Reduce Python Code Duplication#1476
Draft
hunhoffe wants to merge 7 commits into
Draft
Conversation
Adds 5 in-kernel construct helpers to _air_ops_ext.py (auto-exported via
`from air.dialects.air import *`):
- l1_memref_type / l2_memref_type: memory-space-tagged MemRef types
- vec_type: 1D VectorType factory
- identity_map_attr: 1D identity AffineMapAttr for transfer_read/write
- tile_offset_1d: replaces the 12-line AffineMap offset boilerplate
Adds 2 runtime helpers to xrt_runner.py:
- make_air_parser: ArgumentParser with the 4 standard flags (-v, -p,
--compile-mode, --output-format) pre-populated
- run_on_npu: dispatches compile-only vs compile-and-run in one call
Adds programming_examples/utils.py as an alternative import path with
equivalent helpers (make_l1_memref, tiled_1d_offset, etc.) plus
vec_read/vec_write wrappers around subview+transfer_read/write.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
All examples now use helpers from utils.py (or equivalently from air.dialects.air and air.backend.xrt_runner) to eliminate boilerplate: In-kernel patterns eliminated: - MemRefType.get([...], dtype, memory_space=...) -> l1_memref_type/make_l1_memref - VectorType.get([N], dtype) -> vec_type/make_vec_type - AffineMapAttr.get(AffineMap.get_identity(1)) -> identity_map_attr/identity_map_1d - 12-line AffineMap.get/affine_apply offset block -> tile_offset_1d/tiled_1d_offset - subview+transfer_read/write pairs -> vec_read/vec_write Boilerplate eliminated in __main__: - import argparse + 4 add_argument calls -> make_air_parser() - if/elif compile_mode XRTRunner+XRTBackend dispatch -> run_on_npu() - inline sampled_indices/sampled_values block -> stochastic_check() - if args.print_module_only: print(); exit(0) -> check_print_module() Files updated: all 86 Python example scripts in programming_examples/ Note: examples with non-standard XRTRunner params (trace_offset, debug_ir, omit_pingpong, sequential test runs) retain their original runner instantiation; only the import consolidation and argparse refactor apply there. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Fixes from code review (3 FAIL files + 2 additional findings): 1. vector_reduce_add/vector_reduce_add.py: NameError fixes - Replace l1_memref_type() (Style B name) -> make_l1_memref() (Style A) - Add `from air.dialects import arith` (was used but not imported) - Replace identity_map_attr() -> identity_map_1d() (Style A name) 2. rms_norm/rms_norm.py: Add `import numpy as np` (np.* calls in __main__ but no numpy import) 3. layer_norm/layer_norm.py: Add `import numpy as np` (same issue as rms_norm) 4. conv2d/conv2d.py: Add `import numpy as np` (np.int32, np.random, np.zeros used but no numpy import) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The l1_memref_type, l2_memref_type, and tile_offset_1d helpers added in the previous commit referenced MemorySpace and affine_apply which were not imported in _air_ops_ext.py's module scope. Fixed by adding explicit top-level imports: - from ._air_enum_gen import MemorySpace as _MemorySpace - from .affine import apply as _affine_apply This ensures the helpers work correctly when invoked inside a @module_builder context. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
xrt-smi reports "NPU Strix Halo" but the regex only matched single-word NPU names (\w+ stops at the space). The model check also had a typo (run_on_2npu instead of run_on_npu). Fixes: - Regex: (\w+) -> ([\w ]+?) to capture multi-word names like "Strix Halo" - Model check: `model in ["npu4", "Strix"]` -> `"Strix" in model` for substring match, covering "Strix", "Strix Halo", "Strix Point", etc. - Typo fix: run_on_2npu -> run_on_npu (NameError on Strix systems) Without this fix, all ryzen_ai_npu2 tests were UNSUPPORTED on Strix Halo systems despite the NPU being present and XRT detecting it. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Delete the ~1100-line XRTBackend + XRTRunner implementation and replace with a thin layer on top of mlir-aie's CachedXRTRuntime: New public API (python/air/backend/xrt.py): - compile_air(air_module, ...) -> NPUKernel Compiles AIR dialect MLIR via aircc; no pyxrt dependency at compile time. - AirRuntime(CachedXRTRuntime) Richer verify_results(): rtol/atol, mismatch budget, Pearson correlation, stochastic sparse sampling. - get_air_runtime() -> AirRuntime (process-level singleton) - aie.utils.tensor() used throughout; zero raw pyxrt calls in core files. Backward-compat shims for XRTBackend, XRTRunner, XRTCompileArtifact kept so out-of-tree code continues to work during transition. Migrate 140 call sites (python/air/backend/xrt_runner.py, programming_examples/, test/xrt/) to compile_air() + get_air_runtime() / run_on_npu(). Remove all make_xrt_runner(), make_xrt_backend(), and backend.unload() calls. Clean up stale XRTRunner/XRTBackend imports. Format all changed files with black. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
There are still some easy things that could be done to clean up the Python further, but this PR is already large so I will defer those to a later time.