1 /***************************************************************************
2 born_ext.cpp
3 -------------------
4 Trung Dac Nguyen (ORNL)
5
6 Functions for LAMMPS access to born acceleration routines.
7
8 __________________________________________________________________________
9 This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
10 __________________________________________________________________________
11
12 begin :
13 email : nguyentd@ornl.gov
14 ***************************************************************************/
15
16 #include <iostream>
17 #include <cassert>
18 #include <cmath>
19
20 #include "lal_born.h"
21
22 using namespace std;
23 using namespace LAMMPS_AL;
24
25 static Born<PRECISION,ACC_PRECISION> BORNMF;
26
27 // ---------------------------------------------------------------------------
28 // Allocate memory on host and device and copy constants to device
29 // ---------------------------------------------------------------------------
born_gpu_init(const int ntypes,double ** cutsq,double ** host_rhoinv,double ** host_born1,double ** host_born2,double ** host_born3,double ** host_a,double ** host_c,double ** host_d,double ** sigma,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)30 int born_gpu_init(const int ntypes, double **cutsq, double **host_rhoinv,
31 double **host_born1, double **host_born2,
32 double **host_born3, double **host_a, double **host_c,
33 double **host_d, double **sigma,
34 double **offset, double *special_lj, const int inum,
35 const int nall, const int max_nbors, const int maxspecial,
36 const double cell_size, int &gpu_mode, FILE *screen) {
37 BORNMF.clear();
38 gpu_mode=BORNMF.device->gpu_mode();
39 double gpu_split=BORNMF.device->particle_split();
40 int first_gpu=BORNMF.device->first_device();
41 int last_gpu=BORNMF.device->last_device();
42 int world_me=BORNMF.device->world_me();
43 int gpu_rank=BORNMF.device->gpu_rank();
44 int procs_per_gpu=BORNMF.device->procs_per_gpu();
45
46 BORNMF.device->init_message(screen,"born",first_gpu,last_gpu);
47
48 bool message=false;
49 if (BORNMF.device->replica_me()==0 && screen)
50 message=true;
51
52 if (message) {
53 fprintf(screen,"Initializing Device and compiling on process 0...");
54 fflush(screen);
55 }
56
57 int init_ok=0;
58 if (world_me==0)
59 init_ok=BORNMF.init(ntypes, cutsq, host_rhoinv, host_born1, host_born2,
60 host_born3, host_a, host_c, host_d, sigma,
61 offset, special_lj, inum, nall, max_nbors,
62 maxspecial, cell_size, gpu_split, screen);
63
64 BORNMF.device->world_barrier();
65 if (message)
66 fprintf(screen,"Done.\n");
67
68 for (int i=0; i<procs_per_gpu; i++) {
69 if (message) {
70 if (last_gpu-first_gpu==0)
71 fprintf(screen,"Initializing Device %d on core %d...",first_gpu,i);
72 else
73 fprintf(screen,"Initializing Devices %d-%d on core %d...",first_gpu,
74 last_gpu,i);
75 fflush(screen);
76 }
77 if (gpu_rank==i && world_me!=0)
78 init_ok=BORNMF.init(ntypes, cutsq, host_rhoinv, host_born1, host_born2,
79 host_born3, host_a, host_c, host_d, sigma,
80 offset, special_lj, inum, nall, max_nbors,
81 maxspecial, cell_size, gpu_split, screen);
82
83 BORNMF.device->gpu_barrier();
84 if (message)
85 fprintf(screen,"Done.\n");
86 }
87 if (message)
88 fprintf(screen,"\n");
89
90 if (init_ok==0)
91 BORNMF.estimate_gpu_overhead();
92 return init_ok;
93 }
94
95 // ---------------------------------------------------------------------------
96 // Copy updated coeffs from host to device
97 // ---------------------------------------------------------------------------
born_gpu_reinit(const int ntypes,double ** host_rhoinv,double ** host_born1,double ** host_born2,double ** host_born3,double ** host_a,double ** host_c,double ** host_d,double ** offset)98 void born_gpu_reinit(const int ntypes, double **host_rhoinv,
99 double **host_born1, double **host_born2,
100 double **host_born3, double **host_a, double **host_c,
101 double **host_d, double **offset) {
102 int world_me=BORNMF.device->world_me();
103 int gpu_rank=BORNMF.device->gpu_rank();
104 int procs_per_gpu=BORNMF.device->procs_per_gpu();
105
106 if (world_me==0)
107 BORNMF.reinit(ntypes, host_rhoinv, host_born1, host_born2,
108 host_born3, host_a, host_c, host_d, offset);
109
110 BORNMF.device->world_barrier();
111
112 for (int i=0; i<procs_per_gpu; i++) {
113 if (gpu_rank==i && world_me!=0)
114 BORNMF.reinit(ntypes, host_rhoinv, host_born1, host_born2,
115 host_born3, host_a, host_c, host_d, offset);
116
117 BORNMF.device->gpu_barrier();
118 }
119 }
120
born_gpu_clear()121 void born_gpu_clear() {
122 BORNMF.clear();
123 }
124
born_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)125 int ** born_gpu_compute_n(const int ago, const int inum_full,
126 const int nall, double **host_x, int *host_type,
127 double *sublo, double *subhi, tagint *tag, int **nspecial,
128 tagint **special, const bool eflag, const bool vflag,
129 const bool eatom, const bool vatom, int &host_start,
130 int **ilist, int **jnum, const double cpu_time,
131 bool &success) {
132 return BORNMF.compute(ago, inum_full, nall, host_x, host_type, sublo,
133 subhi, tag, nspecial, special, eflag, vflag, eatom,
134 vatom, host_start, ilist, jnum, cpu_time, success);
135 }
136
born_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)137 void born_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) {
142 BORNMF.compute(ago,inum_full,nall,host_x,host_type,ilist,numj,
143 firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success);
144 }
145
born_gpu_bytes()146 double born_gpu_bytes() {
147 return BORNMF.host_memory_usage();
148 }
149
150
151