Skip to content

[LAPACK][CUSOLVER] Inefficient conversion from 32-bit ints to 64-bit ints. #230

Open
@AidanBeltonS

Description

@AidanBeltonS

Summary

The legacy cuSolver API only accepts 32-bit ints as it is a limit of the library. oneMKL uses 64-bit ints, so in cases with ipiv vectors we use sycl kernels to convert from 32-bits to 64-bits. This is inefficient as it dispatches work to the GPU and requires an additional memory allocation. This could be improved by using an inplace expansion using the already 64-bit allocated memory.

The problem was discussed on #209. See #209 (comment)

This issue is to track the problem and its solution.

Metadata

Metadata

Assignees

No one assigned

    Labels

    help wantedTasks, issues or features that could be implemented and contributed to the project

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions