[Perf] Tiles 4d: Consolidate slice error tests using parametrize#508
Conversation
The purity checker was special-casing qd.math via a local import. Replace with a general is_from_quadrants_module() check that covers any quadrants module or class (but not instances, to avoid suppressing mutable state), and remove the qd_math special case. Add tests for both the allowed case (qd.math.pi) and the rejected case (non-quadrants object attribute).
Use == 'quadrants' or .startswith('quadrants.') instead of bare
.startswith('quadrants') to avoid matching hypothetical third-party
packages like quadrantslib.
The merge resolution accidentally dropped this import, which is still used in build_Name. Restores the import alongside is_from_quadrants_module.
Parametrize 16 existing tile16 tests with [qd.ndarray, qd.field] to verify both tensor backends produce identical results. Uses an _ann() helper to select the correct kernel annotation per tensor type.
Pass _TILE as a qd.Template parameter (tile_size) to all 31 kernels, eliminating closure captures that trigger purity violations. All captured fields use qd.Template annotations, captured ints use qd.i32. Removes test_tile16_size_constant_in_kernel (needs public Tile16x16 API).
These were closure-captured and used in qd.static() blocks, which would break fastcache since the cache key wouldn't vary with their values.
Adds the _Tile16x16Proxy class and qd.simt.Tile16x16 lazy accessor, enabling qd.simt.Tile16x16.zeros(dtype=...) and .eye(dtype=...) inside kernels. Includes 18 new tests covering the proxy, vec-proxy patterns, potrf, trsm, partial load/store, f64 roundtrip, CPU error, and reinit dtype survival. SharedArray slice tests are intentionally excluded.
GPU clock cycle ratios can deviate >1 from the ideal linear scaling, especially under CI load. Previous run showed a[12]/a[0] ≈ 11.97 vs expected 13, a deviation of 1.03 that just exceeded the ±1 bound. Made-with: Cursor
Made-with: Cursor
Replace _make_tile16x16/Tile() pattern with qd.simt.Tile16x16.zeros() and qd.simt.Tile16x16.eye() across all tests except test_tile16_make_caching and test_tile16_solve_triangular_upper_raises.
This reverts commit 876a666.
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
LGTM — mechanical test consolidation using , no logic changes.
Extended reasoning...
Overview
This PR modifies a single test file (), consolidating 7 individual slice-error test functions into 3 parametrized tests (, , ). The refactor uses branches inside the kernel to select the bad-slice variant at template-instantiation time, matching the pattern already established elsewhere in the file.
Security Risks
None. This is a test-only change with no production code modifications.
Level of Scrutiny
Low. The change is purely mechanical — same assertions, same error messages, same kernel structure — just deduplicated via parametrize. The parametrize tuples cover the same four cases (neg_row, neg_col, no_start, no_stop) as the original individual tests.
Other Factors
No bugs reported by the automated bug hunter. No prior reviews. The pattern (parametrized kernel with + ) is already used extensively in the file, so this is a consistent, low-risk change.
Was hardcoded to _ATOLS[qd.f32] despite being parametrized over qd_dtype, making the f64 path 100,000x too lenient.
erizmr
left a comment
erizmr
left a comment
There was a problem hiding this comment.
Approving on the basis that I have reviewed the design and the public facing API, tests, and they look reasonable to me.
Made-with: Cursor # Conflicts: # tests/python/test_tile16.py
Made-with: Cursor # Conflicts: # tests/python/test_tile16.py
2 participants
Issue: #
Brief Summary
Summary
Consolidates 7 individual slice-error tests into 3 parametrized tests, reducing boilerplate while adding store-side error coverage that was previously missing.
What's in this PR
test_tile16_load_slice_errors— replacestest_tile16_load_negative_row_raises,test_tile16_load_negative_col_raises,test_tile16_load_missing_start_raises,test_tile16_load_missing_stop_raises. Parametrized over 4 cases (neg_row,neg_col,no_start,no_stop) usingqd.static()branching andbad_slice: qd.Template.test_tile16_store_slice_errors— new parametrized test covering negative-index and missing-start/stop errors on the store side (d[-1:tile_size, ...],d[:tile_size, ...],d[0:, ...]). Previously onlytest_tile16_store_missing_stop_raisesexisted; this adds 3 new error cases.test_tile16_vec_slice_errors— replacestest_tile16_vec_slice_missing_stop_raisesandtest_tile16_vec_slice_missing_start_raises. Parametrized over 2 cases.Net change: 89 lines removed, 50 lines added.
Good
bad_slice: qd.Templateas an explicit kernel parameter (consistent withfastcache=Truepurity)Bad
qd.static()branching pattern in the error tests is less immediately readable than separate test functions — you have to trace the parametrize values to understandwhat each case tests
are error tests)
copilot:summary
Walkthrough
copilot:walkthrough