oneMKL SYCL Sparse

2026.0.0 verified Fri May 01 auth: no python

Intel oneAPI Math Kernel Library (oneMKL) Sparse BLAS routines for SYCL devices. This package provides optimized sparse linear algebra operations (e.g., sparse matrix-vector multiply, sparse triangular solvers) on Intel GPUs and CPUs using SYCL. Current version: 2026.0.0. Released quarterly as part of Intel's oneAPI toolkit.

pip install onemkl-sycl-sparse

Common errors

error ModuleNotFoundError: No module named 'onemkl_sycl_sparse' ↓

cause Package not installed or wrong Python environment.

fix

Install: pip install onemkl-sycl-sparse

error AttributeError: module 'onemkl_sycl_sparse' has no attribute 'sparse' ↓

cause Incorrect import. The correct import is 'from onemkl_sycl_sparse import sparse'.

fix

Use: from onemkl_sycl_sparse import sparse

error dpctl._sycl.SyclQueueCreationError: No device of requested type available ↓

cause No compatible SYCL device (GPU/CPU) found or driver issues.

fix

Check Intel GPU/CPU drivers, install Intel oneAPI runtime, or use 'cpu' selector: dpctl.SyclQueue("cpu")

Warnings

gotcha All input arrays (row, col, val) must be on the same SYCL device as the queue. Do not mix host memory (numpy) directly without copying to USM. ↓

fix Use dpctl.tensor.usm_ndarray to allocate device memory and copy data.

gotcha Matrix handles must be explicitly created and destroyed. Forgetting sparse.destroy_handle leads to memory leaks. ↓

fix Always pair sparse.create_handle with sparse.destroy_handle, ideally in a try-finally block or context manager (if available).

gotcha The sparse module expects Fortran-style (column-major) ordering for some internal operations, but CSR arrays are typically row-major. Ensure row-index array is sorted and unique per row. ↓

fix Pre-process CSR arrays to be in sorted column order per row (e.g., using scipy sparse CSR construction).

deprecated The old import path from 'oneapi.mkl' is deprecated and removed in 2026.0.0. ↓

fix Use 'from onemkl_sycl_sparse import sparse'.

Install

conda install -c intel onemkl-sycl-sparse

Imports

sparse
wrong
```
import onemkl_sycl_sparse
```
correct
```
from onemkl_sycl_sparse import sparse
```
The correct import uses the 'sparse' submodule; importing the top-level package gives no useful functions.

Quickstart

Performs a sparse matrix-vector multiply using CSR format on a SYCL GPU device.

import dpctl
import numpy as np
from onemkl_sycl_sparse import sparse

# Create a SYCL queue (device selector: 0 for gpu, 1 for cpu)
queue = dpctl.SyclQueue("gpu")

# Create a simple CSR matrix
row = np.array([0, 0, 1, 2], dtype=np.int64)
col = np.array([0, 1, 1, 2], dtype=np.int64)
val = np.array([1.0, 2.0, 3.0, 4.0], dtype=np.float64)
nrows, ncols = 3, 3

# Allocate USM memory
d_row = dpctl.tensor.usm_ndarray(row, dtype=row.dtype, queue=queue)
d_col = dpctl.tensor.usm_ndarray(col, dtype=col.dtype, queue=queue)
d_val = dpctl.tensor.usm_ndarray(val, dtype=val.dtype, queue=queue)

# Create handle and perform operation
handle = sparse.create_handle(queue)
sparse_matrix = sparse.init_csr_matrix(handle, nrows, ncols, d_row, d_col, d_val)

# Sparse matrix-vector multiply
x = np.array([1.0, 2.0, 3.0], dtype=np.float64)
d_x = dpctl.tensor.usm_ndarray(x, dtype=x.dtype, queue=queue)
d_y = dpctl.tensor.empty(3, dtype=np.float64, queue=queue)
sparse.omatadd(handle, sparse_matrix, sparse_matrix, 1.0, 1.0, d_x, d_y)
y = d_y.asnumpy()
print(y)

sparse.destroy_handle(handle)