Futhark User’s Guide

Welcome to the documentation for the Futhark compiler and language. For a basic introduction, please see the Futhark website. To get started, read the page on Installation. Once the compiler has been installed, you might want to take a look at Basic Usage. This User’s Guide contains a Language Reference, but new Futhark programmers are probably better served by reading Parallel Programming in Futhark first.

Documentation for the built-in prelude is also available online.

The particularly interested reader may also want to peruse the publications, or the development blog.

Installation

There are two main ways to install the Futhark compiler: using a precompiled tarball or compiling from source. Both methods are discussed below. If you are using Linux, see Installing Futhark on Linux. If you are using Windows, make sure to read Setting up Futhark on Windows. If you are using macOS, read Using OpenCL or CUDA.

Futhark is also available via Nix. If you are using Nix, simply install the futhark derivation from Nixpkgs.

Dependencies

The Linux binaries we distribute are statically linked and should not require any special libraries installed system-wide.

When building from source on Linux and macOS, you will need to have the gmp and tinfo libraries installed. These are pretty common, so you may already have them. On Debian-like systems (e.g. Ubuntu), use:

sudo apt install libtinfo-dev libgmp-dev

If you install Futhark via a package manager (e.g. Homebrew, Nix, or AUR), you shouldn’t need to worry about any of this.

Actually running the output of the Futhark compiler may require additional dependencies, for example an OpenCL library and GPU driver. See the documentation for the respective compiler backends.

Compiling from source

The recommended way to compile Futhark is with the Haskell Tool Stack, which handles dependencies and compilation of the Futhark compiler. You will therefore need to install the stack tool. Fortunately, the stack developers provide ample documentation about installing Stack on a multitude of operating systems. If you’re lucky, it may even be in your local package repository.

You can either retrieve a source release tarball or perform a checkout of our Git repository:

$ git clone https://github.com/diku-dk/futhark.git

This will create a directory futhark, which you must enter:

$ cd futhark

To get all the prerequisites for building the Futhark compiler (including, if necessary, the appropriate version of the Haskell compiler), run:

$ stack setup

Note that this will not install anything system-wide and will have no effect outside the Futhark build directory. Now you can run the following command to build the Futhark compiler, including all dependencies:

$ stack build

The Futhark compiler and its tools will now be built. This step typically requires at least 8GiB of memory. You may be able to build it on a smaller machine by adding the --fast option, although the resulting Futhark compiler binary will run slower.

After building, you can copy the binaries to your $HOME/.local/bin directory by running:

$ stack install

Note that this does not install the Futhark manual pages.

Compiling with cabal

You can also compile Futhark with cabal. If so, you must install an appropriate version of GHC (usually the newest) and cabal yourself, for example through your favourite package manager. On Linux, you can always use ghcup. Then clone the repository as listed above and run:

$ cabal update
$ cabal build

To install the Futhark binaries to a specific location, for example $HOME/.local/bin, run:

$ cabal install --install-method=copy  --overwrite-policy=always --installdir=$HOME/.local/bin/

Installing from a precompiled snapshot

Tarballs of binary releases can be found online, but are available only for very few platforms (as of this writing, only GNU/Linux on x86_64). See the enclosed README.md for installation instructions.

Furthermore, every day a program automatically clones the Git repository, builds the compiler, and packages a simple tarball containing the resulting binaries, built manpages, and a simple Makefile for installing. The implication is that these tarballs are not vetted in any way, nor more stable than Git HEAD at any particular moment in time. They are provided for users who wish to use the most recent code, but are unable to compile Futhark themselves.

We build such binary snapshots for the following operating systems:

Linux (x86_64)

futhark-nightly-linux-x86_64.tar.xz

Windows (x86_64)

futhark-nightly-windows-x86_64.zip

You will still likely need to make a C compiler (such as GCC) available on your own.

Installing Futhark on Linux

• Linuxbrew is a distribution-agnostic package manager that contains a formula for Futhark. If Linuxbrew is installed (which does not require root access), installation is as easy as:

  $ brew install futhark
  

  Note that as of this writing, Linuxbrew is hampered by limited compute resources for building packages, so the Futhark version may be a bit behind.

• Arch Linux users can use a futhark-nightly package or a regular futhark package.

• NixOS users can install the futhark derivation.

Otherwise (or if the version in the package system is too old), your best bet is to install from source or use a tarball, as described above.

Using OpenCL or CUDA

If you wish to use futhark opencl or futhark cuda, you must have the OpenCL or CUDA libraries installed, respectively. Consult your favourite search engine for instructions on how to do this on your distribution. It is usually not terribly difficult if you already have working GPU drivers.

For OpenCL, note that there is a distinction between the general OpenCL host library (OpenCL.so) that Futhark links against, and the Installable Client Driver (ICD) that OpenCL uses to actually talk to the hardware. You will need both. Working display drivers for the GPU does not imply that an ICD has been installed - they are usually in a separate package. Consult your favourite search engine for details.

Installing Futhark on macOS

Futhark is available on Homebrew, and the latest release can be installed via:

$ brew install futhark

Or you can install the unreleased development version with:

$ brew install --HEAD futhark

This has to compile from source, so it takes a little while (20-30 minutes is common).

macOS ships with one OpenCL platform and various devices. One of these devices is always the CPU, which is not fully functional, and is never picked by Futhark by default. You can still select it manually with the usual mechanisms (see Executable Options), but it is unlikely to be able to run most Futhark programs. Depending on the system, there may also be one or more GPU devices, and Futhark will simply pick the first one as always. On multi-GPU MacBooks, this is is the low-power integrated GPU. It should work just fine, but you might have better performance if you use the dedicated GPU instead. On a Mac with an AMD GPU, this is done by passing -dAMD to the generated Futhark executable.

Setting up Futhark on Windows

The Futhark compiler itself is easily installed on Windows via stack (see above). If you are using the default Windows console, you may need to run chcp 65001 to make Unicode characters show up correctly.

It takes a little more work to make the OpenCL and PyOpenCL backends functional. This guide was last updated on the 5th of May 2016, and is for computers using 64-bit Windows along with CUDA 7.5 and Python 2.7 (Anaconda preferred).

Also Git for Windows is required for its Linux command line tools. If you have not marked the option to add them to path, there are instructions below how to do so. The GUI alternative to git, GitHub Desktop is optional and does not come with the required tools.

Setting up Futhark and OpenCL

1. Clone the Futhark repository to your hard drive.

2. Install Stack using the 64-bit installer. Compile the Futhark compiler as described in Installation.

3. For editing environment variables it is strongly recommended that you install the Rapid Environment Editor

4. For a Futhark compatible C/C++ compiler, that you will also need to install pyOpenCL later, install MingWpy. Do this using the pip install -i https://pypi.anaconda.org/carlkl/simple mingwpy command.

5. Assuming you have the latest Anaconda distribution as your primary one, it will get installed to a place such as C:\Users\UserName\Anaconda2\share\mingwpy. The pip installation will not add its bin or include directories to path.

   To do so, open the Rapid Environment Editor and add C:\Users\UserName\Anaconda2\share\mingwpy\bin to the system-wide PATH variable.

   If you have other MingW or GCC distributions, make sure MingWpy takes priority by moving its entry above the other distributions. You can also change which Python distribution is the default one using the same trick should you need so.

   If have done so correctly, typing where gcc in the command prompt should list the aforementioned MingWpy installation at the top or show only it.

   To finish the installation, add the C:\Users\UserName\Anaconda2\share\mingwpy\include to the CPATH environment variable (note: not PATH). Create the variable if necessary.

6. The header files and the .dll for OpenCL that comes with the CUDA 7.5 distribution also need to be installed into MingWpy. Go to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5\include and copy the CL directory into the MingWpy include directory.

   Next, go to C:\Program Files\NVIDIA Corporation\OpenCL and copy the OpenCL64.dll file into the MingWpy lib directory (it is next to include).

   The CUDA distribution also comes with the static OpenCL.lib, but trying to use that one instead of the OpenCL64.dll will cause programs compiled with futhark opencl to crash, so ignore it completely.

Now you should be able to compile with futhark opencl and run Futhark programs on the GPU.

Congratulations!

Setting up PyOpenCL

The following instructions are for how to setup the futhark-pyopencl backend.

First install Mako using pip install mako.

Also install PyPNG using pip install pypng (not stricly necessary, but some examples make use of it).

7. Clone the PyOpenCL repository to your hard drive. Do this instead of downloading the zip, as the zip will not contain some of the other repositories it links to and you will end up with missing header files.

8. If you have ignored the instructions and gotten Python 3.x instead 2.7, you will have to do some extra work.

   Edit .\pyopencl\compyte\ndarray\gen_elemwise.py and .\pyopencl\compyte\ndarray\test_gpu_ndarray.py and convert most Python 2.x style print statements to Python 3 syntax. Basically wrap print arguments in brackets “(..)” and ignore any lines containing StringIO >> operator.

   Otherwise just go to the next point.

9. Go into the repo directory and from the command line execute python configure.py.

   Edit siteconf.py to following:

   CL_TRACE = false
   CL_ENABLE_GL = false
   CL_INC_DIR = ['c:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v7.5\\include']
   CL_LIB_DIR = ['C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v7.5\\lib\\x64']
   CL_LIBNAME = ['OpenCL']
   CXXFLAGS = ['-std=c++0x']
   LDFLAGS = []
   

   Run the following commands:

   > python setup.py build_ext --compiler=mingw32
   > python setup.py install
   

If everything went in order, pyOpenCL should be installed on your machine now.

10. Lastly, Pygame needs to be installed. Again, not stricly necessary, but some examples make use of it. To do so on Windows, download pygame-1.9.2a0-cp27-none-win_amd64.whl from here. cp27 means Python 2.7 and win_amd64 means 64-bit Windows.

    Go to the directory you have downloaded the file and execute pip install pygame-1.9.2a0-cp27-none-win_amd64.whl from the command line.

Now you should be able to run the Game of Life example.

11. To run the makefiles, first setup make by going to the bin directory of MingWpy and making a copy of mingw32-make.exe. Then simply rename mingw32-make – Copy.exe or similar to make.exe. Now you will be able to run the makefiles.

    Also, if you have not selected to add the optional Linux command line tools to PATH during the Git for Windows installation, add the C:\Program Files\Git\usr\bin directory to PATH manually now.

12. This guide has been written off memory, so if you are having difficulties - ask on the issues page. There might be errors in it.

Basic Usage

Futhark contains several code generation backends. Each is provided as subcommand of the futhark binary. For example, futhark c compiles a Futhark program by translating it to sequential C code, while futhark pyopencl generates Python code with calls to the PyOpenCL library. The different compilers all contain the same frontend and optimisation pipeline - only the code generator is different. They all provide roughly the same command line interface, but there may be minor differences and quirks due to characteristics of the specific backends.

There are two main ways of compiling a Futhark program: to an executable (by using --executable, which is the default), and to a library (--library). Executables can be run immediately, but are useful mostly for testing and benchmarking. Libraries can be called from non-Futhark code.

Compiling to Executable

A Futhark program is stored in a file with the extension .fut. It can be compiled to an executable program as follows:

$ futhark c prog.fut

This makes use of the futhark c compiler, but any other will work as well. The compiler will automatically invoke cc to produce an executable binary called prog. If we had used futhark py instead of futhark c, the prog file would instead have contained Python code, along with a shebang for easy execution. In general, when compiling file foo.fut, the result will be written to a file foo (i.e. the extension will be stripped off). This can be overridden using the -o option. For more details on specific compilers, see their individual manual pages.

Executables generated by the various Futhark compilers share a common command-line interface, but may also individually support more options. When a Futhark program is run, execution starts at one of its entry points. By default, the entry point named main is run. An alternative entry point can be indicated by using the -e option. All entry point functions must be declared appropriately in the program (see Entry Points). If the entry point takes any parameters, these will be read from standard input in a subset of the Futhark syntax. A binary input format is also supported; see Binary Data Format. The result of the entry point is printed to standard output.

Only a subset of all Futhark values can be passed to an executable. Specifically, only primitives and arrays of primitive types are supported. In particular, nested tuples and arrays of tuples are not permitted. Non-nested tuples are supported are supported as simply flat values. This restriction is not present for Futhark programs compiled to libraries. If an entry point returns any such value, its printed representation is unspecified. As a special case, an entry point is allowed to return a flat tuple.

Instead of compiling, there is also an interpreter, accessible as futhark run and futhark repl. The latter is an interactive prompt, useful for experimenting with Futhark expressions. Be aware that the interpreter runs code very slowly.

Executable Options

All generated executables support the following options.

-h/--help

Print help text to standard output and exit.

-t FILE

Print the time taken to execute the program to the indicated file, an integral number of microseconds. The time taken to perform setup or teardown, including reading the input or writing the result, is not included in the measurement. See the documentation for specific compilers to see exactly what is measured.

-r RUNS

Run the specified entry point the given number of times (plus a warmup run). The program result is only printed once, after the last run. If combined with -t, one measurement is printed per run. This is a good way to perform benchmarking.

-D

Print debugging information on standard error. Exactly what is printed, and how it looks, depends on which Futhark compiler is used. This option may also enable more conservative (and slower) execution, such as frequently synchronising to check for errors.

-b

Print the result using the binary data format (Binary Data Format). For large outputs, this is significantly faster and takes up less space.

GPU Options

The following options are supported by executables generated with the GPU backends (opencl, pyopencl, and cuda).

-d DEVICE

Pick the first device whose name contains the given string. The special string #k, where k is an integer, can be used to pick the k-th device, numbered from zero.

--tuning=FILE

Load tuning options from the indicated tuning file. The file must contain lines of the form SIZE=VALUE, where each SIZE must be one of the sizes listed by the --print-sizes option (without size class), and the VALUE must be a non-negative integer. Extraneous spaces or blank lines are not allowed. A zero means to use the default size, whatever it may be. In case of duplicate assignments to the same size, the last one takes predecence. This is equivalent to passing each size setting on the command like using the --size option, but more convenient.

--print-sizes

Print a list of tunable sizes followed by their size class in parentheses, which indicates what they are used for.

--size=SIZE=VALUE

Set one of the tunable sizes to the given value. Using the --tuning option is more convenient.

OpenCL-specific Options

The following options are supported by executables generated with the OpenCL backends (opencl, pyopencl):

-P

Measure the time taken by various OpenCL operations (such as kernels) and print a summary at the end. Unfortunately, it is currently nontrivial (and manual) to relate these operations back to source Futhark code.

-p PLATFORM

Pick the first OpenCL platform whose name contains the given string. The special string #k, where k is an integer, can be used to pick the k-th platform, numbered from zero. If used in conjunction with -d, only the devices from matching platforms are considered.

-d DEVICE

Pick the first OpenCL device whose name contains the given string. The special string #k, where k is an integer, can be used to pick the k-th device, numbered from zero. If used in conjunction with -p, only the devices from matching platforms are considered.

--default-group-size INT

The default size of OpenCL workgroups that are launched. Capped to the hardware limit if necessary.

--default-num-groups INT

The default number of OpenCL workgroups that are launched.

--dump-opencl FILE

Don’t run the program, but instead dump the embedded OpenCL program to the indicated file. Useful if you want to see what is actually being executed.

--load-opencl FILE

Instead of using the embedded OpenCL program, load it from the indicated file. This is extremely unlikely to result in succesful execution unless this file is the result of a previous call to --dump-opencl (perhaps lightly modified).

--dump-opencl-binary FILE

Don’t run the program, but instead dump the compiled version of the embedded OpenCL program to the indicated file. On NVIDIA platforms, this will be PTX code. If this option is set, no entry point will be run.

--load-opencl-binary FILE

Load an OpenCL binary from the indicated file.

--build-option OPT

Add an additional build option to the string passed to clBuildProgram(). Refer to the OpenCL documentation for which options are supported. Be careful - some options can easily result in invalid results.

--list-devices

List all OpenCL devices and platforms available on the system.

There is rarely a need to use both -p and -d. For example, to run on the first available NVIDIA GPU, -p NVIDIA is sufficient, as there is likely only a single device associated with this platform. On *nix (including macOS), the clinfo tool (available in many package managers) can be used to determine which OpenCL platforms and devices are available on a given system. On Windows, CPU-z can be used.

CUDA-specific Options

The following options are supported by executables generated by the cuda backend:

--dump-cuda FILE

Don’t run the program, but instead dump the embedded CUDA program to the indicated file. Useful if you want to see what is actually being executed.

--load-cuda FILE

Instead of using the embedded CUDA program, load it from the indicated file. This is extremely unlikely to result in succesful execution unless this file is the result of a previous call to --dump-cuda (perhaps lightly modified).

--dump-ptx FILE

As --dump-cuda, but dumps the compiled PTX code instead.

--load-ptx FILE

Instead of using the embedded CUDA program, load compiled PTX code from the indicated file.

--nvrtc-option=OPT

Add the given option to the command line used to compile CUDA kernels with NVRTC. The list of supported options varies with the CUDA version but can be found in the NVRTC documentation.

For convenience, CUDA executables also accept the same --default-num-groups and --default-group-size options that the OpenCL backend uses. These then refer to grid size and thread block size, respectively.

Multicore options

The following options are supported by executables generated by the multicore backend:

--num-threads=INT

The number of threads used to run parallel operations. If set to a value less than 1, then the runtime system will use one thread per detected core.

Compiling to Library

While compiling a Futhark program to an executable is useful for testing, it is not suitable for production use. Instead, a Futhark program should be compiled into a reusable library in some target language, enabling integration into a larger program. Five of the Futhark compilers support this: futhark c, futhark opencl, futhark cuda, futhark py, and futhark pyopencl.

General Concerns

Futhark entry points are mapped to some form of function or method in the target language. Generally, an entry point taking n parameters will result in a function taking n parameters. Extra parameters may be added to pass in context data, or out-parameters for writing the result, for target languages that do not support multiple return values from functions.

Not all Futhark types can be mapped cleanly to the target language. Arrays of tuples, for example, are a common issue. In such cases, opaque types are used in the generated code. Values of these types cannot be directly inspected, but can be passed back to Futhark entry points. In the general case, these types will be named with a random hash. However, if you insert an explicit type annotation (and the type name contains only characters valid for identifiers for the used backend), the indicated name will be used. Note that arrays contain brackets, which are usually not valid in identifiers. Defining a simple type alias is the best way around this.

Generating C

A Futhark program futlib.fut can be compiled to reusable C code using either:

$ futhark c --library futlib.fut

Or:

$ futhark opencl --library futlib.fut

This produces two files in the current directory: futlib.c and futlib.h. If we wish (and are on a Unix system), we can then compile futlib.c to an object file like this:

$ gcc futlib.c -c

This produces a file futlib.o that can then be linked with the main application. Details of how to link the generated code with other C code is highly system-dependent, and outside the scope of this manual. On Unix, we can simply add futlib.o to the final compiler or linker command line:

$ gcc main.c -o main futlib.o

Depending on the Futhark backend you are using, you may need to add some linker flags. For example, futhark opencl requires -lOpenCL (-framework OpenCL on macOS). See the manual page for each compiler for details.

It is also possible to simply add the generated .c file to the C compiler command line used for compiling our whole program (here main.c):

$ gcc main.c -o main futlib.c

The downside of this approach is that the generated .c file may contain code that causes the C compiler to warn (for example, unused support code that is not needed by the Futhark program).

The generated header file (here, futlib.h) specifies the API, and is intended to be human-readable. See C API Reference for more information.

The basic usage revolves around creating a configuration object, which can then be used to obtain a context object, which must be passed whenever entry points are called.

The configuration object is created using the following function:

struct futhark_context_config *futhark_context_config_new();

Depending on the backend, various functions are generated to modify the configuration. The following is always available:

void futhark_context_config_set_debugging(struct futhark_context_config *cfg,
                                          int flag);

A configuration object can be used to create a context with the following function:

struct futhark_context *futhark_context_new(struct futhark_context_config *cfg);

Context creation may fail. Immediately after futhark_context_new(), call futhark_context_get_error() (see below), which will return a non-NULL error string if context creation failed. The API functions are all thread safe.

Memory management is entirely manual. Deallocation functions are provided for all types defined in the header file. Everything returned by an entry point must be manually deallocated.

For now, many internal errors, such as failure to allocate memory, will cause the function to abort() rather than return an error code. However, all application errors (such as bounds and array size checks) will produce an error code.

C with OpenCL

When generating C code with futhark opencl, you will need to link against the OpenCL library when linking the final binary:

$ gcc main.c -o main futlib.o -lOpenCL

When using the OpenCL backend, extra API functions are provided for directly accessing or providing the OpenCL objects used by Futhark. Take care when using these functions. In particular, a Futhark context can now be provided with the command queue to use:

struct futhark_context *futhark_context_new_with_command_queue(struct futhark_context_config *cfg, cl_command_queue queue);

As a cl_command_queue specifies an OpenCL device, this is also how manual platform and device selection is possible. A function is also provided for retrieving the command queue used by some Futhark context:

cl_command_queue futhark_context_get_command_queue(struct futhark_context *ctx);

This can be used to connect two separate Futhark contexts that have been loaded dynamically.

The raw cl_mem object underlying a Futhark array can be accessed with the function named futhark_values_raw_type, where type depends on the array in question. For example:

cl_mem futhark_values_raw_i32_1d(struct futhark_context *ctx, struct futhark_i32_1d *arr);

The array will be stored in row-major form in the returned memory object. The function performs no copying, so the cl_mem still belongs to Futhark, and may be reused for other purposes when the corresponding array is freed. A dual function can be used to construct a Futhark array from a cl_mem:

struct futhark_i32_1d *futhark_new_raw_i32_1d(struct futhark_context *ctx,
                                              cl_mem data,
                                              int offset,
                                              int dim0);

This function does copy the provided memory into fresh internally allocated memory. The array is assumed to be stored in row-major form offset bytes into the memory region.

See also futhark-opencl.

Generating Python

The futhark py and futhark pyopencl compilers both support generating reusable Python code, although the latter of these generates code of sufficient performance to be worthwhile. The following mentions options and parameters only available for futhark pyopencl. You will need at least PyOpenCL version 2015.2.

We can use futhark pyopencl to translate the program futlib.fut into a Python module futlib.py with the following command:

$ futhark pyopencl --library futlib.fut

This will create a file futlib.py, which contains Python code that defines a class named futlib. This class defines one method for each entry point function (see Entry Points) in the Futhark program. The methods take one parameter for each parameter in the corresponding entry point, and return a tuple containing a value for every value returned by the entry point. For entry points returning a single (non-tuple) value, just that value is returned (that is, single-element tuples are not returned).

After the class has been instantiated, these methods can be invoked to run the corresponding Futhark function. The constructor for the class takes various keyword parameters:

interactive=BOOL

If True (the default is False), show a menu of available OpenCL platforms and devices, and use the one chosen by the user.

platform_pref=STR

Use the first platform that contains the given string. Similar to the -p option for executables.

device_pref=STR

Use the first device that contains the given string. Similar to the -d option for executables.

Futhark arrays are mapped to either the Numpy ndarray type or the pyopencl.array type. Scalars are mapped to Numpy scalar types.

Reproducibility

The Futhark compiler is deterministic by design, meaning that repeatedly compiling the same program with the same compilation flags and using the same version of the compiler will produce identical output every time.

Note that this only applies to the code generated by the Futhark compiler itself. When compiling to an executable with one of the C backends (see Compiling to Executable), Futhark will invoke a C compiler that may not be perfectly reproducible. In such cases the generated .c and .h files will be reproducible, but the final executable may not.

Language Reference

This reference seeks to describe every construct in the Futhark language. It is not presented in a tutorial fashion, but rather intended for quick lookup and documentation of subtleties. For this reason, it is not written in a bottom-up manner, and some concepts may be used before they are fully defined. It is a good idea to have a basic grasp of Futhark (or some other functional programming language) before reading this reference. An ambiguous grammar is given for the full language. The text describes how ambiguities are resolved in practice (for example by applying rules of operator precedence).

This reference describes only the language itself. Documentation for the built-in prelude is available elsewhere.

Identifiers and Keywords

id          ::=  letter (letter | "_" | "'")* | "_" id
quals       ::=  (id ".")+
qualid      ::=  id | quals id
binop       ::=  opstartchar opchar*
qualbinop   ::=  binop | quals binop | "`" qualid "`"
fieldid     ::=  decimal | id
opstartchar ::=  "+" | "-" | "*" | "/" | "%" | "=" | "!" | ">" | "<" | "|" | "&" | "^"
opchar      ::=  opstartchar | "."
constructor ::=  "#" id

Many things in Futhark are named. When we are defining something, we give it an unqualified name (id). When referencing something inside a module, we use a qualified name (qualid). The constructor names of a sum type are identifiers prefixed with #, with no space afterwards. The fields of a record are named with fieldid. Note that a fieldid can be a decimal number. Futhark has three distinct name spaces: terms, module types, and types. Modules (including parametric modules) and values both share the term namespace.

Primitive Types and Values

literal ::=  intnumber | floatnumber | "true" | "false"

Boolean literals are written true and false. The primitive types in Futhark are the signed integer types i8, i16, i32, i64, the unsigned integer types u8, u16, u32, u64, the floating-point types f32, f64, as well as bool. An f32 is always a single-precision float and a f64 is a double-precision float.

int_type   ::=  "i8" | "i16" | "i32" | "i64" | "u8" | "u16" | "u32" | "u64"
float_type ::=  "f32" | "f64"

Numeric literals can be suffixed with their intended type. For example 42i8 is of type i8, and 1337e2f64 is of type f64. If no suffix is given, the type of the literal will be inferred based on its use. If the use is not constrained, integral literals will be assigned type i32, and decimal literals type f64. Hexadecimal literals are supported by prefixing with 0x, and binary literals by prefixing with 0b.

Floats can also be written in hexadecimal format such as 0x1.fp3, instead of the usual decimal notation. Here, 0x1.f evaluates to 1 15/16 and the p3 multiplies it by 2^3 = 8.

intnumber   ::=  (decimal | hexadecimal | binary) [int_type]
decimal     ::=  decdigit (decdigit |"_")*
hexadecimal ::=  0 ("x" | "X") hexdigit (hexdigit |"_")*
binary      ::=  0 ("b" | "B") bindigit (bindigit | "_")*

floatnumber      ::=  (pointfloat | exponentfloat | hexadecimalfloat) [float_type]
pointfloat       ::=  [intpart] fraction
exponentfloat    ::=  (intpart | pointfloat) exponent
hexadecimalfloat ::=  0 ("x" | "X") hexintpart hexfraction ("p"|"P") ["+" | "-"] decdigit+
intpart          ::=  decdigit (decdigit |"_")*
fraction         ::=  "." decdigit (decdigit |"_")*
hexintpart       ::=  hexdigit (hexdigit | "_")*
hexfraction      ::=  "." hexdigit (hexdigit |"_")*
exponent         ::=  ("e" | "E") ["+" | "-"] decdigit+

decdigit ::=  "0"..."9"
hexdigit ::=  decdigit | "a"..."f" | "A"..."F"
bindigit ::=  "0" | "1"

Compound Types and Values

type ::=    qualid
          | array_type
          | tuple_type
          | record_type
          | sum_type
          | function_type
          | type_application

Compound types can be constructed based on the primitive types. The Futhark type system is entirely structural, and type abbreviations are merely shorthands. The only exception is abstract types whose definition has been hidden via the module system (see Module System).

tuple_type ::=  "(" ")" | "(" type ("," type)+ ")"

A tuple value or type is written as a sequence of comma-separated values or types enclosed in parentheses. For example, (0, 1) is a tuple value of type (i32,i32). The elements of a tuple need not have the same type – the value (false, 1, 2.0) is of type (bool, i32, f64). A tuple element can also be another tuple, as in ((1,2),(3,4)), which is of type ((i32,i32),(i32,i32)). A tuple cannot have just one element, but empty tuples are permitted, although they are not very useful. Empty tuples are written () and are of type ().

array_type ::=  "[" [dim] "]" type
dim        ::=  qualid | decimal

An array value is written as a sequence of zero or more comma-separated values enclosed in square brackets: [1,2,3]. An array type is written as [d]t, where t is the element type of the array, and d is an integer or variable indicating the size. We can often elide d and write just [] (an anonymous size), in which case the size will be inferred. As an example, an array of three integers could be written as [1,2,3], and has type [3]i32. An empty array is written as [], and its type is inferred from its use. When writing Futhark values for such uses as futhark test (but not when writing programs), empty arrays are written empty([0]t) for an empty array of type [0]t. When using empty, all dimensions must be given a size, and at least one must be zero, e.g. empty([2][0]i32).

Multi-dimensional arrays are supported in Futhark, but they must be regular, meaning that all inner arrays must have the same shape. For example, [[1,2], [3,4], [5,6]] is a valid array of type [3][2]i32, but [[1,2], [3,4,5], [6,7]] is not, because there we cannot come up with integers m and n such that [m][n]i32 describes the array. The restriction to regular arrays is rooted in low-level concerns about efficient compilation. However, we can understand it in language terms by the inability to write a type with consistent dimension sizes for an irregular array value. In a Futhark program, all array values, including intermediate (unnamed) arrays, must be typeable.

sum_type ::=  constructor type* ("|" constructor type*)*

Sum types are anonymous in Futhark, and are written as the constructors separated by vertical bars. Each constructor consists of a #-prefixed name, followed by zero or more types, called its payload. Note: The current implementation of sum types is fairly inefficient, in that all possible constructors of a sum-typed value will be resident in memory. Avoid using sum types where multiple constructors have large payloads.

record_type ::=  "{" "}" | "{" fieldid ":" type ("," fieldid ":" type)* "}"

Records are mappings from field names to values, with the field names known statically. A tuple behaves in all respects like a record with numeric field names starting from zero, and vice versa. It is an error for a record type to name the same field twice.

type_application ::=  type type_arg | "*" type
type_arg         ::=  "[" [dim] "]" | type

A parametric type abbreviation can be applied by juxtaposing its name and its arguments. The application must provide as many arguments as the type abbreviation has parameters - partial application is presently not allowed. See Type Abbreviations for further details.

function_type ::=  param_type "->" type
param_type    ::=  type | "(" id ":" type ")"

Functions are classified via function types, but they are not fully first class. See Higher-order functions for the details.

stringlit  ::=  '"' stringchar '"'
stringchar ::=  <any source character except "\" or newline or quotes>

String literals are supported, but only as syntactic sugar for UTF-8 encoded arrays of u8 values. There is no character type in Futhark.

Declarations

A Futhark file or module consists of a sequence of declarations. Each declaration is processed in order, and a declaration can only refer to names bound by preceding declarations.

dec ::=    fun_bind | val_bind | type_bind | mod_bind | mod_type_bind
         | "open" mod_exp
         | "import" stringlit
         | "local" dec
         | "#[" attr "]" dec

The open declaration brings names defined in another module into scope (see also Module System). For the meaning of import, see Referring to Other Files. If a declaration is prefixed with local, whatever names it defines will not be visible outside the current module. In particular local open is used to bring names from another module into scope, without making those names available to users of the module being defined. In most cases, using module type ascription is a better idea.

Declaring Functions and Values

fun_bind ::=    ("let" | "entry") (id | "(" binop ")") type_param* pat+ [":" type] "=" exp
              | ("let" | "entry") pat binop pat [":" type] "=" exp

val_bind ::=  "let" id [":" type] "=" exp

Functions and values must be defined before they are used. A function declaration must specify the name, parameters, and body of the function:

let name params...: rettype = body

Hindley-Milner-style type inference is supported. A parameter may be given a type with the notation (name: type). Functions may not be recursive. You may put size annotations in the return type and parameter types; see Size Types. A function can be polymorphic by using type parameters, in the same way as for Type Abbreviations:

let reverse [n] 't (xs: [n]t): [n]t = xs[::-1]

Type parameters for a function do not need to cover the types of all parameters. The type checker will add more if necessary. For example, the following is well typed:

let pair 'a (x: a) y = (x, y)

A new type variable will be invented for the parameter y.

Shape and type parameters are not passed explicitly when calling function, but are automatically derived. If an array value v is passed for a type parameter t, all other arguments passed of type t must have the same shape as v. For example, consider the following definition:

let pair 't (x: t) (y: t) = (x, y)

The application pair [1] [2,3] will fail at run-time.

To simplify the handling of in-place updates (see In-place Updates), the value returned by a function may not alias any global variables.

User-Defined Operators

Infix operators are defined much like functions:

let (p1: t1) op (p2: t2): rt = ...

For example:

let (a:i32,b:i32) +^ (c:i32,d:i32) = (a+c, b+d)

We can also define operators by enclosing the operator name in parentheses and suffixing the parameters, as an ordinary function:

let (+^) (a:i32,b:i32) (c:i32,d:i32) = (a+c, b+d)

This is necessary when defining a polymorphic operator.

A valid operator name is a non-empty sequence of characters chosen from the string "+-*/%=!><&^". The fixity of an operator is determined by its first characters, which must correspond to a built-in operator. Thus, +^ binds like +, whilst *^ binds like *. The longest such prefix is used to determine fixity, so >>= binds like >>, not like >.

It is not permitted to define operators with the names && or || (although these as prefixes are accepted). This is because a user-defined version of these operators would not be short-circuiting. User-defined operators behave exactly like ordinary functions, except for being infix.

A built-in operator can be shadowed (i.e. a new + can be defined). This will result in the built-in polymorphic operator becoming inaccessible, except through the intrinsics module.

An infix operator can also be defined with prefix notation, like an ordinary function, by enclosing it in parentheses:

let (+) (x: i32) (y: i32) = x - y

This is necessary when defining operators that take type or shape parameters.

Entry Points

Apart from declaring a function with the keyword let, it can also be declared with entry. When the Futhark program is compiled any top-level function declared with entry will be exposed as an entry point. If the Futhark program has been compiled as a library, these are the functions that will be exposed. If compiled as an executable, you can use the --entry-point command line option of the generated executable to select the entry point you wish to run.

Any top-level function named main will always be considered an entry point, whether it is declared with entry or not.

The name of an entry point must not contain an apostrophe ('), even though that is normally permitted in Futhark identifiers.

Value Declarations

A named value/constant can be declared as follows:

let name: type = definition

The definition can be an arbitrary expression, including function calls and other values, although they must be in scope before the value is defined. A constant value may not have a unique type (see In-place updates). If the return type contains any anonymous sizes (see Size types), new existential sizes will be constructed for them.

Type Abbreviations

type_bind  ::=  "type" ["^" | "~"] id type_param* "=" type
type_param ::=  "[" id "]" | "'" id | "'~" id | "'^" id

Type abbreviations function as shorthands for the purpose of documentation or brevity. After a type binding type t1 = t2, the name t1 can be used as a shorthand for the type t2. Type abbreviations do not create distinct types: the types t1 and t2 are entirely interchangeable.

If the right-hand side of a type contains anonymous sizes, it must be declared “size-lifted” with type~. If it (potentially) contains a function, it must be declared “fully lifted” with type^. A lifted type can also contain anonymous sizes. Lifted types cannot be put in arrays. Fully lifted types cannot be returned from conditional or loop expressions.

A type abbreviation can have zero or more parameters. A type parameter enclosed with square brackets is a size parameter, and can be used in the definition as an array dimension size, or as a dimension argument to other type abbreviations. When passing an argument for a shape parameter, it must be enclosed in square brackets. Example:

type two_intvecs [n] = ([n]i32, [n]i32)

let x: two_intvecs [2] = (iota 2, replicate 2 0)

Size parameters work much like shape declarations for arrays. Like shape declarations, they can be elided via square brackets containing nothing. All size parameters must be used in the definition of the type abbreviation.

A type parameter prefixed with a single quote is a type parameter. It is in scope as a type in the definition of the type abbreviation. Whenever the type abbreviation is used in a type expression, a type argument must be passed for the parameter. Type arguments need not be prefixed with single quotes:

type two_vecs [n] 't = ([n]t, [n]t)
type two_intvecs [n] = two_vecs [n] i32
let x: two_vecs [2] i32 = (iota 2, replicate 2 0)

A size-lifted type parameter is prefixed with '~, and a fully lifted type parameter with '^. These have the same rules and restrictions as lifted type abbreviations.

Expressions

Expressions are the basic construct of any Futhark program. An expression has a statically determined type, and produces a value at runtime. Futhark is an eager/strict language (“call by value”).

The basic elements of expressions are called atoms, for example literals and variables, but also more complicated forms.

atom     ::=    literal
              | qualid ("." fieldid)*
              | stringlit
              | "(" ")"
              | "(" exp ")" ("." fieldid)*
              | "(" exp ("," exp)* ")"
              | "{" "}"
              | "{" field ("," field)* "}"
              | qualid "[" index ("," index)* "]"
              | "(" exp ")" "[" index ("," index)* "]"
              | quals "." "(" exp ")"
              | "[" exp ("," exp)* "]"
              | "[" exp [".." exp] "..." exp "]"
              | "(" qualbinop ")"
              | "(" exp qualbinop ")"
              | "(" qualbinop exp ")"
              | "(" ( "." field )+ ")"
              | "(" "." "[" index ("," index)* "]" ")"
exp      ::=    atom
              | exp qualbinop exp
              | exp exp
              | constructor exp*
              | exp ":" type
              | exp ":>" type
              | exp [ ".." exp ] "..." exp
              | exp [ ".." exp ] "..<" exp
              | exp [ ".." exp ] "..>" exp
              | "if" exp "then" exp "else" exp
              | "let" pat "=" exp "in" exp
              | "let" id "[" index ("," index)* "]" "=" exp "in" exp
              | "let" id type_param* pat+ [":" type] "=" exp "in" exp
              | "(" "\" pat+ [":" type] "->" exp ")"
              | "loop" pat [("=" exp)] loopform "do" exp
              | "#[" attr "]" exp
              | "unsafe" exp
              | "assert" atom atom
              | exp "with" "[" index ("," index)* "]" "=" exp
              | exp "with" fieldid ("." fieldid)* "=" exp
              | "match" exp ("case" pat "->" exp)+
field    ::=    fieldid "=" exp
              | id
pat      ::=    id
              | literal
              |  "_"
              | "(" ")"
              | "(" pat ")"
              | "(" pat ("," pat)+ ")"
              | "{" "}"
              | "{" fieldid ["=" pat] ("," fieldid ["=" pat])* "}"
              | constructor pat*
              | pat ":" type
loopform ::=    "for" id "<" exp
              | "for" pat "in" exp
              | "while" exp
index    ::=    exp [":" [exp]] [":" [exp]]
              | [exp] ":" exp [":" [exp]]
              | [exp] [":" exp] ":" [exp]

Some of the built-in expression forms have parallel semantics, but it is not guaranteed that the the parallel constructs in Futhark are evaluated in parallel, especially if they are nested in complicated ways. Their purpose is to give the compiler as much freedom and information is possible, in order to enable it to maximise the efficiency of the generated code.

Resolving Ambiguities

The above grammar contains some ambiguities, which in the concrete implementation is resolved via a combination of lexer and grammar transformations. For ease of understanding, they are presented here in natural text.

• An expression x.y may either be a reference to the name y in the module x, or the field y in the record x. Modules and values occupy the same name space, so this is disambiguated by the type of x.

• A type ascription (exp : type) cannot appear as an array index, as it conflicts with the syntax for slicing.

• In f [x], there is am ambiguity between indexing the array f at position x, or calling the function f with the singleton array x. We resolve this the following way:

  • If there is a space between f and the opening bracket, it is treated as a function application.

  • Otherwise, it is an array index operation.

• An expression (-x) is parsed as the variable x negated and enclosed in parentheses, rather than an operator section partially applying the infix operator -.

• The following table describes the precedence and associativity of infix operators. All operators in the same row have the same precedence. The rows are listed in increasing order of precedence. Note that not all operators listed here are used in expressions; nevertheless, they are still used for resolving ambiguities.

  Associativity	Operators
  left	,
  left	:, :>
  left	||
  left	&&
  left	<= >= > < == !=
  left	& ^ |
  left	<< >>
  left	+ -
  left	* / % // %%
  left	|>
  right	<|
  right	->
  left	juxtaposition

Patterns

We say that a pattern is irrefutable if it can never fail to match a value of the appropriate type. Concretely, this means that it does not require any specific sum type constructor (unless the type in question has only a single constructor), or any specific numeric or boolean literal. Patterns used in function parameters and let bindings must be irrefutable. Patterns used in case need not be irrefutable.

A pattern _ matches any value. A pattern consisting of a literal value (e.g. a numeric constant) matches exactly that value.

Semantics of Simple Expressions

literal

Evaluates to itself.

qualid

A variable name; evaluates to its value in the current environment.

stringlit

Evaluates to an array of type []i32 that contains the code points of the characters as integers.

()

Evaluates to an empty tuple.

( e )

Evaluates to the result of e.

(e1, e2, ..., eN)

Evaluates to a tuple containing N values. Equivalent to the record literal {0=e1, 1=e2, ..., N-1=eN}.

{f1, f2, ..., fN}

A record expression consists of a comma-separated sequence of field expressions. Each field expression defines the value of a field in the record. A field expression can take one of two forms:

f = e: defines a field with the name f and the value resulting from evaluating e.

f: defines a field with the name f and the value of the variable f in scope.

Each field may only be defined once.

a[i]

Return the element at the given position in the array. The index may be a comma-separated list of indexes instead of just a single index. If the number of indices given is less than the rank of the array, an array is returned. The index may be of any unsigned integer type.

The array a must be a variable name or a parenthesised expression. Furthermore, there may not be a space between a and the opening bracket. This disambiguates the array indexing a[i], from a [i], which is a function call with a literal array.

a[i:j:s]

Return a slice of the array a from index i to j, the former inclusive and the latter exclusive, taking every s-th element. The s parameter may not be zero. If s is negative, it means to start at i and descend by steps of size s to j (not inclusive). Slicing can be done only with expressions of type i32.

It is generally a bad idea for s to be non-constant. Slicing of multiple dimensions can be done by separating with commas, and may be intermixed freely with indexing.

If s is elided it defaults to 1. If i or j is elided, their value depends on the sign of s. If s is positive, i become 0 and j become the length of the array. If s is negative, i becomes the length of the array minus one, and j becomes minus one. This means that a[::-1] is the reverse of the array a.

In the general case, the size of the array produced by a slice is unknown (see Size types). In a few cases, the size is known statically:

• a[0:n] has size n

• a[:n] has size n

• a[0:n:1] has size n

• a[:n:1] has size n

This holds only if n is a variable or constant.

[x, y, z]

Create an array containing the indicated elements. Each element must have the same type and shape.

x..y...z

Construct a signed integer array whose first element is x and which proceeds with a stride of y-x until reaching z (inclusive). The ..y part can be elided in which case a stride of 1 is used. A run-time error occurs if z is less than x or y, or if x and y are the same value.

In the general case, the size of the array produced by a range is unknown (see Size types). In a few cases, the size is known statically:

• 1..2...n has size n

This holds only if n is a variable or constant.

x..y..<z

Construct a signed integer array whose first elements is x, and which proceeds upwards with a stride of y-x until reaching z (exclusive). The ..y part can be elided in which case a stride of 1 is used. A run-time error occurs if z is less than x or y, or if x and y are the same value.

• 0..1..<n has size n

• 0..<n has size n

This holds only if n is a variable or constant.

x..y..>z

Construct a signed integer array whose first elements is x, and which proceeds downwards with a stride of y-x until reaching z (exclusive). The ..y part can be elided in which case a stride of -1 is used. A run-time error occurs if z is greater than x or y, or if x and y are the same value.

e.f

Access field f of the expression e, which must be a record or tuple.

m.(e)

Evaluate the expression e with the module m locally opened, as if by open. This can make some expressions easier to read and write, without polluting the global scope with a declaration-level open.

x binop y

Apply an operator to x and y. Operators are functions like any other, and can be user-defined. Futhark pre-defines certain “magical” overloaded operators that work on many different types. Overloaded functions cannot be defined by the user. Both operands must have the same type. The predefined operators and their semantics are:

**

Power operator, defined for all numeric types.

//, %%

Division and remainder on integers, with rounding towards zero.

*, /, %, +, -

The usual arithmetic operators, defined for all numeric types. Note that / and % rounds towards negative infinity when used on integers - this is different from in C.

^, &, |, >>, <<

Bitwise operators, respectively bitwise xor, and, or, arithmetic shift right and left, and logical shift right. Shifting is undefined if the right operand is negative, or greater than or equal to the length in bits of the left operand.

Note that, unlike in C, bitwise operators have higher priority than arithmetic operators. This means that x & y == z is understood as (x & y) == z, rather than x & (y == z) as it would in C. Note that the latter is a type error in Futhark anyhow.

==, !=

Compare any two values of builtin or compound type for equality.

<, <=. >, >=

Company any two values of numeric type for equality.

x && y

Short-circuiting logical conjunction; both operands must be of type bool.

x || y

Short-circuiting logical disjunction; both operands must be of type bool.

f x

Apply the function f to the argument x.

#c x y z

Apply the sum type constructor #x to the payload x, y, and z. A constructor application is always assumed to be saturated, i.e. its entire payload provided. This means that constructors may not be partially applied.

e : t

Annotate that e is expected to be of type t, failing with a type error if it is not. If t is an array with shape declarations, the correctness of the shape declarations is checked at run-time.

Due to ambiguities, this syntactic form cannot appear as an array index expression unless it is first enclosed in parentheses. However, as an array index must always be of type i32, there is never a reason to put an explicit type ascription there.

e :> t

Coerce the size of e to t. The type of t must match the type of e, except that the sizes may be statically different. At run-time, it will be verified that the sizes are the same.

! x

Logical negation if x is of type bool. Bitwise negation if x is of integral type.

- x

Numerical negation of x, which must be of numeric type.

#[attr] e

Apply the given attribute to the expression. Attributes are an ad-hoc and optional mechanism for providing extra information, directives, or hints to the compiler. See Attributes for more information.

unsafe e

Elide safety checks and assertions (such as bounds checking) that occur during execution of e. This is useful if the compiler is otherwise unable to avoid bounds checks (e.g. when using indirect indexes), but you really do not want them there. Make very sure that the code is correct; eliding such checks can lead to memory corruption.

This construct is deprecated. Use the #[unsafe] attribute instead.

assert cond e

Terminate execution with an error if cond evaluates to false, otherwise produce the result of evaluating e. Unless e produces a value that is used subsequently (it can just be a variable), dead code elimination may remove the assertion.

a with [i] = e

Return a, but with the element at position i changed to contain the result of evaluating e. Consumes a.

r with f = e

Return the record r, but with field f changed to have value e. The type of the field must remain unchanged. Type inference is limited: r must have a completely known type up to f. This sometimes requires extra type annotations to make the type of r known.

if c then a else b

If c evaluates to true, evaluate a, else evaluate b.

Binding Expressions

let pat = e in body

Evaluate e and bind the result to the irrefutable pattern pat (see Patterns) while evaluating body. The in keyword is optional if body is a let expression.

let a[i] = v in body

Write v to a[i] and evaluate body. The given index need not be complete and can also be a slice, but in these cases, the value of v must be an array of the proper size. This notation is Syntactic sugar for let a = a with [i] = v in a.

let f params... = e in body

Bind f to a function with the given parameters and definition (e) and evaluate body. The function will be treated as aliasing any free variables in e. The function is not in scope of itself, and hence cannot be recursive.

loop pat = initial for x in a do loopbody

1. Bind pat to the initial values given in initial.

2. For each element x in a, evaluate loopbody and rebind pat to the result of the evaluation.

3. Return the final value of pat.

The = initial can be left out, in which case initial values for the pattern are taken from equivalently named variables in the environment. I.e., loop (x) = ... is equivalent to loop (x = x) = ....

loop pat = initial for x < n do loopbody

Equivalent to loop (pat = initial) for x in [0..1..<n] do loopbody.

loop pat = initial while cond do loopbody

1. Bind pat to the initial values given in initial.

2. If cond evaluates to true, bind pat to the result of evaluating loopbody, and repeat the step.

3. Return the final value of pat.

match x case p1 -> e1 case p2 -> e2

Match the value produced by x to each of the patterns in turn, picking the first one that succeeds. The result of the corresponding expression is the value of the entire match expression. All the expressions associated with a case must have the same type (but not necessarily match the type of x). It is a type error if there is not a case for every possible value of x - inexhaustive pattern matching is not allowed.

Function Expressions

\x y z: t -> e

Produces an anonymous function taking parameters x, y, and z, returns type t, and whose body is e. Lambdas do not permit type parameters; use a named function if you want a polymorphic function.

(binop)

An operator section that is equivalent to \x y -> x *binop* y.

(x binop)

An operator section that is equivalent to \y -> x *binop* y.

(binop y)

An operator section that is equivalent to \x -> x *binop* y.

(.a.b.c)

An operator section that is equivalent to \x -> x.a.b.c.

(.[i,j])

An operator section that is equivalent to \x -> x[i,j].

Higher-order functions

At a high level, Futhark functions are values, and can be used as any other value. However, to ensure that the compiler is able to compile the higher-order functions efficiently via defunctionalisation, certain type-driven restrictions exist on how functions can be used. These also apply to any record or tuple containing a function (a functional type):

• Arrays of functions are not permitted.

• A function cannot be returned from an if expression.

• A loop parameter cannot be a function.

Further, type parameters are divided into non-lifted (bound with an apostrophe, e.g. 't), size-lifted ('~t), and fully lifted ('^t). Only fully lifted type parameters may be instantiated with a functional type. Within a function, a lifted type parameter is treated as a functional type.

See also In-place updates for details on how uniqueness types interact with higher-order functions.

Type Inference

Futhark supports Hindley-Milner-style type inference, so in many cases explicit type annotations can be left off. Record field projection cannot in isolation be fully inferred, and may need type annotations where their inputs are bound. The same goes when constructing sum types, as Futhark cannot assume that a given constructor only belongs to a single type. Further, unique types (see In-place updates) must be explicitly annotated.

Type inference processes top-level declared in top-down order, and the type of a top-level function must be completely inferred at its definition site. Specifically, if a top-level function uses overloaded arithmetic operators, the resolution of those overloads cannot be influenced by later uses of the function.

Size Types

Futhark supports a simple system of size-dependent types that statically verifies that the sizes of arrays passed to a function are compatible. The focus is on simplicity, not completeness.

Whenever a pattern occurs (in let, loop, and function parameters), as well as in return types, size annotations may be used to express invariants about the shapes of arrays that are accepted or produced by the function. For example:

let f [n] (a: [n]i32) (b: [n]i32): [n]i32 =
  map2 (+) a b

We use a size parameter, [n], to explicitly quantify sizes. The [n] parameter is not explicitly passed when calling f. Rather, its value is implicitly deduced from the arguments passed for the value parameters. An array can contain anonymous dimensions, e.g. []i32, for which the type checker will invent fresh size parameters, which ensures that all arrays have a (symbolic) size.

A size annotation can also be an integer constant (with no suffix). Size parameters can be used as ordinary variables within the scope of the parameters. The type checker verifies that the program obeys any constraints imposed by size annotations.

Size-dependent types are supported, as the names of parameters can be used in the return type of a function:

let replicate 't (n: i64) (x: t): [n]t = ...

An application replicate 10 0 will have type [10]i32.

Unknown sizes

Since sizes must be constants or variables, there are many cases where the type checker cannot assign a precise size to the result of some operation. For example, the type of concat should conceptually be:

val concat [n] [m] 't : [n]t -> [m]t -> [n+m]t

But this is not presently allowed. Instead, the return type contains an anonymous size:

val concat [n] [m] 't : [n]t -> [m]t -> []t

When an application concat xs ys is found, the result will be of type [k]t, where k is a fresh unknown size variable that is considered distinct from every other size in the program. It is often necessary to perform a size coercion (see Size coercion) to convert an unknown size to a known size.

Generally, unknown sizes are constructed whenever the true size cannot be expressed. The following lists all possible sources of unknown sizes.

Size going out of scope

An unknown size is created when the proper size of an array refers to a name that has gone out of scope:

let c = a + b
in replicate c 0

The type of replicate c 0 is [c]i32, but since c is locally bound, the type of the entire expression is [k]i32 for some fresh k.

Compound expression passed as function argument

Intuitively, the type of replicate (x+y) 0 should be [x+y]i32, but since sizes must be names or constants, this is not expressible. Therefore an unknown size variable is created and the size of the expression becomes [k]i32.

Compound expression used as range bound

While a simple range expression such as 0..<n can be assigned type [n]i32, a range expression 0..<(n+1) will give produce an unknown size.

Complex slicing

Most complex array slicing, such as xs[a:b], will have an unknown size. Exceptions are listed in the reference for slice expressions.

Complex ranges

Most complex ranges, such as a..<b, will have an known size. Exceptions exist for general ranges and “upto” ranges.

Anonymous size in function return type

Whenever the result of a function application would have an anonymous size, that size is replaced with a fresh unknown size variable.

For example, filter has the following type:

val filter [n] 'a : (p: a -> bool) -> (as: [n]a) -> []a

Naively, an application filter f xs seems like it would have type []a, but a fresh unknown size k will be created and the actual type will be [k]a.

Branches of if return arrays of different sizes

When an if (or match) expression has branches that returns array of different sizes, the differing sizes will be replaced with fresh unknown sizes. For example:

if b then [[1,2], [3,4]]
     else [[5,6]]

This expression will have type [k][2]i32, for some fresh k.

Important: The check whether the sizes differ is done when first encountering the if or match during type checking. At this point, the type checker may not realise that the two sizes are actually equal, even though constraints later in the function force them to be. This can always be resolved by adding type annotations.

An array produced by a loop does not have a known size

If the size of some loop parameter is not maintained across a loop iteration, the final result of the loop will contain unknown sizes. For example:

loop xs = [1] for i < n do xs ++ xs

Similar to conditionals, the type checker may sometimes be too cautious in assuming that some size may change during the loop. Adding type annotations to the loop parameter can be used to resolve this.

Size coercion

Size coercion, written with :>, can be used to perform a runtime-checked coercion of one size to another. Since size annotations can refer only to variables and constants, this is necessary when writing more complicated size functions:

let concat_to 'a (m: i32) (a: []a) (b: []a) : [m]a =
  a ++ b :> [m]a

Only expression-level type annotations give rise to run-time checks. Despite their similar syntax, parameter and return type annotations must be valid at compile-time, or type checking will fail.

Causality restriction

Conceptually, size parameters are assigned their value by reading the sizes of concrete values passed along as parameters. This means that any size parameter must be used as the size of some parameter. This is an error:

let f [n] (x: i32) = n

The following is not an error:

let f [n] (g: [n]i32 -> [n]i32) = ...

However, using this function comes with a constraint: whenever an application f x occurs, the value of the size parameter must be inferable. Specifically, this value must have been used as the size of an array before the f x application is encountered. The notion of “before” is subtle, as there is no evaluation ordering of a Futhark expression, except that a let-binding is always evaluated before its body, the argument to a function is always evaluated before the function itself, and the left operand to an operator is evaluated before the right.

The causality restriction only occurs when a function has size parameters whose first use is not as a concrete array size. For example, it does not apply to uses of the following function:

let f [n] (arr: [n]i32) (g: [n]i32 -> [n]i32) = ...

This is because the proper value of n can be read directly from the actual size of the array.

Empty array literals

Just as with size-polymorphic functions, when constructing an empty array, we must know the exact size of the (missing) elements. For example, in the following program we are forcing the elements of a to be the same as the elements of b, but the size of the elements of b are not known at the time a is constructed:

let main (b: bool) (xs: []i32) =
  let a = [] : [][]i32
  let b = [filter (>0) xs]
  in a[0] == b[0]

The result is a type error.

Sum types

When constructing a value of a sum type, the compiler must still be able to determine the size of the constructors that are not used. This is illegal:

type sum = #foo ([]i32) | #bar ([]i32)

let main (xs: *[]i32) =
  let v : sum = #foo xs
  in xs

Modules

When matching a module with a module type (see Module System), a non-lifted abstract type (i.e. one that is declared with type rather than type^) may not be implemented by a type abbreviation that contains any anonymous sizes. This is to ensure that if we have the following:

module m : { type t } = ...

Then we can construct an array of values of type m.t without worrying about constructing an irregular array.

Higher-order functions

When a higher-order function takes a functional argument whose return type is a non-lifted type parameter, any instantiation of that type parameter must have a non-anonymous size. If the return type is a lifted type parameter, then the instantiation may contain anonymous sizes. This is why the type of map guarantees regular arrays:

val map [n] 'a 'b : (a -> b) -> [n]a -> [n]b

The type parameter b can only be replaced with a type that has non-anonymous sizes, which means they must be the same for every application of the function. In contrast, this is the type of the pipeline operator:

val (|>) '^a -> '^b : a -> (a -> b) -> b

The provided function can return something with an anonymous size (such as filter).

A function whose return type has an unknown size

If a function (named or anonymous) is inferred to have a return type that contains an unknown size variable created within the function body, that size variable will be replaced with an anonymous size. In most cases this is not important, but it means that an expression like the following is ill-typed:

map (\xs -> iota (length xs)) (xss : [n][m]i32)

This is because the (length xs) expression gives rise to some fresh size k. The lambda is then assigned the type [n]t -> [k]i32, which is immediately turned into [n]t -> []i32 because k was generated inside its body. A function of this type cannot be passed to map, as explained before. The solution is to bind length to a name before the lambda.

In-place Updates

In-place updates do not provide observable side effects, but they do provide a way to efficiently update an array in-place, with the guarantee that the cost is proportional to the size of the value(s) being written, not the size of the full array.

The a with [i] = v language construct, and derived forms, performs an in-place update. The compiler verifies that the original array (a) is not used on any execution path following the in-place update. This involves also checking that no alias of a is used. Generally, most language constructs produce new arrays, but some (slicing) create arrays that alias their input arrays.

When defining a function parameter or return type, we can mark it as unique by prefixing it with an asterisk. For example:

let modify (a: *[]i32) (i: i32) (x: i32): *[]i32 =
  a with [i] = a[i] + x

For bulk in-place updates with multiple values, use the scatter function in the basis library. In the parameter declaration a: *[i32], the asterisk means that the function modify has been given “ownership” of the array a, meaning that any caller of modify will never reference array a after the call again. This allows the with expression to perform an in-place update.

After a call modify a i x, neither a or any variable that aliases a may be used on any following execution path.

Alias Analysis

The rules used by the Futhark compiler to determine aliasing are intuitive in the intra-procedural case. Aliases are associated with entire arrays. Aliases of a record are tuple are tracked for each element, not for the record or tuple itself. Most constructs produce fresh arrays, with no aliases. The main exceptions are if, loop, function calls, and variable literals.

• After a binding let a = b, that simply assigns a new name to an existing variable, the variable a aliases b. Similarly for record projections and patterns.

• The result of an if aliases the union of the aliases of the components.

• The result of a loop aliases the initial values, as well as any aliases that the merge parameters may assume at the end of an iteration, computed to a fixed point.

• The aliases of a value returned from a function is the most interesting case, and depends on whether the return value is declared unique (with an asterisk *) or not. If it is declared unique, then it has no aliases. Otherwise, it aliases all arguments passed for non-unique parameters.

In-place Updates and Higher-Order Functions

Uniqueness typing generally interacts poorly with higher-order functions. The issue is that we cannot control how many times a function argument is applied, or to what, so it is not safe to pass a function that consumes its argument. The following two conservative rules govern the interaction between uniqueness types and higher-order functions:

1. In the expression let p = e1 in ..., if any in-place update takes place in the expression e1, the value bound by p must not be or contain a function.

2. A function that consumes one of its arguments may not be passed as a higher-order argument to another function.

Module System

mod_bind      ::=  "module" id mod_param* "=" [":" mod_type_exp] "=" mod_exp
mod_param     ::=  "(" id ":" mod_type_exp ")"
mod_type_bind ::=  "module" "type" id type_param* "=" mod_type_exp

Futhark supports an ML-style higher-order module system. Modules can contain types, functions, and other modules and module types. Module types are used to classify the contents of modules, and parametric modules are used to abstract over modules (essentially module-level functions). In Standard ML, modules, module types and parametric modules are called structs, signatures, and functors, respectively. Module names exist in the same name space as values, but module types are their own name space.

Named modules are declared as:

module name = ...

A named module type is defined as:

module type name = ...

Where a module expression can be the name of another module, an application of a parametric module, or a sequence of declarations enclosed in curly braces:

module Vec3 = {
  type t = ( f32 , f32 , f32 )
  let add(a: t) (b: t): t =
    let (a1, a2, a3) = a in
    let (b1, b2, b3) = b in
    (a1 + b1, a2 + b2 , a3 + b3)
}

module AlsoVec3 = Vec3

Functions and types within modules can be accessed using dot notation:

type vector = Vec3.t
let double(v: vector): vector = Vec3.add v v

We can also use open Vec3 to bring the names defined by Vec3 into the current scope. Multiple modules can be opened simultaneously by separating their names with spaces. In case several modules define the same names, the ones mentioned last take precedence. The first argument to open may be a full module expression.

Named module types are defined as:

module type ModuleTypeName = ...

A module type expression can be the name of another module type, or a sequence of specifications, or specs, enclosed in curly braces. A spec can be a value spec, indicating the presence of a function or value, an abstract type spec, or a type abbreviation spec. For example:

module type Addable = {
  type t                 -- abstract type spec
  type two_ts = (t,t)    -- type abbreviation spec
  val add: t -> t -> t   -- value spec
}

This module type specifies the presence of an abstract type t, as well as a function operating on values of type t. We can use module type ascription to restrict a module to what is exposed by some module type:

module AbstractVec = Vec3 : Addable

The definition of AbstractVec.t is now hidden. In fact, with this module type, we can neither construct values of type AbstractVec.T or convert them to anything else, making this a rather useless use of abstraction. As a derived form, we can write module M: S = e to mean module M = e : S.

In a value spec, sizes in types on the left-hand side of a function arrow must not be anonymous. For example, this is forbidden:

val sum: []t -> t

Instead write:

val sum [n]: [n]t -> t

But this is allowed, because the empty size is not to the left of a function arrow:

val evens [n]: [n]i32 -> []i32

Parametric modules allow us to write definitions that abstract over modules. For example:

module Times = \(M: Addable) -> {
  let times (x: M.t) (k: i32): M.t =
    loop x' = x for i < k do
      M.add x' x
}

We can instantiate Times with any module that fulfils the module type Addable and get back a module that defines a function times:

module Vec3Times = Times Vec3

Now Vec3Times.times is a function of type Vec3.t -> int -> Vec3.t. As a derived form, we can write module M p = e to mean module M = \p -> e.

Module Expressions

mod_exp ::=    qualid
             | mod_exp ":" mod_type_exp
             | "\" "(" id ":" mod_type_exp ")" [":" mod_type_exp] "->" mod_exp
             | mod_exp mod_exp
             | "(" mod_exp ")"
             | "{" dec* "}"
             | "import" stringlit

A module expression produces a module. Modules are collections of bindings produced by declarations (dec). In particular, a module may contain other modules or module types.

qualid

Evaluates to the module of the given name.

(mod_exp)

Evaluates to mod_exp.

mod_exp : mod_type_exp

Module ascription evaluates the module expression and the module type expression, verifies that the module implements the module type, then returns a module that exposes only the functionality described by the module type. This is how internal details of a module can be hidden.

\(p: mt1): mt2 -> e

Constructs a parametric module (a function at the module level) that accepts a parameter of module type mt1 and returns a module of type mt2. The latter is optional, but the parameter type is not.

e1 e2

Apply the parametric module m1 to the module m2.

{ decs }

Returns a module that contains the given definitions. The resulting module defines any name defined by any declaration that is not local, in particular including names made available via open.

import "foo"

Returns a module that contains the definitions of the file "foo" relative to the current file. See Referring to Other Files.

Module Type Expressions

mod_type_exp ::=    qualid
                  | "{" spec* "}"
                  | mod_type_exp "with" qualid type_param* "=" type
                  | "(" mod_type_exp ")"
                  | "(" id ":" mod_type_exp ")" "->" mod_type_exp
                  | mod_type_exp "->" mod_type_exp

spec      ::=    "val" id type_param* ":" spec_type
               | "val" binop type_param* ":" spec_type
               | "type" ["^"] id type_param* "=" type
               | "type" ["^"] id type_param*
               | "module" id ":" mod_type_exp
               | "include" mod_type_exp
               | "#[" attr "]" spec
spec_type ::=  type | type "->" spec_type

Module types classify modules, with the only (unimportant) difference in expressivity being that modules can contain module types, but module types cannot specify that a module must contain a specific module type. They can specify of course that a module contains a submodule of a specific module type.

Referring to Other Files

You can refer to external files in a Futhark file like this:

import "file"

The above will include all non-local top-level definitions from file.fut is and make them available in the current file (but will not export them). The .fut extension is implied.

You can also include files from subdirectories:

import "path/to/a/file"

The above will include the file path/to/a/file.fut relative to the including file.

Qualified imports are also possible, where a module is created for the file:

module M = import "file"

In fact, a plain import "file" is equivalent to:

local open import "file"

Attributes

attr ::=    id
          | id "(" [attr ("," attr)*] ")"

An expression, declaration, or module type spec can be prefixed with an attribute, written as #[attr]. This may affect how it is treated by the compiler or other tools. In no case will attributes affect or change the semantics of a program, but it may affect how well it compiles and runs (or in some cases, whether it compiles or runs at all). Unknown attributes are silently ignored. Most have no effect in the interpreter. An attribute can be either an atom, written as just an identifier, or compound, consisting of an identifier and a comma-separated sequence of attributes. The latter is used for grouping and encoding of more complex information.

Expression attributes

Many expression attributes affect second-order array combinators (SOACS). These must be applied to a fully saturated function application or they will have no effect. If two SOACs with contradictory attributes are combined through fusion, it is unspecified which attributes take precedence.

The following expression attributes are supported.

incremental_flattening(no_outer)

When using incremental flattening, do not generate the “only outer parallelism” version for the attributed SOACs.

incremental_flattening(no_intra)

When using incremental flattening, do not generate the “intra-group parallelism” version for the attributed SOACs.

incremental_flattening(only_intra)

When using incremental flattening, only generate the “intra-group parallelism” version of the attributed SOACs. Beware: the resulting program will fail to run if the inner parallelism does not fit on the device.

incremental_flattening(only_inner)

When using incremental flattening, do not generate multiple versions for this SOAC, but do exploit inner parallelism (which may give rise to multiple versions at deeper levels).

noinline

Do not inline the attributed function application. If used within a parallel construct (e.g. map), this will likely prevent the GPU backends from generating working code.

sequential

Fully sequentialise the attributed SOAC.

sequential_outer

Turn the outer parallelism in the attributed SOAC sequential, but preserve any inner parallelism.

sequential_inner

Exploit only outer parallelism in the attributed SOAC.

unroll

Fully unroll the attributed loop. If the compiler cannot determine the exact number of iterations (possibly after other optimisations and simplifications have taken place), then this attribute has no code generation effect, but instead results in a warning. Be very careful with this attribute: it can massively increase program size (possibly crashing the compiler) if the loop has a huge number of iterations.

unsafe

Do not perform any dynamic safety checks (such as bound checks) during execution of the attributed expression.

warn(safety_checks)

Make the compiler issue a warning if the attributed expression (or its subexpressions) requires safety checks (such as bounds checking) at run-time. This is used for performance-critical code where you want to be told when the compiler is unable to statically verify the safety of all operations.

Declaration attributes

The following declaration attributes are supported.

noinline

Do not inline any calls to this function. If the function is then used within a parallel construct (e.g. map), this will likely prevent the GPU backends from generating working code.

Spec attributes

No spec attributes are currently supported by the compiler itself, although they are syntactically permitted and may be used by other tools.

C API Reference

A Futhark program futlib.fut compiled to a C library with the --library command line option produces two files: futlib.c and futlib.h. The API provided in the .h file is documented in the following.

Using the API requires creating a configuration object, which is then used to obtain a context object, which is then used to perform most other operations, such as calling Futhark functions.

Most functions that can fail return an integer: 0 on success and a non-zero value on error. Others return a NULL pointer. Use futhark_context_get_error() to get a (possibly) more precise error message.

FUTHARK_BACKEND_foo

A preprocessor macro identifying that the backend foo was used to generate the code; e.g. c, opencl, or cuda. This can be used for conditional compilation of code that only works with specific backends.

Configuration

Context creation is parameterised by a configuration object. Any changes to the configuration must be made before calling futhark_context_new(). A configuration object must not be freed before any context objects for which it is used. The same configuration may be used for multiple concurrent contexts.

struct futhark_context_config

An opaque struct representing a Futhark configuration.

struct futhark_context_config *futhark_context_config_new(void)

Produce a new configuration object. You must call futhark_context_config_free() when you are done with it.

void futhark_context_config_free(struct futhark_context_config *cfg)

Free the configuration object.

void futhark_context_config_set_debugging(struct futhark_context_config *cfg, int flag)

With a nonzero flag, enable various debugging information, with the details specific to the backend. This may involve spewing copious amounts of information to the standard error stream. It is also likely to make the program run much slower.

void futhark_context_config_set_profiling(struct futhark_context_config *cfg, int flag)

With a nonzero flag, enable the capture of profiling information. This should not significantly impact program performance. Use futhark_context_report() to retrieve captured information, the details of which are backend-specific.

void futhark_context_config_set_logging(struct futhark_context_config *cfg, int flag)

With a nonzero flag, print a running log to standard error of what the program is doing.

Context

struct futhark_context

An opaque struct representing a Futhark context.

struct futhark_context *futhark_context_new(struct futhark_context_config *cfg)

Create a new context object. You must call futhark_context_free() when you are done with it. It is fine for multiple contexts to co-exist within the same process, but you must not pass values between them. They have the same C type, so this is an easy mistake to make.

After you have created a context object, you must immediately call futhark_context_get_error(), which will return non-NULL if initialisation failed. If initialisation has failed, then you still need to call futhark_context_free() to release resources used for the context object, but you may not use the context object for anything else.

void futhark_context_free(struct futhark_context *ctx)

Free the context object. It must not be used again. The configuration must be freed separately with futhark_context_config_free().

int futhark_context_sync(struct futhark_context *ctx)

Block until all outstanding operations, including copies, have finished executing. Many API functions are asynchronous on their own.

void futhark_context_pause_profiling(struct futhark_context *ctx)

Temporarily suspend the collection of profiling information. Has no effect if profiling was not enabled in the configuration.

void futhark_context_unpause_profiling(struct futhark_context *ctx)

Resume the collection of profiling information. Has no effect if profiling was not enabled in the configuration.

char *futhark_context_get_error(struct futhark_context *ctx)

A human-readable string describing the last error, if any. It is the caller’s responsibility to free() the returned string. Any subsequent call to the function returns NULL, until a new error occurs.

char *futhark_context_report(struct futhark_context *ctx)

Produce a human-readable C string with debug and profiling information collected during program runtime. It is the caller’s responsibility to free the returned string. It is likely to only contain interesting information if futhark_context_config_set_debugging() or futhark_context_config_set_profiling() has been called previously.

int futhark_context_clear_caches(struct futhark_context *ctx)

Release any context-internal caches and buffers that may otherwise use computer resources. This is useful for freeing up those resources when no Futhark entry points are expected to run for some time. Particularly relevant when using a GPU backend, due to the relative scarcity of GPU memory.

Values

Primitive types (i32, bool, etc) are mapped directly to their corresponding C type. For each distinct array type (without sizes), an opaque C struct is defined. Complex types (records, nested tuples) are also assigned an opaque C struct. In the general case, these types will be named with a random hash. However, if you insert an explicit type annotation (and the type name contains only characters valid for C identifiers), the indicated name will be used. Note that arrays contain brackets, which are usually not valid in identifiers. Defining a simple type alias is the best way around this.

All values share a similar API, which is illustrated here for the case of the type []i32. The creation/retrieval functions are all asynchronous, so make sure to call futhark_context_sync() when appropriate. Memory management is entirely manual. All values that are created with a new function, or returned from an entry point, must at some point be freed manually. Values are internally reference counted, so even for entry points that return their input unchanged, you should still free both the input and the output - this will not result in a double free.

struct futhark_i32_1d

An opaque struct representing a Futhark value of type []i32.

struct futhark_i32_1d *futhark_new_i32_1d(struct futhark_context *ctx, int32_t *data, int64_t dim0)

Asynchronously create a new array based on the given data. The dimensions express the number of elements. The data is copied into the new value. It is the caller’s responsibility to eventually call futhark_free_i32_1d(). Multi-dimensional arrays are assumed to be in row-major form.

struct futhark_i32_1d *futhark_new_raw_i32_1d(struct futhark_context *ctx, char *data, int offset, int64_t dim0)

Create an array based on raw data, as well as an offset into it. This differs little from futhark_i32_1d() when using the c backend, but when using e.g. the opencl backend, the data parameter will be a cl_mem. It is the caller’s responsibility to eventually call futhark_free_i32_1d().

int futhark_free_i32_1d(struct futhark_context *ctx, struct futhark_i32_1d *arr)

Free the value. In practice, this merely decrements the reference count by one. The value (or at least this reference) may not be used again after this function returns.

int futhark_values_i32_1d(struct futhark_context *ctx, struct futhark_i32_1d *arr, int32_t *data)

Asynchronously copy data from the value into data, which must be of sufficient size. Multi-dimensional arrays are written in row-major form.

const int64_t *futhark_shape_i32_1d(struct futhark_context *ctx, struct futhark_i32_1d *arr)

Return a pointer to the shape of the array, with one element per dimension. The lifetime of the shape is the same as arr, and should not be manually freed.

Entry points

Entry points are mapped 1:1 to C functions. Return values are handled with out-parameters.

For example, this Futhark entry point:

entry sum = i32.sum

Results in the following C function:

int futhark_entry_main(struct futhark_context *ctx, int32_t *out0, const struct futhark_i32_1d *in0)

Asynchronously call the entry point with the given arguments. Make sure to call futhark_context_sync() before using the value of out0.

The exact behaviour of the exit code depends on the backend. For the sequential C backend, errors will always be available when the entry point returns, and futhark_context_sync() will always return success. When using a GPU backend such as cuda or opencl, the entry point may still be running asynchronous operations when it returns, in which case the entry point may return zero successfully, even though execution has already (or will) fail. These problems will be reported when futhark_context_sync() is called. When using GPU backends, be careful to check the return code of both the entry point itself, and futhark_context_sync().

GPU

The following API functions are available when using the opencl or cuda backends.

void futhark_context_config_set_device(struct futhark_context_config *cfg, const char *s)

Use the first device whose name contains the given string. The special string #k, where k is an integer, can be used to pick the k-th device, numbered from zero. If used in conjunction with futhark_context_config_set_platform(), only the devices from matching platforms are considered.

Exotic

The following functions are not interesting to most users.

void futhark_context_config_set_default_group_size(struct futhark_context_config *cfg, int size)

Set the default number of work-items in a work-group.

void futhark_context_config_set_default_num_groups(struct futhark_context_config *cfg, int num)

Set the default number of work-groups used for kernels.

void futhark_context_config_set_default_tile_size(struct futhark_context_config *cfg, int num)

Set the default tile size used when executing kernels that have been block tiled.

void futhark_context_config_dump_program_to(struct futhark_context_config *cfg, const char *path)

During futhark_context_new(), dump the OpenCL or CUDA program source to the given file.

void futhark_context_config_load_program_from(struct futhark_context_config *cfg, const char *path)

During futhark_context_new(), read OpenCL or CUDA program source from the given file instead of using the embedded program.

OpenCL

The following API functions are available only when using the opencl backend.

void futhark_context_config_set_platform(struct futhark_context_config *cfg, const char *s)

Use the first OpenCL platform whose name contains the given string. The special string #k, where k is an integer, can be used to pick the k-th platform, numbered from zero.

void futhark_context_config_select_device_interactively(struct futhark_context_config *cfg)

Immediately conduct an interactive dialogue on standard output to select the platform and device from a list.

struct futhark_context *futhark_context_new_with_command_queue(struct futhark_context_config *cfg, cl_command_queue queue)

Construct a context that uses a pre-existing command queue. This allows the caller to directly customise which device and platform is used.

cl_command_queue futhark_context_get_command_queue(struct futhark_context *ctx)

Retrieve the command queue used by the Futhark context. Be very careful with it - enqueueing your own work is unlikely to go well.

Exotic

The following functions are used for debugging generated code or advanced usage.

void futhark_context_config_add_build_option(struct futhark_context_config *cfg, const char *opt)

Add a build option to the OpenCL kernel compiler. See the OpenCL specification for clBuildProgram for available options.

void futhark_context_config_dump_binary_to(struct futhark_context_config *cfg, const char *path)

During futhark_context_new(), dump the compiled OpenCL binary to the given file.

void futhark_context_config_load_binary_from(struct futhark_context_config *cfg, const char *path)

During futhark_context_new(), read a compiled OpenCL binary from the given file instead of using the embedded program.

CUDA

The following API functions are available when using the cuda backend.

Exotic

The following functions are used for debugging generated code or advanced usage.

void futhark_context_config_add_nvrtc_option(struct futhark_context_config *cfg, const char *opt)

Add a build option to the NVRTC compiler. See the CUDA documentation for nvrtcCompileProgram for available options.

void futhark_context_config_dump_ptx_to(struct futhark_context_config *cfg, const char *path)

During futhark_context_new(), dump the generated PTX code to the given file.

void futhark_context_config_load_ptx_from(struct futhark_context_config *cfg, const char *path)

During futhark_context_new(), read PTX code from the given file instead of using the embedded program.

Multicore

The following API functions are available when using the multicore backend.

void context_config_set_num_threads(struct futhark_context_config *cfg, int n)

The number of threads used to run parallel operations. If set to a value less than 1, then the runtime system will use one thread per detected core.

General guarantees

Calling an entry point, or interacting with Futhark values through the functions listed above, has no system-wide side effects, such as writing to the file system, launching processes, or performing network connections. Defects in the program or Futhark compiler itself can with high probability result only in the consumption of CPU or GPU resources, or a process crash.

Using the #[unsafe] attribute with in-place updates can result in writes to arbitrary memory locations. A malicious program can likely exploit this to obtain arbitrary code execution, just as with any insecure C program. If you must run untrusted code, consider using the --safe command line option to instruct the compiler to disable #[unsafe].

Initialising a Futhark context likewise has no side effects, except if explicitly configured differently, such as by using futhark_context_config_dump_program_to(). In its default configuration, Futhark will not access the file system.

Note that for the GPU backends, the underlying API (such as CUDA or OpenCL) may perform file system operations during startup, and perhaps for caching GPU kernels in some cases. This is beyond Futhark’s control.

Package Management

This document describes futhark pkg, a minimalistic package manager inspired by vgo. A Futhark package is a downloadable collection of .fut files and little more. There is a (not necessarily comprehensive) list of known packages.

Basic Concepts

A package is uniquely identified with a package path, which is similar to a URL, except without a protocol. At the moment, package paths are always links to Git repositories hosted on GitHub or GitLab. In the future, this will become more flexible. As an example, a package path may be github.com/athas/fut-foo.

Packages are versioned with semantic version numbers of the form X.Y.Z. Whenever versions are indicated, all three digits must always be given (that is, 1.0 is not a valid shorthand for 1.0.0).

Most futhark pkg operations involve reading and writing a package manifest, which is always stored in a file called futhark.pkg. The futhark.pkg file is human-editable, but is in day-to-day use mainly modified by futhark pkg automatically.

Using Packages

Required packages can be added by using futhark pkg add, for example:

$ futhark pkg add github.com/athas/fut-foo 0.1.0

This will create a new file futhark.pkg with the following contents:

require {
  github.com/athas/fut-foo 0.1.0 #d285563c25c5152b1ae80fc64de64ff2775fa733
}

This lists one required package, with its package path, minimum version (see Version Selection), and the expected commit hash. The latter is used for verification, to ensure that the contents of a package version cannot be changed silently.

futhark pkg will perform network requests to determine whether a package of the given name and with the given version exists and fail otherwise (but it will not check whether the package is otherwise well-formed). The version number can be elided, in which case futhark pkg will use the newest available version. If the package is already present in futhark.pkg, it will simply have its version requirement changed to the one specified in the command. Any dependencies of the package will not be added to futhark.pkg, but will still be downloaded by futhark pkg sync (see below).

Adding a package with futhark pkg add modifies futhark.pkg, but does not download the package files. This is done with futhark pkg sync (without further options). The contents of each required dependency and any transitive dependencies will be stored in a subdirectory of lib/ corresponding to their package path. As an example:

$ futhark pkg sync
$ tree lib
lib
└── github.com
    └── athas
        └── fut-foo
            └── foo.fut

3 directories, 1 file

Warning: futhark pkg sync will remove any unrecognized files or local modifications to files in lib/ (except of course the package directory of the package path listed in futhark.pkg; see Creating Packages).

Packages can be removed from futhark.pkg with:

$ futhark pkg remove pkgpath

You will need to run futhark pkg sync to actually remove the files in lib/.

The intended usage is that futhark.pkg is added to version control, but lib/ is not, as the contents of lib/ can always be reproduced from futhark.pkg. However, adding lib/ works just fine as well.

Importing Files from Dependencies

futhark pkg sync will populate the lib/ directory, but does not interact with the compiler in any way. The downloaded files can be imported using the usual import mechanism (Referring to Other Files); for example, assuming the package contains a file foo.fut:

import "lib/github.com/athas/fut-foo/foo"

Ultimately, everything boils down to ordinary file system semantics. This has the downside of relatively long and clumsy import paths, but the upside of predictability.

Upgrading Dependencies

The futhark pkg upgrade command will update every version requirement in futhark.pkg to be the most recent available version. You still need to run futhark pkg sync to actually retrieve the new versions. Be careful - while upgrades are safe if semantic versioning is followed correctly, this is not yet properly machine-checked, so human mistakes may occur.

As an example:

$ cat futhark.pkg
require {
  github.com/athas/fut-foo 0.1.0 #d285563c25c5152b1ae80fc64de64ff2775fa733
}
$ futhark pkg upgrade
Upgraded github.com/athas/fut-foo 0.1.0 => 0.2.1.
$ cat futhark.pkg
require {
  github.com/athas/fut-foo 0.2.1 #3ddc9fc93c1d8ce560a3961e55547e5c78bd0f3e
}
$ futhark pkg sync
$ tree lib
lib
└── github.com
    └── athas
        ├── fut-bar
        │   └── bar.fut
        └── fut-foo
            └── foo.fut

4 directories, 2 files

Note that fut-foo 0.2.1 depends on github.com/athas/fut-bar, so it was fetched by futhark pkg sync.

futhark pkg upgrade will never upgrade across a major version number. Due to the principle of Semantic Import Versioning, a new major version is a completely different package from the point of view of the package manager. Thus, to upgrade to a new major version, you will need to use futhark pkg add to add the new version and futhark pkg remove to remove the old version. Or you can keep it around - it is perfectly acceptable to depend on multiple major versions of the same package, because they are really different packages.

Creating Packages

A package is a directory tree (which at the moment must correspond to a Git repository). It must contain two things:

• A file futhark.pkg at the root defining the package path and any required packages.

• A package directory lib/pkg-path, where pkg-path is the full package path.

The contents of the package directory is what will be made available to users of the package. The repository may contain other things (tests, data files, examples, docs, other programs, etc), but these are ignored by futhark pkg. This structure can be created automatically by running for example:

$ futhark pkg init github.com/sturluson/edda

Note again, no https://. The result is this futhark.pkg:

package github.com/sturluson/edda

require {
}

And this file hierarchy:

$ tree lib
lib
└── github.com
    └── sturluson
        └── edda

3 directories, 0 files

Note that futhark pkg init is not necessary simply to use packages, only when creating packages.

When creating a package, the .fut files we are writing will be located inside the lib/ directory. If the package has its own dependencies, whose files we would like to access, we can use relative imports. For example, assume we are creating a package github.com/sturluson/edda and we are writing a Futhark file located at lib/github.com/sturluson/edda/saga.fut. Further, we have a dependency on the package github.com/athas/foo-fut, which is stored in the directory lib/github.com/athas/foo-fut. We can import a file lib/github.com/athas/foo-fut/foo.fut from lib/github.com/sturluson/edda/saga.fut with:

import "../foo-fut/foo"

Releasing a Package

Currently, a package corresponds exactly to a GitHub repository mirroring the package path. A release is done by tagging an appropriate commit with git tag vX.Y.Z and then pushing the tag to GitHub with git push --tags. In the future, this will be generalised to other code hosting sites and version control systems (and possibly self-hosted tarballs). Remember to take semantic versioning into account - unless you bump the major version number (or the major version is 0), the new version must be fully compatible with the old.

When releasing a new package, consider getting it added to the central package list. See this page for details.

Incrementing the Major Version Number

While backwards-incompatible modifications to a package are sometimes unavoidable, it is wise to avoid them as much as possible, as they significantly inconvenience users. To discourage breaking compatibility, futhark pkg tries to ensure that the package developer feels this inconvenience as well. In many cases, an incompatible change can be avoided simply by adding new files to the package rather than incompatibly changing the existing ones.

In the general case, the package path also encodes the major version of the package, separated with a @. For example, version 5.2.1 of a package might have the package path github.com/user/repo@5. For major versions 0 and 1, this can be elided. This means that multiple (major) versions of a package are completely distinct from the point of view of the package manager - this principle is called Semantic Import Versioning, and is intended to facilitate backwards compatibility of packages when new versions are released.

If you really must increment the major version, then you will need to change the package path in futhark.pkg to contain the new major version preceded by @. For example, lib/github.com/sturluson/edda becomes lib/github.com/sturluson/edda@2. As a special case, this is not necessary when moving from major version 0 to 1. Since the package path has changed, you will also need to rename the package directory in lib/. This is painful and awkward, but it is less painful and awkward than what users feel when their dependencies break compatibility.

Renaming a Package

It is likely that the hosting location for a very long-lived package will change from time to time. Since the hosting location is embedded into the package path itself, this causes some issues for futhark pkg.

In simple cases, there is no problem. Consider a package github.com/asgard/loki which is moved to github.com/utgard/loki. If no GitHub-level redirect is set up, all users must update the path by which they import the package. This is unavoidable, unfortunately.

However, the old tagged versions, which contain a futhark.pkg that uses the old package path, will continue to work. This is because the package path indicated in package.pkg merely defines the subdirectory of lib/ where the package files are to be found, while the package path used by dependents in the require section defines where the package files are located after futhark pkg sync. Thus, when we import an old version of github.com/utgard/loki whose futhark.pkg defines the package as github.com/asgard/loki, the package files will be retrieved from the lib/github.com/asgard/loki directory in the repository, but stored at lib/github.com/utgard/loki in the local directory.

The above means that package management remains operational in simple cases of renaming, but it is awkward when a transitive dependency is renamed (or deleted). The Futhark package ecosystem is sufficiently embryonic that we have not yet developed more robust solutions. When such solutions are developed, they will likely involve some form of replace directive that allows transparent local renaming of packages, as well as perhaps a central registry of package paths that does not depend on specific source code hosts.

Version Selection

The package manifest futhark.pkg declares which packages the program depends on. Dependencies are specified as the oldest acceptable version within the given major version. Upper version bounds are not supported, as strict adherence to semantic versioning is assumed, so any later version with the same major version number should work. When futhark pkg sync calculates which version of a given package to download, it will pick the oldest version that still satisfies the minimum version requirements of that package in all transitive dependencies. This means that a version may be used that is newer than the one indicated in futhark.pkg, but only if a dependency requires a more recent version.

Tests and Documentation for Dependencies

Package management has been designed to ensure that the normal development tools work as expected with the contents of the lib/ directory. For example, to ensure that all dependencies do in fact work well (or at least compile) together, run:

futhark test lib

Also, you can generate hyperlinked documentation for all dependencies with:

futhark doc lib -o docs

The file docs/index.html can be opened in a web browser to browse the documentation. Prebuilt documentation is also available via the online package list.

Safety

In contrast to some other package managers, futhark pkg does not run any package-supplied code on installation, upgrade, or removal. This means that all futhark pkg operations are in principle completely safe (barring exploitable bugs in futhark pkg itself, which is unlikely but not impossible). Further, Futhark code itself is also completely pure, so executing it cannot have any unfortunate effects, such as infecting all of your own packages with a worm. The worst it can do is loop infinitely, consume arbitrarily large amounts of memory, or produce wrong results.

The exception is packages that uses unsafe. With some cleverness, unsafe can be combined with in-place updates to perform arbitrary memory reads and writes, which can trivially lead to exploitable behaviour. You should not use untrusted code that employs unsafe (but the --safe compiler option may help). However, this is not any worse than calling external code in a conventional impure language, which generally can perform any conceivable harmful action.

C Porting Guide

This short document contains a collection of tips and tricks for porting simple numerical C code to Futhark. Futhark’s sequential fragment is powerful enough to permit a rather straightforward translation of sequential C code that does not rely on pointer mutation. Additionally, we provide hints on how to recognise C coding patterns that are symptoms of C’s weak type system, and how better to organise it in Futhark.

One intended audience of this document is a programmer who needs to translate a benchmark application written in C, or needs to use a simple numerical algorithm that is already available in the form of C source code.

Where This Guide Falls Short

Some C code makes use of unstructured returns and nonlocal exits (return inside loops, for example). These are not easy to express in Futhark, and will require massaging the control flow a bit. C code that uses goto is likewise not easy to port.

Types

Futhark provides scalar types that match the ones commonly used in C: u8/u16/u32/u64 for the unsigned integers, i8/i16/i32/i64 for the signed, and f32/f64 for float and double respectively. In contrast to C, Futhark does not automatically promote types in expressions - you will have to manually make sure that both operands to e.g. a multiplication are of the exact same type. This means that you will need to understand exactly which types a given expression in original C program operates on, which generally boils down to converting the type of the (type-wise) smaller operand to that of the larger. Note that the Futhark bool type is not considered a number.

Operators

Most of the C operators can be found in Futhark with their usual names. Note however that the Futhark / and % operators for integers round towards negative infinity, whereas their counterparts in C round towards zero. You can write // and %% if you want the C behaviour. There is no difference if both operands are non-zero, but // and %% may be slightly faster. For unsigned numbers, they are exactly the same.

Variable Mutation

As a sequential language, most C programs quite obviously rely heavily on mutating variables. However, in many programs, this is done in a static manner without indirection through pointers (except for arrays; see below), which is conceptually similar to just declaring a new variable of the same name that shadows the old one. If this is the case, a C assignment can generally be translated to just a let-binding. As an example, let us consider the following function for computing the modular multiplicative inverse of a 16-bit unsigned integer (part of the IDEA encryption algorithm):

static uint16_t ideaInv(uint16_t a) {
  uint32_t b;
  uint32_t q;
  uint32_t r;
  int32_t t;
  int32_t u;
  int32_t v;

  b = 0x10001;
  u = 0;
  v = 1;

  while(a > 0)
    {
      q = b / a;
      r = b % a;

      b = a;
      a = r;

      t = v;
      v = u - q * v;
      u = t;
    }

  if(u < 0)
    u += 0x10001;

  return u;
}

Each iteration of the loop mutates the variables a, b, v, and u in ways that are visible to the following iteration. Conversely, the “mutations” of q, r, and t are not truly mutations, and the variable declarations could be moved inside the loop if we wished. Presumably, the C programmer left them outside for aesthetic reasons. When translating such code, it is important to determine exactly how much true mutation is going on, and how much is just reuse of variable space. This can usually be done by checking whether a variable is read before it is written in any given iteration - if not, then it is not true mutation. The variables that change value from one iteration of the loop to the next will need to be maintained as merge parameters of the Futhark do-loop.

The Futhark program resulting from a straightforward port looks as follows:

let main(a: u16): u32 =
  let b = 0x10001u32
  let u = 0i32
  let v = 1i32 in
  let (_,_,u,_) = loop ((a,b,u,v)) while a > 0u16 do
    let q = b / u32.u16(a)
    let r = b % u32.u16(a)

    let b = u32.u16(a)
    let a = u16.u32(r)

    let t = v
    let v = u - i32.u32 (q) * v
    let u = t in
    (a,b,u,v)

  in u32.i32(if u < 0 then u + 0x10001 else u)

Note the heavy use of type conversion and type suffixes for constants. This is necessary due to Futhark’s lack of implicit conversions. Note also the conspicuous way in which the do-loop is written - the result of a loop iteration consists of variables whose names are identical to those of the merge parameters.

This program can still be massaged to make it more idiomatic Futhark - for example, the variable t only serves to store the old value of v that is otherwise clobbered. This can be written more elegantly by simply inlining the expressions in the result part of the loop body.

Arrays

Dynamically sized multidimensional arrays are somewhat awkward in C, and are often simulated via single-dimensional arrays with explicitly calculated indices:

a[i * M + j] = foo;

This indicates a two-dimensional array a whose inner dimension is of size M. We can usually look at where a is allocated to figure out what the size of the outer dimension must be:

a = malloc(N * M * sizeof(int));

We see clearly that a is a two-dimensional integer array of size N times M - or of type [N][M]i32 in Futhark. Thus, the update expression above would be translated as:

let a[i,j] = foo in
...

C programs usually first allocate an array, then enter a loop to provide its initial values. This is not possible in Futhark - consider whether you can write it as a replicate, an iota, or a map. In the worst case, use replicate to obtain an array of the desired size, then use a do-loop with in-place updates to initialise it (but note that this will run stricly sequentially).

Futhark Compared to Other Functional Languages

This guide is intended for programmers who are familiar with other functional languages and want to start working with Futhark.

Futhark is a simple language with a complex compiler. Functional programming is fundamentally well suited to data-parallelism, so Futhark’s syntax and underlying concepts are taken directly from established functional languages; mostly from Haskell and the members of the ML family. While Futhark does add a few small conveniences (built-in array types) and one complicated and unusual feature (in-place updates via uniqueness types, see In-place Updates), a programmer familiar with a common functional language should be able to understand the meaning of a Futhark program, and quickly begin writing their own programs. To speed up this process, we describe here some of the various quirks and unexpected limitations imposed by Futhark. We also recommended reading some of the example programs along with this guide. The guide does not cover all Futhark features worth knowing, so do also skim Language Reference.

Basic Syntax

Futhark uses a keyword-based structure, with optional indentation solely for human readability. This aspect differs from Haskell and F#.

Names are lexically divided into identifiers and symbols:

• Identifiers begin with a letter or underscore and contain letters, numbers, underscores, and apostrophes.

• Symbols contain the characters found in the default operators (+-*/%=!><|&^)

All function and variable names must be identifiers, and all infix operators are symbols. An identifier can be used as an infix operator by enclosing it in backticks, as in Haskell.

Identifiers are case-sensitive, and there is no restriction on the case of the first letter (unlike Haskell and OCaml, but like Standard ML).

User-defined operators are possible, but the fixity of the operator depends on its name. Specifically, the fixity of a user-defined operator op is equal to the fixity of the built-in operator that is the longest prefix of op. For example, <<= would have the same fixity as <<, and =<< the same as =. This rule is the same as the rule found in OCaml and F#.

Top-level functions and values are defined with let, as in OCaml and F#.

Evaluation

Futhark is a completely pure language, with no cheating through monads or anything of the sort.

Evaluation is eager or call-by-value, like most non-Haskell languages. However, there is no defined evaluation order. Furthermore, the Futhark compiler is permitted to turn non-terminating programs into terminating programs, for example by removing dead code that might cause an error. Moreover, there is no way to handle errors within a Futhark program (no exceptions or similar); although errors are gracefully reported to whatever invokes the Futhark program.

The evaluation semantics are entirely sequential, with parallelism being solely an operational detail. Hence, race conditions are impossible. The Futhark compiler does not automatically go looking for parallelism. Only certain special constructs and built-in library functions (in particular map, reduce, scan, and filter) may be executed in parallel.

Currying and partial application work as usual (although functions are not fully first class; see Types). Some Futhark language constructs look like functions, but are not. This means they cannot be partially applied. These include unsafe and assert.

Lambda terms are written as \x -> x + 2, as in Haskell.

A Futhark program is read top-down, and all functions must be declared in the order they are used, like Standard ML. Unlike just about all functional languages, recursive functions are not supported. Most of the time, you will use bulk array operations instead, but there is also a dedicated loop language construct, which is essentially syntactic sugar for tail recursive functions.

Types

Futhark supports a range of integer types, floating point types, and booleans (see Primitive Types and Values). A numeric literal can be suffixed with its desired type, such as 1i8 for an eight-bit signed integer. Un-adorned numerals have their type inferred based on use. This only works for built-in numeric types.

Arrays are a built-in type. The type of an array containing elements of type t is written []t (not [t] as in Haskell), and we may optionally annotate it with a size as [n]t (see Shape Declarations). Array values are written as [1,2,3]. Array indexing is written a[i] with no space allowed between the array name and the brace. Indexing of multi-dimensional arrays is written a[i,j]. Arrays are 0-indexed.

All types can be combined in tuples as usual, as well as in structurally typed records, as in Standard ML. Non-recursive sum types are supported, and are also structurally typed. Abstract types are possible via the module system; see Module System.

If a variable foo is a record of type {a: i32, b: bool}, then we access field a with dot notation: foo.a. Tuples are a special case of records, where all the fields have a 0-indexed numeric label. For example, (i32, bool) is the same as {0: i32, 1: bool}, and can be indexed as foo.1.

Sum types are defined as constructors separated by a vertical bar (|). Constructor names always start with a #. For example, #red | #blue i32 is a sum type with the constructors #red and #blue, where the latter has an i32 as payload. The terms #red and #blue 2 produce values of this type. Constructor applications must always be fully saturated. Due to the structural typing, type annotations are usually necessary to resolve ambiguities. For example, the term #blue 2 can produce a value of any type that has an appropriate constructor.

Function types are supported with the usual a -> b, and functions can be passed as arguments to other functions. However, there are some restrictions:

• A function cannot be put in an array (but a record or tuple is fine).

• A function cannot be returned from a branch.

• A function cannot be used as a loop parameter.

Function types interact with type parameters in a subtle way:

let id 't (x: t) = x

This declaration defines a function id that has a type parameter t. Here, t is an unlifted type parameter, which is guaranteed never to be a function type, and so in the body of the function we could choose to put parameter values of type t in an array. However, it means that this identity function cannot be called on a functional value. Instead, we probably want a lifted type parameter:

let id '^t (x: t) = x

Such lifted type parameters are not restricted from being instantiated with function types. On the other hand, in the function definition they are subject to the same restrictions as functional types.

Futhark supports Hindley-Milner type inference (with some restrictions), so we could also just write it as:

let id x = x

Type abbreviations are possible:

type foo = (i32, i32)

Type parameters are supported as well:

type pair 'a 'b = (a, b)

As with everything else, they are structurally typed, so the types pair i32 bool and (i32, bool) are entirely interchangeable. Most unusually, this is also the case for sum types. The following two types are entirely interchangeable:

type maybe 'a = #just a | #nothing

type option 'a = #nothing | #just a

Only for abstract types, where the definition has been hidden via the module system, do type names have any significance.

Size parameters can also be passed:

type vector [n] t = [n]t
type i32matrix [n][m] = [n] (vector [m] i32)

Note that for an actual array type, the dimensions come before the element type, but with a type abbreviation, a size is just another parameter. This easily becomes hard to read if you are not careful.

Hacking on the Futhark Compiler

The Futhark compiler is a significant body of code with a not entirely straightforward design. The main source of documentation is the Haddock comments in the source code itself. You can generate hyperlinked reference documentation by running stack haddock or cabal haddock, depending on your preference of build system. There is also possibly-outdated documentation on Hackage

If you feel that the documentation is incomplete, or something lacks an explanation, then feel free to report it as an issue on the GitHub page. Documentation bugs are bugs too.

The Futhark compiler is usually built using Stack. It’s a good idea to familiarise yourself with how it works. As a starting point, here are a few hints:

• When testing, pass --fast to stack to disable the GHC optimiser. This speeds up builds considerably (although it still takes a while). The resulting Futhark compiler will run slower, but it is not something you will notice for small test programs.

• When debugging, pass --profile to stack. This will build the Futhark compiler with debugging information (not just profiling). In particular, hard crashes will print a stack trace. You can also get actual profiling information by passing +RTS -pprof-all -RTS to the Futhark compiler. This asks the Haskell runtime to print profiling information to a file. For more information, see the Profiling chapter in the GHC User Guide.

• You may wish to set the environment variable FUTHARK_COMPILER_DEBUGGING=1. Currently this only has the effect of making the frontend print internal names, but it may control more things in the future.

Debugging Internal Type Errors

The Futhark compiler uses a typed core language, and the type checker is run after every pass. If a given pass produces a program with inconsistent typing, the compiler will report an error and abort. While not every compiler bug will manifest itself as a core language type error (unfortunately), many will. To write the erroneous core program to a file in case of type error, pass -v filename to the compiler. This will also enable verbose output, so you can tell which pass fails. The -v option is also useful when the compiler itself crashes, as you can at least tell where in the pipeline it got to.

Checking Generated Code

Hacking on the compiler will often involve inspecting the quality of the generated code. The recommended way to do this is to use futhark-c or futhark-opencl to compile a Futhark program to an executable. These backends insert various forms of instrumentation that can be enabled by passing run-time options to the generated executable.

• As a first resort, use -t option to use the built-in runtime measurements. A nice trick is to pass -t /dev/stderr, while redirecting standard output to /dev/null. This will print the runtime on the screen, but not the execution result.

• Optionally use -r to ask for several runs, e.g. -r 10. If combined with -t, this will cause several runtimes to be printed (one per line). The futhark-bench tool itself uses -t and -r to perform its measurements.

• Pass -D to have the program print information on allocation and deallocation of memory.

• (futhark-opencl only) Use the -D option to enable synchronous execution. clFinish() will be called after most OpenCL operations, and a running log of kernel invocations will be printed. At the end of execution, the program prints a table summarising all kernels and their total runtime and average runtime.

Using futhark dev

For debugging specific compiler passes, the futhark dev subcommand allows you to tailor your own compilation pipeline using command line options. It is also useful for seeing what the AST looks like after specific passes.

When you are about to have a bad day

When using the cuda backend, you can use the --dump-ptx runtime option to dump PTX, a kind of high-level assembly for NVIDIA GPUs, corresponding to the GPU kernels. This can be used to investigate why the generated code isn’t running as fast as you expect (not fun), or even whether NVIDIAs compiler is miscompiling something (extremely not fun). With the OpenCL backend, --dump-opencl-binary does the same thing.

On AMD platforms, --dump-opencl-binary tends to produce an actual binary of some kind, and it is pretty tricky to obtain a debugger for it (they are available and open source, but the documentation and installation instructions are terrible). Instead, AMDs OpenCL kernel compiler accepts a -save-temps=foo build option, which will make it write certain intermediate files, prefixed with foo. In particular, it will write an .s file that contains what appears to be HSA assembly (at least when using ROCm). If you find yourself having to do do this, then you are definitely going to have a bad day, and probably evening and night as well.

Binary Data Format

Futhark programs compiled to an executable support both textual and binary input. Both are read via standard input, and can be mixed, such that one argument to an entry point may be binary, and another may be textual. The binary input format takes up significantly less space on disk, and can be read much faster than the textual format. This chapter describes the binary input format and its current limitations. The input formats (whether textual or binary) are not used for Futhark programs compiled to libraries, which instead use whichever format is supported by their host language.

Currently reading binary input is only supported for programs generated by futhark c/futhark opencl, and futhark py/futhark pyopencl. It is not supported for futhark run.

You can generate random data in the binary format with futhark dataset (futhark-dataset). This tool can also be used to convert between binary and textual data.

Futhark-generated executables can be asked to generate binary output with the -b option.

Specification

Elements that are bigger than one byte are always stored using little endian – we mostly run our code on x86 hardware so this seemed like a reasonable choice.

When reading input for an argument to the entry function, we need to be able to differentiate between text and binary input. If the first non-whitespace character of the input is a b we will parse this argument as binary, otherwise we will parse it in text format. Allowing preceding whitespace characters makes it easy to use binary input for some arguments, and text input for others.

The general format has this header:

b <version> <num_dims> <type>

Where version is a byte containing the version of the binary format used for encoding (currently 2), num_dims is the number of dimensions in the array as a single byte (0 for scalar), and type is a 4 character string describing the type of the values(s) – see below for more details.

Encoding a scalar value is done by treating it as a 0-dimensional array:

b <version> 0 <type> <value>

To encode an array we must encode the number of dimensions n as a single byte, each dimension dim_i as an unsigned 64-bit little endian integer, and finally all the values in their binary little endian representation:

b <version> <n> <type> <dim_1> <dim_2> ... <dim_n> <values>

Type Values

A type is identified by a 4 character ASCII string (four bytes). Valid types are:

"  i8"
" i16"
" i32"
" i64"
"  u8"
" u16"
" u32"
" u64"
" f32"
" f64"
"bool"

Note that unsigned and signed integers have the same byte-level representation.

futhark

SYNOPSIS

futhark <subcommand> options…

DESCRIPTION

Futhark is a data-parallel functional array language. Through various subcommands, the futhark tool provides facilities for compiling, developing, or analysing Futhark programs. Most subcommands are documented in their own manpage. For example, futhark opencl is documented as futhark-opencl. The remaining subcommands are documented in this page.

COMMANDS

futhark check [-w] PROGRAM

Check whether a Futhark program type checks. With -w, no warnings are printed.

futhark datacmp FILE_A FILE_B

Check whether the two files contain the same Futhark values. The files must be formatted using the general Futhark data format that is used by all other executable and tools (such as futhark-dataset). All discrepancies will be reported. This is in contrast to futhark-test, which only reports the first one.

futhark dataget PROGRAM DATASET

Find the test dataset whose description contains DATASET (e.g. #1) and print it in binary representation to standard output.

futhark dev options… PROGRAM

A Futhark compiler development command, intentionally undocumented and intended for use in developing the Futhark compiler, not for programmers writing in Futhark.

futhark imports PROGRAM

Print all non-builtin imported Futhark files to stdout, one per line.

futhark query PROGRAM LINE COL

Print information about the variable at the given position in the program.

SEE ALSO

futhark-opencl, futhark-c, futhark-python, futhark-pyopencl, futhark-dataset, futhark-doc, futhark-test, futhark-bench, futhark-run, futhark-repl

futhark-autotune

SYNOPSIS

futhark autotune [options…] <program.fut>

DESCRIPTION

futhark-autotune attemps to find optimal values for threshold parameters given representative datasets. This is done by repeatedly running running the program through futhark-bench with different values for the threshold parameters. When futhark-autotune finishes tuning a program foo.fut, the results are written to foo.fut.tuning, which will then automatically be picked up by subsequent uses of futhark-bench and futhark-test.

Currently, only the entry point named main is tuned.

OPTIONS

--backend=name

The backend used when compiling Futhark programs (without leading futhark, e.g. just opencl).

--futhark=program

The program used to perform operations (eg. compilation). Defaults to the binary running futhark autotune itself.

--pass-option=opt

Pass an option to programs that are being run. For example, we might want to run OpenCL programs on a specific device:

futhark autotune prog.fut --backend=opencl --pass-option=-dHawaii

--runs=count

The number of runs per data set.

-v, --verbose

Print verbose information about what the tuner is doing. Pass multiple times to increase the amount of information printed.

--tuning=EXTENSION

Change the extension used for tuning files (.tuning by default).

--timeout=seconds

Initial tuning timeout for each dataset in seconds. After running the intitial tuning run on each dataset, the timeout is based on the run time of that initial tuning. Defaults to 60.

A negative timeout means to wait indefinitely.

SEE ALSO

futhark-bench

futhark-bench

SYNOPSIS

futhark bench [options…] programs…

DESCRIPTION

This tool is the recommended way to benchmark Futhark programs. Programs are compiled using the specified backend (futhark c by default), then run a number of times for each test case, and the average runtime printed on standard output. Refer to futhark-test for information on how to format test data. A program will be ignored if it contains no data sets - it will not even be compiled.

If compilation of a program fails, then futhark bench will abort immediately. If execution of a test set fails, an error message will be printed and benchmarking will continue (and --json will write the file), but a non-zero exit code will be returned at the end.

OPTIONS

--backend=name

The backend used when compiling Futhark programs (without leading futhark, e.g. just opencl).

--concurrency=NUM

The number of benchmark programs to prepare concurrently. Defaults to the number of cores available. Prepare means to compile the benchmark, as well as generate any needed datasets. In some cases, this generation can take too much memory, in which case lowering --concurrency may help.

--entry-point=name

Only run entry points with this name.

--exclude-case=TAG

Do not run test cases that contain the given tag. Cases marked with “nobench” or “disable” are ignored by default.

--futhark=program

The program used to perform operations (eg. compilation). Defaults to the binary running futhark bench itself.

--ignore-files=REGEX

Ignore files whose path match the given regular expression.

--json=file

Write raw results in JSON format to the specified file.

--no-tuning

Do not look for tuning files.

--pass-option=opt

Pass an option to benchmark programs that are being run. For example, we might want to run OpenCL programs on a specific device:

futhark bench prog.fut --backend=opencl --pass-option=-dHawaii

--pass-compiler-option=opt

Pass an extra option to the compiler when compiling the programs.

--runner=program

If set to a non-empty string, compiled programs are not run directly, but instead the indicated program is run with its first argument being the path to the compiled Futhark program. This is useful for compilation targets that cannot be executed directly (as with futhark-pyopencl on some platforms), or when you wish to run the program on a remote machine.

--runs=count

The number of runs per data set.

--skip-compilation

Do not run the compiler, and instead assume that each benchmark program has already been compiled. Use with caution.

--timeout=seconds

If the runtime for a dataset exceeds this integral number of seconds, it is aborted. Note that the time is allotted not per run, but for all runs for a dataset. A twenty second limit for ten runs thus means each run has only two seconds (minus initialisation overhead).

A negative timeout means to wait indefinitely.

-v, --verbose

Print verbose information about what the benchmark is doing. Pass multiple times to increase the amount of information printed.

--tuning=EXTENSION

For each program being run, look for a tuning file with this extension, which is suffixed to the name of the program. For example, given --tuning=tuning (the default), the program foo.fut will be passed the tuning file foo.fut.tuning if it exists.

WHAT FUTHARK BENCH MEASURES

futhark bench measures the time it takes to run the given Futhark program by passing the -t FILE option to the generated program. See the man page for the specific compiler to see exactly what is measured.

EXAMPLES

The following program benchmarks how quickly we can sum arrays of different sizes:

-- How quickly can we reduce arrays?
--
-- ==
-- nobench input { 0 }
-- output { 0 }
-- input { 100 }
-- output { 4950 }
-- compiled input { 100000 }
-- output { 704982704 }
-- compiled input { 100000000 }
-- output { 887459712 }

let main(n: i32): i32 =
  reduce (+) 0 (iota n)

SEE ALSO

futhark-c, futhark-test

futhark-c

SYNOPSIS

futhark c [options…] <program.fut>

DESCRIPTION

futhark c translates a Futhark program to sequential C code, and either compiles that C code with a C compiler (see below) to an executable binary program, or produces a .h and .c file that can be linked with other code.. The standard Futhark optimisation pipeline is used, and

The resulting program will read the arguments to the entry point (main by default) from standard input and print its return value on standard output. The arguments are read and printed in Futhark syntax.

OPTIONS

-h

Print help text to standard output and exit.

--library

Generate a library instead of an executable. Appends .c/.h to the name indicated by the -o option to determine output file names.

-o outfile

Where to write the result. If the source program is named foo.fut, this defaults to foo.

--safe

Ignore unsafe in program and perform safety checks unconditionally.

-v verbose

Enable debugging output. If compilation fails due to a compiler error, the result of the last successful compiler step will be printed to standard error.

-V

Print version information on standard output and exit.

-W

Do not print any warnings.

--Werror

Treat warnings as errors.

ENVIRONMENT VARIABLES

CC

The C compiler used to compile the program. Defaults to cc if unset.

CFLAGS

Space-separated list of options passed to the C compiler. Defaults to -O3 -std=c99 if unset.

EXECUTABLE OPTIONS

The following options are accepted by executables generated by futhark c.

-h, --help

Print help text to standard output and exit.

-b, --binary-output

Print the program result in the binary output format. The default is human-readable text, which is very slow.

-D, --debugging

Perform possibly expensive internal correctness checks and verbose logging. Implies -L.

-e, --entry-point=FUN

The entry point to run. Defaults to main.

-L, --log

Print various low-overhead logging information to stderr while running.

-r, --runs=NUM

Perform NUM runs of the program. With -t, the runtime for each individual run will be printed. Additionally, a single leading warmup run will be performed (not counted). Only the final run will have its result written to stdout.

-t, --write-runtime-to=FILE

Print the time taken to execute the program to the indicated file, an integral number of microseconds.

SEE ALSO

futhark-opencl, futhark-cuda, futhark-test

futhark-multicore

SYNOPSIS

futhark multicore [options…] <program.fut>

DESCRIPTION

futhark multicore translates a Futhark program to multithreaded C code, and either compiles that C code with a C compiler to an executable binary program, or produces a .h and .c file that can be linked with other code. The standard Futhark optimisation pipeline is used.

The resulting program will read the arguments to the entry point (main by default) from standard input and print its return value on standard output. The arguments are read and printed in Futhark syntax.

OPTIONS

-h

Print help text to standard output and exit.

--library

Generate a library instead of an executable. Appends .c/.h to the name indicated by the -o option to determine output file names.

-o outfile

Where to write the result. If the source program is named foo.fut, this defaults to foo.

--safe

Ignore unsafe in program and perform safety checks unconditionally.

-v verbose

Enable debugging output. If compilation fails due to a compiler error, the result of the last successful compiler step will be printed to standard error.

-V

Print version information on standard output and exit.

-W

Do not print any warnings.

--Werror

Treat warnings as errors.

ENVIRONMENT VARIABLES

CC

The C compiler used to compile the program. Defaults to cc if unset.

CFLAGS

Space-separated list of options passed to the C compiler. Defaults to -O3 -std=c99 -pthread if unset.

EXECUTABLE OPTIONS

The following options are accepted by executables generated by futhark multicore.

-h, --help

Print help text to standard output and exit.

-b, --binary-output

Print the program result in the binary output format. The default is human-readable text, which is very slow.

-D, --debugging

Perform possibly expensive internal correctness checks and verbose logging. Implies -L.

-e, --entry-point=FUN

The entry point to run. Defaults to main.

-L, --log

Print various low-overhead logging information to stderr while running.

-r, --runs=NUM

Perform NUM runs of the program. With -t, the runtime for each individual run will be printed. Additionally, a single leading warmup run will be performed (not counted). Only the final run will have its result written to stdout.

-t, --write-runtime-to=FILE

Print the time taken to execute the program to the indicated file, an integral number of microseconds.

BUGS

Currently works only on Unix-like systems.

SEE ALSO

futhark-c, futhark-test

futhark-cuda

SYNOPSIS

futhark cuda [options…] <program.fut>

DESCRIPTION

futhark cuda translates a Futhark program to C code invoking CUDA kernels, and either compiles that C code with a C compiler to an executable binary program, or produces a .h and .c file that can be linked with other code. The standard Futhark optimisation pipeline is used.

futhark cuda uses -lcuda -lcudart -lnvrtc to link. If using --library, you will need to do the same when linking the final binary.

The generated CUDA code can be called from multiple CPU threads, as it brackets every API operation with cuCtxPushCurrent() and cuCtxPopCurrent().

OPTIONS

-h

Print help text to standard output and exit.

--library

Generate a library instead of an executable. Appends .c/.h to the name indicated by the -o option to determine output file names.

-o outfile

Where to write the result. If the source program is named foo.fut, this defaults to foo.

--safe

Ignore unsafe in program and perform safety checks unconditionally.

-v verbose

Enable debugging output. If compilation fails due to a compiler error, the result of the last successful compiler step will be printed to standard error.

-V

Print version information on standard output and exit.

-W

Do not print any warnings.

--Werror

Treat warnings as errors.

ENVIRONMENT VARIABLES

CC

The C compiler used to compile the program. Defaults to cc if unset.

CFLAGS

Space-separated list of options passed to the C compiler. Defaults to -O -std=c99 if unset.

EXECUTABLE OPTIONS

Generated executables accept the same options as those generated by futhark-c. The -t option behaves as with futhark-opencl. For commonality, the options use OpenCL nomenclature (“group” instead of “thread block”).

The following additional options are accepted.

-h, --help

Print help text to standard output and exit.

--default-group-size=INT

The default size of thread blocks that are launched. Capped to the hardware limit if necessary.

--default-num-groups=INT

The default number of thread blocks that are launched.

--default-threshold=INT

The default parallelism threshold used for comparisons when selecting between code versions generated by incremental flattening. Intuitively, the amount of parallelism needed to saturate the GPU.

--default-tile-size=INT

The default tile size used when performing two-dimensional tiling (the workgroup size will be the square of the tile size).

--dump-cuda=FILE

Don’t run the program, but instead dump the embedded CUDA kernels to the indicated file. Useful if you want to see what is actually being executed.

--dump-ptx=FILE

Don’t run the program, but instead dump the PTX-compiled version of the embedded kernels to the indicated file.

--load-cuda=FILE

Instead of using the embedded CUDA kernels, load them from the indicated file.

--load-ptx=FILE

Load PTX code from the indicated file.

--nvrtc-option=OPT

Add an additional build option to the string passed to NVRTC. Refer to the CUDA documentation for which options are supported. Be careful - some options can easily result in invalid results.

--print-sizes

Print all sizes that can be set with -size or --tuning.

--size=ASSIGNMENT

Set a configurable run-time parameter to the given value. ASSIGNMENT must be of the form NAME=INT Use --print-sizes to see which names are available.

--tuning=FILE

Read size=value assignments from the given file.

ENVIRONMENT

If run without --library, futhark cuda will invoke a C compiler to compile the generated C program into a binary. This only works if the C compiler can find the necessary CUDA libraries. On most systems, CUDA is installed in /usr/local/cuda, which is usually not part of the default compiler search path. You may need to set the following environment variables before running futhark cuda:

LIBRARY_PATH=/usr/local/cuda/lib64
LD_LIBRARY_PATH=/usr/local/cuda/lib64/
CPATH=/usr/local/cuda/include

SEE ALSO

futhark-opencl

futhark-dataset

SYNOPSIS

futhark dataset [options…]

DESCRIPTION

Generate random values in Futhark syntax, which can be useful when generating input datasets for program testing. All Futhark primitive types are supported. Tuples are not supported. Arrays of specific (non-random) sizes can be generated. You can specify maximum and minimum bounds for values, as well as the random seed used when generating the data. The generated values are written to standard output.

If no -g/--generate options are passed, values are read from standard input, and printed to standard output in the indicated format. The input format (whether textual or binary) is automatically detected.

Returns a nonzero exit code if it fails to write the full output.

OPTIONS

-b, --binary

Output data in binary Futhark format (must precede –generate).

-g type, --generate type

Generate a value of the indicated type, e.g. -g i32 or -g [10]f32.

The type may also be a value, in which case that literal value is generated.

-s int

Set the seed used for the RNG. Zero by default.

--T-bounds=<min:max>

Set inclusive lower and upper bounds on generated values of type T. T is any primitive type, e.g. i32 or f32. The bounds apply to any following uses of the -g option.

You can alter the output format using the following flags. To use them, add them before data generation (–generate):

--text

Output data in text format (must precede –generate). Default.

-t, --type

Output the types of values (textually) instead of the values themselves. Mostly useful when reading values on stdin.

EXAMPLES

Generate a 4 by 2 integer matrix:

futhark dataset -g [4][2]i32

Generate an array of floating-point numbers and an array of indices into that array:

futhark dataset -g [10]f32 --i32-bounds=0:9 -g [100]i32

To generate binary data, the --binary must come before the --generate:

futhark dataset --binary --generate=[42]i32

Create a binary data file from a data file:

futhark dataset --binary < any_data > binary_data

Determine the types of values contained in a data file:

futhark dataset -t < any_data

SEE ALSO

futhark-test, futhark-bench

futhark-doc

SYNOPSIS

futhark doc [options…] dir

DESCRIPTION

futhark doc generates HTML-formatted documentation from Futhark code. One HTML file will be created for each .fut file in the given directory, as well as any file reachable through import expressions. The given Futhark code will be considered as one cohesive whole, and must be type-correct.

Futhark definitions may be documented by prefixing them with a block of line comments starting with -- | (see example below). Simple Markdown syntax is supported within these comments. A link to another identifier is possible with the notation `name`@namespace, where namespace must be either term, type, or mtype (module names are in the term namespace). A file may contain a leading documentation comment, which will be considered the file abstract.

futhark doc will ignore any file whose documentation comment consists solely of the word “ignore”. This is useful for files that contain tests, or are otherwise not relevant to the reader of the documentation.

OPTIONS

-h

Print help text to standard output and exit.

-o outdir

The name of the directory that will contain the generated documentation. This option is mandatory.

-v, --verbose

Print status messages to stderr while running.

-V

Print version information on standard output and exit.

EXAMPLES

-- | Gratuitous re-implementation of `map`@term.
--
-- Does exactly the same.
let mymap = ...

SEE ALSO

futhark-test, futhark-bench

futhark-opencl

SYNOPSIS

futhark opencl [options…] <program.fut>

DESCRIPTION

futhark opencl translates a Futhark program to C code invoking OpenCL kernels, and either compiles that C code with a C compiler to an executable binary program, or produces a .h and .c file that can be linked with other code. The standard Futhark optimisation pipeline is used.

futhark opencl uses -lOpenCL to link (-framework OpenCL on macOS). If using --library, you will need to do the same when linking the final binary.

OPTIONS

-h

Print help text to standard output and exit.

--library

Generate a library instead of an executable. Appends .c/.h to the name indicated by the -o option to determine output file names.

-o outfile

Where to write the result. If the source program is named foo.fut, this defaults to foo.

--safe

Ignore unsafe in program and perform safety checks unconditionally.

-v verbose

Enable debugging output. If compilation fails due to a compiler error, the result of the last successful compiler step will be printed to standard error.

-V

Print version information on standard output and exit.

-W

Do not print any warnings.

--Werror

Treat warnings as errors.

ENVIRONMENT VARIABLES

CC

The C compiler used to compile the program. Defaults to cc if unset.

CFLAGS

Space-separated list of options passed to the C compiler. Defaults to -O -std=c99 if unset.

EXECUTABLE OPTIONS

Generated executables accept the same options as those generated by futhark-c. For the -t option, The time taken to perform device setup or teardown, including writing the input or reading the result, is not included in the measurement. In particular, this means that timing starts after all kernels have been compiled and data has been copied to the device buffers but before setting any kernel arguments. Timing stops after the kernels are done running, but before data has been read from the buffers or the buffers have been released.

The following additional options are accepted.

-h, --help

Print help text to standard output and exit.

--build-option=OPT

Add an additional build option to the string passed to clBuildProgram(). Refer to the OpenCL documentation for which options are supported. Be careful - some options can easily result in invalid results.

--default-group-size=INT

The default size of OpenCL workgroups that are launched. Capped to the hardware limit if necessary.

--default-num-groups=INT

The default number of OpenCL workgroups that are launched.

--default-threshold=INT

The default parallelism threshold used for comparisons when selecting between code versions generated by incremental flattening. Intuitively, the amount of parallelism needed to saturate the GPU.

--default-tile-size=INT

The default tile size used when performing two-dimensional tiling (the workgroup size will be the square of the tile size).

-d, --device=NAME

Use the first OpenCL device whose name contains the given string. The special string #k, where k is an integer, can be used to pick the k-th device, numbered from zero. If used in conjunction with -p, only the devices from matching platforms are considered.

--dump-opencl=FILE

Don’t run the program, but instead dump the embedded OpenCL program to the indicated file. Useful if you want to see what is actually being executed.

--dump-opencl-binary=FILE

Don’t run the program, but instead dump the compiled version of the embedded OpenCL program to the indicated file. On NVIDIA platforms, this will be PTX code.

--load-opencl=FILE

Instead of using the embedded OpenCL program, load it from the indicated file.

--load-opencl-binary=FILE

Load an OpenCL binary from the indicated file.

-p, --platform=NAME

Use the first OpenCL platform whose name contains the given string. The special string #k, where k is an integer, can be used to pick the k-th platform, numbered from zero.

--print-sizes

Print all sizes that can be set with -size or --tuning.

-P, --profile

Gather profiling data while executing and print out a summary at the end. When -r is used, only the last run will be profiled. Implied by -D.

--size=ASSIGNMENT

Set a configurable run-time parameter to the given value. ASSIGNMENT must be of the form NAME=INT Use --print-sizes to see which names are available.

--tuning=FILE

Read size=value assignments from the given file.

--list-devices

List all OpenCL devices and platforms available on the system.

SEE ALSO

futhark-test, futhark-cuda, futhark-c

futhark-pkg

SYNOPSIS

futhark pkg add PKGPATH [X.Y.Z]

futhark pkg check

futhark pkg init PKGPATH

futhark pkg fmt

futhark pkg remove PKGPATH

futhark pkg sync

futhark pkg upgrade

futhark pkg versions

DESCRIPTION

This tool is used to modify the package manifest (futhark.pkg) and download the required packages it describes. futhark pkg is not a build system; you will still need to compile your Futhark code with the usual compilers. The only purpose of futhark pkg is to download code (and perform other package management utility tasks). This manpage is not a general introduction to package management in Futhark; see the User’s Guide for that.

The futhark pkg subcommands will modify only two locations in the file system (relative to the current working directory): the futhark.pkg file, and the contents of lib/. When modifying lib/, futhark pkg constructs the new version in lib~new/ and backs up the old version in lib~old. If futhark pkg should fail for any reason, you can recover the old state by moving lib~old back. These temporary directories are erased if futhark pkg finishes without errors.

The futhark pkg sync and futhark pkg init subcommands are the only ones that actually modifies lib/; the others modify only futhark.pkg and require you to manually run futhark pkg sync afterwards.

Most commands take a -v/--verbose option that makes futhark pkg write running diagnostics to stderr.

Network requests (exclusively HTTP GETs) are done via curl, which must be available on the PATH.

COMMANDS

futhark pkg add PKGPATH [X.Y.Z]

Add the specified package of the given minimum version as a requirement to futhark.pkg. If no version is provided, the newest one is used. If the package is already required in futhark.pkg, the new version requirement will replace the old one.

Note that adding a package does not automatically download it. Run futhark pkg sync to do that.

futhark pkg check

Verify that the futhark.pkg is valid, that all required packages are available in the indicated versions. This command does not check that these versions contain well-formed code. If a package path is defined in futhark.pkg, also checks that .fut files are located at the expected location in the file system.

futhark pkg init PKGPATH

Create a new futhark.pkg defining a package with the given package path, and initially no requirements.

futhark pkg fmt

Reformat the futhark.pkg file, while retaining any comments.

futhark pkg remove PKGPATH

Remove a package from futhark.pkg. Does not remove it from the lib/ directory.

futhark pkg sync

Populate the lib/ directory with the packages listed in futhark.pkg. Warning: this will delete everything in lib/ that does not relate to a file listed in futhark.pkg, as well as any local modifications.

futhark pkg upgrade

Upgrade all package requirements in futhark.pkg to the newest available versions.

futhark pkg versions PKGPATH

Print all available versions for the given package path.

COMMIT VERSIONS

It is possible to use futhark pkg with packages that have not yet made proper releases. This is done via pseudoversions of the form 0.0.0-yyyymmddhhmmss+commitid. The timestamp is not verified against the actual commit. The timestamp ensures that newer commits take precedence if multiple packages depend on a commit version for the same package. If futhark pkg add is given a package with no releases, the most recent commit will be used. In this case, the timestamp is merely set to the current time.

Commit versions are awkward and fragile, and should not be relied upon. Issue proper releases (even experimental 0.x version) as soon as feasible. Released versions also always take precedence over commit versions, since any version number will be greater than 0.0.0.

EXAMPLES

Create a new package that will be hosted at https://github.com/sturluson/edda:

futhark pkg init github.com/sturluson/edda

Add a package dependency:

futhark pkg add github.com/sturluson/hattatal

Download the dependencies:

futhark pkg sync

And then you’re ready to start hacking! (Except that these packages do not actually exist.)

BUGS

Since the lib/ directory is populated with transitive dependencies as well, it is possible for a package to depend unwittingly on one of the dependencies of its dependencies, without the futhark.pkg file reflecting this.

There is no caching of zipballs and version lists between invocations, so the network traffic can be rather heavy.

Only GitHub and GitLab are supported as code hosting sites.

SEE ALSO

futhark-test, futhark-doc

futhark-pyopencl

SYNOPSIS

futhark pyopencl [options…] infile

DESCRIPTION

futhark pyopencl translates a Futhark program to Python code invoking OpenCL kernels, which depends on Numpy and PyOpenCL. By default, the program uses the first device of the first OpenCL platform - this can be changed by passing -p and -d options to the generated program (not to futhark pyopencl itself).

The resulting program will otherwise behave exactly as one compiled with futhark py. While the sequential host-level code is pure Python and just as slow as in futhark py, parallel sections will have been compiled to OpenCL, and runs just as fast as when using futhark opencl. The kernel launch overhead is significantly higher, however, so a good rule of thumb when using futhark pyopencl is to aim for having fewer but longer-lasting parallel sections.

The generated code requires at least PyOpenCL version 2015.2.

OPTIONS

-h

Print help text to standard output and exit.

--library

Instead of compiling to an executable program, generate a Python module that can be imported by other Python code. The module will contain a class of the same name as the Futhark source file with .fut removed. Objects of the class define one method per entry point in the Futhark program, with matching parameters and return value.

-o outfile

Where to write the resulting binary. By default, if the source program is named ‘foo.fut’, the binary will be named ‘foo’.

--safe

Ignore unsafe in program and perform safety checks unconditionally.

-v verbose

Enable debugging output. If compilation fails due to a compiler error, the result of the last successful compiler step will be printed to standard error.

-V

Print version information on standard output and exit.

-W

Do not print any warnings.

--Werror

Treat warnings as errors.

SEE ALSO

futhark-python, futhark-opencl

futhark-python

SYNOPSIS

futhark python [options…] infile

DESCRIPTION

futhark python translates a Futhark program to sequential Python code, which depends on Numpy.

The resulting program will read the arguments to the main function from standard input and print its return value on standard output. The arguments are read and printed in Futhark syntax.

The generated code is very slow, likely too slow to be useful. It is more interesting to use this command’s big brother, futhark-pyopencl.

OPTIONS

-h

Print help text to standard output and exit.

--library

Instead of compiling to an executable program, generate a Python module that can be imported by other Python code. The module will contain a class of the same name as the Futhark source file with .fut removed. Objects of the class define one method per entry point in the Futhark program, with matching parameters and return value.

-o outfile

Where to write the resulting binary. By default, if the source program is named ‘foo.fut’, the binary will be named ‘foo’.

--safe

Ignore unsafe in program and perform safety checks unconditionally.

-v verbose

Enable debugging output. If compilation fails due to a compiler error, the result of the last successful compiler step will be printed to standard error.

-V

Print version information on standard output and exit.

-W

Do not print any warnings.

--Werror

Treat warnings as errors.

SEE ALSO

futhark-pyopencl

futhark-repl

SYNOPSIS

futhark repl [program.fut]

DESCRIPTION

Start an interactive Futhark session. This will let you interactively enter expressions and declarations which are then immediately interpreted. If the entered line can be either a declaration or an expression, it is assumed to be a declaration.

Futhark source files can be loaded using the :load command. This will erase any interactively entered definitions. Use the :help command to see a list of commands. All commands are prefixed with a colon.

futhark repl uses the Futhark interpreter, which grants access to certain special functions. See futhark-run for a description.

OPTIONS

-h

Print help text to standard output and exit.

-V

Print version information on standard output and exit.

SEE ALSO

futhark-run, futhark-test

futhark-run

SYNOPSIS

futhark run <program.fut>

DESCRIPTION

Execute the given program by evaluating the main function with arguments read from standard input, and write the results on standard output.

futhark run is very slow, and in practice only useful for testing, teaching, and experimenting with the language. Certain special debugging functions are available in futhark run:

trace 'a : a -> a

Semantically identity, but prints the value on standard output.

break 'a : a -> a

Semantically identity, but interrupts execution at the calling point, such that the environment can be inspected. Continue execution by entering an empty input line. Breakpoints are only respected when starting a program from the prompt, not when passing a program on the command line.

OPTIONS

-e NAME

Run the given entry point instead of main.

-h

Print help text to standard output and exit.

-V

Print version information on standard output and exit.

-w, --no-warnings

Disable interpreter warnings.

SEE ALSO

futhark-repl, futhark-test

futhark-test

SYNOPSIS

futhark test [options…] infiles…

DESCRIPTION

This tool is used to test Futhark programs based on input/output datasets. If a directory is given, all contained files with a .fut extension are considered.

A Futhark test program is an ordinary Futhark program, with at least one test block describing input/output test cases and possibly other options. A test block consists of commented-out text with the following overall format:

description
==
cases...

The description is an arbitrary (and possibly multiline) human-readable explanation of the test program. It is separated from the test cases by a line containing just ==. Any comment starting at the beginning of the line, and containing a line consisting of just ==, will be considered a test block. The format of a test case is as follows:

[tags { tags... }]
[entry: names...]
[compiled|nobench|random] input ({ values... } | @ filename)
output { values... } | auto output | error: regex

If compiled is present before the input keyword, this test case will never be passed to the interpreter. This is useful for test cases that are annoyingly slow to interpret. The nobench keyword is for data sets that are too small to be worth benchmarking, and only has meaning to futhark-bench.

If input is preceded by random, the text between the curly braces must consist of a sequence of Futhark types, including sizes in the case of arrays. When futhark test is run, a file located in a data/ subdirectory, containing values of the indicated types and shapes is, automatically constructed with futhark-dataset. Apart from sizes, integer constants (with or without type suffix), and floating-point constants (always with type suffix) are also permitted.

If input is followed by an @ and a file name (which must not contain any whitespace) instead of curly braces, values will be read from the indicated file. This is recommended for large data sets. This notation cannot be used with random input.

After the input block, the expected result of the test case is written as either output followed by another block of values, or an expected run-time error, in which a regular expression can be used to specify the exact error message expected. If no regular expression is given, any error message is accepted. If neither output nor error is given, the program will be expected to execute succesfully, but its output will not be validated.

If output is preceded by auto (as in auto output), the expected values are automatically generated by compiling the program with futhark-c and recording its result for the given input (which must not fail). This is usually only useful for testing or benchmarking alternative compilers, and not for testing the correctness of Futhark programs.

Alternatively, instead of input-output pairs, the test cases can simply be a description of an expected compile time type error:

error: regex

This is used to test the type checker.

By default, both the interpreter and compiler is run on all test cases (except those that have specified compiled), although this can be changed with command-line options to futhark test.

Tuple syntax is not supported when specifying input and output values. Instead, you can write an N-tuple as its constituent N values. Beware of syntax errors in the values - the errors reported by futhark test are very poor.

An optional tags specification is permitted in the first test block. This section can contain arbitrary tags that classify the benchmark:

tags { names... }

Tag are sequences of alphanumeric characters, dashes, and underscores, with each tag seperated by whitespace. Any program with the disable tag is ignored by futhark test.

Another optional directive is entry, which specifies the entry point to be used for testing. This is useful for writing programs that test libraries with multiple entry points. Multiple entry points can be specified on the same line by separating them with space, and they will all be tested with the same input/output pairs. The entry directive affects subsequent input-output pairs in the same comment block, and may only be present immediately preceding these input-output pairs. If no entry is given, main is assumed. See below for an example.

For many usage examples, see the tests directory in the Futhark source directory. A simple example can be found in EXAMPLES below.

OPTIONS

--backend=program

The backend used when compiling Futhark programs (without leading futhark, e.g. just opencl).

-c

Only run compiled code - do not run any interpreters.

-C

Compile the programs, but do not run them.

--concurrency=NUM

The number of tests to run concurrently. Defaults to the number of (hyper-)cores available.

--exclude=tag

Do not run test cases that contain the given tag. Cases marked with “disable” are ignored by default.

-i

Only interpret - do not run any compilers.

-t

Type-check the programs, but do not run them.

--futhark=program

The program used to perform operations (eg. compilation). Defaults to the binary running futhark test itself.

--no-terminal

Print each result on a line by itself, without line buffering.

--no-tuning

Do not look for tuning files.

--pass-option=opt

Pass an option to benchmark programs that are being run. For example, we might want to run OpenCL programs on a specific device:

futhark test prog.fut --backend=opencl --pass-option=-dHawaii

--pass-compiler-option=opt

Pass an extra option to the compiler when compiling the programs.

--runner=program

If set to a non-empty string, compiled programs are not run directly, but instead the indicated program is run with its first argument being the path to the compiled Futhark program. This is useful for compilation targets that cannot be executed directly (as with futhark-pyopencl on some platforms), or when you wish to run the program on a remote machine.

--tuning=EXTENSION

For each program being run, look for a tuning file with this extension, which is suffixed to the name of the program. For example, given --tuning=tuning (the default), the program foo.fut will be passed the tuning file foo.fut.tuning if it exists.

EXAMPLES

The following program tests simple indexing and bounds checking:

-- Test simple indexing of an array.
-- ==
-- tags { firsttag secondtag }
-- input { [4,3,2,1] 1 }
-- output { 3 }
-- input { [4,3,2,1] 5 }
-- error: Assertion.*failed

let main (a: []i32) (i: i32): i32 =
  a[i]

The following program contains two entry points, both of which are tested:

let add(x: i32, y: i32): i32 = x + y

-- Test the add1 function.
-- ==
-- entry: add1
-- input { 1 } output { 2 }

entry add1 (x: i32): i32 = add x 1

-- Test the sub1 function.
-- ==
-- entry: sub1
-- input { 1 } output { 0 }

entry sub1 (x: i32): i32 = add x (-1)

The following program contains an entry point that is tested with randomly generated data:

-- ==
-- random input { [100]i32 [100]i32 } auto output
-- random input { [1000]i32 [1000]i32 } auto output

let main xs ys = i32.product (map2 (*) xs ys)

SEE ALSO

futhark-bench, futhark-repl