atl_dmm8x1x120_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 3,266 行 · 第 1/5 页

C
3,266
字号
	addpd	rA0,rC02	movapd	16-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	16-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	16-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	16-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	16-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 4	movapd	32-128(pB0), rB0	movapd	32-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	32-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	32-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	32-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	32-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	32-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	32-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	32-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 6	movapd	48-128(pB0), rB0	movapd	48-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	48-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	48-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	48-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	48-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	48-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	48-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	48-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 8	movapd	64-128(pB0), rB0	movapd	64-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	64-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	64-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	64-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	64-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	64-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	64-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	64-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 10	movapd	80-128(pB0), rB0	movapd	80-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	80-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	80-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	80-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	80-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	80-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	80-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	80-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 12	movapd	96-128(pB0), rB0	movapd	96-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	96-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	96-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	96-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	96-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	96-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	96-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	96-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 14	movapd	112-128(pB0), rB0	movapd	112-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	112-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	112-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	112-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	112-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	112-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	112-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	112-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 16	movapd	128-128(pB0), rB0	movapd	128-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	128-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	128-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	128-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	128-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	128-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	128-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	128-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 18	movapd	144-128(pB0), rB0	movapd	144-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	144-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	144-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	144-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	144-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	144-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	144-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	144-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 20	movapd	160-128(pB0), rB0	movapd	160-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	160-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	160-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	160-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	160-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	160-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	160-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	160-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 22	movapd	176-128(pB0), rB0	movapd	176-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	176-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	176-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	176-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	176-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	176-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	176-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	176-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 24	movapd	192-128(pB0), rB0	movapd	192-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	192-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	192-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	192-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	192-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	192-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	192-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	192-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 26	movapd	208-128(pB0), rB0	movapd	208-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	208-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	208-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	208-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	208-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	208-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	208-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	208-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 28	movapd	224-128(pB0), rB0	movapd	224-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	224-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	224-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	224-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	224-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	224-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	224-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	224-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 30	movapd	240-128(pB0), rB0	movapd	240-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	240-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	240-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	240-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	240-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	240-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	240-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	240-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 32	movapd	256-128(pB0), rB0	movapd	256-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	256-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	256-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	256-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	256-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	256-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	256-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	256-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 34	movapd	272-128(pB0), rB0	movapd	272-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	272-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	272-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	272-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	272-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	272-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	272-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	272-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 36	movapd	288-128(pB0), rB0	movapd	288-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	288-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	288-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	288-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	288-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	288-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	288-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	288-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 38	movapd	304-128(pB0), rB0	movapd	304-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	304-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	304-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	304-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	304-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	304-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	304-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	304-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 40	movapd	320-128(pB0), rB0	movapd	320-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	320-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	320-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	320-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	320-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	320-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	320-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	320-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 42	movapd	336-128(pB0), rB0	movapd	336-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	336-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	336-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	336-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	336-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	336-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	336-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	336-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 44	movapd	352-128(pB0), rB0	movapd	352-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	352-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	352-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	352-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	352-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	352-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	352-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	352-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 46	movapd	368-128(pB0), rB0	movapd	368-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	368-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	368-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	368-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	368-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	368-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	368-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	368-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 48	movapd	384-128(pB0), rB0	movapd	384-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	384-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	384-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	384-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	384-128(pA0,lda,4), rA0	mulpd	rB0,rA0	addpd	rA0,rC04	movapd	384-128(pA0,lda5), rA0	mulpd	rB0,rA0	addpd	rA0,rC05	movapd	384-128(pA0,lda3,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC06	mulpd	384-128(pA0,lda7), rB0	addpd	rB0,rC07#endif#if KB > 50	movapd	400-128(pB0), rB0	movapd	400-128(pA0), rA0	mulpd	rB0,rA0	addpd	rA0,rC00	movapd	400-128(pA0,lda), rA0	mulpd	rB0,rA0	addpd	rA0,rC01	movapd	400-128(pA0,lda,2), rA0	mulpd	rB0,rA0	addpd	rA0,rC02	movapd	400-128(pA0,lda3), rA0	mulpd	rB0,rA0	addpd	rA0,rC03	movapd	400-128(pA0,lda,4), rA0	mulpd	rB0,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?