Support for macOS/iOS with ARM64 CPU

[hjmallon]

This may also help with iOS?

Here is part of the sysctl -a from an Apple DTK. These can be queried with sysctlbyname.

hw.optional.floatingpoint: 1
hw.optional.watchpoint: 4
hw.optional.breakpoint: 6
hw.optional.neon: 1
hw.optional.neon_hpfp: 1
hw.optional.neon_fp16: 1
hw.optional.armv8_1_atomics: 1
hw.optional.armv8_crc32: 1
hw.optional.armv8_2_fhm: 0
hw.optional.amx_version: 0
hw.optional.ucnormal_mem: 0
hw.optional.arm64: 1

[sbehnke]

On my M1 MacBook Pro 13" I have the following:

hw.optional.floatingpoint: 1
hw.optional.watchpoint: 4
hw.optional.breakpoint: 6
hw.optional.neon: 1
hw.optional.neon_hpfp: 1
hw.optional.neon_fp16: 1
hw.optional.armv8_1_atomics: 1
hw.optional.armv8_crc32: 1
hw.optional.armv8_2_fhm: 1
hw.optional.armv8_2_sha512: 1
hw.optional.armv8_2_sha3: 1
hw.optional.amx_version: 2
hw.optional.ucnormal_mem: 1
hw.optional.arm64: 1

I've started adding support to query these on my machine and have it detecting a few of them now, but I'm having a hard time mapping between the Arm Features struct and the Apple optional names. Is there a description for what these mean on the ArmFeature struct? Also, for something like hw.optional.armv8_2_sha512 and hw.optional.armv8_2_sha3 that don't appear to have any support in cpu_feature, should we extend the struct to support them?

[sbehnke]

I figured out why it wasn't detecting as aarch64 and it appears to be matching a bit better now.

❯ ./list_cpu_features
arch            : aarch64
implementer     : 16777228 (0x100000C)
variant         :   2 (0x02)
part            : 458787763 (0x1B588BB3)
revision        :   2 (0x02)
flags           : asimdfhm,atomics,crc32,fp,fphp,sha3,sha512

It's getting there but I still need to fix the part that isn't coming from the flags.

[hjmallon]

There are also tables here https://github.com/pytorch/cpuinfo/blob/master/src/arm/mach/init.c

[sbehnke]

There are also tables here https://github.com/pytorch/cpuinfo/blob/master/src/arm/mach/init.c

I can certainly add a table entry for the M1 CPU, but do I just take the values from the docker instance of cpu_features? I'm not sure where to find a complete list of non-optional features implemented on the CPU.

[gchatelet]

To be clear, we are definitely interested in supporting Apple M1 but we are willing to wait a bit for the M1 build ecosystem to settle as it is unclear yet how to make it work with Universal2 builds.

[dkozel]

Is there any fix possible to resolve the compile issue? It'd be very preferable to report an empty list of CPU features than to fail to compile.