[Perf] (DirectVulkan): snapshot-gated consecutive-draw fast path skips SetupDraw re-resolution

- every render pass declared colorAttachmentCount=8 (the full GL draw-buffer
  slot span) with trailing VK_ATTACHMENT_UNUSED references, and Adreno
  configures its per-pixel render-backend/export path from the DECLARED
  count - so every fragment of every pass paid an 8-render-target export
  cost; this was the bulk of the 1.5x per-pixel gap against
  MobileGlues+ANGLE on the same Qualcomm driver (their subpasses declare
  exactly the used span)
- measured on Adreno 650 / MC 26.2 / 1440x3044: total GPU frame time
  11.9 -> 7.5 ms (-37%, now below ANGLE's 7.87 ms), the single-quad
  swapchain blit pass alone 1.26 -> 0.40 ms, steady in-world FPS 82.8 -> 123
  under the standard cooled-start protocol, matching the
  MobileGlues+ANGLE+system-Vulkan benchmark of 123.8
- trailing UNUSED references are popped before the subpass is built (the
  entry's colorAttachmentCount and every pipeline's colour-blend span follow
  it); interior GL_NONE holes keep their slots so fragment-output locations
  still line up
- the pipeline-side fragmentOutputMask check downgrades from assert to a
  debug log: an output at a location past the trimmed span is discarded,
  which is GL's defined behaviour for a draw buffer set to GL_NONE

MobileGL/MG_Backend/DirectVulkan/Renderer/FrameContext.cpp

@@ -16,18 +16,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_device = device;
         m_commandPool = commandPool;
 
-        Vector<VkCommandBuffer> commandBuffers(frameCount, VK_NULL_HANDLE);
+        Vector<VkCommandBuffer> commandBuffers(frameCount * 2, VK_NULL_HANDLE);
         VkCommandBufferAllocateInfo allocInfo{};
         allocInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;
         allocInfo.commandPool = commandPool;
         allocInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
-        allocInfo.commandBufferCount = frameCount;
+        allocInfo.commandBufferCount = frameCount * 2;
         VkResult result = vkAllocateCommandBuffers(device, &allocInfo, commandBuffers.data());
         if (result != VK_SUCCESS) {
             return result;
         }
         for (Uint32 i = 0; i < frameCount; ++i) {
             m_frames[i].commandBuffer = commandBuffers[i];
+            m_frames[i].preCommandBuffer = commandBuffers[frameCount + i];
         }
 
         VkSemaphoreCreateInfo semaphoreInfo{VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO};
@@ -47,9 +48,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     void FrameContext::Destroy(VkDevice device, VkCommandPool commandPool) {
         const Uint32 frameCount = static_cast<Uint32>(m_frames.size());
-        Vector<VkCommandBuffer> commandBuffers(frameCount, VK_NULL_HANDLE);
+        Vector<VkCommandBuffer> commandBuffers(frameCount * 2, VK_NULL_HANDLE);
         for (Uint32 i = 0; i < frameCount; ++i) {
             commandBuffers[i] = m_frames[i].commandBuffer;
+            commandBuffers[frameCount + i] = m_frames[i].preCommandBuffer;
         }
 
         for (Uint32 i = 0; i < frameCount; ++i) {
@@ -60,7 +62,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             for (auto& frame : m_frames) {
                 FreeRetiredCommandBuffers(frame);
             }
-            vkFreeCommandBuffers(device, commandPool, frameCount, commandBuffers.data());
+            vkFreeCommandBuffers(device, commandPool, frameCount * 2, commandBuffers.data());
         }
         m_frames.clear();
         currentFrameIndex = 0;
@@ -87,6 +89,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         currentFrameIndex = (currentFrameIndex + 1) % static_cast<Uint32>(m_frames.size());
         GetCurrent().isCommandRecording = false;
         GetCurrent().hasCommandBufferRecorded = false;
+        GetCurrent().isPreCommandRecording = false;
+        GetCurrent().hasPreCommandBufferRecorded = false;
     }
 
     VkCommandBuffer& FrameContext::BeginCommandRecording(VkCommandBufferUsageFlags flags,
@@ -118,6 +122,41 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         frame.hasCommandBufferRecorded = true;
     }
 
+    VkCommandBuffer FrameContext::BeginPreCommandRecording() {
+        auto& frame = GetCurrent();
+        if (frame.isPreCommandRecording) {
+            return frame.preCommandBuffer;
+        }
+        MOBILEGL_ASSERT(!frame.hasPreCommandBufferRecorded,
+                        "BeginPreCommandRecording: a recorded pre stream is still awaiting submission");
+        VK_VERIFY(vkResetCommandBuffer(frame.preCommandBuffer, 0), "BeginPreCommandRecording, vkResetCommandBuffer");
+        VkCommandBufferBeginInfo beginInfo{};
+        beginInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_BEGIN_INFO;
+        VK_VERIFY(vkBeginCommandBuffer(frame.preCommandBuffer, &beginInfo),
+                  "BeginPreCommandRecording, vkBeginCommandBuffer");
+        frame.isPreCommandRecording = true;
+        return frame.preCommandBuffer;
+    }
+
+    void FrameContext::EndPreCommandRecordingIfOpen() {
+        auto& frame = GetCurrent();
+        if (!frame.isPreCommandRecording) {
+            return;
+        }
+        VK_VERIFY(vkEndCommandBuffer(frame.preCommandBuffer), "EndPreCommandRecordingIfOpen, vkEndCommandBuffer");
+        frame.isPreCommandRecording = false;
+        frame.hasPreCommandBufferRecorded = true;
+    }
+
+    void FrameContext::AbandonPreCommandRecording() {
+        auto& frame = GetCurrent();
+        if (frame.isPreCommandRecording) {
+            VK_VERIFY(vkEndCommandBuffer(frame.preCommandBuffer), "AbandonPreCommandRecording, vkEndCommandBuffer");
+        }
+        frame.isPreCommandRecording = false;
+        frame.hasPreCommandBufferRecorded = false;
+    }
+
     VkResult FrameContext::InitializeSwapchainSemaphores(VkDevice device, Uint32 swapchainImageCount) {
         DestroySwapchainSemaphores(device);
         if (swapchainImageCount == 0) {
@@ -202,17 +241,27 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                                                Uint32 swapchainImageIndex) const {
         const auto& frame = GetCurrent();
         MOBILEGL_ASSERT(!frame.isCommandRecording, "GetSubmitInfo called while command buffer recording is still active");
+        MOBILEGL_ASSERT(!frame.isPreCommandRecording,
+                        "GetSubmitInfo called while the pre-pass stream is still recording");
         AssertValidSwapchainImageIndex(swapchainImageIndex);
         SubmitInfoPacket packet{};
         packet.waitSemaphore = frame.imageAvailableSemaphore;
         packet.signalSemaphore = m_swapchainImageRenderFinishedSemaphores[swapchainImageIndex];
-        packet.commandBuffer = frame.commandBuffer;
+
+        Uint32 commandBufferCount = 0;
+        // The pre-pass stream executes strictly before the frame's commands.
+        if (frame.hasPreCommandBufferRecorded) {
+            packet.commandBuffers[commandBufferCount++] = frame.preCommandBuffer;
+        }
+        if (shouldSubmitCommandBuffer) {
+            packet.commandBuffers[commandBufferCount++] = frame.commandBuffer;
+        }
 
         packet.submitInfo.waitSemaphoreCount = frame.imageAvailableSemaphoreConsumed ? 0U : 1U;
         packet.submitInfo.pWaitSemaphores = frame.imageAvailableSemaphoreConsumed ? nullptr : &packet.waitSemaphore;
         packet.submitInfo.pWaitDstStageMask = frame.imageAvailableSemaphoreConsumed ? nullptr : &packet.waitDstStageMask;
-        packet.submitInfo.commandBufferCount = shouldSubmitCommandBuffer ? 1U : 0U;
+        packet.submitInfo.commandBufferCount = commandBufferCount;
-        packet.submitInfo.pCommandBuffers = shouldSubmitCommandBuffer ? &packet.commandBuffer : nullptr;
+        packet.submitInfo.pCommandBuffers = commandBufferCount > 0 ? packet.commandBuffers : nullptr;
         packet.submitInfo.signalSemaphoreCount = 1;
         packet.submitInfo.pSignalSemaphores = &packet.signalSemaphore;
         return packet;
@@ -276,12 +325,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_recordingObserver = observer;
     }
 
-    VkResult FrameContext::RetireCurrentCommandBuffer() {
+    VkResult FrameContext::RetireCurrentCommandBuffer(Bool retirePreCommandBuffer) {
         MOBILEGL_ASSERT(m_device != VK_NULL_HANDLE && m_commandPool != VK_NULL_HANDLE,
                         "RetireCurrentCommandBuffer requires an initialized FrameContext");
         auto& frame = GetCurrent();
         MOBILEGL_ASSERT(!frame.isCommandRecording,
                         "RetireCurrentCommandBuffer called while the command buffer is still recording");
+        MOBILEGL_ASSERT(!frame.isPreCommandRecording,
+                        "RetireCurrentCommandBuffer called while the pre-pass stream is still recording");
 
         VkCommandBufferAllocateInfo allocInfo{};
         allocInfo.sType = VK_STRUCTURE_TYPE_COMMAND_BUFFER_ALLOCATE_INFO;
@@ -289,10 +340,20 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         allocInfo.level = VK_COMMAND_BUFFER_LEVEL_PRIMARY;
         allocInfo.commandBufferCount = 1;
         VkCommandBuffer replacement = VK_NULL_HANDLE;
-        const VkResult result = vkAllocateCommandBuffers(m_device, &allocInfo, &replacement);
+        VkResult result = vkAllocateCommandBuffers(m_device, &allocInfo, &replacement);
         if (result != VK_SUCCESS) {
             return result;
         }
+        if (retirePreCommandBuffer) {
+            VkCommandBuffer preReplacement = VK_NULL_HANDLE;
+            result = vkAllocateCommandBuffers(m_device, &allocInfo, &preReplacement);
+            if (result != VK_SUCCESS) {
+                vkFreeCommandBuffers(m_device, m_commandPool, 1, &replacement);
+                return result;
+            }
+            frame.retiredCommandBuffers.push_back({frame.preCommandBuffer, frame.lastSubmitIndex});
+            frame.preCommandBuffer = preReplacement;
+        }
         // lastSubmitIndex was just written by the renderer for the submission
         // that carried this command buffer.
         frame.retiredCommandBuffers.push_back({frame.commandBuffer, frame.lastSubmitIndex});

MobileGL/MG_Backend/DirectVulkan/Renderer/FrameContext.h

@@ -29,7 +29,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkPipelineStageFlags waitDstStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
             VkSemaphore waitSemaphore = VK_NULL_HANDLE;
             VkSemaphore signalSemaphore = VK_NULL_HANDLE;
-            VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
+            // [0] = pre-pass command buffer (when recorded), then the frame
+            // command buffer; submitInfo.pCommandBuffers points here.
+            VkCommandBuffer commandBuffers[2] = {VK_NULL_HANDLE, VK_NULL_HANDLE};
             VkSubmitInfo submitInfo{VK_STRUCTURE_TYPE_SUBMIT_INFO};
         };
 
@@ -52,10 +54,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         struct FrameData {
             VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
+            // Pre-pass work stream: out-of-pass commands (deferred clear
+            // materialization, sampled-layout transitions) for resources the
+            // frame's recording has not touched yet. Submitted immediately
+            // BEFORE commandBuffer in the same vkQueueSubmit, so recording
+            // into it never has to split the frame's active render pass.
+            VkCommandBuffer preCommandBuffer = VK_NULL_HANDLE;
             VkSemaphore imageAvailableSemaphore = VK_NULL_HANDLE;
             VkFence imageInFlightFence = VK_NULL_HANDLE;
             Bool isCommandRecording = false;
             Bool hasCommandBufferRecorded = false;
+            Bool isPreCommandRecording = false;
+            Bool hasPreCommandBufferRecorded = false;
             Bool imageAvailableSemaphoreConsumed = false;
             // Command buffers submitted mid-frame (FlushPendingCommands),
             // appended in submit order; freed once their submission is known
@@ -77,6 +87,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VkCommandBuffer& BeginCommandRecording(VkCommandBufferUsageFlags flags = 0,
                                                const VkCommandBufferInheritanceInfo* pInheritanceInfo = nullptr);
         void EndCommandRecording();
+        // Lazily opens the pre-pass work stream (see FrameData::preCommandBuffer).
+        VkCommandBuffer BeginPreCommandRecording();
+        // Closes the pre stream if open, marking it for submission ahead of the
+        // frame command buffer. Safe to call when it never opened.
+        void EndPreCommandRecordingIfOpen();
+        // Drops an in-progress or recorded-but-unsubmitted pre stream (dropped
+        // frame recordings, swapchain recreation).
+        void AbandonPreCommandRecording();
         VkResult InitializeSwapchainSemaphores(VkDevice device, Uint32 swapchainImageCount);
         void DestroySwapchainSemaphores(VkDevice device);
         Bool TransitionToPresent(VkImage image, VkImageLayout oldLayout,
@@ -91,7 +109,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // can restart while the submitted buffer is still executing. Retired
         // buffers are freed after the slot's fence is next waited, or as soon
         // as their submission is observed complete.
-        VkResult RetireCurrentCommandBuffer();
+        VkResult RetireCurrentCommandBuffer(Bool retirePreCommandBuffer = false);
 
         // Frees every retired command buffer whose tagged submission index is
         // known complete. Driven by the renderer's submit tracker on completion

MobileGL/MG_Backend/DirectVulkan/Renderer/ProgramFactory.cpp

@@ -12,10 +12,7 @@
 #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>
@@ -1102,374 +1099,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             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();
@@ -1499,37 +1128,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             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()) {
@@ -2588,15 +2186,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 }
             }
 
-            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

@@ -47,11 +47,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // level, which makes the two forms produce identical texels (the implicit lambda is
             // clamped into [minLod, maxLod] = [0, 0] regardless of derivatives or bias).
             ExplicitLod0Sampling = 1 << 5,
-            // 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

@@ -262,6 +262,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_images.resize(imageCount, VK_NULL_HANDLE);
         VK_VERIFY(vkGetSwapchainImagesKHR(device, m_swapchain, &imageCount, m_images.data()));
         m_imageLayouts.assign(imageCount, VK_IMAGE_LAYOUT_UNDEFINED);
+        // Fresh swapchain images hold garbage until a render pass stores into them.
+        m_imageContentDefined.assign(imageCount, false);
+        m_depthStencilContentDefined.assign(imageCount, false);
 
         CreateImageViews(device);
         CreateDepthStencilResources(device, physicalDevice);
@@ -433,9 +436,39 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         m_images.clear();
         m_imageLayouts.clear();
+        m_imageContentDefined.clear();
+        m_depthStencilContentDefined.clear();
         m_preTransform = VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR;
     }
 
+    Bool SwapchainObject::IsImageContentDefined(Uint32 index) const {
+        MOBILEGL_ASSERT(index < m_imageContentDefined.size(), "Swapchain image content index out of range");
+        return m_imageContentDefined[index];
+    }
+
+    void SwapchainObject::SetImageContentDefined(Uint32 index, Bool defined) {
+        MOBILEGL_ASSERT(index < m_imageContentDefined.size(), "Swapchain image content index out of range");
+        m_imageContentDefined[index] = defined;
+    }
+
+    Bool SwapchainObject::IsDepthStencilContentDefined(Uint32 index) const {
+        MOBILEGL_ASSERT(index < m_depthStencilContentDefined.size(),
+                        "Swapchain depth/stencil content index out of range");
+        return m_depthStencilContentDefined[index];
+    }
+
+    void SwapchainObject::SetDepthStencilContentDefined(Uint32 index, Bool defined) {
+        MOBILEGL_ASSERT(index < m_depthStencilContentDefined.size(),
+                        "Swapchain depth/stencil content index out of range");
+        m_depthStencilContentDefined[index] = defined;
+    }
+
+    void SwapchainObject::SetAllDepthStencilContentUndefined() {
+        for (SizeT i = 0; i < m_depthStencilContentDefined.size(); ++i) {
+            m_depthStencilContentDefined[i] = false;
+        }
+    }
+
     VkImage SwapchainObject::GetImage(Uint32 index) const {
         MOBILEGL_ASSERT(index < m_images.size(), "Swapchain image index out of range");
         return m_images[index];

MobileGL/MG_Backend/DirectVulkan/Renderer/SwapchainObject.h

@@ -52,6 +52,21 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void SetImageLayout(Uint32 index, VkImageLayout layout);
         SizeT GetImageCount() const { return m_images.size(); }
 
+        // EGL content-validity tracking for the default framebuffer. A color
+        // buffer's content is undefined once its image has been presented
+        // (EGL_BUFFER_DESTROYED swap behaviour, the implementation default),
+        // and every ancillary (depth/stencil) buffer's content is undefined
+        // after ANY swap regardless of swap behaviour (EGL 1.5 §3.10.1). The
+        // render-pass manager turns an undefined attachment's tile load into
+        // LOAD_OP_DONT_CARE. Flags start false (a fresh swapchain image holds
+        // garbage) and a render pass storing into an attachment sets it back
+        // to defined.
+        Bool IsImageContentDefined(Uint32 index) const;
+        void SetImageContentDefined(Uint32 index, Bool defined);
+        Bool IsDepthStencilContentDefined(Uint32 index) const;
+        void SetDepthStencilContentDefined(Uint32 index, Bool defined);
+        void SetAllDepthStencilContentUndefined();
+
     private:
         void CreateImageViews(VkDevice device);
         void CreateDepthStencilResources(VkDevice device, VkPhysicalDevice physicalDevice);
@@ -77,5 +92,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Vector<VkDeviceMemory> m_depthStencilImageMemories;
         Vector<VkImageView> m_depthStencilImageViews;
         Vector<VkImageLayout> m_depthStencilImageLayouts;
+        Vector<Bool> m_imageContentDefined;
+        Vector<Bool> m_depthStencilContentDefined;
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.cpp

@@ -16,9 +16,9 @@
 #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 <algorithm>
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
@@ -205,6 +205,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // The frame's descriptor sets are recycled above, so last frame's reuse target
         // is gone: start the per-draw descriptor-reuse cache fresh this frame.
         m_hasLastDescriptor = false;
+        m_lastBindValid = false;
         // Re-fingerprint the bound sampler set fresh this frame so any GL object address
         // reuse cannot outlive a single frame (see SamplerResolveMemo).
         for (auto& memo : m_samplerResolveMemo) {
@@ -447,64 +448,6 @@ 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;
@@ -1248,6 +1191,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                         bufferInfo.range = ubo.range;
                         dynOffset = static_cast<Uint32>(ubo.dynamicOffset);
                     } else {
+                        // Global-UBO slice reuse (see GlobalUboSliceMemo): unchanged
+                        // uniform bytes re-use the slice already uploaded this frame.
+                        const Bool isGlobalUbo =
+                            programObj.globalUboBinding == static_cast<Int>(binding) && element == 0;
+                        const Uint64 uboFrameSerial = m_bufferManager->GetFrameSerial();
+                        const Uint64 uboProgramLifetimeId = program.GetLifetimeId();
+                        const Uint32 uboContentVersion = program.GetUBOContentVersion();
+                        Bool reusedSlice = false;
+                        if (isGlobalUbo) {
+                            for (const auto& memo : m_globalUboMemo) {
+                                if (memo.buffer != VK_NULL_HANDLE &&
+                                    memo.programLifetimeId == uboProgramLifetimeId &&
+                                    memo.frameSerial == uboFrameSerial &&
+                                    memo.uboContentVersion == uboContentVersion &&
+                                    memo.range == static_cast<VkDeviceSize>(ubo.payloadSize)) {
+                                    bufferInfo.buffer = memo.buffer;
+                                    bufferInfo.range = memo.range;
+                                    dynOffset = static_cast<Uint32>(memo.offset);
+                                    reusedSlice = true;
+                                    break;
+                                }
+                            }
+                        }
+                        if (!reusedSlice) {
                             BufferSlice slice{};
                             if (!m_bufferManager->UploadTransient(BufferKind::Uniform, frameIndex, ubo.payload,
                                                                   ubo.payloadSize, m_minDynamicOffsetAlignment, slice)) {
@@ -1258,6 +1225,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                             bufferInfo.buffer = slice.buffer;
                             bufferInfo.range = ubo.payloadSize;
                             dynOffset = static_cast<Uint32>(slice.offset);
+                            if (isGlobalUbo) {
+                                m_globalUboMemo[m_globalUboMemoNext] = GlobalUboSliceMemo{
+                                    uboProgramLifetimeId, uboFrameSerial, uboContentVersion,
+                                    slice.buffer, slice.offset, static_cast<VkDeviceSize>(ubo.payloadSize)};
+                                m_globalUboMemoNext = (m_globalUboMemoNext + 1) % kGlobalUboMemoSize;
+                            }
+                        }
                     }
                     bufferInfos.push_back(bufferInfo);
                     // Dynamic offsets are consumed in binding order, then array element order,
@@ -1396,8 +1370,34 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_hasLastDescriptor = cacheable;
         }
 
+        // Skip the driver call when this exact binding is already live on the
+        // command buffer (see the bind-dedup shadow in the header).
+        const Uint32 offsetCount = static_cast<Uint32>(dynamicOffsets.size());
+        Bool identicalBind = m_lastBindValid && m_lastBindSet == descriptorSet &&
+                             m_lastBindLayout == programObj.pipelineLayout && m_lastBindPoint == bindPoint &&
+                             m_lastBindOffsetCount == offsetCount && offsetCount <= kMaxShadowedDynamicOffsets;
+        if (identicalBind) {
+            for (Uint32 i = 0; i < offsetCount; ++i) {
+                if (m_lastBindOffsets[i] != dynamicOffsets[i]) {
+                    identicalBind = false;
+                    break;
+                }
+            }
+        }
+        if (!identicalBind) {
             vkCmdBindDescriptorSets(commandBuffer, bindPoint, programObj.pipelineLayout, 0, 1,
-                                &descriptorSet, static_cast<Uint32>(dynamicOffsets.size()), dynamicOffsets.data());
+                                    &descriptorSet, offsetCount, dynamicOffsets.data());
+            if (offsetCount <= kMaxShadowedDynamicOffsets) {
+                m_lastBindValid = true;
+                m_lastBindSet = descriptorSet;
+                m_lastBindLayout = programObj.pipelineLayout;
+                m_lastBindPoint = bindPoint;
+                m_lastBindOffsetCount = offsetCount;
+                std::copy_n(dynamicOffsets.data(), offsetCount, m_lastBindOffsets);
+            } else {
+                m_lastBindValid = false;
+            }
+        }
         return true;
     }
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.h

@@ -39,6 +39,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         void Shutdown();
 
         void BeginFrame(Uint32 frameIndex);
+        // A command buffer (re)began recording: descriptor bindings recorded into
+        // the previous buffer do not carry over, so drop the bind-dedup shadow.
+        void OnCommandBufferBoundary() { m_lastBindValid = false; }
         // A ProgramFactory eviction just destroyed this layout: purge every frame
         // slot's cached descriptor sets for it, so a recycled handle value can never
         // stale-hit sets written for the dead layout's bindings. The sets are
@@ -76,15 +79,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // 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);
 
@@ -194,6 +188,35 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Uint64 m_lastDescriptorSignature = 0;
         Bool m_hasLastDescriptor = false;
 
+        // vkCmdBindDescriptorSets dedup: consecutive draws with a static uniform
+        // block resolve to the same set AND the same dynamic offsets, so the
+        // driver call can be skipped outright. Command-buffer-scope state; reset
+        // via OnCommandBufferBoundary whenever a recording (re)begins. Keyed on
+        // layout+bind point, so a pipeline-layout switch always rebinds.
+        static constexpr Uint32 kMaxShadowedDynamicOffsets = 8;
+        Bool m_lastBindValid = false;
+        VkDescriptorSet m_lastBindSet = VK_NULL_HANDLE;
+        VkPipelineLayout m_lastBindLayout = VK_NULL_HANDLE;
+        VkPipelineBindPoint m_lastBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS;
+        Uint32 m_lastBindOffsetCount = 0;
+        Uint32 m_lastBindOffsets[kMaxShadowedDynamicOffsets] = {};
+
+        // Global-UBO transient-slice reuse: MC leaves the default uniform block
+        // untouched across long GUI/terrain runs, so the per-draw re-upload of
+        // the same bytes can reuse the slice uploaded earlier THIS frame (frame
+        // serial guards arena recycling; the content version guards writes).
+        struct GlobalUboSliceMemo {
+            Uint64 programLifetimeId = 0;
+            Uint64 frameSerial = 0;
+            Uint32 uboContentVersion = 0;
+            VkBuffer buffer = VK_NULL_HANDLE;
+            VkDeviceSize offset = 0;
+            VkDeviceSize range = 0;
+        };
+        static constexpr Uint32 kGlobalUboMemoSize = 4;
+        GlobalUboSliceMemo m_globalUboMemo[kGlobalUboMemoSize];
+        Uint32 m_globalUboMemoNext = 0;
+
         // Per-binding fast path over VkSamplerManager's content-hashed sampler cache, which
         // stays the source of truth: its key hashes all sampler+texture state, so two distinct
         // sampler objects with identical state still resolve to one VkSampler. This memo only

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.cpp

@@ -58,15 +58,27 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
     const VertexInputStateFactory::BackendVertexInputState& VertexInputStateFactory::GetOrCreateVertexInputState(
         const MG_State::GLState::VertexArrayObject& vao) {
-        return GetOrCreateVertexInputState(vao, GetOrComputeHash(vao));
+        // Per-draw fast path: the VAO carries a pointer to its resolved entry,
+        // valid while its config version and the cache's eviction epoch both
+        // match - no re-hash, no map lookup.
+        const void* memoState = nullptr;
+        Uint64 memoEpoch = 0;
+        if (vao.GetBackendStateMemo(memoState, memoEpoch) && memoEpoch == m_evictionEpoch) {
+            const auto* entry = static_cast<const BackendVertexInputState*>(memoState);
+            entry->lastUsedFrameBoundary = m_frameBoundaryCounter;
+            return *entry;
+        }
+        const BackendVertexInputState& entry = GetOrCreateVertexInputState(vao, GetOrComputeHash(vao));
+        vao.SetBackendStateMemo(&entry, m_evictionEpoch);
+        return entry;
     }
 
     const VertexInputStateFactory::BackendVertexInputState& VertexInputStateFactory::GetOrCreateVertexInputState(
         const MG_State::GLState::VertexArrayObject& vao, HashType hash) {
         auto it = m_cache.find(hash);
         if (it != m_cache.end()) {
-            it->second.lastUsedFrameBoundary = m_frameBoundaryCounter;
+            it->second->lastUsedFrameBoundary = m_frameBoundaryCounter;
-            return it->second;
+            return *it->second;
         }
 
         VertexInputStateBuilder builder;
@@ -172,11 +184,37 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const auto& state = builder.Build();
 
-        auto& entry = m_cache[hash];
+        auto& slot = m_cache[hash];
+        if (!slot) {
+            slot = MakeUnique<BackendVertexInputState>();
+        }
+        BackendVertexInputState& entry = *slot;
         entry.hash = hash;
         entry.lastUsedFrameBoundary = m_frameBoundaryCounter;
         entry.bindings = builder.GetBindings();
         entry.attributes = builder.GetAttributes();
+        // See the layoutHash declaration: hash only the resolved layout, never
+        // buffer identities, so identical layouts across VAOs/buffers agree.
+        XXHASH_VERIFY(XXH64_reset(m_hashState, 0));
+        for (const auto& binding : entry.bindings) {
+            XXHASH_VERIFY(XXH64_update(m_hashState, &binding.binding, sizeof(binding.binding)));
+            XXHASH_VERIFY(XXH64_update(m_hashState, &binding.stride, sizeof(binding.stride)));
+            XXHASH_VERIFY(XXH64_update(m_hashState, &binding.inputRate, sizeof(binding.inputRate)));
+        }
+        for (const auto& attribute : entry.attributes) {
+            XXHASH_VERIFY(XXH64_update(m_hashState, &attribute.location, sizeof(attribute.location)));
+            XXHASH_VERIFY(XXH64_update(m_hashState, &attribute.binding, sizeof(attribute.binding)));
+            XXHASH_VERIFY(XXH64_update(m_hashState, &attribute.format, sizeof(attribute.format)));
+            XXHASH_VERIFY(XXH64_update(m_hashState, &attribute.offset, sizeof(attribute.offset)));
+        }
+        XXHASH_VERIFY(XXH64_update(m_hashState, &unsupportedAttribMask, sizeof(unsupportedAttribMask)));
+        entry.layoutHash = XXH64_digest(m_hashState);
+        entry.attributeLocationMask = 0;
+        for (const auto& attribute : entry.attributes) {
+            if (attribute.location < 32u) {
+                entry.attributeLocationMask |= (1u << attribute.location);
+            }
+        }
         entry.bindingBufferKeys = std::move(bindingBufferKeys);
         entry.bindingBaseOffsets = std::move(bindingBaseOffsets);
         entry.bindingAttributeLocations = std::move(bindingAttributeLocations);
@@ -205,8 +243,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
 
         for (auto it = m_cache.begin(); it != m_cache.end();) {
-            if (m_frameBoundaryCounter - it->second.lastUsedFrameBoundary > kRetireAgeBoundaries) {
+            if (m_frameBoundaryCounter - it->second->lastUsedFrameBoundary > kRetireAgeBoundaries) {
                 it = m_cache.erase(it);
+                // Invalidate every VAO's state-pointer memo: the erased node's
+                // address may be reused by a future insert.
+                ++m_evictionEpoch;
             } else {
                 ++it;
             }

MobileGL/MG_Backend/DirectVulkan/Renderer/VertexInputStateFactory.h

@@ -27,9 +27,18 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         struct BackendVertexInputState {
             HashType hash = 0;
+            // Hash of the resolved Vulkan vertex layout only (bindings, attributes,
+            // unsupported mask) - NO buffer identities. `hash` mixes buffer heap
+            // addresses so per-chunk VBOs mint a fresh identity per buffer; keying
+            // pipelines on that minted one VkPipeline per chunk section for an
+            // identical layout, defeating pipeline reuse and the per-draw memo.
+            // Pipelines depend only on the layout, so they key on this instead.
+            HashType layoutHash = 0;
             // Frame boundary of the last cache hit; entries idle past the
             // OnFrameBoundary retirement age are evicted (CPU heap only).
-            Uint64 lastUsedFrameBoundary = 0;
+            // Mutable: the VAO's state-pointer memo fast path stamps it through
+            // a const entry reference.
+            mutable Uint64 lastUsedFrameBoundary = 0;
             Vector<VkVertexInputBindingDescription> bindings;
             Vector<VkVertexInputAttributeDescription> attributes;
             Vector<SizeT> bindingBufferKeys;
@@ -41,6 +50,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             // absent from `attributes`, so without this mask the draw path cannot tell them apart from
             // a genuinely disabled array and would silently feed the shader the current attribute value.
             Uint32 unsupportedAttribMask = 0;
+            // Bitmask of `attributes[i].location` - the draw path needs it up to
+            // three times per draw, so it is baked once at build time.
+            Uint32 attributeLocationMask = 0;
             VkPipelineVertexInputStateCreateInfo state{
                 VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO
             };
@@ -80,9 +92,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
 
         const VulkanRendererConfig& m_config;
         VkPhysicalDevice m_physicalDevice = VK_NULL_HANDLE;
-        UnorderedMap<HashType, BackendVertexInputState> m_cache;
+        // Values are heap-allocated: FastSTL::unordered_map is open-addressing,
+        // so INSERT invalidates references to stored values. The draw path (and
+        // the VAOs' state-pointer memos) hold entry pointers across inserts;
+        // only the unique_ptr cell moves, never the pointee.
+        UnorderedMap<HashType, UniquePtr<BackendVertexInputState>> m_cache;
         // Monotonic frame-boundary counter (bumped in OnFrameBoundary) for cache aging.
         Uint64 m_frameBoundaryCounter = 0;
+        // Bumped whenever any cache entry is erased. VAOs memo a raw pointer to
+        // their heap-allocated entry (stable across map insert/rehash by
+        // construction); a memo is honored only while its recorded epoch
+        // matches, so an evicted entry can never be dereferenced through a
+        // stale memo.
+        Uint64 m_evictionEpoch = 1;
         static inline XXH64_state_t* m_hashState = XXH64_createState();
     };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VkClearManager.cpp

@@ -93,6 +93,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         const std::lock_guard<std::mutex> lock(m_mutex);
         m_pendingClears.clear();
         m_aliveObjects.clear();
+        m_pendingCount.store(static_cast<Uint32>(m_pendingClears.size()), std::memory_order_relaxed);
     }
 
     TextureIdentity VkClearManager::MakeTextureIdentity(MG_State::GLState::ITextureObject* texture) {
@@ -127,6 +128,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_pendingClears.erase(key);
         }
         m_aliveObjects.erase(identity);
+        m_pendingCount.store(static_cast<Uint32>(m_pendingClears.size()), std::memory_order_relaxed);
     }
 
     Bool VkClearManager::LockTextureIdentityLocked(const TextureIdentity& identity,
@@ -221,6 +223,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_aliveObjects[MakeTextureIdentity(texture.get())] = texture;
         auto& pending = m_pendingClears[key];
         MergeClearPayload(pending, clearPayload);
+        m_pendingCount.store(static_cast<Uint32>(m_pendingClears.size()), std::memory_order_relaxed);
     }
 
     void VkClearManager::QueueClear(const ClearAttachmentPayload& clearPayload,
@@ -238,6 +241,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_aliveObjects[MakeTextureIdentity(texture.get())] = texture;
         auto& pending = m_pendingClears[key];
         MergeClearPayload(pending, clearPayload);
+        m_pendingCount.store(static_cast<Uint32>(m_pendingClears.size()), std::memory_order_relaxed);
     }
 
     Bool VkClearManager::HasPendingClear(MG_State::GLState::ITextureObject* texture) {
@@ -245,6 +249,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return false;
         }
 
+        if (m_pendingCount.load(std::memory_order_relaxed) == 0) {
+            return false;  // per-draw hot path: nothing pending anywhere
+        }
+
         const Uint64 lifetimeId = texture->GetLifetimeId();
         const std::lock_guard<std::mutex> lock(m_mutex);
         for (auto it = m_pendingClears.begin(); it != m_pendingClears.end(); ++it) {
@@ -260,6 +268,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (key.texture == nullptr) {
             return false;
         }
+        if (m_pendingCount.load(std::memory_order_relaxed) == 0) {
+            return false;  // per-draw hot path: nothing pending anywhere
+        }
 
         const std::lock_guard<std::mutex> lock(m_mutex);
         if (m_pendingClears.find(key) == m_pendingClears.end()) {
@@ -287,6 +298,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (key.texture == nullptr) {
             return false;
         }
+        if (m_pendingCount.load(std::memory_order_relaxed) == 0) {
+            return false;  // per-draw hot path: nothing pending anywhere
+        }
 
         const std::lock_guard<std::mutex> lock(m_mutex);
         if (!LockTextureLocked(key, outTexture)) {
@@ -325,6 +339,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         if (texture == nullptr) {
             return false;
         }
+        if (m_pendingCount.load(std::memory_order_relaxed) == 0) {
+            return false;  // per-draw hot path: nothing pending anywhere
+        }
 
         const Uint64 lifetimeId = texture->GetLifetimeId();
         const std::lock_guard<std::mutex> lock(m_mutex);
@@ -345,6 +362,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return;
         }
 
+        if (m_pendingCount.load(std::memory_order_relaxed) == 0) {
+            return;  // per-draw hot path: nothing pending anywhere
+        }
         const TextureIdentity identity = MakeTextureIdentity(texture);
         MGLOG_D("%s: Pop all pending clears for texture %d", __func__, texture->GetExternalIndex());
         const std::lock_guard<std::mutex> lock(m_mutex);
@@ -361,6 +381,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             auto it = m_pendingClears.find(key);
             if (it != m_pendingClears.end()) {
                 m_pendingClears.erase(it);
+            m_pendingCount.store(static_cast<Uint32>(m_pendingClears.size()), std::memory_order_relaxed);
             }
         }
 

MobileGL/MG_Backend/DirectVulkan/Renderer/VkClearManager.h

@@ -14,6 +14,7 @@
 #include "MG_Util/Math/VectorTypes.h"
 
 #include <Includes.h>
+#include <atomic>
 #include <unordered_map>
 
 namespace MobileGL::MG_Backend::DirectVulkan {
@@ -120,7 +121,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                SharedPtr<MG_State::GLState::ITextureObject>& outTexture);
 
         Uint8 m_gcCounter = 0;
+    public:
+        // Lock-free probe for the consecutive-draw fast path: any pending clear
+        // forces the full SetupDraw path (which materializes/consumes it).
+        Bool HasAnyPendingClears() const { return m_pendingCount.load(std::memory_order_relaxed) != 0; }
+
+    private:
         mutable std::mutex m_mutex;
+        // Lock-free mirror of m_pendingClears.size(), maintained under m_mutex
+        // by every mutation. The per-draw probes (HasPendingClear/GetPending*)
+        // read it before taking the lock: during draw batches the pending set
+        // is almost always empty, so this turns several locked map probes per
+        // draw into one relaxed load.
+        std::atomic<Uint32> m_pendingCount{0};
         std::unordered_map<PendingClearKey, ClearAttachmentPayload, PendingClearKeyHash> m_pendingClears;
         std::unordered_map<TextureIdentity, WeakPtr<MG_State::GLState::ITextureObject>, TextureIdentityHash> m_aliveObjects;
     };

MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.cpp

@@ -481,7 +481,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     }
 
     VkRenderPassManager::HashType VkRenderPassManager::ComputeHash(
-        const MG_State::GLState::FramebufferObject& fbo, Uint32 swapchainImageIndex, Bool includePendingClear) {
+        const MG_State::GLState::FramebufferObject& fbo, Uint32 swapchainImageIndex, Bool includePendingClear,
+        Bool includeDefaultFboDepthStencil) {
         XXHASH_VERIFY(XXH64_reset(m_hashState, m_config.CacheVersion));
         const Bool isDefaultFbo = fbo.IsDefaultFramebuffer();
         if (isDefaultFbo) {
@@ -560,9 +561,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                          attachment <= FramebufferAttachmentType::BackRight);
                     if (isDefaultColorAttachment) {
                         currentLayout = m_swapchainObject.GetImageLayout(swapchainImageIndex);
+                        // Content validity feeds the attachment's loadOp (see the
+                        // creation path), so it must key the cache as well.
+                        if (!m_swapchainObject.IsImageContentDefined(swapchainImageIndex)) {
+                            currentLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+                        }
                     } else if (attachment == FramebufferAttachmentType::Depth ||
                                attachment == FramebufferAttachmentType::Stencil) {
                         currentLayout = m_swapchainObject.GetDepthStencilImageLayout(swapchainImageIndex);
+                        if (!m_swapchainObject.IsDepthStencilContentDefined(swapchainImageIndex)) {
+                            currentLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+                        }
                     }
                 } else {
                     auto* textureResource = m_textureManager.SyncTextureAndGetDescriptor(*texture);
@@ -617,14 +626,49 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             combineFramebufferAttachmentObjHash(drawbuf);
         }
 
+        // The depth-less default-FBO flavor omits the depth/stencil attachment
+        // entirely, so it must hash differently from the depth-full flavor.
+        const Bool depthStencilIncluded = !isDefaultFbo || includeDefaultFboDepthStencil;
+        XXHASH_VERIFY(XXH64_update(m_hashState, &depthStencilIncluded, sizeof(depthStencilIncluded)));
+        if (depthStencilIncluded) {
             combineFramebufferAttachmentObjHash(FramebufferAttachmentType::Depth);
             combineFramebufferAttachmentObjHash(FramebufferAttachmentType::Stencil);
+        }
 
         return XXH64_digest(m_hashState);
     }
 
     RenderPassEntry& VkRenderPassManager::GetOrCreateRenderPass(const MG_State::GLState::FramebufferObject& fbo,
-                                                                Uint32 swapchainImageIndex) {
+                                                                Uint32 swapchainImageIndex,
+                                                                Bool drawUsesDepthStencil) {
+        // Resolve the default-FBO depth flavor (see the header comment): keep the
+        // depth attachment when the caller needs it, when a depth/stencil clear is
+        // pending, or when the active pass already carries it (escalate-only, so
+        // alternating depth-less draws never split an established depth pass).
+        Bool includeDefaultFboDepthStencil = true;
+        if (fbo.IsDefaultFramebuffer()) {
+            Bool activeDefaultHasDepthStencil = false;
+            if (const auto* active = GetActiveRenderPass()) {
+                Bool activeIsSwapchainPass = false;
+                Bool activeHasSwapchainDepthStencil = false;
+                for (const auto& tracked : active->trackedAttachmentLayouts) {
+                    activeIsSwapchainPass |= tracked.target == TrackedAttachmentTarget::SwapchainColor;
+                    activeHasSwapchainDepthStencil |=
+                        tracked.target == TrackedAttachmentTarget::SwapchainDepthStencil;
+                }
+                activeDefaultHasDepthStencil = activeIsSwapchainPass && activeHasSwapchainDepthStencil;
+            }
+            const auto& defaultDepthAtt = fbo.GetAttachment(FramebufferAttachmentType::Depth);
+            const auto& defaultStencilAtt = fbo.GetAttachment(FramebufferAttachmentType::Stencil);
+            const Bool pendingDepthStencilClear =
+                (defaultDepthAtt.IsTexture() && m_clearManager.HasPendingClear(defaultDepthAtt)) ||
+                HasPendingRenderbufferClear(defaultDepthAtt) ||
+                (defaultStencilAtt.IsTexture() && m_clearManager.HasPendingClear(defaultStencilAtt)) ||
+                HasPendingRenderbufferClear(defaultStencilAtt);
+            includeDefaultFboDepthStencil =
+                drawUsesDepthStencil || activeDefaultHasDepthStencil || pendingDepthStencilClear;
+        }
+
         auto hasPendingClearOnFramebuffer = [&]() -> Bool {
             const auto& drawBuffers = fbo.GetDrawBuffers();
             for (auto attachment : drawBuffers) {
@@ -674,6 +718,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_rpFastFboVersion == fbo.GetObjectVersion() && m_rpFastSwapchainIndex == swapchainImageIndex &&
             m_rpFastTexEpoch == m_textureManager.GetTextureImageEpoch() &&
             m_rpFastRbEpoch == m_renderbufferImageEpoch &&
+            (!fbo.IsDefaultFramebuffer() || m_rpFastHadDepthStencil == includeDefaultFboDepthStencil) &&
             m_rpFastRenderPassHash == activeRenderPass->hash && !hasPendingClearOnFramebuffer()) {
             auto activeIt = m_renderPasses.find(activeRenderPass->hash);
             if (activeIt != m_renderPasses.end()) {
@@ -682,7 +727,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             }
         }
 
-        auto compatibilityHash = ComputeHash(fbo, swapchainImageIndex, false);
+        auto compatibilityHash = ComputeHash(fbo, swapchainImageIndex, false, includeDefaultFboDepthStencil);
         if (activeRenderPass != nullptr &&
             activeRenderPass->CompatibleWith(compatibilityHash) &&
             !hasPendingClearOnFramebuffer()) {
@@ -699,10 +744,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_rpFastTexEpoch = m_textureManager.GetTextureImageEpoch();
             m_rpFastRbEpoch = m_renderbufferImageEpoch;
             m_rpFastRenderPassHash = activeRenderPass->hash;
+            m_rpFastHadDepthStencil = activeIt->second.hasDepthStencilAttachment;
             activeIt->second.lastUsedFrame = m_frameCounter;
             return activeIt->second;
         }
-        auto hash = ComputeHash(fbo, swapchainImageIndex, true);
+        auto hash = ComputeHash(fbo, swapchainImageIndex, true, includeDefaultFboDepthStencil);
         auto it = m_renderPasses.find(hash);
         if (it != m_renderPasses.end()) {
             it->second.lastUsedFrame = m_frameCounter;
@@ -894,6 +940,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                         MOBILEGL_ASSERT(swapchainImageIndex < swapchainViews.size(),
                                         "GetOrCreateRenderPass: swapchain image index out of range");
                         trackedColorLayout = m_swapchainObject.GetImageLayout(swapchainImageIndex);
+                        // EGL: a presented color buffer's content is undefined when its
+                        // image comes back around (EGL_BUFFER_DESTROYED, the default
+                        // swap behaviour) - skip the tile load instead of reloading
+                        // stale pixels nobody may rely on.
+                        if (!hasClear && !m_swapchainObject.IsImageContentDefined(swapchainImageIndex)) {
+                            trackedColorLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+                        }
                         trackedAttachmentLayouts.emplace_back(TrackedAttachmentLayoutInfo {
                             .target = TrackedAttachmentTarget::SwapchainColor,
                             .swapchainImageIndex = swapchainImageIndex,
@@ -913,6 +966,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                         trackedAttachmentLayouts.emplace_back(TrackedAttachmentLayoutInfo {
                             .target = TrackedAttachmentTarget::Texture,
                             .texture = att.GetTexture(),
+                            .textureRaw = att.GetTexture().get(),
                             .textureMipLevel = attachmentMipLevel,
                             .finalLayout = desc.finalLayout,
                         });
@@ -976,6 +1030,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         };
         const auto* selectedDepthStencilAttachment = isUsableDepthStencilAttachment(depthAtt) ? &depthAtt :
                                                      (isUsableDepthStencilAttachment(stencilAtt) ? &stencilAtt : nullptr);
+        // Depth-less default-FBO flavor: nothing in this pass touches depth/stencil
+        // and their content is undefined anyway (EGL swap), so drop the attachment
+        // and its whole tile load + store.
+        if (isDefaultFbo && !includeDefaultFboDepthStencil) {
+            selectedDepthStencilAttachment = nullptr;
+        }
         const Bool hasDistinctDepthAndStencilAttachments =
             isUsableDepthStencilAttachment(depthAtt) && isUsableDepthStencilAttachment(stencilAtt) &&
             !sameDepthStencilAttachmentObject(depthAtt, stencilAtt);
@@ -994,6 +1054,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             VkImageLayout trackedDepthLayout = isDefaultFbo ?
                 m_swapchainObject.GetDepthStencilImageLayout(swapchainImageIndex) :
                 VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL;
+            // EGL 1.5 §3.10.1: every ancillary (depth/stencil) buffer's content is
+            // undefined after a swap, so the first default-FBO pass of a frame can
+            // skip the depth/stencil tile load outright.
+            if (isDefaultFbo && !m_swapchainObject.IsDepthStencilContentDefined(swapchainImageIndex)) {
+                trackedDepthLayout = VK_IMAGE_LAYOUT_UNDEFINED;
+            }
             depthAttachmentDescription.flags = 0;
             VkSampleCountFlagBits depthAttachmentSampleCount = VK_SAMPLE_COUNT_1_BIT;
             Int depthAttachmentId = 0;
@@ -1077,6 +1143,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 trackedAttachmentLayouts.emplace_back(TrackedAttachmentLayoutInfo {
                     .target = TrackedAttachmentTarget::Texture,
                     .texture = selectedDepthStencilAttachment->GetTexture(),
+                    .textureRaw = selectedDepthStencilAttachment->GetTexture().get(),
                     .textureMipLevel = attachmentMipLevel,
                     .finalLayout = depthAttachmentDescription.finalLayout,
                 });
@@ -1122,6 +1189,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         const Bool hasDepthStencilAttachment = depthAttachmentRef.attachment != VK_ATTACHMENT_UNUSED;
 
+        // Declare only the used colour-reference span. The GL draw-buffer array
+        // always spans 8 slots, so passes used to declare colorAttachmentCount=8
+        // with trailing VK_ATTACHMENT_UNUSED holes - and Adreno configures its
+        // per-pixel render-backend/export path from the DECLARED count, so every
+        // fragment of every pass paid the 8-target export cost (measured on
+        // Adreno 650 / MC 26.2: 11.9 -> 7.5 ms of GPU time per frame, with the
+        // single-quad swapchain blit pass alone dropping 1.26 -> 0.40 ms).
+        // Interior GL_NONE holes keep their slots so fragment-output locations
+        // still line up; a fragment output at a location past the trimmed count
+        // is discarded, which is exactly GL's semantic for writing to a draw
+        // buffer set to GL_NONE.
+        while (!colorAttachmentRefs.empty() &&
+               colorAttachmentRefs.back().attachment == VK_ATTACHMENT_UNUSED) {
+            colorAttachmentRefs.pop_back();
+        }
+
         // Subpass
         VkSubpassDescription subpassDesc;
         subpassDesc.flags = 0;
@@ -1330,6 +1413,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         renderPassBeginInfo.pClearValues = clearValues.data();
 
         vkCmdBeginRenderPass(commandBuffer, &renderPassBeginInfo, VK_SUBPASS_CONTENTS_INLINE);
+        // Pre-pass stream bookkeeping: this pass's attachment images are now
+        // referenced by the open frame recording.
+        if (s_textureManager != nullptr) {
+            for (const auto& tracked : renderPassEntry.trackedAttachmentLayouts) {
+                if (tracked.target == TrackedAttachmentTarget::Texture) {
+                    if (const auto texture = tracked.texture.lock()) {
+                        s_textureManager->StampTextureRecordingUse(texture.get());
+                    }
+                }
+            }
+        }
         for (const auto& pending: renderPassEntry.pendingClearAttachments) {
             if (pending.hasInlinePayload) {
                 if (s_renderPassManager != nullptr) {
@@ -1382,11 +1476,15 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                     case TrackedAttachmentTarget::SwapchainColor:
                         MOBILEGL_ASSERT(s_swapchainObject != nullptr, "EndRenderPass: swapchain object is null");
                         s_swapchainObject->SetImageLayout(trackedAttachment.swapchainImageIndex, trackedAttachment.finalLayout);
+                        // The pass stored into the attachment: its content is defined
+                        // until the image is next presented.
+                        s_swapchainObject->SetImageContentDefined(trackedAttachment.swapchainImageIndex, true);
                         break;
                     case TrackedAttachmentTarget::SwapchainDepthStencil:
                         MOBILEGL_ASSERT(s_swapchainObject != nullptr, "EndRenderPass: swapchain object is null");
                         s_swapchainObject->SetDepthStencilImageLayout(trackedAttachment.swapchainImageIndex,
                                                                       trackedAttachment.finalLayout);
+                        s_swapchainObject->SetDepthStencilContentDefined(trackedAttachment.swapchainImageIndex, true);
                         break;
                     default:
                         MOBILEGL_ASSERT(false, "EndRenderPass: unsupported tracked attachment target=%d",

MobileGL/MG_Backend/DirectVulkan/Renderer/VkRenderPassManager.h

@@ -42,6 +42,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     struct TrackedAttachmentLayoutInfo {
         TrackedAttachmentTarget target = TrackedAttachmentTarget::Texture;
         WeakPtr<MG_State::GLState::ITextureObject> texture;
+        // Identity-compare shortcut for the per-draw "does the active pass use
+        // this sampled texture" probe: comparing this against a LIVE texture's
+        // address needs no weak_ptr::lock (two refcount atomics per probe).
+        // May dangle once the texture dies - compare only, never dereference.
+        MG_State::GLState::ITextureObject* textureRaw = nullptr;
         WeakPtr<MG_State::GLState::RenderbufferObject> renderbuffer;
         Uint32 textureMipLevel = 0;
         Uint32 swapchainImageIndex = 0;
@@ -188,8 +193,22 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         HashType ComputeHash(
             const MG_State::GLState::FramebufferObject& fbo,
             Uint32 swapchainImageIndex,
-            Bool includePendingClear = true);
+            Bool includePendingClear = true,
-        RenderPassEntry& GetOrCreateRenderPass(const MG_State::GLState::FramebufferObject& fbo, Uint32 swapchainImageIndex);
+            Bool includeDefaultFboDepthStencil = true);
+        // drawUsesDepthStencil: whether the operation about to run inside the pass
+        // reads or writes the depth/stencil buffer (depth test or stencil test
+        // enabled, or a depth/stencil clear). Only consulted for the DEFAULT
+        // framebuffer: EGL undefines its ancillary buffers at every swap, so a
+        // default-FBO pass whose draws provably never touch depth/stencil is
+        // created WITHOUT the depth attachment - on a tiler that skips the whole
+        // depth tile load AND store. The flavor only escalates: once a pass with
+        // depth is active, later depth-less draws keep using it, and a depth-using
+        // draw against a depth-less active pass resolves to a new (incompatible)
+        // entry, which the caller's compatibility check turns into a pass split;
+        // the new pass's depth loads DONT_CARE (content was undefined all along).
+        RenderPassEntry& GetOrCreateRenderPass(const MG_State::GLState::FramebufferObject& fbo,
+                                               Uint32 swapchainImageIndex,
+                                               Bool drawUsesDepthStencil = true);
         void QueueRenderbufferClear(GLbitfield mask, const ClearFramebufferPayload& clearPayload,
                                     const MG_State::GLState::FramebufferObject& drawFbo);
         void QueueRenderbufferClear(const ClearAttachmentPayload& clearPayload,
@@ -219,6 +238,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // image recreation.
         Uint64 m_renderbufferImageEpoch = 1;
 
+    public:
+        // Bumped whenever a renderbuffer backing is (re)created; consecutive-draw
+        // snapshots include it so an attachment respecify forces a re-resolve.
+        Uint64 GetRenderbufferImageEpoch() const { return m_renderbufferImageEpoch; }
+
+    private:
+
         // Per-draw fast-path memo for GetOrCreateRenderPass (dirty-flag state tracking): when the
         // framebuffer state is provably unchanged since the last resolution, the active render pass
         // is reused WITHOUT recomputing the expensive per-draw hash. Invalidated by FBO switch /
@@ -231,6 +257,10 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Uint64 m_rpFastTexEpoch = 0;
         Uint64 m_rpFastRbEpoch = 0;
         Uint64 m_rpFastRenderPassHash = 0;
+        // Whether the memoized entry carries a depth/stencil attachment; a
+        // default-FBO resolution whose effective depth request differs must
+        // miss the memo (the depth-less/depth-full flavors hash differently).
+        Bool m_rpFastHadDepthStencil = false;
 
     public:
         struct RenderbufferResource {

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.cpp

@@ -607,7 +607,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     }
 
     void VkTextureManager::Shutdown() {
+        if (m_device != VK_NULL_HANDLE) {
+            ReclaimCompletedUploads(/*waitAll=*/true);
+        }
         DestroyDeferredReleases();
+        ++m_resourceEraseEpoch;  // every memoized resource pointer dies with the map
         m_textureResources.clear();
         m_aliveObjects.clear();
         m_storageImageTextures.clear();
@@ -629,6 +633,7 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                         frameIndex, m_deferredViewReleases.size());
         m_currentFrameIndex = frameIndex;
         CollectDeferredReleases(frameIndex);
+        ReclaimCompletedUploads();
 
         // Frame-boundary GC: every 64 frame boundaries (~1 s at 60 fps) bounds the reclaim
         // latency for dead textures regardless of draw traffic — workloads that churn
@@ -659,6 +664,9 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         }
         m_aliveObjects.erase(identity);
         m_storageImageTextures.erase(identity);
+        // Invalidate every cross-draw sampled-texture memo: the erased
+        // resource's address may be reused by a future emplace.
+        ++m_resourceEraseEpoch;
     }
 
     void VkTextureManager::PruneStaleTextureAliases(MG_State::GLState::ITextureObject* texture) {
@@ -714,6 +722,19 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             }
         }
 
+        // Cross-draw memo probe (see SyncedTextureMemoEntry): skips both map
+        // lookups and the (re)registration path for repeat-bound textures.
+        TextureResource* resourcePtr = nullptr;
+        for (Uint32 i = 0; i < kSyncedTextureMemoSize; ++i) {
+            const SyncedTextureMemoEntry& memo = m_syncedTextureMemo[i];
+            if (memo.texture == &texture && memo.lifetimeId == identity.lifetimeId &&
+                memo.eraseEpoch == m_resourceEraseEpoch) {
+                resourcePtr = memo.resource;
+                break;
+            }
+        }
+
+        if (resourcePtr == nullptr) {
             auto aliveIt = m_aliveObjects.find(identity);
             if (aliveIt != m_aliveObjects.end() && aliveIt->second.expired()) {
                 EraseTrackedTexture(aliveIt->first);
@@ -751,8 +772,13 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 auto [insertIt, _] = m_textureResources.emplace(identity, Move(initial));
                 it = insertIt;
             }
+            resourcePtr = &(it->second);
+            m_syncedTextureMemo[m_syncedTextureMemoNext] =
+                SyncedTextureMemoEntry{&texture, identity.lifetimeId, m_resourceEraseEpoch, resourcePtr};
+            m_syncedTextureMemoNext = (m_syncedTextureMemoNext + 1) % kSyncedTextureMemoSize;
+        }
 
-        if (!SyncTexture(texture, it->second)) {
+        if (!SyncTexture(texture, *resourcePtr)) {
             MGLOG_D("%s: Syncing texture %d failed", __func__, texture.GetExternalIndex());
             return nullptr;
         }
@@ -766,11 +792,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 }
             }
             if (!recorded) {
-                m_drawSyncedThisDraw.push_back({identity, &(it->second)});
+                m_drawSyncedThisDraw.push_back({identity, resourcePtr});
             }
         }
 
-        return &(it->second);
+        return resourcePtr;
     }
 
     VkImageView VkTextureManager::GetOrCreateViewAtMipLevel(MG_State::GLState::ITextureObject& texture, Uint32 mipLevel) {
@@ -1049,6 +1075,16 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return view;
     }
 
+    void VkTextureManager::StampTextureRecordingUse(MG_State::GLState::ITextureObject* texture) {
+        if (texture == nullptr) {
+            return;
+        }
+        auto it = m_textureResources.find(MakeTextureIdentity(texture));
+        if (it != m_textureResources.end()) {
+            it->second.lastRecordingGeneration = m_recordingGeneration;
+        }
+    }
+
     void VkTextureManager::UpdateTrackedImageLayout(MG_State::GLState::ITextureObject* texture, VkImageLayout newLayout) {
         MOBILEGL_ASSERT(texture != nullptr, "UpdateTrackedImageLayout: texture is null");
         auto it = m_textureResources.find(MakeTextureIdentity(texture));
@@ -1076,6 +1112,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         MOBILEGL_ASSERT(writtenMipLevel < resource.mipLevels,
                         "UpdateTrackedImageLayoutAfterAttachmentWrite: textureId=%d mipLevel=%u out of range %u",
                         texture->GetExternalIndex(), writtenMipLevel, resource.mipLevels);
+        // Pre-pass stream bookkeeping: the render pass that just ended wrote this image.
+        StampResourceRecordingUse(resource);
 
         if (resource.layout != newLayout && resource.mipLevels > 1) {
             VkPipelineStageFlags srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
@@ -1160,6 +1198,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                               VK_ACCESS_SHADER_READ_BIT, resource->aspect, 0, resource->mipLevels,
                                               resource->arrayLayers);
         MOBILEGL_ASSERT(ok, "TransitionTextureForSampling: transition failed for textureId=%d", texture.GetExternalIndex());
+        // Pre-pass stream bookkeeping: a command referencing the image was recorded.
+        StampResourceRecordingUse(*resource);
         return ok;
     }
 
@@ -1189,6 +1229,8 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                                               resource->aspect, 0, resource->mipLevels, resource->arrayLayers);
         MOBILEGL_ASSERT(ok, "TransitionTextureForStorageImage: transition failed for textureId=%d",
                         texture.GetExternalIndex());
+        // Pre-pass stream bookkeeping: a command referencing the image was recorded.
+        StampResourceRecordingUse(*resource);
         return ok;
     }
 
@@ -1420,8 +1462,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             return false;
         }
         const Bool isMultisampleTexture = IsMultisampleTextureUploadTarget(uploadTarget);
+        // A texture that has only ever defined level 0 gets a single-level backing
+        // (ANGLE's model). Preallocating the full chain put every render target
+        // onto Adreno's multi-mip image layout and grew each texture by a third
+        // for levels most textures never define. Once a second level is defined
+        // the backing is recreated ONE time with the full chain (the
+        // preserve-copy path below carries the pixels over), so sequentially-
+        // defined atlas mips do not recreate per level, and glGenerateMipmap -
+        // which defines every level before syncing - works unchanged.
         const Uint32 backingMipLevels =
-            isMultisampleTexture ? 1u : std::max(mipLevels, ComputeFullMipLevelCount(texelSize));
+            isMultisampleTexture ? 1u
+            : (mipLevels > 1 ? std::max(mipLevels, ComputeFullMipLevelCount(texelSize)) : 1u);
         TextureShapeInfo shapeInfo{};
         const Bool supportedShape = TryResolveTextureShapeInfo(texture, uploadTarget, texelSize, shapeInfo);
         MOBILEGL_ASSERT(supportedShape,
@@ -1683,6 +1734,28 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         m_deferredViewReleases[frameIndex].clear();
     }
 
+    void VkTextureManager::ReclaimCompletedUploads(Bool waitAll) {
+        if (m_pendingUploadReclaims.empty()) {
+            return;
+        }
+
+        SizeT completed = 0;
+        for (; completed < m_pendingUploadReclaims.size(); ++completed) {
+            PendingUploadReclaim& entry = m_pendingUploadReclaims[completed];
+            if (waitAll) {
+                VK_VERIFY(vkWaitForFences(m_device, 1, &entry.fence, VK_TRUE, UINT64_MAX),
+                          "vkWaitForFences(texture upload reclaim)");
+            } else if (vkGetFenceStatus(m_device, entry.fence) != VK_SUCCESS) {
+                break;
+            }
+            vkDestroyFence(m_device, entry.fence, nullptr);
+            vkFreeCommandBuffers(m_device, m_commandPool, 1, &entry.commandBuffer);
+            vmaDestroyBuffer(m_allocator, entry.stagingBuffer, entry.stagingAllocation);
+        }
+        m_pendingUploadReclaims.erase(m_pendingUploadReclaims.begin(),
+                                      m_pendingUploadReclaims.begin() + static_cast<std::ptrdiff_t>(completed));
+    }
+
     void VkTextureManager::DestroyDeferredReleases() {
         for (auto& deferredReleases : m_deferredReleases) {
             deferredReleases.clear();
@@ -1989,11 +2062,23 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         VK_VERIFY(vkCreateFence(m_device, &fenceInfo, nullptr, &uploadFence), "vkCreateFence(texture upload)");
 
         VK_VERIFY(vkQueueSubmit(m_graphicsQueue, 1, &submitInfo, uploadFence), "vkQueueSubmit(texture)");
-        VK_VERIFY(vkWaitForFences(m_device, 1, &uploadFence, VK_TRUE, UINT64_MAX), "vkWaitForFences(texture upload)");
+        // Do NOT wait the fence here: this submit sits behind the previous
-        vkDestroyFence(m_device, uploadFence, nullptr);
+        // frame's rendering on the queue, so a synchronous wait stalls the CPU
-        vkFreeCommandBuffers(m_device, m_commandPool, 1, &commandBuffer);
+        // until the GPU drains - a per-frame vkQueueWaitIdle for any workload
-
+        // with animated textures. Ordering against the current frame's draws is
-        vmaDestroyBuffer(m_allocator, stagingBuffer, stagingAllocation);
+        // already guaranteed (its command buffer is submitted later, at
+        // present), so only the transient objects need to survive execution;
+        // park them until the fence signals.
+        m_pendingUploadReclaims.push_back({uploadFence, commandBuffer, stagingBuffer, stagingAllocation});
+        ReclaimCompletedUploads();
+        // Backstop for pathological upload storms: bound in-flight staging
+        // memory by blocking on the oldest upload only once the list is deep.
+        constexpr SizeT kMaxPendingTextureUploads = 16;
+        if (m_pendingUploadReclaims.size() > kMaxPendingTextureUploads) {
+            VK_VERIFY(vkWaitForFences(m_device, 1, &m_pendingUploadReclaims.front().fence, VK_TRUE, UINT64_MAX),
+                      "vkWaitForFences(texture upload backstop)");
+            ReclaimCompletedUploads();
+        }
 
         if (!ok) {
             MGLOG_D("%s: texture upload cmd failed", __func__);

MobileGL/MG_Backend/DirectVulkan/Renderer/VkTextureManager.h

@@ -28,6 +28,9 @@ public:
     // manager keys its per-draw fast path on this so an attachment's image recreation
     // invalidates the cached render pass (dirty-flag tracking; portable to Vulkan 1.1).
     Uint64 GetTextureImageEpoch() const { return m_textureImageEpoch; }
+    // Bumped whenever any tracked texture resource is erased; cached
+    // TextureResource pointers are valid only while this is unchanged.
+    Uint64 GetResourceEraseEpoch() const { return m_resourceEraseEpoch; }
 
     struct TextureIdentity {
         MG_State::GLState::ITextureObject* texture = nullptr;
@@ -172,6 +175,13 @@ public:
         // NeedsStorageImagePreparation cannot ask for a recreate that will never happen.
         Bool storageUsageResolved = false;
         Uint16 syncedTextureParamsVersion = 0;
+        // Recording generation (VkTextureManager::GetRecordingGeneration) of the last
+        // command referencing this image that was recorded into the CURRENT frame
+        // command buffer. An image untouched by the open recording may have its
+        // out-of-pass work (deferred clears, sampled-layout transitions) recorded
+        // into the frame's PRE command buffer - which executes strictly before the
+        // frame's commands - instead of splitting the active render pass.
+        Uint64 lastRecordingGeneration = 0;
         // Snapshot of ITextureObject::GetContentVersion() at the last successful sync;
         // lets SyncTexture skip the whole re-check/re-upload when content is unchanged.
         Uint64 syncedContentVersion = 0;
@@ -207,6 +217,7 @@ public:
             std::swap(this->usageFlags, that.usageFlags);
             std::swap(this->storageUsageResolved, that.storageUsageResolved);
             std::swap(this->syncedTextureParamsVersion, that.syncedTextureParamsVersion);
+            std::swap(this->lastRecordingGeneration, that.lastRecordingGeneration);
             std::swap(this->syncedContentVersion, that.syncedContentVersion);
             std::swap(this->syncedMipLevelCount, that.syncedMipLevelCount);
         }
@@ -307,6 +318,21 @@ public:
                                                       VkImageLayout newLayout);
     Bool TransitionTextureForSampling(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
     Bool TransitionTextureForStorageImage(VkCommandBuffer commandBuffer, MG_State::GLState::ITextureObject& texture);
+
+    // Recording-generation bookkeeping for the pre-pass command stream. The
+    // generation advances every time the frame command buffer (re)begins
+    // recording; a resource whose stamp does not match was not referenced by
+    // any command in the open recording, so its out-of-pass work may safely
+    // execute ahead of the whole recording (in the pre command buffer).
+    void AdvanceRecordingGeneration() { ++m_recordingGeneration; }
+    void StampResourceRecordingUse(TextureResource& resource) const {
+        resource.lastRecordingGeneration = m_recordingGeneration;
+    }
+    // Map-lookup variant for callers that only hold the GL texture object.
+    void StampTextureRecordingUse(MG_State::GLState::ITextureObject* texture);
+    Bool WasTouchedThisRecording(const TextureResource& resource) const {
+        return resource.lastRecordingGeneration == m_recordingGeneration;
+    }
     // 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
@@ -364,6 +390,9 @@ public:
 private:
     // Bumped in SyncTextureResource right after vmaCreateImage(texture). See GetTextureImageEpoch().
     Uint64 m_textureImageEpoch = 1;
+    // See AdvanceRecordingGeneration. Starts above every resource's default
+    // stamp of 0 so a fresh resource counts as untouched.
+    Uint64 m_recordingGeneration = 1;
 
     Bool SyncTexture(MG_State::GLState::ITextureObject &texture,
                      TextureResource &outResource);
@@ -394,6 +423,11 @@ private:
     void DeferViewRelease(VkImageView view);
     void CollectDeferredReleases(Uint32 frameIndex);
     void DestroyDeferredReleases();
+    // Frees the fence/command buffer/staging buffer of every in-flight texture
+    // upload whose fence has signaled (submission order = completion order on
+    // the single queue, so the scan stops at the first still-pending entry).
+    // waitAll blocks on every entry - Shutdown's drain.
+    void ReclaimCompletedUploads(Bool waitAll = false);
     static TextureIdentity MakeTextureIdentity(MG_State::GLState::ITextureObject* texture);
     void EraseTrackedTexture(const TextureIdentity& identity);
     void PruneStaleTextureAliases(MG_State::GLState::ITextureObject* texture);
@@ -423,6 +457,23 @@ private:
         TextureResource* resource = nullptr;
     };
     Vector<DrawSyncedTexture> m_drawSyncedThisDraw;
+    // Cross-draw sampled-texture memo: the same few textures (atlas, lightmap)
+    // are resolved on every draw, so cache their resource pointers and skip the
+    // alive/resource map lookups. Node-based std::unordered_map keeps the
+    // pointees stable across inserts; erases bump m_resourceEraseEpoch, which
+    // every memo entry must match. SyncTexture still runs on memo hits, so
+    // content/param freshness is unaffected. A dead-then-reused texture address
+    // cannot false-hit: the new object carries a new lifetime id.
+    struct SyncedTextureMemoEntry {
+        const MG_State::GLState::ITextureObject* texture = nullptr;
+        Uint64 lifetimeId = 0;
+        Uint64 eraseEpoch = 0;
+        TextureResource* resource = nullptr;
+    };
+    static constexpr Uint32 kSyncedTextureMemoSize = 8;
+    SyncedTextureMemoEntry m_syncedTextureMemo[kSyncedTextureMemoSize];
+    Uint32 m_syncedTextureMemoNext = 0;
+    Uint64 m_resourceEraseEpoch = 1;
     // Formats whose mutable-image probe failed on this device; their images are created
     // without MUTABLE_FORMAT_BIT so repeat syncs neither re-probe nor flag-mismatch.
     std::unordered_set<VkFormat> m_mutableFormatUnsupported;
@@ -432,5 +483,16 @@ private:
     std::unordered_set<TextureIdentity, TextureIdentityHash> m_storageImageTextures;
     Vector<Vector<TextureResource>> m_deferredReleases;
     Vector<Vector<VkImageView>> m_deferredViewReleases;
+    // Texture uploads are submitted out-of-band but NOT waited on (waiting
+    // behind the queue serialized the CPU against the previous frame's GPU
+    // work every time an animated atlas re-uploaded). Their transient objects
+    // are parked here and reclaimed once the upload fence signals.
+    struct PendingUploadReclaim {
+        VkFence fence = VK_NULL_HANDLE;
+        VkCommandBuffer commandBuffer = VK_NULL_HANDLE;
+        VkBuffer stagingBuffer = VK_NULL_HANDLE;
+        VmaAllocation stagingAllocation = nullptr;
+    };
+    Vector<PendingUploadReclaim> m_pendingUploadReclaims;
 };
 } // namespace MobileGL::MG_Backend::DirectVulkan

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.cpp

@@ -217,6 +217,73 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         return static_cast<Int>((static_cast<Int64>(value) * toExtent + fromExtent / 2) / fromExtent);
     }
 
+    // Redundant dynamic-state elimination for the per-draw hot path: within one
+    // command-buffer recording, a vkCmdSet* whose values already match what the
+    // command buffer holds is skipped. Valid because every PipelineFactory
+    // pipeline declares the same eight dynamic states, so the values persist
+    // across those pipeline binds; the shadow resets whenever a recording
+    // (re)begins, and whenever an auxiliary pipeline with a narrower dynamic
+    // set (blit, depth-mipmap) binds - their static state makes the
+    // corresponding dynamic values undefined per the spec.
+    struct DynamicStateShadow {
+        // Last graphics pipeline bound on the frame command buffer. Pipeline
+        // binds are command-buffer state (they survive render-pass boundaries),
+        // so the same reset points that invalidate dynamic state - recording
+        // (re)begin and the aux blit pipelines' raw binds - are exactly the
+        // points where this becomes unknown.
+        Bool graphicsPipelineValid = false;
+        VkPipeline graphicsPipeline = VK_NULL_HANDLE;
+        // Index/vertex buffer binds are command-buffer state too. Terrain
+        // sections and GUI quads share one sequential index buffer, and GUI
+        // batches often reuse a vertex arena buffer, so skipping identical
+        // rebinds removes a large share of per-draw driver calls.
+        Bool indexBindValid = false;
+        VkBuffer indexBuffer = VK_NULL_HANDLE;
+        VkDeviceSize indexOffset = 0;
+        VkIndexType indexType = VK_INDEX_TYPE_MAX_ENUM;
+        static constexpr Uint32 kMaxShadowedVertexBindings = 8;
+        Bool vertexBindValid = false;
+        Uint32 vertexBindingCount = 0;
+        VkBuffer vertexBuffers[kMaxShadowedVertexBindings] = {};
+        VkDeviceSize vertexOffsets[kMaxShadowedVertexBindings] = {};
+        Bool viewportValid = false;
+        VkViewport viewport{};
+        Bool scissorValid = false;
+        VkRect2D scissor{};
+        Bool blendConstantsValid = false;
+        Float blendConstants[4] = {0.0f, 0.0f, 0.0f, 0.0f};
+        Bool depthBiasValid = false;
+        Float depthBiasConstantFactor = 0.0f;
+        Float depthBiasSlopeFactor = 0.0f;
+        Bool lineWidthValid = false;
+        Float lineWidth = 0.0f;
+        Bool stencilValid = false;
+        Uint32 stencilFrontCompareMask = 0;
+        Uint32 stencilBackCompareMask = 0;
+        Uint32 stencilFrontWriteMask = 0;
+        Uint32 stencilBackWriteMask = 0;
+        Uint32 stencilFrontReference = 0;
+        Uint32 stencilBackReference = 0;
+    };
+    static DynamicStateShadow g_dynamicStateShadow;
+
+    static void ResetDynamicStateShadow() {
+        g_dynamicStateShadow = {};
+    }
+
+    static void ShadowedSetScissor(VkCommandBuffer commandBuffer, const VkRect2D& scissor) {
+        auto& shadow = g_dynamicStateShadow;
+        if (shadow.scissorValid && shadow.scissor.offset.x == scissor.offset.x &&
+            shadow.scissor.offset.y == scissor.offset.y &&
+            shadow.scissor.extent.width == scissor.extent.width &&
+            shadow.scissor.extent.height == scissor.extent.height) {
+            return;
+        }
+        shadow.scissorValid = true;
+        shadow.scissor = scissor;
+        vkCmdSetScissor(commandBuffer, 0, 1, &scissor);
+    }
+
     static void ApplyGLViewportState(VkCommandBuffer commandBuffer,
                                      const IntVec2& framebufferExtent,
                                      VkSurfaceTransformFlagBitsKHR preTransform,
@@ -246,6 +313,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         viewport.height = static_cast<float>(viewportHeight);
         viewport.minDepth = depthRange.x();
         viewport.maxDepth = depthRange.y();
+        auto& shadow = g_dynamicStateShadow;
+        if (shadow.viewportValid && shadow.viewport.x == viewport.x && shadow.viewport.y == viewport.y &&
+            shadow.viewport.width == viewport.width && shadow.viewport.height == viewport.height &&
+            shadow.viewport.minDepth == viewport.minDepth && shadow.viewport.maxDepth == viewport.maxDepth) {
+            return;
+        }
+        shadow.viewportValid = true;
+        shadow.viewport = viewport;
         vkCmdSetViewport(commandBuffer, 0, 1, &viewport);
     }
 
@@ -257,6 +332,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             blendColor.z(),
             blendColor.w(),
         };
+        auto& shadow = g_dynamicStateShadow;
+        if (shadow.blendConstantsValid && shadow.blendConstants[0] == blendConstants[0] &&
+            shadow.blendConstants[1] == blendConstants[1] && shadow.blendConstants[2] == blendConstants[2] &&
+            shadow.blendConstants[3] == blendConstants[3]) {
+            return;
+        }
+        shadow.blendConstantsValid = true;
+        shadow.blendConstants[0] = blendConstants[0];
+        shadow.blendConstants[1] = blendConstants[1];
+        shadow.blendConstants[2] = blendConstants[2];
+        shadow.blendConstants[3] = blendConstants[3];
         vkCmdSetBlendConstants(commandBuffer, blendConstants);
     }
 
@@ -272,8 +358,17 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     }
 
     static void ApplyPolygonOffsetState(VkCommandBuffer commandBuffer) {
-        vkCmdSetDepthBias(commandBuffer, MG_State::pGLContext->GetPolygonOffsetUnits(), 0.0f,
+        const Float constantFactor = MG_State::pGLContext->GetPolygonOffsetUnits();
-                          MG_State::pGLContext->GetPolygonOffsetFactor());
+        const Float slopeFactor = MG_State::pGLContext->GetPolygonOffsetFactor();
+        auto& shadow = g_dynamicStateShadow;
+        if (shadow.depthBiasValid && shadow.depthBiasConstantFactor == constantFactor &&
+            shadow.depthBiasSlopeFactor == slopeFactor) {
+            return;
+        }
+        shadow.depthBiasValid = true;
+        shadow.depthBiasConstantFactor = constantFactor;
+        shadow.depthBiasSlopeFactor = slopeFactor;
+        vkCmdSetDepthBias(commandBuffer, constantFactor, 0.0f, slopeFactor);
     }
 
     static void ApplyLineWidthState(VkCommandBuffer commandBuffer) {
@@ -288,6 +383,12 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                 lineWidth = maxLineWidth;
             }
         }
+        auto& shadow = g_dynamicStateShadow;
+        if (shadow.lineWidthValid && shadow.lineWidth == lineWidth) {
+            return;
+        }
+        shadow.lineWidthValid = true;
+        shadow.lineWidth = lineWidth;
         vkCmdSetLineWidth(commandBuffer, lineWidth);
     }
 
@@ -336,15 +437,31 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     static void ApplyStencilState(VkCommandBuffer commandBuffer) {
         const StencilFaceState& frontStencil = MG_State::pGLContext->GetStencilState(StencilFace::Front);
         const StencilFaceState& backStencil = MG_State::pGLContext->GetStencilState(StencilFace::Back);
+        const Uint32 frontReference = static_cast<Uint32>(std::max(frontStencil.Ref, 0));
+        const Uint32 backReference = static_cast<Uint32>(std::max(backStencil.Ref, 0));
+
+        auto& shadow = g_dynamicStateShadow;
+        if (shadow.stencilValid && shadow.stencilFrontCompareMask == frontStencil.ValueMask &&
+            shadow.stencilBackCompareMask == backStencil.ValueMask &&
+            shadow.stencilFrontWriteMask == frontStencil.WriteMask &&
+            shadow.stencilBackWriteMask == backStencil.WriteMask &&
+            shadow.stencilFrontReference == frontReference && shadow.stencilBackReference == backReference) {
+            return;
+        }
+        shadow.stencilValid = true;
+        shadow.stencilFrontCompareMask = frontStencil.ValueMask;
+        shadow.stencilBackCompareMask = backStencil.ValueMask;
+        shadow.stencilFrontWriteMask = frontStencil.WriteMask;
+        shadow.stencilBackWriteMask = backStencil.WriteMask;
+        shadow.stencilFrontReference = frontReference;
+        shadow.stencilBackReference = backReference;
 
         vkCmdSetStencilCompareMask(commandBuffer, VK_STENCIL_FACE_FRONT_BIT, frontStencil.ValueMask);
         vkCmdSetStencilCompareMask(commandBuffer, VK_STENCIL_FACE_BACK_BIT, backStencil.ValueMask);
         vkCmdSetStencilWriteMask(commandBuffer, VK_STENCIL_FACE_FRONT_BIT, frontStencil.WriteMask);
         vkCmdSetStencilWriteMask(commandBuffer, VK_STENCIL_FACE_BACK_BIT, backStencil.WriteMask);
-        vkCmdSetStencilReference(commandBuffer, VK_STENCIL_FACE_FRONT_BIT,
+        vkCmdSetStencilReference(commandBuffer, VK_STENCIL_FACE_FRONT_BIT, frontReference);
-                                 static_cast<Uint32>(std::max(frontStencil.Ref, 0)));
+        vkCmdSetStencilReference(commandBuffer, VK_STENCIL_FACE_BACK_BIT, backReference);
-        vkCmdSetStencilReference(commandBuffer, VK_STENCIL_FACE_BACK_BIT,
-                                 static_cast<Uint32>(std::max(backStencil.Ref, 0)));
     }
 
     enum class NumericDomain {
@@ -718,16 +835,6 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     static_assert(kMaxVertexAttribs <= ProgramFactory::VkProgramObject::kMaxVertexInputLocations,
                   "vertexInputTypes is indexed by vertex attribute location");
 
-    static Uint32 BuildVertexInputAttributeMask(const Vector<VkVertexInputAttributeDescription>& attributes) {
-        Uint32 attributeMask = 0;
-        for (const auto& attribute : attributes) {
-            if (attribute.location < kMaxVertexAttribs) {
-                attributeMask |= (1u << attribute.location);
-            }
-        }
-        return attributeMask;
-    }
-
     static Bool TryGetCurrentVertexAttributeFormat(GLenum glType, VkFormat& outFormat) {
         switch (glType) {
         case GL_FLOAT:
@@ -878,8 +985,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (trackedAttachment.target != TrackedAttachmentTarget::Texture) {
                 continue;
             }
-            const auto trackedTexture = trackedAttachment.texture.lock();
+            // Raw identity compare (see textureRaw): the caller's texture is
-            if (trackedTexture && trackedTexture.get() == &texture) {
+            // live, so a dangling tracked pointer can never equal its address
+            // unless the allocator reused it - and that false positive merely
+            // ends the render pass early, never misses a genuine use.
+            if (trackedAttachment.textureRaw == &texture) {
                 return true;
             }
         }
@@ -2814,7 +2924,7 @@ void main() {
         // the GetCurrentProgram + GetOrCreateProgram hash lookup every draw.
         auto& vertexInputState = m_vertexInputStateFactory->GetOrCreateVertexInputState(vao);
         const Uint32 activeAttribMask = programObj.activeVertexInputLocationMask;
-        const Uint32 vertexInputAttribMask = BuildVertexInputAttributeMask(vertexInputState.attributes);
+        const Uint32 vertexInputAttribMask = vertexInputState.attributeLocationMask;
         const Uint32 missingAttribMask = activeAttribMask & ~vertexInputAttribMask;
 
         const auto bindingCount = vertexInputState.bindings.size() + static_cast<SizeT>(std::popcount(missingAttribMask));
@@ -3079,8 +3189,29 @@ void main() {
         }
 
         if (bindingCount > 0) {
-            vkCmdBindVertexBuffers(commandBuffer, 0, static_cast<Uint32>(bindingCount), vkBuffers.data(),
+            auto& shadow = g_dynamicStateShadow;
-                                   vkOffsets.data());
+            const Uint32 count = static_cast<Uint32>(bindingCount);
+            Bool identical = shadow.vertexBindValid && shadow.vertexBindingCount == count &&
+                             count <= DynamicStateShadow::kMaxShadowedVertexBindings;
+            if (identical) {
+                for (Uint32 i = 0; i < count; ++i) {
+                    if (shadow.vertexBuffers[i] != vkBuffers[i] || shadow.vertexOffsets[i] != vkOffsets[i]) {
+                        identical = false;
+                        break;
+                    }
+                }
+            }
+            if (!identical) {
+                vkCmdBindVertexBuffers(commandBuffer, 0, count, vkBuffers.data(), vkOffsets.data());
+                if (count <= DynamicStateShadow::kMaxShadowedVertexBindings) {
+                    shadow.vertexBindValid = true;
+                    shadow.vertexBindingCount = count;
+                    std::copy_n(vkBuffers.data(), count, shadow.vertexBuffers);
+                    std::copy_n(vkOffsets.data(), count, shadow.vertexOffsets);
+                } else {
+                    shadow.vertexBindValid = false;
+                }
+            }
         }
         return true;
     }
@@ -3150,8 +3281,17 @@ void main() {
             MGLOG_E("DrawElements skipped: failed to sync resident index buffer");
             return false;
         }
-        vkCmdBindIndexBuffer(frame.commandBuffer, slice.buffer,
+        const VkDeviceSize indexBindOffset =
-                             slice.offset + static_cast<VkDeviceSize>(pIndexBufferView->indexByteOffset), vkIndexType);
+            slice.offset + static_cast<VkDeviceSize>(pIndexBufferView->indexByteOffset);
+        auto& shadow = g_dynamicStateShadow;
+        if (!shadow.indexBindValid || shadow.indexBuffer != slice.buffer ||
+            shadow.indexOffset != indexBindOffset || shadow.indexType != vkIndexType) {
+            vkCmdBindIndexBuffer(frame.commandBuffer, slice.buffer, indexBindOffset, vkIndexType);
+            shadow.indexBindValid = true;
+            shadow.indexBuffer = slice.buffer;
+            shadow.indexOffset = indexBindOffset;
+            shadow.indexType = vkIndexType;
+        }
         return true;
     }
 
@@ -3653,6 +3793,10 @@ void main() {
             vkCmdSetScissor(frame.commandBuffer, 0, 1, &scissor);
 
             vkCmdBindPipeline(frame.commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
+            // The depth-mipmap pipeline's narrower dynamic set (viewport/scissor
+            // only) leaves the other dynamic states undefined; its raw scissor
+            // and viewport writes also bypass the shadow.
+            ResetDynamicStateShadow();
 
             std::fill(depthProgramData,
                       depthProgramData + m_depthMipmapResources.program->GetUBOSize(),
@@ -3714,16 +3858,24 @@ void main() {
         // content hash (folds program identity + link version + transform flags + shader stages),
         // vertex-input hash (VAO layout), render-pass hash (render targets + the draw-buffer/format
         // driven blend & write-mask gating), and the render-state version (all fixed-function state).
-        // Reset per-frame and on pipeline destruction so m_lastPipelineResult can never dangle.
+        // Reset per-frame and on pipeline destruction so a memoized handle can never dangle.
-        const Uint64 vertexInputHash = m_vertexInputStateFactory->GetOrComputeHash(vao);
+        // The identity hash mixes buffer heap addresses (per-chunk VBOs mint a new
+        // one per buffer); the memo and the pipeline payload key on the resolved
+        // LAYOUT hash instead, so draws over identical layouts share one pipeline.
+        // The one-arg fetch rides the VAO's state-pointer memo (no hash, no map).
+        auto& vis = m_vertexInputStateFactory->GetOrCreateVertexInputState(vao);
+        const Uint64 vertexLayoutHash = vis.layoutHash;
         const Uint64 renderPassHash = renderPassEntry.hash;
         const Uint renderStateVersion = MG_State::pGLContext->GetRenderStateParametersVersion();
-        if (m_lastPipelineValid && m_lastPipelineResult != VK_NULL_HANDLE && m_lastPipelineMode == mode &&
+        for (Uint32 i = 0; i < m_pipelineMemoCount; ++i) {
-            m_lastPipelineProgramHash == programObj.hash && m_lastPipelineVertexInputHash == vertexInputHash &&
+            const PipelineMemoEntry& entry = m_pipelineMemo[i];
-            m_lastPipelineRenderPassHash == renderPassHash &&
+            if (entry.pipeline != VK_NULL_HANDLE && entry.mode == mode &&
-            m_lastPipelineRenderStateVersion == renderStateVersion &&
+                entry.programHash == programObj.hash && entry.vertexInputHash == vertexLayoutHash &&
-            m_lastPipelineTransformFlags == transformFlags) {
+                entry.renderPassHash == renderPassHash &&
-            return m_lastPipelineResult;
+                entry.renderStateVersion == renderStateVersion &&
+                entry.transformFlags == transformFlags) {
+                return entry.pipeline;
+            }
         }
 
 #if MOBILEGL_LOG_ACTIVE_LEVEL <= MOBILEGL_LOG_LEVEL_DEBUG
@@ -3764,9 +3916,7 @@ void main() {
         }
 #endif
 
-        // vertexInputHash was computed above for the fast-path key; reuse it here.
+        const Uint32 vertexInputAttribMask = vis.attributeLocationMask;
-        auto& vis = m_vertexInputStateFactory->GetOrCreateVertexInputState(vao, vertexInputHash);
-        const Uint32 vertexInputAttribMask = BuildVertexInputAttributeMask(vis.attributes);
         const Uint32 activeAttribMask = programObj.activeVertexInputLocationMask;
         const Uint32 missingAttribMask = activeAttribMask & ~vertexInputAttribMask;
         auto& patchedAttributes = m_patchedAttributesScratch;
@@ -3876,7 +4026,7 @@ void main() {
 
         PipelineFactory::PipelineCreatePayload payload {
             .programHash = programObj.hash,
-            .vertexInputHash = vertexInputHash,
+            .vertexInputHash = vertexLayoutHash,
             .pipelineLayout = programObj.pipelineLayout,
             .renderPass = renderPassEntry.renderPass,
             .colorAttachmentCount = renderPassEntry.colorAttachmentCount,
@@ -3940,12 +4090,15 @@ void main() {
             payload.backStencilCompareOp = VK_COMPARE_OP_ALWAYS;
         }
         const Uint32 fragmentOutputMask = programObj.activeFragmentOutputLocationMask;
-        MOBILEGL_ASSERT(
+        // Outputs at locations past the render pass's trimmed colour span are
-            (fragmentOutputMask >> payload.colorAttachmentCount) == 0,
+        // simply discarded - GL's semantic for a fragment output whose draw
-            "GetOrCreatePipeline: fragmentOutputMask=0x%x exceeds colorAttachmentCount=%u for program=%u",
+        // buffer is GL_NONE (the trailing UNUSED slots no longer occupy
-            fragmentOutputMask,
+        // references, see GetOrCreateRenderPass).
-            payload.colorAttachmentCount,
+        if ((fragmentOutputMask >> payload.colorAttachmentCount) != 0) {
-            program.GetExternalIndex());
+            MGLOG_D("GetOrCreatePipeline: fragmentOutputMask=0x%x exceeds colorAttachmentCount=%u for program=%u; "
+                    "outputs past the span are discarded",
+                    fragmentOutputMask, payload.colorAttachmentCount, program.GetExternalIndex());
+        }
         MOBILEGL_ASSERT(payload.colorAttachmentCount <= PipelineFactory::PipelineCreatePayload::kMaxColorAttachments,
                         "GetOrCreatePipeline: colorAttachmentCount=%u exceeds payload capacity",
                         payload.colorAttachmentCount);
@@ -4180,14 +4333,16 @@ void main() {
         }
         VkPipeline pipeline = m_pipelineFactory->GetOrCreatePipeline(payload);
         if (pipeline != VK_NULL_HANDLE) {
-            m_lastPipelineValid = true;
+            PipelineMemoEntry& entry = m_pipelineMemo[m_pipelineMemoNext];
-            m_lastPipelineMode = mode;
+            entry.mode = mode;
-            m_lastPipelineProgramHash = programObj.hash;
+            entry.programHash = programObj.hash;
-            m_lastPipelineVertexInputHash = vertexInputHash;
+            entry.vertexInputHash = vertexLayoutHash;
-            m_lastPipelineRenderPassHash = renderPassHash;
+            entry.renderPassHash = renderPassHash;
-            m_lastPipelineRenderStateVersion = renderStateVersion;
+            entry.renderStateVersion = renderStateVersion;
-            m_lastPipelineTransformFlags = transformFlags;
+            entry.transformFlags = transformFlags;
-            m_lastPipelineResult = pipeline;
+            entry.pipeline = pipeline;
+            m_pipelineMemoNext = (m_pipelineMemoNext + 1) % kPipelineMemoSize;
+            m_pipelineMemoCount = std::min(m_pipelineMemoCount + 1, kPipelineMemoSize);
         }
         return pipeline;
     }
@@ -4289,6 +4444,129 @@ void main() {
         return true;
     }
 
+
+    Bool VulkanRenderer::TrySetupDrawFastPath(FrameContext::FrameData& frame, GLenum mode,
+                                              Flags<DrawSetupAspect> aspects, const DrawCmdParam& drawParams,
+                                              const IndexBufferView* pIndexBufferView) {
+        const SetupDrawSnapshot& snap = m_setupDrawSnapshot;
+        if (!snap.valid || !frame.isCommandRecording) {
+            return false;
+        }
+        if (snap.aspects != aspects.GetRaw() || snap.mode != mode) {
+            return false;
+        }
+        if (m_clearManager->HasAnyPendingClears()) {
+            return false;
+        }
+        const auto* activeRenderPass = VkRenderPassManager::GetActiveRenderPass();
+        if (activeRenderPass == nullptr || activeRenderPass->hash != snap.renderPassHash ||
+            snap.imageIndex != m_imageIndexAcquired) {
+            return false;
+        }
+        const auto& program = *MG_State::pGLContext->GetCurrentProgram();
+        if (program.GetLifetimeId() != snap.programLifetimeId ||
+            program.GetBackendStateVersion() != snap.programVersion) {
+            return false;
+        }
+        const auto& vao = *MG_State::pGLContext->GetBoundVertexArray();
+        if (static_cast<const void*>(&vao) != snap.vao || vao.GetConfigVersion() != snap.vaoConfigVersion) {
+            return false;
+        }
+        const auto& drawFbo =
+            MG_State::pGLContext->GetFramebufferBindingSlot(FramebufferTarget::Draw).GetBoundObject();
+        if (static_cast<const void*>(drawFbo.get()) != snap.drawFbo ||
+            drawFbo->GetObjectVersion() != snap.fboVersion) {
+            return false;
+        }
+        if (MG_State::pGLContext->GetRenderStateParametersVersion() != snap.renderStateVersion ||
+            MG_State::pGLContext->GetTextureBindGeneration() != snap.bindGeneration) {
+            return false;
+        }
+        if (GetShaderTransformFlags(m_swapchainObject.GetPreTransform()).GetRaw() != snap.baseTransformFlags) {
+            return false;
+        }
+        if (m_textureManager->GetResourceEraseEpoch() != snap.textureEraseEpoch ||
+            m_textureManager->GetTextureImageEpoch() != snap.textureImageEpoch ||
+            m_renderPassManager->GetRenderbufferImageEpoch() != snap.renderbufferImageEpoch) {
+            return false;
+        }
+
+        // Same sampled set as the snapshotting draw (program/bind keys above);
+        // verify content and params are untouched and every layout is still
+        // sampleable, then stamp recording use exactly as the full path would.
+        // A feedback case (sampled texture written by the active pass) fails the
+        // layout check and falls back to the full path's end-pass handling.
+        const auto& sampledTextures = m_sampledTexturesScratch;
+        const auto& sampledResources = m_sampledResourcesScratch;
+        if (sampledResources.size() != sampledTextures.size()) {
+            return false;
+        }
+        Uint64 contentSum = 0;
+        Uint64 paramsSum = 0;
+        for (SizeT i = 0; i < sampledTextures.size(); ++i) {
+            const auto* sampledTexture = sampledTextures[i];
+            if (sampledTexture == nullptr) {
+                continue;
+            }
+            const auto* resource = sampledResources[i];
+            if (resource == nullptr || !IsValidSampledImageLayout(resource->layout)) {
+                return false;
+            }
+            contentSum += sampledTexture->GetContentVersion();
+            paramsSum += sampledTexture->GetTextureParamsVersion();
+        }
+        if (contentSum != snap.sampledContentSum || paramsSum != snap.sampledParamsSum) {
+            return false;
+        }
+        for (SizeT i = 0; i < sampledTextures.size(); ++i) {
+            if (sampledTextures[i] != nullptr && sampledResources[i] != nullptr) {
+                m_textureManager->StampResourceRecordingUse(*sampledResources[i]);
+            }
+        }
+
+        // Everything the full path would re-resolve is provably unchanged; run
+        // only the per-draw tail.
+        if (!g_dynamicStateShadow.graphicsPipelineValid ||
+            g_dynamicStateShadow.graphicsPipeline != snap.pipeline) {
+            vkCmdBindPipeline(frame.commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, snap.pipeline);
+            g_dynamicStateShadow.graphicsPipelineValid = true;
+            g_dynamicStateShadow.graphicsPipeline = snap.pipeline;
+        }
+        const auto& programObj = m_programFactory->GetOrCreateProgram(
+            program, ProgramFactory::CompileOptionFlags(snap.resolvedTransformFlags));
+        if (!m_uniformManager->BindProgramUniformBuffers(frame.commandBuffer, program, programObj,
+                                                         m_frameContext.GetCurrentFrameIndex())) {
+            return false;
+        }
+        if (!UploadAndBindVertexBuffers(frame.commandBuffer, vao, programObj, drawParams, pIndexBufferView)) {
+            return false;
+        }
+        if (aspects & DrawSetupAspect::IndexBuffer) {
+            const Bool idxUploadOk = UploadAndBindIndexBuffer(frame, vao, pIndexBufferView);
+            MOBILEGL_ASSERT(idxUploadOk, "SetupDraw fast path: failed to upload index buffer");
+        }
+        ApplyGLViewportState(frame.commandBuffer, snap.renderPassExtent, m_swapchainObject.GetPreTransform(),
+                             snap.drawFboIsDefault);
+        ApplyBlendConstants(frame.commandBuffer);
+        ApplyPolygonOffsetState(frame.commandBuffer);
+        ApplyLineWidthState(frame.commandBuffer);
+        ApplyStencilState(frame.commandBuffer);
+        const Bool scissorEnabled = MG_State::pGLContext->IsCapabilityEnabled(CapabilityInput::ScissorTest);
+        VkRect2D scissor{};
+        if (scissorEnabled) {
+            const auto& scissorBox = MG_State::pGLContext->GetScissorBox();
+            scissor = snap.drawFboIsDefault
+                ? MakeDefaultFramebufferScissorRect(scissorBox, snap.renderPassExtent,
+                                                    m_swapchainObject.GetPreTransform())
+                : MakeClampedScissorRect(scissorBox, snap.renderPassExtent);
+        } else {
+            scissor.offset = {0, 0};
+            scissor.extent = { (Uint)snap.renderPassExtent.x(), (Uint)snap.renderPassExtent.y() };
+        }
+        ShadowedSetScissor(frame.commandBuffer, scissor);
+        return true;
+    }
+
     Bool VulkanRenderer::SetupDraw(FrameContext::FrameData& frame, GLenum mode, Flags<DrawSetupAspect> aspects,
                                    const DrawCmdParam& drawParams,
                                    const IndexBufferView* pIndexBufferView) {
@@ -4297,6 +4575,12 @@ void main() {
         // otherwise each re-run the full SyncTexture path on the same textures.
         VkTextureManager::DrawSyncScope drawSyncScope(*m_textureManager);
         m_textureManager->CollectGarbage();
+        if (TrySetupDrawFastPath(frame, mode, aspects, drawParams, pIndexBufferView)) {
+            return true;
+        }
+        // The fast path declined: whatever it saw may be stale. The full path
+        // below re-resolves everything and refreshes the snapshot on success.
+        m_setupDrawSnapshot.valid = false;
         const auto& drawFbo =
                 MG_State::pGLContext->GetFramebufferBindingSlot(FramebufferTarget::Draw).GetBoundObject();
         if (drawFbo != nullptr && IsUnsupportedFramebufferForDirectVulkan(*drawFbo)) {
@@ -4306,25 +4590,52 @@ 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* 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.
+        // single mip level. The probe walks every sampler binding, so its verdict is memoized
-        if (UniformManager::ProgramSamplesOnlySingleLevelTextures(program, *programObjPtr)) {
+        // under the sampled-set memo's key plus the sampled textures' params-version sum (level
+        // range and filter changes live there); the previous draw's texture list is valid for the
+        // sum exactly when that key matches (same program, same binds).
+        {
+            const Uint64 lodProgramLifetimeId = program.GetLifetimeId();
+            const Uint32 lodProgramVersion = program.GetBackendStateVersion();
+            const Uint64 lodBindGeneration = MG_State::pGLContext->GetTextureBindGeneration();
+            Bool lodMemoHit = false;
+            if (m_lastLodDecisionValid && m_lastSampledSetValid &&
+                m_lastLodProgramLifetimeId == lodProgramLifetimeId &&
+                m_lastLodProgramVersion == lodProgramVersion &&
+                m_lastLodBindGeneration == lodBindGeneration && m_lastLodBaseFlags == transformFlags &&
+                m_lastSampledSetProgramLifetimeId == lodProgramLifetimeId &&
+                m_lastSampledSetProgramVersion == lodProgramVersion &&
+                m_lastSampledSetBindGeneration == lodBindGeneration) {
+                Uint64 paramsSum = 0;
+                for (const auto* sampledTexture : m_sampledTexturesScratch) {
+                    if (sampledTexture != nullptr) {
+                        paramsSum += sampledTexture->GetTextureParamsVersion();
+                    }
+                }
+                if (paramsSum == m_lastLodParamsSum) {
+                    transformFlags = m_lastLodResultFlags;
+                    lodMemoHit = true;
+                }
+            }
+            if (!lodMemoHit) {
+                const ProgramFactory::CompileOptionFlags baseFlags = transformFlags;
+                const auto& baseProgramObj = m_programFactory->GetOrCreateProgram(program, transformFlags);
+                if (UniformManager::ProgramSamplesOnlySingleLevelTextures(program, baseProgramObj)) {
                     transformFlags |= ProgramFactory::CompileOptionBit::ExplicitLod0Sampling;
-            programObjPtr = &m_programFactory->GetOrCreateProgram(program, transformFlags);
                 }
-        // fp16 fragment arithmetic is only sound when nothing this draw reads or writes carries
+                m_lastLodDecisionValid = true;
-        // more than 8 normalized bits per channel. A shaderpack's HDR gbuffer, or a data texture
+                m_lastLodProgramLifetimeId = lodProgramLifetimeId;
-        // holding positions, must keep full precision - and SPIR-V cannot tell, since sampler2D
+                m_lastLodProgramVersion = lodProgramVersion;
-        // yields vec4 whatever the bound format is, so the decision has to be made here.
+                m_lastLodBindGeneration = lodBindGeneration;
-        if (UniformManager::ProgramSamplesOnlyLowPrecisionTextures(program, *programObjPtr) &&
+                m_lastLodBaseFlags = baseFlags;
-            UniformManager::DrawTargetIsLowPrecision(drawFbo.get())) {
+                m_lastLodResultFlags = transformFlags;
-            transformFlags |= ProgramFactory::CompileOptionBit::RelaxedFragmentPrecision;
+                m_lastLodParamsSum = 0;  // filled below once the sampled set is known
-            programObjPtr = &m_programFactory->GetOrCreateProgram(program, transformFlags);
             }
-        const auto& programObj = *programObjPtr;
+        }
+        const auto& programObj = m_programFactory->GetOrCreateProgram(program, transformFlags);
 
         // Begin command recording if not yet
         if (!frame.isCommandRecording) {
@@ -4369,6 +4680,18 @@ void main() {
                 m_lastSampledSetTransformFlags = transformFlags;
                 m_lastSampledSetBindGeneration = bindGeneration;
             }
+            // Complete a freshly-made LOD decision (see above): its params sum
+            // can only be taken once the sampled set is known. A genuine
+            // all-zero sum merely re-probes next draw.
+            if (m_lastLodDecisionValid && m_lastLodParamsSum == 0) {
+                Uint64 paramsSum = 0;
+                for (const auto* sampledTexture : sampledTextures) {
+                    if (sampledTexture != nullptr) {
+                        paramsSum += sampledTexture->GetTextureParamsVersion();
+                    }
+                }
+                m_lastLodParamsSum = paramsSum;
+            }
         }
         MGLOG_D("SetupDraw: program=%u drawFbo=%u sampledTextureCount=%zu activeRenderPass=%s",
                 program.GetExternalIndex(), drawFbo ? drawFbo->GetExternalIndex() : 0u, sampledTextures.size(),
@@ -4392,7 +4715,10 @@ void main() {
             activeRenderPass = nullptr;
         }
         Bool needSampledTextureTransitions = false;
-        for (auto* sampledTexture : sampledTextures) {
+        auto& sampledResources = m_sampledResourcesScratch;
+        sampledResources.assign(sampledTextures.size(), nullptr);
+        for (SizeT sampledIndex = 0; sampledIndex < sampledTextures.size(); ++sampledIndex) {
+            auto* sampledTexture = sampledTextures[sampledIndex];
             if (!sampledTexture) {
                 continue;
             }
@@ -4401,13 +4727,36 @@ void main() {
             MOBILEGL_ASSERT(textureResource != nullptr,
                             "%s: SyncTextureAndGetDescriptor failed for textureId=%d",
                             __func__, sampledTexture->GetExternalIndex());
+            sampledResources[sampledIndex] = textureResource;
             MGLOG_D("SetupDraw: sampled textureId=%d layout(before)=%s(%d)",
                     sampledTexture->GetExternalIndex(), VkImageLayoutToString(textureResource->layout),
                     static_cast<Int>(textureResource->layout));
             if (m_clearManager->HasPendingClear(sampledTexture) ||
                 !IsValidSampledImageLayout(textureResource->layout)) {
+                // Out-of-pass work is needed (deferred clear materialization or
+                // a sampled-layout transition). When the open frame recording
+                // has not referenced this image yet, that work can execute
+                // ahead of the WHOLE recording - record it into the pre-pass
+                // stream instead of splitting the active render pass (ANGLE's
+                // outside-render-pass command stream, restricted to the
+                // provably reorderable case).
+                if (activeRenderPass != nullptr &&
+                    !m_frameContext.GetCurrent().hasPreCommandBufferRecorded &&
+                    !m_textureManager->WasTouchedThisRecording(*textureResource)) {
+                    VkCommandBuffer preCommandBuffer = m_frameContext.BeginPreCommandRecording();
+                    const Bool preClearReady =
+                        MaterializePendingClearForTexture(preCommandBuffer, *sampledTexture);
+                    MOBILEGL_ASSERT(preClearReady,
+                                    "%s: pre-pass MaterializePendingClearForTexture failed for textureId=%d",
+                                    __func__, sampledTexture->GetExternalIndex());
+                    const Bool preTransitionReady =
+                        m_textureManager->TransitionTextureForSampling(preCommandBuffer, *sampledTexture);
+                    MOBILEGL_ASSERT(preTransitionReady,
+                                    "%s: pre-pass TransitionTextureForSampling failed for textureId=%d",
+                                    __func__, sampledTexture->GetExternalIndex());
+                    continue;
+                }
                 needSampledTextureTransitions = true;
-                break;
             }
         }
 
@@ -4417,10 +4766,23 @@ void main() {
             activeRenderPass = nullptr;
         }
 
-        for (auto* sampledTexture : sampledTextures) {
+        for (SizeT sampledIndex = 0; sampledIndex < sampledTextures.size(); ++sampledIndex) {
+            auto* sampledTexture = sampledTextures[sampledIndex];
             if (!sampledTexture) {
                 continue;
             }
+            // Fast path: the first loop already resolved this texture, nothing
+            // is pending against it, and its layout is still sampleable (the
+            // layout re-check covers an EndRenderPass between the loops having
+            // rewritten an attachment's layout). Skipping the materialize +
+            // transition + re-resolve chain here is the difference between one
+            // pointer read and three calls per sampled texture per draw.
+            if (auto* fastResource = sampledResources[sampledIndex];
+                fastResource != nullptr && !m_clearManager->HasPendingClear(sampledTexture) &&
+                IsValidSampledImageLayout(fastResource->layout)) {
+                m_textureManager->StampResourceRecordingUse(*fastResource);
+                continue;
+            }
             const Bool clearReady = MaterializePendingClearForTexture(frame.commandBuffer, *sampledTexture);
             MOBILEGL_ASSERT(clearReady, "%s: MaterializePendingClearForTexture failed for textureId=%d",
                             __func__, sampledTexture->GetExternalIndex());
@@ -4431,16 +4793,28 @@ void main() {
             MOBILEGL_ASSERT(transitionedResource != nullptr,
                             "%s: post-transition SyncTextureAndGetDescriptor failed for textureId=%d",
                             __func__, sampledTexture->GetExternalIndex());
+            // Pre-pass stream bookkeeping: the draw about to be recorded reads
+            // this image, so later out-of-pass work on it can no longer jump
+            // ahead of the recording.
+            m_textureManager->StampResourceRecordingUse(*transitionedResource);
             MGLOG_D("SetupDraw: sampled textureId=%d layout(after)=%s(%d)",
                     sampledTexture->GetExternalIndex(), VkImageLayoutToString(transitionedResource->layout),
                     static_cast<Int>(transitionedResource->layout));
         }
 
-        auto* renderPassEntry = &m_renderPassManager->GetOrCreateRenderPass(*drawFbo, m_imageIndexAcquired);
+        // Depth/stencil participation of THIS draw, for the default-FBO depth-less
+        // pass flavor (GL: a disabled depth/stencil test neither reads nor writes
+        // its buffer).
+        const Bool drawUsesDepthStencil =
+            MG_State::pGLContext->IsCapabilityEnabled(CapabilityInput::DepthTest) ||
+            MG_State::pGLContext->IsCapabilityEnabled(CapabilityInput::StencilTest);
+        auto* renderPassEntry =
+            &m_renderPassManager->GetOrCreateRenderPass(*drawFbo, m_imageIndexAcquired, drawUsesDepthStencil);
         if (activeRenderPass && !activeRenderPass->CompatibleWith(*renderPassEntry)) {
             VkRenderPassManager::EndRenderPass(frame.commandBuffer);
             activeRenderPass = nullptr;
-            renderPassEntry = &m_renderPassManager->GetOrCreateRenderPass(*drawFbo, m_imageIndexAcquired);
+            renderPassEntry =
+                &m_renderPassManager->GetOrCreateRenderPass(*drawFbo, m_imageIndexAcquired, drawUsesDepthStencil);
         }
         if (renderPassEntry->attachmentCount == 0 || renderPassEntry->extent.x() <= 0 || renderPassEntry->extent.y() <= 0) {
             MGLOG_D("SetupDraw skipped: drawFbo=%u resolved to an empty render pass (attachmentCount=%u extent=%dx%d)",
@@ -4472,7 +4846,7 @@ void main() {
             // Every genuinely disabled attribute the shader reads must have a current-value type we can
             // synthesize a binding for; otherwise the upload below would push a null payload.
             const Uint32 missingAttribMask =
-                activeAttribMask & ~BuildVertexInputAttributeMask(vertexInputState.attributes);
+                activeAttribMask & ~vertexInputState.attributeLocationMask;
             for (Uint32 location = 0; location < kMaxVertexAttribs; ++location) {
                 if ((missingAttribMask & (1u << location)) == 0) continue;
 
@@ -4500,7 +4874,11 @@ void main() {
             MOBILEGL_ASSERT(ok, "%s: BeginRenderPass failed", __func__);
         }
 
+        if (!g_dynamicStateShadow.graphicsPipelineValid || g_dynamicStateShadow.graphicsPipeline != pipeline) {
             vkCmdBindPipeline(frame.commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
+            g_dynamicStateShadow.graphicsPipelineValid = true;
+            g_dynamicStateShadow.graphicsPipeline = pipeline;
+        }
 
         const Bool boundUniforms = m_uniformManager->BindProgramUniformBuffers(
             frame.commandBuffer, program, programObj, m_frameContext.GetCurrentFrameIndex());
@@ -4540,7 +4918,49 @@ void main() {
             scissor.offset = {0, 0};
             scissor.extent = { (Uint)renderPassEntry->extent.x(), (Uint)renderPassEntry->extent.y() };
         }
-        vkCmdSetScissor(frame.commandBuffer, 0, 1, &scissor);
+        ShadowedSetScissor(frame.commandBuffer, scissor);
+
+        // Snapshot the fully resolved configuration for the consecutive-draw
+        // fast path (see TrySetupDrawFastPath).
+        {
+            auto& snap = m_setupDrawSnapshot;
+            const auto* nowActiveRenderPass = VkRenderPassManager::GetActiveRenderPass();
+            if (nowActiveRenderPass != nullptr && !programObj.hasStorageImages) {
+                snap.valid = true;
+                snap.aspects = aspects.GetRaw();
+                snap.mode = mode;
+                snap.programLifetimeId = program.GetLifetimeId();
+                snap.programVersion = program.GetBackendStateVersion();
+                snap.vao = &vao;
+                snap.vaoConfigVersion = vao.GetConfigVersion();
+                snap.drawFbo = drawFbo.get();
+                snap.fboVersion = drawFbo->GetObjectVersion();
+                snap.drawFboIsDefault = drawFbo->IsDefaultFramebuffer();
+                snap.renderStateVersion = MG_State::pGLContext->GetRenderStateParametersVersion();
+                snap.bindGeneration = MG_State::pGLContext->GetTextureBindGeneration();
+                snap.baseTransformFlags = GetShaderTransformFlags(m_swapchainObject.GetPreTransform()).GetRaw();
+                snap.resolvedTransformFlags = transformFlags.GetRaw();
+                snap.renderPassHash = nowActiveRenderPass->hash;
+                snap.imageIndex = m_imageIndexAcquired;
+                snap.textureEraseEpoch = m_textureManager->GetResourceEraseEpoch();
+                snap.textureImageEpoch = m_textureManager->GetTextureImageEpoch();
+                snap.renderbufferImageEpoch = m_renderPassManager->GetRenderbufferImageEpoch();
+                snap.renderPassExtent = renderPassEntry->extent;
+                snap.pipeline = pipeline;
+                Uint64 snapContentSum = 0;
+                Uint64 snapParamsSum = 0;
+                for (const auto* sampledTexture : sampledTextures) {
+                    if (sampledTexture != nullptr) {
+                        snapContentSum += sampledTexture->GetContentVersion();
+                        snapParamsSum += sampledTexture->GetTextureParamsVersion();
+                    }
+                }
+                snap.sampledContentSum = snapContentSum;
+                snap.sampledParamsSum = snapParamsSum;
+            } else {
+                snap.valid = false;
+            }
+        }
         return true;
     }
 
@@ -5231,8 +5651,12 @@ void main() {
         if (!m_clearManager->GetPendingClears(&texture, pendingClears)) {
             return true;
         }
-        MOBILEGL_ASSERT(VkRenderPassManager::GetActiveRenderPass() == nullptr,
+        // A pass may stay open on the FRAME command buffer while this clear is
-                        "MaterializePendingClearForTexture requires no active render pass");
+        // recorded into the pre-pass stream (a different command buffer that
+        // executes strictly before the frame's commands).
+        MOBILEGL_ASSERT(VkRenderPassManager::GetActiveRenderPass() == nullptr ||
+                            commandBuffer != m_frameContext.GetCurrent().commandBuffer,
+                        "MaterializePendingClearForTexture requires no active render pass on the target buffer");
 
         auto* resource = m_textureManager->SyncTextureAndGetDescriptor(texture);
         MOBILEGL_ASSERT(resource != nullptr,
@@ -5462,7 +5886,10 @@ void main() {
                         "TryBlitToDefaultFramebufferWithShader: failed to create sampled view for textureId=%d mip=%u",
                         sourceTexture->GetExternalIndex(), srcBinding.mipLevel);
 
-        auto& renderPassEntry = m_renderPassManager->GetOrCreateRenderPass(drawFbo, m_imageIndexAcquired);
+        // A color-only blit never touches depth/stencil: let the default-FBO pass
+        // it opens skip the depth attachment (depth-less flavor).
+        auto& renderPassEntry =
+            m_renderPassManager->GetOrCreateRenderPass(drawFbo, m_imageIndexAcquired, /*drawUsesDepthStencil=*/false);
         const Bool ok = VkRenderPassManager::BeginRenderPass(frame.commandBuffer, renderPassEntry);
         MOBILEGL_ASSERT(ok, "%s: BeginRenderPass failed", __func__);
 
@@ -5478,6 +5905,10 @@ void main() {
         const VkPipeline pipeline = GetOrCreateBlitPipeline(renderPassEntry);
         MOBILEGL_ASSERT(pipeline != VK_NULL_HANDLE, "TryBlitToDefaultFramebufferWithShader: blit pipeline is null");
         vkCmdBindPipeline(frame.commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
+        // The blit pipeline's narrower dynamic set (viewport/scissor only)
+        // leaves the other dynamic states undefined; its raw viewport/scissor
+        // writes also bypass the shadow.
+        ResetDynamicStateShadow();
 
         auto* blitProgramData = static_cast<Uint8*>(m_blitResources.program->MapUBO());
         MOBILEGL_ASSERT(blitProgramData != nullptr, "TryBlitToDefaultFramebufferWithShader: blit UBO is null");
@@ -6272,9 +6703,13 @@ void main() {
         if (frame.isCommandRecording) {
             m_frameContext.EndCommandRecording();
             frame.hasCommandBufferRecorded = true;
-            m_lastPipelineValid = false; // command-buffer boundary: drop the pipeline memo
+            InvalidatePipelineMemo(); // command-buffer boundary: drop the pipeline memo
         }
-        if (!frame.hasCommandBufferRecorded) {
+        // The pre-pass stream must never be submitted later than the recording
+        // it was paired with (frame commands recorded after a pre-pass move
+        // rely on the moved work having executed first).
+        m_frameContext.EndPreCommandRecordingIfOpen();
+        if (!frame.hasCommandBufferRecorded && !frame.hasPreCommandBufferRecorded) {
             return true;
         }
 
@@ -7355,8 +7790,7 @@ void main() {
             m_programFactory->OnFrameBoundary();
         }
         if (m_pipelineFactory && m_pipelineFactory->OnFrameBoundary() > 0) {
-            m_lastPipelineValid = false;
+            InvalidatePipelineMemo();
-            m_lastPipelineResult = VK_NULL_HANDLE;
         }
         if (m_vertexInputStateFactory) {
             m_vertexInputStateFactory->OnFrameBoundary();
@@ -7409,8 +7843,18 @@ void main() {
             submitInfo.pWaitSemaphores = &waitSemaphore;
             submitInfo.pWaitDstStageMask = &waitDstStageMask;
         }
-        submitInfo.commandBufferCount = 1;
+        // The pre-pass stream, when recorded, executes strictly before the
-        submitInfo.pCommandBuffers = &frame.commandBuffer;
+        // frame's commands within the same submission.
+        VkCommandBuffer commandBuffers[2] = {VK_NULL_HANDLE, VK_NULL_HANDLE};
+        Uint32 commandBufferCount = 0;
+        if (frame.hasPreCommandBufferRecorded) {
+            commandBuffers[commandBufferCount++] = frame.preCommandBuffer;
+        }
+        if (frame.hasCommandBufferRecorded) {
+            commandBuffers[commandBufferCount++] = frame.commandBuffer;
+        }
+        submitInfo.commandBufferCount = commandBufferCount;
+        submitInfo.pCommandBuffers = commandBuffers;
         const VkResult result = vkQueueSubmit(m_graphicsQueue, 1, &submitInfo, fence);
         if (result != VK_SUCCESS) {
             MGLOG_E("SubmitPendingCommandBuffer: vkQueueSubmit returned %d", result);
@@ -7418,6 +7862,7 @@ void main() {
         }
         frame.imageAvailableSemaphoreConsumed = true;
         frame.hasCommandBufferRecorded = false;
+        frame.hasPreCommandBufferRecorded = false;
         RegisterSubmit(fence, pooledFence);
         frame.lastSubmitIndex = m_submitCounter;
         return true;
@@ -7448,6 +7893,8 @@ void main() {
             }
             m_frameContext.EndCommandRecording();
         }
+        m_frameContext.EndPreCommandRecordingIfOpen();
+        const Bool submittingPreCommandBuffer = frame.hasPreCommandBufferRecorded;
         if (!SubmitPendingCommandBuffer(frame, fence, /*pooledFence=*/true)) {
             // Submit failure (device loss regime): the ended command buffer
             // stays marked recorded so Present can still try to submit it.
@@ -7460,13 +7907,12 @@ void main() {
         // cache and the aging sweep could destroy it while the flushed submission
         // still references it. Mirrors the drops at the readback and Present
         // boundaries; costs one full pipeline lookup on the next draw.
-        m_lastPipelineValid = false;
+        InvalidatePipelineMemo();
-        m_lastPipelineResult = VK_NULL_HANDLE;
 
         // The submitted command buffer may still be executing; recording must
         // restart on a fresh one. If none can be allocated, fall back to
         // draining this submission so reusing the buffer stays legal.
-        const VkResult retireResult = m_frameContext.RetireCurrentCommandBuffer();
+        const VkResult retireResult = m_frameContext.RetireCurrentCommandBuffer(submittingPreCommandBuffer);
         if (retireResult != VK_SUCCESS) {
             MGLOG_E("FlushPendingCommands: RetireCurrentCommandBuffer returned %d; draining submission", retireResult);
             if (vkWaitForFences(m_device, 1, &fence, VK_TRUE, UINT64_MAX) == VK_SUCCESS) {
@@ -7532,6 +7978,17 @@ void main() {
     }
 
     void VulkanRenderer::OnFrameCommandRecordingBegan(VkCommandBuffer commandBuffer) {
+        // Dynamic state does not survive a command-buffer boundary.
+        ResetDynamicStateShadow();
+        m_setupDrawSnapshot.valid = false;
+        if (m_uniformManager) {
+            m_uniformManager->OnCommandBufferBoundary();
+        }
+        // Pre-pass stream bookkeeping: a fresh frame recording references no
+        // textures yet.
+        if (m_textureManager) {
+            m_textureManager->AdvanceRecordingGeneration();
+        }
         if (m_timerQueryManager) {
             m_timerQueryManager->OnFrameCommandRecordingBegan(commandBuffer, m_frameContext.GetCurrentFrameIndex(),
                                                               m_bufferManager.GetFrameSerial());
@@ -7604,9 +8061,10 @@ void main() {
                 if (suspendedFrame.isCommandRecording) {
                     m_frameContext.EndCommandRecording();
                 }
+                m_frameContext.AbandonPreCommandRecording();
                 suspendedFrame.isCommandRecording = false;
                 suspendedFrame.hasCommandBufferRecorded = false;
-                m_lastPipelineValid = false;
+                InvalidatePipelineMemo();
                 // The dropped recording is never submitted, so once the fence
                 // poll shows the pre-suspension submissions complete the frame
                 // transients (descriptor sets, transient arenas, deferred
@@ -7642,8 +8100,11 @@ void main() {
         // performs a real, stamping lookup) and can never age out.
         m_programFactory->OnFrameBoundary();
         if (m_pipelineFactory->OnFrameBoundary() > 0) {
-            m_lastPipelineValid = false; // an aged-out pipeline may still be memoized
+            InvalidatePipelineMemo(); // an aged-out pipeline may still be memoized
-            m_lastPipelineResult = VK_NULL_HANDLE;
+            // A recreated pipeline could reuse a freed handle value and alias
+            // the bind-dedup shadow; force the next draw to re-bind.
+            g_dynamicStateShadow.graphicsPipelineValid = false;
+            m_setupDrawSnapshot.valid = false;
         }
         m_vertexInputStateFactory->OnFrameBoundary();
         m_samplerManager->OnFrameBoundary();
@@ -7666,18 +8127,21 @@ void main() {
         if (frame.isCommandRecording) {
             m_frameContext.EndCommandRecording();
             frame.hasCommandBufferRecorded = true;
-            m_lastPipelineValid = false; // command-buffer boundary: drop the pipeline memo
+            InvalidatePipelineMemo(); // command-buffer boundary: drop the pipeline memo
         }
+        m_frameContext.EndPreCommandRecordingIfOpen();
 
         const Bool shouldSubmitCommandBuffer = frame.hasCommandBufferRecorded;
 
-        // 1) Submit current frame work.
+        // 1) Submit current frame work (the pre-pass stream, when recorded,
+        //    rides the same submission strictly ahead of the frame commands).
         auto submitPacket = m_frameContext.GetSubmitInfo(shouldSubmitCommandBuffer, m_imageIndexAcquired);
         VK_VERIFY(vkQueueSubmit(m_graphicsQueue, 1, &submitPacket.submitInfo, frame.imageInFlightFence));
         RegisterSubmit(frame.imageInFlightFence, /*pooledFence=*/false);
         frame.lastSubmitIndex = m_submitCounter;
         frame.isCommandRecording = false;
         frame.hasCommandBufferRecorded = false;
+        frame.hasPreCommandBufferRecorded = false;
         m_swapchainObject.SetImageLayout(m_imageIndexAcquired, VK_IMAGE_LAYOUT_PRESENT_SRC_KHR);
 
         // 2) Present current frame.
@@ -7705,6 +8169,13 @@ void main() {
             result = VK_SUCCESS;
         }
         VK_VERIFY(result, "Present, vkQueuePresentKHR");
+        // EGL swap semantics: the presented color buffer's content is undefined the
+        // next time this image is acquired (EGL_BUFFER_DESTROYED, the default swap
+        // behaviour), and EVERY ancillary depth/stencil buffer's content is
+        // undefined after any swap. The render-pass manager turns the undefined
+        // attachments' next tile loads into LOAD_OP_DONT_CARE.
+        m_swapchainObject.SetImageContentDefined(m_imageIndexAcquired, false);
+        m_swapchainObject.SetAllDepthStencilContentUndefined();
         // 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
@@ -8659,11 +9130,17 @@ void main() {
         if (m_pipelineFactory) {
             m_pipelineFactory->DestroyAll();
         }
-        m_lastPipelineValid = false; // pipelines freed -> the memoized handle would dangle
+        InvalidatePipelineMemo(); // pipelines freed -> the memoized handle would dangle
+        g_dynamicStateShadow.graphicsPipelineValid = false;
+        m_setupDrawSnapshot.valid = false;
         DestroyComputePipelines();
         if (m_frameContext.GetFrameCount() > 0) {
             m_frameContext.GetCurrent().isCommandRecording = false;
             m_frameContext.GetCurrent().hasCommandBufferRecorded = false;
+            // The pre-pass stream paired with the abandoned recording is
+            // dropped with it (its next Begin resets the buffer).
+            m_frameContext.GetCurrent().isPreCommandRecording = false;
+            m_frameContext.GetCurrent().hasPreCommandBufferRecorded = false;
         }
         const Bool okArena = m_bufferManager.RecreateTransientArenas(m_frameContext.GetFrameCount());
         MOBILEGL_ASSERT(okArena, "RecreateSwapchain: buffer manager transient arena initialization failed");
@@ -8818,8 +9295,7 @@ void main() {
         // destroys them immediately. The memo must drop as well: it can hand out a
         // cached handle without touching the factory.
         if (m_pipelineFactory->EvictByRenderPasses(renderPasses) > 0) {
-            m_lastPipelineValid = false;
+            InvalidatePipelineMemo();
-            m_lastPipelineResult = VK_NULL_HANDLE;
         }
     }
 
@@ -8838,8 +9314,7 @@ void main() {
             m_computePipelines.erase(computeIt);
         }
         if (m_pipelineFactory != nullptr && m_pipelineFactory->EvictByProgramHash(programHash) > 0) {
-            m_lastPipelineValid = false;
+            InvalidatePipelineMemo();
-            m_lastPipelineResult = VK_NULL_HANDLE;
         }
         if (m_uniformManager != nullptr) {
             m_uniformManager->OnDescriptorSetLayoutDestroyed(descriptorSetLayout);

MobileGL/MG_Backend/DirectVulkan/Renderer/VulkanRenderer.h

@@ -151,6 +151,14 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         Bool SetupDraw(FrameContext::FrameData& frame, GLenum mode, Flags<DrawSetupAspect> aspects,
                        const DrawCmdParam& drawParams,
                        const IndexBufferView* pIndexBufferView = nullptr);
+        // ANGLE-style consecutive-draw fast path: SetupDraw snapshots the fully
+        // resolved draw configuration; the next draw whose cheap version/identity
+        // checks all match skips the resolution half (LOD probe, sampled-set
+        // walk, render-pass and pipeline resolution) and jumps straight to the
+        // per-draw tail. Returns false (leaving no side effects that the full
+        // path cannot redo idempotently) whenever anything might have changed.
+        Bool TrySetupDrawFastPath(FrameContext::FrameData& frame, GLenum mode, Flags<DrawSetupAspect> aspects,
+                                  const DrawCmdParam& drawParams, const IndexBufferView* pIndexBufferView);
         void ClearAttachmentsOnActiveRenderPass(VkCommandBuffer commandBuffer,
                                                 const RenderPassEntry& compatibleRenderPassEntry);
 
@@ -455,14 +463,30 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         // gather + synthetic vertex-input rebuild + payload hash + lookup) when the full pipeline
         // state is unchanged from the previous draw. The key provably covers every pipeline field.
         // Reset per-frame and on pipeline destruction so the cached handle can never dangle.
-        Bool m_lastPipelineValid = false;
+        // Small N-way pipeline-resolution memo (round-robin replacement). A
-        GLenum m_lastPipelineMode = 0;
+        // single-entry memo thrashed on draw sequences that alternate a few
-        Uint64 m_lastPipelineProgramHash = 0;
+        // pipelines (GUI text/quad program ping-pong), paying the full
-        Uint64 m_lastPipelineVertexInputHash = 0;
+        // payload-hash lookup per draw; eight entries cover such working sets
-        Uint64 m_lastPipelineRenderPassHash = 0;
+        // while keeping the hit path a trivial linear scan.
-        Uint m_lastPipelineRenderStateVersion = 0;
+        struct PipelineMemoEntry {
-        ProgramFactory::CompileOptionFlags m_lastPipelineTransformFlags = {};
+            GLenum mode = 0;
-        VkPipeline m_lastPipelineResult = VK_NULL_HANDLE;
+            Uint64 programHash = 0;
+            Uint64 vertexInputHash = 0;
+            Uint64 renderPassHash = 0;
+            Uint renderStateVersion = 0;
+            ProgramFactory::CompileOptionFlags transformFlags = {};
+            VkPipeline pipeline = VK_NULL_HANDLE;
+        };
+        static constexpr Uint32 kPipelineMemoSize = 8;
+        PipelineMemoEntry m_pipelineMemo[kPipelineMemoSize];
+        Uint32 m_pipelineMemoCount = 0;
+        Uint32 m_pipelineMemoNext = 0;
+        // Drops every memoized pipeline handle. Required at command-buffer
+        // boundaries and whenever any pipeline may have been destroyed.
+        void InvalidatePipelineMemo() {
+            m_pipelineMemoCount = 0;
+            m_pipelineMemoNext = 0;
+        }
         UnorderedMap<ProgramFactory::HashType, VkPipeline> m_computePipelines;
         UniquePtr<ProgramFactory> m_programFactory;
         UniquePtr<UniformManager> m_uniformManager;
@@ -491,9 +515,61 @@ namespace MobileGL::MG_Backend::DirectVulkan {
         ProgramFactory::CompileOptionFlags m_lastSampledSetTransformFlags = {};
         Uint64 m_lastSampledSetBindGeneration = 0;
 
+        // Memo for the per-draw explicit-LOD-0 eligibility probe
+        // (ProgramSamplesOnlySingleLevelTextures): same key family as the
+        // sampled-set memo, plus the sampled textures' params-version sum so a
+        // level-range or filter change re-probes. On a hit the resolved
+        // transform flags are reused, which also collapses the two
+        // GetOrCreateProgram lookups into one.
+        Bool m_lastLodDecisionValid = false;
+        Uint64 m_lastLodProgramLifetimeId = 0;
+        Uint32 m_lastLodProgramVersion = 0;
+        Uint64 m_lastLodBindGeneration = 0;
+        Uint64 m_lastLodParamsSum = 0;
+        ProgramFactory::CompileOptionFlags m_lastLodBaseFlags = {};
+        ProgramFactory::CompileOptionFlags m_lastLodResultFlags = {};
+
+        // Snapshot behind TrySetupDrawFastPath. Values only: the program and
+        // render-pass caches are open-addressing maps whose entries move on
+        // insert, so no pointers into them are cached; the pipeline handle is
+        // protected by the command-buffer-boundary reset plus the mid-frame
+        // pipeline-destruction resets, and monotonic epochs guard everything
+        // that can be destroyed or recreated between draws.
+        struct SetupDrawSnapshot {
+            Bool valid = false;
+            Uint8 aspects = 0;
+            GLenum mode = 0;
+            Uint64 programLifetimeId = 0;
+            Uint32 programVersion = 0;
+            const void* vao = nullptr;
+            Uint32 vaoConfigVersion = 0;
+            const void* drawFbo = nullptr;
+            Uint16 fboVersion = 0;
+            Bool drawFboIsDefault = false;
+            Uint renderStateVersion = 0;
+            Uint64 bindGeneration = 0;
+            Uint32 baseTransformFlags = 0;
+            Uint32 resolvedTransformFlags = 0;
+            Uint64 renderPassHash = 0;
+            Uint32 imageIndex = 0;
+            Uint64 textureEraseEpoch = 0;
+            Uint64 textureImageEpoch = 0;
+            Uint64 renderbufferImageEpoch = 0;
+            Uint64 sampledContentSum = 0;
+            Uint64 sampledParamsSum = 0;
+            IntVec2 renderPassExtent = {0, 0};
+            VkPipeline pipeline = VK_NULL_HANDLE;
+        };
+        SetupDrawSnapshot m_setupDrawSnapshot;
+
         // Per-draw scratch buffers (clear keeps capacity) — these paths run for every
         // draw call and must not allocate.
         Vector<MG_State::GLState::ITextureObject*> m_sampledTexturesScratch;
+        // Parallel to m_sampledTexturesScratch, refilled by every SetupDraw's
+        // first sampled-texture loop: the resolved backend resources, so the
+        // post-transition loop can skip re-resolving textures whose layout is
+        // already sampleable.
+        Vector<VkTextureManager::TextureResource*> m_sampledResourcesScratch;
         Vector<MG_State::GLState::ITextureObject*> m_storageImageTexturesScratch;
         Vector<VkBuffer> m_vertexBuffersScratch;
         Vector<VkDeviceSize> m_vertexOffsetsScratch;

MobileGL/MG_State/GLState/VertexArrayState/VertexArrayObject.h

@@ -104,6 +104,23 @@ namespace MobileGL {
                     m_backendHashMemoVersion = m_configVersion;
                 }
 
+                // Backend-owned resolved-state memo: an opaque pointer into the
+                // backend's vertex-input-state cache plus the cache's eviction
+                // epoch, valid while the config version matches. Lets the
+                // per-draw path skip the content hash AND the cache lookup; the
+                // epoch guards against the cache evicting the pointee.
+                Bool GetBackendStateMemo(const void*& outState, Uint64& outEpoch) const {
+                    if (m_backendStateMemoVersion != m_configVersion) return false;
+                    outState = m_backendStateMemo;
+                    outEpoch = m_backendStateMemoEpoch;
+                    return true;
+                }
+                void SetBackendStateMemo(const void* state, Uint64 epoch) const {
+                    m_backendStateMemo = state;
+                    m_backendStateMemoEpoch = epoch;
+                    m_backendStateMemoVersion = m_configVersion;
+                }
+
             private:
                 void BumpAttributeFormatVersion(Uint index);
                 void BumpAttributeBufferVersion(Uint index);
@@ -137,6 +154,9 @@ namespace MobileGL {
                 Uint32 m_configVersion = 0;
                 mutable Uint64 m_backendHashMemo = 0;
                 mutable Uint32 m_backendHashMemoVersion = ~0u;
+                mutable const void* m_backendStateMemo = nullptr;
+                mutable Uint64 m_backendStateMemoEpoch = 0;
+                mutable Uint32 m_backendStateMemoVersion = ~0u;
             };
         } // namespace GLState
     } // namespace MG_State