Skip to content

Working with memory in Metal part 2

There are a couple of topics we need to discuss in more depth about working with memory. Last time we have seen that to create MTLBuffer objects we have 3 options: by creating a new memory allocation with new data, by copying data from an existing allocation into a new allocation and by reusing an existing storage allocation which does not copy data. Since we haven’t looked at the memory before, let’s check that is actually true. First we copy data into another allocation:

let count = 2000
let length = count * MemoryLayout< Float >.stride
var myVector = [Float](repeating: 0, count: count)
let myBuffer = device.makeBuffer(bytes: myVector, length: length, options: [])
withUnsafePointer(to: &myVector) { print($0) }

Note: the withUnsafePointer() function gives us the memory address of the actual data on the heap instead of the address of the pointer (from the stack) that wraps that data.

Your output should look similar to this:


The two data buffers are definitely stored at different memory locations. Now let’s use the no-copyoption:

var memory: UnsafeMutableRawPointer? = nil
let alignment = 0x1000
let allocationSize = (length + alignment - 1) & (~(alignment - 1))
posix_memalign(&memory, alignment, allocationSize)
let myBuffer = device.makeBuffer(bytesNoCopy: memory!, 
				 length: allocationSize, 
				 options: [], 
				 deallocator: { (pointer: UnsafeMutableRawPointer, _: Int) in 

First, we create a pointer to our data which (data) we’ll store on the heap. For this we need to page-align it. We set the page size to be 4K (1000 in hexadecimal). We need to also round the buffer size to match the alignment. We used a bitwise AND to avoid division which is a very expensive operation. Otherwise we would just round like this:

let allocationSize = ((length + alignment - 1) / alignment) * alignment

Your output should look similar to this:


Notice the last three digits in the addresses above? Those come from page-alinging the data because an address is determined by 0 mod pageSize, hence the last three 0’s, which makes sense since our page size is 0x1000.

Let’s now move to Storage Modes which we briefly mentioned last time. There are basically only four main rules to keep in mind, one for each of the storage modes:

Shareddefault on macOS buffers, iOS/tvOS resources; not available on macOS textures.
Privatemostly use when data is only accessed by GPU.
Memorylessonly for iOS/tvOS on-chip temporary render targets (textures).
Manageddefault mode for macOS textures; not available on iOS/tvOS resources.

For a better big picture, here is the full cheat sheet in case you might find it easier to use than remembering the rules above:

alt text

The most complicated case is when working with macOS buffers and when the data needs to be accessed by both the CPU and the GPU. We choose the storage mode based on whether one or more of the following conditions are true:

  • Private – for large-sized data that changes at most once, so it is not “dirty” at all. create a source buffer with a Shared mode and then blit its data into a destination buffer with a Private mode. resource coherency is not necessary in this case as the data is only accessed by the GPU. this operation is the least expensive (a one-time cost).
  • Managed – for medium-sized data that changes infrequently (every few frames), so it is partially “dirty”. one copy of the data is stored in system memory for the CPU and another copy is stored in GPU memory. resource coherency is explicitly managed by synchronizing the two copies.
  • Shared – for small-sized data that is updated every frame, so it is fully dirty. data resides in the system memory and is visible and modifyable by both the CPU and the GPU. resource coherency is only guaranteed within command buffer boundaries.

How to make sure coherency is guaranteed? First, make sure that all the modifications done by the CPU finished before the command buffer was committed (check if the command buffer status property is MTLCommandBufferStatusCommitted). After the GPU finishes executing the command buffer, the CPU should only start making modifications again only after the GPU is signaling the CPU that the command buffer finished executing (check if the command buffer status property is MTLCommandBufferStatusCompleted).

Finally, let’s see how synchronization is done for macOS resources. For buffers: after a CPU write use didModifyRange() to inform the GPU of the changes so Metal can update that data region only; after a GPU write use synchronize(resource:) within a blit operation, to refresh the caches so the CPU can access the updated data. For textures: after a CPU write use one of the two replace() region functions to inform the GPU of the changes so Metal can update that data region only; after a GPUwrite use one of the two synchronize() functions within a blit operation to allow Metal to update the system memory copy after the GPU finished modifying the data.

The source code is posted on Github as usual.

Until next time!

Leave a Reply

Your email address will not be published. Required fields are marked *