Command-Line Utilities

DaCe provides a set of command-line utilities to simplify interaction with SDFGs and tools or workflows within the DaCe framework.

sdfgcc - SDFG Command-line Compiler

The SDFG Command-line Compiler sdfgcc enables compilation of SDFG files from the command-line.

Usage:

sdfgcc [-o OUT] [-O] <filepath>

Argument	Required	Description
<filepath>	Yes	Path to the SDFG file to compile.
-O,--optimize		If set, invokes the command-line optimization interface.
-o,--out		If provided, saves the library as the given file or in the specified path, together with a header file.

sdfv - SDFG Viewer

The SDFG Viewer sdfv displays SDFGs in the system’s HTML viewer. It reads an SDFG and creates a temporary HTML file that contains a standalone viewer, which is then opened.

Usage:

sdfv <filepath>

Argument	Required	Description
<filepath>	Yes	Path to the SDFG file to show. If a directory is provided, the tool searches for a file called program.sdfg in said folder

daceprof - Profiler and Report Viewer

The DaCe profiler is a versatile profiling and analysis tool that can provide performance results and performance modeling for DaCe programs. Calling a Python script or module with daceprof instead of python will profile/instrument each individual call to a DaCe program and print the latest report at the end.

The tool can also be used to view a profiling report directly in the console with the -i flag.

If --type is given, performs instrumentation of specific elements in the invoked DaCe program. If nothing is given, the tool will time the entire execution of each program using profile().

Usage:

daceprof [ARGUMENTS] myscript.py [SCRIPT ARGUMENTS]

daceprof [ARGUMENTS] -m package.module [MODULE ARGUMENTS]

daceprof [ARGUMENTS] -i profile.json

Argument	Required	Description
Execution arguments
<scriptpath> OR -m <modulepath> OR -i <filepath>	Yes	Path to the script file, report, or Python module.
Profiling arguments
-r,--repetitions REPETITIONS		Runs each profiled program for the specified number of repetitions (default: 100).
-w,--warmup WARMUP		Number of additional repetitions to run before measuring runtime (default: 0).
-t,--type TYPE		Followed by InstrumentationType, specified which instrumentation type to use. If not given, times the entire SDFG with a wall-clock timer.
--instrument INSTRUMENT		A comma-separated list specifying which SDFG elements to instrument. Can be a comma-separated list of element types from the following: map, tasklet, state, sdfg.
--sequential		Disable CPU multi-threading in code generation.
Data instrumentation
-ds,--save-data		Enable data instrumentation and store all (or filtered) arrays.
-dr,--restore-data		Reproducibly run code by restoring all (or filtered) arrays.
Filtering arguments
-f,--filter		Specifies a filter for elements to instrument.
-df,--filter-data		Specifies a filter for data containers to serialize.
Report arguments
-s,--sort		Sort by a specific criterion. Choices are: min|max|mean|median: Sort by the minimum/maximum/mean/median observed value. counter: Sort by counter name/type. value: Sort by the observed value.
-a,--ascending		If given, sort in ascending order.
-o,--output		If given, saves report in output path.
--csv		Use Comma-Separated Values (CSV) for reporting.

For a more detailed guide on how to profile SDFGs and work with the resulting data, see Profiling and Instrumentation and this tutorial.