Acceleration of DFT with Apple Silicon

Hello,

I would like to know if there is an efficient way to use the Apple Silicon GPU's to calculate a DFT from Matlab and if anybody has compared the performance against the Matlab fft function.

Thank you very much!

Answer 1

[Walter Roberson]

You would need to create mex code that used the Apple Metal2 interface to run the calculation on the gpu.

MATLAB does not provide any support for using the Apple Silicon GPU, and is unlikely to add that support for several years.

[Giorgio Taricco]

Dear Walter,

thank you for your kind reply. I understand that MATLAB is probably not going to support Apple Metal2 but if I make a C program (I am not sure it is possible) following the mex I/O protocol that is compiled by xcode and interfaces through Metal to the Apple GPUs, MATLAB should not even realize (maybe I am wrong) what happens inside the C program as long as there is an entry point.

What I mean is that the compiler is not by MATLAB, MATLAB only supplies header files and libraries to translate the xcode generated code into a mex file.

Again, I may be completely wrong, but I don't understand why MATLAB should prevent this executable to interface with the GPUs.

Thank you again for your reply.

[Walter Roberson]

I have not heard of any compiling conflicts if you were to put in the work yourself to invoke Metal2 -- but I have not been involved in the beta. What I am saying is that Mathworks has not put in any effort to make it easier to use those interfaces.

[Giorgio Taricco]

I understand your point but the questions remain: 1) is it possible to make a mex-compatible C program interfacing with Apple Metal2? 2) any such fft implementation would be faster than native matlab fft?

Besides, I noticed that the fft benchmark with the prerelease of R2023B (which is Apple Silicon native) is much faster than with R2023A. On a Macbook Pro with the M1 Max it takes 0.25 seconds with R2023A and 0.12 seconds on the prerelease of R2023B. I wonder if this can be still improved by using the gpus.

[Walter Roberson]

I have not seen any reason that it would not work, provided that you are using one of the native betas.

I would not count on the metal2 implementation being faster. It might be, perhaps, but there can be a number of tricks for optimal performance, and from what I have read, Apple will not talk about such matters.

[Joss Knight]

I can confirm that we've proved this is possible, but that's about as far as I can go.

[Giorgio Taricco]

Thank you for your reply. Can you tell what is the advantage with respect to the benchmark?