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

MobileGL/MG_Backend/DirectVulkan/Renderer/UniformManager.cpp

@@ -18,6 +18,7 @@
 #include "MG_Util/Converters/MGToVk/TextureEnumConverter.h"
 #include "MG_Util/Metrics/TextureMetrics.h"
 #include <Config.h>
+#include <algorithm>
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
@@ -204,6 +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) {
@@ -1189,16 +1191,47 @@ namespace MobileGL::MG_Backend::DirectVulkan {
                         bufferInfo.range = ubo.range;
                         dynOffset = static_cast<Uint32>(ubo.dynamicOffset);
                     } else {
-                        BufferSlice slice{};
-                        if (!m_bufferManager->UploadTransient(BufferKind::Uniform, frameIndex, ubo.payload,
-                                                              ubo.payloadSize, m_minDynamicOffsetAlignment, slice)) {
-                            MOBILEGL_ASSERT(false, "UniformDescriptorBinder::BindProgramUniformBuffers failed: UBO upload failed on binding %u element %u",
-                                    binding, element);
-                            return false;
+                        // 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)) {
+                                MOBILEGL_ASSERT(false, "UniformDescriptorBinder::BindProgramUniformBuffers failed: UBO upload failed on binding %u element %u",
+                                        binding, element);
+                                return false;
+                            }
+                            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;
+                            }
                         }
-                        bufferInfo.buffer = slice.buffer;
-                        bufferInfo.range = ubo.payloadSize;
-                        dynOffset = static_cast<Uint32>(slice.offset);
                     }
                     bufferInfos.push_back(bufferInfo);
                     // Dynamic offsets are consumed in binding order, then array element order,
@@ -1337,8 +1370,34 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             m_hasLastDescriptor = cacheable;
         }
 
-        vkCmdBindDescriptorSets(commandBuffer, bindPoint, programObj.pipelineLayout, 0, 1,
-                                &descriptorSet, static_cast<Uint32>(dynamicOffsets.size()), dynamicOffsets.data());
+        // 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, 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
@@ -185,6 +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;
-            return it->second;
+            it->second->lastUsedFrameBoundary = m_frameBoundaryCounter;
+            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

@@ -966,6 +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,
                         });
@@ -1142,6 +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,
                 });

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;
@@ -233,6 +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 /

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) {
@@ -714,45 +722,63 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             }
         }
 
-        auto aliveIt = m_aliveObjects.find(identity);
-        if (aliveIt != m_aliveObjects.end() && aliveIt->second.expired()) {
-            EraseTrackedTexture(aliveIt->first);
-            aliveIt = m_aliveObjects.end();
-        }
-
-        // Only (re)register and prune when this (texture, lifetime) pair is new: stale
-        // aliases can only come into existence through an address reuse, which by
-        // construction introduces a new identity. Doing this unconditionally made every
-        // sampled-texture sync scan the entire alive-texture map per draw.
-        if (aliveIt == m_aliveObjects.end()) {
-            WeakPtr<MG_State::GLState::ITextureObject> aliveTexture;
-            const auto& liveTexture = MG_State::pGLContext->GetTextureObject(texture.GetExternalIndex());
-            if (liveTexture && liveTexture.get() == &texture) {
-                aliveTexture = liveTexture;
-            } else {
-                // The name lookup legally fails while the object is alive: the name was
-                // deleted with the texture still attached to an FBO (the attachment's
-                // SharedPtr keeps it alive), or the name was reused by a new texture, or
-                // this is a default texture object (name 0 lives outside the name map).
-                // Register through the object's own control block so the resource created
-                // below still participates in weak-expiry GC instead of becoming an
-                // orphan no reclamation path can reach until Shutdown.
-                aliveTexture = texture.weak_from_this();
-            }
-            if (!aliveTexture.expired()) {
-                m_aliveObjects[identity] = Move(aliveTexture);
-                PruneStaleTextureAliases(&texture);
+        // 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;
             }
         }
 
-        auto it = m_textureResources.find(identity);
-        if (it == m_textureResources.end()) {
-            TextureResource initial{};
-            auto [insertIt, _] = m_textureResources.emplace(identity, Move(initial));
-            it = insertIt;
+        if (resourcePtr == nullptr) {
+            auto aliveIt = m_aliveObjects.find(identity);
+            if (aliveIt != m_aliveObjects.end() && aliveIt->second.expired()) {
+                EraseTrackedTexture(aliveIt->first);
+                aliveIt = m_aliveObjects.end();
+            }
+
+            // Only (re)register and prune when this (texture, lifetime) pair is new: stale
+            // aliases can only come into existence through an address reuse, which by
+            // construction introduces a new identity. Doing this unconditionally made every
+            // sampled-texture sync scan the entire alive-texture map per draw.
+            if (aliveIt == m_aliveObjects.end()) {
+                WeakPtr<MG_State::GLState::ITextureObject> aliveTexture;
+                const auto& liveTexture = MG_State::pGLContext->GetTextureObject(texture.GetExternalIndex());
+                if (liveTexture && liveTexture.get() == &texture) {
+                    aliveTexture = liveTexture;
+                } else {
+                    // The name lookup legally fails while the object is alive: the name was
+                    // deleted with the texture still attached to an FBO (the attachment's
+                    // SharedPtr keeps it alive), or the name was reused by a new texture, or
+                    // this is a default texture object (name 0 lives outside the name map).
+                    // Register through the object's own control block so the resource created
+                    // below still participates in weak-expiry GC instead of becoming an
+                    // orphan no reclamation path can reach until Shutdown.
+                    aliveTexture = texture.weak_from_this();
+                }
+                if (!aliveTexture.expired()) {
+                    m_aliveObjects[identity] = Move(aliveTexture);
+                    PruneStaleTextureAliases(&texture);
+                }
+            }
+
+            auto it = m_textureResources.find(identity);
+            if (it == m_textureResources.end()) {
+                TextureResource initial{};
+                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) {
@@ -1708,6 +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();
@@ -2014,11 +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)");
-        vkDestroyFence(m_device, uploadFence, nullptr);
-        vkFreeCommandBuffers(m_device, m_commandPool, 1, &commandBuffer);
-
-        vmaDestroyBuffer(m_allocator, stagingBuffer, stagingAllocation);
+        // Do NOT wait the fence here: this submit sits behind the previous
+        // frame's rendering on the queue, so a synchronous wait stalls the CPU
+        // until the GPU drains - a per-frame vkQueueWaitIdle for any workload
+        // with animated textures. Ordering against the current frame's draws is
+        // 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;
@@ -420,6 +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);
@@ -449,6 +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;
@@ -458,5 +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

@@ -226,6 +226,26 @@ namespace MobileGL::MG_Backend::DirectVulkan {
     // 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;
@@ -815,16 +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:
@@ -975,8 +985,11 @@ namespace MobileGL::MG_Backend::DirectVulkan {
             if (trackedAttachment.target != TrackedAttachmentTarget::Texture) {
                 continue;
             }
-            const auto trackedTexture = trackedAttachment.texture.lock();
-            if (trackedTexture && trackedTexture.get() == &texture) {
+            // Raw identity compare (see textureRaw): the caller's texture is
+            // 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;
             }
         }
@@ -2911,7 +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));
@@ -3176,8 +3189,29 @@ void main() {
         }
 
         if (bindingCount > 0) {
-            vkCmdBindVertexBuffers(commandBuffer, 0, static_cast<Uint32>(bindingCount), vkBuffers.data(),
-                                   vkOffsets.data());
+            auto& shadow = g_dynamicStateShadow;
+            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;
     }
@@ -3247,8 +3281,17 @@ void main() {
             MGLOG_E("DrawElements skipped: failed to sync resident index buffer");
             return false;
         }
-        vkCmdBindIndexBuffer(frame.commandBuffer, slice.buffer,
-                             slice.offset + static_cast<VkDeviceSize>(pIndexBufferView->indexByteOffset), vkIndexType);
+        const VkDeviceSize indexBindOffset =
+            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;
     }
 
@@ -3815,16 +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.
-        const Uint64 vertexInputHash = m_vertexInputStateFactory->GetOrComputeHash(vao);
+        // Reset per-frame and on pipeline destruction so a memoized handle can never dangle.
+        // 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 &&
-            m_lastPipelineProgramHash == programObj.hash && m_lastPipelineVertexInputHash == vertexInputHash &&
-            m_lastPipelineRenderPassHash == renderPassHash &&
-            m_lastPipelineRenderStateVersion == renderStateVersion &&
-            m_lastPipelineTransformFlags == transformFlags) {
-            return m_lastPipelineResult;
+        for (Uint32 i = 0; i < m_pipelineMemoCount; ++i) {
+            const PipelineMemoEntry& entry = m_pipelineMemo[i];
+            if (entry.pipeline != VK_NULL_HANDLE && entry.mode == mode &&
+                entry.programHash == programObj.hash && entry.vertexInputHash == vertexLayoutHash &&
+                entry.renderPassHash == renderPassHash &&
+                entry.renderStateVersion == renderStateVersion &&
+                entry.transformFlags == transformFlags) {
+                return entry.pipeline;
+            }
         }
 
 #if MOBILEGL_LOG_ACTIVE_LEVEL <= MOBILEGL_LOG_LEVEL_DEBUG
@@ -3865,9 +3916,7 @@ void main() {
         }
 #endif
 
-        // vertexInputHash was computed above for the fast-path key; reuse it here.
-        auto& vis = m_vertexInputStateFactory->GetOrCreateVertexInputState(vao, vertexInputHash);
-        const Uint32 vertexInputAttribMask = BuildVertexInputAttributeMask(vis.attributes);
+        const Uint32 vertexInputAttribMask = vis.attributeLocationMask;
         const Uint32 activeAttribMask = programObj.activeVertexInputLocationMask;
         const Uint32 missingAttribMask = activeAttribMask & ~vertexInputAttribMask;
         auto& patchedAttributes = m_patchedAttributesScratch;
@@ -3977,7 +4026,7 @@ void main() {
 
         PipelineFactory::PipelineCreatePayload payload {
             .programHash = programObj.hash,
-            .vertexInputHash = vertexInputHash,
+            .vertexInputHash = vertexLayoutHash,
             .pipelineLayout = programObj.pipelineLayout,
             .renderPass = renderPassEntry.renderPass,
             .colorAttachmentCount = renderPassEntry.colorAttachmentCount,
@@ -4284,14 +4333,16 @@ void main() {
         }
         VkPipeline pipeline = m_pipelineFactory->GetOrCreatePipeline(payload);
         if (pipeline != VK_NULL_HANDLE) {
-            m_lastPipelineValid = true;
-            m_lastPipelineMode = mode;
-            m_lastPipelineProgramHash = programObj.hash;
-            m_lastPipelineVertexInputHash = vertexInputHash;
-            m_lastPipelineRenderPassHash = renderPassHash;
-            m_lastPipelineRenderStateVersion = renderStateVersion;
-            m_lastPipelineTransformFlags = transformFlags;
-            m_lastPipelineResult = pipeline;
+            PipelineMemoEntry& entry = m_pipelineMemo[m_pipelineMemoNext];
+            entry.mode = mode;
+            entry.programHash = programObj.hash;
+            entry.vertexInputHash = vertexLayoutHash;
+            entry.renderPassHash = renderPassHash;
+            entry.renderStateVersion = renderStateVersion;
+            entry.transformFlags = transformFlags;
+            entry.pipeline = pipeline;
+            m_pipelineMemoNext = (m_pipelineMemoNext + 1) % kPipelineMemoSize;
+            m_pipelineMemoCount = std::min(m_pipelineMemoCount + 1, kPipelineMemoSize);
         }
         return pipeline;
     }
@@ -4393,6 +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) {
@@ -4401,6 +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)) {
@@ -4410,16 +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.
-        if (UniformManager::ProgramSamplesOnlySingleLevelTextures(program, *programObjPtr)) {
-            transformFlags |= ProgramFactory::CompileOptionBit::ExplicitLod0Sampling;
-            programObjPtr = &m_programFactory->GetOrCreateProgram(program, transformFlags);
+        // single mip level. The probe walks every sampler binding, so its verdict is memoized
+        // 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;
+                }
+                m_lastLodDecisionValid = true;
+                m_lastLodProgramLifetimeId = lodProgramLifetimeId;
+                m_lastLodProgramVersion = lodProgramVersion;
+                m_lastLodBindGeneration = lodBindGeneration;
+                m_lastLodBaseFlags = baseFlags;
+                m_lastLodResultFlags = transformFlags;
+                m_lastLodParamsSum = 0;  // filled below once the sampled set is known
+            }
         }
-        const auto& programObj = *programObjPtr;
+        const auto& programObj = m_programFactory->GetOrCreateProgram(program, transformFlags);
 
         // Begin command recording if not yet
         if (!frame.isCommandRecording) {
@@ -4464,6 +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(),
@@ -4487,7 +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;
             }
@@ -4496,6 +4727,7 @@ 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));
@@ -4534,10 +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());
@@ -4601,7 +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;
 
@@ -4629,7 +4874,11 @@ void main() {
             MOBILEGL_ASSERT(ok, "%s: BeginRenderPass failed", __func__);
         }
 
-        vkCmdBindPipeline(frame.commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
+        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());
@@ -4670,6 +4919,48 @@ void main() {
             scissor.extent = { (Uint)renderPassEntry->extent.x(), (Uint)renderPassEntry->extent.y() };
         }
         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;
     }
 
@@ -6412,7 +6703,7 @@ 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
         }
         // The pre-pass stream must never be submitted later than the recording
         // it was paired with (frame commands recorded after a pre-pass move
@@ -7499,8 +7790,7 @@ void main() {
             m_programFactory->OnFrameBoundary();
         }
         if (m_pipelineFactory && m_pipelineFactory->OnFrameBoundary() > 0) {
-            m_lastPipelineValid = false;
-            m_lastPipelineResult = VK_NULL_HANDLE;
+            InvalidatePipelineMemo();
         }
         if (m_vertexInputStateFactory) {
             m_vertexInputStateFactory->OnFrameBoundary();
@@ -7617,8 +7907,7 @@ 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;
-        m_lastPipelineResult = VK_NULL_HANDLE;
+        InvalidatePipelineMemo();
 
         // The submitted command buffer may still be executing; recording must
         // restart on a fresh one. If none can be allocated, fall back to
@@ -7691,6 +7980,10 @@ 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) {
@@ -7771,7 +8064,7 @@ void main() {
                 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
@@ -7807,8 +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
-            m_lastPipelineResult = VK_NULL_HANDLE;
+            InvalidatePipelineMemo(); // an aged-out pipeline may still be memoized
+            // 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();
@@ -7831,7 +8127,7 @@ 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();
 
@@ -8834,7 +9130,9 @@ 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;
@@ -8997,8 +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;
-            m_lastPipelineResult = VK_NULL_HANDLE;
+            InvalidatePipelineMemo();
         }
     }
 
@@ -9017,8 +9314,7 @@ void main() {
             m_computePipelines.erase(computeIt);
         }
         if (m_pipelineFactory != nullptr && m_pipelineFactory->EvictByProgramHash(programHash) > 0) {
-            m_lastPipelineValid = false;
-            m_lastPipelineResult = VK_NULL_HANDLE;
+            InvalidatePipelineMemo();
         }
         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;
-        GLenum m_lastPipelineMode = 0;
-        Uint64 m_lastPipelineProgramHash = 0;
-        Uint64 m_lastPipelineVertexInputHash = 0;
-        Uint64 m_lastPipelineRenderPassHash = 0;
-        Uint m_lastPipelineRenderStateVersion = 0;
-        ProgramFactory::CompileOptionFlags m_lastPipelineTransformFlags = {};
-        VkPipeline m_lastPipelineResult = VK_NULL_HANDLE;
+        // Small N-way pipeline-resolution memo (round-robin replacement). A
+        // single-entry memo thrashed on draw sequences that alternate a few
+        // pipelines (GUI text/quad program ping-pong), paying the full
+        // payload-hash lookup per draw; eight entries cover such working sets
+        // while keeping the hit path a trivial linear scan.
+        struct PipelineMemoEntry {
+            GLenum mode = 0;
+            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