Grouping

Grouping (clustering) the standard cells in the circuit training netlist.

Prerequisite

Before starting, download the following files. They will be referenced by environment variables setup in the first step.

• hmetis binary (University of Minnesota), alternative link
• plc_wrapper_main binary
• Ariane netlist

Unzip the downloaded netlist file:

$ gzip -d ariane.circuit_graph.pb.txt.gz

Note

If the netlist file is larger than 2 GB, split it into small files. This is due to the text proto file size constraint. The smaller files can be passed as follows: --netlist_file=/path/to/netlist_file_1.pb.txt,/path/to/netlist_file_2.pb.txt

You can use the split_proto_netlist_main script from the grouping package to split the netlist:

$ python split_proto_netlist_main \
      --file_name=netlist.pb.txt \
      --output_dir=/dir/to/output/

1. Setup the variables

$ export OUTPUT_DIR=/path/to/output
$ export BLOCK_NAME=ariane
$ export NETLIST_FILE=/path/to/the/netlist.pb.txt
$ export HMETIS_DIR=/path/to/the/hmetis/binary
$ export PLC_WRAPPER_MAIN=/path/to/the/plc_wrapper_main

2. Run the grouping (clustering) code

$ python circuittraining/grouping/grouper_main \
--output_dir=$OUTPUT_DIR \
--netlist_file=$NETLIST_FILE \
--block_name=$BLOCK_NAME \
--hmetis_dir=$HMETIS_DIR \
--plc_wrapper_main=$PLC_WRAPPER_MAIN \
--fixed_macro_names_regex=<Optional first fixed macro regex> \
--fixed_macro_names_regex=<Optional second fixed macro regex> \
--macro_boundary_x_spacing=<Optional macro to boundary X spacing constraint> \
--macro_boundary_y_spacing=<Optional macro to boundary Y spacing constraint> \
--blockage_file=<Optional blockage file>

Example output example

Metrics similar to the following will be output by the grouping code:

grouper.py:336] Wirelength: 4066003.636051
grouper.py:337] Wirelength cost: 0.248466
grouper.py:338] Congestion cost: 1.662051
grouper.py:339] Overlap cost: 0.000034

We expect the metrics of a clustered netlist to be within a small margin of the original unclustered netlist. This gives us confidence that clustering is a good approximation of the standard cells’ placements.

F.A.Q.

Q: Where are the nodes coordinates of the incoming netlist defined?

A: We assume that a netlist that has all its nodes already placed. It does not have to be a good placement. A low effort, fast version of placement results from any tool can be used.

Q: Could you give some hint about how "distance" and "threshold" parameters are determined, for example in the Ariane example?

A "distance" and "threshold" parameters are specific to each block and technology. They need to be adjusted by trial and error. Current defaults are good for Ariane.

Q: What is the relationship between clustered netlist and adjacency matrix?

A: A cluster (group) of standard cells is treated as a macro. We call them soft macros. The original macro is called hard macro. The adjacency matrix is built using the connections between the set of nodes including hard macros and soft macros.

Q: Assuming the number of nets between cluster_a and cluster_b is 100, what is the value of adj[cluster_a][cluster_b] in the adjacency matrix?

A: The adjacency value is computed as the number of connections between clusters, in this case would be 100.

Q: How does the distance of flip flops get involved in the adjacency matrix?

A: It does not.

Q: Are the values in the adjacency matrix normalized?

A: No. We don't normalize them.