GPU-aware singular value decomposition

GPU-aware singular value decomposition

Usage

cuda_svd(
  x,
  nu = min(nrow(x), ncol(x)),
  nv = min(nrow(x), ncol(x)),
  device = c("auto", "cuda", "cpu")
)

Arguments

x

A finite numeric matrix or cudatensor.

nu, nv

Number of left and right singular vectors to return.

device

One of "auto", "cuda", or "cpu".

Value

A list with d, u, v, and the actual device. Matrix row and column names are retained on the corresponding singular vectors.

Details

A native CUDA cudatensor selected on the same backend is validated for finite values on the device and passed directly to cuSOLVER. Float32 and integer storage is converted to float64 on the device. Only the requested decomposition results are materialized in R.

Examples

cuda_svd(matrix(rnorm(30), 10, 3), device = "cpu")
#> <cuda_svd rank=3 device=cpu compute=cpu backend=base>