From 5070134857bed9940c9ab0a3cb2b51c8506c6358 Mon Sep 17 00:00:00 2001
From: CGH0S7 <776459475@qq.com>
Date: Mon, 2 Mar 2026 21:14:35 +0800
Subject: [PATCH] =?UTF-8?q?perf(transfer=5Fcached):=20=E5=B0=86=20per-call?=
 =?UTF-8?q?=20new/delete=20=E7=9A=84=20req=5Fnode/req=5Fis=5Frecv/complete?=
 =?UTF-8?q?d=20=E6=95=B0=E7=BB=84=E7=A7=BB=E5=85=A5=20SyncCache=20?=
 =?UTF-8?q?=E5=A4=8D=E7=94=A8?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

避免 transfer_cached 每次调用分配释放 3 个临时数组，减少堆操作开销。

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 AMSS_NCKU_source/Parallel.C | 31 +++++++++++++++++++++++--------
 AMSS_NCKU_source/Parallel.h |  3 +++
 2 files changed, 26 insertions(+), 8 deletions(-)

diff --git a/AMSS_NCKU_source/Parallel.C b/AMSS_NCKU_source/Parallel.C
index 4e5e4ec..b87ce6d 100644
--- a/AMSS_NCKU_source/Parallel.C
+++ b/AMSS_NCKU_source/Parallel.C
@@ -4320,7 +4320,7 @@ Parallel::SyncCache::SyncCache()
     : valid(false), cpusize(0), combined_src(0), combined_dst(0),
       send_lengths(0), recv_lengths(0), send_bufs(0), recv_bufs(0),
       send_buf_caps(0), recv_buf_caps(0), reqs(0), stats(0), max_reqs(0),
-      lengths_valid(false)
+      lengths_valid(false), tc_req_node(0), tc_req_is_recv(0), tc_completed(0)
 {
 }
 // SyncCache invalidate: free grid segment lists but keep buffers
@@ -4359,11 +4359,15 @@ void Parallel::SyncCache::destroy()
   if (recv_bufs) delete[] recv_bufs;
   if (reqs) delete[] reqs;
   if (stats) delete[] stats;
+  if (tc_req_node) delete[] tc_req_node;
+  if (tc_req_is_recv) delete[] tc_req_is_recv;
+  if (tc_completed) delete[] tc_completed;
   combined_src = combined_dst = 0;
   send_lengths = recv_lengths = 0;
   send_buf_caps = recv_buf_caps = 0;
   send_bufs = recv_bufs = 0;
   reqs = 0; stats = 0;
+  tc_req_node = 0; tc_req_is_recv = 0; tc_completed = 0;
   cpusize = 0; max_reqs = 0;
 }
 // transfer_cached: reuse pre-allocated buffers from SyncCache
@@ -4379,9 +4383,9 @@ void Parallel::transfer_cached(MyList<Parallel::gridseg> **src, MyList<Parallel:
   int req_no = 0;
   int pending_recv = 0;
   int node;
-  int *req_node = new int[cache.max_reqs];
-  int *req_is_recv = new int[cache.max_reqs];
-  int *completed = new int[cache.max_reqs];
+  int *req_node = cache.tc_req_node;
+  int *req_is_recv = cache.tc_req_is_recv;
+  int *completed = cache.tc_completed;
 
   // Post receives first so peers can progress rendezvous early.
   for (node = 0; node < cpusize; node++)
@@ -4466,10 +4470,6 @@ void Parallel::transfer_cached(MyList<Parallel::gridseg> **src, MyList<Parallel:
 
   if (self_len > 0)
     data_packer(cache.recv_bufs[myrank], src[myrank], dst[myrank], myrank, UNPACK, VarList1, VarList2, Symmetry);
-
-  delete[] req_node;
-  delete[] req_is_recv;
-  delete[] completed;
 }
 void Parallel::Sync_cached(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetry, SyncCache &cache)
 {
@@ -4498,6 +4498,9 @@ void Parallel::Sync_cached(MyList<Patch> *PatL, MyList<var> *VarList, int Symmet
       cache.max_reqs = 2 * cpusize;
       cache.reqs = new MPI_Request[cache.max_reqs];
       cache.stats = new MPI_Status[cache.max_reqs];
+      cache.tc_req_node = new int[cache.max_reqs];
+      cache.tc_req_is_recv = new int[cache.max_reqs];
+      cache.tc_completed = new int[cache.max_reqs];
     }
 
     for (int node = 0; node < cpusize; node++)
@@ -4598,6 +4601,9 @@ void Parallel::Sync_start(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetr
       cache.max_reqs = 2 * cpusize;
       cache.reqs = new MPI_Request[cache.max_reqs];
       cache.stats = new MPI_Status[cache.max_reqs];
+      cache.tc_req_node = new int[cache.max_reqs];
+      cache.tc_req_is_recv = new int[cache.max_reqs];
+      cache.tc_completed = new int[cache.max_reqs];
     }
 
     for (int node = 0; node < cpusize; node++)
@@ -5856,6 +5862,9 @@ void Parallel::Restrict_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
       cache.max_reqs = 2 * cpusize;
       cache.reqs = new MPI_Request[cache.max_reqs];
       cache.stats = new MPI_Status[cache.max_reqs];
+      cache.tc_req_node = new int[cache.max_reqs];
+      cache.tc_req_is_recv = new int[cache.max_reqs];
+      cache.tc_completed = new int[cache.max_reqs];
     }
 
     MyList<Parallel::gridseg> *dst = build_complete_gsl(PatcL);
@@ -5902,6 +5911,9 @@ void Parallel::OutBdLow2Hi_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
       cache.max_reqs = 2 * cpusize;
       cache.reqs = new MPI_Request[cache.max_reqs];
       cache.stats = new MPI_Status[cache.max_reqs];
+      cache.tc_req_node = new int[cache.max_reqs];
+      cache.tc_req_is_recv = new int[cache.max_reqs];
+      cache.tc_completed = new int[cache.max_reqs];
     }
 
     MyList<Parallel::gridseg> *dst = build_buffer_gsl(PatfL);
@@ -5948,6 +5960,9 @@ void Parallel::OutBdLow2Himix_cached(MyList<Patch> *PatcL, MyList<Patch> *PatfL,
       cache.max_reqs = 2 * cpusize;
       cache.reqs = new MPI_Request[cache.max_reqs];
       cache.stats = new MPI_Status[cache.max_reqs];
+      cache.tc_req_node = new int[cache.max_reqs];
+      cache.tc_req_is_recv = new int[cache.max_reqs];
+      cache.tc_completed = new int[cache.max_reqs];
     }
 
     MyList<Parallel::gridseg> *dst = build_buffer_gsl(PatfL);
diff --git a/AMSS_NCKU_source/Parallel.h b/AMSS_NCKU_source/Parallel.h
index 5a72797..0ab975c 100644
--- a/AMSS_NCKU_source/Parallel.h
+++ b/AMSS_NCKU_source/Parallel.h
@@ -108,6 +108,9 @@ namespace Parallel
     MPI_Status *stats;
     int max_reqs;
     bool lengths_valid;
+    int *tc_req_node;
+    int *tc_req_is_recv;
+    int *tc_completed;
     SyncCache();
     void invalidate();
     void destroy();