DyRecMul is an original FPGA-based approximate multiplier specifically optimized for machine learning computations. It utilizes dynamically reconfigurable lookup table (LUT) primitives in AMD-Xilinx technology to realize the core part of the computations. Implementation results on an AMD-Xilinx Kintex Ultrascale+ FPGA demonstrate remarkable savings of 64% and 67% in LUT utilization for signed multiplication and multiply-and-accumulation configurations, respectively, when compared to the standard Xilinx multiplier core. Accuracy measurements on four popular deep learning (DL) benchmarks indicate a minimal average accuracy decrease of less than 0.29% during post-training deployment, with the maximum reduction staying less than 0.33%.
DyRecMul has been innovated and developed at the Edge Computing, Communication, and Learning Lab at INRS University in close collaboration with Prof. Pierre Langlois' research group at Polytechnique Montréal.
- VHDL 1993
- AMD_Xilinx Vivado ML 2022.2
- Python 3.10
$ pip install numpy matplotlib
This repository consists of three components: hardware description at the RT level, error analysis, and accuracy measurement for deep learning inference.
DyRecMul_INT8.vhd: INT8 version of DyRecMul
DyRecMul_UINT8.vhd: UNIT8 version of DyRecMul
DyRecMul_MAC.vhd: MAC version of DyRecMul
MAC_Testbench.vhd: Testbench of the MAC version of DyRecMul
The INT8 and UINT8 version of DyRecMul is evaluated through five error metrics, which are Error Probability (EP), Mean Absolute Error (MAE), Mean Relative Error (MRE), Mean Squared Error (MSE), and Normalized Error Distance (NED). All the required modules are available in /Statistical_Tests/stat_test.py, and an example of the evaluation of the designs is presented in /Statistical_Tests/Example.ipynb.
The accuracy of INT8 DyRecMul has been measured using AdaPT tool, available: https://github.com/dimdano/adapt/
Custom approximate multipliers must be provided as C header files which contain a lookup table of all output values that the approximate multiplier generates for each possible combination of inputs.
/Python_Emulator/generate_lut.py generates the lookup table files for INT8 DyRecMul as well as INT5 to INT8 exact multipliers in /Python_Emulator/Output_Files folder.