PR #48 multi-dtype tests: NPU1 (Phoenix/AIE2) results — 15/20 pass, 5 failures

[erwei-xilinx]

Summary

Tested all examples from PR #48 (multi-dtype support) on NPU1 (Phoenix / AIE2) using transform_aie2.mlir. The PR was developed and validated on NPU2 (Strix/AIE2P). This issue documents NPU1 compatibility.

Environment:

• Device: NPU Phoenix (HP EliteBook 845 14 inch G10, Ryzen 7 PRO 7840U)
• XRT: 2.23.0
• NPU Firmware: 1.5.5.391
• mlir-air: 0.0.1.2026032022+f954272
• mlir-aie: 0.0.1.2026040605+e368f3e
• llvm-aie (Peano): 19.0.0.2025071101+b3cd09d3

Results: 15/20 PASS

Passing tests (15)

Example	bf16	f32	i16
vec-add	PASS	PASS	PASS
axpy	PASS	PASS	PASS
relu	PASS	PASS	PASS
leaky_relu	PASS	—	N/A
sigmoid	PASS	—	N/A
silu	PASS	PASS	N/A
swiglu	PASS	PASS	N/A

Note: gelu has no transform_aie2.mlir (NPU2-only), so it was not tested.

Failing tests (5)

1. i8 failures — Peano LLC error (vec-add/i8, axpy/i8, relu/i8)

All three i8 tests compile through MLIR transforms successfully but fail at the Peano LLC backend:

Error running Peano llc
Compilation failed

The installed Peano (llvm-aie 19.0.0.2025071101) cannot lower i8 vector operations to AIE2 machine code. This is consistent with the PR's own notes about i8 limitations on AIE2P (arith.muli/maxsi not supported for i8 vectors).

2. leaky_relu/f32 — Peano LLC error

Error running Peano llc

The leaky_relu kernel uses tl.where(x >= 0, x, 0.01 * x) which generates arith.cmpf + arith.select. With f32 input + bf16-emulation, Peano LLC cannot lower this combination on AIE2.

3. sigmoid/f32 — Transform padding type mismatch

'transform.structured.pad' op expects a padding value of type 'f32', got 0.000000e+00 : bf16

The sigmoid/transform_aie2.mlir hardcodes a bf16 padding value. When f32 input is used (with bf16-emulation), the pad value type mismatches. PR #48 updated transform_aie2p.mlir for sigmoid but did not update transform_aie2.mlir for this case. This is a fixable bug — the AIE2 transform script needs @DTYPE@ placeholder support for the padding value, or a dtype-aware pad constant.

Recommendations

• The sigmoid/f32 failure on AIE2 is a bug in the transform script and should be straightforward to fix
• The i8 and leaky_relu/f32 failures are Peano backend limitations on AIE2 — these should be documented as known unsupported configurations
• Consider adding NPU1 to the PR Add multi-dtype support to elementwise examples and new elementwise_arith ops #48 test plan or CI matrix

Ref: #48