📄 sgemv_n.s
字号:
(p14) lfetch.excl.nt2 [PREB], 16 * 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 (p18) STFD [YST1] = f19, 1 * SIZE (p17) FMA f116 = f10, f54, f116 } ;; { .mfi (p16) LDFPD f52, f53 = [AO3], 2 * SIZE (p17) FMA f119 = f10, f55, f119 } { .mfi (p18) STFD [YST1] = f20, 1 * SIZE (p17) FMA f122 = f10, f56, f122 } ;; { .mfi (p16) LDFPD f54, f55 = [AO3], 2 * SIZE (p17) FMA f16 = f11, f57, f101 } { .mmf (p15) PREFETCH [RPRE4], 16 * SIZE (p16) LDFD f56 = [AO4], 1 * SIZE (p17) FMA f17 = f11, f58, f104 } ;; { .mfi (p16) LDFPD f57, f58 = [AO4], 2 * SIZE (p17) FMA f18 = f11, f59, f107 } { .mfi (p18) STFD [YST1] = f21, 1 * SIZE (p17) FMA f19 = f11, f60, f110 } ;; { .mfi (p16) LDFPD f59, f60 = [AO4], 2 * SIZE (p17) FMA f20 = f11, f61, f113 } { .mfi (p18) STFD [YST1] = f22, 1 * SIZE (p17) FMA f21 = f11, f62, f116 } ;; { .mfi (p16) LDFPD f61, f62 = [AO4], 2 * SIZE (p17) FMA f22 = f11, f63, f119 } { .mfb (p18) STFD [YST1] = f23, 1 * SIZE (p17) FMA f23 = f11, f64, f122 br.ctop.sptk.few .L122 } ;; .align 16.L125: { .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 } ;; { .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 } ;; { .mmi (p14) LDFPD f64, f65 = [AO1], 2 * SIZE (p14) LDFPD f104, f105 = [YLD1], 2 * SIZE } { .mmi (p18) STFD [YST1] = f18, 1 * SIZE } ;; { .mmi (p18) STFD [YST1] = f19, 1 * SIZE (p15) LDFD f80 = [AO1] } { .mmi (p15) LDFD f106 = [YLD1], 1 * SIZE (p13) LDFD f34 = [AO2], 1 * SIZE } ;; { .mmi (p13) LDFPD f35, f50 = [AO2], 2 * SIZE (p13) LDFPD f36, f37 = [AO3], 2 * SIZE } { .mmi (p18) STFD [YST1] = f20, 1 * SIZE } ;; { .mmi (p13) LDFD f51 = [AO2], 1 * SIZE (p13) LDFPD f52, f53 = [AO3], 2 * SIZE } { .mmi (p18) STFD [YST1] = f21, 1 * SIZE } ;; { .mmi (p14) LDFD f66 = [AO2], 1 * SIZE (p14) LDFPD f68, f69 = [AO3], 2 * SIZE } { .mmi (p18) STFD [YST1] = f22, 1 * SIZE } ;; { .mmf (p18) STFD [YST1] = f23, 1 * SIZE (p14) LDFD f67 = [AO2], 1 * SIZE (p13) FMA f100 = f8, f32, f100 } { .mmf (p15) LDFD f82 = [AO3] (p13) LDFD f38 = [AO4], 1 * SIZE (p13) FMA f101 = f8, f33, f101 } ;; ;; { .mmf (p13) LDFPD f39, f54 = [AO4], 2 * SIZE (p15) LDFD f81 = [AO2] (p13) FMA f102 = f8, f48, f102 } { .mfi (p13) FMA f103 = f8, f49, f103 } ;; { .mfi (p13) LDFD f55 = [AO4], 1 * SIZE (p14) FMA f104 = f8, f64, f104 } { .mfi (p14) FMA f105 = f8, f65, f105 } ;; { .mfi (p14) LDFD f70 = [AO4], 1 * SIZE (p15) FMA f106 = f8, f80, f106 } { .mfi (p13) FMA f100 = f9, f34, f100 } ;; { .mfi (p14) LDFD f71 = [AO4], 1 * SIZE (p13) FMA f101 = f9, f35, f101 } { .mfi (p13) FMA f102 = f9, f50, f102 } ;; (p15) LDFD f83 = [AO4] (p13) FMA f103 = f9, f51, f103 (p14) FMA f104 = f9, f66, f104 (p14) FMA f105 = f9, f67, f105 (p15) FMA f106 = f9, f81, f106 ;; (p13) FMA f100 = f10, f36, f100 (p13) FMA f101 = f10, f37, f101 (p13) FMA f102 = f10, f52, f102 (p13) FMA f103 = f10, f53, f103 (p14) FMA f104 = f10, f68, f104 (p14) FMA f105 = f10, f69, f105 (p15) FMA f106 = f10, f82, f106 ;; (p13) FMA f100 = f11, f38, f100 (p13) FMA f101 = f11, f39, f101 ;; (p13) FMA f102 = f11, f54, f102 (p13) STFD [YST1] = f100, 1 * SIZE (p13) FMA f103 = f11, f55, f103 ;; (p13) STFD [YST1] = f101, 1 * SIZE (p14) FMA f104 = f11, f70, f104 ;; (p13) STFD [YST1] = f102, 1 * SIZE (p14) FMA f105 = f11, f71, f105 ;; (p13) STFD [YST1] = f103, 1 * SIZE (p15) FMA f106 = f11, f83, f106 ;; (p14) STFD [YST1] = f104, 1 * SIZE ;; (p14) STFD [YST1] = f105, 1 * SIZE ;; (p15) STFD [YST1] = f106, 1 * SIZE ;; .align 16.L130: { .mmi mov YLD1 = YY mov YST1 = YY tbit.z p6, p0 = N, 1 } ;; { .mib mov AO1 = A mov pr.rot= 0 (p6) br.cond.dpnt .L140 } ;; { .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 A = LDA, 1, A } ;; adds PREB = RPREFETCH * SIZE, YLD1 FMPY f8 = ALPHA, f8 mov ar.ec= 2 adds RPRE1 = RPREFETCH * SIZE, AO1 FMPY f9 = ALPHA, f9 shr I = MM, 3 ;; (p8) LDFD f81 = [AO2], 1 * SIZE cmp.eq p6, p0 = 0, I ;; (p8) FMA f106 = f8, f80, f106 adds RPRE2 = (RPREFETCH + 8) * SIZE, AO2 tbit.nz p13, p0 = MM, 2 ;; (p8) FMA f106 = f9, f81, f106 cmp.eq p16, p0 = r0, r0 adds I = -1, I ;; { .mib (p8) STFD [YST1] = f106, 1 * SIZE mov ar.lc = I (p6) br.cond.dpnt .L135 } ;; .align 16.L132: { .mfi (p17) LDFD f48 = [AO2], 1 * SIZE (p17) FMA f101 = f8, f33, f101 (p16) tbit.nz.unc p14, p15 = I, 0 } { .mmf (p16) LDFPD f100, f103 = [YLD1], 2 * SIZE (p18) STFD [YST1] = f16, 1 * SIZE (p17) FMA f104 = f8, f34, f104 } ;; { .mfi (p16) LDFPD f32, f33 = [AO1], 2 * SIZE (p17) FMA f107 = f8, f35, f107 adds I = -1, I } { .mmf (p14) PREFETCH [RPRE1], 16 * SIZE (p18) STFD [YST1] = f17, 1 * SIZE (p17) FMA f110 = f8, f36, f110 } ;; { .mfi (p16) LDFPD f34, f35 = [AO1], 2 * SIZE (p17) FMA f113 = f8, f37, f113 } { .mmf (p16) LDFPD f106, f109 = [YLD1], 2 * SIZE (p18) STFD [YST1] = f18, 1 * SIZE (p17) FMA f116 = f8, f38, f116 } ;; { .mfi (p16) LDFPD f36, f37 = [AO1], 2 * SIZE (p17) FMA f119 = f8, f39, f119 } { .mmf (p16) LDFPD f112, f115 = [YLD1], 2 * SIZE (p18) STFD [YST1] = f19, 1 * SIZE (p17) FMA f122 = f8, f40, f122 } ;; { .mmf (p16) LDFPD f38, f39 = [AO1], 2 * SIZE (p16) LDFPD f118, f121 = [YLD1], 2 * SIZE (p17) FMA f16 = f9, f41, f101 } { .mmf (p18) STFD [YST1] = f20, 1 * SIZE (p16) LDFD f40 = [AO2], 1 * SIZE (p17) FMA f17 = f9, f42, f104 } ;; { .mfi (p16) LDFPD f41, f42 = [AO2], 2 * SIZE (p17) FMA f18 = f9, f43, f107 } { .mmf (p15) PREFETCH [RPRE2], 16 * SIZE (p18) STFD [YST1] = f21, 1 * SIZE (p17) FMA f19 = f9, f44, f110 } ;; { .mfi (p16) LDFPD f43, f44 = [AO2], 2 * SIZE (p17) FMA f20 = f9, f45, f113 } { .mmf (p14) PREFETCH [PREB], 16 * SIZE (p18) STFD [YST1] = f22, 1 * SIZE (p17) FMA f21 = f9, f46, f116 } ;; { .mfi (p16) LDFPD f45, f46 = [AO2], 2 * SIZE (p17) FMA f22 = f9, f47, f119 } { .mfb (p18) STFD [YST1] = f23, 1 * SIZE (p17) FMA f23 = f9, f48, f122 br.ctop.sptk.few .L132 } ;; .align 16.L135: { .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 } ;; { .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 } ;; { .mmi (p14) LDFPD f64, f65 = [AO1], 2 * SIZE (p14) LDFPD f104, f105 = [YLD1], 2 * SIZE } { .mmi (p18) STFD [YST1] = f18, 1 * SIZE } ;; { .mmi (p15) LDFD f80 = [AO1] (p15) LDFD f106 = [YLD1], 1 * SIZE } { .mmi (p18) STFD [YST1] = f19, 1 * SIZE } ;; { .mmi (p13) LDFD f34 = [AO2], 1 * SIZE (p18) STFD [YST1] = f20, 1 * SIZE } ;; { .mmi (p13) LDFD f35 = [AO2], 1 * SIZE (p18) STFD [YST1] = f21, 1 * SIZE } ;; { .mmi (p13) LDFD f50 = [AO2], 1 * SIZE (p18) STFD [YST1] = f22, 1 * SIZE } ;; { .mmi (p13) LDFD f51 = [AO2], 1 * SIZE (p18) STFD [YST1] = f23, 1 * SIZE } ;; (p14) LDFD f66 = [AO2], 1 * SIZE (p13) FMA f100 = f8, f32, f100 ;; (p14) LDFD f67 = [AO2], 1 * SIZE (p13) FMA f101 = f8, f33, f101 ;; (p15) LDFD f81 = [AO2] (p13) FMA f102 = f8, f48, f102 (p13) FMA f103 = f8, f49, f103 (p14) FMA f104 = f8, f64, f104 (p14) FMA f105 = f8, f65, f105 (p15) FMA f106 = f8, f80, f106 ;; (p13) FMA f100 = f9, f34, f100 (p13) FMA f101 = f9, f35, f101 (p13) FMA f102 = f9, f50, f102 (p13) FMA f103 = f9, f51, f103 (p14) FMA f104 = f9, f66, f104 (p14) FMA f105 = f9, f67, f105 (p15) FMA f106 = f9, f81, 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 ;; .align 16.L140: { .mmi mov YLD1 = YY mov YST1 = YY tbit.z p6, p0 = N, 0 } ;; { .mib mov AO1 = A mov pr.rot= 0 (p6) br.cond.dpnt .L990 } ;; { .mmi LDFD f8 = [X], INCX (p8) LDFD f106 = [YLD1], 1 * SIZE adds RPRE1 = RPREFETCH * SIZE, AO1 } ;; { .mmi (p8) LDFD f80 = [AO1], 1 * SIZE adds PREB = RPREFETCH * SIZE, YLD1 } ;; FMPY f8 = ALPHA, f8 shr I = MM, 3 ;; (p8) FMA f106 = f8, f80, f106 mov ar.ec= 3 ;; { .mmi cmp.eq p6, p0 = 0, I cmp.eq p16, p0 = r0, r0 tbit.nz p14, p15 = r0, 0 } ;; { .mmi adds YST2 = 4 * SIZE, YST1 adds I = -1, I tbit.nz p13, p0 = MM, 2 } ;; { .mmi (p8) STFD [YST1] = f106, 1 * SIZE (p8) adds YST2 = 1 * SIZE, YST2 } { .mib mov ar.lc = I (p6) br.cond.dpnt .L145 } ;; .align 16.L142: { .mmf (p19) STFD [YST1] = f16, 1 * SIZE (p19) STFD [YST2] = f20, 1 * SIZE (p18) FMA f16 = f8, f34, f102 } { .mmf (p16) LDFPD f32, f35 = [AO1], 2 * SIZE (p16) LDFPD f100, f103 = [YLD1], 2 * SIZE (p18) FMA f20 = f8, f46, f114 } ;; { .mmf (p19) STFD [YST1] = f17, 1 * SIZE (p19) STFD [YST2] = f21, 1 * SIZE (p18) FMA f17 = f8, f37, f105 } { .mmf (p16) LDFPD f38, f41 = [AO1], 2 * SIZE (p16) LDFPD f106, f109 = [YLD1], 2 * SIZE (p18) FMA f21 = f8, f49, f117 } ;; { .mmf (p19) STFD [YST1] = f18, 1 * SIZE (p19) STFD [YST2] = f22, 1 * SIZE (p18) FMA f18 = f8, f40, f108 } { .mmf (p16) LDFPD f44, f47 = [AO1], 2 * SIZE (p16) LDFPD f112, f115 = [YLD1], 2 * SIZE (p18) FMA f22 = f8, f52, f120 } ;; { .mmf (p19) STFD [YST1] = f19, 5 * SIZE (p19) STFD [YST2] = f23, 5 * SIZE (p18) FMA f19 = f8, f43, f111 } { .mmf (p16) LDFPD f50, f53 = [AO1], 2 * SIZE (p16) LDFPD f118, f121 = [YLD1], 2 * SIZE (p18) FMA f23 = f8, f55, f123 } ;; { .mmi (p14) PREFETCH [RPRE1], 16 * SIZE (p14) PREFETCH [PREB], 16 * SIZE (p16) tbit.nz.unc p14, p15 = I, 0 } { .mib nop __LINE__ (p16) adds I = -1, I br.ctop.sptk.few .L142 } ;; .align 16.L145: { .mmi (p19) STFD [YST1] = f16, 1 * SIZE (p19) STFD [YST2] = f20, 1 * SIZE tbit.nz p14, p0 = MM, 1 } { .mmi (p13) LDFPD f32, f33 = [AO1], 2 * SIZE (p13) LDFPD f100, f101 = [YLD1], 2 * SIZE } ;; { .mmi (p19) STFD [YST1] = f17, 1 * SIZE (p19) STFD [YST2] = f21, 1 * SIZE tbit.nz p15, p0 = MM, 0 } { .mmi (p13) LDFPD f48, f49 = [AO1], 2 * SIZE (p13) LDFPD f102, f103 = [YLD1], 2 * SIZE } ;; { .mmi (p19) STFD [YST1] = f18, 1 * SIZE (p19) STFD [YST2] = f22, 1 * SIZE } { .mmi (p14) LDFPD f64, f65 = [AO1], 2 * SIZE (p14) LDFPD f104, f105 = [YLD1], 2 * SIZE } ;; { .mmi (p19) STFD [YST1] = f19, 5 * SIZE (p19) STFD [YST2] = f23, 5 * SIZE } { .mmi (p15) LDFD f80 = [AO1] (p15) LDFD f106 = [YLD1], 1 * SIZE } ;; (p13) FMA f100 = f8, f32, f100 (p13) FMA f101 = f8, f33, f101 (p13) FMA f102 = f8, f48, f102 (p13) FMA f103 = f8, f49, f103 (p14) FMA f104 = f8, f64, f104 (p14) FMA f105 = f8, f65, f105 (p15) FMA f106 = f8, f80, 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 ;; .align 16.L990: { .mmi mov YLD1 = YY mov YST1 = Y mov pr.rot= 0 } { .mib mov YST2 = Y shr J = M, 3 (p10) br.cond.dptk .L999 } ;; { .mmi cmp.eq p6, p0 = r0, J adds J = -1, J mov ar.ec = 4 } { .mmi cmp.eq p16, p0 = r0, r0 nop __LINE__ tbit.nz p13, p0 = M, 2 } ;; { .mib nop __LINE__ mov ar.lc = J (p6) br.cond.dpnt .L995 } ;;.L992: { .mfi (p19) STFD [YST2] = f35 (p18) FADD f34 = f34, f66 (p19) add YST2 = YST2, INCY } { .mmi (p16) LDFD f64 = [YLD1], 1 * SIZE (p16) LDFD f32 = [YST1], INCY } ;; { .mfi (p19) STFD [YST2] = f39 (p18) FADD f38 = f38, f70 (p19) add YST2 = YST2, INCY } { .mmi (p16) LDFD f36 = [YST1], INCY (p16) LDFD f68 = [YLD1], 1 * SIZE } ;; { .mfi (p19) STFD [YST2] = f43 (p18) FADD f42 = f42, f74 (p19) add YST2 = YST2, INCY } { .mmi (p16) LDFD f72 = [YLD1], 1 * SIZE (p16) LDFD f40 = [YST1], INCY } ;; { .mfi (p19) STFD [YST2] = f47 (p18) FADD f46 = f46, f78 (p19) add YST2 = YST2, INCY } { .mmi (p16) LDFD f76 = [YLD1], 1 * SIZE (p16) LDFD f44 = [YST1], INCY } ;; { .mfi (p19) STFD [YST2] = f51 (p18) FADD f50 = f50, f82 (p19) add YST2 = YST2, INCY } { .mmi (p16) LDFD f80 = [YLD1], 1 * SIZE (p16) LDFD f48 = [YST1], INCY } ;; { .mfi (p19) STFD [YST2] = f55 (p18) FADD f54 = f54, f86 (p19) add YST2 = YST2, INCY } { .mmi (p16) LDFD f84 = [YLD1], 1 * SIZE (p16) LDFD f52 = [YST1], INCY } ;; { .mfi (p19) STFD [YST2] = f59 (p18) FADD f58 = f58, f90 (p19) add YST2 = YST2, INCY } { .mmi (p16) LDFD f88 = [YLD1], 1 * SIZE (p16) LDFD f56 = [YST1], INCY } ;; { .mfi (p19) STFD [YST2] = f63 (p18) FADD f62 = f62, f94 (p19) add YST2 = YST2, INCY } { .mmb (p16) LDFD f92 = [YLD1], 1 * SIZE (p16) LDFD f60 = [YST1], INCY br.ctop.sptk.few .L992 } ;;.L995: (p13) LDFD f32 = [YST1], INCY (p13) LDFD f40 = [YLD1], 1 * SIZE tbit.nz p14, p0 = M, 1 ;; (p13) LDFD f33 = [YST1], INCY (p13) LDFD f41 = [YLD1], 1 * SIZE tbit.nz p15, p0 = M, 0 ;; (p13) LDFD f34 = [YST1], INCY (p13) LDFD f42 = [YLD1], 1 * SIZE ;; (p13) LDFD f35 = [YST1], INCY (p13) LDFD f43 = [YLD1], 1 * SIZE ;; (p14) LDFD f36 = [YST1], INCY (p14) LDFD f44 = [YLD1], 1 * SIZE ;; (p14) LDFD f37 = [YST1], INCY (p14) LDFD f45 = [YLD1], 1 * SIZE ;; (p15) LDFD f38 = [YST1], INCY (p15) LDFD f46 = [YLD1], 1 * SIZE ;; (p13) FADD f32 = f32, f40 (p13) FADD f33 = f33, f41 (p13) FADD f34 = f34, f42 (p13) FADD f35 = f35, f43 (p14) FADD f36 = f36, f44 (p14) FADD f37 = f37, f45 (p15) FADD f38 = f38, f46 ;; (p13) STFD [YST2] = f32 (p13) add YST2 = YST2, INCY ;; (p13) STFD [YST2] = f33 (p13) add YST2 = YST2, INCY ;; (p13) STFD [YST2] = f34 (p13) add YST2 = YST2, INCY ;; (p13) STFD [YST2] = f35 (p13) add YST2 = YST2, INCY ;; (p14) STFD [YST2] = f36 (p14) add YST2 = YST2, INCY ;; (p14) STFD [YST2] = f37 (p14) add YST2 = YST2, INCY ;; (p15) STFD [YST2] = f38 ;;.L999: mov r8 = r0 adds r9 = 1 * 16, SP ;; ldf.fill f16 = [SP], 32 ldf.fill f17 = [r9], 32 mov ar.lc = ARLC ;; ldf.fill f18 = [SP], 32 ldf.fill f19 = [r9], 32 mov pr = PR, -1 ;; ldf.fill f20 = [SP], 32 ldf.fill f21 = [r9], 32 mov ar.pfs = ARPFS ;; ldf.fill f22 = [SP], 32 ldf.fill f23 = [r9] br.ret.sptk.many b0 ;; EPILOGUE
⌨️ 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 + -