64-BitBrainstorm_2026/AMSS-NCKU
8
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

perf(transfer_cached): 将 per-call new/delete 的 req_node/req_is_recv/completed 数组移入 SyncCache 复用

Browse Source 
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>
This commit is contained in:
CGH0S7
2026-03-02 21:14:35 +08:00
parent 4012e9d068
commit 5070134857
2 changed files with 26 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
AMSS_NCKU_source/Parallel.C
View File

@@ -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);

3
AMSS_NCKU_source/Parallel.h
View File

@@ -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();