ArrayFire offers a number of matrix factorization routines, including Cholesky, LU, QR, and SVD. One notable missing factorization is the LDLT (or Bunch-Kaufman) factorization. It would be useful if this was available.
Description
LDLT factorization corresponds to the LAPACK sytrf function. It is implemented in cuSolver. I don't know about OpenCL. The advantage of this factorization over Cholesky is that it improves stability, possibly at the expense of performance, but it is still faster than alternatives like SVD.
For example, Matlab's backslash operator tries to solve matrix equations for symmetric matrices with a Cholesky decomposition first, and when this fails, falls back to LDLT.
The text was updated successfully, but these errors were encountered:
ArrayFire offers a number of matrix factorization routines, including Cholesky, LU, QR, and SVD. One notable missing factorization is the LDLT (or Bunch-Kaufman) factorization. It would be useful if this was available.
Description
LDLT factorization corresponds to the LAPACK
sytrf
function. It is implemented in cuSolver. I don't know about OpenCL. The advantage of this factorization over Cholesky is that it improves stability, possibly at the expense of performance, but it is still faster than alternatives like SVD.For example, Matlab's backslash operator tries to solve matrix equations for symmetric matrices with a Cholesky decomposition first, and when this fails, falls back to LDLT.
The text was updated successfully, but these errors were encountered: