chore: increase number of cmp

FuelLabs · Jan 1, 2025 · 1e689b3 · 1e689b3
1 parent f14a67e
commit 1e689b3
Showing 1 changed file with 50 additions and 5 deletions.
diff --git a/fuel-vm/src/interpreter/memory.rs b/fuel-vm/src/interpreter/memory.rs
@@ -1104,7 +1104,7 @@ fn slices_equal_avx2(a: &[u8], b: &[u8]) -> bool {
         return false;
     }
 
-    const CHUNK: usize = 32;
+    const CHUNK: usize = 256;
 
     // if the slices are small, we don't need to
     // use SIMD instructions due to overhead
@@ -1117,13 +1117,58 @@ fn slices_equal_avx2(a: &[u8], b: &[u8]) -> bool {
     let mut i = 0;
 
     unsafe {
+        let mut mismatch = false;
+
         while i + CHUNK <= len {
-            let simd_a = _mm256_loadu_si256(a.as_ptr().add(i) as *const _);
-            let simd_b = _mm256_loadu_si256(b.as_ptr().add(i) as *const _);
+            let simd_a1 = _mm256_loadu_si256(a.as_ptr().add(i) as *const _);
+            let simd_b1 = _mm256_loadu_si256(b.as_ptr().add(i) as *const _);
+
+            let cmp = _mm256_cmpeq_epi8(simd_a1, simd_b1);
+            mismatch |= _mm256_movemask_epi8(cmp) != -1;
+
+            let simd_a2 = _mm256_loadu_si256(a.as_ptr().add(i + 32) as *const _);
+            let simd_b2 = _mm256_loadu_si256(b.as_ptr().add(i + 32) as *const _);
+
+            let cmp2 = _mm256_cmpeq_epi8(simd_a2, simd_b2);
+            mismatch |= _mm256_movemask_epi8(cmp2) != -1;
+
+            let simd_a3 = _mm256_loadu_si256(a.as_ptr().add(i + 64) as *const _);
+            let simd_b3 = _mm256_loadu_si256(b.as_ptr().add(i + 64) as *const _);
+
+            let cmp3 = _mm256_cmpeq_epi8(simd_a3, simd_b3);
+            mismatch |= _mm256_movemask_epi8(cmp3) != -1;
+
+            let simd_a4 = _mm256_loadu_si256(a.as_ptr().add(i + 96) as *const _);
+            let simd_b4 = _mm256_loadu_si256(b.as_ptr().add(i + 96) as *const _);
+
+            let cmp4 = _mm256_cmpeq_epi8(simd_a4, simd_b4);
+            mismatch |= _mm256_movemask_epi8(cmp4) != -1;
+
+            let simd_a5 = _mm256_loadu_si256(a.as_ptr().add(i + 128) as *const _);
+            let simd_b5 = _mm256_loadu_si256(b.as_ptr().add(i + 128) as *const _);
+
+            let cmp5 = _mm256_cmpeq_epi8(simd_a5, simd_b5);
+            mismatch |= _mm256_movemask_epi8(cmp5) != -1;
+
+            let simd_a6 = _mm256_loadu_si256(a.as_ptr().add(i + 160) as *const _);
+            let simd_b6 = _mm256_loadu_si256(b.as_ptr().add(i + 160) as *const _);
+
+            let cmp6 = _mm256_cmpeq_epi8(simd_a6, simd_b6);
+            mismatch |= _mm256_movemask_epi8(cmp6) != -1;
+
+            let simd_a7 = _mm256_loadu_si256(a.as_ptr().add(i + 192) as *const _);
+            let simd_b7 = _mm256_loadu_si256(b.as_ptr().add(i + 192) as *const _);
+
+            let cmp7 = _mm256_cmpeq_epi8(simd_a7, simd_b7);
+            mismatch |= _mm256_movemask_epi8(cmp7) != -1;
+
+            let simd_a8 = _mm256_loadu_si256(a.as_ptr().add(i + 224) as *const _);
+            let simd_b8 = _mm256_loadu_si256(b.as_ptr().add(i + 224) as *const _);
 
-            let cmp = _mm256_cmpeq_epi8(simd_a, simd_b);
+            let cmp8 = _mm256_cmpeq_epi8(simd_a8, simd_b8);
+            mismatch |= _mm256_movemask_epi8(cmp8) != -1;
 
-            if _mm256_movemask_epi8(cmp) != -1 {
+            if mismatch {
                 return false;
             }