react-native-webgpu 0.10.2 → 0.10.3

package/cpp/rnwgpu/SurfaceRegistry.h

@@ -1,17 +1,31 @@
 #pragma once
 
 #include <algorithm>
+#include <atomic>
+#include <cerrno>
+#include <chrono>
+#include <condition_variable>
+#include <cstdint>
+#include <deque>
 #include <functional>
 #include <memory>
 #include <mutex>
 #include <shared_mutex>
 #include <stdexcept>
+#include <thread>
 #include <unordered_map>
 #include <utility>
 #include <vector>
 
 #include "webgpu/webgpu_cpp.h"
 
+#if defined(__ANDROID__)
+#include <android/hardware_buffer.h>
+#include <android/log.h>
+#include <poll.h>
+#include <unistd.h>
+#endif
+
 #ifdef __APPLE__
 namespace dawn::native::metal {
 void WaitForCommandsToBeScheduled(WGPUDevice device);
@@ -44,6 +58,64 @@ struct Size {
 // CAMetalLayer on Apple platforms). May run on any thread.
 using NativeSurfaceReleaser = std::function<void(void *)>;
 
+#if defined(__ANDROID__)
+// --- AHB-pool presentation mode (WebGPUHardwareBufferView)
+// -----------------------------
+//
+// A third presentation backend (besides the on-screen wgpu::Surface swapchain
+// and the offscreen texture). WebGPU renders into a small pool of native
+// AHardwareBuffers imported as Dawn SharedTextureMemory; the view draws each
+// finished buffer inline via Bitmap.wrapHardwareBuffer so it behaves like a
+// normal RN view. The pool is sized from the JS canvas drawing buffer
+// (_canvas->getWidth()/Height()) and reallocated when that changes, exactly
+// like the swapchain, so the canvas texture always matches the app's other
+// attachments (e.g. its depth texture).
+
+#define RNWGPU_LOG_TAG "WebGPUHardwareBufferView"
+// Failure / anomaly reporting only (never per-frame).
+#define RNWGPU_POOL_WARN(...)                                                  \
+  __android_log_print(ANDROID_LOG_WARN, RNWGPU_LOG_TAG, __VA_ARGS__)
+
+enum class SlotState : uint8_t {
+  Free,      // available for poolGetCurrentTexture
+  Rendering, // BeginAccess done, JS rendering into it
+  Presented, // EndAccess done, fence(s) queued for the waiter
+  Ready,     // fence signaled, awaiting UI pickup
+  Displayed, // handed to the consumer, held until released (held-ring)
+};
+
+struct PoolSlot {
+  AHardwareBuffer *ahb = nullptr; // owned by the pool (allocated natively)
+  wgpu::SharedTextureMemory memory = nullptr;
+  wgpu::Texture texture = nullptr;
+  SlotState state = SlotState::Free;
+};
+
+// One generation of the pool (one buffer size). Reference-counted (shared_ptr)
+// so an in-flight render / present / ready slot keeps the whole generation
+// alive across a resize. The destructor frees every AHB and tears down the Dawn
+// imports. The Java side keeps its own ref (via wrapHardwareBuffer) for
+// anything it is still drawing, so a generation can be freed here without
+// disturbing a frame still on screen.
+struct AHBPool {
+  std::vector<PoolSlot> slots;
+  uint32_t generation = 0;
+  int width = 0;
+  int height = 0;
+
+  ~AHBPool() {
+    for (auto &slot : slots) {
+      slot.texture = nullptr;
+      slot.memory = nullptr;
+      if (slot.ahb != nullptr) {
+        AHardwareBuffer_release(slot.ahb);
+        slot.ahb = nullptr;
+      }
+    }
+  }
+};
+#endif
+
 // Bridges the asynchronous native surface lifecycle (surfaces appear and
 // disappear on the platform UI thread) with the synchronous WebGPU canvas API
 // (the JS render loop must always be able to acquire a texture).
@@ -79,6 +151,16 @@ public:
       : _gpu(std::move(gpu)), _width(width), _height(height) {}
 
   ~SurfaceInfo() {
+#if defined(__ANDROID__)
+    // Stop the fence waiter before any pool generation is freed.
+    stopWaiter();
+    {
+      std::lock_guard<std::mutex> poolLock(_poolMutex);
+      _presentQueue.clear();
+      poolResetLocked();
+      _poolDevice = nullptr;
+    }
+#endif
     // Drop the Dawn objects before releasing the native surfaces they borrow.
     _surface = nullptr;
 #if defined(__ANDROID__)
@@ -250,6 +332,20 @@ public:
     _config.presentMode = wgpu::PresentMode::Fifo;
     _texture = nullptr;
     _frameEpoch++;
+#if defined(__ANDROID__)
+    {
+      std::lock_guard<std::mutex> poolLock(_poolMutex);
+      if (_poolDevice.Get() != _config.device.Get()) {
+        // Buffers imported for another device cannot be reused.
+        poolResetLocked();
+      }
+      _poolDevice = _config.device;
+      _poolFormat = _config.format;
+      _poolUsage = _config.usage;
+      // A new device may support AHB sharing even if the previous one did not.
+      _poolUnsupported = false;
+    }
+#endif
     _configureLocked();
   }
 
@@ -276,6 +372,13 @@ public:
     _viewFormats.clear();
     _acquiredFromSurface = false;
     _frameEpoch++;
+#if defined(__ANDROID__)
+    {
+      std::lock_guard<std::mutex> poolLock(_poolMutex);
+      poolResetLocked();
+      _poolDevice = nullptr;
+    }
+#endif
   }
 
   bool isConfigured() {
@@ -303,6 +406,16 @@ public:
   void releaseSurfaceForDevice(const wgpu::Device &device) {
 #if defined(__ANDROID__)
     std::unique_lock<std::shared_mutex> lock(_mutex);
+    {
+      // The pool's SharedTextureMemory imports belong to the device too: drop
+      // them while it is alive. The next getCurrentTexture() after a configure
+      // with a live device reallocates the pool.
+      std::lock_guard<std::mutex> poolLock(_poolMutex);
+      if (_poolDevice && _poolDevice.Get() == device.Get()) {
+        poolResetLocked();
+        _poolDevice = nullptr;
+      }
+    }
     if (!_surface || !_surfaceDevice || _surfaceDevice.Get() != device.Get()) {
       return;
     }
@@ -321,6 +434,13 @@ public:
   // otherwise (see RNWebGPU::destroyContext).
   bool hasNativeSurface() {
     std::shared_lock<std::shared_mutex> lock(_mutex);
+#if defined(__ANDROID__)
+    // A WebGPUHardwareBufferView in pool mode owns presentation the same way a
+    // surface-backed view does, and retires the entry on its own teardown.
+    if (_poolMode.load()) {
+      return true;
+    }
+#endif
     return _nativeSurface != nullptr || _hasPendingAttach;
   }
 
@@ -331,6 +451,34 @@ public:
     // Start-of-frame boundary; a new acquire supersedes any previous frame
     // that never presented.
     applyPendingAttach(/* supersedeInFlightFrame = */ true);
+#if defined(__ANDROID__)
+    if (_poolMode.load()) {
+      // The pool tracks the drawing buffer (_config.width/height, kept in sync
+      // with canvas.width/height by reconfigure()), like the swapchain, so the
+      // canvas texture always matches the app's other attachments.
+      int poolW = 0;
+      int poolH = 0;
+      {
+        std::shared_lock<std::shared_mutex> lock(_mutex);
+        if (_config.device != nullptr) {
+          poolW = static_cast<int>(_config.width);
+          poolH = static_cast<int>(_config.height);
+        }
+      }
+      poolResize(poolW, poolH);
+      if (auto texture = poolGetCurrentTexture()) {
+        std::unique_lock<std::shared_mutex> lock(_mutex);
+        _frameInFlight = true;
+        _acquiredFromSurface = false;
+        _frameEpoch++;
+        return texture;
+      }
+      // No slot available (pool not allocated yet, unsupported device, view
+      // detached mid-frame, or a free-slot timeout): fall through to the
+      // offscreen fallback so the render loop survives; this frame is simply
+      // not shown.
+    }
+#endif
     std::unique_lock<std::shared_mutex> lock(_mutex);
     if (_config.device == nullptr) {
       throw std::runtime_error(
@@ -363,6 +511,13 @@ public:
   // acquire failure) are dropped. This is also the end-of-frame boundary: it
   // adopts a surface that attached while the frame was in flight.
   void presentFrame() {
+#if defined(__ANDROID__)
+    if (_poolMode.load()) {
+      // Ends access on the pool slot and hands it to the fence waiter; the
+      // view picks it up once its render-complete fence signals.
+      poolPresent();
+    }
+#endif
 #ifdef __APPLE__
     // Ensure command buffers are scheduled before presenting. Read the device
     // under a shared lock, then wait without holding it (the wait can block).
@@ -404,12 +559,161 @@ public:
     return _config;
   }
 
+#if defined(__ANDROID__)
+  // --- AHB-pool API (called from the JNI layer, see cpp-adapter.cpp) ---------
+
+  // Turn on pool mode for this context. dpW/dpH is the canvas-client (dp) size
+  // reported to JS via getSize(); the actual buffers are sized in pool px from
+  // the canvas drawing buffer in poolResize().
+  void enablePool(int dpW, int dpH) {
+    {
+      std::unique_lock<std::shared_mutex> lock(_mutex);
+      _width = dpW;
+      _height = dpH;
+    }
+    std::lock_guard<std::mutex> poolLock(_poolMutex);
+    _poolMode.store(true);
+    _shutdown = false;
+  }
+
+  // Set the canvas-client (dp) size without changing pool mode. Keeps getSize()
+  // (and thus the JS canvas.clientWidth/Height) in sync on resize.
+  void setPoolClientSize(int dpW, int dpH) {
+    std::unique_lock<std::shared_mutex> lock(_mutex);
+    _width = dpW;
+    _height = dpH;
+  }
+
+  // (Re)allocate the pool to match the canvas drawing buffer (px). Called from
+  // getCurrentTexture() on the rendering thread before acquiring a slot, so the
+  // canvas texture size always tracks the app's other attachments.
+  void poolResize(int pxW, int pxH) {
+    std::lock_guard<std::mutex> poolLock(_poolMutex);
+    if (!_poolMode.load() || _poolDevice == nullptr || pxW <= 0 || pxH <= 0) {
+      return;
+    }
+    if (_pool != nullptr && _pool->width == pxW && _pool->height == pxH) {
+      return;
+    }
+    if (_poolUnsupported) {
+      return;
+    }
+    if (!_poolDevice.HasFeature(
+            wgpu::FeatureName::SharedTextureMemoryAHardwareBuffer)) {
+      _poolUnsupported = true;
+      RNWGPU_POOL_WARN("device lacks SharedTextureMemoryAHardwareBuffer; the "
+                       "transparent canvas cannot present and stays blank");
+      return;
+    }
+    auto pool = allocatePoolLocked(pxW, pxH);
+    if (pool == nullptr) {
+      return;
+    }
+    _pool = pool;
+    _pools[pool->generation] = pool;
+    // Keep current + previous generation (in-flight frames of the previous size
+    // may still be on screen during the cross-fade); drop anything older.
+    for (auto it = _pools.begin(); it != _pools.end();) {
+      if (it->first + 1 < pool->generation) {
+        it = _pools.erase(it);
+      } else {
+        ++it;
+      }
+    }
+    _freeCv.notify_all();
+  }
+
+  // Latest signaled frame awaiting display, encoded as (generation << 32 |
+  // slot). Returns -1 when there is nothing new. Marks the slot Displayed so
+  // the producer will not re-render into it until releaseSlot() is called.
+  // Called on the UI thread after the frame-ready wake-up.
+  int64_t poolPollReady() {
+    std::lock_guard<std::mutex> poolLock(_poolMutex);
+    if (_readySlot < 0 || _readyPool == nullptr) {
+      return -1;
+    }
+    int idx = _readySlot;
+    uint32_t gen = _readyPool->generation;
+    _readyPool->slots[idx].state = SlotState::Displayed;
+    _readyPool = nullptr;
+    _readySlot = -1;
+    return (static_cast<int64_t>(gen) << 32) | static_cast<uint32_t>(idx);
+  }
+
+  // The AHardwareBuffer backing a (generation, slot), for the consumer to wrap
+  // in a Bitmap. Returns nullptr for a retired generation. Called on the UI
+  // thread; the JNI layer converts it to a HardwareBuffer jobject.
+  void *poolBufferForDisplay(uint32_t gen, int slot) {
+    std::lock_guard<std::mutex> poolLock(_poolMutex);
+    auto it = _pools.find(gen);
+    if (it == _pools.end() || slot < 0 ||
+        slot >= static_cast<int>(it->second->slots.size())) {
+      return nullptr;
+    }
+    return it->second->slots[slot].ahb;
+  }
+
+  // The consumer is done displaying (and holding) a slot: return it to the free
+  // list so the producer may render into it again. No-op for a retired
+  // generation (those buffers are never reused). Called on the UI thread.
+  void poolReleaseSlot(uint32_t gen, int idx) {
+    std::lock_guard<std::mutex> poolLock(_poolMutex);
+    if (_pool != nullptr && _pool->generation == gen && idx >= 0 &&
+        idx < static_cast<int>(_pool->slots.size())) {
+      _pool->slots[idx].state = SlotState::Free;
+      _freeCv.notify_one();
+    }
+  }
+
+  // Invoked from the waiter thread (never under _poolMutex) each time a new
+  // frame becomes ready, so the consumer can wake up instead of polling every
+  // vsync. Pass nullptr to unregister.
+  void setPoolFrameReadyCallback(std::function<void()> cb) {
+    std::lock_guard<std::mutex> poolLock(_poolMutex);
+    _poolFrameReady = std::move(cb);
+  }
+#endif
+
 private:
   void detach(bool createFallbackTexture) {
     void *releasedSurfaces[2] = {nullptr, nullptr};
     NativeSurfaceReleaser releasers[2];
+#if defined(__ANDROID__)
+    // Destroyed outside the lock: it releases a JNI reference.
+    std::function<void()> releasedFrameReady;
+#endif
     {
       std::unique_lock<std::shared_mutex> lock(_mutex);
+#if defined(__ANDROID__)
+      // Leaving pool mode (the WebGPUHardwareBufferView detached): drop the
+      // pool. A configured context keeps rendering into the offscreen
+      // fallback, exactly like the surface path.
+      if (_poolMode.load()) {
+        _poolMode.store(false);
+        {
+          std::lock_guard<std::mutex> poolLock(_poolMutex);
+          // A frame between getCurrentTexture and present still has an open
+          // BeginAccess; end it and queue it so the waiter drains its fences
+          // before the generation is freed (presentFrame() no-ops for it
+          // because pool mode is already off).
+          if (_renderPool != nullptr && _currentRenderSlot >= 0 &&
+              _renderPool->slots[_currentRenderSlot].state ==
+                  SlotState::Rendering) {
+            poolQueuePresentLocked();
+          }
+          // _presentQueue is intentionally left alone: the waiter still has
+          // to wait on the queued fences before the generations they pin can
+          // be released; it discards them on publish (no live pool).
+          poolResetLocked();
+          releasedFrameReady = std::move(_poolFrameReady);
+          _poolFrameReady = nullptr;
+        }
+        if (createFallbackTexture && _config.device != nullptr && !_texture) {
+          _texture = createOffscreenTextureLocked();
+        }
+        _frameEpoch++;
+      }
+#endif
       // The platform is tearing surfaces down; a not-yet-adopted attach is
       // stale, cancel it.
       if (_hasPendingAttach) {
@@ -583,6 +887,281 @@ private:
   }
 #endif
 
+#if defined(__ANDROID__)
+  // Drop every pool generation and the ready / in-flight slot bookkeeping.
+  // Caller holds _poolMutex. Generations pinned by a queued present survive
+  // through their PendingPresent until the waiter is done with them.
+  void poolResetLocked() {
+    _pool = nullptr;
+    _pools.clear();
+    _readyPool = nullptr;
+    _readySlot = -1;
+    _renderPool = nullptr;
+    _currentRenderSlot = -1;
+    _freeCv.notify_all();
+  }
+
+  // Allocate + import one generation of the pool. Caller holds _poolMutex.
+  std::shared_ptr<AHBPool> allocatePoolLocked(int w, int h) {
+    auto pool = std::make_shared<AHBPool>();
+    pool->generation = ++_genCounter;
+    pool->width = w;
+    pool->height = h;
+    pool->slots.resize(kPoolSize);
+    for (auto &slot : pool->slots) {
+      AHardwareBuffer_Desc desc = {};
+      desc.width = static_cast<uint32_t>(w);
+      desc.height = static_cast<uint32_t>(h);
+      desc.layers = 1;
+      desc.format = AHARDWAREBUFFER_FORMAT_R8G8B8A8_UNORM;
+      desc.usage = AHARDWAREBUFFER_USAGE_GPU_SAMPLED_IMAGE |
+                   AHARDWAREBUFFER_USAGE_GPU_COLOR_OUTPUT;
+      int err = AHardwareBuffer_allocate(&desc, &slot.ahb);
+      if (err != 0 || slot.ahb == nullptr) {
+        RNWGPU_POOL_WARN("AHardwareBuffer_allocate failed (%d) %dx%d", err, w,
+                         h);
+        return nullptr;
+      }
+      wgpu::SharedTextureMemoryDescriptor memDesc{};
+      wgpu::SharedTextureMemoryAHardwareBufferDescriptor ahbDesc{};
+      ahbDesc.handle = slot.ahb;
+      memDesc.nextInChain = &ahbDesc;
+      slot.memory = _poolDevice.ImportSharedTextureMemory(&memDesc);
+      // A failed import yields a non-null Dawn error object, so probing the
+      // properties is the actual validity check.
+      wgpu::SharedTextureMemoryProperties props{};
+      if (slot.memory == nullptr ||
+          slot.memory.GetProperties(&props) != wgpu::Status::Success) {
+        RNWGPU_POOL_WARN("ImportSharedTextureMemory failed (%dx%d)", w, h);
+        return nullptr;
+      }
+      // The AHB is rgba8unorm; the pool cannot honor another configured format.
+      // The canvas is configured rgba8unorm on Android (the preferred format),
+      // so this only fires for apps hardcoding e.g. bgra8unorm.
+      if (_poolFormat != wgpu::TextureFormat::Undefined &&
+          _poolFormat != wgpu::TextureFormat::RGBA8Unorm) {
+        RNWGPU_POOL_WARN("configured canvas format %d is not supported by the "
+                         "transparent canvas; using rgba8unorm (use "
+                         "navigator.gpu.getPreferredCanvasFormat())",
+                         static_cast<int>(_poolFormat));
+      }
+      // Honor the configured usage flags as far as the imported memory allows.
+      wgpu::TextureUsage wanted =
+          _poolUsage | wgpu::TextureUsage::RenderAttachment;
+      wgpu::TextureUsage usage = wanted & props.usage;
+      if (usage != wanted) {
+        RNWGPU_POOL_WARN("configured canvas usage 0x%llx narrowed to 0x%llx "
+                         "(unsupported by the transparent canvas buffers)",
+                         static_cast<unsigned long long>(wanted),
+                         static_cast<unsigned long long>(usage));
+      }
+      wgpu::TextureDescriptor texDesc{};
+      texDesc.format = props.format;
+      texDesc.usage = usage;
+      texDesc.size.width = static_cast<uint32_t>(w);
+      texDesc.size.height = static_cast<uint32_t>(h);
+      slot.texture = slot.memory.CreateTexture(&texDesc);
+      slot.state = SlotState::Free;
+    }
+    return pool;
+  }
+
+  wgpu::Texture poolGetCurrentTexture() {
+    std::unique_lock<std::mutex> lock(_poolMutex);
+    // Per the WebGPU spec, getCurrentTexture returns the same texture until the
+    // frame is presented; a repeat call must not consume another slot.
+    if (_renderPool != nullptr && _currentRenderSlot >= 0) {
+      return _renderPool->slots[_currentRenderSlot].texture;
+    }
+    if (_pool == nullptr) {
+      return nullptr; // allocation failed / unsupported (already reported)
+    }
+    auto pool = _pool;
+    int idx = -1;
+    // Bounded wait: if no slot frees up (e.g. a wedged fence), skip the frame
+    // instead of parking the JS/worklet thread forever.
+    bool ready = _freeCv.wait_for(lock, std::chrono::seconds(1), [&]() {
+      if (_shutdown || pool != _pool) {
+        return true; // pool replaced or shutting down: bail out
+      }
+      for (size_t i = 0; i < pool->slots.size(); i++) {
+        if (pool->slots[i].state == SlotState::Free) {
+          idx = static_cast<int>(i);
+          return true;
+        }
+      }
+      return false;
+    });
+    if (!ready) {
+      RNWGPU_POOL_WARN("poolGetCurrentTexture: timed out waiting for a free "
+                       "slot; skipping frame");
+      return nullptr;
+    }
+    if (idx < 0 || _shutdown || pool != _pool) {
+      return nullptr;
+    }
+    auto &slot = pool->slots[idx];
+    slot.state = SlotState::Rendering;
+    _renderPool = pool;
+    _currentRenderSlot = idx;
+    poolBeginAccess(slot);
+    return slot.texture;
+  }
+
+  void poolPresent() {
+    std::lock_guard<std::mutex> lock(_poolMutex);
+    if (_currentRenderSlot < 0 || _renderPool == nullptr) {
+      return;
+    }
+    poolQueuePresentLocked();
+  }
+
+  // Ends access on the in-flight render slot and hands it (with its fences) to
+  // the waiter. Caller holds _poolMutex and guarantees _renderPool /
+  // _currentRenderSlot are valid.
+  void poolQueuePresentLocked() {
+    auto pool = _renderPool;
+    int idx = _currentRenderSlot;
+    auto &slot = pool->slots[idx];
+
+    wgpu::SharedTextureMemoryEndAccessState state{};
+    wgpu::SharedTextureMemoryVkImageLayoutEndState vkLayout{};
+    state.nextInChain = &vkLayout;
+    slot.memory.EndAccess(slot.texture, &state);
+
+    std::vector<wgpu::SharedFence> fences;
+    fences.reserve(state.fenceCount);
+    for (size_t i = 0; i < state.fenceCount; i++) {
+      fences.push_back(state.fences[i]);
+    }
+
+    slot.state = SlotState::Presented;
+    _presentQueue.push_back({pool, idx, std::move(fences)});
+    _renderPool = nullptr;
+    _currentRenderSlot = -1;
+    startWaiterLocked();
+    _presentCv.notify_one();
+  }
+
+  void poolBeginAccess(PoolSlot &slot) {
+    wgpu::SharedTextureMemoryBeginAccessDescriptor desc{};
+    desc.initialized = false; // canvas contents are fully redrawn each frame
+    desc.concurrentRead = false;
+    desc.fenceCount = 0;
+    desc.fences = nullptr;
+    desc.signaledValues = nullptr;
+    wgpu::SharedTextureMemoryVkImageLayoutBeginState vkLayout{};
+    vkLayout.oldLayout = 0;
+    vkLayout.newLayout = 0;
+    desc.nextInChain = &vkLayout;
+    slot.memory.BeginAccess(slot.texture, &desc);
+  }
+
+  void startWaiterLocked() {
+    if (_waiterRunning) {
+      return;
+    }
+    _waiterRunning = true;
+    _waiter = std::thread([this]() { waiterLoop(); });
+  }
+
+  void stopWaiter() {
+    {
+      std::lock_guard<std::mutex> lock(_poolMutex);
+      _shutdown = true;
+      _presentCv.notify_all();
+      _freeCv.notify_all();
+    }
+    if (_waiter.joinable()) {
+      _waiter.join();
+    }
+    _waiterRunning = false;
+  }
+
+  // Blocks on each presented frame's render-complete fence (off the UI and JS
+  // threads), then publishes it as the latest ready slot. This is the rigorous
+  // acquire-side wait: HWUI only samples a buffer after the UI thread picks it
+  // up via poolPollReady, which happens strictly after this wait returns.
+  void waiterLoop() {
+    while (true) {
+      PendingPresent pending;
+      {
+        std::unique_lock<std::mutex> lock(_poolMutex);
+        _presentCv.wait(lock,
+                        [&]() { return _shutdown || !_presentQueue.empty(); });
+        if (_shutdown && _presentQueue.empty()) {
+          return;
+        }
+        pending = std::move(_presentQueue.front());
+        _presentQueue.pop_front();
+      }
+
+      // Wait outside the lock. Each fd is owned by its SharedFence and closed
+      // when `pending.fences` is destroyed at the end of this iteration, so we
+      // never dup / double-close (avoids the fdsan abort, see
+      // GPUSharedFence.cpp).
+      for (auto &fence : pending.fences) {
+        wgpu::SharedFenceExportInfo info{};
+        wgpu::SharedFenceSyncFDExportInfo fdInfo{};
+        info.nextInChain = &fdInfo;
+        fence.ExportInfo(&info);
+        if (info.type != wgpu::SharedFenceType::SyncFD) {
+          // fdInfo was not populated (its default handle 0 is NOT a fence fd);
+          // we have no way to wait on this fence type.
+          RNWGPU_POOL_WARN("waiter: unexpected shared fence type %d, cannot "
+                           "wait for render completion",
+                           static_cast<int>(info.type));
+          continue;
+        }
+        if (fdInfo.handle >= 0) {
+          // Sync-fence fds become readable (POLLIN) when signaled; poll()
+          // avoids any libsync linkage concern. Poll in bounded slices so
+          // teardown (which flips _shutdown) cannot hang here on a wedged GPU.
+          struct pollfd pfd;
+          pfd.fd = fdInfo.handle;
+          pfd.events = POLLIN;
+          while (!_shutdown.load()) {
+            pfd.revents = 0;
+            int r = poll(&pfd, 1, 100);
+            if (r > 0) {
+              break; // signaled (or POLLERR): stop waiting
+            }
+            if (r < 0 && errno != EINTR) {
+              break; // real poll error; EINTR just retries
+            }
+          }
+        }
+      }
+
+      std::function<void()> frameReady;
+      {
+        std::lock_guard<std::mutex> lock(_poolMutex);
+        // Pool torn down (switchToOffscreen) while this frame was in flight:
+        // discard it now that its fences have been drained. Publishing it would
+        // pin the retired generation forever (nothing polls anymore).
+        if (_pool == nullptr) {
+          continue;
+        }
+        // Supersede a still-unclaimed ready frame: drop it back to Free if it
+        // belongs to the live pool (a retired pool is simply discarded).
+        if (_readyPool != nullptr && _readySlot >= 0) {
+          if (_readyPool == _pool) {
+            _readyPool->slots[_readySlot].state = SlotState::Free;
+            _freeCv.notify_one();
+          }
+        }
+        pending.pool->slots[pending.slot].state = SlotState::Ready;
+        _readyPool = pending.pool;
+        _readySlot = pending.slot;
+        frameReady = _poolFrameReady;
+      }
+      if (frameReady) {
+        frameReady(); // outside the lock: it calls into Java
+      }
+    }
+  }
+#endif
+
   mutable std::shared_mutex _mutex;
 #if defined(__ANDROID__)
   // DAWN_WORKAROUND_DEVICE_DESTROY_BEFORE_SURFACE_RELEASE
@@ -628,6 +1207,40 @@ private:
   // canvas).
   int _width;
   int _height;
+
+#if defined(__ANDROID__)
+  // Pool state. Guarded by _poolMutex (never nested under _mutex while
+  // blocking).
+  struct PendingPresent {
+    std::shared_ptr<AHBPool> pool;
+    int slot;
+    std::vector<wgpu::SharedFence> fences;
+  };
+
+  static constexpr int kPoolSize =
+      5; // 2 held + 1 rendering + 1 waiting + 1 ready
+
+  std::atomic<bool> _poolMode{false};
+  std::mutex _poolMutex;
+  std::condition_variable _freeCv;    // a slot returned to Free
+  std::condition_variable _presentCv; // a frame was queued for the waiter
+  wgpu::Device _poolDevice = nullptr;
+  wgpu::TextureFormat _poolFormat = wgpu::TextureFormat::Undefined;
+  wgpu::TextureUsage _poolUsage = wgpu::TextureUsage::RenderAttachment;
+  bool _poolUnsupported = false;  // device lacks AHB shared-texture support
+  std::shared_ptr<AHBPool> _pool; // current generation
+  std::unordered_map<uint32_t, std::shared_ptr<AHBPool>> _pools; // gen -> pool
+  std::shared_ptr<AHBPool> _renderPool; // generation of the in-flight render
+  int _currentRenderSlot = -1;
+  std::shared_ptr<AHBPool> _readyPool; // generation of the latest ready frame
+  int _readySlot = -1;
+  std::deque<PendingPresent> _presentQueue;
+  std::thread _waiter;
+  bool _waiterRunning = false;
+  std::atomic<bool> _shutdown{false};
+  uint32_t _genCounter = 0;
+  std::function<void()> _poolFrameReady; // consumer wake-up, see setter
+#endif
 };
 
 class SurfaceRegistry {