BatchLAS documentation

Batched dense and sparse linear algebra for SYCL: thousands of small-to-medium problems in one call. Each call is routed, per op, size and batch, between native SYCL kernels and the vendor library (cuBLAS, cuSOLVER, cuSPARSE, rocBLAS, netlib), whichever measured faster.

User guide

Build, install, and call BatchLAS from C++ or Python.

Get started →

API reference

Every function, type and option, grouped by operation family.

Browse the API →

Architecture

Matrix model, kernel selection, workspaces, errors and the environment.

Read the design →

Algorithms

The mathematics behind the solvers.

See the algorithms →

Performance

Which kernel runs for a shape, and the measurements behind that choice.

See the evidence →

Developer guide

Build, test, benchmark, tune and document the library.

Contribute →

A minimal call

#include <batchlas.hh>
using namespace batchlas;
 
Queue ctx(Device::default_device());          // the backend comes from the device
gemm(ctx, A, B, C, {.alpha = 2.0f});
potrf(ctx, S, PotrfOptions{.uplo = Uplo::Upper});
ctx.wait();                                   // nothing is readable before this

Every entry point enqueues and returns an Event. Matrices are column-major in device-accessible (USM) memory. Use one Queue per thread. The full contract is in The BatchLAS C++ API.

Operations

Family	Operations
Batched BLAS	gemm, gemv, symm, syrk, syr2k, trmm, trsm, hemm, herk, her2k
Factorizations and solves	potrf, posv, getrf, getrs, getri, gesv, geqrf, orgqr, ormqr
Eigenvalues and SVD	syev, syevx (selected range), gesvd (one-sided Jacobi and blocked)
Sparse and spectral	spmm, lanczos, steqr, stedc, ritz_values, ILU(k)
Utilities	ortho, matrix generators, norms, condition numbers
Python	pybind11 package with NumPy and SciPy interop (BatchLAS Python examples)

Performance against cuSOLVER

Float, RTX 4090, CUDA backend, large batch.

Op	BatchLAS	cuSOLVER	Speed-up	Conditions
syev with eigenvectors, n = 320, batch 819	67.8 µs/matrix	203.0 µs/matrix	3.0×	measurement
syev with eigenvectors, n = 448, batch 585	(ratio only)	(ratio only)	2.1×	same
gesvd vs cusolverDnXgesvdjBatched, n = 8, batch 16384	0.0064 µs/matrix	0.339 µs/matrix	53×	comparison
gesvd vs cusolverDnXgesvdjBatched, n = 32	0.468 µs/matrix	0.768 µs/matrix	1.6×	same

Where the vendor library is faster for a shape, it runs instead (Kernel selection tables).

Source, issues and releases: github.com/jonasdelacour/BatchLAS.

User guide

Architecture and design

Algorithms

Routing and performance evidence

Results and evidence database

Developer guide