Actual source code: pipecg.c
1: #include <petsc/private/kspimpl.h>
3: /*
4: KSPSetUp_PIPECG - Sets up the workspace needed by the PIPECG method.
6: This is called once, usually automatically by KSPSolve() or KSPSetUp()
7: but can be called directly by KSPSetUp()
8: */
9: static PetscErrorCode KSPSetUp_PIPECG(KSP ksp)
10: {
11: PetscFunctionBegin;
12: /* get work vectors needed by PIPECG */
13: PetscCall(KSPSetWorkVecs(ksp, 9));
14: PetscFunctionReturn(PETSC_SUCCESS);
15: }
17: /*
18: KSPSolve_PIPECG - This routine actually applies the pipelined conjugate gradient method
19: */
20: static PetscErrorCode KSPSolve_PIPECG(KSP ksp)
21: {
22: PetscInt i;
23: PetscScalar alpha = 0.0, beta = 0.0, gamma = 0.0, gammaold = 0.0, delta = 0.0;
24: PetscReal dp = 0.0;
25: Vec X, B, Z, P, W, Q, U, M, N, R, S;
26: Mat Amat, Pmat;
28: PetscFunctionBegin;
29: X = ksp->vec_sol;
30: B = ksp->vec_rhs;
31: R = ksp->work[0];
32: Z = ksp->work[1];
33: P = ksp->work[2];
34: N = ksp->work[3];
35: W = ksp->work[4];
36: Q = ksp->work[5];
37: U = ksp->work[6];
38: M = ksp->work[7];
39: S = ksp->work[8];
41: PetscCall(PCGetOperators(ksp->pc, &Amat, &Pmat));
43: ksp->its = 0;
44: if (!ksp->guess_zero) {
45: PetscCall(KSP_MatMult(ksp, Amat, X, R)); /* r <- b - Ax */
46: PetscCall(VecAYPX(R, -1.0, B));
47: } else {
48: PetscCall(VecCopy(B, R)); /* r <- b (x is 0) */
49: }
51: PetscCall(KSP_PCApply(ksp, R, U)); /* u <- Br */
53: switch (ksp->normtype) {
54: case KSP_NORM_PRECONDITIONED:
55: PetscCall(VecNormBegin(U, NORM_2, &dp)); /* dp <- u'*u = e'*A'*B'*B*A'*e' */
56: PetscCall(PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)U)));
57: PetscCall(KSP_MatMult(ksp, Amat, U, W)); /* w <- Au */
58: PetscCall(VecNormEnd(U, NORM_2, &dp));
59: break;
60: case KSP_NORM_UNPRECONDITIONED:
61: PetscCall(VecNormBegin(R, NORM_2, &dp)); /* dp <- r'*r = e'*A'*A*e */
62: PetscCall(PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)R)));
63: PetscCall(KSP_MatMult(ksp, Amat, U, W)); /* w <- Au */
64: PetscCall(VecNormEnd(R, NORM_2, &dp));
65: break;
66: case KSP_NORM_NATURAL:
67: PetscCall(VecDotBegin(R, U, &gamma)); /* gamma <- u'*r */
68: PetscCall(PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)R)));
69: PetscCall(KSP_MatMult(ksp, Amat, U, W)); /* w <- Au */
70: PetscCall(VecDotEnd(R, U, &gamma));
71: KSPCheckDot(ksp, gamma);
72: dp = PetscSqrtReal(PetscAbsScalar(gamma)); /* dp <- r'*u = r'*B*r = e'*A'*B*A*e */
73: break;
74: case KSP_NORM_NONE:
75: PetscCall(KSP_MatMult(ksp, Amat, U, W));
76: dp = 0.0;
77: break;
78: default:
79: SETERRQ(PetscObjectComm((PetscObject)ksp), PETSC_ERR_SUP, "%s", KSPNormTypes[ksp->normtype]);
80: }
81: PetscCall(KSPLogResidualHistory(ksp, dp));
82: PetscCall(KSPMonitor(ksp, 0, dp));
83: ksp->rnorm = dp;
84: PetscCall((*ksp->converged)(ksp, 0, dp, &ksp->reason, ksp->cnvP)); /* test for convergence */
85: if (ksp->reason) PetscFunctionReturn(PETSC_SUCCESS);
87: i = 0;
88: do {
89: if (i > 0 && ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
90: PetscCall(VecNormBegin(R, NORM_2, &dp));
91: } else if (i > 0 && ksp->normtype == KSP_NORM_PRECONDITIONED) {
92: PetscCall(VecNormBegin(U, NORM_2, &dp));
93: }
94: if (!(i == 0 && ksp->normtype == KSP_NORM_NATURAL)) PetscCall(VecDotBegin(R, U, &gamma));
95: PetscCall(VecDotBegin(W, U, &delta));
96: PetscCall(PetscCommSplitReductionBegin(PetscObjectComm((PetscObject)R)));
98: PetscCall(KSP_PCApply(ksp, W, M)); /* m <- Bw */
99: PetscCall(KSP_MatMult(ksp, Amat, M, N)); /* n <- Am */
101: if (i > 0 && ksp->normtype == KSP_NORM_UNPRECONDITIONED) {
102: PetscCall(VecNormEnd(R, NORM_2, &dp));
103: } else if (i > 0 && ksp->normtype == KSP_NORM_PRECONDITIONED) {
104: PetscCall(VecNormEnd(U, NORM_2, &dp));
105: }
106: if (!(i == 0 && ksp->normtype == KSP_NORM_NATURAL)) PetscCall(VecDotEnd(R, U, &gamma));
107: PetscCall(VecDotEnd(W, U, &delta));
109: if (i > 0) {
110: if (ksp->normtype == KSP_NORM_NATURAL) dp = PetscSqrtReal(PetscAbsScalar(gamma));
111: else if (ksp->normtype == KSP_NORM_NONE) dp = 0.0;
113: ksp->rnorm = dp;
114: PetscCall(KSPLogResidualHistory(ksp, dp));
115: PetscCall(KSPMonitor(ksp, i, dp));
116: PetscCall((*ksp->converged)(ksp, i, dp, &ksp->reason, ksp->cnvP));
117: if (ksp->reason) PetscFunctionReturn(PETSC_SUCCESS);
118: }
120: if (i == 0) {
121: alpha = gamma / delta;
122: PetscCall(VecCopy(N, Z)); /* z <- n */
123: PetscCall(VecCopy(M, Q)); /* q <- m */
124: PetscCall(VecCopy(U, P)); /* p <- u */
125: PetscCall(VecCopy(W, S)); /* s <- w */
126: } else {
127: beta = gamma / gammaold;
128: alpha = gamma / (delta - beta / alpha * gamma);
129: PetscCall(VecAYPX(Z, beta, N)); /* z <- n + beta * z */
130: PetscCall(VecAYPX(Q, beta, M)); /* q <- m + beta * q */
131: PetscCall(VecAYPX(P, beta, U)); /* p <- u + beta * p */
132: PetscCall(VecAYPX(S, beta, W)); /* s <- w + beta * s */
133: }
134: PetscCall(VecAXPY(X, alpha, P)); /* x <- x + alpha * p */
135: PetscCall(VecAXPY(U, -alpha, Q)); /* u <- u - alpha * q */
136: PetscCall(VecAXPY(W, -alpha, Z)); /* w <- w - alpha * z */
137: PetscCall(VecAXPY(R, -alpha, S)); /* r <- r - alpha * s */
138: gammaold = gamma;
139: i++;
140: ksp->its = i;
142: /* if (i%50 == 0) { */
143: /* PetscCall(KSP_MatMult(ksp,Amat,X,R)); /\* w <- b - Ax *\/ */
144: /* PetscCall(VecAYPX(R,-1.0,B)); */
145: /* PetscCall(KSP_PCApply(ksp,R,U)); */
146: /* PetscCall(KSP_MatMult(ksp,Amat,U,W)); */
147: /* } */
149: } while (i <= ksp->max_it);
150: if (!ksp->reason) ksp->reason = KSP_DIVERGED_ITS;
151: PetscFunctionReturn(PETSC_SUCCESS);
152: }
154: PETSC_INTERN PetscErrorCode KSPBuildResidual_CG(KSP, Vec, Vec, Vec *);
156: /*MC
157: KSPPIPECG - Pipelined conjugate gradient method {cite}`ghyselsvanroose2014`. [](sec_pipelineksp)
159: Level: intermediate
161: Notes:
162: This method has only a single non-blocking reduction per iteration, compared to 2 blocking for standard `KSPPCG`. The
163: non-blocking reduction is overlapped by the matrix-vector product and preconditioner application.
165: See also `KSPPIPECR`, where the reduction is only overlapped with the matrix-vector product and `KSPGROPPCG`
167: MPI configuration may be necessary for reductions to make asynchronous progress, which is important for performance of pipelined methods.
168: See [](doc_faq_pipelined)
170: Contributed by:
171: Pieter Ghysels, Universiteit Antwerpen, Intel Exascience lab Flanders
173: .seealso: [](ch_ksp), [](doc_faq_pipelined), [](sec_pipelineksp), `KSPCreate()`, `KSPSetType()`, `KSPPIPECG2`, `KSPPIPECR`, `KSPGROPPCG`, `KSPPGMRES`, `KSPCG`, `KSPCGUseSingleReduction()`
174: M*/
175: PETSC_EXTERN PetscErrorCode KSPCreate_PIPECG(KSP ksp)
176: {
177: PetscFunctionBegin;
178: PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_UNPRECONDITIONED, PC_LEFT, 2));
179: PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_PRECONDITIONED, PC_LEFT, 2));
180: PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_NATURAL, PC_LEFT, 2));
181: PetscCall(KSPSetSupportedNorm(ksp, KSP_NORM_NONE, PC_LEFT, 1));
183: ksp->ops->setup = KSPSetUp_PIPECG;
184: ksp->ops->solve = KSPSolve_PIPECG;
185: ksp->ops->destroy = KSPDestroyDefault;
186: ksp->ops->view = NULL;
187: ksp->ops->setfromoptions = NULL;
188: ksp->ops->buildsolution = KSPBuildSolutionDefault;
189: ksp->ops->buildresidual = KSPBuildResidual_CG;
190: PetscFunctionReturn(PETSC_SUCCESS);
191: }