Document day 20 and days 22 to 25

maneatingape · maneatingape · commit c1ce2c789195 · 2024-12-28T14:03:44.000Z
diff --git a/README.md b/README.md
@@ -94,7 +94,7 @@ Performance is reasonable even on older hardware, for example a 2011 MacBook Pro
 | 22 | [Monkey Market](https://adventofcode.com/2024/day/22) | [Source](src/year2024/day22.rs) | 1350 |
 | 23 | [LAN Party](https://adventofcode.com/2024/day/23) | [Source](src/year2024/day23.rs) | 43 |
 | 24 | [Crossed Wires](https://adventofcode.com/2024/day/24) | [Source](src/year2024/day24.rs) | 23 |
-| 25 | [Code Chronicle](https://adventofcode.com/2024/day/25) | [Source](src/year2024/day25.rs) | 11 |
+| 25 | [Code Chronicle](https://adventofcode.com/2024/day/25) | [Source](src/year2024/day25.rs) | 8 |
 
 ## 2023
 
diff --git a/src/year2024/day20.rs b/src/year2024/day20.rs
@@ -1,9 +1,42 @@
 //! # Race Condition
+//!
+//! Examining the input shows that there is only a single path from start to finish with
+//! no branches. This simplifies checking for shortcuts as any empty space will be on the shortest
+//! path from start to end. The cheating rules allow us to "warp" up to `n` squares away to any
+//! empty space as measured by [manhattan distance](https://en.wikipedia.org/wiki/Taxicab_geometry).
+//!
+//! For part one this is a distance of 2. When checking surrounding squares we make 2 optimizations:
+//!
+//! * Don't check any squares only 1 away as we can always just move to these normally.
+//! * Checking from point `p => q` is always the negative of `q => p`, e.g if `p = 30, q = 50` then
+//!   `p => q = 20` and `q => p = -20`. This means we only ever need to check any pair once.
+//! * Additionally we only need to check down and right. Previous rows and columns will already
+//!   have checked points above and to the left when we reach an empty square.
+//!
+//! ```none
+//!       #         .
+//!      ###       ...
+//!     ##P## =>  ....#
+//!      ###       ...
+//!       #         #
+//! ```
+//!
+//! For part two the distance is increased to 20 and the shape now resembles a wonky diamond.
+//! This shape ensures complete coverage without duplicating checks.
+//!
+//! ```none
+//!      #        .
+//!     ###      ...
+//!    ##P## => ..P##
+//!     ###      ###
+//!      #        #
+//! ```
 use crate::util::grid::*;
 use crate::util::point::*;
 use crate::util::thread::*;
 use std::sync::atomic::{AtomicU32, Ordering};
 
+/// Create a grid the same size as input with the time taken from start to any location.
 pub fn parse(input: &str) -> Grid<i32> {
     let grid = Grid::parse(input);
     let start = grid.find(b'S').unwrap();
@@ -12,13 +45,15 @@ pub fn parse(input: &str) -> Grid<i32> {
     let mut time = grid.same_size_with(i32::MAX);
     let mut elapsed = 0;
 
+    // Find starting direction, assuming start position is surrounded by 3 walls.
     let mut position = start;
     let mut direction = ORTHOGONAL.into_iter().find(|&o| grid[position + o] != b'#').unwrap();
 
     while position != end {
         time[position] = elapsed;
         elapsed += 1;
 
+        // There are no branches so we only ever need to go straight ahead or turn left or right.
         direction = [direction, direction.clockwise(), direction.counter_clockwise()]
             .into_iter()
             .find(|&d| grid[position + d] != b'#')
@@ -37,6 +72,8 @@ pub fn part1(time: &Grid<i32>) -> u32 {
         for x in 1..time.width - 1 {
             let point = Point::new(x, y);
 
+            // We only need to check 2 points to the right and down as previous empty squares
+            // have already checked up and to the left.
             if time[point] != i32::MAX {
                 cheats += check(time, point, Point::new(2, 0));
                 cheats += check(time, point, Point::new(0, 2));
@@ -47,6 +84,7 @@ pub fn part1(time: &Grid<i32>) -> u32 {
     cheats
 }
 
+/// Searches for all cheats up to distance 20, parallelizing the work over multiple threads.
 pub fn part2(time: &Grid<i32>) -> u32 {
     let mut items = Vec::with_capacity(10_000);
 
@@ -68,13 +106,7 @@ pub fn part2(time: &Grid<i32>) -> u32 {
 fn worker(time: &Grid<i32>, total: &AtomicU32, batch: Vec<Point>) {
     let mut cheats = 0;
 
-    // (p1, p2) is the reciprocal of (p2, p1) so we only need to check each pair once. Checking the
-    // wonky diamond shape on the right ensures complete coverage without duplicating checks.
-    //      #        .
-    //     ###      ...
-    //    ##### => ..###
-    //     ###      ###
-    //      #        #
+    // (p1, p2) is the reciprocal of (p2, p1) so we only need to check each pair once.
     for point in batch {
         for x in 2..21 {
             cheats += check(time, point, Point::new(x, 0));
@@ -87,9 +119,11 @@ fn worker(time: &Grid<i32>, total: &AtomicU32, batch: Vec<Point>) {
         }
     }
 
+    // Update global total.
     total.fetch_add(cheats, Ordering::Relaxed);
 }
 
+// Check if we save enough time warping to another square.
 #[inline]
 fn check(time: &Grid<i32>, first: Point, delta: Point) -> u32 {
     let second = first + delta;
diff --git a/src/year2024/day22.rs b/src/year2024/day22.rs
@@ -1,4 +1,22 @@
 //! # Monkey Market
+//!
+//! Solves both parts simultaneously, parallelizing the work over multiple threads since
+//! each secret number is independent. The process of generating the next secret number is a
+//! [linear feedback shift register](https://en.wikipedia.org/wiki/Linear-feedback_shift_register).
+//! with a cycle of 2²⁴.
+//!
+//! Interestingly this means that with some clever math it's possible to generate the `n`th number
+//! from any starting secret number with only 24 calculations. Unfortunately this doesn't help for
+//! part two since we need to check every possible price change. However to speed things up we can
+//! make several optimizations:
+//!
+//! * First the sequence of 4 prices is converted from -9..9 to a base 19 index of 0..19.
+//! * Whether a monkey has seen a sequence before and the total bananas for each sequence are
+//!   stored in an array. This is much faster than a `HashMap`. Using base 19 gives much better
+//!   cache locality needing only 130321 elements, for example compared to shifting each new cost
+//!   by 5 bits and storing in an array of 2²⁰ = 1048675 elements. Multiplication on modern
+//!   processors is cheap (and several instructions can issue at once) but random memory access
+//!   is expensive.
 use crate::util::parse::*;
 use crate::util::thread::*;
 use std::sync::Mutex;
@@ -53,9 +71,11 @@ fn worker(mutex: &Mutex<Exclusive>, batch: &[usize]) {
             number = hash(number);
             let price = number % 10;
 
+            // Compute index into the array.
             (a, b, c, d) = (b, c, d, 9 + price - previous);
             let index = 6859 * a + 361 * b + 19 * c + d;
 
+            // Only sell the first time we see a sequence.
             if seen[index] != id {
                 part_two[index] += price as u16;
                 seen[index] = id;
@@ -67,11 +87,14 @@ fn worker(mutex: &Mutex<Exclusive>, batch: &[usize]) {
         part_one += number;
     }
 
+    // Merge into global results.
     let mut exclusive = mutex.lock().unwrap();
     exclusive.part_one += part_one;
     exclusive.part_two.iter_mut().zip(part_two).for_each(|(a, b)| *a += b);
 }
 
+/// Compute next secret number using a
+/// [Xorshift LFSR](https://en.wikipedia.org/wiki/Linear-feedback_shift_register#Xorshift_LFSRs).
 fn hash(mut n: usize) -> usize {
     n = (n ^ (n << 6)) & 0xffffff;
     n = (n ^ (n >> 5)) & 0xffffff;
diff --git a/src/year2024/day23.rs b/src/year2024/day23.rs
@@ -1,8 +1,21 @@
 //! # LAN Party
+//!
+//! This is the [Clique problem](https://en.wikipedia.org/wiki/Clique_problem). For part one we
+//! find triangles (cliques of size 3) for each node by checking if there's an edge between any
+//! distinct pair of neighbouring nodes.
+//!
+//! Part two asks to find the maximum clique, for which we could use the
+//! [Bron–Kerbosch algorithm](https://en.wikipedia.org/wiki/Bron%E2%80%93Kerbosch_algorithm).
+//! However the input has a specific structure that enables a simpler approach of finding the
+//! largest *maximal* clique using a greedy algorithm. Nodes are arranged in clusters of 13 and
+//! the maximum clique is size 14. This approach will not necessarily work for any general graph,
+//! but will work for the inputs provided.
 use crate::util::hash::*;
 
 type Input = (FastMap<usize, Vec<usize>>, Vec<[bool; 676]>);
 
+/// Convert each character pair `xy` to an index from 0..676 so that we can use much faster array
+/// lookup instead of a `HashMap`.
 pub fn parse(input: &str) -> Input {
     let mut nodes = FastMap::with_capacity(1_000);
     let mut edges = vec![[false; 676]; 676];
@@ -14,9 +27,11 @@ pub fn parse(input: &str) -> Input {
         let from = to_index(&edge[..2]);
         let to = to_index(&edge[3..]);
 
+        // Graph is undirected so add edges to both nodes.
         nodes.entry(from).or_insert_with(empty).push(to);
         nodes.entry(to).or_insert_with(empty).push(from);
 
+        // https://en.wikipedia.org/wiki/Adjacency_matrix
         edges[from][to] = true;
         edges[to][from] = true;
     }
@@ -29,12 +44,14 @@ pub fn part1(input: &Input) -> usize {
     let mut seen = [false; 676];
     let mut triangles = 0;
 
+    // Only consider nodes starting with `t`.
     for n1 in 494..520 {
         if let Some(neighbours) = nodes.get(&n1) {
             seen[n1] = true;
 
             for (i, &n2) in neighbours.iter().enumerate() {
                 for &n3 in neighbours.iter().skip(i) {
+                    // Skip nodes if we've already seen them.
                     if !seen[n2] && !seen[n3] && edges[n2][n3] {
                         triangles += 1;
                     }
@@ -52,24 +69,29 @@ pub fn part2(input: &Input) -> String {
     let mut clique = Vec::new();
     let mut largest = Vec::new();
 
+    // Greedy algorithm to find *maximal* (not maximum) cliques.
     for (&n1, neighbours) in nodes {
         if !seen[n1] {
             clique.clear();
             clique.push(n1);
 
+            // Add nodes if they're connected to every node already in the clique.
             for &n2 in neighbours {
                 if clique.iter().all(|&c| edges[n2][c]) {
                     seen[n2] = true;
                     clique.push(n2);
                 }
             }
 
+            // For the specific graphs given in the input
+            // finding the largest maximal clique will work.
             if clique.len() > largest.len() {
                 largest.clone_from(&clique);
             }
         }
     }
 
+    // Convert each index back into 2 character identifiers sorted alphabetically.
     let mut result = String::new();
     largest.sort_unstable();
 
diff --git a/src/year2024/day24.rs b/src/year2024/day24.rs
@@ -1,4 +1,55 @@
 //! # Crossed Wires
+//!
+//! Part one is a straightforward simulation of the gates. Part two asks us to fix a broken
+//! [ripple carry adder](https://en.wikipedia.org/wiki/Adder_(electronics)).
+//!
+//! The structure of the adder is:
+//!
+//! * Half adder for bits `x00` and `y00`. Outputs sum to `z00` and carry to `z01`.
+//! * Full adder for bits `x01..x44` and `y01..y44`. Outputs carry to next bit in the chain
+//!   "rippling" up to final bit.
+//! * `z45` is the carry output from `x44` and `y44`.
+//!
+//! Implemented in logic gates this looks like:
+//!
+//! ```none
+//!    Half Adder     Full Adder
+//!    ┌───┐ ┌───┐    ┌───┐ ┌───┐
+//!    |x00| |y00|    |x01| |y01|
+//!    └───┘ └───┘    └───┘ └───┘
+//!     | | ┌─┘ |      | | ┌─┘ |
+//!     | └───┐ |      | └───┐ |
+//!     | ┌-┘ | |      | ┌-┘ | |
+//!    ┌───┐ ┌───┐    ┌───┐ ┌───┐
+//!    |XOR| |AND|    |XOR| |AND|
+//!    └───┘ └───┘    └───┘ └───┘
+//!      |     |    ┌───┴┐     |
+//!      |     └──┬────┐ |     |
+//!      |   Carry| | ┌───┐    |
+//!      |    out | | |AND|    |
+//!      |        | | └───┘    |
+//!      |        | |   └────┐ |
+//!      |        | └────┐   | |
+//!      |        └────┐ |   | |
+//!      |            ┌───┐ ┌───┐
+//!      |            |XOR| |OR |                                  Carry
+//!      |            └───┘ └───┘                                   out
+//!      |              |     |                                      |
+//!    ┌───┐          ┌───┐   |                                    ┌───┐
+//!    |z00|          |z01| Carry    ...repeat for z01 to z44...   |z45|
+//!    └───┘          └───┘  out                                   └───┘
+//! ```
+//!
+//! Then we can deduce some rules for the output of each gate type:
+//!
+//! 1. **XOR** If inputs are `x` and `y` then output must be another XOR gate
+//!    (except for inputs `x00` and `y00`) otherwise output must be `z`.
+//! 2. **AND** Output must be an OR gate (except for inputs `x00` and `y00`).
+//! 3. **OR** Output must be both AND and XOR gate, except for final carry
+//!    which must output to `z45`.
+//!
+//! We only need to find swapped outputs (not fix them) so the result is the labels of gates
+//! that breaks the rules in alphabetical order.
 use crate::util::hash::*;
 use crate::util::iter::*;
 use crate::util::parse::*;
@@ -15,21 +66,27 @@ pub fn parse(input: &str) -> Input<'_> {
 pub fn part1(input: &Input<'_>) -> u64 {
     let (prefix, gates) = input;
 
+    // Using an array to store already computed values is much faster than a `HashMap`.
     let mut todo: VecDeque<_> = gates.iter().copied().collect();
     let mut cache = vec![u8::MAX; 1 << 15];
     let mut result = 0;
 
+    // Convert each character to a 5 bit number from 0..31
+    // then each group of 3 to a 15 bit index from 0..32768.
     let to_index = |s: &str| {
         let b = s.as_bytes();
         ((b[0] as usize & 31) << 10) + ((b[1] as usize & 31) << 5) + (b[2] as usize & 31)
     };
 
+    // Add input signals to cache.
     for line in prefix.lines() {
         let prefix = &line[..3];
         let suffix = &line[5..];
         cache[to_index(prefix)] = suffix.unsigned();
     }
 
+    // If both inputs are available then add gate output to cache
+    // otherwise push back to end of queue for reprocessing later.
     while let Some(gate @ [left, kind, right, _, to]) = todo.pop_front() {
         let left = cache[to_index(left)];
         let right = cache[to_index(right)];
@@ -46,7 +103,8 @@ pub fn part1(input: &Input<'_>) -> u64 {
         }
     }
 
-    for i in (to_index("z00")..to_index("z64")).rev() {
+    // Output 46 bit result.
+    for i in (to_index("z00")..to_index("z46")).rev() {
         if cache[i] != u8::MAX {
             result = (result << 1) | (cache[i] as u64);
         }
@@ -58,18 +116,19 @@ pub fn part1(input: &Input<'_>) -> u64 {
 pub fn part2(input: &Input<'_>) -> String {
     let (_, gates) = input;
 
-    let mut lookup = FastSet::new();
+    let mut output = FastSet::new();
     let mut swapped = FastSet::new();
 
+    // Track the kind of gate that each wire label outputs to.
     for &[left, kind, right, _, _] in gates {
-        lookup.insert((left, kind));
-        lookup.insert((right, kind));
+        output.insert((left, kind));
+        output.insert((right, kind));
     }
 
     for &[left, kind, right, _, to] in gates {
         if kind == "AND" {
             // Check that all AND gates point to an OR, except for first AND.
-            if left != "x00" && right != "x00" && !lookup.contains(&(to, "OR")) {
+            if left != "x00" && right != "x00" && !output.contains(&(to, "OR")) {
                 swapped.insert(to);
             }
         }
@@ -80,15 +139,15 @@ pub fn part2(input: &Input<'_>) -> String {
                 swapped.insert(to);
             }
             // OR can never point to OR.
-            if lookup.contains(&(to, "OR")) {
+            if output.contains(&(to, "OR")) {
                 swapped.insert(to);
             }
         }
 
         if kind == "XOR" {
             if left.starts_with('x') || right.starts_with('x') {
                 // Check that first level XOR points to second level XOR, except for first XOR.
-                if left != "x00" && right != "x00" && !lookup.contains(&(to, "XOR")) {
+                if left != "x00" && right != "x00" && !output.contains(&(to, "XOR")) {
                     swapped.insert(to);
                 }
             } else {
diff --git a/src/year2024/day25.rs b/src/year2024/day25.rs
