This repository contains code used in experiments of the paper "Fixed-Posit: A Floating-Point Representation for Error-Resilient Applications" published in IEEE Transactions on Circuits and Systems II.
The directory structure is as follows:
- pintool: Contains the pintool that replaces single-precision IEEE-754 multiplications with fixed-posit multiplications
- scripts: Contains various scripts used to run experiments
- Download Intel Pin. Our pintool is tested with Pin 3.17.
# Download this repo
git clone https://github.com/COSys-Research/Fixed-Posit.git
# Copy the pintools folder inside Pin's tools folder
# $PIN_HOME is set to the root directot of Intel Pin
cp -r Fixed-Posit/pintool/ $(PIN_HOME)/source/tools/
#Build the fixedpositmul pintool
cd $(PIN_HOME)/source/tools/pintool
make
# This would create a obj-intel64 or obj-ia32 folder with fixedpositmul.so file depending on your machine's architecture.For the experiments we will use a modified version of OpenBLAS repo. The modifications allow OpenBLAS benchmarks to read inputs from the files rather than generating inputs randomly.
#Download this modified OpenBLAS repo
git clone https://github.com/varungohil/OpenBLAS.git
# Compile OpenBLAS and its benchmarks for NEHALEM target
cd $(OPENBLAS_PATH)
make TARGET=NEHALEM
cd $(OPENBLAS_PATH)/benchmark/
make TARGET=NEHALEMCreate two new directories somewhere on your system: one to store the outputs of OpenBLAS benchmarks and the other to store the logs of multiplication operands generated by the pintool.
cd $(OPENBLAS_PATH)/benchmark/
# outputs directory to store the outputs/results of the OpenBLAS benchmarks
mkdir openblas_outputs
# logs directory to store the logs of mutliplication operands generated by the fixedpositmul pintool
mkdir openblas_logsTo setup AxBench, execute the commands given below
#Clone AxBench
git clone https://bitbucket.org/act-lab/axbench.git
cp Fixed-Posit/scripts/compile_axbench.sh $(AXBENCH_PATH)/applications/
cp Fixed-Posit/scripts/qos_blackscholes.py $(AXBENCH_PATH)/applications/blackscholes/scripts/
cp Fixed-Posit/scripts/qos_fft.py $(AXBENCH_PATH)/applications/fft/scripts/
cp Fixed-Posit/scripts/qos_inversek2j.py $(AXBENCH_PATH)/applications/inversek2j/scripts/
cd $(AXBENCH_PATH)/applications/
sh compile_axbench.sh
cd ..
# logs directory to store the logs of mutliplication operands generated by the fixedpositmul pintool
mkdir axbench_logsInstall bitstring python package. Required for generating Verilog codes for fixed-posit multiplier.
pip3 install bitstring
pip3 install opencv-python
# Update pip
python -m pip install -U pip
# Install scikit-image
python -m pip install -U scikit-imageFor OpenBLAS, edit the openblas_benchmark_path, pin_root, pintool_path, log_folder variables in the scripts script_maker_openblas.py and error_script_openblas.py to point to the appropriate directories in your system.
Next run the following:
python3 script_maker_openblas.py
# Running this would generate a shell script run_openblas_exps.sh in the benchmark directory of OpenBLAS
cd (OPENBLAS_PATH)/benchmark/
export OPENBLAS_NUM_THREADS=1
# Since pintool only supports single-threaded applications
sh run_openblas_exps.sh
# Running this shell scripts will run the fixedpositmul pintool with OpenBLAS benchmark and store the results and logs in outputs and logs folder respectively.
python3 error_script_openblas.py
# Running this scripts will compute the average relative error for all workloads for all fixed-posit configurations.The computed errors are reported in the generated comphrehensive_error_openblas.csv file.
For AxBench, edit the axbench_root, pin_root, pintool_path, log_folder variables in the scripts script_maker_openblas.py and error_script_openblas.py to point to the appropriate directories in your system.
Next run the following:
python3 script_maker_axbenchs.py
# Running this would generate a shell script run_axbench_exps.sh
export OPENBLAS_NUM_THREADS=1
# Since pintool only supports single-threaded applications
sh run_axbench_exps.sh
# Running this shell scripts will run the fixedpositmul pintool with OpenBLAS benchmark and store the results and logs in outputs and logs folder respectively.
python3 error_script_axbench.py
# Running this scripts will compute the average relative error for all workloads for all fixed-posit configurations.The computed errors are reported in the generated comphrehensive_error_axbench.csv file.
One can generate the Verilog code for various configurations of fixed-posit multipliers using multiplier_verilog_generator.py. Edit all variables in multiplier_verilog_generator.py as per your needs and run the command below:
python3 multiplier_verilog_generator.pyPass generated Verilog code through synthesis tool (we have used Synopsys Design Compiler) to get area and delay when slack meets. Further, extract the design's netlist from synthesis tool to generate power.
Pass the generated multiplier netlist and input traces of particular application through simulation tool (we have used VCS: Verilog code simulator) and generate Switching Activity Interchange Format (SAIF) file. Generate power by again using the Synopsys Design Compiler and pass netlist, SAIF file and set don't touch rule to check further optimization in the design.
If you use this repository, kindly cite our paper. The bibtex is given below:
@ARTICLE{9399648,
author={Gohil, Varun and Walia, Sumit and Mekie, Joycee and Awasthi, Manu},
journal={IEEE Transactions on Circuits and Systems II: Express Briefs},
title={Fixed-Posit: A Floating-Point Representation for Error-Resilient Applications},
year={2021},
volume={},
number={},
pages={1-1},
doi={10.1109/TCSII.2021.3072217}}