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

Pack sync send buffers directly from GPU state

Browse Source
This commit is contained in:
CGH0S7
2026-04-09 18:49:11 +08:00
parent b0dd069a2b
commit 4484635f0d
4 changed files with 210 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

138
AMSS_NCKU_source/Parallel.C
View File

@@ -3,6 +3,7 @@
#include "fmisc.h"
#include "prolongrestrict.h"
#include "bssn_cuda_ops.h"
#include "bssn_gpu.h"
#include "misc.h"
#include "parameters.h"
#include <cstring>
@@ -43,6 +44,85 @@ struct ParallelTransferContextGuard
}
};
bool parallel_can_gpu_pack_segments(MyList<Parallel::gridseg> *src, MyList<Parallel::gridseg> *dst,
int rank_in, MyList<var> *VarLists, MyList<var> *VarListd)
{
int myrank;
MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
if (!src || !dst)
return false;
if (src->data->Bg->lev != dst->data->Bg->lev)
return false;
while (src && dst)
{
if ((dst->data->Bg->rank == rank_in) && (src->data->Bg->rank == myrank))
{
MyList<var> *varls = VarLists;
MyList<var> *varld = VarListd;
while (varls && varld)
{
(void)varld;
if (!bssn_gpu_find_device_buffer(src->data->Bg->fgfs[varls->data->sgfn]))
return false;
varls = varls->next;
varld = varld->next;
}
if (varls || varld)
return false;
}
src = src->next;
dst = dst->next;
}
return true;
}
bool parallel_gpu_pack_segments(double *data,
MyList<Parallel::gridseg> *src, MyList<Parallel::gridseg> *dst,
int rank_in, MyList<var> *VarLists, MyList<var> *VarListd)
{
if (!data)
return false;
int myrank;
MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
if (!src || !dst)
return false;
if (src->data->Bg->lev != dst->data->Bg->lev)
return false;
int size_out = 0;
while (src && dst)
{
if ((dst->data->Bg->rank == rank_in) && (src->data->Bg->rank == myrank))
{
MyList<var> *varls = VarLists;
MyList<var> *varld = VarListd;
while (varls && varld)
{
(void)varld;
if (bssn_gpu_stage_download_region_to_buffer(src->data->Bg->fgfs[varls->data->sgfn],
src->data->Bg->shape,
src->data->Bg->bbox,
src->data->Bg->bbox + dim,
dst->data->shape,
dst->data->llb,
data + size_out))
return false;
size_out += dst->data->shape[0] * dst->data->shape[1] * dst->data->shape[2];
varls = varls->next;
varld = varld->next;
}
if (varls || varld)
return false;
}
src = src->next;
dst = dst->next;
}
return true;
}
void parallel_report_mpi_error(const char *context, int errcode, int req_no)
{
char errstr[MPI_MAX_ERROR_STRING];
@@ -4843,27 +4923,30 @@ void Parallel::Sync_start(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetr
for (int node = 0; node < cpusize; node++)
{
if (node == myrank)
{
int length;
if (node == myrank)
{
int length;
if (!cache.lengths_valid) {
length = data_packer(0, src[myrank], dst[myrank], node, PACK, VarList, VarList, Symmetry);
cache.recv_lengths[node] = length;
} else {
length = cache.recv_lengths[node];
}
if (length > 0)
{
if (length > cache.recv_buf_caps[node])
{
if (cache.recv_bufs[node]) delete[] cache.recv_bufs[node];
cache.recv_bufs[node] = new double[length];
cache.recv_buf_caps[node] = length;
}
data_packer(cache.recv_bufs[node], src[myrank], dst[myrank], node, PACK, VarList, VarList, Symmetry);
}
}
else
if (length > 0)
{
if (length > cache.recv_buf_caps[node])
{
if (cache.recv_bufs[node]) delete[] cache.recv_bufs[node];
cache.recv_bufs[node] = new double[length];
cache.recv_buf_caps[node] = length;
}
bssn_gpu_prepare_host_buffer(cache.recv_bufs[node], length);
if (!parallel_can_gpu_pack_segments(src[myrank], dst[myrank], node, VarList, VarList) ||
!parallel_gpu_pack_segments(cache.recv_bufs[node], src[myrank], dst[myrank], node, VarList, VarList))
data_packer(cache.recv_bufs[node], src[myrank], dst[myrank], node, PACK, VarList, VarList, Symmetry);
}
}
else
{
int slength;
if (!cache.lengths_valid) {
@@ -4872,17 +4955,20 @@ void Parallel::Sync_start(MyList<Patch> *PatL, MyList<var> *VarList, int Symmetr
} else {
slength = cache.send_lengths[node];
}
if (slength > 0)
{
if (slength > cache.send_buf_caps[node])
{
if (cache.send_bufs[node]) delete[] cache.send_bufs[node];
cache.send_bufs[node] = new double[slength];
cache.send_buf_caps[node] = slength;
}
data_packer(cache.send_bufs[node], src[myrank], dst[myrank], node, PACK, VarList, VarList, Symmetry);
state.req_node[state.req_no] = node;
state.req_is_recv[state.req_no] = 0;
if (slength > 0)
{
if (slength > cache.send_buf_caps[node])
{
if (cache.send_bufs[node]) delete[] cache.send_bufs[node];
cache.send_bufs[node] = new double[slength];
cache.send_buf_caps[node] = slength;
}
bssn_gpu_prepare_host_buffer(cache.send_bufs[node], slength);
if (!parallel_can_gpu_pack_segments(src[myrank], dst[myrank], node, VarList, VarList) ||
!parallel_gpu_pack_segments(cache.send_bufs[node], src[myrank], dst[myrank], node, VarList, VarList))
data_packer(cache.send_bufs[node], src[myrank], dst[myrank], node, PACK, VarList, VarList, Symmetry);
state.req_node[state.req_no] = node;
state.req_is_recv[state.req_no] = 0;
MPI_Isend((void *)cache.send_bufs[node], slength, MPI_DOUBLE, node, state.mpi_tag, MPI_COMM_WORLD, cache.reqs + state.req_no++);
}
int rlength;