From 6e92f207f6b9f8ad02b908c1a2256cc7832888d0 Mon Sep 17 00:00:00 2001
From: primenumber <prime@kmc.gr.jp>
Date: Thu, 9 Nov 2023 09:31:47 +0900
Subject: [PATCH] Optimize for AVX512

---
 src/engine/board.rs | 8 +++++++-
 src/main.rs         | 1 +
 2 files changed, 8 insertions(+), 1 deletion(-)

diff --git a/src/engine/board.rs b/src/engine/board.rs
index 7776d09..2289413 100644
--- a/src/engine/board.rs
+++ b/src/engine/board.rs
@@ -33,6 +33,12 @@ pub struct PlayIterator {
 
 pub const BOARD_SIZE: usize = 64;
 
+unsafe fn avx512_upper_bit(mut x: __m256i) -> __m256i {
+    let y = _mm256_sub_epi64(_mm256_set1_epi64x(63), _mm256_lzcnt_epi64(x));
+    _mm256_sllv_epi64(_mm256_set1_epi64x(1), y)
+}
+
+#[allow(dead_code)]
 unsafe fn smart_upper_bit(mut x: __m256i) -> __m256i {
     x = _mm256_or_si256(x, _mm256_srlv_epi64(x, _mm256_setr_epi64x(8, 1, 7, 9)));
     x = _mm256_or_si256(x, _mm256_srlv_epi64(x, _mm256_setr_epi64x(16, 2, 14, 18)));
@@ -95,7 +101,7 @@ impl Board {
             0x0040201008040201u64 as i64,
         );
         let mut mask = _mm256_srlv_epi64(mask1, _mm256_set1_epi64x((63 - pos) as i64));
-        let mut outflank = _mm256_and_si256(smart_upper_bit(_mm256_andnot_si256(om, mask)), p);
+        let mut outflank = _mm256_and_si256(avx512_upper_bit(_mm256_andnot_si256(om, mask)), p);
         let mut flipped = _mm256_and_si256(
             _mm256_slli_epi64(_mm256_sub_epi64(_mm256_setzero_si256(), outflank), 1),
             mask,
diff --git a/src/main.rs b/src/main.rs
index bd47684..c40f3bd 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -1,4 +1,5 @@
 #![feature(test)]
+#![feature(stdsimd)]
 mod book;
 mod engine;
 mod play;