Improved support of mcxcl and mmc on Apple silicon and Intel iGPUs

[Qianqian Fang]

hi everyone,

as I mentioned in a previous mailing list reply, we have recently tested and optimized the support on the integrated GPUs (iGPU) on Apple arm64 processors (M1/M2/M3/...) and now our nightly build provides native arm64 binaries for Apple machines at

https://mcx.space/nightly/macos64/

because both Intel-based and arm-based MATLAB can be installed on Apple M1/M2 machines, please type "mexext" command first in your matlab to decide what package to use, if you see `mexmaci64`, your matlab is Intel-based, and you should download the mcxlabcl/mmclab x86_64 packages; if it returns `mexmaca64`, you should download the arm64 packages instead.

in addition, I have also further tested and improved mcxcl support on Intel iGPUs. 

now mcxcl can pass all built-in tests (https://github.com/fangq/mcxcl/blob/master/test/testmcx.sh) on pretty much all processors my lab has - including 1) a wide range of CPUs running on Intel's OpenCL runtimes, old intel Apple runtimes,  or AMD/Intel CPUs on POCL runtimes, 2) 11th and 12th generation Intel integrated CPUs, 4) discrete Intel Arc GPU (A770), 5) discrete AMD GPUs ranging from RX480 to 7900GRE, and 6) discrete NVIDIA GPUs ranging from 980Ti to 4090/A100.

you can browse their speed comparisons at the below benchmark page (with the recent M2 data highlighted) - for serious compute, it is still highly recommended to use a discrete GPU due to its substantially higher power and speed

https://mcx.space/computebench/#:~:text=/-,Apple%20M2%20Pro,-/

if you notice any issue on your CPU, iGPU or discrete GPUs, please create a github issue and report it.

cheers

Qianqian