Insufficient datatypes

[QingleiCao]

PaRSEC only defines a limited set of datatypes—BYTE, INT, FLOAT, DOUBLE, COMPLEX_FLOAT, and COMPLEX_DOUBLE. This coverage is too narrow for today’s mixed-precision, accelerator-centric workflows and prevents clean support for commonly used formats (e.g., FP16/BF16/FP8, wider integer variants, and other application-specific scalar types).

[bosilca]

parsec mimics the MPI datatypes. Adding other types will be difficult, especially if they are not supported by MPI. You will need to have an equivalence table, such that the communication layer knows how to deal with these new types, especially their size and alignment. Moreover, as long as MPI does not have support for them, you will be limited to homogeneous environments (which might not be a big issue right now).

[QingleiCao]

parsec mimics the MPI datatypes. Adding other types will be difficult, especially if they are not supported by MPI. You will need to have an equivalence table, such that the communication layer knows how to deal with these new types, especially their size and alignment. Moreover, as long as MPI does not have support for them, you will be limited to homogeneous environments (which might not be a big issue right now).

I see. Maybe (1) we can add some MPI supported datatypes first, e.g., INT16, INT8; (2) for those not supported by MPI, e.g., FP8, FP4, 1bit, add support to declare the matrix structure and use BYTE for communication?

[bosilca]

You already have access to most typical types. Take a look in datatype.h.