Actual source code: pounders.c
1: #include <../src/tao/leastsquares/impls/pounders/pounders.h>
3: static PetscErrorCode pounders_h(Tao subtao, Vec v, Mat H, Mat Hpre, PetscCtx ctx)
4: {
5: PetscFunctionBegin;
6: PetscFunctionReturn(PETSC_SUCCESS);
7: }
9: static PetscErrorCode pounders_fg(Tao subtao, Vec x, PetscReal *f, Vec g, PetscCtx ctx)
10: {
11: TAO_POUNDERS *mfqP = (TAO_POUNDERS *)ctx;
12: PetscReal d1, d2;
14: PetscFunctionBegin;
15: /* g = A*x (add b later)*/
16: PetscCall(MatMult(mfqP->subH, x, g));
18: /* f = 1/2 * x'*(Ax) + b'*x */
19: PetscCall(VecDot(x, g, &d1));
20: PetscCall(VecDot(mfqP->subb, x, &d2));
21: *f = 0.5 * d1 + d2;
23: /* now g = g + b */
24: PetscCall(VecAXPY(g, 1.0, mfqP->subb));
25: PetscFunctionReturn(PETSC_SUCCESS);
26: }
28: static PetscErrorCode pounders_feval(Tao tao, Vec x, Vec F, PetscReal *fsum)
29: {
30: TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
31: PetscInt i, row, col;
32: PetscReal fr, fc;
34: PetscFunctionBegin;
35: PetscCall(TaoComputeResidual(tao, x, F));
36: if (tao->res_weights_v) {
37: PetscCall(VecPointwiseMult(mfqP->workfvec, tao->res_weights_v, F));
38: PetscCall(VecDot(mfqP->workfvec, mfqP->workfvec, fsum));
39: } else if (tao->res_weights_w) {
40: *fsum = 0;
41: for (i = 0; i < tao->res_weights_n; i++) {
42: row = tao->res_weights_rows[i];
43: col = tao->res_weights_cols[i];
44: PetscCall(VecGetValues(F, 1, &row, &fr));
45: PetscCall(VecGetValues(F, 1, &col, &fc));
46: *fsum += tao->res_weights_w[i] * fc * fr;
47: }
48: } else {
49: PetscCall(VecDot(F, F, fsum));
50: }
51: PetscCall(PetscInfo(tao, "Least-squares residual norm: %20.19e\n", (double)*fsum));
52: PetscCheck(!PetscIsInfOrNanReal(*fsum), PETSC_COMM_SELF, PETSC_ERR_USER, "User provided compute function generated infinity or NaN");
53: PetscFunctionReturn(PETSC_SUCCESS);
54: }
56: static PetscErrorCode gqtwrap(Tao tao, PetscReal *gnorm, PetscReal *qmin)
57: {
58: PetscReal atol = PetscDefined(USE_REAL_SINGLE) ? 1.0e-5 : 1.0e-10;
59: PetscInt info, its;
60: TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
62: PetscFunctionBegin;
63: if (!mfqP->usegqt) {
64: PetscReal maxval;
65: PetscInt i, j;
67: PetscCall(VecSetValues(mfqP->subb, mfqP->n, mfqP->indices, mfqP->Gres, INSERT_VALUES));
68: PetscCall(VecAssemblyBegin(mfqP->subb));
69: PetscCall(VecAssemblyEnd(mfqP->subb));
71: PetscCall(VecSet(mfqP->subx, 0.0));
73: PetscCall(VecSet(mfqP->subndel, -1.0));
74: PetscCall(VecSet(mfqP->subpdel, +1.0));
76: /* Complete the lower triangle of the Hessian matrix */
77: for (i = 0; i < mfqP->n; i++) {
78: for (j = i + 1; j < mfqP->n; j++) mfqP->Hres[j + mfqP->n * i] = mfqP->Hres[mfqP->n * j + i];
79: }
80: PetscCall(MatSetValues(mfqP->subH, mfqP->n, mfqP->indices, mfqP->n, mfqP->indices, mfqP->Hres, INSERT_VALUES));
81: PetscCall(MatAssemblyBegin(mfqP->subH, MAT_FINAL_ASSEMBLY));
82: PetscCall(MatAssemblyEnd(mfqP->subH, MAT_FINAL_ASSEMBLY));
84: PetscCall(TaoResetStatistics(mfqP->subtao));
85: /* PetscCall(TaoSetTolerances(mfqP->subtao,*gnorm,*gnorm,PETSC_CURRENT)); */
86: /* enforce bound constraints -- experimental */
87: if (tao->XU && tao->XL) {
88: PetscCall(VecCopy(tao->XU, mfqP->subxu));
89: PetscCall(VecAXPY(mfqP->subxu, -1.0, tao->solution));
90: PetscCall(VecScale(mfqP->subxu, 1.0 / mfqP->delta));
91: PetscCall(VecCopy(tao->XL, mfqP->subxl));
92: PetscCall(VecAXPY(mfqP->subxl, -1.0, tao->solution));
93: PetscCall(VecScale(mfqP->subxl, 1.0 / mfqP->delta));
95: PetscCall(VecPointwiseMin(mfqP->subxu, mfqP->subxu, mfqP->subpdel));
96: PetscCall(VecPointwiseMax(mfqP->subxl, mfqP->subxl, mfqP->subndel));
97: } else {
98: PetscCall(VecCopy(mfqP->subpdel, mfqP->subxu));
99: PetscCall(VecCopy(mfqP->subndel, mfqP->subxl));
100: }
101: /* Make sure xu > xl */
102: PetscCall(VecCopy(mfqP->subxl, mfqP->subpdel));
103: PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subxu));
104: PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
105: PetscCheck(maxval <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "upper bound < lower bound in subproblem");
106: /* Make sure xu > tao->solution > xl */
107: PetscCall(VecCopy(mfqP->subxl, mfqP->subpdel));
108: PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subx));
109: PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
110: PetscCheck(maxval <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "initial guess < lower bound in subproblem");
112: PetscCall(VecCopy(mfqP->subx, mfqP->subpdel));
113: PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subxu));
114: PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
115: PetscCheck(maxval <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "initial guess > upper bound in subproblem");
117: PetscCall(TaoSolve(mfqP->subtao));
118: PetscCall(TaoGetSolutionStatus(mfqP->subtao, NULL, qmin, NULL, NULL, NULL, NULL));
120: /* test bounds post-solution*/
121: PetscCall(VecCopy(mfqP->subxl, mfqP->subpdel));
122: PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subx));
123: PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
124: if (maxval > 1e-5) {
125: PetscCall(PetscInfo(tao, "subproblem solution < lower bound\n"));
126: tao->reason = TAO_DIVERGED_TR_REDUCTION;
127: }
129: PetscCall(VecCopy(mfqP->subx, mfqP->subpdel));
130: PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subxu));
131: PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
132: if (maxval > 1e-5) {
133: PetscCall(PetscInfo(tao, "subproblem solution > upper bound\n"));
134: tao->reason = TAO_DIVERGED_TR_REDUCTION;
135: }
136: } else {
137: PetscCall(gqt(mfqP->n, mfqP->Hres, mfqP->n, mfqP->Gres, 1.0, mfqP->gqt_rtol, atol, mfqP->gqt_maxits, gnorm, qmin, mfqP->Xsubproblem, &info, &its, mfqP->work, mfqP->work2, mfqP->work3));
138: }
139: *qmin *= -1;
140: PetscFunctionReturn(PETSC_SUCCESS);
141: }
143: static PetscErrorCode pounders_update_res(Tao tao)
144: {
145: TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
146: PetscInt i, row, col;
147: PetscBLASInt blasn, blasn2, blasm, ione = 1;
148: PetscReal zero = 0.0, one = 1.0, wii, factor;
150: PetscFunctionBegin;
151: PetscCall(PetscBLASIntCast(mfqP->n, &blasn));
152: PetscCall(PetscBLASIntCast(mfqP->m, &blasm));
153: PetscCall(PetscBLASIntCast(mfqP->n * mfqP->n, &blasn2));
154: for (i = 0; i < mfqP->n; i++) mfqP->Gres[i] = 0;
155: for (i = 0; i < mfqP->n * mfqP->n; i++) mfqP->Hres[i] = 0;
157: /* Compute Gres= sum_ij[wij * (cjgi + cigj)] */
158: if (tao->res_weights_v) {
159: /* Vector(diagonal) weights: gres = sum_i(wii*ci*gi) */
160: for (i = 0; i < mfqP->m; i++) {
161: PetscCall(VecGetValues(tao->res_weights_v, 1, &i, &factor));
162: factor = factor * mfqP->C[i];
163: PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn, &factor, &mfqP->Fdiff[blasn * i], &ione, mfqP->Gres, &ione));
164: }
166: /* compute Hres = sum_ij [wij * (*ci*Hj + cj*Hi + gi gj' + gj gi') ] */
167: /* vector(diagonal weights) Hres = sum_i(wii*(ci*Hi + gi * gi')*/
168: for (i = 0; i < mfqP->m; i++) {
169: PetscCall(VecGetValues(tao->res_weights_v, 1, &i, &wii));
170: if (tao->niter > 1) {
171: factor = wii * mfqP->C[i];
172: /* add wii * ci * Hi */
173: PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn2, &factor, &mfqP->H[i], &blasm, mfqP->Hres, &ione));
174: }
175: /* add wii * gi * gi' */
176: PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &blasn, &blasn, &ione, &wii, &mfqP->Fdiff[blasn * i], &blasn, &mfqP->Fdiff[blasn * i], &blasn, &one, mfqP->Hres, &blasn));
177: }
178: } else if (tao->res_weights_w) {
179: /* General case: .5 * Gres= sum_ij[wij * (cjgi + cigj)] */
180: for (i = 0; i < tao->res_weights_n; i++) {
181: row = tao->res_weights_rows[i];
182: col = tao->res_weights_cols[i];
184: factor = tao->res_weights_w[i] * mfqP->C[col] / 2.0;
185: PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn, &factor, &mfqP->Fdiff[blasn * row], &ione, mfqP->Gres, &ione));
186: factor = tao->res_weights_w[i] * mfqP->C[row] / 2.0;
187: PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn, &factor, &mfqP->Fdiff[blasn * col], &ione, mfqP->Gres, &ione));
188: }
190: /* compute Hres = sum_ij [wij * (*ci*Hj + cj*Hi + gi gj' + gj gi') ] */
191: /* .5 * sum_ij [wij * (*ci*Hj + cj*Hi + gi gj' + gj gi') ] */
192: for (i = 0; i < tao->res_weights_n; i++) {
193: row = tao->res_weights_rows[i];
194: col = tao->res_weights_cols[i];
195: factor = tao->res_weights_w[i] / 2.0;
196: /* add wij * gi gj' + wij * gj gi' */
197: PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &blasn, &blasn, &ione, &factor, &mfqP->Fdiff[blasn * row], &blasn, &mfqP->Fdiff[blasn * col], &blasn, &one, mfqP->Hres, &blasn));
198: PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &blasn, &blasn, &ione, &factor, &mfqP->Fdiff[blasn * col], &blasn, &mfqP->Fdiff[blasn * row], &blasn, &one, mfqP->Hres, &blasn));
199: }
200: if (tao->niter > 1) {
201: for (i = 0; i < tao->res_weights_n; i++) {
202: row = tao->res_weights_rows[i];
203: col = tao->res_weights_cols[i];
205: /* add wij*cj*Hi */
206: factor = tao->res_weights_w[i] * mfqP->C[col] / 2.0;
207: PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn2, &factor, &mfqP->H[row], &blasm, mfqP->Hres, &ione));
209: /* add wij*ci*Hj */
210: factor = tao->res_weights_w[i] * mfqP->C[row] / 2.0;
211: PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn2, &factor, &mfqP->H[col], &blasm, mfqP->Hres, &ione));
212: }
213: }
214: } else {
215: /* Default: Gres= sum_i[cigi] = G*c' */
216: PetscCall(PetscInfo(tao, "Identity weights\n"));
217: PetscCallBLAS("BLASgemv", BLASgemv_("N", &blasn, &blasm, &one, mfqP->Fdiff, &blasn, mfqP->C, &ione, &zero, mfqP->Gres, &ione));
219: /* compute Hres = sum_ij [wij * (*ci*Hj + cj*Hi + gi gj' + gj gi') ] */
220: /* Hres = G*G' + 0.5 sum {F(xkin,i)*H(:,:,i)} */
221: PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &blasn, &blasn, &blasm, &one, mfqP->Fdiff, &blasn, mfqP->Fdiff, &blasn, &zero, mfqP->Hres, &blasn));
223: /* sum(F(xkin,i)*H(:,:,i)) */
224: if (tao->niter > 1) {
225: for (i = 0; i < mfqP->m; i++) {
226: factor = mfqP->C[i];
227: PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn2, &factor, &mfqP->H[i], &blasm, mfqP->Hres, &ione));
228: }
229: }
230: }
231: PetscFunctionReturn(PETSC_SUCCESS);
232: }
234: static PetscErrorCode phi2eval(PetscReal *x, PetscInt n, PetscReal *phi)
235: {
236: /* Phi = .5*[x(1)^2 sqrt(2)*x(1)*x(2) ... sqrt(2)*x(1)*x(n) ... x(2)^2 sqrt(2)*x(2)*x(3) .. x(n)^2] */
237: PetscInt i, j, k;
238: PetscReal sqrt2 = PetscSqrtReal(2.0);
240: PetscFunctionBegin;
241: j = 0;
242: for (i = 0; i < n; i++) {
243: phi[j] = 0.5 * x[i] * x[i];
244: j++;
245: for (k = i + 1; k < n; k++) {
246: phi[j] = x[i] * x[k] / sqrt2;
247: j++;
248: }
249: }
250: PetscFunctionReturn(PETSC_SUCCESS);
251: }
253: static PetscErrorCode getquadpounders(TAO_POUNDERS *mfqP)
254: {
255: /* Computes the parameters of the quadratic Q(x) = c + g'*x + 0.5*x*G*x'
256: that satisfies the interpolation conditions Q(X[:,j]) = f(j)
257: for j=1,...,m and with a Hessian matrix of least Frobenius norm */
259: /* NB --we are ignoring c */
260: PetscInt i, j, k, num, np = mfqP->nmodelpoints;
261: PetscReal one = 1.0, zero = 0.0, negone = -1.0;
262: PetscBLASInt blasnpmax, blasnplus1, blasnp, blasint, blasint2;
263: PetscBLASInt ione = 1;
264: PetscReal sqrt2 = PetscSqrtReal(2.0);
266: PetscFunctionBegin;
267: PetscCall(PetscBLASIntCast(mfqP->npmax, &blasnpmax));
268: PetscCall(PetscBLASIntCast(mfqP->n + 1, &blasnplus1));
269: PetscCall(PetscBLASIntCast(np, &blasnp));
270: PetscCall(PetscBLASIntCast(mfqP->n * (mfqP->n + 1) / 2, &blasint));
271: PetscCall(PetscBLASIntCast(np - mfqP->n - 1, &blasint2));
272: for (i = 0; i < mfqP->n * mfqP->m; i++) mfqP->Gdel[i] = 0;
273: for (i = 0; i < mfqP->n * mfqP->n * mfqP->m; i++) mfqP->Hdel[i] = 0;
275: /* factor M */
276: PetscCallLAPACKInfo("LAPACKgetrf", LAPACKgetrf_(&blasnplus1, &blasnp, mfqP->M, &blasnplus1, mfqP->npmaxiwork, &info));
278: if (np == mfqP->n + 1) {
279: for (i = 0; i < mfqP->npmax - mfqP->n - 1; i++) mfqP->omega[i] = 0.0;
280: for (i = 0; i < mfqP->n * (mfqP->n + 1) / 2; i++) mfqP->beta[i] = 0.0;
281: } else {
282: /* Let Ltmp = (L'*L) */
283: PetscCallBLAS("BLASgemm", BLASgemm_("T", "N", &blasint2, &blasint2, &blasint, &one, &mfqP->L[(mfqP->n + 1) * blasint], &blasint, &mfqP->L[(mfqP->n + 1) * blasint], &blasint, &zero, mfqP->L_tmp, &blasint));
285: /* factor Ltmp */
286: PetscCallLAPACKInfo("LAPACKpotrf", LAPACKpotrf_("L", &blasint2, mfqP->L_tmp, &blasint, &info));
287: }
289: for (k = 0; k < mfqP->m; k++) {
290: if (np != mfqP->n + 1) {
291: /* Solve L'*L*Omega = Z' * RESk*/
292: PetscCallBLAS("BLASgemv", BLASgemv_("T", &blasnp, &blasint2, &one, mfqP->Z, &blasnpmax, &mfqP->RES[mfqP->npmax * k], &ione, &zero, mfqP->omega, &ione));
293: PetscCallLAPACKInfo("LAPACKpotrs", LAPACKpotrs_("L", &blasint2, &ione, mfqP->L_tmp, &blasint, mfqP->omega, &blasint2, &info));
295: /* Beta = L*Omega */
296: PetscCallBLAS("BLASgemv", BLASgemv_("N", &blasint, &blasint2, &one, &mfqP->L[(mfqP->n + 1) * blasint], &blasint, mfqP->omega, &ione, &zero, mfqP->beta, &ione));
297: }
299: /* solve M'*Alpha = RESk - N'*Beta */
300: PetscCallBLAS("BLASgemv", BLASgemv_("T", &blasint, &blasnp, &negone, mfqP->N, &blasint, mfqP->beta, &ione, &one, &mfqP->RES[mfqP->npmax * k], &ione));
301: PetscCallLAPACKInfo("LAPACKgetrs", LAPACKgetrs_("T", &blasnplus1, &ione, mfqP->M, &blasnplus1, mfqP->npmaxiwork, &mfqP->RES[mfqP->npmax * k], &blasnplus1, &info));
303: /* Gdel(:,k) = Alpha(2:n+1) */
304: for (i = 0; i < mfqP->n; i++) mfqP->Gdel[i + mfqP->n * k] = mfqP->RES[mfqP->npmax * k + i + 1];
306: /* Set Hdels */
307: num = 0;
308: for (i = 0; i < mfqP->n; i++) {
309: /* H[i,i,k] = Beta(num) */
310: mfqP->Hdel[(i * mfqP->n + i) * mfqP->m + k] = mfqP->beta[num];
311: num++;
312: for (j = i + 1; j < mfqP->n; j++) {
313: /* H[i,j,k] = H[j,i,k] = Beta(num)/sqrt(2) */
314: mfqP->Hdel[(j * mfqP->n + i) * mfqP->m + k] = mfqP->beta[num] / sqrt2;
315: mfqP->Hdel[(i * mfqP->n + j) * mfqP->m + k] = mfqP->beta[num] / sqrt2;
316: num++;
317: }
318: }
319: }
320: PetscFunctionReturn(PETSC_SUCCESS);
321: }
323: static PetscErrorCode morepoints(TAO_POUNDERS *mfqP)
324: {
325: /* Assumes mfqP->model_indices[0] is minimum index
326: Finishes adding points to mfqP->model_indices (up to npmax)
327: Computes L,Z,M,N
328: np is actual number of points in model (should equal npmax?) */
329: PetscInt point, i, j, offset;
330: PetscInt reject;
331: PetscBLASInt blasn, blasnpmax, blasnplus1, blasnmax, blasint, blasint2, blasnp, blasmaxmn;
332: const PetscReal *x;
333: PetscReal normd;
335: PetscFunctionBegin;
336: PetscCall(PetscBLASIntCast(mfqP->npmax, &blasnpmax));
337: PetscCall(PetscBLASIntCast(mfqP->n, &blasn));
338: PetscCall(PetscBLASIntCast(mfqP->nmax, &blasnmax));
339: PetscCall(PetscBLASIntCast(mfqP->n + 1, &blasnplus1));
340: PetscCall(PetscBLASIntCast(mfqP->n, &blasnp));
341: /* Initialize M,N */
342: for (i = 0; i < mfqP->n + 1; i++) {
343: PetscCall(VecGetArrayRead(mfqP->Xhist[mfqP->model_indices[i]], &x));
344: mfqP->M[(mfqP->n + 1) * i] = 1.0;
345: for (j = 0; j < mfqP->n; j++) mfqP->M[j + 1 + ((mfqP->n + 1) * i)] = (x[j] - mfqP->xmin[j]) / mfqP->delta;
346: PetscCall(VecRestoreArrayRead(mfqP->Xhist[mfqP->model_indices[i]], &x));
347: PetscCall(phi2eval(&mfqP->M[1 + ((mfqP->n + 1) * i)], mfqP->n, &mfqP->N[mfqP->n * (mfqP->n + 1) / 2 * i]));
348: }
350: /* Now we add points until we have npmax starting with the most recent ones */
351: point = mfqP->nHist - 1;
352: mfqP->nmodelpoints = mfqP->n + 1;
353: while (mfqP->nmodelpoints < mfqP->npmax && point >= 0) {
354: /* Reject any points already in the model */
355: reject = 0;
356: for (j = 0; j < mfqP->n + 1; j++) {
357: if (point == mfqP->model_indices[j]) {
358: reject = 1;
359: break;
360: }
361: }
363: /* Reject if norm(d) >c2 */
364: if (!reject) {
365: PetscCall(VecCopy(mfqP->Xhist[point], mfqP->workxvec));
366: PetscCall(VecAXPY(mfqP->workxvec, -1.0, mfqP->Xhist[mfqP->minindex]));
367: PetscCall(VecNorm(mfqP->workxvec, NORM_2, &normd));
368: normd /= mfqP->delta;
369: if (normd > mfqP->c2) reject = 1;
370: }
371: if (reject) {
372: point--;
373: continue;
374: }
376: PetscCall(VecGetArrayRead(mfqP->Xhist[point], &x));
377: mfqP->M[(mfqP->n + 1) * mfqP->nmodelpoints] = 1.0;
378: for (j = 0; j < mfqP->n; j++) mfqP->M[j + 1 + ((mfqP->n + 1) * mfqP->nmodelpoints)] = (x[j] - mfqP->xmin[j]) / mfqP->delta;
379: PetscCall(VecRestoreArrayRead(mfqP->Xhist[point], &x));
380: PetscCall(phi2eval(&mfqP->M[1 + (mfqP->n + 1) * mfqP->nmodelpoints], mfqP->n, &mfqP->N[mfqP->n * (mfqP->n + 1) / 2 * (mfqP->nmodelpoints)]));
382: /* Update QR factorization */
383: /* Copy M' to Q_tmp */
384: for (i = 0; i < mfqP->n + 1; i++) {
385: for (j = 0; j < mfqP->npmax; j++) mfqP->Q_tmp[j + mfqP->npmax * i] = mfqP->M[i + (mfqP->n + 1) * j];
386: }
387: PetscCall(PetscBLASIntCast(mfqP->nmodelpoints + 1, &blasnp));
388: /* Q_tmp,R = qr(M') */
389: PetscCall(PetscBLASIntCast(PetscMax(mfqP->m, mfqP->n + 1), &blasmaxmn));
390: PetscCallLAPACKInfo("LAPACKgeqrf", LAPACKgeqrf_(&blasnp, &blasnplus1, mfqP->Q_tmp, &blasnpmax, mfqP->tau_tmp, mfqP->mwork, &blasmaxmn, &info));
392: /* Reject if min(svd(N*Q(:,n+2:np+1)) <= theta2 */
393: /* L = N*Qtmp */
394: PetscCall(PetscBLASIntCast(mfqP->n * (mfqP->n + 1) / 2, &blasint2));
395: /* Copy N to L_tmp */
396: for (i = 0; i < mfqP->n * (mfqP->n + 1) / 2 * mfqP->npmax; i++) mfqP->L_tmp[i] = mfqP->N[i];
397: /* Copy L_save to L_tmp */
399: /* L_tmp = N*Qtmp' */
400: PetscCallLAPACKInfo("LAPACKormqr", LAPACKormqr_("R", "N", &blasint2, &blasnp, &blasnplus1, mfqP->Q_tmp, &blasnpmax, mfqP->tau_tmp, mfqP->L_tmp, &blasint2, mfqP->npmaxwork, &blasnmax, &info));
402: /* Copy L_tmp to L_save */
403: for (i = 0; i < mfqP->npmax * mfqP->n * (mfqP->n + 1) / 2; i++) mfqP->L_save[i] = mfqP->L_tmp[i];
405: /* Get svd for L_tmp(:,n+2:np+1) (L_tmp is modified in process) */
406: PetscCall(PetscBLASIntCast(mfqP->nmodelpoints - mfqP->n, &blasint));
407: PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
408: PetscCallLAPACKInfo("LAPACKgesvd", LAPACKgesvd_("N", "N", &blasint2, &blasint, &mfqP->L_tmp[(mfqP->n + 1) * blasint2], &blasint2, mfqP->beta, mfqP->work, &blasn, mfqP->work, &blasn, mfqP->npmaxwork, &blasnmax, &info));
409: PetscCall(PetscFPTrapPop());
411: if (mfqP->beta[PetscMin(blasint, blasint2) - 1] > mfqP->theta2) {
412: /* accept point */
413: mfqP->model_indices[mfqP->nmodelpoints] = point;
414: /* Copy Q_tmp to Q */
415: for (i = 0; i < mfqP->npmax * mfqP->npmax; i++) mfqP->Q[i] = mfqP->Q_tmp[i];
416: for (i = 0; i < mfqP->npmax; i++) mfqP->tau[i] = mfqP->tau_tmp[i];
417: mfqP->nmodelpoints++;
418: PetscCall(PetscBLASIntCast(mfqP->nmodelpoints, &blasnp));
420: /* Copy L_save to L */
421: for (i = 0; i < mfqP->npmax * mfqP->n * (mfqP->n + 1) / 2; i++) mfqP->L[i] = mfqP->L_save[i];
422: }
423: point--;
424: }
426: PetscCall(PetscBLASIntCast(mfqP->nmodelpoints, &blasnp));
427: /* Copy Q(:,n+2:np) to Z */
428: /* First set Q_tmp to I */
429: for (i = 0; i < mfqP->npmax * mfqP->npmax; i++) mfqP->Q_tmp[i] = 0.0;
430: for (i = 0; i < mfqP->npmax; i++) mfqP->Q_tmp[i + mfqP->npmax * i] = 1.0;
432: /* Q_tmp = I * Q */
433: PetscCallLAPACKInfo("LAPACKormqr", LAPACKormqr_("R", "N", &blasnp, &blasnp, &blasnplus1, mfqP->Q, &blasnpmax, mfqP->tau, mfqP->Q_tmp, &blasnpmax, mfqP->npmaxwork, &blasnmax, &info));
435: /* Copy Q_tmp(:,n+2:np) to Z) */
436: offset = mfqP->npmax * (mfqP->n + 1);
437: for (i = offset; i < mfqP->npmax * mfqP->npmax; i++) mfqP->Z[i - offset] = mfqP->Q_tmp[i];
439: if (mfqP->nmodelpoints == mfqP->n + 1) {
440: /* Set L to I_{n+1} */
441: for (i = 0; i < mfqP->npmax * mfqP->n * (mfqP->n + 1) / 2; i++) mfqP->L[i] = 0.0;
442: for (i = 0; i < mfqP->n; i++) mfqP->L[(mfqP->n * (mfqP->n + 1) / 2) * i + i] = 1.0;
443: }
444: PetscFunctionReturn(PETSC_SUCCESS);
445: }
447: /* Only call from modelimprove, addpoint() needs ->Q_tmp and ->work to be set */
448: static PetscErrorCode addpoint(Tao tao, TAO_POUNDERS *mfqP, PetscInt index)
449: {
450: PetscFunctionBegin;
451: /* Create new vector in history: X[newidx] = X[mfqP->index] + delta*X[index]*/
452: PetscCall(VecDuplicate(mfqP->Xhist[0], &mfqP->Xhist[mfqP->nHist]));
453: PetscCall(VecSetValues(mfqP->Xhist[mfqP->nHist], mfqP->n, mfqP->indices, &mfqP->Q_tmp[index * mfqP->npmax], INSERT_VALUES));
454: PetscCall(VecAssemblyBegin(mfqP->Xhist[mfqP->nHist]));
455: PetscCall(VecAssemblyEnd(mfqP->Xhist[mfqP->nHist]));
456: PetscCall(VecAYPX(mfqP->Xhist[mfqP->nHist], mfqP->delta, mfqP->Xhist[mfqP->minindex]));
458: /* Project into feasible region */
459: if (tao->XU && tao->XL) PetscCall(VecMedian(mfqP->Xhist[mfqP->nHist], tao->XL, tao->XU, mfqP->Xhist[mfqP->nHist]));
461: /* Compute value of new vector */
462: PetscCall(VecDuplicate(mfqP->Fhist[0], &mfqP->Fhist[mfqP->nHist]));
463: CHKMEMQ;
464: PetscCall(pounders_feval(tao, mfqP->Xhist[mfqP->nHist], mfqP->Fhist[mfqP->nHist], &mfqP->Fres[mfqP->nHist]));
466: /* Add new vector to model */
467: mfqP->model_indices[mfqP->nmodelpoints] = mfqP->nHist;
468: mfqP->nmodelpoints++;
469: mfqP->nHist++;
470: PetscFunctionReturn(PETSC_SUCCESS);
471: }
473: static PetscErrorCode modelimprove(Tao tao, TAO_POUNDERS *mfqP, PetscInt addallpoints)
474: {
475: /* modeld = Q(:,np+1:n)' */
476: PetscInt i, j, minindex = 0;
477: PetscReal dp, half = 0.5, one = 1.0, minvalue = PETSC_INFINITY;
478: PetscBLASInt blasn, blasnpmax, blask;
479: PetscBLASInt blas1 = 1, blasnmax;
481: PetscFunctionBegin;
482: PetscCall(PetscBLASIntCast(mfqP->n, &blasn));
483: PetscCall(PetscBLASIntCast(mfqP->npmax, &blasnpmax));
484: PetscCall(PetscBLASIntCast(mfqP->nmodelpoints, &blask));
485: PetscCall(PetscBLASIntCast(mfqP->nmax, &blasnmax));
487: /* Qtmp = I(n x n) */
488: for (i = 0; i < mfqP->n; i++) {
489: for (j = 0; j < mfqP->n; j++) mfqP->Q_tmp[i + mfqP->npmax * j] = 0.0;
490: }
491: for (j = 0; j < mfqP->n; j++) mfqP->Q_tmp[j + mfqP->npmax * j] = 1.0;
493: /* Qtmp = Q * I */
494: PetscCallLAPACKInfo("LAPACKormqr", LAPACKormqr_("R", "N", &blasn, &blasn, &blask, mfqP->Q, &blasnpmax, mfqP->tau, mfqP->Q_tmp, &blasnpmax, mfqP->npmaxwork, &blasnmax, &info));
496: for (i = mfqP->nmodelpoints; i < mfqP->n; i++) {
497: PetscCallBLAS("BLASdot", dp = BLASdot_(&blasn, &mfqP->Q_tmp[i * mfqP->npmax], &blas1, mfqP->Gres, &blas1));
498: if (dp > 0.0) { /* Model says use the other direction! */
499: for (j = 0; j < mfqP->n; j++) mfqP->Q_tmp[i * mfqP->npmax + j] *= -1;
500: }
501: /* mfqP->work[i] = Cres+Modeld(i,:)*(Gres+.5*Hres*Modeld(i,:)') */
502: for (j = 0; j < mfqP->n; j++) mfqP->work2[j] = mfqP->Gres[j];
503: PetscCallBLAS("BLASgemv", BLASgemv_("N", &blasn, &blasn, &half, mfqP->Hres, &blasn, &mfqP->Q_tmp[i * mfqP->npmax], &blas1, &one, mfqP->work2, &blas1));
504: PetscCallBLAS("BLASdot", mfqP->work[i] = BLASdot_(&blasn, &mfqP->Q_tmp[i * mfqP->npmax], &blas1, mfqP->work2, &blas1));
505: if (i == mfqP->nmodelpoints || mfqP->work[i] < minvalue) {
506: minindex = i;
507: minvalue = mfqP->work[i];
508: }
509: if (addallpoints != 0) PetscCall(addpoint(tao, mfqP, i));
510: }
511: if (!addallpoints) PetscCall(addpoint(tao, mfqP, minindex));
512: PetscFunctionReturn(PETSC_SUCCESS);
513: }
515: static PetscErrorCode affpoints(TAO_POUNDERS *mfqP, PetscReal *xmin, PetscReal c)
516: {
517: PetscInt i, j;
518: PetscBLASInt blasm, blasj, blask, blasn, ione = 1;
519: PetscBLASInt blasnpmax, blasmaxmn;
520: PetscReal proj, normd;
521: const PetscReal *x;
523: PetscFunctionBegin;
524: PetscCall(PetscBLASIntCast(mfqP->npmax, &blasnpmax));
525: PetscCall(PetscBLASIntCast(mfqP->m, &blasm));
526: PetscCall(PetscBLASIntCast(mfqP->n, &blasn));
527: for (i = mfqP->nHist - 1; i >= 0; i--) {
528: PetscCall(VecGetArrayRead(mfqP->Xhist[i], &x));
529: for (j = 0; j < mfqP->n; j++) mfqP->work[j] = (x[j] - xmin[j]) / mfqP->delta;
530: PetscCall(VecRestoreArrayRead(mfqP->Xhist[i], &x));
531: PetscCallBLAS("BLAScopy", BLAScopy_(&blasn, mfqP->work, &ione, mfqP->work2, &ione));
532: PetscCallBLAS("BLASnrm2", normd = BLASnrm2_(&blasn, mfqP->work, &ione));
533: if (normd <= c) {
534: PetscCall(PetscBLASIntCast(PetscMax(mfqP->n - mfqP->nmodelpoints, 0), &blasj));
535: if (!mfqP->q_is_I) {
536: /* project D onto null */
537: PetscCall(PetscBLASIntCast(mfqP->nmodelpoints, &blask));
538: PetscCallLAPACKInfo("LAPACKormqr", LAPACKormqr_("R", "N", &ione, &blasn, &blask, mfqP->Q, &blasnpmax, mfqP->tau, mfqP->work2, &ione, mfqP->mwork, &blasm, &info));
539: }
540: PetscCallBLAS("BLASnrm2", proj = BLASnrm2_(&blasj, &mfqP->work2[mfqP->nmodelpoints], &ione));
542: if (proj >= mfqP->theta1) { /* add this index to model */
543: mfqP->model_indices[mfqP->nmodelpoints] = i;
544: mfqP->nmodelpoints++;
545: PetscCallBLAS("BLAScopy", BLAScopy_(&blasn, mfqP->work, &ione, &mfqP->Q_tmp[mfqP->npmax * (mfqP->nmodelpoints - 1)], &ione));
546: PetscCall(PetscBLASIntCast(mfqP->npmax * (mfqP->nmodelpoints), &blask));
547: PetscCallBLAS("BLAScopy", BLAScopy_(&blask, mfqP->Q_tmp, &ione, mfqP->Q, &ione));
548: PetscCall(PetscBLASIntCast(mfqP->nmodelpoints, &blask));
549: PetscCall(PetscBLASIntCast(PetscMax(mfqP->m, mfqP->n), &blasmaxmn));
550: PetscCallLAPACKInfo("LAPACKgeqrf", LAPACKgeqrf_(&blasn, &blask, mfqP->Q, &blasnpmax, mfqP->tau, mfqP->mwork, &blasmaxmn, &info));
551: mfqP->q_is_I = 0;
552: }
553: if (mfqP->nmodelpoints == mfqP->n) break;
554: }
555: }
556: PetscFunctionReturn(PETSC_SUCCESS);
557: }
559: static PetscErrorCode TaoSolve_POUNDERS(Tao tao)
560: {
561: TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
562: PetscInt i, ii, j, k, l;
563: PetscReal step = 1.0;
564: PetscInt low, high;
565: PetscReal minnorm;
566: PetscReal *x, *f;
567: const PetscReal *xmint, *fmin;
568: PetscReal deltaold;
569: PetscReal gnorm;
570: PetscBLASInt info, ione = 1, iblas;
571: PetscBool valid, same;
572: PetscReal mdec, rho, normxsp;
573: PetscReal one = 1.0, zero = 0.0, ratio;
574: PetscBLASInt blasm, blasn, blasncopy, blasnpmax;
575: static PetscBool set = PETSC_FALSE;
577: /* n = # of parameters
578: m = dimension (components) of function */
579: PetscFunctionBegin;
580: PetscCall(PetscCitationsRegister("@article{UNEDF0,\n"
581: "title = {Nuclear energy density optimization},\n"
582: "author = {Kortelainen, M. and Lesinski, T. and Mor\'e, J. and Nazarewicz, W.\n"
583: " and Sarich, J. and Schunck, N. and Stoitsov, M. V. and Wild, S. },\n"
584: "journal = {Phys. Rev. C},\n"
585: "volume = {82},\n"
586: "number = {2},\n"
587: "pages = {024313},\n"
588: "numpages = {18},\n"
589: "year = {2010},\n"
590: "month = {Aug},\n"
591: "doi = {10.1103/PhysRevC.82.024313}\n}\n",
592: &set));
593: tao->niter = 0;
594: if (tao->XL && tao->XU) {
595: /* Check x0 <= XU */
596: PetscReal val;
598: PetscCall(VecCopy(tao->solution, mfqP->Xhist[0]));
599: PetscCall(VecAXPY(mfqP->Xhist[0], -1.0, tao->XU));
600: PetscCall(VecMax(mfqP->Xhist[0], NULL, &val));
601: PetscCheck(val <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "X0 > upper bound");
603: /* Check x0 >= xl */
604: PetscCall(VecCopy(tao->XL, mfqP->Xhist[0]));
605: PetscCall(VecAXPY(mfqP->Xhist[0], -1.0, tao->solution));
606: PetscCall(VecMax(mfqP->Xhist[0], NULL, &val));
607: PetscCheck(val <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "X0 < lower bound");
609: /* Check x0 + delta < XU -- should be able to get around this eventually */
611: PetscCall(VecSet(mfqP->Xhist[0], mfqP->delta));
612: PetscCall(VecAXPY(mfqP->Xhist[0], 1.0, tao->solution));
613: PetscCall(VecAXPY(mfqP->Xhist[0], -1.0, tao->XU));
614: PetscCall(VecMax(mfqP->Xhist[0], NULL, &val));
615: PetscCheck(val <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "X0 + delta > upper bound");
616: }
618: PetscCall(PetscBLASIntCast(mfqP->m, &blasm));
619: PetscCall(PetscBLASIntCast(mfqP->n, &blasn));
620: PetscCall(PetscBLASIntCast(mfqP->npmax, &blasnpmax));
621: for (i = 0; i < mfqP->n * mfqP->n * mfqP->m; ++i) mfqP->H[i] = 0;
623: PetscCall(VecCopy(tao->solution, mfqP->Xhist[0]));
625: /* This provides enough information to approximate the gradient of the objective */
626: /* using a forward difference scheme. */
628: PetscCall(PetscInfo(tao, "Initialize simplex; delta = %10.9e\n", (double)mfqP->delta));
629: PetscCall(pounders_feval(tao, mfqP->Xhist[0], mfqP->Fhist[0], &mfqP->Fres[0]));
630: mfqP->minindex = 0;
631: minnorm = mfqP->Fres[0];
633: PetscCall(VecGetOwnershipRange(mfqP->Xhist[0], &low, &high));
634: for (i = 1; i < mfqP->n + 1; ++i) {
635: PetscCall(VecCopy(mfqP->Xhist[0], mfqP->Xhist[i]));
637: if (i - 1 >= low && i - 1 < high) {
638: PetscCall(VecGetArray(mfqP->Xhist[i], &x));
639: x[i - 1 - low] += mfqP->delta;
640: PetscCall(VecRestoreArray(mfqP->Xhist[i], &x));
641: }
642: CHKMEMQ;
643: PetscCall(pounders_feval(tao, mfqP->Xhist[i], mfqP->Fhist[i], &mfqP->Fres[i]));
644: if (mfqP->Fres[i] < minnorm) {
645: mfqP->minindex = i;
646: minnorm = mfqP->Fres[i];
647: }
648: }
649: PetscCall(VecCopy(mfqP->Xhist[mfqP->minindex], tao->solution));
650: PetscCall(VecCopy(mfqP->Fhist[mfqP->minindex], tao->ls_res));
651: PetscCall(PetscInfo(tao, "Finalize simplex; minnorm = %10.9e\n", (double)minnorm));
653: /* Gather mpi vecs to one big local vec */
655: /* Begin serial code */
657: /* Disp[i] = Xi-xmin, i=1,..,mfqP->minindex-1,mfqP->minindex+1,..,n */
658: /* Fdiff[i] = (Fi-Fmin)', i=1,..,mfqP->minindex-1,mfqP->minindex+1,..,n */
659: /* (Column oriented for blas calls) */
660: ii = 0;
662: PetscCall(PetscInfo(tao, "Build matrix: %d\n", mfqP->size));
663: if (1 == mfqP->size) {
664: PetscCall(VecGetArrayRead(mfqP->Xhist[mfqP->minindex], &xmint));
665: for (i = 0; i < mfqP->n; i++) mfqP->xmin[i] = xmint[i];
666: PetscCall(VecRestoreArrayRead(mfqP->Xhist[mfqP->minindex], &xmint));
667: PetscCall(VecGetArrayRead(mfqP->Fhist[mfqP->minindex], &fmin));
668: for (i = 0; i < mfqP->n + 1; i++) {
669: if (i == mfqP->minindex) continue;
671: PetscCall(VecGetArray(mfqP->Xhist[i], &x));
672: for (j = 0; j < mfqP->n; j++) mfqP->Disp[ii + mfqP->npmax * j] = (x[j] - mfqP->xmin[j]) / mfqP->delta;
673: PetscCall(VecRestoreArray(mfqP->Xhist[i], &x));
675: PetscCall(VecGetArray(mfqP->Fhist[i], &f));
676: for (j = 0; j < mfqP->m; j++) mfqP->Fdiff[ii + mfqP->n * j] = f[j] - fmin[j];
677: PetscCall(VecRestoreArray(mfqP->Fhist[i], &f));
679: mfqP->model_indices[ii++] = i;
680: }
681: for (j = 0; j < mfqP->m; j++) mfqP->C[j] = fmin[j];
682: PetscCall(VecRestoreArrayRead(mfqP->Fhist[mfqP->minindex], &fmin));
683: } else {
684: PetscCall(VecSet(mfqP->localxmin, 0));
685: PetscCall(VecScatterBegin(mfqP->scatterx, mfqP->Xhist[mfqP->minindex], mfqP->localxmin, INSERT_VALUES, SCATTER_FORWARD));
686: PetscCall(VecScatterEnd(mfqP->scatterx, mfqP->Xhist[mfqP->minindex], mfqP->localxmin, INSERT_VALUES, SCATTER_FORWARD));
688: PetscCall(VecGetArrayRead(mfqP->localxmin, &xmint));
689: for (i = 0; i < mfqP->n; i++) mfqP->xmin[i] = xmint[i];
690: PetscCall(VecRestoreArrayRead(mfqP->localxmin, &xmint));
692: PetscCall(VecScatterBegin(mfqP->scatterf, mfqP->Fhist[mfqP->minindex], mfqP->localfmin, INSERT_VALUES, SCATTER_FORWARD));
693: PetscCall(VecScatterEnd(mfqP->scatterf, mfqP->Fhist[mfqP->minindex], mfqP->localfmin, INSERT_VALUES, SCATTER_FORWARD));
694: PetscCall(VecGetArrayRead(mfqP->localfmin, &fmin));
695: for (i = 0; i < mfqP->n + 1; i++) {
696: if (i == mfqP->minindex) continue;
698: PetscCall(VecScatterBegin(mfqP->scatterx, mfqP->Xhist[ii], mfqP->localx, INSERT_VALUES, SCATTER_FORWARD));
699: PetscCall(VecScatterEnd(mfqP->scatterx, mfqP->Xhist[ii], mfqP->localx, INSERT_VALUES, SCATTER_FORWARD));
700: PetscCall(VecGetArray(mfqP->localx, &x));
701: for (j = 0; j < mfqP->n; j++) mfqP->Disp[ii + mfqP->npmax * j] = (x[j] - mfqP->xmin[j]) / mfqP->delta;
702: PetscCall(VecRestoreArray(mfqP->localx, &x));
704: PetscCall(VecScatterBegin(mfqP->scatterf, mfqP->Fhist[ii], mfqP->localf, INSERT_VALUES, SCATTER_FORWARD));
705: PetscCall(VecScatterEnd(mfqP->scatterf, mfqP->Fhist[ii], mfqP->localf, INSERT_VALUES, SCATTER_FORWARD));
706: PetscCall(VecGetArray(mfqP->localf, &f));
707: for (j = 0; j < mfqP->m; j++) mfqP->Fdiff[ii + mfqP->n * j] = f[j] - fmin[j];
708: PetscCall(VecRestoreArray(mfqP->localf, &f));
710: mfqP->model_indices[ii++] = i;
711: }
712: for (j = 0; j < mfqP->m; j++) mfqP->C[j] = fmin[j];
713: PetscCall(VecRestoreArrayRead(mfqP->localfmin, &fmin));
714: }
716: /* Determine the initial quadratic models */
717: /* G = D(ModelIn,:) \ (F(ModelIn,1:m)-repmat(F(xkin,1:m),n,1)); */
718: /* D (nxn) Fdiff (nxm) => G (nxm) */
719: blasncopy = blasn;
720: PetscCallBLAS("LAPACKgesv", LAPACKgesv_(&blasn, &blasm, mfqP->Disp, &blasnpmax, mfqP->iwork, mfqP->Fdiff, &blasncopy, &info));
721: PetscCall(PetscInfo(tao, "Linear solve return: %" PetscBLASInt_FMT "\n", info));
723: PetscCall(pounders_update_res(tao));
725: valid = PETSC_TRUE;
727: PetscCall(VecSetValues(tao->gradient, mfqP->n, mfqP->indices, mfqP->Gres, INSERT_VALUES));
728: PetscCall(VecAssemblyBegin(tao->gradient));
729: PetscCall(VecAssemblyEnd(tao->gradient));
730: PetscCall(VecNorm(tao->gradient, NORM_2, &gnorm));
731: gnorm *= mfqP->delta;
732: PetscCall(VecCopy(mfqP->Xhist[mfqP->minindex], tao->solution));
734: tao->reason = TAO_CONTINUE_ITERATING;
735: PetscCall(TaoLogConvergenceHistory(tao, minnorm, gnorm, 0.0, tao->ksp_its));
736: PetscCall(TaoMonitor(tao, tao->niter, minnorm, gnorm, 0.0, step));
737: PetscUseTypeMethod(tao, convergencetest, tao->cnvP);
739: mfqP->nHist = mfqP->n + 1;
740: mfqP->nmodelpoints = mfqP->n + 1;
741: PetscCall(PetscInfo(tao, "Initial gradient: %20.19e\n", (double)gnorm));
743: while (tao->reason == TAO_CONTINUE_ITERATING) {
744: PetscReal gnm = 1e-4;
745: /* Call general purpose update function */
746: PetscTryTypeMethod(tao, update, tao->niter, tao->user_update);
747: tao->niter++;
748: /* Solve the subproblem min{Q(s): ||s|| <= 1.0} */
749: PetscCall(gqtwrap(tao, &gnm, &mdec));
750: /* Evaluate the function at the new point */
752: for (i = 0; i < mfqP->n; i++) mfqP->work[i] = mfqP->Xsubproblem[i] * mfqP->delta + mfqP->xmin[i];
753: PetscCall(VecDuplicate(tao->solution, &mfqP->Xhist[mfqP->nHist]));
754: PetscCall(VecDuplicate(tao->ls_res, &mfqP->Fhist[mfqP->nHist]));
755: PetscCall(VecSetValues(mfqP->Xhist[mfqP->nHist], mfqP->n, mfqP->indices, mfqP->work, INSERT_VALUES));
756: PetscCall(VecAssemblyBegin(mfqP->Xhist[mfqP->nHist]));
757: PetscCall(VecAssemblyEnd(mfqP->Xhist[mfqP->nHist]));
759: PetscCall(pounders_feval(tao, mfqP->Xhist[mfqP->nHist], mfqP->Fhist[mfqP->nHist], &mfqP->Fres[mfqP->nHist]));
761: rho = (mfqP->Fres[mfqP->minindex] - mfqP->Fres[mfqP->nHist]) / mdec;
762: mfqP->nHist++;
764: /* Update the center */
765: if ((rho >= mfqP->eta1) || (rho > mfqP->eta0 && valid == PETSC_TRUE)) {
766: /* Update model to reflect new base point */
767: for (i = 0; i < mfqP->n; i++) mfqP->work[i] = (mfqP->work[i] - mfqP->xmin[i]) / mfqP->delta;
768: for (j = 0; j < mfqP->m; j++) {
769: /* C(j) = C(j) + work*G(:,j) + .5*work*H(:,:,j)*work';
770: G(:,j) = G(:,j) + H(:,:,j)*work' */
771: for (k = 0; k < mfqP->n; k++) {
772: mfqP->work2[k] = 0.0;
773: for (l = 0; l < mfqP->n; l++) mfqP->work2[k] += mfqP->H[j + mfqP->m * (k + l * mfqP->n)] * mfqP->work[l];
774: }
775: for (i = 0; i < mfqP->n; i++) {
776: mfqP->C[j] += mfqP->work[i] * (mfqP->Fdiff[i + mfqP->n * j] + 0.5 * mfqP->work2[i]);
777: mfqP->Fdiff[i + mfqP->n * j] += mfqP->work2[i];
778: }
779: }
780: /* Cres += work*Gres + .5*work*Hres*work';
781: Gres += Hres*work'; */
783: PetscCallBLAS("BLASgemv", BLASgemv_("N", &blasn, &blasn, &one, mfqP->Hres, &blasn, mfqP->work, &ione, &zero, mfqP->work2, &ione));
784: for (i = 0; i < mfqP->n; i++) mfqP->Gres[i] += mfqP->work2[i];
785: mfqP->minindex = mfqP->nHist - 1;
786: minnorm = mfqP->Fres[mfqP->minindex];
787: PetscCall(VecCopy(mfqP->Fhist[mfqP->minindex], tao->ls_res));
788: /* Change current center */
789: PetscCall(VecGetArrayRead(mfqP->Xhist[mfqP->minindex], &xmint));
790: for (i = 0; i < mfqP->n; i++) mfqP->xmin[i] = xmint[i];
791: PetscCall(VecRestoreArrayRead(mfqP->Xhist[mfqP->minindex], &xmint));
792: }
794: /* Evaluate at a model-improving point if necessary */
795: if (valid == PETSC_FALSE) {
796: mfqP->q_is_I = 1;
797: mfqP->nmodelpoints = 0;
798: PetscCall(affpoints(mfqP, mfqP->xmin, mfqP->c1));
799: if (mfqP->nmodelpoints < mfqP->n) {
800: PetscCall(PetscInfo(tao, "Model not valid -- model-improving\n"));
801: PetscCall(modelimprove(tao, mfqP, 1));
802: }
803: }
805: /* Update the trust region radius */
806: deltaold = mfqP->delta;
807: normxsp = 0;
808: for (i = 0; i < mfqP->n; i++) normxsp += mfqP->Xsubproblem[i] * mfqP->Xsubproblem[i];
809: normxsp = PetscSqrtReal(normxsp);
810: if (rho >= mfqP->eta1 && normxsp > 0.5 * mfqP->delta) {
811: mfqP->delta = PetscMin(mfqP->delta * mfqP->gamma1, mfqP->deltamax);
812: } else if (valid == PETSC_TRUE) {
813: mfqP->delta = PetscMax(mfqP->delta * mfqP->gamma0, mfqP->deltamin);
814: }
816: /* Compute the next interpolation set */
817: mfqP->q_is_I = 1;
818: mfqP->nmodelpoints = 0;
819: PetscCall(PetscInfo(tao, "Affine Points: xmin = %20.19e, c1 = %20.19e\n", (double)*mfqP->xmin, (double)mfqP->c1));
820: PetscCall(affpoints(mfqP, mfqP->xmin, mfqP->c1));
821: if (mfqP->nmodelpoints == mfqP->n) {
822: valid = PETSC_TRUE;
823: } else {
824: valid = PETSC_FALSE;
825: PetscCall(PetscInfo(tao, "Affine Points: xmin = %20.19e, c2 = %20.19e\n", (double)*mfqP->xmin, (double)mfqP->c2));
826: PetscCall(affpoints(mfqP, mfqP->xmin, mfqP->c2));
827: if (mfqP->n > mfqP->nmodelpoints) {
828: PetscCall(PetscInfo(tao, "Model not valid -- adding geometry points\n"));
829: PetscCall(modelimprove(tao, mfqP, mfqP->n - mfqP->nmodelpoints));
830: }
831: }
832: for (i = mfqP->nmodelpoints; i > 0; i--) mfqP->model_indices[i] = mfqP->model_indices[i - 1];
833: mfqP->nmodelpoints++;
834: mfqP->model_indices[0] = mfqP->minindex;
835: PetscCall(morepoints(mfqP));
836: for (i = 0; i < mfqP->nmodelpoints; i++) {
837: PetscCall(VecGetArray(mfqP->Xhist[mfqP->model_indices[i]], &x));
838: for (j = 0; j < mfqP->n; j++) mfqP->Disp[i + mfqP->npmax * j] = (x[j] - mfqP->xmin[j]) / deltaold;
839: PetscCall(VecRestoreArray(mfqP->Xhist[mfqP->model_indices[i]], &x));
840: PetscCall(VecGetArray(mfqP->Fhist[mfqP->model_indices[i]], &f));
841: for (j = 0; j < mfqP->m; j++) {
842: for (k = 0; k < mfqP->n; k++) {
843: mfqP->work[k] = 0.0;
844: for (l = 0; l < mfqP->n; l++) mfqP->work[k] += mfqP->H[j + mfqP->m * (k + mfqP->n * l)] * mfqP->Disp[i + mfqP->npmax * l];
845: }
846: PetscCallBLAS("BLASdot", mfqP->RES[j * mfqP->npmax + i] = -mfqP->C[j] - BLASdot_(&blasn, &mfqP->Fdiff[j * mfqP->n], &ione, &mfqP->Disp[i], &blasnpmax) - 0.5 * BLASdot_(&blasn, mfqP->work, &ione, &mfqP->Disp[i], &blasnpmax) + f[j]);
847: }
848: PetscCall(VecRestoreArray(mfqP->Fhist[mfqP->model_indices[i]], &f));
849: }
851: /* Update the quadratic model */
852: PetscCall(PetscInfo(tao, "Get Quad, size: %" PetscInt_FMT ", points: %" PetscInt_FMT "\n", mfqP->n, mfqP->nmodelpoints));
853: PetscCall(getquadpounders(mfqP));
854: PetscCall(VecGetArrayRead(mfqP->Fhist[mfqP->minindex], &fmin));
855: PetscCallBLAS("BLAScopy", BLAScopy_(&blasm, fmin, &ione, mfqP->C, &ione));
856: /* G = G*(delta/deltaold) + Gdel */
857: ratio = mfqP->delta / deltaold;
858: iblas = blasm * blasn;
859: PetscCallBLAS("BLASscal", BLASscal_(&iblas, &ratio, mfqP->Fdiff, &ione));
860: PetscCallBLAS("BLASaxpy", BLASaxpy_(&iblas, &one, mfqP->Gdel, &ione, mfqP->Fdiff, &ione));
861: /* H = H*(delta/deltaold)^2 + Hdel */
862: iblas = blasm * blasn * blasn;
863: ratio *= ratio;
864: PetscCallBLAS("BLASscal", BLASscal_(&iblas, &ratio, mfqP->H, &ione));
865: PetscCallBLAS("BLASaxpy", BLASaxpy_(&iblas, &one, mfqP->Hdel, &ione, mfqP->H, &ione));
867: /* Get residuals */
868: PetscCall(pounders_update_res(tao));
870: /* Export solution and gradient residual to TAO */
871: PetscCall(VecCopy(mfqP->Xhist[mfqP->minindex], tao->solution));
872: PetscCall(VecSetValues(tao->gradient, mfqP->n, mfqP->indices, mfqP->Gres, INSERT_VALUES));
873: PetscCall(VecAssemblyBegin(tao->gradient));
874: PetscCall(VecAssemblyEnd(tao->gradient));
875: PetscCall(VecNorm(tao->gradient, NORM_2, &gnorm));
876: gnorm *= mfqP->delta;
877: /* final criticality test */
878: PetscCall(TaoLogConvergenceHistory(tao, minnorm, gnorm, 0.0, tao->ksp_its));
879: PetscCall(TaoMonitor(tao, tao->niter, minnorm, gnorm, 0.0, step));
880: PetscUseTypeMethod(tao, convergencetest, tao->cnvP);
881: /* test for repeated model */
882: if (mfqP->nmodelpoints == mfqP->last_nmodelpoints) {
883: same = PETSC_TRUE;
884: } else {
885: same = PETSC_FALSE;
886: }
887: for (i = 0; i < mfqP->nmodelpoints; i++) {
888: if (same) {
889: if (mfqP->model_indices[i] == mfqP->last_model_indices[i]) {
890: same = PETSC_TRUE;
891: } else {
892: same = PETSC_FALSE;
893: }
894: }
895: mfqP->last_model_indices[i] = mfqP->model_indices[i];
896: }
897: mfqP->last_nmodelpoints = mfqP->nmodelpoints;
898: if (same && mfqP->delta == deltaold) {
899: PetscCall(PetscInfo(tao, "Identical model used in successive iterations\n"));
900: tao->reason = TAO_CONVERGED_STEPTOL;
901: }
902: }
903: PetscFunctionReturn(PETSC_SUCCESS);
904: }
906: static PetscErrorCode TaoSetUp_POUNDERS(Tao tao)
907: {
908: TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
909: IS isfloc, isfglob, isxloc, isxglob;
911: PetscFunctionBegin;
912: if (!tao->gradient) PetscCall(VecDuplicate(tao->solution, &tao->gradient));
913: if (!tao->stepdirection) PetscCall(VecDuplicate(tao->solution, &tao->stepdirection));
914: PetscCall(VecGetSize(tao->solution, &mfqP->n));
915: PetscCall(VecGetSize(tao->ls_res, &mfqP->m));
916: mfqP->c1 = PetscSqrtReal((PetscReal)mfqP->n);
917: if (mfqP->npmax == PETSC_CURRENT) mfqP->npmax = 2 * mfqP->n + 1;
918: mfqP->npmax = PetscMin((mfqP->n + 1) * (mfqP->n + 2) / 2, mfqP->npmax);
919: mfqP->npmax = PetscMax(mfqP->npmax, mfqP->n + 2);
921: PetscCall(PetscMalloc1(tao->max_funcs + 100, &mfqP->Xhist));
922: PetscCall(PetscMalloc1(tao->max_funcs + 100, &mfqP->Fhist));
923: for (PetscInt i = 0; i < mfqP->n + 1; i++) {
924: PetscCall(VecDuplicate(tao->solution, &mfqP->Xhist[i]));
925: PetscCall(VecDuplicate(tao->ls_res, &mfqP->Fhist[i]));
926: }
927: PetscCall(VecDuplicate(tao->solution, &mfqP->workxvec));
928: PetscCall(VecDuplicate(tao->ls_res, &mfqP->workfvec));
929: mfqP->nHist = 0;
931: PetscCall(PetscMalloc1(tao->max_funcs + 100, &mfqP->Fres));
932: PetscCall(PetscMalloc1(mfqP->npmax * mfqP->m, &mfqP->RES));
933: PetscCall(PetscMalloc1(mfqP->n, &mfqP->work));
934: PetscCall(PetscMalloc1(mfqP->n, &mfqP->work2));
935: PetscCall(PetscMalloc1(mfqP->n, &mfqP->work3));
936: PetscCall(PetscMalloc1(PetscMax(mfqP->m, mfqP->n + 1), &mfqP->mwork));
937: PetscCall(PetscMalloc1(mfqP->npmax - mfqP->n - 1, &mfqP->omega));
938: PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2, &mfqP->beta));
939: PetscCall(PetscMalloc1(mfqP->n + 1, &mfqP->alpha));
941: PetscCall(PetscMalloc1(mfqP->n * mfqP->n * mfqP->m, &mfqP->H));
942: PetscCall(PetscMalloc1(mfqP->npmax * mfqP->npmax, &mfqP->Q));
943: PetscCall(PetscMalloc1(mfqP->npmax * mfqP->npmax, &mfqP->Q_tmp));
944: PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2 * (mfqP->npmax), &mfqP->L));
945: PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2 * (mfqP->npmax), &mfqP->L_tmp));
946: PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2 * (mfqP->npmax), &mfqP->L_save));
947: PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2 * (mfqP->npmax), &mfqP->N));
948: PetscCall(PetscMalloc1(mfqP->npmax * (mfqP->n + 1), &mfqP->M));
949: PetscCall(PetscMalloc1(mfqP->npmax * (mfqP->npmax - mfqP->n - 1), &mfqP->Z));
950: PetscCall(PetscMalloc1(mfqP->npmax, &mfqP->tau));
951: PetscCall(PetscMalloc1(mfqP->npmax, &mfqP->tau_tmp));
952: mfqP->nmax = PetscMax(5 * mfqP->npmax, mfqP->n * (mfqP->n + 1) / 2);
953: PetscCall(PetscMalloc1(mfqP->nmax, &mfqP->npmaxwork));
954: PetscCall(PetscMalloc1(mfqP->nmax, &mfqP->npmaxiwork));
955: PetscCall(PetscMalloc1(mfqP->n, &mfqP->xmin));
956: PetscCall(PetscMalloc1(mfqP->m, &mfqP->C));
957: PetscCall(PetscMalloc1(mfqP->m * mfqP->n, &mfqP->Fdiff));
958: PetscCall(PetscMalloc1(mfqP->npmax * mfqP->n, &mfqP->Disp));
959: PetscCall(PetscMalloc1(mfqP->n, &mfqP->Gres));
960: PetscCall(PetscMalloc1(mfqP->n * mfqP->n, &mfqP->Hres));
961: PetscCall(PetscMalloc1(mfqP->n * mfqP->n, &mfqP->Gpoints));
962: PetscCall(PetscMalloc1(mfqP->npmax, &mfqP->model_indices));
963: PetscCall(PetscMalloc1(mfqP->npmax, &mfqP->last_model_indices));
964: PetscCall(PetscMalloc1(mfqP->n, &mfqP->Xsubproblem));
965: PetscCall(PetscMalloc1(mfqP->m * mfqP->n, &mfqP->Gdel));
966: PetscCall(PetscMalloc1(mfqP->n * mfqP->n * mfqP->m, &mfqP->Hdel));
967: PetscCall(PetscMalloc1(PetscMax(mfqP->m, mfqP->n), &mfqP->indices));
968: PetscCall(PetscMalloc1(mfqP->n, &mfqP->iwork));
969: PetscCall(PetscMalloc1(mfqP->m * mfqP->m, &mfqP->w));
970: for (PetscInt i = 0; i < mfqP->m; i++) {
971: for (PetscInt j = 0; j < mfqP->m; j++) {
972: if (i == j) {
973: mfqP->w[i + mfqP->m * j] = 1.0;
974: } else {
975: mfqP->w[i + mfqP->m * j] = 0.0;
976: }
977: }
978: }
979: for (PetscInt i = 0; i < PetscMax(mfqP->m, mfqP->n); i++) mfqP->indices[i] = i;
980: PetscCallMPI(MPI_Comm_size(((PetscObject)tao)->comm, &mfqP->size));
981: if (mfqP->size > 1) {
982: PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->n, &mfqP->localx));
983: PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->n, &mfqP->localxmin));
984: PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->m, &mfqP->localf));
985: PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->m, &mfqP->localfmin));
986: PetscCall(ISCreateStride(MPI_COMM_SELF, mfqP->n, 0, 1, &isxloc));
987: PetscCall(ISCreateStride(MPI_COMM_SELF, mfqP->n, 0, 1, &isxglob));
988: PetscCall(ISCreateStride(MPI_COMM_SELF, mfqP->m, 0, 1, &isfloc));
989: PetscCall(ISCreateStride(MPI_COMM_SELF, mfqP->m, 0, 1, &isfglob));
991: PetscCall(VecScatterCreate(tao->solution, isxglob, mfqP->localx, isxloc, &mfqP->scatterx));
992: PetscCall(VecScatterCreate(tao->ls_res, isfglob, mfqP->localf, isfloc, &mfqP->scatterf));
994: PetscCall(ISDestroy(&isxloc));
995: PetscCall(ISDestroy(&isxglob));
996: PetscCall(ISDestroy(&isfloc));
997: PetscCall(ISDestroy(&isfglob));
998: }
1000: if (!mfqP->usegqt) {
1001: KSP ksp;
1002: PC pc;
1003: PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, mfqP->n, mfqP->n, mfqP->Xsubproblem, &mfqP->subx));
1004: PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->n, &mfqP->subxl));
1005: PetscCall(VecDuplicate(mfqP->subxl, &mfqP->subb));
1006: PetscCall(VecDuplicate(mfqP->subxl, &mfqP->subxu));
1007: PetscCall(VecDuplicate(mfqP->subxl, &mfqP->subpdel));
1008: PetscCall(VecDuplicate(mfqP->subxl, &mfqP->subndel));
1009: PetscCall(TaoCreate(PETSC_COMM_SELF, &mfqP->subtao));
1010: PetscCall(PetscObjectIncrementTabLevel((PetscObject)mfqP->subtao, (PetscObject)tao, 1));
1011: PetscCall(TaoSetType(mfqP->subtao, TAOBNTR));
1012: PetscCall(TaoSetOptionsPrefix(mfqP->subtao, "pounders_subsolver_"));
1013: PetscCall(TaoSetSolution(mfqP->subtao, mfqP->subx));
1014: PetscCall(TaoSetObjectiveAndGradient(mfqP->subtao, NULL, pounders_fg, (void *)mfqP));
1015: PetscCall(TaoSetMaximumIterations(mfqP->subtao, mfqP->gqt_maxits));
1016: PetscCall(TaoSetFromOptions(mfqP->subtao));
1017: PetscCall(TaoGetKSP(mfqP->subtao, &ksp));
1018: if (ksp) {
1019: PetscCall(KSPGetPC(ksp, &pc));
1020: PetscCall(PCSetType(pc, PCNONE));
1021: }
1022: PetscCall(TaoSetVariableBounds(mfqP->subtao, mfqP->subxl, mfqP->subxu));
1023: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, mfqP->n, mfqP->n, mfqP->Hres, &mfqP->subH));
1024: PetscCall(TaoSetHessian(mfqP->subtao, mfqP->subH, mfqP->subH, pounders_h, (void *)mfqP));
1025: }
1026: PetscFunctionReturn(PETSC_SUCCESS);
1027: }
1029: static PetscErrorCode TaoDestroy_POUNDERS(Tao tao)
1030: {
1031: TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
1033: PetscFunctionBegin;
1034: if (!mfqP->usegqt) {
1035: PetscCall(TaoDestroy(&mfqP->subtao));
1036: PetscCall(VecDestroy(&mfqP->subx));
1037: PetscCall(VecDestroy(&mfqP->subxl));
1038: PetscCall(VecDestroy(&mfqP->subxu));
1039: PetscCall(VecDestroy(&mfqP->subb));
1040: PetscCall(VecDestroy(&mfqP->subpdel));
1041: PetscCall(VecDestroy(&mfqP->subndel));
1042: PetscCall(MatDestroy(&mfqP->subH));
1043: }
1044: PetscCall(PetscFree(mfqP->Fres));
1045: PetscCall(PetscFree(mfqP->RES));
1046: PetscCall(PetscFree(mfqP->work));
1047: PetscCall(PetscFree(mfqP->work2));
1048: PetscCall(PetscFree(mfqP->work3));
1049: PetscCall(PetscFree(mfqP->mwork));
1050: PetscCall(PetscFree(mfqP->omega));
1051: PetscCall(PetscFree(mfqP->beta));
1052: PetscCall(PetscFree(mfqP->alpha));
1053: PetscCall(PetscFree(mfqP->H));
1054: PetscCall(PetscFree(mfqP->Q));
1055: PetscCall(PetscFree(mfqP->Q_tmp));
1056: PetscCall(PetscFree(mfqP->L));
1057: PetscCall(PetscFree(mfqP->L_tmp));
1058: PetscCall(PetscFree(mfqP->L_save));
1059: PetscCall(PetscFree(mfqP->N));
1060: PetscCall(PetscFree(mfqP->M));
1061: PetscCall(PetscFree(mfqP->Z));
1062: PetscCall(PetscFree(mfqP->tau));
1063: PetscCall(PetscFree(mfqP->tau_tmp));
1064: PetscCall(PetscFree(mfqP->npmaxwork));
1065: PetscCall(PetscFree(mfqP->npmaxiwork));
1066: PetscCall(PetscFree(mfqP->xmin));
1067: PetscCall(PetscFree(mfqP->C));
1068: PetscCall(PetscFree(mfqP->Fdiff));
1069: PetscCall(PetscFree(mfqP->Disp));
1070: PetscCall(PetscFree(mfqP->Gres));
1071: PetscCall(PetscFree(mfqP->Hres));
1072: PetscCall(PetscFree(mfqP->Gpoints));
1073: PetscCall(PetscFree(mfqP->model_indices));
1074: PetscCall(PetscFree(mfqP->last_model_indices));
1075: PetscCall(PetscFree(mfqP->Xsubproblem));
1076: PetscCall(PetscFree(mfqP->Gdel));
1077: PetscCall(PetscFree(mfqP->Hdel));
1078: PetscCall(PetscFree(mfqP->indices));
1079: PetscCall(PetscFree(mfqP->iwork));
1080: PetscCall(PetscFree(mfqP->w));
1081: for (PetscInt i = 0; i < mfqP->nHist; i++) {
1082: PetscCall(VecDestroy(&mfqP->Xhist[i]));
1083: PetscCall(VecDestroy(&mfqP->Fhist[i]));
1084: }
1085: PetscCall(VecDestroy(&mfqP->workxvec));
1086: PetscCall(VecDestroy(&mfqP->workfvec));
1087: PetscCall(PetscFree(mfqP->Xhist));
1088: PetscCall(PetscFree(mfqP->Fhist));
1090: if (mfqP->size > 1) {
1091: PetscCall(VecDestroy(&mfqP->localx));
1092: PetscCall(VecDestroy(&mfqP->localxmin));
1093: PetscCall(VecDestroy(&mfqP->localf));
1094: PetscCall(VecDestroy(&mfqP->localfmin));
1095: }
1096: PetscCall(PetscFree(tao->data));
1097: PetscFunctionReturn(PETSC_SUCCESS);
1098: }
1100: static PetscErrorCode TaoSetFromOptions_POUNDERS(Tao tao, PetscOptionItems PetscOptionsObject)
1101: {
1102: TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
1104: PetscFunctionBegin;
1105: PetscOptionsHeadBegin(PetscOptionsObject, "POUNDERS method for least-squares optimization");
1106: PetscCall(PetscOptionsReal("-tao_pounders_delta", "initial delta", "", mfqP->delta, &mfqP->delta0, NULL));
1107: mfqP->delta = mfqP->delta0;
1108: PetscCall(PetscOptionsInt("-tao_pounders_npmax", "max number of points in model", "", mfqP->npmax, &mfqP->npmax, NULL));
1109: PetscCall(PetscOptionsBool("-tao_pounders_gqt", "use gqt algorithm for subproblem", "", mfqP->usegqt, &mfqP->usegqt, NULL));
1110: PetscOptionsHeadEnd();
1111: PetscFunctionReturn(PETSC_SUCCESS);
1112: }
1114: static PetscErrorCode TaoView_POUNDERS(Tao tao, PetscViewer viewer)
1115: {
1116: TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
1117: PetscBool isascii;
1119: PetscFunctionBegin;
1120: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1121: if (isascii) {
1122: PetscCall(PetscViewerASCIIPrintf(viewer, "initial delta: %g\n", (double)mfqP->delta0));
1123: PetscCall(PetscViewerASCIIPrintf(viewer, "final delta: %g\n", (double)mfqP->delta));
1124: PetscCall(PetscViewerASCIIPrintf(viewer, "model points: %" PetscInt_FMT "\n", mfqP->nmodelpoints));
1125: if (mfqP->usegqt) {
1126: PetscCall(PetscViewerASCIIPrintf(viewer, "subproblem solver: gqt\n"));
1127: } else {
1128: PetscCall(TaoView(mfqP->subtao, viewer));
1129: }
1130: }
1131: PetscFunctionReturn(PETSC_SUCCESS);
1132: }
1133: /*MC
1134: TAOPOUNDERS - POUNDERS derivate-free model-based algorithm for nonlinear least squares
1136: Options Database Keys:
1137: + -tao_pounders_delta - initial step length
1138: . -tao_pounders_npmax - maximum number of points in model
1139: - -tao_pounders_gqt - use gqt algorithm for subproblem instead of TRON
1141: Level: beginner
1143: .seealso: `Tao`, `TAONTR`, `TAONTL`, `TAONM`, `TaoType`, `TaoCreate()`
1144: M*/
1146: PETSC_EXTERN PetscErrorCode TaoCreate_POUNDERS(Tao tao)
1147: {
1148: TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
1150: PetscFunctionBegin;
1151: tao->ops->setup = TaoSetUp_POUNDERS;
1152: tao->ops->solve = TaoSolve_POUNDERS;
1153: tao->ops->view = TaoView_POUNDERS;
1154: tao->ops->setfromoptions = TaoSetFromOptions_POUNDERS;
1155: tao->ops->destroy = TaoDestroy_POUNDERS;
1157: PetscCall(PetscNew(&mfqP));
1158: tao->data = (void *)mfqP;
1160: /* Override default settings (unless already changed) */
1161: PetscObjectParameterSetDefault(tao, max_it, 2000);
1162: PetscObjectParameterSetDefault(tao, max_funcs, 4000);
1164: mfqP->npmax = PETSC_CURRENT;
1165: mfqP->delta0 = 0.1;
1166: mfqP->delta = 0.1;
1167: mfqP->deltamax = 1e3;
1168: mfqP->deltamin = 1e-6;
1169: mfqP->c2 = 10.0;
1170: mfqP->theta1 = 1e-5;
1171: mfqP->theta2 = 1e-4;
1172: mfqP->gamma0 = 0.5;
1173: mfqP->gamma1 = 2.0;
1174: mfqP->eta0 = 0.0;
1175: mfqP->eta1 = 0.1;
1176: mfqP->usegqt = PETSC_FALSE;
1177: mfqP->gqt_rtol = 0.001;
1178: mfqP->gqt_maxits = 50;
1179: mfqP->workxvec = NULL;
1180: PetscFunctionReturn(PETSC_SUCCESS);
1181: }