Merge pull request #119 from sclarkson/master

Fix tests: call cudaHostUnregister on the host pointer instead of the device pointer.

[ROCm/rccl commit: 286916a1a3]

projects/rccl/test/include/test_utilities.h

@@ -310,7 +310,7 @@ double CheckDelta(const T* results, const T* expected, int N) {
   CUDACHECK(cudaHostGetDevicePointer(&devmax, &maxerr, 0));
   deltaKern<T, 512><<<1, 512>>>(results, devexp, N, devmax);
   CUDACHECK(cudaHostUnregister(&maxerr));
-  CUDACHECK(cudaHostUnregister((void*)devexp));
+  CUDACHECK(cudaHostUnregister((void*)expected));
   return maxerr;
 }