Add 'projects/rccl/' from commit '1f2f9f33bac3e8ecfd84c69af6063d7352c362fc'

git-subtree-dir: projects/rccl git-subtree-mainline: 3fd8a0d393 git-subtree-split: 1f2f9f33ba
2025-12-11 20:46:05 +00:00
parent 3fd8a0d393 1f2f9f33ba
commit 42d84317cf
682 changed files with 293338 additions and 0 deletions
@@ -0,0 +1,115 @@
+/*************************************************************************
+ * Copyright (c) 2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * See LICENSE.txt for license information
+ ************************************************************************/
+
+ // Note: InPlace is not supported for All-To-All
+
+#include "TestBed.hpp"
+
+namespace RcclUnitTesting
+{
+  TEST(AllToAll, OutOfPlace)
+  {
+    TestBed testBed;
+
+    // Configuration
+    std::vector<ncclFunc_t>     const funcTypes       = {ncclCollAllToAll};
+    std::vector<ncclDataType_t> const dataTypes       = {ncclFloat16, ncclFloat32};
+    std::vector<ncclRedOp_t>    const redOps          = {ncclSum};
+    std::vector<int>            const roots           = {0};
+    std::vector<int>            const numElements     = {1048576, 1024};
+    std::vector<bool>           const inPlaceList     = {false};
+    std::vector<bool>           const managedMemList  = {false};
+    std::vector<bool>           const useHipGraphList = {false};
+
+    testBed.RunSimpleSweep(funcTypes, dataTypes, redOps, roots, numElements,
+                           inPlaceList, managedMemList, useHipGraphList);
+    testBed.Finalize();
+  }
+
+  TEST(AllToAll, OutOfPlaceGraph)
+  {
+    TestBed testBed;
+
+    // Configuration
+    std::vector<ncclFunc_t>     const funcTypes       = {ncclCollAllToAll};
+    std::vector<ncclDataType_t> const dataTypes       = {ncclFloat64, ncclBfloat16, ncclFloat8e4m3, ncclFloat8e5m2};
+    std::vector<ncclRedOp_t>    const redOps          = {ncclSum};
+    std::vector<int>            const roots           = {0};
+    std::vector<int>            const numElements     = {5685};
+    std::vector<bool>           const inPlaceList     = {false};
+    std::vector<bool>           const managedMemList  = {false};
+    std::vector<bool>           const useHipGraphList = {true};
+
+    testBed.RunSimpleSweep(funcTypes, dataTypes, redOps, roots, numElements,
+                           inPlaceList, managedMemList, useHipGraphList);
+    testBed.Finalize();
+  }
+
+  TEST(AllToAll, ManagedMem)
+  {
+    TestBed testBed;
+
+    // Configuration
+    std::vector<ncclFunc_t>     const funcTypes       = {ncclCollAllToAll};
+    std::vector<ncclDataType_t> const dataTypes       = {ncclUint8};
+    std::vector<ncclRedOp_t>    const redOps          = {ncclSum};
+    std::vector<int>            const roots           = {0};
+    std::vector<int>            const numElements     = {384 * 1024, 1024};
+    std::vector<bool>           const inPlaceList     = {false};
+    std::vector<bool>           const managedMemList  = {true};
+    std::vector<bool>           const useHipGraphList = {false};
+
+    testBed.RunSimpleSweep(funcTypes, dataTypes, redOps, roots, numElements,
+                           inPlaceList, managedMemList, useHipGraphList);
+    testBed.Finalize();
+  }
+
+  TEST(AllToAll, ManagedMemGraph)
+  {
+    TestBed testBed;
+
+    // Configuration
+    std::vector<ncclFunc_t>     const funcTypes       = {ncclCollAllToAll};
+    std::vector<ncclDataType_t> const dataTypes       = {ncclUint32, ncclUint64};
+    std::vector<ncclRedOp_t>    const redOps          = {ncclSum};
+    std::vector<int>            const roots           = {0};
+    std::vector<int>            const numElements     = {1048576};
+    std::vector<bool>           const inPlaceList     = {false};
+    std::vector<bool>           const managedMemList  = {true};
+    std::vector<bool>           const useHipGraphList = {true};
+
+    testBed.RunSimpleSweep(funcTypes, dataTypes, redOps, roots, numElements,
+                           inPlaceList, managedMemList, useHipGraphList);
+    testBed.Finalize();
+  }
+
+    TEST(AllToAll, Channels)
+  {
+    TestBed testBed;
+    if(testBed.ev.maxGpus >= 8) {
+      if(testBed.ev.isGfx94) {
+        // Configuration
+        std::vector<ncclFunc_t>     const funcTypes       = {ncclCollAllToAll};
+        std::vector<ncclDataType_t> const dataTypes       = {ncclBfloat16};
+        std::vector<ncclRedOp_t>    const redOps          = {ncclSum};
+        std::vector<int>            const roots           = {0};
+        std::vector<int>            const numElements     = {64 * 1024 * 1024, 1024};
+        std::vector<bool>           const inPlaceList     = {false};
+        std::vector<bool>           const managedMemList  = {false};
+        std::vector<bool>           const useHipGraphList = {false, true};
+        std::vector<const char *>   const channelList     = {"112"}; 
+        bool                        const enableSweep     = false;
+        for (auto channel : channelList) {
+          setenv("NCCL_MIN_NCHANNELS", channel, 1);
+          testBed.RunSimpleSweep(funcTypes, dataTypes, redOps, roots, numElements,
+                                inPlaceList, managedMemList, useHipGraphList, enableSweep);
+          testBed.Finalize();
+          unsetenv("NCCL_MIN_NCHANNELS");
+        }
+      }
+    }
+  }
+}