Is DIA or ELL currently supported ? can you provide an example?

Hi,

I am trying to generate code for Spmv kernel when the matrix is in DIA or ELL formats. I wonder if the current version support DIA or ELL formats ? If yes, can you please provide a simple example to demonstrate spmv for these two sparse formats? Thank you in advance.

The current version of TACO (on the master branch) does not support DIA and ELL unfortunately. I'm still hoping to be able to add support for those formats to the master branch, but recently we've been more focused on fixing other aspects of TACO, so it might be a while before we're able to fully integrate the work described in our OOPSLA 2018 paper into master.

Have TACO already supported COO, ELL and DIA? How can I use it?

TACO currently has full support for sorted COO (Format({Compressed(ModeFormat::NOT_UNIQUE), Singleton}) for matrices, Format({Compressed(ModeFormat::NOT_UNIQUE), Singleton(ModeFormat::NOT_UNIQUE), Singleton}) for three-dimensional tensors) as well as some support for ELL (Format({Dense, Dense, Singleton(ModeFormat::PADDED)})). Unfortunately, the master branch of TACO still currently does not support DIA.

Note that ELL is treated as a format for a three-dimensional tensor as opposed to a matrix, with the added (first) dimension representing the relative position of each nonzero within its containing row in the original matrix (refer to Figures 9 and 2d in this paper); the master branch of TACO currently cannot do this remapping automatically, so you would need to do it manually when constructing the tensor. To see a more concrete example of how to compute with ELL, you can take a look at line 962 in test/tests-expr_storage.cpp.

@stephenchouca Thank you!