Adolc tensor refactor #14205
Adolc tensor refactor #14205
Conversation
drwells
added
Automatic differentiation
drwells
force-pushed
the
adolc-tensor-refactor
branch
2 times, most recently
from
e7d8ebc
to
8c8bfb4
Compare
This works correctly with advanced branching.
This test should always pass now, regardless of how adol-c is set up.
drwells
force-pushed
the
adolc-tensor-refactor
branch
from
8c8bfb4
to
efe392b
Compare
| for (unsigned int i = 0; i < dim; ++i) | ||
| sum += std::fabs(t[i][j]); | ||
|
|
||
| condassign(max, (sum > max), sum, max); |
There was a problem hiding this comment.
@drwells For these taped types, one needs to use the condassign() function to switch between branches. An if statement won't work if the evaluation point changes which branch is taken.
There was a problem hiding this comment.
I'll write a test so that I understand this better.
There was a problem hiding this comment.
I had the ADOL-C manual in hand when decomposing and writing these functions, so whenever you come across an unusual implementation then I'd suggest consulting the manual. The requirements of the taping algorithm often dictates how code with branches has to be written in order to be taped AD compatible. It would not surprise me if there is not a test for every single specialised function -- the implementation was probably "best effort".
|
Thanks for working towards optimising all things related to the AD types, @drwells. I'll do my best to give you feedback for both this and your other PRs on the weekend. |
|
These functions really need to be in the header for reasons stated above (condassign in a template function). Furthermore, we don't gain a lot by trying to excise adolc from this file: adolc itself, when installed, only has 6k total lines of header files. Hence lets close this. |
More progress on #13949 - now for adol-c instead of Sacado.
I tested this on my machine with and without advanced branching.