fix(a2a3): correct TCOLEXPAND pipe from PIPE_V to PIPE_MTE1

[georgebisbas]

Summary

TCOLEXPAND (plain broadcast) uses pto_copy_ubuf_to_ubuf (copy engine / PIPE_MTE1) in its a2a3 implementation, not a pure vector-unit op. A pipe_barrier(PIPE_V) after TCOLEXPAND does NOT wait for the copy-engine broadcast to complete, so a subsequent vector read can see stale data.

The race is data-size dependent on real hardware: C=32 fails (consumer wins the race), C=128 passes (copy is slow enough).

Root cause

This is a regression from commit 021789c0 ("refactor: extract Event CRTP base class and replace opPipeList with OpPipeEntry template", 2026-06-23), which tagged all expand-family ops as PIPE_V during the array-to-template conversion. Before that refactoring, TCOLEXPAND was not in the old opPipeList at all and defaulted to PIPE_ALL, so barriers were safe.

What was checked

• Only TCOLEXPAND uses the copy engine — every other TCOLEXPAND* / TROWEXPAND* variant uses vector instructions (vadd, vdiv, vmul, vector_dup, vbrcb, etc.) and remains correctly on PIPE_V.
• No other pipeline-critical structures reference expand ops' pipe assignments outside event.hpp.
• The old opPipeList (pre-refactoring) confirms TCOLEXPAND had no entry.

Test plan

• All existing TCOLEXPAND tests pass (they use explicit set_flag/wait_flag and are unaffected by the pipe declaration)
• MFE kernel included in tests/npu/a2a3/src/st/testcase/tcolexpand_pipe_bug_mfe/ for real-hardware regression testing (cannot reproduce on functional simulator)

Related

• Issue reported by @vloncar: pipe_barrier(PIPE_V) under-synchronization after TCOLEXPAND in GLA chunk_h_prep kernel produces wrong results at C=32 on a2a3 hardware
• Workaround in existing code: chunk_h_prep uses pipe_barrier(PIPE_ALL) instead

/cc @vloncar

[georgebisbas]

@Crystal-wzy please have a look wwhen you have available time, thanks

[Crystal-wzy]

@Crystal-wzy please have a look wwhen you have available time, thanks

Sorry for the late reply. The GitHub pto-isa repository no longer accepts merge requests. Please submit your pull request at https://gitcode.com/cann/pto-isa instead. Thank you!

[georgebisbas]

@Crystal-wzy please have a look wwhen you have available time, thanks

Sorry for the late reply. The GitHub pto-isa repository no longer accepts merge requests. Please submit your pull request at https://gitcode.com/cann/pto-isa instead. Thank you!

Hi we have submitted this already, please have a look when you have time, thanks @Crystal-wzy !
https://gitcode.com/cann/pto-isa/pull/1406

[georgebisbas]

@Crystal-wzy please have a look wwhen you have available time, thanks

Sorry for the late reply. The GitHub pto-isa repository no longer accepts merge requests. Please submit your pull request at https://gitcode.com/cann/pto-isa instead. Thank you!

Hi we have submitted this already, please have a look when you have time, thanks @Crystal-wzy !
https://gitcode.com/cann/pto-isa/pull/1406