M2 Compatibility

[sfjohnson]

Hi,

On my M2 (2022 MacBook Air) I'm getting the following:

AMX_LDX: fail
AMX_LDY: fail
AMX_LDZ: pass
AMX_LDZI: pass
AMX_STX: pass
AMX_STY: pass
AMX_STZ: pass
AMX_STZI: pass
AMX_EXTRX: fail
AMX_EXTRY: fail
AMX_MAC16: pass
AMX_FMA16: pass
AMX_FMA32: pass
AMX_FMA64: pass
AMX_FMS16: pass
AMX_FMS32: pass
AMX_FMS64: pass
AMX_VECINT: fail
AMX_VECFP: fail
AMX_MATINT: pass
AMX_MATFP: fail
AMX_GENLUT: fail

This is with clang 14.0.0 on macOS 13.1. I could be doing something wrong here, or there might be a minor architectural difference between AMX on M1 and M2.

I'm going to investigate further to see if I can get everything to pass on M2, but first I was wondering if there has been any existing work done on M2 yet?

Thanks.

[corsix]

I've just rented an M2 machine in the cloud for a month, and I see the same thing. My money would be on "minor architectural difference"...

[sfjohnson]

Ok cool, I'll explore a bit and let you know if I find anything!

[corsix]

Some of the changes seem to be:

• extrh, extrv, vecfp, matfp, genlut gaining bf16 modes (the CPU cores also gain BF16 support in M2)
• extrh and extrv gaining some f32 -> f16 (or bf16) mixed lane-width modes
• ldx and ldy gaining "load four at a time" mode
• extrh, extrv, vecint, vecfp gaining support for operating on two or four vectors in a single instruction, rather than just a single vector
• vecint and vecfp gaining some new ALU modes (albeit not particularly interesting modes, e.g. z = x * y, z = z + x, z = z + y)

[corsix]

Commits pushed to reflect the changes in the above comment, in both the documentation and the emulation code.

[56789KD]

Add SIMD comment