Skip to content
GitLab
Projects Groups Snippets
  • /
  • Help
    • Help
    • Support
    • Community forum
    • Submit feedback
    • Contribute to GitLab
  • Sign in
  • A aml
  • Project information
    • Project information
    • Activity
    • Labels
    • Members
  • Repository
    • Repository
    • Files
    • Commits
    • Branches
    • Tags
    • Contributors
    • Graph
    • Compare
  • Issues 33
    • Issues 33
    • List
    • Boards
    • Service Desk
    • Milestones
  • Merge requests 7
    • Merge requests 7
  • 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
  • argo
  • aml
  • Merge requests
  • !167

Remove Cuda Layout

  • Review changes

  • Download
  • Email patches
  • Plain diff
Merged Nicolas Denoyelle requested to merge no_layout_cuda into staging Dec 04, 2020
  • Overview 1
  • Commits 2
  • Pipelines 4
  • Changes 9

As the title says, this MR removes cuda layout.

Overall, layouts should not be backend specific and include other informations than the data layout. If specific information relative to the destination needs to be passed to the dma, it should rather be embedded in the copy operator arguments or the dma data structure.

This will greatly simplify crafting higher level blocks involving DMAs. When such a block makes several dma requests, it would require to create several layouts. Therefore, DMAs would need to be associated with src/dst layout constructors. For instance copying to cuda devices would imply a cuda layout. If a higher abstraction needs to instanciate layouts, for instance a deepcopy mapper, layout creation operators need to be passed as argument to be able to perform dma requests on different pieces of data.

Next, the MR also removes the script for cuda device code compilation. It is kept in a separate commit to ease eventual future reintegration.

Edited Dec 04, 2020 by Nicolas Denoyelle
Assignee
Assign to
Reviewers
Request review from
Time tracking
Source branch: no_layout_cuda