Primitive Singular Value Decomposition

Computes the reduced ("economy") singular value decomposition of a matrix operand of shape (m, n): $$A = u \, \mathrm{diag}(d) \, vt,$$ where u has orthonormal columns, vt has orthonormal rows, and d is the length-k (k = min(m, n)) vector of non-negative singular values in descending order.

Note: unlike base::svd(), which returns the right singular vectors as v of shape (n, k) (so that a = u %*% diag(d) %*% t(v)), this primitive returns them already transposed as vt of shape (k, n) (matching the underlying LAPACK / cuSOLVER output and avoiding an extra transpose).

Supports any matrix shape on both the host (LAPACK gesdd) and CUDA (cuSOLVER gesvd) backends. cuSOLVER's m >= n requirement is handled transparently via a layout flip for wide matrices.

Usage

prim_svd(operand)

Arguments

operand: (arrayish)
Matrix of data type floating-point with exactly 2 dimensions.

Value

Named list with elements d (length k), u (shape (m, k)), and vt (shape (k, n)). All have the same dtype as the input.

Implemented Rules

stablehlo

StableHLO

Lowers to stablehlo::hlo_custom_call() with target "svd".

Examples

x <- nv_array(c(1, 0, 0, 1, 0, 1), shape = c(3, 2))
prim_svd(x)
#> $d
#> AnvlArray
#>  1.6180
#>  0.6180
#> [ CPUf32{2} ] 
#> 
#> $u
#> AnvlArray
#>   0.8507  0.5257
#>   0.0000  0.0000
#>   0.5257 -0.8507
#> [ CPUf32{3,2} ] 
#> 
#> $vt
#> AnvlArray
#>   0.5257  0.8507
#>   0.8507 -0.5257
#> [ CPUf32{2,2} ] 
#>