Repository navigation
Feature sparse linalg solvers - #2841
Merged
antonwolfy merged 178 commits intoOct 5, 2026
Merged
Conversation
…oneMKL hooks
- _interface.py: add full operator algebra (.H, .T, +, *, **, neg),
_AdjointLinearOperator, _TransposedLinearOperator, _SumLinearOperator,
_ProductLinearOperator, _ScaledLinearOperator, _PowerLinearOperator,
IdentityOperator, MatrixLinearOperator, _AdjointMatrixOperator,
_CustomLinearOperator factory dispatch; extend aslinearoperator
to handle dpnp sparse and dense arrays
- _iterative.py: add _make_system (dtype validation, preconditioner
wiring, working dtype selection); add _make_fast_matvec CSR/oneMKL
SpMV hook; fix GMRES Arnoldi inner product to single oneMKL BLAS
gemv (dpnp.dot) instead of slow Python vdot loop; offload
Hessenberg lstsq to numpy.linalg.lstsq (CPU, matches CuPy);
fix SciPy host-fallback tol->rtol deprecation via _scipy_tol_kwarg;
add preconditioner support to CG; keep MINRES as SciPy-backed stub
Refs: CuPy v14.0.1 cupyx/scipy/sparse/linalg/_interface.py,
cupyx/scipy/sparse/linalg/_iterative.py"
…gmres, minres
Modeled after CuPy's cupyx_tests/scipy_tests/sparse_tests/test_linalg.py.
Covers:
- LinearOperator: shape, dtype inference, matvec/rmatvec/matmat,
subclassing, __matmul__, __call__, edge cases
- aslinearoperator: dense array, duck-type, identity passthrough,
rmatvec from dense, invalid inputs
- cg: SPD convergence, scipy reference match, x0 warm start, b_ndim=2,
callback, atol, LinearOperator path, invalid inputs,
non-convergence info check
- gmres: diag-dominant convergence, scipy reference match, restart
variants, x0, b_ndim=2, callbacks, complex systems, atol,
non-convergence info check, Hilbert-matrix stress test
- minres: SPD, symmetric-indefinite, scipy reference, shift parameter,
non-square guard, LinearOperator path, callback
- Integration: parametric (n, dtype) cross-solver tests via LinearOperator
- Import smoke tests: __all__ completeness
- Use dpnp.tests.helper: assert_dtype_allclose, generate_random_numpy_array, get_all_dtypes, get_float_complex_dtypes, has_support_aspect64 - Use dpnp.tests.third_party.cupy testing harness (with_requires, etc.) - Use numpy.testing assert_allclose / assert_array_equal / assert_raises - Use dpnp.asnumpy() instead of numpy.asarray() - Use pytest parametrize ids matching existing test conventions - Use is_scipy_available() helper from tests/helper.py - Strict class-per-solver organisation matching TestCholesky / TestDet etc.
…or dtype Two bugs fixed: 1. _init_dtype() was calling dpnp.zeros(n) which defaults to float64, so a float32 matvec would upcast and return float64, making the inferred dtype wrong. Fix: use dpnp.zeros(n, dtype=dpnp.int8) as SciPy/CuPy do — any numeric matvec will promote int8 to its own dtype. 2. _CustomLinearOperator.__init__ called _init_dtype() even when an explicit dtype was already supplied, overwriting the caller's value. Fix: _init_dtype() now short-circuits when self.dtype is already set.
…ption handling Align gemv.cpp with the conventions established in blas/gemm.cpp: Headers added: - ext/common.hpp (dpctl_td_ns, consistent with other extensions) - utils/memory_overlap.hpp (MemoryOverlap guard on x vs y) - utils/output_validation.hpp (CheckWritable + AmpleMemory on y) - utils/type_utils.hpp (validate_type_for_device<T> in impl) - <sstream> (needed for stringstream error_msg) Exception handling added in sparse_gemv_impl(): - try/catch(oneapi::mkl::exception) around all oneMKL sparse calls - try/catch(sycl::exception) around all oneMKL sparse calls - release_matrix_handle cleanup in the exception error path - throw std::runtime_error with descriptive message on catch Input validation added in sparse_gemv(): - ndim checks: x and y must be 1-D - queues_are_compatible() across all 5 USM arrays - MemoryOverlap()(x, y) aliasing guard - CheckWritable::throw_if_not_writable(y) - AmpleMemory::throw_if_not_ample(y, num_rows) - keep_args_alive() at function exit (was missing, returning empty event)
… table
Modeled after blas/gemm.cpp (2-D table: value type x index type) and
blas/gemv.cpp (dispatch vector pattern with ContigFactory + init_dispatch_table).
Changes:
- Add sparse/types_matrix.hpp with SparseGemvTypePairSupportFactory<Tv, Ti>
encoding the 4 supported combinations: {float32,float64} x {int32,int64}
- Rewrite sparse_gemv_impl() to take typeless char* pointers (matching
the blas gemv_impl signature style) — type info flows through template
params only, no runtime branching inside the impl
- Replace the 60-line if/else val_typenum/idx_typenum chain in sparse_gemv()
with a 2-D dispatch table lookup (gemv_dispatch_table[val_id][idx_id])
- Rename init_sparse_gemv_dispatch_vector -> init_sparse_gemv_dispatch_table
and implement it via init_dispatch_table<> from ext/common.hpp
- All validation guards and exception handling from prior commit are preserved
…se_gemv_dispatch_table Follows the rename made in gemv.cpp when the dispatch mechanism was changed from a 1-D vector to a 2-D table (value type x index type). All other declarations (sparse_gemv signature, parameters) are unchanged.
The oneMKL 2025-2 sparse BLAS API deprecated the old 8-argument
set_csr_data(queue, handle, nrows, ncols, index_base, row_ptr, col_ind,
values, deps) overload in favour of a new signature that takes the
sparse matrix handle as `spmat` and adds an explicit `nnz` argument:
set_csr_data(queue, spmat, nrows, ncols, nnz, index_base,
row_ptr, col_ind, values, deps)
Fixes:
- Replace old set_csr_data call with the new nnz-aware signature
- Silences the resulting -Wunused-parameter warning on `nnz` (now used)
- No functional change; all other logic is unchanged
…tring Line 477: `hasattr(A, "rmatmat\")` had a Markdown-escaped backslash leaked into the Python source, causing an unterminated string literal. Fixed to `hasattr(A, "rmatmat")`.
dpnp.ndarray blocks implicit NumPy conversion via __array__ to prevent silent dtype=object arrays. All test assertions must use .asnumpy() to materialize device arrays onto the host explicitly. Also replaces numpy.asarray(x_dp) in _rel_residual helper.
…dation order - _iterative.py: raise NotImplementedError for M != None *before* the _HOST_N_THRESHOLD SciPy fast-path in cg() and gmres(), so the contract is enforced regardless of system size (fixes test_cg_preconditioner_unsupported_raises, test_gmres_preconditioner_unsupported_raises). - _iterative.py: validate callback_type and raise NotImplementedError for 'pr_norm' *before* the _HOST_N_THRESHOLD branch in gmres(), so small-n systems also see the error (fixes test_gmres_callback_type_pr_norm_raises). - _iterative.py: pass callback_type='legacy' to scipy.sparse.linalg.gmres when delegating on the fast path to suppress SciPy DeprecationWarning. - test_scipy_sparse_linalg.py: add dtype=numpy.float64 to expected arange() calls in test_identity_operator and test_gmres_happy_breakdown so strict NumPy 2.0 dtype-equality checks pass (float64 result vs int64 expected).
… port SciPy corner cases
- Replace .asnumpy() method calls with dpnp.asnumpy() module fn (asnumpy is not an ndarray method in dpnp; it is a top-level fn) - Fix dpnp.any(x) ambiguous truth value in x0 zero-check; replace with explicit `x0 is not None` guard for r0 initialisation - Fix V_mat.T.conj() -> dpnp.conj(V_mat.T) in GMRES Arnoldi step - Guard minres beta sqrt against tiny negative floats: sqrt(abs(...)) - Unify GMRES Hessenberg h_np assignment to avoid .real stripping producing wrong dtype for complex systems - Fix float() cast on dpnp scalar norm inside GMRES inner h_j1 line
…failures) The committed code used hypot(gbar, oldb) as delta_k which is the gamma (norm) from the PREVIOUS rotation step, not the correct diagonal entry from applying the previous Givens rotation to the current column. The correct Paige-Saunders (1975) two-rotation recurrence is: oldeps = epsln delta = cs * dbar + sn * alpha # apply previous rotation gbar_k = sn * dbar - cs * alpha # residual -> new rotation input epsln = sn * beta dbar = -cs * beta gamma = hypot(gbar_k, beta) # NEW rotation eliminates beta cs = gbar_k / gamma sn = beta / gamma w_new = (v - oldeps*w - delta*w2) / gamma # three-term update This matches scipy.sparse.linalg.minres and Choi (2006) eq. 6.11. The buggy recurrence produced solutions ~1.08x away from the true solution (rel_err ~1e0) instead of the expected ~1e-13. Co-authored-by: fix-minres-recurrence
Reading usm_ndarray data pointers after py::gil_scoped_release aborted the interpreter, since get_data() calls into the Python C-API; read both pointers before releasing the GIL, matching in_place.tpp.
antonwolfy
reviewed
Aug 30, 2026
Co-authored-by: Anton <100830759+antonwolfy@users.noreply.github.com>
…e bug A prior commit accidentally deleted the entire rmatvec/rmatmat/adjoint/ transpose subsystem from _interface.py while addressing unrelated review comments. Restore it, keeping the two comments that did target this file: - LinearOperator.__init__: validate _isshape(shape) before int() truncation, so non-integer shapes like (3.7, 3.2) raise instead of silently rounding. - MatrixLinearOperator._matvec: unify sparse/dense branches to self.A.dot(x), since dpnp.ndarray.dot already dispatches to gemv. Also fix _ScaledLinearOperator: alpha_dtype used type(alpha) for plain Python scalars, forcing strong dtype promotion (float64/complex128) even when the operator was float32/complex64. This crashed LinearOperator.__init__'s dpnp.empty(dtype=float64) probe on fp64-less devices (e.g. Iris Xe). Promote weakly from the scalar value instead, matching array-API semantics. Verified: dpnp/tests/test_scipy_sparse_linalg.py 393 passed, on both the default device and ONEAPI_DEVICE_SELECTOR=opencl:cpu.
antonwolfy
reviewed
Oct 1, 2026
Co-authored-by: Anton <100830759+antonwolfy@users.noreply.github.com>
- clang-format: rewrap mkl_sparse::spmv calls in sparse/gemv.cpp - gersemi: reformat execute_process COMMAND block in CMakeLists.txt - pylint R0904: disable too-many-public-methods on csr_matrix - pylint C0301: wrap 81-char comment in gmres
antonwolfy
reviewed
Oct 1, 2026
The Conda package workflow builds with rattler-build using conda-recipe/rattler_recipe.yaml, which was missing the onemkl-sycl-sparse run dependency (present only in conda-recipe/meta.yaml). Without it libmkl_sycl_sparse.so.6 is absent at runtime and importing the sparse SpMV extension fails.
- dense path: allocate nnz==0 arrays and indptr via *_like(dense) so usm_type and queue follow the (possibly moved) input - sort_indices: build row ids with the indices' usm_type, so indexing no longer promotes data/indices to "device" - toarray: allocate the result like self._data - fix comments claiming placement kwargs are no-ops for device input
third_party/cupyx/.../sparse_tests/test_linalg.py:
- drop non-upstream TestLinearOperatorSmoke (covered by own-scope tests)
- restore upstream M/N grid [1, 6]/[1, 7] and FutureWarning filters
- wrap the scipy side in aslinearoperator too for use_linear_operator
- use plain with_requires("scipy") and upstream gmres matrix setup
- skip float64/complex128 (not 8-byte dtypes) on fp64-less devices
- add from __future__ import annotations; drop stale notes
test_scipy_sparse_linalg.py:
- use the device default float type in tests hardcoding float64
- black formatting
gemv.cpp: clang-format indent fix
numpy.float64(2.0) * op raised on devices without fp64: the scalar's float64 dtype made the scaled operator float64, which LinearOperator rejects. Map a numpy scalar's dtype to the default device, as dpnp elementwise ops do, so the declared dtype matches matvec's result. The test now also checks that dtype match.
antonwolfy
approved these changes
Oct 5, 2026
antonwolfy
left a comment
antonwolfy
left a comment
Contributor
There was a problem hiding this comment.
Thanks @abagusetty for the great work and for patiently addressing all the review feedback, LGTM!
To get the PR ready for the release I pushed a few follow-up commits directly to the branch:
- fixed pre-commit failures (clang-format, gersemi, pylint)
- added the missing
onemkl-sycl-sparserun dependency toconda-recipe/rattler_recipe.yaml - shortened the alpha/beta and placeholder comments in
sparse/gemv.cpp csr_matrixnow preserves the requestedusm_typein the dense constructor path,sort_indices()andtoarray()- aligned
third_party/cupyx/.../sparse_tests/test_linalg.pywith upstream CuPy (dropped the non-upstream class, restored the size-1M/Ncases and theFutureWarningfilters, wrapped the SciPy side inaslinearoperator) and fixed tests failing on devices without fp64 _ScaledLinearOperatormaps a NumPy scalar's dtype to the device, so e.g.numpy.float64(2.0) * opno longer raises on devices without fp64
The CI failures are unrelated to this PR.
Contributor
Author
|
Thanks @antonwolfy @vlad-perevezentsev for pushing this. Really helps the with the app, we are working with. Many months of effort taking into shape. Appreciate for the patience with this PR! |
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
5 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.
Adds support for
from dpnp.scipy.sparse.linalg import LinearOperator, cg, gmres, minresFixes: #2831