Restore using VMA to create buffers and images

[clayjohn]

Fixes: #101850

VMA handles memory allocation on certain devices better than our custom VK code, so we might as well use it

This PR uses VMA when we don't need the extra memory tracking. When not doing the extra memory tracking, there is no reason to use the raw Vulkan code.

I am tempted to find what VMA does differently and just copy it into our code to avoid the branch. However, it is likely best to always fall back on VMA since it is well-maintained and we can benefit from further updates to it in the future.

[darksylinc]

Something smells:

The VMA branch always sets alloc_create_info.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;; which is not the original behavior. I believe this would restore the original 4.3 behavior:

VmaMemoryUsage vma_usage = VMA_MEMORY_USAGE_UNKNOWN;
uint32_t vma_flags_to_remove = 0;

VmaAllocationCreateInfo alloc_create_info = {};
switch (p_allocation_type) {
	case MEMORY_ALLOCATION_TYPE_CPU: {
		bool is_src = p_usage.has_flag(BUFFER_USAGE_TRANSFER_FROM_BIT);
		bool is_dst = p_usage.has_flag(BUFFER_USAGE_TRANSFER_TO_BIT);
		if (is_src && !is_dst) {
			// Looks like a staging buffer: CPU maps, writes sequentially, then GPU copies to VRAM.
			alloc_create_info.flags = VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT;
			alloc_create_info.preferredFlags |= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
			vma_flags_to_remove |= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
		}
		if (is_dst && !is_src) {
			// Looks like a readback buffer: GPU copies from VRAM, then CPU maps and reads.
			alloc_create_info.flags = VMA_ALLOCATION_CREATE_HOST_ACCESS_RANDOM_BIT;
			alloc_create_info.preferredFlags |= VK_MEMORY_PROPERTY_HOST_CACHED_BIT;
			vma_flags_to_remove |= VK_MEMORY_PROPERTY_HOST_CACHED_BIT;
		}
		vma_usage = VMA_MEMORY_USAGE_AUTO_PREFER_HOST;
		alloc_create_info.requiredFlags = (VK_MEMORY_PROPERTY_HOST_COHERENT_BIT | VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT);
	} break;
	case MEMORY_ALLOCATION_TYPE_GPU: {
		vma_usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
		alloc_create_info.preferredFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
		if (p_size <= SMALL_ALLOCATION_MAX_SIZE) {
			uint32_t mem_type_index = 0;
			vmaFindMemoryTypeIndexForBufferInfo(allocator, &create_info, &alloc_create_info, &mem_type_index);
			alloc_create_info.pool = _find_or_create_small_allocs_pool(mem_type_index);
		}
	} break;
}

VkBuffer vk_buffer = VK_NULL_HANDLE;
VmaAllocation allocation = nullptr;
VmaAllocationInfo alloc_info = {};

if (!Engine::get_singleton()->is_extra_gpu_memory_tracking_enabled()) {
	alloc_create_info.preferredFlags &= ~vma_flags_to_remove;
	alloc_create_info.usage = vma_usage;
	VkResult err = vmaCreateBuffer(allocator, &create_info, &alloc_create_info, &vk_buffer, &allocation, &alloc_info);
	ERR_FAIL_COND_V_MSG(err, BufferID(), "Can't create buffer of size: " + itos(p_size) + ", error " + itos(err) + ".");
} else {
	VkResult err = vkCreateBuffer(vk_device, &create_info, VKC::get_allocation_callbacks(VK_OBJECT_TYPE_BUFFER), &vk_buffer);
	ERR_FAIL_COND_V_MSG(err, BufferID(), "Can't create buffer of size: " + itos(p_size) + ", error " + itos(err) + ".");
	err = vmaAllocateMemoryForBuffer(allocator, vk_buffer, &alloc_create_info, &allocation, &alloc_info);
	ERR_FAIL_COND_V_MSG(err, BufferID(), "Can't allocate memory for buffer of size: " + itos(p_size) + ", error " + itos(err) + ".");
	err = vmaBindBufferMemory2(allocator, allocation, 0, vk_buffer, nullptr);
	ERR_FAIL_COND_V_MSG(err, BufferID(), "Can't bind memory to buffer of size: " + itos(p_size) + ", error " + itos(err) + ".");
}

Although something similar happens with vmaCreateImage, that shouldn't be a problem because we want all textures to be device local.

[clayjohn]

Something smells:

The VMA branch always sets alloc_create_info.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;; which is not the original behavior. I believe this would restore the original 4.3 behavior:

Although something similar happens with vmaCreateImage, that shouldn't be a problem because we want all textures to be device local.

Good catch. Previously we used VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE unconditionally with vmaCreateImage(), so I kept the logic there. Then I forgot to check what we used to do when I copied the changes over for vmaCreateBuffer().

I have applied your change and confirmed that it continues to work fine. Taking a look at #90993 I can confirm that your change restores the old behaviour too.

[akien-mga]

I'll merge pro-actively for the 4.4.beta4 snapshot, but a second review from @darksylinc or @DarioSamo would still be good.

[akien-mga]

Thanks!