atl_smm4x4xurx_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 3,232 行 · 第 1/5 页

					lwc1	ra3, 76(pA3)
	#elif KB == 19
					swc1	rC22, -CMUL(8)(pC2)
	#endif
	madd.s	rC23, rC23, rA2, rB3
	#if KB == 19
					swc1	rC32, -CMUL(4)(pC2)
	#endif
	madd.s	rC33, rC33, rA3, rB3
	#if KB > 19
					lwc1	rB3, 76(pB3)
	#endif
#endif
#if KB > 19
	#if KB > 20
					lwc1	rA1, 80(pA1)
	#elif KB == 20
					daddu	pA0, pA0, incAm
	#endif
	madd.s	rC00, rC00, re0, rb0
	#if KB == 20
					daddu	pA1, pA1, incAm
	#endif
	madd.s	rC10, rC10, ra1, rb0
	#if KB > 22
					lwc1	rE0, 88(pA0)
	#elif KB == 20
					daddu	pA2, pA2, incAm
	#endif
	madd.s	rC20, rC20, ra2, rb0
	#if KB == 20
					daddu	pA3, pA3, incAm
	#endif
	madd.s	rC30, rC30, ra3, rb0
	#if KB > 20
					lwc1	rB0, 80(pB0)
	#elif KB == 20
					swc1	rC00, -CMUL(16)(pC0)
	#endif
	madd.s	rC01, rC01, re0, rB1
	#if KB == 20
					swc1	rC10, -CMUL(12)(pC0)
	#endif
	madd.s	rC11, rC11, ra1, rB1
	#if KB == 20
					swc1	rC20, -CMUL(8)(pC0)
	#endif
	madd.s	rC21, rC21, ra2, rB1
	#if KB == 20
					swc1	rC30, -CMUL(4)(pC0)
	#endif
	madd.s	rC31, rC31, ra3, rB1
	#if KB > 20
					lwc1	rB1, 80(pB1)
	#elif KB == 20
					swc1	rC01, -CMUL(16)(pC1)
	#endif
	madd.s	rC02, rC02, re0, rB2
	#if KB == 20
					swc1	rC11, -CMUL(12)(pC1)
	#endif
	madd.s	rC12, rC12, ra1, rB2
	#if KB > 21
					lwc1	rA2, 84(pA2)
	#elif KB == 20
					swc1	rC21, -CMUL(8)(pC1)
	#endif
	madd.s	rC22, rC22, ra2, rB2
	#if KB == 20
					swc1	rC31, -CMUL(4)(pC1)
	#endif
	madd.s	rC32, rC32, ra3, rB2
	#if KB > 20
					lwc1	rB2, 80(pB2)
	#elif KB == 20
					swc1	rC02, -CMUL(16)(pC2)
	#endif
	madd.s	rC03, rC03, re0, rB3
	#if KB == 20
					swc1	rC12, -CMUL(12)(pC2)
	#endif
	madd.s	rC13, rC13, ra1, rB3
	#if KB > 20
					lwc1	rA3, 80(pA3)
	#elif KB == 20
					swc1	rC22, -CMUL(8)(pC2)
	#endif
	madd.s	rC23, rC23, ra2, rB3
	#if KB == 20
					swc1	rC32, -CMUL(4)(pC2)
	#endif
	madd.s	rC33, rC33, ra3, rB3
	#if KB > 20
					lwc1	rB3, 80(pB3)
	#endif
#endif
#if KB > 20
	#if KB > 21
					lwc1	ra1, 84(pA1)
	#elif KB == 21
					daddu	pA0, pA0, incAm
	#endif
	madd.s	rC00, rC00, rA0, rB0
	#if KB == 21
					daddu	pA1, pA1, incAm
	#endif
	madd.s	rC10, rC10, rA1, rB0
	#if KB > 23
					lwc1	re0, 92(pA0)
	#elif KB == 21
					daddu	pA2, pA2, incAm
	#endif
	madd.s	rC20, rC20, rE2, rB0
	#if KB == 21
					daddu	pA3, pA3, incAm
	#endif
	madd.s	rC30, rC30, rA3, rB0
	#if KB > 21
					lwc1	rb0, 84(pB0)
	#elif KB == 21
					swc1	rC00, -CMUL(16)(pC0)
	#endif
	madd.s	rC01, rC01, rA0, rB1
	#if KB == 21
					swc1	rC10, -CMUL(12)(pC0)
	#endif
	madd.s	rC11, rC11, rA1, rB1
	#if KB == 21
					swc1	rC20, -CMUL(8)(pC0)
	#endif
	madd.s	rC21, rC21, rE2, rB1
	#if KB == 21
					swc1	rC30, -CMUL(4)(pC0)
	#endif
	madd.s	rC31, rC31, rA3, rB1
	#if KB > 21
					lwc1	rB1, 84(pB1)
	#elif KB == 21
					swc1	rC01, -CMUL(16)(pC1)
	#endif
	madd.s	rC02, rC02, rA0, rB2
	#if KB == 21
					swc1	rC11, -CMUL(12)(pC1)
	#endif
	madd.s	rC12, rC12, rA1, rB2
	#if KB > 22
					lwc1	ra2, 88(pA2)
	#elif KB == 21
					swc1	rC21, -CMUL(8)(pC1)
	#endif
	madd.s	rC22, rC22, rE2, rB2
	#if KB == 21
					swc1	rC31, -CMUL(4)(pC1)
	#endif
	madd.s	rC32, rC32, rA3, rB2
	#if KB > 21
					lwc1	rB2, 84(pB2)
	#elif KB == 21
					swc1	rC02, -CMUL(16)(pC2)
	#endif
	madd.s	rC03, rC03, rA0, rB3
	#if KB == 21
					swc1	rC12, -CMUL(12)(pC2)
	#endif
	madd.s	rC13, rC13, rA1, rB3
	#if KB > 21
					lwc1	ra3, 84(pA3)
	#elif KB == 21
					swc1	rC22, -CMUL(8)(pC2)
	#endif
	madd.s	rC23, rC23, rE2, rB3
	#if KB == 21
					swc1	rC32, -CMUL(4)(pC2)
	#endif
	madd.s	rC33, rC33, rA3, rB3
	#if KB > 21
					lwc1	rB3, 84(pB3)
	#endif
#endif
#if KB > 21
	#if KB > 22
					lwc1	rA1, 88(pA1)
	#elif KB == 22
					daddu	pA0, pA0, incAm
	#endif
	madd.s	rC00, rC00, ra0, rb0
	#if KB == 22
					daddu	pA1, pA1, incAm
	#endif
	madd.s	rC10, rC10, ra1, rb0
	#if KB > 24
					lwc1	rA0, 96(pA0)
	#elif KB == 22
					daddu	pA2, pA2, incAm
	#endif
	madd.s	rC20, rC20, rA2, rb0
	#if KB == 22
					daddu	pA3, pA3, incAm
	#endif
	madd.s	rC30, rC30, ra3, rb0
	#if KB > 22
					lwc1	rB0, 88(pB0)
	#elif KB == 22
					swc1	rC00, -CMUL(16)(pC0)
	#endif
	madd.s	rC01, rC01, ra0, rB1
	#if KB == 22
					swc1	rC10, -CMUL(12)(pC0)
	#endif
	madd.s	rC11, rC11, ra1, rB1
	#if KB == 22
					swc1	rC20, -CMUL(8)(pC0)
	#endif
	madd.s	rC21, rC21, rA2, rB1
	#if KB == 22
					swc1	rC30, -CMUL(4)(pC0)
	#endif
	madd.s	rC31, rC31, ra3, rB1
	#if KB > 22
					lwc1	rB1, 88(pB1)
	#elif KB == 22
					swc1	rC01, -CMUL(16)(pC1)
	#endif
	madd.s	rC02, rC02, ra0, rB2
	#if KB == 22
					swc1	rC11, -CMUL(12)(pC1)
	#endif
	madd.s	rC12, rC12, ra1, rB2
	#if KB > 23
					lwc1	rE2, 92(pA2)
	#elif KB == 22
					swc1	rC21, -CMUL(8)(pC1)
	#endif
	madd.s	rC22, rC22, rA2, rB2
	#if KB == 22
					swc1	rC31, -CMUL(4)(pC1)
	#endif
	madd.s	rC32, rC32, ra3, rB2
	#if KB > 22
					lwc1	rB2, 88(pB2)
	#elif KB == 22
					swc1	rC02, -CMUL(16)(pC2)
	#endif
	madd.s	rC03, rC03, ra0, rB3
	#if KB == 22
					swc1	rC12, -CMUL(12)(pC2)
	#endif
	madd.s	rC13, rC13, ra1, rB3
	#if KB > 22
					lwc1	rA3, 88(pA3)
	#elif KB == 22
					swc1	rC22, -CMUL(8)(pC2)
	#endif
	madd.s	rC23, rC23, rA2, rB3
	#if KB == 22
					swc1	rC32, -CMUL(4)(pC2)
	#endif
	madd.s	rC33, rC33, ra3, rB3
	#if KB > 22
					lwc1	rB3, 88(pB3)
	#endif
#endif
#if KB > 22
	#if KB > 23
					lwc1	ra1, 92(pA1)
	#elif KB == 23
					daddu	pA0, pA0, incAm
	#endif
	madd.s	rC00, rC00, rE0, rB0
	#if KB == 23
					daddu	pA1, pA1, incAm
	#endif
	madd.s	rC10, rC10, rA1, rB0
	#if KB > 25
					lwc1	ra0, 100(pA0)
	#elif KB == 23
					daddu	pA2, pA2, incAm
	#endif
	madd.s	rC20, rC20, ra2, rB0
	#if KB == 23
					daddu	pA3, pA3, incAm
	#endif
	madd.s	rC30, rC30, rA3, rB0
	#if KB > 23
					lwc1	rb0, 92(pB0)
	#elif KB == 23
					swc1	rC00, -CMUL(16)(pC0)
	#endif
	madd.s	rC01, rC01, rE0, rB1
	#if KB == 23
					swc1	rC10, -CMUL(12)(pC0)
	#endif
	madd.s	rC11, rC11, rA1, rB1
	#if KB == 23
					swc1	rC20, -CMUL(8)(pC0)
	#endif
	madd.s	rC21, rC21, ra2, rB1
	#if KB == 23
					swc1	rC30, -CMUL(4)(pC0)
	#endif
	madd.s	rC31, rC31, rA3, rB1
	#if KB > 23
					lwc1	rB1, 92(pB1)
	#elif KB == 23
					swc1	rC01, -CMUL(16)(pC1)
	#endif
	madd.s	rC02, rC02, rE0, rB2
	#if KB == 23
					swc1	rC11, -CMUL(12)(pC1)
	#endif
	madd.s	rC12, rC12, rA1, rB2
	#if KB > 24
					lwc1	rA2, 96(pA2)
	#elif KB == 23
					swc1	rC21, -CMUL(8)(pC1)
	#endif
	madd.s	rC22, rC22, ra2, rB2
	#if KB == 23
					swc1	rC31, -CMUL(4)(pC1)
	#endif
	madd.s	rC32, rC32, rA3, rB2
	#if KB > 23
					lwc1	rB2, 92(pB2)
	#elif KB == 23
					swc1	rC02, -CMUL(16)(pC2)
	#endif
	madd.s	rC03, rC03, rE0, rB3
	#if KB == 23
					swc1	rC12, -CMUL(12)(pC2)
	#endif
	madd.s	rC13, rC13, rA1, rB3
	#if KB > 23
					lwc1	ra3, 92(pA3)
	#elif KB == 23
					swc1	rC22, -CMUL(8)(pC2)
	#endif
	madd.s	rC23, rC23, ra2, rB3
	#if KB == 23
					swc1	rC32, -CMUL(4)(pC2)
	#endif
	madd.s	rC33, rC33, rA3, rB3
	#if KB > 23
					lwc1	rB3, 92(pB3)
	#endif
#endif
#if KB > 23
	#if KB > 24
					lwc1	rA1, 96(pA1)
	#elif KB == 24
					daddu	pA0, pA0, incAm
	#endif
	madd.s	rC00, rC00, re0, rb0
	#if KB == 24
					daddu	pA1, pA1, incAm
	#endif
	madd.s	rC10, rC10, ra1, rb0
	#if KB > 26
					lwc1	rE0, 104(pA0)
	#elif KB == 24
					daddu	pA2, pA2, incAm
	#endif
	madd.s	rC20, rC20, rE2, rb0
	#if KB == 24
					daddu	pA3, pA3, incAm
	#endif
	madd.s	rC30, rC30, ra3, rb0
	#if KB > 24
					lwc1	rB0, 96(pB0)
	#elif KB == 24
					swc1	rC00, -CMUL(16)(pC0)
	#endif
	madd.s	rC01, rC01, re0, rB1
	#if KB == 24
					swc1	rC10, -CMUL(12)(pC0)
	#endif
	madd.s	rC11, rC11, ra1, rB1
	#if KB == 24
					swc1	rC20, -CMUL(8)(pC0)
	#endif
	madd.s	rC21, rC21, rE2, rB1
	#if KB == 24
					swc1	rC30, -CMUL(4)(pC0)
	#endif
	madd.s	rC31, rC31, ra3, rB1
	#if KB > 24
					lwc1	rB1, 96(pB1)
	#elif KB == 24
					swc1	rC01, -CMUL(16)(pC1)
	#endif
	madd.s	rC02, rC02, re0, rB2
	#if KB == 24
					swc1	rC11, -CMUL(12)(pC1)
	#endif
	madd.s	rC12, rC12, ra1, rB2
	#if KB > 25
					lwc1	ra2, 100(pA2)
	#elif KB == 24
					swc1	rC21, -CMUL(8)(pC1)
	#endif
	madd.s	rC22, rC22, rE2, rB2
	#if KB == 24
					swc1	rC31, -CMUL(4)(pC1)
	#endif
	madd.s	rC32, rC32, ra3, rB2
	#if KB > 24
					lwc1	rB2, 96(pB2)
	#elif KB == 24
					swc1	rC02, -CMUL(16)(pC2)
	#endif
	madd.s	rC03, rC03, re0, rB3
	#if KB == 24
					swc1	rC12, -CMUL(12)(pC2)
	#endif
	madd.s	rC13, rC13, ra1, rB3
	#if KB > 24
					lwc1	rA3, 96(pA3)
	#elif KB == 24
					swc1	rC22, -CMUL(8)(pC2)
	#endif
	madd.s	rC23, rC23, rE2, rB3
	#if KB == 24
					swc1	rC32, -CMUL(4)(pC2)
	#endif
	madd.s	rC33, rC33, ra3, rB3
	#if KB > 24
					lwc1	rB3, 96(pB3)
	#endif
#endif
#if KB > 24
	#if KB > 25
					lwc1	ra1, 100(pA1)
	#elif KB == 25
					daddu	pA0, pA0, incAm
	#endif
	madd.s	rC00, rC00, rA0, rB0
	#if KB == 25
					daddu	pA1, pA1, incAm
	#endif
	madd.s	rC10, rC10, rA1, rB0
	#if KB > 27
					lwc1	re0, 108(pA0)
	#elif KB == 25
					daddu	pA2, pA2, incAm
	#endif
	madd.s	rC20, rC20, rA2, rB0
	#if KB == 25
					daddu	pA3, pA3, incAm
	#endif
	madd.s	rC30, rC30, rA3, rB0
	#if KB > 25
					lwc1	rb0, 100(pB0)
	#elif KB == 25
					swc1	rC00, -CMUL(16)(pC0)
	#endif
	madd.s	rC01, rC01, rA0, rB1
	#if KB == 25
					swc1	rC10, -CMUL(12)(pC0)
	#endif
	madd.s	rC11, rC11, rA1, rB1
	#if KB == 25
					swc1	rC20, -CMUL(8)(pC0)
	#endif
	madd.s	rC21, rC21, rA2, rB1
	#if KB == 25
					swc1	rC30, -CMUL(4)(pC0)
	#endif
	madd.s	rC31, rC31, rA3, rB1
	#if KB > 25
					lwc1	rB1, 100(pB1)
	#elif KB == 25
					swc1	rC01, -CMUL(16)(pC1)
	#endif
	madd.s	rC02, rC02, rA0, rB2
	#if KB == 25
					swc1	rC11, -CMUL(12)(pC1)
	#endif
	madd.s	rC12, rC12, rA1, rB2
	#if KB > 26
					lwc1	rE2, 104(pA2)
	#elif KB == 25
					swc1	rC21, -CMUL(8)(pC1)
	#endif
	madd.s	rC22, rC22, rA2, rB2
	#if KB == 25
					swc1	rC31, -CMUL(4)(pC1)
	#endif
	madd.s	rC32, rC32, rA3, rB2
	#if KB > 25
					lwc1	rB2, 100(pB2)
	#elif KB == 25
					swc1	rC02, -CMUL(16)(pC2)
	#endif
	madd.s	rC03, rC03, rA0, rB3
	#if KB == 25
					swc1	rC12, -CMUL(12)(pC2)
	#endif
	madd.s	rC13, rC13, rA1, rB3
	#if KB > 25
					lwc1	ra3, 100(pA3)
	#elif KB == 25
					swc1	rC22, -CMUL(8)(pC2)
	#endif
	madd.s	rC23, rC23, rA2, rB3
	#if KB == 25
					swc1	rC32, -CMUL(4)(pC2)
	#endif
	madd.s	rC33, rC33, rA3, rB3
	#if KB > 25
					lwc1	rB3, 100(pB3)
	#endif
#endif
#if KB > 25
	#if KB > 26
					lwc1	rA1, 104(pA1)
	#elif KB == 26
					daddu	pA0, pA0, incAm
	#endif
	madd.s	rC00, rC00, ra0, rb0
	#if KB == 26
					daddu	pA1, pA1, incAm
	#endif
	madd.s	rC10, rC10, ra1, rb0
	#if KB > 28
					lwc1	rA0, 112(pA0)
	#elif KB == 26
					daddu	pA2, pA2, incAm
	#endif
	madd.s	rC20, rC20, ra2, rb0
	#if KB == 26
					daddu	pA3, pA3, incAm
	#endif
	madd.s	rC30, rC30, ra3, rb0
	#if KB > 26
					lwc1	rB0, 104(pB0)
	#elif KB == 26
					swc1	rC00, -CMUL(16)(pC0)
	#endif
	madd.s	rC01, rC01, ra0, rB1
	#if KB == 26
					swc1	rC10, -CMUL(12)(pC0)
	#endif
	madd.s	rC11, rC11, ra1, rB1
	#if KB == 26
					swc1	rC20, -CMUL(8)(pC0)
	#endif
	madd.s	rC21, rC21, ra2, rB1
	#if KB == 26
					swc1	rC30, -CMUL(4)(pC0)
	#endif
	madd.s	rC31, rC31, ra3, rB1
	#if KB > 26
					lwc1	rB1, 104(pB1)
	#elif KB == 26
					swc1	rC01, -CMUL(16)(pC1)
	#endif
	madd.s	rC02, rC02, ra0, rB2
	#if KB == 26
					swc1	rC11, -CMUL(12)(pC1)
	#endif
	madd.s	rC12, rC12, ra1, rB2
	#if KB > 27
					lwc1	rA2, 108(pA2)
	#elif KB == 26
					swc1	rC21, -CMUL(8)(pC1)
	#endif
	madd.s	rC22, rC22, ra2, rB2
	#if KB == 26
					swc1	rC31, -CMUL(4)(pC1)
	#endif
	madd.s	rC32, rC32, ra3, rB2
	#if KB > 26
					lwc1	rB2, 104(pB2)
	#elif KB == 26
					swc1	rC02, -CMUL(16)(pC2)
	#endif
	madd.s	rC03, rC03, ra0, rB3
	#if KB == 26
					swc1	rC12, -CMUL(12)(pC2)
	#endif
	madd.s	rC13, rC13, ra1, rB3
	#if KB > 26
					lwc1	rA3, 104(pA3)
	#elif KB == 26
					swc1	rC22, -CMUL(8)(pC2)
	#endif
	madd.s	rC23, rC23, ra2, rB3
	#if KB == 26
					swc1	rC32, -CMUL(4)(pC2)
	#endif
	madd.s	rC33, rC33, ra3, rB3
	#if KB > 26
					lwc1	rB3, 104(pB3)
	#endif
#endif
#if KB > 26
	#if KB > 27
					lwc1	ra1, 108(pA1)
	#elif KB == 27
					daddu	pA0, pA0, incAm
	#endif
	madd.s	rC00, rC00, rE0, rB0
	#if KB == 27
					daddu	pA1, pA1, incAm
	#endif
	madd.s	rC10, rC10, rA1, rB0
	#if KB > 29
					lwc1	ra0, 116(pA0)
	#elif KB == 27
					daddu	pA2, pA2, incAm
	#endif
	madd.s	rC20, rC20, rE2, rB0
	#if KB == 27
					daddu	pA3, pA3, incAm
	#endif
	madd.s	rC30, rC30, rA3, rB0
	#if KB > 27
					lwc1	rb0, 108(pB0)
	#elif KB == 27
					swc1	rC00, -CMUL(16)(pC0)
	#endif
	madd.s	rC01, rC01, rE0, rB1
	#if KB == 27
					swc1	rC10, -CMUL(12)(pC0)
	#endif
	madd.s	rC11, rC11, rA1, rB1
	#if KB == 27
					swc1	rC20, -CMUL(8)(pC0)
	#endif
	madd.s	rC21, rC21, rE2, rB1
	#if KB == 27
					swc1	rC30, -CMUL(4)(pC0)
	#endif
	madd.s	rC31, rC31, rA3, rB1
	#if KB > 27
					lwc1	rB1, 108(pB1)
	#elif KB == 27
					swc1	rC01, -CMUL(16)(pC1)
	#endif
	madd.s	rC02, rC02, rE0, rB2
	#if KB == 27
					swc1	rC11, -CMUL(12)(pC1)
	#endif
	madd.s	rC12, rC12, rA1, rB2
	#if KB > 28
					lwc1	ra2, 112(pA2)
	#elif KB == 27
					swc1	rC21, -CMUL(8)(pC1)