Summary

• Add A2/A3 VPTO flat-UB lowering infrastructure and gather support.
• Add pto.ub.vgather / pto.ub.vgatherb ops and lower them to packed llvm.hivm.VGATHER{B}.b16/.b32 calls.
• Lower pto.tgather index form through vmuls + vgather, preserving count-mode mask lifetime and chunking wide rows.
• Lower pto.tgatherb with CCE block-gather semantics: one compact 32B block address per output 32B block, padded to 8 address entries.
• Add verifier checks, lit coverage, PTODSL helpers, and NPU e2e tests.

Notes

• tgatherb is a 32B block gather, not scalar gather. The offset tile uses compact i32 32B-aligned block addresses.
• Scalar arbitrary element gather is covered by tgather index form / vgather.
• A2/A3 mgather remains out of scope because it is A5-only.

Validation

• Targeted ptoas/FileCheck gather lit checks: pass.
• NPU e2e on 910B2 devices 6/7:

pytest -q ptodsl/tests/e2e/test_gather.py -s
15 passed in 55.50s

Covered NPU shapes:

• tgatherb f32: 1x8, 1x64, 1x96, 2x64, 1x128, 4x64
• tgatherb f16: 1x16, 1x128, 1x192
• scalar tgather f32/f16: 1x64, 2x64, 2x128