atl_dmm4x4xur2_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 2,473 行 · 第 1/5 页

	madd.d	rC03, rC03, ra0, rb3
	madd.d	rC13, rC13, ra1, rb3
	madd.d	rC23, rC23, ra2, rb3
	madd.d	rC33, rC33, ra3, rb3
#endif
#if KB > 62
					ldc1	rb0, 488(pB0)
	madd.d	rC00, rC00, rA0, rB0
					ldc1	ra0, 488(pA0)
	madd.d	rC10, rC10, rA1, rB0
					ldc1	ra1, 488(pA1)
	madd.d	rC20, rC20, rA2, rB0
					ldc1	ra2, 488(pA2)
	madd.d	rC30, rC30, rA3, rB0
					ldc1	ra3, 488(pA3)
	madd.d	rC01, rC01, rA0, rB1
					ldc1	rb1, 488(pB1)
	madd.d	rC11, rC11, rA1, rB1
					ldc1	rb2, 488(pB2)
	madd.d	rC21, rC21, rA2, rB1
					ldc1	rb3, 488(pB3)
	madd.d	rC31, rC31, rA3, rB1
	madd.d	rC02, rC02, rA0, rB2
	madd.d	rC12, rC12, rA1, rB2
	madd.d	rC22, rC22, rA2, rB2
	madd.d	rC32, rC32, rA3, rB2
	madd.d	rC03, rC03, rA0, rB3
	madd.d	rC13, rC13, rA1, rB3
	madd.d	rC23, rC23, rA2, rB3
	madd.d	rC33, rC33, rA3, rB3
					ldc1	rB0, 496(pB0)
	madd.d	rC00, rC00, ra0, rb0
					ldc1	rA0, 496(pA0)
	madd.d	rC10, rC10, ra1, rb0
					ldc1	rA1, 496(pA1)
	madd.d	rC20, rC20, ra2, rb0
					ldc1	rA2, 496(pA2)
	madd.d	rC30, rC30, ra3, rb0
					ldc1	rA3, 496(pA3)
	madd.d	rC01, rC01, ra0, rb1
					ldc1	rB1, 496(pB1)
	madd.d	rC11, rC11, ra1, rb1
					ldc1	rB2, 496(pB2)
	madd.d	rC21, rC21, ra2, rb1
					ldc1	rB3, 496(pB3)
	madd.d	rC31, rC31, ra3, rb1
	madd.d	rC02, rC02, ra0, rb2
	madd.d	rC12, rC12, ra1, rb2
	madd.d	rC22, rC22, ra2, rb2
	madd.d	rC32, rC32, ra3, rb2
	madd.d	rC03, rC03, ra0, rb3
	madd.d	rC13, rC13, ra1, rb3
	madd.d	rC23, rC23, ra2, rb3
	madd.d	rC33, rC33, ra3, rb3
#endif
#if KB > 64
					ldc1	rb0, 504(pB0)
	madd.d	rC00, rC00, rA0, rB0
					ldc1	ra0, 504(pA0)
	madd.d	rC10, rC10, rA1, rB0
					ldc1	ra1, 504(pA1)
	madd.d	rC20, rC20, rA2, rB0
					ldc1	ra2, 504(pA2)
	madd.d	rC30, rC30, rA3, rB0
					ldc1	ra3, 504(pA3)
	madd.d	rC01, rC01, rA0, rB1
					ldc1	rb1, 504(pB1)
	madd.d	rC11, rC11, rA1, rB1
					ldc1	rb2, 504(pB2)
	madd.d	rC21, rC21, rA2, rB1
					ldc1	rb3, 504(pB3)
	madd.d	rC31, rC31, rA3, rB1
	madd.d	rC02, rC02, rA0, rB2
	madd.d	rC12, rC12, rA1, rB2
	madd.d	rC22, rC22, rA2, rB2
	madd.d	rC32, rC32, rA3, rB2
	madd.d	rC03, rC03, rA0, rB3
	madd.d	rC13, rC13, rA1, rB3
	madd.d	rC23, rC23, rA2, rB3
	madd.d	rC33, rC33, rA3, rB3
					ldc1	rB0, 512(pB0)
	madd.d	rC00, rC00, ra0, rb0
					ldc1	rA0, 512(pA0)
	madd.d	rC10, rC10, ra1, rb0
					ldc1	rA1, 512(pA1)
	madd.d	rC20, rC20, ra2, rb0
					ldc1	rA2, 512(pA2)
	madd.d	rC30, rC30, ra3, rb0
					ldc1	rA3, 512(pA3)
	madd.d	rC01, rC01, ra0, rb1
					ldc1	rB1, 512(pB1)
	madd.d	rC11, rC11, ra1, rb1
					ldc1	rB2, 512(pB2)
	madd.d	rC21, rC21, ra2, rb1
					ldc1	rB3, 512(pB3)
	madd.d	rC31, rC31, ra3, rb1
	madd.d	rC02, rC02, ra0, rb2
	madd.d	rC12, rC12, ra1, rb2
	madd.d	rC22, rC22, ra2, rb2
	madd.d	rC32, rC32, ra3, rb2
	madd.d	rC03, rC03, ra0, rb3
	madd.d	rC13, rC13, ra1, rb3
	madd.d	rC23, rC23, ra2, rb3
	madd.d	rC33, rC33, ra3, rb3
#endif
#if KB > 66
					ldc1	rb0, 520(pB0)
	madd.d	rC00, rC00, rA0, rB0
					ldc1	ra0, 520(pA0)
	madd.d	rC10, rC10, rA1, rB0
					ldc1	ra1, 520(pA1)
	madd.d	rC20, rC20, rA2, rB0
					ldc1	ra2, 520(pA2)
	madd.d	rC30, rC30, rA3, rB0
					ldc1	ra3, 520(pA3)
	madd.d	rC01, rC01, rA0, rB1
					ldc1	rb1, 520(pB1)
	madd.d	rC11, rC11, rA1, rB1
					ldc1	rb2, 520(pB2)
	madd.d	rC21, rC21, rA2, rB1
					ldc1	rb3, 520(pB3)
	madd.d	rC31, rC31, rA3, rB1
	madd.d	rC02, rC02, rA0, rB2
	madd.d	rC12, rC12, rA1, rB2
	madd.d	rC22, rC22, rA2, rB2
	madd.d	rC32, rC32, rA3, rB2
	madd.d	rC03, rC03, rA0, rB3
	madd.d	rC13, rC13, rA1, rB3
	madd.d	rC23, rC23, rA2, rB3
	madd.d	rC33, rC33, rA3, rB3
					ldc1	rB0, 528(pB0)
	madd.d	rC00, rC00, ra0, rb0
					ldc1	rA0, 528(pA0)
	madd.d	rC10, rC10, ra1, rb0
					ldc1	rA1, 528(pA1)
	madd.d	rC20, rC20, ra2, rb0
					ldc1	rA2, 528(pA2)
	madd.d	rC30, rC30, ra3, rb0
					ldc1	rA3, 528(pA3)
	madd.d	rC01, rC01, ra0, rb1
					ldc1	rB1, 528(pB1)
	madd.d	rC11, rC11, ra1, rb1
					ldc1	rB2, 528(pB2)
	madd.d	rC21, rC21, ra2, rb1
					ldc1	rB3, 528(pB3)
	madd.d	rC31, rC31, ra3, rb1
	madd.d	rC02, rC02, ra0, rb2
	madd.d	rC12, rC12, ra1, rb2
	madd.d	rC22, rC22, ra2, rb2
	madd.d	rC32, rC32, ra3, rb2
	madd.d	rC03, rC03, ra0, rb3
	madd.d	rC13, rC13, ra1, rb3
	madd.d	rC23, rC23, ra2, rb3
	madd.d	rC33, rC33, ra3, rb3
#endif
#if KB > 68
					ldc1	rb0, 536(pB0)
	madd.d	rC00, rC00, rA0, rB0
					ldc1	ra0, 536(pA0)
	madd.d	rC10, rC10, rA1, rB0
					ldc1	ra1, 536(pA1)
	madd.d	rC20, rC20, rA2, rB0
					ldc1	ra2, 536(pA2)
	madd.d	rC30, rC30, rA3, rB0
					ldc1	ra3, 536(pA3)
	madd.d	rC01, rC01, rA0, rB1
					ldc1	rb1, 536(pB1)
	madd.d	rC11, rC11, rA1, rB1
					ldc1	rb2, 536(pB2)
	madd.d	rC21, rC21, rA2, rB1
					ldc1	rb3, 536(pB3)
	madd.d	rC31, rC31, rA3, rB1
	madd.d	rC02, rC02, rA0, rB2
	madd.d	rC12, rC12, rA1, rB2
	madd.d	rC22, rC22, rA2, rB2
	madd.d	rC32, rC32, rA3, rB2
	madd.d	rC03, rC03, rA0, rB3
	madd.d	rC13, rC13, rA1, rB3
	madd.d	rC23, rC23, rA2, rB3
	madd.d	rC33, rC33, rA3, rB3
					ldc1	rB0, 544(pB0)
	madd.d	rC00, rC00, ra0, rb0
					ldc1	rA0, 544(pA0)
	madd.d	rC10, rC10, ra1, rb0
					ldc1	rA1, 544(pA1)
	madd.d	rC20, rC20, ra2, rb0
					ldc1	rA2, 544(pA2)
	madd.d	rC30, rC30, ra3, rb0
					ldc1	rA3, 544(pA3)
	madd.d	rC01, rC01, ra0, rb1
					ldc1	rB1, 544(pB1)
	madd.d	rC11, rC11, ra1, rb1
					ldc1	rB2, 544(pB2)
	madd.d	rC21, rC21, ra2, rb1
					ldc1	rB3, 544(pB3)
	madd.d	rC31, rC31, ra3, rb1
	madd.d	rC02, rC02, ra0, rb2
	madd.d	rC12, rC12, ra1, rb2
	madd.d	rC22, rC22, ra2, rb2
	madd.d	rC32, rC32, ra3, rb2
	madd.d	rC03, rC03, ra0, rb3
	madd.d	rC13, rC13, ra1, rb3
	madd.d	rC23, rC23, ra2, rb3
	madd.d	rC33, rC33, ra3, rb3
#endif
#if KB > 70
					ldc1	rb0, 552(pB0)
	madd.d	rC00, rC00, rA0, rB0
					ldc1	ra0, 552(pA0)
	madd.d	rC10, rC10, rA1, rB0
					ldc1	ra1, 552(pA1)
	madd.d	rC20, rC20, rA2, rB0
					ldc1	ra2, 552(pA2)
	madd.d	rC30, rC30, rA3, rB0
					ldc1	ra3, 552(pA3)
	madd.d	rC01, rC01, rA0, rB1
					ldc1	rb1, 552(pB1)
	madd.d	rC11, rC11, rA1, rB1
					ldc1	rb2, 552(pB2)
	madd.d	rC21, rC21, rA2, rB1
					ldc1	rb3, 552(pB3)
	madd.d	rC31, rC31, rA3, rB1
	madd.d	rC02, rC02, rA0, rB2
	madd.d	rC12, rC12, rA1, rB2
	madd.d	rC22, rC22, rA2, rB2
	madd.d	rC32, rC32, rA3, rB2
	madd.d	rC03, rC03, rA0, rB3
	madd.d	rC13, rC13, rA1, rB3
	madd.d	rC23, rC23, rA2, rB3
	madd.d	rC33, rC33, rA3, rB3
					ldc1	rB0, 560(pB0)
	madd.d	rC00, rC00, ra0, rb0
					ldc1	rA0, 560(pA0)
	madd.d	rC10, rC10, ra1, rb0
					ldc1	rA1, 560(pA1)
	madd.d	rC20, rC20, ra2, rb0
					ldc1	rA2, 560(pA2)
	madd.d	rC30, rC30, ra3, rb0
					ldc1	rA3, 560(pA3)
	madd.d	rC01, rC01, ra0, rb1
					ldc1	rB1, 560(pB1)
	madd.d	rC11, rC11, ra1, rb1
					ldc1	rB2, 560(pB2)
	madd.d	rC21, rC21, ra2, rb1
					ldc1	rB3, 560(pB3)
	madd.d	rC31, rC31, ra3, rb1
	madd.d	rC02, rC02, ra0, rb2
	madd.d	rC12, rC12, ra1, rb2
	madd.d	rC22, rC22, ra2, rb2
	madd.d	rC32, rC32, ra3, rb2
	madd.d	rC03, rC03, ra0, rb3
	madd.d	rC13, rC13, ra1, rb3
	madd.d	rC23, rC23, ra2, rb3
	madd.d	rC33, rC33, ra3, rb3
#endif
#if KB > 72
					ldc1	rb0, 568(pB0)
	madd.d	rC00, rC00, rA0, rB0
					ldc1	ra0, 568(pA0)
	madd.d	rC10, rC10, rA1, rB0
					ldc1	ra1, 568(pA1)
	madd.d	rC20, rC20, rA2, rB0
					ldc1	ra2, 568(pA2)
	madd.d	rC30, rC30, rA3, rB0
					ldc1	ra3, 568(pA3)
	madd.d	rC01, rC01, rA0, rB1
					ldc1	rb1, 568(pB1)
	madd.d	rC11, rC11, rA1, rB1
					ldc1	rb2, 568(pB2)
	madd.d	rC21, rC21, rA2, rB1
					ldc1	rb3, 568(pB3)
	madd.d	rC31, rC31, rA3, rB1
	madd.d	rC02, rC02, rA0, rB2
	madd.d	rC12, rC12, rA1, rB2
	madd.d	rC22, rC22, rA2, rB2
	madd.d	rC32, rC32, rA3, rB2
	madd.d	rC03, rC03, rA0, rB3
	madd.d	rC13, rC13, rA1, rB3
	madd.d	rC23, rC23, rA2, rB3
	madd.d	rC33, rC33, rA3, rB3
					ldc1	rB0, 576(pB0)
	madd.d	rC00, rC00, ra0, rb0
					ldc1	rA0, 576(pA0)
	madd.d	rC10, rC10, ra1, rb0
					ldc1	rA1, 576(pA1)
	madd.d	rC20, rC20, ra2, rb0
					ldc1	rA2, 576(pA2)
	madd.d	rC30, rC30, ra3, rb0
					ldc1	rA3, 576(pA3)
	madd.d	rC01, rC01, ra0, rb1
					ldc1	rB1, 576(pB1)
	madd.d	rC11, rC11, ra1, rb1
					ldc1	rB2, 576(pB2)
	madd.d	rC21, rC21, ra2, rb1
					ldc1	rB3, 576(pB3)
	madd.d	rC31, rC31, ra3, rb1
	madd.d	rC02, rC02, ra0, rb2
	madd.d	rC12, rC12, ra1, rb2
	madd.d	rC22, rC22, ra2, rb2
	madd.d	rC32, rC32, ra3, rb2
	madd.d	rC03, rC03, ra0, rb3
	madd.d	rC13, rC13, ra1, rb3
	madd.d	rC23, rC23, ra2, rb3
	madd.d	rC33, rC33, ra3, rb3
#endif
#if KB > 74
					ldc1	rb0, 584(pB0)
	madd.d	rC00, rC00, rA0, rB0
					ldc1	ra0, 584(pA0)
	madd.d	rC10, rC10, rA1, rB0
					ldc1	ra1, 584(pA1)
	madd.d	rC20, rC20, rA2, rB0
					ldc1	ra2, 584(pA2)
	madd.d	rC30, rC30, rA3, rB0
					ldc1	ra3, 584(pA3)
	madd.d	rC01, rC01, rA0, rB1
					ldc1	rb1, 584(pB1)
	madd.d	rC11, rC11, rA1, rB1
					ldc1	rb2, 584(pB2)
	madd.d	rC21, rC21, rA2, rB1
					ldc1	rb3, 584(pB3)
	madd.d	rC31, rC31, rA3, rB1
	madd.d	rC02, rC02, rA0, rB2
	madd.d	rC12, rC12, rA1, rB2
	madd.d	rC22, rC22, rA2, rB2
	madd.d	rC32, rC32, rA3, rB2
	madd.d	rC03, rC03, rA0, rB3
	madd.d	rC13, rC13, rA1, rB3
	madd.d	rC23, rC23, rA2, rB3
	madd.d	rC33, rC33, rA3, rB3
					ldc1	rB0, 592(pB0)
	madd.d	rC00, rC00, ra0, rb0
					ldc1	rA0, 592(pA0)
	madd.d	rC10, rC10, ra1, rb0
					ldc1	rA1, 592(pA1)
	madd.d	rC20, rC20, ra2, rb0
					ldc1	rA2, 592(pA2)
	madd.d	rC30, rC30, ra3, rb0
					ldc1	rA3, 592(pA3)
	madd.d	rC01, rC01, ra0, rb1
					ldc1	rB1, 592(pB1)
	madd.d	rC11, rC11, ra1, rb1
					ldc1	rB2, 592(pB2)
	madd.d	rC21, rC21, ra2, rb1
					ldc1	rB3, 592(pB3)
	madd.d	rC31, rC31, ra3, rb1
	madd.d	rC02, rC02, ra0, rb2
	madd.d	rC12, rC12, ra1, rb2
	madd.d	rC22, rC22, ra2, rb2
	madd.d	rC32, rC32, ra3, rb2
	madd.d	rC03, rC03, ra0, rb3
	madd.d	rC13, rC13, ra1, rb3
	madd.d	rC23, rC23, ra2, rb3
	madd.d	rC33, rC33, ra3, rb3
#endif
#if KB > 76
					ldc1	rb0, 600(pB0)
	madd.d	rC00, rC00, rA0, rB0
					ldc1	ra0, 600(pA0)
	madd.d	rC10, rC10, rA1, rB0
					ldc1	ra1, 600(pA1)
	madd.d	rC20, rC20, rA2, rB0
					ldc1	ra2, 600(pA2)
	madd.d	rC30, rC30, rA3, rB0
					ldc1	ra3, 600(pA3)
	madd.d	rC01, rC01, rA0, rB1
					ldc1	rb1, 600(pB1)
	madd.d	rC11, rC11, rA1, rB1
					ldc1	rb2, 600(pB2)
	madd.d	rC21, rC21, rA2, rB1
					ldc1	rb3, 600(pB3)
	madd.d	rC31, rC31, rA3, rB1
	madd.d	rC02, rC02, rA0, rB2
	madd.d	rC12, rC12, rA1, rB2
	madd.d	rC22, rC22, rA2, rB2
	madd.d	rC32, rC32, rA3, rB2
	madd.d	rC03, rC03, rA0, rB3
	madd.d	rC13, rC13, rA1, rB3
	madd.d	rC23, rC23, rA2, rB3
	madd.d	rC33, rC33, rA3, rB3
					ldc1	rB0, 608(pB0)
	madd.d	rC00, rC00, ra0, rb0
					ldc1	rA0, 608(pA0)
	madd.d	rC10, rC10, ra1, rb0
					ldc1	rA1, 608(pA1)
	madd.d	rC20, rC20, ra2, rb0
					ldc1	rA2, 608(pA2)
	madd.d	rC30, rC30, ra3, rb0
					ldc1	rA3, 608(pA3)
	madd.d	rC01, rC01, ra0, rb1
					ldc1	rB1, 608(pB1)
	madd.d	rC11, rC11, ra1, rb1
					ldc1	rB2, 608(pB2)
	madd.d	rC21, rC21, ra2, rb1
					ldc1	rB3, 608(pB3)
	madd.d	rC31, rC31, ra3, rb1
	madd.d	rC02, rC02, ra0, rb2
	madd.d	rC12, rC12, ra1, rb2
	madd.d	rC22, rC22, ra2, rb2
	madd.d	rC32, rC32, ra3, rb2
	madd.d	rC03, rC03, ra0, rb3
	madd.d	rC13, rC13, ra1, rb3
	madd.d	rC23, rC23, ra2, rb3
	madd.d	rC33, rC33, ra3, rb3
#endif
#if KB > 78
					ldc1	rb0, 616(pB0)
	madd.d	rC00, rC00, rA0, rB0
					ldc1	ra0, 616(pA0)
	madd.d	rC10, rC10, rA1, rB0
					ldc1	ra1, 616(pA1)
	madd.d	rC20, rC20, rA2, rB0
					ldc1	ra2, 616(pA2)
	madd.d	rC30, rC30, rA3, rB0
					ldc1	ra3, 616(pA3)
	madd.d	rC01, rC01, rA0, rB1
					ldc1	rb1, 616(pB1)
	madd.d	rC11, rC11, rA1, rB1
					ldc1	rb2, 616(pB2)
	madd.d	rC21, rC21, rA2, rB1
					ldc1	rb3, 616(pB3)
	madd.d	rC31, rC31, rA3, rB1
	madd.d	rC02, rC02, rA0, rB2
	madd.d	rC12, rC12, rA1, rB2
	madd.d	rC22, rC22, rA2, rB2
	madd.d	rC32, rC32, rA3, rB2
	madd.d	rC03, rC03, rA0, rB3
	madd.d	rC13, rC13, rA1, rB3
	madd.d	rC23, rC23, rA2, rB3
	madd.d	rC33, rC33, rA3, rB3
					ldc1	rB0, 624(pB0)
	madd.d	rC00, rC00, ra0, rb0
					ldc1	rA0, 624(pA0)
	madd.d	rC10, rC10, ra1, rb0
					ldc1	rA1, 624(pA1)
	madd.d	rC20, rC20, ra2, rb0
					ldc1	rA2, 624(pA2)
	madd.d	rC30, rC30, ra3, rb0
					ldc1	rA3, 624(pA3)
	madd.d	rC01, rC01, ra0, rb1
					ldc1	rB1, 624(pB1)
	madd.d	rC11, rC11, ra1, rb1
					ldc1	rB2, 624(pB2)
	madd.d	rC21, rC21, ra2, rb1
					ldc1	rB3, 624(pB3)
	madd.d	rC31, rC31, ra3, rb1
	madd.d	rC02, rC02, ra0, rb2
	madd.d	rC12, rC12, ra1, rb2
	madd.d	rC22, rC22, ra2, rb2
	madd.d	rC32, rC32, ra3, rb2
	madd.d	rC03, rC03, ra0, rb3
	madd.d	rC13, rC13, ra1, rb3
	madd.d	rC23, rC23, ra2, rb3
	madd.d	rC33, rC33, ra3, rb3
#endif
#if KB > 80
					ldc1	rb0, 632(pB0)
	madd.d	rC00, rC00, rA0, rB0
					ldc1	ra0, 632(pA0)
	madd.d	rC10, rC10, rA1, rB0
					ldc1	ra1, 632(pA1)
	madd.d	rC20, rC20, rA2, rB0
					ldc1	ra2, 632(pA2)
	madd.d	rC30, rC30, rA3, rB0
					ldc1	ra3, 632(pA3)
	madd.d	rC01, rC01, rA0, rB1
					ldc1	rb1, 632(pB1)
	madd.d	rC11, rC11, rA1, rB1
					ldc1	rb2, 632(pB2)
	madd.d	rC21, rC21, rA2, rB1
					ldc1	rb3, 632(pB3)
	madd.d	rC31, rC31, rA3, rB1
	madd.d	rC02, rC02, rA0, rB2
	madd.d	rC12, rC12, rA1, rB2
	madd.d	rC22, rC22, rA2, rB2
	madd.d	rC32, rC32, rA3, rB2
	madd.d	rC03, rC03, rA0, rB3
	madd.d	rC13, rC13, rA1, rB3
	madd.d	rC23, rC23, rA2, rB3
	madd.d	rC33, rC33, rA3, rB3
					ldc1	rB0, 640(pB0)
	madd.d	rC00, rC00, ra0, rb0
					ldc1	rA0, 640(pA0)
	madd.d	rC10, rC10, ra1, rb0
					ldc1	rA1, 640(pA1)
	madd.d	rC20, rC20, ra2, rb0
					ldc1	rA2, 640(pA2)
	madd.d	rC30, rC30, ra3, rb0
					ldc1	rA3, 640(pA3)
	madd.d	rC01, rC01, ra0, rb1
					ldc1	rB1, 640(pB1)
	madd.d	rC11, rC11, ra1, rb1
					ldc1	rB2, 640(