feat(ir): actually get toposort working

This commit is contained in:
multisn8 2024-01-19 02:59:15 +01:00
parent bf51aec1e4
commit e1f5d11ab8
Signed by: multisamplednight
GPG key ID: C81EF9B053977241

View file

@ -1,6 +1,5 @@
use std::{collections::BTreeSet, iter, ops::RangeInclusive}; use std::{num::NonZeroUsize, ops::RangeInclusive};
use either::Either;
use instruction::SocketCount; use instruction::SocketCount;
use serde::{Deserialize, Serialize}; use serde::{Deserialize, Serialize};
@ -9,7 +8,7 @@ pub mod instruction;
pub mod semi_human; pub mod semi_human;
pub type Map<K, V> = std::collections::BTreeMap<K, V>; pub type Map<K, V> = std::collections::BTreeMap<K, V>;
pub type Set<V> = std::collections::BTreeSet<V>; pub type Set<T> = std::collections::BTreeSet<T>;
/// Gives you a super well typed graph IR for a given human-readable repr. /// Gives you a super well typed graph IR for a given human-readable repr.
/// ///
@ -41,8 +40,8 @@ pub fn from_ron(source: &str) -> ron::error::SpannedResult<GraphIr> {
/// to come back to an already visited node. /// to come back to an already visited node.
/// ///
/// Here, if an edge points from _A_ to _B_ (`A --> B`), /// Here, if an edge points from _A_ to _B_ (`A --> B`),
/// then _A_ is called a **dependency** of _B_, /// then _A_ is called a **dependency** or an **input source** of _B_,
/// and _B_ is called a **dependent** of _A_. /// and _B_ is called a **dependent** or an **output target** of _A_.
/// ///
/// The DAG also enables another neat operation: /// The DAG also enables another neat operation:
/// [Topological sorting](https://en.wikipedia.org/wiki/Topological_sorting). /// [Topological sorting](https://en.wikipedia.org/wiki/Topological_sorting).
@ -69,33 +68,8 @@ pub struct GraphIr {
rev_edges: Map<id::Input, id::Output>, rev_edges: Map<id::Input, id::Output>,
} }
// TODO: this impl block, but actually the whole module, screams for tests
impl GraphIr { impl GraphIr {
/// Look "forwards" in the graph to see what other instructions this instruction feeds into.
///
/// The output slots represent the top-level iterator,
/// and each one's connections are emitted one level below.
///
/// Just [`Iterator::flatten`] if you are not interested in the slots.
///
/// The same caveats as for [`GraphIr::resolve`] apply.
#[must_use]
pub fn dependents(
&self,
subject: &id::Instruction,
) -> Option<impl Iterator<Item = impl Iterator<Item = &id::Instruction>> + '_> {
let (subject, kind) = self.instructions.get_key_value(subject)?;
let SocketCount { inputs, .. } = kind.socket_count();
Some((0..inputs).map(|idx| {
let output = id::Output(socket(subject, idx));
self.edges
.get(&output)
.map_or(Either::Right(iter::empty()), |targets| {
Either::Left(targets.iter().map(|input| &input.socket().belongs_to))
})
}))
}
/// Look "backwards" in the graph, /// Look "backwards" in the graph,
/// and find out what instructions need to be done before this one. /// and find out what instructions need to be done before this one.
/// The input slots are visited in order. /// The input slots are visited in order.
@ -105,22 +79,41 @@ impl GraphIr {
/// ///
/// The same caveats as for [`GraphIr::resolve`] apply. /// The same caveats as for [`GraphIr::resolve`] apply.
#[must_use] #[must_use]
pub fn dependencies( pub fn input_sources(
&self, &self,
subject: &id::Instruction, subject: &id::Instruction,
) -> Option<impl Iterator<Item = Option<&id::Instruction>> + '_> { ) -> Option<impl Iterator<Item = Option<&id::Output>> + '_> {
let (subject, kind) = self.instructions.get_key_value(subject)?; let (subject, kind) = self.instructions.get_key_value(subject)?;
let SocketCount { inputs, .. } = kind.socket_count(); let SocketCount { inputs, .. } = kind.socket_count();
Some((0..inputs).map(|idx| { Some((0..inputs).map(|idx| {
let input = id::Input(socket(subject, idx)); let input = id::Input(socket(subject, idx));
self.rev_edges self.rev_edges.get(&input)
.get(&input) }))
.map(|output| &output.socket().belongs_to) }
/// Look "forwards" in the graph to see what other instructions this instruction feeds into.
///
/// The output slots represent the top-level iterator,
/// and each one's connections are emitted one level below.
///
/// Just [`Iterator::flatten`] if you are not interested in the slots.
///
/// The same caveats as for [`GraphIr::resolve`] apply.
#[must_use]
pub fn output_targets(
&self,
subject: &id::Instruction,
) -> Option<impl Iterator<Item = Option<&Set<id::Input>>> + '_> {
let (subject, kind) = self.instructions.get_key_value(subject)?;
let SocketCount { outputs, .. } = kind.socket_count();
Some((0..outputs).map(|idx| {
let output = id::Output(socket(subject, idx));
self.edges.get(&output)
})) }))
} }
// TODO: this function, but actually the whole module, screams for tests
/// Returns the instruction corresponding to the given ID. /// Returns the instruction corresponding to the given ID.
/// Returns [`None`] if there is no such instruction in this graph IR. /// Returns [`None`] if there is no such instruction in this graph IR.
/// ///
@ -133,33 +126,14 @@ impl GraphIr {
pub fn resolve<'ir>(&'ir self, subject: &id::Instruction) -> Option<Instruction<'ir>> { pub fn resolve<'ir>(&'ir self, subject: &id::Instruction) -> Option<Instruction<'ir>> {
let (id, kind) = self.instructions.get_key_value(subject)?; let (id, kind) = self.instructions.get_key_value(subject)?;
// just try each slot and see if it's connected let input_sources = self.input_sources(subject)?.collect();
// very crude, but it works for a proof of concept let output_targets = self.output_targets(subject)?.collect();
let SocketCount { inputs, outputs } = kind.socket_count();
let socket = |id: &id::Instruction, idx| id::Socket {
belongs_to: id.clone(),
// impossible since the length is limited to a u16 already
#[allow(clippy::cast_possible_truncation)]
idx: id::SocketIdx(idx as u16),
};
let mut inputs_from = vec![None; inputs.into()];
for (idx, slot) in inputs_from.iter_mut().enumerate() {
let input = id::Input(socket(id, idx));
*slot = self.rev_edges.get(&input);
}
let mut outputs_to = vec![None; outputs.into()];
for (idx, slot) in outputs_to.iter_mut().enumerate() {
let output = id::Output(socket(id, idx));
*slot = self.edges.get(&output);
}
Some(Instruction { Some(Instruction {
id, id,
kind, kind,
inputs_from, input_sources,
outputs_to, output_targets,
}) })
} }
@ -187,15 +161,18 @@ impl GraphIr {
/// ///
/// Panics if there are any cycles in the IR, as it needs to be a DAG. /// Panics if there are any cycles in the IR, as it needs to be a DAG.
#[must_use] #[must_use]
// yes, this function could actually return an iterator and be lazy // yes, this function could probably return an iterator and be lazy
// no, not today // no, not today
pub fn topological_sort(&self) -> Vec<Instruction> { pub fn topological_sort(&self) -> Vec<Instruction> {
// count how many incoming edges each vertex has // count how many incoming edges each vertex has
let nonzero_input_counts: Map<_, usize> = let mut nonzero_input_counts: Map<_, NonZeroUsize> =
self.rev_edges self.rev_edges
.iter() .iter()
.fold(Map::new(), |mut count, (input, _)| { .fold(Map::new(), |mut count, (input, _)| {
*count.entry(input.socket().belongs_to.clone()).or_default() += 1; let _ = *count
.entry(input.socket().belongs_to.clone())
.and_modify(|count| *count = count.saturating_add(1))
.or_insert(NonZeroUsize::MIN);
count count
}); });
@ -204,32 +181,52 @@ impl GraphIr {
let no_inputs: Vec<_> = { let no_inputs: Vec<_> = {
let nonzero: Set<_> = nonzero_input_counts.keys().collect(); let nonzero: Set<_> = nonzero_input_counts.keys().collect();
let all: Set<_> = self.instructions.keys().collect(); let all: Set<_> = self.instructions.keys().collect();
all.difference(&nonzero).copied().collect() all.difference(&nonzero).copied().cloned().collect()
}; };
let mut active_queue = no_inputs;
// then let's find the order! // then let's find the order!
let mut order = Vec::new(); let mut order = Vec::new();
let mut active_queue = no_inputs;
while let Some(current) = active_queue.pop() { while let Some(current) = active_queue.pop() {
// now that this vertex is visited and resolved, // now that this vertex is visited and resolved,
// make sure all dependents notice that // make sure all dependents notice that
for dependent in self let dependents = self
.dependents(current) .output_targets(&current)
.expect("graph to be consistent") .expect("graph to be consistent")
.flatten() .flatten()
{ .flatten();
dbg!(dependent);
for dependent_input in dependents {
let dependent = &dependent_input.socket().belongs_to;
// how many inputs are connected to this dependent without us?
let count = nonzero_input_counts
.get_mut(dependent)
.expect("connected output must refer to non-zero input");
let new = NonZeroUsize::new(count.get() - 1);
if let Some(new) = new {
// aww, still some
*count = new;
continue;
}
// none, that means this one is free now! let's throw it onto the active queue then
let (now_active, _) = nonzero_input_counts
.remove_entry(dependent)
.expect("connected output must refer to non-zero input");
active_queue.push(now_active);
} }
// TODO: check if this instruction is "well-fed", that is, has all the inputs it needs, // TODO: check if this instruction is "well-fed", that is, has all the inputs it needs,
// and if not, panic // and if not, panic
order.push(self.resolve(current).expect("graph to be consistent")); order.push(self.resolve(&current).expect("graph to be consistent"));
} }
assert!( assert!(
!nonzero_input_counts.is_empty(), nonzero_input_counts.is_empty(),
concat!( concat!(
"topological sort didn't cover all instructions\n", "topological sort didn't cover all instructions\n",
"either there are unconnected inputs, or there is a cycle\n", "either there are unconnected inputs, or there is a cycle\n",
@ -250,8 +247,8 @@ pub struct Instruction<'ir> {
pub kind: &'ir instruction::Kind, pub kind: &'ir instruction::Kind,
// can't have these two public since then a user might corrupt their length // can't have these two public since then a user might corrupt their length
inputs_from: Vec<Option<&'ir id::Output>>, input_sources: Vec<Option<&'ir id::Output>>,
outputs_to: Vec<Option<&'ir BTreeSet<id::Input>>>, output_targets: Vec<Option<&'ir Set<id::Input>>>,
} }
impl<'ir> Instruction<'ir> { impl<'ir> Instruction<'ir> {
@ -260,14 +257,14 @@ impl<'ir> Instruction<'ir> {
/// [`None`] means that this input is unfilled, /// [`None`] means that this input is unfilled,
/// and must be filled before the instruction can be ran. /// and must be filled before the instruction can be ran.
#[must_use] #[must_use]
pub fn inputs_from(&self) -> &[Option<&'ir id::Output>] { pub fn input_sources(&self) -> &[Option<&'ir id::Output>] {
&self.inputs_from &self.input_sources
} }
/// To whom outputs are sent. [`None`] means that this output is unused. /// To whom outputs are sent.
#[must_use] #[must_use]
pub fn outputs_to(&self) -> &[Option<&'ir BTreeSet<id::Input>>] { pub fn output_targets(&self) -> &[Option<&'ir Set<id::Input>>] {
&self.outputs_to &self.output_targets
} }
} }