Intel oneMKL SYCL BLAS

2026.0.0 verified Fri May 01 auth: no python

Intel oneAPI Math Kernel Library (oneMKL) BLAS routines for SYCL devices. Provides a SYCL implementation of BLAS operations using Intel oneMKL. Current version 2026.0.0; release follows Intel oneAPI release cadence (usually annual).

pip install onemkl-sycl-blas

Common errors

error ModuleNotFoundError: No module named 'onemkl_sycl_blas' ↓

cause Package was renamed/restructured in v2025.0.0.

fix

Run pip install onemkl-sycl-blas and use from onemkl._blas import blas.

error RuntimeError: Queue must be created with a SYCL device ↓

cause Using a default queue without a SYCL device (e.g., no GPU or Level Zero runtime).

fix

Ensure a SYCL device is available (e.g., dpctl.lsplatform() shows devices) and pass a valid queue.

error ValueError: Input array is not contiguous in memory ↓

cause Array is not in column-major order or is a view.

fix

Use np.asfortranarray() or ensure order='F' when creating arrays.

Warnings

breaking v2025.0.0 changed the internal module path from `onemkl_sycl_blas` to `onemkl._blas`. All imports using the old path will fail. ↓

fix Replace `import onemkl_sycl_blas` with `from onemkl._blas import blas`.

gotcha All BLAS routines expect column-major layout (Fortran order). Passing row-major arrays will produce incorrect results without error. ↓

fix Ensure arrays are column-major (use `np.asfortranarray()`) or transpose correctly.

gotcha The library does not support all BLAS operations (e.g., SPMV, TRSV). Check documentation before assuming availability. ↓

fix Consult the Intel oneMKL documentation for supported routines.

deprecated v2024.x and earlier used `dpnp` arrays. `dpnp` is deprecated; use `numpy` arrays with `dpctl` queue instead. ↓

fix Migrate from `dpnp` to `numpy` + `dpctl`.

Imports

blas
wrong
```
from onemkl_sycl_blas import blas
```
correct
```
from onemkl._blas import blas
```
Direct import from onemkl_sycl_blas may fail; use the underscore-prefixed internal module.
ColumnMajor, RowMajor
wrong
```
from onemkl_sycl_blas.enums import ColumnMajor
```
correct
```
from onemkl._blas import ColumnMajor, RowMajor
```
Enums are located in onemkl._blas, not in a separate enums module.

Quickstart

Basic GEMM operation using oneMKL SYCL BLAS. Note: matrices must be in column-major order.

import dpctl
import numpy as np
from onemkl._blas import blas, ColumnMajor

# Create a SYCL queue (use default GPU)
queue = dpctl.SyclQueue()

# Prepare matrices (ColumnMajor order required!)
m, n, k = 4, 4, 4
alpha = np.float64(1.0)
beta = np.float64(0.0)
A = np.random.rand(m, k).astype(np.float64)
B = np.random.rand(k, n).astype(np.float64)
C = np.zeros((m, n), dtype=np.float64)

# Compute C = alpha * A * B + beta * C
blas.gemm(queue, ColumnMajor, 'N', 'N', m, n, k, alpha, A, m, B, k, beta, C, m)
print(C)