Jay Zhuang

Use sparse LU as a faster coarse solver than Pinv

3

The only available `CoarseSolver` is `Pinv`, which gets very slow if the coarsest-level A is large, for example in a two-grid setting instead of multigrid V-cycle. I wrote a code...

Does STRUMPACK support sparse LDLT factorization for symmetric matrix?

8

From the [Sparse Direct Solver](https://portal.nersc.gov/project/sparse/strumpack/v6.3.1/sparse.html) section, it seems to use the general sparse LU factorization (when no compression is used). Is there a LDLT implementation for symmetric matrix (which would...

Solving MFEM's ComplexHypreParMatrix in Strumpack's native complex arithmetic

1

I am attempting to use Strumpack to solve complex-valued, indefinite Maxwell equation (motivated by https://github.com/mfem/mfem/issues/2869#issuecomment-1123588971). I found the [`mfem::STRUMPACKSolver`](https://docs.mfem.org/html/classmfem_1_1STRUMPACKSolver.html#a23836709a06d0be816e1250efc93394e) interface and its usage in [MFEM's ex11p.cpp](https://github.com/mfem/mfem/blob/v4.2/examples/ex11p.cpp#L265-L275) (thanks for that!). Then...

Parallel executor for upper triangular solve (back-substitution)

9

## Problem description First thanks for releasing this code, great work! I saw various implementions of lower triangular solve (forward substitution), but not for upper triangular solve (back-substitution) . In...

## Problem description With left preconditioner `Pl`, the recorded residual (by setting `log=true`) is computed on the preconditioned equation `norm(Pl*b - Pl*A*x) = norm(Pl*r)`, not the original equation `norm(b -...

Allow reordering for incomplete cholesky

1

Reordering can reduce fill-in for both complete and incomplete factorizations. The `CholeskyPreconditioner` calls `lldl()` in `LimitedLDLFactorizations.jl`, but without suppling the reorder parameter: https://github.com/JuliaLinearAlgebra/Preconditioners.jl/blob/52b3702e9c8c5d42d6092ed2df7efdbe07c81411/src/incompletecholesky.jl#L8-L13 [In lldl's example](https://github.com/JuliaSmoothOptimizers/LimitedLDLFactorizations.jl/blob/main/examples/example.jl), AMD or METIS reordering...

Understanding why TorchInductor cannot speed-up huggingface transformer inference

1

## Problem `torch.compile()` shows an impressive ~2x speed-up for this code repo, but when applying to huggingface transformers there is barely no speed-up. I want to understand why, and then...

Out-of-place version of sptrsv executors?

6

All current [sptrsv executors](https://github.com/sympiler/sympiler/blob/97f3b839be22f3e86de87a43ec51c5e721878b89/sparse_blas/sptrsv.cpp) are in-place: ```cpp void sptrsv_csr(int n, int *Lp, int *Li, double *Lx, double *x) { int i, j; for (i = 0; i < n; i++)...

Any good way to support int64 type for sparse matrix index?

1

All the current [sparse BLAS routines](https://www.sympiler.com/docs/sympiler-lib/) assume `int` type for row/column pointers `Lp` and column/row indices `Li`. However for very large matrices the `nnz` size can exceed 2.1 billion (INT32_MAX=2,147,483,647),...

Code fixes for local-storage-only environment

2

In certain virtualized environment there is no shared storage. Both source code and data are stored (replicated) in each worker node's local storage. The code sections below only load data...