How can I improve the runtime of this linear system solve?

[Oyibo]

Hey @ofk123 ,

Have you tried Numba’s “get_cython_function_address”?

The Lapack functions should be:

Decomposer:
void dpotrf(char *uplo, int *n, d *a, int *lda, int *info)

Solver:
void dpotrs(char *uplo, int *n, int *nrhs, d *a, int *lda, d *b, int *ldb, int *info)

from numba.extending import get_cython_function_address

addr_decom = get_cython_function_address('scipy.linalg.cython_lapack','dpotrf')
addr_solve = get_cython_function_address('scipy.linalg.cython_lapack','dpotrs')

# TODO: 
# Define c-types, c-function
# Retrieve function address

Here are discussions that might help:

Extending LAPACK functions into numba

Hello, I was hoping to include the function scipy.linalg.ordqz into a numba njit function. This function calls two LAPACK functions, and I found this issue on github (Issue 5301, sorry I’m not allowed to include links) showing how to do this, so I was initially optimistic. The first function I tried to implement was zgges, which is the complex generalized Schur decomposition. I’m on windows, numba 0.55.1, python 3.10.4. The windows thing turns out to be relevant, quelle surprise! Anyway, here …

Negative eigenvalues in call to dsyevr

Coming from the as a reference: How to include LAPACK functions with complex inputs · Issue #5301 · numba/numba · GitHub I have the following attempt at calling dsyevr from scipy using numba: import ctypes import numpy as np import numba as nb from numba import vectorize, njit, jit from numba.extending import get_cython_function_address addr = get_cython_function_address('scipy.linalg.cython_lapack','dsyevr') _PTR, _dble, _char, _int = ctypes.POINTER, ctypes.c_double, ctypes.c_char, ctypes.c…