3 * This source code is part of
7 * GROningen MAchine for Chemical Simulations
10 * Written by David van der Spoel, Erik Lindahl, Berk Hess, and others.
11 * Copyright (c) 1991-2000, University of Groningen, The Netherlands.
12 * Copyright (c) 2001-2004, The GROMACS development team,
13 * check out http://www.gromacs.org for more information.
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version 2
18 * of the License, or (at your option) any later version.
20 * If you want to redistribute modifications, please consider that
21 * scientific software is very special. Version control is crucial -
22 * bugs must be traceable. We will be happy to consider code for
23 * inclusion in the official distribution, but derived work must not
24 * be called official GROMACS. Details are found in the README & COPYING
25 * files - if they are missing, get the official version at www.gromacs.org.
27 * To help us fund GROMACS development, we humbly ask that you cite
28 * the papers on the package - you can find them in the top README file.
30 * For more info, check our website at http://www.gromacs.org
33 * GROningen Mixture of Alchemy and Childrens' Stories
40 #include "types/commrec.h"
42 #include "gmx_fatal.h"
60 static const t_nrnb_data nbdata[eNRNB] = {
61 /* These are re-used for different NB kernels, since there are so many.
62 * The actual number of flops is set dynamically.
64 { "NB VdW [V&F]", 1 },
66 { "NB Elec. [V&F]", 1 },
67 { "NB Elec. [F]", 1 },
68 { "NB Elec. [W3,V&F]", 1 },
69 { "NB Elec. [W3,F]", 1 },
70 { "NB Elec. [W3-W3,V&F]", 1 },
71 { "NB Elec. [W3-W3,F]", 1 },
72 { "NB Elec. [W4,V&F]", 1 },
73 { "NB Elec. [W4,F]", 1 },
74 { "NB Elec. [W4-W4,V&F]", 1 },
75 { "NB Elec. [W4-W4,F]", 1 },
76 { "NB VdW & Elec. [V&F]", 1 },
77 { "NB VdW & Elec. [F]", 1 },
78 { "NB VdW & Elec. [W3,V&F]", 1 },
79 { "NB VdW & Elec. [W3,F]", 1 },
80 { "NB VdW & Elec. [W3-W3,V&F]", 1 },
81 { "NB VdW & Elec. [W3-W3,F]", 1 },
82 { "NB VdW & Elec. [W4,V&F]", 1 },
83 { "NB VdW & Elec. [W4,F]", 1 },
84 { "NB VdW & Elec. [W4-W4,V&F]", 1 },
85 { "NB VdW & Elec. [W4-W4,F]", 1 },
87 { "NB Generic kernel", 1 },
88 { "NB Free energy kernel", 1 },
89 { "NB All-vs-all", 1 },
90 { "NB All-vs-all, GB", 1 },
92 { "Pair Search distance check", 9 }, /* nbnxn pair dist. check */
93 /* nbnxn kernel flops are based on inner-loops without exclusion checks.
94 * Plain Coulomb runs through the RF kernels, except with CUDA.
95 * invsqrt is counted as 6 flops: 1 for _mm_rsqt_ps + 5 for iteration.
96 * The flops are equal for plain-C, x86 SIMD and CUDA, except for:
97 * - plain-C kernel uses one flop more for Coulomb-only (F) than listed
98 * - x86 SIMD LJ geom-comb.rule kernels (fastest) use 2 more flops
99 * - x86 SIMD LJ LB-comb.rule kernels (fast) use 3 (8 for F+E) more flops
100 * - GPU always does exclusions, which requires 2-4 flops, but as invsqrt
101 * is always counted as 6 flops, this roughly compensates.
103 { "NxN RF Elec. + VdW [F]", 38 }, /* nbnxn kernel LJ+RF, no ener */
104 { "NxN RF Elec. + VdW [V&F]", 54 },
105 { "NxN QSTab Elec. + VdW [F]", 41 }, /* nbnxn kernel LJ+tab, no en */
106 { "NxN QSTab Elec. + VdW [V&F]", 59 },
107 { "NxN Ewald Elec. + VdW [F]", 66 }, /* nbnxn kernel LJ+Ewald, no en */
108 { "NxN Ewald Elec. + VdW [V&F]", 107 },
109 { "NxN VdW [F]", 33 }, /* nbnxn kernel LJ, no ener */
110 { "NxN VdW [V&F]", 43 },
111 { "NxN RF Electrostatics [F]", 31 }, /* nbnxn kernel RF, no ener */
112 { "NxN RF Electrostatics [V&F]", 36 },
113 { "NxN QSTab Elec. [F]", 34 }, /* nbnxn kernel tab, no ener */
114 { "NxN QSTab Elec. [V&F]", 41 },
115 { "NxN Ewald Elec. [F]", 61 }, /* nbnxn kernel Ewald, no ener */
116 { "NxN Ewald Elec. [V&F]", 84 },
117 { "1,4 nonbonded interactions", 90 },
118 { "Born radii (Still)", 47 },
119 { "Born radii (HCT/OBC)", 183 },
120 { "Born force chain rule", 15 },
121 { "All-vs-All Still radii", 1 },
122 { "All-vs-All HCT/OBC radii", 1 },
123 { "All-vs-All Born chain rule", 1 },
124 { "Calc Weights", 36 },
126 { "Spread Q Bspline", 2 },
128 { "Gather F Bspline", 6 },
130 { "Convolution", 4 },
133 { "Reset In Box", 3 },
139 { "FENE Bonds", 58 },
140 { "Tab. Bonds", 62 },
141 { "Restraint Potential", 86 },
142 { "Linear Angles", 57 },
144 { "G96Angles", 150 },
145 { "Quartic Angles", 160 },
146 { "Tab. Angles", 169 },
148 { "Impropers", 208 },
149 { "RB-Dihedrals", 247 },
150 { "Four. Dihedrals", 247 },
151 { "Tab. Dihedrals", 227 },
152 { "Dist. Restr.", 200 },
153 { "Orient. Restr.", 200 },
154 { "Dihedral Restr.", 200 },
155 { "Pos. Restr.", 50 },
156 { "Flat-bottom posres", 50 },
157 { "Angle Restr.", 191 },
158 { "Angle Restr. Z", 164 },
159 { "Morse Potent.", 83 },
160 { "Cubic Bonds", 54 },
162 { "Polarization", 59 },
163 { "Anharmonic Polarization", 72 },
164 { "Water Pol.", 62 },
165 { "Thole Pol.", 296 },
168 { "Ext.ens. Update", 54 },
175 { "Constraint-V", 8 },
176 { "Shake-Init", 10 },
177 { "Constraint-Vir", 24 },
179 { "Virtual Site 2", 23 },
180 { "Virtual Site 3", 37 },
181 { "Virtual Site 3fd", 95 },
182 { "Virtual Site 3fad", 176 },
183 { "Virtual Site 3out", 87 },
184 { "Virtual Site 4fd", 110 },
185 { "Virtual Site 4fdn", 254 },
186 { "Virtual Site N", 15 },
187 { "Mixed Generalized Born stuff", 10 }
191 void init_nrnb(t_nrnb *nrnb)
195 for(i=0; (i<eNRNB); i++)
199 void cp_nrnb(t_nrnb *dest, t_nrnb *src)
203 for(i=0; (i<eNRNB); i++)
204 dest->n[i]=src->n[i];
207 void add_nrnb(t_nrnb *dest, t_nrnb *s1, t_nrnb *s2)
211 for(i=0; (i<eNRNB); i++)
212 dest->n[i]=s1->n[i]+s2->n[i];
215 void print_nrnb(FILE *out, t_nrnb *nrnb)
219 for(i=0; (i<eNRNB); i++)
221 fprintf(out," %-26s %10.0f.\n",nbdata[i].name,nrnb->n[i]);
224 void _inc_nrnb(t_nrnb *nrnb,int enr,int inc,char *file,int line)
228 printf("nrnb %15s(%2d) incremented with %8d from file %s line %d\n",
229 nbdata[enr].name,enr,inc,file,line);
233 void print_flop(FILE *out,t_nrnb *nrnb,double *nbfs,double *mflop)
236 double mni,frac,tfrac,tflop;
237 const char *myline = "-----------------------------------------------------------------------------";
240 for(i=0; (i<eNR_NBKERNEL_ALLVSALLGB); i++) {
241 if (strstr(nbdata[i].name,"W3-W3") != NULL)
242 *nbfs += 9e-6*nrnb->n[i];
243 else if (strstr(nbdata[i].name,"W3") != NULL)
244 *nbfs += 3e-6*nrnb->n[i];
245 else if (strstr(nbdata[i].name,"W4-W4") != NULL)
246 *nbfs += 10e-6*nrnb->n[i];
247 else if (strstr(nbdata[i].name,"W4") != NULL)
248 *nbfs += 4e-6*nrnb->n[i];
250 *nbfs += 1e-6*nrnb->n[i];
253 for(i=0; (i<eNRNB); i++)
254 tflop+=1e-6*nrnb->n[i]*nbdata[i].flop;
257 fprintf(out,"No MEGA Flopsen this time\n");
261 fprintf(out,"\n\tM E G A - F L O P S A C C O U N T I N G\n\n");
266 fprintf(out," NB=Group-cutoff nonbonded kernels NxN=N-by-N cluster Verlet kernels\n");
267 fprintf(out," RF=Reaction-Field VdW=Van der Waals QSTab=quadratic-spline table\n");
268 fprintf(out," W3=SPC/TIP3p W4=TIP4p (single or pairs)\n");
269 fprintf(out," V&F=Potential and force V=Potential only F=Force only\n\n");
271 fprintf(out," %-32s %16s %15s %7s\n",
272 "Computing:","M-Number","M-Flops","% Flops");
273 fprintf(out,"%s\n",myline);
277 for(i=0; (i<eNRNB); i++) {
278 mni = 1e-6*nrnb->n[i];
279 *mflop += mni*nbdata[i].flop;
280 frac = 100.0*mni*nbdata[i].flop/tflop;
283 fprintf(out," %-32s %16.6f %15.3f %6.1f\n",
284 nbdata[i].name,mni,mni*nbdata[i].flop,frac);
287 fprintf(out,"%s\n",myline);
288 fprintf(out," %-32s %16s %15.3f %6.1f\n",
289 "Total","",*mflop,tfrac);
290 fprintf(out,"%s\n\n",myline);
294 void print_perf(FILE *out,double nodetime,double realtime,int nprocs,
295 gmx_large_int_t nsteps,real delta_t,
296 double nbfs,double mflop,
305 fprintf(out,"%12s %12s %12s %10s\n","","Core t (s)","Wall t (s)","(%)");
306 fprintf(out,"%12s %12.3f %12.3f %10.1f\n","Time:",
307 nodetime, realtime, 100.0*nodetime/realtime);
308 /* only print day-hour-sec format if realtime is more than 30 min */
309 if (realtime > 30*60)
311 fprintf(out,"%12s %12s","","");
312 pr_difftime(out,realtime);
316 mflop = mflop/realtime;
317 runtime = nsteps*delta_t;
319 if (getenv("GMX_DETAILED_PERF_STATS") == NULL)
321 fprintf(out,"%12s %12s %12s\n",
322 "","(ns/day)","(hour/ns)");
323 fprintf(out,"%12s %12.3f %12.3f\n","Performance:",
324 runtime*24*3.6/realtime,1000*realtime/(3600*runtime));
328 fprintf(out,"%12s %12s %12s %12s %12s\n",
329 "","(Mnbf/s)",(mflop > 1000) ? "(GFlops)" : "(MFlops)",
330 "(ns/day)","(hour/ns)");
331 fprintf(out,"%12s %12.3f %12.3f %12.3f %12.3f\n","Performance:",
332 nbfs/realtime,(mflop > 1000) ? (mflop/1000) : mflop,
333 runtime*24*3.6/realtime,1000*realtime/(3600*runtime));
338 if (getenv("GMX_DETAILED_PERF_STATS") == NULL)
340 fprintf(out,"%12s %14s\n",
342 fprintf(out,"%12s %14.1f\n","Performance:",
343 nsteps*3600.0/realtime);
347 fprintf(out,"%12s %12s %12s %14s\n",
348 "","(Mnbf/s)",(mflop > 1000) ? "(GFlops)" : "(MFlops)",
350 fprintf(out,"%12s %12.3f %12.3f %14.1f\n","Performance:",
351 nbfs/realtime,(mflop > 1000) ? (mflop/1000) : mflop,
352 nsteps*3600.0/realtime);
358 int cost_nrnb(int enr)
360 return nbdata[enr].flop;
363 const char *nrnb_str(int enr)
365 return nbdata[enr].name;
368 static const int force_index[]={
369 eNR_BONDS, eNR_ANGLES, eNR_PROPER, eNR_IMPROPER,
370 eNR_RB, eNR_DISRES, eNR_ORIRES, eNR_POSRES,
371 eNR_FBPOSRES, eNR_NS,
373 #define NFORCE_INDEX asize(force_index)
375 static const int constr_index[]={
376 eNR_SHAKE, eNR_SHAKE_RIJ, eNR_SETTLE, eNR_UPDATE, eNR_PCOUPL,
377 eNR_CONSTR_VIR,eNR_CONSTR_V
379 #define NCONSTR_INDEX asize(constr_index)
381 static double pr_av(FILE *log,t_commrec *cr,
382 double fav,double ftot[],const char *title)
389 fav /= cr->nnodes - cr->npmenodes;
390 fprintf(log,"\n %-26s",title);
391 for(i=0; (i<cr->nnodes); i++) {
392 dperc=(100.0*ftot[i])/fav;
395 fprintf(log,"%3d ",perc);
399 fprintf(log,"%6d%%\n\n",perc);
407 void pr_load(FILE *log,t_commrec *cr,t_nrnb nrnb[])
410 double dperc,unb,uf,us;
416 snew(ftot,cr->nnodes);
417 snew(stot,cr->nnodes);
419 for(i=0; (i<cr->nnodes); i++) {
420 add_nrnb(av,av,&(nrnb[i]));
421 /* Cost due to forces */
422 for(j=0; (j<eNR_NBKERNEL_ALLVSALLGB); j++)
423 ftot[i]+=nrnb[i].n[j]*cost_nrnb(j);
424 for(j=0; (j<NFORCE_INDEX); j++)
425 ftot[i]+=nrnb[i].n[force_index[j]]*cost_nrnb(force_index[j]);
427 for(j=0; (j<NCONSTR_INDEX); j++) {
428 stot[i]+=nrnb[i].n[constr_index[j]]*cost_nrnb(constr_index[j]);
431 for(j=0; (j<eNRNB); j++)
432 av->n[j]=av->n[j]/(double)(cr->nnodes - cr->npmenodes);
434 fprintf(log,"\nDetailed load balancing info in percentage of average\n");
436 fprintf(log," Type NODE:");
437 for(i=0; (i<cr->nnodes); i++)
438 fprintf(log,"%3d ",i);
439 fprintf(log,"Scaling\n");
440 fprintf(log,"---------------------------");
441 for(i=0; (i<cr->nnodes); i++)
443 fprintf(log,"-------\n");
445 for(j=0; (j<eNRNB); j++) {
448 fprintf(log," %-26s",nrnb_str(j));
449 for(i=0; (i<cr->nnodes); i++) {
450 dperc=(100.0*nrnb[i].n[j])/av->n[j];
453 fprintf(log,"%3d ",perc);
457 fprintf(log,"%6d%%\n",perc);
464 for(i=0; (i<cr->nnodes); i++) {
468 uf=pr_av(log,cr,fav,ftot,"Total Force");
469 us=pr_av(log,cr,sav,stot,"Total Constr.");
471 unb=(uf*fav+us*sav)/(fav+sav);
474 fprintf(log,"\nTotal Scaling: %.0f%% of max performance\n\n",unb);