NITTA -- Tool for Hard Real-Time CGRA Processors

Overview

Kind: research pet project.

Project status: early prototype.

We develop the tool for generating and programming specialized non-von Neumann CGRA processors used for cyclic execution of control and signal/data processing algorithms. These processors are based on the original Not Instruction Transport Triggered Architecture (NITTA). That allows us to:

• Provide high speed and parallel execution of irregular algorithms (where GPU is not applicable) in hard real-time (clock accuracy).
• Make the processor reconfigurable for different application domains.
• Provide a high-level language for application developers and fast compilation.

Our future users can resolve the following tasks:

• Development of embedded and cyber-physical systems.
• Hardware and software testing and rapid prototyping (HIL and PIL).
• Development of programmable accelerators and coprocessors.

Links

Project page & Publications: https://ryukzak.github.io/projects/nitta/

Maintainer: Aleksandr Penskoi [email protected]

Project chat (telegram): https://t.me/joinchat/BC5sV1GY7ADnCuOsKP-bJw

Project CI chat (telegram): https://t.me/nitta_ci

Contributing

See: CONTRIBUTING.md

Set up development environment

You have at least two options here:

1. Set it up automatically as a docker container using this manual and develop in VS Code.
2. Set it up locally by hand. May be less problematic in some cases.

In this document, we will focus on the local setup.

Mac OS X

Install Stack and required developer tools for Haskell.

brew install ghcup
ghcup install stack

Make sure that PATH contains $HOME/.local/bin.

Make sure that you have up to date version of hlint and fourmolu (as on CI)!

Install icarus-verilog and gtkwave.

brew install icarus-verilog
brew tap homebrew/cask
brew cask install gtkwave

Install npm and required developer tools for UI.

brew install npm yarn
npm install --global tern prettier

Ubuntu

Install Stack and required developer tools for Haskell.

sudo apt-get install haskell-stack
stack install hlint fourmolu

Make sure that PATH contains $HOME/.local/bin.

Make sure that you have up to date version of hlint and fourmolu (as on CI)!

Install icarus-verilog and gtkwave.

sudo apt-get install iverilog
sudo apt-get install gtkwave

Install npm and required developer tools for UI.

sudo apt-get install npm yarn
npm install --global tern prettier yarn

WSL2

Install GHCUP.

curl --proto '=https' --tlsv1.2 -sSf https://get-ghcup.haskell.org | sh

Install icarus-verilog and gtkwave.

sudo apt-get install iverilog
sudo apt-get install gtkwave

Install npm and required developer tools for UI.

sudo apt-get install npm yarn
npm install --global tern prettier yarn

For machine learning base synthesis (optional)

1. Install python package manager: poetry.
2. Install python dependencies: cd ml/synthesis; poetry install

How to Build, Test, Format, Lint & Run

see Makefile as a source of up-to-date command examples. Most general command are represnted as a make targets.

Configuration

• HS_SRC_DIR: Defines the source code directories for Haskell.
• HS_O: Specifies options for Haskell compilation.
• POETRYPATH: Defines the path to the Poetry configuration.
• PYTHONPATH: Specifies the Python path for Poetry.
• POETRY: Command to run Poetry within the specified configuration.
• PYTHON: Command to run Python using Poetry.
• ML_CRAWL_DATA_PATH: Path to store data sets for ML training.
• ML_MODEL_PATH: Path to the machine learning model directory.
• ML_MODEL: Identifies the latest machine learning model in the ML_MODEL_PATH.

nitta (the core of the tool)

• build: Builds the Haskell project using Stack with various options with haddock docs.
• build-prod: Builds the Haskell project with optimized GHC options.
• test: Builds and runs Haskell tests with coverage.
• format: Formats Haskell source code using Fourmolu.
• format-check: Checks Haskell source code formatting.
• lint: Lints Haskell code using HLint and Weeder.
• clean: Cleans the Haskell project with Stack.
• benchmark: Runs a benchmark for the Haskell project.
• benchmark-report: Compares and generates a benchmark report.

UI (to see what actually happening due to synthesis process)

• ui-build: Builds the UI component using Yarn.
• ui-run: Runs the UI component using Yarn.
• ui-format: Formats UI source code using Prettier.
• ui-format-check: Checks UI source code formatting.

Machine Learning based Synthesis

• ml-crawl-data: Crawls data for the machine learning model.
• ml-train-model: Trains the machine learning model.
• ml-format: Formats machine learning code using Black.
• ml-format-check: Checks machine learning code formatting.
• ml-lint: Lints machine learning code using Ruff, MyPy, and Vulture.
• ml-nitta: Runs the machine learning model with Nitta.

Docker image for development

see details in: <./ml/synthesis/README.md>

• docker-dev-build: make dev image for your platform (~15 GB).
• docker-dev-build-with-gpu: make dev image for your platform with GPU support (~25 GB).
• docker-dev-run: run dev image without GPU support.
• docker-dev-run-with-gpu: run dev image with GPU support.

CLI Hints

# Build nitta
$ stack build --fast


# Build frontend
$ stack exec nitta-api-gen
$ yarn --cwd web install
$ yarn --cwd web run build


# Run specific test
$ stack test --test-arguments '-p "pattern for the test name"'

# Run only specific tests in one thread
stack build --test --fast --ta '-p "Divider PU" --num-threads 1'

# Rerun only failed test, if all test passed - run all test
$ stack build nitta:nitta-test --fast --test --test-arguments --rerun

# Run doctest for all files
$ find src -name '*.hs' -exec grep -l '>>>' {} \; | xargs -t -L 1 -P 4 stack exec doctest

# generate haddock for tests
stack exec ghc-pkg unregister interpolate -- --force
stack exec -- haddock test/**/*.hs -odocs -h


# Build only one target and run with UI
$ stack build nitta:nitta --fast && stack exec nitta -- -p=8080 -t=fx32.32 examples/pid.lua

# Run UI without synthesis
$ stack exec nitta -- examples/teacup.lua -p=8080 --method=nosynthesis


# Show profiler report
$ stack build --fast --profile && stack exec nitta --profile -- -t=fx32.32 examples/counter.lua +RTS -p && cat nitta.prof


# Show stack trace if application raise an error
$ stack build --fast --profile && stack exec nitta --profile -- -t=fx32.32 examples/counter.lua +RTS -xc


# Show modules dependency
$ graphmod -q -p src | pbcopy

Usage

$ stack exec nitta -- --help
nitta v0.0.0.1 - tool for hard real-time CGRA processors

nitta [OPTIONS] FILE

Target system configuration:
         --uarch=PATH                   Microarchitecture configuration file
  -a     --auto-uarch                   Use empty microarchitecture and
                                        allocate PUs during synthesis process.
  -t     --type=fxM.B                   Overrides data type specified in
                                        config file
         --io-sync=sync|async|onboard   Overrides IO synchronization mode
                                        specified in config file
         --templates=PATH[:PATH]        Target platform templates (default:
                                        'templates/Icarus:templates/DE0-Nano')
         --frontend-language=Lua|XMILE  Language used to source algorithm
                                        description. (default: decision by file
                                        extension)
Common flags:
  -p     --port=INT                     Run nitta server for UI on specific
                                        port (by default - not run)
  -o     --output-path=PATH             Target system path
Simulation:
  -n=INT                                Number of simulation cycles
  -f     --fsim                         Functional simulation with trace
  -l     --lsim                         Logical (HDL) simulation with trace
         --format=md|json|csv           Simulation output format (default:
                                        'md')
Other:
  -v     --verbose                      Verbose
  -e     --extra-verbose                Extra verbose
  -?     --help                         Display help message
  -V     --version                      Print version information
         --numeric-version              Print just the version number
Synthesis:
  -s     --score=NAME                   Name of the synthesis tree node score
                                        to additionally evaluate. Can be
                                        included multiple times (-s score1 -s
                                        score2). Scores like ml_<model_name>
                                        will enable ML scoring.
  -d     --depth-base=FLOAT             Only for 'TopDownByScore' synthesis:
                                        a [1; +inf) value to be an exponential
                                        base of the depth priority coefficient
                                        (default: 1.4)
  -m     --method=NAME                  Synthesis method
                                        (stateoftheart|topdownbyscore|nosynthesis,
                                        default: stateoftheart). `nosynthesis`
                                        required to run UI without synthesis on
                                        the start.                                        

Logical simulation for a specific algorithm

$ stack exec nitta -- examples/teacup.lua --fsim --format=md -t=fx24.32
| Cycle  | temp_cup_1  | time_0  |
|:-------|:------------|:--------|
| 1      | 180.000     | 0.000   |
| 2      | 178.625     | 0.125   |
| 3      | 177.375     | 0.250   |
| 4      | 176.125     | 0.375   |
| 5      | 174.875     | 0.500   |
| 6      | 173.625     | 0.625   |
| 7      | 172.375     | 0.750   |
| 8      | 171.125     | 0.875   |
| 9      | 169.875     | 1.000   |
| 10     | 168.750     | 1.125   |

Synthesis of a target system for a specific algorithm

$ stack exec nitta -- examples/teacup.lua -v --lsim -t=fx24.32
[NOTICE : NITTA] synthesis process...ok
[NOTICE : NITTA] write target project to: "gen/main"...ok
[NOTICE : NITTA] run testbench (gen/main)...ok
| Cycle  | temp_cup_1  | time_0  |
|:-------|:------------|:--------|
| 1      | 180.000     | 0.000   |
| 2      | 178.625     | 0.125   |
| 3      | 177.375     | 0.250   |
| 4      | 176.125     | 0.375   |
| 5      | 174.875     | 0.500   |
| 6      | 173.625     | 0.625   |
| 7      | 172.375     | 0.750   |
| 8      | 171.125     | 0.875   |
| 9      | 169.875     | 1.000   |
| 10     | 168.750     | 1.125   |