Add divison strategies for for-each api.

lib/pls/include/pls/algorithms/for_each.h

@@ -5,9 +5,21 @@
 namespace pls {
 namespace pls {
 namespace algorithm {
 namespace algorithm {
 
 
+class fixed_strategy;
+class dynamic_strategy;
+
+template<typename Function, typename ExecutionStrategy>
+void for_each_range(unsigned long first,
+                    unsigned long last,
+                    const Function &function,
+                    ExecutionStrategy &execution_strategy);
+
 template<typename Function>
 template<typename Function>
 void for_each_range(unsigned long first, unsigned long last, const Function &function);
 void for_each_range(unsigned long first, unsigned long last, const Function &function);
 
 
+template<typename RandomIt, typename Function, typename ExecutionStrategy>
+void for_each(RandomIt first, RandomIt last, const Function &function, ExecutionStrategy execution_strategy);
+
 template<typename RandomIt, typename Function>
 template<typename RandomIt, typename Function>
 void for_each(RandomIt first, RandomIt last, const Function &function);
 void for_each(RandomIt first, RandomIt last, const Function &function);
 
 

lib/pls/include/pls/algorithms/for_each_impl.h

@@ -4,6 +4,7 @@
 
 
 #include "pls/internal/scheduling/task.h"
 #include "pls/internal/scheduling/task.h"
 #include "pls/internal/scheduling/scheduler.h"
 #include "pls/internal/scheduling/scheduler.h"
+#include "pls/internal/scheduling/thread_state.h"
 
 
 #include "pls/internal/helpers/unique_id.h"
 #include "pls/internal/helpers/unique_id.h"
 #include "pls/internal/helpers/range.h"
 #include "pls/internal/helpers/range.h"
@@ -13,11 +14,10 @@ namespace algorithm {
 namespace internal {
 namespace internal {
 
 
 template<typename RandomIt, typename Function>
 template<typename RandomIt, typename Function>
-void for_each(RandomIt first, RandomIt last, const Function &function) {
+void for_each(const RandomIt first, const RandomIt last, const Function function, const long min_elements) {
   using namespace ::pls::internal::scheduling;
   using namespace ::pls::internal::scheduling;
-  constexpr long min_elements = 1; // TODO: tune this value/allow for execution strategies
 
 
-  long num_elements = std::distance(first, last);
+  const long num_elements = std::distance(first, last);
   if (num_elements <= min_elements) {
   if (num_elements <= min_elements) {
     // calculate last elements in loop to avoid overhead
     // calculate last elements in loop to avoid overhead
     for (auto current = first; current != last; current++) {
     for (auto current = first; current != last; current++) {
@@ -25,15 +25,25 @@ void for_each(RandomIt first, RandomIt last, const Function &function) {
     }
     }
   } else {
   } else {
     // Cut in half recursively
     // Cut in half recursively
-    long middle_index = num_elements / 2;
+    const long middle_index = num_elements / 2;
 
 
     auto second_half_body =
     auto second_half_body =
-        [first, middle_index, last, &function] { internal::for_each(first + middle_index, last, function); };
+        [first, middle_index, last, &function, min_elements] {
+          internal::for_each(first + middle_index,
+                             last,
+                             function,
+                             min_elements);
+        };
     using second_half_t = lambda_task_by_reference<decltype(second_half_body)>;
     using second_half_t = lambda_task_by_reference<decltype(second_half_body)>;
     scheduler::spawn_child<second_half_t>(std::move(second_half_body));
     scheduler::spawn_child<second_half_t>(std::move(second_half_body));
 
 
     auto first_half_body =
     auto first_half_body =
-        [first, middle_index, last, &function] { internal::for_each(first, first + middle_index, function); };
+        [first, middle_index, last, &function, min_elements] {
+          internal::for_each(first,
+                             first + middle_index,
+                             function,
+                             min_elements);
+        };
     using first_half_t = lambda_task_by_reference<decltype(first_half_body)>;
     using first_half_t = lambda_task_by_reference<decltype(first_half_body)>;
     scheduler::spawn_child_and_wait<first_half_t>(std::move(first_half_body));
     scheduler::spawn_child_and_wait<first_half_t>(std::move(first_half_body));
   }
   }
@@ -41,15 +51,53 @@ void for_each(RandomIt first, RandomIt last, const Function &function) {
 
 
 }
 }
 
 
-template<typename Function>
-void for_each_range(unsigned long first, unsigned long last, const Function &function) {
-  auto range = pls::internal::helpers::range(first, last);
-  internal::for_each(range.begin(), range.end(), function);
+class dynamic_strategy {
+ public:
+  explicit dynamic_strategy(const unsigned int tasks_per_thread = 4) : tasks_per_thread_{tasks_per_thread} {};
+
+  long calculate_min_elements(long num_elements) const {
+    const long num_threads = pls::internal::scheduling::thread_state::get()->scheduler_->num_threads();
+    return num_elements / (num_threads * tasks_per_thread_);
+  }
+ private:
+  unsigned const int tasks_per_thread_;
+};
+
+class fixed_strategy {
+ public:
+  explicit fixed_strategy(const long min_elements_per_task) : min_elements_per_task_{min_elements_per_task} {};
+
+  long calculate_min_elements(long /*num_elements*/) const {
+    return min_elements_per_task_;
+  }
+ private:
+  const long min_elements_per_task_;
+};
+
+template<typename RandomIt, typename Function, typename ExecutionStrategy>
+void for_each(RandomIt first, RandomIt last, const Function &function, ExecutionStrategy execution_strategy) {
+  long num_elements = std::distance(first, last);
+  internal::for_each(first, last, function, execution_strategy.calculate_min_elements(num_elements));
 }
 }
 
 
 template<typename RandomIt, typename Function>
 template<typename RandomIt, typename Function>
 void for_each(RandomIt first, RandomIt last, const Function &function) {
 void for_each(RandomIt first, RandomIt last, const Function &function) {
-  internal::for_each(first, last, function);
+  for_each(first, last, function, dynamic_strategy{4});
+}
+
+template<typename Function, typename ExecutionStrategy>
+void for_each_range(unsigned long first,
+                    unsigned long last,
+                    const Function &function,
+                    ExecutionStrategy execution_strategy) {
+  auto range = pls::internal::helpers::range(first, last);
+  for_each(range.begin(), range.end(), function, execution_strategy);
+}
+
+template<typename Function>
+void for_each_range(unsigned long first, unsigned long last, const Function &function) {
+  auto range = pls::internal::helpers::range(first, last);
+  for_each(range.begin(), range.end(), function);
 }
 }
 
 
 }
 }

lib/pls/include/pls/algorithms/scan_impl.h

@@ -69,7 +69,7 @@ class scan_task : public pls::internal::scheduling::task {
       internal::serial_scan(chunk_start, chunk_end, chunk_output, op_, neutral_elem_);
       internal::serial_scan(chunk_start, chunk_end, chunk_output, op_, neutral_elem_);
       auto last_chunk_value = *(chunk_output + chunk_size - 1);
       auto last_chunk_value = *(chunk_output + chunk_size - 1);
       chunk_sums_[i] = last_chunk_value;
       chunk_sums_[i] = last_chunk_value;
-    });
+    }, fixed_strategy{1});
 
 
     // Calculate prefix sums of each chunks sum
     // Calculate prefix sums of each chunks sum
     // (effectively the prefix sum at the end of each chunk, then used to correct the following chunk).
     // (effectively the prefix sum at the end of each chunk, then used to correct the following chunk).
@@ -85,7 +85,7 @@ class scan_task : public pls::internal::scheduling::task {
       for (; chunk_start != chunk_end; chunk_start++) {
       for (; chunk_start != chunk_end; chunk_start++) {
         *chunk_start = op_(*chunk_start, chunk_sums_[i - 1]);
         *chunk_start = op_(*chunk_start, chunk_sums_[i - 1]);
       }
       }
-    });
+    }, fixed_strategy{1});
   }
   }
 };
 };