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

64
AMSS_NCKU_source/bssn_rhs_cuda.h
View File

@@ -1,6 +1,6 @@
#ifndef BSSN_RHS_CUDA_H
#define BSSN_RHS_CUDA_H
#ifndef BSSN_RHS_CUDA_H
#define BSSN_RHS_CUDA_H
#ifdef __cplusplus
extern "C" {
#endif
@@ -9,28 +9,28 @@ enum {
BSSN_CUDA_STATE_COUNT = 24,
BSSN_CUDA_MATTER_COUNT = 10
};
int f_compute_rhs_bssn(int *ex, double &T,
double *X, double *Y, double *Z,
double *chi, double *trK,
double *dxx, double *gxy, double *gxz, double *dyy, double *gyz, double *dzz,
double *Axx, double *Axy, double *Axz, double *Ayy, double *Ayz, double *Azz,
double *Gamx, double *Gamy, double *Gamz,
double *Lap, double *betax, double *betay, double *betaz,
double *dtSfx, double *dtSfy, double *dtSfz,
double *chi_rhs, double *trK_rhs,
double *gxx_rhs, double *gxy_rhs, double *gxz_rhs, double *gyy_rhs, double *gyz_rhs, double *gzz_rhs,
double *Axx_rhs, double *Axy_rhs, double *Axz_rhs, double *Ayy_rhs, double *Ayz_rhs, double *Azz_rhs,
double *Gamx_rhs, double *Gamy_rhs, double *Gamz_rhs,
double *Lap_rhs, double *betax_rhs, double *betay_rhs, double *betaz_rhs,
double *dtSfx_rhs, double *dtSfy_rhs, double *dtSfz_rhs,
double *rho, double *Sx, double *Sy, double *Sz,
double *Sxx, double *Sxy, double *Sxz, double *Syy, double *Syz, double *Szz,
double *Gamxxx, double *Gamxxy, double *Gamxxz, double *Gamxyy, double *Gamxyz, double *Gamxzz,
double *Gamyxx, double *Gamyxy, double *Gamyxz, double *Gamyyy, double *Gamyyz, double *Gamyzz,
double *Gamzxx, double *Gamzxy, double *Gamzxz, double *Gamzyy, double *Gamzyz, double *Gamzzz,
double *Rxx, double *Rxy, double *Rxz, double *Ryy, double *Ryz, double *Rzz,
double *ham_Res, double *movx_Res, double *movy_Res, double *movz_Res,
double *dxx, double *gxy, double *gxz, double *dyy, double *gyz, double *dzz,
double *Axx, double *Axy, double *Axz, double *Ayy, double *Ayz, double *Azz,
double *Gamx, double *Gamy, double *Gamz,
double *Lap, double *betax, double *betay, double *betaz,
double *dtSfx, double *dtSfy, double *dtSfz,
double *chi_rhs, double *trK_rhs,
double *gxx_rhs, double *gxy_rhs, double *gxz_rhs, double *gyy_rhs, double *gyz_rhs, double *gzz_rhs,
double *Axx_rhs, double *Axy_rhs, double *Axz_rhs, double *Ayy_rhs, double *Ayz_rhs, double *Azz_rhs,
double *Gamx_rhs, double *Gamy_rhs, double *Gamz_rhs,
double *Lap_rhs, double *betax_rhs, double *betay_rhs, double *betaz_rhs,
double *dtSfx_rhs, double *dtSfy_rhs, double *dtSfz_rhs,
double *rho, double *Sx, double *Sy, double *Sz,
double *Sxx, double *Sxy, double *Sxz, double *Syy, double *Syz, double *Szz,
double *Gamxxx, double *Gamxxy, double *Gamxxz, double *Gamxyy, double *Gamxyz, double *Gamxzz,
double *Gamyxx, double *Gamyxy, double *Gamyxz, double *Gamyyy, double *Gamyyz, double *Gamyzz,
double *Gamzxx, double *Gamzxy, double *Gamzxz, double *Gamzyy, double *Gamzyz, double *Gamzzz,
double *Rxx, double *Rxy, double *Rxz, double *Ryy, double *Ryz, double *Rzz,
double *ham_Res, double *movx_Res, double *movy_Res, double *movz_Res,
double *Gmx_Res, double *Gmy_Res, double *Gmz_Res,
int &Symmetry, int &Lev, double &eps, int &co);
@@ -104,6 +104,20 @@ int bssn_cuda_unpack_state_batch_from_host_buffer(void *block_tag,
int i0, int j0, int k0,
int sx, int sy, int sz);
int bssn_cuda_pack_state_batch_to_device_buffer(void *block_tag,
int state_count,
double *device_buffer,
int *ex,
int i0, int j0, int k0,
int sx, int sy, int sz);
int bssn_cuda_unpack_state_batch_from_device_buffer(void *block_tag,
int state_count,
double *device_buffer,
int *ex,
int i0, int j0, int k0,
int sx, int sy, int sz);
int bssn_cuda_download_state_subset(void *block_tag,
int *ex,
int subset_count,
@@ -122,6 +136,6 @@ void bssn_cuda_release_step_ctx(void *block_tag);
#ifdef __cplusplus
}
#endif
#endif
#endif
#endif