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

C
3,232
字号
					swc1	rC10, -CMUL(12)(pC0)	#elif !defined(BETA0)		prefC(16(pC2))	#endif	madd.s	rC11, rC11, ra1, rB1	#if KB == 4					swc1	rC20, -CMUL(8)(pC0)	#elif !defined(BETA0)		prefC(16(pC3))	#endif	madd.s	rC21, rC21, rA2, rB1	#if KB == 4					swc1	rC30, -CMUL(4)(pC0)	#endif	madd.s	rC31, rC31, ra3, rB1	#if KB > 4					lwc1	rB1, 16(pB1)	#elif KB == 4					swc1	rC01, -CMUL(16)(pC1)	#endif	madd.s	rC02, rC02, re0, rB2	#if KB == 4					swc1	rC11, -CMUL(12)(pC1)	#endif	madd.s	rC12, rC12, ra1, rB2	#if KB > 5					lwc1	rE2, 20(pA2)	#elif KB == 4					swc1	rC21, -CMUL(8)(pC1)	#endif	madd.s	rC22, rC22, rA2, rB2	#if KB == 4					swc1	rC31, -CMUL(4)(pC1)	#endif	madd.s	rC32, rC32, ra3, rB2	#if KB > 4					lwc1	rB2, 16(pB2)	#elif KB == 4					swc1	rC02, -CMUL(16)(pC2)	#endif	madd.s	rC03, rC03, re0, rB3	#if KB == 4					swc1	rC12, -CMUL(12)(pC2)	#endif	madd.s	rC13, rC13, ra1, rB3	#if KB > 4					lwc1	rA3, 16(pA3)	#elif KB == 4					swc1	rC22, -CMUL(8)(pC2)	#endif	madd.s	rC23, rC23, rA2, rB3	#if KB == 4					swc1	rC32, -CMUL(4)(pC2)	#endif	madd.s	rC33, rC33, ra3, rB3	#if KB > 4					lwc1	rB3, 16(pB3)	#endif#endif#if KB > 4	#if KB > 5					lwc1	ra1, 20(pA1)	#elif KB == 5					daddu	pA0, pA0, incAm	#endif	madd.s	rC00, rC00, rA0, rB0	#if KB == 5					daddu	pA1, pA1, incAm	#endif	madd.s	rC10, rC10, rA1, rB0	#if KB > 7					lwc1	re0, 28(pA0)	#elif KB == 5					daddu	pA2, pA2, incAm	#endif	madd.s	rC20, rC20, ra2, rB0	#if KB == 5					daddu	pA3, pA3, incAm	#endif	madd.s	rC30, rC30, rA3, rB0	#if KB > 5					lwc1	rb0, 20(pB0)	#elif KB == 5					swc1	rC00, -CMUL(16)(pC0)	#endif	madd.s	rC01, rC01, rA0, rB1	#if KB == 5					swc1	rC10, -CMUL(12)(pC0)	#endif	madd.s	rC11, rC11, rA1, rB1	#if KB == 5					swc1	rC20, -CMUL(8)(pC0)	#endif	madd.s	rC21, rC21, ra2, rB1	#if KB == 5					swc1	rC30, -CMUL(4)(pC0)	#endif	madd.s	rC31, rC31, rA3, rB1	#if KB > 5					lwc1	rB1, 20(pB1)	#elif KB == 5					swc1	rC01, -CMUL(16)(pC1)	#endif	madd.s	rC02, rC02, rA0, rB2	#if KB == 5					swc1	rC11, -CMUL(12)(pC1)	#endif	madd.s	rC12, rC12, rA1, rB2	#if KB > 6					lwc1	rA2, 24(pA2)	#elif KB == 5					swc1	rC21, -CMUL(8)(pC1)	#endif	madd.s	rC22, rC22, ra2, rB2	#if KB == 5					swc1	rC31, -CMUL(4)(pC1)	#endif	madd.s	rC32, rC32, rA3, rB2	#if KB > 5					lwc1	rB2, 20(pB2)	#elif KB == 5					swc1	rC02, -CMUL(16)(pC2)	#endif	madd.s	rC03, rC03, rA0, rB3	#if KB == 5					swc1	rC12, -CMUL(12)(pC2)	#endif	madd.s	rC13, rC13, rA1, rB3	#if KB > 5					lwc1	ra3, 20(pA3)	#elif KB == 5					swc1	rC22, -CMUL(8)(pC2)	#endif	madd.s	rC23, rC23, ra2, rB3	#if KB == 5					swc1	rC32, -CMUL(4)(pC2)	#endif	madd.s	rC33, rC33, rA3, rB3	#if KB > 5					lwc1	rB3, 20(pB3)	#endif#endif#if KB > 5	#if KB > 6					lwc1	rA1, 24(pA1)	#elif KB == 6					daddu	pA0, pA0, incAm	#endif	madd.s	rC00, rC00, ra0, rb0	#if KB == 6					daddu	pA1, pA1, incAm	#endif	madd.s	rC10, rC10, ra1, rb0	#if KB > 8					lwc1	rA0, 32(pA0)	#elif KB == 6					daddu	pA2, pA2, incAm	#endif	madd.s	rC20, rC20, rE2, rb0	#if KB == 6					daddu	pA3, pA3, incAm	#endif	madd.s	rC30, rC30, ra3, rb0	#if KB > 6					lwc1	rB0, 24(pB0)	#elif KB == 6					swc1	rC00, -CMUL(16)(pC0)	#endif	madd.s	rC01, rC01, ra0, rB1	#if KB == 6					swc1	rC10, -CMUL(12)(pC0)	#endif	madd.s	rC11, rC11, ra1, rB1	#if KB == 6					swc1	rC20, -CMUL(8)(pC0)	#endif	madd.s	rC21, rC21, rE2, rB1	#if KB == 6					swc1	rC30, -CMUL(4)(pC0)	#endif	madd.s	rC31, rC31, ra3, rB1	#if KB > 6					lwc1	rB1, 24(pB1)	#elif KB == 6					swc1	rC01, -CMUL(16)(pC1)	#endif	madd.s	rC02, rC02, ra0, rB2	#if KB == 6					swc1	rC11, -CMUL(12)(pC1)	#endif	madd.s	rC12, rC12, ra1, rB2	#if KB > 7					lwc1	ra2, 28(pA2)	#elif KB == 6					swc1	rC21, -CMUL(8)(pC1)	#endif	madd.s	rC22, rC22, rE2, rB2	#if KB == 6					swc1	rC31, -CMUL(4)(pC1)	#endif	madd.s	rC32, rC32, ra3, rB2	#if KB > 6					lwc1	rB2, 24(pB2)	#elif KB == 6					swc1	rC02, -CMUL(16)(pC2)	#endif	madd.s	rC03, rC03, ra0, rB3	#if KB == 6					swc1	rC12, -CMUL(12)(pC2)	#endif	madd.s	rC13, rC13, ra1, rB3	#if KB > 6					lwc1	rA3, 24(pA3)	#elif KB == 6					swc1	rC22, -CMUL(8)(pC2)	#endif	madd.s	rC23, rC23, rE2, rB3	#if KB == 6					swc1	rC32, -CMUL(4)(pC2)	#endif	madd.s	rC33, rC33, ra3, rB3	#if KB > 6					lwc1	rB3, 24(pB3)	#endif#endif#if KB > 6	#if KB > 7					lwc1	ra1, 28(pA1)	#elif KB == 7					daddu	pA0, pA0, incAm	#endif	madd.s	rC00, rC00, rE0, rB0	#if KB == 7					daddu	pA1, pA1, incAm	#endif	madd.s	rC10, rC10, rA1, rB0	#if KB > 9					lwc1	ra0, 36(pA0)	#elif KB == 7					daddu	pA2, pA2, incAm	#endif	madd.s	rC20, rC20, rA2, rB0	#if KB == 7					daddu	pA3, pA3, incAm	#endif	madd.s	rC30, rC30, rA3, rB0	#if KB > 7					lwc1	rb0, 28(pB0)	#elif KB == 7					swc1	rC00, -CMUL(16)(pC0)	#endif	madd.s	rC01, rC01, rE0, rB1	#if KB == 7					swc1	rC10, -CMUL(12)(pC0)	#endif	madd.s	rC11, rC11, rA1, rB1	#if KB == 7					swc1	rC20, -CMUL(8)(pC0)	#endif	madd.s	rC21, rC21, rA2, rB1	#if KB == 7					swc1	rC30, -CMUL(4)(pC0)	#endif	madd.s	rC31, rC31, rA3, rB1	#if KB > 7					lwc1	rB1, 28(pB1)	#elif KB == 7					swc1	rC01, -CMUL(16)(pC1)	#endif	madd.s	rC02, rC02, rE0, rB2	#if KB == 7					swc1	rC11, -CMUL(12)(pC1)	#endif	madd.s	rC12, rC12, rA1, rB2	#if KB > 8					lwc1	rE2, 32(pA2)	#elif KB == 7					swc1	rC21, -CMUL(8)(pC1)	#endif	madd.s	rC22, rC22, rA2, rB2	#if KB == 7					swc1	rC31, -CMUL(4)(pC1)	#endif	madd.s	rC32, rC32, rA3, rB2	#if KB > 7					lwc1	rB2, 28(pB2)	#elif KB == 7					swc1	rC02, -CMUL(16)(pC2)	#endif	madd.s	rC03, rC03, rE0, rB3	#if KB == 7					swc1	rC12, -CMUL(12)(pC2)	#endif	madd.s	rC13, rC13, rA1, rB3	#if KB > 7					lwc1	ra3, 28(pA3)	#elif KB == 7					swc1	rC22, -CMUL(8)(pC2)	#endif	madd.s	rC23, rC23, rA2, rB3	#if KB == 7					swc1	rC32, -CMUL(4)(pC2)	#endif	madd.s	rC33, rC33, rA3, rB3	#if KB > 7					lwc1	rB3, 28(pB3)	#endif#endif#if KB > 7	#if KB > 8					lwc1	rA1, 32(pA1)	#elif KB == 8					daddu	pA0, pA0, incAm	#endif	madd.s	rC00, rC00, re0, rb0	#if KB == 8					daddu	pA1, pA1, incAm	#endif	madd.s	rC10, rC10, ra1, rb0	#if KB > 10					lwc1	rE0, 40(pA0)	#elif KB == 8					daddu	pA2, pA2, incAm	#endif	madd.s	rC20, rC20, ra2, rb0	#if KB == 8					daddu	pA3, pA3, incAm	#endif	madd.s	rC30, rC30, ra3, rb0	#if KB > 8					lwc1	rB0, 32(pB0)	#elif KB == 8					swc1	rC00, -CMUL(16)(pC0)	#endif	madd.s	rC01, rC01, re0, rB1	#if KB == 8					swc1	rC10, -CMUL(12)(pC0)	#endif	madd.s	rC11, rC11, ra1, rB1	#if KB == 8					swc1	rC20, -CMUL(8)(pC0)	#endif	madd.s	rC21, rC21, ra2, rB1	#if KB == 8					swc1	rC30, -CMUL(4)(pC0)	#endif	madd.s	rC31, rC31, ra3, rB1	#if KB > 8					lwc1	rB1, 32(pB1)	#elif KB == 8					swc1	rC01, -CMUL(16)(pC1)	#endif	madd.s	rC02, rC02, re0, rB2	#if KB == 8					swc1	rC11, -CMUL(12)(pC1)	#endif					prefB(0(pfB))	madd.s	rC12, rC12, ra1, rB2	#if KB > 9					lwc1	rA2, 36(pA2)	#elif KB == 8					swc1	rC21, -CMUL(8)(pC1)	#endif	madd.s	rC22, rC22, ra2, rB2	#if KB == 8					swc1	rC31, -CMUL(4)(pC1)	#endif	madd.s	rC32, rC32, ra3, rB2	#if KB > 8					lwc1	rB2, 32(pB2)	#elif KB == 8					swc1	rC02, -CMUL(16)(pC2)	#endif	madd.s	rC03, rC03, re0, rB3	#if KB == 8					swc1	rC12, -CMUL(12)(pC2)	#endif	madd.s	rC13, rC13, ra1, rB3	#if KB > 8					lwc1	rA3, 32(pA3)	#elif KB == 8					swc1	rC22, -CMUL(8)(pC2)	#endif	madd.s	rC23, rC23, ra2, rB3	#if KB == 8					swc1	rC32, -CMUL(4)(pC2)	#endif	madd.s	rC33, rC33, ra3, rB3	#if KB > 8					lwc1	rB3, 32(pB3)	#endif#endif#if KB > 8	#if KB > 9					lwc1	ra1, 36(pA1)	#elif KB == 9					daddu	pA0, pA0, incAm	#endif	madd.s	rC00, rC00, rA0, rB0	#if KB == 9					daddu	pA1, pA1, incAm	#endif					prefB(32(pfB))	madd.s	rC10, rC10, rA1, rB0	#if KB > 11					lwc1	re0, 44(pA0)	#elif KB == 9					daddu	pA2, pA2, incAm	#endif	madd.s	rC20, rC20, rE2, rB0	#if KB == 9					daddu	pA3, pA3, incAm	#endif					daddiu  pfB, pfB, 64	madd.s	rC30, rC30, rA3, rB0	#if KB > 9					lwc1	rb0, 36(pB0)	#elif KB == 9					swc1	rC00, -CMUL(16)(pC0)	#endif	madd.s	rC01, rC01, rA0, rB1	#if KB == 9					swc1	rC10, -CMUL(12)(pC0)	#endif	madd.s	rC11, rC11, rA1, rB1	#if KB == 9					swc1	rC20, -CMUL(8)(pC0)	#endif	madd.s	rC21, rC21, rE2, rB1	#if KB == 9					swc1	rC30, -CMUL(4)(pC0)	#endif	madd.s	rC31, rC31, rA3, rB1	#if KB > 9					lwc1	rB1, 36(pB1)	#elif KB == 9					swc1	rC01, -CMUL(16)(pC1)	#endif	madd.s	rC02, rC02, rA0, rB2	#if KB == 9					swc1	rC11, -CMUL(12)(pC1)	#endif	madd.s	rC12, rC12, rA1, rB2	#if KB > 10					lwc1	ra2, 40(pA2)	#elif KB == 9					swc1	rC21, -CMUL(8)(pC1)	#endif	madd.s	rC22, rC22, rE2, rB2	#if KB == 9					swc1	rC31, -CMUL(4)(pC1)	#endif	madd.s	rC32, rC32, rA3, rB2	#if KB > 9					lwc1	rB2, 36(pB2)	#elif KB == 9					swc1	rC02, -CMUL(16)(pC2)	#endif	madd.s	rC03, rC03, rA0, rB3	#if KB == 9					swc1	rC12, -CMUL(12)(pC2)	#endif	madd.s	rC13, rC13, rA1, rB3	#if KB > 9					lwc1	ra3, 36(pA3)	#elif KB == 9					swc1	rC22, -CMUL(8)(pC2)	#endif	madd.s	rC23, rC23, rE2, rB3	#if KB == 9					swc1	rC32, -CMUL(4)(pC2)	#endif	madd.s	rC33, rC33, rA3, rB3	#if KB > 9					lwc1	rB3, 36(pB3)	#endif#endif#if KB > 9	#if KB > 10					lwc1	rA1, 40(pA1)	#elif KB == 10					daddu	pA0, pA0, incAm	#endif	madd.s	rC00, rC00, ra0, rb0	#if KB == 10					daddu	pA1, pA1, incAm	#endif	madd.s	rC10, rC10, ra1, rb0	#if KB > 12					lwc1	rA0, 48(pA0)	#elif KB == 10					daddu	pA2, pA2, incAm	#endif	madd.s	rC20, rC20, rA2, rb0	#if KB == 10					daddu	pA3, pA3, incAm	#endif	madd.s	rC30, rC30, ra3, rb0	#if KB > 10					lwc1	rB0, 40(pB0)	#elif KB == 10					swc1	rC00, -CMUL(16)(pC0)	#endif	madd.s	rC01, rC01, ra0, rB1	#if KB == 10					swc1	rC10, -CMUL(12)(pC0)	#endif	madd.s	rC11, rC11, ra1, rB1	#if KB == 10					swc1	rC20, -CMUL(8)(pC0)	#endif	madd.s	rC21, rC21, rA2, rB1	#if KB == 10					swc1	rC30, -CMUL(4)(pC0)	#endif	madd.s	rC31, rC31, ra3, rB1	#if KB > 10					lwc1	rB1, 40(pB1)	#elif KB == 10					swc1	rC01, -CMUL(16)(pC1)	#endif	madd.s	rC02, rC02, ra0, rB2	#if KB == 10					swc1	rC11, -CMUL(12)(pC1)	#endif	madd.s	rC12, rC12, ra1, rB2	#if KB > 11					lwc1	rE2, 44(pA2)	#elif KB == 10					swc1	rC21, -CMUL(8)(pC1)	#endif	madd.s	rC22, rC22, rA2, rB2	#if KB == 10					swc1	rC31, -CMUL(4)(pC1)	#endif	madd.s	rC32, rC32, ra3, rB2	#if KB > 10					lwc1	rB2, 40(pB2)	#elif KB == 10					swc1	rC02, -CMUL(16)(pC2)	#endif	madd.s	rC03, rC03, ra0, rB3	#if KB == 10					swc1	rC12, -CMUL(12)(pC2)	#endif	madd.s	rC13, rC13, ra1, rB3	#if KB > 10					lwc1	rA3, 40(pA3)	#elif KB == 10					swc1	rC22, -CMUL(8)(pC2)	#endif	madd.s	rC23, rC23, rA2, rB3	#if KB == 10					swc1	rC32, -CMUL(4)(pC2)	#endif	madd.s	rC33, rC33, ra3, rB3	#if KB > 10					lwc1	rB3, 40(pB3)	#endif#endif#if KB > 10	#if KB > 11					lwc1	ra1, 44(pA1)	#elif KB == 11					daddu	pA0, pA0, incAm	#endif	madd.s	rC00, rC00, rE0, rB0	#if KB == 11					daddu	pA1, pA1, incAm	#endif	madd.s	rC10, rC10, rA1, rB0	#if KB > 13					lwc1	ra0, 52(pA0)	#elif KB == 11					daddu	pA2, pA2, incAm	#endif	madd.s	rC20, rC20, ra2, rB0	#if KB == 11					daddu	pA3, pA3, incAm	#endif	madd.s	rC30, rC30, rA3, rB0	#if KB > 11					lwc1	rb0, 44(pB0)	#elif KB == 11					swc1	rC00, -CMUL(16)(pC0)	#endif	madd.s	rC01, rC01, rE0, rB1	#if KB == 11					swc1	rC10, -CMUL(12)(pC0)	#endif	madd.s	rC11, rC11, rA1, rB1	#if KB == 11					swc1	rC20, -CMUL(8)(pC0)	#endif	madd.s	rC21, rC21, ra2, rB1	#if KB == 11					swc1	rC30, -CMUL(4)(pC0)	#endif	madd.s	rC31, rC31, rA3, rB1	#if KB > 11					lwc1	rB1, 44(pB1)	#elif KB == 11					swc1	rC01, -CMUL(16)(pC1)	#endif	madd.s	rC02, rC02, rE0, rB2	#if KB == 11					swc1	rC11, -CMUL(12)(pC1)	#endif	madd.s	rC12, rC12, rA1, rB2	#if KB > 12					lwc1	rA2, 48(pA2)	#elif KB == 11					swc1	rC21, -CMUL(8)(pC1)	#endif	madd.s	rC22, rC22, ra2, rB2	#if KB == 11					swc1	rC31, -CMUL(4)(pC1)	#endif	madd.s	rC32, rC32, rA3, rB2	#if KB > 11					lwc1	rB2, 44(pB2)	#elif KB == 11					swc1	rC02, -CMUL(16)(pC2)	#endif	madd.s	rC03, rC03, rE0, rB3	#if KB == 11					swc1	rC12, -CMUL(12)(pC2)	#endif	madd.s	rC13, rC13, rA1, rB3	#if KB > 11					lwc1	ra3, 44(pA3)	#elif KB == 11					swc1	rC22, -CMUL(8)(pC2)	#endif	madd.s	rC23, rC23, ra2, rB3	#if KB == 11					swc1	rC32, -CMUL(4)(pC2)	#endif	madd.s	rC33, rC33, rA3, rB3	#if KB > 11					lwc1	rB3, 44(pB3)	#endif#endif#if KB > 11	#if KB > 12					lwc1	rA1, 48(pA1)	#elif KB == 12

⌨️ 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?