Beaver đŸĻĢ

PackageDocumentation

Boost the almighty blue-silver dragon with some magical elixir! 🧙🧙‍♀ī¸đŸ§™â€â™‚ī¸

Motivation

In the de-facto way of using MLIR, we need to work with C/C++, TableGen, CMake and Python (in most of cases). Each language or tool here has some functionalities and convenience we want to leverage. There is nothing wrong choosing the most popular and upstream-supported solution, but having alternative ways to build MLIR-based projects is still valuable or at least worth trying.

Elixir could actually be a good fit as a MLIR front end. Elixir has SSA, pattern-matching, pipe-operator. We can use these language features to define MLIR patterns and pass pipeline in a natural and uniformed way. Elixir is strong-typed but not static-typed which makes it a great choice for quickly building prototypes to validate and explore new ideas.

Here is an example to build and verify a piece of IR in Beaver:

mlir do
  module ctx: ctx do
    Func.func some_func(function_type: Type.function([], [Type.i(32)])) do
      region do
        block bb_entry() do
          v0 = Arith.constant(value: Attribute.integer(Type.i(32), 0)) >>> Type.i(32)
          cond0 = Arith.constant(true) >>> Type.i(1)
          CF.cond_br(cond0, Beaver.Env.block(bb1), {Beaver.Env.block(bb2), [v0]}) >>> []
        end

        block bb1() do
          v1 = Arith.constant(value: Attribute.integer(Type.i(32), 0)) >>> Type.i(32)
          _add = Arith.addi(v0, v0) >>> Type.i(32)
          CF.br({Beaver.Env.block(bb2), [v1]}) >>> []
        end

        block bb2(arg >>> Type.i(32)) do
          v2 = Arith.constant(value: Attribute.integer(Type.i(32), 0)) >>> Type.i(32)
          add = Arith.addi(arg, v2) >>> Type.i(32)
          Func.return(add) >>> []
        end
      end
    end
    |> MLIR.Operation.verify!(debug: true)
  end
end
|> MLIR.Operation.verify!(debug: true)

And a small example to showcase what it is like to define and run a pass in Beaver (with some monad magic):

alias Beaver.MLIR.Dialect.Func

defmodule ToyPass do
  use Beaver.MLIR.Pass, on: "func.func"

  defpat replace_add_op() do
    a = value()
    b = value()
    res = type()
    {op, _t} = TOSA.add(a, b) >>> {:op, [res]}

    rewrite op do
      {r, _} = TOSA.sub(a, b) >>> {:op, [res]}
      replace(op, with: r)
    end
  end

  def run(%MLIR.Operation{} = operation) do
    with "func.func" <- Beaver.MLIR.Operation.name(operation),
          attributes <- Beaver.Walker.attributes(operation),
          2 <- Enum.count(attributes),
          {:ok, _} <- MLIR.Pattern.apply_(operation, [replace_add_op(benefit: 2)]) do
      :ok
    end
  end
end

~m"""
module {
  func.func @tosa_add(%arg0: tensor<1x3xf32>, %arg1: tensor<2x1xf32>) -> tensor<2x3xf32> {
    %0 = "tosa.add"(%arg0, %arg1) : (tensor<1x3xf32>, tensor<2x1xf32>) -> tensor<2x3xf32>
    return %0 : tensor<2x3xf32>
  }
}
""".(ctx)
|> MLIR.Pass.Composer.nested("func.func", [
  ToyPass.create()
])
|> canonicalize
|> MLIR.Pass.Composer.run!()

Goals

Why is it called Beaver?

Beaver is an umbrella species increase biodiversity. We hope this project could enable other compilers and applications in the way a beaver pond becomes the habitat of many other creatures. Many Elixir projects also use animal names as their package names and it is often about raising awareness of endangered species. To read more about why beavers are important to our planet, check out this National Geographic article.

Quick introduction

Beaver is essentially LLVM/MLIR on Erlang/Elixir. It is kind of interesting to see a crossover of two well established communities and four sub-communities. Here are some brief information about each of them.

For Erlang/Elixir forks

For LLVM/MLIR forks

Getting started

Installation

If available in Hex, the package can be installed by adding beaver to your list of dependencies in mix.exs:

def deps do
  [
    {:beaver, "~> 0.3.1"}
  ]
end

Documentation can be generated with ExDoc and published on HexDocs. Once published, the docs can be found at https://hexdocs.pm/beaver.

Erlang apps related to the Beaver

LLVM/MLIR is a giant project, and built around that Beaver have thousands of functions. To properly ship LLVM/MLIR and streamline the development process, we need to carefully break the functionalities at different level into different Erlang apps under the same umbrella.

Notes on consuming and development

How it works?

To implement a MLIR toolkit, we at least need these group of APIs:

We implement the IR API and Pass API with the help of the MLIR C API. There are both lower level APIs generated from the C headers and higher level APIs that are more idiomatic in Elixir. The Pattern API is implemented with the help from the PDL dialect. We are using the lower level IR APIs to compile your Elixir code to PDL. Another way to look at this is that Elixir/Erlang pattern matching is serving as a frontend alternative to PDLL.

Design principles

Transformation over builder

It is very common to use builder pattern to construct IR, especially in an OO programming language like C++/Python. One problem this approach has is that the compiler code looks very different from the code it is generating. Because Erlang/Elixir is SSA by its nature, in Beaver a MLIR Op's creation is very declarative and its container will transform it with the correct contextual information. By doing this, we could:

One example:

module do
  v2 = Arith.constant(1) >>> ~t<i32>
end
# module/1 is a macro, it will transformed the SSA `v2= Arith.constant..` to:
v2 =
 %Beaver.SSA{}
  |> Beaver.SSA.put_arguments(value: ~a{1})
  |> Beaver.SSA.put_block(Beaver.Env.block())
  |> Beaver.SSA.put_ctx(Beaver.Env.context())
  |> Beaver.SSA.put_results(~t<i32>)
  |> Arith.constant()

Also, using the declarative way to construct IR, proper dominance and operand reference is formed naturally.

SomeDialect.some_op do
  region do
    block entry() do
      x = Arith.constant(1) >>> ~t<i32>
      y = Arith.constant(1) >>> ~t<i32>
    end
  end
  region do
    block entry() do
      z = Arith.addi(x, y) >>> ~t<i32>
    end
  end
end

# will be transformed to:

SomeDialect.some_op(
  regions: fn -> do
    region = Beaver.Env.region() # first region created
    block = Beaver.Env.block()
    x = Arith.constant(...)
    y = Arith.constant(...)

    region = Beaver.Env.region() # second region created
    block = Beaver.Env.block()
    z = Arith.addi([x, y, ...]) # x and y dominate z
  end
)

Beaver DSL as higher level AST for MLIR

There should be a 1:1 mapping between Beaver SSA DSL to MLIR SSA. It is possible to do a roundtrip parsing MLIR text format and dump it to Beaver DSL which is Elixir AST essentially. This makes it possible to easily debug a piece of IR in a more programmable and readable way.

In Beaver, working with MLIR should be in one format, no matter it is generating, transforming, debugging.

High level API in Erlang/Elixir idiom

When possible, lower level C APIs should be wrapped as Elixir struct with support to common Elixir protocols. For instance the iteration over one MLIR operation's operands, results, successors, attributes, regions should be implemented in Elixir's Enumerable protocol. This enable the possibility to use the rich collection of functions in Elixir standard libraries and Hex packages.

Is Beaver a compiler or binding to LLVM/MLIR?

Elixir is a programming language built for all purposes. There are multiple sub-ecosystems in the general Erlang/Elixir ecosystem. Each sub-ecosystem appears distinct/unrelated to each other, but they actually complement each other in the real world production. To name a few:

Each of these sub-ecosystems starts with a seed project/library. Beaver should evolve to become a sub-ecosystem for compilers built with Elixir and MLIR.

MLIR context management

When calling higher-level APIs, it is ideal not to have MLIR context passing around everywhere. If no MLIR context provided, an attribute and type getter should return an anonymous function with MLIR context as argument. In Erlang, all values are copied, so it is very safe to pass around these anonymous functions. When creating an operation, these functions will be called with the MLIR context in an operation state. With this approach we achieve both succinctness and modularity, not having a global MLIR context. Usually a function accepting a MLIR context to create an operation or type is called a "creator" in Beaver.

Development

  1. Install Elixir, https://elixir-lang.org/install.html
  2. Install Zig, https://ziglang.org/learn/getting-started/#installing-zig
  3. Install LLVM/MLIR
  1. Run tests
  1. debug
  1. Livebook

Release a new version

Update Elixir source

Linux

Mac

Generate checksum-xxx.exs

mix rustler_precompiled.download Beaver.MLIR.CAPI --all --ignore-unavailable --print

Check the version in the output is correct.

Publish to Hex

BEAVER_BUILD_CMAKE=1 mix hex.publish

Run linters/static analysis

mix doctor
mix credo --all
mix gradient

(Optional) Format CMake files

python3 -m pip install cmake-format
cmake-format -i native/**/CMakeLists.txt native/**/*.cmake