mirror of
https://github.com/ceres-solver/ceres-solver.git
synced 2026-08-30 00:50:37 +08:00
Document rationale for not reordering program when using Accelerate.
Change-Id: Id11b3920cfdc706133ee497773a537287adc0a27
This commit is contained in:
@@ -569,9 +569,12 @@ bool ReorderProgramForSparseNormalCholesky(
|
||||
*tsm_block_jacobian_transpose,
|
||||
&ordering[0]);
|
||||
} else if (sparse_linear_algebra_library_type == ACCELERATE_SPARSE) {
|
||||
// TODO(alex): Investigate what analysis on pre-ordered matrices
|
||||
// Accelerate supports, but for now disable.
|
||||
// https://github.com/ceres-solver/ceres-solver/issues/394.
|
||||
// Accelerate does not provide a function to perform reordering without
|
||||
// performing a full symbolic factorisation. As such, we have nothing
|
||||
// to gain from trying to reorder the problem here, as it will happen
|
||||
// in AppleAccelerateCholesky::Factorize() (once) and reordering here
|
||||
// would involve performing two symbolic factorisations instead of one
|
||||
// which would have a negative overall impact on performance.
|
||||
return true;
|
||||
|
||||
} else if (sparse_linear_algebra_library_type == EIGEN_SPARSE) {
|
||||
|
||||
Reference in New Issue
Block a user