Actual source code: pipecr.c
petsc-3.7.7 2017-09-25
1: #include <petsc/private/kspimpl.h>
3: /*
4: KSPSetUp_PIPECR - Sets up the workspace needed by the PIPECR method.
6: This is called once, usually automatically by KSPSolve() or KSPSetUp()
7: but can be called directly by KSPSetUp()
8: */
11: PetscErrorCode KSPSetUp_PIPECR(KSP ksp)
12: {
16: /* get work vectors needed by PIPECR */
17: KSPSetWorkVecs(ksp,7);
18: return(0);
19: }
21: /*
22: KSPSolve_PIPECR - This routine actually applies the pipelined conjugate residual method
24: Input Parameter:
25: . ksp - the Krylov space object that was set to use conjugate gradient, by, for
26: example, KSPCreate(MPI_Comm,KSP *ksp); KSPSetType(ksp,KSPCG);
27: */
30: PetscErrorCode KSPSolve_PIPECR(KSP ksp)
31: {
33: PetscInt i;
34: PetscScalar alpha=0.0,beta=0.0,gamma,gammaold=0.0,delta;
35: PetscReal dp = 0.0;
36: Vec X,B,Z,P,W,Q,U,M,N;
37: Mat Amat,Pmat;
38: PetscBool diagonalscale;
41: PCGetDiagonalScale(ksp->pc,&diagonalscale);
42: if (diagonalscale) SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"Krylov method %s does not support diagonal scaling",((PetscObject)ksp)->type_name);
44: X = ksp->vec_sol;
45: B = ksp->vec_rhs;
46: M = ksp->work[0];
47: Z = ksp->work[1];
48: P = ksp->work[2];
49: N = ksp->work[3];
50: W = ksp->work[4];
51: Q = ksp->work[5];
52: U = ksp->work[6];
54: PCGetOperators(ksp->pc,&Amat,&Pmat);
56: ksp->its = 0;
57: /* we don't have an R vector, so put the (unpreconditioned) residual in w for now */
58: if (!ksp->guess_zero) {
59: KSP_MatMult(ksp,Amat,X,W); /* w <- b - Ax */
60: VecAYPX(W,-1.0,B);
61: } else {
62: VecCopy(B,W); /* w <- b (x is 0) */
63: }
64: KSP_PCApply(ksp,W,U); /* u <- Bw */
66: switch (ksp->normtype) {
67: case KSP_NORM_PRECONDITIONED:
68: VecNormBegin(U,NORM_2,&dp); /* dp <- u'*u = e'*A'*B'*B*A'*e' */
69: PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)U));
70: KSP_MatMult(ksp,Amat,U,W); /* w <- Au */
71: VecNormEnd(U,NORM_2,&dp);
72: break;
73: case KSP_NORM_NONE:
74: KSP_MatMult(ksp,Amat,U,W);
75: dp = 0.0;
76: break;
77: default: SETERRQ1(PetscObjectComm((PetscObject)ksp),PETSC_ERR_SUP,"%s",KSPNormTypes[ksp->normtype]);
78: }
79: KSPLogResidualHistory(ksp,dp);
80: KSPMonitor(ksp,0,dp);
81: ksp->rnorm = dp;
82: (*ksp->converged)(ksp,0,dp,&ksp->reason,ksp->cnvP); /* test for convergence */
83: if (ksp->reason) return(0);
85: i = 0;
86: do {
87: KSP_PCApply(ksp,W,M); /* m <- Bw */
89: if (i > 0 && ksp->normtype == KSP_NORM_PRECONDITIONED) {
90: VecNormBegin(U,NORM_2,&dp);
91: }
92: VecDotBegin(W,U,&gamma);
93: VecDotBegin(M,W,&delta);
94: PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)U));
96: KSP_MatMult(ksp,Amat,M,N); /* n <- Am */
98: if (i > 0 && ksp->normtype == KSP_NORM_PRECONDITIONED) {
99: VecNormEnd(U,NORM_2,&dp);
100: }
101: VecDotEnd(W,U,&gamma);
102: VecDotEnd(M,W,&delta);
104: if (i > 0) {
105: if (ksp->normtype == KSP_NORM_NONE) dp = 0.0;
106: ksp->rnorm = dp;
107: KSPLogResidualHistory(ksp,dp);
108: KSPMonitor(ksp,i,dp);
109: (*ksp->converged)(ksp,i,dp,&ksp->reason,ksp->cnvP);
110: if (ksp->reason) break;
111: }
113: if (i == 0) {
114: alpha = gamma / delta;
115: VecCopy(N,Z); /* z <- n */
116: VecCopy(M,Q); /* q <- m */
117: VecCopy(U,P); /* p <- u */
118: } else {
119: beta = gamma / gammaold;
120: alpha = gamma / (delta - beta / alpha * gamma);
121: VecAYPX(Z,beta,N); /* z <- n + beta * z */
122: VecAYPX(Q,beta,M); /* q <- m + beta * q */
123: VecAYPX(P,beta,U); /* p <- u + beta * p */
124: }
125: VecAXPY(X, alpha,P); /* x <- x + alpha * p */
126: VecAXPY(U,-alpha,Q); /* u <- u - alpha * q */
127: VecAXPY(W,-alpha,Z); /* w <- w - alpha * z */
128: gammaold = gamma;
129: i++;
130: ksp->its = i;
132: /* if (i%50 == 0) { */
133: /* KSP_MatMult(ksp,Amat,X,W); /\* w <- b - Ax *\/ */
134: /* VecAYPX(W,-1.0,B); */
135: /* KSP_PCApply(ksp,W,U); */
136: /* KSP_MatMult(ksp,Amat,U,W); */
137: /* } */
139: } while (i<ksp->max_it);
140: if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
141: return(0);
142: }
144: /*MC
145: KSPPIPECR - Pipelined conjugate residual method
147: This method has only a single non-blocking reduction per iteration, compared to 2 blocking for standard CR. The
148: non-blocking reduction is overlapped by the matrix-vector product, but not the preconditioner application.
150: See also KSPPIPECG, where the reduction is only overlapped with the matrix-vector product.
152: Level: intermediate
154: Notes:
155: MPI configuration may be necessary for reductions to make asynchronous progress, which is important for performance of pipelined methods.
156: See the FAQ on the PETSc website for details.
158: Contributed by:
159: Pieter Ghysels, Universiteit Antwerpen, Intel Exascience lab Flanders
161: Reference:
162: P. Ghysels and W. Vanroose, "Hiding global synchronization latency in the preconditioned Conjugate Gradient algorithm",
163: Submitted to Parallel Computing, 2012.
165: .seealso: KSPCreate(), KSPSetType(), KSPPIPECG, KSPGROPPCG, KSPPGMRES, KSPCG, KSPCGUseSingleReduction()
166: M*/
170: PETSC_EXTERN PetscErrorCode KSPCreate_PIPECR(KSP ksp)
171: {
175: KSPSetSupportedNorm(ksp,KSP_NORM_PRECONDITIONED,PC_LEFT,2);
177: ksp->ops->setup = KSPSetUp_PIPECR;
178: ksp->ops->solve = KSPSolve_PIPECR;
179: ksp->ops->destroy = KSPDestroyDefault;
180: ksp->ops->view = 0;
181: ksp->ops->setfromoptions = 0;
182: ksp->ops->buildsolution = KSPBuildSolutionDefault;
183: ksp->ops->buildresidual = KSPBuildResidualDefault;
184: return(0);
185: }