Actual source code: adr_ex5adj_mf.cxx

  1: static char help[] = "Demonstrates automatic, matrix-free Jacobian generation using ADOL-C for a time-dependent PDE in 2d, solved using implicit timestepping.\n";

  3: /*
  4:    REQUIRES configuration of PETSc with option --download-adolc.

  6:    For documentation on ADOL-C, see
  7:      $PETSC_ARCH/externalpackages/ADOL-C-2.6.0/ADOL-C/doc/adolc-manual.pdf
  8: */
  9: /* ------------------------------------------------------------------------
 10:   See ../advection-diffusion-reaction/ex5 for a description of the problem
 11:   ------------------------------------------------------------------------- */

 13: #include <petscdmda.h>
 14: #include <petscts.h>
 15: #include "adolc-utils/init.cxx"
 16: #include "adolc-utils/matfree.cxx"
 17: #include <adolc/adolc.h>

 19: /* (Passive) field for the two variables */
 20: typedef struct {
 21:   PetscScalar u, v;
 22: } Field;

 24: /* Active field for the two variables */
 25: typedef struct {
 26:   adouble u, v;
 27: } AField;

 29: /* Application context */
 30: typedef struct {
 31:   PetscReal D1, D2, gamma, kappa;
 32:   AField  **u_a, **f_a;
 33:   AdolcCtx *adctx; /* Automatic differentation support */
 34: } AppCtx;

 36: extern PetscErrorCode InitialConditions(DM da, Vec U);
 37: extern PetscErrorCode InitializeLambda(DM da, Vec lambda, PetscReal x, PetscReal y);
 38: extern PetscErrorCode IFunctionLocalPassive(DMDALocalInfo *info, PetscReal t, Field **u, Field **udot, Field **f, void *ptr);
 39: extern PetscErrorCode IFunctionActive(TS ts, PetscReal ftime, Vec U, Vec Udot, Vec F, void *ptr);
 40: extern PetscErrorCode IJacobianMatFree(TS ts, PetscReal t, Vec X, Vec Xdot, PetscReal a, Mat A_shell, Mat B, void *ctx);

 42: int main(int argc, char **argv)
 43: {
 44:   TS          ts;   /* ODE integrator */
 45:   Vec         x, r; /* solution, residual */
 46:   DM          da;
 47:   AppCtx      appctx; /* Application context */
 48:   AdolcMatCtx matctx; /* Matrix (free) context */
 49:   Vec         lambda[1];
 50:   PetscBool   forwardonly = PETSC_FALSE;
 51:   Mat         A; /* (Matrix free) Jacobian matrix */
 52:   PetscInt    gxm, gym;

 54:   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
 55:      Initialize program
 56:      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
 57:   PetscFunctionBeginUser;
 58:   PetscCall(PetscInitialize(&argc, &argv, NULL, help));
 59:   PetscCall(PetscOptionsGetBool(NULL, NULL, "-forwardonly", &forwardonly, NULL));
 60:   appctx.D1    = 8.0e-5;
 61:   appctx.D2    = 4.0e-5;
 62:   appctx.gamma = .024;
 63:   appctx.kappa = .06;
 64:   PetscCall(PetscLogEventRegister("df/dx forward", MAT_CLASSID, &matctx.event1));
 65:   PetscCall(PetscLogEventRegister("df/d(xdot) forward", MAT_CLASSID, &matctx.event2));
 66:   PetscCall(PetscLogEventRegister("df/dx reverse", MAT_CLASSID, &matctx.event3));
 67:   PetscCall(PetscLogEventRegister("df/d(xdot) reverse", MAT_CLASSID, &matctx.event4));

 69:   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
 70:      Create distributed array (DMDA) to manage parallel grid and vectors
 71:   - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
 72:   PetscCall(DMDACreate2d(PETSC_COMM_WORLD, DM_BOUNDARY_PERIODIC, DM_BOUNDARY_PERIODIC, DMDA_STENCIL_STAR, 65, 65, PETSC_DECIDE, PETSC_DECIDE, 2, 1, NULL, NULL, &da));
 73:   PetscCall(DMSetFromOptions(da));
 74:   PetscCall(DMSetUp(da));
 75:   PetscCall(DMDASetFieldName(da, 0, "u"));
 76:   PetscCall(DMDASetFieldName(da, 1, "v"));

 78:   /*  - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
 79:      Extract global vectors from DMDA; then duplicate for remaining
 80:      vectors that are the same types
 81:    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
 82:   PetscCall(DMCreateGlobalVector(da, &x));
 83:   PetscCall(VecDuplicate(x, &r));

 85:   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
 86:     Create matrix-free context and specify usage of PETSc-ADOL-C drivers
 87:     - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
 88:   PetscCall(DMSetMatType(da, MATSHELL));
 89:   PetscCall(DMCreateMatrix(da, &A));
 90:   PetscCall(MatShellSetContext(A, &matctx));
 91:   PetscCall(MatShellSetOperation(A, MATOP_MULT, (void (*)(void))PetscAdolcIJacobianVectorProductIDMass));
 92:   PetscCall(MatShellSetOperation(A, MATOP_MULT_TRANSPOSE, (void (*)(void))PetscAdolcIJacobianTransposeVectorProductIDMass));
 93:   PetscCall(VecDuplicate(x, &matctx.X));
 94:   PetscCall(VecDuplicate(x, &matctx.Xdot));
 95:   PetscCall(DMGetLocalVector(da, &matctx.localX0));

 97:   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
 98:      Create timestepping solver context
 99:      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
100:   PetscCall(TSCreate(PETSC_COMM_WORLD, &ts));
101:   PetscCall(TSSetType(ts, TSCN));
102:   PetscCall(TSSetDM(ts, da));
103:   PetscCall(TSSetProblemType(ts, TS_NONLINEAR));
104:   PetscCall(DMDATSSetIFunctionLocal(da, INSERT_VALUES, (DMDATSIFunctionLocalFn *)IFunctionLocalPassive, &appctx));

106:   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
107:     Some data required for matrix-free context
108:      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
109:   PetscCall(DMDAGetGhostCorners(da, NULL, NULL, NULL, &gxm, &gym, NULL));
110:   matctx.m    = 2 * gxm * gym;
111:   matctx.n    = 2 * gxm * gym; /* Number of dependent and independent variables */
112:   matctx.flg  = PETSC_FALSE;   /* Flag for reverse mode */
113:   matctx.tag1 = 1;             /* Tape identifier */

115:   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
116:      Trace function just once
117:    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
118:   PetscCall(PetscNew(&appctx.adctx));
119:   PetscCall(IFunctionActive(ts, 1., x, matctx.Xdot, r, &appctx));
120:   PetscCall(PetscFree(appctx.adctx));

122:   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
123:      Set Jacobian. In this case, IJacobian simply acts to pass context
124:      information to the matrix-free Jacobian vector product.
125:    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
126:   PetscCall(TSSetIJacobian(ts, A, A, IJacobianMatFree, &appctx));

128:   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
129:      Set initial conditions
130:    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
131:   PetscCall(InitialConditions(da, x));
132:   PetscCall(TSSetSolution(ts, x));

134:   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
135:     Have the TS save its trajectory so that TSAdjointSolve() may be used
136:     and set solver options
137:    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
138:   if (!forwardonly) {
139:     PetscCall(TSSetSaveTrajectory(ts));
140:     PetscCall(TSSetMaxTime(ts, 200.0));
141:     PetscCall(TSSetTimeStep(ts, 0.5));
142:   } else {
143:     PetscCall(TSSetMaxTime(ts, 2000.0));
144:     PetscCall(TSSetTimeStep(ts, 10));
145:   }
146:   PetscCall(TSSetExactFinalTime(ts, TS_EXACTFINALTIME_STEPOVER));
147:   PetscCall(TSSetFromOptions(ts));

149:   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
150:      Solve ODE system
151:      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
152:   PetscCall(TSSolve(ts, x));
153:   if (!forwardonly) {
154:     /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
155:        Start the Adjoint model
156:        - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
157:     PetscCall(VecDuplicate(x, &lambda[0]));
158:     /*   Reset initial conditions for the adjoint integration */
159:     PetscCall(InitializeLambda(da, lambda[0], 0.5, 0.5));
160:     PetscCall(TSSetCostGradients(ts, 1, lambda, NULL));
161:     PetscCall(TSAdjointSolve(ts));
162:     PetscCall(VecDestroy(&lambda[0]));
163:   }

165:   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
166:      Free work space.  All PETSc objects should be destroyed when they
167:      are no longer needed.
168:    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
169:   PetscCall(DMRestoreLocalVector(da, &matctx.localX0));
170:   PetscCall(VecDestroy(&r));
171:   PetscCall(VecDestroy(&matctx.X));
172:   PetscCall(VecDestroy(&matctx.Xdot));
173:   PetscCall(MatDestroy(&A));
174:   PetscCall(VecDestroy(&x));
175:   PetscCall(TSDestroy(&ts));
176:   PetscCall(DMDestroy(&da));

178:   PetscCall(PetscFinalize());
179:   return 0;
180: }

182: PetscErrorCode InitialConditions(DM da, Vec U)
183: {
184:   PetscInt  i, j, xs, ys, xm, ym, Mx, My;
185:   Field   **u;
186:   PetscReal hx, hy, x, y;

188:   PetscFunctionBegin;
189:   PetscCall(DMDAGetInfo(da, PETSC_IGNORE, &Mx, &My, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE));

191:   hx = 2.5 / (PetscReal)Mx;
192:   hy = 2.5 / (PetscReal)My;

194:   /*
195:      Get pointers to vector data
196:   */
197:   PetscCall(DMDAVecGetArray(da, U, &u));

199:   /*
200:      Get local grid boundaries
201:   */
202:   PetscCall(DMDAGetCorners(da, &xs, &ys, NULL, &xm, &ym, NULL));

204:   /*
205:      Compute function over the locally owned part of the grid
206:   */
207:   for (j = ys; j < ys + ym; j++) {
208:     y = j * hy;
209:     for (i = xs; i < xs + xm; i++) {
210:       x = i * hx;
211:       if (PetscApproximateGTE(x, 1.0) && PetscApproximateLTE(x, 1.5) && PetscApproximateGTE(y, 1.0) && PetscApproximateLTE(y, 1.5))
212:         u[j][i].v = PetscPowReal(PetscSinReal(4.0 * PETSC_PI * x), 2.0) * PetscPowReal(PetscSinReal(4.0 * PETSC_PI * y), 2.0) / 4.0;
213:       else u[j][i].v = 0.0;

215:       u[j][i].u = 1.0 - 2.0 * u[j][i].v;
216:     }
217:   }

219:   /*
220:      Restore vectors
221:   */
222:   PetscCall(DMDAVecRestoreArray(da, U, &u));
223:   PetscFunctionReturn(PETSC_SUCCESS);
224: }

226: PetscErrorCode InitializeLambda(DM da, Vec lambda, PetscReal x, PetscReal y)
227: {
228:   PetscInt i, j, Mx, My, xs, ys, xm, ym;
229:   Field  **l;

231:   PetscFunctionBegin;
232:   PetscCall(DMDAGetInfo(da, PETSC_IGNORE, &Mx, &My, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE, PETSC_IGNORE));
233:   /* locate the global i index for x and j index for y */
234:   i = (PetscInt)(x * (Mx - 1));
235:   j = (PetscInt)(y * (My - 1));
236:   PetscCall(DMDAGetCorners(da, &xs, &ys, NULL, &xm, &ym, NULL));

238:   if (xs <= i && i < xs + xm && ys <= j && j < ys + ym) {
239:     /* the i,j vertex is on this process */
240:     PetscCall(DMDAVecGetArray(da, lambda, &l));
241:     l[j][i].u = 1.0;
242:     l[j][i].v = 1.0;
243:     PetscCall(DMDAVecRestoreArray(da, lambda, &l));
244:   }
245:   PetscFunctionReturn(PETSC_SUCCESS);
246: }

248: PetscErrorCode IFunctionLocalPassive(DMDALocalInfo *info, PetscReal t, Field **u, Field **udot, Field **f, void *ptr)
249: {
250:   AppCtx     *appctx = (AppCtx *)ptr;
251:   PetscInt    i, j, xs, ys, xm, ym;
252:   PetscReal   hx, hy, sx, sy;
253:   PetscScalar uc, uxx, uyy, vc, vxx, vyy;

255:   PetscFunctionBegin;
256:   hx = 2.50 / (PetscReal)info->mx;
257:   sx = 1.0 / (hx * hx);
258:   hy = 2.50 / (PetscReal)info->my;
259:   sy = 1.0 / (hy * hy);

261:   /* Get local grid boundaries */
262:   xs = info->xs;
263:   xm = info->xm;
264:   ys = info->ys;
265:   ym = info->ym;

267:   /* Compute function over the locally owned part of the grid */
268:   for (j = ys; j < ys + ym; j++) {
269:     for (i = xs; i < xs + xm; i++) {
270:       uc        = u[j][i].u;
271:       uxx       = (-2.0 * uc + u[j][i - 1].u + u[j][i + 1].u) * sx;
272:       uyy       = (-2.0 * uc + u[j - 1][i].u + u[j + 1][i].u) * sy;
273:       vc        = u[j][i].v;
274:       vxx       = (-2.0 * vc + u[j][i - 1].v + u[j][i + 1].v) * sx;
275:       vyy       = (-2.0 * vc + u[j - 1][i].v + u[j + 1][i].v) * sy;
276:       f[j][i].u = udot[j][i].u - appctx->D1 * (uxx + uyy) + uc * vc * vc - appctx->gamma * (1.0 - uc);
277:       f[j][i].v = udot[j][i].v - appctx->D2 * (vxx + vyy) - uc * vc * vc + (appctx->gamma + appctx->kappa) * vc;
278:     }
279:   }
280:   PetscCall(PetscLogFlops(16.0 * xm * ym));
281:   PetscFunctionReturn(PETSC_SUCCESS);
282: }

284: PetscErrorCode IFunctionActive(TS ts, PetscReal ftime, Vec U, Vec Udot, Vec F, void *ptr)
285: {
286:   AppCtx       *appctx = (AppCtx *)ptr;
287:   DM            da;
288:   DMDALocalInfo info;
289:   Field       **u, **f, **udot;
290:   Vec           localU;
291:   PetscInt      i, j, xs, ys, xm, ym, gxs, gys, gxm, gym;
292:   PetscReal     hx, hy, sx, sy;
293:   adouble       uc, uxx, uyy, vc, vxx, vyy;
294:   AField      **f_a, *f_c, **u_a, *u_c;
295:   PetscScalar   dummy;

297:   PetscFunctionBegin;
298:   PetscCall(TSGetDM(ts, &da));
299:   PetscCall(DMDAGetLocalInfo(da, &info));
300:   PetscCall(DMGetLocalVector(da, &localU));
301:   hx  = 2.50 / (PetscReal)info.mx;
302:   sx  = 1.0 / (hx * hx);
303:   hy  = 2.50 / (PetscReal)info.my;
304:   sy  = 1.0 / (hy * hy);
305:   xs  = info.xs;
306:   xm  = info.xm;
307:   gxs = info.gxs;
308:   gxm = info.gxm;
309:   ys  = info.ys;
310:   ym  = info.ym;
311:   gys = info.gys;
312:   gym = info.gym;

314:   /*
315:      Scatter ghost points to local vector,using the 2-step process
316:         DMGlobalToLocalBegin(),DMGlobalToLocalEnd().
317:      By placing code between these two statements, computations can be
318:      done while messages are in transition.
319:   */
320:   PetscCall(DMGlobalToLocalBegin(da, U, INSERT_VALUES, localU));
321:   PetscCall(DMGlobalToLocalEnd(da, U, INSERT_VALUES, localU));

323:   /*
324:      Get pointers to vector data
325:   */
326:   PetscCall(DMDAVecGetArrayRead(da, localU, &u));
327:   PetscCall(DMDAVecGetArray(da, F, &f));
328:   PetscCall(DMDAVecGetArrayRead(da, Udot, &udot));

330:   /*
331:     Create contiguous 1-arrays of AFields

333:     NOTE: Memory for ADOL-C active variables (such as adouble and AField)
334:           cannot be allocated using PetscMalloc, as this does not call the
335:           relevant class constructor. Instead, we use the C++ keyword `new`.
336:   */
337:   u_c = new AField[info.gxm * info.gym];
338:   f_c = new AField[info.gxm * info.gym];

340:   /* Create corresponding 2-arrays of AFields */
341:   u_a = new AField *[info.gym];
342:   f_a = new AField *[info.gym];

344:   /* Align indices between array types to endow 2d array with ghost points */
345:   PetscCall(GiveGhostPoints(da, u_c, &u_a));
346:   PetscCall(GiveGhostPoints(da, f_c, &f_a));

348:   trace_on(1); /* Start of active section on tape 1 */

350:   /*
351:     Mark independence

353:     NOTE: Ghost points are marked as independent, in place of the points they represent on
354:           other processors / on other boundaries.
355:   */
356:   for (j = gys; j < gys + gym; j++) {
357:     for (i = gxs; i < gxs + gxm; i++) {
358:       u_a[j][i].u <<= u[j][i].u;
359:       u_a[j][i].v <<= u[j][i].v;
360:     }
361:   }

363:   /* Compute function over the locally owned part of the grid */
364:   for (j = ys; j < ys + ym; j++) {
365:     for (i = xs; i < xs + xm; i++) {
366:       uc          = u_a[j][i].u;
367:       uxx         = (-2.0 * uc + u_a[j][i - 1].u + u_a[j][i + 1].u) * sx;
368:       uyy         = (-2.0 * uc + u_a[j - 1][i].u + u_a[j + 1][i].u) * sy;
369:       vc          = u_a[j][i].v;
370:       vxx         = (-2.0 * vc + u_a[j][i - 1].v + u_a[j][i + 1].v) * sx;
371:       vyy         = (-2.0 * vc + u_a[j - 1][i].v + u_a[j + 1][i].v) * sy;
372:       f_a[j][i].u = udot[j][i].u - appctx->D1 * (uxx + uyy) + uc * vc * vc - appctx->gamma * (1.0 - uc);
373:       f_a[j][i].v = udot[j][i].v - appctx->D2 * (vxx + vyy) - uc * vc * vc + (appctx->gamma + appctx->kappa) * vc;
374:     }
375:   }

377:   /*
378:     Mark dependence

380:     NOTE: Marking dependence of dummy variables makes the index notation much simpler when forming
381:           the Jacobian later.
382:   */
383:   for (j = gys; j < gys + gym; j++) {
384:     for (i = gxs; i < gxs + gxm; i++) {
385:       if ((i < xs) || (i >= xs + xm) || (j < ys) || (j >= ys + ym)) {
386:         f_a[j][i].u >>= dummy;
387:         f_a[j][i].v >>= dummy;
388:       } else {
389:         f_a[j][i].u >>= f[j][i].u;
390:         f_a[j][i].v >>= f[j][i].v;
391:       }
392:     }
393:   }
394:   trace_off(); /* End of active section */
395:   PetscCall(PetscLogFlops(16.0 * xm * ym));

397:   /* Restore vectors */
398:   PetscCall(DMDAVecRestoreArray(da, F, &f));
399:   PetscCall(DMDAVecRestoreArrayRead(da, localU, &u));
400:   PetscCall(DMDAVecRestoreArrayRead(da, Udot, &udot));

402:   PetscCall(DMRestoreLocalVector(da, &localU));

404:   /* Destroy AFields appropriately */
405:   f_a += info.gys;
406:   u_a += info.gys;
407:   delete[] f_a;
408:   delete[] u_a;
409:   delete[] f_c;
410:   delete[] u_c;
411:   PetscFunctionReturn(PETSC_SUCCESS);
412: }

414: /*
415:   Simply acts to pass TS information to the AdolcMatCtx
416: */
417: PetscErrorCode IJacobianMatFree(TS ts, PetscReal t, Vec X, Vec Xdot, PetscReal a, Mat A_shell, Mat B, void *ctx)
418: {
419:   AdolcMatCtx *mctx;
420:   DM           da;

422:   PetscFunctionBeginUser;
423:   PetscCall(MatShellGetContext(A_shell, &mctx));

425:   mctx->time  = t;
426:   mctx->shift = a;
427:   if (mctx->ts != ts) mctx->ts = ts;
428:   PetscCall(VecCopy(X, mctx->X));
429:   PetscCall(VecCopy(Xdot, mctx->Xdot));
430:   PetscCall(TSGetDM(ts, &da));
431:   PetscCall(DMGlobalToLocalBegin(da, mctx->X, INSERT_VALUES, mctx->localX0));
432:   PetscCall(DMGlobalToLocalEnd(da, mctx->X, INSERT_VALUES, mctx->localX0));
433:   PetscFunctionReturn(PETSC_SUCCESS);
434: }

436: /*TEST

438:   build:
439:     requires: double !complex adolc

441:   test:
442:     suffix: 1
443:     args: -ts_max_steps 1 -da_grid_x 12 -da_grid_y 12 -snes_test_jacobian
444:     output_file: output/adr_ex5adj_mf_1.out

446:   test:
447:     suffix: 2
448:     nsize: 4
449:     args: -ts_max_steps 10 -da_grid_x 12 -da_grid_y 12 -ts_monitor -ts_adjoint_monitor
450:     output_file: output/adr_ex5adj_mf_2.out

452: TEST*/