Merge pull request #204 from wenkaidu/launch_flags

Set flags when calling hipExtLaunchMultiKernelMultiDevice in hip-clang [ROCm/rccl commit: 94d16c0f0a]
2020-05-08 11:20:50 -07:00
parent 65ad48404b 069322d05a
commit 0d642448d7
1 changed files with 4 additions and 0 deletions
@@ -74,7 +74,11 @@ ncclResult_t ncclLaunchCooperativeKernelMultiDevice(hipLaunchParams *paramsList,
  if (cgMode & 0x01) {
    CUDACHECK(hipExtLaunchMultiKernelMultiDevice(paramsList, numDevices,
            // These flags are to reduce the latency of using this API
+#if __HIP__
+            hipCooperativeLaunchMultiDeviceNoPreSync|hipCooperativeLaunchMultiDeviceNoPostSync));
+#else
            0));
+#endif
    return ncclSuccess;
  }
  int savedDev;