From 8053532039e11d2cd1909df105873cca55f6f711 Mon Sep 17 00:00:00 2001
From: hal8174 <henning.lindemann@pm.me>
Date: Sun, 6 Jul 2025 12:46:48 +0200
Subject: [PATCH] More performance improvements

---
 src/main.rs | 33 ++++++++++++++++++---------------
 1 file changed, 18 insertions(+), 15 deletions(-)

diff --git a/src/main.rs b/src/main.rs
index 3b8ea61..ecba62f 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -121,7 +121,7 @@ fn read_traces(path: impl AsRef<Path>, capacity: usize) -> Vec<Vec<u8>> {
     r
 }
 
-const LANES: usize = 16;
+const LANES: usize = 32;
 
 fn correlation(
     bit: usize,
@@ -130,31 +130,33 @@ fn correlation(
     cyphtertext: &[[Vec<u8>; 16]],
     traces: &[Vec<u8>],
 ) -> f64 {
-    let mut x = Simd::<u64, LANES>::default();
-    let mut y = Simd::<u64, LANES>::default();
-    let mut xy = Simd::<u64, LANES>::default();
-    let mut xsqr = Simd::<u64, LANES>::default();
-    let mut ysqr = Simd::<u64, LANES>::default();
+    assert!(
+        traces[0].len() < (1 << 16) * LANES,
+        "Integers might overflow"
+    );
+
+    let mut x = Simd::<u32, LANES>::default();
+    let mut y = Simd::<u32, LANES>::default();
+    let mut xy = Simd::<u32, LANES>::default();
+    let mut ysqr = Simd::<u32, LANES>::default();
 
     let mask = Simd::<u8, LANES>::splat(1 << (bit % 8));
 
     for i in 0..traces[0].len() / LANES {
         let xi = Simd::<u8, LANES>::from_slice(&cyphtertext[key_hypothesis][bit / 8][i * LANES..]);
-        let xi = (xi & mask).cast::<u64>();
-        let yi = Simd::<u8, LANES>::from_slice(&traces[trace_index][i * LANES..]).cast::<u64>();
+        let xi = (xi & mask).cast();
+        let yi = Simd::<u8, LANES>::from_slice(&traces[trace_index][i * LANES..]).cast();
 
         x += xi;
         y += yi;
         xy += xi * yi;
-        xsqr += xi * xi;
         ysqr += yi * yi;
     }
 
-    let mut x = x.reduce_sum() as i64;
-    let mut y = y.reduce_sum() as i64;
-    let mut xy = xy.reduce_sum() as i64;
-    let mut xsqr = xsqr.reduce_sum() as i64;
-    let mut ysqr = ysqr.reduce_sum() as i64;
+    let mut x = x.cast::<u64>().reduce_sum() as i64;
+    let mut y = y.cast::<u64>().reduce_sum() as i64;
+    let mut xy = xy.cast::<u64>().reduce_sum() as i64;
+    let mut ysqr = ysqr.cast::<u64>().reduce_sum() as i64;
 
     for i in (traces[0].len() / LANES) * LANES..traces[0].len() {
         let xi = (cyphtertext[key_hypothesis][bit / 8][i] & (1 << (bit % 8))) as i64;
@@ -163,10 +165,11 @@ fn correlation(
         x += xi;
         y += yi;
         xy += xi * yi;
-        xsqr += xi * xi;
         ysqr += yi * yi;
     }
 
+    let xsqr = x * (1 << (bit % 8));
+
     let n = traces[0].len() as i64;
     let num = (n * xy - x * y) as f64;
     let denom = f64::sqrt((n * xsqr - x * x) as f64) * f64::sqrt((n * ysqr - y * y) as f64);