This repo shows how to call Metal kernels from Racket.
It's a learning playground, you won't find here any high-level stuff.

I written 4 simple kernels:

• add_arrays
• mult_arrays
• add_matrices
• matmul

There are no optimizations, really, just simple functions.

I tried to extract the minimum that is required to run kernel computations from Racket.
The most important pieces are:

Metal requires the Buffers to be associated with device and library.
Below structure keeps the reference, it's used with pretty much every command.

_metal-config
A struct that holds device and library.
Almost all other functions take metal-config as an argument.

_metal_vector A struct that contains Metal Buffer pointer as well as other informations about the metal vector.
See source code and take a look at metal-ffi.rkt (or MetalAdder.m) to see how it's defined.

list->mvector | mvector->list
Transition lists similar to those you might know from racket (i.e. list->vector). But they additionally take metal-config as first argument and they require the list to be a flat list.
It's the first building block for the communication with Metal, it allocates Metal Buffer with shape of the list.

compute-add | compute-mul Vector operations. They take 2 metal_vector's and return a new one.
You can see in examples.rkt how to use them.

_metal_matrix A struct that contains Metal Buffer pointer as well as other informations about the metal matrix.
See source code and take a look at metal-ffi.rkt (or MetalAdder.m) to see how it's defined.
It's pretty much the same as _metal_vector, but it also contains information about number of rows and columns.

list->mmatix | mmatrix->list
Just like list->mvector and mvector->list, but with matrix in mind.
It requires the list to be 2D of course.

compute-mat-add | compute-mat-mul Matrix operations. They take 2 metal_matrix's and return a new one.
You can see in examples.rkt how to use them.

racket examples.rkt

Build dylib and metallib

xcodebuild -project MetalComputeBasic.xcodeproj -scheme MetalComputeBasic -configuration Debug

It will build the two files under Build/Products/Debug/:

• default.metallib
• libMetalComputeBasic (dylib)

Those are the main ffi bindings to my functions. Together with dylib and metallib mentioned above, they are all that is needed to create simple vectors/matrices and make really basic operations.

racket ./metal-ffi.rkt

raco test test.rkt

Those are not very rigorous. The kernels are not optimized, but just to see how does it compare to CPU-optimized operations, here it is. To run execute that:

racket -l racket/base -e '(require (submod "benchmark.rkt" benchmark))'

Some example results for [1e5, 1e3]x[1e3, 1]:

  Flomat: 195.611083984375
  Vector: 181.384033203125
  Metal:   30.556884765625

  Flomat: 179.22509765625
  Vector: 167.153076171875
  Metal:   52.55712890625

  Flomat: 173.863037109375
  Vector: 168.3330078125
  Metal:   27.2109375

Some example results for [1e3, 1e5]x[1e5, 1]:

  Flomat: 322.506103515625
  Vector: 331.7861328125
  Metal:   40.25390625

  Flomat: 326.198974609375
  Vector: 329.2919921875
  Metal:   44.015869140625

  Flomat: 297.712890625
  Vector: 245.10302734375
  Metal:   35.764892578125

One might expect that to be orders of magnitude better.
They hugely depend on what are the exact operations and dimensions.
I suspect it's because my kernels aren't optimized at all. I just use a really simple one for multiplication.
I'm gonna have to read more about the specifics of the metal language at some point, but it's really to get a grasp of what it's all about.