feat(vmi): VMIPredicateFold — DCE compile-time expert-pad vcmp/vsel

[peanutchan]

Summary

• Adds VMIPredicateFold (-vmi-predicate-fold) to constant-prove VMI index ranges and fold identity expert-pad vcmp_lt+vsel before unified→legacy lowering.
• Extracts shared VMIMaskUtils (isAllActiveSeed / isAllInactiveSeed) used by this pass and VMILowerUnifiedToLegacy.
• Enables frontends (e.g. TileLang topk_gate) to always emit pad masking and drop hand-coded need_pad when num_experts is compile-time known.

Closes #580.

Motivation

Padding after VL-aligned vci / vadds is a common MoE pattern:

score = vsel(vcmp(index, num_experts, lt), score, -inf)

When the compare is provably all-true, this is dead work. Hand-gating with need_pad = (E != aligned_E) misses non-tail chunks and duplicates policy in every kernel.

Design

Item	Choice
Pipeline slot	VMI semantic pipeline, before VMILowerUnifiedToLegacy
Ranges	vci, vadds(vci,C), vbrc(C), affine iv*stride+C + known scf.for bounds
Rewrites	all-true/false vcmp → lattice; identity / constant vsel; pure-op DCE
A/B	--disable-vmi-predicate-fold (PTO_FLAGS)

Lit: test/lit/vmi_new/vmi_predicate_fold_pad.pto.

Performance (camodel 950PR, N=4, pad always-on, fold OFF vs ON)

Config	g	RVECEX Δ	elem/cycle Δ
E16 K2 (real pad, E≪VL)	1/2	0	0%
E64 K8	1	73→70	+1.0%
E64 K8	2	138→134	+1.2%
E384 K9	1	366→353 (−13)	+1.0%
E384 K9	2	714→695 (−19)	+1.2%

Identity-pad cases match the old hand-need_pad=False rooftop after fold; E≪VL correctly keeps pad.

Test plan

• pto-test-opt test/lit/vmi_new/vmi_predicate_fold_pad.pto -vmi-predicate-fold | FileCheck
• topk_gate camodel A/B with/without PTO_FLAGS=--disable-vmi-predicate-fold for E=16/64/384
• Confirm dynamic num_experts does not fold (lit case)

Made with Cursor

[peanutchan]

Superseded for upstream by hw-native-sys#1151 against main (issue hw-native-sys#1150). This feature-vmi PR can stay for the fork branch if still useful.

[peanutchan]

Follow-up generalization (R1–R6 + R4 neutral reduce skip + camodel report) landed on main-based branch:

• PR: feat(vmi): generalize VMIPredicateFold (R1–R6) + neutral reduce skip hw-native-sys/PTOAS#1194
• Branch: peanutchan:feat/vmi-predicate-fold-main
• Report: docs/vmi/vmi_predicate_fold_generalization_report.md

This feature-vmi PR (#581) remains the original pad-only fold; prefer hw-native-sys#1194 for the full optimizer.