Skip to content
GitLab
Projects Groups Topics Snippets
  • /
  • Help
    • Help
    • Support
    • Community forum
    • Submit feedback
    • Contribute to GitLab
  • Sign in
  • pystencils pystencils
  • Project information
    • Project information
    • Activity
    • Labels
    • Members
  • Repository
    • Repository
    • Files
    • Commits
    • Branches
    • Tags
    • Contributor statistics
    • Graph
    • Compare revisions
  • Issues 19
    • Issues 19
    • List
    • Boards
    • Service Desk
    • Milestones
  • Merge requests 5
    • Merge requests 5
  • CI/CD
    • CI/CD
    • Pipelines
    • Jobs
    • Schedules
  • Deployments
    • Deployments
    • Environments
    • Releases
  • Monitor
    • Monitor
    • Incidents
  • Analytics
    • Analytics
    • Value stream
    • CI/CD
    • Repository
  • Wiki
    • Wiki
  • Snippets
    • Snippets
  • Activity
  • Graph
  • Create a new issue
  • Jobs
  • Commits
  • Issue Boards
Collapse sidebar
  • pycodegen
  • pystencilspystencils
  • Issues
  • #34
Closed
Open
Issue created Apr 13, 2021 by Michael Kuron@kuronMaintainer

Use AVX512 scatter/gather

AVX512 has _mm512_i32gather_ps/_mm512_i32gather_pd that can load strided data into a vector and _mm512_i32scatter_ps/_mm512_i32scatter_pd which can write a vector to strided memory. These can be used to enable vectorization without assume_inner_stride_one=True. For LBM, this would permit vectorizing the streaming and collision kernels simultaneously. In general, it would also allow vectorizing with zyxf memory layout. I don't know how big the performance benefits would be, but it's certainly worth trying.

The Fujitsu A64fx also has special hardware support for scatter/gather with SVE instructions and would probably perform even better.

Edited Apr 18, 2021 by Michael Kuron
Assignee
Assign to
Time tracking