[Fix] (MG_State, MG_Impl): GL-order capture for geometry triangle strips

Vulkan transform feedback captures odd strip triangles as (i, i+2, i+1) while GL table 10.1 decomposes them as (i+1, i, i+2). When the capture stage is a triangle-strip geometry shader whose EmitVertex/EndPrimitive sequence is statically knowable (no emission under control flow), link time extracts the per-invocation strip lengths from the glslang AST, and EndTransformFeedback rotates each odd triangle's captured vertex records into GL order in place (bounded by the binding ranges' whole-triangle capacity; raw input primitives tracked per Begin/End). KHR-GL33.transform_feedback.geometry passes - the family is 21/21.
[Fix] (MG_Impl): count multisample texture attachments in GL_SAMPLE_BUFFERS
2026-09-11 21:58:31 +09:00 · 2026-08-01 03:17:30 -04:00 · 2026-08-01 02:50:29 -04:00 · 2026-08-01 02:46:16 -04:00 · 2026-08-01 02:19:07 -04:00 · 2026-08-01 02:02:35 -04:00
53 changed files with 3724 additions and 208 deletions
@@ -220,6 +220,15 @@ namespace MobileGL {
            // and leave the query readable later.
            Bool (*GetQueryResult64)(BackendQueryHandle query, Bool wait, Uint64* outNanoseconds);
            void (*DeleteBackendQuery)(BackendQueryHandle query);
+            // GL_SAMPLES_PASSED occlusion queries (optional; null = unsupported,
+            // the frontend then rejects the target). Results/deletion flow through
+            // GetQueryResult64 / DeleteBackendQuery like timer queries.
+            BackendQueryHandle (*BeginOcclusionQuery)();
+            void (*EndOcclusionQuery)(BackendQueryHandle query);
+            // Transform feedback primitive queries backed by real GPU query pools
+            // (optional; null = frontend falls back to CPU accounting).
+            BackendQueryHandle (*BeginXfbPrimitivesQuery)(Bool generated);
+            void (*EndXfbPrimitivesQuery)(BackendQueryHandle query);
            Int64 (*GetGpuTimestampNs)(); // glGetInteger64v(GL_TIMESTAMP); 0 if unsupported
        };
        struct GlobalBackendFunctionsTable {
@@ -635,6 +635,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                funcsTable.GL.DeleteBackendQuery = DeleteBackendQuery;
                funcsTable.GL.GetGpuTimestampNs = GetGpuTimestampNs;
            }
+            // Occlusion queries share the handle-based result/delete entries, which must
+            // exist even when timer queries are disabled.
+            funcsTable.GL.BeginOcclusionQuery = BeginOcclusionQuery;
+            funcsTable.GL.EndOcclusionQuery = EndOcclusionQuery;
+            funcsTable.GL.BeginXfbPrimitivesQuery = BeginXfbPrimitivesQuery;
+            funcsTable.GL.EndXfbPrimitivesQuery = EndXfbPrimitivesQuery;
+            funcsTable.GL.IsQueryResultAvailable = IsQueryResultAvailable;
+            funcsTable.GL.GetQueryResult64 = GetQueryResult64;
+            funcsTable.GL.DeleteBackendQuery = DeleteBackendQuery;
            funcsTableInitialized = true;
        }
        return funcsTable;
@@ -1250,10 +1250,76 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        pVulkanRenderer->Clear(mask);
    }

+    // Vulkan has no LINE_LOOP topology; rewrite the draw as an indexed LINE_STRIP
+    // whose synthesized index list revisits the first vertex at the end.
+    static void DrawLineLoopAsIndexedStrip(const Vector<Uint32>& closedIndices, GLint basevertex) {
+        DrawIndexedCmd payload{};
+        payload.mode = GL_LINE_STRIP;
+        payload.indexBufferView.indexType = GL_UNSIGNED_INT;
+        payload.indexBufferView.indexByteOffset = reinterpret_cast<SizeT>(closedIndices.data());
+        payload.indexBufferView.indexByteSize = closedIndices.size() * sizeof(Uint32);
+        payload.indexBufferView.forceClientMemory = true;
+        payload.params.indexCount = static_cast<Uint32>(closedIndices.size());
+        payload.params.instanceCount = 1;
+        payload.params.vertexOffset = basevertex;
+        pVulkanRenderer->DrawElements(payload);
+    }
+
+    // Resolve a DrawElements index list (bound element-array buffer or client
+    // memory) into uint32 values with the loop-closing first index appended.
+    static Bool BuildClosedLineLoopIndices(GLsizei count, GLenum type, const void* indices,
+                                           Vector<Uint32>& outIndices) {
+        const SizeT indexSize = MG_Util::GetGLTypeSize(type);
+        if (indexSize == 0 || count < 2) {
+            return false;
+        }
+        const Uint8* indexBytes = nullptr;
+        const auto& vao = *MG_State::pGLContext->GetBoundVertexArray();
+        const auto& indexBufferShared = vao.GetIndexBufferBindingSlot().GetBoundObject();
+        if (indexBufferShared != nullptr) {
+            const SizeT offset = reinterpret_cast<SizeT>(indices);
+            const SizeT bufferSize = indexBufferShared->GetSize();
+            if (indexBufferShared->MappedData() == nullptr || offset > bufferSize ||
+                static_cast<SizeT>(count) * indexSize > bufferSize - offset) {
+                return false;
+            }
+            indexBufferShared->SyncPersistentMappedRange();
+            indexBytes = indexBufferShared->MappedData() + offset;
+        } else {
+            indexBytes = static_cast<const Uint8*>(indices);
+            if (indexBytes == nullptr) {
+                return false;
+            }
+        }
+        outIndices.resize(static_cast<SizeT>(count) + 1);
+        for (GLsizei i = 0; i < count; ++i) {
+            switch (indexSize) {
+            case 1: outIndices[i] = indexBytes[i]; break;
+            case 2: outIndices[i] = reinterpret_cast<const Uint16*>(indexBytes)[i]; break;
+            default: outIndices[i] = reinterpret_cast<const Uint32*>(indexBytes)[i]; break;
+            }
+        }
+        outIndices[count] = outIndices[0];
+        return true;
+    }
+
    void DrawArrays(GLenum mode, GLint first, GLsizei count) {
        MOBILEGL_ASSERT(pVulkanRenderer, "DirectVulkan::DrawArrays called with null VulkanRenderer");
        MOBILEGL_ASSERT(MG_State::pGLContext, "DirectVulkan::DrawArrays called with null GL context");

+        if (mode == GL_LINE_LOOP) {
+            if (count < 2) {
+                return;
+            }
+            Vector<Uint32> closedIndices(static_cast<SizeT>(count) + 1);
+            for (GLsizei i = 0; i < count; ++i) {
+                closedIndices[i] = static_cast<Uint32>(first + i);
+            }
+            closedIndices[count] = static_cast<Uint32>(first);
+            DrawLineLoopAsIndexedStrip(closedIndices, 0);
+            return;
+        }
+
        DrawCmd payload{};
        payload.mode = mode;
        payload.params.firstVertex = first;
@@ -1266,6 +1332,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        MOBILEGL_ASSERT(pVulkanRenderer, "DirectVulkan::DrawElements called with null VulkanRenderer");
        MOBILEGL_ASSERT(MG_State::pGLContext, "DirectVulkan::DrawElements called with null GL context");

+        if (mode == GL_LINE_LOOP) {
+            Vector<Uint32> closedIndices;
+            if (BuildClosedLineLoopIndices(count, type, indices, closedIndices)) {
+                DrawLineLoopAsIndexedStrip(closedIndices, 0);
+            }
+            return;
+        }
+
        DrawIndexedCmd payload{};
        payload.mode = mode;
        payload.indexBufferView.indexType = type;
@@ -1334,6 +1408,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    void DrawElementsBaseVertex(GLenum mode, GLsizei count, GLenum type, const GLvoid* indices, GLint basevertex) {
        MOBILEGL_ASSERT(pVulkanRenderer, "DirectVulkan::DrawElementsBaseVertex called with null VulkanRenderer");
        MOBILEGL_ASSERT(MG_State::pGLContext, "DirectVulkan::DrawElementsBaseVertex called with null GL context");
+        if (mode == GL_LINE_LOOP) {
+            Vector<Uint32> closedIndices;
+            if (BuildClosedLineLoopIndices(count, type, indices, closedIndices)) {
+                DrawLineLoopAsIndexedStrip(closedIndices, basevertex);
+            }
+            return;
+        }
        DrawIndexedCmd payload{};
        payload.mode = mode;
        payload.indexBufferView.indexType = type;
@@ -1483,8 +1564,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // records are shared (SharedPtr) with the owning pool's pending list,
        // so deleting the query while results are still in flight is safe.
        struct VulkanTimerQuery {
+            enum class Kind : Uint8 { Timer, Occlusion, XfbWritten, XfbGenerated };
+            Kind kind = Kind::Timer;
            SharedPtr<VkTimerQueryManager::TimestampRecord> begin;
            SharedPtr<VkTimerQueryManager::TimestampRecord> end;
+            // Kind::Occlusion - pool slots recorded between Begin/End; summed at result time.
+            Vector<Uint32> occlusionSlots;
            // Renderer generation the records were written under (see
            // g_rendererGeneration). A stale generation resolves as available
            // with a final zero result: the records' pool indices and frame
@@ -1574,6 +1659,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            // ever be produced, so resolve with a final 0.
            return true;
        }
+        if (query->kind == VulkanTimerQuery::Kind::Occlusion) {
+            Uint64 samples = 0;
+            if (!pVulkanRenderer->ResolveOcclusionQueryResult(query->occlusionSlots, samples)) {
+                return false;
+            }
+            query->occlusionSlots.clear(); // slots are recycled by the resolve
+            *outNanoseconds = samples;
+            return true;
+        }
+        if (query->kind == VulkanTimerQuery::Kind::XfbWritten ||
+            query->kind == VulkanTimerQuery::Kind::XfbGenerated) {
+            Uint64 primitives = 0;
+            if (!pVulkanRenderer->ResolveXfbQueryResult(query->occlusionSlots,
+                                                        query->kind == VulkanTimerQuery::Kind::XfbGenerated,
+                                                        primitives)) {
+                return false;
+            }
+            *outNanoseconds = primitives;
+            return true;
+        }
        // With wait, mirrors ClientWaitSync: a query ended this frame cannot
        // complete until Present submits the commands, so the wait refuses to
        // block on the current unsubmitted serial. Returning false keeps the
@@ -1606,6 +1711,47 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        delete static_cast<VulkanTimerQuery*>(handle);
    }

+    BackendQueryHandle BeginXfbPrimitivesQuery(Bool generated) {
+        MOBILEGL_ASSERT(pVulkanRenderer, "DirectVulkan::BeginXfbPrimitivesQuery called with null VulkanRenderer");
+        if (!pVulkanRenderer->StartXfbQueryCapture(generated ? 1u : 0u)) {
+            return nullptr;
+        }
+        auto* query = new VulkanTimerQuery{};
+        query->kind = generated ? VulkanTimerQuery::Kind::XfbGenerated : VulkanTimerQuery::Kind::XfbWritten;
+        query->rendererGeneration = GetRendererGeneration();
+        return query;
+    }
+
+    void EndXfbPrimitivesQuery(BackendQueryHandle handle) {
+        auto* query = static_cast<VulkanTimerQuery*>(handle);
+        MOBILEGL_ASSERT(pVulkanRenderer, "DirectVulkan::EndXfbPrimitivesQuery called with null VulkanRenderer");
+        if (query == nullptr || query->rendererGeneration != GetRendererGeneration()) {
+            return;
+        }
+        pVulkanRenderer->StopXfbQueryCapture(
+            query->kind == VulkanTimerQuery::Kind::XfbGenerated ? 1u : 0u, query->occlusionSlots);
+    }
+
+    BackendQueryHandle BeginOcclusionQuery() {
+        MOBILEGL_ASSERT(pVulkanRenderer, "DirectVulkan::BeginOcclusionQuery called with null VulkanRenderer");
+        if (!pVulkanRenderer->StartOcclusionQueryCapture()) {
+            return nullptr;
+        }
+        auto* query = new VulkanTimerQuery{};
+        query->kind = VulkanTimerQuery::Kind::Occlusion;
+        query->rendererGeneration = GetRendererGeneration();
+        return query;
+    }
+
+    void EndOcclusionQuery(BackendQueryHandle handle) {
+        auto* query = static_cast<VulkanTimerQuery*>(handle);
+        MOBILEGL_ASSERT(pVulkanRenderer, "DirectVulkan::EndOcclusionQuery called with null VulkanRenderer");
+        if (query == nullptr || query->rendererGeneration != GetRendererGeneration()) {
+            return;
+        }
+        pVulkanRenderer->StopOcclusionQueryCapture(query->occlusionSlots);
+    }
+
    Int64 GetGpuTimestampNs() {
        // Vulkan cannot synchronously sample the GPU clock: timestamps only
        // exist as vkCmdWriteTimestamp results read back later, and
@@ -123,6 +123,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    // only while a live renderer exists whose device can actually time.
    Bool IsTimerQuerySupported();
    BackendQueryHandle BeginTimeElapsedQuery();
+    BackendQueryHandle BeginXfbPrimitivesQuery(Bool generated);
+    void EndXfbPrimitivesQuery(BackendQueryHandle query);
+    BackendQueryHandle BeginOcclusionQuery();
+    void EndOcclusionQuery(BackendQueryHandle query);
    void EndTimeElapsedQuery(BackendQueryHandle query);
    BackendQueryHandle QueryCounterTimestamp();
    Bool IsQueryResultAvailable(BackendQueryHandle query);
@@ -923,6 +923,163 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            ProgramFactory::CompileOptionFlags m_transformFlags;
        };

+        // Decorates the module's captured varyings for VK_EXT_transform_feedback:
+        // user outputs get XfbBuffer/XfbStride/Offset directly; a captured
+        // gl_Position (a gl_PerVertex member) is mirrored into a dedicated output
+        // variable copied before every OpReturn, BEFORE the position fixup runs,
+        // so the captured value is the shader's own (pre-remap) gl_Position.
+        class XfbCaptureDecoratePass final : public spvtools::opt::Pass {
+        public:
+            struct CapturedVarying {
+                std::string name;
+                Uint32 bufferIndex = 0;
+                Uint32 offsetBytes = 0;
+            };
+            const char* name() const override { return "mobilegl-xfb-capture-decorate"; }
+            XfbCaptureDecoratePass(Vector<CapturedVarying> varyings, Vector<Uint32> strides)
+                : m_varyings(Move(varyings)), m_strides(Move(strides)) {}
+
+            Status Process() override {
+                using namespace spvtools::opt;
+                if (m_varyings.empty()) return Status::SuccessWithoutChange;
+
+                auto entryPointIter = get_module()->entry_points().begin();
+                if (entryPointIter == get_module()->entry_points().end()) return Status::SuccessWithoutChange;
+                spvtools::opt::Instruction* entryPoint = &*entryPointIter;
+                const Uint32 entryFunctionId = entryPoint->GetSingleWordInOperand(1);
+
+                // Name -> result id map from the debug section.
+                std::unordered_map<std::string, Uint32> idsByName;
+                for (auto& debugInst : get_module()->debugs2()) {
+                    if (debugInst.opcode() != spv::Op::OpName) continue;
+                    idsByName[debugInst.GetInOperand(1).AsString()] = debugInst.GetSingleWordInOperand(0);
+                }
+
+                auto* decorationManager = context()->get_decoration_mgr();
+                const auto decorateForXfb = [&](Uint32 targetId, Uint32 bufferIndex, Uint32 offsetBytes) {
+                    const Uint32 stride = bufferIndex < m_strides.size() ? m_strides[bufferIndex] : 0;
+                    decorationManager->AddDecorationVal(targetId, static_cast<Uint32>(spv::Decoration::XfbBuffer),
+                                                        bufferIndex);
+                    decorationManager->AddDecorationVal(targetId, static_cast<Uint32>(spv::Decoration::XfbStride),
+                                                        stride);
+                    decorationManager->AddDecorationVal(targetId, static_cast<Uint32>(spv::Decoration::Offset),
+                                                        offsetBytes);
+                };
+
+                Bool modified = false;
+                Bool needsPositionMirror = false;
+                Uint32 positionBufferIndex = 0;
+                Uint32 positionOffset = 0;
+                for (const auto& varying : m_varyings) {
+                    if (varying.name == "gl_Position") {
+                        needsPositionMirror = true;
+                        positionBufferIndex = varying.bufferIndex;
+                        positionOffset = varying.offsetBytes;
+                        continue;
+                    }
+                    const auto idIt = idsByName.find(varying.name);
+                    if (idIt == idsByName.end()) {
+                        MGLOG_E("XfbCaptureDecoratePass: no SPIR-V variable named '%s'", varying.name.c_str());
+                        continue;
+                    }
+                    decorateForXfb(idIt->second, varying.bufferIndex, varying.offsetBytes);
+                    modified = true;
+                }
+
+                if (needsPositionMirror) {
+                    modified |= MirrorPositionForCapture(entryFunctionId, *entryPoint, positionBufferIndex,
+                                                         positionOffset, decorateForXfb);
+                }
+
+                if (!modified) return Status::SuccessWithoutChange;
+
+                context()->AddCapability(spv::Capability::TransformFeedback);
+                {
+                    auto executionMode = MakeUnique<spvtools::opt::Instruction>(
+                        context(), spv::Op::OpExecutionMode, 0, 0,
+                        std::initializer_list<spvtools::opt::Operand>{
+                            {SPV_OPERAND_TYPE_ID, {entryPoint->GetSingleWordInOperand(1)}},
+                            {SPV_OPERAND_TYPE_EXECUTION_MODE, {static_cast<Uint32>(spv::ExecutionMode::Xfb)}}});
+                    get_module()->AddExecutionMode(Move(executionMode));
+                }
+                context()->InvalidateAnalysesExceptFor(spvtools::opt::IRContext::kAnalysisNone);
+                return Status::SuccessWithChange;
+            }
+
+        private:
+            template <typename DecorateFn>
+            Bool MirrorPositionForCapture(Uint32 entryFunctionId, spvtools::opt::Instruction& entryPoint,
+                                          Uint32 bufferIndex, Uint32 offsetBytes, const DecorateFn& decorateForXfb) {
+                const Uint32 entryPointModel = entryPoint.GetSingleWordInOperand(0);
+                using namespace spvtools::opt;
+                PositionTargetInfo target{};
+                if (!FindPositionTarget(context(), &target)) {
+                    MGLOG_E("XfbCaptureDecoratePass: gl_Position capture requested but no position output found");
+                    return false;
+                }
+                if (!target.isMember) {
+                    // Standalone gl_Position variable: decorate it directly.
+                    decorateForXfb(target.variableId, bufferIndex, offsetBytes);
+                    return true;
+                }
+
+                auto* typeManager = context()->get_type_mgr();
+                const Uint32 mirrorPointerTypeId =
+                    typeManager->FindPointerToType(target.vectorTypeId, spv::StorageClass::Output);
+                if (mirrorPointerTypeId == 0) return false;
+
+                const Uint32 mirrorVariableId = context()->TakeNextId();
+                auto mirrorVariable = MakeUnique<Instruction>(
+                    context(), spv::Op::OpVariable, mirrorPointerTypeId, mirrorVariableId,
+                    std::initializer_list<Operand>{
+                        {SPV_OPERAND_TYPE_STORAGE_CLASS, {static_cast<Uint32>(spv::StorageClass::Output)}}});
+                get_module()->AddGlobalValue(Move(mirrorVariable));
+
+                // A free output location: past every explicitly decorated output.
+                Uint32 mirrorLocation = 0;
+                for (auto& annotation : get_module()->annotations()) {
+                    if (annotation.opcode() != spv::Op::OpDecorate ||
+                        annotation.GetSingleWordInOperand(1) != static_cast<Uint32>(spv::Decoration::Location)) {
+                        continue;
+                    }
+                    mirrorLocation = std::max(mirrorLocation, annotation.GetSingleWordInOperand(2) + 1);
+                }
+                auto* decorationManager = context()->get_decoration_mgr();
+                decorationManager->AddDecorationVal(mirrorVariableId,
+                                                    static_cast<Uint32>(spv::Decoration::Location), mirrorLocation);
+                decorateForXfb(mirrorVariableId, bufferIndex, offsetBytes);
+                entryPoint.AddOperand({SPV_OPERAND_TYPE_ID, {mirrorVariableId}});
+
+                auto* function = context()->GetFunction(entryFunctionId);
+                if (function == nullptr) return false;
+                const auto model = static_cast<spv::ExecutionModel>(entryPointModel);
+                Bool injected = false;
+                for (auto& block : *function) {
+                    for (auto instIter = block.begin(); instIter != block.end(); ++instIter) {
+                        // Geometry stages capture per emitted vertex; other stages at return.
+                        const Bool isInjectionSite =
+                            model == spv::ExecutionModel::Geometry
+                                ? instIter->opcode() == spv::Op::OpEmitVertex
+                                : instIter->opcode() == spv::Op::OpReturn;
+                        if (!isInjectionSite) continue;
+                        InstructionBuilder builder(context(), &*instIter, IRContext::kAnalysisNone);
+                        const Uint32 memberIndexId = builder.GetUintConstantId(target.memberIndex);
+                        auto* access =
+                            builder.AddAccessChain(target.vectorPtrTypeId, target.variableId, {memberIndexId});
+                        if (access == nullptr) return injected;
+                        auto* value = builder.AddLoad(target.vectorTypeId, access->result_id());
+                        if (value == nullptr) return injected;
+                        builder.AddStore(mirrorVariableId, value->result_id());
+                        injected = true;
+                    }
+                }
+                return injected;
+            }
+
+            Vector<CapturedVarying> m_varyings;
+            Vector<Uint32> m_strides;
+        };
+
        // Adreno 650 (driver 512.502) faults the GPU on an implicit-LOD sample of a full-screen
        // colour render target: the texture unit's derivative path reads outside the image's
        // allocation even though the sampler clamps LOD to 0 and the mapping is 1:1. MobileGL's
@@ -1128,6 +1285,41 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            return spvtools::Optimizer::PassToken(MakeUnique<GlToVulkanPositionFixPass>(transformFlags));
        }

+        Bool TransformSpirvForXfbCapture(const Vector<Uint>& input, Vector<Uint>& output,
+                                         const MG_State::GLState::ProgramObject& program) {
+            if (input.empty()) {
+                output.clear();
+                return true;
+            }
+            Vector<XfbCaptureDecoratePass::CapturedVarying> varyings;
+            varyings.reserve(program.GetTransformFeedbackVaryingCount());
+            for (const auto& varying : program.GetTransformFeedbackVaryings()) {
+                varyings.push_back({varying.name, varying.bufferIndex, varying.offsetBytes});
+            }
+            Vector<Uint32> strides;
+            strides.reserve(program.GetTransformFeedbackBufferCount());
+            for (SizeT i = 0; i < program.GetTransformFeedbackBufferCount(); ++i) {
+                strides.push_back(program.GetTransformFeedbackStride(static_cast<Uint32>(i)));
+            }
+
+            spvtools::Optimizer optimizer(SPV_ENV_VULKAN_1_3);
+            spvtools::OptimizerOptions options;
+            options.set_run_validator(false);
+            optimizer.SetMessageConsumer([](spv_message_level_t, const char*, const spv_position_t&,
+                                            const char* message) {
+                MGLOG_E("Vulkan: xfb capture pass: %s", message != nullptr ? message : "");
+            });
+            optimizer.RegisterPass(spvtools::Optimizer::PassToken(
+                MakeUnique<XfbCaptureDecoratePass>(Move(varyings), Move(strides))));
+
+            const Bool success = optimizer.Run(input.data(), input.size(), &output, options);
+            if (!success) {
+                MGLOG_E("Vulkan: xfb capture decoration pass failed; keeping the original module");
+                output = input;
+            }
+            return success;
+        }
+
        Bool TransformSpirvForVulkanPositionFix(const Vector<Uint>& input, Vector<Uint>& output,
                                                ProgramFactory::CompileOptionFlags transformFlags) {
            if (input.empty()) {
@@ -2173,7 +2365,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {

            // Apply position fixup if needed
            if (fixupStage != ShaderStage::Unknown && shaders[i] && shaders[i]->GetShaderStage() == fixupStage) {
-                TransformSpirvForVulkanPositionFix(spv, moduleSpirvs[i], flags);
+                const Vector<Uint>* fixupInput = &spv;
+                Vector<Uint> xfbSpirv;
+                if ((flags & ProgramFactory::CompileOptionBit::XfbCapture) &&
+                    program.GetTransformFeedbackVaryingCount() > 0) {
+                    // Decorate BEFORE the position fixup so a captured gl_Position
+                    // mirror copies the shader's own (pre-remap) value.
+                    if (TransformSpirvForXfbCapture(spv, xfbSpirv, program)) {
+                        fixupInput = &xfbSpirv;
+                    }
+                }
+                TransformSpirvForVulkanPositionFix(*fixupInput, moduleSpirvs[i], flags);
            } else {
                moduleSpirvs[i] = spv;
            }
@@ -47,6 +47,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            // level, which makes the two forms produce identical texels (the implicit lambda is
            // clamped into [minLod, maxLod] = [0, 0] regardless of derivatives or bias).
            ExplicitLod0Sampling = 1 << 5,
+            // Decorates the last vertex-processing stage's captured varyings with
+            // XfbBuffer/XfbStride/Offset (VK_EXT_transform_feedback). Set only for draws
+            // recorded while GL transform feedback is active, so plain draws keep the
+            // undecorated variant.
+            XfbCapture = 1 << 6,
        };
        using CompileOptionFlags = Flags<CompileOptionBit>;
        using HashType = Uint64;
@@ -22,6 +22,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT | VK_BUFFER_USAGE_STORAGE_BUFFER_BIT |
            VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT | VK_BUFFER_USAGE_UNIFORM_TEXEL_BUFFER_BIT |
            VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
+        // Appended to kPersistentBackedUsage when VK_EXT_transform_feedback is enabled
+        // (see VkBufferManagerInitInfo::transformFeedbackUsageEnabled).
+        constexpr VkBufferUsageFlags kTransformFeedbackUsage =
+            VK_BUFFER_USAGE_TRANSFORM_FEEDBACK_BUFFER_BIT_EXT;
        // The app writes into the persistent map with no explicit flush, so its memory must
        // be host-coherent (Adreno host-visible memory is; requiring it keeps us portable).
        constexpr VkMemoryPropertyFlags kPersistentBackedRequiredFlags =
@@ -465,7 +469,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        // it from the current shadow - MappedData() is still the shadow here because the
        // frontend adopts (and drops) the shadow only after this returns.
        DeferRelease(std::move(resource->buffer));
-        if (!CreateResidentStorage(*resource, size, kPersistentBackedUsage, kPersistentBackedRequiredFlags)) {
+        const VkBufferUsageFlags persistentUsage =
+            kPersistentBackedUsage |
+            (m_initInfo.transformFeedbackUsageEnabled ? kTransformFeedbackUsage : 0);
+        if (!CreateResidentStorage(*resource, size, persistentUsage, kPersistentBackedRequiredFlags)) {
            resource->persistentMapped = false;
            resource->storageSize = 0;
            resource->usageFlags = 0;
@@ -31,6 +31,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        VmaMemoryUsage transientMemoryUsage = VMA_MEMORY_USAGE_AUTO;
        VmaAllocationCreateFlags transientAllocationFlags = VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT;
        Bool transientPersistentMapping = false;
+        // VK_EXT_transform_feedback is enabled: persistent-map storage additionally
+        // carries the transform feedback usage so capture targets can bind directly.
+        Bool transformFeedbackUsageEnabled = false;
    };

    // The DirectVulkan storage behind one frontend buffer (pipe_resource analogue).
@@ -16,31 +16,21 @@

 namespace MobileGL::MG_Backend::DirectVulkan {
    static Bool TryResolveSampleCountFlagBits(Int requestedSamples, VkSampleCountFlagBits& outSampleCount) {
-        switch (requestedSamples <= 0 ? 1 : requestedSamples) {
-        case 1:
+        // GL promises "at least the requested samples", so a non-power-of-two
+        // request (legal in GL, e.g. 3) rounds up to the next Vulkan bit.
+        if (requestedSamples <= 1) {
            outSampleCount = VK_SAMPLE_COUNT_1_BIT;
            return true;
-        case 2:
-            outSampleCount = VK_SAMPLE_COUNT_2_BIT;
-            return true;
-        case 4:
-            outSampleCount = VK_SAMPLE_COUNT_4_BIT;
-            return true;
-        case 8:
-            outSampleCount = VK_SAMPLE_COUNT_8_BIT;
-            return true;
-        case 16:
-            outSampleCount = VK_SAMPLE_COUNT_16_BIT;
-            return true;
-        case 32:
-            outSampleCount = VK_SAMPLE_COUNT_32_BIT;
-            return true;
-        case 64:
-            outSampleCount = VK_SAMPLE_COUNT_64_BIT;
-            return true;
-        default:
+        }
+        if (requestedSamples > 64) {
            return false;
        }
+        Uint32 bit = 1;
+        while (bit < static_cast<Uint32>(requestedSamples)) {
+            bit <<= 1;
+        }
+        outSampleCount = static_cast<VkSampleCountFlagBits>(bit);
+        return true;
    }

    static VkImageAspectFlags ResolveImageAspectMaskForFormat(VkFormat format) {
@@ -166,6 +156,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        if (view != VK_NULL_HANDLE) {
            vkDestroyImageView(device, view, nullptr);
        }
+        if (unormTwinView != VK_NULL_HANDLE) {
+            vkDestroyImageView(device, unormTwinView, nullptr);
+        }
        if (image != VK_NULL_HANDLE && allocation != nullptr) {
            vmaDestroyImage(allocator, image, allocation);
        }
@@ -173,6 +166,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        image = VK_NULL_HANDLE;
        allocation = nullptr;
        view = VK_NULL_HANDLE;
+        unormTwinView = VK_NULL_HANDLE;
        layout = VK_IMAGE_LAYOUT_UNDEFINED;
        format = VK_FORMAT_UNDEFINED;
        aspect = VK_IMAGE_ASPECT_NONE;
@@ -231,10 +225,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        if (resource.image == VK_NULL_HANDLE && resource.view == VK_NULL_HANDLE) {
            return;
        }
-        m_deferredRenderbufferReleases.push_back({resource.image, resource.allocation, resource.view, m_frameCounter});
+        m_deferredRenderbufferReleases.push_back(
+            {resource.image, resource.allocation, resource.view, resource.unormTwinView, m_frameCounter});
        resource.image = VK_NULL_HANDLE;
        resource.allocation = nullptr;
        resource.view = VK_NULL_HANDLE;
+        resource.unormTwinView = VK_NULL_HANDLE;
    }

    void VkRenderPassManager::CollectDeferredRenderbufferReleases(Bool destroyAll) {
@@ -249,6 +245,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            if (release.view != VK_NULL_HANDLE) {
                vkDestroyImageView(m_device, release.view, nullptr);
            }
+            if (release.unormTwinView != VK_NULL_HANDLE) {
+                vkDestroyImageView(m_device, release.unormTwinView, nullptr);
+            }
            if (release.image != VK_NULL_HANDLE) {
                vmaDestroyImage(m_allocator, release.image, release.allocation);
            }
@@ -304,7 +303,47 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        }

        const auto internalFormat = renderbuffer->GetInternalFormat();
-        const VkFormat format = MG_Util::ConvertTextureInternalFormatToVkEnum(internalFormat);
+        // Three-channel color formats widen to their RGBA twin exactly like textures do
+        // (VkTextureManager::ResolveTextureFormatInfo): blits/resolves between a
+        // renderbuffer and a texture of the same GL format then see one VkFormat.
+        const VkFormat format = [&]() -> VkFormat {
+            switch (internalFormat) {
+            case TextureInternalFormat::RGB:
+            case TextureInternalFormat::RGB8:
+            case TextureInternalFormat::R3G3B2:
+            case TextureInternalFormat::RGB4:
+            case TextureInternalFormat::RGB5:
+                return VK_FORMAT_R8G8B8A8_UNORM;
+            case TextureInternalFormat::SRGB8:
+                return VK_FORMAT_R8G8B8A8_SRGB;
+            case TextureInternalFormat::RGB8Snorm:
+                return VK_FORMAT_R8G8B8A8_SNORM;
+            case TextureInternalFormat::RGB10:
+            case TextureInternalFormat::RGB12:
+            case TextureInternalFormat::RGB16:
+                return VK_FORMAT_R16G16B16A16_UNORM;
+            case TextureInternalFormat::RGB16Snorm:
+                return VK_FORMAT_R16G16B16A16_SNORM;
+            case TextureInternalFormat::RGB16F:
+                return VK_FORMAT_R16G16B16A16_SFLOAT;
+            case TextureInternalFormat::RGB32F:
+                return VK_FORMAT_R32G32B32A32_SFLOAT;
+            case TextureInternalFormat::RGB8I:
+                return VK_FORMAT_R8G8B8A8_SINT;
+            case TextureInternalFormat::RGB8UI:
+                return VK_FORMAT_R8G8B8A8_UINT;
+            case TextureInternalFormat::RGB16I:
+                return VK_FORMAT_R16G16B16A16_SINT;
+            case TextureInternalFormat::RGB16UI:
+                return VK_FORMAT_R16G16B16A16_UINT;
+            case TextureInternalFormat::RGB32I:
+                return VK_FORMAT_R32G32B32A32_SINT;
+            case TextureInternalFormat::RGB32UI:
+                return VK_FORMAT_R32G32B32A32_UINT;
+            default:
+                return MG_Util::ConvertTextureInternalFormatToVkEnum(internalFormat);
+            }
+        }();
        const VkImageAspectFlags aspect = ResolveImageAspectMaskForFormat(format);
        // Renderbuffers are never sampled (GL has no way to bind one to a sampler), so the
        // usage set is attachment + transfer: transfer covers readback (vkCmdCopyImageToBuffer),
@@ -314,6 +353,46 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                       : VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) |
            VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT;

+        // GL allows the implementation to allocate more samples than requested
+        // (glRenderbufferStorageMultisample only promises "at least"), and devices
+        // like llvmpipe expose 1x/4x but not 2x. Round the request up to the
+        // nearest supported count for this format.
+        if (renderbuffer->GetSamples() > 0) {
+            auto supportedIt = m_attachmentSampleCountsByFormat.find(format);
+            if (supportedIt == m_attachmentSampleCountsByFormat.end()) {
+                VkImageFormatProperties formatProperties{};
+                VkSampleCountFlags supported = VK_SAMPLE_COUNT_1_BIT;
+                if (vkGetPhysicalDeviceImageFormatProperties(m_physicalDevice, format, VK_IMAGE_TYPE_2D,
+                                                             VK_IMAGE_TILING_OPTIMAL, imageUsage, 0,
+                                                             &formatProperties) == VK_SUCCESS) {
+                    supported = formatProperties.sampleCounts;
+                }
+                supportedIt = m_attachmentSampleCountsByFormat.emplace(format, supported).first;
+            }
+            const VkSampleCountFlags supported = supportedIt->second;
+            if ((supported & sampleCount) == 0) {
+                // Smallest supported count above the request, else the largest below it.
+                Uint32 rounded = 0;
+                for (Uint32 bit = static_cast<Uint32>(sampleCount) << 1; bit <= VK_SAMPLE_COUNT_64_BIT; bit <<= 1) {
+                    if ((supported & bit) != 0) {
+                        rounded = bit;
+                        break;
+                    }
+                }
+                if (rounded == 0) {
+                    for (Uint32 bit = static_cast<Uint32>(sampleCount) >> 1; bit != 0; bit >>= 1) {
+                        if ((supported & bit) != 0) {
+                            rounded = bit;
+                            break;
+                        }
+                    }
+                }
+                if (rounded != 0) {
+                    sampleCount = static_cast<VkSampleCountFlagBits>(rounded);
+                }
+            }
+        }
+
        auto& resource = m_renderbufferResources[renderbuffer.get()];
        const Bool needsCreate =
            resource.image == VK_NULL_HANDLE ||
@@ -351,6 +430,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        imageInfo.usage = imageUsage;
        imageInfo.samples = sampleCount;
        imageInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
+        // sRGB renderbuffers attach through their UNORM twin while GL_FRAMEBUFFER_SRGB
+        // is disabled, which needs a format-reinterpreting second view.
+        const Bool hasUnormTwin = ResolveSrgbAttachmentWriteFormat(format, false) != format;
+        if (hasUnormTwin) {
+            imageInfo.flags |= VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
+        }

        VkImageFormatProperties imageFormatProperties{};
        const VkResult imageFormatResult = vkGetPhysicalDeviceImageFormatProperties(
@@ -385,6 +470,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        viewInfo.subresourceRange.layerCount = 1;
        VK_VERIFY(vkCreateImageView(m_device, &viewInfo, nullptr, &resource.view),
                  "vkCreateImageView(renderbuffer)");
+        if (hasUnormTwin) {
+            viewInfo.format = ResolveSrgbAttachmentWriteFormat(format, false);
+            VK_VERIFY(vkCreateImageView(m_device, &viewInfo, nullptr, &resource.unormTwinView),
+                      "vkCreateImageView(renderbuffer unorm twin)");
+        }

        resource.layout = VK_IMAGE_LAYOUT_UNDEFINED;
        resource.format = format;
@@ -488,6 +578,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        if (isDefaultFbo) {
            XXHASH_VERIFY(XXH64_update(m_hashState, &swapchainImageIndex, sizeof(swapchainImageIndex)));
        }
+        // sRGB attachments switch between their sRGB and UNORM-twin views with this
+        // capability (ResolveSrgbAttachmentWriteFormat), changing the render pass formats.
+        const Bool framebufferSrgbEnabled =
+            MG_State::pGLContext->IsCapabilityEnabled(MobileGL::CapabilityInput::FramebufferSrgb);
+        XXHASH_VERIFY(XXH64_update(m_hashState, &framebufferSrgbEnabled, sizeof(framebufferSrgbEnabled)));
        auto& drawBuffers = fbo.GetDrawBuffers();
        XXHASH_VERIFY(XXH64_update(m_hashState, drawBuffers.data(), drawBuffers.size() * sizeof(drawBuffers[0])));
        auto readBuffer = fbo.GetReadBuffer();
@@ -835,8 +930,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                    }

                    const VkImageLayout trackedRbLayout = rbResource->layout;
+                    const Bool rbFramebufferSrgb =
+                        MG_State::pGLContext->IsCapabilityEnabled(MobileGL::CapabilityInput::FramebufferSrgb);
+                    const VkFormat rbAttachmentFormat =
+                        ResolveSrgbAttachmentWriteFormat(rbResource->format, rbFramebufferSrgb);
                    rbDesc.flags = 0;
-                    rbDesc.format = rbResource->format;
+                    rbDesc.format = rbAttachmentFormat;
                    rbDesc.samples = rbResource->sampleCount;
                    rbDesc.loadOp = rbHasClear ? VK_ATTACHMENT_LOAD_OP_CLEAR :
                                    (trackedRbLayout == VK_IMAGE_LAYOUT_UNDEFINED ? VK_ATTACHMENT_LOAD_OP_DONT_CARE
@@ -871,7 +970,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                        .finalLayout = rbDesc.finalLayout,
                    });
                    textureResources.emplace_back(nullptr);
-                    attachmentViews.emplace_back(rbResource->view);
+                    attachmentViews.emplace_back(rbAttachmentFormat != rbResource->format ? rbResource->unormTwinView
+                                                                                          : rbResource->view);
                    MOBILEGL_ASSERT(attachmentViews.back() != VK_NULL_HANDLE,
                                    "GetOrCreateRenderPass: renderbuffer view missing at color attachment %d", i);

@@ -960,7 +1060,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                        MOBILEGL_ASSERT(textureResource,
                                        "GetOrCreateRenderPass: SyncTextureAndGetDescriptor failed at color attachment %d", i);
                        textureResources.emplace_back(textureResource);
-                        desc.format = textureResource->format;
+                        desc.format = ResolveSrgbAttachmentWriteFormat(
+                            textureResource->format,
+                            MG_State::pGLContext->IsCapabilityEnabled(MobileGL::CapabilityInput::FramebufferSrgb));
                        attachmentSampleCount = textureResource->sampleCount;
                        trackedColorLayout = textureResource->layout;
                        trackedAttachmentLayouts.emplace_back(TrackedAttachmentLayoutInfo {
@@ -274,6 +274,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            VkImage image = VK_NULL_HANDLE;
            VmaAllocation allocation = nullptr;
            VkImageView view = VK_NULL_HANDLE;
+            // UNORM reinterpretation of an sRGB image, used as the attachment view while
+            // GL_FRAMEBUFFER_SRGB is disabled (raw writes). Null for non-sRGB formats.
+            VkImageView unormTwinView = VK_NULL_HANDLE;
            VkImageLayout layout = VK_IMAGE_LAYOUT_UNDEFINED;
            VkFormat format = VK_FORMAT_UNDEFINED;
            VkImageAspectFlags aspect = VK_IMAGE_ASPECT_NONE;
@@ -307,12 +310,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            VkImage image = VK_NULL_HANDLE;
            VmaAllocation allocation = nullptr;
            VkImageView view = VK_NULL_HANDLE;
+            VkImageView unormTwinView = VK_NULL_HANDLE;
            Uint64 deferredAtFrame = 0;
        };

        UnorderedMap<MG_State::GLState::RenderbufferObject*, RenderbufferResource> m_renderbufferResources;
        UnorderedMap<MG_State::GLState::RenderbufferObject*, PendingRenderbufferClear> m_pendingRenderbufferClears;
        Vector<DeferredRenderbufferRelease> m_deferredRenderbufferReleases;
+        // Supported sample counts per attachment format, so per-draw resource lookups
+        // do not repeat vkGetPhysicalDeviceImageFormatProperties.
+        UnorderedMap<VkFormat, VkSampleCountFlags> m_attachmentSampleCountsByFormat;

        Bool HasPendingRenderbufferClear(
            const MG_State::GLState::FramebufferAttachmentObject& attachment) const;
@@ -120,31 +120,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    }

    static Bool TryResolveSampleCountFlagBits(Int requestedSamples, VkSampleCountFlagBits& outSampleCount) {
-        switch (requestedSamples) {
-        case 1:
+        // GL promises "at least the requested samples", so a non-power-of-two
+        // request (legal in GL, e.g. 3) rounds up to the next Vulkan bit.
+        if (requestedSamples <= 1) {
            outSampleCount = VK_SAMPLE_COUNT_1_BIT;
            return true;
-        case 2:
-            outSampleCount = VK_SAMPLE_COUNT_2_BIT;
-            return true;
-        case 4:
-            outSampleCount = VK_SAMPLE_COUNT_4_BIT;
-            return true;
-        case 8:
-            outSampleCount = VK_SAMPLE_COUNT_8_BIT;
-            return true;
-        case 16:
-            outSampleCount = VK_SAMPLE_COUNT_16_BIT;
-            return true;
-        case 32:
-            outSampleCount = VK_SAMPLE_COUNT_32_BIT;
-            return true;
-        case 64:
-            outSampleCount = VK_SAMPLE_COUNT_64_BIT;
-            return true;
-        default:
+        }
+        if (requestedSamples > 64) {
            return false;
        }
+        Uint32 bit = 1;
+        while (bit < static_cast<Uint32>(requestedSamples)) {
+            bit <<= 1;
+        }
+        outSampleCount = static_cast<VkSampleCountFlagBits>(bit);
+        return true;
    }

    static Bool IsCubeMapFaceUploadTarget(TextureUploadTarget target) {
@@ -840,7 +830,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            return VK_NULL_HANDLE;
        }

-        if (baseArrayLayer == 0 && layerCount == resource->arrayLayers && viewType == resource->viewType) {
+        const Bool framebufferSrgbEnabled =
+            MG_State::pGLContext->IsCapabilityEnabled(MobileGL::CapabilityInput::FramebufferSrgb);
+        const VkFormat attachmentFormat = ResolveSrgbAttachmentWriteFormat(resource->format, framebufferSrgbEnabled);
+
+        if (attachmentFormat == resource->format && baseArrayLayer == 0 && layerCount == resource->arrayLayers &&
+            viewType == resource->viewType) {
            return GetOrCreateViewAtMipLevel(texture, mipLevel);
        }

@@ -849,6 +844,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            .baseArrayLayer = baseArrayLayer,
            .layerCount = layerCount,
            .viewType = viewType,
+            .viewFormat = attachmentFormat,
        };
        auto it = resource->attachmentViews.find(key);
        if (it == resource->attachmentViews.end()) {
@@ -859,7 +855,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            return attachmentView;
        }

-        attachmentView = CreateImageView(resource->image, resource->format, resource->aspect, viewType,
+        attachmentView = CreateImageView(resource->image, attachmentFormat, resource->aspect, viewType,
                                         mipLevel, 1, baseArrayLayer, layerCount);
        if (attachmentView == VK_NULL_HANDLE) {
            MGLOG_D("%s: CreateImageView failed for textureId=%d mipLevel=%u baseArrayLayer=%u layerCount=%u viewType=%d",
@@ -1448,11 +1444,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                               const IntVec3 &texelSize, SizeT byteSize, Uint32 mipLevels,
                                               TextureResource &resource) {
        const TextureFormatInfo formatInfo = ResolveTextureFormatInfo(texture.GetFormat());
-        const VkFormat format = formatInfo.format;
+        VkFormat format = formatInfo.format;
        if (format == VK_FORMAT_UNDEFINED) {
            MGLOG_D("%s: format == VK_FORMAT_UNDEFINED", __func__);
            return false;
        }
+        // X8_D24 lacks optimal-tiling support on several drivers (lavapipe included);
+        // D32_SFLOAT holds every 24-bit depth value exactly, and the upload path
+        // converts the shadow words to float (see the pure-depth branch below).
+        if (format == VK_FORMAT_X8_D24_UNORM_PACK32) {
+            VkFormatProperties formatProperties{};
+            vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &formatProperties);
+            constexpr VkFormatFeatureFlags kDepthAttachmentAndSample =
+                VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT | VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT;
+            if ((formatProperties.optimalTilingFeatures & kDepthAttachmentAndSample) != kDepthAttachmentAndSample) {
+                format = VK_FORMAT_D32_SFLOAT;
+            }
+        }
        if (texelSize.x() <= 0 || texelSize.y() <= 0 /*|| byteSize == 0*/) {
            MGLOG_D("%s: texelSize or byteSize is zero", __func__);
            return false;
@@ -1524,6 +1532,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            m_mutableFormatUnsupported.find(format) == m_mutableFormatUnsupported.end()) {
            imageCreateFlags |= VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
        }
+        // sRGB color images attach through their UNORM twin while GL_FRAMEBUFFER_SRGB is
+        // disabled (see ResolveSrgbAttachmentWriteFormat), which needs format-reinterpreting
+        // views - multisample sRGB render targets included.
+        if (ResolveSrgbAttachmentWriteFormat(format, false) != format &&
+            (aspect & VK_IMAGE_ASPECT_COLOR_BIT) != 0 &&
+            m_mutableFormatUnsupported.find(format) == m_mutableFormatUnsupported.end()) {
+            imageCreateFlags |= VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
+        }

        VkImageUsageFlags desiredUsage =
            VK_IMAGE_USAGE_SAMPLED_BIT |
@@ -1536,6 +1552,44 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            desiredUsage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
        }

+        // Round a multisample request up to a count the device supports for this
+        // format (GL only promises "at least"), mirroring the renderbuffer path.
+        if (isMultisampleTexture && resolvedSampleCount != VK_SAMPLE_COUNT_1_BIT) {
+            auto supportedIt = m_multisampleCountsByFormat.find(format);
+            if (supportedIt == m_multisampleCountsByFormat.end()) {
+                VkImageFormatProperties imageFormatProperties{};
+                VkSampleCountFlags supported = VK_SAMPLE_COUNT_1_BIT;
+                if (vkGetPhysicalDeviceImageFormatProperties(m_physicalDevice, format, shapeInfo.imageType,
+                                                             VK_IMAGE_TILING_OPTIMAL, desiredUsage, imageCreateFlags,
+                                                             &imageFormatProperties) == VK_SUCCESS) {
+                    supported = imageFormatProperties.sampleCounts;
+                }
+                supportedIt = m_multisampleCountsByFormat.emplace(format, supported).first;
+            }
+            const VkSampleCountFlags supported = supportedIt->second;
+            if ((supported & resolvedSampleCount) == 0) {
+                Uint32 rounded = 0;
+                for (Uint32 bit = static_cast<Uint32>(resolvedSampleCount) << 1; bit <= VK_SAMPLE_COUNT_64_BIT;
+                     bit <<= 1) {
+                    if ((supported & bit) != 0) {
+                        rounded = bit;
+                        break;
+                    }
+                }
+                if (rounded == 0) {
+                    for (Uint32 bit = static_cast<Uint32>(resolvedSampleCount) >> 1; bit != 0; bit >>= 1) {
+                        if ((supported & bit) != 0) {
+                            rounded = bit;
+                            break;
+                        }
+                    }
+                }
+                if (rounded != 0) {
+                    resolvedSampleCount = static_cast<VkSampleCountFlagBits>(rounded);
+                }
+            }
+        }
+
        const Bool compatible = resource.image != VK_NULL_HANDLE && resource.format == format &&
                                resource.extent.width == static_cast<Uint32>(texelSize.x()) &&
                                resource.extent.height == static_cast<Uint32>(texelSize.y()) &&
@@ -1665,8 +1719,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        VmaAllocationCreateInfo allocationInfo{};
        allocationInfo.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
        allocationInfo.requiredFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
-        VK_VERIFY(vmaCreateImage(m_allocator, &imageInfo, &allocationInfo, &resource.image, &resource.allocation, nullptr),
-                  "vmaCreateImage(texture)");
+        // Soft failure like the unsupported-sample-count path above: a driver can pass the
+        // vkGetPhysicalDeviceImageFormatProperties pre-check yet still refuse the creation
+        // (e.g. multisampled depth on lavapipe); the texture simply stays unbacked.
+        const VkResult createImageResult =
+            vmaCreateImage(m_allocator, &imageInfo, &allocationInfo, &resource.image, &resource.allocation, nullptr);
+        if (createImageResult != VK_SUCCESS) {
+            MGLOG_F("SyncTextureResource: vmaCreateImage failed (%d) textureId=%d extent=%ux%u depth=%u layers=%u "
+                    "mips=%u samples=%d format=%d",
+                    createImageResult, texture.GetExternalIndex(), imageInfo.extent.width, imageInfo.extent.height,
+                    imageInfo.extent.depth, imageInfo.arrayLayers, imageInfo.mipLevels,
+                    static_cast<Int>(imageInfo.samples), static_cast<Int>(imageInfo.format));
+            resource.image = VK_NULL_HANDLE;
+            resource.allocation = nullptr;
+            return false;
+        }
        ++m_textureImageEpoch; // a new attachment image invalidates cached render passes

        resource.layout = VK_IMAGE_LAYOUT_UNDEFINED;
@@ -1957,17 +2024,119 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            return true;
        }

-        // Combined depth-stencil images need per-aspect de-interleaved copies (VkBufferImageCopy
-        // aspectMask must have exactly one bit set). Until that is implemented, skip the upload
-        // instead of recording an invalid command buffer that kills the process.
+        // Combined depth-stencil images need per-aspect copies (VkBufferImageCopy aspectMask
+        // must have exactly one bit set), so de-interleave the shadow's GL wire format into
+        // a depth plane followed by a stencil plane per upload item.
        const VkImageAspectFlags uploadAspectMask = GetAspectMaskForFormat(outResource.format);
-        if ((uploadAspectMask & VK_IMAGE_ASPECT_DEPTH_BIT) && (uploadAspectMask & VK_IMAGE_ASPECT_STENCIL_BIT)) {
-            MGLOG_E("UploadDirtyMipLevels: skipping unimplemented depth-stencil data upload for textureId=%d",
-                    mipmapTexture.GetExternalIndex());
-            for (const auto& item : uploadItems) {
-                mipmapTexture.MarkStorageDirty(item.target, item.level, false);
+        const Bool isCombinedDepthStencil =
+            (uploadAspectMask & VK_IMAGE_ASPECT_DEPTH_BIT) && (uploadAspectMask & VK_IMAGE_ASPECT_STENCIL_BIT);
+        if (isCombinedDepthStencil) {
+            const Bool srcIsD24S8 = outResource.format == VK_FORMAT_D24_UNORM_S8_UINT;
+            const Bool srcIsD32FS8 = outResource.format == VK_FORMAT_D32_SFLOAT_S8_UINT;
+            if (!srcIsD24S8 && !srcIsD32FS8) {
+                MGLOG_E("UploadDirtyMipLevels: unsupported combined depth-stencil format %d for textureId=%d",
+                        static_cast<Int>(outResource.format), mipmapTexture.GetExternalIndex());
+                for (const auto& item : uploadItems) {
+                    mipmapTexture.MarkStorageDirty(item.target, item.level, false);
+                }
+                return true;
+            }
+            stagingSize = 0;
+            for (auto& item : uploadItems) {
+                const SizeT texelCount = static_cast<SizeT>(item.texelSize.x()) *
+                                         static_cast<SizeT>(item.texelSize.y()) *
+                                         static_cast<SizeT>(std::max(item.texelSize.z(), 1));
+                const SizeT shadowTexelSize = item.uploadByteSize / std::max<SizeT>(texelCount, 1);
+                MOBILEGL_ASSERT(shadowTexelSize == 4 || shadowTexelSize == 8,
+                                "UploadDirtyMipLevels: unexpected depth-stencil shadow texel size %zu for textureId=%d",
+                                shadowTexelSize, mipmapTexture.GetExternalIndex());
+                // Depth plane as the aspect's buffer-copy format (32-bit word for
+                // D24: low 24 bits; float for D32F), then one stencil byte per texel.
+                Vector<Uint8> deinterleaved(texelCount * 4 + texelCount);
+                Uint8* depthPlane = deinterleaved.data();
+                Uint8* stencilPlane = deinterleaved.data() + texelCount * 4;
+                const Uint8* shadow = static_cast<const Uint8*>(item.source);
+                for (SizeT t = 0; t < texelCount; ++t) {
+                    if (shadowTexelSize == 8) {
+                        // GL_FLOAT_32_UNSIGNED_INT_24_8_REV: float depth, then a word
+                        // with stencil in its low 8 bits.
+                        float depthValue;
+                        Uint32 stencilWord;
+                        std::memcpy(&depthValue, shadow + t * 8, sizeof(depthValue));
+                        std::memcpy(&stencilWord, shadow + t * 8 + 4, sizeof(stencilWord));
+                        if (srcIsD32FS8) {
+                            std::memcpy(depthPlane + t * 4, &depthValue, sizeof(depthValue));
+                        } else {
+                            const float clamped = std::min(std::max(depthValue, 0.0f), 1.0f);
+                            const Uint32 depthWord = static_cast<Uint32>(clamped * 16777215.0f + 0.5f);
+                            std::memcpy(depthPlane + t * 4, &depthWord, sizeof(depthWord));
+                        }
+                        stencilPlane[t] = static_cast<Uint8>(stencilWord & 0xFFu);
+                    } else {
+                        // GL_UNSIGNED_INT_24_8: depth in the high 24 bits, stencil low 8.
+                        Uint32 packed;
+                        std::memcpy(&packed, shadow + t * 4, sizeof(packed));
+                        if (srcIsD24S8) {
+                            const Uint32 depthWord = packed >> 8;
+                            std::memcpy(depthPlane + t * 4, &depthWord, sizeof(depthWord));
+                        } else {
+                            const float depthValue = static_cast<float>(packed >> 8) / 16777215.0f;
+                            std::memcpy(depthPlane + t * 4, &depthValue, sizeof(depthValue));
+                        }
+                        stencilPlane[t] = static_cast<Uint8>(packed & 0xFFu);
+                    }
+                }
+                item.expandedData = Move(deinterleaved);
+                item.source = item.expandedData.data();
+                item.uploadByteSize = item.expandedData.size();
+                item.offset = stagingSize;
+                stagingSize += static_cast<VkDeviceSize>(item.uploadByteSize);
+            }
+        }
+
+        // Pure-depth images whose canonical shadow layout differs from the image texel
+        // layout (the shadow keeps a full-scale 16/32-bit unorm word or a float; the
+        // image may be X8_D24 or a D32_SFLOAT fallback) convert per texel here.
+        if (uploadAspectMask == VK_IMAGE_ASPECT_DEPTH_BIT) {
+            const TextureInternalFormat depthInternal = mipmapTexture.GetFormat();
+            const Bool shadowIsFloat = depthInternal == TextureInternalFormat::DepthComponent32F;
+            const Bool dstIsFloat = outResource.format == VK_FORMAT_D32_SFLOAT;
+            const Bool dstIsD24Word = outResource.format == VK_FORMAT_X8_D24_UNORM_PACK32;
+            stagingSize = 0;
+            for (auto& item : uploadItems) {
+                const SizeT texelCount = static_cast<SizeT>(item.texelSize.x()) *
+                                         static_cast<SizeT>(item.texelSize.y()) *
+                                         static_cast<SizeT>(std::max(item.texelSize.z(), 1));
+                const SizeT shadowTexelSize = item.uploadByteSize / std::max<SizeT>(texelCount, 1);
+                const Bool needsConversion =
+                    (dstIsFloat && !shadowIsFloat) || (dstIsD24Word && shadowTexelSize == 4 && !shadowIsFloat);
+                if (needsConversion) {
+                    Vector<Uint8> converted(texelCount * 4);
+                    const Uint8* shadow = static_cast<const Uint8*>(item.source);
+                    for (SizeT t = 0; t < texelCount; ++t) {
+                        Uint32 wide = 0;
+                        if (shadowTexelSize == 2) {
+                            Uint16 raw = 0;
+                            std::memcpy(&raw, shadow + t * 2, sizeof(raw));
+                            wide = (static_cast<Uint32>(raw) << 16) | raw;
+                        } else {
+                            std::memcpy(&wide, shadow + t * 4, sizeof(wide));
+                        }
+                        if (dstIsFloat) {
+                            const float value = static_cast<float>(static_cast<double>(wide) / 4294967295.0);
+                            std::memcpy(converted.data() + t * 4, &value, sizeof(value));
+                        } else { // X8_D24: depth in the low 24 bits of a 32-bit word
+                            const Uint32 word = wide >> 8;
+                            std::memcpy(converted.data() + t * 4, &word, sizeof(word));
+                        }
+                    }
+                    item.expandedData = Move(converted);
+                    item.source = item.expandedData.data();
+                    item.uploadByteSize = item.expandedData.size();
+                }
+                item.offset = stagingSize;
+                stagingSize += static_cast<VkDeviceSize>(item.uploadByteSize);
            }
-            return true;
        }

        VkBuffer stagingBuffer = VK_NULL_HANDLE;
@@ -2020,7 +2189,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                        aspectMask, 0, outResource.mipLevels, outResource.arrayLayers);
        MOBILEGL_ASSERT(ok, "TransitionImageLayout to VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL failed");

+        // Array textures keep their GL "depth" in VkImage array layers, so the
+        // copy must address layerCount, not imageExtent.depth (which is invalid
+        // for 2D images and silently dropped every layer past the first).
+        const Bool depthSelectsArrayLayer = outResource.viewType == VK_IMAGE_VIEW_TYPE_1D_ARRAY ||
+                                            outResource.viewType == VK_IMAGE_VIEW_TYPE_2D_ARRAY ||
+                                            outResource.viewType == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY;
        for (const auto& item : uploadItems) {
+            const Uint32 depthOrLayers = item.texelSize.z() > 0 ? static_cast<Uint32>(item.texelSize.z()) : 1u;
            VkBufferImageCopy copy{};
            copy.bufferOffset = item.offset;
            copy.bufferRowLength = 0;
@@ -2028,10 +2204,24 @@ namespace MobileGL::MG_Backend::DirectVulkan {
            copy.imageSubresource.aspectMask = aspectMask;
            copy.imageSubresource.mipLevel = item.level;
            copy.imageSubresource.baseArrayLayer = item.baseArrayLayer;
-            copy.imageSubresource.layerCount = 1;
+            copy.imageSubresource.layerCount = depthSelectsArrayLayer ? depthOrLayers : 1;
            copy.imageOffset = {0, 0, 0};
            copy.imageExtent = {static_cast<Uint32>(item.texelSize.x()), static_cast<Uint32>(item.texelSize.y()),
-                                item.texelSize.z() > 0 ? static_cast<Uint32>(item.texelSize.z()) : 1u};
+                                depthSelectsArrayLayer ? 1u : depthOrLayers};
+            if (isCombinedDepthStencil) {
+                const SizeT texelCount = static_cast<SizeT>(item.texelSize.x()) *
+                                         static_cast<SizeT>(item.texelSize.y()) *
+                                         static_cast<SizeT>(std::max(item.texelSize.z(), 1));
+                VkBufferImageCopy depthCopy = copy;
+                depthCopy.imageSubresource.aspectMask = VK_IMAGE_ASPECT_DEPTH_BIT;
+                VkBufferImageCopy stencilCopy = copy;
+                stencilCopy.imageSubresource.aspectMask = VK_IMAGE_ASPECT_STENCIL_BIT;
+                stencilCopy.bufferOffset = item.offset + static_cast<VkDeviceSize>(texelCount) * 4;
+                const VkBufferImageCopy copies[2] = {depthCopy, stencilCopy};
+                vkCmdCopyBufferToImage(commandBuffer, stagingBuffer, outResource.image,
+                                       VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, 2, copies);
+                continue;
+            }
            vkCmdCopyBufferToImage(commandBuffer, stagingBuffer, outResource.image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
                                   1, &copy);
        }
@@ -67,12 +67,16 @@ public:
            Uint32 baseArrayLayer = 0;
            Uint32 layerCount = 1;
            VkImageViewType viewType = VK_IMAGE_VIEW_TYPE_2D;
+            // May differ from the image format: sRGB images attach through their UNORM
+            // twin while GL_FRAMEBUFFER_SRGB is disabled.
+            VkFormat viewFormat = VK_FORMAT_UNDEFINED;

            Bool operator==(const AttachmentViewKey& other) const {
                return mipLevel == other.mipLevel &&
                       baseArrayLayer == other.baseArrayLayer &&
                       layerCount == other.layerCount &&
-                       viewType == other.viewType;
+                       viewType == other.viewType &&
+                       viewFormat == other.viewFormat;
            }
        };

@@ -83,6 +87,8 @@ public:
                hash ^= std::hash<Uint32>{}(key.layerCount) + 0x9e3779b9u + (hash << 6) + (hash >> 2);
                hash ^= std::hash<Uint32>{}(static_cast<Uint32>(key.viewType)) +
                        0x9e3779b9u + (hash << 6) + (hash >> 2);
+                hash ^= std::hash<Uint32>{}(static_cast<Uint32>(key.viewFormat)) +
+                        0x9e3779b9u + (hash << 6) + (hash >> 2);
                return hash;
            }
        };
@@ -481,6 +487,9 @@ private:
    std::unordered_map<TextureIdentity, TextureResource, TextureIdentityHash> m_textureResources;
    // Textures that have been bound to a GL image unit (see MarkStorageImageTexture).
    std::unordered_set<TextureIdentity, TextureIdentityHash> m_storageImageTextures;
+    // Supported multisample counts per format, so repeat texture syncs do not
+    // re-query vkGetPhysicalDeviceImageFormatProperties.
+    std::unordered_map<VkFormat, VkSampleCountFlags> m_multisampleCountsByFormat;
    Vector<Vector<TextureResource>> m_deferredReleases;
    Vector<Vector<VkImageView>> m_deferredViewReleases;
    // Texture uploads are submitted out-of-band but NOT waited on (waiting
@@ -76,6 +76,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
        GLenum indexType = GL_UNSIGNED_SHORT;
        SizeT indexByteOffset = 0;
        SizeT indexByteSize = 0;
+        // Interpret indexByteOffset as a raw client pointer even when an element
+        // array buffer is bound (backend-synthesized index lists, e.g. the
+        // GL_LINE_LOOP -> LINE_STRIP rewrite).
+        Bool forceClientMemory = false;
    };

    struct DrawIndexedCmd {
@@ -196,6 +200,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                              GLsizei srcWidth, GLsizei srcHeight, GLsizei srcDepth);
        void GenerateMipmap(GLenum target);
        void ReadPixels(GLint x, GLint y, GLsizei width, GLsizei height, GLenum format, GLenum type, void* pixels);
+        // GL_DEPTH_COMPONENT / GL_DEPTH_STENCIL / GL_STENCIL_INDEX readback from the
+        // read framebuffer's depth/stencil attachment (per-aspect buffer copies with
+        // CPU repacking into the requested client layout).
+        void ReadDepthStencilPixels(MG_State::GLState::FramebufferObject& readFbo, GLint x, GLint y, GLsizei width,
+                                    GLsizei height, GLenum format, GLenum type, void* pixels);
+        // Copy-and-repack core shared by depth-stencil ReadPixels and GetTexImage;
+        // expects command recording to be active and any render pass already ended.
+        void ReadDepthStencilImageToClient(VkImage image, VkFormat vkFormat, VkImageLayout* trackedLayout,
+                                           VkImageAspectFlags imageAspect, Uint32 mipLevel, Uint32 baseArrayLayer,
+                                           GLint x, GLint y, GLsizei width, GLsizei height, GLenum format, GLenum type,
+                                           void* pixels);
+        // Same-extent depth blit between images of different depth formats: host
+        // round-trip with a per-texel re-encode (see BlitNamedFramebuffer).
+        Bool BlitDepthAcrossFormats(FrameContext::FrameData& frame, VkImage srcImage, VkFormat srcFormat,
+                                    VkImageLayout* srcTrackedLayout, Uint32 srcMipLevel, Uint32 srcBaseArrayLayer,
+                                    VkImage dstImage, VkFormat dstFormat, VkImageLayout* dstTrackedLayout,
+                                    Uint32 dstMipLevel, Uint32 dstBaseArrayLayer, GLint srcX, GLint srcY, GLint dstX,
+                                    GLint dstY, GLint width, GLint height, VkImageLayout srcRestoreLayout,
+                                    VkImageLayout dstRestoreLayout, Bool stencilAspect);
        static SizeT GetReadbackTexelSize(VkFormat sourceFormat);
        static Bool ConvertReadbackPixels(const Uint8* sourcePixels, VkFormat sourceFormat,
                                          GLsizei width, GLsizei height, GLenum destinationFormat,
@@ -281,6 +304,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                      const VkTimerQueryManager::TimestampRecord& end) const;
        Uint64 GetTimerQueryTimestampNs(const VkTimerQueryManager::TimestampRecord& record) const;

+        // GL_SAMPLES_PASSED occlusion queries: every app draw between Start and Stop is
+        // wrapped in a Vulkan occlusion query slot; the result is the slot sum. Requires
+        // hostQueryReset for slot recycling - Start fails (frontend keeps the query
+        // unsupported) when the device lacks it.
+        Bool StartOcclusionQueryCapture();
+        void StopOcclusionQueryCapture(Vector<Uint32>& outSlots);
+        // Flushes pending commands, waits, sums the slots, and recycles them.
+        Bool ResolveOcclusionQueryResult(const Vector<Uint32>& slots, Uint64& outSamples);
+
        void RequestSwapchainResize(Uint32 width, Uint32 height);
        // Re-query the surface and report whether the live swapchain no longer matches it
        // (size or orientation). This - not a VK_SUBOPTIMAL_KHR result - is what decides a
@@ -451,6 +483,59 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                                 Uint32 stride);
        static inline PFNDrawIndexedIndirectCountFunc s_vkCmdDrawIndexedIndirectCount = nullptr;

+        // VK_EXT_transform_feedback (GL transform feedback capture)
+        Bool m_transformFeedbackFeatureEnabled = false;
+        static inline PFN_vkCmdBindTransformFeedbackBuffersEXT s_vkCmdBindTransformFeedbackBuffersEXT = nullptr;
+        static inline PFN_vkCmdBeginTransformFeedbackEXT s_vkCmdBeginTransformFeedbackEXT = nullptr;
+        static inline PFN_vkCmdEndTransformFeedbackEXT s_vkCmdEndTransformFeedbackEXT = nullptr;
+        // Counter buffers (one 4-byte slot per capture binding) let consecutive
+        // draws within one glBeginTransformFeedback append GL-style.
+        VkBufferObject m_xfbCounterBuffer;
+        // Non-zero while inside a GL Begin/End with at least one captured draw
+        // recorded; selects counter-buffer resume on the next captured draw.
+        Bool m_xfbCountersValid = false;
+        Uint64 m_xfbLastSeenGeneration = 0;
+        // Wraps a recorded draw with BeginTransformFeedbackEXT/EndTransformFeedbackEXT
+        // when GL transform feedback is active; binds capture buffers on demand.
+        Bool BeginXfbCaptureForDraw(FrameContext::FrameData& frame);
+        void EndXfbCaptureForDraw(FrameContext::FrameData& frame, Bool began);
+        // Wrap one app draw in an occlusion-query slot while a GL_SAMPLES_PASSED
+        // query is active. Returns whether a slot was begun (End must mirror it).
+        Bool BeginOcclusionForDraw(VkCommandBuffer commandBuffer);
+        void EndOcclusionForDraw(VkCommandBuffer commandBuffer, Bool began);
+        Bool m_occlusionQueryPreciseEnabled = false;
+        Bool m_hostQueryResetEnabled = false;
+        PFN_vkResetQueryPool s_vkResetQueryPool = nullptr;
+        VkQueryPool m_occlusionQueryPool = VK_NULL_HANDLE;
+        static constexpr Uint32 kOcclusionQuerySlots = 8192;
+        Uint32 m_occlusionSlotCursor = 0;
+        Bool m_occlusionCaptureActive = false;
+        Vector<Uint32> m_occlusionActiveSlots;
+        // Transform feedback primitive queries: one pool slot per captured draw yields
+        // the (written, needed) pair; GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN sums the
+        // first, GL_PRIMITIVES_GENERATED the second - exact with geometry shaders,
+        // unlike the CPU fallback accounting.
+        Bool m_xfbQueriesSupported = false;
+        PFN_vkCmdBeginQueryIndexedEXT s_vkCmdBeginQueryIndexedEXT = nullptr;
+        PFN_vkCmdEndQueryIndexedEXT s_vkCmdEndQueryIndexedEXT = nullptr;
+        VkQueryPool m_xfbQueryPool = VK_NULL_HANDLE;
+        static constexpr Uint32 kXfbQuerySlots = 8192;
+        Uint32 m_xfbQuerySlotCursor = 0;
+        Bool m_xfbQueryCaptureActive[2] = {false, false}; // [0]=written, [1]=generated
+        Vector<Uint32> m_xfbQueryActiveSlots[2];
+        Bool m_xfbQuerySlotOpen = false;
+        Uint32 m_xfbQueryOpenSlot = 0;
+
+    public:
+        // kind: 0 = PRIMITIVES_WRITTEN, 1 = PRIMITIVES_GENERATED.
+        Bool StartXfbQueryCapture(Uint32 kind);
+        void StopXfbQueryCapture(Uint32 kind, Vector<Uint32>& outSlots);
+        Bool ResolveXfbQueryResult(const Vector<Uint32>& slots, Bool wantGenerated, Uint64& outPrimitives);
+
+    private:
+        void BeginXfbQueryForDraw(VkCommandBuffer commandBuffer);
+        void EndXfbQueryForDraw(VkCommandBuffer commandBuffer);
+
        VkCommandPool m_commandPool = VK_NULL_HANDLE;

        VkBufferManager m_bufferManager;
@@ -52,19 +52,48 @@ namespace MobileGL::MG_Backend::DirectVulkan {
    }
 } // namespace MobileGL::MG_Backend::DirectVulkan

+namespace MobileGL::MG_Backend::DirectVulkan {
+    // GL renders into sRGB color attachments RAW while GL_FRAMEBUFFER_SRGB is disabled
+    // (the core-profile default); Vulkan sRGB attachments always encode on write. The
+    // attachment view (and render pass format) therefore drops to the UNORM twin
+    // whenever the capability is off. Sampled views keep the sRGB format (decode on
+    // sample is unconditional in GL).
+    inline VkFormat ResolveSrgbAttachmentWriteFormat(VkFormat format, bool framebufferSrgbEnabled) {
+        if (framebufferSrgbEnabled) return format;
+        switch (format) {
+        case VK_FORMAT_R8G8B8A8_SRGB:
+            return VK_FORMAT_R8G8B8A8_UNORM;
+        case VK_FORMAT_B8G8R8A8_SRGB:
+            return VK_FORMAT_B8G8R8A8_UNORM;
+        default:
+            return format;
+        }
+    }
+} // namespace MobileGL::MG_Backend::DirectVulkan
+
+// The context line (__VA_ARGS__ = its own format string + args) must be a SEPARATE log
+// call: appending its format to the base format while its arguments precede the base
+// arguments makes every conversion read the wrong slot (a %s pulling an int crashes).
 #define VK_VERIFY(expr, ...)                                                                                           \
    do {                                                                                                               \
        VkResult _vk_verify_result = (expr);                                                                           \
        if (_vk_verify_result != VK_SUCCESS) {                                                                         \
-            MGLOG_F("Vulkan error %s (%d) at %s:%d" __VA_OPT__(" - ") __VA_ARGS__,                                  \
+            __VA_OPT__(MGLOG_F(__VA_ARGS__);)                                                                          \
+            MGLOG_F("Vulkan error %s (%d) at %s:%d",                                                                   \
                    MobileGL::MG_Backend::DirectVulkan::VkResultToString(_vk_verify_result),                          \
                    _vk_verify_result, __FILE__, __LINE__);                                                            \
        }                                                                                                              \
-        MOBILEGL_ASSERT(_vk_verify_result == VK_SUCCESS, "Vulkan error %s (%d) at %s:%d" __VA_OPT__(" - ") __VA_ARGS__, MobileGL::MG_Backend::DirectVulkan::VkResultToString(_vk_verify_result), _vk_verify_result, __FILE__, __LINE__);  \
+        MOBILEGL_ASSERT(_vk_verify_result == VK_SUCCESS, "Vulkan error %s (%d) at %s:%d",                              \
+                        MobileGL::MG_Backend::DirectVulkan::VkResultToString(_vk_verify_result),                       \
+                        _vk_verify_result, __FILE__, __LINE__);                                                        \
    } while (0)

 #define XXHASH_VERIFY(expr, ...)                                                                                       \
    do {                                                                                                               \
        XXH_errorcode _xxh_verify_result = (expr);                                                                     \
-        MOBILEGL_ASSERT(_xxh_verify_result == XXH_OK, "XXHash error %d at %s:%d" __VA_OPT__(" - ") __VA_ARGS__, _xxh_verify_result, __FILE__, __LINE__);  \
+        if (_xxh_verify_result != XXH_OK) {                                                                            \
+            __VA_OPT__(MGLOG_F(__VA_ARGS__);)                                                                          \
+        }                                                                                                              \
+        MOBILEGL_ASSERT(_xxh_verify_result == XXH_OK, "XXHash error %d at %s:%d", _xxh_verify_result, __FILE__,        \
+                        __LINE__);                                                                                     \
    } while (0)
@@ -1351,6 +1351,14 @@ namespace MobileGL::MG_Impl::GLImpl {
        BufferTarget bufferTarget = MG_Util::ConvertGLEnumToBufferTarget(target);
        if (!BufferImpl::ValidateBufferBindingPointTarget(bufferTarget)) return;
        if (!BufferImpl::ValidateBufferBindingPointIndex(bufferTarget, pointIndex)) return;
+        if (bufferTarget == BufferTarget::TransformFeedback && MG_State::pGLContext->IsTransformFeedbackActive()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "Transform feedback buffer bindings cannot change while transform "
+                                             "feedback is active."));
+            return;
+        }
        MG_State::pGLContext->TouchBufferBindingPoint(bufferTarget, pointIndex);

        auto& point = MG_State::pGLContext->GetBufferBindingPoint(bufferTarget, pointIndex);
@@ -1384,6 +1392,14 @@ namespace MobileGL::MG_Impl::GLImpl {
        BufferTarget bufferTarget = MG_Util::ConvertGLEnumToBufferTarget(target);
        if (!BufferImpl::ValidateBufferBindingPointTarget(bufferTarget)) return;
        if (!BufferImpl::ValidateBufferBindingPointIndex(bufferTarget, index)) return;
+        if (bufferTarget == BufferTarget::TransformFeedback && MG_State::pGLContext->IsTransformFeedbackActive()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "Transform feedback buffer bindings cannot change while transform "
+                                             "feedback is active."));
+            return;
+        }
        MG_State::pGLContext->TouchBufferBindingPoint(bufferTarget, index);

        auto& point = MG_State::pGLContext->GetBufferBindingPoint(bufferTarget, index);
@@ -60,6 +60,11 @@ namespace MobileGL::MG_Impl::GLImpl::BufferImpl {
                MG_Backend::pActiveBackendObject->GetDynamicParameters().MaxShaderStorageBufferBindings;
            pointCount = std::min(pointCount, static_cast<SizeT>(std::max(backendCount, 0)));
        }
+        if (target == BufferTarget::TransformFeedback) {
+            // GL_MAX_TRANSFORM_FEEDBACK_SEPARATE_ATTRIBS bounds the indexed capture
+            // binding points in GL 3.3 (no ARB_transform_feedback3).
+            pointCount = std::min<SizeT>(pointCount, 4);
+        }

        if (index < pointCount) {
            return true;
@@ -48,6 +48,73 @@ namespace MobileGL::MG_Impl::GLImpl {
        return true;
    }

+    // Primitives a draw of `count` vertices in `mode` assembles (0 for
+    // incomplete primitives). Used for the CPU-side transform feedback
+    // primitive accounting.
+    static Uint64 CountPrimitivesForDraw(GLenum mode, GLsizei count) {
+        if (count <= 0) return 0;
+        switch (mode) {
+        case GL_POINTS: return static_cast<Uint64>(count);
+        case GL_LINES: return static_cast<Uint64>(count / 2);
+        case GL_LINE_STRIP: return count >= 2 ? static_cast<Uint64>(count - 1) : 0;
+        case GL_LINE_LOOP: return count >= 2 ? static_cast<Uint64>(count) : 0;
+        case GL_TRIANGLES: return static_cast<Uint64>(count / 3);
+        case GL_TRIANGLE_STRIP:
+        case GL_TRIANGLE_FAN: return count >= 3 ? static_cast<Uint64>(count - 2) : 0;
+        default: return 0;
+        }
+    }
+
+    // Accumulate the transform feedback primitive counter for a captured draw.
+    // Draws without a geometry stage write exactly the primitives they assemble,
+    // clamped by the capture buffers' remaining capacity (a full buffer stops
+    // recording whole primitives, which is what PRIMITIVES_WRITTEN reports).
+    // Geometry amplification is not modelled here.
+    static void AccountTransformFeedbackPrimitives(GLenum mode, GLsizei count) {
+        if (!MG_State::pGLContext->IsTransformFeedbackActive()) return;
+        Uint64 primitives = CountPrimitivesForDraw(mode, count);
+        if (primitives == 0) return;
+        MG_State::pGLContext->AddTransformFeedbackInputPrimitives(primitives);
+
+        Uint64 verticesPerPrimitive = 1;
+        switch (mode) {
+        case GL_LINES:
+        case GL_LINE_STRIP:
+        case GL_LINE_LOOP:
+            verticesPerPrimitive = 2;
+            break;
+        case GL_TRIANGLES:
+        case GL_TRIANGLE_STRIP:
+        case GL_TRIANGLE_FAN:
+            verticesPerPrimitive = 3;
+            break;
+        default:
+            break;
+        }
+
+        const auto& program = MG_State::pGLContext->GetTransformFeedbackProgram();
+        if (program != nullptr) {
+            // Capacity in captured vertices = the tightest bound buffer.
+            Uint64 capacityVertices = ~0ull;
+            for (SizeT i = 0; i < program->GetTransformFeedbackBufferCount(); ++i) {
+                const Uint32 stride = program->GetTransformFeedbackStride(static_cast<Uint32>(i));
+                if (stride == 0) continue;
+                const auto& point = MG_State::pGLContext->GetBufferBindingPoint(BufferTarget::TransformFeedback,
+                                                                                static_cast<Uint>(i));
+                const Range1D range = point.GetRange();
+                const Uint64 bytes = range.end > range.start ? static_cast<Uint64>(range.end - range.start) : 0;
+                capacityVertices = std::min<Uint64>(capacityVertices, bytes / stride);
+            }
+            if (capacityVertices != ~0ull) {
+                const Uint64 usedVertices = MG_State::pGLContext->GetTransformFeedbackCapturedVertices();
+                const Uint64 remainingVertices = capacityVertices > usedVertices ? capacityVertices - usedVertices : 0;
+                primitives = std::min<Uint64>(primitives, remainingVertices / verticesPerPrimitive);
+            }
+        }
+        MG_State::pGLContext->AddTransformFeedbackPrimitives(primitives);
+        MG_State::pGLContext->AddTransformFeedbackCapturedVertices(primitives * verticesPerPrimitive);
+    }
+
    static Bool ValidatePrimitiveModeForBackend(const char* functionName, GLenum mode) {
        const auto& activeBackendObject = MG_Backend::pActiveBackendObject;
        if (!activeBackendObject) {
@@ -57,15 +124,6 @@ namespace MobileGL::MG_Impl::GLImpl {
            return false;
        }

-        if (activeBackendObject->GetBackendType() == BackendType::DirectVulkan && mode == GL_LINE_LOOP) {
-            MG_State::pGLContext->RecordError(
-                ErrorCode::InvalidOperation,
-                MakeUnique<GenericErrorInfo>(
-                    "MG_Impl/GLImpl", functionName,
-                    "Primitive mode GL_LINE_LOOP is not supported by the DirectVulkan backend."));
-            return false;
-        }
-
        const auto& vao = MG_State::pGLContext->GetBoundVertexArray();
        if (vao && vao->GetExternalIndex() == 0 && !MG_State::IsRelaxedSemanticsActive()) {
            MG_State::pGLContext->RecordError(
@@ -75,6 +133,38 @@ namespace MobileGL::MG_Impl::GLImpl {
            return false;
        }

+        // While transform feedback is active the draw's primitive type must match
+        // the feedback primitive mode (GL 3.3 core 13.2.2). With a geometry shader
+        // the constraint moves to the shader's output primitive type instead, so
+        // the draw mode itself is unconstrained here.
+        if (MG_State::pGLContext->IsTransformFeedbackActive() &&
+            !(MG_State::pGLContext->GetTransformFeedbackProgram() &&
+              MG_State::pGLContext->GetTransformFeedbackProgram()->GetShaderIndexByStage(ShaderStage::Geometry) >= 0)) {
+            const GLenum feedbackMode = MG_State::pGLContext->GetTransformFeedbackPrimitiveMode();
+            Bool compatible = false;
+            switch (feedbackMode) {
+            case GL_POINTS:
+                compatible = mode == GL_POINTS;
+                break;
+            case GL_LINES:
+                compatible = mode == GL_LINES || mode == GL_LINE_STRIP || mode == GL_LINE_LOOP;
+                break;
+            case GL_TRIANGLES:
+                compatible = mode == GL_TRIANGLES || mode == GL_TRIANGLE_STRIP || mode == GL_TRIANGLE_FAN;
+                break;
+            default:
+                break;
+            }
+            if (!compatible) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidOperation,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", functionName,
+                        "Primitive mode is incompatible with the active transform feedback primitive mode."));
+                return false;
+            }
+        }
+
        return true;
    }

@@ -402,12 +492,14 @@ namespace MobileGL::MG_Impl::GLImpl {
    void DrawElementsBaseVertex(GLenum mode, GLsizei count, GLenum type, const void* indices, GLint basevertex) {
        if (!ValidateCurrentProgramForExecution(__func__)) return;
        if (!ValidatePrimitiveModeForBackend(__func__, mode)) return;
+        AccountTransformFeedbackPrimitives(mode, count);
        DrawElementsBaseVertex_Backend(mode, count, type, indices, basevertex);
    }

    void DrawArrays(GLenum mode, GLint first, GLsizei count) {
        if (!ValidateCurrentProgramForExecution(__func__)) return;
        if (!ValidatePrimitiveModeForBackend(__func__, mode)) return;
+        AccountTransformFeedbackPrimitives(mode, count);
        DrawArrays_Backend(mode, first, count);
    }

@@ -444,7 +536,133 @@ namespace MobileGL::MG_Impl::GLImpl {
    void DrawElements(GLenum mode, GLsizei count, GLenum type, const void* indices) {
        if (!ValidateCurrentProgramForExecution(__func__)) return;
        if (!ValidatePrimitiveModeForBackend(__func__, mode)) return;
+        AccountTransformFeedbackPrimitives(mode, count);
        DrawElements_Backend(mode, count, type, indices);
    }

+    void BeginTransformFeedback(GLenum primitiveMode) {
+        if (primitiveMode != GL_POINTS && primitiveMode != GL_LINES && primitiveMode != GL_TRIANGLES) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidEnum,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "primitiveMode must be GL_POINTS, GL_LINES or GL_TRIANGLES."));
+            return;
+        }
+        if (MG_State::pGLContext->IsTransformFeedbackActive()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "Transform feedback is already active."));
+            return;
+        }
+        const auto& program = MG_State::pGLContext->GetCurrentProgram();
+        if (!program || !program->GetLinkStatus() || program->GetTransformFeedbackVaryingCount() == 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", __func__,
+                    "No program with transform feedback varyings is active."));
+            return;
+        }
+        // Every capture buffer slot the program's mode uses must have a buffer bound.
+        const SizeT usedBufferCount = program->GetTransformFeedbackBufferCount();
+        for (SizeT i = 0; i < usedBufferCount; ++i) {
+            const auto& point = MG_State::pGLContext->GetBufferBindingPoint(BufferTarget::TransformFeedback,
+                                                                            static_cast<Uint>(i));
+            if (point.GetBoundObject() == nullptr) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidOperation,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", __func__,
+                        "Transform feedback buffer binding point " + std::to_string(i) + " has no buffer bound."));
+                return;
+            }
+        }
+        MG_State::pGLContext->BeginTransformFeedback(primitiveMode, program);
+    }
+
+    // Vulkan transform feedback captures triangle strips in plain (i, i+1, i+2)
+    // vertex order, but GL decomposes odd strip triangles as (i+1, i, i+2)
+    // (GL 4.6 table 10.1). With the geometry stage's statically-known strip
+    // lengths the captured records are reordered in place: swap the first two
+    // vertex records of every odd triangle within each emitted strip.
+    static void FixupGsStripCaptureOrder(const SharedPtr<MG_State::GLState::ProgramObject>& program,
+                                         Uint64 inputPrimitives) {
+        if (program == nullptr || !program->HasGsTriangleStripCaptureFixup() || inputPrimitives == 0) {
+            return;
+        }
+        const auto& stripTriangles = program->GetGsStripTriangles();
+
+        // Global triangle indices whose leading vertex pair must swap.
+        Vector<Uint64> swapTriangles;
+        Uint64 triangleBase = 0;
+        for (Uint64 input = 0; input < inputPrimitives; ++input) {
+            for (const Uint32 stripLength : stripTriangles) {
+                for (Uint32 t = 1; t < stripLength; t += 2) {
+                    swapTriangles.push_back(triangleBase + t);
+                }
+                triangleBase += stripLength;
+            }
+        }
+        if (swapTriangles.empty()) {
+            return;
+        }
+
+        for (SizeT bufferIndex = 0; bufferIndex < program->GetTransformFeedbackBufferCount(); ++bufferIndex) {
+            const Uint32 stride = program->GetTransformFeedbackStride(static_cast<Uint32>(bufferIndex));
+            if (stride == 0) continue;
+            const auto& bindingPoint =
+                MG_State::pGLContext->GetBufferBindingPoint(BufferTarget::TransformFeedback,
+                                                            static_cast<Uint>(bufferIndex));
+            const auto& buffer = bindingPoint.GetBoundObject();
+            if (buffer == nullptr) continue;
+            const Range1D range = bindingPoint.GetRange();
+            const Uint8* mapped = buffer->MappedData();
+            if (mapped == nullptr) continue;
+            // The geometry stage amplifies, so the CPU vertex counter does not bound
+            // the capture; the binding range's whole-triangle capacity does.
+            const Uint64 rangeBytes = range.end > range.start ? static_cast<Uint64>(range.end - range.start) : 0;
+            const Uint64 capturedTriangles = std::min<Uint64>(triangleBase, (rangeBytes / stride) / 3);
+
+            // Observed Vulkan capture order for odd strip triangles is (i, i+2, i+1)
+            // (winding preserved by swapping the trailing pair); GL wants
+            // (i+1, i, i+2), which is one rotation away: (a,b,c) -> (c,a,b).
+            Vector<Uint8> scratch(stride);
+            for (const Uint64 triangle : swapTriangles) {
+                if (triangle >= capturedTriangles) break;
+                const SizeT v0Offset = static_cast<SizeT>(range.start) + static_cast<SizeT>(triangle * 3) * stride;
+                const SizeT v1Offset = v0Offset + stride;
+                const SizeT v2Offset = v1Offset + stride;
+                Memcpy(scratch.data(), mapped + v2Offset, stride);
+                buffer->WritebackFromBackend({const_cast<Uint8*>(mapped) + v1Offset, stride}, v2Offset);
+                buffer->WritebackFromBackend({const_cast<Uint8*>(mapped) + v0Offset, stride}, v1Offset);
+                buffer->WritebackFromBackend({scratch.data(), stride}, v0Offset);
+            }
+        }
+    }
+
+    void EndTransformFeedback(void) {
+        if (!MG_State::pGLContext->IsTransformFeedbackActive()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "Transform feedback is not active."));
+            return;
+        }
+        const auto capturedProgram = MG_State::pGLContext->GetTransformFeedbackProgram();
+        const Uint64 inputPrimitives = MG_State::pGLContext->GetTransformFeedbackInputPrimitives();
+        MG_State::pGLContext->EndTransformFeedback();
+        // Captured results must be visible to MapBuffer/GetBufferSubData after
+        // End; the capture targets are host-coherent GPU memory, so completing
+        // the GPU work is all that is required.
+        auto& backendGL = MG_Backend::gBackendFunctionsTable.GL;
+        if (backendGL.FenceSync && backendGL.ClientWaitSync) {
+            if (auto sync = backendGL.FenceSync()) {
+                backendGL.ClientWaitSync(sync, GL_SYNC_FLUSH_COMMANDS_BIT, ~0ull);
+                if (backendGL.DeleteSync) {
+                    backendGL.DeleteSync(sync);
+                }
+            }
+        }
+        FixupGsStripCaptureOrder(capturedProgram, inputPrimitives);
+    }
+
 } // namespace MobileGL::MG_Impl::GLImpl
@@ -11,6 +11,8 @@

 namespace MobileGL::MG_Impl::GLImpl {
    /* @INSERTION_POINT:FUNCTION_DECLARATION@ */
+    void BeginTransformFeedback(GLenum primitiveMode);
+    void EndTransformFeedback(void);
    void DispatchCompute(GLuint numGroupsX, GLuint numGroupsY, GLuint numGroupsZ);
    void DispatchComputeIndirect(GLintptr indirect);
    void MemoryBarrier(GLbitfield barriers);
@@ -236,12 +236,12 @@ DECLARE_GL_FUNCTION_HEAD(void, DeleteVertexArrays, GLsizei n, const GLuint* arra
 DECLARE_GL_FUNCTION_HEAD(void, GenVertexArrays, GLsizei n, GLuint* arrays) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GenVertexArrays, n, arrays)
 DECLARE_GL_FUNCTION_HEAD(GLboolean, IsVertexArray, GLuint array) DECLARE_GL_FUNCTION_END(GLboolean, IsVertexArray, array)
 DECLARE_GL_FUNCTION_HEAD(void, GetIntegeri_v, GLenum target, GLuint index, GLint* data) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GetIntegeri_v, target, index, data)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, BeginTransformFeedback, GLenum primitiveMode) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, BeginTransformFeedback, primitiveMode)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, EndTransformFeedback) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, EndTransformFeedback)
+DECLARE_GL_FUNCTION_HEAD(void, BeginTransformFeedback, GLenum primitiveMode) DECLARE_GL_FUNCTION_END_NO_RETURN(void, BeginTransformFeedback, primitiveMode)
+DECLARE_GL_FUNCTION_HEAD(void, EndTransformFeedback) DECLARE_GL_FUNCTION_END_NO_RETURN(void, EndTransformFeedback)
 DECLARE_GL_FUNCTION_HEAD(void, BindBufferRange, GLenum target, GLuint index, GLuint buffer, GLintptr offset, GLsizeiptr size) DECLARE_GL_FUNCTION_END_NO_RETURN(void, BindBufferRange, target, index, buffer, offset, size)
 DECLARE_GL_FUNCTION_HEAD(void, BindBufferBase, GLenum target, GLuint index, GLuint buffer) DECLARE_GL_FUNCTION_END_NO_RETURN(void, BindBufferBase, target, index, buffer)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, TransformFeedbackVaryings, GLuint program, GLsizei count, const GLchar* const* varyings, GLenum bufferMode) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, TransformFeedbackVaryings, program, count, varyings, bufferMode)
-DECLARE_GL_FUNCTION_STUB_HEAD(void, GetTransformFeedbackVarying, GLuint program, GLuint index, GLsizei bufSize, GLsizei* length, GLsizei* size, GLenum* type, GLchar* name) DECLARE_GL_FUNCTION_STUB_END_NO_RETURN(void, GetTransformFeedbackVarying, program, index, bufSize, length, size, type, name)
+DECLARE_GL_FUNCTION_HEAD(void, TransformFeedbackVaryings, GLuint program, GLsizei count, const GLchar* const* varyings, GLenum bufferMode) DECLARE_GL_FUNCTION_END_NO_RETURN(void, TransformFeedbackVaryings, program, count, varyings, bufferMode)
+DECLARE_GL_FUNCTION_HEAD(void, GetTransformFeedbackVarying, GLuint program, GLuint index, GLsizei bufSize, GLsizei* length, GLsizei* size, GLenum* type, GLchar* name) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GetTransformFeedbackVarying, program, index, bufSize, length, size, type, name)
 DECLARE_GL_FUNCTION_HEAD(void, VertexAttribIPointer, GLuint index, GLint size, GLenum type, GLsizei stride, const void* pointer) DECLARE_GL_FUNCTION_END_NO_RETURN(void, VertexAttribIPointer, index, size, type, stride, pointer)
 DECLARE_GL_FUNCTION_HEAD(void, GetVertexAttribIiv, GLuint index, GLenum pname, GLint* params) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GetVertexAttribIiv, index, pname, params)
 DECLARE_GL_FUNCTION_HEAD(void, GetVertexAttribIuiv, GLuint index, GLenum pname, GLuint* params) DECLARE_GL_FUNCTION_END_NO_RETURN(void, GetVertexAttribIuiv, index, pname, params)
@@ -45,10 +45,32 @@ namespace MobileGL::MG_Impl::GLImpl {
                   depthAttachment.GetTextureLevel() != stencilAttachment.GetTextureLevel();
        }

+        // Mirrors the renderer-side gate: distinct depth/stencil renderbuffers (or a
+        // renderbuffer paired with a texture) cannot form one Vulkan depth-stencil
+        // attachment, and GL permits reporting such framebuffers as UNSUPPORTED.
+        Bool HasDistinctCompleteDepthStencilRenderbufferAttachments(
+            const MG_State::GLState::FramebufferObject& framebufferObject) {
+            if (framebufferObject.GetExternalIndex() == 0) {
+                return false;
+            }
+
+            const auto& depthAttachment = framebufferObject.GetAttachment(FramebufferAttachmentType::Depth);
+            const auto& stencilAttachment = framebufferObject.GetAttachment(FramebufferAttachmentType::Stencil);
+            if (!depthAttachment.IsComplete() || !stencilAttachment.IsComplete()) {
+                return false;
+            }
+            if (depthAttachment.IsRenderbuffer() && stencilAttachment.IsRenderbuffer()) {
+                return depthAttachment.GetRenderbuffer().get() != stencilAttachment.GetRenderbuffer().get();
+            }
+            return (depthAttachment.IsRenderbuffer() || stencilAttachment.IsRenderbuffer()) &&
+                   (depthAttachment.IsTexture() || stencilAttachment.IsTexture());
+        }
+
        Bool IsUnsupportedFramebufferForDirectVulkan(
            const MG_State::GLState::FramebufferObject& framebufferObject) {
            // TODO: Keep this in sync with DirectVulkan renderbuffer support as color renderbuffer rendering lands.
-            return HasDistinctCompleteDepthStencilTextureAttachments(framebufferObject);
+            return HasDistinctCompleteDepthStencilTextureAttachments(framebufferObject) ||
+                   HasDistinctCompleteDepthStencilRenderbufferAttachments(framebufferObject);
        }

        Bool HasDefinedAttachment(const MG_State::GLState::FramebufferObject& framebufferObject) {
@@ -147,6 +169,148 @@ namespace MobileGL::MG_Impl::GLImpl {
                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", functionName, detail));
        }

+        GLint ClassifyAttachmentComponentType(TextureInternalFormat internalFormat,
+                                              FramebufferAttachmentType attachmentType) {
+            // A stencil value is an unsigned integer index regardless of the depth half
+            // of a packed format.
+            if (attachmentType == FramebufferAttachmentType::Stencil) return GL_UNSIGNED_INT;
+            switch (internalFormat) {
+            case TextureInternalFormat::R16F:
+            case TextureInternalFormat::RG16F:
+            case TextureInternalFormat::RGB16F:
+            case TextureInternalFormat::RGBA16F:
+            case TextureInternalFormat::R32F:
+            case TextureInternalFormat::RG32F:
+            case TextureInternalFormat::RGB32F:
+            case TextureInternalFormat::RGBA32F:
+            case TextureInternalFormat::R11FG11FB10F:
+            case TextureInternalFormat::RGB9E5:
+            case TextureInternalFormat::DepthComponent32F:
+            case TextureInternalFormat::Depth32FStencil8:
+                return GL_FLOAT;
+            case TextureInternalFormat::R8I:
+            case TextureInternalFormat::R16I:
+            case TextureInternalFormat::R32I:
+            case TextureInternalFormat::RG8I:
+            case TextureInternalFormat::RG16I:
+            case TextureInternalFormat::RG32I:
+            case TextureInternalFormat::RGB8I:
+            case TextureInternalFormat::RGB16I:
+            case TextureInternalFormat::RGB32I:
+            case TextureInternalFormat::RGBA8I:
+            case TextureInternalFormat::RGBA16I:
+            case TextureInternalFormat::RGBA32I:
+                return GL_INT;
+            case TextureInternalFormat::R8UI:
+            case TextureInternalFormat::R16UI:
+            case TextureInternalFormat::R32UI:
+            case TextureInternalFormat::RG8UI:
+            case TextureInternalFormat::RG16UI:
+            case TextureInternalFormat::RG32UI:
+            case TextureInternalFormat::RGB8UI:
+            case TextureInternalFormat::RGB16UI:
+            case TextureInternalFormat::RGB32UI:
+            case TextureInternalFormat::RGBA8UI:
+            case TextureInternalFormat::RGBA16UI:
+            case TextureInternalFormat::RGBA32UI:
+            case TextureInternalFormat::RGB10A2UI:
+                return GL_UNSIGNED_INT;
+            case TextureInternalFormat::R8Snorm:
+            case TextureInternalFormat::R16Snorm:
+            case TextureInternalFormat::RG8Snorm:
+            case TextureInternalFormat::RG16Snorm:
+            case TextureInternalFormat::RGB8Snorm:
+            case TextureInternalFormat::RGB16Snorm:
+            case TextureInternalFormat::RGBA8Snorm:
+            case TextureInternalFormat::RGBA16Snorm:
+                return GL_SIGNED_NORMALIZED;
+            case TextureInternalFormat::Unknown:
+                return GL_NONE;
+            default:
+                return GL_UNSIGNED_NORMALIZED;
+            }
+        }
+
+        // Handles the format-derived pnames shared by GetFramebufferAttachmentParameteriv
+        // and its DSA variant. Returns true when pname was one of them.
+        Bool TryAnswerAttachmentFormatQuery(const MG_State::GLState::FramebufferAttachmentObject* attachmentObject,
+                                            FramebufferAttachmentType attachmentType, Bool depthStencilAlias,
+                                            GLenum pname, GLint* params, const char* caller) {
+            switch (pname) {
+            case GL_FRAMEBUFFER_ATTACHMENT_RED_SIZE:
+            case GL_FRAMEBUFFER_ATTACHMENT_GREEN_SIZE:
+            case GL_FRAMEBUFFER_ATTACHMENT_BLUE_SIZE:
+            case GL_FRAMEBUFFER_ATTACHMENT_ALPHA_SIZE:
+            case GL_FRAMEBUFFER_ATTACHMENT_DEPTH_SIZE:
+            case GL_FRAMEBUFFER_ATTACHMENT_STENCIL_SIZE:
+            case GL_FRAMEBUFFER_ATTACHMENT_COMPONENT_TYPE:
+            case GL_FRAMEBUFFER_ATTACHMENT_COLOR_ENCODING:
+                break;
+            default:
+                return false;
+            }
+
+            if (pname == GL_FRAMEBUFFER_ATTACHMENT_COMPONENT_TYPE && depthStencilAlias) {
+                // The depth and stencil components have different types, so the combined
+                // attachment name has no single answer.
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidOperation,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", caller,
+                        "GL_FRAMEBUFFER_ATTACHMENT_COMPONENT_TYPE cannot be queried on "
+                        "GL_DEPTH_STENCIL_ATTACHMENT."));
+                return true;
+            }
+
+            if (attachmentObject == nullptr || attachmentObject->IsEmpty() || !attachmentObject->IsValid()) {
+                // With OBJECT_TYPE == GL_NONE only OBJECT_TYPE and OBJECT_NAME may be queried.
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidOperation,
+                    MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                                 "No image is attached to the queried attachment point."));
+                return true;
+            }
+
+            TextureInternalFormat internalFormat = TextureInternalFormat::Unknown;
+            if (attachmentObject->IsTexture() && attachmentObject->GetTexture()) {
+                internalFormat = attachmentObject->GetTexture()->GetFormat();
+            } else if (attachmentObject->IsRenderbuffer() && attachmentObject->GetRenderbuffer()) {
+                internalFormat = attachmentObject->GetRenderbuffer()->GetInternalFormat();
+            }
+            const auto sizes = MG_Util::GetComponentSizesForInternalFormat(internalFormat);
+
+            switch (pname) {
+            case GL_FRAMEBUFFER_ATTACHMENT_RED_SIZE:
+                *params = sizes.Red;
+                break;
+            case GL_FRAMEBUFFER_ATTACHMENT_GREEN_SIZE:
+                *params = sizes.Green;
+                break;
+            case GL_FRAMEBUFFER_ATTACHMENT_BLUE_SIZE:
+                *params = sizes.Blue;
+                break;
+            case GL_FRAMEBUFFER_ATTACHMENT_ALPHA_SIZE:
+                *params = sizes.Alpha;
+                break;
+            case GL_FRAMEBUFFER_ATTACHMENT_DEPTH_SIZE:
+                *params = sizes.Depth;
+                break;
+            case GL_FRAMEBUFFER_ATTACHMENT_STENCIL_SIZE:
+                *params = sizes.Stencil;
+                break;
+            case GL_FRAMEBUFFER_ATTACHMENT_COLOR_ENCODING:
+                *params = (internalFormat == TextureInternalFormat::SRGB8 ||
+                           internalFormat == TextureInternalFormat::SRGB8Alpha8)
+                    ? GL_SRGB
+                    : GL_LINEAR;
+                break;
+            case GL_FRAMEBUFFER_ATTACHMENT_COMPONENT_TYPE:
+                *params = ClassifyAttachmentComponentType(internalFormat, attachmentType);
+                break;
+            }
+            return true;
+        }
+
        Bool ResolveRepresentableFramebufferTextureUploadTarget(const MG_State::GLState::ITextureObject& textureObject,
                                                                TextureUploadTarget& outUploadTarget,
                                                                Bool& outLayered) {
@@ -412,6 +576,27 @@ namespace MobileGL::MG_Impl::GLImpl {
        FramebufferTarget framebufferTarget = MG_Util::ConvertGLEnumToFramebufferTarget(target);
        if (!FramebufferImpl::ValidateFramebufferTarget(framebufferTarget)) return;

+        // Default-framebuffer attachment names (GL_DEPTH, GL_STENCIL, GL_FRONT/GL_BACK
+        // variants) alias onto the equivalent attachment points.
+        switch (attachment) {
+        case GL_DEPTH:
+            attachment = GL_DEPTH_ATTACHMENT;
+            break;
+        case GL_STENCIL:
+            attachment = GL_STENCIL_ATTACHMENT;
+            break;
+        case GL_FRONT:
+        case GL_FRONT_LEFT:
+        case GL_FRONT_RIGHT:
+        case GL_BACK:
+        case GL_BACK_LEFT:
+        case GL_BACK_RIGHT:
+            attachment = GL_COLOR_ATTACHMENT0;
+            break;
+        default:
+            break;
+        }
+
        const Bool depthStencilAlias = attachment == GL_DEPTH_STENCIL_ATTACHMENT;
        FramebufferAttachmentType attachmentType = depthStencilAlias
            ? FramebufferAttachmentType::Depth
@@ -428,20 +613,40 @@ namespace MobileGL::MG_Impl::GLImpl {
            return;
        }

+        Bool depthStencilMismatch = false;
        const auto* attachmentObject = [&]() -> const MG_State::GLState::FramebufferAttachmentObject* {
            if (!depthStencilAlias) {
                return &framebufferObject->GetAttachment(attachmentType);
            }

            const auto& depthAttachment = framebufferObject->GetAttachment(FramebufferAttachmentType::Depth);
-            if (depthAttachment.IsValid() && !depthAttachment.IsEmpty()) return &depthAttachment;
-
            const auto& stencilAttachment = framebufferObject->GetAttachment(FramebufferAttachmentType::Stencil);
-            if (stencilAttachment.IsValid() && !stencilAttachment.IsEmpty()) return &stencilAttachment;
-
+            const Bool depthLive = depthAttachment.IsValid() && !depthAttachment.IsEmpty();
+            const Bool stencilLive = stencilAttachment.IsValid() && !stencilAttachment.IsEmpty();
+            if (depthLive && stencilLive) {
+                const Bool sameObject = depthAttachment.IsTexture() == stencilAttachment.IsTexture() &&
+                    (!depthAttachment.IsTexture() || depthAttachment.GetTexture() == stencilAttachment.GetTexture()) &&
+                    (!depthAttachment.IsRenderbuffer() ||
+                     depthAttachment.GetRenderbuffer() == stencilAttachment.GetRenderbuffer());
+                depthStencilMismatch = !sameObject;
+            } else {
+                // GL_DEPTH_STENCIL_ATTACHMENT means "both halves"; a lone half does not answer it.
+                depthStencilMismatch = depthLive != stencilLive;
+            }
+            if (depthLive) return &depthAttachment;
+            if (stencilLive) return &stencilAttachment;
            return nullptr;
        }();

+        if (depthStencilMismatch) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", "GetFramebufferAttachmentParameteriv_State",
+                                             "GL_DEPTH_STENCIL_ATTACHMENT query with different depth and stencil "
+                                             "attachment images."));
+            return;
+        }
+
        switch (pname) {
        case GL_FRAMEBUFFER_ATTACHMENT_OBJECT_TYPE:
            if (attachmentObject == nullptr || attachmentObject->IsEmpty() || !attachmentObject->IsValid()) {
@@ -505,6 +710,10 @@ namespace MobileGL::MG_Impl::GLImpl {
                : GL_FALSE;
            break;
        default:
+            if (TryAnswerAttachmentFormatQuery(attachmentObject, attachmentType, depthStencilAlias, pname, params,
+                                               "GetFramebufferAttachmentParameteriv_State")) {
+                return;
+            }
            MG_State::pGLContext->RecordError(
                ErrorCode::InvalidEnum,
                MakeUnique<GenericErrorInfo>(
@@ -1468,20 +1677,40 @@ namespace MobileGL::MG_Impl::GLImpl {
            : MG_Util::ConvertGLEnumToFramebufferAttachmentType(attachment);
        if (!FramebufferImpl::ValidateFramebufferAttachmentType(attachmentType)) return;

+        Bool depthStencilMismatch = false;
        const auto* attachmentObject = [&]() -> const MG_State::GLState::FramebufferAttachmentObject* {
            if (!depthStencilAlias) {
                return &framebufferObject->GetAttachment(attachmentType);
            }

            const auto& depthAttachment = framebufferObject->GetAttachment(FramebufferAttachmentType::Depth);
-            if (depthAttachment.IsValid() && !depthAttachment.IsEmpty()) return &depthAttachment;
-
            const auto& stencilAttachment = framebufferObject->GetAttachment(FramebufferAttachmentType::Stencil);
-            if (stencilAttachment.IsValid() && !stencilAttachment.IsEmpty()) return &stencilAttachment;
-
+            const Bool depthLive = depthAttachment.IsValid() && !depthAttachment.IsEmpty();
+            const Bool stencilLive = stencilAttachment.IsValid() && !stencilAttachment.IsEmpty();
+            if (depthLive && stencilLive) {
+                const Bool sameObject = depthAttachment.IsTexture() == stencilAttachment.IsTexture() &&
+                    (!depthAttachment.IsTexture() || depthAttachment.GetTexture() == stencilAttachment.GetTexture()) &&
+                    (!depthAttachment.IsRenderbuffer() ||
+                     depthAttachment.GetRenderbuffer() == stencilAttachment.GetRenderbuffer());
+                depthStencilMismatch = !sameObject;
+            } else {
+                // GL_DEPTH_STENCIL_ATTACHMENT means "both halves"; a lone half does not answer it.
+                depthStencilMismatch = depthLive != stencilLive;
+            }
+            if (depthLive) return &depthAttachment;
+            if (stencilLive) return &stencilAttachment;
            return nullptr;
        }();

+        if (depthStencilMismatch) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", caller,
+                                             "GL_DEPTH_STENCIL_ATTACHMENT query with different depth and stencil "
+                                             "attachment images."));
+            return;
+        }
+
        switch (pname) {
        case GL_FRAMEBUFFER_ATTACHMENT_OBJECT_TYPE:
            if (attachmentObject == nullptr || attachmentObject->IsEmpty() || !attachmentObject->IsValid()) {
@@ -1527,6 +1756,10 @@ namespace MobileGL::MG_Impl::GLImpl {
                : GL_FALSE;
            break;
        default:
+            if (TryAnswerAttachmentFormatQuery(attachmentObject, attachmentType, depthStencilAlias, pname, params,
+                                               caller)) {
+                return;
+            }
            MG_State::pGLContext->RecordError(
                ErrorCode::InvalidEnum,
                MakeUnique<GenericErrorInfo>(
@@ -213,8 +213,13 @@ namespace MobileGL::MG_Impl::GLImpl {

            GLint maxSamples = 0;
            for (const auto& attachment : drawFbo->GetAllAttachmentObjects()) {
-                if (!attachment.IsRenderbuffer() || !attachment.GetRenderbuffer()) continue;
-                maxSamples = std::max(maxSamples, static_cast<GLint>(attachment.GetRenderbuffer()->GetSamples()));
+                if (attachment.IsRenderbuffer() && attachment.GetRenderbuffer()) {
+                    maxSamples = std::max(maxSamples, static_cast<GLint>(attachment.GetRenderbuffer()->GetSamples()));
+                } else if (attachment.IsTexture() && attachment.GetTexture()) {
+                    // Multisample texture attachments count too (GL_SAMPLE_BUFFERS must
+                    // report 1 for any multisampled draw framebuffer).
+                    maxSamples = std::max(maxSamples, static_cast<GLint>(attachment.GetTexture()->GetSamples()));
+                }
            }
            return maxSamples;
        }
@@ -1922,6 +1927,22 @@ namespace MobileGL::MG_Impl::GLImpl {
        case GL_MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS:
            *params = kFrontendMaxTransformFeedbackSeparateComponents;
            break;
+        // ARB_transform_feedback3 limits. The GL CTS queries these before checking
+        // whether the extension is advertised and requires no GL error; desktop
+        // drivers all accept them, so answer with the separate-attrib capacity and
+        // the single vertex stream the backends provide.
+        case GL_MAX_TRANSFORM_FEEDBACK_BUFFERS:
+            *params = kFrontendMaxTransformFeedbackSeparateAttribs;
+            break;
+        case GL_MAX_VERTEX_STREAMS:
+            *params = 1;
+            break;
+        case GL_TRANSFORM_FEEDBACK_ACTIVE:
+            *params = MG_State::pGLContext->IsTransformFeedbackActive() ? 1 : 0;
+            break;
+        case GL_TRANSFORM_FEEDBACK_PAUSED:
+            *params = 0;
+            break;
        case GL_MAX_TEXTURE_IMAGE_UNITS:
            *params = dynamicParameters.MaxTextureImageUnits;
            break;
@@ -49,10 +49,18 @@ namespace MobileGL::MG_Impl::GLImpl {

    static bool CheckProgramNameValidity(GLuint program) {
        if (!MG_State::pGLContext->ValidateProgramName(program)) {
+            // Programs and shaders share one name space: a name that exists but
+            // belongs to a shader is INVALID_OPERATION, a name GL never handed
+            // out is INVALID_VALUE (GL 3.3 core 2.11.x).
+            const ErrorCode error = MG_State::pGLContext->ValidateShaderName(program)
+                ? ErrorCode::InvalidOperation
+                : ErrorCode::InvalidValue;
            MG_State::pGLContext->RecordError(
-                ErrorCode::InvalidValue,
+                error,
                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
-                                             std::to_string(program) + " is not a valid name."));
+                                             std::to_string(program) +
+                                                 (error == ErrorCode::InvalidOperation ? " is not a program object."
+                                                                                       : " is not a valid name.")));
            return false;
        }
        return true;
@@ -605,6 +613,18 @@ namespace MobileGL::MG_Impl::GLImpl {
            *params = programObject->GetActiveUniformBlocksMaxNameLength() + 1;
            MGLOG_D("%s: %s = %d", __func__, MG_Util::ConvertGLEnumToString(pname).c_str(), *params);
            break;
+        case GL_TRANSFORM_FEEDBACK_VARYINGS:
+            *params = static_cast<GLint>(programObject->GetTransformFeedbackVaryingCount());
+            MGLOG_D("%s: %s = %d", __func__, MG_Util::ConvertGLEnumToString(pname).c_str(), *params);
+            break;
+        case GL_TRANSFORM_FEEDBACK_BUFFER_MODE:
+            *params = static_cast<GLint>(programObject->GetTransformFeedbackBufferMode());
+            MGLOG_D("%s: %s = %d", __func__, MG_Util::ConvertGLEnumToString(pname).c_str(), *params);
+            break;
+        case GL_TRANSFORM_FEEDBACK_VARYING_MAX_LENGTH:
+            *params = programObject->GetTransformFeedbackVaryingMaxLength();
+            MGLOG_D("%s: %s = %d", __func__, MG_Util::ConvertGLEnumToString(pname).c_str(), *params);
+            break;
        case GL_COMPUTE_WORK_GROUP_SIZE: { // GL >= 4.3
            if (!programObject->GetLinkStatus() || programObject->GetShaderIndexByStage(ShaderStage::Compute) < 0) {
                MG_State::pGLContext->RecordError(
@@ -624,9 +644,6 @@ namespace MobileGL::MG_Impl::GLImpl {

        case GL_PROGRAM_BINARY_LENGTH:

-        case GL_TRANSFORM_FEEDBACK_BUFFER_MODE:
-        case GL_TRANSFORM_FEEDBACK_VARYINGS:
-        case GL_TRANSFORM_FEEDBACK_VARYING_MAX_LENGTH:
        case GL_GEOMETRY_VERTICES_OUT:
        case GL_GEOMETRY_INPUT_TYPE:
        case GL_GEOMETRY_OUTPUT_TYPE:
@@ -827,19 +844,13 @@ namespace MobileGL::MG_Impl::GLImpl {
    }

    GLboolean IsProgram_State(GLuint program) {
-        /* FIXME: Handle situations that:
-         * A program object marked for deletion with glDeleteProgram but still in use as part of current
-         * rendering state is still considered a program object and glIsProgram will return GL_TRUE.
-         */
+        // Deletion-flagged names stay valid while the object is still GL-visible (program in
+        // use, shader attached), so name validity is exactly the Is* answer.
        if (program == 0) return GL_FALSE;
        return MG_State::pGLContext->ValidateProgramName(program) ? GL_TRUE : GL_FALSE;
    }

    GLboolean IsShader_State(GLuint shader) {
-        /* FIXME: Handle situations that:
-         * A shader object marked for deletion with glDeleteShader but still attached to a program object is still
-         * considered a shader object and glIsShader will return GL_TRUE.
-         */
        if (shader == 0) return GL_FALSE;
        return MG_State::pGLContext->ValidateShaderName(shader) ? GL_TRUE : GL_FALSE;
    }
@@ -849,6 +860,18 @@ namespace MobileGL::MG_Impl::GLImpl {
        if (!programObject) return;
        MGLOG_D("%s: linking program %d", __func__, program);

+        // Relinking the program an active transform feedback captures from would
+        // invalidate its varyings mid-capture (GL 3.3 core 2.11.3).
+        if (MG_State::pGLContext->IsTransformFeedbackActive() &&
+            MG_State::pGLContext->GetTransformFeedbackProgram().get() == programObject.get()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", __func__,
+                    "The program used by active transform feedback cannot be relinked."));
+            return;
+        }
+
        static Bool allowVSOnlyPrograms;
        static Bool initialized = false;
        if (!initialized) {
@@ -892,6 +915,16 @@ namespace MobileGL::MG_Impl::GLImpl {
    void UseProgram_State(GLuint program) {
        MGLOG_D("UseProgram_State: program=%u", program);

+        // GL 3.3 core 2.11.3: the program in use may not change while transform
+        // feedback is active (there is no pause in 3.3).
+        if (MG_State::pGLContext->IsTransformFeedbackActive()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "The current program cannot change while transform feedback is active."));
+            return;
+        }
+
        if (program == 0) {
            MG_State::pGLContext->UseProgram(0);
            return;
@@ -2217,4 +2250,66 @@ namespace MobileGL::MG_Impl::GLImpl {
    void ValidateProgram(GLuint program) {
        ValidateProgram_State(program);
    }
+
+    void TransformFeedbackVaryings(GLuint program, GLsizei count, const GLchar* const* varyings, GLenum bufferMode) {
+        auto& programObject = TryToGetProgramObject(program);
+        if (!programObject) return;
+        if (bufferMode != GL_INTERLEAVED_ATTRIBS && bufferMode != GL_SEPARATE_ATTRIBS) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidEnum,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "bufferMode is not a valid capture mode."));
+            return;
+        }
+        if (count < 0) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__, "count must be non-negative."));
+            return;
+        }
+        // GL 3.3 core: SEPARATE_ATTRIBS count may not exceed the separate-attrib limit.
+        if (bufferMode == GL_SEPARATE_ATTRIBS && count > 4) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             "count exceeds GL_MAX_TRANSFORM_FEEDBACK_SEPARATE_ATTRIBS."));
+            return;
+        }
+        Vector<String> names;
+        names.reserve(static_cast<SizeT>(count));
+        for (GLsizei i = 0; i < count; ++i) {
+            names.emplace_back(varyings != nullptr && varyings[i] != nullptr ? varyings[i] : "");
+        }
+        programObject->SetTransformFeedbackVaryings(Move(names), bufferMode);
+    }
+
+    void GetTransformFeedbackVarying(GLuint program, GLuint index, GLsizei bufSize, GLsizei* length, GLsizei* size,
+                                     GLenum* type, GLchar* name) {
+        auto& programObject = TryToGetProgramObject(program);
+        if (!programObject) return;
+        if (!programObject->GetLinkStatus()) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidOperation,
+                MakeUnique<GenericErrorInfo>("MG_Impl/GLImpl", __func__,
+                                             std::to_string(program) + " has not been successfully linked."));
+            return;
+        }
+        const auto* varying = programObject->GetTransformFeedbackVarying(index);
+        if (varying == nullptr) {
+            MG_State::pGLContext->RecordError(
+                ErrorCode::InvalidValue,
+                MakeUnique<GenericErrorInfo>(
+                    "MG_Impl/GLImpl", __func__,
+                    "index is not an active transform feedback varying of the program."));
+            return;
+        }
+        if (size != nullptr) *size = varying->size;
+        if (type != nullptr) *type = varying->type;
+        GLsizei written = 0;
+        if (name != nullptr && bufSize > 0) {
+            written = std::min<GLsizei>(bufSize - 1, static_cast<GLsizei>(varying->name.size()));
+            Memcpy(name, varying->name.data(), static_cast<SizeT>(written));
+            name[written] = '\0';
+        }
+        if (length != nullptr) *length = written;
+    }
 } // namespace MobileGL::MG_Impl::GLImpl
@@ -138,4 +138,7 @@ namespace MobileGL::MG_Impl::GLImpl {
    GLint GetProgramResourceLocationIndex(GLuint program, GLenum programInterface, const GLchar* name);
    void ShaderStorageBlockBinding(GLuint program, GLuint storageBlockIndex, GLuint storageBlockBinding);
    void ValidateProgram(GLuint program);
+    void TransformFeedbackVaryings(GLuint program, GLsizei count, const GLchar* const* varyings, GLenum bufferMode);
+    void GetTransformFeedbackVarying(GLuint program, GLuint index, GLsizei bufSize, GLsizei* length, GLsizei* size,
+                                     GLenum* type, GLchar* name);
 } // namespace MobileGL::MG_Impl::GLImpl
@@ -27,6 +27,8 @@ namespace MobileGL::MG_Impl::GLImpl {
            Bool ended = false;
            Bool resultCached = false;
            Uint64 cachedResult = 0;
+            // Transform feedback primitive counter at BeginQuery time.
+            Uint64 counterSnapshot = 0;
        };

        // Query calls may arrive from any thread (launchers migrate the context
@@ -41,6 +43,11 @@ namespace MobileGL::MG_Impl::GLImpl {
        GLuint g_nextQueryId = 1;
        // Id of the query currently active on GL_TIME_ELAPSED (0 = none).
        GLuint g_activeTimeElapsedQueryId = 0;
+        // Ids of the queries active on the transform feedback targets (0 = none).
+        GLuint g_activePrimitivesWrittenQueryId = 0;
+        GLuint g_activePrimitivesGeneratedQueryId = 0;
+        // Id of the query active on GL_SAMPLES_PASSED (0 = none).
+        GLuint g_activeSamplesPassedQueryId = 0;

        Bool TimerQueryDisabled() {
            return MG_Config::Features.DisableTimerQuery;
@@ -126,6 +133,11 @@ namespace MobileGL::MG_Impl::GLImpl {
                        outValue = 0;
                        return true;
                    }
+                    // ANY_SAMPLES_PASSED* report a boolean.
+                    if (queryObject->target == GL_ANY_SAMPLES_PASSED ||
+                        queryObject->target == GL_ANY_SAMPLES_PASSED_CONSERVATIVE) {
+                        result = result != 0 ? 1 : 0;
+                    }
                    // Final value produced (or no GetQueryResult64 hook: the
                    // query degrades to a zero result); the backend handle is
                    // consumed and the value cached for later reads.
@@ -180,7 +192,24 @@ namespace MobileGL::MG_Impl::GLImpl {
            }
            QueryObject* queryObject = it->second;
            if (queryObject->active) {
-                EndTimeElapsedQueryLocked(queryObject); // implicitly end before deletion
+                // Implicitly end before deletion, releasing the matching active slot.
+                if (queryObject->target == GL_SAMPLES_PASSED || queryObject->target == GL_ANY_SAMPLES_PASSED ||
+                    queryObject->target == GL_ANY_SAMPLES_PASSED_CONSERVATIVE) {
+                    if (const auto endOcclusionQuery = MG_Backend::gBackendFunctionsTable.GL.EndOcclusionQuery;
+                        endOcclusionQuery && queryObject->backendHandle) {
+                        endOcclusionQuery(queryObject->backendHandle);
+                    }
+                    queryObject->active = false;
+                    g_activeSamplesPassedQueryId = 0;
+                } else if (queryObject->target == GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN ||
+                           queryObject->target == GL_PRIMITIVES_GENERATED) {
+                    queryObject->active = false;
+                    (queryObject->target == GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN
+                         ? g_activePrimitivesWrittenQueryId
+                         : g_activePrimitivesGeneratedQueryId) = 0;
+                } else {
+                    EndTimeElapsedQueryLocked(queryObject);
+                }
            }
            if (queryObject->backendHandle) {
                if (const auto deleteBackendQuery = MG_Backend::gBackendFunctionsTable.GL.DeleteBackendQuery) {
@@ -204,10 +233,15 @@ namespace MobileGL::MG_Impl::GLImpl {
    }

    void BeginQuery(GLenum target, GLuint id) {
-        if (target != GL_TIME_ELAPSED) {
-            // Only GL_TIME_ELAPSED timer queries are implemented (occlusion and
-            // primitive queries remain stubs); GL_TIMESTAMP is not a valid
-            // BeginQuery target either.
+        const Bool isTransformFeedbackQuery =
+            target == GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN || target == GL_PRIMITIVES_GENERATED;
+        const Bool isOcclusionQuery =
+            (target == GL_SAMPLES_PASSED || target == GL_ANY_SAMPLES_PASSED ||
+             target == GL_ANY_SAMPLES_PASSED_CONSERVATIVE) &&
+            MG_Backend::gBackendFunctionsTable.GL.BeginOcclusionQuery != nullptr;
+        if (target != GL_TIME_ELAPSED && !isTransformFeedbackQuery && !isOcclusionQuery) {
+            // GL_TIMESTAMP is not a valid BeginQuery target; the occlusion targets
+            // need backend support.
            RecordQueryError(ErrorCode::InvalidEnum, __FUNCTION__, "Query target is not supported.");
            return;
        }
@@ -221,9 +255,13 @@ namespace MobileGL::MG_Impl::GLImpl {
            RecordQueryError(ErrorCode::InvalidOperation, __FUNCTION__, "Query object does not exist.");
            return;
        }
-        if (g_activeTimeElapsedQueryId != 0) {
+        GLuint& activeQueryId = isTransformFeedbackQuery
+            ? (target == GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN ? g_activePrimitivesWrittenQueryId
+                                                                  : g_activePrimitivesGeneratedQueryId)
+            : (isOcclusionQuery ? g_activeSamplesPassedQueryId : g_activeTimeElapsedQueryId);
+        if (activeQueryId != 0) {
            RecordQueryError(ErrorCode::InvalidOperation, __FUNCTION__,
-                             "A query is already active on GL_TIME_ELAPSED.");
+                             "A query is already active on this target.");
            return;
        }
        if (queryObject->active) {
@@ -239,25 +277,72 @@ namespace MobileGL::MG_Impl::GLImpl {
        ResetQueryObjectLocked(queryObject); // discard any previous result
        queryObject->target = target;
        queryObject->active = true;
-        const auto beginTimeElapsedQuery = MG_Backend::gBackendFunctionsTable.GL.BeginTimeElapsedQuery;
-        queryObject->backendHandle =
-            (!TimerQueryDisabled() && beginTimeElapsedQuery) ? beginTimeElapsedQuery() : nullptr;
-        g_activeTimeElapsedQueryId = id;
+        if (isTransformFeedbackQuery) {
+            // Prefer real GPU transform-feedback queries (exact with geometry shaders);
+            // the CPU accounting delta stays as the fallback when the backend lacks them.
+            const auto beginXfbPrimitivesQuery = MG_Backend::gBackendFunctionsTable.GL.BeginXfbPrimitivesQuery;
+            queryObject->backendHandle =
+                beginXfbPrimitivesQuery ? beginXfbPrimitivesQuery(target == GL_PRIMITIVES_GENERATED) : nullptr;
+            queryObject->counterSnapshot = MG_State::pGLContext->GetTransformFeedbackPrimitiveCounter();
+        } else if (isOcclusionQuery) {
+            queryObject->backendHandle = MG_Backend::gBackendFunctionsTable.GL.BeginOcclusionQuery();
+        } else {
+            const auto beginTimeElapsedQuery = MG_Backend::gBackendFunctionsTable.GL.BeginTimeElapsedQuery;
+            queryObject->backendHandle =
+                (!TimerQueryDisabled() && beginTimeElapsedQuery) ? beginTimeElapsedQuery() : nullptr;
+        }
+        activeQueryId = id;
    }

    void EndQuery(GLenum target) {
-        if (target != GL_TIME_ELAPSED) {
+        const Bool isTransformFeedbackQuery =
+            target == GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN || target == GL_PRIMITIVES_GENERATED;
+        const Bool isOcclusionQuery =
+            (target == GL_SAMPLES_PASSED || target == GL_ANY_SAMPLES_PASSED ||
+             target == GL_ANY_SAMPLES_PASSED_CONSERVATIVE) &&
+            MG_Backend::gBackendFunctionsTable.GL.BeginOcclusionQuery != nullptr;
+        if (target != GL_TIME_ELAPSED && !isTransformFeedbackQuery && !isOcclusionQuery) {
            RecordQueryError(ErrorCode::InvalidEnum, __FUNCTION__, "Query target is not supported.");
            return;
        }
        const std::lock_guard<std::mutex> lock(g_queryObjectsMutex);
-        if (g_activeTimeElapsedQueryId == 0) {
-            RecordQueryError(ErrorCode::InvalidOperation, __FUNCTION__, "No query is active on GL_TIME_ELAPSED.");
+        GLuint& activeQueryId = isTransformFeedbackQuery
+            ? (target == GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN ? g_activePrimitivesWrittenQueryId
+                                                                  : g_activePrimitivesGeneratedQueryId)
+            : (isOcclusionQuery ? g_activeSamplesPassedQueryId : g_activeTimeElapsedQueryId);
+        if (activeQueryId == 0) {
+            RecordQueryError(ErrorCode::InvalidOperation, __FUNCTION__, "No query is active on this target.");
            return;
        }
-        auto* queryObject = FindQueryObjectLocked(g_activeTimeElapsedQueryId);
+        auto* queryObject = FindQueryObjectLocked(activeQueryId);
        if (!queryObject) {
-            g_activeTimeElapsedQueryId = 0; // should not happen; keep state consistent
+            activeQueryId = 0; // should not happen; keep state consistent
+            return;
+        }
+        if (isTransformFeedbackQuery) {
+            if (queryObject->backendHandle) {
+                if (const auto endXfbPrimitivesQuery = MG_Backend::gBackendFunctionsTable.GL.EndXfbPrimitivesQuery) {
+                    endXfbPrimitivesQuery(queryObject->backendHandle);
+                }
+                // Result comes from the GPU query at read time.
+            } else {
+                queryObject->cachedResult =
+                    MG_State::pGLContext->GetTransformFeedbackPrimitiveCounter() - queryObject->counterSnapshot;
+                queryObject->resultCached = true;
+            }
+            queryObject->active = false;
+            queryObject->ended = true;
+            activeQueryId = 0;
+            return;
+        }
+        if (isOcclusionQuery) {
+            if (const auto endOcclusionQuery = MG_Backend::gBackendFunctionsTable.GL.EndOcclusionQuery;
+                endOcclusionQuery && queryObject->backendHandle) {
+                endOcclusionQuery(queryObject->backendHandle);
+            }
+            queryObject->active = false;
+            queryObject->ended = true;
+            activeQueryId = 0;
            return;
        }
        EndTimeElapsedQueryLocked(queryObject);
@@ -303,9 +388,25 @@ namespace MobileGL::MG_Impl::GLImpl {
        switch (pname) {
        case GL_CURRENT_QUERY: {
            const std::lock_guard<std::mutex> lock(g_queryObjectsMutex);
-            // Only GL_TIME_ELAPSED queries can be active; GL_TIMESTAMP queries
-            // never are, and other targets remain unimplemented.
-            *params = target == GL_TIME_ELAPSED ? static_cast<GLint>(g_activeTimeElapsedQueryId) : 0;
+            switch (target) {
+            case GL_TIME_ELAPSED:
+                *params = static_cast<GLint>(g_activeTimeElapsedQueryId);
+                break;
+            case GL_SAMPLES_PASSED:
+            case GL_ANY_SAMPLES_PASSED:
+            case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
+                *params = static_cast<GLint>(g_activeSamplesPassedQueryId);
+                break;
+            case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
+                *params = static_cast<GLint>(g_activePrimitivesWrittenQueryId);
+                break;
+            case GL_PRIMITIVES_GENERATED:
+                *params = static_cast<GLint>(g_activePrimitivesGeneratedQueryId);
+                break;
+            default:
+                *params = 0;
+                break;
+            }
            return;
        }
        case GL_QUERY_COUNTER_BITS: {
@@ -313,7 +414,13 @@ namespace MobileGL::MG_Impl::GLImpl {
            // time: IsTimerQuerySupported is the dynamic truth (extension /
            // entry points / timestamp valid bits at call time, not at table
            // init), and the MOBILEGL_DISABLE_TIMERQUERY kill switch always
-            // wins. Non-timer targets remain unimplemented and report 0.
+            // wins.
+            if (target == GL_SAMPLES_PASSED || target == GL_ANY_SAMPLES_PASSED ||
+                target == GL_ANY_SAMPLES_PASSED_CONSERVATIVE) {
+                const Bool occlusionSupported = MG_Backend::gBackendFunctionsTable.GL.BeginOcclusionQuery != nullptr;
+                *params = occlusionSupported ? (target == GL_SAMPLES_PASSED ? 32 : 1) : 0;
+                return;
+            }
            const Bool timerTarget = target == GL_TIME_ELAPSED || target == GL_TIMESTAMP;
            const auto isTimerQuerySupported = MG_Backend::gBackendFunctionsTable.GL.IsTimerQuerySupported;
            const Bool supported =
@@ -185,6 +185,11 @@ namespace MobileGL::MG_Impl::GLImpl {
        static thread_local Vector<GLuint> names;
        MG_State::pGLContext->GenSamplerNames(count, names);
        Memcpy(samplers, names.data(), count * sizeof(GLuint));
+        // Unlike textures/buffers, glGenSamplers CREATES the sampler objects: each name
+        // is immediately a sampler (glIsSampler == GL_TRUE before any bind).
+        for (GLsizei i = 0; i < count; ++i) {
+            MG_State::pGLContext->CreateSamplerObject(names[i]);
+        }
    }

    void DeleteSamplers_State(GLsizei count, const GLuint* samplers) {
@@ -2820,7 +2820,22 @@ namespace MobileGL::MG_Impl::GLImpl {
                                         "The attachment specified by the read buffer is incomplete."));               \
        return false;                                                                                                  \
    }
-        if (isDepth) {
+        if (isDepth && isStencil) {
+            // A combined internalformat copies both halves, so the read framebuffer
+            // must populate both attachment points.
+            const auto& stencilAttachment = currentReadFBO->GetAttachment(FramebufferAttachmentType::Stencil);
+            const auto& depthAttachment = currentReadFBO->GetAttachment(FramebufferAttachmentType::Depth);
+            if (!depthAttachment.IsValid() || depthAttachment.IsEmpty() || !stencilAttachment.IsValid() ||
+                stencilAttachment.IsEmpty()) {
+                MG_State::pGLContext->RecordError(
+                    ErrorCode::InvalidOperation,
+                    MakeUnique<GenericErrorInfo>(
+                        "MG_Impl/GLImpl", "CopyTexImage2D_State",
+                        "DEPTH_STENCIL copy requires both depth and stencil attachments in the read framebuffer."));
+                return false;
+            }
+            GET_SRC_INTERNAL_FORMAT(FramebufferAttachmentType::Depth);
+        } else if (isDepth) {
            GET_SRC_INTERNAL_FORMAT(FramebufferAttachmentType::Depth);
        } else if (isStencil) {
            GET_SRC_INTERNAL_FORMAT(FramebufferAttachmentType::Stencil);
@@ -231,6 +231,21 @@ namespace MobileGL::MG_State::GLState {
        return m_resource.Bytes();
    }

+    Bool BufferObject::EnsureGpuResidentStorage() {
+        if (m_resource.IsGpuResident()) {
+            return true;
+        }
+        if (m_size == 0 || g_bufferBackendOps == nullptr || g_bufferBackendOps->AcquirePersistentMap == nullptr) {
+            return false;
+        }
+        void* base = g_bufferBackendOps->AcquirePersistentMap(*this);
+        if (base == nullptr) {
+            return false;
+        }
+        m_resource.AdoptPersistentMap(base);
+        return true;
+    }
+
    void* BufferObject::AcquireMemoryRange(Range1D range, Flags<BufferMappingAccessBit> access) {
        MOBILEGL_ASSERT(range.end <= m_size && range.start <= range.end,
                        "AcquireMemoryRange out of bounds: range (%zu, %zu) exceeds m_size (%zu)", range.start,
@@ -133,6 +133,11 @@ namespace MobileGL {

            void* AcquireMemory(Bool markMapped, Bool read, Bool write);
            void* AcquireMemoryRange(Range1D range, Flags<BufferMappingAccessBit> access);
+            // Adopt backend host-visible coherent GPU storage as the source of truth
+            // (used for GPU-written targets like transform feedback capture, so
+            // MapBuffer/GetBufferSubData read real GPU results). No-op when already
+            // resident or when the backend declines.
+            Bool EnsureGpuResidentStorage();
            void ReleaseMemory();
            void FlushMemoryRange(SizeT offset, SizeT length);

@@ -211,6 +211,47 @@ namespace MobileGL {
                void SetScissorBox(IntVec4 box);      // x, y, width, height
                const IntVec4& GetScissorBox() const; // x, y, width, height

+                // Transform feedback (GL 3.0 core Begin/End; no feedback objects yet)
+                void BeginTransformFeedback(GLenum primitiveMode, const SharedPtr<ProgramObject>& program) {
+                    m_transformFeedbackActive = true;
+                    m_transformFeedbackPrimitiveMode = primitiveMode;
+                    m_transformFeedbackProgram = program;
+                    ++m_transformFeedbackGeneration;
+                    m_transformFeedbackCapturedVertices = 0;
+                    m_transformFeedbackInputPrimitives = 0;
+                }
+                void EndTransformFeedback() {
+                    m_transformFeedbackActive = false;
+                    m_transformFeedbackProgram.reset();
+                }
+                Bool IsTransformFeedbackActive() const { return m_transformFeedbackActive; }
+                GLenum GetTransformFeedbackPrimitiveMode() const { return m_transformFeedbackPrimitiveMode; }
+                const SharedPtr<ProgramObject>& GetTransformFeedbackProgram() const {
+                    return m_transformFeedbackProgram;
+                }
+                // Bumped on every BeginTransformFeedback; the backend uses it to
+                // distinguish "resume appending" from "fresh capture".
+                Uint64 GetTransformFeedbackGeneration() const { return m_transformFeedbackGeneration; }
+                // CPU-side primitive accounting for the transform feedback queries:
+                // every captured draw adds its primitive count (draws without a
+                // geometry stage write exactly what they generate).
+                void AddTransformFeedbackPrimitives(Uint64 primitives) {
+                    m_transformFeedbackPrimitiveCounter += primitives;
+                }
+                Uint64 GetTransformFeedbackPrimitiveCounter() const { return m_transformFeedbackPrimitiveCounter; }
+                // Vertices already captured since BeginTransformFeedback (drives the
+                // buffer-capacity clamp on the primitives-written accounting).
+                void AddTransformFeedbackCapturedVertices(Uint64 vertices) {
+                    m_transformFeedbackCapturedVertices += vertices;
+                }
+                Uint64 GetTransformFeedbackCapturedVertices() const { return m_transformFeedbackCapturedVertices; }
+                // Raw assembled input primitives fed to the capture stage since Begin
+                // (pre-clamp; drives the GS strip capture-order fixup at EndTF).
+                void AddTransformFeedbackInputPrimitives(Uint64 primitives) {
+                    m_transformFeedbackInputPrimitives += primitives;
+                }
+                Uint64 GetTransformFeedbackInputPrimitives() const { return m_transformFeedbackInputPrimitives; }
+
                // Framebuffer
                void GenFramebufferNames(Uint number, Vector<Uint>& framebuffers);
                const SharedPtr<FramebufferObject>& GetFramebufferObject(Uint index);
@@ -243,6 +284,13 @@ namespace MobileGL {
                BufferState m_bufferState;
                VertexArrayState m_vertexArrayState;
                Array<CurrentVertexAttributeValue, VertexArrayObject::MAX_VERTEX_ATTRIBS> m_currentVertexAttributes{};
+                Bool m_transformFeedbackActive = false;
+                GLenum m_transformFeedbackPrimitiveMode = GL_POINTS;
+                SharedPtr<ProgramObject> m_transformFeedbackProgram;
+                Uint64 m_transformFeedbackGeneration = 0;
+                Uint64 m_transformFeedbackPrimitiveCounter = 0;
+                Uint64 m_transformFeedbackCapturedVertices = 0;
+                Uint64 m_transformFeedbackInputPrimitives = 0;
                TextureState m_textureState;
                ProgramState m_programState;
                RenderState m_renderState;
@@ -170,9 +170,230 @@ namespace MobileGL::MG_State::GLState {
        m_uniformNameMaxLength = 0;
        m_attribInNameMaxLength = 0;
        m_uniformBlockNameMaxLength = 0;
+        m_xfbVaryings.clear();
+        m_xfbStrides.clear();
+        m_xfbBufferMode = GL_INTERLEAVED_ATTRIBS;
+        m_xfbVaryingNameMaxLength = 0;
        m_linkStatus = false;
    }

+    namespace {
+        // GL type enum for a vertex-stage output symbol captured by transform
+        // feedback. Covers the scalar/vector/matrix float+integer types transform
+        // feedback may legally capture in GL 3.3.
+        Bool ResolveXfbSymbolType(const glslang::TType& type, GLenum& outType, GLint& outArraySize,
+                                  Uint32& outBytesPerElement) {
+            outArraySize = type.isArray() ? type.getOuterArraySize() : 1;
+            const Int columns = type.isMatrix() ? type.getMatrixCols() : 1;
+            const Int components = type.isMatrix() ? type.getMatrixRows()
+                                                   : (type.isVector() ? type.getVectorSize() : 1);
+            const glslang::TBasicType basic = type.getBasicType();
+            static constexpr GLenum kFloatTypes[5] = {0, GL_FLOAT, GL_FLOAT_VEC2, GL_FLOAT_VEC3, GL_FLOAT_VEC4};
+            static constexpr GLenum kIntTypes[5] = {0, GL_INT, GL_INT_VEC2, GL_INT_VEC3, GL_INT_VEC4};
+            static constexpr GLenum kUintTypes[5] = {0, GL_UNSIGNED_INT, GL_UNSIGNED_INT_VEC2, GL_UNSIGNED_INT_VEC3,
+                                                     GL_UNSIGNED_INT_VEC4};
+            if (type.isMatrix()) {
+                if (basic != glslang::EbtFloat) return false;
+                static constexpr GLenum kMatTypes[5][5] = {
+                    {}, {},
+                    {0, 0, GL_FLOAT_MAT2, GL_FLOAT_MAT2x3, GL_FLOAT_MAT2x4},
+                    {0, 0, GL_FLOAT_MAT3x2, GL_FLOAT_MAT3, GL_FLOAT_MAT3x4},
+                    {0, 0, GL_FLOAT_MAT4x2, GL_FLOAT_MAT4x3, GL_FLOAT_MAT4},
+                };
+                if (columns < 2 || columns > 4 || components < 2 || components > 4) return false;
+                outType = kMatTypes[columns][components];
+            } else if (components >= 1 && components <= 4) {
+                switch (basic) {
+                case glslang::EbtFloat: outType = kFloatTypes[components]; break;
+                case glslang::EbtInt: outType = kIntTypes[components]; break;
+                case glslang::EbtUint: outType = kUintTypes[components]; break;
+                default: return false;
+                }
+            } else {
+                return false;
+            }
+            outBytesPerElement = static_cast<Uint32>(columns * components) * 4u;
+            return true;
+        }
+    } // namespace
+
+    Bool ProgramObject::ResolveTransformFeedbackVaryings() {
+        m_xfbVaryings.clear();
+        m_xfbStrides.clear();
+        m_xfbBufferMode = m_requestedXfbBufferMode;
+        m_xfbVaryingNameMaxLength = 0;
+        if (m_requestedXfbVaryings.empty()) {
+            return true;
+        }
+
+        // Capture happens at the last vertex-processing stage (geometry, then
+        // tessellation evaluation, then vertex).
+        const glslang::TIntermediate* captureIntermediate = nullptr;
+        for (EShLanguage stage : {EShLangGeometry, EShLangTessEvaluation, EShLangVertex}) {
+            captureIntermediate = m_program->getIntermediate(stage);
+            if (captureIntermediate != nullptr) {
+                break;
+            }
+        }
+        if (captureIntermediate == nullptr) {
+            m_infoLog = "Transform feedback varyings requested but the program has no vertex-processing stage.";
+            return false;
+        }
+        const glslang::TIntermAggregate* linkerObjects = captureIntermediate->findLinkerObjects();
+
+        const Bool interleaved = m_xfbBufferMode == GL_INTERLEAVED_ATTRIBS;
+        Uint32 interleavedOffset = 0;
+        for (SizeT i = 0; i < m_requestedXfbVaryings.size(); ++i) {
+            const String& name = m_requestedXfbVaryings[i];
+            for (SizeT j = 0; j < i; ++j) {
+                if (m_requestedXfbVaryings[j] == name) {
+                    m_infoLog = "Transform feedback varying '" + name + "' is specified more than once.";
+                    return false;
+                }
+            }
+
+            XfbVarying varying;
+            varying.name = name;
+            Uint32 bytesPerElement = 0;
+            Bool resolved = false;
+            if (name == "gl_Position") {
+                varying.type = GL_FLOAT_VEC4;
+                varying.size = 1;
+                bytesPerElement = 16;
+                resolved = true;
+            } else if (name == "gl_PointSize") {
+                varying.type = GL_FLOAT;
+                varying.size = 1;
+                bytesPerElement = 4;
+                resolved = true;
+            } else if (linkerObjects != nullptr) {
+                for (const auto* node : linkerObjects->getSequence()) {
+                    const glslang::TIntermSymbol* symbol = node->getAsSymbolNode();
+                    if (symbol == nullptr || symbol->getType().getQualifier().storage != glslang::EvqVaryingOut) {
+                        continue;
+                    }
+                    if (symbol->getName() != name.c_str()) {
+                        continue;
+                    }
+                    resolved = ResolveXfbSymbolType(symbol->getType(), varying.type, varying.size, bytesPerElement);
+                    break;
+                }
+            }
+            if (!resolved) {
+                m_infoLog = "Transform feedback varying '" + name + "' is not an output of the vertex stage.";
+                return false;
+            }
+
+            varying.byteSize = bytesPerElement * static_cast<Uint32>(varying.size);
+            if (interleaved) {
+                varying.bufferIndex = 0;
+                varying.offsetBytes = interleavedOffset;
+                interleavedOffset += varying.byteSize;
+            } else {
+                varying.bufferIndex = static_cast<Uint32>(i);
+                varying.offsetBytes = 0;
+            }
+            m_xfbVaryingNameMaxLength =
+                std::max(m_xfbVaryingNameMaxLength, static_cast<Int>(name.size()) + 1);
+            m_xfbVaryings.push_back(Move(varying));
+        }
+
+        constexpr Uint32 kMaxSeparateAttribs = 4;
+        constexpr Uint32 kMaxSeparateComponents = 4;
+        constexpr Uint32 kMaxInterleavedComponents = 64;
+        if (interleaved) {
+            if (interleavedOffset > kMaxInterleavedComponents * 4) {
+                m_infoLog = "Transform feedback interleaved capture exceeds "
+                            "GL_MAX_TRANSFORM_FEEDBACK_INTERLEAVED_COMPONENTS.";
+                return false;
+            }
+            m_xfbStrides.assign(1, interleavedOffset);
+        } else {
+            if (m_xfbVaryings.size() > kMaxSeparateAttribs) {
+                m_infoLog = "Transform feedback separate capture exceeds "
+                            "GL_MAX_TRANSFORM_FEEDBACK_SEPARATE_ATTRIBS.";
+                return false;
+            }
+            m_xfbStrides.resize(m_xfbVaryings.size());
+            for (SizeT i = 0; i < m_xfbVaryings.size(); ++i) {
+                if (m_xfbVaryings[i].byteSize > kMaxSeparateComponents * 4) {
+                    m_infoLog = "Transform feedback varying '" + m_xfbVaryings[i].name +
+                                "' exceeds GL_MAX_TRANSFORM_FEEDBACK_SEPARATE_COMPONENTS.";
+                    return false;
+                }
+                m_xfbStrides[i] = m_xfbVaryings[i].byteSize;
+            }
+        }
+
+        ResolveGsTriangleStripCapture(captureIntermediate);
+        return true;
+    }
+
+    namespace {
+        // Extracts a geometry shader's per-invocation EmitVertex/EndPrimitive sequence
+        // when it is statically knowable (no emit inside selection/loop/switch). Vulkan
+        // transform feedback captures triangle strips in plain (i, i+1, i+2) order while
+        // GL decomposes odd strip triangles as (i+1, i, i+2) (GL 4.6 table 10.1); with
+        // the static strip lengths the capture buffer can be reordered after EndTF.
+        class GsEmitSequenceTraverser final : public glslang::TIntermTraverser {
+        public:
+            bool visitAggregate(glslang::TVisit, glslang::TIntermAggregate* node) override {
+                if (node->getOp() == glslang::EOpEmitVertex) {
+                    ++emitCount;
+                    hasEmit = true;
+                } else if (node->getOp() == glslang::EOpEndPrimitive) {
+                    FlushStrip();
+                }
+                return true;
+            }
+            bool visitSelection(glslang::TVisit, glslang::TIntermSelection*) override {
+                inControlFlow = true;
+                return true;
+            }
+            bool visitLoop(glslang::TVisit, glslang::TIntermLoop*) override {
+                inControlFlow = true;
+                return true;
+            }
+            bool visitSwitch(glslang::TVisit, glslang::TIntermSwitch*) override {
+                inControlFlow = true;
+                return true;
+            }
+            void FlushStrip() {
+                if (emitCount >= 3) {
+                    stripTriangles.push_back(static_cast<Uint32>(emitCount - 2));
+                }
+                emitCount = 0;
+            }
+
+            Vector<Uint32> stripTriangles;
+            Uint32 emitCount = 0;
+            Bool hasEmit = false;
+            Bool inControlFlow = false;
+        };
+    } // namespace
+
+    void ProgramObject::ResolveGsTriangleStripCapture(const glslang::TIntermediate* captureIntermediate) {
+        m_gsStripTriangles.clear();
+        m_gsStripCaptureFixup = false;
+        if (captureIntermediate == nullptr || m_program == nullptr) {
+            return;
+        }
+        if (m_program->getIntermediate(EShLangGeometry) != captureIntermediate) {
+            return;
+        }
+        if (captureIntermediate->getOutputPrimitive() != glslang::ElgTriangleStrip) {
+            return;
+        }
+        GsEmitSequenceTraverser traverser;
+        const_cast<glslang::TIntermediate*>(captureIntermediate)->getTreeRoot()->traverse(&traverser);
+        traverser.FlushStrip(); // the invocation end acts as an implicit EndPrimitive
+        if (!traverser.hasEmit || traverser.inControlFlow || traverser.stripTriangles.empty()) {
+            return;
+        }
+        m_gsStripTriangles = Move(traverser.stripTriangles);
+        m_gsStripCaptureFixup = true;
+    }
+
    bool ProgramObject::ShaderIsAttached(const SharedPtr<ShaderObject>& shader) {
        MGLOG_D("ProgramObject %u: ShaderIsAttached check for shader %p", m_externalIndex, shader.get());
        auto it = std::find_if(m_shaders.begin(), m_shaders.end(),
@@ -327,6 +548,12 @@ namespace MobileGL::MG_State::GLState {
        if (!ValidateFragmentOutputLocations()) {
            return;
        }
+        if (!ResolveTransformFeedbackVaryings()) {
+            m_linkStatus = false;
+            MGLOG_E("ProgramObject %u: transform feedback varying resolution failed: %s", m_externalIndex,
+                    m_infoLog.c_str());
+            return;
+        }

        MGLOG_D("ProgramObject %u: Starting binary generation", m_externalIndex);
        GenerateBinary();
@@ -465,6 +465,39 @@ namespace MobileGL::MG_State::GLState {
            return it == m_shaders.end() ? -1 : (Int)std::distance(m_shaders.begin(), it);
        }

+        // Transform feedback (GL 3.0 core: glTransformFeedbackVaryings applies on
+        // the NEXT link; the linked snapshot below is what draws and queries see).
+        struct XfbVarying {
+            String name;
+            GLenum type = GL_FLOAT;
+            GLint size = 1;           // array element count
+            Uint32 bufferIndex = 0;   // capture buffer slot
+            Uint32 offsetBytes = 0;   // offset within the capture buffer
+            Uint32 byteSize = 0;      // bytes captured per vertex for this varying
+        };
+        void SetTransformFeedbackVaryings(Vector<String>&& names, GLenum bufferMode) {
+            m_requestedXfbVaryings = Move(names);
+            m_requestedXfbBufferMode = bufferMode;
+        }
+        GLenum GetTransformFeedbackBufferMode() const { return m_xfbBufferMode; }
+        SizeT GetTransformFeedbackVaryingCount() const { return m_xfbVaryings.size(); }
+        const XfbVarying* GetTransformFeedbackVarying(SizeT index) const {
+            return index < m_xfbVaryings.size() ? &m_xfbVaryings[index] : nullptr;
+        }
+        const Vector<XfbVarying>& GetTransformFeedbackVaryings() const { return m_xfbVaryings; }
+        // Stride of one captured vertex in the given capture buffer slot.
+        Uint32 GetTransformFeedbackStride(Uint32 bufferIndex) const {
+            return bufferIndex < m_xfbStrides.size() ? m_xfbStrides[bufferIndex] : 0;
+        }
+        SizeT GetTransformFeedbackBufferCount() const { return m_xfbStrides.size(); }
+        Int GetTransformFeedbackVaryingMaxLength() const { return m_xfbVaryingNameMaxLength; }
+        // True when the capture stage is a triangle-strip geometry shader with a
+        // statically-known emit sequence: the Vulkan capture order then needs the GL
+        // odd-triangle vertex swap after EndTransformFeedback.
+        Bool HasGsTriangleStripCaptureFixup() const { return m_gsStripCaptureFixup; }
+        // Triangles per strip, in emission order, for ONE geometry invocation.
+        const Vector<Uint32>& GetGsStripTriangles() const { return m_gsStripTriangles; }
+
        Uint GetExternalIndex() const { return m_externalIndex; }
        // Globally-unique, never-reused id for this program object's lifetime. Unlike the GL
        // name (external index), which is freed to a LIFO list and immediately handed back by
@@ -475,6 +508,11 @@ namespace MobileGL::MG_State::GLState {
    private:
        void ResetLinkArtifacts();
        void DoReflection();
+        // Resolves the requested transform feedback varyings against the linked
+        // vertex stage; fails the link (GL semantics) on unknown or duplicate
+        // names or exceeded capture limits.
+        Bool ResolveTransformFeedbackVaryings();
+        void ResolveGsTriangleStripCapture(const glslang::TIntermediate* captureIntermediate);
        void GenerateBinary();
        void WaitUntilGenerationCompleted() const;
        void AddDefaultFragmentShaderIfMissing();
@@ -558,5 +596,15 @@ namespace MobileGL::MG_State::GLState {
        mutable Uint32 m_backendHashMemoVersion = ~0u;
        Uint32 m_uboContentVersion = 0;
        Uint32 m_linkVersion = 0;
+
+        // Transform feedback: request (applies at next link) and linked snapshot.
+        Vector<String> m_requestedXfbVaryings;
+        GLenum m_requestedXfbBufferMode = GL_INTERLEAVED_ATTRIBS;
+        Vector<XfbVarying> m_xfbVaryings;
+        Vector<Uint32> m_xfbStrides;
+        Vector<Uint32> m_gsStripTriangles;
+        Bool m_gsStripCaptureFixup = false;
+        GLenum m_xfbBufferMode = GL_INTERLEAVED_ATTRIBS;
+        Int m_xfbVaryingNameMaxLength = 0;
    };
 } // namespace MobileGL::MG_State::GLState
@@ -29,17 +29,25 @@ namespace MobileGL::MG_State::GLState {
        if (!CheckIndexAvail(program, m_programObjects)) return; // FIXME: add error reporting here
        auto& programObject = m_programObjects[program];
        if (programObject != nullptr) {
-            // Snapshot the attachments: deleting the program is a detach point for shaders
-            // that were flagged with glDeleteShader while still attached.
-            const Vector<SharedPtr<ShaderObject>> attachedShaders = programObject->GetAttachedShaders();
            programObject->MarkAsDeleted();
-            programObject.reset();
-            m_programIndexGenerator.Delete(program);
-            for (const auto& shader : attachedShaders) {
-                const Uint shaderName = shader->GetExternalIndex();
-                if (CheckIndexAvail(shaderName, m_shaderObjects) && m_shaderObjects[shaderName] == shader) {
-                    ReleaseShaderNameIfOrphaned(shaderName);
-                }
+            // A program in use is only FLAGGED: its name (and every program query) stays
+            // valid until it stops being current, at which point UseProgram finishes the job.
+            if (programObject == m_currentProgram) return;
+            DestroyProgramSlot(program);
+        }
+    }
+
+    void ProgramState::DestroyProgramSlot(const Uint program) {
+        auto& programObject = m_programObjects[program];
+        // Snapshot the attachments: deleting the program is a detach point for shaders
+        // that were flagged with glDeleteShader while still attached.
+        const Vector<SharedPtr<ShaderObject>> attachedShaders = programObject->GetAttachedShaders();
+        programObject.reset();
+        m_programIndexGenerator.Delete(program);
+        for (const auto& shader : attachedShaders) {
+            const Uint shaderName = shader->GetExternalIndex();
+            if (CheckIndexAvail(shaderName, m_shaderObjects) && m_shaderObjects[shaderName] == shader) {
+                ReleaseShaderNameIfOrphaned(shaderName);
            }
        }
    }
@@ -49,10 +57,22 @@ namespace MobileGL::MG_State::GLState {
    }

    void ProgramState::UseProgram(Uint program) {
+        const SharedPtr<ProgramObject> previous = m_currentProgram;
+
        if (program == 0) m_currentProgram.reset();

-        if (!CheckIndexAvail(program, m_programObjects)) return;
-        m_currentProgram = m_programObjects[program];
+        if (CheckIndexAvail(program, m_programObjects)) {
+            m_currentProgram = m_programObjects[program];
+        }
+
+        // A deletion flagged while the program was current takes effect the moment it
+        // stops being current.
+        if (previous != nullptr && previous != m_currentProgram && previous->GetDeleteStatus()) {
+            const Uint previousName = previous->GetExternalIndex();
+            if (CheckIndexAvail(previousName, m_programObjects) && m_programObjects[previousName] == previous) {
+                DestroyProgramSlot(previousName);
+            }
+        }
    }

    Uint ProgramState::CreateShader(ShaderStage stage) {
@@ -35,6 +35,9 @@ namespace MobileGL::MG_State::GLState {

    private:
        Bool ShaderHasGLVisibleAttachment(const SharedPtr<ShaderObject>& shaderObject) const;
+        // Frees the name slot and releases orphaned attached shaders; the immediate half
+        // of glDeleteProgram (deferred while the program is current).
+        void DestroyProgramSlot(Uint program);

        template <typename T>
        static Bool CheckIndexAvail(const SizeT idx, const Vector<T>& vec) {
@@ -169,6 +169,15 @@ namespace MobileGL::MG_State::GLState {
            }
        }

+        const std::optional<String> reservedError =
+            MG_Util::ShaderTranspiler::FindReservedIdentifierViolation(compileSource);
+        if (reservedError) {
+            m_compileStatus = false;
+            m_shader.reset();
+            m_infoLog = *reservedError;
+            return;
+        }
+
        // Compile for OpenGL here, so that we can do validation and link
        // like a real OpenGL driver at linking stage
        // Will compile for other backends later.
@@ -162,6 +162,7 @@ namespace MobileGL {
        DepthComponent32F,
        Depth24Stencil8,
        Depth32FStencil8,
+        StencilIndex8,

        DepthComponent,
        DepthStencil,
@@ -28,6 +28,7 @@ namespace MobileGL {

        bool IsStencilFormatInternalFormat(TextureInternalFormat internalformat) {
            switch (internalformat) {
+            case TextureInternalFormat::StencilIndex8:
            case TextureInternalFormat::Depth24Stencil8:
            case TextureInternalFormat::Depth32FStencil8:
            case TextureInternalFormat::DepthStencil:
@@ -251,6 +251,8 @@ namespace MobileGL {
                return TextureInternalFormat::Depth24Stencil8;
            case GL_DEPTH32F_STENCIL8:
                return TextureInternalFormat::Depth32FStencil8;
+            case GL_STENCIL_INDEX8:
+                return TextureInternalFormat::StencilIndex8;
            case GL_DEPTH_COMPONENT:
                return TextureInternalFormat::DepthComponent;
            case GL_DEPTH_STENCIL:
@@ -233,6 +233,8 @@ namespace MobileGL {
                return GL_DEPTH24_STENCIL8;
            case TextureInternalFormat::Depth32FStencil8:
                return GL_DEPTH32F_STENCIL8;
+            case TextureInternalFormat::StencilIndex8:
+                return GL_STENCIL_INDEX8;
            case TextureInternalFormat::DepthComponent32:
                return GL_DEPTH_COMPONENT32;
            case TextureInternalFormat::DepthStencil:
@@ -234,6 +234,8 @@ namespace MobileGL {
                return "Depth24Stencil8";
            case TextureInternalFormat::Depth32FStencil8:
                return "Depth32FStencil8";
+            case TextureInternalFormat::StencilIndex8:
+                return "StencilIndex8";
            case TextureInternalFormat::Red:
                return "Red";
            case TextureInternalFormat::RG:
@@ -25,6 +25,19 @@ namespace MobileGL {
            case GL_TRIANGLE_FAN:
                return VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN;
            case GL_LINE_LOOP:
+                // DrawArrays/DrawElements rewrite line loops into closed indexed
+                // strips; entry points without that rewrite (instanced/indirect)
+                // degrade to an open strip, which only misses the closing segment.
+                MGLOG_W("GL_LINE_LOOP without index rewrite; drawing as LINE_STRIP");
+                return VK_PRIMITIVE_TOPOLOGY_LINE_STRIP;
+            case GL_LINES_ADJACENCY:
+                return VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY;
+            case GL_LINE_STRIP_ADJACENCY:
+                return VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY;
+            case GL_TRIANGLES_ADJACENCY:
+                return VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY;
+            case GL_TRIANGLE_STRIP_ADJACENCY:
+                return VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY;
            default:
                MGLOG_W("Unrecognized primitive topology");
                return VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST;
@@ -226,6 +226,8 @@ namespace MobileGL {
                return VK_FORMAT_D24_UNORM_S8_UINT;
            case TextureInternalFormat::Depth32FStencil8:
                return VK_FORMAT_D32_SFLOAT_S8_UINT;
+            case TextureInternalFormat::StencilIndex8:
+                return VK_FORMAT_S8_UINT;
            case TextureInternalFormat::DepthComponent32:
                return VK_FORMAT_D32_SFLOAT;
            case TextureInternalFormat::DepthStencil:
@@ -18,6 +18,7 @@ namespace MobileGL {
            case TextureInternalFormat::Red: // UNorm8 shadow layout
            case TextureInternalFormat::R8Snorm:
            case TextureInternalFormat::R8I:
+            case TextureInternalFormat::StencilIndex8:
            case TextureInternalFormat::R8UI:
                return 1;

@@ -43,8 +44,11 @@ namespace MobileGL {
            case TextureInternalFormat::SRGB8:
            case TextureInternalFormat::RGB8I:
            case TextureInternalFormat::RGB8UI:
-            case TextureInternalFormat::DepthComponent24:
                return 3;
+            // Canonical depth shadow is a full 32-bit unorm word (see PixelStoreProcessor),
+            // converted at upload to the image's own 24/32-bit layout.
+            case TextureInternalFormat::DepthComponent24:
+                return 4;

            case TextureInternalFormat::RGBA2:
            case TextureInternalFormat::RGBA4:
@@ -91,6 +95,9 @@ namespace MobileGL {
            case TextureInternalFormat::RG32F:
            case TextureInternalFormat::RG32I:
            case TextureInternalFormat::RG32UI:
+            // Shadow bytes hold the GL_FLOAT_32_UNSIGNED_INT_24_8_REV wire format
+            // (float depth + a word whose low 8 bits are stencil), 8 bytes/texel.
+            case TextureInternalFormat::Depth32FStencil8:
                return 8;

            case TextureInternalFormat::RGB32F:
@@ -101,7 +108,6 @@ namespace MobileGL {
            case TextureInternalFormat::RGBA32F:
            case TextureInternalFormat::RGBA32I:
            case TextureInternalFormat::RGBA32UI:
-            case TextureInternalFormat::Depth32FStencil8:
                return 16;

            case TextureInternalFormat::R11FG11FB10F:
@@ -253,8 +259,10 @@ namespace MobileGL {
            case TexturePixelDataType::UnsignedInt101111Rev:
            case TexturePixelDataType::UnsignedInt5999Rev:
            case TexturePixelDataType::UnsignedInt248:
-            case TexturePixelDataType::Float32UnsignedInt248Rev:
                return 4;
+            case TexturePixelDataType::Float32UnsignedInt248Rev:
+                // A 32-bit float depth word followed by a 32-bit word holding stencil.
+                return 8;
            default:
                return 0;
            }
@@ -501,9 +509,16 @@ namespace MobileGL {
                s.Depth = 32;
                s.Stencil = 8;
                break;
+            case TextureInternalFormat::StencilIndex8:
+                s.Stencil = 8;
+                break;
+            case TextureInternalFormat::Unknown:
+                // Queried for attachments that have no storage yet (e.g. framebuffer
+                // parameter queries on the initial state); every size stays 0.
+                break;
            default:
-                MOBILEGL_ASSERT(false, "Unimplemented internal format in GetComponentSizesForInternalFormat: %d",
-                                static_cast<Int>(internal));
+                MGLOG_W("Unimplemented internal format in GetComponentSizesForInternalFormat: %d",
+                        static_cast<Int>(internal));
                break;
            }

@@ -220,11 +220,14 @@ namespace MobileGL {
                auto result = ParseShaderSource(lang, shaderType, source, attrib.flags);
                if (result) return result;

-                // Legacy desktop sources are normalized to "#version 330 core", which parses under
-                // stricter rules than the 460 they used to be forced to: a shader declaring 330 while
-                // using e.g. layout(binding=...) without the matching #extension line compiles on real
-                // drivers but is rejected here. Retry once at 460 before reporting failure; a genuinely
-                // broken shader fails both attempts and keeps its original diagnostics.
+                // Legacy desktop sources are normalized to "#version 330 core" (with a marker on the
+                // directive), which parses under stricter rules than the 460 they used to be forced
+                // to: a shader declaring 110-150 while using e.g. layout(binding=...) without the
+                // matching #extension line compiles on real drivers but is rejected here. Retry once
+                // at 460 before reporting failure; a genuinely broken shader fails both attempts and
+                // keeps its original diagnostics. Application-declared 330+ sources carry no marker
+                // and keep their declared version's strict rules (the GL CTS negative-compile cases
+                // depend on that).
                String retrySource = source;
                if (!MG_Util::ShaderTranspiler::RetargetLegacyVersionDirectiveTo460(retrySource)) {
                    return result;
@@ -688,6 +688,10 @@ namespace {
        return info;
    }

+    // Stamped onto the normalized directive when a legacy (or absent) desktop
+    // version was rewritten to 330; consumed by RetargetLegacyVersionDirectiveTo460.
+    constexpr const char* kNormalizedLegacyMarker = "/*mobilegl-normalized-legacy*/";
+
    MobileGL::String GetNormalizedVersionDirective(const ShaderLanguageInfo& info) {
        if (info.profile == MobileGL::ShaderProfile::ES) {
            // Preserve the pre-existing behavior for standard lowercase "es" directives. MobileGL's Vulkan
@@ -702,9 +706,23 @@ namespace {
            return "#version 460 compatibility\n";
        }

+        // An explicitly declared modern core version keeps its number: the GL CTS
+        // negative-compile cases (reserved names, layout-qualifier forms, missing
+        // overloads) rely on the declared version's rules, and raising it would
+        // silently legalize them. gpu_shader5 opt-ins keep the 460 escalation -
+        // Vulkan glslang's ARB_gpu_shader5 support is not complete enough alone.
+        if (info.hasValidVersionDirective && info.version >= 330 && !info.enablesGpuShader5) {
+            return "#version " + std::to_string(info.version) + " core\n";
+        }
+
        const bool useLegacyDesktopVersion =
            info.version < 400 && !info.enablesGpuShader5;
-        return useLegacyDesktopVersion ? "#version 330 core\n" : "#version 460 core\n";
+        // The trailing marker records that this 330 came from a legacy declaration
+        // (or none at all), so the compile-failure retry may re-raise it to 460.
+        // An application's own "#version 330" never carries it and keeps strict
+        // 3.30 semantics.
+        return useLegacyDesktopVersion ? MobileGL::String("#version 330 core ") + kNormalizedLegacyMarker + "\n"
+                                       : "#version 460 core\n";
    }

    void NormalizeVersionDirective(MobileGL::String& source, const ShaderLanguageInfo& info) {
@@ -1306,12 +1324,123 @@ namespace MobileGL {
                // Only the set NormalizeVersionDirective downgraded: desktop core below 400. ES and
                // compatibility shaders keep whatever they declared.
                if (info.profile != ShaderProfile::Core || info.version >= 400) return false;
+                // Only rescue MobileGL's own legacy normalization (marked on the directive line).
+                // An application-declared "#version 330" keeps strict 3.30 semantics: raising it
+                // would re-legalize the CTS negative-compile cases (reserved names, arrays of
+                // arrays, missing overloads).
+                SizeT lineEnd = source.find('\n', info.versionDirectiveStart);
+                if (lineEnd == MobileGL::String::npos) {
+                    lineEnd = source.size();
+                }
+                const SizeT markerPos = source.find(kNormalizedLegacyMarker, info.versionDirectiveStart);
+                if (markerPos == MobileGL::String::npos || markerPos > lineEnd) {
+                    return false;
+                }

                source.replace(info.versionDirectiveStart, info.versionDirectiveEnd - info.versionDirectiveStart,
                               "#version 460 core\n");
                return true;
            }

+            std::optional<String> FindReservedIdentifierViolation(const String& source) {
+                // Reserved anywhere; glslang accepts them as plain identifiers.
+                static constexpr const char* kAlwaysReserved[] = {
+                    "image1DShadow",
+                    "image2DShadow",
+                    "image1DArrayShadow",
+                    "image2DArrayShadow",
+                };
+                // Keywords legal only inside a layout(...) qualifier list.
+                static constexpr const char* kLayoutOnlyKeywords[] = {
+                    "packed",
+                    "row_major",
+                };
+
+                const auto isIdentChar = [](char c) {
+                    return (c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z') || (c >= '0' && c <= '9') || c == '_';
+                };
+
+                const SizeT length = source.size();
+                SizeT i = 0;
+                Int layoutParenDepth = 0;   // >0 while inside layout(...)
+                Bool pendingLayoutParen = false; // saw "layout", awaiting its '('
+                while (i < length) {
+                    const char c = source[i];
+                    // Comments.
+                    if (c == '/' && i + 1 < length && source[i + 1] == '/') {
+                        while (i < length && source[i] != '\n') ++i;
+                        continue;
+                    }
+                    if (c == '/' && i + 1 < length && source[i + 1] == '*') {
+                        i += 2;
+                        while (i + 1 < length && !(source[i] == '*' && source[i + 1] == '/')) ++i;
+                        i = (i + 1 < length) ? i + 2 : length;
+                        continue;
+                    }
+                    // Preprocessor lines stay out of scope (macro names may shadow anything).
+                    if (c == '#' && (i == 0 || source[i - 1] == '\n' ||
+                                     source.find_last_not_of(" \t", i - 1) == MobileGL::String::npos ||
+                                     source[source.find_last_not_of(" \t", i - 1)] == '\n')) {
+                        while (i < length && source[i] != '\n') {
+                            if (source[i] == '\\' && i + 1 < length && source[i + 1] == '\n') ++i;
+                            ++i;
+                        }
+                        continue;
+                    }
+                    if (c == '(') {
+                        if (pendingLayoutParen) {
+                            layoutParenDepth = 1;
+                            pendingLayoutParen = false;
+                        } else if (layoutParenDepth > 0) {
+                            ++layoutParenDepth;
+                        }
+                        ++i;
+                        continue;
+                    }
+                    if (c == ')') {
+                        if (layoutParenDepth > 0) --layoutParenDepth;
+                        ++i;
+                        continue;
+                    }
+                    if (c == ' ' || c == '\t' || c == '\r' || c == '\n') {
+                        ++i;
+                        continue;
+                    }
+                    if (isIdentChar(c) && !(c >= '0' && c <= '9')) {
+                        const SizeT start = i;
+                        while (i < length && isIdentChar(source[i])) ++i;
+                        const StringView word(source.data() + start, i - start);
+                        if (word == "layout") {
+                            pendingLayoutParen = true;
+                            continue;
+                        }
+                        pendingLayoutParen = false;
+                        for (const char* reserved : kAlwaysReserved) {
+                            if (word == reserved) {
+                                return String("ERROR: reserved identifier '") + reserved + "' may not be used.";
+                            }
+                        }
+                        if (layoutParenDepth == 0) {
+                            for (const char* keyword : kLayoutOnlyKeywords) {
+                                if (word == keyword) {
+                                    return String("ERROR: '") + keyword +
+                                        "' is a keyword and may not be used as an identifier.";
+                                }
+                            }
+                        }
+                        continue;
+                    }
+                    if (isIdentChar(c)) { // digit-led token: skip the whole number/identifier tail
+                        while (i < length && isIdentChar(source[i])) ++i;
+                        pendingLayoutParen = false;
+                        continue;
+                    }
+                    pendingLayoutParen = false;
+                    ++i;
+                }
+                return std::nullopt;
+            }
+
        } // namespace ShaderTranspiler
    } // namespace MG_Util
 } // namespace MobileGL
@@ -40,6 +40,11 @@ namespace MobileGL {
            // uses 420-era syntax without the matching #extension line, which real drivers tend to
            // accept - can be retried instead of failing to compile.
            Bool RetargetLegacyVersionDirectiveTo460(String& source);
+
+            // GLSL reserves a few names glslang happily accepts as identifiers ("packed",
+            // "row_major" outside a layout(...) list, the image*Shadow family). Returns the
+            // compile-error text for the first violation, or nullopt for a clean source.
+            std::optional<String> FindReservedIdentifierViolation(const String& source);
        } // namespace ShaderTranspiler
    } // namespace MG_Util
 } // namespace MobileGL
@@ -95,6 +95,7 @@ namespace MobileGL::MG_Util::PixelStoreProcessor {
            Int32,
            Half,
            Float32,
+            UNorm32, // 32-bit fixed-point depth shadow
        };

        struct InternalShadowLayout {
@@ -123,6 +124,21 @@ namespace MobileGL::MG_Util::PixelStoreProcessor {

        Bool GetInternalShadowLayout(TextureInternalFormat internal, InternalShadowLayout& out) {
            switch (internal) {
+            // Depth shadows follow TextureFormatProcessor::NormalizePixelFormat: 16-bit
+            // unorm for DEPTH_COMPONENT16, 32-bit unorm for the 24/32-bit fixed-point
+            // depths, float for DEPTH_COMPONENT32F.
+            case TextureInternalFormat::DepthComponent16:
+                out = {1, ShadowComponent::UNorm16, false};
+                return true;
+            case TextureInternalFormat::DepthComponent24:
+            case TextureInternalFormat::DepthComponent32:
+            case TextureInternalFormat::DepthComponent:
+                out = {1, ShadowComponent::UNorm32, false};
+                return true;
+            case TextureInternalFormat::DepthComponent32F:
+                out = {1, ShadowComponent::Float32, false};
+                return true;
+
            case TextureInternalFormat::R8:
            case TextureInternalFormat::Red:     out = {1, ShadowComponent::UNorm8, false}; return true;
            case TextureInternalFormat::RG8:
@@ -299,8 +315,10 @@ namespace MobileGL::MG_Util::PixelStoreProcessor {
            case TextureInputFormat::RGBAInteger:  out = {{0, 1, 2, 3}, 4, true};     return true;
            case TextureInputFormat::BGRA:         out = {{2, 1, 0, 3}, 4, false};    return true;
            case TextureInputFormat::BGRAInteger:  out = {{2, 1, 0, 3}, 4, true};     return true;
+            // A depth value converts like a single normalized/float channel.
+            case TextureInputFormat::DepthComponent: out = {{0, -1, -1, -1}, 1, false}; return true;
            default:
-                return false; // depth / stencil / unknown
+                return false; // stencil / packed depth-stencil / unknown
            }
        }

@@ -346,8 +364,7 @@ namespace MobileGL::MG_Util::PixelStoreProcessor {
                out = isInteger ? ShadowComponent::Int16 : ShadowComponent::SNorm16;
                return true;
            case TexturePixelDataType::UnsignedInt:
-                if (!isInteger) return false; // no 32-bit normalized shadow layout
-                out = ShadowComponent::UInt32;
+                out = isInteger ? ShadowComponent::UInt32 : ShadowComponent::UNorm32;
                return true;
            case TexturePixelDataType::Int:
                if (!isInteger) return false;
@@ -617,6 +634,12 @@ namespace MobileGL::MG_Util::PixelStoreProcessor {
            case ShadowComponent::Float32:
                Memcpy(dst, &v, sizeof(v));
                break;
+            case ShadowComponent::UNorm32: {
+                const auto out = static_cast<Uint32>(
+                    std::llround(static_cast<double>(std::clamp(v, 0.0f, 1.0f)) * 4294967295.0));
+                Memcpy(dst, &out, sizeof(out));
+                break;
+            }
            default:
                break; // integer components never reach the float encoder
            }
@@ -771,6 +794,64 @@ namespace MobileGL::MG_Util::PixelStoreProcessor {
        const Int effectiveHeight = (params.ImageHeight > 0) ? params.ImageHeight : height;
        const SizeT inputRowStride = CalculateRowStride(effectiveWidth, pixelSize, params.Alignment);

+        // GL_DEPTH_COMPONENT client data may populate a packed depth-stencil internal
+        // format (the stencil half becomes zero); the generic channel converter cannot
+        // express the packed shadow words, so convert here.
+        const Bool packedDepthStencilInternal = targetInternalFormat == TextureInternalFormat::Depth24Stencil8 ||
+            targetInternalFormat == TextureInternalFormat::DepthStencil ||
+            targetInternalFormat == TextureInternalFormat::Depth32FStencil8;
+        if (!isBitmap && packedDepthStencilInternal && textureInputFormat == TextureInputFormat::DepthComponent &&
+            (inputDataType == TexturePixelDataType::Float || inputDataType == TexturePixelDataType::UnsignedInt ||
+             inputDataType == TexturePixelDataType::UnsignedShort)) {
+            const Bool floatShadow = targetInternalFormat == TextureInternalFormat::Depth32FStencil8;
+            const SizeT outPixelSize = floatShadow ? 8 : 4;
+            outSize = static_cast<SizeT>(width) * height * std::max(depth, 1) * outPixelSize;
+            Uint8* outputPixels = static_cast<Uint8*>(malloc(outSize));
+            if (!outputPixels) {
+                outSize = 0;
+                return nullptr;
+            }
+            const Uint8* srcBase = static_cast<const Uint8*>(inputPixels) +
+                static_cast<SizeT>(params.SkipImages) * static_cast<SizeT>(effectiveHeight) * inputRowStride +
+                static_cast<SizeT>(params.SkipRows) * inputRowStride +
+                static_cast<SizeT>(params.SkipPixels) * pixelSize;
+            Uint8* dst = outputPixels;
+            for (Int z = 0; z < std::max(depth, 1); ++z) {
+                for (Int y = 0; y < height; ++y) {
+                    const Uint8* srcRow = srcBase +
+                        static_cast<SizeT>(z) * static_cast<SizeT>(effectiveHeight) * inputRowStride +
+                        static_cast<SizeT>(y) * inputRowStride;
+                    for (Int x = 0; x < width; ++x) {
+                        Float depthValue = 0.0f;
+                        if (inputDataType == TexturePixelDataType::Float) {
+                            Memcpy(&depthValue, srcRow + static_cast<SizeT>(x) * 4, sizeof(depthValue));
+                        } else if (inputDataType == TexturePixelDataType::UnsignedInt) {
+                            Uint32 raw = 0;
+                            Memcpy(&raw, srcRow + static_cast<SizeT>(x) * 4, sizeof(raw));
+                            depthValue = static_cast<Float>(static_cast<double>(raw) / 4294967295.0);
+                        } else {
+                            Uint16 raw = 0;
+                            Memcpy(&raw, srcRow + static_cast<SizeT>(x) * 2, sizeof(raw));
+                            depthValue = static_cast<Float>(raw) / 65535.0f;
+                        }
+                        if (floatShadow) {
+                            const Uint32 stencilWord = 0;
+                            Memcpy(dst, &depthValue, sizeof(depthValue));
+                            Memcpy(dst + 4, &stencilWord, sizeof(stencilWord));
+                            dst += 8;
+                        } else {
+                            const Uint32 depth24 = static_cast<Uint32>(
+                                std::llround(static_cast<double>(std::clamp(depthValue, 0.0f, 1.0f)) * 16777215.0));
+                            const Uint32 word = depth24 << 8;
+                            Memcpy(dst, &word, sizeof(word));
+                            dst += 4;
+                        }
+                    }
+                }
+            }
+            return outputPixels;
+        }
+
        UnpackConversionSpec conversion{};
        const Bool needConversion =
            !isBitmap && GetUnpackConversionSpec(targetInternalFormat, textureInputFormat, inputDataType, conversion);
@@ -972,6 +1053,11 @@ namespace MobileGL::MG_Util::PixelStoreProcessor {
                Memcpy(&v, p, sizeof(v));
                return v;
            }
+            case ShadowComponent::UNorm32: {
+                Uint32 v;
+                Memcpy(&v, p, sizeof(v));
+                return static_cast<Float>(static_cast<double>(v) / 4294967295.0);
+            }
            default:
                return 0.0f;
            }
@@ -1,5 +1,5 @@
 /*-------------------------------------------------------------------------
- * dEQP platform port for MobileGL on Android
+ * dEQP platform port for MobileGL (Android and desktop Linux)
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -28,14 +28,17 @@
 *  - A real surface is always created. MobileGL rejects EGL_NO_SURFACE with
 *    EGL_BAD_MATCH, and --deqp-surface-type=fbo asks the platform for
 *    SURFACETYPE_DONT_CARE, so "no surface" is not an option.
- *  - Window surfaces are backed by an AImageReader rather than an Activity,
- *    which is what lets the suite run as a plain adb-shell binary. DirectVulkan
- *    needs this: its pbuffer path requires VK_EXT_headless_surface, which
- *    Adreno's Android driver does not expose.
+ *  - On Android, window surfaces are backed by an AImageReader rather than an
+ *    Activity, which is what lets the suite run as a plain adb-shell binary.
+ *    DirectVulkan needs this: its pbuffer path requires VK_EXT_headless_surface,
+ *    which Adreno's Android driver does not expose.
+ *  - On desktop Linux, only pbuffer surfaces are offered. DirectVulkan's
+ *    pbuffer path works there because desktop Vulkan loaders expose
+ *    VK_EXT_headless_surface.
 *
 * Environment:
 *   MOBILEGL_CTS_LIB      path/soname of the MobileGL library (default libMobileGL.so)
- *   MOBILEGL_CTS_SURFACE  "window" (default) or "pbuffer"
+ *   MOBILEGL_CTS_SURFACE  "window" (Android default) or "pbuffer" (desktop default/only)
 *   MOBILEGL_BACKEND_TYPE read by MobileGL itself; set it before launching
 *//*--------------------------------------------------------------------*/

@@ -58,9 +61,11 @@
 #include "tcuPlatform.hpp"
 #include "tcuRenderTarget.hpp"

+#if defined(__ANDROID__)
 #include <android/hardware_buffer.h>
 #include <android/native_window.h>
 #include <media/NdkImageReader.h>
+#endif

 using std::string;
 using std::vector;
@@ -88,13 +93,24 @@ static string getLibraryName(void)
    return (env && env[0]) ? string(env) : string("libMobileGL.so");
 }

-//! Window surfaces default on: they are the only kind DirectVulkan can use.
+#if defined(__ANDROID__)
+//! Window surfaces default on: they are the only kind DirectVulkan can use
+//! on Android (its pbuffer path needs VK_EXT_headless_surface).
 static bool useWindowSurface(void)
 {
    const char *env = std::getenv("MOBILEGL_CTS_SURFACE");
    return !(env && string(env) == "pbuffer");
 }
+#else
+//! Desktop: pbuffer only. VK_EXT_headless_surface is available there and no
+//! Activity-free native window abstraction exists.
+static bool useWindowSurface(void)
+{
+    return false;
+}
+#endif

+#if defined(__ANDROID__)
 /*--------------------------------------------------------------------*//*!
 * \brief A real ANativeWindow with no Activity behind it.
 *
@@ -160,6 +176,12 @@ private:
    AImageReader *m_reader;
    ANativeWindow *m_window;
 };
+#else
+//! Never instantiated on desktop; keeps EglRenderContext's member deletable.
+class ImageReaderWindow
+{
+};
+#endif

 class GetProcFuncLoader : public glw::FunctionLoader
 {
@@ -352,6 +374,7 @@ EglRenderContext::EglRenderContext(const glu::RenderConfig &config, const tcu::C

    if (wantWindow)
    {
+#if defined(__ANDROID__)
        m_window = new ImageReaderWindow(width, height);

        eglw::EGLint visualId = 0;
@@ -361,6 +384,9 @@ EglRenderContext::EglRenderContext(const glu::RenderConfig &config, const tcu::C
        m_eglSurface = m_egl.createWindowSurface(m_eglDisplay, eglConfig,
                                                 (eglw::EGLNativeWindowType)m_window->getWindow(), nullptr);
        EGLU_CHECK_MSG(m_egl, "eglCreateWindowSurface()");
+#else
+        throw tcu::NotSupportedError("Window surfaces are not supported by the desktop MobileGL platform");
+#endif
    }
    else
    {
@@ -0,0 +1,208 @@
+#!/usr/bin/env python
+"""Drive a glcts run on the local host, resuming across crashes.
+
+Local-host counterpart of run_cts.py: MobileGL crashes on some cases and glcts
+takes the whole process down with it, so a single invocation stops at the first
+crash. This runner re-invokes glcts with only the cases that have no result
+yet, records the case that was open when the process died as "Crash" (or
+"Hang" on a timeout), and repeats until the list is exhausted.
+
+Usage:
+  python run_cts_local.py --backend DirectVulkan \\
+      --glcts <path-to-glcts-binary> --lib <path-to-libMobileGL.so> \\
+      --caselist <mustpass.txt> --outdir <dir> [--env K=V ...]
+"""
+
+import argparse
+import glob
+import os
+import re
+import resource
+import subprocess
+import sys
+import time
+
+CASE_START = re.compile(r"^#beginTestCaseResult\s+(\S+)")
+CASE_END = re.compile(r"^#endTestCaseResult")
+CASE_TERM = re.compile(r"^#terminateTestCaseResult")
+
+
+def completed_cases(qpa_path):
+    """Return (finished_case_names, last_started_case_or_None)."""
+    finished = []
+    current = None
+    if not os.path.exists(qpa_path):
+        return finished, None
+    with open(qpa_path, "r", encoding="utf-8", errors="replace") as fh:
+        for line in fh:
+            m = CASE_START.match(line)
+            if m:
+                current = m.group(1)
+                continue
+            if current is not None and (CASE_END.match(line) or CASE_TERM.match(line)):
+                finished.append(current)
+                current = None
+    return finished, current
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--backend", required=True, choices=["DirectGLES", "DirectVulkan"])
+    ap.add_argument("--glcts", required=True, help="path to the glcts binary")
+    ap.add_argument("--lib", required=True, help="path to libMobileGL.so")
+    ap.add_argument("--caselist", required=True)
+    ap.add_argument("--outdir", required=True)
+    ap.add_argument("--surface", default="fbo", help="--deqp-surface-type value")
+    # Without an explicit size, dEQP's FboRenderContext sizes the wrapper FBO to
+    # GL_MAX_RENDERBUFFER_SIZE (16384^2 here) and size-derived test allocations
+    # explode (a 4-sample 16K depth texture alone is 4 GiB).
+    ap.add_argument("--surface-size", type=int, default=256,
+                    help="--deqp-surface-width/height value")
+    ap.add_argument("--max-rounds", type=int, default=4000)
+    ap.add_argument("--max-empty-streak", type=int, default=64,
+                    help="abort after this many consecutive chunks that produce no log at all")
+    ap.add_argument("--chunk-timeout", type=int, default=1800,
+                    help="seconds before killing one glcts invocation (a wedged case never returns)")
+    ap.add_argument("--skip-file", default=None,
+                    help="file of case names to exclude, e.g. cases known to wedge the host")
+    ap.add_argument("--env", action="append", default=[], metavar="K=V",
+                    help="extra environment variable for glcts (repeatable)")
+    args = ap.parse_args()
+
+    os.makedirs(args.outdir, exist_ok=True)
+    glcts = os.path.abspath(args.glcts)
+    lib = os.path.abspath(args.lib)
+    # glcts resolves its gl_cts data tree relative to the binary's directory.
+    workdir = os.path.dirname(glcts)
+
+    with open(args.caselist, "r", encoding="utf-8") as fh:
+        remaining = [l.strip() for l in fh if l.strip() and not l.strip().startswith("#")]
+
+    skipped = []
+    if args.skip_file and os.path.isfile(args.skip_file):
+        with open(args.skip_file, "r", encoding="utf-8") as fh:
+            skip = {l.strip() for l in fh if l.strip() and not l.strip().startswith("#")}
+        skipped = [c for c in remaining if c in skip]
+        remaining = [c for c in remaining if c not in skip]
+        print(f"[run_cts_local] skipping {len(skipped)} case(s) from {args.skip_file}")
+
+    total = len(remaining)
+    print(f"[run_cts_local] {args.backend}: {total} cases")
+
+    env = dict(os.environ)
+    env["MOBILEGL_BACKEND_TYPE"] = args.backend
+    env["MOBILEGL_CTS_LIB"] = lib
+    for kv in args.env:
+        k, _, v = kv.partition("=")
+        env[k] = v
+
+    crashed = []
+    hung = []
+    done = set()
+    chunk = 0
+    started = time.time()
+    empty_streak = 0
+
+    # Resume: results in chunk files from an interrupted run still count. The
+    # case that was open when that run died is re-tried rather than assumed bad.
+    prior_chunks = sorted(glob.glob(os.path.join(args.outdir, "chunk*.qpa")))
+    for prior in prior_chunks:
+        finished, _ = completed_cases(prior)
+        done.update(finished)
+    if prior_chunks:
+        chunk = int(re.search(r"chunk(\d+)\.qpa$", prior_chunks[-1]).group(1)) + 1
+        remaining = [c for c in remaining if c not in done]
+        print(f"[run_cts_local] resuming: {len(done)} case(s) already measured, "
+              f"{len(remaining)} to go")
+
+    while remaining and chunk < args.max_rounds:
+        listfile = os.path.abspath(os.path.join(args.outdir, "remaining.txt"))
+        with open(listfile, "w", encoding="utf-8", newline="\n") as fh:
+            fh.write("\n".join(remaining) + "\n")
+
+        qpa = os.path.abspath(os.path.join(args.outdir, f"chunk{chunk:04d}.qpa"))
+        cmd = [
+            glcts,
+            f"--deqp-caselist-file={listfile}",
+            f"--deqp-surface-type={args.surface}",
+            f"--deqp-surface-width={args.surface_size}",
+            f"--deqp-surface-height={args.surface_size}",
+            "--deqp-terminate-on-device-lost=disable",
+            # A wedged case aborts the process instead of stalling the chunk;
+            # the runner then records it as Crash and resumes past it.
+            "--deqp-watchdog=enable",
+            "--deqp-log-images=disable",
+            "--deqp-log-shader-sources=disable",
+            f"--deqp-log-filename={qpa}",
+        ]
+        timed_out = False
+        try:
+            # RLIMIT_CORE=0: MobileGL asserts abort with a core dump, and writing
+            # a multi-GB glcts core image after every crash dominates wall time.
+            subprocess.run(cmd, cwd=workdir, env=env, timeout=args.chunk_timeout,
+                           stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
+                           start_new_session=True,
+                           preexec_fn=lambda: resource.setrlimit(resource.RLIMIT_CORE, (0, 0)))
+        except subprocess.TimeoutExpired:
+            timed_out = True
+            print(f"[run_cts_local] chunk {chunk:04d} timed out after {args.chunk_timeout}s",
+                  file=sys.stderr)
+
+        finished, in_flight = completed_cases(qpa)
+        for c in finished:
+            done.add(c)
+
+        progressed = len(finished)
+        if progressed > 0:
+            empty_streak = 0
+        if in_flight is not None:
+            if timed_out:
+                print(f"[run_cts_local] HANG in {in_flight} - quarantining it")
+                hung.append(in_flight)
+            else:
+                crashed.append(in_flight)
+            done.add(in_flight)
+            progressed += 1
+        elif progressed == 0:
+            empty_streak += 1
+            if empty_streak >= args.max_empty_streak:
+                print(f"[run_cts_local] ABORTING: {empty_streak} consecutive chunks produced no "
+                      f"output. Something systemic is wrong; refusing to label the rest of the "
+                      f"suite as crashes.", file=sys.stderr)
+                break
+            victim = remaining[0]
+            label = "Hang" if timed_out else "Crash"
+            print(f"[run_cts_local] no output at all; recording {victim} as {label}")
+            (hung if timed_out else crashed).append(victim)
+            done.add(victim)
+            progressed = 1
+
+        remaining = [c for c in remaining if c not in done]
+        elapsed = time.time() - started
+        print(
+            f"[run_cts_local] chunk {chunk:04d}: +{progressed} (done {len(done)}/{total}, "
+            f"crashes {len(crashed)}, hangs {len(hung)}, {elapsed / 60:.1f} min)"
+        )
+        chunk += 1
+
+    with open(os.path.join(args.outdir, "crashed.txt"), "w", encoding="utf-8", newline="\n") as fh:
+        fh.write("\n".join(crashed) + ("\n" if crashed else ""))
+    with open(os.path.join(args.outdir, "hung.txt"), "w", encoding="utf-8", newline="\n") as fh:
+        fh.write("\n".join(hung) + ("\n" if hung else ""))
+    with open(os.path.join(args.outdir, "unrun.txt"), "w", encoding="utf-8", newline="\n") as fh:
+        fh.write("\n".join(remaining) + ("\n" if remaining else ""))
+    if skipped:
+        with open(os.path.join(args.outdir, "skipped.txt"), "w", encoding="utf-8", newline="\n") as fh:
+            fh.write("\n".join(skipped) + "\n")
+
+    if remaining:
+        print(f"[run_cts_local] WARNING: {len(remaining)} cases were never run (see unrun.txt)",
+              file=sys.stderr)
+    print(f"[run_cts_local] finished: {len(done)}/{total} cases, {len(crashed)} crashes, "
+          f"{len(hung)} hangs, {chunk} invocations")
+    print(f"[run_cts_local] qpa chunks in {args.outdir}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
@@ -21,6 +21,7 @@ CTS_TOOLS = os.path.dirname(HERE)
 COPIES = [
    (os.path.join(CTS_TOOLS, "platform"), "framework/platform/mobilegl", None),
    (os.path.join(CTS_TOOLS, "targets"), "targets/mobilegl", ["mobilegl.cmake", "ndk-modern.cmake"]),
+    (os.path.join(CTS_TOOLS, "targets"), "targets/mobilegl-desktop", ["mobilegl-desktop.cmake"]),
 ]


@@ -0,0 +1,29 @@
+#-------------------------------------------------------------------------
+# VK-GL-CTS target: MobileGL on desktop Linux
+#
+# Builds glcts as a normal host executable that reaches OpenGL exclusively
+# through libMobileGL.so, loaded at runtime via the eglw dynamic wrapper.
+# Nothing here links libEGL or libGL: a conformance result must be
+# unambiguously MobileGL's, never the system GL stack's.
+#
+# The platform port only offers pbuffer surfaces on desktop; DirectVulkan's
+# pbuffer path works because desktop Vulkan exposes VK_EXT_headless_surface.
+#-------------------------------------------------------------------------
+
+message("*** Using MobileGL desktop target")
+
+set(DEQP_TARGET_NAME "MobileGL")
+
+# EGL comes from libMobileGL.so via the eglw dynamic wrapper, so the support
+# flag is on but no import library is supplied.
+set(DEQP_SUPPORT_EGL ON)
+set(DEQP_EGL_LIBRARIES)
+set(DEQP_GLES2_LIBRARIES)
+set(DEQP_GLES3_LIBRARIES)
+
+set(TCUTIL_PLATFORM_SRCS
+	mobilegl/tcuMobileGLPlatform.cpp
+	mobilegl/tcuMobileGLPlatform.hpp
+	)
+
+list(APPEND TCUTIL_PLATFORM_LIBS dl pthread)