📄 sgemv_n.s
字号:
LDFD f10 = [X], INCX ;; LDFD f11 = [X], INCX ;; LDFD f12 = [X], INCX ;; LDFD f13 = [X], INCX ;; LDFD f14 = [X], INCX ;; LDFD f15 = [X], INCX ;; FMPY f8 = ALPHA, f8 FMPY f9 = ALPHA, f9 FMPY f10 = ALPHA, f10 FMPY f11 = ALPHA, f11 FMPY f12 = ALPHA, f12 FMPY f13 = ALPHA, f13 FMPY f14 = ALPHA, f14 FMPY f15 = ALPHA, f15 ;; mov AO1 = A add AO2 = LDA, A ;; shladd AO3 = LDA, 1, A shladd AO4 = LDA, 1, AO2 ;; shladd AO5 = LDA, 1, AO3 shladd AO6 = LDA, 1, AO4 ;; shladd AO7 = LDA, 1, AO5 shladd AO8 = LDA, 1, AO6 shladd A = LDA, 3, A ;; ;; adds PREB = RPREFETCH * SIZE, YLD1 adds RPRE1 = RPREFETCH * SIZE, AO1 adds RPRE2 = (RPREFETCH + 8) * SIZE, AO2 adds RPRE3 = RPREFETCH * SIZE, AO3 adds RPRE4 = (RPREFETCH + 8) * SIZE, AO4 adds RPRE5 = RPREFETCH * SIZE, AO5 adds RPRE6 = (RPREFETCH + 8) * SIZE, AO6 adds RPRE7 = RPREFETCH * SIZE, AO7 adds RPRE8 = (RPREFETCH + 8) * SIZE, AO8 (p8) LDFD f80 = [AO1], 1 * SIZE (p8) LDFD f81 = [AO2], 1 * SIZE (p8) LDFD f82 = [AO3], 1 * SIZE (p8) LDFD f83 = [AO4], 1 * SIZE (p8) LDFD f84 = [AO5], 1 * SIZE (p8) LDFD f85 = [AO6], 1 * SIZE (p8) LDFD f86 = [AO7], 1 * SIZE (p8) LDFD f87 = [AO8], 1 * SIZE (p8) LDFD f106 = [YLD1], 1 * SIZE ;; (p8) FMPY f32 = f8, f80 (p8) FMPY f33 = f9, f81 (p8) FMPY f34 = f10, f82 (p8) FMA f35 = f11, f83, f106 ;; (p8) FMA f32 = f12, f84, f32 (p8) FMA f33 = f13, f85, f33 (p8) FMA f34 = f14, f86, f34 (p8) FMA f35 = f15, f87, f35 ;; (p8) FADD f32 = f32, f33 (p8) FADD f34 = f34, f35 ;; (p8) FADD f32 = f32, f34 ;; (p8) STFD [YST1] = f32, 1 * SIZE shr I = MM, 3 mov pr.rot= 0 ;; cmp.eq p6, p0 = 0, I cmp.eq p16, p0 = r0, r0 ;; adds I = -1, I tbit.nz p13, p0 = MM, 2 ;; mov ar.lc = I mov ar.ec= 2 (p6) br.cond.dpnt .L115 ;; .align 16.L112: { .mfi (p17) LDFD f96 = [AO8], 1 * SIZE (p17) FMA f101 = f8, f33, f101 (p16) tbit.nz.unc p14, p15 = I, 0 } { .mfi (p17) FMA f104 = f8, f34, f104 } ;; { .mfi (p16) LDFPD f32, f33 = [AO1], 2 * SIZE (p17) FMA f107 = f8, f35, f107 } { .mfi (p14) PREFETCH [RPRE1], 16 * SIZE (p17) FMA f110 = f8, f36, f110 } ;; { .mfi (p16) LDFPD f34, f35 = [AO1], 2 * SIZE (p17) FMA f113 = f8, f37, f113 } { .mfi (p16) LDFPD f100, f103 = [YLD1], 2 * SIZE (p17) FMA f116 = f8, f38, f116 } ;; { .mfi (p16) LDFPD f36, f37 = [AO1], 2 * SIZE (p17) FMA f119 = f8, f39, f119 } { .mfi (p16) LDFPD f106, f109 = [YLD1], 2 * SIZE (p17) FMA f122 = f8, f40, f122 } ;; { .mfi (p16) LDFPD f38, f39 = [AO1], 2 * SIZE (p17) FMA f101 = f9, f41, f101 } { .mmf (p16) LDFPD f112, f115 = [YLD1], 2 * SIZE (p16) LDFD f40 = [AO2], 1 * SIZE (p17) FMA f104 = f9, f42, f104 } ;; { .mfi (p16) LDFPD f41, f42 = [AO2], 2 * SIZE (p17) FMA f107 = f9, f43, f107 } { .mfi (p15) PREFETCH [RPRE2], 16 * SIZE (p17) FMA f110 = f9, f44, f110 } ;; { .mfi (p16) LDFPD f43, f44 = [AO2], 2 * SIZE (p17) FMA f113 = f9, f45, f113 } { .mfi (p16) LDFPD f118, f121 = [YLD1], 2 * SIZE (p17) FMA f116 = f9, f46, f116 } ;; { .mfi (p16) LDFPD f45, f46 = [AO2], 2 * SIZE (p17) FMA f119 = f9, f47, f119 } { .mfi (p18) STFD [YST1] = f16, 1 * SIZE (p17) FMA f122 = f9, f48, f122 } ;; { .mfi (p16) LDFD f47 = [AO2], 1 * SIZE (p17) FMA f101 = f10, f49, f101 } { .mfi (p18) STFD [YST1] = f17, 1 * SIZE (p17) FMA f104 = f10, f50, f104 } ;; { .mfi (p16) LDFPD f48, f49 = [AO3], 2 * SIZE (p17) FMA f107 = f10, f51, f107 } { .mfi (p14) PREFETCH [RPRE3], 16 * SIZE (p17) FMA f110 = f10, f52, f110 } ;; { .mfi (p16) LDFPD f50, f51 = [AO3], 2 * SIZE (p17) FMA f113 = f10, f53, f113 } { .mfi (p17) FMA f116 = f10, f54, f116 } ;; { .mfi (p16) LDFPD f52, f53 = [AO3], 2 * SIZE (p17) FMA f119 = f10, f55, f119 } { .mfi (p18) STFD [YST1] = f18, 1 * SIZE (p17) FMA f122 = f10, f56, f122 } ;; { .mfi (p16) LDFPD f54, f55 = [AO3], 2 * SIZE (p17) FMA f101 = f11, f57, f101 } { .mmf (p18) STFD [YST1] = f19, 1 * SIZE (p16) LDFD f56 = [AO4], 1 * SIZE (p17) FMA f104 = f11, f58, f104 } ;; { .mfi (p16) LDFPD f57, f58 = [AO4], 2 * SIZE (p17) FMA f107 = f11, f59, f107 } { .mfi (p15) PREFETCH [RPRE4], 16 * SIZE (p17) FMA f110 = f11, f60, f110 } ;; { .mfi (p16) LDFPD f59, f60 = [AO4], 2 * SIZE (p17) FMA f113 = f11, f61, f113 } { .mfi (p17) FMA f116 = f11, f62, f116 } ;; { .mfi (p16) LDFPD f61, f62 = [AO4], 2 * SIZE (p17) FMA f119 = f11, f63, f119 } { .mfi (p17) FMA f122 = f11, f64, f122 } ;; { .mfi (p16) LDFD f63 = [AO4], 1 * SIZE (p17) FMA f101 = f12, f65, f101 } { .mfi (p18) STFD [YST1] = f20, 1 * SIZE (p17) FMA f104 = f12, f66, f104 } ;; { .mfi (p16) LDFPD f64, f65 = [AO5], 2 * SIZE (p17) FMA f107 = f12, f67, f107 } { .mfi (p18) STFD [YST1] = f21, 1 * SIZE (p17) FMA f110 = f12, f68, f110 } ;; { .mfi (p16) LDFPD f66, f67 = [AO5], 2 * SIZE (p17) FMA f113 = f12, f69, f113 } { .mfi (p14) PREFETCH [RPRE5], 16 * SIZE (p17) FMA f116 = f12, f70, f116 } ;; { .mfi (p16) LDFPD f68, f69 = [AO5], 2 * SIZE (p17) FMA f119 = f12, f71, f119 } { .mfi (p18) STFD [YST1] = f22, 1 * SIZE (p17) FMA f122 = f12, f72, f122 } ;; { .mfi (p16) LDFPD f70, f71 = [AO5], 2 * SIZE (p17) FMA f101 = f13, f73, f101 } { .mmf (p18) STFD [YST1] = f23, 1 * SIZE (p16) LDFD f72 = [AO6], 1 * SIZE (p17) FMA f104 = f13, f74, f104 } ;; { .mfi (p16) LDFPD f73, f74 = [AO6], 2 * SIZE (p17) FMA f107 = f13, f75, f107 } { .mfi (p15) PREFETCH [RPRE6], 16 * SIZE (p17) FMA f110 = f13, f76, f110 } ;; { .mfi (p16) LDFPD f75, f76 = [AO6], 2 * SIZE (p17) FMA f113 = f13, f77, f113 } { .mfi (p17) FMA f116 = f13, f78, f116 } ;; { .mfi (p16) LDFPD f77, f78 = [AO6], 2 * SIZE (p17) FMA f119 = f13, f79, f119 } { .mfi (p17) FMA f122 = f13, f80, f122 } ;; { .mfi (p16) LDFD f79 = [AO6], 1 * SIZE (p17) FMA f101 = f14, f81, f101 } { .mfi (p17) FMA f104 = f14, f82, f104 } ;; { .mfi (p16) LDFPD f80, f81 = [AO7], 2 * SIZE (p17) FMA f107 = f14, f83, f107 } { .mfi (p14) PREFETCH [RPRE7], 16 * SIZE (p17) FMA f110 = f14, f84, f110 } ;; { .mfi (p16) LDFPD f82, f83 = [AO7], 2 * SIZE (p17) FMA f113 = f14, f85, f113 } { .mfi (p17) FMA f116 = f14, f86, f116 } ;; { .mfi (p16) LDFPD f84, f85 = [AO7], 2 * SIZE (p17) FMA f119 = f14, f87, f119 } { .mfi (p17) FMA f122 = f14, f88, f122 } ;; { .mfi (p16) LDFPD f86, f87 = [AO7], 2 * SIZE (p17) FMA f16 = f15, f89, f101 } { .mfi (p16) LDFD f88 = [AO8], 1 * SIZE (p17) FMA f17 = f15, f90, f104 } ;; { .mfi (p16) LDFPD f89, f90 = [AO8], 2 * SIZE (p17) FMA f18 = f15, f91, f107 } { .mfi (p15) PREFETCH [RPRE8], 16 * SIZE (p17) FMA f19 = f15, f92, f110 } ;; { .mfi (p16) LDFPD f91, f92 = [AO8], 2 * SIZE (p17) FMA f20 = f15, f93, f113 } { .mfi (p14) lfetch.excl.nt2 [PREB], 16 * SIZE (p17) FMA f21 = f15, f94, f116 } ;; { .mfi (p16) LDFPD f93, f94 = [AO8], 2 * SIZE (p17) FMA f22 = f15, f95, f119 } { .mfb (p16) adds I = -1, I (p17) FMA f23 = f15, f96, f122 br.ctop.sptk.few .L112 } ;; .align 16.L115: { .mmi (p13) LDFPD f32, f33 = [AO1], 2 * SIZE (p13) LDFPD f100, f101 = [YLD1], 2 * SIZE tbit.nz p14, p0 = MM, 1 } { .mmi (p18) STFD [YST1] = f16, 1 * SIZE cmp.lt p6, p0 = 1, J adds J = -1, J } ;; { .mmi (p13) LDFPD f48, f49 = [AO1], 2 * SIZE (p13) LDFPD f102, f103 = [YLD1], 2 * SIZE tbit.nz p15, p0 = MM, 0 } { .mmi (p18) STFD [YST1] = f17, 1 * SIZE nop __LINE__ nop __LINE__ } ;; { .mmi (p14) LDFPD f64, f65 = [AO1], 2 * SIZE (p14) LDFPD f104, f105 = [YLD1], 2 * SIZE nop __LINE__ } { .mmi (p18) STFD [YST1] = f18, 1 * SIZE (p13) LDFD f34 = [AO2], 1 * SIZE nop __LINE__ } ;; { .mmi (p13) LDFPD f35, f50 = [AO2], 2 * SIZE (p13) LDFPD f36, f37 = [AO3], 2 * SIZE nop __LINE__ } { .mmi (p18) STFD [YST1] = f19, 1 * SIZE nop __LINE__ nop __LINE__ } ;; { .mmi (p15) LDFD f80 = [AO1] (p15) LDFD f106 = [YLD1], 1 * SIZE nop __LINE__ } { .mmi (p18) STFD [YST1] = f20, 1 * SIZE nop __LINE__ nop __LINE__ } ;; { .mmi (p13) LDFD f51 = [AO2], 1 * SIZE (p13) LDFPD f52, f53 = [AO3], 2 * SIZE nop __LINE__ } { .mmi (p18) STFD [YST1] = f21, 1 * SIZE nop __LINE__ nop __LINE__ } ;; { .mmi (p14) LDFD f66 = [AO2], 1 * SIZE (p14) LDFPD f68, f69 = [AO3], 2 * SIZE nop __LINE__ } { .mmi (p18) STFD [YST1] = f22, 1 * SIZE nop __LINE__ nop __LINE__ } ;; { .mmi (p14) LDFD f67 = [AO2], 1 * SIZE (p15) LDFD f82 = [AO3] nop __LINE__ } { .mmi (p18) STFD [YST1] = f23, 1 * SIZE nop __LINE__ } ;; { .mmf (p15) LDFD f81 = [AO2] (p13) LDFD f38 = [AO4], 1 * SIZE (p13) FMA f100 = f8, f32, f100 } { .mfi (p13) LDFPD f40, f41 = [AO5], 2 * SIZE (p13) FMA f101 = f8, f33, f101 nop __LINE__ } ;; { .mfi (p13) LDFPD f39, f54 = [AO4], 2 * SIZE (p13) FMA f102 = f8, f48, f102 nop __LINE__ } { .mfi (p13) LDFPD f56, f57 = [AO5], 2 * SIZE (p13) FMA f103 = f8, f49, f103 nop __LINE__ } ;; { .mfi (p13) LDFD f55 = [AO4], 1 * SIZE (p14) FMA f104 = f8, f64, f104 nop __LINE__ } { .mfi (p14) LDFPD f72, f73 = [AO5], 2 * SIZE (p14) FMA f105 = f8, f65, f105 nop __LINE__ } ;; { .mfi (p14) LDFD f70 = [AO4], 1 * SIZE (p15) FMA f106 = f8, f80, f106 nop __LINE__ } { .mmi (p15) LDFD f84 = [AO5] (p13) LDFD f42 = [AO6], 1 * SIZE nop __LINE__ } ;; { .mmf (p13) LDFPD f43, f58 = [AO6], 2 * SIZE (p14) LDFD f71 = [AO4], 1 * SIZE (p13) FMA f100 = f9, f34, f100 } { .mfi (p13) LDFPD f44, f45 = [AO7], 2 * SIZE (p13) FMA f101 = f9, f35, f101 nop __LINE__ } ;; { .mmf (p13) LDFD f59 = [AO6], 1 * SIZE (p15) LDFD f83 = [AO4] (p13) FMA f102 = f9, f50, f102 } { .mfi (p13) LDFPD f60, f61 = [AO7], 2 * SIZE (p13) FMA f103 = f9, f51, f103 nop __LINE__ } ;; { .mfi (p14) LDFD f74 = [AO6], 1 * SIZE (p14) FMA f104 = f9, f66, f104 nop __LINE__ } { .mfi (p14) LDFPD f76, f77 = [AO7], 2 * SIZE (p14) FMA f105 = f9, f67, f105 nop __LINE__ } ;; { .mfi (p14) LDFD f75 = [AO6], 1 * SIZE (p15) FMA f106 = f9, f81, f106 nop __LINE__ } { .mmi (p15) LDFD f86 = [AO7] (p13) LDFD f46 = [AO8], 1 * SIZE nop __LINE__ } ;; { .mmf (p13) LDFPD f47, f62 = [AO8], 2 * SIZE (p15) LDFD f85 = [AO6] (p13) FMA f100 = f10, f36, f100 } { .mfi (p13) FMA f101 = f10, f37, f101 nop __LINE__ } ;; { .mfi (p13) LDFD f63 = [AO8], 1 * SIZE (p13) FMA f102 = f10, f52, f102 nop __LINE__ } { .mfi (p13) FMA f103 = f10, f53, f103 nop __LINE__ } ;; { .mfi (p14) LDFD f78 = [AO8], 1 * SIZE (p14) FMA f104 = f10, f68, f104 nop __LINE__ } { .mfi (p14) FMA f105 = f10, f69, f105 nop __LINE__ } ;; { .mfi (p14) LDFD f79 = [AO8], 1 * SIZE (p15) FMA f106 = f10, f82, f106 nop __LINE__ } ;; (p15) LDFD f87 = [AO8] (p13) FMA f100 = f11, f38, f100 (p13) FMA f101 = f11, f39, f101 (p13) FMA f102 = f11, f54, f102 (p13) FMA f103 = f11, f55, f103 (p14) FMA f104 = f11, f70, f104 (p14) FMA f105 = f11, f71, f105 (p15) FMA f106 = f11, f83, f106 ;; (p13) FMA f100 = f12, f40, f100 (p13) FMA f101 = f12, f41, f101 (p13) FMA f102 = f12, f56, f102 (p13) FMA f103 = f12, f57, f103 (p14) FMA f104 = f12, f72, f104 (p14) FMA f105 = f12, f73, f105 (p15) FMA f106 = f12, f84, f106 ;; (p13) FMA f100 = f13, f42, f100 (p13) FMA f101 = f13, f43, f101 (p13) FMA f102 = f13, f58, f102 (p13) FMA f103 = f13, f59, f103 (p14) FMA f104 = f13, f74, f104 (p14) FMA f105 = f13, f75, f105 (p15) FMA f106 = f13, f85, f106 ;; (p13) FMA f100 = f14, f44, f100 (p13) FMA f101 = f14, f45, f101 (p13) FMA f102 = f14, f60, f102 (p13) FMA f103 = f14, f61, f103 (p14) FMA f104 = f14, f76, f104 (p14) FMA f105 = f14, f77, f105 (p15) FMA f106 = f14, f86, f106 ;; (p13) FMA f100 = f15, f46, f100 (p13) FMA f101 = f15, f47, f101 (p13) FMA f102 = f15, f62, f102 (p13) FMA f103 = f15, f63, f103 (p14) FMA f104 = f15, f78, f104 (p14) FMA f105 = f15, f79, f105 (p15) FMA f106 = f15, f87, f106 ;; (p13) STFD [YST1] = f100, 1 * SIZE ;; (p13) STFD [YST1] = f101, 1 * SIZE ;; (p13) STFD [YST1] = f102, 1 * SIZE ;; (p13) STFD [YST1] = f103, 1 * SIZE ;; (p14) STFD [YST1] = f104, 1 * SIZE ;; (p14) STFD [YST1] = f105, 1 * SIZE ;; (p15) STFD [YST1] = f106, 1 * SIZE (p6) br.cond.dptk .L111 ;; .align 16.L120: { .mmi mov YLD1 = YY mov YST1 = YY tbit.z p6, p0 = N, 2 } ;; { .mib mov AO1 = A mov pr.rot= 0 (p6) br.cond.dpnt .L130 } ;; { .mmi LDFD f8 = [X], INCX (p8) LDFD f106 = [YLD1], 1 * SIZE add AO2 = LDA, A } ;; { .mmi LDFD f9 = [X], INCX (p8) LDFD f80 = [AO1], 1 * SIZE shladd AO4 = LDA, 1, AO2 } ;; { .mmi LDFD f10 = [X], INCX (p8) LDFD f81 = [AO2], 1 * SIZE shladd AO3 = LDA, 1, A } ;; { .mmi LDFD f11 = [X], INCX (p8) LDFD f82 = [AO3], 1 * SIZE } ;; { .mfi (p8) LDFD f83 = [AO4], 1 * SIZE FMPY f8 = ALPHA, f8 adds PREB = RPREFETCH * SIZE, YLD1 } { .mfi adds RPRE1 = RPREFETCH * SIZE, AO1 FMPY f9 = ALPHA, f9 adds RPRE2 = (RPREFETCH + 8) * SIZE, AO2 } ;; FMPY f10 = ALPHA, f10 shladd A = LDA, 2, A FMPY f11 = ALPHA, f11 ;; { .mfi adds RPRE3 = RPREFETCH * SIZE, AO3 (p8) FMA f106 = f8, f80, f106 mov ar.ec= 2 } ;; adds RPRE4 = (RPREFETCH + 8) * SIZE, AO4 (p8) FMA f106 = f9, f81, f106 shr I = MM, 3 ;; { .mmf cmp.eq p6, p0 = 0, I cmp.eq p16, p0 = r0, r0 (p8) FMA f106 = f10, f82, f106 } ;; { .mfi adds I = -1, I (p8) FMA f106 = f11, f83, f106 tbit.nz p13, p0 = MM, 2 } ;; { .mib (p8) STFD [YST1] = f106, 1 * SIZE mov ar.lc = I (p6) br.cond.dpnt .L125 } ;; .align 16.L122: { .mfi (p17) LDFD f64 = [AO4], 1 * SIZE (p17) FMA f101 = f8, f33, f101 (p16) tbit.nz.unc p14, p15 = I, 0 } { .mfi (p16) LDFPD f100, f103 = [YLD1], 2 * SIZE (p17) FMA f104 = f8, f34, f104 } ;; { .mfi (p16) LDFPD f32, f33 = [AO1], 2 * SIZE (p17) FMA f107 = f8, f35, f107 (p16) adds I = -1, I } { .mfi (p14) PREFETCH [RPRE1], 16 * SIZE (p17) FMA f110 = f8, f36, f110 } ;; { .mfi (p16) LDFPD f34, f35 = [AO1], 2 * SIZE (p17) FMA f113 = f8, f37, f113 } { .mfi (p16) LDFPD f106, f109 = [YLD1], 2 * SIZE (p17) FMA f116 = f8, f38, f116 } ;; { .mfi (p16) LDFPD f36, f37 = [AO1], 2 * SIZE (p17) FMA f119 = f8, f39, f119 } { .mfi (p16) LDFPD f112, f115 = [YLD1], 2 * SIZE (p17) FMA f122 = f8, f40, f122 } ;; { .mfi (p16) LDFPD f38, f39 = [AO1], 2 * SIZE (p17) FMA f101 = f9, f41, f101 } { .mmf (p16) LDFPD f118, f121 = [YLD1], 2 * SIZE (p16) LDFD f40 = [AO2], 1 * SIZE (p17) FMA f104 = f9, f42, f104 } ;; { .mmf (p16) LDFPD f41, f42 = [AO2], 2 * SIZE (p15) PREFETCH [RPRE2], 16 * SIZE (p17) FMA f107 = f9, f43, f107 } { .mfi (p18) STFD [YST1] = f16, 1 * SIZE (p17) FMA f110 = f9, f44, f110 } ;; { .mfi (p16) LDFPD f43, f44 = [AO2], 2 * SIZE (p17) FMA f113 = f9, f45, f113 } { .mfi (p18) STFD [YST1] = f17, 1 * SIZE (p17) FMA f116 = f9, f46, f116 } ;; { .mfi (p16) LDFPD f45, f46 = [AO2], 2 * SIZE (p17) FMA f119 = f9, f47, f119 } { .mfi (p18) STFD [YST1] = f18, 1 * SIZE (p17) FMA f122 = f9, f48, f122 } ;; { .mfi (p16) LDFD f47 = [AO2], 1 * SIZE (p17) FMA f101 = f10, f49, f101 } { .mfi
⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -