If your game oversubscribes video memory, if may work OK in previous-generation graphics APIs (DirectX 9, 10, 11, OpenGL) because resources are automatically paged to system RAM. In Vulkan you can't do it because when you run out of memory, an allocation just fails. If you have more data (e.g. textures) that can fit into VRAM and you don't need it all at once, you may want to upload them to GPU on demand and "push out" ones that are not used for a long time to make room for the new ones, effectively using VRAM (or a cartain memory pool) as a form of cache. Vulkan Memory Allocator can help you with that by supporting a concept of "lost allocations".

To create an allocation that can become lost, include VMA_ALLOCATION_CREATE_CAN_BECOME_LOST_BIT flag in VmaAllocationCreateInfo::flags. Before using a buffer or image bound to such allocation in every new frame, you need to query it if it's not lost. To check it, call vmaTouchAllocation(). If the allocation is lost, you should not use it or buffer/image bound to it. You mustn't forget to destroy this allocation and this buffer/image. vmaGetAllocationInfo() can also be used for checking status of the allocation. Allocation is lost when returned VmaAllocationInfo::deviceMemory == VK_NULL_HANDLE.

To create an allocation that can make some other allocations lost to make room for it, use VMA_ALLOCATION_CREATE_CAN_MAKE_OTHER_LOST_BIT flag. You will usually use both flags VMA_ALLOCATION_CREATE_CAN_MAKE_OTHER_LOST_BIT and VMA_ALLOCATION_CREATE_CAN_BECOME_LOST_BIT at the same time.

Warning! Current implementation uses quite naive, brute force algorithm, which can make allocation calls that use VMA_ALLOCATION_CREATE_CAN_MAKE_OTHER_LOST_BIT flag quite slow. A new, more optimal algorithm and data structure to speed this up is planned for the future.

Q: When interleaving creation of new allocations with usage of existing ones, how do you make sure that an allocation won't become lost while it's used in the current frame?

It is ensured because vmaTouchAllocation() / vmaGetAllocationInfo() not only returns allocation status/parameters and checks whether it's not lost, but when it's not, it also atomically marks it as used in the current frame, which makes it impossible to become lost in that frame. It uses lockless algorithm, so it works fast and doesn't involve locking any internal mutex.

Q: What if my allocation may still be in use by the GPU when it's rendering a previous frame while I already submit new frame on the CPU?

You can make sure that allocations "touched" by vmaTouchAllocation() / vmaGetAllocationInfo() will not become lost for a number of additional frames back from the current one by specifying this number as VmaAllocatorCreateInfo::frameInUseCount (for default memory pool) and VmaPoolCreateInfo::frameInUseCount (for custom pool).

Q: How do you inform the library when new frame starts?

You need to call function vmaSetCurrentFrameIndex().

Example code:

struct MyBuffer
{
    VkBuffer m_Buf = nullptr;
    VmaAllocation m_Alloc = nullptr;
 
    // Called when the buffer is really needed in the current frame.
    void EnsureBuffer();
};
 
void MyBuffer::EnsureBuffer()
{
    // Buffer has been created.
    if(m_Buf != VK_NULL_HANDLE)
    {
        // Check if its allocation is not lost + mark it as used in current frame.
        if(vmaTouchAllocation(allocator, m_Alloc))
        {
            // It's all OK - safe to use m_Buf.
            return;
        }
    }
 
    // Buffer not yet exists or lost - destroy and recreate it.
 
    vmaDestroyBuffer(allocator, m_Buf, m_Alloc);
 
    VkBufferCreateInfo bufCreateInfo = { VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO };
    bufCreateInfo.size = 1024;
    bufCreateInfo.usage = VK_BUFFER_USAGE_UNIFORM_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT;
 
    VmaAllocationCreateInfo allocCreateInfo = {};
    allocCreateInfo.usage = VMA_MEMORY_USAGE_GPU_ONLY;
    allocCreateInfo.flags = VMA_ALLOCATION_CREATE_CAN_BECOME_LOST_BIT |
        VMA_ALLOCATION_CREATE_CAN_MAKE_OTHER_LOST_BIT;
 
    vmaCreateBuffer(allocator, &bufCreateInfo, &allocCreateInfo, &m_Buf, &m_Alloc, nullptr);
}

When using lost allocations, you may see some Vulkan validation layer warnings about overlapping regions of memory bound to different kinds of buffers and images. This is still valid as long as you implement proper handling of lost allocations (like in the example above) and don't use them.

You can create an allocation that is already in lost state from the beginning using function vmaCreateLostAllocation(). It may be useful if you need a "dummy" allocation that is not null.

You can call function vmaMakePoolAllocationsLost() to set all eligible allocations in a specified custom pool to lost state. Allocations that have been "touched" in current frame or VmaPoolCreateInfo::frameInUseCount frames back cannot become lost.

Q: Can I touch allocation that cannot become lost?

Yes, although it has no visible effect. Calls to vmaGetAllocationInfo() and vmaTouchAllocation() update last use frame index also for allocations that cannot become lost, but the only way to observe it is to dump internal allocator state using vmaBuildStatsString(). You can use this feature for debugging purposes to explicitly mark allocations that you use in current frame and then analyze JSON dump to see for how long each allocation stays unused.