# SPDX-FileCopyrightText: 2026 Edward Kmett <ekmett@gmail.com>
# SPDX-License-Identifier: BSD-2-Clause OR Apache-2.0
# Benchmark-only SAD-first prototype, based on aeaaa9d9896db9aec1a003a0a5ba7b0b6174bd06.
# Result SHA-256: ed125be1898573ec2b926df930c96b909921fc506536d9c2e1c8b0363e0fb686
--- qword_rank15.h
+++ sad_rank15.h
@@ -80 +80,4 @@
-      return result + sum_bytes(pairs);
+      // Widen before reducing: four 16-bit lanes and their total (at most
+      // 1920) fit without carries into the product's high 16 bits.
+      pairs = (pairs & 0x00ff00ff00ff00ffull) + ((pairs >> 8) & 0x00ff00ff00ff00ffull);
+      return result + ((pairs * 0x0001000100010001ull) >> 48);
@@ -101,3 +104,3 @@
-      // Reduce eight qwords first: byte lanes total at most 240, so they
-      // cannot carry into each other. Exactly 64 bytes are readable.
-      return sum_bytes(std::uint64_t(_mm512_reduce_add_epi64(pairs)));
+      // Each byte is at most 30. SAD widens groups of eight bytes before the
+      // qword reduction, whose maximum is 1920. Exactly 64 bytes are readable.
+      return unsigned(_mm512_reduce_add_epi64(_mm512_sad_epu8(pairs, _mm512_setzero_si512())));
@@ -123,3 +126,3 @@
-      // Two vectors contribute at most 60 per byte. Reduce four qwords first;
-      // byte lanes total at most 240. Exactly 64 bytes are readable.
-      auto totals = _mm256_add_epi8(a, b);
+      // Two vectors contribute at most 60 per byte. Widen before reducing the
+      // total, which fits in 32 bits (at most 1920). Exactly 64 bytes are readable.
+      auto totals = _mm256_sad_epu8(_mm256_add_epi8(a, b), _mm256_setzero_si256());
@@ -127,3 +130 @@
-      std::uint64_t value;
-      _mm_storel_epi64(reinterpret_cast<__m128i *>(&value), _mm_add_epi64(halves, _mm_srli_si128(halves, 8)));
-      return sum_bytes(value);
+      return unsigned(_mm_cvtsi128_si32(_mm_add_epi64(halves, _mm_srli_si128(halves, 8))));
@@ -160,7 +160,0 @@
-    static unsigned sum_bytes(std::uint64_t value) noexcept {
-      // Byte lanes are at most 240. Widen before the horizontal sum: four
-      // 16-bit lanes and their total (at most 1920) fit without carries.
-      value = (value & 0x00ff00ff00ff00ffull) + ((value >> 8) & 0x00ff00ff00ff00ffull);
-      return unsigned((value * 0x0001000100010001ull) >> 48);
-    }
-