atl_dmm6x1x72_sse2.c

来自「基于Blas CLapck的.用过的人知道是干啥的」· C语言 代码 · 共 3,260 行 · 第 1/5 页

C
3,260
字号
	addpd	rB0, rC5#endif#if KB > 6	movapd	48-120(pB0), rB0	movapd	48-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	48-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	48-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	48-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	48-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	48-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 8	movapd	64-120(pB0), rB0	movapd	64-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	64-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	64-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	64-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	64-120(pA3,ldab), rA0	mulpd	rB0, rA0						prefB(256(pB0,ldab))   #if KB > 48						prefB(384(pB0,ldab))   #endif	addpd	rA0, rC4	mulpd	64-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 10	movapd	80-120(pB0), rB0	movapd	80-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	80-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	80-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	80-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	80-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	80-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 12	movapd	96-120(pB0), rB0	movapd	96-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	96-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	96-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	96-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	96-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	96-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 14	movapd	112-120(pB0), rB0	movapd	112-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	112-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	112-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	112-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	112-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	112-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 16	movapd	128-120(pB0), rB0	movapd	128-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	128-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	128-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	128-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	128-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	128-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 18	movapd	144-120(pB0), rB0	movapd	144-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	144-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	144-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	144-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	144-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	144-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 20	movapd	160-120(pB0), rB0	movapd	160-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	160-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	160-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	160-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	160-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	160-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 22	movapd	176-120(pB0), rB0	movapd	176-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	176-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	176-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	176-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	176-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	176-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 24	movapd	192-120(pB0), rB0	movapd	192-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	192-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	192-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	192-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	192-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	192-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 26	movapd	208-120(pB0), rB0	movapd	208-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	208-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	208-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	208-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	208-120(pA3,ldab), rA0	mulpd	rB0, rA0   #if KB > 56						prefB(448(pB0,ldab))   #endif	addpd	rA0, rC4	mulpd	208-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 28	movapd	224-120(pB0), rB0	movapd	224-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	224-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	224-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	224-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	224-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	224-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 30	movapd	240-120(pB0), rB0	movapd	240-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	240-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	240-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	240-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	240-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	240-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 32	movapd	256-120(pB0), rB0	movapd	256-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	256-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	256-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	256-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	256-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	256-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 34	movapd	272-120(pB0), rB0	movapd	272-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	272-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	272-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	272-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	272-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	272-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 36	movapd	288-120(pB0), rB0	movapd	288-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	288-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	288-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	288-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	288-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	288-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 38	movapd	304-120(pB0), rB0	movapd	304-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	304-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	304-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	304-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	304-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	304-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 40	movapd	320-120(pB0), rB0	movapd	320-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	320-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	320-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	320-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	320-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	320-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 42	movapd	336-120(pB0), rB0	movapd	336-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	336-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	336-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	336-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	336-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	336-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 44	movapd	352-120(pB0), rB0	movapd	352-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	352-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	352-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	352-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	352-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	352-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 46	movapd	368-120(pB0), rB0	movapd	368-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	368-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	368-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	368-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	368-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	368-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 48	movapd	384-120(pB0), rB0	movapd	384-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	384-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	384-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	384-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	384-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	384-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 50	movapd	400-120(pB0), rB0	movapd	400-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	400-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	400-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	400-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	400-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	400-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 52	movapd	416-120(pB0), rB0	movapd	416-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	416-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	416-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	416-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	416-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	416-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 54	movapd	432-120(pB0), rB0	movapd	432-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	432-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	432-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	432-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	432-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	432-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 56	movapd	448-120(pB0), rB0	movapd	448-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	448-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	448-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	448-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	448-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	448-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 58	movapd	464-120(pB0), rB0	movapd	464-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	464-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	464-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	464-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	464-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	464-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 60	movapd	480-120(pB0), rB0	movapd	480-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	480-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	480-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	480-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	480-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	480-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 62	movapd	496-120(pB0), rB0	movapd	496-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	496-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	496-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	496-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	496-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	496-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 64	movapd	512-120(pB0), rB0	movapd	512-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	512-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	512-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	512-120(pA3), rA0	mulpd	rB0, rA0	addpd	rA0, rC3	movapd	512-120(pA3,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC4	mulpd	512-120(pA3,ldab,2), rB0	addpd	rB0, rC5#endif#if KB > 66	movapd	528-120(pB0), rB0	movapd	528-120(pA0), rA0	mulpd	rB0, rA0	addpd	rA0, rC0	movapd	528-120(pA0,ldab), rA0	mulpd	rB0, rA0	addpd	rA0, rC1	movapd	528-120(pA0,ldab,2), rA0	mulpd	rB0, rA0	addpd	rA0, rC2	movapd	528-120(pA3), rA0

⌨️ 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?