Statheros is a Probabilistic programming language with fixed-point inference by Jacob Laurel, Rem Yang, Atharva Sehgal, Shubham Ugare, Sasa Misailovic at UIUC. This work was published at the 58th Design Automation Conference paper.

@inproceedings{statheros, title={Statheros: Compiler for Efficient Low-Precision Probabilistic Programming}, author={Laurel, Jacob and Yang, Rem and Sehgal, Atharva and Ugare, Shubham and Misailovic, Sasa}, booktitle={2021 58th ACM/IEEE Design Automation Conference (DAC)}, pages={787--792}, year={2021}, organization={IEEE} }

The codebase extends from the first stable version of AutoPPL by James Yang et al. and utilizes a modified version of the fpm fixed-point library by Mike Lankamp.

In addition to the two aforementioned libraries which our codebase extends, to run Statheros we require the following libraries:

• Boost (version 1.63+)

The language itself is in the autoppl folder, which needs the libraries in the include folder to run. The experiments folder consists of the benchmarks we implemented (in the Models subfolder), the experimental driver to run these models (driver_desktop.py), and a notebook to plot the results (Plot_Results.ipynb). To see how to write a probabilistic program, see the benchmarks in experiments/Models. To compile one simple program, see below:

• g++ -std=c++17 -O3 -w -I./ -I./include test.cpp for inference on floating-point numbers (make sure to include mh.hpp)
• g++ -std=c++17 -O3 -w -I./ -I./include test.cpp -D FP -D PREC=? -D DIST_VALUE_PREC=? for inference on fixed-point numbers (make sure to include mh_fixed.hpp). The PREC and DIST_VALUE_PREC macros are optional; the default for both is 16. To manually define either, replace ? with a number in the range [1, 31]. There are also two additional macros that may be defined at compile-time: BIG_DATA allows the language to use a more approximate version of log to speed up computation, and PRINT_ERROR_CHECK prints an indication on whether the inferred values were erroneous.

To run Statheros on a Raspberry Pi or PocketBeagle, gcc/g++-7 is sufficient.

To run Statheros on an Arduino Due, one must download a more recent version of the gcc-arm compiler from https://developer.arm.com/tools-and-software/open-source-software/developer-tools/gnu-toolchain/gnu-rm. We used 'arm-none-eabi-gcc-9.3.1'. Hence, instead of using the standard gcc/g++ compiler, the command to compile will be

• arm-none-eabi-gcc-9.3.1 -std=c++17 -O3 -w -I./ -I./include test.cpp -D FP -D PREC=? -D DIST_VALUE_PREC=?

Inside the Arduino IDE, one can configure the settings to automatically use the downloaded arm-gcc compiler instead of the default version when compiling code for the Due. One does this by going into /.arduino15/packages/arduino/hardware/sam/1.6.12/ and updating the platforms.txt to use the downloaded compiler. Examples of these updated versions can be seen in the ArduinoCompilerConfigs directory. For more information about how to customize the Arduino settings checkout https://arduino.github.io/arduino-cli/0.19/platform-specification/

To run a single example navigate to the Examples folder which contains FixedPoint and FloatingPoint sub-directories. Each contains their respective version of the classic Burglar Alarm model. Use the following compiler flags

• g++ -std=c++17 -O3 -w -I./ -I./include BurglarAlarm_double.cpp (floating point version)
• g++ -std=c++17 -O3 -w -I./ -I./include BurglarAlarm_fixed.cpp -D FP -D PREC=16 -D DIST_VALUE_PREC=16 (Fixed point version with 16 fractional bits for both PREC and DIST_VALUE_PREC)

Compiling this way will produce an executable that will run 20 separate instances of inference on this model. The results can be the results.txt files.

See the LICENSE file. Licenses for the libraries we build upon are available here and here