View page source

warp.tile_cholesky_solve_inplace#

warp.tile_cholesky_solve_inplace(
L: Tile[Float, tuple[int, int]],
y: Tile[Float, tuple[int]],
fill_mode: str,
) → None#

  • Kernel

Solve for x in Ax = y by overwriting y with x.

When fill_mode="lower" (default), L is lower-triangular such that LL^T = A. When fill_mode="upper", L is upper-triangular U such that U^T U = A.

The fill_mode parameter must be a compile-time constant.

Note: This inplace variant does not support automatic differentiation (adjoint computation), but avoids allocating shared memory for the output x by reusing y’s memory.

Supported datatypes are:

  • float32

  • float64

Parameters:

  • L – A square triangular Cholesky factor of A.

  • y – A 1D or 2D tile of length M that gets overwritten by x where Ax = y.

  • fill_mode – "lower" (default) or "upper". Must be a compile-time constant.