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

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

CMakeLists.txt

@@ -455,8 +455,21 @@ if (ANDROID)
 endif()
 
 if (APPLE AND NOT MOBILEGL_IOS)
+    # MobileGL statically embeds glslang, SPIRV-Tools, and SPIRV-Cross. When
+    # this dylib is injected with DYLD_INSERT_LIBRARIES, exporting those C++
+    # symbols interposes incompatible copies embedded by host libraries such
+    # as shaderc. Keep only the public GL/EGL/CGL loader surface globally
+    # visible; GetProcAddress can still return pointers to hidden internals.
+    set(MOBILEGL_MACOS_EXPORTED_SYMBOLS
+        "${CMAKE_CURRENT_SOURCE_DIR}/MobileGL/MG_Impl/DyldInterpose/ExportedSymbols.txt")
+    target_link_options(${CMAKE_PROJECT_NAME} PRIVATE
+        "LINKER:-exported_symbols_list,${MOBILEGL_MACOS_EXPORTED_SYMBOLS}")
+    set_property(TARGET ${CMAKE_PROJECT_NAME} APPEND PROPERTY
+        LINK_DEPENDS "${MOBILEGL_MACOS_EXPORTED_SYMBOLS}")
+
     target_link_libraries(${CMAKE_PROJECT_NAME} PUBLIC
         "-framework Cocoa"
+        "-framework CoreVideo"
         "-framework QuartzCore"
         "-framework Foundation"
         "-framework OpenGL"
@@ -464,6 +477,7 @@ if (APPLE AND NOT MOBILEGL_IOS)
     if(TARGET ${CMAKE_PROJECT_NAME}_s)
         target_link_libraries(${CMAKE_PROJECT_NAME}_s PUBLIC
             "-framework Cocoa"
+            "-framework CoreVideo"
             "-framework QuartzCore"
             "-framework Foundation"
             "-framework OpenGL"

MobileGL/Init.cpp

@@ -107,9 +107,11 @@ namespace MobileGL {
     // (EGL/WGL/CGL): initialization happens lazily on the first entry point
     // via EnsureInitialized(), and full teardown happens deterministically
     // when the last EGL display is terminated with nothing current (EGLImpl
-    // calls Destroy()). There is intentionally no static constructor, no
+    // calls Destroy()). There is intentionally no backend-initializing static
-    // static destructor, and no DllMain: the global singletons use
+    // constructor, no static destructor, and no DllMain: the global singletons
-    // leak-at-exit storage (see GlobalObjects.cpp), so a process that exits
+    // use leak-at-exit storage (see GlobalObjects.cpp), so a process that exits
     // without eglTerminate simply leaks them to the OS instead of running
-    // backend destructors during static teardown.
+    // backend destructors during static teardown. macOS has a lightweight
+    // dyld constructor that installs NSOpenGL dispatch hooks only; full backend
+    // initialization still enters here from the first hooked CGL context.
 } // namespace MobileGL

MobileGL/Init.h

@@ -13,9 +13,10 @@ namespace MobileGL {
     void Initialize();
     // Thread-safe, idempotent, and re-entrant wrapper around Initialize().
     // Host layers (EGL/WGL/CGL entry points) call this lazily on first use so
-    // MobileGL's lifecycle never depends on ELF/DLL static constructors, and
+    // full backend initialization never depends on ELF/DLL static constructors,
-    // so a fresh init can follow a full Destroy() (e.g. after the last
+    // and so a fresh init can follow a full Destroy() (e.g. after the last
-    // eglTerminate).
+    // eglTerminate). The macOS dyld bootstrap installs only lightweight
+    // NSOpenGL method hooks.
     void EnsureInitialized();
     void Destroy();
 

MobileGL/MG_Backend/DirectVulkan/Renderer/FrameContext.cpp

@@ -150,12 +150,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     Bool FrameContext::TransitionToPresent(VkImage image, VkImageLayout oldLayout, VkImageLayout presentLayout) {
         auto& frame = GetCurrent();
-        if (frame.hasCommandBufferRecorded || frame.isCommandRecording || oldLayout == presentLayout ||
+        if (oldLayout == presentLayout || oldLayout == VK_IMAGE_LAYOUT_SHARED_PRESENT_KHR) {
-            oldLayout == VK_IMAGE_LAYOUT_SHARED_PRESENT_KHR) {
             return false;
         }
 
-        auto& commandBuffer = BeginCommandRecording();
+        // The barrier belongs in the frame's own recording. Bailing out because
+        // something was already recorded (the previous behaviour) dropped the
+        // transition entirely for every frame that never ran a default-framebuffer
+        // render pass - the only other thing that carries the image to
+        // PRESENT_SRC_KHR, via that pass's finalLayout - so the swapchain image was
+        // handed to the WSI still in the layout it was acquired in.
+        // A closed-but-unsubmitted buffer can only come from a submit that already
+        // failed (SubmitPendingCommandBuffer leaves the flag set on error), and
+        // appending to it is illegal while reopening would reset the frame's own
+        // commands away. The device is gone on that path anyway - stay silent-safe
+        // rather than trade a lost device for a barrier into a closed buffer.
+        if (frame.hasCommandBufferRecorded) {
+            MGLOG_E("TransitionToPresent: command buffer already closed; skipping the present barrier");
+            return false;
+        }
+
+        // Reopening a recording here would vkResetCommandBuffer this frame's own
+        // commands away, so append to the open one and let the caller close it.
+        const Bool openedRecording = !frame.isCommandRecording;
+        VkCommandBuffer commandBuffer = openedRecording ? BeginCommandRecording() : frame.commandBuffer;
 
         VkImageMemoryBarrier presentBarrier{};
         presentBarrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
@@ -174,7 +192,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         vkCmdPipelineBarrier(commandBuffer, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT, 0, 0,
                              nullptr, 0, nullptr, 1, &presentBarrier);
 
+        if (openedRecording) {
             EndCommandRecording();
+        }
         return true;
     }
 
@@ -227,12 +247,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         result = vkAcquireNextImageKHR(device, swapchain, timeout, frame.imageAvailableSemaphore, acquireFence,
                                        &outImageIndex);
-        if (result != VK_SUCCESS) {
+        // VK_SUBOPTIMAL_KHR is a success code: an image *was* acquired and
+        // imageAvailableSemaphore *will* be signaled. Bailing out on it skipped both
+        // the consumed-flag reset (leaving a stale "already consumed", so the next
+        // submit never waited on the pending signal) and the fence reset (leaving
+        // the slot's fence signaled for the next submit to reuse). Only a genuine
+        // failure - VK_ERROR_OUT_OF_DATE_KHR and friends, where nothing is acquired
+        // and nothing is signaled - skips the bookkeeping.
+        if (result != VK_SUCCESS && result != VK_SUBOPTIMAL_KHR) {
             return result;
         }
 
         frame.imageAvailableSemaphoreConsumed = false;
-        return vkResetFences(device, 1, &frame.imageInFlightFence);
+        const VkResult resetResult = vkResetFences(device, 1, &frame.imageInFlightFence);
+        // Hand the acquire's own code back so the caller can schedule a rebuild.
+        return resetResult == VK_SUCCESS ? result : resetResult;
     }
 
     Uint32 FrameContext::GetCurrentFrameIndex() const {

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.cpp

@@ -12,7 +12,10 @@
 #include "MG_Util/ShaderTranspiler/ShaderCompiler.h"
 #include "MG_Util/ShaderTranspiler/SpvcSession.h"
 #include "MG_Util/ShaderTranspiler/Types.h"
+#include <cmath>
+#include <cstdio>
 #include <cstring>
+#include <unordered_set>
 #include <spirv-tools/libspirv.h>
 #include <spirv-tools/optimizer.hpp>
 #include <source/opt/build_module.h>
@@ -923,11 +926,610 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             ProgramFactory::CompileOptionFlags m_transformFlags;
         };
 
+        // 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
+        // own default-framebuffer blit shader works around it with textureLod, but an
+        // application's shader (Minecraft's blit.fsh is `texture(InSampler, texCoord)`) cannot be
+        // edited - so rewrite the sample at the SPIR-V level instead.
+        //
+        // The rewrite is only requested for draws whose every sampler binding is clamped to one
+        // mip level, where explicit LOD 0 is exactly what the implicit form must already produce:
+        // lambda' = clamp(lambda + bias, minLod, maxLod) with minLod = maxLod = 0. Bias and MinLod
+        // operands are therefore dropped rather than translated.
+        class ForceExplicitLod0SamplePass final : public spvtools::opt::Pass {
+        public:
+            const char* name() const override { return "force-explicit-lod0-sample"; }
+
+            Status Process() override {
+                Bool isFragment = false;
+                for (auto& entryPoint : get_module()->entry_points()) {
+                    if (entryPoint.opcode() != spv::Op::OpEntryPoint) continue;
+                    if (static_cast<spv::ExecutionModel>(entryPoint.GetSingleWordInOperand(0)) ==
+                        spv::ExecutionModel::Fragment) {
+                        isFragment = true;
+                        break;
+                    }
+                }
+                if (!isFragment) return Status::SuccessWithoutChange;
+
+                // Plan first, mutate second. Materializing the LOD constant is itself a module
+                // change, so it must not happen unless at least one rewrite is going to follow -
+                // otherwise the pass would grow the binary while reporting SuccessWithoutChange.
+                Vector<RewritePlan> plans;
+                for (auto& function : *get_module()) {
+                    for (auto& block : function) {
+                        for (auto& inst : block) {
+                            RewritePlan plan{};
+                            if (PlanRewrite(&inst, plan)) plans.push_back(Move(plan));
+                        }
+                    }
+                }
+                if (plans.empty()) return Status::SuccessWithoutChange;
+
+                const Uint32 zeroId = GetFloatZeroId();
+                if (zeroId == 0) return Status::SuccessWithoutChange;
+
+                for (auto& plan : plans) {
+                    plan.operands.push_back({SPV_OPERAND_TYPE_ID, {zeroId}});
+                    for (auto& operand : plan.trailingOperands) {
+                        plan.operands.push_back(operand);
+                    }
+                    plan.instruction->SetOpcode(plan.opcode);
+                    plan.instruction->SetInOperands(Move(plan.operands));
+                }
+                // Opcodes and operand lists changed underneath every cached analysis.
+                context()->InvalidateAnalysesExceptFor(spvtools::opt::IRContext::kAnalysisNone);
+                return Status::SuccessWithChange;
+            }
+
+        private:
+            struct RewritePlan {
+                spvtools::opt::Instruction* instruction = nullptr;
+                spv::Op opcode = spv::Op::OpNop;
+                // Everything up to and including the Image Operands mask; the Lod id and the
+                // trailing operand values are appended once the constant exists.
+                Vector<spvtools::opt::Operand> operands;
+                Vector<spvtools::opt::Operand> trailingOperands;
+            };
+
+            // Image Operands bits that may accompany an implicit-LOD sample, in the canonical
+            // ascending order SPIR-V requires the operand values to appear in.
+            static constexpr Uint32 kBias = 0x1;
+            static constexpr Uint32 kLod = 0x2;
+            static constexpr Uint32 kGrad = 0x4;
+            static constexpr Uint32 kConstOffset = 0x8;
+            static constexpr Uint32 kOffset = 0x10;
+            static constexpr Uint32 kConstOffsets = 0x20;
+            static constexpr Uint32 kSample = 0x40;
+            static constexpr Uint32 kMinLod = 0x80;
+            static constexpr Uint32 kKnownMask = 0xFF;
+
+            Uint32 GetFloatZeroId() {
+                // Reuse a 32-bit float type already in the module; a shader that samples always has
+                // one, and looking it up avoids depending on type-creation API details.
+                Uint32 floatTypeId = 0;
+                for (auto& inst : get_module()->types_values()) {
+                    if (inst.opcode() == spv::Op::OpTypeFloat && inst.NumInOperands() >= 1 &&
+                        inst.GetSingleWordInOperand(0) == 32) {
+                        floatTypeId = inst.result_id();
+                        break;
+                    }
+                }
+                if (floatTypeId == 0) return 0;
+
+                const auto* floatType = context()->get_type_mgr()->GetType(floatTypeId);
+                if (floatType == nullptr) return 0;
+                const auto zeroBits = std::bit_cast<Uint32>(0.0f);
+                const auto* zeroConst = context()->get_constant_mgr()->GetConstant(floatType, {zeroBits});
+                if (zeroConst == nullptr) return 0;
+                auto* zeroInst = context()->get_constant_mgr()->GetDefiningInstruction(zeroConst);
+                return zeroInst != nullptr ? zeroInst->result_id() : 0;
+            }
+
+            static Bool MapOpcode(spv::Op op, spv::Op& outOpcode, Uint32& outFixedOperandCount) {
+                switch (op) {
+                case spv::Op::OpImageSampleImplicitLod:
+                    outOpcode = spv::Op::OpImageSampleExplicitLod;
+                    outFixedOperandCount = 2; // sampled image, coordinate
+                    return true;
+                case spv::Op::OpImageSampleProjImplicitLod:
+                    outOpcode = spv::Op::OpImageSampleProjExplicitLod;
+                    outFixedOperandCount = 2;
+                    return true;
+                case spv::Op::OpImageSampleDrefImplicitLod:
+                    outOpcode = spv::Op::OpImageSampleDrefExplicitLod;
+                    outFixedOperandCount = 3; // sampled image, coordinate, Dref
+                    return true;
+                case spv::Op::OpImageSampleProjDrefImplicitLod:
+                    outOpcode = spv::Op::OpImageSampleProjDrefExplicitLod;
+                    outFixedOperandCount = 3;
+                    return true;
+                default:
+                    return false;
+                }
+            }
+
+            static Bool PlanRewrite(spvtools::opt::Instruction* inst, RewritePlan& outPlan) {
+                spv::Op newOpcode = spv::Op::OpNop;
+                Uint32 fixedCount = 0;
+                if (!MapOpcode(inst->opcode(), newOpcode, fixedCount)) return false;
+                if (inst->NumInOperands() < fixedCount) return false;
+
+                Uint32 mask = 0;
+                Uint32 next = fixedCount;
+                if (inst->NumInOperands() > fixedCount) {
+                    mask = inst->GetSingleWordInOperand(fixedCount);
+                    next = fixedCount + 1;
+                }
+                // An operand this pass does not model would be silently reordered or dropped, and
+                // Grad cannot legally accompany an implicit-LOD sample: leave such an instruction be.
+                if ((mask & ~kKnownMask) != 0 || (mask & kGrad) != 0) return false;
+
+                Vector<spvtools::opt::Operand> fixedOperands;
+                fixedOperands.reserve(fixedCount + 1);
+                for (Uint32 i = 0; i < fixedCount; ++i) {
+                    fixedOperands.push_back(inst->GetInOperand(i));
+                }
+
+                // Collect the surviving operand values in the same ascending-bit order they were
+                // encoded in, so the rebuilt list stays canonical.
+                Uint32 keptMask = kLod;
+                Vector<spvtools::opt::Operand> keptOperands;
+                static constexpr Uint32 kOrderedBits[] = {kBias,   kLod,          kGrad,   kConstOffset,
+                                                          kOffset, kConstOffsets, kSample, kMinLod};
+                for (const Uint32 bit : kOrderedBits) {
+                    if ((mask & bit) == 0) continue;
+                    if (next >= inst->NumInOperands()) return false;
+                    const spvtools::opt::Operand value = inst->GetInOperand(next++);
+                    // Bias and MinLod only shift a lambda that is already clamped to 0, and any
+                    // original Lod is replaced by the constant the caller appends.
+                    if (bit == kBias || bit == kMinLod || bit == kLod) continue;
+                    keptMask |= bit;
+                    keptOperands.push_back(value);
+                }
+
+                fixedOperands.push_back({SPV_OPERAND_TYPE_IMAGE, {keptMask}});
+                outPlan.instruction = inst;
+                outPlan.opcode = newOpcode;
+                outPlan.operands = Move(fixedOperands);
+                outPlan.trailingOperands = Move(keptOperands);
+                return true;
+            }
+        };
+
+        spvtools::Optimizer::PassToken CreateForceExplicitLod0SamplePass() {
+            return spvtools::Optimizer::PassToken(MakeUnique<ForceExplicitLod0SamplePass>());
+        }
+
+        // TEMP-PERFDIAG: measure what fragment-stage fp32 costs on this GPU. Desktop GLSL carries
+        // no precision qualifiers, so everything reaches the driver as full fp32 while Adreno runs
+        // fp16 at twice the rate. Decorating every float-typed result in a fragment entry point
+        // with RelaxedPrecision is the blunt "all mediump" upper bound - it changes results, so it
+        // is a probe, not a shipping transform. Toggled by /sdcard/MG/exp_relaxed_precision.
+        class RelaxedPrecisionProbePass final : public spvtools::opt::Pass {
+        public:
+            const char* name() const override { return "relaxed-precision-probe"; }
+
+            Status Process() override {
+                Bool isFragment = false;
+                for (auto& entryPoint : get_module()->entry_points()) {
+                    if (entryPoint.opcode() != spv::Op::OpEntryPoint) continue;
+                    if (static_cast<spv::ExecutionModel>(entryPoint.GetSingleWordInOperand(0)) ==
+                        spv::ExecutionModel::Fragment) {
+                        isFragment = true;
+                        break;
+                    }
+                }
+                if (!isFragment) return Status::SuccessWithoutChange;
+
+                // Every 32-bit-float scalar/vector/matrix type in the module. Anything wider (f64)
+                // or narrower is left alone: RelaxedPrecision only has meaning for 32-bit floats.
+                std::unordered_set<Uint32> relaxableTypes;
+                for (auto& type : get_module()->types_values()) {
+                    const Uint32 typeId = type.result_id();
+                    if (typeId == 0) continue;
+                    switch (type.opcode()) {
+                    case spv::Op::OpTypeFloat:
+                        if (type.GetSingleWordInOperand(0) == 32) relaxableTypes.insert(typeId);
+                        break;
+                    case spv::Op::OpTypeVector:
+                    case spv::Op::OpTypeMatrix:
+                        if (relaxableTypes.count(type.GetSingleWordInOperand(0)) != 0) {
+                            relaxableTypes.insert(typeId);
+                        }
+                        break;
+                    default:
+                        break;
+                    }
+                }
+                if (relaxableTypes.empty()) return Status::SuccessWithoutChange;
+
+                Vector<Uint32> targets;
+                for (auto& function : *get_module()) {
+                    for (auto& block : function) {
+                        for (auto& inst : block) {
+                            const Uint32 resultId = inst.result_id();
+                            if (resultId == 0) continue;
+                            if (relaxableTypes.count(inst.type_id()) == 0) continue;
+                            targets.push_back(resultId);
+                        }
+                    }
+                }
+                if (targets.empty()) return Status::SuccessWithoutChange;
+
+                for (const Uint32 id : targets) {
+                    context()->get_decoration_mgr()->AddDecoration(
+                        id, static_cast<Uint32>(spv::Decoration::RelaxedPrecision));
+                }
+                context()->InvalidateAnalysesExceptFor(spvtools::opt::IRContext::kAnalysisNone);
+                return Status::SuccessWithChange;
+            }
+        };
+
+        // Relax fragment-stage arithmetic that provably came out of a texture read. Desktop GLSL
+        // has no precision qualifiers, so every fragment value reaches the driver as fp32 while
+        // Adreno runs fp16 at twice the rate - and a texel is at most 8 bits per channel, which
+        // fp16's 11-bit mantissa carries exactly. Seeding at image reads and propagating only
+        // through operations whose every input is already relaxed keeps everything the shader
+        // computes from other sources (screen coordinates, depth, wide-range uniforms) at full
+        // precision, which is where fp16 would actually go wrong: fp16 cannot even represent a
+        // 3044-pixel gl_FragCoord.x exactly.
+        class RelaxTextureDerivedPrecisionPass final : public spvtools::opt::Pass {
+        public:
+            const char* name() const override { return "relax-texture-derived-precision"; }
+
+            Status Process() override {
+                if (!IsFragmentEntryPoint()) return Status::SuccessWithoutChange;
+                // A shader that drives depth or coverage itself is out of scope: those values must
+                // stay exact, and proving which computations feed them is not worth it here.
+                if (WritesDepthOrSampleMask()) return Status::SuccessWithoutChange;
+
+                CollectRelaxableFloatTypes();
+                if (m_relaxableTypes.empty()) return Status::SuccessWithoutChange;
+
+                // Whitelisting from texture reads captures nothing in practice: MC's fragment
+                // shaders multiply every texel by an interpolated colour and a UBO value, so one
+                // un-relaxed operand vetoes the whole expression (measured: no fps change).
+                // Taint the few genuinely precision-critical sources instead and relax the rest.
+                std::unordered_set<Uint32> tainted;
+                CollectPrecisionCriticalSeeds(tainted);
+                Bool grew = true;
+                while (grew) {
+                    grew = false;
+                    for (auto& function : *get_module()) {
+                        for (auto& block : function) {
+                            for (auto& inst : block) {
+                                const Uint32 resultId = inst.result_id();
+                                if (resultId == 0 || tainted.count(resultId) != 0) continue;
+                                if (!AnyOperandTainted(inst, tainted)) continue;
+                                tainted.insert(resultId);
+                                grew = true;
+                            }
+                        }
+                    }
+                }
+
+                std::unordered_set<Uint32> relaxed;
+                for (auto& function : *get_module()) {
+                    for (auto& block : function) {
+                        for (auto& inst : block) {
+                            const Uint32 resultId = inst.result_id();
+                            if (resultId == 0 || tainted.count(resultId) != 0) continue;
+                            if (m_relaxableTypes.count(inst.type_id()) == 0) continue;
+                            relaxed.insert(resultId);
+                        }
+                    }
+                }
+                if (relaxed.empty()) return Status::SuccessWithoutChange;
+
+                for (const Uint32 id : relaxed) {
+                    context()->get_decoration_mgr()->AddDecoration(
+                        id, static_cast<Uint32>(spv::Decoration::RelaxedPrecision));
+                }
+                context()->InvalidateAnalysesExceptFor(spvtools::opt::IRContext::kAnalysisNone);
+                return Status::SuccessWithChange;
+            }
+
+        private:
+            std::unordered_set<Uint32> m_relaxableTypes;
+
+            Bool IsFragmentEntryPoint() const {
+                for (auto& entryPoint : get_module()->entry_points()) {
+                    if (entryPoint.opcode() != spv::Op::OpEntryPoint) continue;
+                    if (static_cast<spv::ExecutionModel>(entryPoint.GetSingleWordInOperand(0)) ==
+                        spv::ExecutionModel::Fragment) {
+                        return true;
+                    }
+                }
+                return false;
+            }
+
+            Bool WritesDepthOrSampleMask() const {
+                for (auto& annotation : get_module()->annotations()) {
+                    if (annotation.opcode() != spv::Op::OpDecorate) continue;
+                    if (static_cast<spv::Decoration>(annotation.GetSingleWordInOperand(1)) !=
+                        spv::Decoration::BuiltIn) {
+                        continue;
+                    }
+                    const auto builtIn = static_cast<spv::BuiltIn>(annotation.GetSingleWordInOperand(2));
+                    if (builtIn == spv::BuiltIn::FragDepth || builtIn == spv::BuiltIn::SampleMask) {
+                        return true;
+                    }
+                }
+                return false;
+            }
+
+            void CollectRelaxableFloatTypes() {
+                m_relaxableTypes.clear();
+                for (auto& type : get_module()->types_values()) {
+                    const Uint32 typeId = type.result_id();
+                    if (typeId == 0) continue;
+                    switch (type.opcode()) {
+                    case spv::Op::OpTypeFloat:
+                        if (type.GetSingleWordInOperand(0) == 32) m_relaxableTypes.insert(typeId);
+                        break;
+                    case spv::Op::OpTypeVector:
+                        if (m_relaxableTypes.count(type.GetSingleWordInOperand(0)) != 0) {
+                            m_relaxableTypes.insert(typeId);
+                        }
+                        break;
+                    default:
+                        break;
+                    }
+                }
+            }
+
+            void CollectImageReadSeeds(std::unordered_set<Uint32>& relaxed) const {
+                for (auto& function : *get_module()) {
+                    for (auto& block : function) {
+                        for (auto& inst : block) {
+                            const Uint32 resultId = inst.result_id();
+                            if (resultId == 0 || m_relaxableTypes.count(inst.type_id()) == 0) continue;
+                            // Interpolated user varyings seed too, or propagation dies at the
+                            // first `texel * vertexColour`: the load of an Input can never be
+                            // relaxed by the rule below (its operand is a pointer), so a single
+                            // varying vetoes every downstream operation. This is what ESSL's
+                            // mediump varyings already mean. Built-ins are excluded - gl_FragCoord
+                            // carries pixel coordinates that fp16 cannot represent exactly.
+                            if (inst.opcode() == spv::Op::OpLoad && IsNonBuiltInFragmentInput(inst)) {
+                                relaxed.insert(resultId);
+                                continue;
+                            }
+                            switch (inst.opcode()) {
+                            case spv::Op::OpImageSampleImplicitLod:
+                            case spv::Op::OpImageSampleExplicitLod:
+                            case spv::Op::OpImageSampleProjImplicitLod:
+                            case spv::Op::OpImageSampleProjExplicitLod:
+                            case spv::Op::OpImageSampleDrefImplicitLod:
+                            case spv::Op::OpImageSampleDrefExplicitLod:
+                            case spv::Op::OpImageFetch:
+                            case spv::Op::OpImageRead:
+                            case spv::Op::OpImageGather:
+                                relaxed.insert(resultId);
+                                break;
+                            default:
+                                break;
+                            }
+                        }
+                    }
+                }
+            }
+
+            // OpLoad straight out of a fragment Input variable that carries no BuiltIn decoration.
+            // Only a direct load counts: a load through an access chain could be indexing a
+            // structure whose other members are not interpolated colour data.
+            Bool IsNonBuiltInFragmentInput(const spvtools::opt::Instruction& load) const {
+                const Uint32 pointerId = load.GetSingleWordInOperand(0);
+                const auto* pointer = context()->get_def_use_mgr()->GetDef(pointerId);
+                if (pointer == nullptr || pointer->opcode() != spv::Op::OpVariable) return false;
+                if (static_cast<spv::StorageClass>(pointer->GetSingleWordInOperand(0)) !=
+                    spv::StorageClass::Input) {
+                    return false;
+                }
+                Bool isBuiltIn = false;
+                context()->get_decoration_mgr()->ForEachDecoration(
+                    pointerId, static_cast<Uint32>(spv::Decoration::BuiltIn),
+                    [&isBuiltIn](const spvtools::opt::Instruction&) { isBuiltIn = true; });
+                return !isBuiltIn;
+            }
+
+            // A float constant small enough that fp16 represents it without surprise. Colour math
+            // constants (0, 1, 0.5, 255, gamma exponents) all live here; anything larger is
+            // treated as unknown so it stops propagation.
+            Bool IsBoundedFloatConstant(Uint32 id) const {
+                const auto* constant = context()->get_constant_mgr()->FindDeclaredConstant(id);
+                if (constant == nullptr) return false;
+                if (const auto* scalar = constant->AsFloatConstant()) {
+                    const float value = scalar->GetFloat();
+                    return std::isfinite(value) && std::fabs(value) <= 1024.0f;
+                }
+                if (const auto* composite = constant->AsVectorConstant()) {
+                    for (const auto* component : composite->GetComponents()) {
+                        const auto* scalar = component->AsFloatConstant();
+                        if (scalar == nullptr) return false;
+                        const float value = scalar->GetFloat();
+                        if (!std::isfinite(value) || std::fabs(value) > 1024.0f) return false;
+                    }
+                    return true;
+                }
+                return false;
+            }
+
+            // Precision-critical sources: a built-in fragment input. gl_FragCoord is the one that
+            // matters - fp16 cannot represent a 3044-pixel x coordinate exactly, and anything
+            // derived from it (screen-space effects, manual depth reconstruction) would visibly
+            // quantise. Everything else a fragment shader reads is colour-range data.
+            void CollectPrecisionCriticalSeeds(std::unordered_set<Uint32>& tainted) const {
+                for (auto& function : *get_module()) {
+                    for (auto& block : function) {
+                        for (auto& inst : block) {
+                            if (inst.opcode() != spv::Op::OpLoad || inst.result_id() == 0) continue;
+                            if (IsBuiltInInputLoad(inst)) tainted.insert(inst.result_id());
+                        }
+                    }
+                }
+            }
+
+            Bool IsBuiltInInputLoad(const spvtools::opt::Instruction& load) const {
+                const Uint32 pointerId = load.GetSingleWordInOperand(0);
+                const auto* pointer = context()->get_def_use_mgr()->GetDef(pointerId);
+                if (pointer == nullptr || pointer->opcode() != spv::Op::OpVariable) return false;
+                if (static_cast<spv::StorageClass>(pointer->GetSingleWordInOperand(0)) !=
+                    spv::StorageClass::Input) {
+                    return false;
+                }
+                Bool isBuiltIn = false;
+                context()->get_decoration_mgr()->ForEachDecoration(
+                    pointerId, static_cast<Uint32>(spv::Decoration::BuiltIn),
+                    [&isBuiltIn](const spvtools::opt::Instruction&) { isBuiltIn = true; });
+                return isBuiltIn;
+            }
+
+            Bool AnyOperandTainted(const spvtools::opt::Instruction& inst,
+                                   const std::unordered_set<Uint32>& tainted) const {
+                const Uint32 operandCount = inst.NumInOperands();
+                for (Uint32 i = 0; i < operandCount; ++i) {
+                    const auto& operand = inst.GetInOperand(i);
+                    if (!spvIsIdType(operand.type)) continue;
+                    if (IsNonNumericOperand(inst, i)) continue;
+                    if (tainted.count(operand.words[0]) != 0) return true;
+                }
+                return false;
+            }
+
+            Bool AllValueOperandsRelaxed(const spvtools::opt::Instruction& inst,
+                                         const std::unordered_set<Uint32>& relaxed) const {
+                switch (inst.opcode()) {
+                // Pointer-typed plumbing: relaxing the loaded value would say nothing about the
+                // memory it came from, and the pointer operand can never be in the set.
+                case spv::Op::OpLoad:
+                case spv::Op::OpStore:
+                case spv::Op::OpAccessChain:
+                case spv::Op::OpInBoundsAccessChain:
+                case spv::Op::OpFunctionCall:
+                    return false;
+                default:
+                    break;
+                }
+
+                Bool sawValueOperand = false;
+                Bool allRelaxed = true;
+                const Uint32 operandCount = inst.NumInOperands();
+                for (Uint32 i = 0; i < operandCount; ++i) {
+                    const auto& operand = inst.GetInOperand(i);
+                    if (!spvIsIdType(operand.type)) continue; // literals: selectors, swizzle indices
+                    const Uint32 id = operand.words[0];
+                    // OpPhi's block labels, OpSelect's condition and OpExtInst's instruction-set id
+                    // are ids that carry no numeric precision; skip them rather than let them veto.
+                    if (IsNonNumericOperand(inst, i)) continue;
+                    sawValueOperand = true;
+                    if (relaxed.count(id) != 0) continue;
+                    if (IsBoundedFloatConstant(id)) continue;
+                    allRelaxed = false;
+                    break;
+                }
+                return sawValueOperand && allRelaxed;
+            }
+
+            static Bool IsNonNumericOperand(const spvtools::opt::Instruction& inst, Uint32 index) {
+                switch (inst.opcode()) {
+                case spv::Op::OpPhi:
+                    return (index % 2) == 1; // parent block labels
+                case spv::Op::OpSelect:
+                    return index == 0; // condition
+                case spv::Op::OpExtInst:
+                    return index == 0; // extended instruction set
+                default:
+                    return false;
+                }
+            }
+        };
+
+        // TEMP-PERFDIAG: A/B switch between the scoped transform and the all-float upper bound.
+        Bool PerfDiagRelaxAllPrecision() {
+            static const Bool enabled = [] {
+                std::FILE* probe = std::fopen("/sdcard/MG/exp_relaxed_precision_all", "rb");
+                if (probe == nullptr) return false;
+                std::fclose(probe);
+                MGLOG_I("[PERFDIAG] fragment RelaxedPrecision: ALL floats (upper-bound probe)");
+                return true;
+            }();
+            return enabled;
+        }
+
+        // TEMP-PERFDIAG: lets a run turn the transform off entirely for an A/B baseline.
+        Bool PerfDiagRelaxedPrecisionEnabled() {
+            static const Bool disabled = [] {
+                std::FILE* probe = std::fopen("/sdcard/MG/exp_no_relaxed_precision", "rb");
+                if (probe == nullptr) return false;
+                std::fclose(probe);
+                MGLOG_I("[PERFDIAG] fragment RelaxedPrecision DISABLED");
+                return true;
+            }();
+            return !disabled;
+        }
+
+        Bool TransformSpirvForExplicitLod0Sampling(const Vector<Uint>& input, Vector<Uint>& output) {
+            if (input.empty()) {
+                output.clear();
+                return true;
+            }
+            spvtools::Optimizer optimizer(SPV_ENV_VULKAN_1_3);
+            spvtools::OptimizerOptions options;
+            // Matches the position-fix pass: this build of spirv-tools asserts rather than
+            // reporting, so validation stays off in the shipping path.
+            options.set_run_validator(false);
+            optimizer.SetMessageConsumer([](spv_message_level_t, const char*, const spv_position_t&,
+                                            const char* message) {
+                MGLOG_E("Vulkan: explicit-LOD0 pass: %s", message != nullptr ? message : "");
+            });
+            optimizer.RegisterPass(CreateForceExplicitLod0SamplePass());
+
+            const Bool success = optimizer.Run(input.data(), input.size(), &output, options);
+            if (!success) {
+                MGLOG_E("Vulkan: explicit-LOD0 sampling pass failed; keeping the original module");
+                output = input;
+            }
+            return success;
+        }
+
         spvtools::Optimizer::PassToken CreateGlToVulkanPositionFixPass(
             ProgramFactory::CompileOptionFlags transformFlags) {
             return spvtools::Optimizer::PassToken(MakeUnique<GlToVulkanPositionFixPass>(transformFlags));
         }
 
+        // TEMP-PERFDIAG
+        Bool TransformSpirvForRelaxedPrecisionProbe(const Vector<Uint>& input, Vector<Uint>& output) {
+            if (input.empty()) {
+                output.clear();
+                return true;
+            }
+            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: relaxed-precision probe: %s", message != nullptr ? message : "");
+            });
+            // SSA promotion first: glslang emits function-local variables with stores and loads,
+            // and a load can never be relaxed (its operand is a pointer), so without this the
+            // propagation below dies at the first temporary.
+            optimizer.RegisterPass(spvtools::CreateLocalMultiStoreElimPass());
+            if (PerfDiagRelaxAllPrecision()) {
+                optimizer.RegisterPass(spvtools::Optimizer::PassToken(MakeUnique<RelaxedPrecisionProbePass>()));
+            } else {
+                optimizer.RegisterPass(
+                    spvtools::Optimizer::PassToken(MakeUnique<RelaxTextureDerivedPrecisionPass>()));
+            }
+            const Bool success = optimizer.Run(input.data(), input.size(), &output, options);
+            if (!success) {
+                MGLOG_E("Vulkan: relaxed-precision probe failed; keeping the original module");
+                output = input;
+            }
+            return success;
+        }
+
         Bool TransformSpirvForVulkanPositionFix(const Vector<Uint>& input, Vector<Uint>& output,
                                                 ProgramFactory::CompileOptionFlags transformFlags) {
             if (input.empty()) {
@@ -1978,6 +2580,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 moduleSpirvs[i] = spv;
             }
 
+            if ((flags & ProgramFactory::CompileOptionBit::ExplicitLod0Sampling) && shaders[i] &&
+                shaders[i]->GetShaderStage() == ShaderStage::Fragment) {
+                Vector<Uint> explicitLodSpirv;
+                if (TransformSpirvForExplicitLod0Sampling(moduleSpirvs[i], explicitLodSpirv)) {
+                    moduleSpirvs[i] = Move(explicitLodSpirv);
+                }
+            }
+
+            if ((flags & ProgramFactory::CompileOptionBit::RelaxedFragmentPrecision) &&
+                PerfDiagRelaxedPrecisionEnabled() && shaders[i] &&
+                shaders[i]->GetShaderStage() == ShaderStage::Fragment) {
+                Vector<Uint> relaxedSpirv;
+                if (TransformSpirvForRelaxedPrecisionProbe(moduleSpirvs[i], relaxedSpirv)) {
+                    moduleSpirvs[i] = Move(relaxedSpirv);
+                }
+            }
+
             // GL apps depend on cross-program position invariance for multi-pass equality
             // depth tests (MC 26.3's OIT re-draws the cloud geometry with GEQUAL against the
             // depth its own first pass wrote); decorate Position outputs Invariant so

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.h

@@ -42,6 +42,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             SurfaceRotate90 = 1 << 2,
             SurfaceRotate180 = 1 << 3,
             SurfaceRotate270 = 1 << 4,
+            // Rewrites the fragment stage's implicit-LOD image samples to explicit LOD 0.
+            // Only ever set for a draw whose every sampler binding is clamped to a single mip
+            // level, which makes the two forms produce identical texels (the implicit lambda is
+            // clamped into [minLod, maxLod] = [0, 0] regardless of derivatives or bias).
+            ExplicitLod0Sampling = 1 << 5,
+            // Fragment arithmetic may run at relaxed (fp16) precision. Only requested for draws
+            // where every sampled texture and every colour attachment is an 8-bit-or-less
+            // normalized format, so nothing the shader reads or writes carries more precision
+            // than fp16 already represents exactly.
+            RelaxedFragmentPrecision = 1 << 6,
         };
         using CompileOptionFlags = Flags<CompileOptionBit>;
         using HashType = Uint64;

MobileGL/MG_Backend/DirectVulkan/Renderer/SwapchainObject.cpp

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

MobileGL/MG_Backend/DirectVulkan/Renderer/SwapchainObject.h

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

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.cpp

@@ -16,6 +16,7 @@
 #include "MG_Util/Converters/GLToMG/TextureEnumConverter.h"
 #include "MG_Util/Converters/MGToStr/FramebufferEnumConverter.h"
 #include "MG_Util/Converters/MGToVk/TextureEnumConverter.h"
+#include <vulkan/utility/vk_format_utils.h>
 #include "MG_Util/Metrics/TextureMetrics.h"
 #include <Config.h>
 #include <cstdio>
@@ -384,24 +385,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const Uint16 samplerVersion = samplerToUse->GetVersion();
             const Uint64 textureLifetimeId = texture->GetLifetimeId();
             const Uint16 textureParamsVersion = texture->GetTextureParamsVersion();
+            // The sampler's LOD clamp depends on how many levels the sampled view exposes, and that
+            // follows uploads as well as GL parameters - so it belongs in the memo key too.
+            const Uint32 viewLevelCount = resource->sampledLevelCount;
             if (memo.valid && memo.samplerLifetimeId == samplerLifetimeId && memo.samplerVersion == samplerVersion &&
                 memo.textureLifetimeId == textureLifetimeId && memo.textureParamsVersion == textureParamsVersion &&
-                memo.forceNearestFiltering == forceNearestFiltering) {
+                memo.forceNearestFiltering == forceNearestFiltering && memo.viewLevelCount == viewLevelCount) {
                 resolvedSampler = memo.sampler;
             } else {
-                resolvedSampler =
+                resolvedSampler = m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture,
-                    m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture, forceNearestFiltering);
+                                                                       forceNearestFiltering, viewLevelCount);
                 memo.samplerLifetimeId = samplerLifetimeId;
                 memo.samplerVersion = samplerVersion;
                 memo.textureLifetimeId = textureLifetimeId;
                 memo.textureParamsVersion = textureParamsVersion;
                 memo.forceNearestFiltering = forceNearestFiltering;
+                memo.viewLevelCount = viewLevelCount;
                 memo.sampler = resolvedSampler;
                 memo.valid = true;
             }
         } else {
-            resolvedSampler =
+            resolvedSampler = m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture, forceNearestFiltering,
-                m_samplerManager->GetOrCreateSampler(*samplerToUse, *texture, forceNearestFiltering);
+                                                                   resource->sampledLevelCount);
         }
         outImageInfo = {
             .sampler = resolvedSampler,
@@ -442,6 +447,106 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return outImageInfo.sampler != VK_NULL_HANDLE;
     }
 
+    namespace {
+        // fp16 carries an 11-bit mantissa, so an 8-bit normalized channel round-trips exactly.
+        // Anything wider - 16-bit normalized, half float, full float, and every packed HDR
+        // encoding - holds precision or range that relaxing the arithmetic would throw away.
+        Bool IsLowPrecisionNormalizedFormat(VkFormat format) {
+            if (format == VK_FORMAT_UNDEFINED) return false;
+            if (!vkuFormatIsUNORM(format) && !vkuFormatIsSNORM(format) && !vkuFormatIsSRGB(format)) {
+                return false;
+            }
+            const struct VKU_FORMAT_INFO info = vkuGetFormatInfo(format);
+            for (Uint32 i = 0; i < info.component_count; ++i) {
+                if (info.components[i].size > 8) return false;
+            }
+            return info.component_count > 0;
+        }
+    } // namespace
+
+    Bool UniformManager::DrawTargetIsLowPrecision(const MG_State::GLState::FramebufferObject* drawFramebuffer) {
+        // Default framebuffer: the swapchain is an 8-bit normalized surface.
+        if (drawFramebuffer == nullptr) return true;
+
+        Bool sawColour = false;
+        for (Int i = static_cast<Int>(FramebufferAttachmentType::Color0);
+             i < static_cast<Int>(FramebufferAttachmentType::FramebufferAttachmentTypeCount);
+             ++i) {
+            const auto& attachment =
+                drawFramebuffer->GetAttachment(static_cast<FramebufferAttachmentType>(i));
+            VkFormat format = VK_FORMAT_UNDEFINED;
+            if (const auto& texture = attachment.GetTexture()) {
+                format = MG_Util::ConvertTextureInternalFormatToVkEnum(texture->GetFormat());
+            } else if (const auto& renderbuffer = attachment.GetRenderbuffer()) {
+                format = MG_Util::ConvertTextureInternalFormatToVkEnum(
+                    renderbuffer->GetInternalFormat());
+            } else {
+                continue;
+            }
+            if (!IsLowPrecisionNormalizedFormat(format)) return false;
+            sawColour = true;
+        }
+        return sawColour;
+    }
+
+    Bool UniformManager::ProgramSamplesOnlyLowPrecisionTextures(
+        const MG_State::GLState::ProgramObject& program, const ProgramFactory::VkProgramObject& programObj) {
+        for (Uint32 binding = 0; binding < programObj.bindingKinds.size(); ++binding) {
+            if (programObj.bindingKinds[binding] != ProgramFactory::DescriptorBindingKind::CombinedImageSampler) {
+                continue;
+            }
+            const auto* texture = ResolveSamplerTextureRaw(program, programObj, binding);
+            // An unresolvable binding is unknown territory, not licence to relax.
+            if (texture == nullptr) return false;
+            const VkFormat format =
+                MG_Util::ConvertTextureInternalFormatToVkEnum(texture->GetFormat());
+            if (!IsLowPrecisionNormalizedFormat(format)) return false;
+        }
+        return true;
+    }
+
+    Bool UniformManager::ProgramSamplesOnlySingleLevelTextures(
+        const MG_State::GLState::ProgramObject& program, const ProgramFactory::VkProgramObject& programObj) {
+        Bool sawSampler = false;
+        for (Uint32 binding = 0; binding < programObj.bindingKinds.size(); ++binding) {
+            if (programObj.bindingKinds[binding] != ProgramFactory::DescriptorBindingKind::CombinedImageSampler) {
+                continue;
+            }
+            const auto* texture = ResolveSamplerTextureRaw(program, programObj, binding);
+            if (texture == nullptr) return false;
+            const auto& levelRange = texture->GetLevelRange();
+            if (levelRange.x() != levelRange.y()) return false;
+
+            // An explicit-LOD sample is a single filtered tap, so it also gives up anisotropic
+            // filtering - which a single-level view can still have. Resolve the sampler exactly
+            // the way ResolveSamplerDescriptor does and bail if anisotropy would apply.
+            const Int location = programObj.samplerUniformLocationByBinding[binding];
+            const Int unit = ResolveSamplerUnitIndex(program, location, binding);
+            const auto& samplerOverride = MG_State::pGLContext->GetTextureUnitObject(unit).GetSamplerObject();
+            const auto* effectiveSampler =
+                samplerOverride ? samplerOverride.get() : texture->GetSamplerObject().get();
+            if (effectiveSampler == nullptr) return false;
+            if (effectiveSampler->GetMaxAnisotropy() > 1.0f &&
+                effectiveSampler->GetMinFilter() == SamplerFilterMode::Linear &&
+                effectiveSampler->GetMagFilter() == SamplerFilterMode::Linear) {
+                return false;
+            }
+
+            // An explicit LOD 0 makes lambda exactly 0, which is the magnification side of the
+            // min/mag decision. That only matches the implicit form when lambda could not have been
+            // positive anyway (the LOD clamp already pins it at or below 0), or when the two
+            // filters are the same and the choice cannot be observed.
+            const Float effectiveMaxLod = effectiveSampler->GetMipmapMode() == SamplerMipmapMode::None
+                                              ? 0.0f
+                                              : effectiveSampler->GetMaxLod();
+            if (effectiveMaxLod > 0.0f && effectiveSampler->GetMinFilter() != effectiveSampler->GetMagFilter()) {
+                return false;
+            }
+            sawSampler = true;
+        }
+        return sawSampler;
+    }
+
     Bool UniformManager::ResolveSamplerTexture(const MG_State::GLState::ProgramObject& program,
                                                          const ProgramFactory::VkProgramObject& programObj, Uint32 binding,
                                                          SharedPtr<MG_State::GLState::ITextureObject>& outTexture) {

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -69,6 +69,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         static VkFormat ResolveStorageImageViewFormat(VkFormat reflectedFormat, GLenum bindingFormat,
                                                       VkFormat resourceFormat, Bool useBindingFormat);
 
+        // True when the program reads at least one sampler and every one of them is bound to a
+        // texture whose GL level range is a single level. Such a sampler resolves to
+        // minLod = maxLod = 0 (see VkSamplerManager::GetOrCreateSampler), so an implicit-LOD sample
+        // and an explicit LOD 0 sample must read the same texel - which is what makes the
+        // ExplicitLod0Sampling SPIR-V rewrite safe to request. Deliberately conservative: it reads
+        // only GL state, so a texture that ends up single-level for another reason (one uploaded
+        // level under a wide level range) merely misses the rewrite.
+        // True when every texture this program samples is an 8-bit-or-less normalized format, so
+        // relaxing the fragment stage to fp16 cannot lose a bit the texel ever carried. Says
+        // nothing about the render target - the caller must check that too.
+        static Bool ProgramSamplesOnlyLowPrecisionTextures(const MG_State::GLState::ProgramObject& program,
+                                                          const ProgramFactory::VkProgramObject& programObj);
+        // True when every colour attachment the draw writes is an 8-bit-or-less normalized
+        // format (nullptr = default framebuffer, which is). Blending happens at attachment
+        // precision, so a wider target must keep the fragment stage at full precision.
+        static Bool DrawTargetIsLowPrecision(const MG_State::GLState::FramebufferObject* drawFramebuffer);
+        static Bool ProgramSamplesOnlySingleLevelTextures(const MG_State::GLState::ProgramObject& program,
+                                                          const ProgramFactory::VkProgramObject& programObj);
+
     private:
         struct DescriptorPoolBucket {
             VkDescriptorPool handle = VK_NULL_HANDLE;
@@ -191,6 +210,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             Uint64 samplerLifetimeId = 0;
             Uint64 textureLifetimeId = 0;
             VkSampler sampler = VK_NULL_HANDLE;
+            Uint32 viewLevelCount = 0;
             Uint16 samplerVersion = 0;
             Uint16 textureParamsVersion = 0;
             Bool forceNearestFiltering = false;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkSamplerManager.cpp

@@ -51,6 +51,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Float ResolveEffectiveMinLod(const MG_State::GLState::SamplerObject& sampler, Float effectiveMaxLod) {
             return std::min(sampler.GetMinLod(), effectiveMaxLod);
         }
+
+        // A single-level view can only ever deliver the base level, but the LOD clamp must not be
+        // collapsed to exactly 0: both GL and Vulkan pick magFilter over minFilter from the
+        // *clamped* lambda, so maxLod = 0 would make every fragment magnify and quietly retire the
+        // min filter. 0.25 is the value VkSamplerCreateInfo's own note prescribes for emulating
+        // GL's non-mipmapped minification - large enough for lambda to stay positive, small enough
+        // that a NEAREST mip mode still rounds down to level 0. Clamped rather than assigned, so a
+        // texture whose GL_TEXTURE_MAX_LOD really is 0 keeps magnifying as GL says it must.
+        Float ResolveSingleLevelMaxLod(const MG_State::GLState::SamplerObject& sampler, Bool singleLevelView) {
+            const Float maxLod = ResolveEffectiveMaxLod(sampler);
+            return singleLevelView ? std::min(maxLod, 0.25f) : maxLod;
+        }
     } // namespace
 
     Bool VkSamplerManager::Initialize(const InitInfo& initInfo) {
@@ -120,11 +132,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     Uint64 VkSamplerManager::BuildSamplerKey(const MG_State::GLState::SamplerObject& sampler,
                                              const MG_State::GLState::ITextureObject& texture,
-                                             Bool forceNearestFiltering) const {
+                                             Bool forceNearestFiltering, Bool singleLevelView) const {
         MOBILEGL_ASSERT(m_config != nullptr, "VkSamplerManager::BuildSamplerKey: m_config is null");
         XXHASH_VERIFY(XXH64_reset(m_hashState, m_config->CacheVersion));
 
         XXHASH_VERIFY(XXH64_update(m_hashState, &forceNearestFiltering, sizeof(forceNearestFiltering)));
+        XXHASH_VERIFY(XXH64_update(m_hashState, &singleLevelView, sizeof(singleLevelView)));
 
         const auto minFilter = sampler.GetMinFilter();
         XXHASH_VERIFY(XXH64_update(m_hashState, &minFilter, sizeof(minFilter)));
@@ -138,7 +151,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         XXHASH_VERIFY(XXH64_update(m_hashState, &wrapT, sizeof(wrapT)));
         const auto wrapR = sampler.GetWrapR();
         XXHASH_VERIFY(XXH64_update(m_hashState, &wrapR, sizeof(wrapR)));
-        const auto maxLod = ResolveEffectiveMaxLod(sampler);
+        const auto maxLod = ResolveSingleLevelMaxLod(sampler, singleLevelView);
         const auto minLod = ResolveEffectiveMinLod(sampler, maxLod);
         XXHASH_VERIFY(XXH64_update(m_hashState, &minLod, sizeof(minLod)));
         XXHASH_VERIFY(XXH64_update(m_hashState, &maxLod, sizeof(maxLod)));
@@ -160,8 +173,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     VkSampler VkSamplerManager::GetOrCreateSampler(const MG_State::GLState::SamplerObject& sampler,
                                                    const MG_State::GLState::ITextureObject& texture,
-                                                   Bool forceNearestFiltering) {
+                                                   Bool forceNearestFiltering, Uint32 viewLevelCount) {
-        const Uint64 key = BuildSamplerKey(sampler, texture, forceNearestFiltering);
+        // A view that exposes a single mip level has no second level to blend with, so GL's
+        // *_MIPMAP_* minification filters degenerate to plain filtering on the base level -
+        // sampling is unchanged by pinning the Vulkan sampler to NEAREST mip mode at LOD 0.
+        // It is not cosmetic: MobileGL backs such a view with a fully allocated mip chain whose
+        // tail is never written, and a LINEAR mip mode lets the texture unit issue the level+1
+        // fetch anyway. On Adreno that fetch lands in uninitialized UBWC pages (or past the
+        // allocation for a genuinely single-level image) and faults the GPU - the same failure
+        // the default-framebuffer blit shader had to work around with an explicit-LOD sample.
+        const Bool singleLevelView = viewLevelCount == 1;
+        const Uint64 key = BuildSamplerKey(sampler, texture, forceNearestFiltering, singleLevelView);
         auto it = m_samplers.find(key);
         if (it != m_samplers.end()) {
             it->second.lastUsedFrameBoundary = m_frameBoundaryCounter;
@@ -172,7 +194,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         samplerInfo.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO;
         samplerInfo.magFilter = forceNearestFiltering ? VK_FILTER_NEAREST : ToVkFilter(sampler.GetMagFilter());
         samplerInfo.minFilter = forceNearestFiltering ? VK_FILTER_NEAREST : ToVkFilter(sampler.GetMinFilter());
-        samplerInfo.mipmapMode = forceNearestFiltering ? VK_SAMPLER_MIPMAP_MODE_NEAREST
+        samplerInfo.mipmapMode = (forceNearestFiltering || singleLevelView)
+                                     ? VK_SAMPLER_MIPMAP_MODE_NEAREST
                                      : ToVkMipmapMode(sampler.GetMipmapMode());
         samplerInfo.addressModeU = ToVkAddressMode(sampler.GetWrapS());
         samplerInfo.addressModeV = ToVkAddressMode(sampler.GetWrapT());
@@ -185,7 +208,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         samplerInfo.maxAnisotropy = maxAnisotropy;
         samplerInfo.compareEnable = sampler.GetCompareMode() == SamplerCompareMode::CompareToTexture ? VK_TRUE : VK_FALSE;
         samplerInfo.compareOp = ToVkCompareOp(ResolveCompareFunc(sampler, texture));
-        samplerInfo.maxLod = ResolveEffectiveMaxLod(sampler);
+        // Must match BuildSamplerKey's resolution exactly.
+        samplerInfo.maxLod = ResolveSingleLevelMaxLod(sampler, singleLevelView);
         samplerInfo.minLod = ResolveEffectiveMinLod(sampler, samplerInfo.maxLod);
         samplerInfo.borderColor = ResolveVkBorderColor(sampler, texture);
         samplerInfo.unnormalizedCoordinates = VK_FALSE;

MobileGL/MG_Backend/DirectVulkan/Renderer/VkSamplerManager.h

@@ -33,9 +33,12 @@ public:
     Bool Initialize(const InitInfo& initInfo);
     void Shutdown();
 
+    // viewLevelCount is the mip-level count of the image view this sampler will be paired
+    // with; 0 means "unknown, do not narrow". See GetOrCreateSampler for why it matters.
     VkSampler GetOrCreateSampler(const MG_State::GLState::SamplerObject& sampler,
                                  const MG_State::GLState::ITextureObject& texture,
-                                 Bool forceNearestFiltering = false);
+                                 Bool forceNearestFiltering = false,
+                                 Uint32 viewLevelCount = 0);
     // Frame boundary hook: ages the sampler cache and destroys samplers not used
     // for many frames. The key hashes continuous float state (lodBias, LOD clamps,
     // anisotropy), so an app animating those would otherwise mint an unbounded
@@ -61,7 +64,7 @@ private:
 
     Uint64 BuildSamplerKey(const MG_State::GLState::SamplerObject& sampler,
                            const MG_State::GLState::ITextureObject& texture,
-                           Bool forceNearestFiltering) const;
+                           Bool forceNearestFiltering, Bool singleLevelView) const;
     static VkFilter ToVkFilter(SamplerFilterMode mode);
     static VkSamplerMipmapMode ToVkMipmapMode(SamplerMipmapMode mode);
     static VkSamplerAddressMode ToVkAddressMode(SamplerWrapMode mode);

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.cpp

@@ -587,6 +587,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_allocator = initInfo.allocator;
         m_commandPool = initInfo.commandPool;
         m_graphicsQueue = initInfo.graphicsQueue;
+        m_imageFormatListSupported = initInfo.imageFormatListSupported;
         m_currentFrameIndex = 0;
         m_deferredReleases.clear();
         m_deferredReleases.resize(initInfo.frameCount);
@@ -609,6 +610,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         DestroyDeferredReleases();
         m_textureResources.clear();
         m_aliveObjects.clear();
+        m_storageImageTextures.clear();
 
         m_device = VK_NULL_HANDLE;
         m_physicalDevice = VK_NULL_HANDLE;
@@ -656,6 +658,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_textureResources.erase(resourceIt);
         }
         m_aliveObjects.erase(identity);
+        m_storageImageTextures.erase(identity);
     }
 
     void VkTextureManager::PruneStaleTextureAliases(MG_State::GLState::ITextureObject* texture) {
@@ -1189,8 +1192,25 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return ok;
     }
 
+    void VkTextureManager::MarkStorageImageTexture(MG_State::GLState::ITextureObject& texture) {
+        m_storageImageTextures.insert(MakeTextureIdentity(&texture));
+    }
+
+    Bool VkTextureManager::NeedsStorageUsageUpgrade(MG_State::GLState::ITextureObject& texture) const {
+        const TextureIdentity identity = MakeTextureIdentity(&texture);
+        if (m_storageImageTextures.find(identity) == m_storageImageTextures.end()) {
+            return false;
+        }
+        const auto it = m_textureResources.find(identity);
+        // No image yet: the first sync creates it with STORAGE straight away, so there is nothing
+        // to preserve and nothing to order against.
+        return it != m_textureResources.end() && it->second.image != VK_NULL_HANDLE &&
+               !it->second.storageUsageResolved;
+    }
+
     Bool VkTextureManager::NeedsStorageImagePreparation(MG_State::GLState::ITextureObject& texture) const {
-        const auto it = m_textureResources.find(MakeTextureIdentity(&texture));
+        const TextureIdentity identity = MakeTextureIdentity(&texture);
+        const auto it = m_textureResources.find(identity);
         if (it == m_textureResources.end()) {
             return true;
         }
@@ -1198,6 +1218,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (resource.image == VK_NULL_HANDLE || resource.layout != VK_IMAGE_LAYOUT_GENERAL) {
             return true;
         }
+        // The image predates this texture's first image-unit binding, so it was created without
+        // STORAGE usage and has to be recreated - which is illegal inside a render pass.
+        if (!resource.storageUsageResolved &&
+            m_storageImageTextures.find(identity) != m_storageImageTextures.end()) {
+            return true;
+        }
         // Mirror SyncTexture's cross-draw skip condition: any version drift means the sync
         // path may upload or rebuild, both of which need the render pass ended first.
         const auto* mipTexture = MG_State::GLState::AsMipmapTexture(&texture);
@@ -1304,7 +1330,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const auto* syncingMipTexture = MG_State::GLState::AsMipmapTexture(&texture);
         const Uint32 syncingMipLevelCount =
             syncingMipTexture != nullptr ? syncingMipTexture->GetMipmapLevelCount() : 0u;
-        if (outResource.image != VK_NULL_HANDLE &&
+        // A pending storage-usage upgrade also has to bust the skip: nothing about the texture's
+        // content or params changed, but the image itself must be recreated with STORAGE usage
+        // before it can back an image-unit descriptor.
+        const Bool storageUpgradePending =
+            !outResource.storageUsageResolved &&
+            m_storageImageTextures.find(MakeTextureIdentity(&texture)) != m_storageImageTextures.end();
+        if (outResource.image != VK_NULL_HANDLE && !storageUpgradePending &&
             outResource.syncedContentVersion == syncingContentVersion &&
             outResource.syncedTextureParamsVersion == texture.GetTextureParamsVersion() &&
             outResource.syncedMipLevelCount == syncingMipLevelCount) {
@@ -1415,16 +1447,44 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const VkImageAspectFlags aspect = GetAspectMaskForFormat(format);
         VkFormatProperties formatProperties{};
         vkGetPhysicalDeviceFormatProperties(m_physicalDevice, format, &formatProperties);
-        const Bool supportsStorageImage =
+        // Only textures that have actually been bound to a GL image unit get STORAGE usage (and
+        // the MUTABLE_FORMAT it drags in for format-reinterpreting image views). Requesting it
+        // for every storage-capable colour texture costs real bandwidth: Adreno cannot keep UBWC
+        // compression on an image that may be written through a storage descriptor, so the whole
+        // render target - MC's included - runs uncompressed. MarkStorageImageTexture upgrades a
+        // texture before its first image-unit draw, and the usage below feeds the compatibility
+        // check so the upgrade recreates the image.
+        const Bool markedAsStorageImage =
+            m_storageImageTextures.find(MakeTextureIdentity(
+                const_cast<MG_State::GLState::ITextureObject*>(&texture))) != m_storageImageTextures.end();
+        // Storage-image CAPABILITY (does the format allow it at all) is deliberately separate from
+        // whether this texture actually needs the usage. MUTABLE_FORMAT keys off capability, as
+        // before: format-reinterpreting views are not a storage-only concern - the SAMPLED path
+        // needs them too (GetOrCreateSampledImageView bails out without it, see ~line 892), so
+        // tying MUTABLE_FORMAT to the image-unit mark would break sampled format reinterpretation
+        // for every texture that never becomes a storage image.
+        const Bool storageImageCapable =
             !isMultisampleTexture &&
             (aspect & VK_IMAGE_ASPECT_COLOR_BIT) != 0 &&
             (formatProperties.optimalTilingFeatures & VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT) != 0;
+        const Bool supportsStorageImage = storageImageCapable && markedAsStorageImage;
         VkImageCreateFlags imageCreateFlags = shapeInfo.imageFlags;
-        if (supportsStorageImage && IsMutableStorageImageFormat(format) &&
+        if (storageImageCapable && IsMutableStorageImageFormat(format) &&
             m_mutableFormatUnsupported.find(format) == m_mutableFormatUnsupported.end()) {
             imageCreateFlags |= VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
         }
 
+        VkImageUsageFlags desiredUsage =
+            VK_IMAGE_USAGE_SAMPLED_BIT |
+            (supportsStorageImage ? VK_IMAGE_USAGE_STORAGE_BIT : 0) |
+            ((aspect & VK_IMAGE_ASPECT_COLOR_BIT) ? VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT : 0) |
+            (((aspect & VK_IMAGE_ASPECT_DEPTH_BIT) || (aspect & VK_IMAGE_ASPECT_STENCIL_BIT)) ?
+                 VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT :
+                 0);
+        if (!isMultisampleTexture) {
+            desiredUsage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
+        }
+
         const Bool compatible = resource.image != VK_NULL_HANDLE && resource.format == format &&
                                 resource.extent.width == static_cast<Uint32>(texelSize.x()) &&
                                 resource.extent.height == static_cast<Uint32>(texelSize.y()) &&
@@ -1433,6 +1493,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                 resource.viewType == shapeInfo.viewType &&
                                 resource.sampleCount == resolvedSampleCount &&
                                 resource.imageCreateFlags == imageCreateFlags &&
+                                resource.usageFlags == desiredUsage &&
                                 resource.mipLevels == backingMipLevels;
         if (compatible) {
             if (resource.perMipViews.size() != backingMipLevels) {
@@ -1441,6 +1502,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (resource.perMipSampledViews.size() != backingMipLevels) {
                 resource.perMipSampledViews.resize(backingMipLevels, VK_NULL_HANDLE);
             }
+            // Keeping the image is itself the answer to the mark: either it already carries
+            // STORAGE, or this format can never carry it. Either way there is nothing left to
+            // recreate, so stop reporting the texture as needing preparation.
+            resource.storageUsageResolved = markedAsStorageImage;
             return true;
         }
 
@@ -1455,7 +1520,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             resource.sampleCount == resolvedSampleCount &&
             resource.imageCreateFlags == imageCreateFlags &&
             resolvedSampleCount == VK_SAMPLE_COUNT_1_BIT &&
-            resource.mipLevels < backingMipLevels &&
+            // '<=' rather than '<': a storage-usage upgrade recreates the image with an
+            // unchanged mip count, and its contents (a render target's pixels live only on the
+            // GPU) still have to survive. The vkCmdCopyImage below copies min(mipLevels).
+            resource.mipLevels <= backingMipLevels &&
             resource.layout != VK_IMAGE_LAYOUT_UNDEFINED;
 
         std::unique_ptr<TextureResource> preservedResource;
@@ -1477,16 +1545,37 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         imageInfo.format = format;
         imageInfo.tiling = VK_IMAGE_TILING_OPTIMAL;
         imageInfo.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
-        imageInfo.usage = VK_IMAGE_USAGE_SAMPLED_BIT |
+        imageInfo.usage = desiredUsage;
-                          (supportsStorageImage ? VK_IMAGE_USAGE_STORAGE_BIT : 0) |
-                          ((aspect & VK_IMAGE_ASPECT_COLOR_BIT) ? VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT : 0) |
-                          (((aspect & VK_IMAGE_ASPECT_DEPTH_BIT) || (aspect & VK_IMAGE_ASPECT_STENCIL_BIT)) ?
-                               VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT :
-                               0);
-        if (!isMultisampleTexture) {
-            imageInfo.usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT;
-        }
         imageInfo.samples = resolvedSampleCount;
+
+        // Bound the mutability. A blindly-mutable image has to be laid out so that ANY format in
+        // its compatibility class can be viewed, which costs bandwidth compression on tilers;
+        // naming the exact set instead lets the driver keep it. Only safe when that set really is
+        // exhaustive, so it is restricted to textures that are not image-unit bound: sampled views
+        // can only ever ask for ResolveSampledImageViewFormat's output, whereas glBindImageTexture
+        // may name any compatible format, which nothing here can enumerate ahead of time.
+        Vector<VkFormat> viewFormats;
+        VkImageFormatListCreateInfo formatListInfo{};
+        if (m_imageFormatListSupported && !supportsStorageImage &&
+            (imageInfo.flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) != 0) {
+            viewFormats.push_back(format);
+            for (const SamplerNumericDomain domain : {SamplerNumericDomain::Float,
+                                                      SamplerNumericDomain::SignedInteger,
+                                                      SamplerNumericDomain::UnsignedInteger}) {
+                const VkFormat viewFormat = ResolveSampledImageViewFormat(format, domain);
+                if (viewFormat == VK_FORMAT_UNDEFINED) {
+                    continue;
+                }
+                if (std::find(viewFormats.begin(), viewFormats.end(), viewFormat) == viewFormats.end()) {
+                    viewFormats.push_back(viewFormat);
+                }
+            }
+            formatListInfo.sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_LIST_CREATE_INFO;
+            formatListInfo.viewFormatCount = static_cast<Uint32>(viewFormats.size());
+            formatListInfo.pViewFormats = viewFormats.data();
+            imageInfo.pNext = &formatListInfo;
+        }
+
         if (isMultisampleTexture || (imageInfo.flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) != 0) {
             VkImageFormatProperties imageFormatProperties{};
             VkResult imageFormatResult = vkGetPhysicalDeviceImageFormatProperties(
@@ -1543,6 +1632,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         resource.viewType = shapeInfo.viewType;
         resource.sampleCount = resolvedSampleCount;
         resource.imageCreateFlags = imageCreateFlags;
+        resource.usageFlags = imageInfo.usage;
+        resource.storageUsageResolved = markedAsStorageImage;
         resource.syncedTextureParamsVersion = 0;
 
         if (preservedResource) {

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.h

@@ -53,6 +53,9 @@ public:
         VkCommandPool commandPool = VK_NULL_HANDLE;
         VkQueue graphicsQueue = VK_NULL_HANDLE;
         Uint32 frameCount = 0;
+        // VK_KHR_image_format_list is enabled: MUTABLE_FORMAT images can name the exact set of
+        // formats they will be viewed as, which is what lets a tiler keep them compressed.
+        Bool imageFormatListSupported = false;
     };
 
     struct TextureResource {
@@ -157,6 +160,17 @@ public:
         VkImageViewType viewType = VK_IMAGE_VIEW_TYPE_2D;
         VkSampleCountFlagBits sampleCount = VK_SAMPLE_COUNT_1_BIT;
         VkImageCreateFlags imageCreateFlags = 0;
+        // Usage the live image was created with. STORAGE is only requested for textures that
+        // have actually been bound to a GL image unit, because on Adreno a storage-capable
+        // image loses UBWC bandwidth compression; a later image binding upgrades the usage
+        // and recreates the image, so the resolved usage has to be part of the compatibility
+        // check that decides whether the existing image can be kept.
+        VkImageUsageFlags usageFlags = 0;
+        // True once this image was (re)resolved while the texture was already marked as an
+        // image-unit texture. Distinguishes "not upgraded yet" from "cannot be upgraded"
+        // (a format whose optimalTilingFeatures lack STORAGE_IMAGE never gains the bit), so
+        // NeedsStorageImagePreparation cannot ask for a recreate that will never happen.
+        Bool storageUsageResolved = false;
         Uint16 syncedTextureParamsVersion = 0;
         // Snapshot of ITextureObject::GetContentVersion() at the last successful sync;
         // lets SyncTexture skip the whole re-check/re-upload when content is unchanged.
@@ -190,6 +204,8 @@ public:
             std::swap(this->viewType, that.viewType);
             std::swap(this->sampleCount, that.sampleCount);
             std::swap(this->imageCreateFlags, that.imageCreateFlags);
+            std::swap(this->usageFlags, that.usageFlags);
+            std::swap(this->storageUsageResolved, that.storageUsageResolved);
             std::swap(this->syncedTextureParamsVersion, that.syncedTextureParamsVersion);
             std::swap(this->syncedContentVersion, that.syncedContentVersion);
             std::swap(this->syncedMipLevelCount, that.syncedMipLevelCount);
@@ -251,6 +267,8 @@ public:
             viewType = VK_IMAGE_VIEW_TYPE_2D;
             sampleCount = VK_SAMPLE_COUNT_1_BIT;
             imageCreateFlags = 0;
+            usageFlags = 0;
+            storageUsageResolved = false;
             syncedTextureParamsVersion = 0;
             syncedContentVersion = 0;
             syncedMipLevelCount = 0;
@@ -289,6 +307,17 @@ public:
                                                       VkImageLayout newLayout);
     Bool TransitionTextureForSampling(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
     Bool TransitionTextureForStorageImage(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
+    // Records that this texture is bound to a GL image unit, so its image must carry
+    // VK_IMAGE_USAGE_STORAGE_BIT. Must be called before NeedsStorageImagePreparation, and
+    // therefore before the render pass is committed: an image that has to be upgraded is
+    // recreated, which is illegal inside a render pass. Sticky for the texture's lifetime -
+    // GL lets an image binding come and go, and re-creating the image every time it does
+    // would cost far more than the compression it wins back.
+    void MarkStorageImageTexture(MG_State::GLState::ITextureObject& texture);
+    // True when this texture is marked but its live image predates the mark, i.e. the next sync
+    // will recreate it with STORAGE usage and copy the old contents forward. Callers use this to
+    // submit their pending recording first, so that copy cannot read pre-flush content.
+    Bool NeedsStorageUsageUpgrade(MG_State::GLState::ITextureObject& texture) const;
     // Non-mutating probe for the per-draw storage-image fast path: true when preparing this
     // texture as a storage image may need work that is illegal inside a render pass (resource
     // creation, dirty-content upload, or a layout transition to GENERAL). Unknown state reports
@@ -375,6 +404,7 @@ private:
     VmaAllocator m_allocator = nullptr;
     VkCommandPool m_commandPool = VK_NULL_HANDLE;
     VkQueue m_graphicsQueue = VK_NULL_HANDLE;
+    Bool m_imageFormatListSupported = false;
     Uint32 m_currentFrameIndex = 0;
 
     Uint8 m_gcCounter = 0;
@@ -398,6 +428,8 @@ private:
     std::unordered_set<VkFormat> m_mutableFormatUnsupported;
     std::unordered_map<TextureIdentity, WeakPtr<MG_State::GLState::ITextureObject>, TextureIdentityHash> m_aliveObjects;
     std::unordered_map<TextureIdentity, TextureResource, TextureIdentityHash> m_textureResources;
+    // Textures that have been bound to a GL image unit (see MarkStorageImageTexture).
+    std::unordered_set<TextureIdentity, TextureIdentityHash> m_storageImageTextures;
     Vector<Vector<TextureResource>> m_deferredReleases;
     Vector<Vector<VkImageView>> m_deferredViewReleases;
 };

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.cpp

@@ -1016,7 +1016,18 @@ layout(location = 0) in vec2 vTexCoord;
 layout(location = 0) out vec4 outColor;
 
 void main() {
-    outColor = texture(uSource, vTexCoord);
+    // Explicit LOD, not texture(): a blit reads exactly the selected level, so
+    // derivative-based mip selection has no business here. It is also load-bearing:
+    // on Adreno 650 (driver 512.502) an implicit-LOD sample of this single-mip
+    // UBWC render target through the pre-rotation (ROTATE_90) mapping reads past
+    // the image's allocation - despite the sampler's maxLod=0 and a nominal 1:1
+    // texel mapping whose LOD is 0, so the driver's implicit-LOD path itself is at
+    // fault - and page-faults the GPU once the neighbouring memory is returned to
+    // the kernel (frame 2 of Minecraft 26.2's resource reload; the kernel then
+    // invalidates the context and the next submit dies with EDEADLK ->
+    // VK_ERROR_DEVICE_LOST at Present). Verified on device: texture() faults on
+    // the second frame every run, textureLod survives with identical state.
+    outColor = textureLod(uSource, vTexCoord, 0.0);
 }
 )";
 
@@ -2480,7 +2491,7 @@ void main() {
         MOBILEGL_ASSERT(m_textureManager != nullptr, "VkTextureManager creation failed.");
         succeeded = m_textureManager->Initialize(
             {m_device, m_physicalDevice.handle, m_allocator, m_commandPool, m_graphicsQueue,
-             m_frameContext.GetFrameCount()});
+             m_frameContext.GetFrameCount(), m_imageFormatListExtensionEnabled});
         MOBILEGL_ASSERT(succeeded, "VkTextureManager initialization failed.");
         m_clearManager = MakeUnique<VkClearManager>();
         MOBILEGL_ASSERT(m_clearManager != nullptr, "VkClearManager creation failed.");
@@ -2563,11 +2574,20 @@ void main() {
         if (m_swapchainObject.GetHandle() != VK_NULL_HANDLE) {
             VkResult acquireResult =
                 m_frameContext.WaitAndAcquireNextImage(m_device, m_swapchainObject.GetHandle(), m_imageIndexAcquired);
-            if (acquireResult == VK_ERROR_OUT_OF_DATE_KHR || acquireResult == VK_SUBOPTIMAL_KHR) {
+            if (acquireResult == VK_ERROR_OUT_OF_DATE_KHR) {
+                // Nothing was acquired and no semaphore signal was armed, so
+                // rebuilding and re-acquiring on the same semaphore is safe.
                 MGLOG_D("Initialize, vkAcquireNextImageKHR got %d, recreating swapchain", acquireResult);
                 RecreateSwapchain();
                 acquireResult =
                     m_frameContext.WaitAndAcquireNextImage(m_device, m_swapchainObject.GetHandle(), m_imageIndexAcquired);
+            } else if (acquireResult == VK_SUBOPTIMAL_KHR) {
+                // The image is usable, and its acquire signal is already armed on
+                // imageAvailableSemaphore. Re-acquiring here would arm a second signal on a
+                // binary semaphore whose first one nobody has waited on yet; keep the image.
+                // Only a real surface change schedules a rebuild.
+                m_swapchainResizeRequested = m_swapchainResizeRequested || SwapchainIsOutOfDate();
+                acquireResult = VK_SUCCESS;
             }
             VK_VERIFY(acquireResult, "Initialize, WaitAndAcquireNextImage");
         } else {
@@ -2693,6 +2713,7 @@ void main() {
             DestroyDebugMessenger();
             m_debugMessenger = VK_NULL_HANDLE;
         }
+        DestroyDebugReportCallback();
 
         if (m_instance != VK_NULL_HANDLE) {
             vkDestroyInstance(m_instance, nullptr);
@@ -4172,7 +4193,7 @@ void main() {
     }
 
     Bool VulkanRenderer::PrepareStorageImageTextures(
-        VkCommandBuffer commandBuffer,
+        FrameContext::FrameData& frame,
         const MG_State::GLState::ProgramObject& program,
         const ProgramFactory::VkProgramObject& programObj) {
         if (!programObj.hasStorageImages) {
@@ -4193,9 +4214,18 @@ void main() {
         // keep the render pass alive instead of splitting it on every storage-image draw (on
         // tiled GPUs each split is a full tile load/store). GL makes cross-draw image-store
         // coherence the app's job (glMemoryBarrier), so no implicit barrier is owed here.
-        Bool anyNeedsPreparation = false;
+        // Record every image-unit binding before probing anything: a texture whose image was
+        // created without STORAGE usage (the default - it costs UBWC compression on Adreno)
+        // needs a recreate, and the probe below is what ends the render pass so that recreate
+        // lands here rather than mid-pass. This cannot be folded into the probe loop, which
+        // stops at the first texture that needs work and would leave the rest unmarked.
         for (auto* texture : storageTextures) {
             MOBILEGL_ASSERT(texture != nullptr, "%s: collected a null storage texture", __func__);
+            m_textureManager->MarkStorageImageTexture(*texture);
+        }
+
+        Bool anyNeedsPreparation = false;
+        for (auto* texture : storageTextures) {
             if (m_textureManager->NeedsStorageImagePreparation(*texture) ||
                 m_clearManager->HasPendingClear(texture)) {
                 anyNeedsPreparation = true;
@@ -4206,21 +4236,51 @@ void main() {
             return true;
         }
 
+        // A first-time storage-usage upgrade recreates the image and carries the old contents
+        // forward with an out-of-band, immediately-submitted copy (PreserveTextureContentsOnRecreate).
+        // Whatever this frame already recorded into the old image is still sitting unsubmitted in
+        // this command buffer, so that copy would read pre-frame content and this frame's rendering
+        // into the texture would be lost - precisely the render-target-then-image-unit case this
+        // whole path exists for. Submit what is recorded first; the copy then queues behind it.
+        Bool anyNeedsStorageUpgrade = false;
+        for (auto* texture : storageTextures) {
+            if (m_textureManager->NeedsStorageUsageUpgrade(*texture)) {
+                anyNeedsStorageUpgrade = true;
+                break;
+            }
+        }
+        if (anyNeedsStorageUpgrade && HasPendingRecordedWork()) {
+            if (FlushPendingCommands()) {
+                // Fresh command buffer: the sampled-descriptor-set memo describes bindings that
+                // only existed in the retired one. FlushPendingCommands drops the pipeline memo
+                // itself; this is the other command-buffer-scoped cache.
+                m_lastSampledSetValid = false;
+            } else {
+                // Best effort: the upgrade still produces a correct image, only its preserved
+                // contents may predate this frame's writes. Dropping the draw would be worse.
+                MGLOG_E("%s: flush before a storage-usage image upgrade failed; preserved contents "
+                        "may be stale for one frame", __func__);
+            }
+        }
+        if (!frame.isCommandRecording) {
+            m_frameContext.BeginCommandRecording();
+        }
+
         // Image uploads, deferred-clear materialization, and layout barriers are illegal inside
         // a classic render pass. Do this before sampler preparation as well: a texture used by
         // both a sampler and an image must stay in GENERAL, and both descriptors must name that
         // same layout independent of SPIR-V reflection/binding order.
         if (VkRenderPassManager::GetActiveRenderPass() != nullptr) {
-            VkRenderPassManager::EndRenderPass(commandBuffer);
+            VkRenderPassManager::EndRenderPass(frame.commandBuffer);
         }
 
         for (auto* texture : storageTextures) {
-            if (!MaterializePendingClearForTexture(commandBuffer, *texture)) {
+            if (!MaterializePendingClearForTexture(frame.commandBuffer, *texture)) {
                 MGLOG_E("%s: failed to materialize pending clear for storage textureId=%d",
                         __func__, texture->GetExternalIndex());
                 return false;
             }
-            if (!m_textureManager->TransitionTextureForStorageImage(commandBuffer, *texture)) {
+            if (!m_textureManager->TransitionTextureForStorageImage(frame.commandBuffer, *texture)) {
                 MGLOG_E("%s: failed to prepare storage textureId=%d",
                         __func__, texture->GetExternalIndex());
                 return false;
@@ -4246,7 +4306,25 @@ void main() {
         const auto& vao = *MG_State::pGLContext->GetBoundVertexArray();
         const auto& program = *MG_State::pGLContext->GetCurrentProgram();
         ProgramFactory::CompileOptionFlags transformFlags = GetShaderTransformFlags(m_swapchainObject.GetPreTransform());
-        const auto& programObj = m_programFactory->GetOrCreateProgram(program, transformFlags);
+        const auto* programObjPtr = &m_programFactory->GetOrCreateProgram(program, transformFlags);
+        // Sampling a colour render target through the driver's implicit-LOD path faults the GPU on
+        // Adreno 650 (see ForceExplicitLod0SamplePass); ask for the explicit-LOD variant when doing
+        // so cannot change a texel, i.e. when every sampler this program reads is pinned to a
+        // single mip level.
+        if (UniformManager::ProgramSamplesOnlySingleLevelTextures(program, *programObjPtr)) {
+            transformFlags |= ProgramFactory::CompileOptionBit::ExplicitLod0Sampling;
+            programObjPtr = &m_programFactory->GetOrCreateProgram(program, transformFlags);
+        }
+        // fp16 fragment arithmetic is only sound when nothing this draw reads or writes carries
+        // more than 8 normalized bits per channel. A shaderpack's HDR gbuffer, or a data texture
+        // holding positions, must keep full precision - and SPIR-V cannot tell, since sampler2D
+        // yields vec4 whatever the bound format is, so the decision has to be made here.
+        if (UniformManager::ProgramSamplesOnlyLowPrecisionTextures(program, *programObjPtr) &&
+            UniformManager::DrawTargetIsLowPrecision(drawFbo.get())) {
+            transformFlags |= ProgramFactory::CompileOptionBit::RelaxedFragmentPrecision;
+            programObjPtr = &m_programFactory->GetOrCreateProgram(program, transformFlags);
+        }
+        const auto& programObj = *programObjPtr;
 
         // Begin command recording if not yet
         if (!frame.isCommandRecording) {
@@ -4256,7 +4334,7 @@ void main() {
             m_lastSampledSetValid = false;
         }
 
-        if (!PrepareStorageImageTextures(frame.commandBuffer, program, programObj)) {
+        if (!PrepareStorageImageTextures(frame, program, programObj)) {
             MGLOG_E("SetupDraw skipped: storage image preparation failed");
             return false;
         }
@@ -4481,7 +4559,7 @@ void main() {
             VkRenderPassManager::EndRenderPass(frame.commandBuffer);
         }
 
-        if (!PrepareStorageImageTextures(frame.commandBuffer, program, programObj)) {
+        if (!PrepareStorageImageTextures(frame, program, programObj)) {
             MGLOG_E("DispatchCompute skipped: storage image preparation failed");
             return;
         }
@@ -4521,7 +4599,7 @@ void main() {
             VkRenderPassManager::EndRenderPass(frame.commandBuffer);
         }
 
-        if (!PrepareStorageImageTextures(frame.commandBuffer, program, programObj)) {
+        if (!PrepareStorageImageTextures(frame, program, programObj)) {
             MGLOG_E("DispatchComputeIndirect skipped: storage image preparation failed");
             return;
         }
@@ -7542,7 +7620,11 @@ void main() {
             m_presentSuspended = false;
             const VkResult acquireResult =
                 m_frameContext.WaitAndAcquireNextImage(m_device, m_swapchainObject.GetHandle(), m_imageIndexAcquired);
-            if (acquireResult != VK_SUBOPTIMAL_KHR) {
+            if (acquireResult == VK_SUBOPTIMAL_KHR) {
+                // Usable image with its acquire signal already armed; a rebuild is scheduled
+                // only if the surface genuinely no longer matches (see step 4 of Present).
+                m_swapchainResizeRequested = m_swapchainResizeRequested || SwapchainIsOutOfDate();
+            } else {
                 VK_VERIFY(acquireResult, "Present, deferred first WaitAndAcquireNextImage");
             }
         }
@@ -7570,16 +7652,24 @@ void main() {
         if (activeRenderPass)
             VkRenderPassManager::EndRenderPass(frame.commandBuffer);
 
+        // Transition while this frame's recording is still open. A frame that
+        // rendered only into FBOs has no default-framebuffer render pass, and that
+        // pass's finalLayout is the only other thing that carries the swapchain
+        // image to PRESENT_SRC_KHR - so closing the buffer first, which made
+        // TransitionToPresent refuse to record, handed the image to
+        // vkQueuePresentKHR in the layout it was acquired in (UNDEFINED on a fresh
+        // swapchain). The SetImageLayout below then made the tracker's
+        // disagreement with reality permanent for that image index.
+        const auto acquiredImageLayout = m_swapchainObject.GetImageLayout(m_imageIndexAcquired);
+        m_frameContext.TransitionToPresent(m_swapchainObject.GetImage(m_imageIndexAcquired), acquiredImageLayout);
+
         if (frame.isCommandRecording) {
             m_frameContext.EndCommandRecording();
             frame.hasCommandBufferRecorded = true;
             m_lastPipelineValid = false; // command-buffer boundary: drop the pipeline memo
         }
 
-        const auto acquiredImageLayout = m_swapchainObject.GetImageLayout(m_imageIndexAcquired);
+        const Bool shouldSubmitCommandBuffer = frame.hasCommandBufferRecorded;
-        const Bool needsLayoutTransitionForPresent =
-            m_frameContext.TransitionToPresent(m_swapchainObject.GetImage(m_imageIndexAcquired), acquiredImageLayout);
-        const Bool shouldSubmitCommandBuffer = frame.hasCommandBufferRecorded || needsLayoutTransitionForPresent;
 
         // 1) Submit current frame work.
         auto submitPacket = m_frameContext.GetSubmitInfo(shouldSubmitCommandBuffer, m_imageIndexAcquired);
@@ -7593,7 +7683,15 @@ void main() {
         // 2) Present current frame.
         auto presentPacket = m_frameContext.GetPresentInfo(m_swapchainObject.GetHandle(), m_imageIndexAcquired);
         auto result = vkQueuePresentKHR(m_presentQueue, &presentPacket.presentInfo);
-        if (result == VK_ERROR_OUT_OF_DATE_KHR || result == VK_SUBOPTIMAL_KHR) {
+        if (result == VK_SUBOPTIMAL_KHR) {
+            // Suboptimal is not a reason to rebuild on its own: a driver may report it for a
+            // surface whose size and orientation still match what we built from (Android does
+            // this routinely), and rebuilding on it alone destroys every pipeline and
+            // reallocates the default framebuffer once per frame - flicker, then garbage.
+            // Defer to the surface-capabilities comparison below.
+            result = VK_SUCCESS;
+        }
+        if (result == VK_ERROR_OUT_OF_DATE_KHR) {
             MGLOG_D("Present, vkQueuePresentKHR got %d, recreating swapchain", result);
             if (!RecreateSwapchain()) {
                 // Window went zero-area (minimize) with the swapchain out of date:
@@ -7607,6 +7705,13 @@ void main() {
             result = VK_SUCCESS;
         }
         VK_VERIFY(result, "Present, vkQueuePresentKHR");
+        // The authoritative check, done here - after the frame is presented, before the next
+        // acquire. This is what makes a launcher-side resolution change take effect: shrinking
+        // the window's buffer (SurfaceHolder.setFixedSize) moves currentExtent, the swapchain
+        // follows, and the compositor scales the smaller image up to the view for free.
+        if (!m_swapchainResizeRequested && SwapchainIsOutOfDate()) {
+            m_swapchainResizeRequested = true;
+        }
         if (m_swapchainResizeRequested) {
             MGLOG_D("Present, processing requested swapchain resize");
             if (!RecreateSwapchain()) {
@@ -7623,7 +7728,16 @@ void main() {
 
         // 4) Wait/reset/acquire for next frame.
         result = m_frameContext.WaitAndAcquireNextImage(m_device, m_swapchainObject.GetHandle(), m_imageIndexAcquired);
-        if (result == VK_ERROR_OUT_OF_DATE_KHR || result == VK_SUBOPTIMAL_KHR) {
+        if (result == VK_SUBOPTIMAL_KHR) {
+            // An image WAS acquired and its signal is armed on this slot's
+            // imageAvailableSemaphore, so the frame proceeds normally. Whether a rebuild is
+            // actually needed is decided by the surface-capabilities comparison at the next
+            // Present - suboptimal alone must not schedule one, or a driver that reports it
+            // every frame would rebuild every frame.
+            m_swapchainResizeRequested = m_swapchainResizeRequested || SwapchainIsOutOfDate();
+            result = VK_SUCCESS;
+        } else if (result == VK_ERROR_OUT_OF_DATE_KHR) {
+            // Nothing acquired, nothing signaled: safe to rebuild and re-acquire.
             MGLOG_D("Present, vkAcquireNextImageKHR got %d, recreating swapchain", result);
             if (!RecreateSwapchain()) {
                 m_presentSuspended = true;
@@ -7671,6 +7785,24 @@ void main() {
 
         m_validationLayersEnabled = m_config.EnableValidationLayers && validationLayerAvailable;
 
+        // The debug messenger is a VK_EXT_debug_utils object, but a driver can ship
+        // the validation layers while exposing only the older VK_EXT_debug_report
+        // (Adreno 650 / Vulkan 1.1.128 does exactly that). Requesting the extension
+        // unconditionally tripped the required-extension assert below, aborting every
+        // validation-enabled build in CreateInstance. Keep the layers - they still
+        // validate, and on Android they report to logcat on their own - and drop only
+        // the messenger.
+        const Bool debugUtilsAvailable =
+            m_validationLayersEnabled && IsExtensionSupported(m_extensions, VK_EXT_DEBUG_UTILS_EXTENSION_NAME);
+        // Without a reporting channel the layers validate but say nothing, so fall
+        // back to VK_EXT_debug_report when debug_utils is missing.
+        const Bool debugReportAvailable = m_validationLayersEnabled && !debugUtilsAvailable &&
+                                          IsExtensionSupported(m_extensions, VK_EXT_DEBUG_REPORT_EXTENSION_NAME);
+        if (m_validationLayersEnabled && !debugUtilsAvailable) {
+            MGLOG_I("%s not available; validation reports via %s instead.", VK_EXT_DEBUG_UTILS_EXTENSION_NAME,
+                    debugReportAvailable ? VK_EXT_DEBUG_REPORT_EXTENSION_NAME : "(no channel)");
+        }
+
         // ---------------- App info -------------------
         VkApplicationInfo appInfo = {};
         appInfo.sType = VK_STRUCTURE_TYPE_APPLICATION_INFO;
@@ -7721,8 +7853,10 @@ void main() {
         }
 #endif
 
-        if (m_validationLayersEnabled) {
+        if (debugUtilsAvailable) {
             exts.push_back(VK_EXT_DEBUG_UTILS_EXTENSION_NAME);
+        } else if (debugReportAvailable) {
+            exts.push_back(VK_EXT_DEBUG_REPORT_EXTENSION_NAME);
         }
 
         MGLOG_I("Enabling %d Vulkan instance extensions:", exts.size());
@@ -7747,7 +7881,8 @@ void main() {
             MGLOG_I("Enabling validation layer...");
             instanceInfo.enabledLayerCount = static_cast<uint32_t>(std::size(s_validationLayerNames));
             instanceInfo.ppEnabledLayerNames = s_validationLayerNames;
-            instanceInfo.pNext = &debugMessengerCreateInfo;
+            // Chaining the messenger create-info is only legal with the extension on.
+            instanceInfo.pNext = debugUtilsAvailable ? &debugMessengerCreateInfo : nullptr;
         } else {
             instanceInfo.enabledLayerCount = 0;
             instanceInfo.pNext = nullptr;
@@ -7755,7 +7890,40 @@ void main() {
 
         VK_VERIFY(vkCreateInstance(&instanceInfo, nullptr, &m_instance), "vkCreateInstance failed");
 
-        if (m_validationLayersEnabled) VK_VERIFY(SetupDebugMessenger());
+        if (debugUtilsAvailable) {
+            VK_VERIFY(SetupDebugMessenger());
+        } else if (debugReportAvailable) {
+            VK_VERIFY(SetupDebugReportCallback());
+        }
+    }
+
+    static VKAPI_ATTR VkBool32 VKAPI_CALL DebugReportCallback(VkDebugReportFlagsEXT flags, VkDebugReportObjectTypeEXT,
+                                                              Uint64, size_t, Int32 messageCode, const char* pLayerPrefix,
+                                                              const char* pMessage, void*) {
+        if ((flags & (VK_DEBUG_REPORT_ERROR_BIT_EXT | VK_DEBUG_REPORT_WARNING_BIT_EXT |
+                      VK_DEBUG_REPORT_PERFORMANCE_WARNING_BIT_EXT)) != 0) {
+            MGLOG_F("[Vulkan %s %d] %s", pLayerPrefix ? pLayerPrefix : "?", messageCode, pMessage ? pMessage : "");
+        }
+        return VK_FALSE;
+    }
+
+    VkResult VulkanRenderer::SetupDebugReportCallback() {
+        auto vkCreateDebugReportCallbackEXT =
+            (PFN_vkCreateDebugReportCallbackEXT)vkGetInstanceProcAddr(m_instance, "vkCreateDebugReportCallbackEXT");
+        if (!vkCreateDebugReportCallbackEXT) return VK_ERROR_EXTENSION_NOT_PRESENT;
+        VkDebugReportCallbackCreateInfoEXT createInfo{VK_STRUCTURE_TYPE_DEBUG_REPORT_CALLBACK_CREATE_INFO_EXT};
+        createInfo.flags = VK_DEBUG_REPORT_ERROR_BIT_EXT | VK_DEBUG_REPORT_WARNING_BIT_EXT |
+                           VK_DEBUG_REPORT_PERFORMANCE_WARNING_BIT_EXT;
+        createInfo.pfnCallback = &DebugReportCallback;
+        return vkCreateDebugReportCallbackEXT(m_instance, &createInfo, nullptr, &m_debugReportCallback);
+    }
+
+    void VulkanRenderer::DestroyDebugReportCallback() {
+        if (m_debugReportCallback == VK_NULL_HANDLE) return;
+        auto func = (PFN_vkDestroyDebugReportCallbackEXT)vkGetInstanceProcAddr(m_instance,
+                                                                               "vkDestroyDebugReportCallbackEXT");
+        if (func != nullptr) func(m_instance, m_debugReportCallback, nullptr);
+        m_debugReportCallback = VK_NULL_HANDLE;
     }
 
     VkResult VulkanRenderer::SetupDebugMessenger() {
@@ -8037,6 +8205,18 @@ void main() {
 
         const Vector<VkExtensionProperties> availableExtensions = EnumerateDeviceExtensions(m_physicalDevice.handle);
         ResolveOptionalDeviceExtensions(availableExtensions, enabledDeviceExtensions);
+
+        // VK_KHR_image_format_list lets a MUTABLE_FORMAT image declare exactly which formats it
+        // may be viewed as. Adreno drops UBWC bandwidth compression on a blindly-mutable image
+        // (measured: 65 -> 80 fps in MC 26.2 once mutability is not requested); an explicit,
+        // compression-compatible format list is the portable way to keep both.
+        m_imageFormatListExtensionEnabled =
+            IsExtensionSupported(availableExtensions, VK_KHR_IMAGE_FORMAT_LIST_EXTENSION_NAME);
+        if (m_imageFormatListExtensionEnabled) {
+            enabledDeviceExtensions.push_back(VK_KHR_IMAGE_FORMAT_LIST_EXTENSION_NAME);
+        }
+        MGLOG_I("VK_KHR_image_format_list enabled: %s",
+                m_imageFormatListExtensionEnabled ? "true" : "false");
         MGLOG_I("VK_KHR_draw_indirect_count enabled: %s", m_drawIndirectCountExtensionEnabled ? "true" : "false");
 
         m_indexTypeUint8ExtensionEnabled = false;
@@ -8501,6 +8681,38 @@ void main() {
         return m_physicalDevice;
     }
 
+    Bool VulkanRenderer::SwapchainIsOutOfDate() {
+        if (m_surface == VK_NULL_HANDLE || m_swapchainObject.GetHandle() == VK_NULL_HANDLE) {
+            return false;
+        }
+        VkSurfaceCapabilitiesKHR surfaceCaps{};
+        if (vkGetPhysicalDeviceSurfaceCapabilitiesKHR(m_physicalDevice.handle, m_surface, &surfaceCaps) !=
+            VK_SUCCESS) {
+            return false;
+        }
+        // A driver-defined currentExtent (UINT32_MAX) means the surface takes its size from the
+        // swapchain, so there is nothing to compare against - the app's requested size wins and
+        // only an explicit RequestSwapchainResize can change it.
+        if (surfaceCaps.currentExtent.width == UINT32_MAX || surfaceCaps.currentExtent.height == UINT32_MAX) {
+            return false;
+        }
+        // Compare in SURFACE space against the extent the live swapchain was created from. Using
+        // the swapchain's own (quarter-turn swapped) extent here would report a difference on
+        // every rotated frame and rebuild forever.
+        const VkExtent2D builtFrom = m_swapchainObject.GetSurfaceExtent();
+        const Bool extentChanged = surfaceCaps.currentExtent.width != builtFrom.width ||
+                                   surfaceCaps.currentExtent.height != builtFrom.height;
+        const Bool transformChanged = surfaceCaps.currentTransform != m_swapchainObject.GetPreTransform();
+        if (!extentChanged && !transformChanged) {
+            return false;
+        }
+        MGLOG_I("Swapchain out of date: surface %ux%u transform %u -> %ux%u transform %u",
+                builtFrom.width, builtFrom.height, static_cast<Uint32>(m_swapchainObject.GetPreTransform()),
+                surfaceCaps.currentExtent.width, surfaceCaps.currentExtent.height,
+                static_cast<Uint32>(surfaceCaps.currentTransform));
+        return true;
+    }
+
     void VulkanRenderer::RequestSwapchainResize(Uint32 width, Uint32 height) {
         width = std::max<Uint32>(width, 1);
         height = std::max<Uint32>(height, 1);

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h

@@ -274,6 +274,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Uint64 GetTimerQueryTimestampNs(const VkTimerQueryManager::TimestampRecord& record) const;
 
         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
+        // rebuild, so a surface the driver merely considers suboptimal cannot thrash.
+        Bool SwapchainIsOutOfDate();
         // Returns false when the surface is zero-area (minimized/hidden window):
         // no new swapchain is installed and presentation must stay suspended.
         Bool RecreateSwapchain();
@@ -399,6 +403,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Vector<VkExtensionProperties> m_extensions;
         VkInstance m_instance = VK_NULL_HANDLE;
         VkDebugUtilsMessengerEXT m_debugMessenger = VK_NULL_HANDLE;
+        // Fallback reporting channel for drivers that ship the validation layers but
+        // only expose the older VK_EXT_debug_report (Adreno 650 / Vulkan 1.1.128).
+        VkDebugReportCallbackEXT m_debugReportCallback = VK_NULL_HANDLE;
         PhysicalDevice m_physicalDevice;
         VkDevice m_device = VK_NULL_HANDLE;
         VmaAllocator m_allocator = nullptr;
@@ -549,6 +556,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void CreateInstance();
         VkResult SetupDebugMessenger();
         VkResult DestroyDebugMessenger();
+        VkResult SetupDebugReportCallback();
+        void DestroyDebugReportCallback();
         VkDebugUtilsMessengerCreateInfoEXT PopulateDebugMessengerCreateInfo();
         void CreateSurface();
         void PickPhysicalDevice();
@@ -567,8 +576,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             const RenderPassEntry& renderPassEntry);
         VkPipeline GetOrCreateComputePipeline(const ProgramFactory::VkProgramObject& programObj);
         void DestroyComputePipelines();
+        // Takes the frame rather than a command buffer: a first-time storage-usage upgrade has to
+        // flush the pending recording (see the body), which retires the current command buffer.
         Bool PrepareStorageImageTextures(
-            VkCommandBuffer commandBuffer,
+            FrameContext::FrameData& frame,
             const MG_State::GLState::ProgramObject& program,
             const ProgramFactory::VkProgramObject& programObj);
 
@@ -630,6 +641,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                  const PhysicalDevice& compareWithDevice,
                                                  PhysicalDevice& outBetterDevice);
         static constexpr const char* s_validationLayerNames[] = {"VK_LAYER_KHRONOS_validation"};
+        // VK_KHR_image_format_list: lets MUTABLE_FORMAT images declare their exact view-format
+        // set so the driver can keep bandwidth compression (see CreateLogicalDeviceAndQueues).
+        Bool m_imageFormatListExtensionEnabled = false;
+
         static constexpr const char* s_deviceExtensionNames[] = {VK_KHR_SWAPCHAIN_EXTENSION_NAME};
         static Bool CheckValidationLayerSupport();
 

MobileGL/MG_Impl/CGLImpl/CGLImpl.cpp

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

MobileGL/MG_Impl/CGLImpl/CGLImpl.h

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

MobileGL/MG_Impl/CGLImpl/Exporting/Definitions.cpp

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

MobileGL/MG_Impl/DyldInterpose/DyldInterpose.cpp

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

MobileGL/MG_Impl/DyldInterpose/ExportedSymbols.txt

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

MobileGL/MG_Impl/GetProcAddress.cpp

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

MobileGL/MG_Impl/NSOpenGLImpl/NSOpenGLImpl.cpp

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

MobileGL/MG_State/GLState/ProgramState/ProgramObject.h

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