Batch up scheduling of deferred deletions (#3030)

src/mbgl/renderer/tile_pyramid.cpp

@@ -82,6 +82,7 @@ void TilePyramid::update(const std::vector<Immutable<style::LayerProperties>>& l
 
         tiles.clear();
         renderedTiles.clear();
+        cache.deferPendingReleases();
 
         return;
     }
@@ -279,6 +280,8 @@ void TilePyramid::update(const std::vector<Immutable<style::LayerProperties>>& l
             tile.usedByRenderedLayers |= tile.layerPropertiesUpdated(layerProperties);
         }
     }
+
+    cache.deferPendingReleases();
 }
 
 void TilePyramid::handleWrapJump(float lng) {

src/mbgl/tile/tile_cache.cpp

@@ -1,10 +1,22 @@
 #include <mbgl/tile/tile_cache.hpp>
+
 #include <mbgl/actor/scheduler.hpp>
 #include <mbgl/util/instrumentation.hpp>
+
 #include <cassert>
 
 namespace mbgl {
 
+TileCache::~TileCache() {
+    MLN_TRACE_FUNC();
+
+    clear();
+    pendingReleases.clear();
+
+    std::unique_lock counterLock{deferredSignalLock};
+    deferredSignal.wait(counterLock, [&]() { return deferredDeletionsPending == 0; });
+}
+
 void TileCache::setSize(size_t size_) {
     MLN_TRACE_FUNC();
 
@@ -26,16 +38,20 @@ void TileCache::setSize(size_t size_) {
 }
 
 namespace {
-
 /// This exists solely to prevent a problem where temporary lambda captures
 /// are retained for the duration of the scope instead of being destroyed immediately.
-template <typename T>
 struct CaptureWrapper {
-    CaptureWrapper(std::unique_ptr<T>&& item_)
-        : item(std::move(item_)) {}
+    CaptureWrapper(std::vector<std::unique_ptr<Tile>>&& items_)
+        : items(items_.size()) {
+        std::ranges::move(items_, items.begin());
+    }
+    CaptureWrapper(CaptureWrapper&&) = default;
+
+    /// This copy constructor is required to build, but doesn't seem to be called.
     CaptureWrapper(const CaptureWrapper& other)
-        : item(other.item) {}
-    std::shared_ptr<T> item;
+        : items(other.items) {}
+
+    std::vector<std::shared_ptr<Tile>> items;
 };
 } // namespace
 
@@ -43,6 +59,25 @@ void TileCache::deferredRelease(std::unique_ptr<Tile>&& tile) {
     MLN_TRACE_FUNC();
 
     tile->cancel();
+    pendingReleases.push_back(std::move(tile));
+}
+
+void TileCache::deferPendingReleases() {
+    MLN_TRACE_FUNC();
+
+    constexpr std::size_t scheduleThreshold = 1;
+    if (pendingReleases.size() < scheduleThreshold) {
+        return;
+    }
+
+    // Block destruction until the cleanup task is complete
+    {
+        std::lock_guard counterLock{deferredSignalLock};
+        deferredDeletionsPending++;
+    }
+
+    CaptureWrapper wrap{std::move(pendingReleases)};
+    pendingReleases.clear();
 
     // The `std::function` must be created in a separate statement from the `schedule` call.
     // Creating a `std::function` from a lambda involves a copy, which is why we must use
@@ -51,17 +86,16 @@ void TileCache::deferredRelease(std::unique_ptr<Tile>&& tile) {
     // If this temporary outlives the `schedule` call, and the function is executed immediately
     // by a waiting thread and is already complete, that temporary reference ends up being the
     // last one and the destruction actually occurs here on this thread.
-    std::function<void()> func{[tile_{CaptureWrapper<Tile>{std::move(tile)}}, this]() mutable {
-        tile_.item = {};
+    std::function<void()> func{[tile_{CaptureWrapper{std::move(wrap)}}, this]() mutable {
+        MLN_TRACE_ZONE(deferPendingReleases lambda);
+        MLN_ZONE_VALUE(wrap_.releases.size());
+        tile_.items.clear();
 
         std::lock_guard<std::mutex> counterLock(deferredSignalLock);
         deferredDeletionsPending--;
         deferredSignal.notify_all();
     }};
 
-    std::unique_lock<std::mutex> counterLock(deferredSignalLock);
-    deferredDeletionsPending++;
-
     threadPool.schedule(std::move(func));
 }
 

src/mbgl/tile/tile_cache.hpp

@@ -18,15 +18,7 @@ public:
     TileCache(const TaggedScheduler& threadPool_, size_t size_ = 0)
         : threadPool(threadPool_),
           size(size_) {}
-
-    ~TileCache() {
-        clear();
-
-        std::unique_lock<std::mutex> counterLock(deferredSignalLock);
-        while (deferredDeletionsPending != 0) {
-            deferredSignal.wait(counterLock);
-        }
-    }
+    ~TileCache();
 
     /// Change the maximum size of the cache.
     void setSize(size_t);
@@ -43,13 +35,17 @@ public:
     bool has(const OverscaledTileID& key);
     void clear();
 
-    /// Destroy a tile without blocking
+    /// Set aside a tile to be destroyed later, without blocking
     void deferredRelease(std::unique_ptr<Tile>&&);
 
+    /// Schedule any accumulated deferred tiles to be destroyed
+    void deferPendingReleases();
+
 private:
     std::map<OverscaledTileID, std::unique_ptr<Tile>> tiles;
     std::list<OverscaledTileID> orderedKeys;
     TaggedScheduler threadPool;
+    std::vector<std::unique_ptr<Tile>> pendingReleases;
     size_t deferredDeletionsPending{0};
     std::mutex deferredSignalLock;
     std::condition_variable deferredSignal;

src/mbgl/vulkan/context.cpp

@@ -596,7 +596,7 @@ void Context::buildUniformDescriptorSetLayout(vk::UniqueDescriptorSetLayout& lay
 
     for (size_t i = 0; i < uniformCount; ++i) {
         bindings.push_back(vk::DescriptorSetLayoutBinding()
-                               .setBinding(i)
+                               .setBinding(static_cast<uint32_t>(i))
                                .setStageFlags(stageFlags)
                                .setDescriptorType(vk::DescriptorType::eUniformBuffer)
                                .setDescriptorCount(1));

src/mbgl/vulkan/descriptor_set.cpp

@@ -50,7 +50,7 @@ void DescriptorSet::createDescriptorPool(DescriptorPoolGrowable& growablePool) {
     const auto descriptorPoolInfo = vk::DescriptorPoolCreateInfo(poolFlags).setPoolSizes(size).setMaxSets(maxSets);
 
     growablePool.pools.emplace_back(device->createDescriptorPoolUnique(descriptorPoolInfo), maxSets);
-    growablePool.currentPoolIndex = growablePool.pools.size() - 1;
+    growablePool.currentPoolIndex = static_cast<int32_t>(growablePool.pools.size() - 1);
 };
 
 void DescriptorSet::allocate() {
@@ -73,7 +73,8 @@ void DescriptorSet::allocate() {
             growablePool.pools.begin(), growablePool.pools.end(), [&](const auto& p) { return !p.unusedSets.empty(); });
 
         if (unusedPoolIt != growablePool.pools.end()) {
-            growablePool.currentPoolIndex = std::distance(growablePool.pools.begin(), unusedPoolIt);
+            growablePool.currentPoolIndex = static_cast<int32_t>(
+                std::distance(growablePool.pools.begin(), unusedPoolIt));
         } else
 #endif
         {
@@ -83,7 +84,8 @@ void DescriptorSet::allocate() {
                                                   [&](const auto& p) { return p.remainingSets >= layouts.size(); });
 
             if (freePoolIt != growablePool.pools.end()) {
-                growablePool.currentPoolIndex = std::distance(growablePool.pools.begin(), freePoolIt);
+                growablePool.currentPoolIndex = static_cast<int32_t>(
+                    std::distance(growablePool.pools.begin(), freePoolIt));
             } else {
                 createDescriptorPool(growablePool);
             }
@@ -161,7 +163,7 @@ void UniformDescriptorSet::update(const gfx::UniformBufferArray& uniforms,
                                             .setBufferInfo(descriptorBufferInfo)
                                             .setDescriptorCount(1)
                                             .setDescriptorType(vk::DescriptorType::eUniformBuffer)
-                                            .setDstBinding(index)
+                                            .setDstBinding(static_cast<uint32_t>(index))
                                             .setDstSet(descriptorSets[frameIndex]);
 
         device->updateDescriptorSets(writeDescriptorSet, nullptr);
@@ -197,7 +199,7 @@ void ImageDescriptorSet::update(const std::array<gfx::Texture2DPtr, shaders::max
                                             .setImageInfo(descriptorImageInfo)
                                             .setDescriptorCount(1)
                                             .setDescriptorType(vk::DescriptorType::eCombinedImageSampler)
-                                            .setDstBinding(id)
+                                            .setDstBinding(static_cast<uint32_t>(id))
                                             .setDstSet(descriptorSets[frameIndex]);
 
         device->updateDescriptorSets(writeDescriptorSet, nullptr);