| #
2083db08
|
| 21-Sep-2019 |
Barry Smith <bsmith@mcs.anl.gov> |
Merge branch 'hongzh/fix-ex20opt-p' into 'master'
Fix a bug in ex20opt_p
See merge request petsc/petsc!2077
|
| #
29b3c8a1
|
| 20-Sep-2019 |
Hong Zhang <hongzhang@anl.gov> |
Make the adjoint info more informative
Funded-by:
Project: FASTMath adjoint
Time: 0.5 hours
Reported-by:
Thanks-to:
|
| #
ed0ed7c6
|
| 19-Sep-2019 |
Barry Smith <bsmith@mcs.anl.gov> |
Merge branch 'master' into 'barry/2019-08-22/fix-fblaslapack-for-nag'
# Conflicts:
# lib/petsc/conf/test
|
| #
2eee7f68
|
| 17-Sep-2019 |
Barry Smith <bsmith@mcs.anl.gov> |
Merge branch 'hongzh/fd-for-adjoint' into 'master'
Support Jacobian approximation with FDColoring in adjoint
See merge request petsc/petsc!2037
|
| #
04f7482e
|
| 14-Sep-2019 |
Hong Zhang <hongzhang@anl.gov> |
Use the new API instead of a private function for adjoint Theta
Funded-by:
Project: FASTMath adjoint
Time: 0.5 hours
Reported-by:
Thanks-to: Barry Smith <bsmith@mcs.anl.gov>
|
| #
494ce9fc
|
| 09-Sep-2019 |
Hong Zhang <hongzhang@anl.gov> |
Fix SNESTSFormFunction_Theta in order to use FDColoring
VecAXPBYPCZ complains about aliasing vectors in the input argument. This can be seen by
./ex5adj -ts_max_steps 10 -ts_monitor -ts_adjoint_mon
Fix SNESTSFormFunction_Theta in order to use FDColoring
VecAXPBYPCZ complains about aliasing vectors in the input argument. This can be seen by
./ex5adj -ts_max_steps 10 -ts_monitor -ts_adjoint_monitor -ksp_monitor_short -da_grid_x 15 -da_grid_y 15 -snes_fd -implicitform 0 -ts_type cn
Funded-by:
Project: FASTMath adjoint
Time: 1.0 hours
Reported-by:
Thanks-to:
show more ...
|
| #
c586f2e8
|
| 06-Sep-2019 |
Hong Zhang <hongzhang@anl.gov> |
Use FD Jacobian in adjoint Theta methods
Funded-by:
Project: FASTMath adjoint
Time: 1.0 hours
Reported-by:
Thanks-to: Barry Smith <bsmith@mcs.anl.gov>
|
| #
3e05ceb1
|
| 06-Sep-2019 |
Hong Zhang <hongzhang@anl.gov> |
Add the shift variable to the TSTheta context
So the shift can be changed dynamically for SNES function and Jacobian evaluations. Previously it is fixed.
Funded-by:
Project: FASTMath adjoint
Time:
Add the shift variable to the TSTheta context
So the shift can be changed dynamically for SNES function and Jacobian evaluations. Previously it is fixed.
Funded-by:
Project: FASTMath adjoint
Time: 1.0 hours
Reported-by:
Thanks-to: Barry Smith <bsmith@mcs.anl.gov>
show more ...
|
| #
6980b705
|
| 05-Sep-2019 |
Hong Zhang <hongzhang@anl.gov> |
Initial attempt to use FD Jacobian in adjoint BEuler
-snes_fd_color has more aggressive performance optimizations than -snes_fd; it takes advantage of the fact that a base vector may be ready before
Initial attempt to use FD Jacobian in adjoint BEuler
-snes_fd_color has more aggressive performance optimizations than -snes_fd; it takes advantage of the fact that a base vector may be ready before the Jacobian evaluation and the function evaluation can be skipped for the base vector. This commit makes -snes_fd_colr work for BEuler.
Funded-by:
Project: FASTMath adjoint
Time: 4.0 hours
Reported-by:
Thanks-to: Barry Smith <bsmith@mcs.anl.gov>
show more ...
|
| #
35140aeb
|
| 05-Sep-2019 |
Pierre Jolivet <pierre.jolivet@enseeiht.fr> |
Merge branch 'master' into this one.
|
| #
c771d45a
|
| 04-Sep-2019 |
Satish Balay <balay@mcs.anl.gov> |
Merge branch 'hongzh/improve-adjoint-logging' into 'master'
Hongzh/improve adjoint logging
See merge request petsc/petsc!1984
|
| #
077c4fea
|
| 05-Aug-2019 |
Hong Zhang <hongzhang@anl.gov> |
Performance improvement for adjoint BEuler
Adjoint BEuler can use one Jacobian evaluatation per adjoint time step instead of two for explicit ODEs. But the equation type needs to be set by the user.
Performance improvement for adjoint BEuler
Adjoint BEuler can use one Jacobian evaluatation per adjoint time step instead of two for explicit ODEs. But the equation type needs to be set by the user.
Funded-by:
Project: FASTMath adjoint
Time: 1.5 hours
Reported-by:
Thanks-to:
show more ...
|
| #
7e33834a
|
| 10-May-2019 |
Hong Zhang <hongzhang@anl.gov> |
Merged in hongzh/ts-adjoint2 (pull request #1562)
Second-order adjoint ODE solvers
Approved-by: BarryFSmith <bsmith@mcs.anl.gov>
|
| #
b1e111eb
|
| 24-Apr-2019 |
Hong Zhang <hongzhang@anl.gov> |
Change names for Hessian functions
Funded-by:
Project: FASTMath adjoint
Time: 0.5 hours
Reported-by:
Thanks-to:
|
| #
7207e0fd
|
| 05-Mar-2019 |
Hong Zhang <hongzhang@anl.gov> |
Use quadrature ts to hold tlm sensitivities
- Need guards for integral sensitivity
- Remove unneeded check since TLM of quadrature TS use a different TS
Funded-by:
Project: FASTMath adjoint
Time: 6
Use quadrature ts to hold tlm sensitivities
- Need guards for integral sensitivity
- Remove unneeded check since TLM of quadrature TS use a different TS
Funded-by:
Project: FASTMath adjoint
Time: 6.0 hours
Reported-by:
Thanks-to:
show more ...
|
| #
cd4cee2d
|
| 11-Dec-2018 |
Hong Zhang <hongzhang@anl.gov> |
Refactor quadrature
- Use a separate TS for quadrature
- Initially focus on Theta and Beuler
- Tested on ex3opt and ex9opt
- RK support is not complete
- Fixed memory leak. log_view for ex3sa shows
Refactor quadrature
- Use a separate TS for quadrature
- Initially focus on Theta and Beuler
- Tested on ex3opt and ex9opt
- RK support is not complete
- Fixed memory leak. log_view for ex3sa shows that there is one matrix that is not destroyed.
Funded-by:
Project: FASTMath adjoint
Time: 9.0 hours
Reported-by:
Thanks-to:
show more ...
|
| #
67633408
|
| 12-Apr-2019 |
Hong Zhang <hongzhang@anl.gov> |
TSTheta: hack Hessian computation for explicit ODE
The adjoint of Theta methods does not work if users provide callbacks for an explicit ODE (RHSFunction,RHSJacobian,RHSHessian,etc.).
TSComputeXXX h
TSTheta: hack Hessian computation for explicit ODE
The adjoint of Theta methods does not work if users provide callbacks for an explicit ODE (RHSFunction,RHSJacobian,RHSHessian,etc.).
TSComputeXXX handles the conversion implicitly, which is convoluted.
Funded-by:
Project: FASTMATH adjoint
Time: 1.0 hours
Reported-by:
Thanks-to:
show more ...
|
| #
33f72c5d
|
| 07-May-2018 |
Hong Zhang <hongzhang@anl.gov> |
Refactor TSAdjoint
- Add TSComputeIJacobianP() and TSSetIJacobianP() to pave the way for IMEX adjoint.
- Working on the second-order adjoint to compute parametric sensitivities using backward Euler.
Refactor TSAdjoint
- Add TSComputeIJacobianP() and TSSetIJacobianP() to pave the way for IMEX adjoint.
- Working on the second-order adjoint to compute parametric sensitivities using backward Euler.
- Remove redundant code in TSReset_Theta. TSReset_Theta should not do what TSForwardReset_Theta or TSAdjointReset_Theta do
Funded-by:
Project: FASTMath adjoint
Time: 7.0 hours
Reported-by:
Thanks-to:
Fix TSAdjointInitializeForward for second-order adjoint
To calculate second-order parametric sensitivity, the TLM variable should be initialized as 2*dIdP*dir.
Funded-by:
Project: FASTMath adjoint
Time: 2.0 hours
Reported-by:
Thanks-to:
show more ...
|
| #
1d14d03b
|
| 30-Apr-2018 |
Hong Zhang <hongzhang@anl.gov> |
Second order adjoint works for theta
|
| #
05c9950e
|
| 25-Apr-2018 |
Hong Zhang <hongzhang@anl.gov> |
Add check for convergence in adjoint run
Funded-by:
Project: FASTMATH adjoint
Time: 0.5 hours
Reported-by:
Thanks-to:
|
| #
c9681e5d
|
| 19-Apr-2018 |
Hong Zhang <hongzhang@anl.gov> |
Separate adjoint of BEuler from Theta
BEuler corresponds to a special case: no endpoint + no intermediate stage
Funded-by:
Project: FASTMATH adjoint
Time: 6.0 hours
Reported-by:
Thanks-to:
|
| #
7adebdde
|
| 20-Apr-2018 |
Hong Zhang <hongzhang@anl.gov> |
Add reset function for TSForward
Funded-by:
Project: FASTMath adjoint
Time: 2.0 hours
Reported-by:
Thanks-to:
|
| #
ece46509
|
| 19-Apr-2018 |
Hong Zhang <hongzhang@anl.gov> |
Add reset function for TSAdjoint
This allows TSAdjoint to be turned on/off between different runs using the same TS
Funded-by:
Project: FASTMATH adjoint
Time: 2.0 hours
Reported-by:
Thanks-to:
|
| #
6affb6f8
|
| 05-Apr-2018 |
Hong Zhang <hongzhang@anl.gov> |
Checkpoint the TLM sensitivity for second-order adjoint
Add TSForwardGetStages to acces the stage values for TLM sensitivities
Funded-by:
Project: FASTMath adjoint
Time: 8 hours
Reported-by:
Thanks
Checkpoint the TLM sensitivity for second-order adjoint
Add TSForwardGetStages to acces the stage values for TLM sensitivities
Funded-by:
Project: FASTMath adjoint
Time: 8 hours
Reported-by:
Thanks-to:
show more ...
|
| #
0e11c21f
|
| 01-Apr-2018 |
Hong Zhang <hongzhang@anl.gov> |
Fix the error in TLM when JacobianP is not provided
When calculating sensitivity to initial values, the JacobianP should not be required. In this case, users can either provide an all-zero JacobianP
Fix the error in TLM when JacobianP is not provided
When calculating sensitivity to initial values, the JacobianP should not be required. In this case, users can either provide an all-zero JacobianP or not provide it.
Funded-by:
Project: FASTMATH adjoint
Time: 1.0 hours
Reported-by:
Thanks-to:
show more ...
|