Skip to contents

GPU-accelerated truncated SVD via irlba

Source: R/irlba.R
irlba.Rd

Wraps irlba::irlba() with an adgeMatrix input so that every Lanczos matrix-vector product routes through the amatrix GPU dispatch path. The matrix A is kept resident on device; consecutive matvecs in the Lanczos loop avoid host round-trips.

Usage

irlba(
  A,
  nv = 5,
  nu = nv,
  ...,
  mode = "fast",
  backend = NULL,
  implementation = c("compat", "block"),
  block_size = NULL,
  n_steps = NULL
)

Arguments

A

A matrix, adgeMatrix, or adgCMatrix. Plain matrices are coerced via adgeMatrix(A, mode=mode, backend=backend).

nv

Number of right singular vectors.

nu

Number of left singular vectors. Defaults to nv.

...

Additional arguments forwarded to irlba::irlba(). fastpath is always forced to FALSE — the C fastpath bypasses S4 dispatch and cannot be GPU-accelerated.

mode

Execution mode passed to adgeMatrix() when coercing. "fast" permits float32 and enables GPU routing. Ignored if A is already an adgeMatrix or adgCMatrix.

backend

Backend name (e.g. "mlx", "arrayfire"). Ignored if A is already an amatrix object.

implementation

Lanczos implementation to use. "compat" preserves the current irlba::irlba() wrapper behavior. "block" routes to block_lanczos for a GEMM-oriented approximation.

block_size

Block size passed to am_block_lanczos() when implementation = "block". Defaults to a small MLX-friendly block size derived from the requested rank.

n_steps

Number of block Krylov steps passed to am_block_lanczos() when implementation = "block".

Value

Same structure as irlba::irlba(): a list with components d, u, v, iter, mprod.

Details

The hot loop in irlba is two matrix-vector products per Lanczos step: A %*% v and w %*% A. Both route through am_matmul() when A is an adgeMatrix, giving GPU acceleration on the dominant cost. Orthogonalization, svd(B), and convergence tests remain on CPU where they belong (the subspace dimension work is always small).

Do not pass mult= — it is deprecated in irlba and forces a non-standard dispatch path. Pass an adgeMatrix instead.

See also

adgeMatrix, svd_factor