-
Notifications
You must be signed in to change notification settings - Fork 60
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
1 parent
8c1f8da
commit ab31965
Showing
2 changed files
with
292 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,291 @@ | ||
use std::{ | ||
collections::{BTreeMap, BTreeSet}, | ||
fmt::Display, | ||
}; | ||
|
||
use thiserror::Error; | ||
|
||
#[derive(Debug, Clone, Copy, PartialEq, PartialOrd, Eq, Ord)] | ||
pub struct Subscript(char); | ||
|
||
#[derive(Debug, Clone)] | ||
pub enum Subscripts { | ||
Indexes(Vec<Subscript>), | ||
Ellipsis { | ||
start: Vec<Subscript>, | ||
end: Vec<Subscript>, | ||
}, | ||
} | ||
|
||
/// Represents an Einstein summation expression following the notation described [https://onnx.ai/onnx/operators/onnx__Einsum.html](here). | ||
#[derive(Debug, Clone)] | ||
pub struct Einsum { | ||
inputs: Vec<Subscripts>, | ||
output: Option<Subscripts>, | ||
} | ||
|
||
#[derive(Error, Debug)] | ||
pub enum EinsumError { | ||
#[error("invalid character encountered: {0}")] | ||
InvalidCharacter(char), | ||
|
||
#[error("the formula has no inputs")] | ||
MissingInputs, | ||
} | ||
|
||
impl Subscript { | ||
pub fn from(c: char) -> Subscript { | ||
assert!(c.is_alphabetic()); | ||
Subscript(c) | ||
} | ||
} | ||
|
||
fn count_indices(inputs: &[Subscripts]) -> BTreeMap<Subscript, u32> { | ||
let mut count = BTreeMap::new(); | ||
for input in inputs { | ||
for c in input.subscripts() { | ||
count.entry(c).and_modify(|n| *n += 1).or_insert(1); | ||
} | ||
} | ||
count | ||
} | ||
|
||
impl Subscripts { | ||
fn push(&mut self, index: Subscript) { | ||
match self { | ||
Subscripts::Indexes(idxs) => idxs.push(index), | ||
Subscripts::Ellipsis { end, .. } => { | ||
end.push(index); | ||
} | ||
} | ||
} | ||
|
||
fn is_empty(&self) -> bool { | ||
match self { | ||
Subscripts::Indexes(idx) => idx.is_empty(), | ||
Subscripts::Ellipsis { start, end } => start.is_empty() && end.is_empty(), | ||
} | ||
} | ||
|
||
fn subscripts(&self) -> Vec<Subscript> { | ||
match &self { | ||
Subscripts::Indexes(indices) => indices.clone(), | ||
Subscripts::Ellipsis { start, end } => { | ||
start.iter().chain(end.iter()).cloned().collect() | ||
} | ||
} | ||
} | ||
} | ||
|
||
impl Display for Subscript { | ||
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { | ||
write!(f, "{}", self.0) | ||
} | ||
} | ||
|
||
impl Display for Subscripts { | ||
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { | ||
match self { | ||
Subscripts::Indexes(idxs) => { | ||
for i in idxs { | ||
write!(f, "{}", i)?; | ||
} | ||
Ok(()) | ||
} | ||
Subscripts::Ellipsis { start, end } => { | ||
for i in start { | ||
write!(f, "{}", i)?; | ||
} | ||
write!(f, "...")?; | ||
for i in end { | ||
write!(f, "{}", i)?; | ||
} | ||
Ok(()) | ||
} | ||
} | ||
} | ||
} | ||
|
||
impl Einsum { | ||
#[allow(dead_code)] | ||
pub fn from(str: &str) -> Result<Einsum, EinsumError> { | ||
let mut sum = Einsum { | ||
inputs: vec![], | ||
output: None, | ||
}; | ||
|
||
// Parse up to arrow | ||
let mut chars = str.chars(); | ||
let mut current_subscripts = Subscripts::Indexes(vec![]); | ||
let mut after_arrow = false; | ||
while let Some(character) = &chars.next() { | ||
match character { | ||
'-' if chars.next() == Some('>') => { | ||
// Arrow: switch from inputs to outputs | ||
if !current_subscripts.is_empty() { | ||
sum.inputs.push(current_subscripts); | ||
current_subscripts = Subscripts::Indexes(vec![]); | ||
} | ||
if sum.inputs.is_empty() { | ||
return Err(EinsumError::MissingInputs); | ||
} | ||
after_arrow = true; | ||
} | ||
'.' if chars.next() == Some('.') && chars.next() == Some('.') => { | ||
// Ellipsis | ||
current_subscripts = match current_subscripts { | ||
Subscripts::Indexes(idxs) => Subscripts::Ellipsis { | ||
start: idxs, | ||
end: vec![], | ||
}, | ||
Subscripts::Ellipsis { .. } => { | ||
return Err(EinsumError::InvalidCharacter('.')) | ||
} | ||
} | ||
} | ||
' ' => {} | ||
',' if !after_arrow => { | ||
// Next input (cannot occur in output) | ||
sum.inputs.push(current_subscripts); | ||
current_subscripts = Subscripts::Indexes(vec![]); | ||
} | ||
c if c.is_alphabetic() => { | ||
current_subscripts.push(Subscript::from(*c)); | ||
} | ||
_ => return Err(EinsumError::InvalidCharacter(*character)), | ||
} | ||
} | ||
|
||
// If we still have subscripts, they are either the last input or the output | ||
if !current_subscripts.is_empty() | ||
|| matches!(current_subscripts, Subscripts::Ellipsis { .. }) && after_arrow | ||
{ | ||
if after_arrow { | ||
sum.output = Some(current_subscripts); | ||
} else { | ||
sum.inputs.push(current_subscripts); | ||
} | ||
} | ||
|
||
Ok(sum) | ||
} | ||
|
||
fn output_or_implicit_subscripts(&self) -> Vec<Subscript> { | ||
match &self.output { | ||
Some(o) => o.subscripts(), | ||
None => { | ||
// In implicit mode output indices are set to the alphabetically sorted sequence of indices | ||
// appearing exactly once in the equation. | ||
let counts = count_indices(&self.inputs); | ||
let mut keys: Vec<Subscript> = counts | ||
.into_iter() | ||
.filter_map(|(k, v)| if v == 1 { Some(k) } else { None }) | ||
.collect(); | ||
keys.sort(); | ||
keys | ||
} | ||
} | ||
} | ||
|
||
fn contraction_indices(&self) -> Vec<Subscript> { | ||
let count = count_indices(&self.inputs); | ||
let mut subscripts: BTreeSet<Subscript> = count | ||
.into_iter() | ||
.filter_map(|(key, value)| if value > 1 { Some(key) } else { None }) | ||
.collect(); | ||
for c in &self.output_or_implicit_subscripts() { | ||
subscripts.remove(c); | ||
} | ||
subscripts.into_iter().collect() | ||
} | ||
} | ||
|
||
impl Display for Einsum { | ||
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { | ||
write!( | ||
f, | ||
"{}", | ||
self.inputs | ||
.iter() | ||
.map(|x| x.to_string()) | ||
.collect::<Vec<String>>() | ||
.join(",") | ||
)?; | ||
|
||
if let Some(output) = &self.output { | ||
write!(f, " -> {}", output)?; | ||
} | ||
Ok(()) | ||
} | ||
} | ||
|
||
#[cfg(test)] | ||
mod tests { | ||
use super::{count_indices, Einsum, Subscript}; | ||
|
||
pub fn compare_after_reserialize(formula: &str, expected: &str) { | ||
assert_eq!(Einsum::from(formula).unwrap().to_string(), expected); | ||
} | ||
|
||
pub fn expect_fail(formula: &str) { | ||
assert!(Einsum::from(formula).is_err()) | ||
} | ||
|
||
#[test] | ||
pub fn test_parse_einsum() { | ||
compare_after_reserialize("ij,jk->ik", "ij,jk -> ik"); | ||
compare_after_reserialize(" i j, j k -> i k", "ij,jk -> ik"); | ||
compare_after_reserialize(" i j-> i k", "ij -> ik"); | ||
|
||
compare_after_reserialize("a ...d,x... z->a ...z", "a...d,x...z -> a...z"); | ||
compare_after_reserialize(" ...d,x... z->a ...", "...d,x...z -> a..."); | ||
compare_after_reserialize("a...", "a..."); | ||
compare_after_reserialize("a ...d,x... z->...", "a...d,x...z -> ..."); | ||
|
||
expect_fail("ij- >ik"); | ||
expect_fail("->ik"); | ||
expect_fail("a ...d,x... z->a . ..z"); | ||
expect_fail("a...b...c"); | ||
expect_fail("a....b...c"); | ||
expect_fail("a..b...c"); | ||
} | ||
|
||
#[test] | ||
pub fn test_indices() { | ||
let es = Einsum::from("ij,jk->ik").unwrap(); | ||
let out = count_indices(&es.inputs); | ||
assert_eq!(out.len(), 3); | ||
assert_eq!(out[&Subscript::from('i')], 1); | ||
assert_eq!(out[&Subscript::from('j')], 2); | ||
assert_eq!(out[&Subscript::from('k')], 1); | ||
|
||
let es = Einsum::from("i...k,k...m->i...m").unwrap(); | ||
let out = count_indices(&es.inputs); | ||
println!("{:?}", out); | ||
assert_eq!(out.len(), 5); | ||
assert_eq!(out[&Subscript::from('i')], 1); | ||
assert_eq!(out[&Subscript::from('j')], 1); | ||
assert_eq!(out[&Subscript::from('k')], 2); | ||
assert_eq!(out[&Subscript::from('l')], 1); | ||
assert_eq!(out[&Subscript::from('m')], 1); | ||
} | ||
|
||
#[test] | ||
pub fn test_analysis() { | ||
let es = Einsum::from("ij,jk->ik").unwrap(); | ||
assert_eq!(es.contraction_indices(), vec![Subscript::from('j')]); | ||
|
||
let es = Einsum::from("ij,jk").unwrap(); | ||
assert_eq!( | ||
es.output_or_implicit_subscripts(), | ||
vec![Subscript::from('i'), Subscript::from('k')] | ||
); | ||
assert_eq!(es.contraction_indices(), vec![Subscript::from('j')]); | ||
|
||
let transpose = Einsum::from("ba").unwrap(); | ||
assert_eq!( | ||
transpose.output_or_implicit_subscripts(), | ||
vec![Subscript::from('a'), Subscript::from('b')] | ||
); | ||
} | ||
} |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -1,4 +1,5 @@ | ||
mod compiler; | ||
mod einsum; | ||
mod gpu; | ||
mod ir; | ||
pub mod onnx; | ||
|