\(\newcommand{\W}[1]{ \; #1 \; }\) \(\newcommand{\R}[1]{ {\rm #1} }\) \(\newcommand{\B}[1]{ {\bf #1} }\) \(\newcommand{\D}[2]{ \frac{\partial #1}{\partial #2} }\) \(\newcommand{\DD}[3]{ \frac{\partial^2 #1}{\partial #2 \partial #3} }\) \(\newcommand{\Dpow}[2]{ \frac{\partial^{#1}}{\partial {#2}^{#1}} }\) \(\newcommand{\dpow}[2]{ \frac{ {\rm d}^{#1}}{{\rm d}\, {#2}^{#1}} }\)
Faq¶
View page sourceFrequently Asked Questions and Answers¶
Assignment and Independent¶
Why does the code sequence
Independent
( u );behave differently from the code sequence
Independent
( u );Before the call to Independent , u [0] is a Parameter and after the call it is a variable. Thus in the first case, v is a variable and in the second case it is a parameter.
Bugs¶
What should I do if I suspect that there is a bug in CppAD ?
The first step is to check currently open issues on github. If it is an open issue, and you want to hurry it along, you can add a comment to the effect that it is holding you up.
The next step is to search the whats_new sections for mention of a related bug fix between the date of the version you are using and the current date. It the bug has been fixed, obtain a more recent release that has the fix and see if that works for you.
The next step is to create a simple demonstration of the bug; see the file
bug/template.sh
for a template that you can edit for that purpose. The smaller the program, the better the bug report.The next step is open a new issue on github and provide your simple example so that the problem can be reproduced.
CompareChange¶
If you attempt to use the
CompareChange
function when NDEBUG
is true,
you will get an error message stating that
CompareChange
is not a member of the
ADFun template class.
Complex Types¶
Which of the following complex types is better:
AD< std::complex<
Base > >std::complex< AD<
Base > >The complex abs function is differentiable with respect to its real and imaginary parts, but it is not complex differentiable. Thus one would prefer to use
std::complex< AD<
Base > >
On the other hand, the C++ standard only specifies
std::complex<
Type > where Type is
float
, double
, or lone double
.
The effect of instantiating the template complex for any other type
is unspecified.
Exceptions¶
Why, in all the examples, do you pass back a boolean variable instead of throwing an exception ?
The examples are also used to test the correctness of CppAD and to check your installation. For these two uses, it is helpful to run all the tests and to know which ones failed. The actual code in CppAD uses the ErrorHandler utility to signal exceptions. Specifications for redefining this action are provided.
Independent Variables¶
Is it possible to evaluate the same tape recording with different values for the independent variables ?
Yes (see forward_zero ).
Matrix Inverse¶
Is it possible to differentiate (with respect to the matrix elements) the computation of the inverse of a matrix where the computation of the inverse uses pivoting ?
LuSolve¶
The example routine LuSolve can be used to do this because the inverse is a special case of the solution of linear equations. The examples jac_lu_det.cpp and hes_lu_det.cpp use LuSolve to compute derivatives of the determinant with respect to the components of the matrix.
Atomic Operation¶
One can also do this by making the inversion of the matrix an atomic operation; e.g., see atomic_two_eigen_mat_inv.cpp .
Mode: Forward or Reverse¶
When evaluating derivatives, one always has a choice between forward and reverse mode. How does one decide which mode to use ?
In general, the best mode depends on the number of domain and range
components in the function that your are differentiating.
Each call to Forward computes the derivative of all
the range directions with respect to one domain direction.
Each call to Reverse computes the derivative of one
range direction with respect to all the domain directions.
The times required for (speed of)
calls Forward
and Reverse
are about equal.
The Parameter
function can be used to quickly determine that
some range directions have derivative zero.
Namespace¶
Test Vector Preprocessor Symbol¶
Why do you use CPPAD_TESTVECTOR
instead of a namespace for
the CppAD testvector class ?
The preprocessor symbol CPPAD_TESTVECTOR
( see testvector )
determines which
SimpleVector template class is used for extensive testing.
The default definition for CPPAD_TESTVECTOR
is the
CppAD::vector template class, but it can be changed.
Note that all the preprocessor symbols that are defined or used
by CppAD begin with either CPPAD
(some old deprecated symbols begin with CppAD
).
Speed¶
How do I get the best speed performance out of CppAD ?
NDEBUG¶
You should compile your code with optimization, without debugging,
and with the preprocessor symbol NDEBUG
defined.
(The speed_cppad tests do this.)
Note that defining NDEBUG
will turn off all of the error checking and reporting that
is done using ErrorHandler .
Optimize¶
It is also possible that preforming a tape optimization will improve the speed of evaluation more than the time required for the optimization.
Memory Allocation¶
You may also increase execution speed
by calling hold_memory
with
value equal to true.
Tape Storage: Disk or Memory¶
Does CppAD store the tape on disk or in memory ?
CppAD uses memory to store a different tape for recording operations
for each AD
< Base > type that is used.
If you have a very large number calculations that are recorded
on a tape, the tape will keep growing to hold the necessary information.
Eventually, virtual memory may be used to store the tape
and the calculations may slow down because of necessary disk access.