xref: /petsc/src/ksp/ksp/impls/bcgs/fbcgsr/fbcgsr.c (revision 447bcd8fae0acafb34f76e22d0a980ee3af1ea6c)
1 
2 /*
3     This file implements FBiCGStab-R.
4     FBiCGStab-R is a mathematically equivalent variant of FBiCGStab. Differences are:
5       (1) There are fewer MPI_Allreduce calls.
6       (2) The convergence occasionally is much faster than that of FBiCGStab.
7 */
8 #include <../src/ksp/ksp/impls/bcgs/bcgsimpl.h> /*I  "petscksp.h"  I*/
9 #include <petsc/private/vecimpl.h>
10 
11 static PetscErrorCode KSPSetUp_FBCGSR(KSP ksp)
12 {
13   PetscFunctionBegin;
14   PetscCall(KSPSetWorkVecs(ksp, 8));
15   PetscFunctionReturn(0);
16 }
17 
18 static PetscErrorCode KSPSolve_FBCGSR(KSP ksp)
19 {
20   PetscInt                    i, j, N;
21   PetscScalar                 tau, sigma, alpha, omega, beta;
22   PetscReal                   rho;
23   PetscScalar                 xi1, xi2, xi3, xi4;
24   Vec                         X, B, P, P2, RP, R, V, S, T, S2;
25   PetscScalar *PETSC_RESTRICT rp, *PETSC_RESTRICT r, *PETSC_RESTRICT p;
26   PetscScalar *PETSC_RESTRICT v, *PETSC_RESTRICT s, *PETSC_RESTRICT t, *PETSC_RESTRICT s2;
27   PetscScalar insums[4], outsums[4];
28   KSP_BCGS   *bcgs = (KSP_BCGS *)ksp->data;
29   PC          pc;
30   Mat         mat;
31 
32   PetscFunctionBegin;
33   PetscCheck(ksp->vec_rhs->petscnative, PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "Only coded for PETSc vectors");
34   PetscCall(VecGetLocalSize(ksp->vec_sol, &N));
35 
36   X  = ksp->vec_sol;
37   B  = ksp->vec_rhs;
38   P2 = ksp->work[0];
39 
40   /* The followings are involved in modified inner product calculations and vector updates */
41   RP = ksp->work[1];
42   PetscCall(VecGetArray(RP, (PetscScalar **)&rp));
43   PetscCall(VecRestoreArray(RP, NULL));
44   R = ksp->work[2];
45   PetscCall(VecGetArray(R, (PetscScalar **)&r));
46   PetscCall(VecRestoreArray(R, NULL));
47   P = ksp->work[3];
48   PetscCall(VecGetArray(P, (PetscScalar **)&p));
49   PetscCall(VecRestoreArray(P, NULL));
50   V = ksp->work[4];
51   PetscCall(VecGetArray(V, (PetscScalar **)&v));
52   PetscCall(VecRestoreArray(V, NULL));
53   S = ksp->work[5];
54   PetscCall(VecGetArray(S, (PetscScalar **)&s));
55   PetscCall(VecRestoreArray(S, NULL));
56   T = ksp->work[6];
57   PetscCall(VecGetArray(T, (PetscScalar **)&t));
58   PetscCall(VecRestoreArray(T, NULL));
59   S2 = ksp->work[7];
60   PetscCall(VecGetArray(S2, (PetscScalar **)&s2));
61   PetscCall(VecRestoreArray(S2, NULL));
62 
63   /* Only supports right preconditioning */
64   PetscCheck(ksp->pc_side == PC_RIGHT, PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "KSP fbcgsr does not support %s", PCSides[ksp->pc_side]);
65   if (!ksp->guess_zero) {
66     if (!bcgs->guess) PetscCall(VecDuplicate(X, &bcgs->guess));
67     PetscCall(VecCopy(X, bcgs->guess));
68   } else {
69     PetscCall(VecSet(X, 0.0));
70   }
71 
72   /* Compute initial residual */
73   PetscCall(KSPGetPC(ksp, &pc));
74   PetscCall(PCSetUp(pc));
75   PetscCall(PCGetOperators(pc, &mat, NULL));
76   if (!ksp->guess_zero) {
77     PetscCall(KSP_MatMult(ksp, mat, X, P2)); /* P2 is used as temporary storage */
78     PetscCall(VecCopy(B, R));
79     PetscCall(VecAXPY(R, -1.0, P2));
80   } else {
81     PetscCall(VecCopy(B, R));
82   }
83 
84   /* Test for nothing to do */
85   PetscCall(VecNorm(R, NORM_2, &rho));
86   PetscCall(PetscObjectSAWsTakeAccess((PetscObject)ksp));
87   ksp->its = 0;
88   if (ksp->normtype != KSP_NORM_NONE) ksp->rnorm = rho;
89   else ksp->rnorm = 0;
90   PetscCall(PetscObjectSAWsGrantAccess((PetscObject)ksp));
91   PetscCall(KSPLogResidualHistory(ksp, ksp->rnorm));
92   PetscCall(KSPMonitor(ksp, 0, ksp->rnorm));
93   PetscCall((*ksp->converged)(ksp, 0, ksp->rnorm, &ksp->reason, ksp->cnvP));
94   if (ksp->reason) PetscFunctionReturn(0);
95 
96   /* Initialize iterates */
97   PetscCall(VecCopy(R, RP)); /* rp <- r */
98   PetscCall(VecCopy(R, P));  /* p <- r */
99 
100   /* Big loop */
101   for (i = 0; i < ksp->max_it; i++) {
102     /* matmult and pc */
103     PetscCall(KSP_PCApply(ksp, P, P2));      /* p2 <- K p */
104     PetscCall(KSP_MatMult(ksp, mat, P2, V)); /* v <- A p2 */
105 
106     /* inner prodcuts */
107     if (i == 0) {
108       tau = rho * rho;
109       PetscCall(VecDot(V, RP, &sigma)); /* sigma <- (v,rp) */
110     } else {
111       PetscCall(PetscLogEventBegin(VEC_ReduceArithmetic, 0, 0, 0, 0));
112       tau = sigma = 0.0;
113       for (j = 0; j < N; j++) {
114         tau += r[j] * rp[j];   /* tau <- (r,rp) */
115         sigma += v[j] * rp[j]; /* sigma <- (v,rp) */
116       }
117       PetscCall(PetscLogFlops(4.0 * N));
118       PetscCall(PetscLogEventEnd(VEC_ReduceArithmetic, 0, 0, 0, 0));
119       insums[0] = tau;
120       insums[1] = sigma;
121       PetscCall(PetscLogEventBegin(VEC_ReduceCommunication, 0, 0, 0, 0));
122       PetscCall(MPIU_Allreduce(insums, outsums, 2, MPIU_SCALAR, MPIU_SUM, PetscObjectComm((PetscObject)ksp)));
123       PetscCall(PetscLogEventEnd(VEC_ReduceCommunication, 0, 0, 0, 0));
124       tau   = outsums[0];
125       sigma = outsums[1];
126     }
127 
128     /* scalar update */
129     alpha = tau / sigma;
130 
131     /* vector update */
132     PetscCall(VecWAXPY(S, -alpha, V, R)); /* s <- r - alpha v */
133 
134     /* matmult and pc */
135     PetscCall(KSP_PCApply(ksp, S, S2));      /* s2 <- K s */
136     PetscCall(KSP_MatMult(ksp, mat, S2, T)); /* t <- A s2 */
137 
138     /* inner prodcuts */
139     PetscCall(PetscLogEventBegin(VEC_ReduceArithmetic, 0, 0, 0, 0));
140     xi1 = xi2 = xi3 = xi4 = 0.0;
141     for (j = 0; j < N; j++) {
142       xi1 += s[j] * s[j];  /* xi1 <- (s,s) */
143       xi2 += t[j] * s[j];  /* xi2 <- (t,s) */
144       xi3 += t[j] * t[j];  /* xi3 <- (t,t) */
145       xi4 += t[j] * rp[j]; /* xi4 <- (t,rp) */
146     }
147     PetscCall(PetscLogFlops(8.0 * N));
148     PetscCall(PetscLogEventEnd(VEC_ReduceArithmetic, 0, 0, 0, 0));
149 
150     insums[0] = xi1;
151     insums[1] = xi2;
152     insums[2] = xi3;
153     insums[3] = xi4;
154 
155     PetscCall(PetscLogEventBegin(VEC_ReduceCommunication, 0, 0, 0, 0));
156     PetscCall(MPIU_Allreduce(insums, outsums, 4, MPIU_SCALAR, MPIU_SUM, PetscObjectComm((PetscObject)ksp)));
157     PetscCall(PetscLogEventEnd(VEC_ReduceCommunication, 0, 0, 0, 0));
158     xi1 = outsums[0];
159     xi2 = outsums[1];
160     xi3 = outsums[2];
161     xi4 = outsums[3];
162 
163     /* test denominator */
164     if ((xi3 == 0.0) || (sigma == 0.0)) {
165       PetscCheck(!ksp->errorifnotconverged, PetscObjectComm((PetscObject)ksp), PETSC_ERR_NOT_CONVERGED, "KSPSolve has failed due to zero inner product");
166       ksp->reason = KSP_DIVERGED_BREAKDOWN;
167       PetscCall(PetscInfo(ksp, "KSPSolve has failed due to zero inner product\n"));
168       break;
169     }
170 
171     /* scalar updates */
172     omega = xi2 / xi3;
173     beta  = -xi4 / sigma;
174     rho   = PetscSqrtReal(PetscAbsScalar(xi1 - omega * xi2)); /* residual norm */
175 
176     /* vector updates */
177     PetscCall(VecAXPBYPCZ(X, alpha, omega, 1.0, P2, S2)); /* x <- alpha * p2 + omega * s2 + x */
178 
179     /* convergence test */
180     PetscCall(PetscObjectSAWsTakeAccess((PetscObject)ksp));
181     ksp->its++;
182     if (ksp->normtype != KSP_NORM_NONE) ksp->rnorm = rho;
183     else ksp->rnorm = 0;
184     PetscCall(PetscObjectSAWsGrantAccess((PetscObject)ksp));
185     PetscCall(KSPLogResidualHistory(ksp, ksp->rnorm));
186     PetscCall(KSPMonitor(ksp, i + 1, ksp->rnorm));
187     PetscCall((*ksp->converged)(ksp, i + 1, ksp->rnorm, &ksp->reason, ksp->cnvP));
188     if (ksp->reason) break;
189 
190     /* vector updates */
191     PetscCall(PetscLogEventBegin(VEC_Ops, 0, 0, 0, 0));
192     for (j = 0; j < N; j++) {
193       r[j] = s[j] - omega * t[j];                 /* r <- s - omega t */
194       p[j] = r[j] + beta * (p[j] - omega * v[j]); /* p <- r + beta * (p - omega v) */
195     }
196     PetscCall(PetscLogFlops(6.0 * N));
197     PetscCall(PetscLogEventEnd(VEC_Ops, 0, 0, 0, 0));
198   }
199 
200   if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
201   PetscFunctionReturn(0);
202 }
203 
204 /*MC
205      KSPFBCGSR - Implements a mathematically equivalent variant of flexible bi-CG-stab, `KSPFBCGS`. [](sec_flexibleksp)
206 
207    Level: beginner
208 
209    Notes:
210    This implementation requires fewer `MPI_Allreduce()` calls than `KSPFBCGS` and may converge faster
211 
212    See `KSPPIPEBCGS` for a pipelined version of the algorithm
213 
214    Flexible BiCGStab, unlike most Krylov methods, allows the preconditioner to be nonlinear, that is the action of the preconditioner to a vector need not be linear
215    in the vector entries.
216 
217    Only supports right preconditioning
218 
219 .seealso: [](chapter_ksp),  [](sec_flexibleksp), `KSPFBCGSR`, `KSPPIPEBCGS`, `KSPBCGSL`, `KSPBCGS`, `KSPCreate()`, `KSPSetType()`, `KSPType`, `KSP`, `KSPBICG`, `KSPFBCGSL`, `KSPSetPCSide()`
220 M*/
221 PETSC_EXTERN PetscErrorCode KSPCreate_FBCGSR(KSP ksp)
222 {
223   KSP_BCGS *bcgs;
224 
225   PetscFunctionBegin;
226   PetscCall(PetscNew(&bcgs));
227 
228   ksp->data                = bcgs;
229   ksp->ops->setup          = KSPSetUp_FBCGSR;
230   ksp->ops->solve          = KSPSolve_FBCGSR;
231   ksp->ops->destroy        = KSPDestroy_BCGS;
232   ksp->ops->reset          = KSPReset_BCGS;
233   ksp->ops->buildsolution  = KSPBuildSolution_BCGS;
234   ksp->ops->buildresidual  = KSPBuildResidualDefault;
235   ksp->ops->setfromoptions = KSPSetFromOptions_BCGS;
236   ksp->pc_side             = PC_RIGHT; /* set default PC side */
237 
238   PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_PRECONDITIONED, PC_LEFT, 3));
239   PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_UNPRECONDITIONED, PC_RIGHT, 2));
240   PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_NONE, PC_RIGHT, 1));
241   PetscFunctionReturn(0);
242 }
243