Mosaic: An interoperable compiler for tensor algebra

Manya Bansal, Olivia Hsu, Kunle Olukotun, Fredrik Kjolstad
[doi] [Google Scholar] [DBLP] [Citeseer] [url]

Proc. ACM Program. Lang. 7(PLDI)
Association for Computing Machinery
New York, NY, USA
jun 2023
Note(s): machine learning, tensor, sparse model, loop fusion

Tensor algebra compilers are important because we want to target a very broad range of hardware (with new designs coming out all the time) and to cope with many different combinations of sparse and dense tensors (where each dimension of each operand can be dense or sparse).

The big contribution of this paper is that it can exploit both vector instructions (like existing tensor algebra compilers) and vector libraries (hand-written or generated by other tools).