A5 target hangs on RMSNorm vector kernel that passes on A2A3

[zhangqi-chen]

Problem

A PTO-IR vector kernel (decode_projection_incore_0.pto) compiles and runs correctly when targeting A2A3, but hangs at runtime when compiled for A5.

Environment

• PTOAS version: 0.22

Reproduction

PTO file attached below.

decode_projection_incore_0.pto.txt
rmsnorm_incore_0.pto.txt

Steps:

1. Change pto.target_arch from "a2a3" to "a5" in the module attributes
2. Compile with ptoas
3. Run on A5 platform — program hangs indefinitely (no crash, no error)

Behavior

Target	Compile	Run
A2A3	OK	OK
A5	OK	Hangs

Kernel Summary

This is a RMSNorm vector kernel (decode_projection_incore_0) from the Qwen3-32B decode layer projection. It operates on [16, 5120] BF16 input with K_CHUNK=128 (40 iterations):

1. Loop 1 — accumulate squared partial sums: tload → tcvt(bf16→f32) → tmul(x²) → trowsum → tadd (accumulate) → tmov
2. Post-loop — compute inv_rms: tmuls(÷5120) → tadds(+ε) → trsqrt
3. Loop 2 — apply normalization: tload → tcvt → trowexpandmul(×inv_rms) → tcolexpandmul(×γ) → tcvt(f32→bf16) → tstore

Operations Used

tload, tstore, tcvt, tmul, trowsum, tadd, tmov, tmuls, tadds, trsqrt, trowexpandmul, tcolexpandmul, texpands

PTO File

Context

Discovered during E2E validation of pypto-lib Qwen3-32B decode tilelet (hw-native-sys/pypto-lib#58, Scope 1).

[zhangstevenunity]

本issue也跟pto-isa的tmov在16x1的col major的tile上的实现存在问题相关。

[TaoTao-real]

Cross-repo tracking: related ISA-side investigation is now tracked at hw-native-sys/pto-isa#58
Link: hw-native-sys/pto-isa#58
This issue focuses on potential A5 TMOV col_major 16x1 vec->vec alignment risk (error 340).