[Perf] Tiles 4: add SharedArray slice support#482
Merged
hughperkins merged 157 commits intomainfrom Apr 20, 2026
Merged
Conversation
TypeCheck has allow_undefined_visitor=true, so removing this override is a no-op. The original code was actively wrong (overwriting correct ret_type with i32), but nothing downstream relied on ret_type from the type_check pass for InternalFuncStmt, so the bug was latent. Removing the override eliminates the misleading TODO and prevents a future pass from accidentally depending on the wrong type.
…tests - Parametrize ger_sub and cholesky tests over f32/f64 dtypes - Use tighter tolerance (1e-10) for f64, 1e-5 for f32 - Parametrize cholesky over src_offset (0, 5, 32) and dst_delta (0, 3, 16) - Verify untouched regions of dst array remain at sentinel value
…olesky_ _ger_sub: 34 lines → 4 lines cholesky_: 224 lines → 22 lines
Quadrants DSL types don't carry operator overloads in their stubs, so pyright can't verify +=, *, /, > on shuffled values.
Fix scipy reference computation: solve X @ L^T = B requires solve_triangular(L, B.T) not solve_triangular(L, B, trans='T'). Add type: ignore[reportOperatorIssue] for DSL operator expressions in _trsm that pyright can't verify.
…_eye_
Replace 16-way explicit register access with loop-based _get_col/_set_col
helpers, eliminating ~145 lines of boilerplate. Also fix trsm test reference
computation (was computing B^T @ L^{-1} instead of B @ L^{-T}).
…e16x16 - Slice-based load/store: arr[r0:r1, c0:c1] for 2D and arr[b, r0:r1, c0:c1] for 3D - qd.outer(a, b) deferred proxy for augmented assignment (t -= qd.outer(a, b)) - _TileSliceProxy, _VecSliceProxy, _TileRefProxy for deferred subscript evaluation - AST purity exemptions for quadrants-internal code - _quadrants_internal flag on Tile struct
The purity checker only flags int/float/Field attribute accesses inside kernels. Tile() and Tile.zeros() don't trigger this, and Tile.SIZE is only accessed in plain Python, not inside kernels.
Add tile16.md covering tile creation, slice load/store, rank-1 updates, Cholesky, and triangular solve. Remove the unused _quadrants_internal getattr escape hatch from the purity checker (file-level check suffices). Add test for Tile.SIZE access inside a kernel.
…top in slices Update tile16 doc examples to use Tile.SIZE. Enforce that both start and stop indices are provided in tile/vec slice syntax (previously stop defaulted to start+16 when omitted). Reorder _TileSliceProxy.__init__ args to match (row_start, row_stop, col_start, col_stop).
The if-condition checking all four slice indices was too long for black. The test_tile16_load_missing_start_raises regex didn't match the updated error message after requiring both start and stop indices.
- test_tile16_outer_symmetric_same_variable: passes same variable for both args to qd.outer(v, v) - test_tile16_vec_proxy_ger_sub_3d: column vector load from 3D array via arr[batch, r0:r1, col] - test_outer_composition_raises: verifies qd.outer(a,b) + qd.outer(c,d) raises TypeError
…errors Only intercept 2D/1D array slices for tile proxy creation when _tile16_cache is non-empty (i.e., _make_tile16x16 has been called). Non-tile programs now get the original "does not support slice" error. Add _DeferredProxyMixin with __add__, __sub__, __mul__, __getitem__, and __repr__ that raise clear TypeErrors explaining the proxy is only valid in tile operations.
Check isinstance(value, t) against cached tile types before creating a _TileRefProxy for struct[:] subscripts. Non-tile structs now fall through to normal subscript handling.
# Conflicts: # tests/python/test_tile16.py
# Conflicts: # docs/source/user_guide/tile16.md
This reverts commit 876a666.
hughperkins
changed the title
Was hardcoded to _ATOLS[qd.f32] despite being parametrized over qd_dtype, making the f64 path 100,000x too lenient.
Collaborator
Author
|
Opus: Done. All 6 SharedArray tests are now parametrized over qd_dtype with _QD_DTYPES (f32, f64): Each test now includes skip_if_f64_unsupported, uses _NP_DTYPES[qd_dtype] for numpy dtypes, _ATOLS/_EPS_VALS where applicable, and passes qd_dtype: qd.Template as a kernel |
Syncs with main branch commit 62af01b ("[CI] Increase clock test tolerance to 20%"). The 10% threshold was flaky on CI GPU runners. Made-with: Cursor
Made-with: Cursor
…ure) Made-with: Cursor
Made-with: Cursor # Conflicts: # tests/python/test_tile16.py
erizmr
reviewed
Apr 17, 2026
| ratio = a[i] / a[0] | ||
| expected = i + 1 | ||
| assert abs(ratio - expected) / expected < 0.1 # 10% tolerance | ||
| assert abs(ratio - expected) / expected < 0.2 # 20% tolerance |
Collaborator
There was a problem hiding this comment.
How do we determine this tolearnce? why we change it from 10% to 20% here
Collaborator
Author
There was a problem hiding this comment.
This is not part of this PR.
The Manylinux wheel Test (3.12, ubuntu-22.04) hung for 6h and was cancelled. Previous run of the same commit range passed in ~10 min. Made-with: Cursor
erizmr
approved these changes
Apr 19, 2026
Collaborator
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.
# Conflicts: # docs/source/user_guide/tile16.md # tests/python/test_tile16.py
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
2 participants
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.
Issue: #
Brief Summary
Summary
Adds Tile16x16 ↔ SharedArray interop: tiles can now load from and store to
qd.simt.block.SharedArrayusing the same slice syntax as device arrays.What's in this PR
6 new SharedArray tests
test_tile16_shared_array_roundtrip— field → tile → SharedArray → tile → field, verify data survivestest_tile16_shared_array_partial_cols— partial-column load/store through SharedArray, parametrized over(partial_store, partial_load)combinationstest_tile16_shared_array_cholesky— Cholesky factorization with L stored in SharedArray, verify reconstructiontest_tile16_shared_array_clamp_store— store tile to SharedArray narrower than 16 cols, verify auto-clampingtest_tile16_shared_array_clamp_load— load tile from SharedArray narrower than 16 cols, verify extra registers are zerotest_tile16_vec_proxy_shared_array— symmetric rank-1 subtract via vec proxy loaded from SharedArray at non-zero row offsetDocs
tile16.mdwith code example and clamping behavior notesGood
v = sh[K0:K1, col]) also works with SharedArray, enabling lookback patterns in blocked CholeskyBad
test_tile16_vec_proxy_shared_arraytest manually copies data into SharedArray element-by-element with a loop, which is verbose but necessary since there's no bulk copyAPI
copilot:summary
Walkthrough
copilot:walkthrough