📄 zgemm_kernel.s
字号:
#if! defined(TRMMKERNEL) && !defined(BETAZERO) (p5) LDFD f122 = [C4], SIZE#else nop __LINE__#endif (p3) FMA_B f99 = f42, f61, f99 // A3 * B6 nop __LINE__ } ;; { .mfb#if! defined(TRMMKERNEL) && !defined(BETAZERO) (p5) LDFD f107 = [C3], -3 * SIZE#else nop __LINE__#endif (p3) FMA f114 = f42, f62, f114 // A3 * B7 nop __LINE__ } { .mfb#if! defined(TRMMKERNEL) && !defined(BETAZERO) (p5) LDFD f123 = [C4], -3 * SIZE#else nop __LINE__#endif (p3) FMA_B f115 = f42, f63, f115 // A3 * B8 nop __LINE__ } ;; { .mfb nop __LINE__ (p3) FMA f67 = f43, f56, f67 // A4 * B1 nop __LINE__ } { .mfb nop __LINE__ (p3) FMA_A f66 = f43, f57, f66 // A4 * B2 nop __LINE__ } ;; { .mfb nop __LINE__ (p3) FMA f83 = f43, f58, f83 // A4 * B3 nop __LINE__ } { .mfb nop __LINE__ (p3) FMA_A f82 = f43, f59, f82 // A4 * B4 nop __LINE__ } ;; { .mfb nop __LINE__ (p3) FMA f99 = f43, f60, f99 // A4 * B5 nop __LINE__ } { .mfb nop __LINE__ (p3) FMA_A f98 = f43, f61, f98 // A4 * B6 nop __LINE__ } ;; { .mfi nop __LINE__ (p3) FMA f115 = f43, f62, f115 // A4 * B7 adds L = -1, L } { .mfb nop __LINE__ (p3) FMA_A f114 = f43, f63, f114 // A4 * B8 br.cloop.sptk.few .L022 } ;;#if! defined(TRMMKERNEL) && !defined(BETAZERO) { .mfb nop __LINE__ FMA f72 = ALPHA_R, f64, f72 nop __LINE__ } { .mfb nop __LINE__ FMA f88 = ALPHA_R, f80, f88 nop __LINE__ } ;; { .mfb nop __LINE__ FCALC_C f73 = ALPHA_R, f65, f73 nop __LINE__ } { .mfb nop __LINE__ FCALC_C f89 = ALPHA_R, f81, f89 nop __LINE__ } ;; { .mfb nop __LINE__ FMA f74 = ALPHA_R, f66, f74 nop __LINE__ } { .mfb nop __LINE__ FMA f90 = ALPHA_R, f82, f90 nop __LINE__ } ;; { .mfb nop __LINE__ FCALC_C f75 = ALPHA_R, f67, f75 nop __LINE__ } { .mfb nop __LINE__ FCALC_C f91 = ALPHA_R, f83, f91 nop __LINE__ } ;; { .mfb nop __LINE__ FCALC_D f72 = ALPHA_I, f65, f72 nop __LINE__ } { .mfb nop __LINE__ FCALC_D f88 = ALPHA_I, f81, f88 nop __LINE__ } { .mfb nop __LINE__ FMA f73 = ALPHA_I, f64, f73 nop __LINE__ } { .mfb FMA f89 = ALPHA_I, f80, f89 nop __LINE__ } { .mfb nop __LINE__ FCALC_D f74 = ALPHA_I, f67, f74 nop __LINE__ } { .mfb FCALC_D f90 = ALPHA_I, f83, f90 nop __LINE__ } { .mfb nop __LINE__ FMA f75 = ALPHA_I, f66, f75 nop __LINE__ } { .mfb nop __LINE__ FMA f91 = ALPHA_I, f82, f91 nop __LINE__ } ;; { .mfb STFD [C1] = f72, SIZE FMA f104 = ALPHA_R, f96, f104 nop __LINE__ } { .mfb STFD [C2] = f88, SIZE FMA f120 = ALPHA_R, f112, f120 nop __LINE__ } ;; { .mfb STFD [C1] = f73, SIZE FCALC_C f105 = ALPHA_R, f97, f105 nop __LINE__ } { .mfb STFD [C2] = f89, SIZE FCALC_C f121 = ALPHA_R, f113, f121 nop __LINE__ } ;; { .mfb STFD [C1] = f74, SIZE FMA f106 = ALPHA_R, f98, f106 nop __LINE__ } { .mfb STFD [C2] = f90, SIZE FMA f122 = ALPHA_R, f114, f122 nop __LINE__ } ;; { .mfb STFD [C1] = f75, SIZE FCALC_C f107 = ALPHA_R, f99, f107 nop __LINE__ } { .mfb STFD [C2] = f91, SIZE FCALC_C f123 = ALPHA_R, f115, f123 nop __LINE__ } ;; { .mfb nop __LINE__ FCALC_D f104 = ALPHA_I, f97, f104 nop __LINE__ } { .mfb nop __LINE__ FCALC_D f120 = ALPHA_I, f113, f120 nop __LINE__ } ;; { .mfb nop __LINE__ FMA f105 = ALPHA_I, f96, f105 nop __LINE__ } { .mfb nop __LINE__ FMA f121 = ALPHA_I, f112, f121 nop __LINE__ } ;; { .mfb nop __LINE__ FCALC_D f106 = ALPHA_I, f99, f106 nop __LINE__ } { .mfb nop __LINE__ FCALC_D f122 = ALPHA_I, f115, f122 nop __LINE__ } ;; { .mfb nop __LINE__ FMA f107 = ALPHA_I, f98, f107 nop __LINE__ } { .mfb nop __LINE__ FMA f123 = ALPHA_I, f114, f123 nop __LINE__ } ;; { .mfb STFD [C3] = f104, SIZE mov f64 = f0 nop __LINE__ } { .mfi STFD [C4] = f120, SIZE mov f65 = f0 } ;; { .mfb STFD [C3] = f105, SIZE mov f80 = f0 nop __LINE__ } { .mfi STFD [C4] = f121, SIZE mov f81 = f0 } ;; { .mfb STFD [C3] = f106, SIZE mov f96 = f0 nop __LINE__ } { .mfi STFD [C4] = f122, SIZE mov f97 = f0 } ;; { .mfi STFD [C3] = f107, SIZE mov f112 = f0 } { .mfb STFD [C4] = f123, SIZE mov f113 = f0 nop __LINE__ } ;;#else { .mfb nop __LINE__ FMPY f72 = ALPHA_R, f64 nop __LINE__ } { .mfb nop __LINE__ FMPY f88 = ALPHA_R, f80 nop __LINE__ } ;; { .mfb nop __LINE__ FCALC_C f73 = ALPHA_R, f65, f0 nop __LINE__ } { .mfb nop __LINE__ FCALC_C f89 = ALPHA_R, f81, f0 nop __LINE__ } ;; { .mfb nop __LINE__ FMPY f74 = ALPHA_R, f66 nop __LINE__ } { .mfb nop __LINE__ FMPY f90 = ALPHA_R, f82 nop __LINE__ } ;; { .mfb nop __LINE__ FCALC_C f75 = ALPHA_R, f67, f0 nop __LINE__ } { .mfb nop __LINE__ FCALC_C f91 = ALPHA_R, f83, f0 nop __LINE__ } ;; { .mfb nop __LINE__ FCALC_D f72 = ALPHA_I, f65, f72 nop __LINE__ } { .mfb nop __LINE__ FCALC_D f88 = ALPHA_I, f81, f88 nop __LINE__ } ;; { .mfb nop __LINE__ FMA f73 = ALPHA_I, f64, f73 nop __LINE__ } { .mfb FMA f89 = ALPHA_I, f80, f89 nop __LINE__ } ;; { .mfb nop __LINE__ FCALC_D f74 = ALPHA_I, f67, f74 nop __LINE__ } { .mfb FCALC_D f90 = ALPHA_I, f83, f90 nop __LINE__ } ;; { .mfb nop __LINE__ FMA f75 = ALPHA_I, f66, f75 nop __LINE__ } { .mfb nop __LINE__ FMA f91 = ALPHA_I, f82, f91 nop __LINE__ } ;; { .mfb STFD [C1] = f72, SIZE FMPY f104 = ALPHA_R, f96 nop __LINE__ } { .mfb STFD [C2] = f88, SIZE FMPY f120 = ALPHA_R, f112 nop __LINE__ } ;; { .mfb STFD [C1] = f73, SIZE FCALC_C f105 = ALPHA_R, f97, f0 nop __LINE__ } { .mfb STFD [C2] = f89, SIZE FCALC_C f121 = ALPHA_R, f113, f0 nop __LINE__ } ;; { .mfb STFD [C1] = f74, SIZE FMPY f106 = ALPHA_R, f98 nop __LINE__ } { .mfb STFD [C2] = f90, SIZE FMPY f122 = ALPHA_R, f114 nop __LINE__ } ;; { .mfb STFD [C1] = f75, SIZE FCALC_C f107 = ALPHA_R, f99, f0 nop __LINE__ } { .mfb STFD [C2] = f91, SIZE FCALC_C f123 = ALPHA_R, f115, f0 nop __LINE__ } ;; { .mfb nop __LINE__ FCALC_D f104 = ALPHA_I, f97, f104 nop __LINE__ } { .mfb nop __LINE__ FCALC_D f120 = ALPHA_I, f113, f120 nop __LINE__ } ;; { .mfb nop __LINE__ FMA f105 = ALPHA_I, f96, f105 nop __LINE__ } { .mfb nop __LINE__ FMA f121 = ALPHA_I, f112, f121 nop __LINE__ } ;; { .mfi nop __LINE__ FCALC_D f106 = ALPHA_I, f99, f106#if defined(TRMMKERNEL) && \ ((defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))) sub L = K, KK#else nop __LINE__#endif } { .mfb nop __LINE__ FCALC_D f122 = ALPHA_I, f115, f122 nop __LINE__ } ;; { .mfi nop __LINE__ FMA f107 = ALPHA_I, f98, f107#if defined(TRMMKERNEL) && (defined(LEFT) && defined(TRANSA)) adds L = -2, L#else nop __LINE__#endif } { .mfi nop __LINE__ FMA f123 = ALPHA_I, f114, f123#if defined(TRMMKERNEL) && (!defined(LEFT) && !defined(TRANSA)) adds L = -4, L#else nop __LINE__#endif } ;; { .mfi STFD [C3] = f104, SIZE mov f64 = f0#if defined(TRMMKERNEL) && \ ((defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))) shladd KK8 = L, ZBASE_SHIFT, r0#else nop __LINE__#endif } { .mfi STFD [C4] = f120, SIZE mov f65 = f0 nop __LINE__ } ;; { .mfi STFD [C3] = f105, SIZE mov f80 = f0#if defined(TRMMKERNEL) && \ ((defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))) shladd AOFFSET = KK8, 1, AOFFSET#else nop __LINE__#endif } { .mfi STFD [C4] = f121, SIZE mov f81 = f0#if defined(TRMMKERNEL) && \ ((defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))) shladd BOFFSET = KK8, 2, BOFFSET#else nop __LINE__#endif } ;; { .mfi STFD [C3] = f106, SIZE mov f96 = f0#if defined(TRMMKERNEL) && defined(LEFT) adds KK = 2, KK#else nop __LINE__#endif } { .mfi STFD [C4] = f122, SIZE mov f97 = f0 nop __LINE__ } ;; { .mfi STFD [C3] = f107, SIZE mov f112 = f0#ifdef TRMMKERNEL shladd KK8 = KK, ZBASE_SHIFT, r0#else nop __LINE__#endif } { .mfb STFD [C4] = f123, SIZE mov f113 = f0 nop __LINE__ } ;;#endif .align 16.L030: { .mib#ifndef TRMMKERNEL nop __LINE__#else#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) sub L = K, KK#elif defined(LEFT) adds L = 1, KK#else adds L = 4, KK#endif#endif tbit.z p6, p7 = M, 0 (p6) br.cond.dptk .L049 } ;;#if !defined(TRMMKERNEL) || \ defined(TRMMKERNEL) && \ ((defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))) { .mfb LDFPD f48, f49 = [B] mov f72 = f0 nop __LINE__ } { .mfi adds BOFFSET = 2 * SIZE, B mov f73 = f0#ifndef TRMMKERNEL adds L = 1, K#else adds L = 1, L#endif } ;;#else { .mfi shladd BOFFSET = KK8, 2, B mov f72 = f0 add AOFFSET = KK8, AOFFSET } ;; { .mfi LDFPD f48, f49 = [BOFFSET], 2 * SIZE mov f73 = f0#ifndef TRMMKERNEL adds L = 1, K#else adds L = 1, L#endif } ;;#endif { .mmi nop __LINE__ adds PREB = (PREFETCHSIZE + 0) * SIZE, BOFFSET tbit.z p12, p0 = L, 0 } ;; { .mfi LDFPD f50, f51 = [BOFFSET], 2 * SIZE mov f88 = f0 shr L = L, 1 } { .mfi (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE mov f89 = f0 nop __LINE__ } ;; { .mfi LDFPD f52, f53 = [BOFFSET], 2 * SIZE mov f104 = f0 adds L = -1, L } { .mfb adds PREA = (PREFETCHSIZE + 0) * SIZE, AOFFSET mov f105 = f0 nop __LINE__ } ;; { .mfi LDFPD f54, f55 = [BOFFSET], 2 * SIZE mov f120 = f0 mov ar.lc = L } { .mfi cmp.eq p3, p0 = r0, r0 mov f121 = f0 nop __LINE__ } ;; .align 16.L032: { .mfb lfetch.nt1 [PREA], 4 * SIZE FMA f64 = f32, f48, f64 // A1 * B1 nop __LINE__ } { .mfi nop __LINE__ FMA_B f65 = f32, f49, f65 // A1 * B2 (p12) cmp.ne p3, p0 = 0, L } ;; { .mfi lfetch.nt1 [PREB], 16 * SIZE FMA f80 = f32, f50, f80 // A1 * B3 cmp.ne p4, p5 = 0, L } { .mfb nop __LINE__ FMA_B f81 = f32, f51, f81 // A1 * B4 nop __LINE__ } ;; { .mfb (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE FMA f96 = f32, f52, f96 // A1 * B5 nop __LINE__ } { .mfb nop __LINE__ FMA_B f97 = f32, f53, f97 // A1 * B6 nop __LINE__ } ;; { .mfb (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE FMA f112 = f32, f54, f112 // A1 * B7 nop __LINE__ } { .mfb nop __LINE__ FMA_B f113 = f32, f55, f113 // A1 * B8 nop __LINE__ } ;; { .mfb (p3) LDFPD f58, f59 = [BOFFSET], 2 * SIZE FMA f65 = f33, f48, f65 // A2 * B1 nop __LINE__ } { .mfb nop __LINE__ FMA_A f64 = f33, f49, f64 // A2 * B2 nop __LINE__ } ;; { .mfb (p3) LDFPD f60, f61 = [BOFFSET], 2 * SIZE FMA f81 = f33, f50, f81 // A2 * B3 nop __LINE__ } { .mfb nop __LINE__ FMA_A f80 = f33, f51, f80 // A2 * B4 nop __LINE__ } ;; { .mfb (p3) LDFPD f62, f63 = [BOFFSET], 2 * SIZE FMA f97 = f33, f52, f97 // A2 * B5 nop __LINE__ } { .mfb nop __LINE__ FMA_A f96 = f33, f53, f96 // A2 * B6 nop __LINE__ } ;; { .mfb (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE FMA f113 = f33, f54, f113 // A2 * B7 nop __LINE__ } { .mfb nop __LINE__ FMA_A f112 = f33, f55, f112 // A2 * B8 nop __LINE__ } ;; { .mfb (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE (p3) FMA f64 = f40, f56, f64 // A1 * B1 nop __LINE__ } { .mfb nop __LINE__ (p3) FMA_B f65 = f40, f57, f65 // A1 * B2 nop __LINE__ } ;; { .mfb (p4) LDFPD f50, f51 = [BOFFSET], 2 * SIZE (p3) FMA f80 = f40, f58, f80 // A1 * B3 nop __LINE__ } { .mfb nop __LINE__ (p3) FMA_B f81 = f40, f59, f81 // A1 * B4 nop __LINE__ } ;; { .mfb (p4) LDFPD f52, f53 = [BOFFSET], 2 * SIZE (p3) FMA f96 = f40, f60, f96 // A1 * B5 nop __LINE__ } { .mfb nop __LINE__ (p3) FMA_B f97 = f40, f61, f97 // A1 * B6 nop __LINE__ } ;; { .mfb#if! defined(TRMMKERNEL) && !defined(BETAZERO) (p5) LDFD f72 = [C1], SIZE#else nop __LINE__#endif (p3) FMA f112 = f40, f62, f112 // A1 * B7 nop __LINE__ } { .mfb#if! defined(TRMMKERNEL) && !defined(BETAZERO) (p5) LDFD f88 = [C2], SIZE#else nop __LINE__#endif (p3) FMA_B f113 = f40, f63, f113 // A1 * B8 nop __LINE__ } ;; { .mfb (p4) LDFPD f54, f55 = [BOFFSET], 2 * SIZE (p3) FMA f65 = f41, f56, f65 // A2 * B1 nop __LINE__ } { .mfb nop __LINE__ (p3) FMA_A f64 = f41, f57, f64 // A2 * B2 nop __LINE__ } ;; { .mfb#if! defined(TRMMKERNEL) && !defined(BETAZERO) (p5) LDFD f73 = [C1], - SIZE#else nop __LINE__#endif (p3) FMA f81 = f41, f58, f81 // A2 * B3 nop __LINE__ } { .mfb#if! defined(TRMMKERNEL) && !defined(BETAZERO) (p5) LDFD f89 = [C2], - SIZE#else nop __LINE__#endif (p3) FMA_A f80 = f41, f59, f80 // A2 * B4 nop __LINE__ } ;; { .mfb#if! defined(TRMMKERNEL) && !defined(BETAZERO) (p5) LDFD f104 = [C3], SIZE#else nop __LINE__#endif (p3) FMA f97 = f41, f60, f97 // A2 * B5 nop __LINE__ } { .mfb#if! defined(TRMMKERNEL) && !defined(BETAZERO) (p5) LDFD f120 = [C4], SIZE#else nop __LINE__#endif (p3) FMA_A f96 = f41, f61, f96 // A2 * B6
⌨️ 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 + -