muIR - Library

muIR is a library of hardware components for auto generating highly configurable parallel dataflow accelerator. muIR provides the implementation of the following hardware units:

1. A set of highly configurable and parameterizable computation nodes.
2. A set of control units to support arbitrary control path.
3. A collection configurable Memory structures like Cache, Scratchpad memory, and such.
4. A set of standard flexible set of junctions and interfaces to connect different pieces of the design.

Getting Started on a Local Ubuntu Machine

This quick start guide will walk you through installation of Chisel and necessary dependencies:

• sbt: which is the preferred Scala build system and what Chisel uses.

• Verilator:, which compiles Verilog down to C++ for simulation. The included unit testing infrastructure uses this.

(Ubuntu-like) Linux

Install Java

sudo apt-get install default-jdk

Install sbt, which isn't available by default in the system package manager:

echo "deb https://dl.bintray.com/sbt/debian /" | sudo tee -a /etc/apt/sources.list.d/sbt.list
sudo apt-key adv --keyserver hkp://keyserver.ubuntu.com:80 --recv 642AC823
sudo apt-get update
sudo apt-get install sbt

Install Verilator.

We currently recommend Verilator version 4.016.

Follow these instructions to compile it from source.

1. Install prerequisites (if not installed already):

   sudo apt-get install git make autoconf g++ flex bison libfl2 libfl-dev zlibc zlib1g zlib1g-dev

2. Clone the Verilator repository:

   git clone http://git.veripool.org/git/verilator

3. In the Verilator repository directory, check out a known good version:

   git pull
   git checkout verilator_4_016

4. In the Verilator repository directory, build and install:

   unset VERILATOR_ROOT # For bash, unsetenv for csh
   autoconf # Create ./configure script
   ./configure
   make
   sudo make install

Please remember that verialtor should be installed in the default system path, otherwise, chisel-iotesters won't find Verilator and the simulation can not be executed

muIR's dependencies

muIR depends on Berkeley Hardware Floating-Point Units for floating nodes. Therefore, before building muIR you need to clone hardfloat project, build it and then publish it locally on your system. Hardfloat repository has all the necessary information about how to build the project, here we only briefly mention how to build it and then publish it.

git clone https://github.com/ucb-bar/berkeley-hardfloat.git
cd berkeley-hardfloat
sbt "publishLocal"

Testing muIR

muIR is an active project that we students in our lab are actively working on it. Hence, we needed to have a solid testing framework to test any new changes in the repo. Under src/test/main/scala we have provided more than 90 test cases to test different pieces of the project after any changes. To run all the test cases you can run:

    sbt test

It is highly encouraged that if you add any new module to the design you have to write one or two test case for two reason:

1. If there is any new changes, we make sure that the new change doesn't affect your implementation
2. The test case can be used as an example of how someone can use your new module in the rest of the design

What's in the muIR repository?

The muIR library repository contains code that is used to implement library modular hardware components to build hardware accelerators. Hardware generation is done using Chisel, a hardware construction language embedded in Scala. The IR-Lib code base is itself factored into several Scala packages. These packages are all found within the src/main/scala directory. Some of these packages provide Scala utilities for generator configuration, while others contain the actual Chisel RTL generators themselves. Here is a brief description of what can be found in each package:

Please find each package under src/main/scala folder

accel:

src/main/scala/accel

This RTL package contains all the accelerator code used to wrap a dataflow scala file. The top level file is Accelerator.scala. It instantiates and connects three helper blocks: SimpleReg.scala, Cache.scala and Core.scala.

To read more about IR-lib accelerator design and SoC interface, please read the following document:

• Accelerator Interface

node:

src/main/scala/node

This RTL package resembles LLVM IR instruction nodes in our design so that we can target arbitrary dataflow of computation nodes from our input software IR. All these nodes use Handhsaking interface from our interface package to talk with other nodes. The logic within each node can vary from a simple ALU operation like addition to more complicated operations like memory address calculation and so on.

To read more about IR-lib node's design, please read the following document:

• IR Node design

control:

src/main/scala/control

In this package, we have implemented all of our control logic to support arbitrary dataflow between IR-lib's nodes.

• Controlflow

loop:

src/main/scala/loop

This RTL package contains the implementation of our special loop nodes for controlling repetitive execution of the same dataflow. Under this package, we have different implementation of the notion of software loops. The loop structure not only should capture all the live-ins and live-outs to the loop but also it has been able to have support for a different type of loops like serial, parallel and different patterns.

• Parallel Loop
• Nested Parallelism

arbiters:

src/main/scala/arbiters

This RTL package contains a parametrizable set of arbiter implementation that's been used in other packages like memory or junctions.

concurrent:

src/main/scala/concurrent

This RTL package contains the implementation of our concurrent modules to support higher task level parallelism. For example, our task manager implementation exists under this package. Different implementation of task controller can be found under this package.

config:

src/main/scala/config

This utility package provides Scala interfaces for configuring a generator via a dynamically-scoped parameterization library.

dataflow:

src/main/scala/dataflow

This RTL package contains different small dataflow to test the correctness of IR-lib accelerator's design.

dnn:

src/main/scala/dnn

This RTL package contains our computation nodes to implement application-specific accelerators for dnn workloads. For instance, we have different implementation of Systolic arraies in this branch. Another type of computation nodes which exist in this package is our Typed node computation nodes.

FPU:

src/main/scala/FPU

This RTL package provides wrappers around Floating point operations to be integrated with IR-lib design. At this moment, there are two different wrappers in this package. One is a warper for Berkeley hard floating point unit. Second, is a wrapper for embedding Alter's IP cores in our design during the FPGA mapping process.

generator:

src/main/scala/generator

This RTL package contains different test cases generate using our front-end to target IR-lib. There is a wide range of examples to hammer IR-lib for testing different software test cases. The test cases test IR-lib logics in different scenarios such as arbitrary control-flow, nested loops, nested parallel loops, nested parallel loops with arbitrary , and so on.

interfaces:

src/main/scala/interfaces

This package contains all the different definition of our interfaces between nodes. These interfaces connect different types of the node. For instance, for connecting two compute node Databundle is provided, alternatively for connecting control signals ControlBundle is provided.

junctions:

src/main/scala/junctions

This RTL package includes the implementation of generic 1: N, N:1, and M: N connections. All the memory operations in the task are time multiplexed over the junction. There are different implementation of junction in this package. For instance, one possible implementation of junctions is a static tree network or a local bus. We have parameterized the junction implementations in IR-lib so that the designer can control the physical network that the junction is lowered into.

memory:

src/main/scala/memory

This RTL package contains our different memory unit implementation of memory units. These units vary from the different implementation of Caches to Scratchpad memories. There is also a different implementation of IR-lib memory controllers for a different type of data like Scalar, Tensor2D and so such.

utility:

src/main/scala/utility

This utility package provides a variety of common Scala and Chisel constructs that are re-used across multiple other packages,

How to debug your design?

In this document, we have explained how someone can start debugging his generated accelerator using simulation tools.

• Debug design

How to fit design on an FPGA?

In this document, we explain how someone can build an accelerator using IR-Lib and then using template files; he can instantiate a new project and fit the design on and FPGA. Moreover, we explain the software interface to talk with the accelerator.

• Test Fitting

Authors:

• Amirali Sharifian ([email protected])
• Arrvindh Shriraman ([email protected])
• Steve Margerm ([email protected])
• Reza Hojabr ([email protected])
• Parmida Vahdatnia ([email protected])