Is TACO capable of doing caculation with extra itervar?

[qelk123]

Hi all,
I'm a beginer in TACO , and I am trying to use it to compile some stencil caculations.However,I didn't find a way to declare the extent and other attributes of axis which not included in the tensor dim.So,I am wondering is TACO capable of doing caculation for something like stencil?
Mike

[rohany]

TACO cannot currently be used for direct stencil calculations (i.e. A(i + 1, j + 1)), but it could be used through indirect stencil operations by constructing an indirect higher order stencil matrix and multiplying by that. If you're interested in sparse stencil operations, talk to @willow-ahrens about some work she's been up to!

[willow-ahrens]

It's been a while, but I wanted to give the best response I could on sparse tensor compiler convolution support.

https://dl.acm.org/doi/abs/10.1145/3485505 describes TACO support for static affine expressions (A[i + 1]), which would allow you to represent each point in a stencil separately. If there are only a few points in the stencil this could be a good approach.

https://drive.google.com/file/d/1_VIvuiBV7HHDOMnKppHbAMiNLOYJU7DG/view?usp=sharing is very recent work which describes TACO support for arbitrary affine expressions (A[i + j]), which would be a better fit for bigger stencils.

Finally, https://github.com/willow-ahrens/Finch.jl is a Julia-to-Julia sparse tensor compiler I've been working on which currently supports both of the above (for a very wide variety of operations and formats) using named index modifiers like A[offset[i, j]], you can read more about it at https://dl.acm.org/doi/10.1145/3579990.3580020.