Announcement

Collapse
No announcement yet.

Testing The LLVM SI Machine Instruction Scheduler

Collapse
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Testing The LLVM SI Machine Instruction Scheduler

    Phoronix: Testing The LLVM SI Machine Instruction Scheduler

    Landing last month in the LLVM SVN/Git code-base was the SI machine scheduler for the AMDGPU LLVM back-end. This scheduler has the potential to improve the performance for some hardware/workloads, but not by the wide margins originally reported by some early testers...

    http://www.phoronix.com/scan.php?pag...sisched-Test-1

  • #2
    i think there might be something odd about these graphs.... Xonotic seriously produces more framerates when multiplying the amount of pixels by 4? in Ultra settings???

    Comment


    • #3
      Would be interesting to see radeontop numbers of xonotic on fury. It is probably idling at 1080p :P

      Comment


      • #4
        Originally posted by jakubo View Post
        i think there might be something odd about these graphs.... Xonotic seriously produces more framerates when multiplying the amount of pixels by 4? in Ultra settings???
        It could be CPU bound...

        Comment


        • #5
          How does the SI scheduler work on SI hardware? I heard somewhere it had to be manually enabled? And does this only exist for LLVM 3.8+ on SI hardware? Does the scheduler only really apply to AMDGPU?

          Comment


          • #6
            Originally posted by jakubo View Post
            i think there might be something odd about these graphs.... Xonotic seriously produces more framerates when multiplying the amount of pixels by 4? in Ultra settings???
            It's just random noise, the test is CPU bound so you don't see any real differences at all.

            I'm confused why Michael didn't post the users other test which was a Unigine benchmark. That one did show some improvements and wasn't CPU bound.

            Comment


            • #7
              I didn't see any improvement either, but this is noted on the actual llvm page for it.
              http://reviews.llvm.org/D11885
              It is off by default, but can be used
              with --misched=si

              Comment

              Working...
              X