Merge pull request #2575 from PDoakORNL/improved_walker_count_handling

Fix crowds input tag and implement solid walker count input handling in batched drivers

src/QMCDrivers/Crowd.cpp

@@ -2,7 +2,7 @@
 // This file is distributed under the University of Illinois/NCSA Open Source License.
 // See LICENSE file in top directory for details.
 //
-// Copyright (c) 2019 QMCPACK developers.
+// Copyright (c) 2020 QMCPACK developers.
 //
 // File developed by: Peter Doak, doakpw@ornl.gov, Oak Ridge National Laboratory
 //////////////////////////////////////////////////////////////////////////////////////
@@ -14,6 +14,7 @@ namespace qmcplusplus
 {
 void Crowd::clearWalkers()
 {
+  // We're clearing the refs to the objects not the referred to objects.
   mcp_walkers_.clear();
   mcp_wfbuffers_.clear();
   walker_elecs_.clear();

src/QMCDrivers/Crowd.h

@@ -5,8 +5,6 @@
 // Copyright (c) 2019 developers.
 //
 // File developed by: Peter Doak, doakpw@ornl.gov, Oak Ridge National Laboratory
-//
-// File refactored from VMC.h
 //////////////////////////////////////////////////////////////////////////////////////
 
 #ifndef QMCPLUSPLUS_CROWD_H

src/QMCDrivers/DMC/DMCBatched.cpp

@@ -2,7 +2,7 @@
 // This file is distributed under the University of Illinois/NCSA Open Source License.
 // See LICENSE file in top directory for details.
 //
-// Copyright (c) 2019 QMCPACK developers.
+// Copyright (c) 2020 QMCPACK developers.
 //
 // File developed by: Peter Doak, doakpw@ornl.gov, Oak Ridge National Laboratory
 //
@@ -28,7 +28,9 @@ using WP = WalkerProperties::Indexes;
 
 // clang-format off
 /** Constructor maintains proper ownership of input parameters
-   */
+ *
+ *  Note you must call the Base constructor before the derived class sets QMCType
+ */
 DMCBatched::DMCBatched(QMCDriverInput&& qmcdriver_input,
                        DMCDriverInput&& input,
                        MCPopulation& pop,
@@ -45,26 +47,6 @@ DMCBatched::DMCBatched(QMCDriverInput&& qmcdriver_input,
 }
 // clang-format on
 
-QMCDriverNew::AdjustedWalkerCounts DMCBatched::calcDefaultLocalWalkers(QMCDriverNew::AdjustedWalkerCounts awc) const
-{
-  checkNumCrowdsLTNumThreads();
-  int num_threads(Concurrency::maxThreads<>());
-
-  int orig_walkers_per_rank = awc.walkers_per_rank;
-
-  if (awc.num_crowds == 0)
-    awc.num_crowds = std::min(num_threads, awc.walkers_per_rank);
-  awc.walkers_per_crowd = (awc.walkers_per_rank % awc.num_crowds) ? awc.walkers_per_rank / awc.num_crowds + 1
-                                                                  : awc.walkers_per_rank / awc.num_crowds;
-
-  awc.walkers_per_rank = awc.walkers_per_crowd * awc.num_crowds;
-  if (awc.walkers_per_rank != orig_walkers_per_rank)
-    app_warning() << "DMCBatched driver has adjusted walkers per rank to: " << awc.walkers_per_rank << '\n';
-
-  app_log() << "DMCBatched walkers per crowd " << awc.walkers_per_crowd << std::endl;
-  return awc;
-}
-
 void DMCBatched::setNonLocalMoveHandler(QMCHamiltonian& golden_hamiltonian)
 {
   golden_hamiltonian.setNonLocalMoves(dmcdriver_input_.get_non_local_move(), qmcdriver_input_.get_tau(),
@@ -519,22 +501,10 @@ void DMCBatched::runDMCStep(int crowd_id,
 void DMCBatched::process(xmlNodePtr node)
 {
   QMCDriverNew::AdjustedWalkerCounts awc =
-      adjustGlobalWalkerCount(myComm, qmcdriver_input_.get_total_walkers(), qmcdriver_input_.get_walkers_per_rank(),
-                              dmcdriver_input_.get_reserve(), get_num_crowds());
-
-  walkers_per_rank_  = awc.walkers_per_rank;
-  walkers_per_crowd_ = awc.walkers_per_crowd;
-  num_crowds_        = awc.num_crowds;
-
-  app_log() << "DMCBatched Driver running with target_walkers=" << awc.global_walkers << '\n'
-            << "                               walkers_per_rank=" << walkers_per_rank_ << '\n'
-            << "                               num_crowds=" << num_crowds_ << '\n';
-
-  // side effect updates walkers_per_crowd_;
-  makeLocalWalkers(awc.walkers_per_rank, awc.reserve_walkers,
-                   ParticleAttrib<TinyVector<QMCTraits::RealType, 3>>(population_.get_num_particles()));
-  population_.syncWalkersPerNode(myComm);
-  Base::process(node);
+      adjustGlobalWalkerCount(myComm->size(), myComm->rank(), qmcdriver_input_.get_total_walkers(), qmcdriver_input_.get_walkers_per_rank(),
+                              dmcdriver_input_.get_reserve(), qmcdriver_input_.get_num_crowds());
+
+  Base::startup(node, awc);
 }
 
 bool DMCBatched::run()
@@ -553,11 +523,11 @@ bool DMCBatched::run()
 
   { // walker initialization
     ScopedTimer local_timer(&(timers_.init_walkers_timer));
-    TasksOneToOne<> section_start_task(num_crowds_);
+    TasksOneToOne<> section_start_task(crowds_.size());
     section_start_task(initialLogEvaluation, std::ref(crowds_), std::ref(step_contexts_));
   }
 
-  TasksOneToOne<> crowd_task(num_crowds_);
+  TasksOneToOne<> crowd_task(crowds_.size());
 
   for (int block = 0; block < num_blocks; ++block)
   {
@@ -582,7 +552,8 @@ bool DMCBatched::run()
 
       for (UPtr<Crowd>& crowd_ptr : crowds_)
         crowd_ptr->clearWalkers();
-      population_.distributeWalkers(crowds_.begin(), crowds_.end(), walkers_per_crowd_);
+
+      population_.distributeWalkers(crowds_);
 
       // Accumulate on the whole population
       // But it is now visible in the algorithm not hidden in the BranchEngine::branch.

src/QMCDrivers/DMC/DMCBatched.h

@@ -102,8 +102,6 @@ class DMCBatched : public QMCDriverNew
 
   void setNonLocalMoveHandler(QMCHamiltonian& golden_hamiltonian);
 
-  QMCDriverNew::AdjustedWalkerCounts calcDefaultLocalWalkers(QMCDriverNew::AdjustedWalkerCounts awc) const;
-
 private:
   DMCDriverInput dmcdriver_input_;
   /// Interval between branching

src/QMCDrivers/MCPopulation.h

@@ -25,7 +25,7 @@
 #include "OhmmsPETE/OhmmsVector.h"
 #include "QMCWaveFunctions/TrialWaveFunction.h"
 #include "QMCHamiltonians/QMCHamiltonian.h"
-
+#include "Utilities/FairDivide.h"
 namespace qmcplusplus
 {
 class MCPopulation
@@ -146,48 +146,18 @@ class MCPopulation
    *  walkers are distributed one by one to crowds.
    *
    */
-  template<typename ITER, typename = RequireInputIterator<ITER>>
-  void distributeWalkers(ITER it_group_start, ITER group_end, int walkers_per_group)
+  template<typename WTTV>
+  void distributeWalkers(WTTV& walker_consumer)
   {
-    auto it_group               = it_group_start;
-    auto it_walkers             = walkers_.begin();
-    auto it_walker_elecs        = walker_elec_particle_sets_.begin();
-    auto it_walker_twfs         = walker_trial_wavefunctions_.begin();
-    auto it_walker_hamiltonians = walker_hamiltonians_.begin();
-
-    assert((group_end - it_group < walkers_.size()) || walkers_per_group == 1);
-
-    // while (it_group != group_end)
-    // {
-    //   for (int i = 0; i < walkers_per_group; ++i)
-    //   {
-    //     // possible that walkers_all < walkers_per_group * group_size
-    //     if (it_walkers == walkers_.end())
-    //       break;
-    //     (**it_group).addWalker(**it_walkers, **it_walker_elecs, **it_walker_twfs, **it_walker_hamiltonians);
-    //     ++it_walkers;
-    //     ++it_walker_elecs;
-    //     ++it_walker_twfs;
-    //     ++it_walker_hamiltonians;
-    //   }
-    //   ++it_group;
-    // }
-
-    // For now ignore the requesting walkers_per_group
-
-    while (it_walkers != walkers_.end())
+    auto walkers_per_crowd = fairDivide(walkers_.size(), walker_consumer.size());
+
+    auto walker_index = 0;
+    for (int i = 0; i < walker_consumer.size(); ++i)
     {
-      it_group = it_group_start;
-      while (it_group != group_end)
+      for(int j = 0; j < walkers_per_crowd[i]; ++j)
       {
-        if (it_walkers == walkers_.end())
-          break;
-        (**it_group).addWalker(**it_walkers, **it_walker_elecs, **it_walker_twfs, **it_walker_hamiltonians);
-        ++it_walkers;
-        ++it_walker_elecs;
-        ++it_walker_twfs;
-        ++it_walker_hamiltonians;
-        ++it_group;
+        walker_consumer[i]->addWalker(*walkers_[walker_index], *walker_elec_particle_sets_[walker_index], *walker_trial_wavefunctions_[walker_index], *walker_hamiltonians_[walker_index]);
+        ++walker_index;
       }
     }
   }

src/QMCDrivers/QMCDriverInput.cpp

@@ -52,11 +52,10 @@ void QMCDriverInput::readXML(xmlNodePtr cur)
   parameter_set.add(warmup_steps_, "warmupsteps", "int");
   parameter_set.add(warmup_steps_, "warmup_steps", "int");
   parameter_set.add(num_crowds_, "crowds", "int");
-  parameter_set.add(walkers_per_rank_, "walkers", "int");
+  parameter_set.add(walkers_per_rank_, "walkers_per_rank", "int");
   parameter_set.add(total_walkers_, "total_walkers", "int");
   parameter_set.add(steps_between_samples_, "stepsbetweensamples", "int");
   parameter_set.add(samples_per_thread_, "samplesperthread", "real");
-  parameter_set.add(samples_per_thread_, "dmcwalkersperthread", "real");
   parameter_set.add(requested_samples_, "samples", "real");
   parameter_set.add(tau_, "timestep", "AU");
   parameter_set.add(tau_, "time_step", "AU");