kfac_util.hpp

Go to the documentation of this file.

// Copyright (c) 2014-2023, Lawrence Livermore National Security, LLC.
// Produced at the Lawrence Livermore National Laboratory.
// Written by the LBANN Research Team (B. Van Essen, et al.) listed in
// the CONTRIBUTORS file. <lbann-dev@llnl.gov>
//
// LLNL-CODE-697807.
// All rights reserved.
//
// This file is part of LBANN: Livermore Big Artificial Neural Network
// Toolkit. For details, see http://software.llnl.gov/LBANN or
// https://github.com/LLNL/LBANN.
//
// Licensed under the Apache License, Version 2.0 (the "Licensee"); you
// may not use this file except in compliance with the License.  You may
// obtain a copy of the License at:
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
// implied. See the License for the specific language governing
// permissions and limitations under the license.

#ifndef LBANN_EXECUTION_ALGORITHMS_KFAC_KFAC_UTIL_HPP_INCLUDED
#define LBANN_EXECUTION_ALGORITHMS_KFAC_KFAC_UTIL_HPP_INCLUDED

#include "lbann/base.hpp"
#include "lbann/execution_algorithms/kfac/kfac_block.hpp"
#include "lbann/layers/learning/convolution.hpp"
#include "lbann/layers/learning/fully_connected.hpp"
#include "lbann/layers/learning/gru.hpp"
#include "lbann/layers/regularizers/batch_normalization.hpp"

// Forward declarations
namespace lbann {
class KFAC;
template <El::Device Device>
class kfac_block;
} // namespace lbann

namespace lbann {

namespace kfac {

#if defined AL_HAS_NCCL
using BackendT = ::Al::NCCLBackend;
#elif defined AL_HAS_HOST_TRANSFER
using BackendT = ::Al::HostTransferBackend;
#else
using BackendT = ::Al::MPIBackend;
#endif

using ReqT = typename BackendT::req_type;

enum class kfac_inverse_strategy
{
  ALL,  // Apply round-robin assingment to all of the layers. may cause load
        // imbalance.
  EACH, // Apply round-robin assingment to every type of layers. may
  // not work well for small networks.
  ROOT, // Use only the root GPU. This is only for testing.
};

enum class kfac_reduce_scatter_mode
{
  ALLREDUCE,      // Use lbann_comm::allreduce
  REDUCE_SCATTER, // Use El::ReduceScatter
  REDUCE,         // Use El::Reduce for each block
};

enum class kfac_allgather_mode
{
  ALLREDUCE, // Use lbann_comm::allreduce
  ALLGATHER, // Use El::ReduceScatter
  BROADCAST  // Use El::Broadcast for each block
};

template <El::Device Device>
void get_matrix_inverse(El::AbstractMatrix<DataType>& Ainv,
                        El::AbstractMatrix<DataType>& Linv,
                        const El::AbstractMatrix<DataType>& A,
                        bool report_time,
                        DataType damping,
                        DataType damping_bn_err,
                        bool is_bn,
                        const El::SyncInfo<Device>& sync_info);

template <El::Device Device>
void get_matrix_inverse_eigen(El::AbstractMatrix<DataType>& Ainv,
                              El::AbstractMatrix<DataType>& Linv,
                              const El::AbstractMatrix<DataType>& A,
                              bool report_time,
                              DataType damping,
                              DataType damping_bn_err,
                              bool is_bn,
                              const El::SyncInfo<Device>& sync_info);

template <El::Device Device>
std::string get_matrix_stat(const El::Matrix<DataType, Device>& X,
                            const char* name);

template <El::Device Device>
void allreduce_lower_tri(El::AbstractMatrix<DataType>& A,
                         El::AbstractMatrix<DataType>& AL,
                         lbann_comm* comm,
                         const El::SyncInfo<Device>& sync_info);

bool is_reduce_scatter_buffer_required(kfac_reduce_scatter_mode mode);

template <El::Device Device>
void reduce_scatter_blocks(
  const std::vector<std::pair<size_t, El::AbstractMatrix<DataType>*>>& blocks,
  El::Matrix<DataType, Device>& global_buffer,
  lbann_comm* comm,
  kfac_reduce_scatter_mode mode);

std::pair<bool, bool> is_allgather_buffer_required(kfac_allgather_mode mode);

template <El::Device Device>
void allgather_blocks(
  const std::vector<std::pair<size_t, El::AbstractMatrix<DataType>*>>& blocks,
  El::Matrix<DataType, Device>& send_buffer,
  El::Matrix<DataType, Device>& recv_buffer,
  lbann_comm* comm,
  kfac_allgather_mode mode);

template <El::Device Device>
void allgather_inverse_matrices(
  const std::vector<std::shared_ptr<kfac_block<Device>>>& blocks,
  El::Matrix<DataType, Device>& global_buffer,
  lbann_comm* comm);

template <El::Device Device>
void allgather_inverse_matrices_sizes(
  const std::vector<std::shared_ptr<kfac_block<Device>>>& blocks,
  El::Matrix<double, El::Device::CPU>& global_buffer,
  lbann_comm* comm);

template <El::Device Device>
void add_to_diagonal(El::Matrix<DataType, Device>& A,
                     DataType value,
                     DataType value_bn_err,
                     bool is_bn,
                     const El::SyncInfo<Device>& sync_info);

template <El::Device Device>
void make_diagonal(El::Matrix<DataType, Device>& A,
                   El::Matrix<DataType, Device>& B,
                   DataType value,
                   DataType value_bn_err,
                   bool is_bn,
                   const El::SyncInfo<Device>& sync_info);

template <El::Device Device>
void get_matrix_entrywise_inverse(El::Matrix<DataType, Device>& input,
                                  El::Matrix<DataType, Device>& output,
                                  const El::SyncInfo<Device>& sync_info);

template <El::Device Device>
void fill_upper_tri(El::Matrix<DataType, Device>& A,
                    const El::SyncInfo<Device>& sync_info);

template <El::Device Device>
void update_kronecker_average(El::Matrix<DataType, Device>& Aave,
                              const El::Matrix<DataType, Device>& A,
                              size_t count,
                              double decay,
                              const El::SyncInfo<Device>& sync_info);

template <El::Device Device>
void identity(El::Matrix<DataType, Device>& A,
              const El::SyncInfo<Device>& sync_info);

template <El::Device Device>
void pack_lower_tri(El::Matrix<DataType, Device>& L,
                    const El::Matrix<DataType, Device>& A,
                    const El::SyncInfo<Device>& sync_info);

template <El::Device Device>
void unpack_lower_tri(El::Matrix<DataType, Device>& A,
                      const El::Matrix<DataType, Device>& L,
                      const El::SyncInfo<Device>& sync_info);

template <typename T, El::Device Device>
void TranslateBetweenGridsVCAsync(
  const El::DistMatrix<T, El::STAR, El::VC, El::ELEMENT, Device>& A,
  El::DistMatrix<T, El::STAR, El::VC, El::ELEMENT, Device>& B,
  El::DistMatrix<T, El::STAR, El::VC, El::ELEMENT, Device>& subset,
  std::vector<ReqT>& Requests);

template <typename T, El::Device Device>
void TranslateBetweenGridsVCAsyncDirect(
  const El::DistMatrix<T, El::STAR, El::VC, El::ELEMENT, Device>& A,
  El::DistMatrix<T, El::STAR, El::VC, El::ELEMENT, Device>& B,
  El::Int featureSize,
  El::Int currentBatchSize,
  std::vector<ReqT>& Requests);

template <typename T, El::Device Device>
void TranslateBetweenGridsSTARAsync(
  const El::DistMatrix<T, El::STAR, El::STAR, El::ELEMENT, Device>& A,
  El::DistMatrix<T, El::STAR, El::STAR, El::ELEMENT, Device>& B,
  std::vector<ReqT>& Requests);

template <typename T, El::Device Device>
void TranslateBetweenGridsKFACAsync(
  const El::DistMatrix<T, El::STAR, El::VC, El::ELEMENT, Device>& A,
  El::DistMatrix<T, El::STAR, El::VC, El::ELEMENT, Device>& B,
  std::vector<ReqT>& Requests);

template <typename T, El::Device Device>
void TranslateBetweenGridsVC(
  El::DistMatrix<T, El::STAR, El::VC, El::ELEMENT, Device> const& A,
  El::DistMatrix<T, El::STAR, El::VC, El::ELEMENT, Device>& B);

} // namespace kfac
} // namespace lbann

#endif // LBANN_EXECUTION_ALGORITHMS_KFAC_KFAC_UTIL_HPP_INCLUDED