3 Compiling a Stan Program
A Stan program must be in a file with extension .stan
.
The CmdStan makefile rules specify all necessary steps to
translate files with suffix .stan
to a CmdStan executable program.
This is a two-stage process:
- first the Stan program is translated to C++ by the
stanc
compiler - then the C++ compiler compiles all C++ sources and links them together with the CmdStan interface program and the Stan and math libraries.
3.1 Invoking the Make utility
To compile Stan programs, you must invoke the Make program from
the <cmdstan-home>
directory.
The Stan program can be in a different directory, but the directory path
names cannot contain spaces - this limitation is imposed by Make.
> cd <cmdstan_home>
In the call to the Make program, the target is name of the CmdStan executable
corresponding to the Stan program file.
On Mac and Linux, this is the name of the Stan program with the .stan
omitted. On Windows, replace .stan
with .exe
, and make
sure that the path is given with slashes and not backslashes.
To build the Bernoulli example, on Mac and Linux:
> make examples/bernoulli/bernoulli
On Windows, the command is the same with the addition of .exe
at the end of the target (note the use of forward slashes):
> make examples/bernoulli/bernoulli.exe
The generated C++ code (bernoulli.hpp
), object file (bernoulli.o
)
and the compiled executable will be placed in the same directory as the Stan program.
The compiled executable consists of the Stan model and the CmdStan command line interface which provides inference algorithms to do MCMC sampling, optimization, and variational inference. The following sections provide examples of doing inference using each method on the example model and data file.
3.2 Dependencies
When executing a Make target, all its dependencies are checked
to see if they are up to date, and if they are not, they are rebuilt.
If the you call Make with target bernoulli
twice in a row, without any
editing bernoulli.stan
or otherwise changing the system,
on the second invocation, Make will determine that the executable is already newer
than the Stan source file and will not recompile the program:
> make examples/bernoulli/bernoulli
make: `examples/bernoulli/bernoulli' is up to date.
If the file containing the Stan program is updated,
the next call to make
will rebuild the CmdStan executable.
3.3 Compiler errors
The Stan probabilistic programming language is a programming language with a rich syntax, as such, it is often the case that a carefully written program contains errors.
The simplest class of errors are simple syntax errors such as forgetting
the semi-colon statement termination marker at the end of a line,
or typos such as a misspelled variable name.
For example, if in the bernoulli.stan
program, we introduce a typo on line \(9\)
by writing thata
instead of theta
, the Make command fails with the following
--- Translating Stan model to C++ code ---
bin/stanc --o=bernoulli.hpp bernoulli.stan
Semantic error in 'bernoulli.stan', line 9, column 2 to column 7:
-------------------------------------------------
7: }
8: model {
9: thata ~ beta(1, 1); // uniform prior on interval 0, 1
^
10: y ~ bernoulli(theta);
11: }
-------------------------------------------------
Identifier 'thata' not in scope.
make: *** [bernoulli.hpp] Error 1
Stan is a strongly-typed language;
and the compiler will throw an error if statements or expressions violate the type rules.
The following trivial program foo.stan
contains an illegal assignment statement:
data {
real x;
}
transformed data {
int y = x;
}
The Make command fails with the following:
Semantic error in 'foo.stan', line 5, column 2 to column 12:
-------------------------------------------------
3: }
4: transformed data {
5: int y = x;
^
6: }
-------------------------------------------------
Ill-typed arguments supplied to assignment operator =: lhs has type int and rhs has type real
The Stan Reference Manual provides a complete specification of the Stan programming language. The Stan User’s Guide also contains a full description of the errors and warnings stanc can emit.
3.4 Troubleshooting C++ compiler or linker errors
If the stanc compiler successfully translates a Stan program to C++, the resulting C++ code should be valid C++ which can be compiled into an executable. The stanc compiler is also a program, and while it has been extensively tested, it may still contain errors such that the generated C++ code fails to compile.
The Make command prints the following message to the terminal at the point when it compiles and links the C++ file:
--- Compiling, linking C++ code ---
If the program fails to compile for any reason, the C++ compiler and linker will most likely print a long series of error messages to the console.
If this happens, please report the error, together with the Stan program on either the Stan Forums or on the Stan compiler GitHub issues tracker.
3.5 C++ compilation and linking flags
Users can set flags for the C++ compiler and linker and compiler to optimize their executables. We advise users to only do this once they are sure their basic setup of Cmdstan without flags works.
The CXXFLAGS
and LDFLAGS
makefile variables can be used to set compiler and linker flags respecitvely. We
recommend setting these in the make/local
file.
For example:
CXXFLAGS = -O2
A recommend a set of CXXFLAGS
and LDFLAGS
flags can be turned on by setting STAN_CPP_OPTIMS=true
in the make/local
file. These are tested compiler and link-time optimizations that can speed up execution
of certain models. We have observed speedups up to 15 percent, but this depends on the model, operating system
and hardware used. The use of these flags does considerably slow down compilation, so they are not used by default.
3.5.1 Optimizing by ignoring range checks
When assigning or reading from with vectors, row_vectors, matrices or arrays using indexing, Stan performs cheks that a supplied index is valid (not out of range). These check avoids segmentation faults and other runtime errors that can be difficult to debug.
For some models these checks can represent a significant part of the models execution time. By setting the
STAN_NO_RANGE_CHECKS=true
makefile flag in the make/local
file the range checks can be removed.
Use this flag with caution (only once the indexing has been validated). In case of any unexpected behavior
remove the flag for easier debugging.