xref: /petsc/src/tao/leastsquares/impls/pounders/pounders.c (revision 4e8208cbcbc709572b8abe32f33c78b69c819375)
1 #include <../src/tao/leastsquares/impls/pounders/pounders.h>
2 
pounders_h(Tao subtao,Vec v,Mat H,Mat Hpre,PetscCtx ctx)3 static PetscErrorCode pounders_h(Tao subtao, Vec v, Mat H, Mat Hpre, PetscCtx ctx)
4 {
5   PetscFunctionBegin;
6   PetscFunctionReturn(PETSC_SUCCESS);
7 }
8 
pounders_fg(Tao subtao,Vec x,PetscReal * f,Vec g,PetscCtx ctx)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;
13 
14   PetscFunctionBegin;
15   /* g = A*x  (add b later)*/
16   PetscCall(MatMult(mfqP->subH, x, g));
17 
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;
22 
23   /* now  g = g + b */
24   PetscCall(VecAXPY(g, 1.0, mfqP->subb));
25   PetscFunctionReturn(PETSC_SUCCESS);
26 }
27 
pounders_feval(Tao tao,Vec x,Vec F,PetscReal * fsum)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;
33 
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 }
55 
gqtwrap(Tao tao,PetscReal * gnorm,PetscReal * qmin)56 static PetscErrorCode gqtwrap(Tao tao, PetscReal *gnorm, PetscReal *qmin)
57 {
58 #if defined(PETSC_USE_REAL_SINGLE)
59   PetscReal atol = 1.0e-5;
60 #else
61   PetscReal atol = 1.0e-10;
62 #endif
63   PetscInt      info, its;
64   TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
65 
66   PetscFunctionBegin;
67   if (!mfqP->usegqt) {
68     PetscReal maxval;
69     PetscInt  i, j;
70 
71     PetscCall(VecSetValues(mfqP->subb, mfqP->n, mfqP->indices, mfqP->Gres, INSERT_VALUES));
72     PetscCall(VecAssemblyBegin(mfqP->subb));
73     PetscCall(VecAssemblyEnd(mfqP->subb));
74 
75     PetscCall(VecSet(mfqP->subx, 0.0));
76 
77     PetscCall(VecSet(mfqP->subndel, -1.0));
78     PetscCall(VecSet(mfqP->subpdel, +1.0));
79 
80     /* Complete the lower triangle of the Hessian matrix */
81     for (i = 0; i < mfqP->n; i++) {
82       for (j = i + 1; j < mfqP->n; j++) mfqP->Hres[j + mfqP->n * i] = mfqP->Hres[mfqP->n * j + i];
83     }
84     PetscCall(MatSetValues(mfqP->subH, mfqP->n, mfqP->indices, mfqP->n, mfqP->indices, mfqP->Hres, INSERT_VALUES));
85     PetscCall(MatAssemblyBegin(mfqP->subH, MAT_FINAL_ASSEMBLY));
86     PetscCall(MatAssemblyEnd(mfqP->subH, MAT_FINAL_ASSEMBLY));
87 
88     PetscCall(TaoResetStatistics(mfqP->subtao));
89     /* PetscCall(TaoSetTolerances(mfqP->subtao,*gnorm,*gnorm,PETSC_CURRENT)); */
90     /* enforce bound constraints -- experimental */
91     if (tao->XU && tao->XL) {
92       PetscCall(VecCopy(tao->XU, mfqP->subxu));
93       PetscCall(VecAXPY(mfqP->subxu, -1.0, tao->solution));
94       PetscCall(VecScale(mfqP->subxu, 1.0 / mfqP->delta));
95       PetscCall(VecCopy(tao->XL, mfqP->subxl));
96       PetscCall(VecAXPY(mfqP->subxl, -1.0, tao->solution));
97       PetscCall(VecScale(mfqP->subxl, 1.0 / mfqP->delta));
98 
99       PetscCall(VecPointwiseMin(mfqP->subxu, mfqP->subxu, mfqP->subpdel));
100       PetscCall(VecPointwiseMax(mfqP->subxl, mfqP->subxl, mfqP->subndel));
101     } else {
102       PetscCall(VecCopy(mfqP->subpdel, mfqP->subxu));
103       PetscCall(VecCopy(mfqP->subndel, mfqP->subxl));
104     }
105     /* Make sure xu > xl */
106     PetscCall(VecCopy(mfqP->subxl, mfqP->subpdel));
107     PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subxu));
108     PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
109     PetscCheck(maxval <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "upper bound < lower bound in subproblem");
110     /* Make sure xu > tao->solution > xl */
111     PetscCall(VecCopy(mfqP->subxl, mfqP->subpdel));
112     PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subx));
113     PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
114     PetscCheck(maxval <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "initial guess < lower bound in subproblem");
115 
116     PetscCall(VecCopy(mfqP->subx, mfqP->subpdel));
117     PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subxu));
118     PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
119     PetscCheck(maxval <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "initial guess > upper bound in subproblem");
120 
121     PetscCall(TaoSolve(mfqP->subtao));
122     PetscCall(TaoGetSolutionStatus(mfqP->subtao, NULL, qmin, NULL, NULL, NULL, NULL));
123 
124     /* test bounds post-solution*/
125     PetscCall(VecCopy(mfqP->subxl, mfqP->subpdel));
126     PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subx));
127     PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
128     if (maxval > 1e-5) {
129       PetscCall(PetscInfo(tao, "subproblem solution < lower bound\n"));
130       tao->reason = TAO_DIVERGED_TR_REDUCTION;
131     }
132 
133     PetscCall(VecCopy(mfqP->subx, mfqP->subpdel));
134     PetscCall(VecAXPY(mfqP->subpdel, -1.0, mfqP->subxu));
135     PetscCall(VecMax(mfqP->subpdel, NULL, &maxval));
136     if (maxval > 1e-5) {
137       PetscCall(PetscInfo(tao, "subproblem solution > upper bound\n"));
138       tao->reason = TAO_DIVERGED_TR_REDUCTION;
139     }
140   } else {
141     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));
142   }
143   *qmin *= -1;
144   PetscFunctionReturn(PETSC_SUCCESS);
145 }
146 
pounders_update_res(Tao tao)147 static PetscErrorCode pounders_update_res(Tao tao)
148 {
149   TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
150   PetscInt      i, row, col;
151   PetscBLASInt  blasn, blasn2, blasm, ione = 1;
152   PetscReal     zero = 0.0, one = 1.0, wii, factor;
153 
154   PetscFunctionBegin;
155   PetscCall(PetscBLASIntCast(mfqP->n, &blasn));
156   PetscCall(PetscBLASIntCast(mfqP->m, &blasm));
157   PetscCall(PetscBLASIntCast(mfqP->n * mfqP->n, &blasn2));
158   for (i = 0; i < mfqP->n; i++) mfqP->Gres[i] = 0;
159   for (i = 0; i < mfqP->n * mfqP->n; i++) mfqP->Hres[i] = 0;
160 
161   /* Compute Gres= sum_ij[wij * (cjgi + cigj)] */
162   if (tao->res_weights_v) {
163     /* Vector(diagonal) weights: gres = sum_i(wii*ci*gi) */
164     for (i = 0; i < mfqP->m; i++) {
165       PetscCall(VecGetValues(tao->res_weights_v, 1, &i, &factor));
166       factor = factor * mfqP->C[i];
167       PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn, &factor, &mfqP->Fdiff[blasn * i], &ione, mfqP->Gres, &ione));
168     }
169 
170     /* compute Hres = sum_ij [wij * (*ci*Hj + cj*Hi + gi gj' + gj gi') ] */
171     /* vector(diagonal weights) Hres = sum_i(wii*(ci*Hi + gi * gi')*/
172     for (i = 0; i < mfqP->m; i++) {
173       PetscCall(VecGetValues(tao->res_weights_v, 1, &i, &wii));
174       if (tao->niter > 1) {
175         factor = wii * mfqP->C[i];
176         /* add wii * ci * Hi */
177         PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn2, &factor, &mfqP->H[i], &blasm, mfqP->Hres, &ione));
178       }
179       /* add wii * gi * gi' */
180       PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &blasn, &blasn, &ione, &wii, &mfqP->Fdiff[blasn * i], &blasn, &mfqP->Fdiff[blasn * i], &blasn, &one, mfqP->Hres, &blasn));
181     }
182   } else if (tao->res_weights_w) {
183     /* General case: .5 * Gres= sum_ij[wij * (cjgi + cigj)] */
184     for (i = 0; i < tao->res_weights_n; i++) {
185       row = tao->res_weights_rows[i];
186       col = tao->res_weights_cols[i];
187 
188       factor = tao->res_weights_w[i] * mfqP->C[col] / 2.0;
189       PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn, &factor, &mfqP->Fdiff[blasn * row], &ione, mfqP->Gres, &ione));
190       factor = tao->res_weights_w[i] * mfqP->C[row] / 2.0;
191       PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn, &factor, &mfqP->Fdiff[blasn * col], &ione, mfqP->Gres, &ione));
192     }
193 
194     /* compute Hres = sum_ij [wij * (*ci*Hj + cj*Hi + gi gj' + gj gi') ] */
195     /* .5 * sum_ij [wij * (*ci*Hj + cj*Hi + gi gj' + gj gi') ] */
196     for (i = 0; i < tao->res_weights_n; i++) {
197       row    = tao->res_weights_rows[i];
198       col    = tao->res_weights_cols[i];
199       factor = tao->res_weights_w[i] / 2.0;
200       /* add wij * gi gj' + wij * gj gi' */
201       PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &blasn, &blasn, &ione, &factor, &mfqP->Fdiff[blasn * row], &blasn, &mfqP->Fdiff[blasn * col], &blasn, &one, mfqP->Hres, &blasn));
202       PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &blasn, &blasn, &ione, &factor, &mfqP->Fdiff[blasn * col], &blasn, &mfqP->Fdiff[blasn * row], &blasn, &one, mfqP->Hres, &blasn));
203     }
204     if (tao->niter > 1) {
205       for (i = 0; i < tao->res_weights_n; i++) {
206         row = tao->res_weights_rows[i];
207         col = tao->res_weights_cols[i];
208 
209         /* add  wij*cj*Hi */
210         factor = tao->res_weights_w[i] * mfqP->C[col] / 2.0;
211         PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn2, &factor, &mfqP->H[row], &blasm, mfqP->Hres, &ione));
212 
213         /* add wij*ci*Hj */
214         factor = tao->res_weights_w[i] * mfqP->C[row] / 2.0;
215         PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn2, &factor, &mfqP->H[col], &blasm, mfqP->Hres, &ione));
216       }
217     }
218   } else {
219     /* Default: Gres= sum_i[cigi] = G*c' */
220     PetscCall(PetscInfo(tao, "Identity weights\n"));
221     PetscCallBLAS("BLASgemv", BLASgemv_("N", &blasn, &blasm, &one, mfqP->Fdiff, &blasn, mfqP->C, &ione, &zero, mfqP->Gres, &ione));
222 
223     /* compute Hres = sum_ij [wij * (*ci*Hj + cj*Hi + gi gj' + gj gi') ] */
224     /*  Hres = G*G' + 0.5 sum {F(xkin,i)*H(:,:,i)}  */
225     PetscCallBLAS("BLASgemm", BLASgemm_("N", "T", &blasn, &blasn, &blasm, &one, mfqP->Fdiff, &blasn, mfqP->Fdiff, &blasn, &zero, mfqP->Hres, &blasn));
226 
227     /* sum(F(xkin,i)*H(:,:,i)) */
228     if (tao->niter > 1) {
229       for (i = 0; i < mfqP->m; i++) {
230         factor = mfqP->C[i];
231         PetscCallBLAS("BLASaxpy", BLASaxpy_(&blasn2, &factor, &mfqP->H[i], &blasm, mfqP->Hres, &ione));
232       }
233     }
234   }
235   PetscFunctionReturn(PETSC_SUCCESS);
236 }
237 
phi2eval(PetscReal * x,PetscInt n,PetscReal * phi)238 static PetscErrorCode phi2eval(PetscReal *x, PetscInt n, PetscReal *phi)
239 {
240   /* 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] */
241   PetscInt  i, j, k;
242   PetscReal sqrt2 = PetscSqrtReal(2.0);
243 
244   PetscFunctionBegin;
245   j = 0;
246   for (i = 0; i < n; i++) {
247     phi[j] = 0.5 * x[i] * x[i];
248     j++;
249     for (k = i + 1; k < n; k++) {
250       phi[j] = x[i] * x[k] / sqrt2;
251       j++;
252     }
253   }
254   PetscFunctionReturn(PETSC_SUCCESS);
255 }
256 
getquadpounders(TAO_POUNDERS * mfqP)257 static PetscErrorCode getquadpounders(TAO_POUNDERS *mfqP)
258 {
259   /* Computes the parameters of the quadratic Q(x) = c + g'*x + 0.5*x*G*x'
260    that satisfies the interpolation conditions Q(X[:,j]) = f(j)
261    for j=1,...,m and with a Hessian matrix of least Frobenius norm */
262 
263   /* NB --we are ignoring c */
264   PetscInt     i, j, k, num, np = mfqP->nmodelpoints;
265   PetscReal    one = 1.0, zero = 0.0, negone = -1.0;
266   PetscBLASInt blasnpmax, blasnplus1, blasnp, blasint, blasint2;
267   PetscBLASInt info, ione = 1;
268   PetscReal    sqrt2 = PetscSqrtReal(2.0);
269 
270   PetscFunctionBegin;
271   PetscCall(PetscBLASIntCast(mfqP->npmax, &blasnpmax));
272   PetscCall(PetscBLASIntCast(mfqP->n + 1, &blasnplus1));
273   PetscCall(PetscBLASIntCast(np, &blasnp));
274   PetscCall(PetscBLASIntCast(mfqP->n * (mfqP->n + 1) / 2, &blasint));
275   PetscCall(PetscBLASIntCast(np - mfqP->n - 1, &blasint2));
276   for (i = 0; i < mfqP->n * mfqP->m; i++) mfqP->Gdel[i] = 0;
277   for (i = 0; i < mfqP->n * mfqP->n * mfqP->m; i++) mfqP->Hdel[i] = 0;
278 
279   /* factor M */
280   PetscCallBLAS("LAPACKgetrf", LAPACKgetrf_(&blasnplus1, &blasnp, mfqP->M, &blasnplus1, mfqP->npmaxiwork, &info));
281   PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine getrf returned with value %" PetscBLASInt_FMT, info);
282 
283   if (np == mfqP->n + 1) {
284     for (i = 0; i < mfqP->npmax - mfqP->n - 1; i++) mfqP->omega[i] = 0.0;
285     for (i = 0; i < mfqP->n * (mfqP->n + 1) / 2; i++) mfqP->beta[i] = 0.0;
286   } else {
287     /* Let Ltmp = (L'*L) */
288     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));
289 
290     /* factor Ltmp */
291     PetscCallBLAS("LAPACKpotrf", LAPACKpotrf_("L", &blasint2, mfqP->L_tmp, &blasint, &info));
292     PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine potrf returned with value %" PetscBLASInt_FMT, info);
293   }
294 
295   for (k = 0; k < mfqP->m; k++) {
296     if (np != mfqP->n + 1) {
297       /* Solve L'*L*Omega = Z' * RESk*/
298       PetscCallBLAS("BLASgemv", BLASgemv_("T", &blasnp, &blasint2, &one, mfqP->Z, &blasnpmax, &mfqP->RES[mfqP->npmax * k], &ione, &zero, mfqP->omega, &ione));
299       PetscCallBLAS("LAPACKpotrs", LAPACKpotrs_("L", &blasint2, &ione, mfqP->L_tmp, &blasint, mfqP->omega, &blasint2, &info));
300       PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine potrs returned with value %" PetscBLASInt_FMT, info);
301 
302       /* Beta = L*Omega */
303       PetscCallBLAS("BLASgemv", BLASgemv_("N", &blasint, &blasint2, &one, &mfqP->L[(mfqP->n + 1) * blasint], &blasint, mfqP->omega, &ione, &zero, mfqP->beta, &ione));
304     }
305 
306     /* solve M'*Alpha = RESk - N'*Beta */
307     PetscCallBLAS("BLASgemv", BLASgemv_("T", &blasint, &blasnp, &negone, mfqP->N, &blasint, mfqP->beta, &ione, &one, &mfqP->RES[mfqP->npmax * k], &ione));
308     PetscCallBLAS("LAPACKgetrs", LAPACKgetrs_("T", &blasnplus1, &ione, mfqP->M, &blasnplus1, mfqP->npmaxiwork, &mfqP->RES[mfqP->npmax * k], &blasnplus1, &info));
309     PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine getrs returned with value %" PetscBLASInt_FMT, info);
310 
311     /* Gdel(:,k) = Alpha(2:n+1) */
312     for (i = 0; i < mfqP->n; i++) mfqP->Gdel[i + mfqP->n * k] = mfqP->RES[mfqP->npmax * k + i + 1];
313 
314     /* Set Hdels */
315     num = 0;
316     for (i = 0; i < mfqP->n; i++) {
317       /* H[i,i,k] = Beta(num) */
318       mfqP->Hdel[(i * mfqP->n + i) * mfqP->m + k] = mfqP->beta[num];
319       num++;
320       for (j = i + 1; j < mfqP->n; j++) {
321         /* H[i,j,k] = H[j,i,k] = Beta(num)/sqrt(2) */
322         mfqP->Hdel[(j * mfqP->n + i) * mfqP->m + k] = mfqP->beta[num] / sqrt2;
323         mfqP->Hdel[(i * mfqP->n + j) * mfqP->m + k] = mfqP->beta[num] / sqrt2;
324         num++;
325       }
326     }
327   }
328   PetscFunctionReturn(PETSC_SUCCESS);
329 }
330 
morepoints(TAO_POUNDERS * mfqP)331 static PetscErrorCode morepoints(TAO_POUNDERS *mfqP)
332 {
333   /* Assumes mfqP->model_indices[0]  is minimum index
334    Finishes adding points to mfqP->model_indices (up to npmax)
335    Computes L,Z,M,N
336    np is actual number of points in model (should equal npmax?) */
337   PetscInt         point, i, j, offset;
338   PetscInt         reject;
339   PetscBLASInt     blasn, blasnpmax, blasnplus1, info, blasnmax, blasint, blasint2, blasnp, blasmaxmn;
340   const PetscReal *x;
341   PetscReal        normd;
342 
343   PetscFunctionBegin;
344   PetscCall(PetscBLASIntCast(mfqP->npmax, &blasnpmax));
345   PetscCall(PetscBLASIntCast(mfqP->n, &blasn));
346   PetscCall(PetscBLASIntCast(mfqP->nmax, &blasnmax));
347   PetscCall(PetscBLASIntCast(mfqP->n + 1, &blasnplus1));
348   PetscCall(PetscBLASIntCast(mfqP->n, &blasnp));
349   /* Initialize M,N */
350   for (i = 0; i < mfqP->n + 1; i++) {
351     PetscCall(VecGetArrayRead(mfqP->Xhist[mfqP->model_indices[i]], &x));
352     mfqP->M[(mfqP->n + 1) * i] = 1.0;
353     for (j = 0; j < mfqP->n; j++) mfqP->M[j + 1 + ((mfqP->n + 1) * i)] = (x[j] - mfqP->xmin[j]) / mfqP->delta;
354     PetscCall(VecRestoreArrayRead(mfqP->Xhist[mfqP->model_indices[i]], &x));
355     PetscCall(phi2eval(&mfqP->M[1 + ((mfqP->n + 1) * i)], mfqP->n, &mfqP->N[mfqP->n * (mfqP->n + 1) / 2 * i]));
356   }
357 
358   /* Now we add points until we have npmax starting with the most recent ones */
359   point              = mfqP->nHist - 1;
360   mfqP->nmodelpoints = mfqP->n + 1;
361   while (mfqP->nmodelpoints < mfqP->npmax && point >= 0) {
362     /* Reject any points already in the model */
363     reject = 0;
364     for (j = 0; j < mfqP->n + 1; j++) {
365       if (point == mfqP->model_indices[j]) {
366         reject = 1;
367         break;
368       }
369     }
370 
371     /* Reject if norm(d) >c2 */
372     if (!reject) {
373       PetscCall(VecCopy(mfqP->Xhist[point], mfqP->workxvec));
374       PetscCall(VecAXPY(mfqP->workxvec, -1.0, mfqP->Xhist[mfqP->minindex]));
375       PetscCall(VecNorm(mfqP->workxvec, NORM_2, &normd));
376       normd /= mfqP->delta;
377       if (normd > mfqP->c2) reject = 1;
378     }
379     if (reject) {
380       point--;
381       continue;
382     }
383 
384     PetscCall(VecGetArrayRead(mfqP->Xhist[point], &x));
385     mfqP->M[(mfqP->n + 1) * mfqP->nmodelpoints] = 1.0;
386     for (j = 0; j < mfqP->n; j++) mfqP->M[j + 1 + ((mfqP->n + 1) * mfqP->nmodelpoints)] = (x[j] - mfqP->xmin[j]) / mfqP->delta;
387     PetscCall(VecRestoreArrayRead(mfqP->Xhist[point], &x));
388     PetscCall(phi2eval(&mfqP->M[1 + (mfqP->n + 1) * mfqP->nmodelpoints], mfqP->n, &mfqP->N[mfqP->n * (mfqP->n + 1) / 2 * (mfqP->nmodelpoints)]));
389 
390     /* Update QR factorization */
391     /* Copy M' to Q_tmp */
392     for (i = 0; i < mfqP->n + 1; i++) {
393       for (j = 0; j < mfqP->npmax; j++) mfqP->Q_tmp[j + mfqP->npmax * i] = mfqP->M[i + (mfqP->n + 1) * j];
394     }
395     PetscCall(PetscBLASIntCast(mfqP->nmodelpoints + 1, &blasnp));
396     /* Q_tmp,R = qr(M') */
397     PetscCall(PetscBLASIntCast(PetscMax(mfqP->m, mfqP->n + 1), &blasmaxmn));
398     PetscCallBLAS("LAPACKgeqrf", LAPACKgeqrf_(&blasnp, &blasnplus1, mfqP->Q_tmp, &blasnpmax, mfqP->tau_tmp, mfqP->mwork, &blasmaxmn, &info));
399     PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine geqrf returned with value %" PetscBLASInt_FMT, info);
400 
401     /* Reject if min(svd(N*Q(:,n+2:np+1)) <= theta2 */
402     /* L = N*Qtmp */
403     PetscCall(PetscBLASIntCast(mfqP->n * (mfqP->n + 1) / 2, &blasint2));
404     /* Copy N to L_tmp */
405     for (i = 0; i < mfqP->n * (mfqP->n + 1) / 2 * mfqP->npmax; i++) mfqP->L_tmp[i] = mfqP->N[i];
406     /* Copy L_save to L_tmp */
407 
408     /* L_tmp = N*Qtmp' */
409     PetscCallBLAS("LAPACKormqr", LAPACKormqr_("R", "N", &blasint2, &blasnp, &blasnplus1, mfqP->Q_tmp, &blasnpmax, mfqP->tau_tmp, mfqP->L_tmp, &blasint2, mfqP->npmaxwork, &blasnmax, &info));
410     PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine ormqr returned with value %" PetscBLASInt_FMT, info);
411 
412     /* Copy L_tmp to L_save */
413     for (i = 0; i < mfqP->npmax * mfqP->n * (mfqP->n + 1) / 2; i++) mfqP->L_save[i] = mfqP->L_tmp[i];
414 
415     /* Get svd for L_tmp(:,n+2:np+1) (L_tmp is modified in process) */
416     PetscCall(PetscBLASIntCast(mfqP->nmodelpoints - mfqP->n, &blasint));
417     PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF));
418     PetscCallBLAS("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));
419     PetscCall(PetscFPTrapPop());
420     PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine gesvd returned with value %" PetscBLASInt_FMT, info);
421 
422     if (mfqP->beta[PetscMin(blasint, blasint2) - 1] > mfqP->theta2) {
423       /* accept point */
424       mfqP->model_indices[mfqP->nmodelpoints] = point;
425       /* Copy Q_tmp to Q */
426       for (i = 0; i < mfqP->npmax * mfqP->npmax; i++) mfqP->Q[i] = mfqP->Q_tmp[i];
427       for (i = 0; i < mfqP->npmax; i++) mfqP->tau[i] = mfqP->tau_tmp[i];
428       mfqP->nmodelpoints++;
429       PetscCall(PetscBLASIntCast(mfqP->nmodelpoints, &blasnp));
430 
431       /* Copy L_save to L */
432       for (i = 0; i < mfqP->npmax * mfqP->n * (mfqP->n + 1) / 2; i++) mfqP->L[i] = mfqP->L_save[i];
433     }
434     point--;
435   }
436 
437   PetscCall(PetscBLASIntCast(mfqP->nmodelpoints, &blasnp));
438   /* Copy Q(:,n+2:np) to Z */
439   /* First set Q_tmp to I */
440   for (i = 0; i < mfqP->npmax * mfqP->npmax; i++) mfqP->Q_tmp[i] = 0.0;
441   for (i = 0; i < mfqP->npmax; i++) mfqP->Q_tmp[i + mfqP->npmax * i] = 1.0;
442 
443   /* Q_tmp = I * Q */
444   PetscCallBLAS("LAPACKormqr", LAPACKormqr_("R", "N", &blasnp, &blasnp, &blasnplus1, mfqP->Q, &blasnpmax, mfqP->tau, mfqP->Q_tmp, &blasnpmax, mfqP->npmaxwork, &blasnmax, &info));
445   PetscCheck(info == 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "LAPACK routine ormqr returned with value %" PetscBLASInt_FMT, info);
446 
447   /* Copy Q_tmp(:,n+2:np) to Z) */
448   offset = mfqP->npmax * (mfqP->n + 1);
449   for (i = offset; i < mfqP->npmax * mfqP->npmax; i++) mfqP->Z[i - offset] = mfqP->Q_tmp[i];
450 
451   if (mfqP->nmodelpoints == mfqP->n + 1) {
452     /* Set L to I_{n+1} */
453     for (i = 0; i < mfqP->npmax * mfqP->n * (mfqP->n + 1) / 2; i++) mfqP->L[i] = 0.0;
454     for (i = 0; i < mfqP->n; i++) mfqP->L[(mfqP->n * (mfqP->n + 1) / 2) * i + i] = 1.0;
455   }
456   PetscFunctionReturn(PETSC_SUCCESS);
457 }
458 
459 /* Only call from modelimprove, addpoint() needs ->Q_tmp and ->work to be set */
addpoint(Tao tao,TAO_POUNDERS * mfqP,PetscInt index)460 static PetscErrorCode addpoint(Tao tao, TAO_POUNDERS *mfqP, PetscInt index)
461 {
462   PetscFunctionBegin;
463   /* Create new vector in history: X[newidx] = X[mfqP->index] + delta*X[index]*/
464   PetscCall(VecDuplicate(mfqP->Xhist[0], &mfqP->Xhist[mfqP->nHist]));
465   PetscCall(VecSetValues(mfqP->Xhist[mfqP->nHist], mfqP->n, mfqP->indices, &mfqP->Q_tmp[index * mfqP->npmax], INSERT_VALUES));
466   PetscCall(VecAssemblyBegin(mfqP->Xhist[mfqP->nHist]));
467   PetscCall(VecAssemblyEnd(mfqP->Xhist[mfqP->nHist]));
468   PetscCall(VecAYPX(mfqP->Xhist[mfqP->nHist], mfqP->delta, mfqP->Xhist[mfqP->minindex]));
469 
470   /* Project into feasible region */
471   if (tao->XU && tao->XL) PetscCall(VecMedian(mfqP->Xhist[mfqP->nHist], tao->XL, tao->XU, mfqP->Xhist[mfqP->nHist]));
472 
473   /* Compute value of new vector */
474   PetscCall(VecDuplicate(mfqP->Fhist[0], &mfqP->Fhist[mfqP->nHist]));
475   CHKMEMQ;
476   PetscCall(pounders_feval(tao, mfqP->Xhist[mfqP->nHist], mfqP->Fhist[mfqP->nHist], &mfqP->Fres[mfqP->nHist]));
477 
478   /* Add new vector to model */
479   mfqP->model_indices[mfqP->nmodelpoints] = mfqP->nHist;
480   mfqP->nmodelpoints++;
481   mfqP->nHist++;
482   PetscFunctionReturn(PETSC_SUCCESS);
483 }
484 
modelimprove(Tao tao,TAO_POUNDERS * mfqP,PetscInt addallpoints)485 static PetscErrorCode modelimprove(Tao tao, TAO_POUNDERS *mfqP, PetscInt addallpoints)
486 {
487   /* modeld = Q(:,np+1:n)' */
488   PetscInt     i, j, minindex = 0;
489   PetscReal    dp, half = 0.5, one = 1.0, minvalue = PETSC_INFINITY;
490   PetscBLASInt blasn, blasnpmax, blask, info;
491   PetscBLASInt blas1 = 1, blasnmax;
492 
493   PetscFunctionBegin;
494   PetscCall(PetscBLASIntCast(mfqP->n, &blasn));
495   PetscCall(PetscBLASIntCast(mfqP->npmax, &blasnpmax));
496   PetscCall(PetscBLASIntCast(mfqP->nmodelpoints, &blask));
497   PetscCall(PetscBLASIntCast(mfqP->nmax, &blasnmax));
498 
499   /* Qtmp = I(n x n) */
500   for (i = 0; i < mfqP->n; i++) {
501     for (j = 0; j < mfqP->n; j++) mfqP->Q_tmp[i + mfqP->npmax * j] = 0.0;
502   }
503   for (j = 0; j < mfqP->n; j++) mfqP->Q_tmp[j + mfqP->npmax * j] = 1.0;
504 
505   /* Qtmp = Q * I */
506   PetscCallBLAS("LAPACKormqr", LAPACKormqr_("R", "N", &blasn, &blasn, &blask, mfqP->Q, &blasnpmax, mfqP->tau, mfqP->Q_tmp, &blasnpmax, mfqP->npmaxwork, &blasnmax, &info));
507 
508   for (i = mfqP->nmodelpoints; i < mfqP->n; i++) {
509     PetscCallBLAS("BLASdot", dp = BLASdot_(&blasn, &mfqP->Q_tmp[i * mfqP->npmax], &blas1, mfqP->Gres, &blas1));
510     if (dp > 0.0) { /* Model says use the other direction! */
511       for (j = 0; j < mfqP->n; j++) mfqP->Q_tmp[i * mfqP->npmax + j] *= -1;
512     }
513     /* mfqP->work[i] = Cres+Modeld(i,:)*(Gres+.5*Hres*Modeld(i,:)') */
514     for (j = 0; j < mfqP->n; j++) mfqP->work2[j] = mfqP->Gres[j];
515     PetscCallBLAS("BLASgemv", BLASgemv_("N", &blasn, &blasn, &half, mfqP->Hres, &blasn, &mfqP->Q_tmp[i * mfqP->npmax], &blas1, &one, mfqP->work2, &blas1));
516     PetscCallBLAS("BLASdot", mfqP->work[i] = BLASdot_(&blasn, &mfqP->Q_tmp[i * mfqP->npmax], &blas1, mfqP->work2, &blas1));
517     if (i == mfqP->nmodelpoints || mfqP->work[i] < minvalue) {
518       minindex = i;
519       minvalue = mfqP->work[i];
520     }
521     if (addallpoints != 0) PetscCall(addpoint(tao, mfqP, i));
522   }
523   if (!addallpoints) PetscCall(addpoint(tao, mfqP, minindex));
524   PetscFunctionReturn(PETSC_SUCCESS);
525 }
526 
affpoints(TAO_POUNDERS * mfqP,PetscReal * xmin,PetscReal c)527 static PetscErrorCode affpoints(TAO_POUNDERS *mfqP, PetscReal *xmin, PetscReal c)
528 {
529   PetscInt         i, j;
530   PetscBLASInt     blasm, blasj, blask, blasn, ione = 1, info;
531   PetscBLASInt     blasnpmax, blasmaxmn;
532   PetscReal        proj, normd;
533   const PetscReal *x;
534 
535   PetscFunctionBegin;
536   PetscCall(PetscBLASIntCast(mfqP->npmax, &blasnpmax));
537   PetscCall(PetscBLASIntCast(mfqP->m, &blasm));
538   PetscCall(PetscBLASIntCast(mfqP->n, &blasn));
539   for (i = mfqP->nHist - 1; i >= 0; i--) {
540     PetscCall(VecGetArrayRead(mfqP->Xhist[i], &x));
541     for (j = 0; j < mfqP->n; j++) mfqP->work[j] = (x[j] - xmin[j]) / mfqP->delta;
542     PetscCall(VecRestoreArrayRead(mfqP->Xhist[i], &x));
543     PetscCallBLAS("BLAScopy", BLAScopy_(&blasn, mfqP->work, &ione, mfqP->work2, &ione));
544     PetscCallBLAS("BLASnrm2", normd = BLASnrm2_(&blasn, mfqP->work, &ione));
545     if (normd <= c) {
546       PetscCall(PetscBLASIntCast(PetscMax(mfqP->n - mfqP->nmodelpoints, 0), &blasj));
547       if (!mfqP->q_is_I) {
548         /* project D onto null */
549         PetscCall(PetscBLASIntCast(mfqP->nmodelpoints, &blask));
550         PetscCallBLAS("LAPACKormqr", LAPACKormqr_("R", "N", &ione, &blasn, &blask, mfqP->Q, &blasnpmax, mfqP->tau, mfqP->work2, &ione, mfqP->mwork, &blasm, &info));
551         PetscCheck(info >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "ormqr returned value %" PetscBLASInt_FMT, info);
552       }
553       PetscCallBLAS("BLASnrm2", proj = BLASnrm2_(&blasj, &mfqP->work2[mfqP->nmodelpoints], &ione));
554 
555       if (proj >= mfqP->theta1) { /* add this index to model */
556         mfqP->model_indices[mfqP->nmodelpoints] = i;
557         mfqP->nmodelpoints++;
558         PetscCallBLAS("BLAScopy", BLAScopy_(&blasn, mfqP->work, &ione, &mfqP->Q_tmp[mfqP->npmax * (mfqP->nmodelpoints - 1)], &ione));
559         PetscCall(PetscBLASIntCast(mfqP->npmax * (mfqP->nmodelpoints), &blask));
560         PetscCallBLAS("BLAScopy", BLAScopy_(&blask, mfqP->Q_tmp, &ione, mfqP->Q, &ione));
561         PetscCall(PetscBLASIntCast(mfqP->nmodelpoints, &blask));
562         PetscCall(PetscBLASIntCast(PetscMax(mfqP->m, mfqP->n), &blasmaxmn));
563         PetscCallBLAS("LAPACKgeqrf", LAPACKgeqrf_(&blasn, &blask, mfqP->Q, &blasnpmax, mfqP->tau, mfqP->mwork, &blasmaxmn, &info));
564         PetscCheck(info >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "geqrf returned value %" PetscBLASInt_FMT, info);
565         mfqP->q_is_I = 0;
566       }
567       if (mfqP->nmodelpoints == mfqP->n) break;
568     }
569   }
570   PetscFunctionReturn(PETSC_SUCCESS);
571 }
572 
TaoSolve_POUNDERS(Tao tao)573 static PetscErrorCode TaoSolve_POUNDERS(Tao tao)
574 {
575   TAO_POUNDERS    *mfqP = (TAO_POUNDERS *)tao->data;
576   PetscInt         i, ii, j, k, l;
577   PetscReal        step = 1.0;
578   PetscInt         low, high;
579   PetscReal        minnorm;
580   PetscReal       *x, *f;
581   const PetscReal *xmint, *fmin;
582   PetscReal        deltaold;
583   PetscReal        gnorm;
584   PetscBLASInt     info, ione = 1, iblas;
585   PetscBool        valid, same;
586   PetscReal        mdec, rho, normxsp;
587   PetscReal        one = 1.0, zero = 0.0, ratio;
588   PetscBLASInt     blasm, blasn, blasncopy, blasnpmax;
589   static PetscBool set = PETSC_FALSE;
590 
591   /* n = # of parameters
592      m = dimension (components) of function  */
593   PetscFunctionBegin;
594   PetscCall(PetscCitationsRegister("@article{UNEDF0,\n"
595                                    "title = {Nuclear energy density optimization},\n"
596                                    "author = {Kortelainen, M.  and Lesinski, T.  and Mor\'e, J.  and Nazarewicz, W.\n"
597                                    "          and Sarich, J.  and Schunck, N.  and Stoitsov, M. V. and Wild, S. },\n"
598                                    "journal = {Phys. Rev. C},\n"
599                                    "volume = {82},\n"
600                                    "number = {2},\n"
601                                    "pages = {024313},\n"
602                                    "numpages = {18},\n"
603                                    "year = {2010},\n"
604                                    "month = {Aug},\n"
605                                    "doi = {10.1103/PhysRevC.82.024313}\n}\n",
606                                    &set));
607   tao->niter = 0;
608   if (tao->XL && tao->XU) {
609     /* Check x0 <= XU */
610     PetscReal val;
611 
612     PetscCall(VecCopy(tao->solution, mfqP->Xhist[0]));
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 > upper bound");
616 
617     /* Check x0 >= xl */
618     PetscCall(VecCopy(tao->XL, mfqP->Xhist[0]));
619     PetscCall(VecAXPY(mfqP->Xhist[0], -1.0, tao->solution));
620     PetscCall(VecMax(mfqP->Xhist[0], NULL, &val));
621     PetscCheck(val <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "X0 < lower bound");
622 
623     /* Check x0 + delta < XU  -- should be able to get around this eventually */
624 
625     PetscCall(VecSet(mfqP->Xhist[0], mfqP->delta));
626     PetscCall(VecAXPY(mfqP->Xhist[0], 1.0, tao->solution));
627     PetscCall(VecAXPY(mfqP->Xhist[0], -1.0, tao->XU));
628     PetscCall(VecMax(mfqP->Xhist[0], NULL, &val));
629     PetscCheck(val <= 1e-10, PetscObjectComm((PetscObject)tao), PETSC_ERR_ARG_OUTOFRANGE, "X0 + delta > upper bound");
630   }
631 
632   PetscCall(PetscBLASIntCast(mfqP->m, &blasm));
633   PetscCall(PetscBLASIntCast(mfqP->n, &blasn));
634   PetscCall(PetscBLASIntCast(mfqP->npmax, &blasnpmax));
635   for (i = 0; i < mfqP->n * mfqP->n * mfqP->m; ++i) mfqP->H[i] = 0;
636 
637   PetscCall(VecCopy(tao->solution, mfqP->Xhist[0]));
638 
639   /* This provides enough information to approximate the gradient of the objective */
640   /* using a forward difference scheme. */
641 
642   PetscCall(PetscInfo(tao, "Initialize simplex; delta = %10.9e\n", (double)mfqP->delta));
643   PetscCall(pounders_feval(tao, mfqP->Xhist[0], mfqP->Fhist[0], &mfqP->Fres[0]));
644   mfqP->minindex = 0;
645   minnorm        = mfqP->Fres[0];
646 
647   PetscCall(VecGetOwnershipRange(mfqP->Xhist[0], &low, &high));
648   for (i = 1; i < mfqP->n + 1; ++i) {
649     PetscCall(VecCopy(mfqP->Xhist[0], mfqP->Xhist[i]));
650 
651     if (i - 1 >= low && i - 1 < high) {
652       PetscCall(VecGetArray(mfqP->Xhist[i], &x));
653       x[i - 1 - low] += mfqP->delta;
654       PetscCall(VecRestoreArray(mfqP->Xhist[i], &x));
655     }
656     CHKMEMQ;
657     PetscCall(pounders_feval(tao, mfqP->Xhist[i], mfqP->Fhist[i], &mfqP->Fres[i]));
658     if (mfqP->Fres[i] < minnorm) {
659       mfqP->minindex = i;
660       minnorm        = mfqP->Fres[i];
661     }
662   }
663   PetscCall(VecCopy(mfqP->Xhist[mfqP->minindex], tao->solution));
664   PetscCall(VecCopy(mfqP->Fhist[mfqP->minindex], tao->ls_res));
665   PetscCall(PetscInfo(tao, "Finalize simplex; minnorm = %10.9e\n", (double)minnorm));
666 
667   /* Gather mpi vecs to one big local vec */
668 
669   /* Begin serial code */
670 
671   /* Disp[i] = Xi-xmin, i=1,..,mfqP->minindex-1,mfqP->minindex+1,..,n */
672   /* Fdiff[i] = (Fi-Fmin)', i=1,..,mfqP->minindex-1,mfqP->minindex+1,..,n */
673   /* (Column oriented for blas calls) */
674   ii = 0;
675 
676   PetscCall(PetscInfo(tao, "Build matrix: %d\n", mfqP->size));
677   if (1 == mfqP->size) {
678     PetscCall(VecGetArrayRead(mfqP->Xhist[mfqP->minindex], &xmint));
679     for (i = 0; i < mfqP->n; i++) mfqP->xmin[i] = xmint[i];
680     PetscCall(VecRestoreArrayRead(mfqP->Xhist[mfqP->minindex], &xmint));
681     PetscCall(VecGetArrayRead(mfqP->Fhist[mfqP->minindex], &fmin));
682     for (i = 0; i < mfqP->n + 1; i++) {
683       if (i == mfqP->minindex) continue;
684 
685       PetscCall(VecGetArray(mfqP->Xhist[i], &x));
686       for (j = 0; j < mfqP->n; j++) mfqP->Disp[ii + mfqP->npmax * j] = (x[j] - mfqP->xmin[j]) / mfqP->delta;
687       PetscCall(VecRestoreArray(mfqP->Xhist[i], &x));
688 
689       PetscCall(VecGetArray(mfqP->Fhist[i], &f));
690       for (j = 0; j < mfqP->m; j++) mfqP->Fdiff[ii + mfqP->n * j] = f[j] - fmin[j];
691       PetscCall(VecRestoreArray(mfqP->Fhist[i], &f));
692 
693       mfqP->model_indices[ii++] = i;
694     }
695     for (j = 0; j < mfqP->m; j++) mfqP->C[j] = fmin[j];
696     PetscCall(VecRestoreArrayRead(mfqP->Fhist[mfqP->minindex], &fmin));
697   } else {
698     PetscCall(VecSet(mfqP->localxmin, 0));
699     PetscCall(VecScatterBegin(mfqP->scatterx, mfqP->Xhist[mfqP->minindex], mfqP->localxmin, INSERT_VALUES, SCATTER_FORWARD));
700     PetscCall(VecScatterEnd(mfqP->scatterx, mfqP->Xhist[mfqP->minindex], mfqP->localxmin, INSERT_VALUES, SCATTER_FORWARD));
701 
702     PetscCall(VecGetArrayRead(mfqP->localxmin, &xmint));
703     for (i = 0; i < mfqP->n; i++) mfqP->xmin[i] = xmint[i];
704     PetscCall(VecRestoreArrayRead(mfqP->localxmin, &xmint));
705 
706     PetscCall(VecScatterBegin(mfqP->scatterf, mfqP->Fhist[mfqP->minindex], mfqP->localfmin, INSERT_VALUES, SCATTER_FORWARD));
707     PetscCall(VecScatterEnd(mfqP->scatterf, mfqP->Fhist[mfqP->minindex], mfqP->localfmin, INSERT_VALUES, SCATTER_FORWARD));
708     PetscCall(VecGetArrayRead(mfqP->localfmin, &fmin));
709     for (i = 0; i < mfqP->n + 1; i++) {
710       if (i == mfqP->minindex) continue;
711 
712       PetscCall(VecScatterBegin(mfqP->scatterx, mfqP->Xhist[ii], mfqP->localx, INSERT_VALUES, SCATTER_FORWARD));
713       PetscCall(VecScatterEnd(mfqP->scatterx, mfqP->Xhist[ii], mfqP->localx, INSERT_VALUES, SCATTER_FORWARD));
714       PetscCall(VecGetArray(mfqP->localx, &x));
715       for (j = 0; j < mfqP->n; j++) mfqP->Disp[ii + mfqP->npmax * j] = (x[j] - mfqP->xmin[j]) / mfqP->delta;
716       PetscCall(VecRestoreArray(mfqP->localx, &x));
717 
718       PetscCall(VecScatterBegin(mfqP->scatterf, mfqP->Fhist[ii], mfqP->localf, INSERT_VALUES, SCATTER_FORWARD));
719       PetscCall(VecScatterEnd(mfqP->scatterf, mfqP->Fhist[ii], mfqP->localf, INSERT_VALUES, SCATTER_FORWARD));
720       PetscCall(VecGetArray(mfqP->localf, &f));
721       for (j = 0; j < mfqP->m; j++) mfqP->Fdiff[ii + mfqP->n * j] = f[j] - fmin[j];
722       PetscCall(VecRestoreArray(mfqP->localf, &f));
723 
724       mfqP->model_indices[ii++] = i;
725     }
726     for (j = 0; j < mfqP->m; j++) mfqP->C[j] = fmin[j];
727     PetscCall(VecRestoreArrayRead(mfqP->localfmin, &fmin));
728   }
729 
730   /* Determine the initial quadratic models */
731   /* G = D(ModelIn,:) \ (F(ModelIn,1:m)-repmat(F(xkin,1:m),n,1)); */
732   /* D (nxn) Fdiff (nxm)  => G (nxm) */
733   blasncopy = blasn;
734   PetscCallBLAS("LAPACKgesv", LAPACKgesv_(&blasn, &blasm, mfqP->Disp, &blasnpmax, mfqP->iwork, mfqP->Fdiff, &blasncopy, &info));
735   PetscCall(PetscInfo(tao, "Linear solve return: %" PetscBLASInt_FMT "\n", info));
736 
737   PetscCall(pounders_update_res(tao));
738 
739   valid = PETSC_TRUE;
740 
741   PetscCall(VecSetValues(tao->gradient, mfqP->n, mfqP->indices, mfqP->Gres, INSERT_VALUES));
742   PetscCall(VecAssemblyBegin(tao->gradient));
743   PetscCall(VecAssemblyEnd(tao->gradient));
744   PetscCall(VecNorm(tao->gradient, NORM_2, &gnorm));
745   gnorm *= mfqP->delta;
746   PetscCall(VecCopy(mfqP->Xhist[mfqP->minindex], tao->solution));
747 
748   tao->reason = TAO_CONTINUE_ITERATING;
749   PetscCall(TaoLogConvergenceHistory(tao, minnorm, gnorm, 0.0, tao->ksp_its));
750   PetscCall(TaoMonitor(tao, tao->niter, minnorm, gnorm, 0.0, step));
751   PetscUseTypeMethod(tao, convergencetest, tao->cnvP);
752 
753   mfqP->nHist        = mfqP->n + 1;
754   mfqP->nmodelpoints = mfqP->n + 1;
755   PetscCall(PetscInfo(tao, "Initial gradient: %20.19e\n", (double)gnorm));
756 
757   while (tao->reason == TAO_CONTINUE_ITERATING) {
758     PetscReal gnm = 1e-4;
759     /* Call general purpose update function */
760     PetscTryTypeMethod(tao, update, tao->niter, tao->user_update);
761     tao->niter++;
762     /* Solve the subproblem min{Q(s): ||s|| <= 1.0} */
763     PetscCall(gqtwrap(tao, &gnm, &mdec));
764     /* Evaluate the function at the new point */
765 
766     for (i = 0; i < mfqP->n; i++) mfqP->work[i] = mfqP->Xsubproblem[i] * mfqP->delta + mfqP->xmin[i];
767     PetscCall(VecDuplicate(tao->solution, &mfqP->Xhist[mfqP->nHist]));
768     PetscCall(VecDuplicate(tao->ls_res, &mfqP->Fhist[mfqP->nHist]));
769     PetscCall(VecSetValues(mfqP->Xhist[mfqP->nHist], mfqP->n, mfqP->indices, mfqP->work, INSERT_VALUES));
770     PetscCall(VecAssemblyBegin(mfqP->Xhist[mfqP->nHist]));
771     PetscCall(VecAssemblyEnd(mfqP->Xhist[mfqP->nHist]));
772 
773     PetscCall(pounders_feval(tao, mfqP->Xhist[mfqP->nHist], mfqP->Fhist[mfqP->nHist], &mfqP->Fres[mfqP->nHist]));
774 
775     rho = (mfqP->Fres[mfqP->minindex] - mfqP->Fres[mfqP->nHist]) / mdec;
776     mfqP->nHist++;
777 
778     /* Update the center */
779     if ((rho >= mfqP->eta1) || (rho > mfqP->eta0 && valid == PETSC_TRUE)) {
780       /* Update model to reflect new base point */
781       for (i = 0; i < mfqP->n; i++) mfqP->work[i] = (mfqP->work[i] - mfqP->xmin[i]) / mfqP->delta;
782       for (j = 0; j < mfqP->m; j++) {
783         /* C(j) = C(j) + work*G(:,j) + .5*work*H(:,:,j)*work';
784          G(:,j) = G(:,j) + H(:,:,j)*work' */
785         for (k = 0; k < mfqP->n; k++) {
786           mfqP->work2[k] = 0.0;
787           for (l = 0; l < mfqP->n; l++) mfqP->work2[k] += mfqP->H[j + mfqP->m * (k + l * mfqP->n)] * mfqP->work[l];
788         }
789         for (i = 0; i < mfqP->n; i++) {
790           mfqP->C[j] += mfqP->work[i] * (mfqP->Fdiff[i + mfqP->n * j] + 0.5 * mfqP->work2[i]);
791           mfqP->Fdiff[i + mfqP->n * j] += mfqP->work2[i];
792         }
793       }
794       /* Cres += work*Gres + .5*work*Hres*work';
795        Gres += Hres*work'; */
796 
797       PetscCallBLAS("BLASgemv", BLASgemv_("N", &blasn, &blasn, &one, mfqP->Hres, &blasn, mfqP->work, &ione, &zero, mfqP->work2, &ione));
798       for (i = 0; i < mfqP->n; i++) mfqP->Gres[i] += mfqP->work2[i];
799       mfqP->minindex = mfqP->nHist - 1;
800       minnorm        = mfqP->Fres[mfqP->minindex];
801       PetscCall(VecCopy(mfqP->Fhist[mfqP->minindex], tao->ls_res));
802       /* Change current center */
803       PetscCall(VecGetArrayRead(mfqP->Xhist[mfqP->minindex], &xmint));
804       for (i = 0; i < mfqP->n; i++) mfqP->xmin[i] = xmint[i];
805       PetscCall(VecRestoreArrayRead(mfqP->Xhist[mfqP->minindex], &xmint));
806     }
807 
808     /* Evaluate at a model-improving point if necessary */
809     if (valid == PETSC_FALSE) {
810       mfqP->q_is_I       = 1;
811       mfqP->nmodelpoints = 0;
812       PetscCall(affpoints(mfqP, mfqP->xmin, mfqP->c1));
813       if (mfqP->nmodelpoints < mfqP->n) {
814         PetscCall(PetscInfo(tao, "Model not valid -- model-improving\n"));
815         PetscCall(modelimprove(tao, mfqP, 1));
816       }
817     }
818 
819     /* Update the trust region radius */
820     deltaold = mfqP->delta;
821     normxsp  = 0;
822     for (i = 0; i < mfqP->n; i++) normxsp += mfqP->Xsubproblem[i] * mfqP->Xsubproblem[i];
823     normxsp = PetscSqrtReal(normxsp);
824     if (rho >= mfqP->eta1 && normxsp > 0.5 * mfqP->delta) {
825       mfqP->delta = PetscMin(mfqP->delta * mfqP->gamma1, mfqP->deltamax);
826     } else if (valid == PETSC_TRUE) {
827       mfqP->delta = PetscMax(mfqP->delta * mfqP->gamma0, mfqP->deltamin);
828     }
829 
830     /* Compute the next interpolation set */
831     mfqP->q_is_I       = 1;
832     mfqP->nmodelpoints = 0;
833     PetscCall(PetscInfo(tao, "Affine Points: xmin = %20.19e, c1 = %20.19e\n", (double)*mfqP->xmin, (double)mfqP->c1));
834     PetscCall(affpoints(mfqP, mfqP->xmin, mfqP->c1));
835     if (mfqP->nmodelpoints == mfqP->n) {
836       valid = PETSC_TRUE;
837     } else {
838       valid = PETSC_FALSE;
839       PetscCall(PetscInfo(tao, "Affine Points: xmin = %20.19e, c2 = %20.19e\n", (double)*mfqP->xmin, (double)mfqP->c2));
840       PetscCall(affpoints(mfqP, mfqP->xmin, mfqP->c2));
841       if (mfqP->n > mfqP->nmodelpoints) {
842         PetscCall(PetscInfo(tao, "Model not valid -- adding geometry points\n"));
843         PetscCall(modelimprove(tao, mfqP, mfqP->n - mfqP->nmodelpoints));
844       }
845     }
846     for (i = mfqP->nmodelpoints; i > 0; i--) mfqP->model_indices[i] = mfqP->model_indices[i - 1];
847     mfqP->nmodelpoints++;
848     mfqP->model_indices[0] = mfqP->minindex;
849     PetscCall(morepoints(mfqP));
850     for (i = 0; i < mfqP->nmodelpoints; i++) {
851       PetscCall(VecGetArray(mfqP->Xhist[mfqP->model_indices[i]], &x));
852       for (j = 0; j < mfqP->n; j++) mfqP->Disp[i + mfqP->npmax * j] = (x[j] - mfqP->xmin[j]) / deltaold;
853       PetscCall(VecRestoreArray(mfqP->Xhist[mfqP->model_indices[i]], &x));
854       PetscCall(VecGetArray(mfqP->Fhist[mfqP->model_indices[i]], &f));
855       for (j = 0; j < mfqP->m; j++) {
856         for (k = 0; k < mfqP->n; k++) {
857           mfqP->work[k] = 0.0;
858           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];
859         }
860         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]);
861       }
862       PetscCall(VecRestoreArray(mfqP->Fhist[mfqP->model_indices[i]], &f));
863     }
864 
865     /* Update the quadratic model */
866     PetscCall(PetscInfo(tao, "Get Quad, size: %" PetscInt_FMT ", points: %" PetscInt_FMT "\n", mfqP->n, mfqP->nmodelpoints));
867     PetscCall(getquadpounders(mfqP));
868     PetscCall(VecGetArrayRead(mfqP->Fhist[mfqP->minindex], &fmin));
869     PetscCallBLAS("BLAScopy", BLAScopy_(&blasm, fmin, &ione, mfqP->C, &ione));
870     /* G = G*(delta/deltaold) + Gdel */
871     ratio = mfqP->delta / deltaold;
872     iblas = blasm * blasn;
873     PetscCallBLAS("BLASscal", BLASscal_(&iblas, &ratio, mfqP->Fdiff, &ione));
874     PetscCallBLAS("BLASaxpy", BLASaxpy_(&iblas, &one, mfqP->Gdel, &ione, mfqP->Fdiff, &ione));
875     /* H = H*(delta/deltaold)^2 + Hdel */
876     iblas = blasm * blasn * blasn;
877     ratio *= ratio;
878     PetscCallBLAS("BLASscal", BLASscal_(&iblas, &ratio, mfqP->H, &ione));
879     PetscCallBLAS("BLASaxpy", BLASaxpy_(&iblas, &one, mfqP->Hdel, &ione, mfqP->H, &ione));
880 
881     /* Get residuals */
882     PetscCall(pounders_update_res(tao));
883 
884     /* Export solution and gradient residual to TAO */
885     PetscCall(VecCopy(mfqP->Xhist[mfqP->minindex], tao->solution));
886     PetscCall(VecSetValues(tao->gradient, mfqP->n, mfqP->indices, mfqP->Gres, INSERT_VALUES));
887     PetscCall(VecAssemblyBegin(tao->gradient));
888     PetscCall(VecAssemblyEnd(tao->gradient));
889     PetscCall(VecNorm(tao->gradient, NORM_2, &gnorm));
890     gnorm *= mfqP->delta;
891     /*  final criticality test */
892     PetscCall(TaoLogConvergenceHistory(tao, minnorm, gnorm, 0.0, tao->ksp_its));
893     PetscCall(TaoMonitor(tao, tao->niter, minnorm, gnorm, 0.0, step));
894     PetscUseTypeMethod(tao, convergencetest, tao->cnvP);
895     /* test for repeated model */
896     if (mfqP->nmodelpoints == mfqP->last_nmodelpoints) {
897       same = PETSC_TRUE;
898     } else {
899       same = PETSC_FALSE;
900     }
901     for (i = 0; i < mfqP->nmodelpoints; i++) {
902       if (same) {
903         if (mfqP->model_indices[i] == mfqP->last_model_indices[i]) {
904           same = PETSC_TRUE;
905         } else {
906           same = PETSC_FALSE;
907         }
908       }
909       mfqP->last_model_indices[i] = mfqP->model_indices[i];
910     }
911     mfqP->last_nmodelpoints = mfqP->nmodelpoints;
912     if (same && mfqP->delta == deltaold) {
913       PetscCall(PetscInfo(tao, "Identical model used in successive iterations\n"));
914       tao->reason = TAO_CONVERGED_STEPTOL;
915     }
916   }
917   PetscFunctionReturn(PETSC_SUCCESS);
918 }
919 
TaoSetUp_POUNDERS(Tao tao)920 static PetscErrorCode TaoSetUp_POUNDERS(Tao tao)
921 {
922   TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
923   PetscInt      i, j;
924   IS            isfloc, isfglob, isxloc, isxglob;
925 
926   PetscFunctionBegin;
927   if (!tao->gradient) PetscCall(VecDuplicate(tao->solution, &tao->gradient));
928   if (!tao->stepdirection) PetscCall(VecDuplicate(tao->solution, &tao->stepdirection));
929   PetscCall(VecGetSize(tao->solution, &mfqP->n));
930   PetscCall(VecGetSize(tao->ls_res, &mfqP->m));
931   mfqP->c1 = PetscSqrtReal((PetscReal)mfqP->n);
932   if (mfqP->npmax == PETSC_CURRENT) mfqP->npmax = 2 * mfqP->n + 1;
933   mfqP->npmax = PetscMin((mfqP->n + 1) * (mfqP->n + 2) / 2, mfqP->npmax);
934   mfqP->npmax = PetscMax(mfqP->npmax, mfqP->n + 2);
935 
936   PetscCall(PetscMalloc1(tao->max_funcs + 100, &mfqP->Xhist));
937   PetscCall(PetscMalloc1(tao->max_funcs + 100, &mfqP->Fhist));
938   for (i = 0; i < mfqP->n + 1; i++) {
939     PetscCall(VecDuplicate(tao->solution, &mfqP->Xhist[i]));
940     PetscCall(VecDuplicate(tao->ls_res, &mfqP->Fhist[i]));
941   }
942   PetscCall(VecDuplicate(tao->solution, &mfqP->workxvec));
943   PetscCall(VecDuplicate(tao->ls_res, &mfqP->workfvec));
944   mfqP->nHist = 0;
945 
946   PetscCall(PetscMalloc1(tao->max_funcs + 100, &mfqP->Fres));
947   PetscCall(PetscMalloc1(mfqP->npmax * mfqP->m, &mfqP->RES));
948   PetscCall(PetscMalloc1(mfqP->n, &mfqP->work));
949   PetscCall(PetscMalloc1(mfqP->n, &mfqP->work2));
950   PetscCall(PetscMalloc1(mfqP->n, &mfqP->work3));
951   PetscCall(PetscMalloc1(PetscMax(mfqP->m, mfqP->n + 1), &mfqP->mwork));
952   PetscCall(PetscMalloc1(mfqP->npmax - mfqP->n - 1, &mfqP->omega));
953   PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2, &mfqP->beta));
954   PetscCall(PetscMalloc1(mfqP->n + 1, &mfqP->alpha));
955 
956   PetscCall(PetscMalloc1(mfqP->n * mfqP->n * mfqP->m, &mfqP->H));
957   PetscCall(PetscMalloc1(mfqP->npmax * mfqP->npmax, &mfqP->Q));
958   PetscCall(PetscMalloc1(mfqP->npmax * mfqP->npmax, &mfqP->Q_tmp));
959   PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2 * (mfqP->npmax), &mfqP->L));
960   PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2 * (mfqP->npmax), &mfqP->L_tmp));
961   PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2 * (mfqP->npmax), &mfqP->L_save));
962   PetscCall(PetscMalloc1(mfqP->n * (mfqP->n + 1) / 2 * (mfqP->npmax), &mfqP->N));
963   PetscCall(PetscMalloc1(mfqP->npmax * (mfqP->n + 1), &mfqP->M));
964   PetscCall(PetscMalloc1(mfqP->npmax * (mfqP->npmax - mfqP->n - 1), &mfqP->Z));
965   PetscCall(PetscMalloc1(mfqP->npmax, &mfqP->tau));
966   PetscCall(PetscMalloc1(mfqP->npmax, &mfqP->tau_tmp));
967   mfqP->nmax = PetscMax(5 * mfqP->npmax, mfqP->n * (mfqP->n + 1) / 2);
968   PetscCall(PetscMalloc1(mfqP->nmax, &mfqP->npmaxwork));
969   PetscCall(PetscMalloc1(mfqP->nmax, &mfqP->npmaxiwork));
970   PetscCall(PetscMalloc1(mfqP->n, &mfqP->xmin));
971   PetscCall(PetscMalloc1(mfqP->m, &mfqP->C));
972   PetscCall(PetscMalloc1(mfqP->m * mfqP->n, &mfqP->Fdiff));
973   PetscCall(PetscMalloc1(mfqP->npmax * mfqP->n, &mfqP->Disp));
974   PetscCall(PetscMalloc1(mfqP->n, &mfqP->Gres));
975   PetscCall(PetscMalloc1(mfqP->n * mfqP->n, &mfqP->Hres));
976   PetscCall(PetscMalloc1(mfqP->n * mfqP->n, &mfqP->Gpoints));
977   PetscCall(PetscMalloc1(mfqP->npmax, &mfqP->model_indices));
978   PetscCall(PetscMalloc1(mfqP->npmax, &mfqP->last_model_indices));
979   PetscCall(PetscMalloc1(mfqP->n, &mfqP->Xsubproblem));
980   PetscCall(PetscMalloc1(mfqP->m * mfqP->n, &mfqP->Gdel));
981   PetscCall(PetscMalloc1(mfqP->n * mfqP->n * mfqP->m, &mfqP->Hdel));
982   PetscCall(PetscMalloc1(PetscMax(mfqP->m, mfqP->n), &mfqP->indices));
983   PetscCall(PetscMalloc1(mfqP->n, &mfqP->iwork));
984   PetscCall(PetscMalloc1(mfqP->m * mfqP->m, &mfqP->w));
985   for (i = 0; i < mfqP->m; i++) {
986     for (j = 0; j < mfqP->m; j++) {
987       if (i == j) {
988         mfqP->w[i + mfqP->m * j] = 1.0;
989       } else {
990         mfqP->w[i + mfqP->m * j] = 0.0;
991       }
992     }
993   }
994   for (i = 0; i < PetscMax(mfqP->m, mfqP->n); i++) mfqP->indices[i] = i;
995   PetscCallMPI(MPI_Comm_size(((PetscObject)tao)->comm, &mfqP->size));
996   if (mfqP->size > 1) {
997     PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->n, &mfqP->localx));
998     PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->n, &mfqP->localxmin));
999     PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->m, &mfqP->localf));
1000     PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->m, &mfqP->localfmin));
1001     PetscCall(ISCreateStride(MPI_COMM_SELF, mfqP->n, 0, 1, &isxloc));
1002     PetscCall(ISCreateStride(MPI_COMM_SELF, mfqP->n, 0, 1, &isxglob));
1003     PetscCall(ISCreateStride(MPI_COMM_SELF, mfqP->m, 0, 1, &isfloc));
1004     PetscCall(ISCreateStride(MPI_COMM_SELF, mfqP->m, 0, 1, &isfglob));
1005 
1006     PetscCall(VecScatterCreate(tao->solution, isxglob, mfqP->localx, isxloc, &mfqP->scatterx));
1007     PetscCall(VecScatterCreate(tao->ls_res, isfglob, mfqP->localf, isfloc, &mfqP->scatterf));
1008 
1009     PetscCall(ISDestroy(&isxloc));
1010     PetscCall(ISDestroy(&isxglob));
1011     PetscCall(ISDestroy(&isfloc));
1012     PetscCall(ISDestroy(&isfglob));
1013   }
1014 
1015   if (!mfqP->usegqt) {
1016     KSP ksp;
1017     PC  pc;
1018     PetscCall(VecCreateSeqWithArray(PETSC_COMM_SELF, mfqP->n, mfqP->n, mfqP->Xsubproblem, &mfqP->subx));
1019     PetscCall(VecCreateSeq(PETSC_COMM_SELF, mfqP->n, &mfqP->subxl));
1020     PetscCall(VecDuplicate(mfqP->subxl, &mfqP->subb));
1021     PetscCall(VecDuplicate(mfqP->subxl, &mfqP->subxu));
1022     PetscCall(VecDuplicate(mfqP->subxl, &mfqP->subpdel));
1023     PetscCall(VecDuplicate(mfqP->subxl, &mfqP->subndel));
1024     PetscCall(TaoCreate(PETSC_COMM_SELF, &mfqP->subtao));
1025     PetscCall(PetscObjectIncrementTabLevel((PetscObject)mfqP->subtao, (PetscObject)tao, 1));
1026     PetscCall(TaoSetType(mfqP->subtao, TAOBNTR));
1027     PetscCall(TaoSetOptionsPrefix(mfqP->subtao, "pounders_subsolver_"));
1028     PetscCall(TaoSetSolution(mfqP->subtao, mfqP->subx));
1029     PetscCall(TaoSetObjectiveAndGradient(mfqP->subtao, NULL, pounders_fg, (void *)mfqP));
1030     PetscCall(TaoSetMaximumIterations(mfqP->subtao, mfqP->gqt_maxits));
1031     PetscCall(TaoSetFromOptions(mfqP->subtao));
1032     PetscCall(TaoGetKSP(mfqP->subtao, &ksp));
1033     if (ksp) {
1034       PetscCall(KSPGetPC(ksp, &pc));
1035       PetscCall(PCSetType(pc, PCNONE));
1036     }
1037     PetscCall(TaoSetVariableBounds(mfqP->subtao, mfqP->subxl, mfqP->subxu));
1038     PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, mfqP->n, mfqP->n, mfqP->Hres, &mfqP->subH));
1039     PetscCall(TaoSetHessian(mfqP->subtao, mfqP->subH, mfqP->subH, pounders_h, (void *)mfqP));
1040   }
1041   PetscFunctionReturn(PETSC_SUCCESS);
1042 }
1043 
TaoDestroy_POUNDERS(Tao tao)1044 static PetscErrorCode TaoDestroy_POUNDERS(Tao tao)
1045 {
1046   TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
1047   PetscInt      i;
1048 
1049   PetscFunctionBegin;
1050   if (!mfqP->usegqt) {
1051     PetscCall(TaoDestroy(&mfqP->subtao));
1052     PetscCall(VecDestroy(&mfqP->subx));
1053     PetscCall(VecDestroy(&mfqP->subxl));
1054     PetscCall(VecDestroy(&mfqP->subxu));
1055     PetscCall(VecDestroy(&mfqP->subb));
1056     PetscCall(VecDestroy(&mfqP->subpdel));
1057     PetscCall(VecDestroy(&mfqP->subndel));
1058     PetscCall(MatDestroy(&mfqP->subH));
1059   }
1060   PetscCall(PetscFree(mfqP->Fres));
1061   PetscCall(PetscFree(mfqP->RES));
1062   PetscCall(PetscFree(mfqP->work));
1063   PetscCall(PetscFree(mfqP->work2));
1064   PetscCall(PetscFree(mfqP->work3));
1065   PetscCall(PetscFree(mfqP->mwork));
1066   PetscCall(PetscFree(mfqP->omega));
1067   PetscCall(PetscFree(mfqP->beta));
1068   PetscCall(PetscFree(mfqP->alpha));
1069   PetscCall(PetscFree(mfqP->H));
1070   PetscCall(PetscFree(mfqP->Q));
1071   PetscCall(PetscFree(mfqP->Q_tmp));
1072   PetscCall(PetscFree(mfqP->L));
1073   PetscCall(PetscFree(mfqP->L_tmp));
1074   PetscCall(PetscFree(mfqP->L_save));
1075   PetscCall(PetscFree(mfqP->N));
1076   PetscCall(PetscFree(mfqP->M));
1077   PetscCall(PetscFree(mfqP->Z));
1078   PetscCall(PetscFree(mfqP->tau));
1079   PetscCall(PetscFree(mfqP->tau_tmp));
1080   PetscCall(PetscFree(mfqP->npmaxwork));
1081   PetscCall(PetscFree(mfqP->npmaxiwork));
1082   PetscCall(PetscFree(mfqP->xmin));
1083   PetscCall(PetscFree(mfqP->C));
1084   PetscCall(PetscFree(mfqP->Fdiff));
1085   PetscCall(PetscFree(mfqP->Disp));
1086   PetscCall(PetscFree(mfqP->Gres));
1087   PetscCall(PetscFree(mfqP->Hres));
1088   PetscCall(PetscFree(mfqP->Gpoints));
1089   PetscCall(PetscFree(mfqP->model_indices));
1090   PetscCall(PetscFree(mfqP->last_model_indices));
1091   PetscCall(PetscFree(mfqP->Xsubproblem));
1092   PetscCall(PetscFree(mfqP->Gdel));
1093   PetscCall(PetscFree(mfqP->Hdel));
1094   PetscCall(PetscFree(mfqP->indices));
1095   PetscCall(PetscFree(mfqP->iwork));
1096   PetscCall(PetscFree(mfqP->w));
1097   for (i = 0; i < mfqP->nHist; i++) {
1098     PetscCall(VecDestroy(&mfqP->Xhist[i]));
1099     PetscCall(VecDestroy(&mfqP->Fhist[i]));
1100   }
1101   PetscCall(VecDestroy(&mfqP->workxvec));
1102   PetscCall(VecDestroy(&mfqP->workfvec));
1103   PetscCall(PetscFree(mfqP->Xhist));
1104   PetscCall(PetscFree(mfqP->Fhist));
1105 
1106   if (mfqP->size > 1) {
1107     PetscCall(VecDestroy(&mfqP->localx));
1108     PetscCall(VecDestroy(&mfqP->localxmin));
1109     PetscCall(VecDestroy(&mfqP->localf));
1110     PetscCall(VecDestroy(&mfqP->localfmin));
1111   }
1112   PetscCall(PetscFree(tao->data));
1113   PetscFunctionReturn(PETSC_SUCCESS);
1114 }
1115 
TaoSetFromOptions_POUNDERS(Tao tao,PetscOptionItems PetscOptionsObject)1116 static PetscErrorCode TaoSetFromOptions_POUNDERS(Tao tao, PetscOptionItems PetscOptionsObject)
1117 {
1118   TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
1119 
1120   PetscFunctionBegin;
1121   PetscOptionsHeadBegin(PetscOptionsObject, "POUNDERS method for least-squares optimization");
1122   PetscCall(PetscOptionsReal("-tao_pounders_delta", "initial delta", "", mfqP->delta, &mfqP->delta0, NULL));
1123   mfqP->delta = mfqP->delta0;
1124   PetscCall(PetscOptionsInt("-tao_pounders_npmax", "max number of points in model", "", mfqP->npmax, &mfqP->npmax, NULL));
1125   PetscCall(PetscOptionsBool("-tao_pounders_gqt", "use gqt algorithm for subproblem", "", mfqP->usegqt, &mfqP->usegqt, NULL));
1126   PetscOptionsHeadEnd();
1127   PetscFunctionReturn(PETSC_SUCCESS);
1128 }
1129 
TaoView_POUNDERS(Tao tao,PetscViewer viewer)1130 static PetscErrorCode TaoView_POUNDERS(Tao tao, PetscViewer viewer)
1131 {
1132   TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
1133   PetscBool     isascii;
1134 
1135   PetscFunctionBegin;
1136   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1137   if (isascii) {
1138     PetscCall(PetscViewerASCIIPrintf(viewer, "initial delta: %g\n", (double)mfqP->delta0));
1139     PetscCall(PetscViewerASCIIPrintf(viewer, "final delta: %g\n", (double)mfqP->delta));
1140     PetscCall(PetscViewerASCIIPrintf(viewer, "model points: %" PetscInt_FMT "\n", mfqP->nmodelpoints));
1141     if (mfqP->usegqt) {
1142       PetscCall(PetscViewerASCIIPrintf(viewer, "subproblem solver: gqt\n"));
1143     } else {
1144       PetscCall(TaoView(mfqP->subtao, viewer));
1145     }
1146   }
1147   PetscFunctionReturn(PETSC_SUCCESS);
1148 }
1149 /*MC
1150   TAOPOUNDERS - POUNDERS derivate-free model-based algorithm for nonlinear least squares
1151 
1152   Options Database Keys:
1153 + -tao_pounders_delta - initial step length
1154 . -tao_pounders_npmax - maximum number of points in model
1155 - -tao_pounders_gqt - use gqt algorithm for subproblem instead of TRON
1156 
1157   Level: beginner
1158 
1159 M*/
1160 
TaoCreate_POUNDERS(Tao tao)1161 PETSC_EXTERN PetscErrorCode TaoCreate_POUNDERS(Tao tao)
1162 {
1163   TAO_POUNDERS *mfqP = (TAO_POUNDERS *)tao->data;
1164 
1165   PetscFunctionBegin;
1166   tao->ops->setup          = TaoSetUp_POUNDERS;
1167   tao->ops->solve          = TaoSolve_POUNDERS;
1168   tao->ops->view           = TaoView_POUNDERS;
1169   tao->ops->setfromoptions = TaoSetFromOptions_POUNDERS;
1170   tao->ops->destroy        = TaoDestroy_POUNDERS;
1171 
1172   PetscCall(PetscNew(&mfqP));
1173   tao->data = (void *)mfqP;
1174 
1175   /* Override default settings (unless already changed) */
1176   PetscCall(TaoParametersInitialize(tao));
1177   PetscObjectParameterSetDefault(tao, max_it, 2000);
1178   PetscObjectParameterSetDefault(tao, max_funcs, 4000);
1179 
1180   mfqP->npmax      = PETSC_CURRENT;
1181   mfqP->delta0     = 0.1;
1182   mfqP->delta      = 0.1;
1183   mfqP->deltamax   = 1e3;
1184   mfqP->deltamin   = 1e-6;
1185   mfqP->c2         = 10.0;
1186   mfqP->theta1     = 1e-5;
1187   mfqP->theta2     = 1e-4;
1188   mfqP->gamma0     = 0.5;
1189   mfqP->gamma1     = 2.0;
1190   mfqP->eta0       = 0.0;
1191   mfqP->eta1       = 0.1;
1192   mfqP->usegqt     = PETSC_FALSE;
1193   mfqP->gqt_rtol   = 0.001;
1194   mfqP->gqt_maxits = 50;
1195   mfqP->workxvec   = NULL;
1196   PetscFunctionReturn(PETSC_SUCCESS);
1197 }
1198