https://github.com/Microsoft/CNTK
Tip revision: 9771dc4c66d7aa00b85f0cab5bd3122b71f837d7 authored by Eldar Akchurin on 29 August 2017, 17:28:23 UTC
Updating 1bitsgd
Updating 1bitsgd
Tip revision: 9771dc4
V2SimpleDistGradAggregator.h
//
// Copyright (c) Microsoft. All rights reserved.
// Copyright (c) 2017, NVIDIA CORPORATION. All rights reserved.
// Licensed under the MIT license. See LICENSE.md file in the project root for full license information.
//
#pragma once
#include <future>
#undef _SCL_SECURE_NO_WARNINGS
#include "CNTKLibrary.h"
#include "IDistGradAggregator.h"
#include "TimerUtility.h"
#include "MatrixQuantizerImpl.h"
#include "Utils.h"
#include "NcclComm.h"
namespace Microsoft { namespace MSR { namespace CNTK {
template <class ElemType>
class V2SimpleDistGradAggregator : public IDistGradAggregator<ElemType>
{
UsingIDistGradAggregatorMembers;
::CNTK::DistributedCommunicatorPtr m_communicator;
NcclComm m_nccl;
public:
V2SimpleDistGradAggregator(const MPIWrapperPtr& mpi, bool useAsyncAggregation, int deviceId, int syncStatsTrace, ::CNTK::DistributedCommunicatorPtr communicator)
: IDistGradAggregator<ElemType>(mpi), m_useAsyncAggregation(useAsyncAggregation), m_initialized(false), m_bufferedGradHeader(nullptr), m_syncStatsTrace(syncStatsTrace), m_iterationCount(0),
m_communicator(communicator), m_nccl(deviceId, mpi)
{}
~V2SimpleDistGradAggregator()
{
if (m_bufferedGradHeader != nullptr)
DistGradHeader::Destroy(m_bufferedGradHeader);
}
// Aggregate the gradient matrices across all nodes
bool AggregateGradients(const std::vector<Matrix<ElemType>*>& gradients, DistGradHeader* headerCPU, bool resetState) override
{
if (!IsInitialized())
Initialize(gradients, headerCPU->numEvalNode);
else if (resetState)
ResetState(gradients);
bool showSyncPerfStats = (m_syncStatsTrace > 0) && ((m_iterationCount % m_syncStatsTrace) == 0);
m_iterationCount++;
if (!m_useAsyncAggregation) // In case we do not use asyn aggregation, simply aggregate.
{
AggregateGradientsImpl(gradients, headerCPU, showSyncPerfStats);
return (headerCPU->numSamples != 0);
}
// If we are performing async gradient aggregation, let's wait for the pending gradient aggregation to finish
// then swap the contents of the buffered gradients and the new gradient matrices and fire an async aggreagation
// of the new gradient matrices
if (m_pendingAsyncAggregation.valid())
{
Timer aggregationTimer;
if (showSyncPerfStats)
aggregationTimer.Start();
m_pendingAsyncAggregation.get();
if (showSyncPerfStats)
{
aggregationTimer.Stop();
double gradientAggregationTime = aggregationTimer.ElapsedSeconds();
fprintf(stderr, "Async gradient aggregation wait time: %.6g\n", gradientAggregationTime);
}
}
std::vector<Matrix<ElemType>*> newGradients;
size_t numGradMatrices = gradients.size();
for (size_t i = 0; i < numGradMatrices; i++)
{
Matrix<ElemType>* bufferedGradientMatrix = m_bufferedGradients[gradients[i]].get();
if ((bufferedGradientMatrix == nullptr) ||
(bufferedGradientMatrix->GetNumCols() != gradients[i]->GetNumCols()) ||
(bufferedGradientMatrix->GetNumRows() != gradients[i]->GetNumRows()) ||
(bufferedGradientMatrix->GetDeviceId() != gradients[i]->GetDeviceId()))
{
LogicError("No buffered gradient matrix found corresponding to a gradient matrix to be aggregated!");
}
// Swap the gradient matrix contents with the buffered matrices
std::swap(*(gradients[i]), *bufferedGradientMatrix);
newGradients.push_back(bufferedGradientMatrix);
}
// Swap the grad header contents with the buffered grad header
swap(*headerCPU, *m_bufferedGradHeader);
// Initiate aggregation only if any samples were processed in previous iteration
if (resetState || (headerCPU->numSamples != 0))
{
int deviceId = gradients[0]->GetDeviceId();
DistGradHeader* newGradHeader = m_bufferedGradHeader;
MatrixComputeStreamEvent* mainStreamSyncEvent = MatrixComputeStreamEvent::Create(deviceId);
m_pendingAsyncAggregation = std::async(std::launch::async, [=] {
// We are starting on a new thread. Make sure the new thread is
// setup to use the right device
Matrix<ElemType>::SetDevice(deviceId);
// Synchronize the Quantization compute stream with the completion of
// compute of the gradient matrices on the main compute stream
mainStreamSyncEvent->SynchronizeDataTransferFetchStreamWithEvent<ElemType>();
delete mainStreamSyncEvent;
AggregateGradientsImpl(newGradients, newGradHeader, showSyncPerfStats);
});
return true;
}
return false;
}
private:
bool IsInitialized() const { return m_initialized; }
void Initialize(const std::vector<Matrix<ElemType>*>& gradients, int numEvalNodes)
{
int deviceId = gradients[0]->GetDeviceId();
for (size_t i = 0; i < gradients.size(); i++)
{
// Make sure none of the gradient matrixes are sparse - we currently do not support aggregation of sparse gradient matrices
if (gradients[i]->GetMatrixType() != DENSE)
RuntimeError("Gradient aggregation for sparse gradient matrices is currently unsupported!");
if (m_useAsyncAggregation)
m_bufferedGradients[gradients[i]].reset(new Matrix<ElemType>(gradients[i]->GetNumRows(), gradients[i]->GetNumCols(), deviceId));
}
if (m_useAsyncAggregation)
{
m_bufferedGradHeader = DistGradHeader::Create(numEvalNodes);
m_bufferedGradHeader->Clear();
}
m_initialized = true;
}
void ResetState(const std::vector<Matrix<ElemType>*>& gradients)
{
if (!m_useAsyncAggregation)
return;
// Make sure there is no pending async aggregation
if (m_pendingAsyncAggregation.valid())
LogicError("Unexpected pending async gradient aggregation found when resetting aggregator state!");
// Zero out the buffered gradients if resetting state
for (size_t i = 0; i < gradients.size(); i++)
m_bufferedGradients[gradients[i]]->SetValue(0);
m_bufferedGradHeader->Clear();
}
void AggregateGradientsImpl(const std::vector<Matrix<ElemType>*>& gradients, DistGradHeader* headerCPU, bool showSyncPerfStats)
{
Timer aggregationTimer;
int deviceId = gradients.front()->GetDeviceId();
if (showSyncPerfStats)
{
std::unique_ptr<MatrixComputeStreamEvent> mainStreamSyncEvent(MatrixComputeStreamEvent::Create(deviceId));
mainStreamSyncEvent->SynchronizeEvent();
aggregationTimer.Start();
}
if (headerCPU->numSamples == 0)
{
assert(headerCPU->criterion == 0.0);
assert(headerCPU->numSamplesWithLabel == 0);
for (int i = 0; i < headerCPU->numEvalNode; ++i)
assert(headerCPU->evalErrors[i].first == 0 && headerCPU->evalErrors[i].second == 0);
// If the current node did not process any samples, the gradients should be zero'd
for (size_t i = 0; i < gradients.size(); ++i)
gradients[i]->SetValue(0);
if (m_useAsyncAggregation)
{
std::unique_ptr<MatrixComputeStreamEvent> mainStreamSyncEvent(MatrixComputeStreamEvent::Create(deviceId));
mainStreamSyncEvent->SynchronizeDataTransferFetchStreamWithEvent<ElemType>();
}
}
// Prepare gradients.
std::vector<::CNTK::NDArrayViewPtr> valuesToAggregate;
if (m_nccl.IsSupported()) // nccl is only enabled if all ranks have net on GPUs.
{ // we assume in this case all grad layers are on the GPU too.
m_nccl.AllReduce(gradients);
}
else
{
for (size_t i = 0; i < gradients.size(); ++i)
{
if (gradients[i]->Data() == nullptr) // Hack in case of eval.
continue;
::CNTK::NDShape shape{ gradients[i]->GetNumElements() };
auto data = ::CNTK::MakeSharedObject<::CNTK::NDArrayView>(::CNTK::AsDataType<ElemType>(), shape, gradients[i]->Data(), gradients[i]->GetNumElements() * sizeof(ElemType), ::CNTK::AsDeviceDescriptor(gradients[i]->GetDeviceId()));
valuesToAggregate.push_back(data);
}
}
// Prepare header.
size_t numberOfElements = 1 + 1 + 1 + headerCPU->numEvalNode * 2;
std::unique_ptr<double[]> headerBuffer(new double[numberOfElements]);
headerBuffer[0] = headerCPU->criterion;
headerBuffer[1] = static_cast<double>(headerCPU->numSamples);
headerBuffer[2] = static_cast<double>(headerCPU->numSamplesWithLabel);
for (size_t i = 0; i < headerCPU->numEvalNode; ++i)
{
headerBuffer[3 + 2 * i] = headerCPU->evalErrors[i].first;
headerBuffer[3 + 2 * i + 1] = static_cast<double>(headerCPU->evalErrors[i].second);
}
auto headerData = ::CNTK::MakeSharedObject<::CNTK::NDArrayView>(::CNTK::DataType::Double, ::CNTK::NDShape{ numberOfElements }, headerBuffer.get(), numberOfElements * sizeof(double), ::CNTK::DeviceDescriptor::CPUDevice());
valuesToAggregate.push_back(headerData);
m_communicator->AggregateInPlace(valuesToAggregate, m_communicator->Workers());
if (m_nccl.IsSupported())
m_nccl.Sync();
// Copy data back to the header
headerCPU->criterion = headerBuffer[0];
headerCPU->numSamples = static_cast<size_t>(headerBuffer[1]);
headerCPU->numSamplesWithLabel = static_cast<size_t>(headerBuffer[2]);
for (size_t i = 0; i < headerCPU->numEvalNode; ++i)
{
headerCPU->evalErrors[i].first = headerBuffer[3 + 2 * i];
headerCPU->evalErrors[i].second = static_cast<size_t>(headerBuffer[3 + 2 * i + 1]);
}
if (showSyncPerfStats)
{
aggregationTimer.Stop();
double gradientAggregationTime = aggregationTimer.ElapsedSeconds();
fprintf(stderr, "Actual gradient aggregation time: %.6g\n", gradientAggregationTime);
}
}
private:
// Perform aysnchronous gradient aggregation using double buffering of the gradient matrices
bool m_useAsyncAggregation;
// Future corresponding to the current in-flight async gradient aggregation
std::future<void> m_pendingAsyncAggregation;
// Buffered gradients that we asynchronously aggregate
std::unordered_map<Matrix<ElemType>*, std::unique_ptr<Matrix<ElemType>>> m_bufferedGradients;
DistGradHeader* m_bufferedGradHeader;
// Only used for controlling frequency of measuring/showing gradient aggregation perf stats
int m_syncStatsTrace;
size_t m_iterationCount;
bool m_initialized;
};
}}}