Removing OpenMP from unit tests (#163)

test/test_Reduce.cpp

@@ -5,7 +5,6 @@
  ************************************************************************/
 
 #include "test_Reduce.hpp"
-#include <omp.h>
 
 namespace CorrectnessTests
 {
@@ -25,7 +24,7 @@ namespace CorrectnessTests
             ComputeExpectedResults(dataset, op, root);
 
             // Launch the reduction (1 thread per GPU)
-            #pragma omp parallel for num_threads(numDevices)
+            ncclGroupStart();
             for (int i = 0; i < numDevices; i++)
             {
                 ncclReduce(dataset.inputs[i],
@@ -33,6 +32,7 @@ namespace CorrectnessTests
                            numElements, dataType, op,
                            root, comms[i], streams[i]);
             }
+            ncclGroupEnd();
 
             // Wait for reduction to complete
             Synchronize();