Marzieh Berenjkoub
348 baris
14 KiB
C++
348 baris
14 KiB
C++
/*************************************************************************
|
|
* Copyright (c) 2023 Advanced Micro Devices, Inc. All rights reserved.
|
|
*
|
|
* See LICENSE.txt for license information
|
|
************************************************************************/
|
|
#include "TestBed.hpp"
|
|
|
|
namespace RcclUnitTesting
|
|
{
|
|
// Test identical collectives within the same group call
|
|
TEST(GroupCall, Identical)
|
|
{
|
|
TestBed testBed;
|
|
|
|
// Configuration
|
|
std::vector<ncclFunc_t> const funcTypes = {ncclCollAllReduce, ncclCollAllReduce, ncclCollAllReduce};
|
|
std::vector<ncclRedOp_t> const testRedOps = {ncclSum, ncclSum, ncclSum};
|
|
std::vector<ncclDataType_t> const testDataTypes = {ncclFloat, ncclFloat, ncclFloat};
|
|
std::vector<int> const numElements = {1048576, 384 * 1024, 384};
|
|
|
|
int const numCollPerGroup = numElements.size();
|
|
bool const inPlace = false;
|
|
bool const useManagedMem = false;
|
|
|
|
std::vector<ncclDataType_t> dataTypes;
|
|
testBed.GetSupportedDataTypes(dataTypes, testDataTypes);
|
|
if (dataTypes.empty()) {
|
|
GTEST_SKIP() << "Skipping... test datatypes excluded by UT_DATATYPES.";
|
|
}
|
|
|
|
std::vector<ncclRedOp_t> redOps;
|
|
testBed.GetSupportedRedOps(redOps, testRedOps);
|
|
if (redOps.empty()) {
|
|
GTEST_SKIP() << "Skipping... test reduction operations excluded by UT_REDOPS.";
|
|
}
|
|
|
|
bool isCorrect = true;
|
|
for (int totalRanks : testBed.ev.GetNumGpusList())
|
|
for (int isMultiProcess : testBed.ev.GetIsMultiProcessList())
|
|
{
|
|
// Test either single process all GPUs, or 1 process per GPU
|
|
int const numProcesses = isMultiProcess ? totalRanks : 1;
|
|
const std::vector<int>& gpuPriorityOrder = testBed.ev.GetGpuPriorityOrder();
|
|
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks, gpuPriorityOrder), numCollPerGroup);
|
|
|
|
if (testBed.ev.showNames)
|
|
INFO("%s %d-ranks GroupCall Identical\n", isMultiProcess ? "MP" : "SP", totalRanks);
|
|
|
|
// Set up the different collectives within the group
|
|
for (int collIdx = 0; collIdx < numCollPerGroup; ++collIdx)
|
|
{
|
|
OptionalColArgs options;
|
|
options.redOp = redOps[collIdx];
|
|
testBed.SetCollectiveArgs(funcTypes[collIdx],
|
|
dataTypes[collIdx],
|
|
numElements[collIdx],
|
|
numElements[collIdx],
|
|
options,
|
|
collIdx);
|
|
}
|
|
|
|
testBed.AllocateMem(inPlace, useManagedMem);
|
|
testBed.PrepareData();
|
|
testBed.ExecuteCollectives();
|
|
testBed.ValidateResults(isCorrect);
|
|
testBed.DeallocateMem();
|
|
testBed.DestroyComms();
|
|
}
|
|
testBed.Finalize();
|
|
}
|
|
|
|
// Test different collectives within the same group call
|
|
TEST(GroupCall, Different)
|
|
{
|
|
TestBed testBed;
|
|
|
|
// Configuration
|
|
std::vector<ncclFunc_t> const funcTypes = {ncclCollBroadcast,
|
|
ncclCollAllGather,
|
|
ncclCollReduceScatter,
|
|
ncclCollAllReduce,
|
|
ncclCollGather,
|
|
ncclCollScatter,
|
|
ncclCollAlltoAll};
|
|
int const numCollPerGroup = funcTypes.size();
|
|
int const numElements = 1048576;
|
|
bool const inPlace = false;
|
|
bool const useManagedMem = false;
|
|
|
|
OptionalColArgs options;
|
|
options.redOp = ncclSum;
|
|
options.root = 0;
|
|
|
|
bool isCorrect = true;
|
|
for (int totalRanks : testBed.ev.GetNumGpusList())
|
|
for (int isMultiProcess : testBed.ev.GetIsMultiProcessList())
|
|
{
|
|
// Test either single process all GPUs, or 1 process per GPU
|
|
int const numProcesses = isMultiProcess ? totalRanks : 1;
|
|
const std::vector<int>& gpuPriorityOrder = testBed.ev.GetGpuPriorityOrder();
|
|
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks, gpuPriorityOrder), numCollPerGroup);
|
|
|
|
if (testBed.ev.showNames)
|
|
INFO("%s %d-ranks GroupCall Different\n", isMultiProcess ? "MP" : "SP", totalRanks);
|
|
|
|
// Set up the different collectives within the group
|
|
for (int collIdx = 0; collIdx < numCollPerGroup; ++collIdx)
|
|
{
|
|
int numInputElements;
|
|
int numOutputElements;
|
|
CollectiveArgs::GetNumElementsForFuncType(funcTypes[collIdx],
|
|
numElements,
|
|
totalRanks,
|
|
&numInputElements,
|
|
&numOutputElements);
|
|
|
|
testBed.SetCollectiveArgs(funcTypes[collIdx],
|
|
ncclFloat,
|
|
numInputElements,
|
|
numOutputElements,
|
|
options,
|
|
collIdx);
|
|
}
|
|
|
|
testBed.AllocateMem(inPlace, useManagedMem);
|
|
testBed.PrepareData();
|
|
testBed.ExecuteCollectives();
|
|
testBed.ValidateResults(isCorrect);
|
|
testBed.DeallocateMem();
|
|
testBed.DestroyComms();
|
|
}
|
|
testBed.Finalize();
|
|
}
|
|
|
|
// Test identical collectives with different data type
|
|
TEST(GroupCall, MixedDataType)
|
|
{
|
|
TestBed testBed;
|
|
|
|
// Configuration
|
|
std::vector<ncclFunc_t> const funcTypes = {ncclCollAllReduce, ncclCollAllReduce, ncclCollAllReduce};
|
|
std::vector<ncclRedOp_t> const testRedOps = {ncclSum, ncclSum, ncclSum};
|
|
std::vector<ncclDataType_t> const testDataTypes = {ncclFloat16, ncclFloat32, ncclFloat64};
|
|
std::vector<int> const numElements = {1048576, 384 * 1024, 384};
|
|
|
|
int const numCollPerGroup = numElements.size();
|
|
bool const inPlace = false;
|
|
bool const useManagedMem = false;
|
|
|
|
std::vector<ncclDataType_t> dataTypes;
|
|
testBed.GetSupportedDataTypes(dataTypes, testDataTypes);
|
|
if (dataTypes.empty()) {
|
|
GTEST_SKIP() << "Skipping... test datatypes excluded by UT_DATATYPES.";
|
|
}
|
|
|
|
std::vector<ncclRedOp_t> redOps;
|
|
testBed.GetSupportedRedOps(redOps, testRedOps);
|
|
if (redOps.empty()) {
|
|
GTEST_SKIP() << "Skipping... test reduction operations excluded by UT_REDOPS.";
|
|
}
|
|
|
|
bool isCorrect = true;
|
|
for (int totalRanks : testBed.ev.GetNumGpusList())
|
|
for (int isMultiProcess : testBed.ev.GetIsMultiProcessList())
|
|
{
|
|
// Test either single process all GPUs, or 1 process per GPU
|
|
int const numProcesses = isMultiProcess ? totalRanks : 1;
|
|
const std::vector<int>& gpuPriorityOrder = testBed.ev.GetGpuPriorityOrder();
|
|
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks, gpuPriorityOrder), numCollPerGroup);
|
|
|
|
if (testBed.ev.showNames)
|
|
INFO("%s %d-ranks GroupCall MixedDataType\n", isMultiProcess ? "MP" : "SP", totalRanks);
|
|
|
|
// Set up the different collectives within the group
|
|
for (int collIdx = 0; collIdx < numCollPerGroup; ++collIdx)
|
|
{
|
|
OptionalColArgs options;
|
|
options.redOp = redOps[collIdx];
|
|
testBed.SetCollectiveArgs(funcTypes[collIdx],
|
|
dataTypes[collIdx],
|
|
numElements[collIdx],
|
|
numElements[collIdx],
|
|
options,
|
|
collIdx);
|
|
}
|
|
|
|
testBed.AllocateMem(inPlace, useManagedMem);
|
|
testBed.PrepareData();
|
|
testBed.ExecuteCollectives();
|
|
testBed.ValidateResults(isCorrect);
|
|
testBed.DeallocateMem();
|
|
testBed.DestroyComms();
|
|
}
|
|
testBed.Finalize();
|
|
}
|
|
|
|
TEST(GroupCall, Multistream)
|
|
{
|
|
TestBed testBed;
|
|
|
|
// Configuration
|
|
int const numElements = 1048576;
|
|
bool const inPlace = false;
|
|
bool const useManagedMem = false;
|
|
|
|
OptionalColArgs options;
|
|
|
|
// This test runs multiple AllReduce collectives on different streams within the same group call
|
|
bool isCorrect = true;
|
|
for (int totalRanks : testBed.ev.GetNumGpusList())
|
|
for (int isMultiProcess : testBed.ev.GetIsMultiProcessList())
|
|
{
|
|
// Test either single process all GPUs, or 1 process per GPU
|
|
int const numProcesses = isMultiProcess ? totalRanks : 1;
|
|
|
|
for (int numCollPerGroup = 2; numCollPerGroup <= 6; numCollPerGroup += 2)
|
|
{
|
|
for (int numStreamsPerGroup = numCollPerGroup; numStreamsPerGroup >= 2; numStreamsPerGroup -= 3)
|
|
{
|
|
if (testBed.ev.showNames)
|
|
INFO("%s %d-ranks Multistream %d-Group Calls across %d streams\n",
|
|
isMultiProcess ? "MP" : "SP", totalRanks, numCollPerGroup, numStreamsPerGroup);
|
|
|
|
const std::vector<int>& gpuPriorityOrder = testBed.ev.GetGpuPriorityOrder();
|
|
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks, gpuPriorityOrder),
|
|
numCollPerGroup, numStreamsPerGroup);
|
|
|
|
// Set up each collective in group in different stream (modulo numStreamsPerGroup)
|
|
options.redOp = ncclSum;
|
|
for (int collIdx = 0; collIdx < numCollPerGroup; ++collIdx)
|
|
{
|
|
testBed.SetCollectiveArgs(ncclCollAllReduce, ncclFloat, numElements, numElements,
|
|
options, collIdx, 0, -1, collIdx % numStreamsPerGroup);
|
|
}
|
|
|
|
testBed.AllocateMem(inPlace, useManagedMem);
|
|
testBed.PrepareData();
|
|
testBed.ExecuteCollectives();
|
|
testBed.ValidateResults(isCorrect);
|
|
testBed.DeallocateMem();
|
|
testBed.DestroyComms();
|
|
}
|
|
}
|
|
}
|
|
testBed.Finalize();
|
|
}
|
|
|
|
TEST(GroupCall, MultiGroupCall)
|
|
{
|
|
TestBed testBed;
|
|
|
|
// Configuration
|
|
std::vector<std::vector<ncclFunc_t>> const groupCalls = {{ncclCollAllReduce, ncclCollAllGather},
|
|
{ncclCollAlltoAll, ncclCollGather},
|
|
{ncclCollBroadcast, ncclCollReduceScatter}};
|
|
std::vector<std::vector<int>> const numElements = {{1250, 1048576}, {384, 384 * 1024}, {1048576, 127}};
|
|
std::vector<ncclDataType_t> const testDataTypes = {ncclFloat16, ncclFloat32, ncclBfloat16};
|
|
std::vector<ncclRedOp_t> const testRedOps = {ncclSum, ncclProd, ncclMax};
|
|
std::vector<int> const numCollsPerGroup = {2, 2, 2};
|
|
std::vector<int> const numStreamsPerGroup = {1, 1, 1};
|
|
std::vector<bool> const useHipGraphList = {true, false, true};
|
|
bool const inPlace = false;
|
|
bool const useManagedMem = false;
|
|
bool const useBlocking = true;
|
|
int const numGroupCalls = groupCalls.size();
|
|
int const numIterations = 10;
|
|
|
|
std::vector<ncclDataType_t> dataTypes;
|
|
testBed.GetSupportedDataTypes(dataTypes, testDataTypes);
|
|
if (dataTypes.empty()) {
|
|
GTEST_SKIP() << "Skipping... test datatypes excluded by UT_DATATYPES.";
|
|
}
|
|
|
|
std::vector<ncclRedOp_t> redOps;
|
|
testBed.GetSupportedRedOps(redOps, testRedOps);
|
|
if (redOps.empty()) {
|
|
GTEST_SKIP() << "Skipping... test reduction operations excluded by UT_REDOPS.";
|
|
}
|
|
|
|
bool isCorrect = true;
|
|
for (int totalRanks : testBed.ev.GetNumGpusList())
|
|
for (int isMultiProcess : testBed.ev.GetIsMultiProcessList())
|
|
{
|
|
int const numProcesses = isMultiProcess ? totalRanks : 1;
|
|
|
|
// Initialize comms by specifying the # of group calls
|
|
const std::vector<int>& gpuPriorityOrder = testBed.ev.GetGpuPriorityOrder();
|
|
testBed.InitComms(TestBed::GetDeviceIdsList(numProcesses, totalRanks, gpuPriorityOrder), numCollsPerGroup, numStreamsPerGroup, numGroupCalls, useBlocking);
|
|
|
|
if (testBed.ev.showNames)
|
|
INFO("%s %d-ranks GroupCall MultiGroupCall\n", isMultiProcess ? "MP" : "SP", totalRanks);
|
|
|
|
for (int groupCallIdx = 0; groupCallIdx < groupCalls.size(); ++groupCallIdx)
|
|
{
|
|
std::vector<ncclFunc_t> funcTypes = groupCalls[groupCallIdx];
|
|
OptionalColArgs options;
|
|
options.redOp = redOps[groupCallIdx];
|
|
options.root = 0;
|
|
|
|
for (int collIdx = 0; collIdx < numCollsPerGroup[groupCallIdx]; ++collIdx)
|
|
{
|
|
int numInputElements;
|
|
int numOutputElements;
|
|
CollectiveArgs::GetNumElementsForFuncType(funcTypes[collIdx],
|
|
numElements[groupCallIdx][collIdx],
|
|
totalRanks,
|
|
&numInputElements,
|
|
&numOutputElements);
|
|
|
|
testBed.SetCollectiveArgs(funcTypes[collIdx],
|
|
dataTypes[groupCallIdx],
|
|
numInputElements,
|
|
numOutputElements,
|
|
options,
|
|
collIdx,
|
|
groupCallIdx);
|
|
}
|
|
|
|
testBed.AllocateMem(inPlace, useManagedMem, groupCallIdx);
|
|
testBed.PrepareData(groupCallIdx);
|
|
|
|
// Stream capture in advance for HIP graph enabled collective groups
|
|
if (useHipGraphList[groupCallIdx])
|
|
{
|
|
testBed.ExecuteCollectives({}, groupCallIdx, useHipGraphList[groupCallIdx]);
|
|
}
|
|
}
|
|
|
|
// Execute collectives based on groupIdx
|
|
for (int i = 0; i < numIterations; ++i)
|
|
{
|
|
// Select a random group call
|
|
int groupCallIdx = i % groupCalls.size();
|
|
|
|
// Use graphs if enabled otherwise execute the collective
|
|
if (useHipGraphList[groupCallIdx]) testBed.LaunchGraphs(groupCallIdx);
|
|
else testBed.ExecuteCollectives({}, groupCallIdx);
|
|
testBed.ValidateResults(isCorrect, groupCallIdx);
|
|
}
|
|
|
|
testBed.DeallocateMem();
|
|
testBed.DestroyGraphs();
|
|
testBed.DestroyComms();
|
|
}
|
|
testBed.Finalize();
|
|
}
|
|
}
|