reverted the last 'optimization'

sol/levensol.hpp

@@ -201,21 +201,21 @@ if constexpr (std::is_same<policy, policy_avx512>::value) {
 				cycles = std::min(a_size - x, b_size - y) / multiplier;
 
 				for (; cycles != 0; x += multiplier, y += multiplier, d += multiplier, --cycles) {
-					auto tmp = _mm512_maskz_set1_epi32(
-						_mm512_cmpeq_epu32_mask(
-							_mm512_loadu_si512(a_v + x),
-							_mm512_loadu_si512(b_v + y)),
-						-1);
+					const auto mask = _mm512_cmpeq_epu32_mask(
+						_mm512_loadu_si512(a_v + x),
+						_mm512_loadu_si512(b_v + y));
 
 					if (odd) {
-						tmp = _mm512_add_epi32(tmp, *(data_pointer)(odd_v + d));
+						data_type tmp = *(data_pointer)(odd_v + d);
+						tmp = _mm512_mask_sub_epi32(tmp, mask, tmp, _mm512_set1_epi32(1));
 						tmp = _mm512_min_epu32(*(data_pointer)(even_v + d), tmp);
 						const data_type tmp2 = _mm512_loadu_si512(even_v + d + 1);
 						tmp = _mm512_min_epu32(tmp, tmp2);
 
 						*(data_pointer)(odd_v + d) = _mm512_add_epi32(tmp, _mm512_set1_epi32(1));
 					} else {
-						tmp = _mm512_add_epi32(tmp, *(data_pointer)(even_v + d));
+						data_type tmp = *(data_pointer)(even_v + d);
+						tmp = _mm512_mask_sub_epi32(tmp, mask, tmp, _mm512_set1_epi32(1));
 						tmp = _mm512_min_epu32(*(data_pointer)(odd_v + d), tmp);
 						const data_type tmp2 = _mm512_loadu_si512(odd_v + d - 1);
 						tmp = _mm512_min_epu32(tmp, tmp2);