Home | History | Annotate | Download | only in products
      1 // This file is part of Eigen, a lightweight C++ template library
      2 // for linear algebra.
      3 //
      4 // Copyright (C) 2010 Gael Guennebaud <gael.guennebaud (at) inria.fr>
      5 //
      6 // This Source Code Form is subject to the terms of the Mozilla
      7 // Public License v. 2.0. If a copy of the MPL was not distributed
      8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
      9 
     10 #ifndef EIGEN_PARALLELIZER_H
     11 #define EIGEN_PARALLELIZER_H
     12 
     13 namespace Eigen {
     14 
     15 namespace internal {
     16 
     17 /** \internal */
     18 inline void manage_multi_threading(Action action, int* v)
     19 {
     20   static EIGEN_UNUSED int m_maxThreads = -1;
     21 
     22   if(action==SetAction)
     23   {
     24     eigen_internal_assert(v!=0);
     25     m_maxThreads = *v;
     26   }
     27   else if(action==GetAction)
     28   {
     29     eigen_internal_assert(v!=0);
     30     #ifdef EIGEN_HAS_OPENMP
     31     if(m_maxThreads>0)
     32       *v = m_maxThreads;
     33     else
     34       *v = omp_get_max_threads();
     35     #else
     36     *v = 1;
     37     #endif
     38   }
     39   else
     40   {
     41     eigen_internal_assert(false);
     42   }
     43 }
     44 
     45 }
     46 
     47 /** Must be call first when calling Eigen from multiple threads */
     48 inline void initParallel()
     49 {
     50   int nbt;
     51   internal::manage_multi_threading(GetAction, &nbt);
     52   std::ptrdiff_t l1, l2, l3;
     53   internal::manage_caching_sizes(GetAction, &l1, &l2, &l3);
     54 }
     55 
     56 /** \returns the max number of threads reserved for Eigen
     57   * \sa setNbThreads */
     58 inline int nbThreads()
     59 {
     60   int ret;
     61   internal::manage_multi_threading(GetAction, &ret);
     62   return ret;
     63 }
     64 
     65 /** Sets the max number of threads reserved for Eigen
     66   * \sa nbThreads */
     67 inline void setNbThreads(int v)
     68 {
     69   internal::manage_multi_threading(SetAction, &v);
     70 }
     71 
     72 namespace internal {
     73 
     74 template<typename Index> struct GemmParallelInfo
     75 {
     76   GemmParallelInfo() : sync(-1), users(0), lhs_start(0), lhs_length(0) {}
     77 
     78   Index volatile sync;
     79   int volatile users;
     80 
     81   Index lhs_start;
     82   Index lhs_length;
     83 };
     84 
     85 template<bool Condition, typename Functor, typename Index>
     86 void parallelize_gemm(const Functor& func, Index rows, Index cols, Index depth, bool transpose)
     87 {
     88   // TODO when EIGEN_USE_BLAS is defined,
     89   // we should still enable OMP for other scalar types
     90 #if !(defined (EIGEN_HAS_OPENMP)) || defined (EIGEN_USE_BLAS)
     91   // FIXME the transpose variable is only needed to properly split
     92   // the matrix product when multithreading is enabled. This is a temporary
     93   // fix to support row-major destination matrices. This whole
     94   // parallelizer mechanism has to be redisigned anyway.
     95   EIGEN_UNUSED_VARIABLE(depth);
     96   EIGEN_UNUSED_VARIABLE(transpose);
     97   func(0,rows, 0,cols);
     98 #else
     99 
    100   // Dynamically check whether we should enable or disable OpenMP.
    101   // The conditions are:
    102   // - the max number of threads we can create is greater than 1
    103   // - we are not already in a parallel code
    104   // - the sizes are large enough
    105 
    106   // compute the maximal number of threads from the size of the product:
    107   // This first heuristic takes into account that the product kernel is fully optimized when working with nr columns at once.
    108   Index size = transpose ? rows : cols;
    109   Index pb_max_threads = std::max<Index>(1,size / Functor::Traits::nr);
    110 
    111   // compute the maximal number of threads from the total amount of work:
    112   double work = static_cast<double>(rows) * static_cast<double>(cols) *
    113       static_cast<double>(depth);
    114   double kMinTaskSize = 50000;  // FIXME improve this heuristic.
    115   pb_max_threads = std::max<Index>(1, std::min<Index>(pb_max_threads, work / kMinTaskSize));
    116 
    117   // compute the number of threads we are going to use
    118   Index threads = std::min<Index>(nbThreads(), pb_max_threads);
    119 
    120   // if multi-threading is explicitely disabled, not useful, or if we already are in a parallel session,
    121   // then abort multi-threading
    122   // FIXME omp_get_num_threads()>1 only works for openmp, what if the user does not use openmp?
    123   if((!Condition) || (threads==1) || (omp_get_num_threads()>1))
    124     return func(0,rows, 0,cols);
    125 
    126   Eigen::initParallel();
    127   func.initParallelSession(threads);
    128 
    129   if(transpose)
    130     std::swap(rows,cols);
    131 
    132   ei_declare_aligned_stack_constructed_variable(GemmParallelInfo<Index>,info,threads,0);
    133 
    134   #pragma omp parallel num_threads(threads)
    135   {
    136     Index i = omp_get_thread_num();
    137     // Note that the actual number of threads might be lower than the number of request ones.
    138     Index actual_threads = omp_get_num_threads();
    139 
    140     Index blockCols = (cols / actual_threads) & ~Index(0x3);
    141     Index blockRows = (rows / actual_threads);
    142     blockRows = (blockRows/Functor::Traits::mr)*Functor::Traits::mr;
    143 
    144     Index r0 = i*blockRows;
    145     Index actualBlockRows = (i+1==actual_threads) ? rows-r0 : blockRows;
    146 
    147     Index c0 = i*blockCols;
    148     Index actualBlockCols = (i+1==actual_threads) ? cols-c0 : blockCols;
    149 
    150     info[i].lhs_start = r0;
    151     info[i].lhs_length = actualBlockRows;
    152 
    153     if(transpose) func(c0, actualBlockCols, 0, rows, info);
    154     else          func(0, rows, c0, actualBlockCols, info);
    155   }
    156 #endif
    157 }
    158 
    159 } // end namespace internal
    160 
    161 } // end namespace Eigen
    162 
    163 #endif // EIGEN_PARALLELIZER_H
    164