[PATCH] D11885: AMDGPU/SI: Add SI Machine Scheduler

Tom Stellard via llvm-commits llvm-commits at lists.llvm.org
Thu Sep 10 10:44:43 PDT 2015

tstellarAMD added inline comments.

Comment at: lib/Target/AMDGPU/AMDGPUTargetMachine.cpp:134-135
@@ -123,2 +133,4 @@
       return createR600MachineScheduler(C);
+    //else //(uncomment to turn default for SI)
+    //  return createSIMachineScheduler(C);
     return nullptr;
axeldavy wrote:
> tstellarAMD wrote:
> > Ok.  We'll need to remove these commented lines before the patch is committed.
> Then all the SI tests should be changed to use the converge scheduler (which is the current default one), else lots
> of them will fail (because instruction order differs).
> Since I'm not sure the scheduler is good for opencl tasks (perhaps the converge scheduler is better then),
> I was thinking having mesa explicitly ask for the scheduler was better, but another option is to that opencl programs ask for the converge scheduler (or whatever else) if this scheduler is not good for them.
While the converge scheduler is still the default, if Mesa wants to use this scheduler, then it should explicitly ask for it using the command line options.


More information about the llvm-commits mailing list