rocm-systems/projects/rccl/test/AllToAllVTests.cpp

/*************************************************************************
 * Copyright (c) 2023 Advanced Micro Devices, Inc. All rights reserved.
 *
 * See LICENSE.txt for license information
 ************************************************************************/

 // Note: InPlace is not supported for All-To-Allv

#include "TestBed.hpp"

namespace RcclUnitTesting
{
  // Prepare sendcount/recvcounts, sdispls/rdispls arrays within options
  void PrepareCounts(int const totalRanks, int const chunkSize,
                     OptionalColArgs& options,
                     std::vector<size_t>& numInputElements,
                     std::vector<size_t>& numOutputElements,
                     int percentZeroElement = 0)
  {
    numInputElements.clear();
    numOutputElements.clear();
    numInputElements.resize(totalRanks, 0);
    numOutputElements.resize(totalRanks, 0);

    // Decide how many elements each pair send/recv
    for (int sendRank = 0; sendRank < totalRanks; ++sendRank)
    for (int recvRank = 0; recvRank < totalRanks; ++recvRank)
    {
      // Get linear indices into sendcounts/recvcounts array
      int const sendIdx = sendRank * totalRanks + recvRank;
      int const recvIdx = recvRank * totalRanks + sendRank;

      // Each pair sends slightly different amounts of elements (based on chunkSize)
      int numElements = (1 + sendRank + recvRank) * chunkSize;
      options.sendcounts[sendIdx]  = options.recvcounts[recvIdx] = numElements;
    }

    // Psuedo-randomly zero out some of the sends
    int s = 0, r = 0;
    double zeroStride = (percentZeroElement <= 0)   ? (totalRanks * totalRanks + 1) :
                        (percentZeroElement >= 100) ? 1.0
                                                    : 100.0 / percentZeroElement;
    double zeroTarget = zeroStride;
    for (int i = 1; i <= totalRanks * totalRanks; i++) {
      if (i >= zeroTarget) {
        options.sendcounts[s * totalRanks + r] = options.recvcounts[r * totalRanks + s] = 0;
        zeroTarget += zeroStride;
      }
      int next = ((s*totalRanks+(r+(s*s)%7/2)%totalRanks) + totalRanks-1) % (totalRanks*totalRanks);
      s = next / totalRanks;
      r = ((next % totalRanks) - (s*s)%7/2) % totalRanks;
      if (r < 0) r += totalRanks;
    }

    // Compute displacements
    for (int sendRank = 0; sendRank < totalRanks; ++sendRank)
    {
      int totalSend = 0;
      int totalRecv = 0;

      for (int recvRank = 0; recvRank < totalRanks; ++recvRank)
      {
        int const pairIdx = sendRank * totalRanks + recvRank;

        options.sdispls[pairIdx] = totalSend;
        options.rdispls[pairIdx] = totalRecv;

        totalSend += options.sendcounts[pairIdx];
        totalRecv += options.recvcounts[pairIdx];
      }

      numInputElements[sendRank] = totalSend;
      numOutputElements[sendRank] = totalRecv;
    }
  }

  TEST(AlltoAllv, OutOfPlace)
  {
    TestBed testBed;

    // Configuration
    std::vector<ncclDataType_t> const& testDataTypes   = {ncclInt32, ncclFloat64, ncclFloat16};
    bool                        const  inPlace         = false;
    bool                        const  useManagedMem   = false;
    bool                        const  useHipGraph     = false;

    OptionalColArgs options;

    std::vector<ncclDataType_t> dataTypes;
    testBed.GetSupportedDataTypes(dataTypes, testDataTypes);
    if (dataTypes.empty()) {
      GTEST_SKIP() << "Skipping... test datatypes excluded by UT_DATATYPES.";
    }

    bool isCorrect = true;
    for (int totalRanks : testBed.ev.GetNumGpusList())
    for (int isMultiProcess : testBed.ev.GetIsMultiProcessList())
    {
      int const numProcesses = isMultiProcess ? totalRanks : 1;
      const std::vector<int>& gpuPriorityOrder = testBed.ev.GetGpuPriorityOrder();
      testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks, gpuPriorityOrder));

      // Prepare AlltoAllV options
      std::vector<size_t> numInputElements;
      std::vector<size_t> numOutputElements;
      PrepareCounts(totalRanks, 256, options, numInputElements, numOutputElements, 40);

      for (int dataIdx = 0; dataIdx < dataTypes.size() && isCorrect; ++dataIdx)
      {
        if (testBed.ev.showNames)
        {
          std::string name = testBed.GetTestCaseName(totalRanks, isMultiProcess,
                                                     ncclCollAlltoAllv, dataTypes[dataIdx],
                                                     ncclSum, -1, inPlace, useManagedMem, useHipGraph);
          INFO("%s\n", name.c_str());
        }

        for (int rank = 0; rank < totalRanks; ++rank)
        {
          testBed.SetCollectiveArgs(ncclCollAlltoAllv,
                                    dataTypes[dataIdx],
                                    numInputElements[rank],
                                    numOutputElements[rank],
                                    options,
                                    -1,
                                    0,
                                    rank);
        }
        testBed.AllocateMem(inPlace, useManagedMem);
        testBed.PrepareData();
        testBed.ExecuteCollectives({}, useHipGraph);
        testBed.ValidateResults(isCorrect);
        testBed.DeallocateMem();
      }
      testBed.DestroyComms();
    }
    testBed.Finalize();
  }


  TEST(AlltoAllv, OutOfPlaceGraph)
  {
    TestBed testBed;

    // Configuration
    std::vector<ncclDataType_t> const& testDataTypes   = {ncclFloat32, ncclInt8};
    bool                        const  inPlace         = false;
    bool                        const  useManagedMem   = false;
    bool                        const  useHipGraph     = false;

    OptionalColArgs options;

    std::vector<ncclDataType_t> dataTypes;
    testBed.GetSupportedDataTypes(dataTypes, testDataTypes);
    if (dataTypes.empty()) {
      GTEST_SKIP() << "Skipping... test datatypes excluded by UT_DATATYPES.";
    }

    bool isCorrect = true;
    for (int totalRanks : testBed.ev.GetNumGpusList())
    for (int isMultiProcess : testBed.ev.GetIsMultiProcessList())
    {
      int const numProcesses = isMultiProcess ? totalRanks : 1;
      const std::vector<int>& gpuPriorityOrder = testBed.ev.GetGpuPriorityOrder();
      testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks, gpuPriorityOrder));

      // Prepare AlltoAllV options
      std::vector<size_t> numInputElements;
      std::vector<size_t> numOutputElements;
      PrepareCounts(totalRanks, 256, options, numInputElements, numOutputElements, 60);

      for (int dataIdx = 0; dataIdx < dataTypes.size() && isCorrect; ++dataIdx)
      {
        if (testBed.ev.showNames)
        {
          std::string name = testBed.GetTestCaseName(totalRanks, isMultiProcess,
                                                     ncclCollAlltoAllv, dataTypes[dataIdx],
                                                     ncclSum, -1, inPlace, useManagedMem, useHipGraph);
          INFO("%s\n", name.c_str());
        }

        for (int rank = 0; rank < totalRanks; ++rank)
        {
          testBed.SetCollectiveArgs(ncclCollAlltoAllv,
                                    dataTypes[dataIdx],
                                    numInputElements[rank],
                                    numOutputElements[rank],
                                    options,
                                    -1,
                                    0,
                                    rank);
        }
        testBed.AllocateMem(inPlace, useManagedMem);
        testBed.PrepareData();
        testBed.ExecuteCollectives({}, useHipGraph);
        testBed.ValidateResults(isCorrect);
        testBed.DeallocateMem();
      }
      testBed.DestroyComms();
    }
    testBed.Finalize();
  }
}