Fix pool DeviceMemory allocations and rek

include/vsg/vk/MemoryBufferPools.h

@@ -58,6 +58,11 @@ namespace vsg
         VkResult reserve(const BufferInfoList& bufferInfoList, VkDeviceSize alignment, VkBufferUsageFlags bufferUsageFlags, VkSharingMode sharingMode, VkMemoryPropertyFlags memoryProperties);
         VkResult reserve(ResourceRequirements& requirements);
 
+        /// Release wholly unused Buffer and DeviceMemory pool entries, returning their memory to the driver.
+        /// Pools normally retain freed capacity for reuse, so call this on application memory pressure rather than routinely.
+        /// Returns the number of bytes of device memory released.
+        VkDeviceSize releaseUnusedPools();
+
         void report(LogOutput& out) const;
 
     protected:

src/vsg/state/Buffer.cpp

@@ -166,6 +166,11 @@ ref_ptr<Buffer> vsg::createBufferAndMemory(Device* device, VkDeviceSize size, Vk
         if (deviceMemoryBufferPools)
         {
             auto [memory, offset] = deviceMemoryBufferPools->reserveMemory(memRequirements, memoryProperties);
+            if (!memory)
+            {
+                warn("vsg::createBufferAndMemory(.., size = ", size, ", ..) failed to reserve DeviceMemory.");
+                return {};
+            }
 
             buffer->bind(memory, offset);
             return buffer;

src/vsg/vk/MemoryBufferPools.cpp

@@ -82,7 +82,7 @@ ref_ptr<BufferInfo> MemoryBufferPools::reserveBuffer(VkDeviceSize totalSize, VkD
         std::scoped_lock<std::mutex> lock(_mutex);
         for (auto& bufferFromPool : bufferPools)
         {
-            if (bufferFromPool->usage == bufferUsageFlags && bufferFromPool->size >= totalSize)
+            if (bufferFromPool->usage == bufferUsageFlags && bufferFromPool->size >= totalSize && bufferFromPool->maximumAvailableSpace() >= totalSize)
             {
                 MemorySlots::OptionalOffset reservedBufferSlot = bufferFromPool->reserve(totalSize, alignment);
                 if (reservedBufferSlot.first)
@@ -151,12 +151,20 @@ MemoryBufferPools::DeviceMemoryOffset MemoryBufferPools::reserveMemory(VkMemoryR
     ref_ptr<DeviceMemory> deviceMemory;
     MemorySlots::OptionalOffset reservedSlot(false, 0);
 
+    // Pools can hold linear buffers and optimal tiling images side by side, so every slot is aligned
+    // to bufferImageGranularity to satisfy the Vulkan granularity rule for adjacent resources.
+    const VkDeviceSize granularity = device->getPhysicalDevice()->getProperties().limits.bufferImageGranularity;
+    const VkDeviceSize alignment = std::max(memRequirements.alignment, granularity);
+
     for (auto& memoryPool : memoryPools)
     {
-        if (((memoryPool->getMemoryRequirements().memoryTypeBits & memRequirements.memoryTypeBits) == memRequirements.memoryTypeBits) &&
+        // memoryTypeBits and property flags equality guarantees the pool's chosen memory type index
+        // is also valid for the new resource.
+        if (memoryPool->getMemoryRequirements().memoryTypeBits == memRequirements.memoryTypeBits &&
+            memoryPool->getMemoryPropertyFlags() == memoryPropertiesFlags &&
             memoryPool->maximumAvailableSpace() >= totalSize)
         {
-            reservedSlot = memoryPool->reserve(totalSize, memRequirements.alignment);
+            reservedSlot = memoryPool->reserve(totalSize, alignment);
             if (reservedSlot.first)
             {
                 deviceMemory = memoryPool;
@@ -179,10 +187,15 @@ MemoryBufferPools::DeviceMemoryOffset MemoryBufferPools::reserveMemory(VkMemoryR
 
         if (deviceSize <= availableMemory)
         {
+            // Allocate a pool sized block so subsequent resources suballocate from it rather than
+            // each getting a dedicated vkAllocateMemory. Explicit dedicated allocations
+            // (pNextAllocInfo) must keep their exact size.
+            VkMemoryRequirements poolRequirements = memRequirements;
+            if (!pNextAllocInfo) poolRequirements.size = deviceSize;
 
             try
             {
-                deviceMemory = vsg::DeviceMemory::create(device, memRequirements, memoryPropertiesFlags, pNextAllocInfo);
+                deviceMemory = vsg::DeviceMemory::create(device, poolRequirements, memoryPropertiesFlags, pNextAllocInfo);
             }
             catch (...)
             {
@@ -191,8 +204,13 @@ MemoryBufferPools::DeviceMemoryOffset MemoryBufferPools::reserveMemory(VkMemoryR
 
             if (deviceMemory)
             {
-                reservedSlot = deviceMemory->reserve(totalSize);
-                // if (!deviceMemory->full())
+                reservedSlot = deviceMemory->reserve(totalSize, alignment);
+
+                // Only blocks inflated beyond the request have spare capacity worth sharing. An
+                // exact fit block is wholly consumed by its one resource, so pooling it would just
+                // keep the memory alive after the resource is destroyed; leaving it out means the
+                // resource's own reference returns it to the driver.
+                if (poolRequirements.size > totalSize)
                 {
                     //debug("  inserting DeviceMemory into memoryPool ", deviceMemory.get());
                     memoryPools.push_back(deviceMemory);
@@ -336,6 +354,42 @@ VkResult MemoryBufferPools::reserve(ResourceRequirements& requirements)
     }
 }
 
+VkDeviceSize MemoryBufferPools::releaseUnusedPools()
+{
+    std::scoped_lock<std::mutex> lock(_mutex);
+
+    // release empty Buffers first so their DeviceMemory slots are returned before the memory pools are checked
+    for (auto itr = bufferPools.begin(); itr != bufferPools.end();)
+    {
+        auto& buffer = *itr;
+        if (buffer->totalReservedSize() == 0 && buffer->referenceCount() == 1)
+        {
+            itr = bufferPools.erase(itr);
+        }
+        else
+        {
+            ++itr;
+        }
+    }
+
+    VkDeviceSize totalReleased = 0;
+    for (auto itr = memoryPools.begin(); itr != memoryPools.end();)
+    {
+        auto& memoryPool = *itr;
+        if (memoryPool->totalReservedSize() == 0 && memoryPool->referenceCount() == 1)
+        {
+            totalReleased += memoryPool->totalMemorySize();
+            itr = memoryPools.erase(itr);
+        }
+        else
+        {
+            ++itr;
+        }
+    }
+
+    return totalReleased;
+}
+
 void MemoryBufferPools::report(LogOutput& out) const
 {
     out.enter("MemoryBufferPools::report(..)");