futhark profile JSONFILE
This tool produces human-readable profiling information based on information collected with futhark bench. Futhark has only rudimentary support for profiling. While the system can collect information about the run-time behaviour of the program, there is currently no automatic way to connect the information to the program source code. However, the collected information can still be useful for estimating the source of inefficiencies.
The first step is to run futhark bench on
your program, while passing
--json. This will
produce a JSON file containing runtime measurements, as well as
collected profiling information. If you neglect to pass
the profiling information will be missing. The information in the JSON
file is complete, but it is difficult for humans to read.
The next step is to run
futhark profile on the JSON file. For a
prog.json, this will create a top level directory
prog.prof that contains files with human-readable profiling
information. A set of files will be created for each benchmark
dataset. If the original invocation of
futhark bench included
multiple programs, then
futhark profile will create subdirectories
for each program (although all inside the same top level directory).
Supposing a dataset
futhark profile will produce the
following files in the top level directory.
foo.log: the running log produced during execution. Contains many details on dynamic behaviour, depending on the exact backend.
foo.summary: a summary of memory usage and cost centres. For the GPU backends, the cost centres are kernel executions and memory copies.
foo.timeline: a list of all recorded profiling events, in the order in which they occurred, along with their runtime and other available information
The profiling information, including the log, is collected from a
final run performed after all the measured runs. Profiling
information is not collected during the runs that contribute to the
runtime measurement reported by
futhark bench. However, enabling
profiling may still affect performance, as it changes the
behaviour of the run time system.
Alternatively, the JSON can also contain a raw profiling report as
produced by the C API function
directory is still created, but it will only contain a single set of
files, and it will not contain a log.
This shows the sequence of commands one might use to profile the
LocVolCalib.fut, which has three datasets associated with
it, using the
$ futhark bench --backend=hip --profile --json result.json LocVolCalib.fut
$ futhark profile result.json
$ tree result.prof/
Only the C-based backends currently support profiling.
c backend does not actually record useful profiling information.