NcclComm.cpp
//
// Copyright (c) 2016-2017, NVIDIA CORPORATION. All rights reserved.
// Licensed under the MIT license. See LICENSE.md file in the project root for full license information.
//
#include "NcclComm.h"
#ifdef USE_NCCL
#include "GPUMatrix.h"
#include <nccl.h>
#include <cuda_runtime.h>
namespace Microsoft { namespace MSR { namespace CNTK {
// allows to write cudaFunction() || "error" (CUDA runtime)
static void operator||(cudaError_t rc, const char *msg)
{
if (rc != cudaSuccess)
RuntimeError("%s: %s (cuda error %d)", msg, cudaGetErrorString(rc), (int) rc);
}
ncclRedOp_t ncclRedOpFromMpiOp(MPI_Op op)
{
if (op == MPI_SUM) return ncclSum;
else if (op == MPI_MAX) return ncclMax;
else if (op == MPI_MIN) return ncclMin;
else if (op == MPI_PROD) return ncclProd;
else RuntimeError("Invalid MPI_Op");
}
NcclComm::NcclComm(int deviceId, const MPIWrapperPtr& mpi)
: m_ncclComm(nullptr), m_stream(nullptr)
{
cudaDeviceSynchronize();
size_t numRanks = mpi->NumNodesInUse();
std::vector<int> allDevs(numRanks);
std::vector<std::array<char, MPI_MAX_PROCESSOR_NAME>> allHosts(numRanks);
std::array<char, MPI_MAX_PROCESSOR_NAME> procName {};
int nameLen;
MPI_Get_processor_name(procName.data(), &nameLen);
mpi->Allgather(&deviceId, 1, MPI_INT, allDevs.data(), 1, MPI_INT);
mpi->Allgather(procName.data(), MPI_MAX_PROCESSOR_NAME, MPI_CHAR, allHosts[0].data(), MPI_MAX_PROCESSOR_NAME, MPI_CHAR);
for (size_t r = 0; r < numRanks; r++)
{
if (allDevs[r] == CPUDEVICE)
{
fprintf(stderr, "NcclComm: disabled, at least one rank using CPU device\n");
return;
}
for (size_t s = 0; s < r; s++)
{
if (allHosts[r] == allHosts[s] && allDevs[r] == allDevs[s])
{
fprintf(stderr, "NcclComm: disabled, same device used by more than one rank\n");
return;
}
}
}
ncclUniqueId ncclId;
ncclResult_t res;
if (mpi->IsMainNode())
{
res = ncclGetUniqueId(&ncclId);
if (res != ncclSuccess)
RuntimeError("NcclComm failed to obtain ncclUniqueId: %s", ncclGetErrorString(res));
}
mpi->Bcast(&ncclId, NCCL_UNIQUE_ID_BYTES, MPI_CHAR, 0);
PrepareDevice(deviceId);
res = ncclCommInitRank(&m_ncclComm, numRanks, ncclId, mpi->CurrentNodeRank());
if (res != ncclSuccess)
RuntimeError("NcclComm failed to initialize: %s. Set the ENV \"NCCL_DEBUG=INFO\" for more information.", ncclGetErrorString(res));
cudaStreamCreateWithFlags(&m_stream, cudaStreamDefault)
|| "cudaStreamCreateWithFlags failed";
fprintf(stderr, "NcclComm: initialized\n");
}
NcclComm::~NcclComm()
{
if (m_stream != nullptr)
cudaStreamDestroy(m_stream);
if (m_ncclComm != nullptr)
ncclCommDestroy(m_ncclComm);
}
bool NcclComm::IsSupported()
{
return m_ncclComm != nullptr;
}
void NcclComm::AllReduceImpl(void* inputbuffer, void *outputbuffer, size_t count, DataType dtype, MPI_Op op)
{
ncclResult_t res;
class NcclTypeLookup
{
ncclDataType_t ncclTypes[(int)DataType::COUNT];
public:
NcclTypeLookup()
{
ncclTypes[(int)DataType::FLOAT] = ncclFloat;
ncclTypes[(int)DataType::DOUBLE] = ncclDouble;
ncclTypes[(int)DataType::INT] = ncclInt;
}
ncclDataType_t Lookup(DataType dtype)
{
return ncclTypes[(int)dtype];
}
};
static NcclTypeLookup s_ncclTypeLookup;
res = ncclAllReduce(inputbuffer, outputbuffer, count, s_ncclTypeLookup.Lookup(dtype), ncclRedOpFromMpiOp(op), m_ncclComm, m_stream);
if (res != ncclSuccess)
RuntimeError("NcclComm ncclAllReduce failed: %s", ncclGetErrorString(res));
}
void NcclComm::BroadcastImpl(void* buffer, size_t count, MPI_Datatype dtype, int root)
{
ncclResult_t res;
if (dtype == MPI_CHAR)
{
res = ncclBcast(buffer, count, ncclChar, root, m_ncclComm, m_stream);
}
else
{
RuntimeError("NcclComm Broadcast supports Char type only");
}
if (res != ncclSuccess)
{
RuntimeError("NcclComm ncclBcast failed: %s", ncclGetErrorString(res));
}
}
void NcclComm::Sync()
{
cudaStreamSynchronize(m_stream) || "NcclComm: cudaStreamSynchronize failed";
}
}}} // end namespaces
#else // !USE_NCCL
namespace Microsoft { namespace MSR { namespace CNTK {
NcclComm::NcclComm(int /*deviceId*/, const MPIWrapperPtr& /*mpi*/) { }
NcclComm::~NcclComm() { }
bool NcclComm::IsSupported()
{
return false;
}
void NcclComm::Sync() { }
}}} // end namespaces
#endif