To run the OpenVX* tracer tool from the command line, invoke the
ovx_tracer.sh script, which is located in the VAD installation directory:
./ovx_tracer.sh <application_path> ?application_switches? – executions <execution-times> -trace_path <path-to-tracer-log> -dump_path <path-to-dumped-data> -output_dir <path-to-output-directory>
application_path– the input application
application_switches– the list of arguments which are being used by the application
executions- the number of times to execute the given application. This gives better accuracy for the performance report (traceML file). Default is
trace_path- the trace path in which a tracer log will be generated. If not specified, tracer log will not be generated. Default is empty.
dump_path- the path in which intermediate nodes outputs (images/tensors) will be created. If not specified, outputs will not be generated. Default is empty.
output_dir- the path to which
tracemlfiles will be generated. Default is
Example of Running the GraphML Creator
Let us take an AlexNet application that has few switches. The syntax of the application:
./alexnet-fp16 1 5 ../../images/Miniskirt1.jpg
In order to run the graphML creator on this application, run the following command:
<VAD_DIRECTORY>/openvx_tracer.sh ./alexnet-fp16 -application_switches "1 5 ../../images/Miniskirt1.jpg"
This command generates the matching
<application> intermediate file in the
Since no switches are defined, the application is executed 1 time, application’s file are generated in the
/tmp/ directory, a tracer report is not generated, and also intermediate nodes tensors/images are not created.
You can view the outcome by opening the generated graphML in the Vision Algorithm Design tool.