dovado-rtl 0.10.4

RTL Design Space Exploration on top of Vivado

Table of Contents

1. Installation
2. Usage
   1. Defining Custom Metrics
   2. Examples
      1. neorv32 (VHDL)
      2. neorv32 with boolean (VHDL)
      3. corundum (VERILOG)
      4. CICERO (VERILOG/SYSTEM-VERILOG)
      5. cv32e40p (SYSTEM-VERILOG)
3. RAW'22 Experiments
4. Associated Publication DoVado is a RTL design automation and exploration CLI tool.

Installation

DoVado needs python 3.6 or higher. Install it through pip, on many Linux systems use pip3 to force python 3 installation. Dovado has been tested on Vivado 2018.3.

pip3 install --user --no-cache dovado-rtl

Usage

Dovado has two modes:

• points: design automation mode in which a file containing parameter values must be given and a file containing all the evaluations is returned for some given metrics,
• space: design exploration mode in which parameters and their ranges must be given together with some target metrics and the pareto set of design points with respect to the given metrics is returned.

Table 1: dovado general parameters

parameter	description	mandatory
–file-path	path to the target file	yes
–board	vivado descriptor of a board	yes
–parameters	parameters to use either for points/space (integers and booleans supported)	yes
–clock-port	RTL identifier of the clock port	yes
–implementation	switch to evaluate designs after implementation (default is after synthesis)	no
–incremental	switch to use incremental synthesis/implementation	no
–directives	list of directives to pass to synthesis, place and route (default is RuntimeOptimized for all three)	no
–target-clock	clock (Mhz) to give as a constraint to Vivado (default=1000)	no
–metrics	list of metrics to target using their integer identifier (default mode is interactive, you will be asked after first synthesis/implementation)	no

After those parameters specify points/space both these modes take an argument:

• points argument: specify the path to the csv file containing the design points to be analyzed. The csv file must contain on each line the value for each of the parameters stated through –parameters in the same order,
• space argument: a list of ranges stated as 1 2 3 4 where this way we would be defining two ranges (1, 2) for the first parameter and (3, 4) for the second parameter

No further parameters can be passed to points

parameter	description	mandatory
–power-of-2	list of ’y/n’ to state whether a parameter must be explored stepping power of 2s	no
–many-objective	replace default ga implementation (NSGA-II) with AGEMOEA which is a variant better suited for many objective (e.g. >4) optimization	no
–param-initial-values	parameter values which are guaranteed to be synthesizable to retrieve metric mapping	no
–optimization-runtime	set as a termination condition a timeout as hh:mm:ss	no
–read-design-values	read design values from a csv	no
–disable-approximate	disable approximation	no
–estimation-model	choose Hoeffding Adaptive Tree (default) or Kernel Ridge regressor or Shadow to not use the controller but log anyways	no
–controller-model	choose Mab or Distance-based (default) controller	no
–disable-controller-mab-weight	disable loss weighting in distance controller	no
–n-controllers	set the number of voting controllers (default is 500, too high for many applications)	no

Directory structure is vital for the functioning of the tool:

• VHDL: if a package is used the corresponding folder must be named exactly as the package; if one wants to analyse a module in a project with multiple packages each file belonging to a given package must reside in a subfolder with the same name as the package it belongs to:
  • package-name (top folder must have the name of the top package if it exists or any name if it does not exist)
    • file-1 (belonging to package-name)
    • file-2 (belonging to package-name)
    • subpackage1-name
      • file-1 (belonging to subpackage-name)
      • file-2 (belonging to subpackage-name)
      • …
    • subpackage2-name
      • …
    • …
• VERILOG/SYSTEM-VERILOG: include directives are not supported all files must be in the same folder, no subfolders allowed; no import package allowed.

In order to inspect the tool's work you have several files at hand:

• dovado_work/point_evaluation.csv which has one design evaluation per line, mapped with the points you give dovado; written online during tool's execution in points mode
• dovado_work/space_exploration.csv which has all the design points explored together with the design values; written online during tool's execution both in points mode and in space mode
• dovado_work/design_space.csv and dovado_work/objective_space.csv are written at the end of the design exploration process (space mode) and contain respectively the pareto set of design parameters and the corresponding evaluations

Defining Custom Metrics

Procedure:

1. Create a folder named custom_metrics in the same folder where you are running dovado

   mkdir custom_metrics
   

2. Create the python file which will contain your custom metric

   touch test_metric.py
   

3. Write your metric function, any function you need to carry out the computation and any import for the libraries

   # here any import works
   # e.g. import numpy as np
   import numpy as np
   
   
   def test_metric(**kwargs) -> float:
       # only one metric per file is admitted
       # if you want another custom metric create a new file
       print(kwargs)
       return float(__helper_function(kwargs["frequency"]))
   
   
   def __helper_function(a):
       # Care the underscores '__' are mandatory for helper functions
       # This function won't show as a metric is here only for helping purposes
       return a + 1000
   

4. Run dovado without metric selected:

   dovado --file-path <path to "neorv32/rtl/neorv32/neorv32_top.vhd"> --board xc7k70tfbv676-1 --parameters MEM_INT_IMEM_SIZE --parameters MEM_INT_DMEM_SIZE --clock-port clk_i space 16384 131072 8129 65536 --power-of-2 y --power-of-2 y
   

5. Select your metrics, you will now find your custom metrics after all utilisation metrics provided by your board of choice:

General advice:

• the function must return float (highly recommended to annotate the return type)

• all helper functions must start with double underscore “__”

• relative imports are not supported, use only absolute imports

• all subfolders of custom_metrics are ignored.

• from the **kwargs you can access all the other board metrics, the frequency and all the parameters you are using for explorationo by using either “frequency”, the name you find above or the parameter name in dovado's call e.g:

  kwargs["frequency"]
  kwargs["Slice LUTs*"]
  kwargs["MEM_INT_IMEM_SIZE"]
  

Examples

neorv32 (VHDL)

neorv32 is an embedded RISC-V core.

git clone https://github.com/stnolting/neorv32
cd neorv32/rtl
mkdir neorv32
cp core/* neorv32/
cp  core/mem/neorv32_*.default.vhd neorv32/
sed -i "s/CLOCK_FREQUENCY\s*: natural;\s*-- clock frequency of clk_i in Hz/CLOCK_FREQUENCY              :natural:=100000000; -- clock frequency of clk_i in Hz/" neorv32/neorv32_top.vhd

Changing the name of the core folder, which contains all vhdl files, to the name of the package which is used along the files is mandatory to make dovado get ’use’ directives right. Exploring the parameter space of the top module:

dovado --file-path <path to "neorv32/rtl/neorv32/neorv32_top.vhd"> --board xc7k70tfbv676-1 --parameters MEM_INT_IMEM_SIZE --parameters MEM_INT_DMEM_SIZE --clock-port clk_i --metrics 0 --metrics 1 --metrics 4 --metrics 9 space 16384 131072 8129 65536 --power-of-2 y --power-of-2 y

Above we are optimizing two memory parameters (MEM_INTIMEMSIZE, MEM_INTDMEMSIZE) with clk_i as the clock port with metrics chosen:

• frequency (0)
• LUT occupation (1)
• REGISTER occupation (4)
• BRAM occupation (9)

Ranges are specified after space and we also specify that we want to search only among power of 2’s solutions.

neorv32 boolean (VHDL)

Here an example of exploring boolean parameters, the trick here is to explore them as normal parameters but use as range [0, 1] obviously they can be mixed up with non-boolean parameters during exploration:

dovado --file-path <path to "neorv32/rtl/neorv32/neorv32_top.vhd"> --board xc7k70tfbv676-1 --parameters BOOTLOADER_EN --parameters CPU_EXTENSION_RISCV_A --parameters CPU_EXTENSION_RISCV_B --parameters CPU_EXTENSION_RISCV_C --clock-port clk_i --metrics 0 --metrics 1 --metrics 4 --metrics 9 space 0 1 0 1 0 1 0 1 --disable-approximate

corundum (VERILOG)

corundum is an open-source 100Gbps-NIC.

git clone https://github.com/corundum/corundum
cd corundum/

Exploring the parameter space of the top module:

dovado --file-path <path to "corundum/fpga/common/rtl/cpl_queue_manager.v"> --board xc7k70tfbv676-1 --target-clock 100000 --parameters OP_TABLE_SIZE --parameters QUEUE_INDEX_WIDTH --parameters PIPELINE --clock-port clk --metrics 0 --metrics 1 --metrics 4 --metrics 9 space 8 64 4 11 2 32 --disable-approximate

Using approximation parameters:

dovado --file-path <path to "corundum/fpga/common/rtl/cpl_queue_manager.v"> --board xc7k70tfbv676-1 --target-clock 100000 --parameters OP_TABLE_SIZE --parameters QUEUE_INDEX_WIDTH --parameters PIPELINE --clock-port clk --metrics 0 --metrics 1 --metrics 4 --metrics 9 space 8 64 4 11 2 32

CICERO (VERILOG/SYSTEM-VERILOG)

CICERO is a Domain-Specific Architecture (DSA) for Regular Expression matching. We take off-the-shelf DSA and apply a DSE to maximize Frequency and custom metrics, and minimize LUTs usage according to some metrics. The custom metrics employed are performance of the multi-engine architecture and $I cache size by changing Core#, Internal Parallelism, $I lines.

git clone https://github.com/necst/cicero -b feature/dse
mkdir custom_metrics
cp cicero/hdl_src/cicero_core/custom_metrics/avg_perf.py  custom_metrics/
cp cicero/hdl_src/cicero_core/custom_metrics/isize.py   custom_metrics/

For the exact version run:

dovado --file-path cicero/hdl_src/cicero_core/cicero_core.v --board xcku060-ffva1156-2-i --parameters BB_N --parameters CC_ID_BITS --parameters PC_WIDTH --clock-port s00_axi_aclk --metrics 0 --metrics 1 --metrics 37 --metrics 38 space 2 72 1 3 9 10 --many-objective --disable-approximate

For the approximated version run:

dovado --file-path cicero/hdl_src/cicero_core/cicero_core.v --board xcku060-ffva1156-2-i --parameters BB_N --parameters CC_ID_BITS --parameters PC_WIDTH --clock-port s00_axi_aclk --metrics 0 --metrics 1 --metrics 37 --metrics 38 space 2 72 1 3 9 10 --many-objective 

cv32e40p (SYSTEM-VERILOG)

Vivado does not support for 2018.3 release a complex SystemVerilog top-module with a hierarchy. Therefore, we use a simpler module for showcasing purposes.

git clone https://github.com/openhwgroup/cv32e40p
cd rtl
mkdir testing
cp cv32e40p_fifo.sv testing/

In this project an include directory is used but dovado does not currently support it thus we create a subfolder, name may be whatever, where to isolate the module we are interested in studying. This workaround is only possible if the module one wants to study works standalone without include directives.

dovado --file-path <path to "cv32e40p/rtl/testing/cv32e40p_fifo.sv"> --board xc7k70tfbv676-1 --target-clock 100000 --parameters DEPTH --parameters DATA_WIDTH --clock-port clk_i --metrics 0 --metrics 1 --metrics 4 --metrics 9 space 2 4096 2 64 --power-of-2 y --power-of-2 y --disable-approximate

RAW'22 Experiments

We combined Dovado with Movado capabilities and run DSEs for different purposes and showcasing Movado approximation capabilities. We explore corundum with the corresponding section commands .

Then we explore the RISCV extensions of neorv32, modelling the combination of such extensions with the following custom metric (already written in custom_metrics/extension_score.py):

def extension_score(**kwargs) -> float:
	print("kwargs:")
	print(kwargs)
	return float(__helper_function([int(kwargs["CPU_EXTENSION_RISCV_A"]), int(kwargs["CPU_EXTENSION_RISCV_C"]), int(kwargs["CPU_EXTENSION_RISCV_E"]), int(kwargs["CPU_EXTENSION_RISCV_M"]), int(kwargs["CPU_EXTENSION_RISCV_U"]), int(kwargs["CPU_EXTENSION_RISCV_Zfinx"]), int(kwargs["CPU_EXTENSION_RISCV_Zicsr"]), int(kwargs["CPU_EXTENSION_RISCV_Zifencei"])]))

def __helper_function(variants):
	print(variants)
	return sum(variants) * -1000.00

and run with

dovado --file-path neorv32/rtl/neorv32/neorv32_ProcessorTop_Minimal.vhd --board xa7a12tcpg238-2I --parameters CPU_EXTENSION_RISCV_A --parameters CPU_EXTENSION_RISCV_C --parameters CPU_EXTENSION_RISCV_E --parameters CPU_EXTENSION_RISCV_M --parameters CPU_EXTENSION_RISCV_U --parameters CPU_EXTENSION_RISCV_Zfinx --parameters CPU_EXTENSION_RISCV_Zicsr --parameters CPU_EXTENSION_RISCV_Zifencei --clock-port clk_i space 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1  --disable-approximate

You will be prompted to choose the objective metrics to optimize, where you will be able to select the custom metric. Please remove the final --disable-approximate to run the exact version.

We then explore two DSA for Regular Expressions. We scale up TiReX single core with its internal parallelism, but the project is not open-sourced, hence not replicable here. We then scale out CICERO engines number according to the corresponding examples section procedures.

Associated Publication

If you find this repository useful, please use the following citation:

Dovado with custom metrics and Movado:

@inproceedings{paletti2021online,
  title={Online Learning RTL Synthesis for Automated Design Space Exploration},
  author={Paletti, Daniele and Peverelli, Francesco and Conficconi, Davide and Santambrogio, Marco D},
  booktitle={2022 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW)},
  year={2022},
  organization={IEEE}
}

Original Dovado Publication:

@inproceedings{paletti2021dovado,
  title={Dovado: An Open-Source Design Space Exploration Framework},
  author={Paletti, Daniele and Conficconi, Davide and Santambrogio, Marco D},
  booktitle={2021 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW)},
  pages={128--135},
  year={2021},
  organization={IEEE}
}