compile | R Documentation |
Compile a C++ template into a shared object file. OpenMP flag is set if the template is detected to be parallel.
compile(
file,
flags = "",
safebounds = TRUE,
safeunload = TRUE,
openmp = isParallelTemplate(file[1]),
libtmb = TRUE,
libinit = TRUE,
tracesweep = FALSE,
framework = getOption("tmb.ad.framework"),
supernodal = FALSE,
longint = FALSE,
eigen.disable.warnings = TRUE,
max.order = NULL,
...
)
file |
C++ file. |
flags |
Character with compile flags. |
safebounds |
Turn on preprocessor flag for bound checking? |
safeunload |
Turn on preprocessor flag for safe DLL unloading? |
openmp |
Turn on openmp flag? Auto detected for parallel templates. |
libtmb |
Use precompiled TMB library if available (to speed up compilation)? |
libinit |
Turn on preprocessor flag to register native routines? |
tracesweep |
Turn on preprocessor flag to trace AD sweeps? (Silently disables |
framework |
Which AD framework to use ('TMBad' or 'CppAD') |
supernodal |
Turn on preprocessor flag to use supernodal sparse Cholesky/Inverse from system wide suitesparse library |
longint |
Turn on preprocessor flag to use long integers for Eigen's SparseMatrix StorageIndex |
eigen.disable.warnings |
Turn on preprocessor flag to disable nuisance warnings. Note that this is not allowed for code to be compiled on CRAN. |
max.order |
Maximum derivative order of compiler generated atomic special functions - see details. |
... |
Passed as Makeconf variables. |
TMB relies on R's built in functionality to create shared libraries independent of the platform.
A template is compiled by compile("template.cpp")
, which will call R's makefile with appropriate
preprocessor flags.
Compiler and compiler flags can be stored in a configuration file. In order of precedence either via
the file pointed at by R_MAKEVARS_USER or the file ~/.R/Makevars if it exists.
Additional configuration variables can be set with the flags
and ...
arguments, which will override any
previous selections.
Sparse matrix calculations play an important role in TMB. By default TMB uses a small subset of SuiteSparse
available through the R package Matrix
. This is sufficient for most use cases, however for some very large models the following extra features are worth considering:
Some large models benefit from an extended set of graph reordering algorithms (especially METIS) not part of Matrix
. It is common that these orderings can provide quite big speedups.
Some large models need sparse matrices with number of nonzeros exceeding the current 32 bit limitation of Matrix
. Normally such cases will result in the cholmod error 'problem too large'. SuiteSparse
includes 64 bit integer routines to address this problem.
Experimental support for linking to a custom SuiteSparse
installation is available through two arguments to the compile
function. The first argument supernodal=TRUE
tells TMB to use the supernodal Cholesky factorization from the system wide SuiteSparse
on the C++ side. This will affect the speed of the Laplace approximation when run internally (using arguments intern
or integrate
to MakeADFun
).
The second argument longint=TRUE
tells TMB to use 64 bit integers for sparse matrices on the C++ side. This works in combination with supernodal=TRUE
from Eigen version 3.4.
On Windows a SuiteSparse
installation can be obtained using the Rtools
package manager. Start 'Rtools Bash' terminal and run:
pacman -Sy pacman -S mingw-w64-{i686,x86_64}-suitesparse
On Linux one should look for the package libsuitesparse-dev
.
TMB supports two different AD libraries 'CppAD' and 'TMBad' selected via the argument framework
which works as a switch to set one of two C++ preprocessor flags: 'CPPAD_FRAMEWORK' or 'TMBAD_FRAMEWORK'. The default value of framework
can be set from R by options("tmb.ad.framework")
or alternatively from the shell via the environment variable 'TMB_AD_FRAMEWORK'. Packages linking to TMB should set one of the two C++ preprocessor flags in Makevars.
The argument max.order
controls the maximum derivative order of special functions (e.g. pbeta
) generated by the compiler. By default the value is set to 3 which is sufficient to obtain the Laplace approximation (order 2) and its derivatives (order 3). However, sometimes a higher value may be needed. For example framework='TMBad'
allows one to calculate the Hessian of the Laplace approximation, but that requires 4th order derivatives of special functions in use. A too small value will cause the runtime error 'increase TMB_MAX_ORDER'. Note that compilation time and binary size increases with max.order
.
precompile
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.