Commit 726ee40f authored by Johannes Doerfert's avatar Johannes Doerfert
Browse files

[OpenMP] Move the recording code to account for KernelLaunchEnvironment

We need to record late to account for the kernel launch environment as
well as the potential changes in block and thread count.
parent b6f2597a
Loading
Loading
Loading
Loading
+6 −6
Original line number Diff line number Diff line
@@ -504,6 +504,12 @@ Error GenericKernelTy::launch(GenericDeviceTy &GenericDevice, void **ArgPtrs,
          printLaunchInfo(GenericDevice, KernelArgs, NumThreads, NumBlocks))
    return Err;

  if (RecordReplay.isRecording())
    RecordReplay.saveKernelInputInfo(
        getName(), getImage(), ArgPtrs, ArgOffsets,
        KernelArgs.NumArgs - /* KernelLaunchEnvironment */ 1, NumBlocks,
        NumThreads, KernelArgs.Tripcount);

  return launchImpl(GenericDevice, NumThreads, NumBlocks, KernelArgs,
                    KernelArgsPtr, AsyncInfoWrapper);
}
@@ -1405,12 +1411,6 @@ Error GenericDeviceTy::launchKernel(void *EntryPtr, void **ArgPtrs,
  GenericKernelTy &GenericKernel =
      *reinterpret_cast<GenericKernelTy *>(EntryPtr);

  if (RecordReplay.isRecording())
    RecordReplay.saveKernelInputInfo(
        GenericKernel.getName(), GenericKernel.getImage(), ArgPtrs, ArgOffsets,
        KernelArgs.NumArgs, KernelArgs.NumTeams[0], KernelArgs.ThreadLimit[0],
        KernelArgs.Tripcount);

  if (RecordReplay.isRecording())
    RecordReplay.saveImage(GenericKernel.getName(), GenericKernel.getImage());