Format files

eth-cscs · Sep 6, 2024 · a3d6493 · a3d6493
1 parent b9df191
commit a3d6493
Show file tree

Hide file tree

Showing 3 changed files with 33 additions and 13 deletions.
diff --git a/include/dlaf/memory/memory_chunk.h b/include/dlaf/memory/memory_chunk.h
@@ -27,11 +27,15 @@ namespace memory {
 
 namespace internal {
 umpire::Allocator& getUmpireHostAllocator();
-void initializeUmpireHostAllocator(std::size_t initial_bytes, std::size_t next_bytes, std::size_t alignment_bytes, double coalesce_free_ratio, double coalesce_reallocation_ratio);
+void initializeUmpireHostAllocator(std::size_t initial_bytes, std::size_t next_bytes,
+                                   std::size_t alignment_bytes, double coalesce_free_ratio,
+                                   double coalesce_reallocation_ratio);
 void finalizeUmpireHostAllocator();
 
 #ifdef DLAF_WITH_GPU
-void initializeUmpireDeviceAllocator(std::size_t initial_bytes, std::size_t next_bytes, std::size_t alignment_bytes, double coalesce_free_ratio, double coalesce_reallocation_ratio);
+void initializeUmpireDeviceAllocator(std::size_t initial_bytes, std::size_t next_bytes,
+                                     std::size_t alignment_bytes, double coalesce_free_ratio,
+                                     double coalesce_reallocation_ratio);
 void finalizeUmpireDeviceAllocator();
 umpire::Allocator& getUmpireDeviceAllocator();
 #endif

diff --git a/src/init.cpp b/src/init.cpp
@@ -64,7 +64,10 @@ struct Init {
 template <>
 struct Init<Backend::MC> {
   static void initialize(const configuration& cfg) {
-    memory::internal::initializeUmpireHostAllocator(cfg.umpire_host_memory_pool_initial_bytes, cfg.umpire_host_memory_pool_next_bytes, cfg.umpire_host_memory_pool_alignment_bytes, cfg.umpire_host_memory_pool_coalescing_free_ratio, cfg.umpire_host_memory_pool_coalescing_reallocation_ratio);
+    memory::internal::initializeUmpireHostAllocator(
+        cfg.umpire_host_memory_pool_initial_bytes, cfg.umpire_host_memory_pool_next_bytes,
+        cfg.umpire_host_memory_pool_alignment_bytes, cfg.umpire_host_memory_pool_coalescing_free_ratio,
+        cfg.umpire_host_memory_pool_coalescing_reallocation_ratio);
   }
 
   static void finalize() {
@@ -104,7 +107,10 @@ template <>
 struct Init<Backend::GPU> {
   static void initialize(const configuration& cfg) {
     const int device = 0;
-    memory::internal::initializeUmpireDeviceAllocator(cfg.umpire_device_memory_pool_initial_bytes, cfg.umpire_device_memory_pool_initial_bytes, cfg.umpire_device_memory_pool_alignment_bytes, cfg.umpire_host_memory_pool_coalescing_free_ratio, cfg.umpire_host_memory_pool_coalescing_reallocation_ratio);
+    memory::internal::initializeUmpireDeviceAllocator(
+        cfg.umpire_device_memory_pool_initial_bytes, cfg.umpire_device_memory_pool_initial_bytes,
+        cfg.umpire_device_memory_pool_alignment_bytes, cfg.umpire_host_memory_pool_coalescing_free_ratio,
+        cfg.umpire_host_memory_pool_coalescing_reallocation_ratio);
     initializeGpuPool(device, cfg.num_np_gpu_streams_per_thread, cfg.num_hp_gpu_streams_per_thread);
     pika::cuda::experimental::detail::register_polling(pika::resource::get_thread_pool("default"));
   }
@@ -195,6 +201,7 @@ void updateConfigurationValue(const pika::program_options::variables_map& vm, T&
 }
 
 void updateConfiguration(const pika::program_options::variables_map& vm, configuration& cfg) {
+  // clang-format off
   updateConfigurationValue(vm, cfg.num_np_gpu_streams_per_thread, "NUM_NP_GPU_STREAMS_PER_THREAD", "num-np-gpu-streams-per-thread");
   updateConfigurationValue(vm, cfg.num_hp_gpu_streams_per_thread, "NUM_HP_GPU_STREAMS_PER_THREAD", "num-hp-gpu-streams-per-thread");
   updateConfigurationValue(vm, cfg.umpire_host_memory_pool_initial_bytes, "UMPIRE_HOST_MEMORY_POOL_INITIAL_BYTES", "umpire-host-memory-pool-initial-bytes");
@@ -207,6 +214,7 @@ void updateConfiguration(const pika::program_options::variables_map& vm, configu
   updateConfigurationValue(vm, cfg.umpire_device_memory_pool_alignment_bytes, "UMPIRE_DEVICE_MEMORY_POOL_ALIGNMENT_BYTES", "umpire-device-memory-pool-alignment-bytes");
   updateConfigurationValue(vm, cfg.umpire_device_memory_pool_coalescing_free_ratio, "UMPIRE_DEVICE_MEMORY_POOL_COALESCING_FREE_RATIO", "umpire-device-memory-pool-coalescing-free-ratio");
   updateConfigurationValue(vm, cfg.umpire_device_memory_pool_coalescing_reallocation_ratio, "UMPIRE_DEVICE_MEMORY_POOL_COALESCING_REALLOCATION_RATIO", "umpire-device-memory-pool-coalescing-reallocation-ratio");
+  // clang-format on
   cfg.mpi_pool = (pika::resource::pool_exists("mpi")) ? "mpi" : "default";
 
   // Warn if not using MPI pool without --dlaf:no-mpi-pool

diff --git a/src/memory/memory_chunk.cpp b/src/memory/memory_chunk.cpp
@@ -60,18 +60,22 @@ using CoalesceHeuristicType = umpire::strategy::PoolCoalesceHeuristic<PoolType>;
 // this can lead to strange reallocations, so we simply avoid that case. Two or
 // more blocks are always coalesced to one block, so no reallocation will
 // happen immediately after coalescing two or more blocks.
-static CoalesceHeuristicType get_coalesce_heuristic(double coalesce_free_ratio, double coalesce_reallocation_ratio) {
+static CoalesceHeuristicType get_coalesce_heuristic(double coalesce_free_ratio,
+                                                    double coalesce_reallocation_ratio) {
   return [=](const PoolType& pool) {
     std::size_t threshold = static_cast<std::size_t>(coalesce_free_ratio * pool.getActualSize());
     if (pool.getReleasableBlocks() >= 2 && pool.getReleasableSize() >= threshold) {
       return static_cast<std::size_t>(coalesce_reallocation_ratio * pool.getActualSize());
-    } else {
+    }
+    else {
       return static_cast<std::size_t>(0);
     }
   };
 }
 
-void initializeUmpireHostAllocator(std::size_t initial_bytes, std::size_t next_bytes, std::size_t alignment_bytes, double coalesce_free_ratio, double coalesce_reallocation_ratio) {
+void initializeUmpireHostAllocator(std::size_t initial_bytes, std::size_t next_bytes,
+                                   std::size_t alignment_bytes, double coalesce_free_ratio,
+                                   double coalesce_reallocation_ratio) {
 #ifdef DLAF_WITH_GPU
   static bool initialized = false;
 
@@ -80,9 +84,9 @@ void initializeUmpireHostAllocator(std::size_t initial_bytes, std::size_t next_b
   if (!initialized) {
     auto host_allocator = umpire::ResourceManager::getInstance().getAllocator("PINNED");
     auto pooled_host_allocator =
-        umpire::ResourceManager::getInstance().makeAllocator<umpire::strategy::QuickPool>("DLAF_PINNED_pool",
-                                                                                          host_allocator,
-                                                                                          initial_bytes, next_bytes, alignment_bytes, get_coalesce_heuristic(coalesce_free_ratio, coalesce_reallocation_ratio));
+        umpire::ResourceManager::getInstance().makeAllocator<umpire::strategy::QuickPool>(
+            "DLAF_PINNED_pool", host_allocator, initial_bytes, next_bytes, alignment_bytes,
+            get_coalesce_heuristic(coalesce_free_ratio, coalesce_reallocation_ratio));
     auto thread_safe_pooled_host_allocator =
         umpire::ResourceManager::getInstance().makeAllocator<umpire::strategy::ThreadSafeAllocator>(
             "DLAF_PINNED_thread_safe_pool", pooled_host_allocator);
@@ -92,14 +96,17 @@ void initializeUmpireHostAllocator(std::size_t initial_bytes, std::size_t next_b
     initialized = true;
   }
 #else
-  dlaf::internal::silenceUnusedWarningFor(initial_bytes, next_bytes, alignment_bytes, coalesce_free_ratio, coalesce_reallocation_ratio);
+  dlaf::internal::silenceUnusedWarningFor(initial_bytes, next_bytes, alignment_bytes,
+                                          coalesce_free_ratio, coalesce_reallocation_ratio);
 #endif
 }
 
 void finalizeUmpireHostAllocator() {}
 
 #ifdef DLAF_WITH_GPU
-void initializeUmpireDeviceAllocator(std::size_t initial_bytes, std::size_t next_bytes, std::size_t alignment_bytes, double coalesce_free_ratio, double coalesce_reallocation_ratio) {
+void initializeUmpireDeviceAllocator(std::size_t initial_bytes, std::size_t next_bytes,
+                                     std::size_t alignment_bytes, double coalesce_free_ratio,
+                                     double coalesce_reallocation_ratio) {
   static bool initialized = false;
 
   // Umpire pools cannot be released, so we keep the pools around even when
@@ -108,7 +115,8 @@ void initializeUmpireDeviceAllocator(std::size_t initial_bytes, std::size_t next
     auto device_allocator = umpire::ResourceManager::getInstance().getAllocator("DEVICE");
     auto pooled_device_allocator =
         umpire::ResourceManager::getInstance().makeAllocator<umpire::strategy::QuickPool>(
-            "DLAF_DEVICE_pool", device_allocator, initial_bytes, next_bytes, alignment_bytes, get_coalesce_heuristic(coalesce_free_ratio, coalesce_reallocation_ratio));
+            "DLAF_DEVICE_pool", device_allocator, initial_bytes, next_bytes, alignment_bytes,
+            get_coalesce_heuristic(coalesce_free_ratio, coalesce_reallocation_ratio));
     auto thread_safe_pooled_device_allocator =
         umpire::ResourceManager::getInstance().makeAllocator<umpire::strategy::ThreadSafeAllocator>(
             "DLAF_DEVICE_thread_safe_pool", pooled_device_allocator);