xref: /petsc/src/ksp/ksp/impls/gmres/agmres/agmresdeflation.c (revision be37439ebbbdb2f81c3420c175a94aa72e59929c)
1 /*
2   This file computes data for the deflated restarting in the Newton-basis GMRES.
3   At each restart (or at each detected stagnation in the adaptive strategy), a basis of an
4   (approximated)invariant subspace corresponding to the smallest eigenvalues is extracted from the Krylov subspace.
5   It is then used to augment the Newton basis.
6 */
7 
8 #include <../src/ksp/ksp/impls/gmres/agmres/agmresimpl.h>
9 
10 /* Quicksort algorithm to sort  the eigenvalues in increasing orders
11    val_r - real part of eigenvalues, unchanged on exit.
12    val_i - Imaginary part of eigenvalues unchanged on exit.
13    size - Number of eigenvalues (with complex conjugates)
14    perm - contains on exit the permutation vector to reorder the vectors val_r and val_i.
15 */
16 #define DEPTH 500
KSPAGMRESQuickSort(PetscScalar * val_r,PetscScalar * val_i,PetscInt size,PetscInt * perm)17 static PetscErrorCode KSPAGMRESQuickSort(PetscScalar *val_r, PetscScalar *val_i, PetscInt size, PetscInt *perm)
18 {
19   PetscInt    deb[DEPTH], fin[DEPTH];
20   PetscInt    ipivot;
21   PetscScalar pivot_r, pivot_i;
22   PetscInt    i, L, R, j;
23   PetscScalar abs_pivot;
24   PetscScalar abs_val;
25 
26   PetscFunctionBegin;
27   /* initialize perm vector */
28   for (j = 0; j < size; j++) perm[j] = j;
29 
30   deb[0] = 0;
31   fin[0] = size;
32   i      = 0;
33   while (i >= 0) {
34     L = deb[i];
35     R = fin[i] - 1;
36     if (L < R) {
37       pivot_r   = val_r[L];
38       pivot_i   = val_i[L];
39       abs_pivot = PetscSqrtReal(pivot_r * pivot_r + pivot_i * pivot_i);
40       ipivot    = perm[L];
41       PetscCheck(i != DEPTH - 1, PETSC_COMM_SELF, PETSC_ERR_MEM, "Could cause stack overflow: Try to increase the value of DEPTH ");
42       while (L < R) {
43         abs_val = PetscSqrtReal(val_r[R] * val_r[R] + val_i[R] * val_i[R]);
44         while (abs_val >= abs_pivot && L < R) {
45           R--;
46           abs_val = PetscSqrtReal(val_r[R] * val_r[R] + val_i[R] * val_i[R]);
47         }
48         if (L < R) {
49           val_r[L] = val_r[R];
50           val_i[L] = val_i[R];
51           perm[L]  = perm[R];
52           L += 1;
53         }
54         abs_val = PetscSqrtReal(val_r[L] * val_r[L] + val_i[L] * val_i[L]);
55         while (abs_val <= abs_pivot && L < R) {
56           L++;
57           abs_val = PetscSqrtReal(val_r[L] * val_r[L] + val_i[L] * val_i[L]);
58         }
59         if (L < R) {
60           val_r[R] = val_r[L];
61           val_i[R] = val_i[L];
62           perm[R]  = perm[L];
63           R -= 1;
64         }
65       }
66       val_r[L]   = pivot_r;
67       val_i[L]   = pivot_i;
68       perm[L]    = ipivot;
69       deb[i + 1] = L + 1;
70       fin[i + 1] = fin[i];
71       fin[i]     = L;
72       i += 1;
73       PetscCheck(i != DEPTH - 1, PETSC_COMM_SELF, PETSC_ERR_MEM, "Could cause stack overflow: Try to increase the value of DEPTH ");
74     } else i--;
75   }
76   PetscFunctionReturn(PETSC_SUCCESS);
77 }
78 
79 /*
80  Compute the Schur vectors from the generalized eigenvalue problem A.u =\lambda.B.u
81  KspSize -  rank of the matrices A and B, size of the current Krylov basis
82  A - Left matrix
83  B - Right matrix
84  ldA - first dimension of A as declared  in the calling program
85  ldB - first dimension of B as declared  in the calling program
86  IsReduced - specifies if the matrices are already in the reduced form,
87  i.e A is a Hessenberg matrix and B is upper triangular.
88  Sr - on exit, the extracted Schur vectors corresponding
89  the smallest eigenvalues (with complex conjugates)
90  CurNeig - Number of extracted eigenvalues
91 */
KSPAGMRESSchurForm(KSP ksp,PetscBLASInt KspSize,PetscScalar * A,PetscBLASInt ldA,PetscScalar * B,PetscBLASInt ldB,PetscBool IsReduced,PetscScalar * Sr,PetscInt * CurNeig)92 static PetscErrorCode KSPAGMRESSchurForm(KSP ksp, PetscBLASInt KspSize, PetscScalar *A, PetscBLASInt ldA, PetscScalar *B, PetscBLASInt ldB, PetscBool IsReduced, PetscScalar *Sr, PetscInt *CurNeig)
93 {
94   KSP_AGMRES   *agmres = (KSP_AGMRES *)ksp->data;
95   PetscInt      max_k  = agmres->max_k;
96   PetscBLASInt  r;
97   PetscInt      neig   = agmres->neig;
98   PetscScalar  *wr     = agmres->wr;
99   PetscScalar  *wi     = agmres->wi;
100   PetscScalar  *beta   = agmres->beta;
101   PetscScalar  *Q      = agmres->Q;
102   PetscScalar  *Z      = agmres->Z;
103   PetscScalar  *work   = agmres->work;
104   PetscBLASInt *select = agmres->select;
105   PetscInt     *perm   = agmres->perm;
106   PetscBLASInt  sdim   = 0;
107   PetscInt      i, j;
108   PetscBLASInt  info;
109   PetscBLASInt *iwork = agmres->iwork;
110   PetscBLASInt  N;
111   PetscBLASInt  lwork, liwork;
112   PetscBLASInt  ilo;
113   PetscBLASInt  ijob, wantQ, wantZ;
114   PetscScalar   Dif[2];
115 
116   PetscFunctionBegin;
117   ijob  = 2;
118   wantQ = 1;
119   wantZ = 1;
120   PetscCall(PetscBLASIntCast(MAXKSPSIZE, &N));
121   PetscCall(PetscBLASIntCast(PetscMax(8 * N + 16, 4 * neig * (N - neig)), &lwork));
122   PetscCall(PetscBLASIntCast(2 * N * neig, &liwork));
123   ilo = 1;
124 
125   /* Compute the Schur form */
126   if (IsReduced) { /* The eigenvalue problem is already in reduced form, meaning that A is upper Hessenberg and B is triangular */
127     PetscCallBLAS("LAPACKhgeqz", LAPACKhgeqz_("S", "I", "I", &KspSize, &ilo, &KspSize, A, &ldA, B, &ldB, wr, wi, beta, Q, &N, Z, &N, work, &lwork, &info));
128     PetscCheck(!info, PetscObjectComm((PetscObject)ksp), PETSC_ERR_PLIB, "Error while calling LAPACK routine xhgeqz_");
129   } else {
130     PetscCallBLAS("LAPACKgges", LAPACKgges_("V", "V", "N", NULL, &KspSize, A, &ldA, B, &ldB, &sdim, wr, wi, beta, Q, &N, Z, &N, work, &lwork, NULL, &info));
131     PetscCheck(!info, PetscObjectComm((PetscObject)ksp), PETSC_ERR_PLIB, "Error while calling LAPACK routine xgges_");
132   }
133 
134   /* We should avoid computing these ratio...  */
135   for (i = 0; i < KspSize; i++) {
136     if (beta[i] != 0.0) {
137       wr[i] /= beta[i];
138       wi[i] /= beta[i];
139     }
140   }
141 
142   /* Sort the eigenvalues to extract the smallest ones */
143   PetscCall(KSPAGMRESQuickSort(wr, wi, KspSize, perm));
144 
145   /* Count the number of extracted eigenvalues (with complex conjugates) */
146   r = 0;
147   while (r < neig) {
148     if (wi[r] != 0) r += 2;
149     else r += 1;
150   }
151   /* Reorder the Schur decomposition so that the cluster of smallest/largest eigenvalues appears in the leading diagonal blocks of A (and B)*/
152   PetscCall(PetscArrayzero(select, N));
153   if (!agmres->GreatestEig) {
154     for (j = 0; j < r; j++) select[perm[j]] = 1;
155   } else {
156     for (j = 0; j < r; j++) select[perm[KspSize - j - 1]] = 1;
157   }
158   PetscCallBLAS("LAPACKtgsen", LAPACKtgsen_(&ijob, &wantQ, &wantZ, select, &KspSize, A, &ldA, B, &ldB, wr, wi, beta, Q, &N, Z, &N, &r, NULL, NULL, &Dif[0], work, &lwork, iwork, &liwork, &info));
159   PetscCheck(info != 1, PetscObjectComm((PetscObject)ksp), PETSC_ERR_PLIB, "UNABLE TO REORDER THE EIGENVALUES WITH THE LAPACK ROUTINE : ILL-CONDITIONED PROBLEM");
160   /* Extract the Schur vectors associated to the r smallest eigenvalues */
161   PetscCall(PetscArrayzero(Sr, (N + 1) * r));
162   for (j = 0; j < r; j++) {
163     for (i = 0; i < KspSize; i++) Sr[j * (N + 1) + i] = Z[j * N + i];
164   }
165 
166   /* Broadcast Sr to all other processes to have consistent data;
167    * FIXME should investigate how to get unique Schur vectors (unique QR factorization, probably the sign of rotations) */
168   PetscCallMPI(MPI_Bcast(Sr, (N + 1) * r, MPIU_SCALAR, agmres->First, PetscObjectComm((PetscObject)ksp)));
169   /* Update the Shift values for the Newton basis. This is surely necessary when applying the DeflationPrecond */
170   if (agmres->DeflPrecond) PetscCall(KSPAGMRESLejaOrdering(wr, wi, agmres->Rshift, agmres->Ishift, max_k));
171   *CurNeig = r; /* Number of extracted eigenvalues */
172   PetscFunctionReturn(PETSC_SUCCESS);
173 }
174 
175 /*
176   Forms the matrices for the generalized eigenvalue problem,
177   it then compute the Schur vectors needed to augment the Newton basis.
178 */
KSPAGMRESComputeDeflationData(KSP ksp)179 PetscErrorCode KSPAGMRESComputeDeflationData(KSP ksp)
180 {
181   KSP_AGMRES  *agmres  = (KSP_AGMRES *)ksp->data;
182   Vec         *U       = agmres->U;
183   Vec         *TmpU    = agmres->TmpU;
184   PetscScalar *MatEigL = agmres->MatEigL;
185   PetscScalar *MatEigR = agmres->MatEigR;
186   PetscScalar *Sr      = agmres->Sr;
187   PetscScalar  alpha, beta;
188   PetscInt     i, j;
189   PetscInt     max_k = agmres->max_k; /* size of the non - augmented subspace */
190   PetscInt     CurNeig;               /* Current number of extracted eigenvalues */
191   PetscInt     N = MAXKSPSIZE;
192   PetscBLASInt bN, iKspSize;
193   PetscInt     lC      = N + 1;
194   PetscInt     KspSize = KSPSIZE;
195   PetscBLASInt blC, bKspSize;
196   PetscInt     PrevNeig = agmres->r;
197 
198   PetscFunctionBegin;
199   PetscCall(PetscLogEventBegin(KSP_AGMRESComputeDeflationData, ksp, 0, 0, 0));
200   if (agmres->neig <= 1) PetscFunctionReturn(PETSC_SUCCESS);
201   /* Explicitly form MatEigL = H^T*H, It can also be formed as H^T+h_{N+1,N}H^-1e^T */
202   alpha = 1.0;
203   beta  = 0.0;
204   PetscCall(PetscBLASIntCast(KspSize, &bKspSize));
205   PetscCall(PetscBLASIntCast(lC, &blC));
206   PetscCall(PetscBLASIntCast(N, &bN));
207   PetscCallBLAS("BLASgemm", BLASgemm_("T", "N", &bKspSize, &bKspSize, &blC, &alpha, agmres->hes_origin, &blC, agmres->hes_origin, &blC, &beta, MatEigL, &bN));
208   if (!agmres->ritz) {
209     /* Form TmpU = V*H where V is the Newton basis orthogonalized  with roddec*/
210     for (j = 0; j < KspSize; j++) {
211       /* Apply the elementary reflectors (stored in Qloc) on H */
212       PetscCall(KSPAGMRESRodvec(ksp, KspSize + 1, &agmres->hes_origin[j * lC], TmpU[j]));
213     }
214     /* Now form MatEigR = TmpU^T*W where W is [VEC_V(1:max_k); U] */
215     for (j = 0; j < max_k; j++) PetscCall(VecMDot(VEC_V(j), KspSize, TmpU, &MatEigR[j * N]));
216     for (j = max_k; j < KspSize; j++) PetscCall(VecMDot(U[j - max_k], KspSize, TmpU, &MatEigR[j * N]));
217   } else { /* Form H^T */
218     for (j = 0; j < N; j++) {
219       for (i = 0; i < N; i++) MatEigR[j * N + i] = agmres->hes_origin[i * lC + j];
220     }
221   }
222   /* Obtain the Schur form of  the generalized eigenvalue problem MatEigL*y = \lambda*MatEigR*y */
223   PetscCall(PetscBLASIntCast(KspSize, &iKspSize));
224   if (agmres->DeflPrecond) {
225     PetscCall(KSPAGMRESSchurForm(ksp, iKspSize, agmres->hes_origin, blC, agmres->Rloc, blC, PETSC_TRUE, Sr, &CurNeig));
226   } else {
227     PetscCall(KSPAGMRESSchurForm(ksp, iKspSize, MatEigL, bN, MatEigR, bN, PETSC_FALSE, Sr, &CurNeig));
228   }
229 
230   if (agmres->DeflPrecond) { /* Switch to DGMRES to improve the basis of the invariant subspace associated to the deflation */
231     agmres->HasSchur = PETSC_TRUE;
232     PetscCall(KSPDGMRESComputeDeflationData(ksp, &CurNeig));
233     PetscFunctionReturn(PETSC_SUCCESS);
234   }
235   /* Form the Schur vectors in the entire subspace: U = W * Sr where W = [VEC_V(1:max_k); U]*/
236   for (j = 0; j < PrevNeig; j++) { /* First, copy U to a temporary place */
237     PetscCall(VecCopy(U[j], TmpU[j]));
238   }
239 
240   for (j = 0; j < CurNeig; j++) {
241     PetscCall(VecMAXPBY(U[j], max_k, &Sr[j * (N + 1)], 0, &VEC_V(0)));
242     PetscCall(VecMAXPY(U[j], PrevNeig, &Sr[j * (N + 1) + max_k], TmpU));
243   }
244   agmres->r = CurNeig;
245   PetscCall(PetscLogEventEnd(KSP_AGMRESComputeDeflationData, ksp, 0, 0, 0));
246   PetscFunctionReturn(PETSC_SUCCESS);
247 }
248