Avoid allocating unused memory

doc-sources/reference/lincs.rst

@@ -865,10 +865,15 @@
 
                     Indexed by ``[model_index][boundary_index][criterion_index]``. The current rank of each low profile, for each model and criterion.
 
+                .. property:: high_profile_rank_indexes
+                    :type: list[unsigned]
+
+                    Indexed by ``[criterion_index]``. The index in ``high_profile_ranks``, for each single-peaked criterion.
+
                 .. property:: high_profile_ranks
                     :type: list[list[list[int]]]
 
-                    Indexed by ``[model_index][boundary_index][criterion_index]``. The current rank of each high profile, for each model and criterion.
+                    Indexed by ``[model_index][boundary_index][high_profile_rank_indexes[criterion_index]]``. The current rank of each high profile, for each model and single-peaked criterion.
 
                 .. property:: accuracies
                     :type: list[int]

doc-sources/reference/lincs.yml

@@ -360,6 +360,8 @@ children:
               type: list[list[int]]
             - name: low_profile_ranks
               type: list[list[list[int]]]
+            - name: high_profile_rank_indexes
+              type: list[unsigned]
             - name: high_profile_ranks
               type: list[list[list[int]]]
             - name: accuracies

integration-tests/dir-all/liblincs.txt

@@ -204,6 +204,7 @@ liblincs: module
             criteria_count: property
             get_best_accuracy: instancemethod
             get_best_model: instancemethod
+            high_profile_rank_indexes: property
             high_profile_ranks: property
             iteration_index: property
             low_profile_ranks: property

integration-tests/dir-all/lincs.txt

@@ -184,6 +184,7 @@ lincs: module
                 criteria_count: property
                 get_best_accuracy: instancemethod
                 get_best_model: instancemethod
+                high_profile_rank_indexes: property
                 high_profile_ranks: property
                 iteration_index: property
                 low_profile_ranks: property

lincs/liblincs/learning/mrsort-by-weights-profiles-breed.cpp

@@ -26,7 +26,8 @@ LearnMrsortByWeightsProfilesBreed::LearningData::LearningData(
   model_indexes(models_count),
   accuracies(models_count, zeroed),
   low_profile_ranks(models_count, boundaries_count, criteria_count, uninitialized),
-  high_profile_ranks(models_count, boundaries_count, criteria_count, uninitialized),
+  high_profile_rank_indexes(criteria_count, uninitialized),
+  high_profile_ranks(models_count, boundaries_count, count_single_peaked_criteria(), uninitialized),
   weights(models_count, criteria_count, uninitialized)
 {
   CHRONE();
@@ -36,6 +37,25 @@ LearnMrsortByWeightsProfilesBreed::LearningData::LearningData(
   for (unsigned model_index = 0; model_index != models_count; ++model_index) {
     random_generators[model_index].seed(random_seed * (model_index + 1));
   }
+
+  unsigned count = 0;
+  for (unsigned criterion_index = 0; criterion_index != criteria_count; ++criterion_index) {
+    if (single_peaked[criterion_index]) {
+      high_profile_rank_indexes[criterion_index] = count;
+      ++count;
+    }
+  }
+  assert(high_profile_ranks.s0() == count);
+}
+
+unsigned LearnMrsortByWeightsProfilesBreed::LearningData::count_single_peaked_criteria() const {
+  unsigned count = 0;
+  for (unsigned criterion_index = 0; criterion_index != criteria_count; ++criterion_index) {
+    if (single_peaked[criterion_index]) {
+      ++count;
+    }
+  }
+  return count;
 }
 
 Model LearnMrsortByWeightsProfilesBreed::LearningData::get_model(const unsigned model_index) const {
@@ -57,8 +77,8 @@ Model LearnMrsortByWeightsProfilesBreed::LearningData::get_model(const unsigned
     boundary_profile.reserve(criteria_count);
     for (unsigned criterion_index = 0; criterion_index != criteria_count; ++criterion_index) {
       const unsigned low_profile_rank = low_profile_ranks[model_index][boundary_index][criterion_index];
-      const unsigned high_profile_rank = high_profile_ranks[model_index][boundary_index][criterion_index];
       if (single_peaked[criterion_index]) {
+        const unsigned high_profile_rank = high_profile_ranks[model_index][boundary_index][high_profile_rank_indexes[criterion_index]];
         boundary_profile.push_back(std::make_pair(low_profile_rank, high_profile_rank));
       } else {
         boundary_profile.push_back(low_profile_rank);
@@ -176,7 +196,7 @@ bool LearnMrsortByWeightsProfilesBreed::is_accepted(
   const unsigned alternative_rank = learning_data.performance_ranks[criterion_index][alternative_index];
   const unsigned low_profile_rank = learning_data.low_profile_ranks[model_index][boundary_index][criterion_index];
   if (learning_data.single_peaked[criterion_index]) {
-    const unsigned high_profile_rank = learning_data.high_profile_ranks[model_index][boundary_index][criterion_index];
+    const unsigned high_profile_rank = learning_data.high_profile_ranks[model_index][boundary_index][learning_data.high_profile_rank_indexes[criterion_index]];
     return low_profile_rank <= alternative_rank && alternative_rank <= high_profile_rank;
   } else {
     return low_profile_rank <= alternative_rank;

lincs/liblincs/learning/mrsort-by-weights-profiles-breed.hpp

@@ -71,7 +71,8 @@ struct LearnMrsortByWeightsProfilesBreed::LearningData : public PreProcessedLear
   std::vector<unsigned> model_indexes;  // [model_index_index]: this is a reordering of the models' indexes
   Array1D<Host, unsigned> accuracies;  // [model_index]
   Array3D<Host, unsigned> low_profile_ranks;  // [model_index][boundary_index][criterion_index]
-  Array3D<Host, unsigned> high_profile_ranks;  // [model_index][boundary_index][criterion_index]
+  Array1D<Host, unsigned> high_profile_rank_indexes;  // [criterion_index], meaningful only for single-peaked criteria (i.e. when single_peaked[criterion_index] is true)
+  Array3D<Host, unsigned> high_profile_ranks;  // [model_index][boundary_index][high_profile_rank_indexes[criterion_index]]
   Array2D<Host, float> weights;  // [model_index][criterion_index]
   // @todo(Performance, later) Add models' ages
 
@@ -86,6 +87,9 @@ struct LearnMrsortByWeightsProfilesBreed::LearningData : public PreProcessedLear
   bool model_is_correct(unsigned model_index) const;
   bool models_are_correct() const;
   #endif
+
+ private:
+  unsigned count_single_peaked_criteria() const;
 };
 
 struct LearnMrsortByWeightsProfilesBreed::ProfilesInitializationStrategy {

lincs/liblincs/learning/mrsort-by-weights-profiles-breed/improve-profiles/accuracy-heuristic-on-cpu.cpp

@@ -57,6 +57,8 @@ void ImproveProfilesWithAccuracyHeuristicOnCpu::improve_low_profile_then_high_pr
   const unsigned boundary_index,
   const unsigned criterion_index
 ) {
+  assert(learning_data.single_peaked[criterion_index]);
+
   improve_low_profile(
     model_index,
     boundary_index,
@@ -65,7 +67,7 @@ void ImproveProfilesWithAccuracyHeuristicOnCpu::improve_low_profile_then_high_pr
       0 :
       learning_data.low_profile_ranks[model_index][boundary_index - 1][criterion_index],
     boundary_index == learning_data.boundaries_count - 1 ?
-      learning_data.high_profile_ranks[model_index][boundary_index][criterion_index]:
+      learning_data.high_profile_ranks[model_index][boundary_index][learning_data.high_profile_rank_indexes[criterion_index]]:
       learning_data.low_profile_ranks[model_index][boundary_index + 1][criterion_index]
   );
 
@@ -75,10 +77,10 @@ void ImproveProfilesWithAccuracyHeuristicOnCpu::improve_low_profile_then_high_pr
     criterion_index,
     boundary_index == learning_data.boundaries_count - 1 ?
       learning_data.low_profile_ranks[model_index][boundary_index][criterion_index] :
-      learning_data.high_profile_ranks[model_index][boundary_index + 1][criterion_index],
+      learning_data.high_profile_ranks[model_index][boundary_index + 1][learning_data.high_profile_rank_indexes[criterion_index]],
     boundary_index == 0 ?
       learning_data.values_counts[criterion_index] - 1 :
-      learning_data.high_profile_ranks[model_index][boundary_index - 1][criterion_index]
+      learning_data.high_profile_ranks[model_index][boundary_index - 1][learning_data.high_profile_rank_indexes[criterion_index]]
   );
 }
 
@@ -307,11 +309,12 @@ void ImproveProfilesWithAccuracyHeuristicOnCpu::improve_high_profile(
   const unsigned lowest_destination_rank,
   const unsigned highest_destination_rank
 ) {
+  assert(learning_data.single_peaked[criterion_index]);
   assert(lowest_destination_rank <= highest_destination_rank);
   if (lowest_destination_rank == highest_destination_rank) {
-    assert(learning_data.high_profile_ranks[model_index][boundary_index][criterion_index] == lowest_destination_rank);
+    assert(learning_data.high_profile_ranks[model_index][boundary_index][learning_data.high_profile_rank_indexes[criterion_index]] == lowest_destination_rank);
   } else {
-    unsigned best_destination_rank = learning_data.high_profile_ranks[model_index][boundary_index][criterion_index];
+    unsigned best_destination_rank = learning_data.high_profile_ranks[model_index][boundary_index][learning_data.high_profile_rank_indexes[criterion_index]];
     float best_desirability = Desirability().value();
 
     if (highest_destination_rank - lowest_destination_rank >= max_destinations_count) {
@@ -342,7 +345,7 @@ void ImproveProfilesWithAccuracyHeuristicOnCpu::improve_high_profile(
     }
 
     if (std::uniform_real_distribution<float>(0, 1)(learning_data.random_generators[model_index]) <= best_desirability) {
-      learning_data.high_profile_ranks[model_index][boundary_index][criterion_index] = best_destination_rank;
+      learning_data.high_profile_ranks[model_index][boundary_index][learning_data.high_profile_rank_indexes[criterion_index]] = best_destination_rank;
     }
   }
 }
@@ -376,7 +379,8 @@ void ImproveProfilesWithAccuracyHeuristicOnCpu::update_move_desirability_for_hig
   const unsigned alternative_index,
   Desirability* desirability
 ) {
-  const unsigned current_rank = learning_data.high_profile_ranks[model_index][boundary_index][criterion_index];
+  assert(learning_data.single_peaked[criterion_index]);
+  const unsigned current_rank = learning_data.high_profile_ranks[model_index][boundary_index][learning_data.high_profile_rank_indexes[criterion_index]];
   const float weight = learning_data.weights[model_index][criterion_index];
 
   const unsigned alternative_rank = learning_data.performance_ranks[criterion_index][alternative_index];