSubgroups ? 0xffffffffu : (1u << numSubgroups) - 1u;
+        }
+
+        [[nodiscard]] std::string JoinRequirements(const std::vector<std::string>& requirements) {
+            std::ostringstream output;
+            for (std::size_t i = 0; i < requirements.size(); ++i) {
+                if (i != 0u) output << "; ";
+                output << requirements[i];
+            }
+            return output.str();
+        }
+    } // namespace
+
+    Program203WitnessEligibilityResult
+    EvaluateProgram203WitnessEligibility(const Program203WitnessLimits& limits) {
+        // This classification deliberately precedes numeric limits. An absent native
+        // compute/basic/arithmetic subgroup contract means there is nothing to witness,
+        // whereas every resource/entry-point failure on a capable device is a POST FAIL.
+        if (!limits.computeStageSupported || !limits.basicSubgroupSupported || !limits.arithmeticSubgroupSupported) {
+            std::vector<std::string> missing;
+            if (!limits.computeStageSupported) missing.emplace_back("VK_SHADER_STAGE_COMPUTE_BIT");
+            if (!limits.basicSubgroupSupported) missing.emplace_back("VK_SUBGROUP_FEATURE_BASIC_BIT");
+            if (!limits.arithmeticSubgroupSupported) missing.emplace_back("VK_SUBGROUP_FEATURE_ARITHMETIC_BIT");
+            return {Program203WitnessEligibility::SkipUnsupportedNativeFeatureSet,
+                    "skipped because the native compute/basic/arithmetic subgroup feature set is unsupported (missing " +
+                        JoinRequirements(missing) + ')'};
+        }
+
+        std::vector<std::string> inadequate;
+        if (limits.subgroupSize == 0u) {
+            inadequate.emplace_back("subgroupSize == 0");
+        }
+        if (limits.maxComputeWorkGroupInvocations < kProgram203WitnessInvocationCount) {
+            inadequate.emplace_back("maxComputeWorkGroupInvocations < 512");
+        }
+        if (limits.maxComputeWorkGroupSize[0] < 32u || limits.maxComputeWorkGroupSize[1] < 16u ||
+            limits.maxComputeWorkGroupSize[2] < 1u) {
+            inadequate.emplace_back("maxComputeWorkGroupSize does not cover 32x16x1");
+        }
+        if (limits.maxComputeSharedMemorySize < kProgram203WitnessSharedMemoryBytes) {
+            inadequate.emplace_back("maxComputeSharedMemorySize < " +
+                                    std::to_string(kProgram203WitnessSharedMemoryBytes));
+        }
+        if (limits.maxPerStageDescriptorStorageBuffers < 1u) {
+            inadequate.emplace_back("maxPerStageDescriptorStorageBuffers < 1");
+        }
+        if (limits.maxDescriptorSetStorageBuffers < 1u) {
+            inadequate.emplace_back("maxDescriptorSetStorageBuffers < 1");
+        }
+        if (limits.maxBoundDescriptorSets < 1u) {
+            inadequate.emplace_back("maxBoundDescriptorSets < 1");
+        }
+        if (limits.maxStorageBufferRange < sizeof(Program203WitnessOutput)) {
+            inadequate.emplace_back("maxStorageBufferRange < " +
+                                    std::to_string(sizeof(Program203WitnessOutput)));
+        }
+        if (!inadequate.empty()) {
+            return {Program203WitnessEligibility::FailInadequateLimits,
+                    "insufficient Vulkan limits for a 32x16x1 workgroup, one output SSBO, and " +
+                        std::to_string(kProgram203WitnessSharedMemoryBytes) + " bytes of shared memory: " +
+                        JoinRequirements(inadequate)};
+        }
+        return {Program203WitnessEligibility::Execute, {}};
+    }
+
+    std::uint32_t ComputeProgram203WitnessLoopLength(std::uint32_t numSubgroups) {
+        if (numSubgroups < 2u || numSubgroups > kProgram203WitnessMaxSubgroups) return 0u;
+
+        // Exact C++ spelling of the source's findMSB-based calculation. In
+        // particular, its final iteration for powers of two is intentional.
+        std::uint32_t loopLength = 0u;
+        for (std::uint32_t value = numSubgroups; value > 1u; value >>= 1u) {
+            ++loopLength;
+        }
+        loopLength += static_cast<std::uint32_t>(numSubgroups - (1u << (loopLength - 1u)) > 0u);
+        return loopLength;
+    }
+
+    Program203WitnessValidationResult ValidateProgram203Witness(const Program203WitnessOutput& output) {
+        // 1. Completion. A poisoned or unwritten result must never turn into a
+        // topology diagnosis, because it says nothing about execution.
+        if (output.magic != kProgram203WitnessMagic) {
+            std::ostringstream detail;
+            detail << "completion: magic was 0x" << std::hex << output.magic << ", expected 0x"
+                   << kProgram203WitnessMagic;
+            return Failure(Program203WitnessValidationFailure::Completion, detail.str());
+        }
+
+        // 2. Observed topology. All checks consume observations written by the
+        // shader, rather than inferring subgroup layout from invocation indices.
+        const std::uint32_t numSubgroups = output.numSubgroups;
+        if (numSubgroups < 2u || numSubgroups > kProgram203WitnessMaxSubgroups) {
+            std::ostringstream detail;
+            detail << "topology: canonical gl_NumSubgroups=" << numSubgroups << " is outside [2, 32]";
+            return Failure(Program203WitnessValidationFailure::Topology, detail.str());
+        }
+        if ((output.topologyFlags & Program203WitnessNonuniformNumSubgroups) != 0u) {
+            return Failure(Program203WitnessValidationFailure::Topology,
+                           "topology: gl_NumSubgroups differed across workgroup");
+        }
+        if ((output.topologyFlags & Program203WitnessInvalidNumSubgroups) != 0u) {
+            return Failure(Program203WitnessValidationFailure::Topology,
+                           "topology: an invocation reported gl_NumSubgroups outside [2, 32]");
+        }
+        if ((output.topologyFlags & Program203WitnessInvalidSubgroupId) != 0u) {
+            return Failure(Program203WitnessValidationFailure::Topology,
+                           "topology: an invocation reported an invalid gl_SubgroupID");
+        }
+        if ((output.topologyFlags & Program203WitnessInvalidSubgroupLane) != 0u) {
+            return Failure(Program203WitnessValidationFailure::Topology,
+                           "topology: an invocation reported an invalid subgroup lane");
+        }
+        if ((output.topologyFlags & ~(Program203WitnessNonuniformNumSubgroups |
+                                      Program203WitnessInvalidNumSubgroups |
+                                      Program203WitnessInvalidSubgroupId |
+                                      Program203WitnessInvalidSubgroupLane)) != 0u) {
+            std::ostringstream detail;
+            detail << "topology: unknown topology flags 0x" << std::hex << output.topologyFlags;
+            return Failure(Program203WitnessValidationFailure::Topology, detail.str());
+        }
+        const std::uint32_t expectedMask = ExpectedSeenSubgroupMask(numSubgroups);
+        if (output.seenSubgroupMask != expectedMask) {
+            std::ostringstream detail;
+            detail << "topology: seen subgroup-ID mask was 0x" << std::hex << output.seenSubgroupMask
+                   << ", expected 0x" << expectedMask;
+            return Failure(Program203WitnessValidationFailure::Topology, detail.str());
+        }
+        const std::uint32_t expectedLoopLength = ComputeProgram203WitnessLoopLength(numSubgroups);
+        if (output.loopLength != expectedLoopLength) {
+            std::ostringstream detail;
+            detail << "topology: loopLength was " << std::dec << output.loopLength << ", expected "
+                   << expectedLoopLength;
+            return Failure(Program203WitnessValidationFailure::Topology, detail.str());
+        }
+        for (std::uint32_t subgroup = 0u; subgroup < numSubgroups; ++subgroup) {
+            if (output.lastLaneWriterCount[subgroup] != 1u) {
+                std::ostringstream detail;
+                detail << "topology: subgroup " << subgroup << " has "
+                       << output.lastLaneWriterCount[subgroup] << " source last-lane writers, expected exactly 1";
+                return Failure(Program203WitnessValidationFailure::Topology, detail.str(), 0u, subgroup);
+            }
+        }
+        if (output.owner511.y != numSubgroups) {
+            std::ostringstream detail;
+            detail << "final owner: invocation 511 reported gl_NumSubgroups=" << output.owner511.y << ", expected "
+                   << numSubgroups;
+            return Failure(Program203WitnessValidationFailure::FinalOwner, detail.str());
+        }
+        if (output.owner511.z != numSubgroups - 1u) {
+            std::ostringstream detail;
+            detail << "final owner: invocation 511 is not in the highest subgroup (id" << output.owner511.z
+                   << ", expected id" << (numSubgroups - 1u) << ')';
+            return Failure(Program203WitnessValidationFailure::FinalOwner, detail.str());
+        }
+        if (output.owner511.x == 0u || output.owner511.w != output.owner511.x - 1u) {
+            std::ostringstream detail;
+            detail << "final owner: invocation 511 is not the last lane of highest subgroup (size "
+                   << output.owner511.x << ", lane " << output.owner511.w << ')';
+            return Failure(Program203WitnessValidationFailure::FinalOwner, detail.str());
+        }
+
+        // 3. Initial subgroup handoff. The atomic scalar totals are independent
+        // of subgroupInclusiveAdd; their sum and the cache values establish that
+        // the final lanes handed off the native vector inclusive-add results.
+        std::uint64_t indexedTotal = 0u;
+        for (std::uint32_t subgroup = 0u; subgroup < numSubgroups; ++subgroup) {
+            indexedTotal += output.indexedInputTotal[subgroup];
+        }
+        if (indexedTotal != 131328u) {
+            std::ostringstream detail;
+            detail << "initial subgroup handoff: indexed input total was " << indexedTotal << ", expected 131328";
+            return Failure(Program203WitnessValidationFailure::InitialSubgroupHandoff, detail.str());
+        }
+        for (std::uint32_t subgroup = 0u; subgroup < numSubgroups; ++subgroup) {
+            const Program203WitnessVec2 expected = {static_cast<float>(output.indexedInputTotal[subgroup]), 0.0f};
+            if (!SameBits(output.rawPrefix[subgroup], expected)) {
+                std::ostringstream detail;
+                detail << "initial subgroup handoff: subgroup " << subgroup << " rawPrefix observed "
+                       << Vec2String(output.rawPrefix[subgroup]) << ", expected " << Vec2String(expected);
+                return Failure(Program203WitnessValidationFailure::InitialSubgroupHandoff, detail.str(), 0u,
+                               subgroup);
+            }
+        }
+
+        // 4. Source scan. Do not substitute a conventional scan: this reproduces
+        // the source cache index expression and stage ordering word for word.
+        std::array<Program203WitnessVec2, kProgram203WitnessMaxSubgroups> expectedCache = output.rawPrefix;
+        for (std::uint32_t scanStage = 0u; scanStage < expectedLoopLength; ++scanStage) {
+            auto cacheAfterStage = expectedCache;
+            for (std::uint32_t subgroup = 0u; subgroup < numSubgroups; ++subgroup) {
+                if ((subgroup & (1u << scanStage)) > 0u) {
+                    const std::uint32_t sourceCacheIndex = (subgroup >> scanStage << scanStage) - 1u;
+                    cacheAfterStage[subgroup].x += expectedCache[sourceCacheIndex].x;
+                    cacheAfterStage[subgroup].y += expectedCache[sourceCacheIndex].y;
+                }
+            }
+            expectedCache = cacheAfterStage;
+            for (std::uint32_t subgroup = 0u; subgroup < numSubgroups; ++subgroup) {
+                if (!SameBits(output.scanCache[scanStage][subgroup], expectedCache[subgroup])) {
+                    std::ostringstream detail;
+                    detail << "source scan stage " << scanStage << ", subgroup " << subgroup << ": observed "
+                           << Vec2String(output.scanCache[scanStage][subgroup]) << ", expected "
+                           << Vec2String(expectedCache[subgroup]);
+                    return Failure(Program203WitnessValidationFailure::SourceScan, detail.str(), scanStage, subgroup);
+                }
+            }
+        }
+
+        // 5. The owner contract was checked above with the other topology facts;
+        // this final result remains a separate exact-vector check.
+        const Program203WitnessVec2 expectedAverage = {256.5f, 0.0f};
+        if (!SameBits(output.finalAverage, expectedAverage)) {
+            std::ostringstream detail;
+            detail << "final average: observed " << Vec2String(output.finalAverage) << ", expected "
+                   << Vec2String(expectedAverage);
+            return Failure(Program203WitnessValidationFailure::FinalAverage, detail.str());
+        }
+
+        std::ostringstream detail;
+        detail << "N=" << numSubgroups << ", owner511=id" << output.owner511.z << "/lane" << output.owner511.w
+               << ", " << expectedLoopLength << " scan stages, average=" << Vec2String(output.finalAverage);
+        Program203WitnessValidationResult result;
+        result.ok = true;
+        result.failure = Program203WitnessValidationFailure::None;
+        result.detail = detail.str();
+        return result;
+    }
+} // namespace MobileGL::MG_Util::SelfTest

MobileGL/MG_Util/SelfTest/DriverPostProgram203Witness.h

@@ -0,0 +1,151 @@
+// MobileGL - MobileGL/MG_Util/SelfTest/DriverPostProgram203Witness.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
+// End of Source File Header
+//
+// Compact, native-Vulkan Program-203 first-reduction witness ABI and its pure
+// validator. The types below deliberately mirror DriverPostProgram203Witness.comp's
+// single std430 storage block; changing either side requires updating the static
+// layout assertions here.
+
+#pragma once
+
+#include <array>
+#include <cstddef>
+#include <cstdint>
+#include <string>
+#include <type_traits>
+
+namespace MobileGL::MG_Util::SelfTest {
+    constexpr std::uint32_t kProgram203WitnessMagic = 0x50323033u; // "P203"
+    constexpr std::uint32_t kProgram203WitnessInvocationCount = 512u;
+    constexpr std::uint32_t kProgram203WitnessMaxSubgroups = 32u;
+    constexpr std::uint32_t kProgram203WitnessMaxScanStages = 6u;
+
+    // These bit values are shared with the GLSL source. They document failures in
+    // topology observations rather than guessing a topology from local IDs on the host.
+    enum Program203WitnessTopologyFlag : std::uint32_t {
+        Program203WitnessNonuniformNumSubgroups = 1u << 0u,
+        Program203WitnessInvalidNumSubgroups = 1u << 1u,
+        Program203WitnessInvalidSubgroupId = 1u << 2u,
+        Program203WitnessInvalidSubgroupLane = 1u << 3u,
+    };
+
+    struct alignas(8) Program203WitnessVec2 {
+        float x;
+        float y;
+    };
+
+    struct alignas(16) Program203WitnessUVec4 {
+        std::uint32_t x;
+        std::uint32_t y;
+        std::uint32_t z;
+        std::uint32_t w;
+    };
+
+    // std430 layout of DriverPostProgram203Witness.comp's Program203WitnessOutput block.
+    struct alignas(16) Program203WitnessOutput {
+        std::uint32_t magic;
+        std::uint32_t topologyFlags;
+        std::uint32_t numSubgroups;
+        std::uint32_t loopLength;
+        std::uint32_t seenSubgroupMask;
+
+        Program203WitnessUVec4 owner511;
+
+        std::array<std::uint32_t, kProgram203WitnessMaxSubgroups> lastLaneWriterCount;
+        std::array<std::uint32_t, kProgram203WitnessMaxSubgroups> indexedInputTotal;
+
+        std::array<Program203WitnessVec2, kProgram203WitnessMaxSubgroups> rawPrefix;
+        std::array<std::array<Program203WitnessVec2, kProgram203WitnessMaxSubgroups>,
+                   kProgram203WitnessMaxScanStages>
+            scanCache;
+        Program203WitnessVec2 finalAverage;
+    };
+
+    static_assert(std::is_standard_layout_v<Program203WitnessVec2>);
+    static_assert(std::is_standard_layout_v<Program203WitnessUVec4>);
+    static_assert(std::is_standard_layout_v<Program203WitnessOutput>);
+    static_assert(sizeof(Program203WitnessVec2) == 8u);
+    static_assert(alignof(Program203WitnessVec2) == 8u);
+    static_assert(sizeof(Program203WitnessUVec4) == 16u);
+    static_assert(alignof(Program203WitnessUVec4) == 16u);
+    static_assert(offsetof(Program203WitnessOutput, magic) == 0u);
+    static_assert(offsetof(Program203WitnessOutput, topologyFlags) == 4u);
+    static_assert(offsetof(Program203WitnessOutput, numSubgroups) == 8u);
+    static_assert(offsetof(Program203WitnessOutput, loopLength) == 12u);
+    static_assert(offsetof(Program203WitnessOutput, seenSubgroupMask) == 16u);
+    static_assert(offsetof(Program203WitnessOutput, owner511) == 32u);
+    static_assert(offsetof(Program203WitnessOutput, lastLaneWriterCount) == 48u);
+    static_assert(offsetof(Program203WitnessOutput, indexedInputTotal) == 176u);
+    static_assert(offsetof(Program203WitnessOutput, rawPrefix) == 304u);
+    static_assert(offsetof(Program203WitnessOutput, scanCache) == 560u);
+    static_assert(offsetof(Program203WitnessOutput, finalAverage) == 2096u);
+    static_assert(sizeof(Program203WitnessOutput) == 2112u);
+
+    // The witness uses prefixSumCache[32], three scalar shared diagnostics, and
+    // two 32-entry scalar diagnostic arrays in the GLSL source. Keep this
+    // independent of the output SSBO size.
+    constexpr std::uint32_t kProgram203WitnessSharedMemoryBytes =
+        kProgram203WitnessMaxSubgroups * sizeof(Program203WitnessVec2) +
+        3u * sizeof(std::uint32_t) +
+        2u * kProgram203WitnessMaxSubgroups * sizeof(std::uint32_t);
+
+    enum class Program203WitnessEligibility {
+        Execute,
+        SkipUnsupportedNativeFeatureSet,
+        FailInadequateLimits,
+    };
+
+    // The raw physical-device conditions needed by the native witness. This is
+    // intentionally distinct from MobileGL's advertised-extension policy.
+    struct Program203WitnessLimits {
+        bool computeStageSupported = false;
+        bool basicSubgroupSupported = false;
+        bool arithmeticSubgroupSupported = false;
+        std::uint32_t subgroupSize = 0u;
+
+        std::uint32_t maxComputeWorkGroupInvocations = 0u;
+        std::array<std::uint32_t, 3> maxComputeWorkGroupSize{};
+        std::uint32_t maxComputeSharedMemorySize = 0u;
+        std::uint32_t maxPerStageDescriptorStorageBuffers = 0u;
+        std::uint32_t maxDescriptorSetStorageBuffers = 0u;
+        std::uint32_t maxBoundDescriptorSets = 0u;
+        std::uint64_t maxStorageBufferRange = 0u;
+    };
+
+    struct Program203WitnessEligibilityResult {
+        Program203WitnessEligibility eligibility = Program203WitnessEligibility::FailInadequateLimits;
+        std::string detail;
+    };
+
+    enum class Program203WitnessValidationFailure {
+        None,
+        Completion,
+        Topology,
+        InitialSubgroupHandoff,
+        SourceScan,
+        FinalOwner,
+        FinalAverage,
+    };
+
+    struct Program203WitnessValidationResult {
+        bool ok = false;
+        Program203WitnessValidationFailure failure = Program203WitnessValidationFailure::Completion;
+        std::uint32_t scanStage = 0u;
+        std::uint32_t subgroup = 0u;
+        std::string detail;
+    };
+
+    [[nodiscard]] Program203WitnessEligibilityResult
+    EvaluateProgram203WitnessEligibility(const Program203WitnessLimits& limits);
+
+    // Mirrors the source's findMSB expression for valid N in [2, 32].
+    [[nodiscard]] std::uint32_t ComputeProgram203WitnessLoopLength(std::uint32_t numSubgroups);
+
+    [[nodiscard]] Program203WitnessValidationResult
+    ValidateProgram203Witness(const Program203WitnessOutput& output);
+} // namespace MobileGL::MG_Util::SelfTest

MobileGL/MG_Util/SelfTest/DriverPostProgram203WitnessSpv.h

@@ -0,0 +1,291 @@
+// MobileGL - MobileGL/MG_Util/SelfTest/DriverPostProgram203WitnessSpv.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
+// End of Source File Header
+//
+// Generated from DriverPostProgram203Witness.comp with:
+// glslangValidator --target-env vulkan1.1 -V DriverPostProgram203Witness.comp
+// Validated with spirv-val --target-env vulkan1.1. Do not edit words by hand.
+
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+
+namespace MobileGL::MG_Util::SelfTest {
+    inline constexpr std::uint32_t kDriverPostProgram203WitnessSpv[] = {
+        0x07230203u, 0x00010300u, 0x0008000bu, 0x00000145u, 0x00000000u, 0x00020011u, 0x00000001u, 0x00020011u,
+        0x0000003du, 0x00020011u, 0x0000003fu, 0x0006000bu, 0x00000001u, 0x4c534c47u, 0x6474732eu, 0x3035342eu,
+        0x00000000u, 0x0003000eu, 0x00000000u, 0x00000001u, 0x000a000fu, 0x00000005u, 0x00000004u, 0x6e69616du,
+        0x00000000u, 0x0000000au, 0x0000002au, 0x00000050u, 0x0000005eu, 0x00000064u, 0x00060010u, 0x00000004u,
+        0x00000011u, 0x00000020u, 0x00000010u, 0x00000001u, 0x00030003u, 0x00000002u, 0x000001c2u, 0x000a0004u,
+        0x4b5f4c47u, 0x735f5248u, 0x65646168u, 0x75735f72u, 0x6f726762u, 0x615f7075u, 0x68746972u, 0x6974656du,
+        0x00000063u, 0x00090004u, 0x4b5f4c47u, 0x735f5248u, 0x65646168u, 0x75735f72u, 0x6f726762u, 0x625f7075u,
+        0x63697361u, 0x00000000u, 0x00040005u, 0x00000004u, 0x6e69616du, 0x00000000u, 0x00080005u, 0x00000008u,
+        0x61636f6cu, 0x766e496cu, 0x7461636fu, 0x496e6f69u, 0x7865646eu, 0x00000000u, 0x00080005u, 0x0000000au,
+        0x4c5f6c67u, 0x6c61636fu, 0x6f766e49u, 0x69746163u, 0x6e496e6fu, 0x00786564u, 0x00080005u, 0x00000013u,
+        0x6f6e6163u, 0x6163696eu, 0x6d754e6cu, 0x67627553u, 0x70756f72u, 0x00000073u, 0x00070005u, 0x00000014u,
+        0x6f706f74u, 0x79676f6cu, 0x67616c46u, 0x61685373u, 0x00646572u, 0x00080005u, 0x00000015u, 0x6e656573u,
+        0x67627553u, 0x70756f72u, 0x6b73614du, 0x72616853u, 0x00006465u, 0x00090005u, 0x0000001du, 0x7473616cu,
+        0x656e614cu, 0x74697257u, 0x6f437265u, 0x53746e75u, 0x65726168u, 0x00000064u, 0x00080005u, 0x00000020u,
+        0x65646e69u, 0x49646578u, 0x7475706eu, 0x61746f54u, 0x6168536cu, 0x00646572u, 0x00060005u, 0x0000002au,
+        0x4e5f6c67u, 0x75536d75u, 0x6f726762u, 0x00737075u, 0x00080005u, 0x00000035u, 0x676f7250u, 0x326d6172u,
+        0x69573330u, 0x73656e74u, 0x74754f73u, 0x00747570u, 0x00050006u, 0x00000035u, 0x00000000u, 0x6967616du,
+        0x00000063u, 0x00070006u, 0x00000035u, 0x00000001u, 0x6f706f74u, 0x79676f6cu, 0x67616c46u, 0x00000073u,
+        0x00070006u, 0x00000035u, 0x00000002u, 0x536d756eu, 0x72676275u, 0x7370756fu, 0x00000000u, 0x00060006u,
+        0x00000035u, 0x00000003u, 0x706f6f6cu, 0x676e654cu, 0x00006874u, 0x00080006u, 0x00000035u, 0x00000004u,
+        0x6e656573u, 0x67627553u, 0x70756f72u, 0x6b73614du, 0x00000000u, 0x00060006u, 0x00000035u, 0x00000005u,
+        0x656e776fu, 0x31313572u, 0x00000000u, 0x00080006u, 0x00000035u, 0x00000006u, 0x7473616cu, 0x656e614cu,
+        0x74697257u, 0x6f437265u, 0x00746e75u, 0x00080006u, 0x00000035u, 0x00000007u, 0x65646e69u, 0x49646578u,
+        0x7475706eu, 0x61746f54u, 0x0000006cu, 0x00060006u, 0x00000035u, 0x00000008u, 0x50776172u, 0x69666572u,
+        0x00000078u, 0x00060006u, 0x00000035u, 0x00000009u, 0x6e616373u, 0x68636143u, 0x00000065u, 0x00070006u,
+        0x00000035u, 0x0000000au, 0x616e6966u, 0x6576416cu, 0x65676172u, 0x00000000u, 0x00050005u, 0x00000037u,
+        0x5774756fu, 0x656e7469u, 0x00007373u, 0x00050005u, 0x0000003du, 0x6f6e6163u, 0x6163696eu, 0x00004e6cu,
+        0x00060005u, 0x00000050u, 0x535f6c67u, 0x72676275u, 0x4970756fu, 0x00000044u, 0x00060005u, 0x0000005eu,
+        0x535f6c67u, 0x72676275u, 0x5370756fu, 0x00657a69u, 0x00080005u, 0x00000064u, 0x535f6c67u, 0x72676275u,
+        0x4970756fu, 0x636f766eu, 0x6f697461u, 0x0044496eu, 0x00060005u, 0x0000006eu, 0x6f6e6163u, 0x6163696eu,
+        0x6d6f446cu, 0x006e6961u, 0x00070005u, 0x00000074u, 0x6e496469u, 0x6f6e6143u, 0x6163696eu, 0x6d6f446cu,
+        0x006e6961u, 0x00060005u, 0x000000acu, 0x72756f73u, 0x6f446563u, 0x6e69616du, 0x00000000u, 0x00060005u,
+        0x000000b7u, 0x706d6173u, 0x754c656cu, 0x616e696du, 0x0065636eu, 0x00060005u, 0x000000c8u, 0x66657270u,
+        0x75537869u, 0x6361436du, 0x00006568u, 0x00050005u, 0x000000d9u, 0x706f6f6cu, 0x676e654cu, 0x00006874u,
+        0x00050005u, 0x000000edu, 0x6e616373u, 0x67617453u, 0x00000065u, 0x00040047u, 0x0000000au, 0x0000000bu,
+        0x0000001du, 0x00040047u, 0x0000002au, 0x0000000bu, 0x00000026u, 0x00040047u, 0x0000002du, 0x00000006u,
+        0x00000004u, 0x00040047u, 0x0000002eu, 0x00000006u, 0x00000004u, 0x00040047u, 0x00000031u, 0x00000006u,
+        0x00000008u, 0x00040047u, 0x00000032u, 0x00000006u, 0x00000008u, 0x00040047u, 0x00000034u, 0x00000006u,
+        0x00000100u, 0x00030047u, 0x00000035u, 0x00000002u, 0x00050048u, 0x00000035u, 0x00000000u, 0x00000023u,
+        0x00000000u, 0x00050048u, 0x00000035u, 0x00000001u, 0x00000023u, 0x00000004u, 0x00050048u, 0x00000035u,
+        0x00000002u, 0x00000023u, 0x00000008u, 0x00050048u, 0x00000035u, 0x00000003u, 0x00000023u, 0x0000000cu,
+        0x00050048u, 0x00000035u, 0x00000004u, 0x00000023u, 0x00000010u, 0x00050048u, 0x00000035u, 0x00000005u,
+        0x00000023u, 0x00000020u, 0x00050048u, 0x00000035u, 0x00000006u, 0x00000023u, 0x00000030u, 0x00050048u,
+        0x00000035u, 0x00000007u, 0x00000023u, 0x000000b0u, 0x00050048u, 0x00000035u, 0x00000008u, 0x00000023u,
+        0x00000130u, 0x00050048u, 0x00000035u, 0x00000009u, 0x00000023u, 0x00000230u, 0x00050048u, 0x00000035u,
+        0x0000000au, 0x00000023u, 0x00000830u, 0x00040047u, 0x00000037u, 0x00000021u, 0x00000000u, 0x00040047u,
+        0x00000037u, 0x00000022u, 0x00000000u, 0x00040047u, 0x00000050u, 0x0000000bu, 0x00000028u, 0x00030047u,
+        0x0000005eu, 0x00000000u, 0x00040047u, 0x0000005eu, 0x0000000bu, 0x00000024u, 0x00030047u, 0x0000005fu,
+        0x00000000u, 0x00030047u, 0x00000064u, 0x00000000u, 0x00040047u, 0x00000064u, 0x0000000bu, 0x00000029u,
+        0x00030047u, 0x00000065u, 0x00000000u, 0x00030047u, 0x00000066u, 0x00000000u, 0x00030047u, 0x00000087u,
+        0x00000000u, 0x00030047u, 0x0000008bu, 0x00000000u, 0x00030047u, 0x0000008cu, 0x00000000u, 0x00030047u,
+        0x0000008du, 0x00000000u, 0x00030047u, 0x000000c0u, 0x00000000u, 0x00030047u, 0x000000c1u, 0x00000000u,
+        0x00030047u, 0x000000c2u, 0x00000000u, 0x00030047u, 0x00000107u, 0x00000000u, 0x00030047u, 0x00000108u,
+        0x00000000u, 0x00030047u, 0x00000109u, 0x00000000u, 0x00030047u, 0x0000012fu, 0x00000000u, 0x00030047u,
+        0x00000132u, 0x00000000u, 0x00040047u, 0x00000144u, 0x0000000bu, 0x00000019u, 0x00020013u, 0x00000002u,
+        0x00030021u, 0x00000003u, 0x00000002u, 0x00040015u, 0x00000006u, 0x00000020u, 0x00000000u, 0x00040020u,
+        0x00000007u, 0x00000007u, 0x00000006u, 0x00040020u, 0x00000009u, 0x00000001u, 0x00000006u, 0x0004003bu,
+        0x00000009u, 0x0000000au, 0x00000001u, 0x0004002bu, 0x00000006u, 0x0000000du, 0x00000000u, 0x00020014u,
+        0x0000000eu, 0x00040020u, 0x00000012u, 0x00000004u, 0x00000006u, 0x0004003bu, 0x00000012u, 0x00000013u,
+        0x00000004u, 0x0004003bu, 0x00000012u, 0x00000014u, 0x00000004u, 0x0004003bu, 0x00000012u, 0x00000015u,
+        0x00000004u, 0x0004002bu, 0x00000006u, 0x00000017u, 0x00000020u, 0x0004001cu, 0x0000001bu, 0x00000006u,
+        0x00000017u, 0x00040020u, 0x0000001cu, 0x00000004u, 0x0000001bu, 0x0004003bu, 0x0000001cu, 0x0000001du,
+        0x00000004u, 0x0004003bu, 0x0000001cu, 0x00000020u, 0x00000004u, 0x0004002bu, 0x00000006u, 0x00000023u,
+        0x00000001u, 0x0004002bu, 0x00000006u, 0x00000024u, 0x00000108u, 0x0004002bu, 0x00000006u, 0x00000025u,
+        0x00000002u, 0x0004003bu, 0x00000009u, 0x0000002au, 0x00000001u, 0x00040017u, 0x0000002cu, 0x00000006u,
+        0x00000004u, 0x0004001cu, 0x0000002du, 0x00000006u, 0x00000017u, 0x0004001cu, 0x0000002eu, 0x00000006u,
+        0x00000017u, 0x00030016u, 0x0000002fu, 0x00000020u, 0x00040017u, 0x00000030u, 0x0000002fu, 0x00000002u,
+        0x0004001cu, 0x00000031u, 0x00000030u, 0x00000017u, 0x0004001cu, 0x00000032u, 0x00000030u, 0x00000017u,
+        0x0004002bu, 0x00000006u, 0x00000033u, 0x00000006u, 0x0004001cu, 0x00000034u, 0x00000032u, 0x00000033u,
+        0x000d001eu, 0x00000035u, 0x00000006u, 0x00000006u, 0x00000006u, 0x00000006u, 0x00000006u, 0x0000002cu,
+        0x0000002du, 0x0000002eu, 0x00000031u, 0x00000034u, 0x00000030u, 0x00040020u, 0x00000036u, 0x0000000cu,
+        0x00000035u, 0x0004003bu, 0x00000036u, 0x00000037u, 0x0000000cu, 0x00040015u, 0x00000038u, 0x00000020u,
+        0x00000001u, 0x0004002bu, 0x00000038u, 0x00000039u, 0x00000002u, 0x00040020u, 0x0000003bu, 0x0000000cu,
+        0x00000006u, 0x0004003bu, 0x00000009u, 0x00000050u, 0x00000001u, 0x0004002bu, 0x00000006u, 0x0000005cu,
+        0x00000004u, 0x0004003bu, 0x00000009u, 0x0000005eu, 0x00000001u, 0x0004003bu, 0x00000009u, 0x00000064u,
+        0x00000001u, 0x0004002bu, 0x00000006u, 0x0000006bu, 0x00000008u, 0x00040020u, 0x0000006du, 0x00000007u,
+        0x0000000eu, 0x0004002bu, 0x00000038u, 0x00000099u, 0x00000004u, 0x0004002bu, 0x00000038u, 0x000000a0u,
+        0x00000006u, 0x0004002bu, 0x00000038u, 0x000000a6u, 0x00000007u, 0x00040020u, 0x000000b6u, 0x00000007u,
+        0x00000030u, 0x0004002bu, 0x0000002fu, 0x000000bbu, 0x00000000u, 0x0004002bu, 0x00000006u, 0x000000beu,
+        0x00000003u, 0x0004001cu, 0x000000c6u, 0x00000030u, 0x00000017u, 0x00040020u, 0x000000c7u, 0x00000004u,
+        0x000000c6u, 0x0004003bu, 0x000000c7u, 0x000000c8u, 0x00000004u, 0x00040020u, 0x000000cbu, 0x00000004u,
+        0x00000030u, 0x0004002bu, 0x00000038u, 0x000000d2u, 0x00000008u, 0x00040020u, 0x000000d7u, 0x0000000cu,
+        0x00000030u, 0x0004002bu, 0x00000038u, 0x000000eau, 0x00000003u, 0x0004002bu, 0x00000038u, 0x00000115u,
+        0x00000009u, 0x0004002bu, 0x00000038u, 0x0000011du, 0x00000001u, 0x0004002bu, 0x00000006u, 0x00000120u,
+        0x000001ffu, 0x0004002bu, 0x00000038u, 0x00000124u, 0x00000000u, 0x0004002bu, 0x0000002fu, 0x00000126u,
+        0x44000000u, 0x0004002bu, 0x00000038u, 0x0000012eu, 0x00000005u, 0x00040020u, 0x00000134u, 0x0000000cu,
+        0x0000002cu, 0x0004002bu, 0x00000038u, 0x00000136u, 0x0000000au, 0x0004002bu, 0x00000006u, 0x00000140u,
+        0x50323033u, 0x00040017u, 0x00000142u, 0x00000006u, 0x00000003u, 0x0004002bu, 0x00000006u, 0x00000143u,
+        0x00000010u, 0x0006002cu, 0x00000142u, 0x00000144u, 0x00000017u, 0x00000143u, 0x00000023u, 0x00050036u,
+        0x00000002u, 0x00000004u, 0x00000000u, 0x00000003u, 0x000200f8u, 0x00000005u, 0x0004003bu, 0x00000007u,
+        0x00000008u, 0x00000007u, 0x0004003bu, 0x00000007u, 0x0000003du, 0x00000007u, 0x0004003bu, 0x0000006du,
+        0x0000006eu, 0x00000007u, 0x0004003bu, 0x0000006du, 0x00000074u, 0x00000007u, 0x0004003bu, 0x0000006du,
+        0x000000acu, 0x00000007u, 0x0004003bu, 0x000000b6u, 0x000000b7u, 0x00000007u, 0x0004003bu, 0x00000007u,
+        0x000000d9u, 0x00000007u, 0x0004003bu, 0x00000007u, 0x000000edu, 0x00000007u, 0x0004003du, 0x00000006u,
+        0x0000000bu, 0x0000000au, 0x0003003eu, 0x00000008u, 0x0000000bu, 0x0004003du, 0x00000006u, 0x0000000cu,
+        0x00000008u, 0x000500aau, 0x0000000eu, 0x0000000fu, 0x0000000cu, 0x0000000du, 0x000300f7u, 0x00000011u,
+        0x00000000u, 0x000400fau, 0x0000000fu, 0x00000010u, 0x00000011u, 0x000200f8u, 0x00000010u, 0x0003003eu,
+        0x00000013u, 0x0000000du, 0x0003003eu, 0x00000014u, 0x0000000du, 0x0003003eu, 0x00000015u, 0x0000000du,
+        0x000200f9u, 0x00000011u, 0x000200f8u, 0x00000011u, 0x0004003du, 0x00000006u, 0x00000016u, 0x00000008u,
+        0x000500b0u, 0x0000000eu, 0x00000018u, 0x00000016u, 0x00000017u, 0x000300f7u, 0x0000001au, 0x00000000u,
+        0x000400fau, 0x00000018u, 0x00000019u, 0x0000001au, 0x000200f8u, 0x00000019u, 0x0004003du, 0x00000006u,
+        0x0000001eu, 0x00000008u, 0x00050041u, 0x00000012u, 0x0000001fu, 0x0000001du, 0x0000001eu, 0x0003003eu,
+        0x0000001fu, 0x0000000du, 0x0004003du, 0x00000006u, 0x00000021u, 0x00000008u, 0x00050041u, 0x00000012u,
+        0x00000022u, 0x00000020u, 0x00000021u, 0x0003003eu, 0x00000022u, 0x0000000du, 0x000200f9u, 0x0000001au,
+        0x000200f8u, 0x0000001au, 0x000300e1u, 0x00000023u, 0x00000024u, 0x000400e0u, 0x00000025u, 0x00000025u,
+        0x00000024u, 0x0004003du, 0x00000006u, 0x00000026u, 0x00000008u, 0x000500aau, 0x0000000eu, 0x00000027u,
+        0x00000026u, 0x0000000du, 0x000300f7u, 0x00000029u, 0x00000000u, 0x000400fau, 0x00000027u, 0x00000028u,
+        0x00000029u, 0x000200f8u, 0x00000028u, 0x0004003du, 0x00000006u, 0x0000002bu, 0x0000002au, 0x0003003eu,
+        0x00000013u, 0x0000002bu, 0x0004003du, 0x00000006u, 0x0000003au, 0x0000002au, 0x00050041u, 0x0000003bu,
+        0x0000003cu, 0x00000037u, 0x00000039u, 0x0003003eu, 0x0000003cu, 0x0000003au, 0x000200f9u, 0x00000029u,
+        0x000200f8u, 0x00000029u, 0x000400e0u, 0x00000025u, 0x00000025u, 0x00000024u, 0x0004003du, 0x00000006u,
+        0x0000003eu, 0x00000013u, 0x0003003eu, 0x0000003du, 0x0000003eu, 0x0004003du, 0x00000006u, 0x0000003fu,
+        0x0000002au, 0x0004003du, 0x00000006u, 0x00000040u, 0x0000003du, 0x000500abu, 0x0000000eu, 0x00000041u,
+        0x0000003fu, 0x00000040u, 0x000300f7u, 0x00000043u, 0x00000000u, 0x000400fau, 0x00000041u, 0x00000042u,
+        0x00000043u, 0x000200f8u, 0x00000042u, 0x000700f1u, 0x00000006u, 0x00000044u, 0x00000014u, 0x00000023u,
+        0x0000000du, 0x00000023u, 0x000200f9u, 0x00000043u, 0x000200f8u, 0x00000043u, 0x0004003du, 0x00000006u,
+        0x00000045u, 0x0000002au, 0x000500b0u, 0x0000000eu, 0x00000046u, 0x00000045u, 0x00000025u, 0x000400a8u,
+        0x0000000eu, 0x00000047u, 0x00000046u, 0x000300f7u, 0x00000049u, 0x00000000u, 0x000400fau, 0x00000047u,
+        0x00000048u, 0x00000049u, 0x000200f8u, 0x00000048u, 0x0004003du, 0x00000006u, 0x0000004au, 0x0000002au,
+        0x000500acu, 0x0000000eu, 0x0000004bu, 0x0000004au, 0x00000017u, 0x000200f9u, 0x00000049u, 0x000200f8u,
+        0x00000049u, 0x000700f5u, 0x0000000eu, 0x0000004cu, 0x00000046u, 0x00000043u, 0x0000004bu, 0x00000048u,
+        0x000300f7u, 0x0000004eu, 0x00000000u, 0x000400fau, 0x0000004cu, 0x0000004du, 0x0000004eu, 0x000200f8u,
+        0x0000004du, 0x000700f1u, 0x00000006u, 0x0000004fu, 0x00000014u, 0x00000023u, 0x0000000du, 0x00000025u,
+        0x000200f9u, 0x0000004eu, 0x000200f8u, 0x0000004eu, 0x0004003du, 0x00000006u, 0x00000051u, 0x00000050u,
+        0x0004003du, 0x00000006u, 0x00000052u, 0x0000003du, 0x000500aeu, 0x0000000eu, 0x00000053u, 0x00000051u,
+        0x00000052u, 0x000400a8u, 0x0000000eu, 0x00000054u, 0x00000053u, 0x000300f7u, 0x00000056u, 0x00000000u,
+        0x000400fau, 0x00000054u, 0x00000055u, 0x00000056u, 0x000200f8u, 0x00000055u, 0x0004003du, 0x00000006u,
+        0x00000057u, 0x00000050u, 0x000500aeu, 0x0000000eu, 0x00000058u, 0x00000057u, 0x00000017u, 0x000200f9u,
+        0x00000056u, 0x000200f8u, 0x00000056u, 0x000700f5u, 0x0000000eu, 0x00000059u, 0x00000053u, 0x0000004eu,
+        0x00000058u, 0x00000055u, 0x000300f7u, 0x0000005bu, 0x00000000u, 0x000400fau, 0x00000059u, 0x0000005au,
+        0x0000005bu, 0x000200f8u, 0x0000005au, 0x000700f1u, 0x00000006u, 0x0000005du, 0x00000014u, 0x00000023u,
+        0x0000000du, 0x0000005cu, 0x000200f9u, 0x0000005bu, 0x000200f8u, 0x0000005bu, 0x0004003du, 0x00000006u,
+        0x0000005fu, 0x0000005eu, 0x000500aau, 0x0000000eu, 0x00000060u, 0x0000005fu, 0x0000000du, 0x000400a8u,
+        0x0000000eu, 0x00000061u, 0x00000060u, 0x000300f7u, 0x00000063u, 0x00000000u, 0x000400fau, 0x00000061u,
+        0x00000062u, 0x00000063u, 0x000200f8u, 0x00000062u, 0x0004003du, 0x00000006u, 0x00000065u, 0x00000064u,
+        0x0004003du, 0x00000006u, 0x00000066u, 0x0000005eu, 0x000500aeu, 0x0000000eu, 0x00000067u, 0x00000065u,
+        0x00000066u, 0x000200f9u, 0x00000063u, 0x000200f8u, 0x00000063u, 0x000700f5u, 0x0000000eu, 0x00000068u,
+        0x00000060u, 0x0000005bu, 0x00000067u, 0x00000062u, 0x000300f7u, 0x0000006au, 0x00000000u, 0x000400fau,
+        0x00000068u, 0x00000069u, 0x0000006au, 0x000200f8u, 0x00000069u, 0x000700f1u, 0x00000006u, 0x0000006cu,
+        0x00000014u, 0x00000023u, 0x0000000du, 0x0000006bu, 0x000200f9u, 0x0000006au, 0x000200f8u, 0x0000006au,
+        0x0004003du, 0x00000006u, 0x0000006fu, 0x0000003du, 0x000500aeu, 0x0000000eu, 0x00000070u, 0x0000006fu,
+        0x00000025u, 0x0004003du, 0x00000006u, 0x00000071u, 0x0000003du, 0x000500b2u, 0x0000000eu, 0x00000072u,
+        0x00000071u, 0x00000017u, 0x000500a7u, 0x0000000eu, 0x00000073u, 0x00000070u, 0x00000072u, 0x0003003eu,
+        0x0000006eu, 0x00000073u, 0x0004003du, 0x0000000eu, 0x00000075u, 0x0000006eu, 0x000300f7u, 0x00000077u,
+        0x00000000u, 0x000400fau, 0x00000075u, 0x00000076u, 0x00000077u, 0x000200f8u, 0x00000076u, 0x0004003du,
+        0x00000006u, 0x00000078u, 0x00000050u, 0x0004003du, 0x00000006u, 0x00000079u, 0x0000003du, 0x000500b0u,
+        0x0000000eu, 0x0000007au, 0x00000078u, 0x00000079u, 0x000200f9u, 0x00000077u, 0x000200f8u, 0x00000077u,
+        0x000700f5u, 0x0000000eu, 0x0000007bu, 0x00000075u, 0x0000006au, 0x0000007au, 0x00000076u, 0x0003003eu,
+        0x00000074u, 0x0000007bu, 0x0004003du, 0x0000000eu, 0x0000007cu, 0x00000074u, 0x000300f7u, 0x0000007eu,
+        0x00000000u, 0x000400fau, 0x0000007cu, 0x0000007du, 0x0000007eu, 0x000200f8u, 0x0000007du, 0x0004003du,
+        0x00000006u, 0x0000007fu, 0x00000050u, 0x000500c4u, 0x00000006u, 0x00000080u, 0x00000023u, 0x0000007fu,
+        0x000700f1u, 0x00000006u, 0x00000081u, 0x00000015u, 0x00000023u, 0x0000000du, 0x00000080u, 0x0004003du,
+        0x00000006u, 0x00000082u, 0x00000050u, 0x00050041u, 0x00000012u, 0x00000083u, 0x00000020u, 0x00000082u,
+        0x0004003du, 0x00000006u, 0x00000084u, 0x00000008u, 0x00050080u, 0x00000006u, 0x00000085u, 0x00000084u,
+        0x00000023u, 0x000700eau, 0x00000006u, 0x00000086u, 0x00000083u, 0x00000023u, 0x0000000du, 0x00000085u,
+        0x0004003du, 0x00000006u, 0x00000087u, 0x0000005eu, 0x000500abu, 0x0000000eu, 0x00000088u, 0x00000087u,
+        0x0000000du, 0x000300f7u, 0x0000008au, 0x00000000u, 0x000400fau, 0x00000088u, 0x00000089u, 0x0000008au,
+        0x000200f8u, 0x00000089u, 0x0004003du, 0x00000006u, 0x0000008bu, 0x00000064u, 0x0004003du, 0x00000006u,
+        0x0000008cu, 0x0000005eu, 0x00050082u, 0x00000006u, 0x0000008du, 0x0000008cu, 0x00000023u, 0x000500aau,
+        0x0000000eu, 0x0000008eu, 0x0000008bu, 0x0000008du, 0x000200f9u, 0x0000008au, 0x000200f8u, 0x0000008au,
+        0x000700f5u, 0x0000000eu, 0x0000008fu, 0x00000088u, 0x0000007du, 0x0000008eu, 0x00000089u, 0x000300f7u,
+        0x00000091u, 0x00000000u, 0x000400fau, 0x0000008fu, 0x00000090u, 0x00000091u, 0x000200f8u, 0x00000090u,
+        0x0004003du, 0x00000006u, 0x00000092u, 0x00000050u, 0x00050041u, 0x00000012u, 0x00000093u, 0x0000001du,
+        0x00000092u, 0x000700eau, 0x00000006u, 0x00000094u, 0x00000093u, 0x00000023u, 0x0000000du, 0x00000023u,
+        0x000200f9u, 0x00000091u, 0x000200f8u, 0x00000091u, 0x000200f9u, 0x0000007eu, 0x000200f8u, 0x0000007eu,
+        0x000300e1u, 0x00000023u, 0x00000024u, 0x000400e0u, 0x00000025u, 0x00000025u, 0x00000024u, 0x0004003du,
+        0x00000006u, 0x00000095u, 0x00000008u, 0x000500aau, 0x0000000eu, 0x00000096u, 0x00000095u, 0x0000000du,
+        0x000300f7u, 0x00000098u, 0x00000000u, 0x000400fau, 0x00000096u, 0x00000097u, 0x00000098u, 0x000200f8u,
+        0x00000097u, 0x0004003du, 0x00000006u, 0x0000009au, 0x00000015u, 0x00050041u, 0x0000003bu, 0x0000009bu,
+        0x00000037u, 0x00000099u, 0x0003003eu, 0x0000009bu, 0x0000009au, 0x000200f9u, 0x00000098u, 0x000200f8u,
+        0x00000098u, 0x0004003du, 0x00000006u, 0x0000009cu, 0x00000008u, 0x000500b0u, 0x0000000eu, 0x0000009du,
+        0x0000009cu, 0x00000017u, 0x000300f7u, 0x0000009fu, 0x00000000u, 0x000400fau, 0x0000009du, 0x0000009eu,
+        0x0000009fu, 0x000200f8u, 0x0000009eu, 0x0004003du, 0x00000006u, 0x000000a1u, 0x00000008u, 0x0004003du,
+        0x00000006u, 0x000000a2u, 0x00000008u, 0x00050041u, 0x00000012u, 0x000000a3u, 0x0000001du, 0x000000a2u,
+        0x0004003du, 0x00000006u, 0x000000a4u, 0x000000a3u, 0x00060041u, 0x0000003bu, 0x000000a5u, 0x00000037u,
+        0x000000a0u, 0x000000a1u, 0x0003003eu, 0x000000a5u, 0x000000a4u, 0x0004003du, 0x00000006u, 0x000000a7u,
+        0x00000008u, 0x0004003du, 0x00000006u, 0x000000a8u, 0x00000008u, 0x00050041u, 0x00000012u, 0x000000a9u,
+        0x00000020u, 0x000000a8u, 0x0004003du, 0x00000006u, 0x000000aau, 0x000000a9u, 0x00060041u, 0x0000003bu,
+        0x000000abu, 0x00000037u, 0x000000a6u, 0x000000a7u, 0x0003003eu, 0x000000abu, 0x000000aau, 0x000200f9u,
+        0x0000009fu, 0x000200f8u, 0x0000009fu, 0x0004003du, 0x0000000eu, 0x000000adu, 0x0000006eu, 0x000300f7u,
+        0x000000afu, 0x00000000u, 0x000400fau, 0x000000adu, 0x000000aeu, 0x000000afu, 0x000200f8u, 0x000000aeu,
+        0x0004003du, 0x00000006u, 0x000000b0u, 0x00000014u, 0x000500aau, 0x0000000eu, 0x000000b1u, 0x000000b0u,
+        0x0000000du, 0x000200f9u, 0x000000afu, 0x000200f8u, 0x000000afu, 0x000700f5u, 0x0000000eu, 0x000000b2u,
+        0x000000adu, 0x0000009fu, 0x000000b1u, 0x000000aeu, 0x0003003eu, 0x000000acu, 0x000000b2u, 0x0004003du,
+        0x0000000eu, 0x000000b3u, 0x000000acu, 0x000300f7u, 0x000000b5u, 0x00000000u, 0x000400fau, 0x000000b3u,
+        0x000000b4u, 0x000000b5u, 0x000200f8u, 0x000000b4u, 0x0004003du, 0x00000006u, 0x000000b8u, 0x0000000au,
+        0x00050080u, 0x00000006u, 0x000000b9u, 0x000000b8u, 0x00000023u, 0x00040070u, 0x0000002fu, 0x000000bau,
+        0x000000b9u, 0x00050050u, 0x00000030u, 0x000000bcu, 0x000000bau, 0x000000bbu, 0x0003003eu, 0x000000b7u,
+        0x000000bcu, 0x0004003du, 0x00000030u, 0x000000bdu, 0x000000b7u, 0x0006015eu, 0x00000030u, 0x000000bfu,
+        0x000000beu, 0x00000001u, 0x000000bdu, 0x0003003eu, 0x000000b7u, 0x000000bfu, 0x0004003du, 0x00000006u,
+        0x000000c0u, 0x00000064u, 0x0004003du, 0x00000006u, 0x000000c1u, 0x0000005eu, 0x00050082u, 0x00000006u,
+        0x000000c2u, 0x000000c1u, 0x00000023u, 0x000500aau, 0x0000000eu, 0x000000c3u, 0x000000c0u, 0x000000c2u,
+        0x000300f7u, 0x000000c5u, 0x00000000u, 0x000400fau, 0x000000c3u, 0x000000c4u, 0x000000c5u, 0x000200f8u,
+        0x000000c4u, 0x0004003du, 0x00000006u, 0x000000c9u, 0x00000050u, 0x0004003du, 0x00000030u, 0x000000cau,
+        0x000000b7u, 0x00050041u, 0x000000cbu, 0x000000ccu, 0x000000c8u, 0x000000c9u, 0x0003003eu, 0x000000ccu,
+        0x000000cau, 0x000200f9u, 0x000000c5u, 0x000200f8u, 0x000000c5u, 0x000400e0u, 0x00000025u, 0x00000025u,
+        0x00000024u, 0x0004003du, 0x00000006u, 0x000000cdu, 0x0000000au, 0x0004003du, 0x00000006u, 0x000000ceu,
+        0x0000002au, 0x000500b0u, 0x0000000eu, 0x000000cfu, 0x000000cdu, 0x000000ceu, 0x000300f7u, 0x000000d1u,
+        0x00000000u, 0x000400fau, 0x000000cfu, 0x000000d0u, 0x000000d1u, 0x000200f8u, 0x000000d0u, 0x0004003du,
+        0x00000006u, 0x000000d3u, 0x0000000au, 0x0004003du, 0x00000006u, 0x000000d4u, 0x0000000au, 0x00050041u,
+        0x000000cbu, 0x000000d5u, 0x000000c8u, 0x000000d4u, 0x0004003du, 0x00000030u, 0x000000d6u, 0x000000d5u,
+        0x00060041u, 0x000000d7u, 0x000000d8u, 0x00000037u, 0x000000d2u, 0x000000d3u, 0x0003003eu, 0x000000d8u,
+        0x000000d6u, 0x000200f9u, 0x000000d1u, 0x000200f8u, 0x000000d1u, 0x000400e0u, 0x00000025u, 0x00000025u,
+        0x00000024u, 0x0004003du, 0x00000006u, 0x000000dau, 0x0000002au, 0x0006000cu, 0x00000038u, 0x000000dbu,
+        0x00000001u, 0x0000004bu, 0x000000dau, 0x0004007cu, 0x00000006u, 0x000000dcu, 0x000000dbu, 0x0003003eu,
+        0x000000d9u, 0x000000dcu, 0x0004003du, 0x00000006u, 0x000000ddu, 0x0000002au, 0x0004003du, 0x00000006u,
+        0x000000deu, 0x000000d9u, 0x00050082u, 0x00000006u, 0x000000dfu, 0x000000deu, 0x00000023u, 0x000500c4u,
+        0x00000006u, 0x000000e0u, 0x00000023u, 0x000000dfu, 0x00050082u, 0x00000006u, 0x000000e1u, 0x000000ddu,
+        0x000000e0u, 0x000500acu, 0x0000000eu, 0x000000e2u, 0x000000e1u, 0x0000000du, 0x000600a9u, 0x00000006u,
+        0x000000e3u, 0x000000e2u, 0x00000023u, 0x0000000du, 0x0004003du, 0x00000006u, 0x000000e4u, 0x000000d9u,
+        0x00050080u, 0x00000006u, 0x000000e5u, 0x000000e4u, 0x000000e3u, 0x0003003eu, 0x000000d9u, 0x000000e5u,
+        0x0004003du, 0x00000006u, 0x000000e6u, 0x0000000au, 0x000500aau, 0x0000000eu, 0x000000e7u, 0x000000e6u,
+        0x0000000du, 0x000300f7u, 0x000000e9u, 0x00000000u, 0x000400fau, 0x000000e7u, 0x000000e8u, 0x000000e9u,
+        0x000200f8u, 0x000000e8u, 0x0004003du, 0x00000006u, 0x000000ebu, 0x000000d9u, 0x00050041u, 0x0000003bu,
+        0x000000ecu, 0x00000037u, 0x000000eau, 0x0003003eu, 0x000000ecu, 0x000000ebu, 0x000200f9u, 0x000000e9u,
+        0x000200f8u, 0x000000e9u, 0x0003003eu, 0x000000edu, 0x0000000du, 0x000200f9u, 0x000000eeu, 0x000200f8u,
+        0x000000eeu, 0x000400f6u, 0x000000f0u, 0x000000f1u, 0x00000000u, 0x000200f9u, 0x000000f2u, 0x000200f8u,
+        0x000000f2u, 0x0004003du, 0x00000006u, 0x000000f3u, 0x000000edu, 0x0004003du, 0x00000006u, 0x000000f4u,
+        0x000000d9u, 0x000500b0u, 0x0000000eu, 0x000000f5u, 0x000000f3u, 0x000000f4u, 0x000400fau, 0x000000f5u,
+        0x000000efu, 0x000000f0u, 0x000200f8u, 0x000000efu, 0x0004003du, 0x00000006u, 0x000000f6u, 0x00000050u,
+        0x0004003du, 0x00000006u, 0x000000f7u, 0x000000edu, 0x000500c4u, 0x00000006u, 0x000000f8u, 0x00000023u,
+        0x000000f7u, 0x000500c7u, 0x00000006u, 0x000000f9u, 0x000000f6u, 0x000000f8u, 0x000500acu, 0x0000000eu,
+        0x000000fau, 0x000000f9u, 0x0000000du, 0x000300f7u, 0x000000fcu, 0x00000000u, 0x000400fau, 0x000000fau,
+        0x000000fbu, 0x000000fcu, 0x000200f8u, 0x000000fbu, 0x0004003du, 0x00000006u, 0x000000fdu, 0x00000050u,
+        0x0004003du, 0x00000006u, 0x000000feu, 0x000000edu, 0x000500c2u, 0x00000006u, 0x000000ffu, 0x000000fdu,
+        0x000000feu, 0x0004003du, 0x00000006u, 0x00000100u, 0x000000edu, 0x000500c4u, 0x00000006u, 0x00000101u,
+        0x000000ffu, 0x00000100u, 0x00050082u, 0x00000006u, 0x00000102u, 0x00000101u, 0x00000023u, 0x00050041u,
+        0x000000cbu, 0x00000103u, 0x000000c8u, 0x00000102u, 0x0004003du, 0x00000030u, 0x00000104u, 0x00000103u,
+        0x0004003du, 0x00000030u, 0x00000105u, 0x000000b7u, 0x00050081u, 0x00000030u, 0x00000106u, 0x00000105u,
+        0x00000104u, 0x0003003eu, 0x000000b7u, 0x00000106u, 0x0004003du, 0x00000006u, 0x00000107u, 0x00000064u,
+        0x0004003du, 0x00000006u, 0x00000108u, 0x0000005eu, 0x00050082u, 0x00000006u, 0x00000109u, 0x00000108u,
+        0x00000023u, 0x000500aau, 0x0000000eu, 0x0000010au, 0x00000107u, 0x00000109u, 0x000300f7u, 0x0000010cu,
+        0x00000000u, 0x000400fau, 0x0000010au, 0x0000010bu, 0x0000010cu, 0x000200f8u, 0x0000010bu, 0x0004003du,
+        0x00000006u, 0x0000010du, 0x00000050u, 0x0004003du, 0x00000030u, 0x0000010eu, 0x000000b7u, 0x00050041u,
+        0x000000cbu, 0x0000010fu, 0x000000c8u, 0x0000010du, 0x0003003eu, 0x0000010fu, 0x0000010eu, 0x000200f9u,
+        0x0000010cu, 0x000200f8u, 0x0000010cu, 0x000200f9u, 0x000000fcu, 0x000200f8u, 0x000000fcu, 0x000400e0u,
+        0x00000025u, 0x00000025u, 0x00000024u, 0x0004003du, 0x00000006u, 0x00000110u, 0x0000000au, 0x0004003du,
+        0x00000006u, 0x00000111u, 0x0000002au, 0x000500b0u, 0x0000000eu, 0x00000112u, 0x00000110u, 0x00000111u,
+        0x000300f7u, 0x00000114u, 0x00000000u, 0x000400fau, 0x00000112u, 0x00000113u, 0x00000114u, 0x000200f8u,
+        0x00000113u, 0x0004003du, 0x00000006u, 0x00000116u, 0x000000edu, 0x0004003du, 0x00000006u, 0x00000117u,
+        0x0000000au, 0x0004003du, 0x00000006u, 0x00000118u, 0x0000000au, 0x00050041u, 0x000000cbu, 0x00000119u,
+        0x000000c8u, 0x00000118u, 0x0004003du, 0x00000030u, 0x0000011au, 0x00000119u, 0x00070041u, 0x000000d7u,
+        0x0000011bu, 0x00000037u, 0x00000115u, 0x00000116u, 0x00000117u, 0x0003003eu, 0x0000011bu, 0x0000011au,
+        0x000200f9u, 0x00000114u, 0x000200f8u, 0x00000114u, 0x000400e0u, 0x00000025u, 0x00000025u, 0x00000024u,
+        0x000200f9u, 0x000000f1u, 0x000200f8u, 0x000000f1u, 0x0004003du, 0x00000006u, 0x0000011cu, 0x000000edu,
+        0x00050080u, 0x00000006u, 0x0000011eu, 0x0000011cu, 0x0000011du, 0x0003003eu, 0x000000edu, 0x0000011eu,
+        0x000200f9u, 0x000000eeu, 0x000200f8u, 0x000000f0u, 0x0004003du, 0x00000006u, 0x0000011fu, 0x0000000au,
+        0x000500aau, 0x0000000eu, 0x00000121u, 0x0000011fu, 0x00000120u, 0x000300f7u, 0x00000123u, 0x00000000u,
+        0x000400fau, 0x00000121u, 0x00000122u, 0x00000123u, 0x000200f8u, 0x00000122u, 0x0004003du, 0x00000030u,
+        0x00000125u, 0x000000b7u, 0x00050050u, 0x00000030u, 0x00000127u, 0x00000126u, 0x00000126u, 0x00050088u,
+        0x00000030u, 0x00000128u, 0x00000125u, 0x00000127u, 0x00050041u, 0x000000cbu, 0x00000129u, 0x000000c8u,
+        0x00000124u, 0x0003003eu, 0x00000129u, 0x00000128u, 0x000200f9u, 0x00000123u, 0x000200f8u, 0x00000123u,
+        0x000400e0u, 0x00000025u, 0x00000025u, 0x00000024u, 0x0004003du, 0x00000006u, 0x0000012au, 0x0000000au,
+        0x000500aau, 0x0000000eu, 0x0000012bu, 0x0000012au, 0x00000120u, 0x000300f7u, 0x0000012du, 0x00000000u,
+        0x000400fau, 0x0000012bu, 0x0000012cu, 0x0000012du, 0x000200f8u, 0x0000012cu, 0x0004003du, 0x00000006u,
+        0x0000012fu, 0x0000005eu, 0x0004003du, 0x00000006u, 0x00000130u, 0x0000002au, 0x0004003du, 0x00000006u,
+        0x00000131u, 0x00000050u, 0x0004003du, 0x00000006u, 0x00000132u, 0x00000064u, 0x00070050u, 0x0000002cu,
+        0x00000133u, 0x0000012fu, 0x00000130u, 0x00000131u, 0x00000132u, 0x00050041u, 0x00000134u, 0x00000135u,
+        0x00000037u, 0x0000012eu, 0x0003003eu, 0x00000135u, 0x00000133u, 0x00050041u, 0x000000cbu, 0x00000137u,
+        0x000000c8u, 0x00000124u, 0x0004003du, 0x00000030u, 0x00000138u, 0x00000137u, 0x00050041u, 0x000000d7u,
+        0x00000139u, 0x00000037u, 0x00000136u, 0x0003003eu, 0x00000139u, 0x00000138u, 0x000200f9u, 0x0000012du,
+        0x000200f8u, 0x0000012du, 0x000200f9u, 0x000000b5u, 0x000200f8u, 0x000000b5u, 0x000400e0u, 0x00000025u,
+        0x00000025u, 0x00000024u, 0x0004003du, 0x00000006u, 0x0000013au, 0x00000008u, 0x000500aau, 0x0000000eu,
+        0x0000013bu, 0x0000013au, 0x0000000du, 0x000300f7u, 0x0000013du, 0x00000000u, 0x000400fau, 0x0000013bu,
+        0x0000013cu, 0x0000013du, 0x000200f8u, 0x0000013cu, 0x0004003du, 0x00000006u, 0x0000013eu, 0x00000014u,
+        0x00050041u, 0x0000003bu, 0x0000013fu, 0x00000037u, 0x0000011du, 0x0003003eu, 0x0000013fu, 0x0000013eu,
+        0x00050041u, 0x0000003bu, 0x00000141u, 0x00000037u, 0x00000124u, 0x0003003eu, 0x00000141u, 0x00000140u,
+        0x000200f9u, 0x0000013du, 0x000200f8u, 0x0000013du, 0x000100fdu, 0x00010038u,
+    };
+    inline constexpr std::size_t kDriverPostProgram203WitnessSpvWordCount =
+        sizeof(kDriverPostProgram203WitnessSpv) / sizeof(kDriverPostProgram203WitnessSpv[0]);
+} // namespace MobileGL::MG_Util::SelfTest

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

@@ -25,6 +25,7 @@
 #include "SpirvPasses/SplitArrayVertexInputsPass.h"
 #include "SpirvPasses/RebaseInstanceIndexPass.h"
 #include "SpirvPasses/ZeroBaseVertexPass.h"
+#include "SpirvPasses/DeriveNumSubgroupsPass.h"
 #include "SpirvPasses/NormalizeRectCoordinatesPass.h"
 #include "SpirvPasses/Lower1DArrayImagesPass.h"
 #include "SpirvPasses/BakeImageFormatsPass.h"
@@ -369,12 +370,6 @@ 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 +378,6 @@ 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 +418,6 @@ 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 +446,10 @@ 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)) {
@@ -502,40 +470,24 @@ 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) {
+            void ShaderCompiler::PrepareSpirvValidation() {
                 PinValidatorTablesForProcessExit();
             }
-                }
-                return state == 1;
-            }
-
-            void ShaderCompiler::SetSpirvValidationEnabled(bool enabled) {
-                g_validateSpirv.store(enabled ? 1 : 0, std::memory_order_release);
-                if (enabled) {
-                    PinValidatorTablesForProcessExit();
-                }
-            }
 
             Uint64 ShaderCompiler::NoteSpirvValidationFailure() {
                 return g_spirvValidationFailures.fetch_add(1, std::memory_order_relaxed) + 1;
@@ -604,16 +556,19 @@ 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 +618,41 @@ 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 +676,8 @@ 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 +685,7 @@ namespace MobileGL {
                     blockNames, &flattenedBlockNames));
 
                 return RunOptimizerChecked("FlattenXfbInterfaceBlocksForEssl", optimizer, inputBinary,
-                                           outputBinary);
+                                           outputBinary, true, enableSpirvValidation);
             }
 
             bool ShaderCompiler::RewriteXfbCaptureNameForFlattenedBlock(
@@ -736,48 +695,53 @@ 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 +774,7 @@ 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 +793,7 @@ namespace MobileGL {
                 lowerer.RegisterPass(CreateAggressiveDCEPass(false));
 
                 if (!RunOptimizerChecked("LegalizeFragmentOutputIndexingForEssl.lower", lowerer, folded,
-                                         outputBinary) ||
+                                         outputBinary, true, enableSpirvValidation) ||
                     outputBinary.empty()) {
                     outputBinary = folded;
                     return true;
@@ -846,16 +810,17 @@ 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 +862,52 @@ 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::DeriveNumSubgroupsForVulkan(const Vector<Uint32>& inputBinary,
+                                                             Vector<uint32_t>& outputBinary,
+                                                             const bool enableSpirvValidation) {
+                using namespace spvtools;
+                Optimizer optimizer(SPV_ENV_VULKAN_1_1);
+                optimizer.RegisterPass(DeriveNumSubgroupsPass::CreateDeriveNumSubgroupsPass());
+
+                return RunOptimizerChecked("DeriveNumSubgroupsForVulkan", 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 +1035,8 @@ 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 +136,38 @@ 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);
+                // Replaces compute gl_NumSubgroups loads with the value derived from the local
+                // workgroup dimensions and gl_SubgroupSize. DirectVulkan only; this avoids a
+                // driver builtin that can disagree with the subgroup IDs the same dispatch emits.
+                // See DeriveNumSubgroupsPass.
+                static bool DeriveNumSubgroupsForVulkan(const Vector<Uint32>& inputBinary,
+                                                        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 +175,15 @@ 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 +210,16 @@ 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/DeriveNumSubgroupsPass.cpp

@@ -0,0 +1,251 @@
+// MobileGL - MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DeriveNumSubgroupsPass.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
+// End of Source File Header
+
+#include "DeriveNumSubgroupsPass.h"
+
+#include "spirv.hpp"
+#include "source/opt/constants.h"
+#include "source/opt/def_use_manager.h"
+#include "source/opt/instruction.h"
+#include "source/opt/ir_context.h"
+#include "source/opt/module.h"
+#include "source/util/make_unique.h"
+
+#include <vector>
+
+namespace MobileGL {
+    namespace MG_Util {
+        namespace ShaderTranspiler {
+            namespace {
+                using spvtools::opt::Instruction;
+                using spvtools::opt::IRContext;
+                using spvtools::opt::Operand;
+
+                Instruction* FindBuiltinDefinition(IRContext* context, spv::BuiltIn builtin) {
+                    auto* defUseMgr = context->get_def_use_mgr();
+                    for (auto& annotation : context->annotations()) {
+                        if (annotation.opcode() != spv::Op::OpDecorate || annotation.NumInOperands() < 3) {
+                            continue;
+                        }
+                        if (static_cast<spv::Decoration>(annotation.GetSingleWordInOperand(1)) !=
+                            spv::Decoration::BuiltIn) {
+                            continue;
+                        }
+                        if (static_cast<spv::BuiltIn>(annotation.GetSingleWordInOperand(2)) != builtin) {
+                            continue;
+                        }
+                        return defUseMgr->GetDef(annotation.GetSingleWordInOperand(0));
+                    }
+                    return nullptr;
+                }
+
+                bool IsInputPointerTo(IRContext* context, const Instruction* variable, uint32_t pointeeTypeId) {
+                    if (variable == nullptr || variable->opcode() != spv::Op::OpVariable ||
+                        variable->NumInOperands() < 1 ||
+                        static_cast<spv::StorageClass>(variable->GetSingleWordInOperand(0)) !=
+                            spv::StorageClass::Input) {
+                        return false;
+                    }
+                    const Instruction* pointerType = context->get_def_use_mgr()->GetDef(variable->type_id());
+                    return pointerType != nullptr && pointerType->opcode() == spv::Op::OpTypePointer &&
+                           pointerType->NumInOperands() >= 2 &&
+                           static_cast<spv::StorageClass>(pointerType->GetSingleWordInOperand(0)) ==
+                               spv::StorageClass::Input &&
+                           pointerType->GetSingleWordInOperand(1) == pointeeTypeId;
+                }
+
+                bool IsUnsignedInt32(IRContext* context, uint32_t typeId) {
+                    const Instruction* type = context->get_def_use_mgr()->GetDef(typeId);
+                    return type != nullptr && type->opcode() == spv::Op::OpTypeInt &&
+                           type->NumInOperands() >= 2 && type->GetSingleWordInOperand(0) == 32u &&
+                           type->GetSingleWordInOperand(1) == 0u;
+                }
+
+                uint32_t SynthesizeSubgroupSizeVariable(IRContext* context, uint32_t pointerTypeId) {
+                    const uint32_t variableId = context->TakeNextId();
+                    context->AddGlobalValue(spvtools::MakeUnique<Instruction>(
+                        context, spv::Op::OpVariable, pointerTypeId, variableId,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_STORAGE_CLASS,
+                             {static_cast<uint32_t>(spv::StorageClass::Input)}}}));
+                    context->AddAnnotationInst(spvtools::MakeUnique<Instruction>(
+                        context, spv::Op::OpDecorate, 0, 0,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_ID, {variableId}},
+                            {SPV_OPERAND_TYPE_DECORATION,
+                             {static_cast<uint32_t>(spv::Decoration::BuiltIn)}},
+                            {SPV_OPERAND_TYPE_LITERAL_INTEGER,
+                             {static_cast<uint32_t>(spv::BuiltIn::SubgroupSize)}}}));
+
+                    for (Instruction& entryPoint : context->module()->entry_points()) {
+                        entryPoint.AddOperand({SPV_OPERAND_TYPE_ID, {variableId}});
+                    }
+                    return variableId;
+                }
+            } // namespace
+
+            spvtools::opt::Pass::Status DeriveNumSubgroupsPass::Process() {
+                auto* irContext = context();
+                auto* defUseMgr = irContext->get_def_use_mgr();
+
+                Instruction* numSubgroupsVar = FindBuiltinDefinition(irContext, spv::BuiltIn::NumSubgroups);
+                if (numSubgroupsVar == nullptr) {
+                    return Status::SuccessWithoutChange;
+                }
+
+                std::vector<Instruction*> numSubgroupsLoads;
+                bool sawUnexpectedUser = false;
+                const uint32_t numSubgroupsVarId = numSubgroupsVar->result_id();
+                defUseMgr->ForEachUser(numSubgroupsVar, [&](Instruction* user) {
+                    switch (user->opcode()) {
+                    case spv::Op::OpLoad:
+                        if (user->NumInOperands() >= 1 &&
+                            user->GetSingleWordInOperand(0) == numSubgroupsVarId) {
+                            numSubgroupsLoads.push_back(user);
+                        } else {
+                            sawUnexpectedUser = true;
+                        }
+                        return;
+                    case spv::Op::OpDecorate:
+                    case spv::Op::OpDecorateId:
+                    case spv::Op::OpDecorateString:
+                    case spv::Op::OpName:
+                    case spv::Op::OpEntryPoint:
+                        return;
+                    default:
+                        sawUnexpectedUser = true;
+                        return;
+                    }
+                });
+                if (sawUnexpectedUser) {
+                    return Status::Failure;
+                }
+                if (numSubgroupsLoads.empty()) {
+                    return Status::SuccessWithoutChange;
+                }
+
+                const uint32_t valueTypeId = numSubgroupsLoads.front()->type_id();
+                if (!IsUnsignedInt32(irContext, valueTypeId) ||
+                    !IsInputPointerTo(irContext, numSubgroupsVar, valueTypeId)) {
+                    return Status::Failure;
+                }
+                for (const Instruction* load : numSubgroupsLoads) {
+                    if (load->type_id() != valueTypeId) {
+                        return Status::Failure;
+                    }
+                }
+
+                Instruction* workgroupSize = FindBuiltinDefinition(irContext, spv::BuiltIn::WorkgroupSize);
+                if (workgroupSize == nullptr ||
+                    (workgroupSize->opcode() != spv::Op::OpConstantComposite &&
+                     workgroupSize->opcode() != spv::Op::OpSpecConstantComposite)) {
+                    return Status::Failure;
+                }
+                const Instruction* workgroupSizeType = defUseMgr->GetDef(workgroupSize->type_id());
+                if (workgroupSizeType == nullptr || workgroupSizeType->opcode() != spv::Op::OpTypeVector ||
+                    workgroupSizeType->NumInOperands() < 2 ||
+                    workgroupSizeType->GetSingleWordInOperand(0) != valueTypeId ||
+                    workgroupSizeType->GetSingleWordInOperand(1) != 3u) {
+                    return Status::Failure;
+                }
+
+                Instruction* subgroupSizeVar = FindBuiltinDefinition(irContext, spv::BuiltIn::SubgroupSize);
+                if (subgroupSizeVar != nullptr &&
+                    !IsInputPointerTo(irContext, subgroupSizeVar, valueTypeId)) {
+                    return Status::Failure;
+                }
+
+                auto* constantMgr = irContext->get_constant_mgr();
+                auto* typeMgr = irContext->get_type_mgr();
+                const auto* valueType = typeMgr->GetType(valueTypeId);
+                if (valueType == nullptr) {
+                    return Status::Failure;
+                }
+                const auto* one = constantMgr->GetConstant(valueType, {1u});
+                const Instruction* oneInst =
+                    one != nullptr ? constantMgr->GetDefiningInstruction(one, valueTypeId) : nullptr;
+                if (oneInst == nullptr) {
+                    return Status::Failure;
+                }
+                const uint32_t oneId = oneInst->result_id();
+
+                const uint32_t subgroupSizeVarId = subgroupSizeVar != nullptr
+                    ? subgroupSizeVar->result_id()
+                    : SynthesizeSubgroupSizeVariable(irContext, numSubgroupsVar->type_id());
+                const uint32_t workgroupSizeId = workgroupSize->result_id();
+
+                // The pipeline never enables ALLOW_VARYING_SUBGROUP_SIZE, so Vulkan's fixed
+                // subgroup partition is exactly ceil(local invocation count / SubgroupSize).
+                // `(count - 1) / size + 1` avoids an addition overflow at count + size - 1.
+                for (Instruction* load : numSubgroupsLoads) {
+                    const uint32_t localSizeXId = irContext->TakeNextId();
+                    const uint32_t localSizeYId = irContext->TakeNextId();
+                    const uint32_t localSizeZId = irContext->TakeNextId();
+                    const uint32_t localSizeXYId = irContext->TakeNextId();
+                    const uint32_t invocationCountId = irContext->TakeNextId();
+                    const uint32_t adjustedCountId = irContext->TakeNextId();
+                    const uint32_t subgroupSizeId = irContext->TakeNextId();
+                    const uint32_t quotientId = irContext->TakeNextId();
+
+                    load->InsertBefore(spvtools::MakeUnique<Instruction>(
+                        irContext, spv::Op::OpCompositeExtract, valueTypeId, localSizeXId,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_ID, {workgroupSizeId}},
+                            {SPV_OPERAND_TYPE_LITERAL_INTEGER, {0u}}}));
+                    load->InsertBefore(spvtools::MakeUnique<Instruction>(
+                        irContext, spv::Op::OpCompositeExtract, valueTypeId, localSizeYId,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_ID, {workgroupSizeId}},
+                            {SPV_OPERAND_TYPE_LITERAL_INTEGER, {1u}}}));
+                    load->InsertBefore(spvtools::MakeUnique<Instruction>(
+                        irContext, spv::Op::OpCompositeExtract, valueTypeId, localSizeZId,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_ID, {workgroupSizeId}},
+                            {SPV_OPERAND_TYPE_LITERAL_INTEGER, {2u}}}));
+                    load->InsertBefore(spvtools::MakeUnique<Instruction>(
+                        irContext, spv::Op::OpIMul, valueTypeId, localSizeXYId,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_ID, {localSizeXId}},
+                            {SPV_OPERAND_TYPE_ID, {localSizeYId}}}));
+                    load->InsertBefore(spvtools::MakeUnique<Instruction>(
+                        irContext, spv::Op::OpIMul, valueTypeId, invocationCountId,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_ID, {localSizeXYId}},
+                            {SPV_OPERAND_TYPE_ID, {localSizeZId}}}));
+                    load->InsertBefore(spvtools::MakeUnique<Instruction>(
+                        irContext, spv::Op::OpISub, valueTypeId, adjustedCountId,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_ID, {invocationCountId}},
+                            {SPV_OPERAND_TYPE_ID, {oneId}}}));
+                    load->InsertBefore(spvtools::MakeUnique<Instruction>(
+                        irContext, spv::Op::OpLoad, valueTypeId, subgroupSizeId,
+                        std::initializer_list<Operand>{{SPV_OPERAND_TYPE_ID, {subgroupSizeVarId}}}));
+                    load->InsertBefore(spvtools::MakeUnique<Instruction>(
+                        irContext, spv::Op::OpUDiv, valueTypeId, quotientId,
+                        std::initializer_list<Operand>{
+                            {SPV_OPERAND_TYPE_ID, {adjustedCountId}},
+                            {SPV_OPERAND_TYPE_ID, {subgroupSizeId}}}));
+
+                    // Preserve the original result id so every downstream use automatically sees
+                    // the derived value instead of the driver's NumSubgroups builtin.
+                    load->SetOpcode(spv::Op::OpIAdd);
+                    load->SetInOperands(Instruction::OperandList{
+                        {SPV_OPERAND_TYPE_ID, {quotientId}},
+                        {SPV_OPERAND_TYPE_ID, {oneId}}});
+                }
+
+                irContext->InvalidateAnalysesExceptFor(IRContext::kAnalysisNone);
+                return Status::SuccessWithChange;
+            }
+
+            spvtools::Optimizer::PassToken DeriveNumSubgroupsPass::CreateDeriveNumSubgroupsPass() {
+                return spvtools::Optimizer::PassToken(MakeUnique<DeriveNumSubgroupsPass>());
+            }
+        } // namespace ShaderTranspiler
+    } // namespace MG_Util
+} // namespace MobileGL

MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DeriveNumSubgroupsPass.h

@@ -0,0 +1,37 @@
+// MobileGL - MobileGL/MG_Util/ShaderTranspiler/SpirvPasses/DeriveNumSubgroupsPass.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
+// End of Source File Header
+
+#pragma once
+
+#include "source/opt/pass.h"
+#include "spirv-tools/optimizer.hpp"
+
+#include <Includes.h>
+
+namespace MobileGL {
+    namespace MG_Util {
+        namespace ShaderTranspiler {
+            // Replaces compute-stage NumSubgroups builtin loads with
+            // ceil(WorkgroupSize.x * WorkgroupSize.y * WorkgroupSize.z / SubgroupSize).
+            //
+            // That is the value Vulkan defines for NumSubgroups when the pipeline does not
+            // enable varying subgroup sizes, which MobileGL never does. Deriving it avoids
+            // drivers that expose the real SubgroupId topology but return an inconsistent
+            // NumSubgroups value. This is a DirectVulkan semantic repair, not a source-shader
+            // rewrite; the application's subgroup arithmetic and shared-memory logic remain
+            // unchanged.
+            class DeriveNumSubgroupsPass : public spvtools::opt::Pass {
+            public:
+                const char* name() const override { return "derive-num-subgroups"; }
+                Status Process() override;
+
+                static spvtools::Optimizer::PassToken CreateDeriveNumSubgroupsPass();
+            };
+        } // namespace ShaderTranspiler
+    } // namespace MG_Util
+} // namespace MobileGL

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";
-
-                Bool IsNamedSamplerFunctionParameter(spvtools::opt::IRContext* context,
-                                                     spvtools::opt::Instruction& nameInst) {
-                    if (nameInst.opcode() != spv::Op::OpName || nameInst.NumInOperands() < 2) {
-                        return false;
+                const char* GetCompatName(StringView name) {
+                    if (name == "sampler") return "MGL_COMPAT_sampler";
+                    if (name == "new") return "MGL_COMPAT_new";
+                    return nullptr;
                 }
 
-                    if (nameInst.GetInOperand(1).AsString() != kConflictingName) {
-                        return false;
+                const char* GetConflictingFunctionParameterCompatName(spvtools::opt::IRContext* context,
+                                                                       spvtools::opt::Instruction& nameInst) {
+                    if (nameInst.opcode() != spv::Op::OpName || nameInst.NumInOperands() < 2) {
+                        return nullptr;
+                    }
+
+                    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/src/trace/cpp/CMakeLists.txt

@@ -223,7 +223,8 @@ target_include_directories(glretrace_common PUBLIC
         "${APITRACE_GENERATED_DIR}"
         "${APITRACE_ROOT}/dispatch"
         "${APITRACE_ROOT}/helpers"
-        "${APITRACE_ROOT}/retrace")
+        "${APITRACE_ROOT}/retrace"
+        "${CMAKE_CURRENT_LIST_DIR}/../../../../../tools/trace_replay")
 target_compile_definitions(glretrace_common PRIVATE
         main=mobilegl_apitrace_main)
 target_redirect_exit(glretrace_common)
@@ -231,14 +232,16 @@ target_link_libraries(glretrace_common PUBLIC retrace_common glhelpers glproc)
 
 add_library(trace_replay_runner SHARED
         trace_replay_core.cpp
-        trace_replay_jni.cpp)
+        trace_replay_jni.cpp
+        "${CMAKE_CURRENT_LIST_DIR}/../../../../../tools/trace_replay/apitrace_fbo_dump.cpp")
 
 target_compile_features(trace_replay_runner PRIVATE cxx_std_17)
 target_compile_definitions(trace_replay_runner PRIVATE
         MOBILEGL_APITRACE_RETRACE_MAIN=mobilegl_apitrace_main)
 
 target_include_directories(trace_replay_runner PRIVATE
-        "${APITRACE_ROOT}/lib/image")
+        "${APITRACE_ROOT}/lib/image"
+        "${CMAKE_CURRENT_LIST_DIR}/../../../../../tools/trace_replay")
 
 target_link_libraries(trace_replay_runner
         glretrace_common

android-plugin/app/src/trace/cpp/apitrace_glws_android.cpp

@@ -1,3 +1,4 @@
+#include "apitrace_fbo_dump.hpp"
 #include "glws.hpp"
 #include "retrace.hpp"
 
@@ -375,6 +376,7 @@ bool makeCurrentInternal(Drawable *drawable, Drawable *readable, Context *contex
     }
     gCurrentDrawable = drawable;
     gCurrentContext = eglContext;
+    mobilegl_trace_dump::InstallIfRequested();
     return true;
 }
 

android-plugin/app/src/trace/cpp/trace_replay_core.cpp

@@ -164,6 +164,11 @@ bool LoadMobileGL(const Request& request, std::string& error) {
     } else {
         unsetenv("MOBILEGL_COHERENT_AS_FLUSH");
     }
+    if (request.numSubgroupsQuirk) {
+        setenv("MOBILEGL_NUM_SUBGROUPS_QUIRK", "1", 1);
+    } else {
+        unsetenv("MOBILEGL_NUM_SUBGROUPS_QUIRK");
+    }
     if (request.fboAttachmentDumps.empty()) {
         unsetenv("MOBILEGL_TRACE_DUMP_FBO_ATTACHMENTS");
     } else {
@@ -176,6 +181,18 @@ bool LoadMobileGL(const Request& request, std::string& error) {
         }
         setenv("MOBILEGL_TRACE_DUMP_FBO_ATTACHMENTS", dumpPoints.c_str(), 1);
     }
+    if (request.texture2dDumps.empty()) {
+        unsetenv("MOBILEGL_TRACE_DUMP_TEXTURE_2D");
+    } else {
+        std::string dumpPoints;
+        for (const std::string& dumpPoint : request.texture2dDumps) {
+            if (!dumpPoints.empty()) {
+                dumpPoints += ';';
+            }
+            dumpPoints += dumpPoint;
+        }
+        setenv("MOBILEGL_TRACE_DUMP_TEXTURE_2D", dumpPoints.c_str(), 1);
+    }
 
     void* handle = dlopen(request.mobileGlLibrary.c_str(), RTLD_NOW | RTLD_GLOBAL);
     if (handle == nullptr) {
@@ -383,6 +400,13 @@ std::string SnapshotCallSet(const Request& request) {
             callSet += "," + call;
         }
     }
+    for (const std::string& dumpPoint : request.texture2dDumps) {
+        const std::size_t separator = dumpPoint.find(',');
+        const std::string call = dumpPoint.substr(0, separator);
+        if (!call.empty() && call != std::to_string(request.targetCall)) {
+            callSet += "," + call;
+        }
+    }
     return callSet;
 }
 
@@ -799,6 +823,7 @@ bool WriteResultJson(const Request& request, const Result& result) {
          << (request.avoidAngleLlvmpipeSamplerMipmapMinFilter ? "true" : "false") << ",\n";
     file << "  \"avoidAngleLlvmpipeExplicitLodBias\": "
          << (request.avoidAngleLlvmpipeExplicitLodBias ? "true" : "false") << ",\n";
+    file << "  \"numSubgroupsQuirk\": " << (request.numSubgroupsQuirk ? "true" : "false") << ",\n";
     file << "  \"holdMs\": " << request.holdMs << ",\n";
     file << "  \"mismatchPixels\": " << result.mismatchPixels << "\n";
     file << "}\n";

android-plugin/app/src/trace/cpp/trace_replay_core.hpp

@@ -27,6 +27,9 @@ struct Request {
     // Framebuffer-attachment dump points, each `CALL:DIR[:FBO,FBO,...]`. Debug-only; the
     // replay behaves exactly as before when this is empty.
     std::vector<std::string> fboAttachmentDumps;
+    // Named GL_TEXTURE_2D dump points, each `CALL,TEXTURE,LEVEL,DIR`. Debug-only; the replay
+    // behaves exactly as before when this is empty.
+    std::vector<std::string> texture2dDumps;
     int targetFrame = -1;
     long long targetCall = -1;
     int width = 0;
@@ -41,6 +44,7 @@ struct Request {
     bool avoidAngleLlvmpipeSamplerMipmapMinFilter = false;
     bool avoidAngleLlvmpipeExplicitLodBias = false;
     bool coherentAsFlush = false;
+    bool numSubgroupsQuirk = false;
     int holdMs = 0;
 };
 

android-plugin/app/src/trace/cpp/trace_replay_jni.cpp

@@ -6,6 +6,7 @@
 
 #include <exception>
 #include <string>
+#include <vector>
 #include <sys/stat.h>
 
 extern "C" void mobilegl_trace_set_native_window(ANativeWindow *window);
@@ -25,6 +26,23 @@ std::string ToString(JNIEnv* env, jstring value) {
     return out;
 }
 
+std::vector<std::string> SplitSemicolonList(const std::string& value) {
+    std::vector<std::string> values;
+    std::size_t begin = 0;
+    while (begin < value.size()) {
+        const std::size_t end = value.find(';', begin);
+        const std::string entry = value.substr(begin, end - begin);
+        if (!entry.empty()) {
+            values.push_back(entry);
+        }
+        if (end == std::string::npos) {
+            break;
+        }
+        begin = end + 1;
+    }
+    return values;
+}
+
 jobject MakeResult(JNIEnv* env, const mobilegl_trace::Result& result) {
     jclass clazz = env->FindClass("top/mobilegl/plugin/trace/TraceReplayActivity$TraceReplayResult");
     if (clazz == nullptr) {
@@ -103,7 +121,9 @@ Java_top_mobilegl_plugin_trace_TraceReplayActivity_nativeRunTraceReplay(JNIEnv*
                                                                         jboolean usePbuffer,
                                                                         jboolean avoidAngleLlvmpipeSamplerMipmapMinFilter,
                                                                         jboolean avoidAngleLlvmpipeExplicitLodBias,
-                                                                        jboolean coherentAsFlush) {
+                                                                        jboolean coherentAsFlush,
+                                                                        jboolean numSubgroupsQuirk,
+                                                                        jstring texture2dDumps) {
     mobilegl_trace::Request request;
     request.tracePath = ToString(env, tracePath);
     request.goldenPath = ToString(env, goldenPath);
@@ -115,6 +135,7 @@ Java_top_mobilegl_plugin_trace_TraceReplayActivity_nativeRunTraceReplay(JNIEnv*
     request.diffPath = ToString(env, diffPath);
     request.backend = ToString(env, backend);
     request.angleVariant = ToString(env, angleVariant);
+    request.texture2dDumps = SplitSemicolonList(ToString(env, texture2dDumps));
     request.targetFrame = targetFrame;
     request.targetCall = targetCall;
     request.width = width;
@@ -130,6 +151,7 @@ Java_top_mobilegl_plugin_trace_TraceReplayActivity_nativeRunTraceReplay(JNIEnv*
         avoidAngleLlvmpipeSamplerMipmapMinFilter == JNI_TRUE;
     request.avoidAngleLlvmpipeExplicitLodBias = avoidAngleLlvmpipeExplicitLodBias == JNI_TRUE;
     request.coherentAsFlush = coherentAsFlush == JNI_TRUE;
+    request.numSubgroupsQuirk = numSubgroupsQuirk == JNI_TRUE;
 
     ScopedTraceReplayState replayState;
     mobilegl_trace_set_requested_size(request.width, request.height);

android-plugin/app/src/trace/java/top/mobilegl/plugin/trace/TraceReplayActivity.java

@@ -115,7 +115,9 @@ public final class TraceReplayActivity extends Activity {
                 request.usePbuffer,
                 request.avoidAngleLlvmpipeSamplerMipmapMinFilter,
                 request.avoidAngleLlvmpipeExplicitLodBias,
-                request.coherentAsFlush
+                request.coherentAsFlush,
+                request.numSubgroupsQuirk,
+                request.texture2dDumps
         );
         Log.i(TAG, result.toString());
         TraceReplayResult finalResult = result;
@@ -147,7 +149,9 @@ public final class TraceReplayActivity extends Activity {
             boolean usePbuffer,
             boolean avoidAngleLlvmpipeSamplerMipmapMinFilter,
             boolean avoidAngleLlvmpipeExplicitLodBias,
-            boolean coherentAsFlush
+            boolean coherentAsFlush,
+            boolean numSubgroupsQuirk,
+            String texture2dDumps
     );
 
     private static final class TraceReplayRequest {
@@ -172,6 +176,8 @@ public final class TraceReplayActivity extends Activity {
         final boolean avoidAngleLlvmpipeSamplerMipmapMinFilter;
         final boolean avoidAngleLlvmpipeExplicitLodBias;
         final boolean coherentAsFlush;
+        final boolean numSubgroupsQuirk;
+        final String texture2dDumps;
 
         private TraceReplayRequest(
                 String tracePath,
@@ -194,7 +200,9 @@ public final class TraceReplayActivity extends Activity {
                 boolean usePbuffer,
                 boolean avoidAngleLlvmpipeSamplerMipmapMinFilter,
                 boolean avoidAngleLlvmpipeExplicitLodBias,
-                boolean coherentAsFlush
+                boolean coherentAsFlush,
+                boolean numSubgroupsQuirk,
+                String texture2dDumps
         ) {
             this.tracePath = tracePath;
             this.goldenPath = goldenPath;
@@ -217,6 +225,8 @@ public final class TraceReplayActivity extends Activity {
             this.avoidAngleLlvmpipeSamplerMipmapMinFilter = avoidAngleLlvmpipeSamplerMipmapMinFilter;
             this.avoidAngleLlvmpipeExplicitLodBias = avoidAngleLlvmpipeExplicitLodBias;
             this.coherentAsFlush = coherentAsFlush;
+            this.numSubgroupsQuirk = numSubgroupsQuirk;
+            this.texture2dDumps = texture2dDumps;
         }
 
         static TraceReplayRequest from(Intent intent, File filesDir, String defaultBackend) {
@@ -243,7 +253,9 @@ public final class TraceReplayActivity extends Activity {
                     intent.getBooleanExtra("use_pbuffer", false),
                     intent.getBooleanExtra("avoid_angle_llvmpipe_sampler_mipmap_min_filter", false),
                     intent.getBooleanExtra("avoid_angle_llvmpipe_explicit_lod_bias", false),
-                    intent.getBooleanExtra("coherent_as_flush", false)
+                    intent.getBooleanExtra("coherent_as_flush", false),
+                    intent.getBooleanExtra("num_subgroups_quirk", false),
+                    readString(intent, "texture_2d_dumps", "")
             );
         }
 

android-plugin/trace-replay-ci.sh

@@ -31,6 +31,8 @@ Usage:
     [--avoid-angle-llvmpipe-sampler-mipmap-min-filter] \
     [--avoid-angle-llvmpipe-explicit-lod-bias] \
     [--coherent-as-flush] \
+    [--num-subgroups-quirk] \
+    [--dump-texture-2d CALL,TEXTURE,LEVEL,DIR] \
     --timeout-seconds N
 
 Set MOBILEGL_USE_ANGLE=1 to run DirectGLES replay with packaged ANGLE
@@ -46,6 +48,8 @@ sample with an explicit LOD that ANGLE llvmpipe cannot take a LOD bias on
 (MOBILEGL_AVOID_EXPLICIT_LOD_BIAS=1).
 Pass --coherent-as-flush for traces whose engine writes persistent
 GL_MAP_FLUSH_EXPLICIT_BIT maps it never flushes (MOBILEGL_COHERENT_AS_FLUSH=1).
+Pass --num-subgroups-quirk to derive compute gl_NumSubgroups instead of reading
+the Vulkan builtin (MOBILEGL_NUM_SUBGROUPS_QUIRK=1).
 EOF
 }
 
@@ -104,6 +108,8 @@ use_pbuffer=0
 avoid_angle_llvmpipe_sampler_mipmap_min_filter=0
 avoid_angle_llvmpipe_explicit_lod_bias=0
 coherent_as_flush=0
+num_subgroups_quirk=0
+texture_2d_dumps=""
 timeout_seconds=""
 
 while [ "$#" -gt 0 ]; do
@@ -144,6 +150,8 @@ while [ "$#" -gt 0 ]; do
       shift 1
       ;;
     --coherent-as-flush) coherent_as_flush=1; shift 1 ;;
+    --num-subgroups-quirk) num_subgroups_quirk=1; shift 1 ;;
+    --dump-texture-2d) texture_2d_dumps="$(next_arg "$@")"; shift 2 ;;
     --timeout-seconds) timeout_seconds="$(next_arg "$@")"; shift 2 ;;
     -h|--help) usage; exit 0 ;;
     *) die "unknown argument: $1" ;;
@@ -262,6 +270,27 @@ copy_app_artifact() {
   fi
 }
 
+copy_texture_2d_dumps() {
+  [ -n "${texture_2d_dumps}" ] || return 0
+  saved_ifs="${IFS}"
+  IFS=';'
+  set -- ${texture_2d_dumps}
+  IFS="${saved_ifs}"
+  for dump_point in "$@"; do
+    dump_dir="${dump_point#*,}"
+    dump_dir="${dump_dir#*,}"
+    dump_dir="${dump_dir#*,}"
+    [ -n "${dump_dir}" ] || continue
+    dump_name="$(basename "${dump_dir}")"
+    destination_dir="${result_dir}/${dump_name}"
+    mkdir -p "${destination_dir}"
+    if ! adb_device_path exec-out run-as "${package_name}" tar -C "${dump_dir}" -cf - . | tar -xf - -C "${destination_dir}"; then
+      echo "trace-replay-ci.sh: warning: failed to copy texture dump ${dump_dir}" >&2
+      rm -rf "${destination_dir}"
+    fi
+  done
+}
+
 prepare_fixture() {
   fixture_dir="${fixture_root}/${safe_case}"
   rm -rf "${fixture_dir}"
@@ -339,6 +368,12 @@ run_retrace() {
   if [ "${coherent_as_flush}" -eq 1 ]; then
     set -- "$@" --ez coherent_as_flush true
   fi
+  if [ "${num_subgroups_quirk}" -eq 1 ]; then
+    set -- "$@" --ez num_subgroups_quirk true
+  fi
+  if [ -n "${texture_2d_dumps}" ]; then
+    set -- "$@" --es texture_2d_dumps "${texture_2d_dumps}"
+  fi
   set -- "$@" \
     --es output_dir "${app_dir}/output" \
     --es diff_path "${app_dir}/output/${safe_case}-diff.png" \
@@ -399,6 +434,7 @@ run_retrace() {
   copy_app_artifact "${app_dir}/output/${safe_case}-diff.png" "${result_dir}/${safe_case}-${backend}-diff.png"
   copy_app_artifact "${app_dir}/output/retrace.log" "${result_dir}/retrace.log"
   copy_app_artifact "${app_dir}/output/mobilegl.log" "${result_dir}/mobilegl.log"
+  copy_texture_2d_dumps
 
   # A replay that wrote result.json but did not pass used to print nothing but
   # the JSON, which for a non-zero statusCode says only "retrace failed with

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/apitrace_fbo_dump.cpp

@@ -13,8 +13,13 @@
 #include <cstring>
 #include <fstream>
 #include <iostream>
+#include <limits>
 #include <string>
+#if defined(_WIN32)
+#include <direct.h>
+#else
 #include <sys/stat.h>
+#endif
 #include <vector>
 
 // Dumps every colour attachment (and the depth attachment) of every live framebuffer
@@ -34,6 +39,7 @@ using PfnGetIntegerv = void (*)(GLenum, GLint *);
 using PfnGetError = GLenum (*)(void);
 
 constexpr const char *kDumpPointsEnv = "MOBILEGL_TRACE_DUMP_FBO_ATTACHMENTS";
+constexpr const char *kTexture2dDumpPointsEnv = "MOBILEGL_TRACE_DUMP_TEXTURE_2D";
 constexpr const char *kScanLimitEnv = "MOBILEGL_TRACE_DUMP_FBO_SCAN_LIMIT";
 constexpr unsigned kDefaultScanLimit = 1024;
 
@@ -45,6 +51,16 @@ struct DumpPoint {
     bool done = false;
 };
 
+// CALL,TEXTURE,LEVEL,DIR entries, separated by ';'. This deliberately does not share the
+// framebuffer dump grammar: an absolute Windows directory contains ':' but not ','.
+struct Texture2dDumpPoint {
+    unsigned call = 0;
+    unsigned texture = 0;
+    unsigned level = 0;
+    std::string directory;
+    bool done = false;
+};
+
 struct AttachmentDesc {
     GLint objectType = GL_NONE;
     GLint objectName = 0;
@@ -56,6 +72,7 @@ struct AttachmentDesc {
 };
 
 std::vector<DumpPoint> gDumpPoints;
+std::vector<Texture2dDumpPoint> gTexture2dDumpPoints;
 bool gInstalled = false;
 bool gConfigured = false;
 retrace::Dumper *gInnerDumper = nullptr;
@@ -105,13 +122,27 @@ bool MakeDirectories(const std::string &path) {
     for (std::size_t i = 0; i < path.size(); ++i) {
         partial.push_back(path[i]);
         const bool last = i + 1 == path.size();
-        if (path[i] != '/' && !last) {
+#if defined(_WIN32)
+        const bool separator = path[i] == '/' || path[i] == '\\';
+#else
+        const bool separator = path[i] == '/';
+#endif
+        if (!separator && !last) {
             continue;
         }
         if (partial == "/") {
             continue;
         }
+#if defined(_WIN32)
+        if (separator && partial.size() == 3 && partial[1] == ':') {
+            continue;
+        }
+#endif
+#if defined(_WIN32)
+        if (_mkdir(partial.c_str()) != 0 && errno != EEXIST) {
+#else
         if (mkdir(partial.c_str(), 0755) != 0 && errno != EEXIST) {
+#endif
             return false;
         }
     }
@@ -161,6 +192,52 @@ void ParseDumpPoints(const char *spec) {
     }
 }
 
+bool IsDecimal(const std::string &value) {
+    return !value.empty() && value.find_first_not_of("0123456789") == std::string::npos;
+}
+
+void ParseTexture2dDumpPoints(const char *spec) {
+    for (const std::string &entry : Split(spec, ';')) {
+        if (entry.empty()) {
+            continue;
+        }
+        const std::vector<std::string> fields = Split(entry, ',');
+        if (fields.size() != 4 || !IsDecimal(fields[0]) || !IsDecimal(fields[1]) ||
+            !IsDecimal(fields[2]) || fields[3].empty()) {
+            std::cerr << "warning: ignoring malformed " << kTexture2dDumpPointsEnv
+                      << " entry: " << entry << "\n";
+            continue;
+        }
+
+        char *end = nullptr;
+        const unsigned long call = std::strtoul(fields[0].c_str(), &end, 10);
+        if (*end != '\0' || call > std::numeric_limits<unsigned>::max()) {
+            std::cerr << "warning: ignoring malformed " << kTexture2dDumpPointsEnv
+                      << " call: " << entry << "\n";
+            continue;
+        }
+        const unsigned long texture = std::strtoul(fields[1].c_str(), &end, 10);
+        if (*end != '\0' || texture == 0 || texture > std::numeric_limits<unsigned>::max()) {
+            std::cerr << "warning: ignoring malformed " << kTexture2dDumpPointsEnv
+                      << " texture: " << entry << "\n";
+            continue;
+        }
+        const unsigned long level = std::strtoul(fields[2].c_str(), &end, 10);
+        if (*end != '\0' || level > static_cast<unsigned long>(std::numeric_limits<GLint>::max())) {
+            std::cerr << "warning: ignoring malformed " << kTexture2dDumpPointsEnv
+                      << " level: " << entry << "\n";
+            continue;
+        }
+
+        Texture2dDumpPoint point;
+        point.call = static_cast<unsigned>(call);
+        point.texture = static_cast<unsigned>(texture);
+        point.level = static_cast<unsigned>(level);
+        point.directory = fields[3];
+        gTexture2dDumpPoints.push_back(point);
+    }
+}
+
 unsigned ScanLimit() {
     const char *value = std::getenv(kScanLimitEnv);
     if (value == nullptr || value[0] == '\0') {
@@ -235,6 +312,45 @@ bool DescribeAttachment(GLenum attachment, AttachmentDesc &desc) {
     return desc.width > 0 && desc.height > 0;
 }
 
+bool DescribeTexture2D(GLuint texture, GLint level, AttachmentDesc &desc) {
+    const GLint savedTexture = GetInteger(GL_TEXTURE_BINDING_2D);
+    glBindTexture(GL_TEXTURE_2D, texture);
+    desc.objectType = GL_TEXTURE;
+    desc.objectName = static_cast<GLint>(texture);
+    desc.level = level;
+    glGetTexLevelParameteriv(GL_TEXTURE_2D, level, GL_TEXTURE_WIDTH, &desc.width);
+    glGetTexLevelParameteriv(GL_TEXTURE_2D, level, GL_TEXTURE_HEIGHT, &desc.height);
+    glGetTexLevelParameteriv(GL_TEXTURE_2D, level, GL_TEXTURE_INTERNAL_FORMAT, &desc.internalFormat);
+    glGetTexLevelParameteriv(GL_TEXTURE_2D, level, GL_TEXTURE_RED_TYPE, &desc.componentType);
+    glBindTexture(GL_TEXTURE_2D, static_cast<GLuint>(savedTexture));
+    return DrainErrors() == 0 && desc.width > 0 && desc.height > 0;
+}
+
+bool ReadTexture2DFloats(const AttachmentDesc &desc, std::vector<float> &pixels) {
+    const std::size_t count = static_cast<std::size_t>(desc.width) * desc.height * 4;
+    pixels.assign(count, 0.0f);
+    const GLint savedTexture = GetInteger(GL_TEXTURE_BINDING_2D);
+    glBindTexture(GL_TEXTURE_2D, static_cast<GLuint>(desc.objectName));
+    if (desc.componentType == GL_INT || desc.componentType == GL_UNSIGNED_INT) {
+        std::vector<std::int32_t> raw(count, 0);
+        const GLenum type = desc.componentType == GL_INT ? GL_INT : GL_UNSIGNED_INT;
+        glGetTexImage(GL_TEXTURE_2D, desc.level, GL_RGBA_INTEGER, type, raw.data());
+        glBindTexture(GL_TEXTURE_2D, static_cast<GLuint>(savedTexture));
+        if (DrainErrors() != 0) {
+            return false;
+        }
+        for (std::size_t i = 0; i < count; ++i) {
+            pixels[i] = desc.componentType == GL_INT
+                                ? static_cast<float>(raw[i])
+                                : static_cast<float>(static_cast<std::uint32_t>(raw[i]));
+        }
+        return true;
+    }
+    glGetTexImage(GL_TEXTURE_2D, desc.level, GL_RGBA, GL_FLOAT, pixels.data());
+    glBindTexture(GL_TEXTURE_2D, static_cast<GLuint>(savedTexture));
+    return DrainErrors() == 0;
+}
+
 // Reads the attachment as floats regardless of its storage: normalised and float targets
 // convert on the way out, integer targets are read as integers and widened. The float view
 // keeps out-of-[0,1] accumulation buffers legible in the statistics even though the PNG
@@ -308,6 +424,17 @@ std::string FormatStatistics(const std::vector<float> &pixels, unsigned channels
                       static_cast<double>(minimum[c]), static_cast<double>(maximum[c]), mean);
         text += buffer;
     }
+    if (pixelCount > 0) {
+        text += " first=(";
+        for (unsigned c = 0; c < channels; ++c) {
+            if (c > 0) {
+                text += ",";
+            }
+            std::snprintf(buffer, sizeof(buffer), "%.9g", static_cast<double>(pixels[c]));
+            text += buffer;
+        }
+        text += ")";
+    }
     std::snprintf(buffer, sizeof(buffer), " nonfinite=%zu hash=%016llx", nonFinite,
                   static_cast<unsigned long long>(hash));
     text += buffer;
@@ -325,8 +452,8 @@ bool WriteFloatPng(const std::string &path, const AttachmentDesc &desc, unsigned
     return snapshot.writePNG(path.c_str());
 }
 
-void DumpOneAttachment(std::ofstream &manifest, const std::string &directory, unsigned framebuffer,
-                       GLenum attachment, const char *label, bool depth) {
+void DumpOneAttachment(std::ofstream &manifest, const std::string &directory,
+                       const std::string &identity, GLenum attachment, const char *label, bool depth) {
     AttachmentDesc desc;
     if (!DescribeAttachment(attachment, desc)) {
         return;
@@ -343,14 +470,13 @@ void DumpOneAttachment(std::ofstream &manifest, const std::string &directory, un
     std::vector<float> pixels;
     const bool read = ReadAttachmentFloats(desc, depth, channels, pixels);
 
-    const std::string path =
-            directory + "/fbo" + std::to_string(framebuffer) + "-" + label + ".png";
+    const std::string path = directory + "/" + identity + "-" + label + ".png";
     const bool wrote = read && WriteFloatPng(path, desc, channels, pixels);
 
     char header[512];
     std::snprintf(header, sizeof(header),
-                  "fbo %u %s object=%s name=%d level=%d size=%dx%d internalformat=0x%04x component=%s",
-                  framebuffer, label,
+                  "%s %s object=%s name=%d level=%d size=%dx%d internalformat=0x%04x component=%s",
+                  identity.c_str(), label,
                   desc.objectType == GL_RENDERBUFFER ? "renderbuffer" : "texture", desc.objectName,
                   desc.level, desc.width, desc.height, static_cast<unsigned>(desc.internalFormat),
                   ComponentTypeName(desc.componentType));
@@ -381,13 +507,14 @@ void DumpFramebuffer(std::ofstream &manifest, const std::string &directory, unsi
     }
 
     const GLint savedReadBuffer = GetInteger(GL_READ_BUFFER);
+    const std::string identity = "fbo" + std::to_string(framebuffer);
     for (GLint index = 0; index < maxColorAttachments; ++index) {
         char label[32];
         std::snprintf(label, sizeof(label), "att%d", index);
-        DumpOneAttachment(manifest, directory, framebuffer,
+        DumpOneAttachment(manifest, directory, identity,
                           static_cast<GLenum>(GL_COLOR_ATTACHMENT0 + index), label, false);
     }
-    DumpOneAttachment(manifest, directory, framebuffer, GL_DEPTH_ATTACHMENT, "depth", true);
+    DumpOneAttachment(manifest, directory, identity, GL_DEPTH_ATTACHMENT, "depth", true);
 
     // The read buffer is per-framebuffer state the trace goes on using; put it back.
     if (framebuffer != 0 && savedReadBuffer != 0) {
@@ -416,6 +543,7 @@ void RunDumpPoint(DumpPoint &point) {
     const GLint savedPackSkipRows = GetInteger(GL_PACK_SKIP_ROWS);
     const GLint savedPackImageHeight = GetInteger(GL_PACK_IMAGE_HEIGHT);
     const GLint savedPackSkipImages = GetInteger(GL_PACK_SKIP_IMAGES);
+    const GLint savedPackSwapBytes = GetInteger(GL_PACK_SWAP_BYTES);
     DrainErrors();
 
     if (savedPackBuffer != 0) {
@@ -427,6 +555,7 @@ void RunDumpPoint(DumpPoint &point) {
     glPixelStorei(GL_PACK_SKIP_ROWS, 0);
     glPixelStorei(GL_PACK_IMAGE_HEIGHT, 0);
     glPixelStorei(GL_PACK_SKIP_IMAGES, 0);
+    glPixelStorei(GL_PACK_SWAP_BYTES, GL_FALSE);
     DrainErrors();
 
     const GLint maxColorAttachments = GetInteger(GL_MAX_COLOR_ATTACHMENTS);
@@ -462,6 +591,7 @@ void RunDumpPoint(DumpPoint &point) {
     glPixelStorei(GL_PACK_SKIP_ROWS, savedPackSkipRows);
     glPixelStorei(GL_PACK_IMAGE_HEIGHT, savedPackImageHeight);
     glPixelStorei(GL_PACK_SKIP_IMAGES, savedPackSkipImages);
+    glPixelStorei(GL_PACK_SWAP_BYTES, savedPackSwapBytes);
     DrainErrors();
 
     std::cerr << "MOBILEGL_TRACE_FBO_DUMP: call " << retrace::callNo << " -> " << manifestPath
@@ -469,12 +599,98 @@ void RunDumpPoint(DumpPoint &point) {
     point.done = true;
 }
 
+void RunTexture2dDumpPoint(Texture2dDumpPoint &point) {
+    if (!MakeDirectories(point.directory)) {
+        std::cerr << "warning: failed to create texture dump directory " << point.directory << "\n";
+        point.done = true;
+        return;
+    }
+
+    // glGetError is destructive. This debug-only snapshot hook deliberately starts from a
+    // clean error state so diagnostics below identify the dump rather than an earlier trace call.
+    DrainErrors();
+    const GLint savedReadFramebuffer = GetInteger(GL_READ_FRAMEBUFFER_BINDING);
+    const GLint savedPackBuffer = GetInteger(GL_PIXEL_PACK_BUFFER_BINDING);
+    const GLint savedPackAlignment = GetInteger(GL_PACK_ALIGNMENT);
+    const GLint savedPackRowLength = GetInteger(GL_PACK_ROW_LENGTH);
+    const GLint savedPackSkipPixels = GetInteger(GL_PACK_SKIP_PIXELS);
+    const GLint savedPackSkipRows = GetInteger(GL_PACK_SKIP_ROWS);
+    const GLint savedPackImageHeight = GetInteger(GL_PACK_IMAGE_HEIGHT);
+    const GLint savedPackSkipImages = GetInteger(GL_PACK_SKIP_IMAGES);
+    const GLint savedPackSwapBytes = GetInteger(GL_PACK_SWAP_BYTES);
+    DrainErrors();
+
+    if (savedPackBuffer != 0) {
+        glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+    }
+    glPixelStorei(GL_PACK_ALIGNMENT, 1);
+    glPixelStorei(GL_PACK_ROW_LENGTH, 0);
+    glPixelStorei(GL_PACK_SKIP_PIXELS, 0);
+    glPixelStorei(GL_PACK_SKIP_ROWS, 0);
+    glPixelStorei(GL_PACK_IMAGE_HEIGHT, 0);
+    glPixelStorei(GL_PACK_SKIP_IMAGES, 0);
+    glPixelStorei(GL_PACK_SWAP_BYTES, GL_FALSE);
+    DrainErrors();
+
+    const std::string identity = "texture" + std::to_string(point.texture) +
+                                 "-level" + std::to_string(point.level);
+    const std::string manifestPath = point.directory + "/manifest.txt";
+    std::ofstream manifest(manifestPath, std::ios::trunc);
+    manifest << "call " << retrace::callNo << " texture " << point.texture << " level "
+             << point.level << "\n";
+    AttachmentDesc desc;
+    if (glIsTexture(point.texture) == GL_FALSE ||
+        !DescribeTexture2D(point.texture, static_cast<GLint>(point.level), desc)) {
+        manifest << identity << " skipped=not-live-2d-texture\n";
+    } else {
+        std::vector<float> pixels;
+        const bool read = ReadTexture2DFloats(desc, pixels);
+        const bool wrote = read && WriteFloatPng(point.directory + "/" + identity + ".png", desc, 4, pixels);
+        manifest << identity << " object=texture name=" << desc.objectName << " level=" << desc.level
+                 << " size=" << desc.width << "x" << desc.height << " internalformat=0x" << std::hex
+                 << static_cast<unsigned>(desc.internalFormat) << std::dec
+                 << " component=" << ComponentTypeName(desc.componentType);
+        if (read) {
+            manifest << FormatStatistics(pixels, 4);
+        } else {
+            manifest << " read=failed";
+        }
+        if (!wrote) {
+            manifest << " png=failed";
+        }
+        manifest << "\n";
+    }
+    manifest.flush();
+
+    glBindFramebuffer(GL_READ_FRAMEBUFFER, static_cast<GLuint>(savedReadFramebuffer));
+    if (savedPackBuffer != 0) {
+        glBindBuffer(GL_PIXEL_PACK_BUFFER, static_cast<GLuint>(savedPackBuffer));
+    }
+    glPixelStorei(GL_PACK_ALIGNMENT, savedPackAlignment);
+    glPixelStorei(GL_PACK_ROW_LENGTH, savedPackRowLength);
+    glPixelStorei(GL_PACK_SKIP_PIXELS, savedPackSkipPixels);
+    glPixelStorei(GL_PACK_SKIP_ROWS, savedPackSkipRows);
+    glPixelStorei(GL_PACK_IMAGE_HEIGHT, savedPackImageHeight);
+    glPixelStorei(GL_PACK_SKIP_IMAGES, savedPackSkipImages);
+    glPixelStorei(GL_PACK_SWAP_BYTES, savedPackSwapBytes);
+    DrainErrors();
+
+    std::cerr << "MOBILEGL_TRACE_TEXTURE_2D_DUMP: call " << retrace::callNo << " texture "
+              << point.texture << " level " << point.level << " -> " << manifestPath << "\n";
+    point.done = true;
+}
+
 void RunPendingDumps() {
     for (DumpPoint &point : gDumpPoints) {
         if (!point.done && point.call == retrace::callNo) {
             RunDumpPoint(point);
         }
     }
+    for (Texture2dDumpPoint &point : gTexture2dDumpPoints) {
+        if (!point.done && point.call == retrace::callNo) {
+            RunTexture2dDumpPoint(point);
+        }
+    }
 }
 
 class DumpingDumper final : public retrace::Dumper {
@@ -511,8 +727,12 @@ void InstallIfRequested() {
         if (spec != nullptr && spec[0] != '\0') {
             ParseDumpPoints(spec);
         }
+        const char *textureSpec = std::getenv(kTexture2dDumpPointsEnv);
+        if (textureSpec != nullptr && textureSpec[0] != '\0') {
+            ParseTexture2dDumpPoints(textureSpec);
         }
-    if (gDumpPoints.empty()) {
+    }
+    if (gDumpPoints.empty() && gTexture2dDumpPoints.empty()) {
         gInstalled = true;
         return;
     }
@@ -528,6 +748,11 @@ void InstallIfRequested() {
         std::cerr << "MOBILEGL_TRACE_FBO_DUMP: armed for call " << point.call << " -> "
                   << point.directory << "\n";
     }
+    for (const Texture2dDumpPoint &point : gTexture2dDumpPoints) {
+        std::cerr << "MOBILEGL_TRACE_TEXTURE_2D_DUMP: armed for call " << point.call
+                  << " texture " << point.texture << " level " << point.level << " -> "
+                  << point.directory << "\n";
+    }
 }
 
 } // namespace mobilegl_trace_dump

tools/trace_replay/apitrace_fbo_dump.hpp

@@ -2,9 +2,10 @@
 
 namespace mobilegl_trace_dump {
 
-// Installs the framebuffer-attachment dump hook when MOBILEGL_TRACE_DUMP_FBO_ATTACHMENTS
-// describes at least one dump point. Safe and cheap to call on every makeCurrent: the
-// environment is consulted once and the hook is installed at most once.
+// Installs the opt-in framebuffer-attachment and named-2D-texture dump hooks. The respective
+// environments are MOBILEGL_TRACE_DUMP_FBO_ATTACHMENTS and MOBILEGL_TRACE_DUMP_TEXTURE_2D.
+// Safe and cheap to call on every makeCurrent: the environment is consulted once and the hook is
+// installed at most once.
 void InstallIfRequested();
 
 } // namespace mobilegl_trace_dump

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 +277,15 @@
     },
     {
       "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,
+      "num_subgroups_quirk": true
     },
     {
       "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)