1 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 11 #undef __FUNCT__ 12 #define __FUNCT__ "KSPSetUp_FBCGSR" 13 PetscErrorCode KSPSetUp_FBCGSR(KSP ksp) 14 { 15 PetscErrorCode ierr; 16 17 PetscFunctionBegin; 18 ierr = KSPDefaultGetWork(ksp,8);CHKERRQ(ierr); 19 PetscFunctionReturn(0); 20 } 21 22 #include <petsc-private/pcimpl.h> /*I "petscksp.h" I*/ 23 #undef __FUNCT__ 24 #define __FUNCT__ "KSPSolve_FBCGSR" 25 PetscErrorCode KSPSolve_FBCGSR(KSP ksp) 26 { 27 PetscErrorCode ierr; 28 PetscInt i,j,N; 29 PetscScalar tau,sigma,alpha,omega,beta; 30 PetscReal rho; 31 PetscScalar xi1,xi2,xi3,xi4; 32 Vec X,B,P,P2,RP,R,V,S,T,S2; 33 PetscScalar *PETSC_RESTRICT rp, *PETSC_RESTRICT r, *PETSC_RESTRICT p; 34 PetscScalar *PETSC_RESTRICT v, *PETSC_RESTRICT s, *PETSC_RESTRICT t, *PETSC_RESTRICT s2; 35 PetscScalar insums[4],outsums[4]; 36 KSP_BCGS *bcgs = (KSP_BCGS*)ksp->data; 37 PC pc; 38 39 PetscFunctionBegin; 40 if (!ksp->vec_rhs->petscnative) SETERRQ(((PetscObject)ksp)->comm,PETSC_ERR_SUP,"Only coded for PETSc vectors"); 41 ierr = VecGetLocalSize(ksp->vec_sol,&N);CHKERRQ(ierr); 42 43 X = ksp->vec_sol; 44 B = ksp->vec_rhs; 45 P2 = ksp->work[0]; 46 47 /* The followings are involved in modified inner product calculations and vector updates */ 48 RP = ksp->work[1]; ierr = VecGetArray(RP,(PetscScalar**)&rp);CHKERRQ(ierr); ierr = VecRestoreArray(RP,PETSC_NULL);CHKERRQ(ierr); 49 R = ksp->work[2]; ierr = VecGetArray(R,(PetscScalar**)&r);CHKERRQ(ierr); ierr = VecRestoreArray(R,PETSC_NULL);CHKERRQ(ierr); 50 P = ksp->work[3]; ierr = VecGetArray(P,(PetscScalar**)&p);CHKERRQ(ierr); ierr = VecRestoreArray(P,PETSC_NULL);CHKERRQ(ierr); 51 V = ksp->work[4]; ierr = VecGetArray(V,(PetscScalar**)&v);CHKERRQ(ierr); ierr = VecRestoreArray(V,PETSC_NULL);CHKERRQ(ierr); 52 S = ksp->work[5]; ierr = VecGetArray(S,(PetscScalar**)&s);CHKERRQ(ierr); ierr = VecRestoreArray(S,PETSC_NULL);CHKERRQ(ierr); 53 T = ksp->work[6]; ierr = VecGetArray(T,(PetscScalar**)&t);CHKERRQ(ierr); ierr = VecRestoreArray(T,PETSC_NULL);CHKERRQ(ierr); 54 S2 = ksp->work[7]; ierr = VecGetArray(S2,(PetscScalar**)&s2);CHKERRQ(ierr); ierr = VecRestoreArray(S2,PETSC_NULL);CHKERRQ(ierr); 55 56 /* Only supports right preconditioning */ 57 if (ksp->pc_side != PC_RIGHT) 58 SETERRQ1(((PetscObject)ksp)->comm,PETSC_ERR_SUP,"KSP fbcgsr does not support %s",PCSides[ksp->pc_side]); 59 if (!ksp->guess_zero) { 60 if (!bcgs->guess) { 61 ierr = VecDuplicate(X,&bcgs->guess);CHKERRQ(ierr); 62 } 63 ierr = VecCopy(X,bcgs->guess);CHKERRQ(ierr); 64 } else { 65 ierr = VecSet(X,0.0);CHKERRQ(ierr); 66 } 67 68 /* Compute initial residual */ 69 ierr = KSPGetPC(ksp,&pc);CHKERRQ(ierr); 70 ierr = PCSetUp(pc);CHKERRQ(ierr); 71 if (!ksp->guess_zero) { 72 ierr = MatMult(pc->mat,X,P2);CHKERRQ(ierr); /* P2 is used as temporary storage */ 73 ierr = VecCopy(B,R);CHKERRQ(ierr); 74 ierr = VecAXPY(R,-1.0,P2);CHKERRQ(ierr); 75 } else { 76 ierr = VecCopy(B,R);CHKERRQ(ierr); 77 } 78 79 /* Test for nothing to do */ 80 if (ksp->normtype != KSP_NORM_NONE) { 81 ierr = VecNorm(R,NORM_2,&rho);CHKERRQ(ierr); 82 } 83 ierr = PetscObjectTakeAccess(ksp);CHKERRQ(ierr); 84 ksp->its = 0; 85 ksp->rnorm = rho; 86 ierr = PetscObjectGrantAccess(ksp);CHKERRQ(ierr); 87 KSPLogResidualHistory(ksp,rho); 88 ierr = KSPMonitor(ksp,0,rho);CHKERRQ(ierr); 89 ierr = (*ksp->converged)(ksp,0,rho,&ksp->reason,ksp->cnvP);CHKERRQ(ierr); 90 if (ksp->reason) PetscFunctionReturn(0); 91 92 /* Initialize iterates */ 93 ierr = VecCopy(R,RP);CHKERRQ(ierr); /* rp <- r */ 94 ierr = VecCopy(R,P);CHKERRQ(ierr); /* p <- r */ 95 96 /* Big loop */ 97 for (i=0; i<ksp->max_it; i++) { 98 99 /* matmult and pc */ 100 ierr = PCApply(pc,P,P2);CHKERRQ(ierr); /* p2 <- K p */ 101 ierr = MatMult(pc->mat,P2,V);CHKERRQ(ierr); /* v <- A p2 */ 102 103 /* inner prodcuts */ 104 if (i==0) { 105 tau = rho*rho; 106 ierr = VecDot(V,RP,&sigma);CHKERRQ(ierr); /* sigma <- (v,rp) */ 107 } else { 108 ierr = PetscLogEventBegin(VEC_ReduceArithmetic,0,0,0,0);CHKERRQ(ierr); 109 tau = sigma = 0.0; 110 for (j=0; j<N; j++) { 111 tau += r[j]*rp[j]; /* tau <- (r,rp) */ 112 sigma += v[j]*rp[j]; /* sigma <- (v,rp) */ 113 } 114 PetscLogFlops(4.0*N); 115 ierr = PetscLogEventEnd(VEC_ReduceArithmetic,0,0,0,0);CHKERRQ(ierr); 116 insums[0] = tau; 117 insums[1] = sigma; 118 ierr = PetscLogEventBarrierBegin(VEC_ReduceBarrier,0,0,0,0,((PetscObject)ksp)->comm);CHKERRQ(ierr); 119 ierr = MPI_Allreduce(insums,outsums,2,MPIU_SCALAR,MPIU_SUM,((PetscObject)ksp)->comm);CHKERRQ(ierr); 120 ierr = PetscLogEventBarrierEnd(VEC_ReduceBarrier,0,0,0,0,((PetscObject)ksp)->comm);CHKERRQ(ierr); 121 tau = outsums[0]; 122 sigma = outsums[1]; 123 } 124 125 /* scalar update */ 126 alpha = tau / sigma; 127 128 /* vector update */ 129 ierr = VecWAXPY(S,-alpha,V,R);CHKERRQ(ierr); /* s <- r - alpha v */ 130 131 /* matmult and pc */ 132 ierr = PCApply(pc,S,S2);CHKERRQ(ierr); /* s2 <- K s */ 133 ierr = MatMult(pc->mat,S2,T);CHKERRQ(ierr); /* t <- A s2 */ 134 135 /* inner prodcuts */ 136 ierr = PetscLogEventBegin(VEC_ReduceArithmetic,0,0,0,0);CHKERRQ(ierr); 137 xi1 = xi2 = xi3 = xi4 = 0.0; 138 for (j=0; j<N; j++) { 139 xi1 += s[j]*s[j]; /* xi1 <- (s,s) */ 140 xi2 += t[j]*s[j]; /* xi2 <- (t,s) */ 141 xi3 += t[j]*t[j]; /* xi3 <- (t,t) */ 142 xi4 += t[j]*rp[j]; /* xi4 <- (t,rp) */ 143 } 144 PetscLogFlops(8.0*N); 145 ierr = PetscLogEventEnd(VEC_ReduceArithmetic,0,0,0,0);CHKERRQ(ierr); 146 insums[0] = xi1; 147 insums[1] = xi2; 148 insums[2] = xi3; 149 insums[3] = xi4; 150 ierr = PetscLogEventBarrierBegin(VEC_ReduceBarrier,0,0,0,0,((PetscObject)ksp)->comm);CHKERRQ(ierr); 151 ierr = MPI_Allreduce(insums,outsums,4,MPIU_SCALAR,MPIU_SUM,((PetscObject)ksp)->comm);CHKERRQ(ierr); 152 ierr = PetscLogEventBarrierEnd(VEC_ReduceBarrier,0,0,0,0,((PetscObject)ksp)->comm);CHKERRQ(ierr); 153 xi1 = outsums[0]; 154 xi2 = outsums[1]; 155 xi3 = outsums[2]; 156 xi4 = outsums[3]; 157 158 /* test denominator */ 159 if (xi3 == 0.0) SETERRQ(((PetscObject)ksp)->comm,PETSC_ERR_PLIB,"Divide by zero"); 160 if (sigma == 0.0) SETERRQ(((PetscObject)ksp)->comm,PETSC_ERR_PLIB,"Divide by zero"); 161 162 /* scalar updates */ 163 omega = xi2 / xi3; 164 beta = - xi4 / sigma; 165 rho = PetscSqrtReal(PetscAbsScalar(xi1 - omega * xi2)); /* residual norm */ 166 167 /* vector updates */ 168 ierr = VecAXPBYPCZ(X,alpha,omega,1.0,P2,S2);CHKERRQ(ierr); /* x <- alpha * p2 + omega * s2 + x */ 169 170 /* convergence test */ 171 ierr = PetscObjectTakeAccess(ksp);CHKERRQ(ierr); 172 ksp->its++; 173 ksp->rnorm = rho; 174 ierr = PetscObjectGrantAccess(ksp);CHKERRQ(ierr); 175 KSPLogResidualHistory(ksp,rho); 176 ierr = KSPMonitor(ksp,i+1,rho);CHKERRQ(ierr); 177 ierr = (*ksp->converged)(ksp,i+1,rho,&ksp->reason,ksp->cnvP);CHKERRQ(ierr); 178 if (ksp->reason) break; 179 180 /* vector updates */ 181 ierr = PetscLogEventBegin(VEC_Ops,0,0,0,0);CHKERRQ(ierr); 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 ierr = PetscLogEventEnd(VEC_Ops,0,0,0,0);CHKERRQ(ierr); 188 189 } 190 191 if (i >= ksp->max_it) { 192 ksp->reason = KSP_DIVERGED_ITS; 193 } 194 195 PetscFunctionReturn(0); 196 } 197 198 /*MC 199 KSPFBCGSR - Implements a mathematically equivalent variant of FBiCGSTab. 200 201 Options Database Keys: 202 . see KSPSolve() 203 204 Level: beginner 205 206 Notes: Only allow right preconditioning 207 208 .seealso: KSPCreate(), KSPSetType(), KSPType (for list of available types), KSP, KSPBICG, KSPFBCGSL, KSPSetPCSide() 209 M*/ 210 EXTERN_C_BEGIN 211 #undef __FUNCT__ 212 #define __FUNCT__ "KSPCreate_FBCGSR" 213 PetscErrorCode KSPCreate_FBCGSR(KSP ksp) 214 { 215 PetscErrorCode ierr; 216 KSP_BCGS *bcgs; 217 218 PetscFunctionBegin; 219 ierr = PetscNewLog(ksp,KSP_BCGS,&bcgs);CHKERRQ(ierr); 220 ksp->data = bcgs; 221 ksp->ops->setup = KSPSetUp_FBCGSR; 222 ksp->ops->solve = KSPSolve_FBCGSR; 223 ksp->ops->destroy = KSPDestroy_BCGS; 224 ksp->ops->reset = KSPReset_BCGS; 225 ksp->ops->buildsolution = KSPBuildSolution_BCGS; 226 ksp->ops->buildresidual = KSPDefaultBuildResidual; 227 ksp->ops->setfromoptions = KSPSetFromOptions_BCGS; 228 ksp->ops->view = KSPView_BCGS; 229 ksp->pc_side = PC_RIGHT; /* set default PC side */ 230 231 ierr = KSPSetSupportedNorm(ksp,KSP_NORM_PRECONDITIONED,PC_LEFT,2);CHKERRQ(ierr); 232 ierr = KSPSetSupportedNorm(ksp,KSP_NORM_UNPRECONDITIONED,PC_RIGHT,1);CHKERRQ(ierr); 233 PetscFunctionReturn(0); 234 } 235 EXTERN_C_END 236