Actual source code: snesnoise.c

  1: #include <petsc/private/snesimpl.h>

  3: PETSC_INTERN PetscErrorCode SNESDiffParameterCreate_More(SNES, Vec, void **);
  4: PETSC_INTERN PetscErrorCode SNESDiffParameterCompute_More(SNES, void *, Vec, Vec, PetscReal *, PetscReal *);
  5: PETSC_INTERN PetscErrorCode SNESDiffParameterDestroy_More(void *);

  7: /* Data used by Jorge's diff parameter computation method */
  8: typedef struct {
  9:   Vec     *workv;          /* work vectors */
 10:   FILE    *fp;             /* output file */
 11:   PetscInt function_count; /* count of function evaluations for diff param estimation */
 12:   double   fnoise_min;     /* minimum allowable noise */
 13:   double   hopt_min;       /* minimum allowable hopt */
 14:   double   h_first_try;    /* first try for h used in diff parameter estimate */
 15:   PetscInt fnoise_resets;  /* number of times we've reset the noise estimate */
 16:   PetscInt hopt_resets;    /* number of times we've reset the hopt estimate */
 17: } DIFFPAR_MORE;

 19: PETSC_INTERN PetscErrorCode SNESUnSetMatrixFreeParameter(SNES snes);
 20: PETSC_INTERN PetscErrorCode SNESNoise_dnest_(PetscInt *, PetscScalar *, PetscScalar *, PetscScalar *, PetscScalar *, PetscScalar *, PetscInt *, PetscScalar *);

 22: static PetscErrorCode JacMatMultCompare(SNES, Vec, Vec, double);

 24: PetscErrorCode SNESDiffParameterCreate_More(SNES snes, Vec x, void **outneP)
 25: {
 26:   DIFFPAR_MORE *neP;
 27:   Vec           w;
 28:   PetscRandom   rctx; /* random number generator context */
 29:   PetscBool     flg;
 30:   char          noise_file[PETSC_MAX_PATH_LEN];

 32:   PetscFunctionBegin;
 33:   PetscCall(PetscNew(&neP));

 35:   neP->function_count = 0;
 36:   neP->fnoise_min     = 1.0e-20;
 37:   neP->hopt_min       = 1.0e-8;
 38:   neP->h_first_try    = 1.0e-3;
 39:   neP->fnoise_resets  = 0;
 40:   neP->hopt_resets    = 0;

 42:   /* Create work vectors */
 43:   PetscCall(VecDuplicateVecs(x, 3, &neP->workv));
 44:   w = neP->workv[0];

 46:   /* Set components of vector w to random numbers */
 47:   PetscCall(PetscRandomCreate(PetscObjectComm((PetscObject)snes), &rctx));
 48:   PetscCall(PetscRandomSetFromOptions(rctx));
 49:   PetscCall(VecSetRandom(w, rctx));
 50:   PetscCall(PetscRandomDestroy(&rctx));

 52:   /* Open output file */
 53:   PetscCall(PetscOptionsGetString(((PetscObject)snes)->options, ((PetscObject)snes)->prefix, "-snes_mf_noise_file", noise_file, sizeof(noise_file), &flg));
 54:   if (flg) neP->fp = fopen(noise_file, "w");
 55:   else neP->fp = fopen("noise.out", "w");
 56:   PetscCheck(neP->fp, PETSC_COMM_SELF, PETSC_ERR_FILE_OPEN, "Cannot open file");
 57:   PetscCall(PetscInfo(snes, "Creating Jorge's differencing parameter context\n"));

 59:   *outneP = neP;
 60:   PetscFunctionReturn(PETSC_SUCCESS);
 61: }

 63: PetscErrorCode SNESDiffParameterDestroy_More(void *nePv)
 64: {
 65:   DIFFPAR_MORE *neP = (DIFFPAR_MORE *)nePv;
 66:   int           err;

 68:   PetscFunctionBegin;
 69:   /* Destroy work vectors and close output file */
 70:   PetscCall(VecDestroyVecs(3, &neP->workv));
 71:   err = fclose(neP->fp);
 72:   PetscCheck(!err, PETSC_COMM_SELF, PETSC_ERR_SYS, "fclose() failed on file");
 73:   PetscCall(PetscFree(neP));
 74:   PetscFunctionReturn(PETSC_SUCCESS);
 75: }

 77: PetscErrorCode SNESDiffParameterCompute_More(SNES snes, void *nePv, Vec x, Vec p, double *fnoise, double *hopt)
 78: {
 79:   DIFFPAR_MORE *neP = (DIFFPAR_MORE *)nePv;
 80:   Vec           w, xp, fvec; /* work vectors to use in computing h */
 81:   double        zero = 0.0, hl, hu, h, fnoise_s, fder2_s;
 82:   PetscScalar   alpha;
 83:   PetscScalar   fval[7], tab[7][7], eps[7], f = -1;
 84:   double        rerrf = -1., fder2;
 85:   PetscInt      iter, k, i, j, info;
 86:   PetscInt      nf = 7; /* number of function evaluations */
 87:   PetscInt      fcount;
 88:   MPI_Comm      comm;
 89:   FILE         *fp;
 90:   PetscBool     noise_test = PETSC_FALSE;

 92:   PetscFunctionBegin;
 93:   PetscCall(PetscObjectGetComm((PetscObject)snes, &comm));
 94:   /* Call to SNESSetUp() just to set data structures in SNES context */
 95:   if (!snes->setupcalled) PetscCall(SNESSetUp(snes));

 97:   w    = neP->workv[0];
 98:   xp   = neP->workv[1];
 99:   fvec = neP->workv[2];
100:   fp   = neP->fp;

102:   /* Initialize parameters */
103:   hl       = zero;
104:   hu       = zero;
105:   h        = neP->h_first_try;
106:   fnoise_s = zero;
107:   fder2_s  = zero;
108:   fcount   = neP->function_count;

110:   /* We have 5 tries to attempt to compute a good hopt value */
111:   PetscCall(SNESGetIterationNumber(snes, &i));
112:   PetscCall(PetscFPrintf(comm, fp, "\n ------- SNES iteration %" PetscInt_FMT " ---------\n", i));
113:   for (iter = 0; iter < 5; iter++) {
114:     neP->h_first_try = h;

116:     /* Compute the nf function values needed to estimate the noise from
117:        the difference table */
118:     for (k = 0; k < nf; k++) {
119:       alpha = h * (k + 1 - (nf + 1) / 2);
120:       PetscCall(VecWAXPY(xp, alpha, p, x));
121:       PetscCall(SNESComputeFunction(snes, xp, fvec));
122:       neP->function_count++;
123:       PetscCall(VecDot(fvec, w, &fval[k]));
124:     }
125:     f = fval[(nf + 1) / 2 - 1];

127:     /* Construct the difference table */
128:     for (i = 0; i < nf; i++) tab[i][0] = fval[i];

130:     for (j = 0; j < nf - 1; j++) {
131:       for (i = 0; i < nf - j - 1; i++) tab[i][j + 1] = tab[i + 1][j] - tab[i][j];
132:     }

134:     /* Print the difference table */
135:     PetscCall(PetscFPrintf(comm, fp, "Difference Table: iter = %" PetscInt_FMT "\n", iter));
136:     for (i = 0; i < nf; i++) {
137:       for (j = 0; j < nf - i; j++) PetscCall(PetscFPrintf(comm, fp, " %10.2e ", tab[i][j]));
138:       PetscCall(PetscFPrintf(comm, fp, "\n"));
139:     }

141:     /* Call the noise estimator */
142:     PetscCall(SNESNoise_dnest_(&nf, fval, &h, fnoise, &fder2, hopt, &info, eps));

144:     /* Output statements */
145:     rerrf = *fnoise / PetscAbsScalar(f);
146:     if (info == 1) PetscCall(PetscFPrintf(comm, fp, "%s\n", "Noise detected"));
147:     if (info == 2) PetscCall(PetscFPrintf(comm, fp, "%s\n", "Noise not detected; h is too small"));
148:     if (info == 3) PetscCall(PetscFPrintf(comm, fp, "%s\n", "Noise not detected; h is too large"));
149:     if (info == 4) PetscCall(PetscFPrintf(comm, fp, "%s\n", "Noise detected, but unreliable hopt"));
150:     PetscCall(PetscFPrintf(comm, fp, "Approximate epsfcn %g  %g  %g  %g  %g  %g\n", (double)eps[0], (double)eps[1], (double)eps[2], (double)eps[3], (double)eps[4], (double)eps[5]));
151:     PetscCall(PetscFPrintf(comm, fp, "h = %g, fnoise = %g, fder2 = %g, rerrf = %g, hopt = %g\n\n", (double)h, (double)*fnoise, (double)fder2, (double)rerrf, (double)*hopt));

153:     /* Save fnoise and fder2. */
154:     if (*fnoise) fnoise_s = *fnoise;
155:     if (fder2) fder2_s = fder2;

157:     /* Check for noise detection. */
158:     if (fnoise_s && fder2_s) {
159:       *fnoise = fnoise_s;
160:       fder2   = fder2_s;
161:       *hopt   = 1.68 * sqrt(*fnoise / PetscAbsScalar(fder2));
162:       goto theend;
163:     } else {
164:       /* Update hl and hu, and determine new h */
165:       if (info == 2 || info == 4) {
166:         hl = h;
167:         if (hu == zero) h = 100 * h;
168:         else h = PetscMin(100 * h, 0.1 * hu);
169:       } else if (info == 3) {
170:         hu = h;
171:         h  = PetscMax(1.0e-3, sqrt(hl / hu)) * hu;
172:       }
173:     }
174:   }
175: theend:

177:   if (*fnoise < neP->fnoise_min) {
178:     PetscCall(PetscFPrintf(comm, fp, "Resetting fnoise: fnoise1 = %g, fnoise_min = %g\n", (double)*fnoise, (double)neP->fnoise_min));
179:     *fnoise = neP->fnoise_min;
180:     neP->fnoise_resets++;
181:   }
182:   if (*hopt < neP->hopt_min) {
183:     PetscCall(PetscFPrintf(comm, fp, "Resetting hopt: hopt1 = %g, hopt_min = %g\n", (double)*hopt, (double)neP->hopt_min));
184:     *hopt = neP->hopt_min;
185:     neP->hopt_resets++;
186:   }

188:   PetscCall(PetscFPrintf(comm, fp, "Errors in derivative:\n"));
189:   PetscCall(PetscFPrintf(comm, fp, "f = %g, fnoise = %g, fder2 = %g, hopt = %g\n", (double)f, (double)*fnoise, (double)fder2, (double)*hopt));

191:   /* For now, compute h **each** MV Mult!! */
192:   /*
193:   PetscCall(PetscOptionsHasName(NULL,"-matrix_free_jorge_each_mvp",&flg));
194:   if (!flg) {
195:     Mat mat;
196:     PetscCall(SNESGetJacobian(snes,&mat,NULL,NULL));
197:     PetscCall(MatSNESMFMoreSetParameters(mat,PETSC_DEFAULT,PETSC_DEFAULT,*hopt));
198:   }
199:   */
200:   fcount = neP->function_count - fcount;
201:   PetscCall(PetscInfo(snes, "fct_now = %" PetscInt_FMT ", fct_cum = %" PetscInt_FMT ", rerrf=%g, sqrt(noise)=%g, h_more=%g\n", fcount, neP->function_count, (double)rerrf, (double)PetscSqrtReal(*fnoise), (double)*hopt));

203:   PetscCall(PetscOptionsGetBool(NULL, NULL, "-noise_test", &noise_test, NULL));
204:   if (noise_test) PetscCall(JacMatMultCompare(snes, x, p, *hopt));
205:   PetscFunctionReturn(PETSC_SUCCESS);
206: }

208: static PetscErrorCode JacMatMultCompare(SNES snes, Vec x, Vec p, double hopt)
209: {
210:   Vec         yy1, yy2; /* work vectors */
211:   PetscViewer view2;    /* viewer */
212:   Mat         J;        /* analytic Jacobian (set as preconditioner matrix) */
213:   Mat         Jmf;      /* matrix-free Jacobian (set as true system matrix) */
214:   double      h;        /* differencing parameter */
215:   Vec         f;
216:   PetscScalar alpha;
217:   PetscReal   yy1n, yy2n, enorm;
218:   PetscInt    i;
219:   PetscBool   printv = PETSC_FALSE;
220:   char        filename[32];
221:   MPI_Comm    comm;

223:   PetscFunctionBegin;
224:   PetscCall(PetscObjectGetComm((PetscObject)snes, &comm));
225:   /* Compute function and analytic Jacobian at x */
226:   PetscCall(SNESGetJacobian(snes, &Jmf, &J, NULL, NULL));
227:   PetscCall(SNESComputeJacobian(snes, x, Jmf, J));
228:   PetscCall(SNESGetFunction(snes, &f, NULL, NULL));
229:   PetscCall(SNESComputeFunction(snes, x, f));

231:   /* Duplicate work vectors */
232:   PetscCall(VecDuplicate(x, &yy2));
233:   PetscCall(VecDuplicate(x, &yy1));

235:   /* Compute true matrix-vector product */
236:   PetscCall(MatMult(J, p, yy1));
237:   PetscCall(VecNorm(yy1, NORM_2, &yy1n));

239:   /* View product vector if desired */
240:   PetscCall(PetscOptionsGetBool(NULL, NULL, "-print_vecs", &printv, NULL));
241:   if (printv) {
242:     PetscCall(PetscViewerASCIIOpen(comm, "y1.out", &view2));
243:     PetscCall(PetscViewerPushFormat(view2, PETSC_VIEWER_ASCII_COMMON));
244:     PetscCall(VecView(yy1, view2));
245:     PetscCall(PetscViewerPopFormat(view2));
246:     PetscCall(PetscViewerDestroy(&view2));
247:   }

249:   /* Test Jacobian-vector product computation */
250:   alpha = -1.0;
251:   h     = 0.01 * hopt;
252:   for (i = 0; i < 5; i++) {
253:     /* Set differencing parameter for matrix-free multiplication */
254:     PetscCall(MatSNESMFMoreSetParameters(Jmf, PETSC_DEFAULT, PETSC_DEFAULT, h));

256:     /* Compute matrix-vector product via differencing approximation */
257:     PetscCall(MatMult(Jmf, p, yy2));
258:     PetscCall(VecNorm(yy2, NORM_2, &yy2n));

260:     /* View product vector if desired */
261:     if (printv) {
262:       PetscCall(PetscSNPrintf(filename, PETSC_STATIC_ARRAY_LENGTH(filename), "y2.%d.out", (int)i));
263:       PetscCall(PetscViewerASCIIOpen(comm, filename, &view2));
264:       PetscCall(PetscViewerPushFormat(view2, PETSC_VIEWER_ASCII_COMMON));
265:       PetscCall(VecView(yy2, view2));
266:       PetscCall(PetscViewerPopFormat(view2));
267:       PetscCall(PetscViewerDestroy(&view2));
268:     }

270:     /* Compute relative error */
271:     PetscCall(VecAXPY(yy2, alpha, yy1));
272:     PetscCall(VecNorm(yy2, NORM_2, &enorm));
273:     enorm = enorm / yy1n;
274:     PetscCall(PetscFPrintf(comm, stdout, "h = %g: relative error = %g\n", (double)h, (double)enorm));
275:     h *= 10.0;
276:   }
277:   PetscFunctionReturn(PETSC_SUCCESS);
278: }