Jay Zhuang

> having sparse tensors backing a quantum tensor network calculation is very useful. > having an ndarray backed by sparse tensors is useful to express batched SpMV and SpMM operations,...

Further speaking of HPC use cases, considering the two biggest domains in traditional HPC -- fluid mechanics (weather, climate, engineering) and molecular dynamics (chemistry, biology, combustion). Those physical problems are...

> In terms of number of use cases: `2-D > batched 2-D & block-sparse > 3-D > n-D`. > the semantics should match those of dense/arrays tensors, and that statement...

> Isn't this just `a @ b` or `matmul(a, b)`? Yes, for just SpMV. If the goal is just to build another CuPyx-like solver library (backed by one of PyTorch/MXNet/TVM...

> I expect there to be some follow-up yes within the next months. Thanks for the information -- what will be the right place to track the follow-ups on accelerating...

FYI, PyTorch 1.9 (released today) reimplements `numpy.linalg` and `scipy.special`: - https://pytorch.org/docs/1.9.0/linalg.html - https://pytorch.org/docs/1.9.0/special.html

> If you don't import face label info you will only be able to apply boundary conditions on the entire boundary The MFEM mesh file does contain labels, for both...

@fverdugo The MFEM `*.mesh` is a rather structured and simple text file, so I write a parser in Julia, by almost line-by-line translating from MFEM's C++ parser ([Mesh::Loader](https://github.com/mfem/mfem/blob/v4.4/mesh/mesh.cpp#L3954) and [Mesh::ReadMFEMMesh](https://github.com/mfem/mfem/blob/v4.4/mesh/mesh_readers.cpp#L39-L106))....

`sparse * sparse` is supported by cuSPARSE: https://github.com/JuliaGPU/CUDA.jl/issues/1264#issuecomment-992038154, but just not added to CUDA.jl `CUDA.CUSPARSE` yet. As an alternative to warning, setting `CUDA.allowscalar(false)` prevents performing these unimplemented sparse operations siliently...

The cppyy interface might be helpful here: https://github.com/wlav/cppyy