xref: /petsc/src/ts/tutorials/autodiff/ex16adj_tl.cxx (revision 4e8208cbcbc709572b8abe32f33c78b69c819375)
1 static char help[] = "Demonstrates tapeless automatic Jacobian generation using ADOL-C for an adjoint sensitivity analysis of the van der Pol equation.\n\
2 Input parameters include:\n\
3       -mu : stiffness parameter\n\n";
4 
5 /*
6    REQUIRES configuration of PETSc with option --download-adolc.
7 
8    For documentation on ADOL-C, see
9      $PETSC_ARCH/externalpackages/ADOL-C-2.6.0/ADOL-C/doc/adolc-manual.pdf
10 */
11 /* ------------------------------------------------------------------------
12    See ex16adj for a description of the problem being solved.
13   ------------------------------------------------------------------------- */
14 
15 #include <petscts.h>
16 #include <petscmat.h>
17 
18 #define ADOLC_TAPELESS
19 #define NUMBER_DIRECTIONS 3
20 #include "adolc-utils/drivers.cxx"
21 #include <adolc/adtl.h>
22 using namespace adtl;
23 
24 typedef struct _n_User *User;
25 struct _n_User {
26   PetscReal mu;
27   PetscReal next_output;
28   PetscReal tprev;
29 
30   /* Automatic differentiation support */
31   AdolcCtx *adctx;
32   Vec       F;
33 };
34 
35 /*
36   Residual evaluation templated, so as to allow for PetscScalar or adouble
37   arguments.
38 */
39 template <class T>
EvaluateResidual(const T * x,T mu,T * f)40 PetscErrorCode EvaluateResidual(const T *x, T mu, T *f)
41 {
42   PetscFunctionBegin;
43   f[0] = x[1];
44   f[1] = mu * (1. - x[0] * x[0]) * x[1] - x[0];
45   PetscFunctionReturn(PETSC_SUCCESS);
46 }
47 
48 /*
49   'Passive' RHS function, used in residual evaluations during the time integration.
50 */
RHSFunctionPassive(TS ts,PetscReal t,Vec X,Vec F,PetscCtx ctx)51 static PetscErrorCode RHSFunctionPassive(TS ts, PetscReal t, Vec X, Vec F, PetscCtx ctx)
52 {
53   User               user = (User)ctx;
54   PetscScalar       *f;
55   const PetscScalar *x;
56 
57   PetscFunctionBeginUser;
58   PetscCall(VecGetArrayRead(X, &x));
59   PetscCall(VecGetArray(F, &f));
60   PetscCall(EvaluateResidual(x, user->mu, f));
61   PetscCall(VecRestoreArrayRead(X, &x));
62   PetscCall(VecRestoreArray(F, &f));
63   PetscFunctionReturn(PETSC_SUCCESS);
64 }
65 
66 /*
67   Compute the Jacobian w.r.t. x using tapeless mode of ADOL-C.
68 */
RHSJacobian(TS ts,PetscReal t,Vec X,Mat A,Mat B,PetscCtx ctx)69 static PetscErrorCode RHSJacobian(TS ts, PetscReal t, Vec X, Mat A, Mat B, PetscCtx ctx)
70 {
71   User               user = (User)ctx;
72   PetscScalar      **J;
73   const PetscScalar *x;
74   adouble            f_a[2];       /* 'active' double for dependent variables */
75   adouble            x_a[2], mu_a; /* 'active' doubles for independent variables */
76   PetscInt           i, j;
77 
78   PetscFunctionBeginUser;
79   /* Set values for independent variables and parameters */
80   PetscCall(VecGetArrayRead(X, &x));
81   x_a[0].setValue(x[0]);
82   x_a[1].setValue(x[1]);
83   mu_a.setValue(user->mu);
84   PetscCall(VecRestoreArrayRead(X, &x));
85 
86   /* Set seed matrix as 3x3 identity matrix */
87   x_a[0].setADValue(0, 1.);
88   x_a[0].setADValue(1, 0.);
89   x_a[0].setADValue(2, 0.);
90   x_a[1].setADValue(0, 0.);
91   x_a[1].setADValue(1, 1.);
92   x_a[1].setADValue(2, 0.);
93   mu_a.setADValue(0, 0.);
94   mu_a.setADValue(1, 0.);
95   mu_a.setADValue(2, 1.);
96 
97   /* Evaluate residual (on active variables) */
98   PetscCall(EvaluateResidual(x_a, mu_a, f_a));
99 
100   /* Extract derivatives */
101   PetscCall(PetscMalloc1(user->adctx->n, &J));
102   J[0] = (PetscScalar *)f_a[0].getADValue();
103   J[1] = (PetscScalar *)f_a[1].getADValue();
104 
105   /* Set matrix values */
106   for (i = 0; i < user->adctx->m; i++) {
107     for (j = 0; j < user->adctx->n; j++) PetscCall(MatSetValues(A, 1, &i, 1, &j, &J[i][j], INSERT_VALUES));
108   }
109   PetscCall(PetscFree(J));
110   PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
111   PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
112   if (A != B) {
113     PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
114     PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
115   }
116   PetscFunctionReturn(PETSC_SUCCESS);
117 }
118 
119 /*
120   Compute the Jacobian w.r.t. mu using tapeless mode of ADOL-C.
121 */
RHSJacobianP(TS ts,PetscReal t,Vec X,Mat A,PetscCtx ctx)122 static PetscErrorCode RHSJacobianP(TS ts, PetscReal t, Vec X, Mat A, PetscCtx ctx)
123 {
124   User          user = (User)ctx;
125   PetscScalar **J;
126   PetscScalar  *x;
127   adouble       f_a[2];       /* 'active' double for dependent variables */
128   adouble       x_a[2], mu_a; /* 'active' doubles for independent variables */
129   PetscInt      i, j = 0;
130 
131   PetscFunctionBeginUser;
132   /* Set values for independent variables and parameters */
133   PetscCall(VecGetArray(X, &x));
134   x_a[0].setValue(x[0]);
135   x_a[1].setValue(x[1]);
136   mu_a.setValue(user->mu);
137   PetscCall(VecRestoreArray(X, &x));
138 
139   /* Set seed matrix as 3x3 identity matrix */
140   x_a[0].setADValue(0, 1.);
141   x_a[0].setADValue(1, 0.);
142   x_a[0].setADValue(2, 0.);
143   x_a[1].setADValue(0, 0.);
144   x_a[1].setADValue(1, 1.);
145   x_a[1].setADValue(2, 0.);
146   mu_a.setADValue(0, 0.);
147   mu_a.setADValue(1, 0.);
148   mu_a.setADValue(2, 1.);
149 
150   /* Evaluate residual (on active variables) */
151   PetscCall(EvaluateResidual(x_a, mu_a, f_a));
152 
153   /* Extract derivatives */
154   PetscCall(PetscMalloc1(2, &J));
155   J[0] = (PetscScalar *)f_a[0].getADValue();
156   J[1] = (PetscScalar *)f_a[1].getADValue();
157 
158   /* Set matrix values */
159   for (i = 0; i < user->adctx->m; i++) PetscCall(MatSetValues(A, 1, &i, 1, &j, &J[i][user->adctx->n], INSERT_VALUES));
160   PetscCall(PetscFree(J));
161   PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
162   PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
163   PetscFunctionReturn(PETSC_SUCCESS);
164 }
165 
166 /*
167   Monitor timesteps and use interpolation to output at integer multiples of 0.1
168 */
Monitor(TS ts,PetscInt step,PetscReal t,Vec X,PetscCtx ctx)169 static PetscErrorCode Monitor(TS ts, PetscInt step, PetscReal t, Vec X, PetscCtx ctx)
170 {
171   const PetscScalar *x;
172   PetscReal          tfinal, dt, tprev;
173   User               user = (User)ctx;
174 
175   PetscFunctionBeginUser;
176   PetscCall(TSGetTimeStep(ts, &dt));
177   PetscCall(TSGetMaxTime(ts, &tfinal));
178   PetscCall(TSGetPrevTime(ts, &tprev));
179   PetscCall(VecGetArrayRead(X, &x));
180   PetscCall(PetscPrintf(PETSC_COMM_WORLD, "[%.1f] %" PetscInt_FMT " TS %.6f (dt = %.6f) X % 12.6e % 12.6e\n", (double)user->next_output, step, (double)t, (double)dt, (double)PetscRealPart(x[0]), (double)PetscRealPart(x[1])));
181   PetscCall(PetscPrintf(PETSC_COMM_WORLD, "t %.6f (tprev = %.6f) \n", (double)t, (double)tprev));
182   PetscCall(VecRestoreArrayRead(X, &x));
183   PetscFunctionReturn(PETSC_SUCCESS);
184 }
185 
main(int argc,char ** argv)186 int main(int argc, char **argv)
187 {
188   TS             ts;   /* nonlinear solver */
189   Vec            x;    /* solution, residual vectors */
190   Mat            A;    /* Jacobian matrix */
191   Mat            Jacp; /* JacobianP matrix */
192   PetscInt       steps;
193   PetscReal      ftime   = 0.5;
194   PetscBool      monitor = PETSC_FALSE;
195   PetscScalar   *x_ptr;
196   PetscMPIInt    size;
197   struct _n_User user;
198   AdolcCtx      *adctx;
199   Vec            lambda[2], mu[2];
200 
201   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
202      Initialize program
203      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
204   PetscFunctionBeginUser;
205   PetscCall(PetscInitialize(&argc, &argv, NULL, help));
206   PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size));
207   PetscCheck(size == 1, PETSC_COMM_WORLD, PETSC_ERR_WRONG_MPI_SIZE, "This is a uniprocessor example only!");
208 
209   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
210     Set runtime options and create AdolcCtx
211     - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
212   PetscCall(PetscNew(&adctx));
213   user.mu          = 1;
214   user.next_output = 0.0;
215   adctx->m         = 2;
216   adctx->n         = 2;
217   adctx->p         = 2;
218   user.adctx       = adctx;
219   adtl::setNumDir(adctx->n + 1); /* #indep. variables, plus parameters */
220 
221   PetscCall(PetscOptionsGetReal(NULL, NULL, "-mu", &user.mu, NULL));
222   PetscCall(PetscOptionsGetBool(NULL, NULL, "-monitor", &monitor, NULL));
223 
224   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
225     Create necessary matrix and vectors, solve same ODE on every process
226     - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
227   PetscCall(MatCreate(PETSC_COMM_WORLD, &A));
228   PetscCall(MatSetSizes(A, PETSC_DECIDE, PETSC_DECIDE, 2, 2));
229   PetscCall(MatSetFromOptions(A));
230   PetscCall(MatSetUp(A));
231   PetscCall(MatCreateVecs(A, &x, NULL));
232 
233   PetscCall(MatCreate(PETSC_COMM_WORLD, &Jacp));
234   PetscCall(MatSetSizes(Jacp, PETSC_DECIDE, PETSC_DECIDE, 2, 1));
235   PetscCall(MatSetFromOptions(Jacp));
236   PetscCall(MatSetUp(Jacp));
237 
238   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
239      Create timestepping solver context
240      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
241   PetscCall(TSCreate(PETSC_COMM_WORLD, &ts));
242   PetscCall(TSSetType(ts, TSRK));
243   PetscCall(TSSetRHSFunction(ts, NULL, RHSFunctionPassive, &user));
244 
245   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
246      Set initial conditions
247    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
248   PetscCall(VecGetArray(x, &x_ptr));
249   x_ptr[0] = 2;
250   x_ptr[1] = 0.66666654321;
251   PetscCall(VecRestoreArray(x, &x_ptr));
252 
253   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
254      Set RHS Jacobian for the adjoint integration
255      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
256   PetscCall(TSSetRHSJacobian(ts, A, A, RHSJacobian, &user));
257   PetscCall(TSSetMaxTime(ts, ftime));
258   PetscCall(TSSetExactFinalTime(ts, TS_EXACTFINALTIME_MATCHSTEP));
259   if (monitor) PetscCall(TSMonitorSet(ts, Monitor, &user, NULL));
260   PetscCall(TSSetTimeStep(ts, .001));
261 
262   /*
263     Have the TS save its trajectory so that TSAdjointSolve() may be used
264   */
265   PetscCall(TSSetSaveTrajectory(ts));
266 
267   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
268      Set runtime options
269    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
270   PetscCall(TSSetFromOptions(ts));
271 
272   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
273      Solve nonlinear system
274      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
275   PetscCall(TSSolve(ts, x));
276   PetscCall(TSGetSolveTime(ts, &ftime));
277   PetscCall(TSGetStepNumber(ts, &steps));
278   PetscCall(PetscPrintf(PETSC_COMM_WORLD, "mu %g, steps %" PetscInt_FMT ", ftime %g\n", (double)user.mu, steps, (double)ftime));
279   PetscCall(VecView(x, PETSC_VIEWER_STDOUT_WORLD));
280 
281   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
282      Start the Adjoint model
283      - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
284   PetscCall(MatCreateVecs(A, &lambda[0], NULL));
285   PetscCall(MatCreateVecs(A, &lambda[1], NULL));
286   /*   Reset initial conditions for the adjoint integration */
287   PetscCall(VecGetArray(lambda[0], &x_ptr));
288   x_ptr[0] = 1.0;
289   x_ptr[1] = 0.0;
290   PetscCall(VecRestoreArray(lambda[0], &x_ptr));
291   PetscCall(VecGetArray(lambda[1], &x_ptr));
292   x_ptr[0] = 0.0;
293   x_ptr[1] = 1.0;
294   PetscCall(VecRestoreArray(lambda[1], &x_ptr));
295 
296   PetscCall(MatCreateVecs(Jacp, &mu[0], NULL));
297   PetscCall(MatCreateVecs(Jacp, &mu[1], NULL));
298   PetscCall(VecGetArray(mu[0], &x_ptr));
299   x_ptr[0] = 0.0;
300   PetscCall(VecRestoreArray(mu[0], &x_ptr));
301   PetscCall(VecGetArray(mu[1], &x_ptr));
302   x_ptr[0] = 0.0;
303   PetscCall(VecRestoreArray(mu[1], &x_ptr));
304   PetscCall(TSSetCostGradients(ts, 2, lambda, mu));
305 
306   /*   Set RHS JacobianP */
307   PetscCall(TSSetRHSJacobianP(ts, Jacp, RHSJacobianP, &user));
308 
309   PetscCall(TSAdjointSolve(ts));
310 
311   PetscCall(VecView(lambda[0], PETSC_VIEWER_STDOUT_WORLD));
312   PetscCall(VecView(lambda[1], PETSC_VIEWER_STDOUT_WORLD));
313   PetscCall(VecView(mu[0], PETSC_VIEWER_STDOUT_WORLD));
314   PetscCall(VecView(mu[1], PETSC_VIEWER_STDOUT_WORLD));
315 
316   /* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
317      Free work space.  All PETSc objects should be destroyed when they
318      are no longer needed.
319    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
320   PetscCall(MatDestroy(&A));
321   PetscCall(MatDestroy(&Jacp));
322   PetscCall(VecDestroy(&x));
323   PetscCall(VecDestroy(&lambda[0]));
324   PetscCall(VecDestroy(&lambda[1]));
325   PetscCall(VecDestroy(&mu[0]));
326   PetscCall(VecDestroy(&mu[1]));
327   PetscCall(TSDestroy(&ts));
328   PetscCall(PetscFree(adctx));
329   PetscCall(PetscFinalize());
330   return 0;
331 }
332 
333 /*TEST
334 
335   build:
336     requires: double !complex adolc
337 
338   test:
339     suffix: 1
340     args: -ts_max_steps 10 -ts_monitor -ts_adjoint_monitor
341     output_file: output/ex16adj_tl_1.out
342 
343   test:
344     suffix: 2
345     args: -ts_max_steps 10 -ts_monitor -ts_adjoint_monitor -mu 5
346     output_file: output/ex16adj_tl_2.out
347 
348   test:
349     suffix: 3
350     args: -ts_max_steps 10 -monitor
351     output_file: output/ex16adj_tl_3.out
352 
353   test:
354     suffix: 4
355     args: -ts_max_steps 10 -monitor -mu 5
356     output_file: output/ex16adj_tl_4.out
357 
358 TEST*/
359