jlrs

jlrs is a crate that provides access to most of the Julia C API, it can be used to embed Julia in Rust applications and to use functionality it provides when writing ccallable functions in Rust. Currently this crate is only tested in combination with Julia 1.6 and 1.9, but also supports Julia 1.7, 1.8 and 1.10. Using the current stable version is highly recommended. The minimum supported Rust version is currently 1.65.

The documentation assumes you're already familiar with the Julia and Rust programming languages.

An incomplete list of features that are currently supported by jlrs:

• Access arbitrary Julia modules and their content.
• Call Julia functions, including functions that take keyword arguments.
• Handle exceptions or convert them to an error message, optionally with color.
• Include and call your own Julia code.
• Use a custom system image.
• Create values that Julia can use, and convert them back to Rust, from Rust.
• Access the type information and fields of values. The contents of inline and bits-union fields can be accessed directly.
• Create and use n-dimensional arrays. The jlrs-ndarray feature can be enabled for integration with ndarray.
• Map Julia structs to Rust structs, the Rust implementation can be generated with the JlrsCore package.
• Structs that can be mapped to Rust include those with type parameters and bits unions.
• Use Julia from multiple threads with an async runtime, these runtimes support scheduling Julia Tasks and awaiting them without blocking the runtime thread.
• Export Rust types, methods and functions to Julia with the julia_module macro; libraries that use this macro can be compiled with BinaryBuilder and distributed as JLLs.

NB: Active development happens on the dev branch, the master branch points to the most recently released version.

Prerequisites

Julia must be installed before jlrs can be used, jlrs is compatible with Julia 1.6 up to and including Julia 1.10. The JlrsCore package must also have been installed, if this is not the case it will automatically be added when jlrs is initialized by default. jlrs has not been tested with juliaup yet on Linux and macOS.

Linux

The recommended way to install Julia is to download the binaries from the official website, which is distributed in an archive containing a directory called julia-x.y.z. This directory contains several other directories, including a bin directory containing the julia executable.

During compilation, the paths to the header and library are normally detected automatically by executing the command which julia. The path to julia.h must be $(which julia)/../include/julia/julia.h and the path to the library $(which julia)/../lib/libjulia.so. If you want to override this default behaviour the JULIA_DIR environment variable must be set to the path to the appropriate julia.x-y-z directory, in this case $JULIA_DIR/include/julia/julia.h and $JULIA_DIR/lib/libjulia.so are used instead.

In order to be able to load libjulia.so this file must be on the library search path. If this is not the case you must add /path/to/julia-x.y.z/lib to the LD_LIBRARY_PATH environment variable. When the uv feature is enabled, /path/to/julia-x.y.z/lib/julia must also be added to LD_LIBRARY_PATH. The latter path should not be added to the default path because this can break tools installed on your system.

Windows

Julia can be installed using juliaup, or with the installer or portable installation downloaded from the official website. In the first case, Julia has been likely installed in %USERPROFILE%\.julia\juliaup\julia-x.y.z+0~x64, while using the installer or extracting allows you to pick the destination. After installation or extraction a folder called Julia-x.y.z exists, which contains several folders including a bin folder containing julia.exe. The path to the bin folder must be added to the Path environment variable.

Julia is automatically detected by executing the command where julia. If this returns multiple locations the first one is used. The default can be overridden by setting the JULIA_DIR environment variable. This doesn't work correctly with juliaup, in this case the environment variable must be set.

macOS

Follow the instructions for Linux, but replace LD_LIBRARY_PATH with DYLD_LIBRARY_PATH.

Features

Most functionality of jlrs is only available if the proper features are enabled. These features generally belong to one of three categories: versions, runtimes and utilities.

Versions

There are minor incompatibilities between different versions of Julia, to ensure the correct bindings are used for a particular version of Julia you must enable a version features to use jlrs. The following version features currently exist:

• julia-1-6
• julia-1-7
• julia-1-8
• julia-1-9
• julia-1-10

Exactly one version feature must be enabled. If no version is enabled, or multiple are, jl-sys will fail to compile.

If you want your crate to be compatible with multiple versions of Julia, you should reexport these version features:

[features]
julia-1-6 = ["jlrs/julia-1-6"]
julia-1-7 = ["jlrs/julia-1-7"]
julia-1-8 = ["jlrs/julia-1-8"]
julia-1-9 = ["jlrs/julia-1-9"]
julia-1-10 = ["jlrs/julia-1-10"]

In this case you must provide this feature when you build or run your crate: cargo (build,run) --feature julia-1-8.

Runtimes

A runtime lets you embed Julia in a Rust application, the following features enable a runtime:

• sync-rt

  Enables the sync runtime, Julia. The sync runtime provides single-threaded, blocking access to the Julia C API.

• async-rt

  Enables the async runtime, AsyncJulia. The async runtime runs on a separate thread and can be used from multiple threads. Since Julia 1.9 it's possible to start the async runtime with multiple worker threads.

• tokio-rt and async-std-rt

  These features provide a backing runtime for the async runtime. The first uses tokio, the second async-std. The async-rt feature is automatically enabled when one of these features is enabled.

If you're writing a library, either one that will be called from Julia or one that will be used by a Rust application that embeds Julia, no runtime is required.

Utilities

In addition to these runtimes, the following utility features are available:

• prelude

  Provides a prelude module, jlrs::prelude. This feature is enabled by default.

• async

  Enable the features of the async runtime which don't depend on the backing runtime. This can be used in libraries which provide implementations of tasks that the async runtime can handle.

• jlrs-derive

  This feature should be used in combination with the code generation provided by the Reflect module in the JlrsCore package. This module lets you generate Rust implementations of Julia structs, this generated code uses custom derive macros made available with this feature to enable the safe conversion of data from Julia to Rust, and from Rust to Julia in some cases.

• jlrs-ndarray

  Access the content of a Julia array as an ArrayView or ArrayViewMut from ndarray.

• f16

  Adds support for working with Julia's Float16 type from Rust using half's f16 type.

• ccall

  Julia's ccall interface can be used to call functions written in Rust from Julia. No runtime can be used in this case because Julia has already been initialized, when this feature is enabled the CCall struct is available which offers the same functionality as the sync runtime without initializing Julia. The julia_module macro is provided to easily export functions, types, and data in combination with the macros from the Wrap module in the JlrsCore package.

• uv

  This feature enables the method CCall::uv_async_send, which can be used to wake a Julia AsyncCondition from Rust. The ccall feature is automically enabled when this feature is used.

• pyplot

  This feature lets you plot data using the Pyplot package and Gtk 3 from Rust.

• internal-types

  Provide extra managed types for types that are mostly used internally by Julia.

• extra-fields

  Provide extra field accessor methods for managed types.

• i686

  Link with a 32-bit build of Julia on Linux, only used for cross-compilation.

• windows

  Flag that must be enabled when cross-compiling for Windows from Linux.

• debug

  Link with a debug build of Julia on Linux.

• no-link

  Don't link Julia, linking can be skipped when writing libraries that will be loaded by Julia.

• yggdrasil

  Flag that must be enabled when compiling with BinaryBuilder.

You can enable all features except debug, i686, windows, no-link and yggdrasil by enabling the full feature. If you don't want to enable any runtimes either, you can use full-no-rt.

Using this crate

If you want to embed Julia in a Rust application, you must enable a runtime and a version feature:

jlrs = {version = "0.19.0", features = ["sync-rt", "julia-1-8"]}

jlrs = {version = "0.19.0", features = ["tokio-rt", "julia-1-8"]}

jlrs = {version = "0.19.0", features = ["async-std-rt", "julia-1-8"]}

When Julia is embedded in an application, it must be initialized before it can be used. The following snippet initializes the sync runtime:

use jlrs::prelude::*;

// Initializing Julia is unsafe because this can load arbitrary
// Julia code, and because it can race with other crates unrelated
// to jlrs. It returns an error if Julia has already been
// initialized.
let mut julia = unsafe { RuntimeBuilder::new().start().unwrap() };

// A StackFrame must be provided to ensure Julia's GC can be made aware
// of references to Julia data that exist in Rust code.
let mut frame = StackFrame::new();
let _instance = julia.instance(&mut frame);

To use the async runtime you must upgrade the RuntimeBuilder to an AsyncRuntimeBuilder by providing a backing runtime. Implementations for tokio and async-std are available if these features have been enabled. When starting the async runtime, you must declare the maximum number of concurrent tasks as a const generic.

For example, an async runtime backed by tokio and an unbounded channel, that supports 3 concurrent task can be initialized as follows if the tokio-rt feature is enabled:

use jlrs::prelude::*;

// Initializing Julia is unsafe for the same reasons as the sync runtime.
let (_julia, _task_handle) = unsafe {
    RuntimeBuilder::new()
        .async_runtime::<Tokio>()
        .start::<3>()
        .unwrap()
};

The async runtime can also be spawned as a blocking task on an existing executor:

use jlrs::prelude::*;

#[tokio::main]
async fn main() {
    // Initializing Julia is unsafe for the same reasons as the sync runtime.
    let (_julia, _task_handle) = unsafe {
        RuntimeBuilder::new()
            .async_runtime::<Tokio>()
            .start_async::<3>()
            .unwrap()
    };
}

If you're calling Rust from Julia everything has already been initialized. If the ccall feature is enabled CCall is available which provides the same functionality as the sync runtime.

Calling Julia from Rust

This section will focus on some topics that are common between the sync and async runtimes.

After initialization you have an instance of Julia or AsyncJulia, both provide a method called include that lets you include files with custom Julia code. In order to create Julia data and call Julia functions, a scope must be created first.

When the sync runtime is used this can be done by calling the method Julia::scope. This method takes a closure with a single argument, a GcFrame (frame). This frame can be used to access Julia data, and ensure it's not freed by the GC while it's accessible from Rust.

The async runtime can't create a new scope directly, AsyncJulia is a handle to the async runtime which runs on another thread. Instead, the async runtime deals with tasks, each task runs in its own scope. The simplest kind of task is a blocking task, which can be executed by calling AsyncJulia::blocking_task. This method accepts any closure Julia::scope can handle with the additional requirement that it must be Send and Sync. It's called a blocking task because the thread that executes this task is blocked while executing it. The other kinds of tasks that the async runtime can handle will be introduced later.

Inside the closure provided to Julia::scope or AsyncJulia::blocking_task it's possible to interact with Julia. Global Julia data can be accessed through its module system, the methods Module::main, Module::base, and Module::core can be used to access the Main, Base, and Core modules respectively. The contents of these modules can then be accessed by calling Module::function which returns a Function, Module::global which returns a Value, and Module::submodule which returns another Module. These types are examples of managed types, handles to data owned by Julia's GC. Most functionality in jlrs is provided through methods implemented by managed types.

The most generic managed type is Value, all other managed types can always be converted to a Value. It provides several methods to allocate new Julia data. The simplest one is Value::eval_string, which evaluates the contents of the string passed to it and returns the result as a Value. For example, you can evaluate 2 to convert it to Value. In practice, this method should rarely be used. It can be used to evaluate simple function calls like sqrt(2), but it must be parsed, compiled, and can't take any non-literal arguments. Its most important use-case is importing installed and standard library packages by evaluating an import or using statement.

A more interesting method, Value::new, can be used with data of any type that implements IntoJulia. This trait is implemented by primitive types like i8 and char. Any type that implements IntoJulia also implements Unbox which is used to extract the contents of a Value. Managed types like Array don't implement IntoJulia or Unbox, if they can be created from Rust they provide methods to do so.

As a simple example, let's convert two numbers to Julia values and add them:

use jlrs::prelude::*;

// Initializing Julia is unsafe because it can race with another crate that does
// the same.
let mut julia = unsafe { RuntimeBuilder::new().start().unwrap() };
let mut frame = StackFrame::new();
let mut julia = julia.instance(&mut frame);

let res = julia.scope(|mut frame| {
    // Create the two arguments.
    let i = Value::new(&mut frame, 2u64);
    let j = Value::new(&mut frame, 1u32);

    // The `+` function can be found in the base module.
    let func = Module::base(&frame).function(&mut frame, "+")?;

    // Call the function and unbox the result as a `u64`. The result of the function
    // call is a nested `Result`; the outer error doesn't contain to any Julia
    // data, while the inner error contains the exception if one is thrown. Here the
    // exception is converted to the outer error type by calling `into_jlrs_result`, this new
    // error contains the error message Julia would have shown.
    unsafe { func.call2(&mut frame, i, j) }
        .into_jlrs_result()?
        .unbox::<u64>()
}).unwrap();

assert_eq!(res, 3);

Evaluating raw code and calling Julia functions is always unsafe. Nothing prevents you from calling a function like nasaldemons() = unsafe_load(Ptr{Float64}(0x05391A445)). Similarly, mutating Julia data is unsafe because nothing prevents you from mutating data that shouldn't be mutated, e.g. a DataType. A full overview of the rules that you should keep in mind can be found in the safety module.

Async and persistent tasks

In addition to blocking tasks, the async runtime lets you execute async tasks which implement the AsyncTask trait, and persistent tasks which implement PersistentTask. Both of these traits are async traits.

An async task is similar to a blocking task, except that you must implement the async run method instead of providing a closure. This method takes an AsyncGcFrame. This new frame type not only provides access to the same features as GcFrame, it can also be used to call async methods provided by the CallAsync trait. These methods schedule a function call as a new Julia Task and can be awaited until this task has completed. The async runtime can switch to another task while the result is pending, allowing multiple tasks to run concurrently on a single thread.

The previous example can be rewritten as an async task:

use jlrs::prelude::*;

struct AdditionTask {
    a: u64,
    b: u32,
}

// Only the runtime thread can call the Julia C API, so the async
// trait methods of `AsyncTask` must not return a future that
// implements `Send` or `Sync`.
#[async_trait(?Send)]
impl AsyncTask for AdditionTask {
    // The type of the result of this task if it succeeds.
    type Output = u64;

    // The affinity of the task. Setting it to `DispatchAny` allows the
    // task to be dispatched to both the main thread and worker threads
    // if they are available.
    type Affinity = DispatchAny;

    // This async method replaces the closure from the previous examples,
    // an `AsyncGcFrame` can be used the same way as a `GcFrame` but also
    // can be used in combination with methods from the `CallAsync` trait.
    async fn run<'frame>(
        &mut self,
        mut frame: AsyncGcFrame<'frame>,
    ) -> JlrsResult<Self::Output> {
        let a = Value::new(&mut frame, self.a);
        let b = Value::new(&mut frame, self.b);

        let func = Module::base(&frame).function(&mut frame, "+")?;

        // CallAsync::call_async schedules the function call on another
        // thread and returns a Future that resolves when the scheduled
        // function has returned or thrown an error.
        unsafe { func.call_async(&mut frame, &mut [a, b]) }
            .await
            .into_jlrs_result()?
            .unbox::<u64>()
    }
}

While blocking and async tasks run once and return their result, a persistent task returns a handle to the task. This handle can be shared across threads and used to call its run method. In addition to a global and an async frame, this method can use the state and input data provided by the caller.

As an example, let's accumulate some number of values in a Julia array and return the sum of its contents:

use jlrs::prelude::*;

struct AccumulatorTask {
    n_values: usize,
}

struct AccumulatorTaskState<'state> {
    array: TypedArray<'state, 'static, usize>,
    offset: usize,
}

// Only the runtime thread can call the Julia C API, so the async trait
// methods of `PersistentTask` must not return a future that implements
// `Send` or `Sync`.
#[async_trait(?Send)]
impl PersistentTask for AccumulatorTask {
    // The type of the result of the task if it succeeds.
    type Output = usize;

    // The type of the task's internal state.
    type State<'state> = AccumulatorTaskState<'state>;

    // The type of the additional data that the task must be called with.
    type Input = usize;

    // The affinity of the task. Setting it to `DispatchAny` allows the
    // task to be dispatched to both the main thread and worker threads
    // if they are available.
    type Affinity = DispatchAny;

    // This method is called before the task can be called. Note that the
    // frame is not dropped until the task has completed, so the task's
    // internal state can contain Julia data rooted in this frame.
    async fn init<'frame>(
        &mut self,
        mut frame: AsyncGcFrame<'frame>,
    ) -> JlrsResult<Self::State<'frame>> {
        // A `Vec` can be moved from Rust to Julia if the element type
        // implements `IntoJulia`.
        let data = vec![0usize; self.n_values];
        let array = TypedArray::from_vec(&mut frame, data, self.n_values)?
            .into_jlrs_result()?;

        Ok(AccumulatorTaskState { array, offset: 0 })
    }

    // Whenever the task is called through its handle this method
    // is called. Unlike `init`, the frame that this method can use
    // is dropped after `run` returns.
    async fn run<'frame, 'state: 'frame>(
        &mut self,
        mut frame: AsyncGcFrame<'frame>,
        state: &mut Self::State<'state>,
        input: Self::Input,
    ) -> JlrsResult<Self::Output> {
        {
            // Array data can be directly accessed from Rust.
            // The data is tracked first to ensure it's not
            // already borrowed from Rust.
            unsafe {
                let mut tracked = state.array.track_exclusive()?;
                let mut data = tracked.bits_data_mut()?;
                data[state.offset] = input;
            };

            state.offset += 1;
            if (state.offset == self.n_values) {
                state.offset = 0;
            }
        }

        // Return the sum of the contents of `state.array`.
        unsafe {
            Module::base(&frame)
                .function(&mut frame, "sum")?
                .call1(&mut frame, state.array.as_value())
                .into_jlrs_result()?
                .unbox::<usize>()
        }
    }
}

Calling Rust from Julia

Julia's ccall interface can be used to call extern "C" functions defined in Rust. A function pointer can be cast to a void pointer and converted to a Value:

use jlrs::prelude::*;

// This function will be provided to Julia as a pointer, so its name can be mangled.
unsafe extern "C" fn call_me(arg: bool) -> isize {
    if arg {
        1
    } else {
        -1
    }
}

let mut frame = StackFrame::new();
let mut julia = unsafe { RuntimeBuilder::new().start().unwrap() };
let mut julia = julia.instance(&mut frame);

julia
    .scope(|mut frame| unsafe {
        // Cast the function to a void pointer
        let call_me_val = Value::new(&mut frame, call_me as *mut std::ffi::c_void);

        // Value::eval_string can be used to create new functions.
        let func = Value::eval_string(
            &mut frame,
            "myfunc(callme::Ptr{Cvoid})::Int = ccall(callme, Int, (Bool,), true)",
        )
        .into_jlrs_result()?;

        // Call the function and unbox the result.
        let result = func
            .call1(&mut frame, call_me_val)
            .into_jlrs_result()?
            .unbox::<isize>()?;

        assert_eq!(result, 1);

        Ok(())
    })
    .unwrap();

You can also use functions defined in cdylib libraries. In order to create such a library you need to add

[lib]
crate-type = ["cdylib"]

to your crate's Cargo.toml. It's also recommended to abort on panic:

[profile.release]
panic = "abort"

You must not enable any runtime features.

The easiest way to export Rust functions like call_me from the previous example is by using the julia_module macro. The content of the macro is converted to an initialization function that can be called from Julia to generate the module.

In Rust, the macro can be used like this:

julia_module! {
    become callme_init_fn;
    fn call_me(arg: bool) -> isize;
}

while on the Julia side things look like this:

module CallMe
using JlrsCore.Wrap

@wrapmodule("./path/to/libcallme.so", :callme_init_fn)

function __init__()
    @initjlrs
end
end

All Julia functions are automatically generated and have the same name as the exported function:

@assert CallMe.call_me(true) == 1
@assert CallMe.call_me(false) == -1

This macro has many more capabilities than just exporting functions, for more information see the documentation. A practical example that uses this macro is the rustfft-jl crate, which uses this macro to expose RustFFT to Julia. The recipe for BinaryBuilder can be found here.

While call_me doesn't call back into Julia, it is possible to call arbitrary functions from jlrs from a ccalled function. This will often require a Target, to create a target you must create an instance of CCall first.

Testing

The restriction that Julia can be initialized once must be taken into account when running tests that use jlrs. Because tests defined in a single crate are not guaranteed to be run from the same thread you must guarantee that each crate has only one test that initializes Julia. It's recommended you only use jlrs in integration tests because each top-level integration test file is treated as a separate crate.

use jlrs::prelude::*;

fn test_1(julia: &mut Julia) {
    // use instance
}
fn test_2(julia: &mut Julia) {
    // use instance
}

#[test]
fn call_tests() {
    let mut pending = unsafe { RuntimeBuilder::new().start().unwrap() };
    let mut frame = StackFrame::new();
    let mut julia = pending.instance(&mut frame);

    test_1(&mut julia);
    test_2(&mut julia);
}

Because AsyncJulia is thread-safe, it is possible to have multiple tests in a single crate when the async runtime is used:

use std::{num::NonZeroUsize, sync::Arc};

use jlrs::prelude::*;
use once_cell::sync::OnceCell;

fn init() -> Arc<AsyncJulia<Tokio>> {
    unsafe {
        Arc::new(
            RuntimeBuilder::new()
                .async_runtime::<Tokio>()
                .n_threads(4)
                .channel_capacity(NonZeroUsize::new_unchecked(32))
                .start::<4>()
                .expect("Could not init Julia")
                .0,
        )
    }
}

pub static JULIA: OnceCell<Arc<AsyncJulia<Tokio>>> = OnceCell::new();

#[test]
fn test_1() {
    let julia = JULIA.get_or_init(init);

    // use instance
}

#[test]
fn test_2() {
    let julia = JULIA.get_or_init(init);

    // use instance
}

Custom types

In order to map a struct in Rust to one in Julia you can derive several traits. You normally shouldn't need to implement these structs or traits manually. The reflect function defined in the JlrsCore.Reflect module can generate Rust structs whose layouts match their counterparts in Julia and automatically derive the supported traits.

The main restriction is that structs with atomic fields, and tuple or union fields with type parameters are not supported. The reason for this restriction is that the layout of such fields can be very different depending on the parameters in a way that can't be easily represented in Rust.

These custom types can also be used when you call Rust from Julia with ccall.