> From: Haochen Jiang <[email protected]> > Sent: Thursday, July 2, 2026 2:49 PM > > Hi all, > > As mentioned in previous ACEv1 thread, we implemented it with a > different way. We used internal pattern instead of inline assmebly > by treating tmm as a fake register. Users still need to manage the > register allocation for now by passing register number, but this > will help compiler know the dependency between insts and maybe > more convenient for potential future tmm register allocation since > all the patterns are there for reference. Since in ACE, tmms are > accumulator unit rather than calculation unit, this is acceptable. > > Also since ACE and legacy AMX should not be used together, we use > different intrin name for those shared insts. See patch detail > description for how we consider those names. >
One thing forgot to mention in the cover letter. The patch series depends on AVX10_V2_AUX patches and must be committed after AVX10_V2_AUX due to the CPUID emulation part. Thus, the patch has our avx10v2aux implementation base. We will rebase the patch after AVX10_V2_AUX patches landing on trunk. Thx, Haochen > Bootstrapped and regtested on x86_64-pc-linux-gnu. Discussions are > welcomed on this patch series.
