algorithms_cpp: using block size partitioner in Scan

8a46bcb3 · Tobias Fuchs · 9a3b71bb · 8a46bcb3 · 8a46bcb3
Commit 8a46bcb3 authored Feb 23, 2015 by Tobias Fuchs
Hide whitespace changes
Inline Side-by-side

Showing with 82 additions and 61 deletions

algorithms_cpp/include/embb/algorithms/internal/for_each-inl.h
+5 -7

algorithms_cpp/include/embb/algorithms/internal/scan-inl.h
+77 -54

No files found.
--- a/algorithms_cpp/include/embb/algorithms/internal/for_each-inl.h
+++ b/algorithms_cpp/include/embb/algorithms/internal/for_each-inl.h
@@ -50,7 +50,7 @@ class ForEachFunctor {
  ForEachFunctor(size_t chunk_first, size_t chunk_last, Function unary,
                 const embb::mtapi::ExecutionPolicy& policy, 
                 const BlockSizePartitioner<RAI>& partitioner)
-  : chunk_first_(chunk_first), chunk_last_(chunk_last), 
+  : chunk_first_(chunk_first), chunk_last_(chunk_last),
    unary_(unary), policy_(policy), partitioner_(partitioner) {
  }
@@ -76,13 +76,11 @@ class ForEachFunctor {
                       unary_, policy_, partitioner_);
      mtapi::Task task_l = mtapi::Node::GetInstance().Spawn(
        mtapi::Action(
-          base::MakeFunction(
+          base::MakeFunction(functor_l, &self_t::Action),
-          functor_l, &self_t::Action),
          policy_));
      mtapi::Task task_r = mtapi::Node::GetInstance().Spawn(
        mtapi::Action(
-          base::MakeFunction(
+          base::MakeFunction(functor_r, &self_t::Action),
-          functor_r, &self_t::Action),
          policy_));
      task_l.Wait(MTAPI_INFINITE);
      task_r.Wait(MTAPI_INFINITE);
@@ -120,8 +118,8 @@ void ForEachRecursive(RAI first, RAI last, Function unary,
  }
  // Perform check of task number sufficiency
  if (((distance / block_size) * 2) + 1 > MTAPI_NODE_MAX_TASKS_DEFAULT) {
-    EMBB_THROW(embb::base::ErrorException, "Not enough MTAPI tasks available "
+    EMBB_THROW(embb::base::ErrorException,
-               "to perform the parallel foreach loop");
+               "Not enough MTAPI tasks available for parallel foreach");
  }
  BlockSizePartitioner<RAI> partitioner(first, last, block_size);

--- a/algorithms_cpp/include/embb/algorithms/internal/scan-inl.h
+++ b/algorithms_cpp/include/embb/algorithms/internal/scan-inl.h
@@ -41,70 +41,86 @@ template<typename RAIIn, typename RAIOut, typename ReturnType,
 typename ScanFunction, typename TransformationFunction>
 class ScanFunctor {
 public:
-  ScanFunctor(RAIIn first, RAIIn last, RAIOut output_iterator,
+  ScanFunctor(size_t chunk_first, size_t chunk_last, RAIOut output_iterator,
              ReturnType neutral, ScanFunction scan,
              TransformationFunction transformation,
              const embb::mtapi::ExecutionPolicy& policy,
-              size_t block_size, ReturnType* tree_values, size_t node_id,
+              const BlockSizePartitioner<RAIIn>& partitioner, 
+              ReturnType* tree_values, size_t node_id,
              bool going_down)
-    : policy_(policy), first_(first), last_(last),
+    : policy_(policy), chunk_first_(chunk_first), chunk_last_(chunk_last),
      output_iterator_(output_iterator), scan_(scan),
      transformation_(transformation),
-      neutral_(neutral), block_size_(block_size), tree_values_(tree_values),
+      neutral_(neutral), partitioner_(partitioner), tree_values_(tree_values),
      node_id_(node_id), parent_value_(neutral), is_first_pass_(going_down)  {
  }
-  void Action(mtapi::TaskContext& context) {
+  void Action(mtapi::TaskContext&) {
-    if (first_ == last_) {
+    if (chunk_first_ == chunk_last_) {
-      return;
+      // leaf case -> do work      
-    }
-    size_t distance = static_cast<size_t>(std::distance(first_, last_));
-    if (distance <= block_size_) {  // leaf case -> do work
      if (is_first_pass_) {
-        RAIIn iter_in = first_;
+        ChunkDescriptor<RAIIn> chunk = partitioner_[chunk_first_];
+        RAIIn iter_in = chunk.GetFirst();
+        RAIIn last_in = chunk.GetLast();
        RAIOut iter_out = output_iterator_;
-        ReturnType result = transformation_(*first_);
+        ReturnType result = transformation_(*iter_in);
        *iter_out = result;
        ++iter_in;
        ++iter_out;
-        for (; iter_in != last_; ++iter_in, ++iter_out) {
+        for (; iter_in != last_in; ++iter_in, ++iter_out) {
          result = scan_(result, transformation_(*iter_in));
          *iter_out = result;
        }
        SetTreeValue(result);
-      } else { // Second pass
+      }
-        RAIIn iter_in = first_;
+      else {
+        // Second pass
+        ChunkDescriptor<RAIIn> chunk = partitioner_[chunk_first_];
+        RAIIn iter_in = chunk.GetFirst();
+        RAIIn last_in = chunk.GetLast();
        RAIOut iter_out = output_iterator_;
-        for (; iter_in != last_; ++iter_in, ++iter_out) {
+        for (; iter_in != last_in; ++iter_in, ++iter_out) {
          *iter_out = scan_(parent_value_, *iter_out);
        }
      }
-    } else {  // recurse further
+    }
-      internal::ChunkPartitioner<RAIIn> partitioner(first_, last_, 2);
+    else {
-      ScanFunctor functor_l(partitioner[0].GetFirst(), partitioner[0].GetLast(),
+      // recurse further
-                            output_iterator_, neutral_, scan_, transformation_,
+      size_t chunk_split_index = (chunk_first_ + chunk_last_) / 2;
-                            policy_, block_size_, tree_values_, node_id_,
+      // Split chunks into left / right branches:
-                            is_first_pass_);
+      ScanFunctor functor_l(
-      ScanFunctor functor_r(partitioner[1].GetFirst(), partitioner[1].GetLast(),
+        chunk_first_, chunk_split_index,
-                            output_iterator_, neutral_, scan_, transformation_,
+        output_iterator_, neutral_, scan_, transformation_,
-                            policy_, block_size_, tree_values_, node_id_,
+        policy_, partitioner_, tree_values_, node_id_,
-                            is_first_pass_);
+        is_first_pass_);
-      functor_l.SetID(1);
+      ScanFunctor functor_r(
-      functor_r.SetID(2);
+        chunk_split_index + 1, chunk_last_,
-      std::advance(functor_r.output_iterator_,
+        output_iterator_, neutral_, scan_, transformation_,
-                   std::distance(functor_l.first_, functor_r.first_));
+        policy_, partitioner_, tree_values_, node_id_,
+        is_first_pass_);
+      functor_l.SetID(LEFT);
+      functor_r.SetID(RIGHT);
+      // Advance output iterator of right branch:
+      ChunkDescriptor<RAIIn> chunk_left  = partitioner_[chunk_first_];
+      ChunkDescriptor<RAIIn> chunk_right = partitioner_[chunk_split_index + 1];
+      long long dist = std::distance(chunk_left.GetFirst(), chunk_right.GetFirst());
+      std::advance(functor_r.output_iterator_, dist);
      if (!is_first_pass_) {
        functor_l.parent_value_ = parent_value_;
        functor_r.parent_value_ = functor_l.GetTreeValue() + parent_value_;
      }
-      mtapi::Node& node = mtapi::Node::GetInstance();
+      // Spawn tasks to recurse:
-      // Spawn tasks for right partition first:
+      mtapi::Node& node = mtapi::Node::GetInstance();      
-      mtapi::Task task_r = node.Spawn(mtapi::Action(base::MakeFunction(
+      mtapi::Task task_l = node.Spawn(
-                                      functor_r, &ScanFunctor::Action),
+        mtapi::Action(
-                                      policy_));
+          base::MakeFunction(functor_l, &ScanFunctor::Action),
-      // Recurse on left partition:
+          policy_));
-      functor_l.Action(context);
+      mtapi::Task task_r = node.Spawn(
-      // Wait for tasks on right partition to complete:
+        mtapi::Action(
+          base::MakeFunction(functor_r, &ScanFunctor::Action),
+          policy_));
+      // Wait for tasks to complete:
+      task_l.Wait(MTAPI_INFINITE);
      task_r.Wait(MTAPI_INFINITE);
      SetTreeValue(scan_(functor_l.GetTreeValue(), functor_r.GetTreeValue()));
    }
@@ -119,23 +135,26 @@ class ScanFunctor {
  }
 private:
+  static const int LEFT  = 1;
+  static const int RIGHT = 2;
  const embb::mtapi::ExecutionPolicy& policy_;
-  RAIIn first_;
+  size_t chunk_first_;
-  RAIIn last_;
+  size_t chunk_last_;
  RAIOut output_iterator_;
  ScanFunction scan_;
  TransformationFunction transformation_;
  ReturnType neutral_;
-  size_t block_size_;
+  const BlockSizePartitioner<RAIIn>& partitioner_;
  ReturnType* tree_values_;
  size_t node_id_;
  ReturnType parent_value_;
  bool is_first_pass_;
-  void SetID(int is_left) {
+  void SetID(int branch) {
-    if (is_left == 1) {
+    if (branch == LEFT) {
      node_id_ = 2 * node_id_ + 1;
-    } else if (is_left == 2) {
+    } 
+    else if (branch == RIGHT) {
      node_id_ = 2 * node_id_ + 2;
    }
  }
@@ -166,23 +185,25 @@ void ScanIteratorCheck(RAIIn first, RAIIn last, RAIOut output_iterator,
  }
  mtapi::Node& node = mtapi::Node::GetInstance();
  ReturnType values[MTAPI_NODE_MAX_TASKS_DEFAULT];
-  size_t used_block_size = block_size;
  if (block_size == 0) {
-    used_block_size = static_cast<size_t>(distance) / node.GetCoreCount();
+    block_size = static_cast<size_t>(distance) / node.GetCoreCount();
-    if (used_block_size == 0) used_block_size = 1;
+    if (block_size == 0) {
+      block_size = 1;
+    }
  }
-  if (((distance / used_block_size) * 2) + 1 > MTAPI_NODE_MAX_TASKS_DEFAULT) {
+  if (((distance / block_size) * 2) + 1 > MTAPI_NODE_MAX_TASKS_DEFAULT) {
    EMBB_THROW(embb::base::ErrorException,
-               "Number of computation tasks required in scan "
+               "Not enough MTAPI tasks available for parallel scan");
-               "exceeds MTAPI maximum number of tasks");
  }
  // first pass. Calculates prefix sums for leaves and when recursion returns
  // it creates the tree.
  typedef ScanFunctor<RAIIn, RAIOut, ReturnType, ScanFunction,
                      TransformationFunction> Functor;
-  Functor functor_down(first, last, output_iterator, neutral, scan,
-                       transformation, policy, used_block_size, values, 0,
+  BlockSizePartitioner<RAIIn> partitioner_down(first, last, block_size);
+  Functor functor_down(0, partitioner_down.Size() - 1, output_iterator, neutral, 
+                       scan, transformation, policy, partitioner_down, values, 0,
                       true);
  mtapi::Task task_down = node.Spawn(mtapi::Action(base::MakeFunction(
                          functor_down, &Functor::Action),
@@ -190,8 +211,10 @@ void ScanIteratorCheck(RAIIn first, RAIIn last, RAIOut output_iterator,
  task_down.Wait(MTAPI_INFINITE);
  // Second pass. Gives to each leaf the part of the prefix missing
-  Functor functor_up(first, last, output_iterator, neutral, scan,
+  BlockSizePartitioner<RAIIn> partitioner_up(first, last, block_size);
-                     transformation, policy, used_block_size, values, 0, false);
+  Functor functor_up(0, partitioner_up.Size() - 1, output_iterator, neutral,
+                     scan, transformation, policy, partitioner_up, values, 0, 
+                     false);
  mtapi::Task task_up = node.Spawn(mtapi::Action(base::MakeFunction(
                        functor_up, &Functor::Action),
                        policy));