Enable all reduce P2P add through iteration & fix mem leak & refacter…

… the mem using to decrease the GPU mem using Signed-off-by: Zhai, Xuejun <xuejun.zhai@intel.com>
zhaixuejun1993 · Oct 17, 2024 · 25de131 · 25de131
1 parent 66f88b5
commit 25de131
Show file tree

Hide file tree

Showing 3 changed files with 1,331 additions and 142 deletions.
diff --git a/src/plugins/intel_gpu/include/intel_gpu/plugin/sub_memory_manager.hpp b/src/plugins/intel_gpu/include/intel_gpu/plugin/sub_memory_manager.hpp
@@ -20,9 +20,15 @@ class SubMemoryManager {
         void* send_buf;
         std::shared_ptr<void> buf;
         bool flag;
+        int32_t last_rec_part;
         bool last_used;
         std::shared_ptr<cldnn::stream> stream_ptr;
         std::vector<cldnn::memory::ptr> recv_bufs;
+        std::vector<void*> remote_mems;
+        std::vector<void*> remote_mems_p2p;
+        std::vector<bool> recv_flag;
+        std::vector<bool> recv_flag_concat;
+        std::vector<bool> add_flag;
         std::vector<cldnn::event::ptr> events;
         cldnn::memory::ptr output;
         cldnn::layout layout;
@@ -33,9 +39,15 @@ class SubMemoryManager {
          _num_sub_streams = num_sub_streams;
         MemoryInfo memory_info;
         memory_info.flag = false;
+        memory_info.last_rec_part = -1;
         memory_info.last_used = false;
         memory_info.layout = cldnn::layout();
         memory_info.recv_bufs.assign(_num_sub_streams, nullptr);
+        memory_info.remote_mems.assign(_num_sub_streams, nullptr);
+        memory_info.remote_mems_p2p.assign(_num_sub_streams, nullptr);
+        memory_info.recv_flag.assign(_num_sub_streams, false);
+        memory_info.recv_flag_concat.assign(_num_sub_streams, false);
+        memory_info.add_flag.assign(_num_sub_streams, false);
         memory_info.events.assign(_num_sub_streams, nullptr);
         std::vector<MemoryInfo> memorys;
         memorys.assign(_num_sub_streams, memory_info);