day21

benches/bench.rs

@@ -36,17 +36,17 @@ macro_rules! benches {
 
 #[rustfmt::skip]
 benches!(
-    9, "9.txt",
-    10, "10.txt",
-    11, "11.txt",
-    12, "12.txt",
-    13, "13.txt", "13_simd", "13.txt",
-    14, "14.txt",
-    15, "15.txt",
-    16, "16.txt",
-    17, "17.txt",
-    18, "18.txt",
-    19, "19.txt",
-    20, "20.txt",
+    // 9, "9.txt",
+    // 10, "10.txt",
+    // 11, "11.txt",
+    // 12, "12.txt",
+    // 13, "13.txt", "13_simd", "13.txt",
+    // 14, "14.txt",
+    // 15, "15.txt",
+    // 16, "16.txt",
+    // 17, "17.txt",
+    // 18, "18.txt",
+    // 19, "19.txt",
+    // 20, "20.txt",
     21, "21.txt"
 );

src/day21.rs

@@ -1,45 +1,29 @@
 use std::mem::transmute;
-use std::simd::prelude::*;
 
 const LUT1: [u64; 1000] = unsafe { transmute(*include_bytes!("../luts/day21_1.bin")) };
 const LUT2: [u64; 1000] = unsafe { transmute(*include_bytes!("../luts/day21_2.bin")) };
 
-#[cfg_attr(
-    target_arch = "x86_64",
-    target_feature(enable = "popcnt,avx2,ssse3,bmi1,bmi2,lzcnt")
-)]
-#[cfg_attr(avx512_available, target_feature(enable = "avx512vl"))]
-unsafe fn solve(input: &str, lut: &[u64]) -> u64 {
-    let input = input.as_bytes();
-
-    let u1 = usizex8::splat(1);
-    let u10 = usizex8::splat(10);
-    let u100 = usizex8::splat(100);
-    let u5328 = usizex8::splat(5328); // 100 * 48 + 10 * 48 + 48
-
-    let first_buttons_idxs = usizex8::from_array([0, 5, 10, 15, 20, 25, 30, 35]);
-    let second_buttons_idxs = first_buttons_idxs + u1;
-    let third_buttons_idxs = second_buttons_idxs + u1;
-
-    let first_buttons = u8x8::gather_or_default(&input, first_buttons_idxs);
-    let second_buttons = u8x8::gather_or_default(&input, second_buttons_idxs);
-    let third_buttons = u8x8::gather_or_default(&input, third_buttons_idxs);
-
-    let first_buttons = first_buttons.cast();
-    let second_buttons = second_buttons.cast();
-    let third_buttons = third_buttons.cast();
-
-    let num = first_buttons * u100 + second_buttons * u10 + third_buttons - u5328;
+unsafe fn solve(input: &[u8], lut: &[u64; 1000]) -> u64 {
+    macro_rules! solve {
+        ($a:expr, $b:expr, $c:expr) => {
+            lut.get_unchecked(
+                (*input.get_unchecked($a) as usize) * 100
+                    + (*input.get_unchecked($b) as usize) * 10
+                    + (*input.get_unchecked($c) as usize)
+                    - 5328,
+            )
+        };
+    }
 
-    u64x8::gather_or_default(lut, num).reduce_sum()
+    solve!(0, 1, 2) + solve!(5, 6, 7) + solve!(10, 11, 12) + solve!(15, 16, 17) + solve!(20, 21, 22)
 }
 
 pub fn part1(input: &str) -> u64 {
-    unsafe { solve(input, &LUT1) }
+    unsafe { solve(input.as_bytes(), &LUT1) }
 }
 
 pub fn part2(input: &str) -> u64 {
-    unsafe { solve(input, &LUT2) }
+    unsafe { solve(input.as_bytes(), &LUT2) }
 }
 
 #[cfg(test)]

src/lib.rs

@@ -1,25 +1,25 @@
 #![feature(portable_simd)]
 
-pub mod day1;
-pub mod day2;
-pub mod day3;
-pub mod day4;
-pub mod day5;
-pub mod day6;
-pub mod day7;
-pub mod day8;
-pub mod day9;
+// pub mod day1;
+// pub mod day2;
+// pub mod day3;
+// pub mod day4;
+// pub mod day5;
+// pub mod day6;
+// pub mod day7;
+// pub mod day8;
+// pub mod day9;
 
-pub mod day10;
-pub mod day11;
-pub mod day12;
-pub mod day13;
-pub mod day13_simd;
-pub mod day14;
-pub mod day15;
-pub mod day16;
-pub mod day17;
-pub mod day18;
-pub mod day19;
-pub mod day20;
+// pub mod day10;
+// pub mod day11;
+// pub mod day12;
+// pub mod day13;
+// pub mod day13_simd;
+// pub mod day14;
+// pub mod day15;
+// pub mod day16;
+// pub mod day17;
+// pub mod day18;
+// pub mod day19;
+// pub mod day20;
 pub mod day21;