cuNumeric.jl

Appearance

Standard functions ​

Matrix multiply ​

For two 2D arrays, * performs matrix multiplication. Use .* for an elementwise product.

julia
using LinearAlgebra
using cuNumeric

A = cuNumeric.rand(Float32, 128, 128)
B = cuNumeric.rand(Float32, 128, 128)

C = A * B                 # allocates a new result
mul!(similar(C), A, B)    # in-place GEMM into an existing array
LinearAlgebra.mul! Method
julia
LinearAlgebra.mul!(out::NDArray, arr1::NDArray, arr2::NDArray)

Compute the matrix multiplication of arr1 and arr2, storing the result in out.

This function performs the operation in-place, modifying out.

Examples

julia
a = cuNumeric.ones(2, 3)
b = cuNumeric.ones(3, 2)
out = cuNumeric.zeros(2, 2)
LinearAlgebra.mul!(out, a, b)
source

Vector operations ​

Supported LinearAlgebra operations include:

  • mul!(y, A, x) and mul!(y, A, x, α, β) for matrix-vector products. A * x is also supported.

  • mul!(C, A, B, α, β) for matrix-matrix products.

  • dot(x, y) and entrywise norm(A, p) (dense-array norms currently require a GPU).

  • axpy!, axpby!, lmul!, and rmul!.

  • ldiv!(y, D, x) and ldiv!(D, x) for an NDArray-backed Diagonal.

dot and dense-array norm return device scalars.

Solve ​

cuNumeric.solve(A, b) solves a linear system and returns an array with the same shape as b. A is a square m × m matrix and b has shape (m,) or (m, n). A \ b is equivalent.

julia
A = cuNumeric.rand(Float32, 64, 64)
b = cuNumeric.rand(Float32, 64)
x = cuNumeric.solve(A, b)     # or: A \ b

B = cuNumeric.rand(Float32, 64, 4)
X = A \ B

Helpers ​

These helpers live on NDArray and are also listed in the Public API:

  • cuNumeric.transpose

  • cuNumeric.diag (2D to 1D)

  • cuNumeric.trace

Iterative solvers ​

See the Krylov.jl example for conjugate gradient, workspace reuse, and the scalar-handling macros.