Actual source code: fbcgsr.c
petsc-3.7.5 2017-01-01
2: /*
3: This file implements FBiCGStab-R.
4: Only allow right preconditioning.
5: FBiCGStab-R is a mathematically equivalent variant of FBiCGStab. Differences are:
6: (1) There are fewer MPI_Allreduce calls.
7: (2) The convergence occasionally is much faster than that of FBiCGStab.
8: */
9: #include <../src/ksp/ksp/impls/bcgs/bcgsimpl.h> /*I "petscksp.h" I*/
10: #include <petsc/private/vecimpl.h>
14: static PetscErrorCode KSPSetUp_FBCGSR(KSP ksp)
15: {
19: KSPSetWorkVecs(ksp,8);
20: return(0);
21: }
23: #include <petsc/private/pcimpl.h> /*I "petscksp.h" I*/
26: static PetscErrorCode KSPSolve_FBCGSR(KSP ksp)
27: {
28: PetscErrorCode ierr;
29: PetscInt i,j,N;
30: PetscScalar tau,sigma,alpha,omega,beta;
31: PetscReal rho;
32: PetscScalar xi1,xi2,xi3,xi4;
33: Vec X,B,P,P2,RP,R,V,S,T,S2;
34: PetscScalar *PETSC_RESTRICT rp, *PETSC_RESTRICT r, *PETSC_RESTRICT p;
35: PetscScalar *PETSC_RESTRICT v, *PETSC_RESTRICT s, *PETSC_RESTRICT t, *PETSC_RESTRICT s2;
36: PetscScalar insums[4],outsums[4];
37: KSP_BCGS *bcgs = (KSP_BCGS*)ksp->data;
38: PC pc;
41: if (!ksp->vec_rhs->petscnative) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"Only coded for PETSc vectors");
42: VecGetLocalSize(ksp->vec_sol,&N);
44: X = ksp->vec_sol;
45: B = ksp->vec_rhs;
46: P2 = ksp->work[0];
48: /* The followings are involved in modified inner product calculations and vector updates */
49: RP = ksp->work[1]; VecGetArray(RP,(PetscScalar**)&rp); VecRestoreArray(RP,NULL);
50: R = ksp->work[2]; VecGetArray(R,(PetscScalar**)&r); VecRestoreArray(R,NULL);
51: P = ksp->work[3]; VecGetArray(P,(PetscScalar**)&p); VecRestoreArray(P,NULL);
52: V = ksp->work[4]; VecGetArray(V,(PetscScalar**)&v); VecRestoreArray(V,NULL);
53: S = ksp->work[5]; VecGetArray(S,(PetscScalar**)&s); VecRestoreArray(S,NULL);
54: T = ksp->work[6]; VecGetArray(T,(PetscScalar**)&t); VecRestoreArray(T,NULL);
55: S2 = ksp->work[7]; VecGetArray(S2,(PetscScalar**)&s2); VecRestoreArray(S2,NULL);
57: /* Only supports right preconditioning */
58: if (ksp->pc_side != PC_RIGHT) SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"KSP fbcgsr does not support %s",PCSides[ksp->pc_side]);
59: if (!ksp->guess_zero) {
60: if (!bcgs->guess) {
61: VecDuplicate(X,&bcgs->guess);
62: }
63: VecCopy(X,bcgs->guess);
64: } else {
65: VecSet(X,0.0);
66: }
68: /* Compute initial residual */
69: KSPGetPC(ksp,&pc);
70: PCSetUp(pc);
71: if (!ksp->guess_zero) {
72: KSP_MatMult(ksp,pc->mat,X,P2); /* P2 is used as temporary storage */
73: VecCopy(B,R);
74: VecAXPY(R,-1.0,P2);
75: } else {
76: VecCopy(B,R);
77: }
79: /* Test for nothing to do */
80: VecNorm(R,NORM_2,&rho);
81: PetscObjectSAWsTakeAccess((PetscObject)ksp);
82: ksp->its = 0;
83: ksp->rnorm = rho;
84: PetscObjectSAWsGrantAccess((PetscObject)ksp);
85: KSPLogResidualHistory(ksp,rho);
86: KSPMonitor(ksp,0,rho);
87: (*ksp->converged)(ksp,0,rho,&ksp->reason,ksp->cnvP);
88: if (ksp->reason) return(0);
90: /* Initialize iterates */
91: VecCopy(R,RP); /* rp <- r */
92: VecCopy(R,P); /* p <- r */
94: /* Big loop */
95: for (i=0; i<ksp->max_it; i++) {
97: /* matmult and pc */
98: KSP_PCApply(ksp,P,P2); /* p2 <- K p */
99: KSP_MatMult(ksp,pc->mat,P2,V); /* v <- A p2 */
101: /* inner prodcuts */
102: if (i==0) {
103: tau = rho*rho;
104: VecDot(V,RP,&sigma); /* sigma <- (v,rp) */
105: } else {
106: PetscLogEventBegin(VEC_ReduceArithmetic,0,0,0,0);
107: tau = sigma = 0.0;
108: for (j=0; j<N; j++) {
109: tau += r[j]*rp[j]; /* tau <- (r,rp) */
110: sigma += v[j]*rp[j]; /* sigma <- (v,rp) */
111: }
112: PetscLogFlops(4.0*N);
113: PetscLogEventEnd(VEC_ReduceArithmetic,0,0,0,0);
114: insums[0] = tau;
115: insums[1] = sigma;
116: PetscLogEventBarrierBegin(VEC_ReduceBarrier,0,0,0,0,PetscObjectComm((PetscObject)ksp));
117: MPIU_Allreduce(insums,outsums,2,MPIU_SCALAR,MPIU_SUM,PetscObjectComm((PetscObject)ksp));
118: PetscLogEventBarrierEnd(VEC_ReduceBarrier,0,0,0,0,PetscObjectComm((PetscObject)ksp));
119: tau = outsums[0];
120: sigma = outsums[1];
121: }
123: /* scalar update */
124: alpha = tau / sigma;
126: /* vector update */
127: VecWAXPY(S,-alpha,V,R); /* s <- r - alpha v */
129: /* matmult and pc */
130: KSP_PCApply(ksp,S,S2); /* s2 <- K s */
131: KSP_MatMult(ksp,pc->mat,S2,T); /* t <- A s2 */
133: /* inner prodcuts */
134: PetscLogEventBegin(VEC_ReduceArithmetic,0,0,0,0);
135: xi1 = xi2 = xi3 = xi4 = 0.0;
136: for (j=0; j<N; j++) {
137: xi1 += s[j]*s[j]; /* xi1 <- (s,s) */
138: xi2 += t[j]*s[j]; /* xi2 <- (t,s) */
139: xi3 += t[j]*t[j]; /* xi3 <- (t,t) */
140: xi4 += t[j]*rp[j]; /* xi4 <- (t,rp) */
141: }
142: PetscLogFlops(8.0*N);
143: PetscLogEventEnd(VEC_ReduceArithmetic,0,0,0,0);
145: insums[0] = xi1;
146: insums[1] = xi2;
147: insums[2] = xi3;
148: insums[3] = xi4;
150: PetscLogEventBarrierBegin(VEC_ReduceBarrier,0,0,0,0,PetscObjectComm((PetscObject)ksp));
151: MPIU_Allreduce(insums,outsums,4,MPIU_SCALAR,MPIU_SUM,PetscObjectComm((PetscObject)ksp));
152: PetscLogEventBarrierEnd(VEC_ReduceBarrier,0,0,0,0,PetscObjectComm((PetscObject)ksp));
153: xi1 = outsums[0];
154: xi2 = outsums[1];
155: xi3 = outsums[2];
156: xi4 = outsums[3];
158: /* test denominator */
159: if (xi3 == 0.0) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_PLIB,"Divide by zero");
160: if (sigma == 0.0) SETERRQ(PetscObjectComm((PetscObject)ksp),PETSC_ERR_PLIB,"Divide by zero");
162: /* scalar updates */
163: omega = xi2 / xi3;
164: beta = -xi4 / sigma;
165: rho = PetscSqrtReal(PetscAbsScalar(xi1 - omega * xi2)); /* residual norm */
167: /* vector updates */
168: VecAXPBYPCZ(X,alpha,omega,1.0,P2,S2); /* x <- alpha * p2 + omega * s2 + x */
170: /* convergence test */
171: PetscObjectSAWsTakeAccess((PetscObject)ksp);
172: ksp->its++;
173: ksp->rnorm = rho;
174: PetscObjectSAWsGrantAccess((PetscObject)ksp);
175: KSPLogResidualHistory(ksp,rho);
176: KSPMonitor(ksp,i+1,rho);
177: (*ksp->converged)(ksp,i+1,rho,&ksp->reason,ksp->cnvP);
178: if (ksp->reason) break;
180: /* vector updates */
181: PetscLogEventBegin(VEC_Ops,0,0,0,0);
182: for (j=0; j<N; j++) {
183: r[j] = s[j] - omega * t[j]; /* r <- s - omega t */
184: p[j] = r[j] + beta * (p[j] - omega * v[j]); /* p <- r + beta * (p - omega v) */
185: }
186: PetscLogFlops(6.0*N);
187: PetscLogEventEnd(VEC_Ops,0,0,0,0);
189: }
191: if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
192: return(0);
193: }
195: /*MC
196: KSPFBCGSR - Implements a mathematically equivalent variant of FBiCGSTab.
198: Options Database Keys:
199: . see KSPSolve()
201: Level: beginner
203: Notes: Only allow right preconditioning
205: .seealso: KSPCreate(), KSPSetType(), KSPType (for list of available types), KSP, KSPBICG, KSPFBCGSL, KSPSetPCSide()
206: M*/
209: PETSC_EXTERN PetscErrorCode KSPCreate_FBCGSR(KSP ksp)
210: {
212: KSP_BCGS *bcgs;
215: PetscNewLog(ksp,&bcgs);
217: ksp->data = bcgs;
218: ksp->ops->setup = KSPSetUp_FBCGSR;
219: ksp->ops->solve = KSPSolve_FBCGSR;
220: ksp->ops->destroy = KSPDestroy_BCGS;
221: ksp->ops->reset = KSPReset_BCGS;
222: ksp->ops->buildsolution = KSPBuildSolution_BCGS;
223: ksp->ops->buildresidual = KSPBuildResidualDefault;
224: ksp->ops->setfromoptions = KSPSetFromOptions_BCGS;
225: ksp->pc_side = PC_RIGHT; /* set default PC side */
227: KSPSetSupportedNorm(ksp,KSP_NORM_PRECONDITIONED,PC_LEFT,3);
228: KSPSetSupportedNorm(ksp,KSP_NORM_UNPRECONDITIONED,PC_RIGHT,2);
229: return(0);
230: }