软件优化 例子.pptVIP

C6000软件优化 实例 1. 定点点积 通过软件流水等技术，N点点积的时间?N/2。 循环展开。一次迭代出两次结果。但是，必须保证循环次数为偶数次。如果不是循环偶数次, 程序员自己补充memory pad. 定点点积的相关图(.sa) L1: ; PIPED LOOP PROLOG LDW .D2T2 *B5++,B6 ; |16| (P) 0,0 load ai ai+1 from memory || [ A0] BDEC .S1 LOOP,A0 ; |23| (P) 0,2 branch to loop || LDW .D1T1 *A5++,A4 ; |17| (P) 0,0 load bi bi+1 from memory LDW .D2T2 *B5++,B6 ; |16| (P) 1,0 load ai ai+1 from memory || [ A0] BDEC .S1 LOOP,A0 ; |23| (P) 1,2 branch to loop || LDW .D1T1 *A5++,A4 ; |17| (P) 1,0 load bi bi+1 from memory LDW .D2T2 *B5++,B6 ; |16| (P) 2,0 load ai ai+1 from memory || [ A0] BDEC .S1 LOOP,A0 ; |23| (P) 2,2 branch to loop || LDW .D1T1 *A5++,A4 ; |17| (P) 2,0 load bi bi+1 from memory LDW .D2T2 *B5++,B6 ; |16| (P) 3,0 load ai ai+1 from memory || [ A0] BDEC .S1 LOOP,A0 ; |23| (P) 3,2 branch to loop || LDW .D1T1 *A5++,A4 ; |17| (P) 3,0 load bi bi+1 from memory MV .S2X A6,B4 || MVK .L2 0x2,B0 ; init prolog collapse predicate || LDW .D2T2 *B5++,B6 ; |16| (P) 4,0 load ai ai+1 from memory || [ A0] BDEC .S1 LOOP,A0 ; |23| (P) 4,2 branch to loop || LDW .D1T1 *A5++,A4 ; |17| (P) 4,0 load bi bi+1 from memory ;** --------------------------------------------------------------------------* LOOP: ; PIPED LOOP KERNEL [ B0] SUB .S2 B0,1,B0 ; 0,7 || [!B0] ADD .L2 B7,B4,B4 ; |20| 0,7 sum0 += (ai * bi) || [!B0] ADD .L1 A6,A3,A3 ; |21| 0,7 sum1 += (ai+1 * bi+1) || MPY .M2X B6,A4,B7 ; |18| 2,5 ai * bi || MPYH .M1X B6,A4,A6 ; |19| 2,5 ai+1 * bi+1 || [ A0] BDEC .S1 LOOP,A