Generate NetList (gnl) is a research tool that allows the generation of logic level synthetic benchmark circuits (netlists). The circuit generation is based on a bottom-up recursive clustering of blocks or modules (subcircuits) into larger modules, which finally results in the complete circuit. The blocks or modules are combined by connecting net terminals and thus creating the nets between the blocks. This net generation process is done according to Rent's rule. A mechanism is provided to prevent the generation of combinational loops, and the maximum path length during net generation can be controlled. Details on the algorithm can be found in various papers.
Copyright (c) 1998-2024 Peter Verplaetse, Dirk Stroobandt
Contact author: [email protected] [email protected]
Affiliation: Ghent University
Department of Electronics and Information Systems
Technologiepark-Zwijnaarde 126,
9052 Gent, Belgium
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
To compile the program, you will need CMake version 3.0 or higher.
- Create a
builddirectory and navigate into it:mkdir build cd build - Run the CMake command to generate the Makefile (require version 3.24 at least):
cmake .. - Compile the program using the generated Makefile:
make
Here, <XX.gnl> is the circuit specification file for GNL. The format of this file will be discussed in the following section.
The program has the following command line interface (set the command line to gnl):
gnl [options] <XX.gnl>
The input file <XX.gnl> is the circuit specification file for gnl. The format of this file will be discussed in the following section. The command line options will be discussed in the next section. It might be useful to have a look at the examples that come with the distribution as well.
The gnl specification file consists of one or multiple library sections, possible module sections and finally one circuit section. Each section starts with a header line between square brackets. Comments can be included in the gnl file, and must be preceded by the comment prefix character '#'.
The library section looks like this:
[library]
name=lib
latch=dff 1 1
latch=dffq 1 2
gate=inv 1 1
gate=and2 2 1
gate=nand3 3 1
gate=or4 4 1
gate=xor2 2 1
The library name must be specified, and is followed by one or multiple lines of the form
{gate|latch} = <cell name> <number of inputs> <number of outputs>
The keyword 'gate' is used for combinational blocks. Sequential blocks are marked by the keyword 'latch'. The cell name should be unique for each library. The number of inputs and number of outputs must be a non-negative number.
The circuit or module section looks like this:
[circuit]
name=large
libraries=lib
distribution=500 500 3000 2000 1500 1000 500
size=30
p=0.7
q=0.6
size=3000
p=0.55
q=0.5
size=9000
I=350
O=150
Circuits start with the header [circuit], whereas modules start with the header [module]. The circuit name is optional; in case no name is specified, the default 'top' is used. Module names are not optional, since they act as library names for parent modules.
The libraries are a list of used libraries or child modules. Hence, a module appears as a library that contains a single (macro-)cell. The distribution is a list of numbers that corresponds to the number of instances of each cell in the libraries. Multiple distribution lines can be specified, and it is advised to use a single distribution entry for each library or child module.
The different Rent regions are specified by the size keyword (which represents the upper bound for that region), and optionally followed by the following entries (note: T is the number of terminals, thus T=I+O with I being the number of inputs and O being the number of outputs. The parameter g is the output fraction, i.e. g=O/T=O/(I+O)):
meanT = <mean T at end of region>
p = <Rent exponent, default: 0.6>
sigmaT = <standard deviation for T at end of region>
q = <exponent for standard deviation for T, default: p-0.1>
meanG = <mean g at end of region, default: 0.3>
sigmaG = <standard deviation for g for the complete region, default: 0.15>
It is not allowed to specify both p and meanT, or both q and sigmaT, since one can be calculated from the other. At least one Rent region must be specified: the end region, for which the region bound equals the total circuit size. This total circuit size must correspond to the size as specified by the library distribution. For this end region, the number of primary inputs and/or outputs can be specified as
I = <number of primary inputs>
O = <number of primary outputs>
If both are specified, this supersedes meanT and meanG. Optionally, a region with boundary size=1 can be specified as well. For this special start region, no p or q value can be specified. It is only useful for multi-level netlist generation, since in that case the calculation of the start region parameters from the library distribution will give bogus results.
This version of gnl provides enhanced control of the maximum path length and the path length distribution. The following method is being used:
- for each combination of nets, the resulting maximum path length is checked against a certain maximum path length. This maximum length is sampled from a user specified distribution, or the global maximum length if no distribution is specified;
- for blocks smaller than the cutoff size (on a logarithmic scale), the maximum path length is scaled proportional to the logarithm of the block size;
By default, the maximum path length is set to 40 blocks, and the cutoff size is set to 70 percent.
Other features are added to increase control over the Rent behavior of the generated circuits. Both options are enabled by default:
- Flipflop insertion: for blocks larger than the cutoff size (on a logarithmic scale), a flip flop may be inserted (with a certain probability) to ensure that a connection can be made without violating path length restrictions.
- Local connections: for blocks larger than the cutoff size (on a logarithmic scale), it is allowed to make connections between in- and outputs of the same module. At higher levels of the hierarchy, this can greatly enhance the control over the generated Rent behavior, especially in combination with the previous method.
Sweep Rent and run VPR
./rent_sweep/sweep.sh
Examples:
./rent_sweep/sweep.sh -p test1 -v on -r "0.3 0.98 0.02"
./rent_sweep/sweep.sh -p <output_dir_path> -v [on|off] -s [on|off] -r <p_start p_end p_step>
Parameters:
-p: Mandatory. The output directory path where generated files will be saved.
-v: Optional. Whether VPR (Verilog-to-Routing) is run ('on') or not ('off'). Default is 'off'.
-s: Optional. Sweep over Rent parameters using the GNL tool. Default is 'on'.
-r: Optional. Range for Rent's exponents to sweep over, given as "start end step". Default is "0.3 0.98 0.02".
./rent_sweep/post_process.sh
Examples Usage:
./rent_sweep/post_process.sh <work_dir_path> -n [on|off] -v [on|off] -r [on|off]
Parameters:
<work_dir_path>: Mandatory. The path to the working directory containing BLIF files and where output files will be generated.
-n --net2rent [on|off]: blif to rent hypergraph, default 'on'. Optional. whether to run netlist2rent.py.
-v --viz [on|off]: rent hypergraph to figure, default 'on'. Optional. whether to run rent2viz.py.
recommend: <work_dir_path> will search .rent files in the current folder and sub-folder
-r -- read [on|off]: Read VPR results for runtime critical pathlength ..., default is 'off'. Optional. whether to run readvpr.py to process VPR results. Default is 'off'.
-m --norm [on|off]: Using normalization rent_graph. Default is off.
-h help
1. General options:
-------------------
-f <file> Options & arguments file
Use this option to specify a file that contains further
command line options and arguments.
-v Verbose mode on
Write log to stdout as well.
-d <level> Debug mode on
For debug purposes. A gnl.debug file will be written. See
debug.h for details on debug levels.
-log <file> Log filename
Concise logging information is written to the default log
file gnl.log. If the environment variable PVLOGS is set and
points to an existing directory, the log file is written to
that directory. A different log file name can be specified
with the -log option.
2. Generation options:
----------------------
-sp Show progress
A decrementing counter is shown, and represents the number
of modules that need to be combined.
-eP <%error> Warn if error on number of pins too big [20]
Show a warning if the final number of pins of a circuit or
module is off by a certain amount.
-eg <%error> Warn if error on final output fraction too big [20]
Show a warning if the final output fraction of a circuit or
module is off by a certain amount.
-nw Don't show warnings
Don't warn if the number of pins or output fraction deviates
too much.
3. Output options:
------------------
-w <formats> Output formats (hnl,netD,nets,info,plot,rtd,dat,blif) [hnl]
Write the circuit netlist or details to the following output
formats (multiple formats can be specified):
hnl: hierarchical netlist format (internal format)
netD: netD format, as used for ISPD98 benchmark circuits
nets: GSRC generic hypergraph format (nets and nodes)
info: information about the netlist
plot: gnuplot file (rtd and dat is written as well)
rtd: list of module size and terminal counts
dat: target and actual average parameters
blif: Berkeley Logic Interchange Format (VTR tools support)
-wm <formats> Output formats for internal macrocells
Write the netlist or details for the modules in case of
multilevel netlist generation.
-wa Write output for all modules (-wm identical to -w)
The -wm option is ignored, and the -w option is also used
for the internal modules that are generated during
multilevel netlist generation.
4. Loop and delay parameters:
-----------------------------
-mpl Maximum path length [40]
Absolute maximum path length
-plc Path length cut-off percentage [70]
Maximum module size (on a logarithmic scale) to scale path
lengths
-dsd Delay shape distribution [ 0 .. 0 1 ]
Distribution from which the maximum delays are sampled
(scaled to the absolute maximum path length). When no
distribution is specified, all maximum path lenths are set to
the absolute maximum path length
-msb Minimum number of blocks for sequential combination [0]
Minimum module size to allow a combination with a sequential
block
-mipl Minimum path length [0]
Minimum path length to allow a net combination
-ap Allow long paths
Deactivates path length control.
-al Allow combinational loops (this implies -ap)
Deactivates combinational loop prevention and path length
control.
5. Parameters for improved control:
-----------------------------------
-nlc Do not allow local connections
Do not allows connections from modules to themselves
-lcc Local connection cut-off percentage [80]
Minimum module size (on a logarithmic scale) to allow local
connections.
-nfi Do not allow flop insertion
No flip flop insertion (equivalent to -fip 0)
-fic Flop insertion cut-off percentage [70]
Minimum module size (on a logarithmic scale) to allow flop
insertion.
-fip Flop insertion cut-off probability [0.1]
Probability that a flip flop is inserted when no connection
is allowed without exceeding the maximum path length
-dbf <factor> Distribution correction bucket factor [1.3]
The terminal distribution is corrected using a feedback
mechanism. The actual distribution mean is calculated using
buckets with a certain width, which appears as a factor in
the logarithmic domain. Use this option to change the
default bucket factor.
-dct <num> Distribution correction threshold [2]
Terminal count distribution correction is applied as soon as
the number of modules in the bucket exceeds this value.
-dtc <fraction> Distribution mean T correction factor [0.2]
The target mean terminal count is corrected by this fraction
of the error between the target and actual mean value.
-dgc <error> Distribution mean G correction factor [0.01]
The target mean output fraction is corrected by this
fraction of the error between the target and actual mean
value.
-mstf <factor> Minimum sigma T factor [2.0]
A cell can not be selected for combination with a module
unless the number of terminals is less than
MeanT+<factor>*SigmaT that corresponds to the number of
blocks of the other module.
6. Other parameters:
--------------------
-cms Combine modules according to module size
By default, the modules are combined in a random order. This
option forces combination of the smaller modules first. This
results in a more regular and balanced partitioning tree.
-mini <min in> Minimum number of intermediate input pins [1]
Minimum number of intermediate input pins.
-mino <min out> Minimum number of intermediate output pins [1]
Minimum number of intermediate output pins.
-seed <seed> Initial seed [1]
Initial seed for random number generation.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent