FAQs

Here, we will maintain a list of FAQs.

  1. What compressed storage formats are supported to represent sparse tensors? Sparse tensors can be represented using CSR (2D tensors) or CSF (> 2D tensors). Internally, COMET uses a standard representation (see reference) that can express most compressed storage formats like CSR, DCSR, or COO.

  2. How does COMET populate sparse tensors? COMET DSL supports reading of sparse matrices from .mtx (matrix market format) files. Whereas, .tns (FROSTT file format) files are used for populating sparse tensors. The .mtx and .tns files are human readable text files where each line represents a non-zero element. The runtime function gets an integer input (read_from_file(0)) that is correlated with the user-defined environment variable SPARSE_FILE_NAME0 appended with integer input provided as argument to the runtime function.

  3. Where can one find examples of sparse matrices and tensors? The SuiteSparse Matrix Collection has an ample collection of sparse matrices. The Formidable Repository of Open Sparse Tensors and Tools (FROSTT) contains some higher order tensors.