ucb-bar / dsptools

A Library of Chisel3 Tools for Digital Signal Processing

GitHub

DSP Tools Development Environment

Build Status

This repository serves as a good starting point for making and easily testing your various DSP generators in Chisel (1 generator at a time). See UC Berkeley Chisel homepage for more information about Chisel.

For a list of common errors, check out the wiki page. Feel free to add your own!


Key Enhancements

Dsptools is a library that can be used with any Chisel library. Some of the goals of dsptools are to enable:

  1. Pipeline delay checking (Isn't it annoying when the delays of two signals into an operation don't line up because you forgot to delay a corresponding signal in your haste to close timing?)

  2. Enhanced support for designing and testing DSP with generic types (i.e. switching between DSPReal for verifying functional correctness with double-precision floating point and FixedPoint for evaluating fixed-point design metrics by changing a single parameter).

  3. More useful and universal testing platform for numeric types!

Numbers are displayed in their correct formats instead of hex for peek, poke, and expect operations. Additionally, if your tester extends DSPTester, you can optionally dump your test sequence to a Verilog testbench that replays the test for functional verification on all simulation platforms (i.e. Xilinx, Altera, etc. instead of only VCS). The tolerance of comparisons with expected values can also be changed via DSPTester.setTol(floTol = decimal_tolerance, fixedTol = number_of_bits).

  1. Miscellaneous additional features
  • Wide range of LUT modules for ease of generating lookup tables from pre-calculated constants (no intermediate representation)
  • Memory modules that abstract out confusion associated with Chisel Mem
  • Generates useful helper files with each Verilog output (constraints, generator parameters used, etc.).
  • Easier to rename modules & signals and have renaming actually succeed.
  • Expanding Support for non-base-2 math.
  • Adds support for numerical processing in the Chisel Environment via Breeze.

Getting Started

Dsptools is published alongside Chisel, FIRRTL, and the other related projects. It can be used by adding

libraryDependencies += "edu.berkeley.cs" %% "dsptools" % "XXXX"

to your build.sbt, where XXXX is the desired version. See Github for the latest release. Snapshots are also published on Sonatype, which are beneficial if you want to use the latest features.

Projects that dsptools depends on are:


Numeric Typeclasses

This library defines a number of typeclasses for numeric types. A brief explanation of how typeclasses work in scala can be found here and here. Our DSP-specific typeclasses are built on top of spire.

The goal of these typeclasses is to make it easy to write chisel modules that treat the number representation as a parameter. For example, using typeclasses you can write chisel that generates an FIR filter for both real and complex numbers. You can also use typeclasses to write chisel that generates a circuit implementation using floating point (via Verilog's real type). After testing that your circuit implementation works with floating point, you can use the same code to generate a fixed point version of the circuit suitable for synthesis.

For a additional, more detailed description of the Numeric classes in dsptools: see The Numbers ReadMe

A generic function in scala is defined like so:

def func[T](in: T): T

This means that you can call func(obj) for an object of any type. If obj is of type Q, you can write func[Q](obj) to specify that we want the Q version of the generic function func, but this is only necessary if the scala compiler can't figure out what Q is supposed to be.

You can also write

class SomeClass[T]

and use T like it is a real type for any member functions of variables. To write a generic chisel Module, we might try to write

class Passthrough[T](gen: T) extends Module {
  val io = new IO(Bundle {
    val in = Input(gen)
    val out = Output(gen)
  })
  io.out := io.in
}

Here, gen is a parameter specifying the type you want to use for your IO's, so you could write Module(new Passthrough(SInt(width=10))) or Module(new Passthrough(new Bundle { ... })). Unfortunately, there's a problem with this. T can be any type, and a lot of types don't make sense, like String or ()=>Unit. This will not compile, because Input(), Output(), and := are functions defined on chisel types. We can fix this problem by writing

class Passthrough[T<:Data](gen: T) extends Module

This type constraint means that we have to choose T to be a subtype of the chisel type Data. Things like UInt, SInt, and Bundle are subtypes of Data. Now the example above should compile. This example isn't very interesting, though. Data lets you do basic things like assignment and make registers, but doesn't define any mathematical operations, so if we write

class Doubler[T<:Data](gen: T) extends Module {
  val io = IO(new Bundle {
    val in = Input(gen)
    val out = Output(gen)
  })
  io.out := io.in + io.in
}

it won't compile. This is where typeclasses come in. This library defines a trait

trait Real[T] {
  ...
  def plus(x: T, y: T): T
  ...
}

as well as an implicit conversion so that a+b gets converted to Real[T].plus(a,b). Real[T] is a typeclass. Typeclasses are a useful pattern in scala, so there is syntactic sugar to make using them easy:

import dsptools.numbers._
class Doubler[T<:Data:Real](gen: T) extends Module

Note: If you don't include the :Real at the end, the scala compiler will think io.in + io.in is string concatenation and you'll get a weird error saying

[error]  found   : T
[error]  required: String

Some useful typeclasses:

  • Ring

    • defines +, *, -, **, zero, one
    • defined in Spire
    • Read: https://en.wikipedia.org/wiki/Ring_(mathematics)
    • Note: We chose to restrict ourselves to Ring rather than Field because division is particularly expensive and nuanced in hardware. Rather than typing a / b we think it is better to require users to instantiate a module and think about what's going on.
  • Eq

    • defines === and =/= (returning chisel Bools!)
  • PartialOrder

    • extends Eq
    • defines >, <, <=, >= (returning a ValidIO[ComparisonBundle] that has valid false if the objects are not comparable
  • Order

    • extends PartialOrder
    • defines >, <, <=, >=, min, max
  • Sign

    • defines abs, isSignZero, isSignPositive, isSignNegative, isSignNonZero, isSignNonPositive, isSignNonNegative
  • Real

    • extends Ring with Order with Sign
    • defines ceil, round, floor, isWhole
    • defines a bunch of conversion methods from ConvertableTo, e.g. fromDouble, fromInt
  • Integer

    • extends Real
    • defines mod

Rocket-chip

This repo contains a set of tools useful for DSP projects involving rocket-chip:

  1. DspBlock, a generic DSP building block with streaming and memory mapped IO. The type of memory interface is generic and can be selected after the block has been designed. Abstractions are built on top of DspBlock, for example a chain of blocks is a kind of DspBlock.
  2. A diplomatic implementation of AXI4-Stream
  3. Some DMA (stream <-> mem) blocks
  4. Simulation models for chisel-testers for memory interfaces (AHB, APB, AXI4, TileLink) and AXI4-Stream.

This code is maintained by Chick, Angie and Paul. Let us know if you have any questions/feedback!

Copyright (c) 2015 - 2019 The Regents of the University of California. Released under the Modified (3-clause) BSD license.