Clique tuning upgrade (#352)

* Enabling clique for any XGMI-connected topology, adding tuning
* Updating CHANGELOG for clique tuning
* Re-working clique barrier system to work on multi-process / multi-gpu

test/test_AllReduce.cpp

@@ -8,56 +8,56 @@
 
 namespace CorrectnessTests
 {
-    TEST_P(AllReduceCorrectnessTest, Correctness)
+  TEST_P(AllReduceCorrectnessTest, Correctness)
+  {
+    if (numDevices > numDevicesAvailable) return;
+
+    // Prepare input / output / expected results
+    Dataset dataset;
+    dataset.Initialize(numDevices, numElements, dataType, inPlace);
+    FillDatasetWithPattern(dataset);
+    ComputeExpectedResults(dataset, op);
+
+    // Launch the reduction (1 thread per GPU)
+    ncclGroupStart();
+    for (int i = 0; i < numDevices; i++)
     {
-        if (numDevices > numDevicesAvailable) return;
-
-        // Prepare input / output / expected results
-        Dataset dataset;
-        dataset.Initialize(numDevices, numElements, dataType, inPlace);
-        FillDatasetWithPattern(dataset);
-        ComputeExpectedResults(dataset, op);
-
-        // Launch the reduction (1 thread per GPU)
-        ncclGroupStart();
-        for (int i = 0; i < numDevices; i++)
-        {
-            ncclAllReduce(dataset.inputs[i], dataset.outputs[i],
-                          numElements, dataType, op, comms[i], streams[i]);
-        }
-        ncclGroupEnd();
-
-        // Wait for reduction to complete
-        Synchronize();
-
-        // Check results
-        ValidateResults(dataset);
-
-        dataset.Release();
+      ncclAllReduce(dataset.inputs[i], dataset.outputs[i],
+                    numElements, dataType, op, comms[i], streams[i]);
     }
+    ncclGroupEnd();
 
-    INSTANTIATE_TEST_SUITE_P(AllReduceCorrectnessSweep,
-                            AllReduceCorrectnessTest,
-                            testing::Combine(
-                                // Reduction operator
-                                testing::Values(ncclSum, ncclProd, ncclMax, ncclMin),
-                                // Data types
-                                testing::Values(ncclInt8,
-                                                ncclUint8,
-                                                ncclInt32,
-                                                ncclUint32,
-                                                ncclInt64,
-                                                ncclUint64,
-                                                //ncclFloat16,
-                                                ncclFloat32,
-                                                ncclFloat64,
-                                                ncclBfloat16),
-                                // Number of elements
-                                testing::Values(1024, 1048576),
-                                // Number of devices
-                                testing::Values(2,3,4,5,6,7,8),
-                                // In-place or not
-                                testing::Values(false, true),
-                                testing::Values("RCCL_ENABLE_CLIQUE=0", "RCCL_ENABLE_CLIQUE=1")),
-                            CorrectnessTest::PrintToStringParamName());
+    // Wait for reduction to complete
+    Synchronize();
+
+    // Check results
+    ValidateResults(dataset);
+
+    dataset.Release();
+  }
+
+  INSTANTIATE_TEST_SUITE_P(AllReduceCorrectnessSweep,
+                           AllReduceCorrectnessTest,
+                           testing::Combine(
+                             // Reduction operator
+                             testing::Values(ncclSum, ncclProd, ncclMax, ncclMin),
+                             // Data types
+                             testing::Values(ncclInt8,
+                                             ncclUint8,
+                                             ncclInt32,
+                                             ncclUint32,
+                                             ncclInt64,
+                                             ncclUint64,
+                                             //ncclFloat16,
+                                             ncclFloat32,
+                                             ncclFloat64,
+                                             ncclBfloat16),
+                             // Number of elements
+                             testing::Values(1024, 1048576),
+                             // Number of devices
+                             testing::Values(2,3,4,5,6,7,8),
+                             // In-place or not
+                             testing::Values(false, true),
+                             testing::Values("RCCL_ENABLE_CLIQUE=0", "RCCL_ENABLE_CLIQUE=1")),
+                           CorrectnessTest::PrintToStringParamName());
 } // namespace