FØCAL

Architecture-aware profiling for computer vision applications.

For each library that your application uses, f0cal provides a set of tracepoints that perform efficient runtime inspection of the call stack. Other profiling tools can time function calls; f0cal can show you the relationship between call latency, data size, scheduler events, memory caching, and much more.

f0cal runs in a development sandbox built on top of python3-venv and conan. Use it inside the sandbox, just like you would any other Python command line tool. The profiler output is a set of pre-baked reports for your CI/CD pipeline. And if you really want to nerd out, you can access the trace data directly with pandas.

Quick start

We've made compiling and configuring the profiling infrastructure as easy as possible by providing a bootstrap script.

From the bootstrap URL (recommended):

VENV_DIR=_venv
curl bootstrap.f0cal.com/master | python3 - -v ${VENV_DIR} git
source ${VENV_DIR}/bin/activate
eval $(f0cal env activate -a) # only do this the first time
conan install f0cal-opencv/4.0.1@f0cal/testing --build=missing
f0cal pr add -- <your-exe> <your-exe-args> <...>
f0cal pr view

From git:

VENV_DIR=_venv
git clone https://github.com/f0cal/f0cal && python3 f0cal/scripts/bootstrap.py -v ${VENV_DIR} local
source ${VENV_DIR}/bin/activate
eval $(f0cal env activate -a) # only do this the first time
conan install f0cal-opencv/4.0.1@f0cal/testing --build=missing
f0cal pr add -- <your-exe> <your-exe-args> <...>
f0cal pr view

Here's a complete example you can run to test the workflow:

conan install f0cal-opencv/4.0.1@f0cal/testing --build=missing
conan install invisible_demo/0.1@f0cal/testing
f0cal pr add -- Invisibility_Cloak
f0cal pr view

Output

Simple terminal output:

                                    Latency (ms)             Throughput (fps)                     
                              count          min   max  mean              max       min       mean
prefix           width height                                                                     
cv::add          640   480     1158         0.02  0.09  0.05         48461.35  10861.18   19247.80
cv::addWeighted  640   480      772         0.20  1.09  0.47          5059.68    919.17    2142.63
cv::bitwise_and  640   480     2316         0.29  1.42  0.74          3456.45    704.70    1354.75
cv::bitwise_not  640   480      772         0.01  0.18  0.04         69132.39   5417.38   25808.52
cv::countNonZero 3     3       1158         0.00  0.05  0.00       1221001.22  18233.54  302150.20
cv::cvtColor     640   480     2050         0.03  8.16  1.16         33158.70    122.55     859.75
cv::dilate       3     3        772         0.06  0.38  0.17         17585.82   2637.09    5760.23
                 640   480     1544         0.06  0.38  0.17         17585.82   2637.09    5760.23
cv::erode        3     3        386         0.07  0.34  0.19         14511.05   2926.26    5187.10
                 640   480      772         0.07  0.34  0.19         14511.05   2926.26    5187.10
cv::inRange      1     4       1544         0.25  1.17  0.80          3983.87    852.97    1253.34
                 640   480      772         0.25  1.17  0.80          3983.87    852.97    1253.34
cv::morphologyEx 3     3        772         0.06  0.74  0.28         16902.17   1343.41    3574.38
                 640   480     1544         0.06  0.74  0.28         16902.17   1343.41    3574.38

HTML output:

Library support

To view a complete list of the AI/ML/CV libraries that f0cal provides tracepoints for:

conan search -r f0cal