Solution 4.04 (Speed: 1.45)

sol/matrixsol.hpp

@@ -80,7 +80,7 @@ public:
 		assert(a.HS == b.VS);
 
 		for (size_t j = 0; j < HS; ++j) {
-						// This is useless - compiles vectorizes normal loop here
+			// This is useless - compiles vectorizes normal loop here
 			const size_t B_COPY_SIZE = (a.HS + VEC_CNT - 1) / VEC_CNT * VEC_CNT;
 			matrix_element b_copy[B_COPY_SIZE];
 			for (size_t i = 0; i < B_COPY_SIZE / VEC_CNT; ++i) {
@@ -101,11 +101,8 @@ public:
 					res_vec	      = _mm_min_epu16(res_vec, val);
 				}
 
-				_mm_store_si128((__m128i*)aggr_arr, res_vec);
-				matrix_element res = res_init_val;
-				for (size_t k = 0; k < VEC_CNT; ++k) res = std::min(res, aggr_arr[k]);
-
-				addr(i, j) = res;
+				__m128i mn = _mm_minpos_epu16(res_vec);
+				addr(i, j) = _mm_extract_epi16(mn, 0);
 			}
 		}
 	}