📄 gemm_tcopy.s
字号:
(p12) LD f77 = [A2], TEMP1 (p18) add B2 = B2, TEMP2 } ;; { .mmi (p13) ST [B1] = f58, 1 * SIZE (p13) ST [B2] = f70, 1 * SIZE adds TEMP3 = 5 * SIZE, TEMP3 } { .mmi (p16) LD f80 = [A1], SIZE (p16) LD f92 = [A2], SIZE adds TEMP1 = -3 * SIZE, LDA } ;; { .mmi (p13) ST [B1] = f61, 1 * SIZE (p13) ST [B2] = f73, 1 * SIZE nop __LINE__ } { .mmi (p16) LD f83 = [A1], SIZE (p16) LD f95 = [A2], SIZE (p14) mov TEMP1 = TEMP3 } ;; { .mmi (p13) ST [B1] = f64, 1 * SIZE (p13) ST [B2] = f76, 1 * SIZE nop __LINE__ } { .mmi (p16) LD f86 = [A1], SIZE (p16) LD f98 = [A2], SIZE (p12) mov TEMP1 = 5 * SIZE } ;; { .mmi (p13) ST [B1] = f67, 5 * SIZE (p13) ST [B2] = f79, 5 * SIZE (p14) mov LCOUNT = 0 } { .mmi (p16) LD f89 = [A1], TEMP1 (p16) LD f101 = [A2], TEMP1 (p15) mov SCOUNT = 0 } ;; { .mmi (p13) ST [B1] = f106, 1 * SIZE (p13) ST [B2] = f118, 1 * SIZE mov TEMP2 = 5 * SIZE } { .mmi (p12) LD f104 = [A1], SIZE (p12) LD f116 = [A2], SIZE nop __LINE__ } ;; { .mmi (p13) ST [B1] = f109, 1 * SIZE (p13) ST [B2] = f121, 1 * SIZE sub TEMP2 = TEMP2, LDB } { .mmi (p12) LD f107 = [A1], SIZE (p12) LD f119 = [A2], SIZE adds TEMP1 = -11 * SIZE, LDA } ;; { .mmi (p13) ST [B1] = f112, 1 * SIZE (p13) ST [B2] = f124, 1 * SIZE (p15) adds TEMP2 = -27 * SIZE, LDB } { .mmi (p12) LD f110 = [A1], SIZE (p12) LD f122 = [A2], SIZE (p14) mov TEMP1 = TEMP3 } ;; { .mmi (p13) ST [B1] = f115 (p13) ST [B2] = f127 (p13) add B1 = B1, TEMP2 } { .mmi (p12) LD f113 = [A1], TEMP1 (p12) LD f125 = [A2], TEMP1 (p13) add B2 = B2, TEMP2 } ;; { .mmb (p14) adds I = -2, I (p15) adds II = -2, II br.ctop.sptk .L112 } ;; .align 32.L120: { .mmi add A2 = A1, LDA nop __LINE__ tbit.nz p7, p0 = N, 2 } ;; { .mmi (p7) LD f32 = [A1], SIZE (p7) LD f36 = [A2], SIZE tbit.nz p8, p0 = N, 1 } ;; { .mmi (p7) LD f33 = [A1], SIZE (p7) LD f37 = [A2], SIZE adds TEMP1 = -3 * SIZE, LDA } ;; { .mmi (p7) LD f34 = [A1], SIZE (p7) LD f38 = [A2], SIZE add TEMP1 = TEMP1, LDA } ;; { .mmi (p7) LD f35 = [A1], TEMP1 (p7) LD f39 = [A2], TEMP1 tbit.nz p9, p0 = N, 0 } ;; { .mmi (p7) LD f40 = [A1], SIZE (p7) LD f44 = [A2], SIZE mov TEMP2 = -1 * SIZE } ;; { .mmi (p7) LD f41 = [A1], SIZE (p7) LD f45 = [A2], SIZE shladd TEMP2 = LDA, 1, TEMP2 } ;; { .mmi (p7) LD f42 = [A1], SIZE (p7) LD f46 = [A2], SIZE sub TEMP2 = 0, TEMP2 } ;; { .mmi (p7) LD f43 = [A1], TEMP2 (p7) LD f47 = [A2] nop __LINE__ } ;; { .mmi add A2 = A1, LDA adds TEMP1 = -1 * SIZE, LDA mov TEMP2 = -1 * SIZE } ;; { .mmi (p8) LD f48 = [A1], SIZE (p8) LD f50 = [A2], SIZE add TEMP1 = TEMP1, LDA } ;; { .mmi (p8) LD f49 = [A1], TEMP1 (p8) LD f51 = [A2], TEMP1 shladd TEMP2 = LDA, 1, TEMP2 } ;; { .mmi (p8) LD f52 = [A1], SIZE (p8) LD f54 = [A2], SIZE sub TEMP2 = r0, TEMP2 } ;; { .mmi (p8) LD f53 = [A1], TEMP2 (p8) LD f55 = [A2], TEMP2 nop __LINE__ } ;; { .mmi add A2 = A1, LDA adds B2 = 4 * SIZE, BO2 nop __LINE__ } ;; { .mmi (p9) LD f56 = [A1] nop __LINE__ (p9) shladd A1 = LDA, 1, A1 } { .mmi (p9) LD f57 = [A2] nop __LINE__ (p9) shladd A2 = LDA, 1, A2 } ;; { .mmi (p7) ST [BO2] = f32, 1 * SIZE (p7) ST [B2 ] = f36, 1 * SIZE nop __LINE__ } { .mmi (p9) LD f58 = [A1] (p9) LD f59 = [A2] nop __LINE__ } ;; ;; { .mmi (p7) ST [BO2] = f33, 1 * SIZE (p7) ST [B2 ] = f37, 1 * SIZE nop __LINE__ } ;; { .mmi (p7) ST [BO2] = f34, 1 * SIZE (p7) ST [B2 ] = f38, 1 * SIZE nop __LINE__ } ;; { .mmi (p7) ST [BO2] = f35, 5 * SIZE (p7) ST [B2 ] = f39, 5 * SIZE nop __LINE__ } ;; { .mmi (p7) ST [BO2] = f40, 1 * SIZE (p7) ST [B2 ] = f44, 1 * SIZE nop __LINE__ } ;; { .mmi (p7) ST [BO2] = f41, 1 * SIZE (p7) ST [B2 ] = f45, 1 * SIZE nop __LINE__ } ;; { .mmi (p7) ST [BO2] = f42, 1 * SIZE (p7) ST [B2 ] = f46, 1 * SIZE nop __LINE__ } ;; { .mmi (p7) ST [BO2] = f43, 5 * SIZE (p7) ST [B2 ] = f47 adds B2 = 4 * SIZE, BO3 } ;; { .mmi (p8) ST [BO3] = f48, 1 * SIZE (p8) ST [B2 ] = f52, 1 * SIZE nop __LINE__ } ;; { .mmi (p8) ST [BO3] = f49, 1 * SIZE (p8) ST [B2 ] = f53, 1 * SIZE nop __LINE__ } ;; { .mmi (p8) ST [BO3] = f50, 1 * SIZE (p8) ST [B2 ] = f54, 1 * SIZE nop __LINE__ } ;; { .mmi (p8) ST [BO3] = f51, 5 * SIZE (p8) ST [B2 ] = f55 adds B2 = 2 * SIZE, BO4 } ;; { .mmi (p9) ST [BO4] = f56, 1 * SIZE (p9) ST [B2 ] = f58, 1 * SIZE nop __LINE__ } ;; { .mmi (p9) ST [BO4] = f57, 3 * SIZE (p9) ST [B2 ] = f59 nop __LINE__ } ;; .align 32.L200: { .mmi add I = 8, N mov A1 = A mov pr.rot = 0 } { .mmi adds A2 = 4 * SIZE, A nop __LINE__ tbit.z p6, p0 = M, 1 } ;; { .mmi mov B1 = B cmp.eq p16, p0 = r0, r0 mov ar.ec = 3 } { .mib adds B2 = 4 * SIZE, B shr I = I, 4 (p6) br.cond.dpnt .L300 } ;; { .mmi shladd A = LDA, 1, A adds B = 16 * SIZE, B shr II = N, 3 } { .mmi cmp.eq p8, p0 = 0, I adds I = -1, I nop __LINE__ } ;; { .mmi nop __LINE__ nop __LINE__ mov ar.lc = I } { .mib mov I = II nop __LINE__ (p8) br.cond.dpnt .L220 } ;; .align 32.L212: { .mmi (p18) ST [B1] = f34, 1 * SIZE (p18) ST [B2] = f46, 1 * SIZE (p16) cmp.ne.unc p12, p0 = 1, I } { .mmi (p16) LD f32 = [A1], SIZE (p16) LD f44 = [A2], SIZE (p18) cmp.ne.unc p13, p0 = 1, II } ;; { .mmi (p18) ST [B1] = f37, 1 * SIZE (p18) ST [B2] = f49, 1 * SIZE adds TEMP1 = -3 * SIZE, LDA } { .mmi (p16) LD f35 = [A1], SIZE (p16) LD f47 = [A2], SIZE nop __LINE__ } ;; { .mmi (p18) ST [B1] = f40, 1 * SIZE (p18) ST [B2] = f52, 1 * SIZE (p12) mov TEMP1 = 5 * SIZE } { .mmi (p16) LD f38 = [A1], SIZE (p16) LD f50 = [A2], SIZE nop __LINE__ } ;; { .mmi (p18) ST [B1] = f43, 5 * SIZE (p18) ST [B2] = f55, 5 * SIZE nop __LINE__ } { .mmi (p16) LD f41 = [A1], TEMP1 (p16) LD f53 = [A2], TEMP1 nop __LINE__ } ;; { .mmi (p18) ST [B1] = f82, 1 * SIZE (p18) ST [B2] = f94, 1 * SIZE nop __LINE__ } { .mmi (p12) LD f56 = [A1], SIZE (p12) LD f68 = [A2], SIZE nop __LINE__ } ;; { .mmi (p18) ST [B1] = f85, 1 * SIZE (p18) ST [B2] = f97, 1 * SIZE mov TEMP2 = 5 * SIZE } { .mmi (p12) LD f59 = [A1], SIZE (p12) LD f71 = [A2], SIZE nop __LINE__ } ;; { .mmi (p18) ST [B1] = f88, 1 * SIZE (p18) ST [B2] = f100, 1 * SIZE (p13) adds TEMP2 = - 11 * SIZE, LDB } { .mmi (p12) LD f62 = [A1], SIZE (p12) LD f74 = [A2], SIZE (p12) adds TEMP1 = - 11 * SIZE, LDA } ;; { .mmi (p18) ST [B1] = f91 (p18) ST [B2] = f103 (p18) add B1 = B1, TEMP2 } { .mmi (p12) LD f65 = [A1], TEMP1 (p12) LD f77 = [A2], TEMP1 (p18) add B2 = B2, TEMP2 } ;; { .mmi (p13) ST [B1] = f58, 1 * SIZE (p13) ST [B2] = f70, 1 * SIZE nop __LINE__ } { .mmi (p16) LD f80 = [A1], SIZE (p16) LD f92 = [A2], SIZE sub TEMP1 = r0, LDA } ;; { .mmi (p13) ST [B1] = f61, 1 * SIZE (p13) ST [B2] = f73, 1 * SIZE nop __LINE__ } { .mmi (p16) LD f83 = [A1], SIZE (p16) LD f95 = [A2], SIZE (p16) adds TEMP1 = 5 * SIZE, TEMP1 } ;; { .mmi (p13) ST [B1] = f64, 1 * SIZE (p13) ST [B2] = f76, 1 * SIZE nop __LINE__ } { .mmi (p16) LD f86 = [A1], SIZE (p16) LD f98 = [A2], SIZE (p12) mov TEMP1 = 5 * SIZE } ;; { .mmi (p13) ST [B1] = f67, 5 * SIZE (p13) ST [B2] = f79, 5 * SIZE nop __LINE__ } { .mmi (p16) LD f89 = [A1], TEMP1 (p16) LD f101 = [A2], TEMP1 adds TEMP1 = -11 * SIZE, LDA } ;; { .mmi (p13) ST [B1] = f106, 1 * SIZE (p13) ST [B2] = f118, 1 * SIZE mov TEMP2 = 5 * SIZE } { .mmi (p12) LD f104 = [A1], SIZE (p12) LD f116 = [A2], SIZE (p16) shladd TEMP1 = LDA, 1, r0 } ;; { .mmi (p13) ST [B1] = f109, 1 * SIZE (p13) ST [B2] = f121, 1 * SIZE sub TEMP2 = TEMP2, LDB } { .mmi (p12) LD f107 = [A1], SIZE (p12) LD f119 = [A2], SIZE (p16) sub TEMP1 = LDA, TEMP1 } ;; { .mmi (p13) ST [B1] = f112, 1 * SIZE (p13) ST [B2] = f124, 1 * SIZE (p18) adds TEMP2 = -11 * SIZE, LDB } { .mmi (p12) LD f110 = [A1], SIZE (p12) LD f122 = [A2], SIZE (p16) adds TEMP1 = 5 * SIZE, TEMP1 } ;; { .mmi (p13) ST [B1] = f115 (p13) ST [B2] = f127 (p13) add B1 = B1, TEMP2 } { .mmi (p12) LD f113 = [A1], TEMP1 (p12) LD f125 = [A2], TEMP1 (p13) add B2 = B2, TEMP2 } ;; { .mmb (p16) adds I = -2, I (p18) adds II = -2, II br.ctop.sptk .L212 } ;; .align 32.L220: { .mmi add A2 = A1, LDA nop __LINE__ tbit.nz p7, p0 = N, 2 } ;; { .mmi (p7) LD f32 = [A1], SIZE (p7) LD f36 = [A2], SIZE tbit.nz p8, p0 = N, 1 } ;; { .mmi (p7) LD f33 = [A1], SIZE (p7) LD f37 = [A2], SIZE tbit.nz p9, p0 = N, 0 } ;; { .mmi (p7) LD f34 = [A1], SIZE (p7) LD f38 = [A2], SIZE nop __LINE__ } ;; { .mmi (p7) LD f35 = [A1], SIZE (p7) LD f39 = [A2] nop __LINE__ } ;; { .mmi add A2 = A1, LDA nop __LINE__ nop __LINE__ } ;; { .mmi (p8) LD f40 = [A1], SIZE (p8) LD f42 = [A2], SIZE nop __LINE__ } ;; { .mmi (p8) LD f41 = [A1], SIZE (p8) LD f43 = [A2] nop __LINE__ } ;; { .mmi add A2 = A1, LDA nop __LINE__ nop __LINE__ } ;; { .mmi (p9) LD f44 = [A1] (p9) LD f45 = [A2] adds B2 = 4 * SIZE, BO2 } ;; { .mmi (p7) ST [BO2] = f32, 1 * SIZE (p7) ST [B2 ] = f36, 1 * SIZE nop __LINE__ } ;; { .mmi (p7) ST [BO2] = f33, 1 * SIZE (p7) ST [B2 ] = f37, 1 * SIZE nop __LINE__ } ;; { .mmi (p7) ST [BO2] = f34, 1 * SIZE (p7) ST [B2 ] = f38, 1 * SIZE nop __LINE__ } ;; { .mmi (p7) ST [BO2] = f35, 5 * SIZE (p7) ST [B2 ] = f39 adds B2 = 2 * SIZE, BO3 } ;; { .mmi (p8) ST [BO3] = f40, 1 * SIZE (p8) ST [B2 ] = f42, 1 * SIZE nop __LINE__ } ;; { .mmi (p8) ST [BO3] = f41, 3 * SIZE (p8) ST [B2 ] = f43 adds B2 = 1 * SIZE, BO4 } ;; { .mmi (p9) ST [BO4] = f44, 2 * SIZE (p9) ST [B2 ] = f45 nop __LINE__ } ;; .align 32.L300: { .mmi add I = 8, N mov A1 = A mov pr.rot = 0 } { .mmi mov B1 = B adds A2 = 4 * SIZE, A tbit.z p6, p0 = M, 0 } ;; { .mmi adds B2 = 4 * SIZE, B cmp.eq p16, p0 = r0, r0 mov ar.ec = 3 } { .mib nop __LINE__ shr I = I, 4 (p6) br.cond.dpnt .L999 } ;; { .mmi cmp.eq p8, p0 = 0, I adds I = -1, I shr II = N, 3 } ;; { .mmi nop __LINE__ nop __LINE__ mov ar.lc = I } { .mib nop __LINE__ mov I = II (p8) br.cond.dpnt .L320 } ;; .align 32.L312: { .mmi (p18) ST [B1] = f34, 1 * SIZE (p18) ST [B2] = f46, 1 * SIZE (p16) cmp.ne.unc p12, p0 = 1, I } { .mmi (p16) LD f32 = [A1], SIZE (p16) LD f44 = [A2], SIZE (p18) cmp.ne.unc p13, p0 = 1, II } ;; { .mmi (p18) ST [B1] = f37, 1 * SIZE (p18) ST [B2] = f49, 1 * SIZE adds TEMP2 = - 3 * SIZE, LDB } { .mmi (p16) LD f35 = [A1], SIZE (p16) LD f47 = [A2], SIZE nop __LINE__ } ;; { .mmi (p18) ST [B1] = f40, 1 * SIZE (p18) ST [B2] = f52, 1 * SIZE nop __LINE__ } { .mmi (p16) LD f38 = [A1], SIZE (p16) LD f50 = [A2], SIZE nop __LINE__ } ;; { .mmi (p18) ST [B1] = f43 (p18) ST [B2] = f55 (p18) add B1 = B1, TEMP2 } { .mmi (p16) LD f41 = [A1], 5 * SIZE (p16) LD f53 = [A2], 5 * SIZE (p18) add B2 = B2, TEMP2 } ;; { .mmi (p13) ST [B1] = f58, 1 * SIZE (p13) ST [B2] = f70, 1 * SIZE (p16) adds I = -2, I } { .mmi (p12) LD f56 = [A1], SIZE (p12) LD f68 = [A2], SIZE (p18) adds II = -2, II } ;; { .mmi (p13) ST [B1] = f61, 1 * SIZE (p13) ST [B2] = f73, 1 * SIZE nop __LINE__ } { .mmi (p12) LD f59 = [A1], SIZE (p12) LD f71 = [A2], SIZE nop __LINE__ } ;; { .mmi (p13) ST [B1] = f64, 1 * SIZE (p13) ST [B2] = f76, 1 * SIZE nop __LINE__ } { .mmi (p12) LD f62 = [A1], SIZE (p12) LD f74 = [A2], SIZE nop __LINE__ } ;; { .mmi (p13) ST [B1] = f67 (p13) ST [B2] = f79 (p13) add B1 = B1, TEMP2 } { .mmi (p12) LD f65 = [A1], 5 * SIZE (p12) LD f77 = [A2], 5 * SIZE (p13) add B2 = B2, TEMP2 } ;; { .mmb nop __LINE__ nop __LINE__ br.ctop.sptk .L312 } ;; .align 32.L320: { .mmi adds A2 = 2 * SIZE, A1 adds B2 = 2 * SIZE, BO2 tbit.nz p7, p0 = N, 2 } ;; { .mmi (p7) LD f32 = [A1], SIZE (p7) LD f34 = [A2], SIZE tbit.nz p8, p0 = N, 1 } ;; { .mmi (p7) LD f33 = [A1], 3 * SIZE (p7) LD f35 = [A2] nop __LINE__ } ;; { .mmi adds A2 = SIZE, A1 nop __LINE__ nop __LINE__ } ;; { .mmi (p8) LD f36 = [A1], 2 * SIZE (p8) LD f37 = [A2] tbit.nz p9, p0 = N, 0 } ;; { .mmi (p9) LD f38 = [A1] nop __LINE__ nop __LINE__ } ;; { .mmi (p7) ST [BO2] = f32, 1 * SIZE (p7) ST [B2 ] = f34, 1 * SIZE nop __LINE__ } ;; { .mmi (p7) ST [BO2] = f33, 3 * SIZE (p7) ST [B2 ] = f35 adds B2 = SIZE, BO3 } ;; { .mmi (p8) ST [BO3] = f36, 2 * SIZE (p8) ST [B2 ] = f37 nop __LINE__ } ;; { .mmi (p9) ST [BO4] = f38, 1 * SIZE nop __LINE__ nop __LINE__ } ;; .align 32.L999: mov pr = PR, -1 mov ar.lc = ARLC 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 + -