Metal programming in Julia

Build Status

With Metal.jl it's possible to program GPUs on macOS using the Metal programming framework.

The package is a work-in-progress. There are bugs, functionality is missing, and performance hasn't been optimized. Expect to have to make changes to this package if you want to use it. PRs are very welcome!

• Mac device with M-series chip
• Julia 1.8
• macOS 13 (Ventura)

These requirements are fairly strict, and are due to our limited development resources (manpower, hardware). Technically, they can be relaxed. If you are interested in contributing to this, see this issue for more details. In practice, Metal.jl will probably work on any macOS 10.15+, and other GPUs that are supported by Metal might also function (if only partially), but such combinations are unsupported for now.

Metal.jl can be installed with the Julia package manager. From the Julia REPL, type ] to enter the Pkg REPL mode and run:

pkg> add Metal

Or, equivalently, via the Pkg API:

julia> import Pkg; Pkg.add("Metal")

For an overview of the toolchain in use, you can run the following command after importing the package:

julia> using Metal

julia> Metal.versioninfo()
macOS 12.2.0, Darwin 21.3.0

Toolchain:
- Julia: 1.8.0-beta3
- LLVM: 13.0.1

1 device:
- Apple M1 Pro (64.000 KiB allocated)

The easiest way to work with Metal.jl, is by using its array abstraction. The MtlArray type is both meant to be a convenient container for device memory, as well as provide a data-parallel abstraction for using the GPU without writing your own kernels:

julia> a = MtlArray([1])
1-element MtlArray{Int64, 1}:
 1

julia> a .+ 1
1-element MtlArray{Int64, 1}:
 2

The above array abstractions are all implemented using Metal kernels written in Julia. These kernels follow a similar programming style to Julia's other GPU back-ends, and with that deviate from how kernels are implemented in Metal C (i.e., indexing intrinsics are functions not arguments, arbitrary aggregate arguments are supported, etc):

julia> function vadd(a, b, c)
           i = thread_position_in_grid_1d()
           c[i] = a[i] + b[i]
           return
       end
vadd (generic function with 1 method)

julia> a = MtlArray([1]); b = MtlArray([2]); c = similar(a);

julia> @metal threads=length(c) vadd(a, b, c)

julia> Array(c)
1-element Vector{Int64}:
 3

Finally, all of the above functionality is made possible by interfacing with the Metal libraries through a small C library that wraps the ObjectiveC APIs.

These low-level wrappers, along with some slightly higher-level Julia wrappers, are available in the MTL submodule exported by Metal.jl. All wrapped C functions and types start with the mt prefix, whereas the Julia wrappers are prefixed with Mtl:

julia> dev = MtlDevice(1)
MtlDevice:
 name:             Apple M1 Pro
 lowpower:         false
 headless:         false
 removable:        false
 unified memory:   true
 registry id:      4294969448
 transfer rate:    0

julia> dev.name
"Apple M1 Pro"

Metal.jl relies on two binary dependencies (provided as JLLs):

• cmt
• LLVM with an AIR back-end

Normally, these dependencies are built on Yggdrasil. If you need to make changes to these dependencies, have a look at the build_cmt.jl and build_llvm.jl scripts in the deps/ folder. These scripts build a local version of the dependency, and configure a local preference such that any environment depending on the corresponding JLLs will pick-up the modified version (i.e., do julia --project in a clone of Metal.jl):

$ julia --project -e 'using Metal; @show MTL.libcmt'
MTL.libcmt = "/Users/tim/Julia/depot/artifacts/6adc0ed9a8370ff1e3bb8fbaf36e8519ee11fd96/lib/libcmt.dylib"

$ julia --project=deps deps/build_cmt.jl
...
[100%] Built target cmt

$ julia --project -e 'using Metal; @show MTL.libcmt'
MTL.libcmt = "/Users/tim/Julia/depot/scratchspaces/dde4c033-4e86-420c-a63e-0dd931031962/cmt/lib/libcmt.dylib"

These scripts are integrated with our CI, and will be triggered if the ci.build_cmt or ci.build_llvm labels are set on a pull request.

The C library started by forking rcp/cmt, to whom goes the original credit. This package builds upon the experience of several Julia contributors to CUDA.jl, AMDGPU.jl and oneAPI.jl.