Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

feat: add verification to constant folding #1030

Merged
merged 6 commits into from
May 14, 2024
Merged
Show file tree
Hide file tree
Changes from 5 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 31 additions & 0 deletions hugr/src/algorithm.rs
Original file line number Diff line number Diff line change
Expand Up @@ -4,3 +4,34 @@ pub mod const_fold;
mod half_node;
pub mod merge_bbs;
pub mod nest_cfgs;

#[derive(Debug, Clone, Copy, Ord, Eq, PartialOrd, PartialEq)]
/// A type for algorithms to take as configuration, specifying how much
/// verification they should do. Algorithms that accept this configuration
/// should at least verify that input HUGRs are valid, and that output HUGRs are
/// valid.
///
/// The default level is `None` because verification can be expensive.
pub enum VerifyLevel {
/// Do no verification.
None,
/// Verify using [HugrView::validate_no_extensions]. This is useful when you
/// do not expect valid Extension annotations on Nodes.
///
/// [HugrView::validate_no_extensions]: crate::HugrView::validate_no_extensions
WithoutExtensions,
/// Verify using [HugrView::validate].
///
/// [HugrView::validate]: crate::HugrView::validate
WithExtensions,
}

impl Default for VerifyLevel {
fn default() -> Self {
if cfg!(test) {
Self::WithoutExtensions
} else {
Self::None
}
}
}
227 changes: 191 additions & 36 deletions hugr/src/algorithm/const_fold.rs
Original file line number Diff line number Diff line change
Expand Up @@ -3,8 +3,11 @@
use std::collections::{BTreeSet, HashMap};

use itertools::Itertools;
use thiserror::Error;

use crate::hugr::{SimpleReplacementError, ValidationError};
use crate::types::SumType;
use crate::Direction;
use crate::{
builder::{DFGBuilder, Dataflow, DataflowHugr},
extension::{ConstFoldResult, ExtensionRegistry},
Expand All @@ -19,6 +22,94 @@ use crate::{
Hugr, HugrView, IncomingPort, Node, SimpleReplacement,
};

use super::VerifyLevel;

#[derive(Error, Debug)]
#[allow(missing_docs)]
pub enum ConstFoldError {
#[error("Failed to verify {label} HUGR: {err}")]
VerifyError {
label: String,
#[source]
err: ValidationError,
},
#[error(transparent)]
SimpleReplaceError(#[from] SimpleReplacementError),
}

impl ConstFoldError {
fn verify_err(label: impl Into<String>, err: ValidationError) -> Self {
Self::VerifyError {
label: label.into(),
err,
}
}
}

#[derive(Debug, Clone, Copy, Default)]
/// A configuration for the Constant Folding pass.
pub struct ConstFoldConfig {
verify: VerifyLevel,
}

impl ConstFoldConfig {
/// Create a new `ConstFoldConfig` with default configuration.
pub fn new() -> Self {
Self::default()
}

/// Build a `ConstFoldConfig` with the given [VerifyLevel].
pub fn with_verify(mut self, verify: VerifyLevel) -> Self {
self.verify = verify;
self
}

fn verify_impl(
&self,
label: &str,
h: &impl HugrView,
reg: &ExtensionRegistry,
) -> Result<(), ConstFoldError> {
match self.verify {
VerifyLevel::None => Ok(()),
VerifyLevel::WithoutExtensions => h.validate_no_extensions(reg),
VerifyLevel::WithExtensions => h.validate(reg),
}
.map_err(|err| ConstFoldError::verify_err(label, err))
}

/// Run the Constant Folding pass.
pub fn run(&self, h: &mut impl HugrMut, reg: &ExtensionRegistry) -> Result<(), ConstFoldError> {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

not sure I like having a run method on something with the suffix Config. There's the beginnings of a general "algorithm" interface here (configuration, transformation, verfification). Is it worth deferring some of these changes in to a generic follow up PR?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, I can do that. I didn't want to through away verifying before and after the pass, but I can change that to be guarded behind #[cfg(test)].

self.verify_impl("input", h, reg)?;
loop {
// We can only safely apply a single replacement. Applying a
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yes, the original code was based on the assumption that find_consts would only return rewrites for ops that only had const inputs (which would I think be safe to apply all of them?) but that's not what find_consts does

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes I think that would be safe, if that's what find_consts did.

// replacement removes nodes and edges which may be referenced by
// further replacements returned by find_consts. Even worse, if we
// attempted to apply those replacements, expecting them to fail if
// the nodes and edges they reference had been deleted, they may
// succeed because new nodes and edges reused the ids.
//
// We could be a lot smarter here, keeping track of `LoadConstant`
// nodes and only looking at their out neighbours.
let Some((replace, removes)) = find_consts(h, h.nodes(), reg).next() else {
break;
};
h.apply_rewrite(replace)?;
for rem in removes {
// We are optimistically applying these [RemoveLoadConstant] and
// [RemoveConst] rewrites without checking whether the nodes
// they attempt to remove have remaining uses. If they do, then
// the rewrite fails and we move on.
if let Ok(const_node) = h.apply_rewrite(rem) {
// if the LoadConst was removed, try removing the Const too.
let _ = h.apply_rewrite(RemoveConst(const_node));
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is it safe to ignore all errors here?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes it is, but I'll add a comment to explain why.

}
}
}
self.verify_impl("output", h, reg)
}
}

/// Tag some output constants with [`OutgoingPort`] inferred from the ordering.
fn out_row(consts: impl IntoIterator<Item = Value>) -> ConstFoldResult {
let vec = consts
Expand All @@ -43,9 +134,10 @@ pub(crate) fn sorted_consts(consts: &[(IncomingPort, Value)]) -> Vec<&Value> {
.map(|(_, c)| c)
.collect()
}

/// For a given op and consts, attempt to evaluate the op.
pub fn fold_leaf_op(op: &OpType, consts: &[(IncomingPort, Value)]) -> ConstFoldResult {
match op {
let fold_result = match op {
OpType::Noop { .. } => out_row([consts.first()?.1.clone()]),
OpType::MakeTuple { .. } => {
out_row([Value::tuple(sorted_consts(consts).into_iter().cloned())])
Expand All @@ -69,7 +161,10 @@ pub fn fold_leaf_op(op: &OpType, consts: &[(IncomingPort, Value)]) -> ConstFoldR
ext_op.constant_fold(consts)
}
_ => None,
}
};
assert!(fold_result.as_ref().map_or(true, |x| x.len()
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

should this be a debug assert?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

lol, I thought that's what assert was. will change.

== op.value_port_count(Direction::Outgoing)));
fold_result
}

/// Generate a graph that loads and outputs `consts` in order, validating
Expand Down Expand Up @@ -140,18 +235,16 @@ fn fold_op(
})
.unzip();
// attempt to evaluate op
let folded = fold_leaf_op(neighbour_op, &in_consts)?;
let (op_outs, consts): (Vec<_>, Vec<_>) = folded.into_iter().unzip();
let nu_out = op_outs
let (nu_out, consts): (HashMap<_, _>, Vec<_>) = fold_leaf_op(neighbour_op, &in_consts)?
.into_iter()
.enumerate()
.filter_map(|(i, out)| {
// map from the ports the op was linked to, to the output ports of
// the replacement.
hugr.single_linked_input(op_node, out)
.map(|np| (np, i.into()))
.filter_map(|(i, (op_out, konst))| {
// for each used port of the op give the nu_out entry and the
// corresponding Value
hugr.single_linked_input(op_node, op_out)
.map(|np| ((np, i.into()), konst))
})
.collect();
.unzip();
let replacement = const_graph(consts, reg);
let sibling_graph = SiblingSubgraph::try_from_nodes([op_node], hugr)
.expect("Operation should form valid subgraph.");
Expand All @@ -172,39 +265,17 @@ fn get_const(hugr: &impl HugrView, op_node: Node, in_p: IncomingPort) -> Option<
let (load_n, _) = hugr.single_linked_output(op_node, in_p)?;
let load_op = hugr.get_optype(load_n).as_load_constant()?;
let const_node = hugr
.linked_outputs(load_n, load_op.constant_port())
.exactly_one()
.ok()?
.single_linked_output(load_n, load_op.constant_port())?
.0;

let const_op = hugr.get_optype(const_node).as_const()?;

// TODO avoid const clone here
Some((const_op.as_ref().clone(), load_n))
}

/// Exhaustively apply constant folding to a HUGR.
pub fn constant_fold_pass(h: &mut impl HugrMut, reg: &ExtensionRegistry) {
loop {
// would be preferable if the candidates were updated to be just the
// neighbouring nodes of those added.
let rewrites = find_consts(h, h.nodes(), reg).collect_vec();
if rewrites.is_empty() {
break;
}
for (replace, removes) in rewrites {
h.apply_rewrite(replace).unwrap();
for rem in removes {
if let Ok(const_node) = h.apply_rewrite(rem) {
// if the LoadConst was removed, try removing the Const too.
if h.apply_rewrite(RemoveConst(const_node)).is_err() {
// const cannot be removed - no problem
continue;
}
}
}
}
}
pub fn constant_fold_pass<H: HugrMut>(h: &mut H, reg: &ExtensionRegistry) {
ConstFoldConfig::default().run(h, reg).unwrap()
}

#[cfg(test)]
Expand Down Expand Up @@ -395,4 +466,88 @@ mod test {
let expected = Value::false_val();
assert_fully_folded(&h, &expected);
}

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is it worth also including the test from #996 ?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes, will do.

#[test]
fn orphan_output() {
// pseudocode:
// x0 := bool(true)
// x1 := not(x0)
// x2 := or(x0,x1)
// output x2 == true;
//
// We arange things so that the `or` folds away first, leaving the not
// with no outputs.
use crate::hugr::NodeType;
use crate::ops::handle::NodeHandle;

let mut build = DFGBuilder::new(FunctionType::new(type_row![], vec![BOOL_T])).unwrap();
let true_wire = build.add_load_value(Value::true_val());
// this Not will be manually replaced
let orig_not = build.add_dataflow_op(NotOp, [true_wire]).unwrap();
let r = build
.add_dataflow_op(
NaryLogic::Or.with_n_inputs(2),
[true_wire, orig_not.out_wire(0)],
)
.unwrap();
let or_node = r.node();
let parent = build.dfg_node;
let reg =
ExtensionRegistry::try_new([PRELUDE.to_owned(), logic::EXTENSION.to_owned()]).unwrap();
let mut h = build.finish_hugr_with_outputs(r.outputs(), &reg).unwrap();

// we delete the original Not and create a new One. This means it will be
// traversed by `constant_fold_pass` after the Or.
let new_not = h.add_node_with_parent(parent, NodeType::new_auto(NotOp));
h.connect(true_wire.node(), true_wire.source(), new_not, 0);
h.disconnect(or_node, IncomingPort::from(1));
h.connect(new_not, 0, or_node, 1);
h.remove_node(orig_not.node());
constant_fold_pass(&mut h, &reg);
assert_fully_folded(&h, &Value::true_val())
}

#[test]
fn test_folding_pass_issue_996() {
// pseudocode:
//
// x0 := 3.0
// x1 := 4.0
// x2 := fne(x0, x1); // true
// x3 := flt(x0, x1); // true
// x4 := and(x2, x3); // true
// x5 := -10.0
// x6 := flt(x0, x5) // false
// x7 := or(x4, x6) // true
// output x7
let mut build = DFGBuilder::new(FunctionType::new(type_row![], vec![BOOL_T])).unwrap();
let x0 = build.add_load_const(Value::extension(ConstF64::new(3.0)));
let x1 = build.add_load_const(Value::extension(ConstF64::new(4.0)));
let x2 = build.add_dataflow_op(FloatOps::fne, [x0, x1]).unwrap();
let x3 = build.add_dataflow_op(FloatOps::flt, [x0, x1]).unwrap();
let x4 = build
.add_dataflow_op(
NaryLogic::And.with_n_inputs(2),
x2.outputs().chain(x3.outputs()),
)
.unwrap();
let x5 = build.add_load_const(Value::extension(ConstF64::new(-10.0)));
let x6 = build.add_dataflow_op(FloatOps::flt, [x0, x5]).unwrap();
let x7 = build
.add_dataflow_op(
NaryLogic::Or.with_n_inputs(2),
x4.outputs().chain(x6.outputs()),
)
.unwrap();
let reg = ExtensionRegistry::try_new([
PRELUDE.to_owned(),
logic::EXTENSION.to_owned(),
arithmetic::float_types::EXTENSION.to_owned(),
])
.unwrap();
let mut h = build.finish_hugr_with_outputs(x7.outputs(), &reg).unwrap();
constant_fold_pass(&mut h, &reg);
let expected = Value::true_val();
assert_fully_folded(&h, &expected);
}
}
17 changes: 16 additions & 1 deletion hugr/src/hugr/views.rs
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,10 @@ use itertools::{Itertools, MapInto};
use portgraph::render::{DotFormat, MermaidFormat};
use portgraph::{multiportgraph, LinkView, MultiPortGraph, PortView};

use super::{Hugr, HugrError, NodeMetadata, NodeMetadataMap, NodeType, DEFAULT_NODETYPE};
use super::{
Hugr, HugrError, NodeMetadata, NodeMetadataMap, NodeType, ValidationError, DEFAULT_NODETYPE,
};
use crate::extension::ExtensionRegistry;
use crate::ops::handle::NodeHandle;
use crate::ops::{OpParent, OpTag, OpTrait, OpType};

Expand Down Expand Up @@ -460,6 +463,18 @@ pub trait HugrView: sealed::HugrInternals {
self.value_types(node, Direction::Outgoing)
.map(|(p, t)| (p.as_outgoing().unwrap(), t))
}

/// Check the validity of the underlying HUGR.
fn validate(&self, reg: &ExtensionRegistry) -> Result<(), ValidationError> {
self.base_hugr().validate(reg)
}

/// Check the validity of the underlying HUGR, but don't check consistency
/// of extension requirements between connected nodes or between parents and
/// children.
fn validate_no_extensions(&self, reg: &ExtensionRegistry) -> Result<(), ValidationError> {
self.base_hugr().validate_no_extensions(reg)
}
}

/// Wraps an iterator over [Port]s that are known to be [OutgoingPort]s
Expand Down
Loading