1 #include <petsc/private/kspimpl.h>
2
3 /*
4 KSPSetUp_PIPECR - Sets up the workspace needed by the PIPECR method.
5
6 This is called once, usually automatically by KSPSolve() or KSPSetUp()
7 but can be called directly by KSPSetUp()
8 */
KSPSetUp_PIPECR(KSP ksp)9 static PetscErrorCode KSPSetUp_PIPECR(KSP ksp)
10 {
11 PetscFunctionBegin;
12 /* get work vectors needed by PIPECR */
13 PetscCall(KSPSetWorkVecs(ksp, 7));
14 PetscFunctionReturn(PETSC_SUCCESS);
15 }
16
17 /*
18 KSPSolve_PIPECR - This routine actually applies the pipelined conjugate residual method
19 */
KSPSolve_PIPECR(KSP ksp)20 static PetscErrorCode KSPSolve_PIPECR(KSP ksp)
21 {
22 PetscInt i;
23 PetscScalar alpha = 0.0, beta = 0.0, gamma, gammaold = 0.0, delta;
24 PetscReal dp = 0.0;
25 Vec X, B, Z, P, W, Q, U, M, N;
26 Mat Amat, Pmat;
27 PetscBool diagonalscale;
28
29 PetscFunctionBegin;
30 PetscCall(PCGetDiagonalScale(ksp->pc, &diagonalscale));
31 PetscCheck(!diagonalscale, PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "Krylov method %s does not support diagonal scaling", ((PetscObject)ksp)->type_name);
32
33 X = ksp->vec_sol;
34 B = ksp->vec_rhs;
35 M = ksp->work[0];
36 Z = ksp->work[1];
37 P = ksp->work[2];
38 N = ksp->work[3];
39 W = ksp->work[4];
40 Q = ksp->work[5];
41 U = ksp->work[6];
42
43 PetscCall(PCGetOperators(ksp->pc, &Amat, &Pmat));
44
45 ksp->its = 0;
46 /* we don't have an R vector, so put the (unpreconditioned) residual in w for now */
47 if (!ksp->guess_zero) {
48 PetscCall(KSP_MatMult(ksp, Amat, X, W)); /* w <- b - Ax */
49 PetscCall(VecAYPX(W, -1.0, B));
50 } else {
51 PetscCall(VecCopy(B, W)); /* w <- b (x is 0) */
52 }
53 PetscCall(KSP_PCApply(ksp, W, U)); /* u <- Bw */
54
55 switch (ksp->normtype) {
56 case KSP_NORM_PRECONDITIONED:
57 PetscCall(VecNormBegin(U, NORM_2, &dp)); /* dp <- u'*u = e'*A'*B'*B*A'*e' */
58 PetscCall(PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)U)));
59 PetscCall(KSP_MatMult(ksp, Amat, U, W)); /* w <- Au */
60 PetscCall(VecNormEnd(U, NORM_2, &dp));
61 break;
62 case KSP_NORM_NONE:
63 PetscCall(KSP_MatMult(ksp, Amat, U, W));
64 dp = 0.0;
65 break;
66 default:
67 SETERRQ(PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "%s", KSPNormTypes[ksp->normtype]);
68 }
69 PetscCall(KSPLogResidualHistory(ksp, dp));
70 PetscCall(KSPMonitor(ksp, 0, dp));
71 ksp->rnorm = dp;
72 PetscCall((*ksp->converged)(ksp, 0, dp, &ksp->reason, ksp->cnvP)); /* test for convergence */
73 if (ksp->reason) PetscFunctionReturn(PETSC_SUCCESS);
74
75 i = 0;
76 do {
77 PetscCall(KSP_PCApply(ksp, W, M)); /* m <- Bw */
78
79 if (i > 0 && ksp->normtype == KSP_NORM_PRECONDITIONED) PetscCall(VecNormBegin(U, NORM_2, &dp));
80 PetscCall(VecDotBegin(W, U, &gamma));
81 PetscCall(VecDotBegin(M, W, &delta));
82 PetscCall(PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)U)));
83
84 PetscCall(KSP_MatMult(ksp, Amat, M, N)); /* n <- Am */
85
86 if (i > 0 && ksp->normtype == KSP_NORM_PRECONDITIONED) PetscCall(VecNormEnd(U, NORM_2, &dp));
87 PetscCall(VecDotEnd(W, U, &gamma));
88 PetscCall(VecDotEnd(M, W, &delta));
89
90 if (i > 0) {
91 if (ksp->normtype == KSP_NORM_NONE) dp = 0.0;
92 ksp->rnorm = dp;
93 PetscCall(KSPLogResidualHistory(ksp, dp));
94 PetscCall(KSPMonitor(ksp, i, dp));
95 PetscCall((*ksp->converged)(ksp, i, dp, &ksp->reason, ksp->cnvP));
96 if (ksp->reason) PetscFunctionReturn(PETSC_SUCCESS);
97 }
98
99 if (i == 0) {
100 alpha = gamma / delta;
101 PetscCall(VecCopy(N, Z)); /* z <- n */
102 PetscCall(VecCopy(M, Q)); /* q <- m */
103 PetscCall(VecCopy(U, P)); /* p <- u */
104 } else {
105 beta = gamma / gammaold;
106 alpha = gamma / (delta - beta / alpha * gamma);
107 PetscCall(VecAYPX(Z, beta, N)); /* z <- n + beta * z */
108 PetscCall(VecAYPX(Q, beta, M)); /* q <- m + beta * q */
109 PetscCall(VecAYPX(P, beta, U)); /* p <- u + beta * p */
110 }
111 PetscCall(VecAXPY(X, alpha, P)); /* x <- x + alpha * p */
112 PetscCall(VecAXPY(U, -alpha, Q)); /* u <- u - alpha * q */
113 PetscCall(VecAXPY(W, -alpha, Z)); /* w <- w - alpha * z */
114 gammaold = gamma;
115 i++;
116 ksp->its = i;
117
118 /* if (i%50 == 0) { */
119 /* PetscCall(KSP_MatMult(ksp,Amat,X,W)); /\* w <- b - Ax *\/ */
120 /* PetscCall(VecAYPX(W,-1.0,B)); */
121 /* PetscCall(KSP_PCApply(ksp,W,U)); */
122 /* PetscCall(KSP_MatMult(ksp,Amat,U,W)); */
123 /* } */
124
125 } while (i <= ksp->max_it);
126 if (i >= ksp->max_it) ksp->reason = KSP_DIVERGED_ITS;
127 PetscFunctionReturn(PETSC_SUCCESS);
128 }
129
130 /*MC
131 KSPPIPECR - Pipelined conjugate residual method {cite}`ghyselsvanroose2014`. [](sec_pipelineksp)
132
133 Level: intermediate
134
135 Notes:
136 This method has only a single non-blocking reduction per iteration, compared to 2 for standard `KSPCR`. The
137 non-blocking reduction is overlapped by the matrix-vector product, but not the preconditioner application.
138
139 See also `KSPPIPECG`, where the reduction is overlapped with the matrix-vector product.
140
141 MPI configuration may be necessary for reductions to make asynchronous progress, which is important for performance of pipelined methods.
142 See [](doc_faq_pipelined)
143
144 Contributed by:
145 Pieter Ghysels, Universiteit Antwerpen, Intel Exascience lab Flanders
146
147 .seealso: [](ch_ksp), [](sec_pipelineksp), [](doc_faq_pipelined), `KSPCreate()`, `KSPSetType()`, `KSPPIPECG`, `KSPGROPPCG`, `KSPPGMRES`, `KSPCG`, `KSPCGUseSingleReduction()`
148 M*/
149
KSPCreate_PIPECR(KSP ksp)150 PETSC_EXTERN PetscErrorCode KSPCreate_PIPECR(KSP ksp)
151 {
152 PetscFunctionBegin;
153 PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_PRECONDITIONED, PC_LEFT, 2));
154 PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_NONE, PC_LEFT, 1));
155
156 ksp->ops->setup = KSPSetUp_PIPECR;
157 ksp->ops->solve = KSPSolve_PIPECR;
158 ksp->ops->destroy = KSPDestroyDefault;
159 ksp->ops->view = NULL;
160 ksp->ops->setfromoptions = NULL;
161 ksp->ops->buildsolution = KSPBuildSolutionDefault;
162 ksp->ops->buildresidual = KSPBuildResidualDefault;
163 PetscFunctionReturn(PETSC_SUCCESS);
164 }
165