1 /***************************************************************************
2 lj_coul_long_ext.cpp
3 -------------------
4 W. Michael Brown (ORNL)
5
6 Functions for LAMMPS access to lj/cut/coul/long acceleration routines.
7
8 __________________________________________________________________________
9 This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
10 __________________________________________________________________________
11
12 begin :
13 email : brownw@ornl.gov
14 ***************************************************************************/
15
16 #include <iostream>
17 #include <cassert>
18 #include <cmath>
19
20 #include "lal_lj_coul_long.h"
21
22 using namespace std;
23 using namespace LAMMPS_AL;
24
25 static LJCoulLong<PRECISION,ACC_PRECISION> LJCLMF;
26
27 // ---------------------------------------------------------------------------
28 // Allocate memory on host and device and copy constants to device
29 // ---------------------------------------------------------------------------
ljcl_gpu_init(const int ntypes,double ** cutsq,double ** host_lj1,double ** host_lj2,double ** host_lj3,double ** host_lj4,double ** offset,double * special_lj,const int inum,const int nall,const int max_nbors,const int maxspecial,const double cell_size,int & gpu_mode,FILE * screen,double ** host_cut_ljsq,double host_cut_coulsq,double * host_special_coul,const double qqrd2e,const double g_ewald)30 int ljcl_gpu_init(const int ntypes, double **cutsq, double **host_lj1,
31 double **host_lj2, double **host_lj3, double **host_lj4,
32 double **offset, double *special_lj, const int inum,
33 const int nall, const int max_nbors, const int maxspecial,
34 const double cell_size, int &gpu_mode, FILE *screen,
35 double **host_cut_ljsq, double host_cut_coulsq,
36 double *host_special_coul, const double qqrd2e,
37 const double g_ewald) {
38 LJCLMF.clear();
39 gpu_mode=LJCLMF.device->gpu_mode();
40 double gpu_split=LJCLMF.device->particle_split();
41 int first_gpu=LJCLMF.device->first_device();
42 int last_gpu=LJCLMF.device->last_device();
43 int world_me=LJCLMF.device->world_me();
44 int gpu_rank=LJCLMF.device->gpu_rank();
45 int procs_per_gpu=LJCLMF.device->procs_per_gpu();
46
47 LJCLMF.device->init_message(screen,"lj/cut/coul/long",first_gpu,last_gpu);
48
49 bool message=false;
50 if (LJCLMF.device->replica_me()==0 && screen)
51 message=true;
52
53 if (message) {
54 fprintf(screen,"Initializing Device and compiling on process 0...");
55 fflush(screen);
56 }
57
58 int init_ok=0;
59 if (world_me==0)
60 init_ok=LJCLMF.init(ntypes, cutsq, host_lj1, host_lj2, host_lj3, host_lj4,
61 offset, special_lj, inum, nall, max_nbors, maxspecial,
62 cell_size, gpu_split, screen, host_cut_ljsq,
63 host_cut_coulsq, host_special_coul, qqrd2e, g_ewald);
64
65 LJCLMF.device->world_barrier();
66 if (message)
67 fprintf(screen,"Done.\n");
68
69 for (int i=0; i<procs_per_gpu; i++) {
70 if (message) {
71 if (last_gpu-first_gpu==0)
72 fprintf(screen,"Initializing Device %d on core %d...",first_gpu,i);
73 else
74 fprintf(screen,"Initializing Devices %d-%d on core %d...",first_gpu,
75 last_gpu,i);
76 fflush(screen);
77 }
78 if (gpu_rank==i && world_me!=0)
79 init_ok=LJCLMF.init(ntypes, cutsq, host_lj1, host_lj2, host_lj3, host_lj4,
80 offset, special_lj, inum, nall, max_nbors, maxspecial,
81 cell_size, gpu_split, screen, host_cut_ljsq,
82 host_cut_coulsq, host_special_coul, qqrd2e, g_ewald);
83
84 LJCLMF.device->gpu_barrier();
85 if (message)
86 fprintf(screen,"Done.\n");
87 }
88 if (message)
89 fprintf(screen,"\n");
90
91 if (init_ok==0)
92 LJCLMF.estimate_gpu_overhead();
93 return init_ok;
94 }
95
96 // ---------------------------------------------------------------------------
97 // Copy updated coeffs from host to device
98 // ---------------------------------------------------------------------------
ljcl_gpu_reinit(const int ntypes,double ** cutsq,double ** host_lj1,double ** host_lj2,double ** host_lj3,double ** host_lj4,double ** offset,double ** host_cut_ljsq)99 void ljcl_gpu_reinit(const int ntypes, double **cutsq, double **host_lj1,
100 double **host_lj2, double **host_lj3, double **host_lj4,
101 double **offset, double **host_cut_ljsq) {
102 int world_me=LJCLMF.device->world_me();
103 int gpu_rank=LJCLMF.device->gpu_rank();
104 int procs_per_gpu=LJCLMF.device->procs_per_gpu();
105
106 if (world_me==0)
107 LJCLMF.reinit(ntypes, cutsq, host_lj1, host_lj2, host_lj3, host_lj4,
108 offset, host_cut_ljsq);
109 LJCLMF.device->world_barrier();
110
111 for (int i=0; i<procs_per_gpu; i++) {
112 if (gpu_rank==i && world_me!=0)
113 LJCLMF.reinit(ntypes, cutsq, host_lj1, host_lj2, host_lj3, host_lj4,
114 offset, host_cut_ljsq);
115 LJCLMF.device->gpu_barrier();
116 }
117 }
118
ljcl_gpu_clear()119 void ljcl_gpu_clear() {
120 LJCLMF.clear();
121 }
122
ljcl_gpu_compute_n(const int ago,const int inum_full,const int nall,double ** host_x,int * host_type,double * sublo,double * subhi,tagint * tag,int ** nspecial,tagint ** special,const bool eflag,const bool vflag,const bool eatom,const bool vatom,int & host_start,int ** ilist,int ** jnum,const double cpu_time,bool & success,double * host_q,double * boxlo,double * prd)123 int** ljcl_gpu_compute_n(const int ago, const int inum_full,
124 const int nall, double **host_x, int *host_type,
125 double *sublo, double *subhi, tagint *tag, int **nspecial,
126 tagint **special, const bool eflag, const bool vflag,
127 const bool eatom, const bool vatom, int &host_start,
128 int **ilist, int **jnum, const double cpu_time,
129 bool &success, double *host_q, double *boxlo,
130 double *prd) {
131 return LJCLMF.compute(ago, inum_full, nall, host_x, host_type, sublo,
132 subhi, tag, nspecial, special, eflag, vflag, eatom,
133 vatom, host_start, ilist, jnum, cpu_time, success,
134 host_q, boxlo, prd);
135 }
136
ljcl_gpu_compute(const int ago,const int inum_full,const int nall,double ** host_x,int * host_type,int * ilist,int * numj,int ** firstneigh,const bool eflag,const bool vflag,const bool eatom,const bool vatom,int & host_start,const double cpu_time,bool & success,double * host_q,const int nlocal,double * boxlo,double * prd)137 void ljcl_gpu_compute(const int ago, const int inum_full, const int nall,
138 double **host_x, int *host_type, int *ilist, int *numj,
139 int **firstneigh, const bool eflag, const bool vflag,
140 const bool eatom, const bool vatom, int &host_start,
141 const double cpu_time, bool &success, double *host_q,
142 const int nlocal, double *boxlo, double *prd) {
143 LJCLMF.compute(ago,inum_full,nall,host_x,host_type,ilist,numj,
144 firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success,
145 host_q,nlocal,boxlo,prd);
146 }
147
ljcl_gpu_bytes()148 double ljcl_gpu_bytes() {
149 return LJCLMF.host_memory_usage();
150 }
151
152
153