llvm icon indicating copy to clipboard operation
llvm copied to clipboard
verified publisher icon intel

[SYCL][Graph] Optimize graph enqueue for in-order queues

Open fabiomestre opened this issue 7 months ago • 2 comments

Optimizes the enqueue() function of sycl graphs to bypass the scheduler whenever possible and avoid creating events when not needed.

  • Refactors the executable graph enqueue() to have different paths depending on workload:
    • The direct path will be used when there are no host-tasks or accessor requirements in the graph and the execution dependencies are considered safe to bypass the scheduler.
    • The scheduler path will be used when there are requirements in the graph but no host-tasks or, if the execution dependencies require using the scheduler.
    • The multiple partitions path will be used when the graph contains host-tasks which requires scheduling multiple graph partitions. The implementation was also changed to avoid adding unnecessary event dependencies to partition executions and avoiding copying CGData when possible.
  • Extends the changes in https://github.com/intel/llvm/pull/18277 to sycl graphs. This means that no implicit events will be created when using in-order queues and graphs without host-tasks. Also updates the handler to only request events from the graph enqueue() when they are needed.

fabiomestre avatar Jun 03 '25 16:06 fabiomestre

@cperkinsintel , final review is on you: image

aelovikov-intel avatar Jun 12 '25 17:06 aelovikov-intel

@intel/llvm-gatekeepers Can this PR be merged? The existing failures seem to be CI issues. This PR was green for those jobs before:

Intel-Arc jobs:

  • https://github.com/intel/llvm/actions/runs/15681443933/job/44175220366?pr=18792

  • https://github.com/intel/llvm/actions/runs/15681443933/job/44175220336?pr=18792

Cuda UR Job:

  • https://github.com/intel/llvm/actions/runs/15681443929/job/44174565853?pr=18792

There were only unrelated changes to the HIP UR adapter made since then.

Edit: Nevermind, recently merged commits broke this PR, needs to be rebased.

fabiomestre avatar Jun 17 '25 12:06 fabiomestre

@intel/llvm-gatekeepers This PR is ready to merge. The CI failure on PVC is unrelated (I have seen it in other PR's).

fabiomestre avatar Jun 18 '25 18:06 fabiomestre

@intel/llvm-gatekeepers This PR is ready to merge. The CI failure on PVC is unrelated (I have seen it in other PR's).

Failing tests are unrelated and tracked by https://github.com/intel/llvm/issues/18932

uditagarwal97 avatar Jun 18 '25 19:06 uditagarwal97