From b78a82cd6443ab41d846f1920748008709f646a1 Mon Sep 17 00:00:00 2001
From: Stephane Del Pino <stephane.delpino44@gmail.com>
Date: Wed, 27 Feb 2019 18:30:53 +0100
Subject: [PATCH] Use few parallel_for instead of standard for
This is mainly cosmetic. Performances issue are related to the recalculation of
synchronization info.
---
src/mesh/ItemValueSynchronizer.hpp | 49 ++++++++++++++++++------------
1 file changed, 30 insertions(+), 19 deletions(-)
diff --git a/src/mesh/ItemValueSynchronizer.hpp b/src/mesh/ItemValueSynchronizer.hpp
index 4a2c9939d..6e73d5a2c 100644
--- a/src/mesh/ItemValueSynchronizer.hpp
+++ b/src/mesh/ItemValueSynchronizer.hpp
@@ -4,7 +4,7 @@
#include <ItemValue.hpp>
#include <Connectivity.hpp>
-#include <unordered_map>
+#include <map>
class ItemValueSynchronizer
{
@@ -18,15 +18,26 @@ class ItemValueSynchronizer
{
static_assert(not std::is_abstract_v<ConnectivityType>,
"_synchronize must be called on a concrete connectivity");
- const auto& item_owner = connectivity.template owner<item_type>();
using ItemId = ItemIdT<item_type>;
- std::vector<std::vector<ItemId>> ghost_items_per_proc(parallel::size());
- for (ItemId item_id=0; item_id<item_value.size(); ++item_id) {
- if (const size_t owner = item_owner[item_id]; owner != parallel::rank()) {
- ghost_items_per_proc[owner].emplace_back(item_id);
- }
- }
+
+ const auto& item_owner = connectivity.template owner<item_type>();
+
+ std::vector<Array<const ItemId>> ghost_items_per_proc
+ = [&] () {
+ std::vector<std::vector<ItemId>> ghost_items_vector_per_proc(parallel::size());
+ for (ItemId item_id=0; item_id<item_value.size(); ++item_id) {
+ if (const size_t owner = item_owner[item_id]; owner != parallel::rank()) {
+ ghost_items_vector_per_proc[owner].emplace_back(item_id);
+ }
+ }
+ std::vector<Array<const ItemId>> ghost_items_per_proc(parallel::size());
+ for (size_t i_rank=0; i_rank<parallel::size(); ++i_rank) {
+ const auto& ghost_items_vector = ghost_items_vector_per_proc[i_rank];
+ ghost_items_per_proc[i_rank] = convert_to_array(ghost_items_vector);
+ }
+ return ghost_items_per_proc;
+ }();
Array<unsigned int> local_number_of_requested_values(parallel::size());
for (size_t i_rank=0; i_rank<parallel::size(); ++i_rank) {
@@ -41,9 +52,9 @@ class ItemValueSynchronizer
for (size_t i_rank=0; i_rank<parallel::size(); ++i_rank) {
const auto& ghost_items = ghost_items_per_proc[i_rank];
Array<int> item_number_list(ghost_items.size());
- for (size_t i_item = 0; i_item<ghost_items.size(); ++i_item) {
- item_number_list[i_item] = item_number[ghost_items[i_item]];
- }
+ parallel_for (ghost_items.size(), PASTIS_LAMBDA(size_t i_item) {
+ item_number_list[i_item] = item_number[ghost_items[i_item]];
+ });
requested_item_number_list_by_proc[i_rank] = item_number_list;
}
@@ -54,7 +65,7 @@ class ItemValueSynchronizer
parallel::exchange(requested_item_number_list_by_proc, to_send_item_number_list_by_proc);
- std::unordered_map<int, ItemId> item_number_to_id_correspondance(connectivity.template numberOf<item_type>());
+ std::map<int, ItemId> item_number_to_id_correspondance;
for (ItemId item_id=0; item_id<item_number.size(); ++item_id) {
item_number_to_id_correspondance[item_number[item_id]] = item_id;
}
@@ -64,7 +75,7 @@ class ItemValueSynchronizer
Array<ItemId> to_send_item_id{local_number_of_values_to_send[i_rank]};
const Array<int>& to_send_item_number = to_send_item_number_list_by_proc[i_rank];
for (size_t i=0; i<to_send_item_number.size(); ++i) {
- to_send_item_id[i] = item_number_to_id_correspondance[to_send_item_number[i]];
+ to_send_item_id[i] = item_number_to_id_correspondance.find(to_send_item_number[i])->second;
}
to_send_item_id_list_by_proc[i_rank] = to_send_item_id;
}
@@ -73,9 +84,9 @@ class ItemValueSynchronizer
for (size_t i_rank=0; i_rank<parallel::size(); ++i_rank) {
Array<DataType> to_send_data{local_number_of_values_to_send[i_rank]};
const Array<const ItemId>& to_send_item_id = to_send_item_id_list_by_proc[i_rank];
- for (size_t i=0; i<to_send_item_id.size(); ++i) {
- to_send_data[i] = item_value[to_send_item_id[i]];
- }
+ parallel_for(to_send_item_id.size(), PASTIS_LAMBDA(size_t i) {
+ to_send_data[i] = item_value[to_send_item_id[i]];
+ });
to_send_data_by_proc[i_rank] = to_send_data;
}
@@ -90,9 +101,9 @@ class ItemValueSynchronizer
for (size_t i_rank=0; i_rank<parallel::size(); ++i_rank) {
const auto& ghost_items = ghost_items_per_proc[i_rank];
const auto& requested_data = requested_data_list_by_proc[i_rank];
- for (size_t i=0; i<ghost_items.size(); ++i) {
- item_value[ghost_items[i]] = requested_data[i];
- }
+ parallel_for(ghost_items.size(), PASTIS_LAMBDA(size_t i) {
+ item_value[ghost_items[i]] = requested_data[i];
+ });
requested_data_list_by_proc[i_rank] = Array<DataType>{ghost_items.size()};
}
}
--
GitLab