oprecomp/StaticFPTuner

SmartFPTuner (aka StaticFPTuner version n.2)

SmartFPTuner is a python module for statically tuning the precision of floating point variables.

It takes as input the name of the application whose variables need to be tuned, the input set to be fed to the application and the desired maximum error ratio. The error ratio measures the difference between the output generated with maximum variable precision and the output obtained at tuned precision; smaller ratios are preferable.

SmartFPTuner takes a desired error ratio as input and assigns the minimal number of bit to each benchmark FP variable while ensuring that the corresponding error is smaller than the desired one.

SmartFPTuner has been built on top of StaticFPTuner.

Git repository overview:

• README.md: this README file
• benchmarks: folder containing the benchmarks available for testing
• precision_and_errors: folder containing the already computed data set used for training the ML components of SmartFPTuner and StaticFPTuner_v1
• v1: source code for StaticFPTuner version n.1
• v2: source code for SmartFPTuner, that is, StaticFPTuner versione n.2

Building SmartFPTuner

Requires python > 3.6 Python modules required:

• Empirical Model Learning library (EML)
  • download EML at https://github.com/emlopt/emllib
  • unzip it in <EML_download_destination_folder>
  • mv <EML_downloaded_unzipped> <SmartFPTuner_dir>/eml
  • this exact path is required the correct functioning of SmartFPTuner (this behaviour can be modified by changing the value of the variable <eml_path> in tune_variable_precision.py)
• Tensorflow 1.x
  • instructions at https://www.tensorflow.org/install
  • e.g. pip install tensorflow (CPU only)
• keras
  • instructions at https://keras.io/
• numpy
• scikit-learn
• pandas
• networkX
• yaml

Requires also:

• ILOG CPLEX Optimization Studio > 12
  • installation instructions: https://www.ibm.com/support/knowledgecenter/SSSA5P_12.7.1/ilog.odms.studio.help/Optimization_Studio/topics/PLUGINS_ROOT/ilog.odms.cplex.help/CPLEX/GettingStarted/topics/set_up/setup_synopsis.html

Base usage

• cd <SmartFPTuner_dir>
• python3 tune_variable_precision.py <target_error> <input_set_id>
  • : the application whose variable precisions are to be tuned (supported benchmarks can be found in <SmartFPTuner_dir>/benchmarks/)
  • <target_error>: desired bound on the error -- SmartFPTuner expects the exponent of the desired error ratio. For instance, assume the user wants to bound the error obtained with the tuned precision to be lower than 0.001; this value can be expressed in the form: 1^-{exp}, e.g. 0.001 = 1^{-3}; to impose this bound the value passed to SmartFPTuner as <target_error> must be 3 (the negative of the exponent)
  • <input_set_id>: input set for the benchmark; an integer in the range [0, 29], where each values indexes a specific input set

Scientific experiments example

SmartFPTuner has been extensively used to perform experiments in the transprecision computing area, e.g. to determine the potential energy gains obtainable via static FP-variables fine-tuning with the help of AI techniques. More detail can be found in "Combining Learning and Optimization for Transprecision Computing", Andrea Borghesi, Giuseppe Tagliavini, Michele Lombardi, Luca Benini, Michela Milano, CF2020

• The paper is available at: https://arxiv.org/pdf/2002.10890.pdf