Detect DEVICE_LOCAL | HOST_VISIBLE machines and apply VMA_ALLOCATION_CREATE_HOST_ACCESS_* automatically

[spencer-lunarg]

Something I have found is some drivers (like Lavapipe or Intel's new Metor Lake chip) will have all their memory types as DEVICE_LOCAL | HOST_VISIBLE

What this leads to is going

VmaAllocationCreateInfo alloc_ci = {};
alloc_ci.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
alloc_ci.flags = 0;

and then later our code will go

VkMemoryPropertyFlags mem_prop_flags = {};
vmaGetAllocationMemoryProperties(vma_allocator, allocation, &mem_prop_flags);
if (mem_prop_flags & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT) {
    vmaMapMemory(gpuav.vma_allocator_, allocation, &mapped_ptr);

and hit a

vk_mem_alloc.h:11057: void VmaAllocation_T::BlockAllocMap(): Assertion `IsMappingAllowed() && "Mapping is not allowed on this allocation! Please use one of the new VMA_ALLOCATION_CREATE_HOST_ACCESS_* flags when creating it."' failed.

---

Our work around is detecting if there is no dedicated device local machine and then setting

alloc_ci.flags = VMA_ALLOCATION_CREATE_HOST_ACCESS_RANDOM_BIT;

---

Raising this issue as not sure if this is a feature request to have VMA detect everything is mappable and set FLAG_MAPPING_ALLOWED for us

... or if this is us not using VMA correctly and there is a better pattern to this

[sawickiap]

Yeah, I know that some devices have all their memory HOST_VISIBLE. At the same time, I've heard that mapping every memory block may not be good, especially when capturing tools are used like RenderDoc or PIX. That's why VMA defines explicit VMA_ALLOCATION_CREATE_HOST_ACCESS_* flags to be used for those allocations that are going to be mapped.

If you look at the library source code, you will see in VmaBlockVector::AllocatePage() that the algorithm tries to keep mappable allocations and non-mappable allocations in separate block, in order to minimize the number of mapped blocks. Your proposal goes against this logic.

You can try using flags VMA_ALLOCATION_CREATE_HOST_ACCESS_* | VMA_ALLOCATION_CREATE_HOST_ACCESS_ALLOW_TRANSFER_INSTEAD_BIT to express the intent that you would like to map the allocation, but you agree to create it in a non-HOST_VISIBLE memory if that is more optimal on a current GPU, so you will then perform manual transfers from/to it.

Please note that flags like VMA_MEMORY_USAGE_AUTO* and VMA_ALLOCATION_CREATE_HOST_ACCESS_* express a high-level intent about how you want to use the allocation, leaving the decision to VMA. What you describe looks like you are manually inspecting what memory types are available and you want to make your decision based on that. In this case, maybe you should select your memory type manually and specify HOST_ACCESS flag when you found your memory type to be HOST_VISIBLE.

VmaAllocationCreateInfo allocInfo = {};
allocInfo.memoryTypeBits = 1U << memoryTypeIndex;