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

Optimize BSSN CUDA resident state and CUDA-aware MPI

Browse Source
This commit is contained in:
CGH0S7
2026-04-29 17:05:10 +08:00
parent a0dab90bcb
commit 22c1e7168b
7 changed files with 7461 additions and 6727 deletions
Download Patch File Download Diff File Expand all files Collapse all files

125
AMSS_NCKU_source/Block.C
View File

@@ -6,14 +6,68 @@
#include <cstdio>
#include <string>
#include <cmath>
#include <new>
using namespace std;
#include "Block.h"
#include "misc.h"
Block::Block(int DIM, int *shapei, double *bboxi, int ranki, int ingfsi, int fngfsi, int levi, const int cgpui) : rank(ranki), ingfs(ingfsi), fngfs(fngfsi), lev(levi), cgpu(cgpui)
{
#include <new>
using namespace std;
#include "Block.h"
#include "misc.h"
#if USE_CUDA_BSSN || USE_CUDA_Z4C
#include <cuda_runtime_api.h>
#endif
namespace {
bool cuda_pin_gridfuncs_enabled()
{
static int enabled = -1;
if (enabled < 0)
{
const char *env = getenv("AMSS_CUDA_PIN_GRIDFUNCS");
enabled = (env && atoi(env) != 0) ? 1 : 0;
}
return enabled != 0;
}
double *alloc_gridfunc(size_t count, unsigned char &pinned)
{
pinned = 0;
#if USE_CUDA_BSSN || USE_CUDA_Z4C
if (cuda_pin_gridfuncs_enabled())
{
double *ptr = 0;
cudaError_t err = cudaMallocHost((void **)&ptr, count * sizeof(double));
if (err == cudaSuccess)
{
pinned = 1;
return ptr;
}
cudaGetLastError();
}
#endif
return (double *)malloc(sizeof(double) * count);
}
void free_gridfunc(double *ptr, unsigned char pinned)
{
if (!ptr)
return;
#if USE_CUDA_BSSN || USE_CUDA_Z4C
if (pinned)
{
cudaFreeHost(ptr);
return;
}
#else
(void)pinned;
#endif
free(ptr);
}
}
Block::Block(int DIM, int *shapei, double *bboxi, int ranki, int ingfsi, int fngfsi, int levi, const int cgpui) : rank(ranki), lev(levi), cgpu(cgpui), ingfs(ingfsi), fngfs(fngfsi), igfs(0), fgfs(0), fgfs_pinned(0)
{
for (int i = 0; i < dim; i++)
X[i] = 0;
@@ -68,14 +122,15 @@ Block::Block(int DIM, int *shapei, double *bboxi, int ranki, int ingfsi, int fng
#endif
}
int nn = shape[0] * shape[1] * shape[2];
fgfs = new double *[fngfs];
for (int i = 0; i < fngfs; i++)
{
fgfs[i] = (double *)malloc(sizeof(double) * nn);
if (!(fgfs[i]))
{
cout << "on node#" << rank << ", out of memory when constructing Block." << endl;
int nn = shape[0] * shape[1] * shape[2];
fgfs = new double *[fngfs];
fgfs_pinned = new unsigned char[fngfs];
for (int i = 0; i < fngfs; i++)
{
fgfs[i] = alloc_gridfunc((size_t)nn, fgfs_pinned[i]);
if (!(fgfs[i]))
{
cout << "on node#" << rank << ", out of memory when constructing Block." << endl;
MPI_Abort(MPI_COMM_WORLD, 1);
}
memset(fgfs[i], 0, sizeof(double) * nn);
@@ -103,17 +158,19 @@ Block::~Block()
{
for (int i = 0; i < dim; i++)
delete[] X[i];
for (int i = 0; i < ingfs; i++)
free(igfs[i]);
delete[] igfs;
for (int i = 0; i < fngfs; i++)
free(fgfs[i]);
delete[] fgfs;
X[0] = X[1] = X[2] = 0;
igfs = 0;
fgfs = 0;
}
}
for (int i = 0; i < ingfs; i++)
free(igfs[i]);
delete[] igfs;
for (int i = 0; i < fngfs; i++)
free_gridfunc(fgfs[i], fgfs_pinned ? fgfs_pinned[i] : 0);
delete[] fgfs;
delete[] fgfs_pinned;
X[0] = X[1] = X[2] = 0;
igfs = 0;
fgfs = 0;
fgfs_pinned = 0;
}
}
void Block::checkBlock()
{
int myrank;
@@ -184,12 +241,14 @@ void Block::swapList(MyList<var> *VarList1, MyList<var> *VarList2, int myrank)
if (rank == myrank)
{
MyList<var> *varl1 = VarList1, *varl2 = VarList2;
while (varl1 && varl2)
{
misc::swap<double *>(fgfs[varl1->data->sgfn], fgfs[varl2->data->sgfn]);
varl1 = varl1->next;
varl2 = varl2->next;
}
while (varl1 && varl2)
{
misc::swap<double *>(fgfs[varl1->data->sgfn], fgfs[varl2->data->sgfn]);
if (fgfs_pinned)
misc::swap<unsigned char>(fgfs_pinned[varl1->data->sgfn], fgfs_pinned[varl2->data->sgfn]);
varl1 = varl1->next;
varl2 = varl2->next;
}
if (varl1 || varl2)
{
cout << "error in Block::swaplist, var lists does not match." << endl;