futhark bench [options…] programs…


This tool is the recommended way to benchmark Futhark programs. Programs are compiled using the specified backend (c by default), then run a number of times for each test case, and the arithmetic mean runtime and 95% confidence interval 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.


For each program and dataset, futhark bench first does a single “warmup” run that is discarded. After that it uses a two-phase technique.

  1. The initial phase performs ten runs (change with -r), or perform runs for at least half a second, whichever takes longer. If the resulting measurements are sufficiently statistically robust (determined using standard deviation and autocorrelation metrics), the results are produced and the second phase is not entered. Otherwise, the results are discarded and the second phase entered.

  2. The convergence phase keeps performing runs until a measurement of sufficient statistical quality is reached.

The notion of “sufficient statistical quality” is based on heuristics. The intent is that futhark bench will in most cases do the right thing by default, both when benchmarking both long-running programs and short-running programs. If you want complete control, disable the convergence phase with --no-convergence-phase and set the number of runs you want with -r.



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


For a program foo.fut, pass --cache-file foo.fut.EXTENSION. By default, --cache-file is not passed.


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.


Don’t run the convergence phase for longer than this. This does not mean that the measurements have converged. Defaults to 300 seconds (five minutes).


Only run entry points with this name.


Do not run test cases that contain the given tag. Cases marked with “nobench”, “disable”, or “no_foo” (where foo is the backend used) are ignored by default.


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


Ignore files whose path match the given regular expression.


Write raw results in JSON format to the specified file.


Do not look for tuning files.


Do not run the convergence phase.


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 an extra option to the compiler when compiling the programs.


Enable profiling for the binary (by passing --profiling and --logging) and store the recorded information in the file indicated by --json (which is required), along with the other benchmarking results.


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.


The number of runs per data set.


Do not run the compiler, and instead assume that each benchmark program has already been compiled into a server-mode executable. Use with caution.


Ignore the test specification in the program file(s), and instead load them from this other file. These external test specifications use the same syntax as normal, but without line comment prefixes. A == is still expected.


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.


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.


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

-- How quickly can we reduce arrays?
-- ==
-- nobench input { 0i64 }
-- output { 0i64 }
-- input { 100i64 }
-- output { 4950i64 }
-- compiled input { 10000i64 }
-- output { 49995000i64 }
-- compiled input { 1000000i64 }
-- output { 499999500000i64 }

let main(n: i64): i64 =
  reduce (+) 0 (iota n)


futhark-c, futhark-test