Solution 4.06 (Speed: 1.30)

sol/matrixsol.hpp

@@ -68,13 +68,8 @@ private:
 	}
 };
 
-#define gen_switch_label(res, mn, pos)                                      \
-	case pos:                                                           \
-		res = _mm_insert_epi16(res, _mm_extract_epi16(mn, 0), pos); \
-		break;
-
 #define gen_asgn(pos, wpos) \
-	writes##pos = _mm_insert_epi16(writes##pos, _mm_extract_epi16(row, pos), wpos)
+	row##pos = _mm_insert_epi16(row##pos, _mm_extract_epi16(row, pos), wpos)
 
 #define gen_for(wpos)                                                                   \
 	row = _mm_load_si128((__m128i*)&b.addr(i * VEC_CNT + wpos, j_block * VEC_CNT)); \
@@ -98,12 +93,14 @@ public:
 		assert(b.HS == HS);
 		assert(a.HS == b.VS);
 
+		const __m128i MASK = _mm_setr_epi16(0xffff, 0, 0, 0, 0, 0, 0, 0);
+
 		const size_t J_BLOCK_LIMIT = (HS + VEC_CNT - 1) / VEC_CNT;
 		for (size_t j_block = 0; j_block < J_BLOCK_LIMIT; ++j_block) {
 			const size_t B_COPY_SIZE = (a.HS + VEC_CNT - 1) / VEC_CNT * VEC_CNT;
 			matrix_element b_copy[VEC_CNT][B_COPY_SIZE];
 
-			__m128i writes0, writes1, writes2, writes3, writes4, writes5, writes6, writes7;
+			__m128i row0, row1, row2, row3, row4, row5, row6, row7;
 			for (size_t i = 0; i < B_COPY_SIZE / VEC_CNT; ++i) {
 				__m128i row;
 				gen_for(0);
@@ -115,34 +112,22 @@ public:
 				gen_for(6);
 				gen_for(7);
 
-				_mm_store_si128((__m128i*)&b_copy[0][i * VEC_CNT], writes0);
-				_mm_store_si128((__m128i*)&b_copy[1][i * VEC_CNT], writes1);
-				_mm_store_si128((__m128i*)&b_copy[2][i * VEC_CNT], writes2);
-				_mm_store_si128((__m128i*)&b_copy[3][i * VEC_CNT], writes3);
-				_mm_store_si128((__m128i*)&b_copy[4][i * VEC_CNT], writes4);
-				_mm_store_si128((__m128i*)&b_copy[5][i * VEC_CNT], writes5);
-				_mm_store_si128((__m128i*)&b_copy[6][i * VEC_CNT], writes6);
-				_mm_store_si128((__m128i*)&b_copy[7][i * VEC_CNT], writes7);
+				_mm_store_si128((__m128i*)&b_copy[0][i * VEC_CNT], row0);
+				_mm_store_si128((__m128i*)&b_copy[1][i * VEC_CNT], row1);
+				_mm_store_si128((__m128i*)&b_copy[2][i * VEC_CNT], row2);
+				_mm_store_si128((__m128i*)&b_copy[3][i * VEC_CNT], row3);
+				_mm_store_si128((__m128i*)&b_copy[4][i * VEC_CNT], row4);
+				_mm_store_si128((__m128i*)&b_copy[5][i * VEC_CNT], row5);
+				_mm_store_si128((__m128i*)&b_copy[6][i * VEC_CNT], row6);
+				_mm_store_si128((__m128i*)&b_copy[7][i * VEC_CNT], row7);
 			}
 
-			for (size_t j = 0; j < VEC_CNT; ++j) {
-				size_t j_abs = j + (VEC_CNT * j_block);
-#if 0
-			// 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) {
-				const size_t k = i * VEC_CNT;
-				__m128i vec    = _mm_setr_epi16(b.addr(k + 0, j), b.addr(k + 1, j), b.addr(k + 2, j), b.addr(k + 3, j), b.addr(k + 4, j), b.addr(k + 5, j), b.addr(k + 6, j), b.addr(k + 7, j));
-				_mm_storeu_si128((__m128i*)&b_copy[i * VEC_CNT], vec);
-			}
-#else
-				//for (size_t i = 0; i < B_COPY_SIZE; ++i) {
-				//	b_copy[j][i] = b.addr(i, j_abs);
-				//}
-#endif
+			for (size_t i = 0; i < VS; ++i) {
+				//__m128i to_write = _mm_set1_epi16(0);
+
+				for (size_t j = 0; j < VEC_CNT; ++j) {
+					size_t j_abs = j + (VEC_CNT * j_block);
 
-				for (size_t i = 0; i < VS; ++i) {
 					const matrix_element res_init_val = std::numeric_limits<matrix_element>::max();
 					__m128i res_vec			  = _mm_set1_epi16(res_init_val);
 
@@ -155,30 +140,35 @@ public:
 					}
 
 					__m128i mn = _mm_minpos_epu16(res_vec);
-
+					mn	   = _mm_and_si128(mn, MASK);
 #if 0
-				const size_t iter = j % VEC_CNT;
-				switch (iter) {
-					gen_switch_label(write_vec, mn, 0);
-					gen_switch_label(write_vec, mn, 1);
-					gen_switch_label(write_vec, mn, 2);
-					gen_switch_label(write_vec, mn, 3);
-					gen_switch_label(write_vec, mn, 4);
-					gen_switch_label(write_vec, mn, 5);
-					gen_switch_label(write_vec, mn, 6);
-					gen_switch_label(write_vec, mn, 7);
-				default:
-					assert(false);
-					break;
-				}
 
-				if ((iter + 1) == VEC_CNT) {
-					_mm_store_si128((__m128i*)&addr(i, j - iter), write_vec);
-				}
+#define get_to_write_assign_label(pos)            \
+	case pos:                                 \
+		mn = _mm_slli_si128(mn, pos * 2); \
+		break
+
+					switch (j) {
+						get_to_write_assign_label(0);
+						get_to_write_assign_label(1);
+						get_to_write_assign_label(2);
+						get_to_write_assign_label(3);
+						get_to_write_assign_label(4);
+						get_to_write_assign_label(5);
+						get_to_write_assign_label(6);
+						get_to_write_assign_label(7);
+					default:
+						assert(false);
+						break;
+					}
+
+					to_write = _mm_or_si128(to_write, mn);
 #endif
 
 					addr(i, j_abs) = _mm_extract_epi16(mn, 0);
 				}
+
+				//_mm_store_si128((__m128i*)&addr(i, VEC_CNT * j_block), to_write);
 			}
 		}
 	}