HEDALS

HEDALS: highly efficient delay-driven approximate logic synthesis.

HEDALS optimizes the delay of approximate circuits under a given error constraint, such as error rate (ER) constraint, mean error distance (MED) constraint, and mean hamming distance (MHD) constraint. Its diagram is as follows.

Publication

C. Meng, Z. Zhou, Y. Yao, S. Huang, Y. Chen, W. Qian, “HEDALS: Highly Efficient Delay-driven Approximate Logic Synthesis,” in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD), 2023.

Dependencies

Reference environment, Ubuntu 20.04 LTS with the following tools and libraries:
- gcc 10.3.0 & g++ 10.3.0
  
  You can install these tools with the following command:
```
sudo apt install gcc-10
sudo apt install g++-10
```
  You also need to check whether the default versions of gcc and g++ are 10.3.0:
```
gcc --version
g++ --version
```
  If the default versions of gcc and g++ are not 10.3.0, please change them to 10.3.0.
- cmake 3.16.3
  
  You can install the tool using the following command:
```
sudo apt install cmake
```
- libboost 1.75.0
  
  You can download libboost 1.75.0, manually compile it, and then install it.
- libreadline 8.0-4
  
  You can install the library using the following command
```
sudo apt install libreadline-dev
```
Alternatively, we package a docker image containing the above dependencies:

https://hub.docker.com/r/changmeng/als_min

How to Download

HEDALS relies on three submodules:

Open-source logic synthesis and verification tool abc
Open-source logic synthesis tool espresso
Open-source approximate circuit library EvoApproxLib

There are two ways of downloading HEDALS:

Clone the HEDALS project, and then update the submodules:

git clone https://github.com/SJTU-ECTL/HEDALS.git
git submodule init
git submodule update

Alternatively, clone the HEDALS project as well as the submodules:

git clone --recursive https://github.com/SJTU-ECTL/HEDALS.git

How to Build

To build, go to the root directory of the project, and then execute:

mkdir build
cd build
cmake ..
make
cd ..

An executable program called hedals.out will be generated at the project root directory.

To clean up, go to the root directory of the project, and then execute:

rm -r build

How to Run

Example command:

./hedals.out -i ./input/benchmark/bacs/mult8.blif -l ./input/standard-cell/nangate_45nm_typ.lib --lacType RAC -f 100032 --metrType MED -e 128 -m 0 -o ./tmp

In this example,

HEDALS inputs the accurate circuit "./input/benchmark/bacs/mult8.blif".
It uses the standard cell library "./input/standard-cell/nangate_45nm_typ.lib" for technology mapping.
It uses the "RAC" local approximate change to simplify the circuit.
It uses 100032 random input patterns to measure the error.
The error constraint is $MED \le 128$
It uses mode 0, that is, applying the priority cut-based method proposed in our paper.
The approximate circuits will be outputted to the folder ./tmp

Configurations

Use the following command to get help:

./hedals.out -h

The following information will be returned:

usage: ./hedals.out --accCirc=string [options] ... 
options:
  -i, --accCirc        path to accurate circuit (string)
      --appCirc        path to approximate circuit (string [=])
  -l, --standCell      path to standard cell library (string [=input/standard-cell/mcnc.genlib])
  -o, --outpPath       path to approximate circuits (string [=tmp])
      --lacType        lac type: CONST, SASIMI, RAC (string [=CONST])
      --metrType       error metric type: ER, MED, MHD (string [=ER])
      --distrType      error distribution type: UNIF, ENUM (string [=UNIF])
      --mapType        mapping type: SCL, LUT (string [=SCL])
  -s, --sourceSeed     seed for randomness (unsigned int [=0])
  -f, --nFrame         #Monte Carlo samples, nFrame should be an integer multiple of 64 (int [=100032])
  -m, --mode           mode selection, 0 or 1
			0: priority cut
			1: maximum flow (int [=0])
      --usePostProc    whether to use post processing for further reducing the area (int [=0])
      --cutSizeLim     size limit of priority cuts (int [=16])
  -e, --errUppBound    error upper bound (double [=0.15])
  -h, --help           print this message

Long parameter	Short parameter	Default value	Function
--accCirc	-i	None	Path to accurate circuit, e.g., ./input/benchmark/bacs/mult8.blif
--appCirc		""	Path to approximate circuit. When "--appCirc" specifies a non-empty string, then the program will iteratively simplify the accurate circuit provided by "--accCirc". Otherwise, if "--appCirc" specifies an empty string "", then the program will report the error of the approximate circuit provided by "--appCirc", compared to the accurate circuit provided by "--accCirc".
--standCell	-l	input/standard-cell/mcnc.genlib	Path to standard cell library, e.g., ./input/standard-cell/nangate_45nm_typ.lib
--outpPath	-o	tmp	Specify the directory where the approximate circuits are outputed.
--lacType		CONST	The type of local approximate changes. There are three types: CONST, SASIMI, RAC. CONST means constant replacement. SASIMI means signal replacement. RAC means approximate resubstitution. Please refer to our paper to see more details about the LAC type.
--metrType		ER	The type of error metrics. There are three metrics: ER, MED, MHD. Please refer to our paper to see their detailed definition.
--distrType		UNIF	The type of input distribution. There are two types. UNIF means uniform distribution. ENUM means enumerating all possible input patterns (only for small circuits).
--mapType		SCL	The type of final mapping target. There are two types. SCL means mapping into standard cell library. LUT means mapping into look up tables.
--sourceSeed	-s	0	The seed used for generating random input patterns for logic simulation. When seed "0" is used, then HEDALS randomly picks a seed. Otherwise, HEDALS uses the specified seed.
--nFrame	-f	100032	The number of random input patterns used for error measurement.
--mode	-m	0	Mode selection. There are two modes: 0, 1. Mode 0 corresponds to the priority cut method proposed in our paper. Mode 1 corresponds to the maximum flow method proposed in our paper.
--usePostProc		0	Whether to use post processing for further reducing the area. When it is 0, post processing is disabled. When it is 1, post processing is enabled.
--cutSizeLim		16	Size limit of priority cuts. Its detailed meaning is defined in our paper.
--errUppBound	-e	0.15	Error upper bound used for approximate logic synthesis. HEDALS will generate an approximate circuits satisfying the bound.
--help	-h	None	Print help.

SJTU-ECTL / HEDALS