source: pcp/src/helpers.c@ 774ae8

Last change on this file since 774ae8 was 9275a3, checked in by Frederik Heber <heber@…>, 17 years ago

Free(): now takes a debug string to know where free error occured
debug(): now depends on HAVE_DEBUG
  • Property mode set to 100644
File size: 21.7 KB
Line 
1/** \file helpers.c
2 * Malloc-, Realloc-, Free-wrappers and other small routines.
3 *
4 * This file contains small helpful routines that do such omnipresent tasks
5 * as allocation memory Malloc(), Malloci(), Mallocii(), reallocating Realloc(),
6 * Realloci(), Reallocii() or disallocating Free() with checks error messages
7 * in the case of failure. Speed Measuring is initiated InitSpeedMeasure() by setting
8 * SpeedStruct entries to zero,
9 * completed for output to file CompleteSpeedMeasure() and actually performed
10 * SpeedMeasure() herein.\n
11 * RemoveEverything() frees all memory that was alloceted during the course of the
12 * programme.\n
13 * StartParallel() initializes process groups and ranks, while WaitForOtherProcs()
14 * waits for the other running processes between the end of init phase and the begin
15 * of the calculation.\n
16 * Finally, GetRGB() does a colour translation of an RGB value from integer but it's
17 * not used anymore.
18 *
19 Project: ParallelCarParrinello
20 \author Jan Hamaekers
21 \date 2000
22
23 File: helpers.c
24 $Id: helpers.c,v 1.39 2007-02-12 09:44:55 foo Exp $
25*/
26
27#include <stdlib.h>
28#include <stdio.h>
29#include <stddef.h>
30#include <math.h>
31#include <string.h>
32#include "data.h"
33#include "errors.h"
34#include "helpers.h"
35#include "grad.h"
36#include "pseudo.h"
37#include "ions.h"
38#include "output.h"
39#include "run.h"
40#include "gramsch.h"
41#include "mymath.h"
42#include "myfft.h"
43#include "perturbed.h"
44
45/** Output of a debug message to stderr.
46 * \param *P Problem at hand, points to ParallelSimulationData#me
47 * \param output output string
48 */
49#ifdef HAVE_DEBUG
50void debug_in(struct Problem *P, const char *output, const char *file, const int line) {
51 if (output) fprintf(stderr,"(%i) DEBUG: in %s at line %i: %s\n",P->Par.me, file, line, output);
52}
53#else
54void debug_in(struct Problem *P, const char *output, const char *file, const int line) {} // print nothing
55#endif
56
57/** Malloc with error check.
58 * malloc() wrapper, printing error on null pointer
59 * \param size size of memory to be allocated
60 * \param output error message to be given on failure
61 */
62void* Malloc(size_t size, const char* output)
63{
64 void* dummy = malloc(size);
65 if (!dummy)
66 Error(MallocError, output);
67 return dummy;
68}
69
70/** Malloc string array and set its length to the allocated length.
71 * \param *output message if malloc fails
72 * \param size number of chars to alloc for \a *buffer
73 * \return pointer to string array
74 */
75char* MallocString(size_t size, const char* output) {
76 int i;
77 char *buffer;
78 buffer = Malloc(sizeof(char) * (size+1), output); // alloc
79 for (i=0;i<size;i++) // reset
80 buffer[i] = i % 2 == 0 ? 'p': 'c';
81 buffer[size] = '\0'; // and set length marker on its end
82 return(buffer);
83}
84
85
86/** Malloc with error check and 1 int.
87 * malloc() wrapper, printing error on null pointer
88 * \param size size of memory to be allocated
89 * \param output error message to be given on failure (containing one %d or similar)
90 * \param i integer value to be put into error message (%d in output)
91 */
92void* Malloci(size_t size, const char* output, int i)
93{
94 char dummyoutput[MAXDUMMYSTRING];
95 void* dummy = (void *) malloc(size);
96 if (!dummy) {
97 sprintf(dummyoutput,output,i);
98 Error(MallocError, dummyoutput);
99 }
100 return dummy;
101}
102
103/** Malloc with error check and 2 ints.
104 * malloc() wrapper, printing error on null pointer
105 * \param size size of memory to be allocated
106 * \param output error message to be given on failure (containing two %d or similar)
107 * \param i first integer value to be put into error message (%d in output)
108 * \param j second integer value to be put into error message (%d in output)
109 */
110void* Mallocii(size_t size, const char* output, int i, int j)
111{
112 char dummyoutput[MAXDUMMYSTRING];
113 void* dummy = (void *) malloc(size);
114 if (!dummy) {
115 sprintf(dummyoutput,output,i,j);
116 Error(MallocError, dummyoutput);
117 }
118 return dummy;
119}
120
121/** Change size of allocated memory range.
122 * Wrapper for realloc(), giving error message on failure
123 * \param pointer points to allocated memory
124 * \param size new desired size
125 * \param output error message on failure
126 */
127void* Realloc(void* pointer, size_t size, const char* output)
128{
129 void *dummy = (void *) realloc(pointer, size);
130 if (!dummy)
131 Error(MallocError, output);
132 return dummy;
133}
134
135/** Change size of allocated memory range, 1 int in error msg.
136 * Wrapper for realloc(), giving error message on failure
137 * \param pointer points to allocated memory
138 * \param size new desired size
139 * \param output error message on failure (containing one %d or similar)
140 * \param i integer in error message
141 */
142void* Realloci(void* pointer, size_t size, const char* output, int i)
143{
144 char dummyoutput[MAXDUMMYSTRING];
145 void *dummy = (void *) realloc(pointer, size);
146 if (!dummy) {
147 sprintf(dummyoutput, output, i);
148 Error(MallocError, dummyoutput);
149 }
150 return dummy;
151}
152
153/** Change size of allocated memory range, 2 int in error msg.
154 * Wrapper for realloc(), giving error message on failure
155 * \param pointer points to allocated memory
156 * \param size new desired size
157 * \param output error message on failure (containing two %d or similar)
158 * \param i first integer in error message
159 * \param j second integer in error message
160 */
161void* Reallocii(void* pointer, size_t size, const char* output, int i, int j)
162{
163 char dummyoutput[MAXDUMMYSTRING];
164 void *dummy = (void *) realloc(pointer, size);
165 if (!dummy) {
166 sprintf(dummyoutput,output,i,j);
167 Error(MallocError, dummyoutput);
168 }
169 return dummy;
170}
171
172/** Free memory with error check.
173 * Wrapper for free(), free is only called on given non-null pointer
174 * \param *ptr pointer to allocated memory region
175 * \param *output message to print on failure
176 */
177void Free (void *ptr, const char* output) {
178 if (ptr != NULL)
179 free(ptr);
180 else
181 fprintf(stderr,"Free not necessary: %s\n", output);
182}
183
184/** Translates a given integer into color code
185 * The value is interpreted by powers of the range into an rgb set of
186 * values.
187 * \param i specifies value within the range
188 * \param max specifies color range
189 * \param rgb pointer to double[3]
190 */
191void GetRGB(int i, unsigned int max, double* rgb) {
192 int basis,basis2,basis3;
193 int maxRGB,RGBint;
194 int rgbint[3];
195 i++;
196 if (max <= 1) max = 2; /* keine Division durch 0 */
197 basis = ceil(pow((double)(max+1),1./3.));
198 basis2 = basis*basis;
199 basis3 = basis2*basis;
200 maxRGB = basis3-1;
201 /* Koordinatentrans von 1..max -> 1..maxRGB */
202 RGBint = floor(1. + (double)(maxRGB - 1)/(double)(max-1)*(double)(i-1));
203 /* Transferiere RGBint jetzt in Basisdarstellung */
204 rgbint[2] = RGBint/basis2;
205 RGBint %= basis2;
206 rgbint[1] = RGBint/basis;
207 rgbint[0] = RGBint%basis;
208 /* Jetzt Basisdarstellung in rgb darstellung */
209 rgb[2] = rgbint[2]*(1./(basis-1));
210 rgb[1] = rgbint[1]*(1./(basis-1));
211 rgb[0] = rgbint[0]*(1./(basis-1));
212}
213
214/** Waits for other processes to finish.
215 * Simply waits until all processes reach this routine for the MPI_Gather to work
216 * \param *P Problem at hand
217 * \param out 1 - output ready note, 0 - no output
218 */
219void WaitForOtherProcs(struct Problem *P, int out) {
220 /* Warte, bis alle fertig sind! */
221 int buffint = 0;
222 MPI_Gather(&buffint, 0, MPI_INT, &buffint, 0, MPI_INT, 0, P->Par.comm);
223 if(out && P->Call.out[MinOut]) fprintf(stderr,"\nAll procs ready: %i\n",P->Par.me);
224}
225
226/** Initialization for parallel computing.
227 * determines process ids and defines mpi groups (aka Communicators), being
228 * the ones working on the same Psi (when doing fft) and those working on the
229 * same section (transposed GramSchmidt)
230 * \param *P Problem at hand
231 * \param **argv is actually unnecessary
232 */
233void StartParallel (struct Problem *P, char **argv) {
234 int info;
235 char processor_name[MPI_MAX_PROCESSOR_NAME];
236 char dummy[MAXDUMMYSTRING];
237 P->Par.procs = P->Par.proc[PEPsi]*P->Par.proc[PEGamma];
238 P->Par.me = -1; /* eigene Position noch unklar */
239 /* MPI Initialisierung */
240 argv = argv; /* wird nicht benoetigt */
241 /* mpi Gruppen definieren */
242 P->Par.world = MPI_COMM_WORLD;
243 MPI_Comm_dup(P->Par.world, &P->Par.comm);
244 MPI_Comm_size(P->Par.comm, &info);
245 if (info != P->Par.procs) {
246 sprintf(dummy,"StartParallel. Wrong number of processes started %i != %i",info, P->Par.procs);
247 Error(SomeError, dummy);
248 }
249 MPI_Comm_rank(P->Par.comm, &P->Par.mytid);
250 MPI_Get_processor_name(processor_name, &info);
251 P->Par.me = P->Par.mytid;
252 P->Par.mypos[PEGamma] = P->Par.me % P->Par.proc[PEGamma];
253 P->Par.mypos[PEPsi] = P->Par.me / P->Par.proc[PEGamma];
254 if(P->Call.out[NormalOut]) fprintf(stderr,"Process %d(%i,%i) on %s \n", P->Par.mytid, P->Par.mypos[PEPsi], P->Par.mypos[PEGamma], processor_name);
255 if (P->Par.me == 0) { /* Master */
256 if (P->Par.procs > 9999)
257 fprintf(stderr, "(%d)Warning: procs>9999 will cause trouble with output files\n", P->Par.mytid);
258 }
259 if(P->Call.out[MinOut]) fprintf(stderr,"Info: %i processes started\n", P->Par.procs); /* Prozesse gestartet */
260
261
262 /* Erstellung der Communicatoren */
263 switch (P->Lat.Psi.Use) {
264 case UseSpinDouble:
265 P->Par.my_color_comm_ST = (int)SpinDouble;
266 P->Par.me_comm_ST = P->Par.me;
267 P->Par.Max_my_color_comm_ST = 1;
268 P->Par.Max_me_comm_ST = P->Par.procs;
269 break;
270 case UseSpinUpDown:
271 P->Par.my_color_comm_ST = P->Par.me / (P->Par.procs/2);
272 P->Par.me_comm_ST = P->Par.me % (P->Par.procs/2);
273 P->Par.Max_my_color_comm_ST = 2;
274 P->Par.Max_me_comm_ST = P->Par.procs/2;
275 break;
276 }
277 MPI_Comm_split (P->Par.comm, P->Par.my_color_comm_ST, P->Par.me_comm_ST, &P->Par.comm_ST);
278 if (P->Lat.Psi.Use == UseSpinUpDown) {
279 MPI_Intercomm_create ( P->Par.comm_ST, 0, P->Par.comm,
280 (P->Par.my_color_comm_ST ? 0 : P->Par.procs/2)
281 , InterTag1, &P->Par.comm_STInter);
282 }
283
284 /* Alle Procs mit gleichen Psis (Sind alle an einer fft beteiligt)*/
285 P->Par.my_color_comm_ST_Psi = P->Par.me_comm_ST / P->Par.proc[PEGamma];
286 P->Par.me_comm_ST_Psi = P->Par.me_comm_ST % P->Par.proc[PEGamma];
287 P->Par.Max_my_color_comm_ST_Psi = P->Par.Max_me_comm_ST/P->Par.proc[PEGamma];
288 P->Par.Max_me_comm_ST_Psi = P->Par.proc[PEGamma];
289 MPI_Comm_split (P->Par.comm_ST, P->Par.my_color_comm_ST_Psi, P->Par.me_comm_ST_Psi, &P->Par.comm_ST_Psi);
290
291 /* Alle Procs mit gleichen PsisAbschnitt (Transposed Psi - GramSch)*/
292 P->Par.my_color_comm_ST_PsiT = P->Par.me_comm_ST % P->Par.proc[PEGamma];
293 P->Par.me_comm_ST_PsiT = P->Par.me_comm_ST / P->Par.proc[PEGamma];
294 P->Par.Max_my_color_comm_ST_PsiT = P->Par.proc[PEGamma];
295 P->Par.Max_me_comm_ST_PsiT = P->Par.Max_me_comm_ST/P->Par.proc[PEGamma];
296 MPI_Comm_split (P->Par.comm_ST, P->Par.my_color_comm_ST_PsiT, P->Par.me_comm_ST_PsiT, &P->Par.comm_ST_PsiT);
297}
298
299
300/** Removes everything from memory.
301 * Frees memory and exits program
302 * \param *P Problem at hand
303 */
304void RemoveEverything(struct Problem *P)
305{
306 struct Lattice *Lat = &P->Lat;
307 struct Psis *Psi = &Lat->Psi;
308 int i,d, type;
309 if (P->Call.out[NormalOut]) fprintf(stderr,"(%i)RemoveEverything !\n",P->Par.me);
310 RemoveGradients(P, &P->Grad);
311 RemovePseudoRead(P);
312 RemoveIonsRead(&P->Ion);
313 fft_3d_destroy_plan(P,P->Lat.plan);
314 if (P->Files.MeOutVis) {
315 Free(P->Files.PosTemp, "RemoveEverything: P->Files.PosTemp");
316 Free(P->Files.OutputPosType, "RemoveEverything: P->Files.OutputPosType");
317 //Free(P->Files.OutVisStep, "RemoveEverything: P->Files.OutVisStep");
318 }
319 Free(P->Files.mainname, "RemoveEverything: P->Files.mainname");
320 Free(P->Files.mainpath, "RemoveEverything: P->Files.mainpath");
321 Free(P->Call.MainParameterFile, "RemoveEverything: P->Call.MainParameterFile");
322 if (P->Call.ForcesFile != NULL)
323 Free(P->Call.ForcesFile, "RemoveEverything: P->Call.ForcesFile");
324 Free(Lat->MaxNoOfnFields, "RemoveEverything: Lat->MaxNoOfnFields");
325 for (i=0; i < P->Lat.MaxLevel; i++) {
326 Free(Lat->Lev[i].AllMaxG, "RemoveEverything: Lat->Lev[i].AllMaxG");
327 Free(Lat->NFields[i], "RemoveEverything: Lat->NFields");
328 Free(Lat->Lev[i].GArray, "RemoveEverything: Lat->Lev[i].GArray");
329 Free(Lat->Lev[i].HashG, "RemoveEverything: Lat->Lev[i].HashG");
330 if (Lat->Lev[i].MaxDoubleG)
331 Free(Lat->Lev[i].DoubleG, "RemoveEverything: Lat->Lev[i].DoubleG");
332 if (i != 0) {
333 Free(Lat->Lev[i].PosFactorUp, "RemoveEverything: Lat->Lev[i].PosFactorUp");
334 } else {
335 for (d=0; d < (P->R.DoPerturbation == 1 ? MaxDensityTypes : MaxInitDensityTypes); d++) {
336 if (Lat->Lev[i].Dens->DensityArray[d] != NULL) Free(Lat->Lev[i].Dens->DensityArray[d], "RemoveEverything: Lat->Lev[i].Dens->DensityArray[d]");
337 if (Lat->Lev[i].Dens->DensityCArray[d] != NULL) Free(Lat->Lev[i].Dens->DensityCArray[d], "RemoveEverything: Lat->Lev[i].Dens->DensityCArray[d]");
338 }
339 }
340 /* if (i != Lat->MaxLevel-1) */
341 Free(Lat->Lev[i].Dens, "RemoveEverything: Lat->Lev[i].Dens");
342 }
343 for (i=1; i < Lat->MaxLevel; i++) {
344 Free(Lat->Lev[i].LPsi->LocalPsi, "RemoveEverything: Lat->Lev[i].LPsi->LocalPsi");
345 Free(Lat->Lev[i].LPsi->OldLocalPsi, "RemoveEverything: Lat->Lev[i].LPsi->OldLocalPsi");
346 if (i == 1) {
347 Free(Lat->Lev[i].LPsi->PsiDat, "RemoveEverything: Lat->Lev[i].LPsi->PsiDat");
348 Free(Lat->Lev[i].LPsi->OldPsiDat, "RemoveEverything: Lat->Lev[i].LPsi->OldPsiDat");
349 }
350 Free(Lat->Lev[i].LPsi, "RemoveEverything: Lat->Lev[i].LPsi");
351 }
352 for (i=0;i<Lat->Psi.NoOfTotalPsis;i++) {
353 Free(Psi->lambda[i], "RemoveEverything: Psi->lambda[i]");
354 }
355 for (i=0;i<Lat->Psi.NoOfPsis;i++) {
356 Free(Psi->Overlap[i], "RemoveEverything: Psi->Overlap[i]");
357 }
358 Free(Psi->lambda, "RemoveEverything: Psi->lambda");
359 Free(Psi->Overlap, "RemoveEverything: Psi->Overlap");
360 Free(Psi->AllPsiStatus, "RemoveEverything: Psi->AllPsiStatus");
361 Free(Psi->AllPsiStatusForSort, "RemoveEverything: Psi->AllPsiStatusForSort");
362 Free(Psi->LocalPsiStatus, "RemoveEverything: Psi->LocalPsiStatus");
363 Free(Psi->AllLocalNo, "RemoveEverything: Psi->AllLocalNo");
364 Free(Psi->RealAllLocalNo, "RemoveEverything: Psi->RealAllLocalNo");
365 Free(Psi->TempSendA, "RemoveEverything: Psi->TempSendA");
366 Free(Psi->AddData, "RemoveEverything: Psi->AddData");
367 Free(Psi->AllActualLocalPsiNo, "RemoveEverything: Psi->AllActualLocalPsiNo");
368 Free(Psi->AllOldActualLocalPsiNo, "RemoveEverything: Psi->AllOldActualLocalPsiNo");
369 Free(Lat->Lev, "RemoveEverything: Lat->Lev");
370 Free(Lat->LevelSizes, "RemoveEverything: Lat->LevelSizes");
371 Free(Lat->NFields, "RemoveEverything: Lat->NFields");
372 if (Lat->RT.Use == UseRT) {
373 Free(Lat->RT.Coeff, "RemoveEverything: Lat->RT.Coeff");
374 Free(Lat->RT.RTLaplaceS, "RemoveEverything: Lat->RT.RTLaplaceS");
375 Free(Lat->RT.RTLaplace0, "RemoveEverything: Lat->RT.RTLaplace0");
376 for (i=0; i< MAXRTPOSFAC; i++) Free(Lat->RT.PosFactor[i], "RemoveEverything: Lat->RT.PosFactor[i]");
377 for (i=0; i< MAXRTARRAYS; i++) {
378 Free(Lat->RT.DensityC[i], "RemoveEverything: Lat->RT.DensityC[i]");
379 }
380 Free(Lat->RT.TempC, "RemoveEverything: Lat->RT.TempC");
381 }
382 for (type=Occupied;type<Extra;type++)
383 for (i=0; i<MAXALLPSIENERGY; i++) {
384 Free(Lat->Energy[type].PsiEnergy[i], "RemoveEverything: Lat->Energy[type].PsiEnergy[i]");
385 }
386 for (i=Occupied;i<Extra;i++)
387 Free(P->R.MinimisationName[i], "RemoveEverything: P->R.MinimisationName[i]");
388 Free(P->R.MinimisationName, "RemoveEverything: P->R.MinimisationName");
389 MPI_Comm_free(&P->Par.comm_ST_PsiT);
390 //debug(P,"comm_ST_PsiT");
391 MPI_Comm_free(&P->Par.comm_ST_Psi);
392 //debug(P,"comm_ST_Psi");
393 if (Psi->Use == UseSpinUpDown) MPI_Comm_free(&P->Par.comm_STInter);
394 MPI_Comm_free(&P->Par.comm_ST);
395 //debug(P,"comm_ST");
396 MPI_Comm_free(&P->Par.comm);
397 //debug(P,"comm");
398 Free(P, "RemoveEverything: P");
399 FreeMPI_OnePsiElement();
400 //debug(P,"OnePsiElement");
401
402 // Free string names from Files structure
403 Free(P->Files.filename, "RemoveEverything: P->Files.filename");
404 Free(P->Files.default_path, "RemoveEverything: P->Files.default_path");
405 Free(P->Files.pseudopot_path, "RemoveEverything: P->Files.pseudopot_path");
406}
407
408static const char suffixspeed[] = ".speed"; //!< suffix of the FileData#SpeedFile where the timings are written to
409
410
411/** Initializes the timer array.
412 * Sets every time, average, min/max and deviation to zero.
413 * SpeedStep is set to 1.
414 * Opens Filedata::SpeedFile for appended writing
415 * \param *P Problem at hand
416 */
417void InitSpeedMeasure(struct Problem *P)
418{
419 struct FileData *F = &P->Files;
420 struct SpeedStruct *S = &P->Speed;
421 int i;
422 F->SpeedFile = NULL;
423 if (P->Par.me == 0)
424 OpenFile(P, &F->SpeedFile, suffixspeed, "a",P->Call.out[ReadOut]);
425 for (i=0; i < MAXTIMETYPES; i++)
426 S->SpeedStep[i] = 1;
427 SetArrayToDouble0(S->time1,MAXTIMETYPES);
428 SetArrayToDouble0(S->time2,MAXTIMETYPES);
429 SetArrayToDouble0(S->time,MAXTIMETYPES);
430 SetArrayToDouble0(S->average,MAXTIMETYPES);
431 SetArrayToDouble0(S->min,MAXTIMETYPES);
432 SetArrayToDouble0(S->max,MAXTIMETYPES);
433 SetArrayToDouble0(S->stddev,MAXTIMETYPES);
434 S->InitSteps = 0;
435 S->LevUpSteps = 0;
436 S->Steps = 0;
437}
438
439/** Measure speed of a routine.
440 * Points to SpeedStruct of the Problem, gets current time and depending on
441 * \a TOT (whether we start or stop the watch) and the timing group \a TT
442 * puts time in SpeedStruct::time1 (start) or in SpeedStruct::time2 (stop)
443 * and in SpeedStruct::time the difference between the two is added.
444 * \param *P Problem at hand containing SpeedStruct
445 * \param TT TimeTypes
446 * \param TOT TimeDotypes
447 */
448void SpeedMeasure(struct Problem *P, enum TimeTypes TT, enum TimeDoTypes TOT)
449{
450 struct SpeedStruct *S = &P->Speed;
451 double timeA = GetTime();
452 switch (TOT) {
453 case StartTimeDo:
454 S->time1[TT] = timeA;
455 break;
456 case StopTimeDo:
457 S->time2[TT] = timeA;
458 S->time[TT] += (S->time2[TT]-S->time1[TT]);
459 break;
460 }
461}
462
463/** Final Measuring and overall time.
464 * Does the last measurement calculation and combines results from all processes
465 * to calculate min, max, average and writes it to FileData::SpeedFile
466 * \param *P Problem at hand
467 */
468void CompleteSpeedMeasure(struct Problem *P)
469{ /* gibt letzte Messung aus und benoetigte Gesamtzeit */
470 struct SpeedStruct *S = &P->Speed;
471 struct FileData *F = &P->Files;
472 struct Lattice *Lat = &P->Lat;
473 struct Psis *Psi = &Lat->Psi;
474 struct RunStruct *R = &P->R;
475 struct Ions *I = &P->Ion;
476 double average[MAXTIMETYPES];
477 double stddev[MAXTIMETYPES];
478 double time[MAXTIMETYPES];
479 int i;
480 S->InitSteps += S->LevUpSteps;
481 for (i=0; i<MAXTIMETYPES; i++)
482 switch ((enum TimeTypes)i) {
483 case InitSimTime:
484 case InitGramSchTime:
485 case InitLocTime:
486 case InitNonLocTime:
487 case InitDensityTime:
488 case LocFTime:
489 case NonLocFTime:
490 case EwaldTime:
491 case GapTime:
492 S->SpeedStep[i] = S->InitSteps;
493 break;
494 case SimTime:
495 case GramSchTime:
496 case LocTime:
497 case NonLocTime:
498 case DensityTime:
499 case WannierTime:
500 case ReadnWriteTime:
501 S->SpeedStep[i] = S->Steps*P->Par.proc[0]/(double)P->Lat.Psi.GlobalNo[PsiMaxNo];
502 break;
503 default:
504 S->SpeedStep[i] = 1;
505 }
506 S->time[LevSMaxG] = R->LevS->MaxG;
507 MPI_Allreduce( S->time, time, MAXTIMETYPES, MPI_DOUBLE, MPI_SUM, P->Par.comm);
508 for (i=0; i<MAXTIMETYPES; i++)
509 time[i] /= (double)P->Par.procs;
510 time[LevSMaxG] = R->LevS->TotalAllMaxG;
511 for (i=0; i<MAXTIMETYPES; i++)
512 average[i] = S->time[i]/S->SpeedStep[i];
513 MPI_Allreduce( average, S->average, MAXTIMETYPES, MPI_DOUBLE, MPI_SUM, P->Par.comm);
514 for (i=0; i<MAXTIMETYPES; i++)
515 S->average[i] /= (double)P->Par.procs;
516 S->average[LevSMaxG] = R->LevS->TotalAllMaxG/(double)P->Par.proc[1];
517 for (i=0; i<MAXTIMETYPES; i++)
518 stddev[i] = (average[i]-S->average[i])*(average[i]-S->average[i]);
519 MPI_Allreduce( stddev, S->stddev, MAXTIMETYPES, MPI_DOUBLE, MPI_SUM, P->Par.comm);
520 MPI_Allreduce( &stddev[LevSMaxG], &S->stddev[LevSMaxG], 1, MPI_DOUBLE, MPI_SUM, P->Par.comm_ST_Psi);
521 for (i=0; i<MAXTIMETYPES-1; i++)
522 S->stddev[i] /= (double)P->Par.procs;
523 S->stddev[LevSMaxG] /= (double)P->Par.proc[1];
524 for (i=0; i<MAXTIMETYPES; i++)
525 S->stddev[i] = sqrt(S->stddev[i]);
526
527 MPI_Allreduce( average, S->max, MAXTIMETYPES, MPI_DOUBLE, MPI_MAX, P->Par.comm);
528 MPI_Allreduce( average, S->min, MAXTIMETYPES, MPI_DOUBLE, MPI_MIN, P->Par.comm);
529 if (P->Par.me != 0) return;
530
531 fprintf(F->SpeedFile, "LevNo\tMaxG\tRMaxG\tMaxN\tN[0]\tN[1]\tN[2]\tRLevSStep\n");
532 for (i=0; i < Lat->MaxLevel; i++)
533 fprintf(F->SpeedFile, "%i\t%i\t%i\t%i\t%i\t%i\t%i\t%i\n", Lat->Lev[i].LevelNo, Lat->Lev[i].TotalAllMaxG, Lat->Lev[i].TotalRealAllMaxG, Lat->Lev[i].MaxN, Lat->Lev[i].N[0], Lat->Lev[i].N[1], Lat->Lev[i].N[2], Lat->Lev[i].Step);
534 fprintf(F->SpeedFile, "procs\tproc[0]\tproc[1]\tPsiMax\tPsiDoub\tPsiUp\tPsiDown\tNRStep\tTotalIons\n");
535 fprintf(F->SpeedFile, "%i\t%i\t%i\t%i\t%i+%i\t%i+%i\t%i+%i\t%i\t%i\n", P->Par.procs, P->Par.proc[0], P->Par.proc[1], Psi->GlobalNo[PsiMaxNo], Psi->GlobalNo[PsiMaxNoDouble], Psi->GlobalNo[PsiMaxAdd], Psi->GlobalNo[PsiMaxNoUp], Psi->GlobalNo[PsiMaxAdd], Psi->GlobalNo[PsiMaxNoDown], Psi->GlobalNo[PsiMaxAdd], R->NewRStep, I->Max_TotalIons);
536
537 fprintf(F->SpeedFile, "Type\tSimTime\t\tInitSimTime\tInitTime\tInitGramSchTime\tInitLocTime\tInitNonLocTime\tInitDensTime\tGramSchTime\tLocTime\t\tNonLocTime\tDensTime\tLocFTime\tNonLocFTime\tEwaldTime\tGapTime\tCurrDensTime\tWannierTime\tReadnWriteTime\tLevSMaxG\n");
538 fprintf(F->SpeedFile, "Steps");
539 for(i=0;i<MAXTIMETYPES;i++)
540 fprintf(F->SpeedFile, "\t%e", S->SpeedStep[i]);
541 fprintf(F->SpeedFile, "\n");
542
543 fprintf(F->SpeedFile, "total");
544 for(i=0;i<MAXTIMETYPES;i++)
545 fprintf(F->SpeedFile, "\t%e", time[i]);
546 fprintf(F->SpeedFile, "\n");
547
548 fprintf(F->SpeedFile, "average");
549 for(i=0;i<MAXTIMETYPES;i++)
550 fprintf(F->SpeedFile, "\t%e", S->average[i]);
551 fprintf(F->SpeedFile, "\n");
552
553 fprintf(F->SpeedFile, "stddev");
554 for(i=0;i<MAXTIMETYPES;i++)
555 fprintf(F->SpeedFile, "\t%e", S->stddev[i]);
556 fprintf(F->SpeedFile, "\n");
557
558 fprintf(F->SpeedFile, "min");
559 for(i=0;i<MAXTIMETYPES;i++)
560 fprintf(F->SpeedFile, "\t%e", S->min[i]);
561 fprintf(F->SpeedFile, "\n");
562
563 fprintf(F->SpeedFile, "max");
564 for(i=0;i<MAXTIMETYPES;i++)
565 fprintf(F->SpeedFile, "\t%e", S->max[i]);
566 fprintf(F->SpeedFile, "\n");
567
568 fclose(F->SpeedFile);
569}
Note: See TracBrowser for help on using the repository browser.