_   _ ______ 
     /\     | \ | |  ____|
    /  \   _|  \| | |__   
   / /\ \ (_) . ` |  __|  
  / ____ \ _| |\  | |     
 /_/    \_\ |_| \_|_|     
          |_|

Compiling with Arrays

(This is the code corresponding to the paper 'Compiling with Arrays', to be published at ECOOP 2024.)

This artifact contains the Polara language and AINF intermediate representation implemented as a DSL in Lean4.

Arxiv: https://arxiv.org/abs/2405.18242

DOI: ...

Overview: What does the artifact comprise?

Previously you should have downloaded the .tar.gz archive specified via hotcrp, and extracted the .tar.gz in an empty folder. You should now have:

• Type: Artifact Description

  Format: Markdown and PDF

  Location: README.md and README.pdf

• Type: Docker Image with Dependencies

  Format: Docker Image

  Location: docker-image.tar

• Type: Command Line Interface

  Format: Shell Script

  Location: lake.sh

• Type: Source code

  Format: Lean source code

  Location: src/*

Quick-start guide (Kick-the-tires phase)

Please describe how reviewers can check the artifact's integrity and basic functionality.

1. Load the Docker image. Depending on whether you are using Docker or Podman run:

   • docker load -i docker-image.tar, or
   • podman load -i docker-image.tar

2. The Docker image contains an instance of the Lean4 compiler and the lake build system. We provide a thin wrapper for invoking the docker container as if it was a local lake instance, using the lake.sh file.

   Now, you can compile, check proofs, and run the tests:

   • sh lake.sh exe polara (check proofs)

   • sh lake.sh exe test (run tests)

   This will run the Lean compiler on the provided sources. Verify that no errors appear. The expected results are described in further detail below.

(Remark: Depending on how you installed and configured docker, you might need to execute sudo docker load -i and the docker command in the sudo lake.sh instead. Podman does not require sudo.)

For authors claiming an available badge

We offer to publish the artifact on DARTS, in which case the available badge will be issued automatically. If you agree to publishing your artifact under a Creative Commons license, please indicate this here.

For the camera ready, we agree to publishing via DARTS under a creative commons license. (For the artifact evaluation, we will temporarily host the file anonymously via figshare.)

For authors claiming a functional or reusable badge

To verify the functional badge, the reviewers are requested to perform the following steps:

1. Run sh lake.sh exe polara, the output should look like this:

   info: [0/12] Building Polara.Syntax
   info: [1/12] Compiling Polara.Syntax
   info: [1/12] Building Polara.Codegen
   info: [1/12] Building Polara.NbE
   info: [1/12] Building Polara.CSE
   info: [3/12] Compiling Polara.Codegen
   info: [5/12] Compiling Polara.NbE
   info: [7/12] Compiling Polara.CSE
   info: [7/12] Building Polara
   info: [8/12] Compiling Polara
   info: [8/12] Building Main
   info: stdout:
   'Tm.toAINF' does not depend on any axioms
   'AINF.cse' depends on axioms: [Classical.choice, Quot.sound, propext]
   'Tm.norm' depends on axioms: [Classical.choice, Quot.sound, propext]
   info: [10/12] Compiling Main
   info: [12/12] Linking polara
   Success!
   

   What does this mean? At the end of the src/Main.lean file we print out all assumptions and axioms used in the development using:

   #print axioms Tm.toAINF
   #print axioms AINF.cse
   #print axioms Tm.norm
   

   As the reviewer, you should verify that the only axioms we use are Classical.choice, Quot.sound and propext. These axioms are built into Lean and almost unavoidable. You should verify that these are exactly the axioms used, e.g., we did not "cheat" by using additional axioms.

   Also, you could verify that all functions defined in the source files do not use noncomputable def, or sorry. Functions which are not marked as noncomputable are computable, meaning the compiler generates machine code for them. sorry is an escape hatch to avoid having to define a function or proof. We believe that usage of noncomputable or sorry would show up as warnings during compilation.

2. Run sh lake.sh exe test to run the tests, the output should list the tests, all of which succeed, e.g:

   Running tests
   * OK ...
   * OK ...
   ...
   * === 16 / 16 tests passed ===
   

Relation of Artifact and Paper

Section 4 "Mechanization" is based on the contents of the src/ folder. In particular:

• Figure 2+3: src/Polara/Syntax.lean
• Figure 4a: Ty.de in src/Polara/NbE.lean
• Figure 4b+c: Const0.de, Const1.de, Const2.de, Tm.de in src/Polara/NbE.lean
• Figure 4d: quote, splice, Tm.norm in src/Polara/NbE.lean
• Figure 5a: Tm.toAINF in src/Polara/CSE.lean
• Figure 5b: AINF.smart_bnd in src/Polara/CSE.lean
• Figure 6: AINF.cse in src/Polara/CSE.lean
• Theorem 1: Tm.norm in src/Polara/NbE.lean
• Theorem 2: Tm.toAINF in src/Polara/CSE.lean

(Remark: Our development uses intrinsic proofs. An extrinsic proof is a separate function and a proof of a property over the function. An intrinsic proof unifies the function and the proof, meaning that proof checking is performed by type checking. Therefore, you won't find theorems in the source.)

Further

• The two examples from Section 3.3 (a dense neural network layer and convolution) are implemented in src/Polara/Examples.lean as dense and conv.

Differences to the paper

Comparing the functions in the paper to those in the artifact, one can see that, in the latter, functions relating to parametric higher-order abstract syntax (PHOAS) terms as well as the type constructor for terms itself take an additional argument $\Gamma$. This represents the denotation of the variables and is what distinguishes PHOAS from normal HOAS. We omit this technical detail in the paper. The type of terms Tm features constructors for variables and constants (var, cst0 etc.). In the paper, we do not write these constructors explicitly, so we would write x rather than var x. In the artifact, the language has a nat type of natural numbers. This is omitted in the paper for brevity, which means that if-then-else takes a nat as the condition in the artifact, but a fin 2 in the paper.

The smart_bnd function takes an additional number argument, wrapped inside a reader monad, which is used for creating fresh variables. In the paper, we leave this out and just stipulate that the variable is fresh. The same applies to toAINF.

When discussing CSE in the paper, we describe renamings. In the CSE.lean file, the rename functions define how a renaming is actually applied. CSE also requires us to check equality of expressions, which is done with the beq functions. The CSE function in the paper also calls lookup, which is not defined there. It corresponds to the built-in ListMap.lookup.

Not mentioned in the paper are the pretty-printing routines pretty and toString in Syntax.lean and the functions for generating Lean code from AINF in Codegen.lean.

Our code also contains a function Env.or, which merges two environments. This is used to allow CSE to remove redundancies which appear in different, but compatible, environments. In the paper, this is omitted.

For authors claiming a reusable badge

If some parts of your artifacts contains software:

• is your implementation open-source?
• how to recompile the software? If you use benchmarks:
• are the benchmarks public and open-source?

Our implementations and benchmark is licensed under the Apache 2.0 License.

You can recompile the software and test from source by running the Docker container.

Please list any reuse scenarios that you envision for your artifact, i.e., state how the artifact can be reused or repurposed beyond its role in the submitted paper. Example:

We give an example of how the artifact can be reused on additional programs. We describe a simple change to demonstrate how the case studies can be modified and how the new code can be compiled:

Besides the commands above, you can also

• compile and check proofs: sh lake.sh exe polara

• compile and run tests: sh lake.sh exe tests

To demonstrate that our code is not set in stone, e.g., can be modified and reused, the interested reviewer may try following one or more of the following usage scenarios.

1. Open the file src/Polara/Test.lean. This file contains tests that print * OK ... on success and * ERROR ... on failure.

   You can duplicate one of the tests, for example, the file contains a test codegen that creates an expression using the function egypt, transforms it to AINF, then generates Lean code from AINF, and finally evaluates the Lean code.

   In the test, try setting base and expo to other values. This will generate different code producing different outputs, but the result should still be equal to the reference output from egyptLean (in Examples.lean) and the test should still pass.

2. Open the file src/Polara/Examples.lean. This file contains the functions that are tested by the file above.

   In Polara, an expression is either

   • a constant natural .cst0 (.litn n)
   • a constant floating point .cst0 (.litf f)
   • a constant index .cst0 (liti i)
   • an operator, for example cst2 addn a b or cst2 app f x
   • an array construction bld fun i : Gamma (idx n) => e. (The array construction constructs an array of length n by repeatedly evaluating e, with i bound to the values from 0 to n-1. For example, bld fun i => cst2 muln 10 (i2n i) evaluates to #[0, 10, 20].)
   • something else, as described by the inductive Tm in the file src/Polara/Syntax.lean ...

   The egyptLean function is defined as a Lean function calculating base ^ (2 ^ n). The egypt function is defined as a Polara function that calculates the same result, but does so using multiplication instead of powers. This is a good example of how partial evaluation and common subexpression elimination together can optimize a function well:

   def egyptLean (n: Nat) (x: Nat) :=
     n ^ (2 ^ x)
   
   def egypt (n: Nat) {Gamma : Ty → Type} : Tm Gamma (Ty.nat ~> Ty.nat) :=
     let rec foo' (x : Gamma Ty.nat) : Nat → Tm Gamma Ty.nat
     | 0   => var x
     | n+1 => cst2 app (abs fun y => cst2 muln (var y) (var y)) (foo' x n)
     abs fun x => foo' x n
   

   To simulate writing your own Polara function, you could perform the following steps. We define a new variant of egyptLean that performs multiplication instead of powering, and an equivalent variant of the egypt function that performs repeated addition. For this replace the first of the power functions in egypt by multiplication, and the muln operator in egypt by addn (in Examples.lean):

   def egyptLean2 (n: Nat) (x: Nat) :=
     n * (2 ^ x)
   
   def egypt2 (n: Nat) {Gamma : Ty → Type} : Tm Gamma (Ty.nat ~> Ty.nat) :=
     let rec foo' (x : Gamma Ty.nat) : Nat → Tm Gamma Ty.nat
     | 0   => var x
     | n+1 => cst2 app (abs fun y => cst2 addn (var y) (var y)) (foo' x n)
     abs fun x => foo' x n
   

   To tests this we can also duplicate the test function, rewritten such that it invokes egypt2 and egyptLean2 instead, but still compares the two functions for equality (in Test.lean, within the Test namespace):

   def codegen2 : IO Unit := group "codegen" <| do
     let base := 3
     let expo := 5
   
     let e1 := (Tm.norm (fun _ => Tm.cst2 Const2.app
       (egypt2 base) (.cst0 (.litn expo)))).toAINF.cse [] [] |>.codegen id
     let e2 := (Tm.norm (fun _ => (egypt2 base))).toAINF.cse [] [] |>.codegen
       fun x => x ++ s!" {expo}"
     let actual1 ← evalStr e1
     let actual2 ← evalStr e2
     let expected := egyptLean2 expo base
     assertEq "egypt2: generated code evaluates correctly"
       actual1.trim s!"{expected}"
     assertEq "egypt2: generated code evaluates correctly"
       actual2.trim s!"ok: {expected}"
   

   Don't forget to add the test to the main function in Test.lean:

   def main := do
     IO.println ""
     IO.println "Running tests"
     Test.codegen
     Test.codegen2 -- new!
     Test.toAINF
     Test.CSE
     let total := (← success.get).size + (← failure.get).size
     IO.println s!"=== {(← success.get).size} / {total} tests passed"
   

   You can now check that an additional test runs successfully, using sh lake.sh exe test.

3. The implementation of conversion to AINF, i.e. fission, is clean and simple in a few lines. We envision a usage scenario of this artifact is its purpose as a tutorial reference for people who would like to reproduce it.

License

Copyright 2024 the AINF team

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.