atl_dmm4x4xurx_mips.c
来自「基于Blas CLapck的.用过的人知道是干啥的」· C语言 代码 · 共 3,133 行 · 第 1/5 页
C
3,133 行
#endif madd.d rC20, rC20, rE2, rb0 madd.d rC30, rC30, ra3, rb0 #if KB > 6 ldc1 rB0, 48(pB0) #endif madd.d rC01, rC01, ra0, rB1 madd.d rC11, rC11, ra1, rB1 madd.d rC21, rC21, rE2, rB1 madd.d rC31, rC31, ra3, rB1 #if KB > 6 ldc1 rB1, 48(pB1) #endif madd.d rC02, rC02, ra0, rB2 madd.d rC12, rC12, ra1, rB2 #if KB > 7 ldc1 ra2, 56(pA2) #endif madd.d rC22, rC22, rE2, rB2 madd.d rC32, rC32, ra3, rB2 #if KB > 6 ldc1 rB2, 48(pB2) #endif madd.d rC03, rC03, ra0, rB3 madd.d rC13, rC13, ra1, rB3 #if KB > 6 ldc1 rA3, 48(pA3) #endif madd.d rC23, rC23, rE2, rB3 madd.d rC33, rC33, ra3, rB3 #if KB > 6 ldc1 rB3, 48(pB3) #endif#endif#if KB > 6 #if KB > 7 ldc1 ra1, 56(pA1) #endif madd.d rC00, rC00, rE0, rB0 prefB(KB*8*4(pB0)) madd.d rC10, rC10, rA1, rB0 #if KB > 9 ldc1 ra0, 72(pA0) #endif madd.d rC20, rC20, rA2, rB0 prefB(KB*8*4(pB1)) madd.d rC30, rC30, rA3, rB0 #if KB > 7 ldc1 rb0, 56(pB0) #endif madd.d rC01, rC01, rE0, rB1 madd.d rC11, rC11, rA1, rB1 prefB(KB*8*4(pB2)) madd.d rC21, rC21, rA2, rB1 madd.d rC31, rC31, rA3, rB1 #if KB > 7 ldc1 rB1, 56(pB1) #endif madd.d rC02, rC02, rE0, rB2 madd.d rC12, rC12, rA1, rB2 #if KB > 8 ldc1 rE2, 64(pA2) #endif madd.d rC22, rC22, rA2, rB2 prefB(KB*8*4(pB3)) madd.d rC32, rC32, rA3, rB2 #if KB > 7 ldc1 rB2, 56(pB2) #endif madd.d rC03, rC03, rE0, rB3 madd.d rC13, rC13, rA1, rB3 #if KB > 7 ldc1 ra3, 56(pA3) #endif madd.d rC23, rC23, rA2, rB3 madd.d rC33, rC33, rA3, rB3 #if KB > 7 ldc1 rB3, 56(pB3) #endif#endif#if KB > 7 #if KB > 8 ldc1 rA1, 64(pA1) #endif madd.d rC00, rC00, re0, rb0 prefB(32+KB*8*4(pB0)) madd.d rC10, rC10, ra1, rb0 #if KB > 10 ldc1 rE0, 80(pA0) #endif madd.d rC20, rC20, ra2, rb0 prefB(32+KB*8*4(pB1)) madd.d rC30, rC30, ra3, rb0 #if KB > 8 ldc1 rB0, 64(pB0) #endif madd.d rC01, rC01, re0, rB1 madd.d rC11, rC11, ra1, rB1 prefB(32+KB*8*4(pB2)) madd.d rC21, rC21, ra2, rB1 madd.d rC31, rC31, ra3, rB1 #if KB > 8 ldc1 rB1, 64(pB1) #endif madd.d rC02, rC02, re0, rB2 prefB(32+KB*8*4(pB3)) madd.d rC12, rC12, ra1, rB2 #if KB > 9 ldc1 rA2, 72(pA2) #endif madd.d rC22, rC22, ra2, rB2 madd.d rC32, rC32, ra3, rB2 #if KB > 8 ldc1 rB2, 64(pB2) #endif madd.d rC03, rC03, re0, rB3 madd.d rC13, rC13, ra1, rB3 #if KB > 8 ldc1 rA3, 64(pA3) #endif madd.d rC23, rC23, ra2, rB3 madd.d rC33, rC33, ra3, rB3 #if KB > 8 ldc1 rB3, 64(pB3) #endif#endif#if KB > 8 #if KB > 9 ldc1 ra1, 72(pA1) #endif madd.d rC00, rC00, rA0, rB0 prefB(64+KB*8*4(pB0)) madd.d rC10, rC10, rA1, rB0 #if KB > 11 ldc1 re0, 88(pA0) #endif madd.d rC20, rC20, rE2, rB0 prefB(64+KB*8*4(pB1)) madd.d rC30, rC30, rA3, rB0 #if KB > 9 ldc1 rb0, 72(pB0) #endif madd.d rC01, rC01, rA0, rB1 madd.d rC11, rC11, rA1, rB1 prefB(64+KB*8*4(pB2)) madd.d rC21, rC21, rE2, rB1 madd.d rC31, rC31, rA3, rB1 #if KB > 9 ldc1 rB1, 72(pB1) #endif madd.d rC02, rC02, rA0, rB2 prefB(64+KB*8*4(pB3)) madd.d rC12, rC12, rA1, rB2 #if KB > 10 ldc1 ra2, 80(pA2) #endif madd.d rC22, rC22, rE2, rB2 madd.d rC32, rC32, rA3, rB2 #if KB > 9 ldc1 rB2, 72(pB2) #endif madd.d rC03, rC03, rA0, rB3 madd.d rC13, rC13, rA1, rB3 #if KB > 9 ldc1 ra3, 72(pA3) #endif madd.d rC23, rC23, rE2, rB3 madd.d rC33, rC33, rA3, rB3 #if KB > 9 ldc1 rB3, 72(pB3) #endif#endif#if KB > 9 #if KB > 10 ldc1 rA1, 80(pA1) #endif madd.d rC00, rC00, ra0, rb0 prefB(96+KB*8*4(pB0)) madd.d rC10, rC10, ra1, rb0 #if KB > 12 ldc1 rA0, 96(pA0) #endif madd.d rC20, rC20, rA2, rb0 prefB(96+KB*8*4(pB1)) madd.d rC30, rC30, ra3, rb0 #if KB > 10 ldc1 rB0, 80(pB0) #endif madd.d rC01, rC01, ra0, rB1 madd.d rC11, rC11, ra1, rB1 prefB(96+KB*8*4(pB2)) madd.d rC21, rC21, rA2, rB1 madd.d rC31, rC31, ra3, rB1 #if KB > 10 ldc1 rB1, 80(pB1) #endif madd.d rC02, rC02, ra0, rB2 prefB(96+KB*8*4(pB3)) madd.d rC12, rC12, ra1, rB2 #if KB > 11 ldc1 rE2, 88(pA2) #endif madd.d rC22, rC22, rA2, rB2 madd.d rC32, rC32, ra3, rB2 #if KB > 10 ldc1 rB2, 80(pB2) #endif madd.d rC03, rC03, ra0, rB3 madd.d rC13, rC13, ra1, rB3 #if KB > 10 ldc1 rA3, 80(pA3) #endif madd.d rC23, rC23, rA2, rB3 madd.d rC33, rC33, ra3, rB3 #if KB > 10 ldc1 rB3, 80(pB3) #endif#endif#if KB > 10 #if KB > 11 ldc1 ra1, 88(pA1) #endif madd.d rC00, rC00, rE0, rB0 prefB(128+KB*8*4(pB0)) madd.d rC10, rC10, rA1, rB0 #if KB > 13 ldc1 ra0, 104(pA0) #endif madd.d rC20, rC20, ra2, rB0 prefB(128+KB*8*4(pB1)) madd.d rC30, rC30, rA3, rB0 #if KB > 11 ldc1 rb0, 88(pB0) #endif madd.d rC01, rC01, rE0, rB1 madd.d rC11, rC11, rA1, rB1 prefB(128+KB*8*4(pB2)) madd.d rC21, rC21, ra2, rB1 madd.d rC31, rC31, rA3, rB1 #if KB > 11 ldc1 rB1, 88(pB1) #endif madd.d rC02, rC02, rE0, rB2 prefB(128+KB*8*4(pB3)) madd.d rC12, rC12, rA1, rB2 #if KB > 12 ldc1 rA2, 96(pA2) #endif madd.d rC22, rC22, ra2, rB2 madd.d rC32, rC32, rA3, rB2 #if KB > 11 ldc1 rB2, 88(pB2) #endif madd.d rC03, rC03, rE0, rB3 madd.d rC13, rC13, rA1, rB3 #if KB > 11 ldc1 ra3, 88(pA3) #endif madd.d rC23, rC23, ra2, rB3 madd.d rC33, rC33, rA3, rB3 #if KB > 11 ldc1 rB3, 88(pB3) #endif#endif#if KB > 11 #if KB > 12 ldc1 rA1, 96(pA1) #endif madd.d rC00, rC00, re0, rb0 prefB(160+KB*8*4(pB0)) madd.d rC10, rC10, ra1, rb0 #if KB > 14 ldc1 rE0, 112(pA0) #endif madd.d rC20, rC20, rE2, rb0 prefB(160+KB*8*4(pB1)) madd.d rC30, rC30, ra3, rb0 #if KB > 12 ldc1 rB0, 96(pB0) #endif madd.d rC01, rC01, re0, rB1 madd.d rC11, rC11, ra1, rB1 prefB(160+KB*8*4(pB2)) madd.d rC21, rC21, rE2, rB1 madd.d rC31, rC31, ra3, rB1 #if KB > 12 ldc1 rB1, 96(pB1) #endif madd.d rC02, rC02, re0, rB2 prefB(160+KB*8*4(pB3)) madd.d rC12, rC12, ra1, rB2 #if KB > 13 ldc1 ra2, 104(pA2) #endif madd.d rC22, rC22, rE2, rB2 madd.d rC32, rC32, ra3, rB2 #if KB > 12 ldc1 rB2, 96(pB2) #endif madd.d rC03, rC03, re0, rB3 madd.d rC13, rC13, ra1, rB3 #if KB > 12 ldc1 rA3, 96(pA3) #endif madd.d rC23, rC23, rE2, rB3 madd.d rC33, rC33, ra3, rB3 #if KB > 12 ldc1 rB3, 96(pB3) #endif#endif#if KB > 12 #if KB > 13 ldc1 ra1, 104(pA1) #endif madd.d rC00, rC00, rA0, rB0 prefB(192+KB*8*4(pB0)) madd.d rC10, rC10, rA1, rB0 #if KB > 15 ldc1 re0, 120(pA0) #endif madd.d rC20, rC20, rA2, rB0 prefB(192+KB*8*4(pB1)) madd.d rC30, rC30, rA3, rB0 #if KB > 13 ldc1 rb0, 104(pB0) #endif madd.d rC01, rC01, rA0, rB1 madd.d rC11, rC11, rA1, rB1 prefB(192+KB*8*4(pB2)) madd.d rC21, rC21, rA2, rB1 madd.d rC31, rC31, rA3, rB1 #if KB > 13 ldc1 rB1, 104(pB1) #endif madd.d rC02, rC02, rA0, rB2 prefB(192+KB*8*4(pB3)) madd.d rC12, rC12, rA1, rB2 #if KB > 14 ldc1 rE2, 112(pA2) #endif madd.d rC22, rC22, rA2, rB2 madd.d rC32, rC32, rA3, rB2 #if KB > 13 ldc1 rB2, 104(pB2) #endif madd.d rC03, rC03, rA0, rB3 madd.d rC13, rC13, rA1, rB3 #if KB > 13 ldc1 ra3, 104(pA3) #endif madd.d rC23, rC23, rA2, rB3 madd.d rC33, rC33, rA3, rB3 #if KB > 13 ldc1 rB3, 104(pB3) #endif#endif#if KB > 13 #if KB > 14 ldc1 rA1, 112(pA1) #endif madd.d rC00, rC00, ra0, rb0 prefB(224+KB*8*4(pB0)) madd.d rC10, rC10, ra1, rb0 #if KB > 16 ldc1 rA0, 128(pA0) #endif madd.d rC20, rC20, ra2, rb0 prefB(224+KB*8*4(pB1)) madd.d rC30, rC30, ra3, rb0 #if KB > 14 ldc1 rB0, 112(pB0) #endif madd.d rC01, rC01, ra0, rB1 madd.d rC11, rC11, ra1, rB1 prefB(224+KB*8*4(pB2)) madd.d rC21, rC21, ra2, rB1 madd.d rC31, rC31, ra3, rB1 #if KB > 14 ldc1 rB1, 112(pB1) #endif madd.d rC02, rC02, ra0, rB2 prefB(224+KB*8*4(pB3)) madd.d rC12, rC12, ra1, rB2 #if KB > 15 ldc1 rA2, 120(pA2) #endif madd.d rC22, rC22, ra2, rB2 madd.d rC32, rC32, ra3, rB2 #if KB > 14 ldc1 rB2, 112(pB2) #endif madd.d rC03, rC03, ra0, rB3 madd.d rC13, rC13, ra1, rB3 #if KB > 14 ldc1 rA3, 112(pA3) #endif madd.d rC23, rC23, ra2, rB3 madd.d rC33, rC33, ra3, rB3 #if KB > 14 ldc1 rB3, 112(pB3) #endif#endif#if KB > 14 #if KB > 15 ldc1 ra1, 120(pA1) #endif madd.d rC00, rC00, rE0, rB0 prefB(256+KB*8*4(pB0)) madd.d rC10, rC10, rA1, rB0 #if KB > 17 ldc1 ra0, 136(pA0) #endif madd.d rC20, rC20, rE2, rB0 prefB(256+KB*8*4(pB1)) madd.d rC30, rC30, rA3, rB0 #if KB > 15 ldc1 rb0, 120(pB0) #endif madd.d rC01, rC01, rE0, rB1 madd.d rC11, rC11, rA1, rB1 prefB(256+KB*8*4(pB2)) madd.d rC21, rC21, rE2, rB1 madd.d rC31, rC31, rA3, rB1 #if KB > 15 ldc1 rB1, 120(pB1) #endif madd.d rC02, rC02, rE0, rB2 prefB(256+KB*8*4(pB3)) madd.d rC12, rC12, rA1, rB2 #if KB > 16 ldc1 ra2, 128(pA2) #endif madd.d rC22, rC22, rE2, rB2 madd.d rC32, rC32, rA3, rB2 #if KB > 15 ldc1 rB2, 120(pB2) #endif madd.d rC03, rC03, rE0, rB3 madd.d rC13, rC13, rA1, rB3 #if KB > 15 ldc1 ra3, 120(pA3) #endif madd.d rC23, rC23, rE2, rB3 madd.d rC33, rC33, rA3, rB3 #if KB > 15 ldc1 rB3, 120(pB3) #endif#endif#if KB > 15 #if KB > 16 ldc1 rA1, 128(pA1) #endif madd.d rC00, rC00, re0, rb0 prefB(288+KB*8*4(pB0)) madd.d rC10, rC10, ra1, rb0 #if KB > 18 ldc1 rE0, 144(pA0) #endif madd.d rC20, rC20, rA2, rb0 prefB(288+KB*8*4(pB1)) madd.d rC30, rC30, ra3, rb0 #if KB > 16 ldc1 rB0, 128(pB0) #endif madd.d rC01, rC01, re0, rB1 madd.d rC11, rC11, ra1, rB1 prefB(288+KB*8*4(pB2)) madd.d rC21, rC21, rA2, rB1 madd.d rC31, rC31, ra3, rB1 #if KB > 16 ldc1 rB1, 128(pB1) #endif madd.d rC02, rC02, re0, rB2 prefB(288+KB*8*4(pB3)) madd.d rC12, rC12, ra1, rB2 #if KB > 17 ldc1 rE2, 136(pA2) #endif madd.d rC22, rC22, rA2, rB2 madd.d rC32, rC32, ra3, rB2 #if KB > 16 ldc1 rB2, 128(pB2) #endif madd.d rC03, rC03, re0, rB3 madd.d rC13, rC13, ra1, rB3 #if KB > 16 ldc1 rA3, 128(pA3) #endif madd.d rC23, rC23, rA2, rB3 madd.d rC33, rC33, ra3, rB3 #if KB > 16 ldc1 rB3, 128(pB3) #endif#endif#if KB > 16 #if KB > 17 ldc1 ra1, 136(pA1) #endif madd.d rC00, rC00, rA0, rB0 prefB(320+KB*8*4(pB0)) madd.d rC10, rC10, rA1, rB0 #if KB > 19 ldc1 re0, 152(pA0) #endif madd.d rC20, rC20, ra2, rB0 prefB(320+KB*8*4(pB1)) madd.d rC30, rC30, rA3, rB0 #if KB > 17 ldc1 rb0, 136(pB0) #endif madd.d rC01, rC01, rA0, rB1 madd.d rC11, rC11, rA1, rB1 prefB(320+KB*8*4(pB2)) madd.d rC21, rC21, ra2, rB1 madd.d rC31, rC31, rA3, rB1 #if KB > 17 ldc1 rB1, 136(pB1) #endif madd.d rC02, rC02, rA0, rB2 prefB(320+KB*8*4(pB3)) madd.d rC12, rC12, rA1, rB2 #if KB > 18 ldc1 rA2, 144(pA2) #endif madd.d rC22, rC22, ra2, rB2 madd.d rC32, rC32, rA3, rB2 #if KB > 17 ldc1 rB2, 136(pB2) #endif madd.d rC03, rC03, rA0, rB3 madd.d rC13, rC13, rA1, rB3 #if KB > 17 ldc1 ra3, 136(pA3) #endif madd.d rC23, rC23, ra2, rB3 madd.d rC33, rC33, rA3, rB3 #if KB > 17 ldc1 rB3, 136(pB3) #endif#endif#if KB > 17 #if KB > 18 ldc1 rA1, 144(pA1) #endif madd.d rC00, rC00, ra0, rb0 prefB(352+KB*8*4(pB0)) madd.d rC10, rC10, ra1, rb0 #if KB > 20 ldc1 rA0, 160(pA0) #endif madd.d rC20, rC20, rE2, rb0 prefB(352+KB*8*4(pB1)) madd.d rC30, rC30, ra3, rb0 #if KB > 18 ldc1 rB0, 144(pB0) #endif madd.d rC01, rC01, ra0, rB1 madd.d rC11, rC11, ra1, rB1 prefB(352+KB*8*4(pB2)) madd.d rC21, rC21, rE2, rB1 madd.d rC31, rC31, ra3, rB1 #if KB > 18 ldc1 rB1, 144(pB1) #endif madd.d rC02, rC02, ra0, rB2 prefB(352+KB*8*4(pB3)) madd.d rC12, rC12, ra1, rB2 #if KB > 19 ldc1 ra2, 152(pA2) #endif madd.d rC22, rC22, rE2, rB2 madd.d rC32, rC32, ra3, rB2 #if KB > 18 ldc1 rB2, 144(pB2) #endif madd.d rC03, rC03, ra0, rB3 madd.d rC13, rC13, ra1, rB3 #if KB > 18 ldc1 rA3, 144(pA3) #endif madd.d rC23, rC23, rE2, rB3 madd.d rC33, rC33, ra3, rB3 #if KB > 18 ldc1 rB3, 144(pB3) #endif#endif#if KB > 18 #if KB > 19 ldc1 ra1, 152(pA1) #endif madd.d rC00, rC00, rE0, rB0 prefB(384+KB*8*4(pB0)) madd.d rC10, rC10, rA1, rB0 #if KB > 21 ldc1 ra0, 168(pA0) #endif madd.d rC20, rC20, rA2, rB0 prefB(384+KB*8*4(pB1)) madd.d rC30, rC30, rA3, rB0 #if KB > 19 ldc1 rb0, 152(pB0) #endif madd.d rC01, rC01, rE0, rB1 madd.d rC11, rC11, rA1, rB1 prefB(384+KB*8*4(pB2)) madd.d rC21, rC21, rA2, rB1 madd.d rC31, rC31, rA3, rB1 #if KB > 19 ldc1 rB1, 152(pB1) #endif madd.d rC02, rC02, rE0, rB2 prefB(384+KB*8*4(pB3)) madd.d rC12, rC12, rA1, rB2 #if KB > 20 ldc1 rE2, 160(pA2) #endif madd.d rC22, rC22, rA2, rB2 madd.d rC32, rC32, rA3, rB2 #if KB > 19 ldc1 rB2, 152(pB2) #endif madd.d rC03, rC03, rE0, rB3 madd.d rC13, rC13, rA1, rB3 #if KB > 19 ldc1 ra3, 152(pA3) #endif madd.d rC23, rC23, rA2, rB3
⌨️ 快捷键说明
复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?