1 /***************************************************************************
2 coul_long_cs_ext.cpp
3 -------------------
4 Trung Nguyen (Northwestern)
5
6 Functions for LAMMPS access to coul/long/cs acceleration routines.
7
8 __________________________________________________________________________
9 This file is part of the LAMMPS Accelerator Library (LAMMPS_AL)
10 __________________________________________________________________________
11
12 begin : June 2018
13 email : ndactrung@gmail.com
14 ***************************************************************************/
15
16 #include <iostream>
17 #include <cassert>
18 #include <cmath>
19
20 #include "lal_coul_long_cs.h"
21
22 using namespace std;
23 using namespace LAMMPS_AL;
24
25 static CoulLongCS<PRECISION,ACC_PRECISION> CLCSMF;
26
27 // ---------------------------------------------------------------------------
28 // Allocate memory on host and device and copy constants to device
29 // ---------------------------------------------------------------------------
clcs_gpu_init(const int ntypes,double ** host_scale,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_coulsq,double * host_special_coul,const double qqrd2e,const double g_ewald)30 int clcs_gpu_init(const int ntypes, double **host_scale,
31 const int inum, const int nall, const int max_nbors,
32 const int maxspecial, const double cell_size, int &gpu_mode,
33 FILE *screen, double host_cut_coulsq, double *host_special_coul,
34 const double qqrd2e, const double g_ewald) {
35 CLCSMF.clear();
36 gpu_mode=CLCSMF.device->gpu_mode();
37 double gpu_split=CLCSMF.device->particle_split();
38 int first_gpu=CLCSMF.device->first_device();
39 int last_gpu=CLCSMF.device->last_device();
40 int world_me=CLCSMF.device->world_me();
41 int gpu_rank=CLCSMF.device->gpu_rank();
42 int procs_per_gpu=CLCSMF.device->procs_per_gpu();
43
44 CLCSMF.device->init_message(screen,"coul/long/cs",first_gpu,last_gpu);
45
46 bool message=false;
47 if (CLCSMF.device->replica_me()==0 && screen)
48 message=true;
49
50 if (message) {
51 fprintf(screen,"Initializing Device and compiling on process 0...");
52 fflush(screen);
53 }
54
55 int init_ok=0;
56 if (world_me==0)
57 init_ok=CLCSMF.init(ntypes, host_scale, inum, nall, max_nbors, maxspecial,
58 cell_size, gpu_split, screen, host_cut_coulsq,
59 host_special_coul, qqrd2e, g_ewald);
60
61 CLCSMF.device->world_barrier();
62 if (message)
63 fprintf(screen,"Done.\n");
64
65 for (int i=0; i<procs_per_gpu; i++) {
66 if (message) {
67 if (last_gpu-first_gpu==0)
68 fprintf(screen,"Initializing Device %d on core %d...",first_gpu,i);
69 else
70 fprintf(screen,"Initializing Devices %d-%d on core %d...",first_gpu,
71 last_gpu,i);
72 fflush(screen);
73 }
74 if (gpu_rank==i && world_me!=0)
75 init_ok=CLCSMF.init(ntypes, host_scale, inum, nall, max_nbors, maxspecial,
76 cell_size, gpu_split, screen, host_cut_coulsq,
77 host_special_coul, qqrd2e, g_ewald);
78
79 CLCSMF.device->gpu_barrier();
80 if (message)
81 fprintf(screen,"Done.\n");
82 }
83 if (message)
84 fprintf(screen,"\n");
85
86 if (init_ok==0)
87 CLCSMF.estimate_gpu_overhead();
88 return init_ok;
89 }
90
91 // ---------------------------------------------------------------------------
92 // Copy updated coeffs from host to device
93 // ---------------------------------------------------------------------------
clcs_gpu_reinit(const int ntypes,double ** host_scale)94 void clcs_gpu_reinit(const int ntypes, double **host_scale) {
95 int world_me=CLCSMF.device->world_me();
96 int gpu_rank=CLCSMF.device->gpu_rank();
97 int procs_per_gpu=CLCSMF.device->procs_per_gpu();
98
99 if (world_me==0)
100 CLCSMF.reinit(ntypes, host_scale);
101
102 CLCSMF.device->world_barrier();
103
104 for (int i=0; i<procs_per_gpu; i++) {
105 if (gpu_rank==i && world_me!=0)
106 CLCSMF.reinit(ntypes, host_scale);
107
108 CLCSMF.device->gpu_barrier();
109 }
110 }
111
clcs_gpu_clear()112 void clcs_gpu_clear() {
113 CLCSMF.clear();
114 }
115
clcs_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)116 int** clcs_gpu_compute_n(const int ago, const int inum_full,
117 const int nall, double **host_x, int *host_type,
118 double *sublo, double *subhi, tagint *tag, int **nspecial,
119 tagint **special, const bool eflag, const bool vflag,
120 const bool eatom, const bool vatom, int &host_start,
121 int **ilist, int **jnum, const double cpu_time,
122 bool &success, double *host_q, double *boxlo,
123 double *prd) {
124 return CLCSMF.compute(ago, inum_full, nall, host_x, host_type, sublo,
125 subhi, tag, nspecial, special, eflag, vflag, eatom,
126 vatom, host_start, ilist, jnum, cpu_time, success,
127 host_q, boxlo, prd);
128 }
129
clcs_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)130 void clcs_gpu_compute(const int ago, const int inum_full, const int nall,
131 double **host_x, int *host_type, int *ilist, int *numj,
132 int **firstneigh, const bool eflag, const bool vflag,
133 const bool eatom, const bool vatom, int &host_start,
134 const double cpu_time, bool &success, double *host_q,
135 const int nlocal, double *boxlo, double *prd) {
136 CLCSMF.compute(ago,inum_full,nall,host_x,host_type,ilist,numj,
137 firstneigh,eflag,vflag,eatom,vatom,host_start,cpu_time,success,
138 host_q,nlocal,boxlo,prd);
139 }
140
clcs_gpu_bytes()141 double clcs_gpu_bytes() {
142 return CLCSMF.host_memory_usage();
143 }
144
145
146