Actual source code: badbrdn.c

  1: #include <../src/ksp/ksp/utils/lmvm/brdn/brdn.h>

  3: /*
  4:   The solution method is the matrix-free implementation of the inverse Hessian in
  5:   Equation 6 on page 312 of Griewank "Broyden Updating, The Good and The Bad!"
  6:   (http://www.emis.ams.org/journals/DMJDMV/vol-ismp/45_griewank-andreas-broyden.pdf).

  8:   Q[i] = (B_i)^{-1}*S[i] terms are computed ahead of time whenever
  9:   the matrix is updated with a new (S[i], Y[i]) pair. This allows
 10:   repeated calls of MatSolve without incurring redundant computation.

 12:   dX <- J0^{-1} * F

 14:   for i=0,1,2,...,k
 15:     # Q[i] = (B_i)^{-1} * Y[i]
 16:     tau = (Y[i]^T F) / (Y[i]^T Y[i])
 17:     dX <- dX + (tau * (S[i] - Q[i]))
 18:   end
 19:  */

 21: static PetscErrorCode MatSolve_LMVMBadBrdn(Mat B, Vec F, Vec dX)
 22: {
 23:   Mat_LMVM   *lmvm = (Mat_LMVM *)B->data;
 24:   Mat_Brdn   *lbb  = (Mat_Brdn *)lmvm->ctx;
 25:   PetscInt    i, j;
 26:   PetscScalar yjtyi, ytf;

 28:   PetscFunctionBegin;
 29:   VecCheckSameSize(F, 2, dX, 3);
 30:   VecCheckMatCompatible(B, dX, 3, F, 2);

 32:   if (lbb->needQ) {
 33:     /* Pre-compute (Q[i] = (B_i)^{-1} * Y[i]) */
 34:     for (i = 0; i <= lmvm->k; ++i) {
 35:       PetscCall(MatLMVMApplyJ0Inv(B, lmvm->Y[i], lbb->Q[i]));
 36:       for (j = 0; j <= i - 1; ++j) {
 37:         PetscCall(VecDot(lmvm->Y[j], lmvm->Y[i], &yjtyi));
 38:         PetscCall(VecAXPBYPCZ(lbb->Q[i], PetscRealPart(yjtyi) / lbb->yty[j], -PetscRealPart(yjtyi) / lbb->yty[j], 1.0, lmvm->S[j], lbb->Q[j]));
 39:       }
 40:     }
 41:     lbb->needQ = PETSC_FALSE;
 42:   }

 44:   PetscCall(MatLMVMApplyJ0Inv(B, F, dX));
 45:   for (i = 0; i <= lmvm->k; ++i) {
 46:     PetscCall(VecDot(lmvm->Y[i], F, &ytf));
 47:     PetscCall(VecAXPBYPCZ(dX, PetscRealPart(ytf) / lbb->yty[i], -PetscRealPart(ytf) / lbb->yty[i], 1.0, lmvm->S[i], lbb->Q[i]));
 48:   }
 49:   PetscFunctionReturn(PETSC_SUCCESS);
 50: }

 52: /*
 53:   The forward product is the matrix-free implementation of the direct update in
 54:   Equation 6 on page 302 of Griewank "Broyden Updating, The Good and The Bad!"
 55:   (http://www.emis.ams.org/journals/DMJDMV/vol-ismp/45_griewank-andreas-broyden.pdf).

 57:   P[i] = (B_i)*S[i] terms are computed ahead of time whenever
 58:   the matrix is updated with a new (S[i], Y[i]) pair. This allows
 59:   repeated calls of MatMult inside KSP solvers without unnecessarily
 60:   recomputing P[i] terms in expensive nested-loops.

 62:   Z <- J0 * X

 64:   for i=0,1,2,...,k
 65:     # P[i] = B_i * S[i]
 66:     tau = (Y[i]^T X) / (Y[i]^T S[i])
 67:     dX <- dX + (tau * (Y[i] - P[i]))
 68:   end
 69:  */

 71: static PetscErrorCode MatMult_LMVMBadBrdn(Mat B, Vec X, Vec Z)
 72: {
 73:   Mat_LMVM   *lmvm = (Mat_LMVM *)B->data;
 74:   Mat_Brdn   *lbb  = (Mat_Brdn *)lmvm->ctx;
 75:   PetscInt    i, j;
 76:   PetscScalar yjtsi, ytx;

 78:   PetscFunctionBegin;
 79:   VecCheckSameSize(X, 2, Z, 3);
 80:   VecCheckMatCompatible(B, X, 2, Z, 3);

 82:   if (lbb->needP) {
 83:     /* Pre-compute (P[i] = (B_i) * S[i]) */
 84:     for (i = 0; i <= lmvm->k; ++i) {
 85:       PetscCall(MatLMVMApplyJ0Fwd(B, lmvm->S[i], lbb->P[i]));
 86:       for (j = 0; j <= i - 1; ++j) {
 87:         PetscCall(VecDot(lmvm->Y[j], lmvm->S[i], &yjtsi));
 88:         PetscCall(VecAXPBYPCZ(lbb->P[i], PetscRealPart(yjtsi) / lbb->yts[j], -PetscRealPart(yjtsi) / lbb->yts[j], 1.0, lmvm->Y[j], lbb->P[j]));
 89:       }
 90:     }
 91:     lbb->needP = PETSC_FALSE;
 92:   }

 94:   PetscCall(MatLMVMApplyJ0Fwd(B, X, Z));
 95:   for (i = 0; i <= lmvm->k; ++i) {
 96:     PetscCall(VecDot(lmvm->Y[i], X, &ytx));
 97:     PetscCall(VecAXPBYPCZ(Z, PetscRealPart(ytx) / lbb->yts[i], -PetscRealPart(ytx) / lbb->yts[i], 1.0, lmvm->Y[i], lbb->P[i]));
 98:   }
 99:   PetscFunctionReturn(PETSC_SUCCESS);
100: }

102: static PetscErrorCode MatUpdate_LMVMBadBrdn(Mat B, Vec X, Vec F)
103: {
104:   Mat_LMVM   *lmvm = (Mat_LMVM *)B->data;
105:   Mat_Brdn   *lbb  = (Mat_Brdn *)lmvm->ctx;
106:   PetscInt    old_k, i;
107:   PetscScalar yty, yts;

109:   PetscFunctionBegin;
110:   if (!lmvm->m) PetscFunctionReturn(PETSC_SUCCESS);
111:   if (lmvm->prev_set) {
112:     /* Compute the new (S = X - Xprev) and (Y = F - Fprev) vectors */
113:     PetscCall(VecAYPX(lmvm->Xprev, -1.0, X));
114:     PetscCall(VecAYPX(lmvm->Fprev, -1.0, F));
115:     /* Accept the update */
116:     lbb->needP = lbb->needQ = PETSC_TRUE;
117:     old_k                   = lmvm->k;
118:     PetscCall(MatUpdateKernel_LMVM(B, lmvm->Xprev, lmvm->Fprev));
119:     /* If we hit the memory limit, shift the yty and yts arrays */
120:     if (old_k == lmvm->k) {
121:       for (i = 0; i <= lmvm->k - 1; ++i) {
122:         lbb->yty[i] = lbb->yty[i + 1];
123:         lbb->yts[i] = lbb->yts[i + 1];
124:       }
125:     }
126:     /* Accumulate the latest yTy and yTs dot products */
127:     PetscCall(VecDotBegin(lmvm->Y[lmvm->k], lmvm->Y[lmvm->k], &yty));
128:     PetscCall(VecDotBegin(lmvm->Y[lmvm->k], lmvm->S[lmvm->k], &yts));
129:     PetscCall(VecDotEnd(lmvm->Y[lmvm->k], lmvm->Y[lmvm->k], &yty));
130:     PetscCall(VecDotEnd(lmvm->Y[lmvm->k], lmvm->S[lmvm->k], &yts));
131:     lbb->yty[lmvm->k] = PetscRealPart(yty);
132:     lbb->yts[lmvm->k] = PetscRealPart(yts);
133:   }
134:   /* Save the solution and function to be used in the next update */
135:   PetscCall(VecCopy(X, lmvm->Xprev));
136:   PetscCall(VecCopy(F, lmvm->Fprev));
137:   lmvm->prev_set = PETSC_TRUE;
138:   PetscFunctionReturn(PETSC_SUCCESS);
139: }

141: static PetscErrorCode MatCopy_LMVMBadBrdn(Mat B, Mat M, MatStructure str)
142: {
143:   Mat_LMVM *bdata = (Mat_LMVM *)B->data;
144:   Mat_Brdn *bctx  = (Mat_Brdn *)bdata->ctx;
145:   Mat_LMVM *mdata = (Mat_LMVM *)M->data;
146:   Mat_Brdn *mctx  = (Mat_Brdn *)mdata->ctx;
147:   PetscInt  i;

149:   PetscFunctionBegin;
150:   mctx->needP = bctx->needP;
151:   mctx->needQ = bctx->needQ;
152:   for (i = 0; i <= bdata->k; ++i) {
153:     mctx->yty[i] = bctx->yty[i];
154:     mctx->yts[i] = bctx->yts[i];
155:     PetscCall(VecCopy(bctx->P[i], mctx->P[i]));
156:     PetscCall(VecCopy(bctx->Q[i], mctx->Q[i]));
157:   }
158:   PetscFunctionReturn(PETSC_SUCCESS);
159: }

161: static PetscErrorCode MatReset_LMVMBadBrdn(Mat B, PetscBool destructive)
162: {
163:   Mat_LMVM *lmvm = (Mat_LMVM *)B->data;
164:   Mat_Brdn *lbb  = (Mat_Brdn *)lmvm->ctx;

166:   PetscFunctionBegin;
167:   lbb->needP = lbb->needQ = PETSC_TRUE;
168:   if (destructive && lbb->allocated) {
169:     PetscCall(PetscFree2(lbb->yty, lbb->yts));
170:     PetscCall(VecDestroyVecs(lmvm->m, &lbb->P));
171:     PetscCall(VecDestroyVecs(lmvm->m, &lbb->Q));
172:     lbb->allocated = PETSC_FALSE;
173:   }
174:   PetscCall(MatReset_LMVM(B, destructive));
175:   PetscFunctionReturn(PETSC_SUCCESS);
176: }

178: static PetscErrorCode MatAllocate_LMVMBadBrdn(Mat B, Vec X, Vec F)
179: {
180:   Mat_LMVM *lmvm = (Mat_LMVM *)B->data;
181:   Mat_Brdn *lbb  = (Mat_Brdn *)lmvm->ctx;

183:   PetscFunctionBegin;
184:   PetscCall(MatAllocate_LMVM(B, X, F));
185:   if (!lbb->allocated) {
186:     PetscCall(PetscMalloc2(lmvm->m, &lbb->yty, lmvm->m, &lbb->yts));
187:     if (lmvm->m > 0) {
188:       PetscCall(VecDuplicateVecs(X, lmvm->m, &lbb->P));
189:       PetscCall(VecDuplicateVecs(X, lmvm->m, &lbb->Q));
190:     }
191:     lbb->allocated = PETSC_TRUE;
192:   }
193:   PetscFunctionReturn(PETSC_SUCCESS);
194: }

196: static PetscErrorCode MatDestroy_LMVMBadBrdn(Mat B)
197: {
198:   Mat_LMVM *lmvm = (Mat_LMVM *)B->data;
199:   Mat_Brdn *lbb  = (Mat_Brdn *)lmvm->ctx;

201:   PetscFunctionBegin;
202:   if (lbb->allocated) {
203:     PetscCall(PetscFree2(lbb->yty, lbb->yts));
204:     PetscCall(VecDestroyVecs(lmvm->m, &lbb->P));
205:     PetscCall(VecDestroyVecs(lmvm->m, &lbb->Q));
206:     lbb->allocated = PETSC_FALSE;
207:   }
208:   PetscCall(PetscFree(lmvm->ctx));
209:   PetscCall(MatDestroy_LMVM(B));
210:   PetscFunctionReturn(PETSC_SUCCESS);
211: }

213: static PetscErrorCode MatSetUp_LMVMBadBrdn(Mat B)
214: {
215:   Mat_LMVM *lmvm = (Mat_LMVM *)B->data;
216:   Mat_Brdn *lbb  = (Mat_Brdn *)lmvm->ctx;

218:   PetscFunctionBegin;
219:   PetscCall(MatSetUp_LMVM(B));
220:   if (!lbb->allocated) {
221:     PetscCall(PetscMalloc2(lmvm->m, &lbb->yty, lmvm->m, &lbb->yts));
222:     if (lmvm->m > 0) {
223:       PetscCall(VecDuplicateVecs(lmvm->Xprev, lmvm->m, &lbb->P));
224:       PetscCall(VecDuplicateVecs(lmvm->Xprev, lmvm->m, &lbb->Q));
225:     }
226:     lbb->allocated = PETSC_TRUE;
227:   }
228:   PetscFunctionReturn(PETSC_SUCCESS);
229: }

231: PetscErrorCode MatCreate_LMVMBadBrdn(Mat B)
232: {
233:   Mat_LMVM *lmvm;
234:   Mat_Brdn *lbb;

236:   PetscFunctionBegin;
237:   PetscCall(MatCreate_LMVM(B));
238:   PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATLMVMBADBROYDEN));
239:   B->ops->setup   = MatSetUp_LMVMBadBrdn;
240:   B->ops->destroy = MatDestroy_LMVMBadBrdn;
241:   B->ops->solve   = MatSolve_LMVMBadBrdn;

243:   lmvm                = (Mat_LMVM *)B->data;
244:   lmvm->square        = PETSC_TRUE;
245:   lmvm->ops->allocate = MatAllocate_LMVMBadBrdn;
246:   lmvm->ops->reset    = MatReset_LMVMBadBrdn;
247:   lmvm->ops->mult     = MatMult_LMVMBadBrdn;
248:   lmvm->ops->solve    = MatSolve_LMVMBadBrdn;
249:   lmvm->ops->update   = MatUpdate_LMVMBadBrdn;
250:   lmvm->ops->copy     = MatCopy_LMVMBadBrdn;

252:   PetscCall(PetscNew(&lbb));
253:   lmvm->ctx      = (void *)lbb;
254:   lbb->allocated = PETSC_FALSE;
255:   lbb->needP = lbb->needQ = PETSC_TRUE;
256:   PetscFunctionReturn(PETSC_SUCCESS);
257: }

259: /*@
260:   MatCreateLMVMBadBroyden - Creates a limited-memory modified (aka "bad") Broyden-type
261:   approximation matrix used for a Jacobian. L-BadBrdn is not guaranteed to be
262:   symmetric or positive-definite.

264:   To use the L-BadBrdn matrix with other vector types, the matrix must be
265:   created using `MatCreate()` and `MatSetType()`, followed by `MatLMVMAllocate()`.
266:   This ensures that the internal storage and work vectors are duplicated from the
267:   correct type of vector.

269:   Collective

271:   Input Parameters:
272: + comm - MPI communicator
273: . n    - number of local rows for storage vectors
274: - N    - global size of the storage vectors

276:   Output Parameter:
277: . B - the matrix

279:   Options Database Keys:
280: + -mat_lmvm_scale_type - (developer) type of scaling applied to J0 (none, scalar, diagonal)
281: . -mat_lmvm_theta      - (developer) convex ratio between BFGS and DFP components of the diagonal J0 scaling
282: . -mat_lmvm_rho        - (developer) update limiter for the J0 scaling
283: . -mat_lmvm_alpha      - (developer) coefficient factor for the quadratic subproblem in J0 scaling
284: . -mat_lmvm_beta       - (developer) exponential factor for the diagonal J0 scaling
285: - -mat_lmvm_sigma_hist - (developer) number of past updates to use in J0 scaling

287:   Level: intermediate

289:   Note:
290:   It is recommended that one use the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`
291:   paradigm instead of this routine directly.

293: .seealso: [](ch_ksp), `MatCreate()`, `MATLMVM`, `MATLMVMBADBRDN`, `MatCreateLMVMDFP()`, `MatCreateLMVMSR1()`,
294:           `MatCreateLMVMBFGS()`, `MatCreateLMVMBrdn()`, `MatCreateLMVMSymBrdn()`
295: @*/
296: PetscErrorCode MatCreateLMVMBadBroyden(MPI_Comm comm, PetscInt n, PetscInt N, Mat *B)
297: {
298:   PetscFunctionBegin;
299:   PetscCall(KSPInitializePackage());
300:   PetscCall(MatCreate(comm, B));
301:   PetscCall(MatSetSizes(*B, n, n, N, N));
302:   PetscCall(MatSetType(*B, MATLMVMBADBROYDEN));
303:   PetscCall(MatSetUp(*B));
304:   PetscFunctionReturn(PETSC_SUCCESS);
305: }