Embedding Julia

As we have seen in Calling C and Fortran Code, Julia has a simple and efficient way to call functions written in C. But there are situations where the opposite is needed: calling Julia functions from C code. This can be used to integrate Julia code into a larger C/C++ project, without the need to rewrite everything in C/C++. Julia has a C API to make this possible. As almost all programming languages have some way to call C functions, the Julia C API can also be used to build further language bridges (e.g. calling Julia from Python, Rust or C#). Even though Rust and C++ can use the C embedding API directly, both have packages helping with it, for C++ Jluna is useful.

High-Level Embedding

Note: This section covers embedding Julia code in C on Unix-like operating systems. For doing this on Windows, please see the section following this, High-Level Embedding on Windows with Visual Studio.

We start with a simple C program that initializes Julia and calls some Julia code:

#include <julia.h>
JULIA_DEFINE_FAST_TLS // only define this once, in an executable (not in a shared library) if you want fast code.

int main(int argc, char *argv[])
{
    /* required: setup the Julia context */
    jl_init();

    /* run Julia commands */
    jl_eval_string("print(sqrt(2.0))");

    /* strongly recommended: notify Julia that the
         program is about to terminate. this allows
         Julia time to cleanup pending write requests
         and run all finalizers
    */
    jl_atexit_hook(0);
    return 0;
}

In order to build this program you must add the path to the Julia header to the include path and link against libjulia. For instance, when Julia is installed to $JULIA_DIR, one can compile the above test program test.c with gcc using:

gcc -o test -fPIC -I$JULIA_DIR/include/julia -L$JULIA_DIR/lib -Wl,-rpath,$JULIA_DIR/lib test.c -ljulia

Alternatively, look at the embedding.c program in the Julia source tree in the test/embedding/ folder. The file cli/loader_exe.c program is another simple example of how to set jl_options options while linking against libjulia.

The first thing that must be done before calling any other Julia C function is to initialize Julia. This is done by calling jl_init, which tries to automatically determine Julia's install location. If you need to specify a custom location, or specify which system image to load, use jl_init_with_image instead.

The second statement in the test program evaluates a Julia statement using a call to jl_eval_string.

Before the program terminates, it is strongly recommended that jl_atexit_hook is called. The above example program calls this just before returning from main.

>>> julia=CDLL('./libjulia.dylib',RTLD_GLOBAL)
>>> julia.jl_init.argtypes = []
>>> julia.jl_init()
250593296

Using julia-config to automatically determine build parameters

The script julia-config.jl was created to aid in determining what build parameters are required by a program that uses embedded Julia. This script uses the build parameters and system configuration of the particular Julia distribution it is invoked by to export the necessary compiler flags for an embedding program to interact with that distribution. This script is located in the Julia shared data directory.

Example

#include <julia.h>

int main(int argc, char *argv[])
{
    jl_init();
    (void)jl_eval_string("println(sqrt(2.0))");
    jl_atexit_hook(0);
    return 0;
}

On the command line

A simple use of this script is from the command line. Assuming that julia-config.jl is located in /usr/local/julia/share/julia, it can be invoked on the command line directly and takes any combination of three flags:

/usr/local/julia/share/julia/julia-config.jl
Usage: julia-config [--cflags|--ldflags|--ldlibs]

If the above example source is saved in the file embed_example.c, then the following command will compile it into an executable program on Linux and Windows (MSYS2 environment). On macOS, substitute clang for gcc.:

/usr/local/julia/share/julia/julia-config.jl --cflags --ldflags --ldlibs | xargs gcc embed_example.c

Use in Makefiles

In general, embedding projects will be more complicated than the above example, and so the following allows general makefile support as well – assuming GNU make because of the use of the shell macro expansions. Furthermore, although julia-config.jl is usually in the /usr/local directory, if it isn't, then Julia itself can be used to find julia-config.jl, and the makefile can take advantage of this. The above example is extended to use a makefile:

JL_SHARE = $(shell julia -e 'print(joinpath(Sys.BINDIR, Base.DATAROOTDIR, "julia"))')
CFLAGS   += $(shell $(JL_SHARE)/julia-config.jl --cflags)
CXXFLAGS += $(shell $(JL_SHARE)/julia-config.jl --cflags)
LDFLAGS  += $(shell $(JL_SHARE)/julia-config.jl --ldflags)
LDLIBS   += $(shell $(JL_SHARE)/julia-config.jl --ldlibs)

all: embed_example

Now the build command is simply make.

High-Level Embedding on Windows with Visual Studio

If the JULIA_DIR environment variable hasn't been setup, add it using the System panel before starting Visual Studio. The bin folder under JULIA_DIR should be on the system PATH.

We start by opening Visual Studio and creating a new Console Application project. Open the 'stdafx.h' header file, and add the following lines at the end:

#include <julia.h>

Then, replace the main() function in the project with this code:

int main(int argc, char *argv[])
{
    /* required: setup the Julia context */
    jl_init();

    /* run Julia commands */
    jl_eval_string("print(sqrt(2.0))");

    /* strongly recommended: notify Julia that the
         program is about to terminate. this allows
         Julia time to cleanup pending write requests
         and run all finalizers
    */
    jl_atexit_hook(0);
    return 0;
}

The next step is to set up the project to find the Julia include files and the libraries. It's important to know whether the Julia installation is 32- or 64-bit. Remove any platform configuration that doesn't correspond to the Julia installation before proceeding.

Using the project Properties dialog, go to C/C++ | General and add $(JULIA_DIR)\include\julia\ to the Additional Include Directories property. Then, go to the Linker | General section and add $(JULIA_DIR)\lib to the Additional Library Directories property. Finally, under Linker | Input, add libjulia.dll.a;libopenlibm.dll.a; to the list of libraries.

At this point, the project should build and run.

Converting Types

Real applications will not only need to execute expressions, but also return their values to the host program. jl_eval_string returns a jl_value_t*, which is a pointer to a heap-allocated Julia object. Storing simple data types like Float64 in this way is called boxing, and extracting the stored primitive data is called unboxing. Our improved sample program that calculates the square root of 2 in Julia and reads back the result in C has a body that now contains this code:

jl_value_t *ret = jl_eval_string("sqrt(2.0)");

if (jl_typeis(ret, jl_float64_type)) {
    double ret_unboxed = jl_unbox_float64(ret);
    printf("sqrt(2.0) in C: %e \n", ret_unboxed);
}
else {
    printf("ERROR: unexpected return type from sqrt(::Float64)\n");
}

In order to check whether ret is of a specific Julia type, we can use the jl_isa, jl_typeis, or jl_is_... functions. By typing typeof(sqrt(2.0)) into the Julia shell we can see that the return type is Float64 (double in C). To convert the boxed Julia value into a C double the jl_unbox_float64 function is used in the above code snippet.

Corresponding jl_box_... functions are used to convert the other way:

jl_value_t *a = jl_box_float64(3.0);
jl_value_t *b = jl_box_float32(3.0f);
jl_value_t *c = jl_box_int32(3);

As we will see next, boxing is required to call Julia functions with specific arguments.

Calling Julia Functions

While jl_eval_string allows C to obtain the result of a Julia expression, it does not allow passing arguments computed in C to Julia. For this you will need to invoke Julia functions directly, using jl_call:

jl_function_t *func = jl_get_function(jl_base_module, "sqrt");
jl_value_t *argument = jl_box_float64(2.0);
jl_value_t *ret = jl_call1(func, argument);

In the first step, a handle to the Julia function sqrt is retrieved by calling jl_get_function. The first argument passed to jl_get_function is a pointer to the Base module in which sqrt is defined. Then, the double value is boxed using jl_box_float64. Finally, in the last step, the function is called using jl_call1. jl_call0, jl_call2, and jl_call3 functions also exist, to conveniently handle different numbers of arguments. To pass more arguments, use jl_call:

jl_value_t *jl_call(jl_function_t *f, jl_value_t **args, int32_t nargs)

Its second argument args is an array of jl_value_t* arguments and nargs is the number of arguments.

There is also an alternative, possibly simpler, way of calling Julia functions and that is via @cfunction. Using @cfunction allows you to do the type conversions on the Julia side, which is typically easier than doing it on the C side. The sqrt example above would with @cfunction be written as:

double (*sqrt_jl)(double) = jl_unbox_voidpointer(jl_eval_string("@cfunction(sqrt, Float64, (Float64,))"));
double ret = sqrt_jl(2.0);

where we first define a C callable function in Julia, extract the function pointer from it, and finally call it. In addition to simplifying type conversions by doing them in the higher-level language, calling Julia functions via @cfunction pointers eliminates the dynamic-dispatch overhead required by jl_call (for which all of the arguments are "boxed"), and should have performance equivalent to native C function pointers.

Memory Management

As we have seen, Julia objects are represented in C as pointers of type jl_value_t*. This raises the question of who is responsible for freeing these objects.

Typically, Julia objects are freed by the garbage collector (GC), but the GC does not automatically know that we are holding a reference to a Julia value from C. This means the GC can free objects out from under you, rendering pointers invalid.

The GC will only run when new Julia objects are being allocated. Calls like jl_box_float64 perform allocation, but allocation might also happen at any point in running Julia code.

When writing code that embeds Julia, it is generally safe to use jl_value_t* values in between jl_... calls (as GC will only get triggered by those calls). But in order to make sure that values can survive jl_... calls, we have to tell Julia that we still hold a reference to Julia root values, a process called "GC rooting". Rooting a value will ensure that the garbage collector does not accidentally identify this value as unused and free the memory backing that value. This can be done using the JL_GC_PUSH macros:

jl_value_t *ret = jl_eval_string("sqrt(2.0)");
JL_GC_PUSH1(&ret);
// Do something with ret
JL_GC_POP();

The JL_GC_POP call releases the references established by the previous JL_GC_PUSH. Note that JL_GC_PUSH stores references on the C stack, so it must be exactly paired with a JL_GC_POP before the scope is exited. That is, before the function returns, or control flow otherwise leaves the block in which the JL_GC_PUSH was invoked.

Several Julia values can be pushed at once using the JL_GC_PUSH2 to JL_GC_PUSH6 macros:

JL_GC_PUSH2(&ret1, &ret2);
// ...
JL_GC_PUSH6(&ret1, &ret2, &ret3, &ret4, &ret5, &ret6);

To push an array of Julia values one can use the JL_GC_PUSHARGS macro, which can be used as follows:

jl_value_t **args;
JL_GC_PUSHARGS(args, 2); // args can now hold 2 `jl_value_t*` objects
args[0] = some_value;
args[1] = some_other_value;
// Do something with args (e.g. call jl_... functions)
JL_GC_POP();

Each scope must have only one call to JL_GC_PUSH*, and should be paired with only a single JL_GC_POP call. If all necessary variables you want to root cannot be pushed by a one single call to JL_GC_PUSH*, or if there are more than 6 variables to be pushed and using an array of arguments is not an option, then one can use inner blocks:

jl_value_t *ret1 = jl_eval_string("sqrt(2.0)");
JL_GC_PUSH1(&ret1);
jl_value_t *ret2 = 0;
{
    jl_function_t *func = jl_get_function(jl_base_module, "exp");
    ret2 = jl_call1(func, ret1);
    JL_GC_PUSH1(&ret2);
    // Do something with ret2.
    JL_GC_POP();    // This pops ret2.
}
JL_GC_POP();    // This pops ret1.

Note that it is not necessary to have valid jl_value_t* values before calling JL_GC_PUSH*. It is fine to have a number of them initialized to NULL, pass those to JL_GC_PUSH* and then create the actual Julia values. For example:

jl_value_t *ret1 = NULL, *ret2 = NULL;
JL_GC_PUSH2(&ret1, &ret2);
ret1 = jl_eval_string("sqrt(2.0)");
ret2 = jl_eval_string("sqrt(3.0)");
// Use ret1 and ret2
JL_GC_POP();

If it is required to hold the pointer to a variable between functions (or block scopes), then it is not possible to use JL_GC_PUSH*. In this case, it is necessary to create and keep a reference to the variable in the Julia global scope. One simple way to accomplish this is to use a global IdDict that will hold the references, avoiding deallocation by the GC. However, this method will only work properly with mutable types.

// This functions shall be executed only once, during the initialization.
jl_value_t* refs = jl_eval_string("refs = IdDict()");
jl_function_t* setindex = jl_get_function(jl_base_module, "setindex!");

...

// `var` is the variable we want to protect between function calls.
jl_value_t* var = 0;

...

// `var` is a `Vector{Float64}`, which is mutable.
var = jl_eval_string("[sqrt(2.0); sqrt(4.0); sqrt(6.0)]");

// To protect `var`, add its reference to `refs`.
jl_call3(setindex, refs, var, var);

If the variable is immutable, then it needs to be wrapped in an equivalent mutable container or, preferably, in a RefValue{Any} before it is pushed to IdDict. In this approach, the container has to be created or filled in via C code using, for example, the function jl_new_struct. If the container is created by jl_call*, then you will need to reload the pointer to be used in C code.

// This functions shall be executed only once, during the initialization.
jl_value_t* refs = jl_eval_string("refs = IdDict()");
jl_function_t* setindex = jl_get_function(jl_base_module, "setindex!");
jl_datatype_t* reft = (jl_datatype_t*)jl_eval_string("Base.RefValue{Any}");

...

// `var` is the variable we want to protect between function calls.
jl_value_t* var = 0;

...

// `var` is a `Float64`, which is immutable.
var = jl_eval_string("sqrt(2.0)");

// Protect `var` until we add its reference to `refs`.
JL_GC_PUSH1(&var);

// Wrap `var` in `RefValue{Any}` and push to `refs` to protect it.
jl_value_t* rvar = jl_new_struct(reft, var);
JL_GC_POP();

jl_call3(setindex, refs, rvar, rvar);

The GC can be allowed to deallocate a variable by removing the reference to it from refs using the function delete!, provided that no other reference to the variable is kept anywhere:

jl_function_t* delete = jl_get_function(jl_base_module, "delete!");
jl_call2(delete, refs, rvar);

As an alternative for very simple cases, it is possible to just create a global container of type Vector{Any} and fetch the elements from that when necessary, or even to create one global variable per pointer using

jl_module_t *mod = jl_main_module;
jl_sym_t *var = jl_symbol("var");
jl_binding_t *bp = jl_get_binding_wr(mod, var, 1);
jl_checked_assignment(bp, mod, var, val);

Updating fields of GC-managed objects

The garbage collector also operates under the assumption that it is aware of every older-generation object pointing to a younger-generation one. Any time a pointer is updated breaking that assumption, it must be signaled to the collector with the jl_gc_wb (write barrier) function like so:

jl_value_t *parent = some_old_value, *child = some_young_value;
((some_specific_type*)parent)->field = child;
jl_gc_wb(parent, child);

It is in general impossible to predict which values will be old at runtime, so the write barrier must be inserted after all explicit stores. One notable exception is if the parent object has just been allocated and no garbage collection has run since then. Note that most jl_... functions can sometimes invoke garbage collection.

The write barrier is also necessary for arrays of pointers when updating their data directly. Calling jl_array_ptr_set is usually much preferred. But direct updates can be done. For example:

jl_array_t *some_array = ...; // e.g. a Vector{Any}
void **data = jl_array_data(some_array, void*);
jl_value_t *some_value = ...;
data[0] = some_value;
jl_gc_wb(jl_array_owner(some_array), some_value);

Controlling the Garbage Collector

There are some functions to control the GC. In normal use cases, these should not be necessary.

Working with Arrays

Julia and C can share array data without copying. The next example will show how this works.

Julia arrays are represented in C by the datatype jl_array_t*. Basically, jl_array_t is a struct that contains:

• Information about the datatype
• A pointer to the data block
• Information about the sizes of the array

To keep things simple, we start with a 1D array. Creating an array containing Float64 elements of length 10 can be done like this:

jl_value_t* array_type = jl_apply_array_type((jl_value_t*)jl_float64_type, 1);
jl_array_t* x          = jl_alloc_array_1d(array_type, 10);

Alternatively, if you have already allocated the array you can generate a thin wrapper around its data:

double *existingArray = (double*)malloc(sizeof(double)*10);
jl_array_t *x = jl_ptr_to_array_1d(array_type, existingArray, 10, 0);

The last argument is a boolean indicating whether Julia should take ownership of the data. If this argument is non-zero, the GC will call free on the data pointer when the array is no longer referenced.

In order to access the data of x, we can use jl_array_data:

double *xData = jl_array_data(x, double);

Now we can fill the array:

for (size_t i = 0; i < jl_array_nrows(x); i++)
    xData[i] = i;

Now let us call a Julia function that performs an in-place operation on x:

jl_function_t *func = jl_get_function(jl_base_module, "reverse!");
jl_call1(func, (jl_value_t*)x);

By printing the array, one can verify that the elements of x are now reversed.

Accessing Returned Arrays

If a Julia function returns an array, the return value of jl_eval_string and jl_call can be cast to a jl_array_t*:

jl_function_t *func  = jl_get_function(jl_base_module, "reverse");
jl_array_t *y = (jl_array_t*)jl_call1(func, (jl_value_t*)x);

Now the content of y can be accessed as before using jl_array_data. As always, be sure to keep a reference to the array while it is in use.

Multidimensional Arrays

Julia's multidimensional arrays are stored in memory in column-major order. Here is some code that creates a 2D array and accesses its properties:

// Create 2D array of float64 type
jl_value_t *array_type = jl_apply_array_type((jl_value_t*)jl_float64_type, 2);
int dims[] = {10,5};
jl_array_t *x  = jl_alloc_array_nd(array_type, dims, 2);

// Get array pointer
double *p = jl_array_data(x, double);
// Get number of dimensions
int ndims = jl_array_ndims(x);
// Get the size of the i-th dim
size_t size0 = jl_array_dim(x,0);
size_t size1 = jl_array_dim(x,1);

// Fill array with data
for(size_t i=0; i<size1; i++)
    for(size_t j=0; j<size0; j++)
        p[j + size0*i] = i + j;

Notice that while Julia arrays use 1-based indexing, the C API uses 0-based indexing (for example in calling jl_array_dim) in order to read as idiomatic C code.

Exceptions

Julia code can throw exceptions. For example, consider:

jl_eval_string("this_function_does_not_exist()");

This call will appear to do nothing. However, it is possible to check whether an exception was thrown:

if (jl_exception_occurred())
    printf("%s \n", jl_typeof_str(jl_exception_occurred()));

If you are using the Julia C API from a language that supports exceptions (e.g. Python, C#, C++), it makes sense to wrap each call into libjulia with a function that checks whether an exception was thrown, and then rethrows the exception in the host language.

Throwing Julia Exceptions

When writing Julia callable functions, it might be necessary to validate arguments and throw exceptions to indicate errors. A typical type check looks like:

if (!jl_typeis(val, jl_float64_type)) {
    jl_type_error(function_name, (jl_value_t*)jl_float64_type, val);
}

General exceptions can be raised using the functions:

void jl_error(const char *str);
void jl_errorf(const char *fmt, ...);

jl_error takes a C string, and jl_errorf is called like printf:

jl_errorf("argument x = %d is too large", x);

where in this example x is assumed to be an integer.

Thread-safety

In general, the Julia C API is not fully thread-safe. When embedding Julia in a multi-threaded application care needs to be taken not to violate the following restrictions:

• jl_init() may only be called once in the application life-time. The same applies to jl_atexit_hook(), and it may only be called after jl_init().
• jl_...() API functions may only be called from the thread in which jl_init() was called, or from threads started by the Julia runtime. Calling Julia API functions from user-started threads is not supported, and may lead to undefined behaviour and crashes.

The second condition above implies that you can not safely call jl_...() functions from threads that were not started by Julia (the thread calling jl_init() being the exception). For example, the following is not supported and will most likely segfault:

void *func(void*)
{
    // Wrong, jl_eval_string() called from thread that was not started by Julia
    jl_eval_string("println(Threads.threadid())");
    return NULL;
}

int main()
{
    pthread_t t;

    jl_init();

    // Start a new thread
    pthread_create(&t, NULL, func, NULL);
    pthread_join(t, NULL);

    jl_atexit_hook(0);
}

Instead, performing all Julia calls from the same user-created thread will work:

void *func(void*)
{
    // Okay, all jl_...() calls from the same thread,
    // even though it is not the main application thread
    jl_init();
    jl_eval_string("println(Threads.threadid())");
    jl_atexit_hook(0);
    return NULL;
}

int main()
{
    pthread_t t;
    // Create a new thread, which runs func()
    pthread_create(&t, NULL, func, NULL);
    pthread_join(t, NULL);
}

An example of calling the Julia C API from a thread started by Julia itself:

#include <julia/julia.h>
JULIA_DEFINE_FAST_TLS

double c_func(int i)
{
    printf("[C %08x] i = %d\n", pthread_self(), i);

    // Call the Julia sqrt() function to compute the square root of i, and return it
    jl_function_t *sqrt = jl_get_function(jl_base_module, "sqrt");
    jl_value_t* arg = jl_box_int32(i);
    double ret = jl_unbox_float64(jl_call1(sqrt, arg));

    return ret;
}

int main()
{
    jl_init();

    // Define a Julia function func() that calls our c_func() defined in C above
    jl_eval_string("func(i) = ccall(:c_func, Float64, (Int32,), i)");

    // Call func() multiple times, using multiple threads to do so
    jl_eval_string("println(Threads.threadpoolsize())");
    jl_eval_string("use(i) = println(\"[J $(Threads.threadid())] i = $(i) -> $(func(i))\")");
    jl_eval_string("Threads.@threads for i in 1:5 use(i) end");

    jl_atexit_hook(0);
}

If we run this code with 2 Julia threads we get the following output (note: the output will vary per run and system):

$ JULIA_NUM_THREADS=2 ./thread_example
2
[C 3bfd9c00] i = 1
[C 23938640] i = 4
[J 1] i = 1 -> 1.0
[C 3bfd9c00] i = 2
[J 1] i = 2 -> 1.4142135623730951
[C 3bfd9c00] i = 3
[J 2] i = 4 -> 2.0
[C 23938640] i = 5
[J 1] i = 3 -> 1.7320508075688772
[J 2] i = 5 -> 2.23606797749979

As can be seen, Julia thread 1 corresponds to pthread ID 3bfd9c00, and Julia thread 2 corresponds to ID 23938640, showing that indeed multiple threads are used at the C level, and that we can safely call Julia C API routines from those threads.