📄 ztrsm_kernel_ppc440_ln.s
字号:
LFD B3, 2 * SIZE(BO) LFD B4, 3 * SIZE(BO) LFD B5, 4 * SIZE(BO) LFD B6, 8 * SIZE(BO) LFD B7, 12 * SIZE(BO) lfs f0, FZERO fmr f1, f0 fmr f2, f0 fmr f3, f0 fmr f4, f0 fmr f5, f0 fmr f6, f0 fmr f7, f0 fmr f8, f0 fmr f9, f0 fmr f10, f0 fmr f11, f0 fmr f12, f0 fmr f13, f0 fmr f14, f0 fmr f15, f0 srawi. r0, TEMP, 2 mtspr CTR, r0#endif ble .L15 .align 4.L12: FMADD f0, A1, B1, f0 LFD A3, 2 * SIZE(AO) FMADD f4, A1, B2, f4 LFD A6, 12 * SIZE(AO) FMADD f8, A1, B3, f8 nop FMADD f12, A1, B4, f12 nop FMADD f1, A2, B1, f1 LFD A1, 3 * SIZE(AO) FMADD f5, A2, B2, f5 nop FMADD f9, A2, B3, f9 nop FMADD f13, A2, B4, f13 nop FMADD f2, A3, B1, f2 nop FMADD f6, A3, B2, f6 LFD B8, 5 * SIZE(BO) FMADD f10, A3, B3, f10 LFD B9, 6 * SIZE(BO) FMADD f14, A3, B4, f14 LFD B10, 7 * SIZE(BO) FMADD f3, A1, B1, f3 LFD A2, 5 * SIZE(AO) FMADD f7, A1, B2, f7 LFD B1, 16 * SIZE(BO) FMADD f11, A1, B3, f11 nop FMADD f15, A1, B4, f15 nop FMADD f0, A4, B5, f0 LFD A3, 6 * SIZE(AO) FMADD f4, A4, B8, f4 LFD A1, 16 * SIZE(AO) FMADD f8, A4, B9, f8 nop FMADD f12, A4, B10, f12 nop FMADD f1, A2, B5, f1 LFD A4, 7 * SIZE(AO) FMADD f5, A2, B8, f5 nop FMADD f9, A2, B9, f9 nop FMADD f13, A2, B10, f13 nop FMADD f2, A3, B5, f2 nop FMADD f6, A3, B8, f6 LFD B2, 9 * SIZE(BO) FMADD f10, A3, B9, f10 LFD B3, 10 * SIZE(BO) FMADD f14, A3, B10, f14 LFD B4, 11 * SIZE(BO) FMADD f3, A4, B5, f3 LFD A2, 9 * SIZE(AO) FMADD f7, A4, B8, f7 LFD B5, 20 * SIZE(BO) FMADD f11, A4, B9, f11 nop FMADD f15, A4, B10, f15 nop FMADD f0, A5, B6, f0 LFD A3, 10 * SIZE(AO) FMADD f4, A5, B2, f4 LFD A4, 20 * SIZE(AO) FMADD f8, A5, B3, f8 nop FMADD f12, A5, B4, f12 nop FMADD f1, A2, B6, f1 LFD A5, 11 * SIZE(AO) FMADD f5, A2, B2, f5 nop FMADD f9, A2, B3, f9 nop FMADD f13, A2, B4, f13 nop FMADD f2, A3, B6, f2 nop FMADD f6, A3, B2, f6 LFD B8, 13 * SIZE(BO) FMADD f10, A3, B3, f10 LFD B9, 14 * SIZE(BO) FMADD f14, A3, B4, f14 LFD B10,15 * SIZE(BO) FMADD f3, A5, B6, f3 LFD A2, 13 * SIZE(AO) FMADD f7, A5, B2, f7 LFD B6, 24 * SIZE(BO) FMADD f11, A5, B3, f11 nop FMADD f15, A5, B4, f15 nop FMADD f0, A6, B7, f0 LFD A3, 14 * SIZE(AO) FMADD f4, A6, B8, f4 LFD A5, 24 * SIZE(AO) FMADD f8, A6, B9, f8 nop FMADD f12, A6, B10, f12 nop FMADD f1, A2, B7, f1 LFD A6, 15 * SIZE(AO) FMADD f5, A2, B8, f5 nop FMADD f9, A2, B9, f9 nop FMADD f13, A2, B10, f13 nop FMADD f2, A3, B7, f2 addi AO, AO, 16 * SIZE FMADD f6, A3, B8, f6 LFD B2, 17 * SIZE(BO) FMADD f10, A3, B9, f10 LFD B3, 18 * SIZE(BO) FMADD f14, A3, B10, f14 LFD B4, 19 * SIZE(BO) FMADD f3, A6, B7, f3 LFD A2, 1 * SIZE(AO) FMADD f7, A6, B8, f7 LFD B7, 28 * SIZE(BO) FMADD f11, A6, B9, f11 addi BO, BO, 16 * SIZE FMADD f15, A6, B10, f15 bdnz .L12 .align 4.L15:#if defined(LT) || defined(RN) andi. r0, KK, 3#else andi. r0, TEMP, 3#endif mtspr CTR, r0 ble .LKERNEL_MainFinish .align 4.L16: FMADD f0, A1, B1, f0 LFD A3, 2 * SIZE(AO) FMADD f4, A1, B2, f4 FMADD f8, A1, B3, f8 FMADD f12, A1, B4, f12 LFD A4, 3 * SIZE(AO) FMADD f1, A2, B1, f1 FMADD f5, A2, B2, f5 FMADD f9, A2, B3, f9 FMADD f13, A2, B4, f13 LFDU A1, 4 * SIZE(AO) FMADD f2, A3, B1, f2 FMADD f6, A3, B2, f6 FMADD f10, A3, B3, f10 FMADD f14, A3, B4, f14 LFD A2, 1 * SIZE(AO) FMADD f3, A4, B1, f3 LFDU B1, 4 * SIZE(BO) FMADD f7, A4, B2, f7 LFD B2, 1 * SIZE(BO) FMADD f11, A4, B3, f11 LFD B3, 2 * SIZE(BO) FMADD f15, A4, B4, f15 LFD B4, 3 * SIZE(BO) bdnz .L16 .align 4.LKERNEL_MainFinish:#ifndef CONJ FSUB f0, f0, f5 FADD f1, f1, f4 FSUB f2, f2, f7 FADD f3, f3, f6 FSUB f8, f8, f13 FADD f9, f9, f12 FSUB f10, f10, f15 FADD f11, f11, f14#else FADD f0, f0, f5 FSUB f1, f4, f1 FADD f2, f2, f7 FSUB f3, f6, f3 FADD f8, f8, f13 FSUB f9, f12, f9 FADD f10, f10, f15 FSUB f11, f14, f11#endif#if defined(LN) || defined(RT) subi r0, KK, 2 slwi r0, r0, 1 + ZBASE_SHIFT add AO, AORIG, r0 add BO, B, r0#endif#if defined(LN) || defined(LT) LFD f16, 0 * SIZE(BO) LFD f17, 1 * SIZE(BO) LFD f18, 2 * SIZE(BO) LFD f19, 3 * SIZE(BO) LFD f20, 4 * SIZE(BO) LFD f21, 5 * SIZE(BO) LFD f22, 6 * SIZE(BO) LFD f23, 7 * SIZE(BO) FSUB f0, f16, f0 FSUB f1, f17, f1 FSUB f8, f18, f8 FSUB f9, f19, f9 FSUB f2, f20, f2 FSUB f3, f21, f3 FSUB f10, f22, f10 FSUB f11, f23, f11#else LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) LFD f18, 2 * SIZE(AO) LFD f19, 3 * SIZE(AO) LFD f20, 4 * SIZE(AO) LFD f21, 5 * SIZE(AO) LFD f22, 6 * SIZE(AO) LFD f23, 7 * SIZE(AO)#ifndef CONJ FSUB f0, f16, f0 FSUB f1, f17, f1 FSUB f2, f18, f2 FSUB f3, f19, f3 FSUB f8, f20, f8 FSUB f9, f21, f9 FSUB f10, f22, f10 FSUB f11, f23, f11#else FSUB f0, f16, f0 FADD f1, f17, f1 FSUB f2, f18, f2 FADD f3, f19, f3 FSUB f8, f20, f8 FADD f9, f21, f9 FSUB f10, f22, f10 FADD f11, f23, f11#endif#endif#ifdef LN LFD f16, 6 * SIZE(AO) LFD f17, 7 * SIZE(AO) LFD f18, 4 * SIZE(AO) LFD f19, 5 * SIZE(AO) LFD f20, 0 * SIZE(AO) LFD f21, 1 * SIZE(AO) FMUL f6, f17, f3 FMUL f7, f17, f2 FMUL f14, f17, f11 FMUL f15, f17, f10#ifndef CONJ FMSUB f2, f16, f2, f6 FMADD f3, f16, f3, f7 FMSUB f10, f16, f10, f14 FMADD f11, f16, f11, f15 FMADD f0, f19, f3, f0 FNMSUB f1, f19, f2, f1 FMADD f8, f19, f11, f8 FNMSUB f9, f19, f10, f9 FNMSUB f0, f18, f2, f0 FNMSUB f1, f18, f3, f1 FNMSUB f8, f18, f10, f8 FNMSUB f9, f18, f11, f9 FMUL f4, f21, f1 FMUL f5, f21, f0 FMUL f12, f21, f9 FMUL f13, f21, f8 FMSUB f0, f20, f0, f4 FMADD f1, f20, f1, f5 FMSUB f8, f20, f8, f12 FMADD f9, f20, f9, f13#else FMADD f2, f16, f2, f6 FMSUB f3, f16, f3, f7 FMADD f10, f16, f10, f14 FMSUB f11, f16, f11, f15 FMSUB f0, f19, f3, f0 FNMADD f1, f19, f2, f1 FMSUB f8, f19, f11, f8 FNMADD f9, f19, f10, f9 FNMADD f0, f18, f2, f0 FNMADD f1, f18, f3, f1 FNMADD f8, f18, f10, f8 FNMADD f9, f18, f11, f9 FMUL f4, f21, f1 FMUL f5, f21, f0 FMUL f12, f21, f9 FMUL f13, f21, f8 FMADD f0, f20, f0, f4 FMSUB f1, f20, f1, f5 FMADD f8, f20, f8, f12 FMSUB f9, f20, f9, f13#endif#endif#ifdef LT LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) LFD f18, 2 * SIZE(AO) LFD f19, 3 * SIZE(AO) LFD f20, 6 * SIZE(AO) LFD f21, 7 * SIZE(AO) FMUL f4, f17, f1 FMUL f5, f17, f0 FMUL f12, f17, f9 FMUL f13, f17, f8#ifndef CONJ FMSUB f0, f16, f0, f4 FMADD f1, f16, f1, f5 FMSUB f8, f16, f8, f12 FMADD f9, f16, f9, f13 FMADD f2, f19, f1, f2 FNMSUB f3, f19, f0, f3 FMADD f10, f19, f9, f10 FNMSUB f11, f19, f8, f11 FNMSUB f2, f18, f0, f2 FNMSUB f3, f18, f1, f3 FNMSUB f10, f18, f8, f10 FNMSUB f11, f18, f9, f11 FMUL f4, f21, f3 FMUL f5, f21, f2 FMUL f12, f21, f11 FMUL f13, f21, f10 FMSUB f2, f20, f2, f4 FMADD f3, f20, f3, f5 FMSUB f10, f20, f10, f12 FMADD f11, f20, f11, f13#else FMADD f0, f16, f0, f4 FMSUB f1, f16, f1, f5 FMADD f8, f16, f8, f12 FMSUB f9, f16, f9, f13 FMSUB f2, f19, f1, f2 FNMADD f3, f19, f0, f3 FMSUB f10, f19, f9, f10 FNMADD f11, f19, f8, f11 FNMADD f2, f18, f0, f2 FNMADD f3, f18, f1, f3 FNMADD f10, f18, f8, f10 FNMADD f11, f18, f9, f11 FMUL f4, f21, f3 FMUL f5, f21, f2 FMUL f12, f21, f11 FMUL f13, f21, f10 FMADD f2, f20, f2, f4 FMSUB f3, f20, f3, f5 FMADD f10, f20, f10, f12 FMSUB f11, f20, f11, f13#endif#endif#ifdef RN LFD f16, 0 * SIZE(BO) LFD f17, 1 * SIZE(BO) LFD f18, 2 * SIZE(BO) LFD f19, 3 * SIZE(BO) LFD f20, 6 * SIZE(BO) LFD f21, 7 * SIZE(BO) FMUL f4, f17, f1 FMUL f5, f17, f0 FMUL f6, f17, f3 FMUL f7, f17, f2#ifndef CONJ FMSUB f0, f16, f0, f4 FMADD f1, f16, f1, f5 FMSUB f2, f16, f2, f6 FMADD f3, f16, f3, f7 FMADD f8, f19, f1, f8 FNMSUB f9, f19, f0, f9 FMADD f10, f19, f3, f10 FNMSUB f11, f19, f2, f11 FNMSUB f8, f18, f0, f8 FNMSUB f9, f18, f1, f9 FNMSUB f10, f18, f2, f10 FNMSUB f11, f18, f3, f11 FMUL f4, f21, f9 FMUL f5, f21, f8 FMUL f6, f21, f11 FMUL f7, f21, f10 FMSUB f8, f20, f8, f4 FMADD f9, f20, f9, f5 FMSUB f10, f20, f10, f6 FMADD f11, f20, f11, f7#else FMADD f0, f16, f0, f4 FMSUB f1, f16, f1, f5 FMADD f2, f16, f2, f6 FMSUB f3, f16, f3, f7 FMSUB f8, f19, f1, f8 FNMADD f9, f19, f0, f9 FMSUB f10, f19, f3, f10 FNMADD f11, f19, f2, f11 FNMADD f8, f18, f0, f8 FNMADD f9, f18, f1, f9 FNMADD f10, f18, f2, f10 FNMADD f11, f18, f3, f11 FMUL f4, f21, f9 FMUL f5, f21, f8 FMUL f6, f21, f11 FMUL f7, f21, f10 FMADD f8, f20, f8, f4 FMSUB f9, f20, f9, f5 FMADD f10, f20, f10, f6 FMSUB f11, f20, f11, f7#endif#endif#ifdef RT LFD f16, 6 * SIZE(BO) LFD f17, 7 * SIZE(BO) LFD f18, 4 * SIZE(BO) LFD f19, 5 * SIZE(BO) LFD f20, 0 * SIZE(BO) LFD f21, 1 * SIZE(BO) FMUL f12, f17, f9 FMUL f13, f17, f8 FMUL f14, f17, f11 FMUL f15, f17, f10#ifndef CONJ FMSUB f8, f16, f8, f12 FMADD f9, f16, f9, f13 FMSUB f10, f16, f10, f14 FMADD f11, f16, f11, f15 FMADD f0, f19, f9, f0 FNMSUB f1, f19, f8, f1 FMADD f2, f19, f11, f2 FNMSUB f3, f19, f10, f3 FNMSUB f0, f18, f8, f0 FNMSUB f1, f18, f9, f1 FNMSUB f2, f18, f10, f2 FNMSUB f3, f18, f11, f3 FMUL f4, f21, f1 FMUL f5, f21, f0 FMUL f6, f21, f3 FMUL f7, f21, f2 FMSUB f0, f20, f0, f4 FMADD f1, f20, f1, f5 FMSUB f2, f20, f2, f6 FMADD f3, f20, f3, f7#else FMADD f8, f16, f8, f12 FMSUB f9, f16, f9, f13 FMADD f10, f16, f10, f14 FMSUB f11, f16, f11, f15 FMSUB f0, f19, f9, f0 FNMADD f1, f19, f8, f1 FMSUB f2, f19, f11, f2 FNMADD f3, f19, f10, f3 FNMADD f0, f18, f8, f0 FNMADD f1, f18, f9, f1 FNMADD f2, f18, f10, f2 FNMADD f3, f18, f11, f3 FMUL f4, f21, f1 FMUL f5, f21, f0 FMUL f6, f21, f3 FMUL f7, f21, f2 FMADD f0, f20, f0, f4 FMSUB f1, f20, f1, f5 FMADD f2, f20, f2, f6 FMSUB f3, f20, f3, f7#endif#endif#ifdef LN subi CO1, CO1, 4 * SIZE subi CO2, CO2, 4 * SIZE#endif#if defined(LN) || defined(LT) STFD f0, 0 * SIZE(BO) STFD f1, 1 * SIZE(BO) STFD f8, 2 * SIZE(BO) STFD f9, 3 * SIZE(BO) STFD f2, 4 * SIZE(BO) STFD f3, 5 * SIZE(BO) STFD f10, 6 * SIZE(BO) STFD f11, 7 * SIZE(BO)#else STFD f0, 0 * SIZE(AO) STFD f1, 1 * SIZE(AO) STFD f2, 2 * SIZE(AO) STFD f3, 3 * SIZE(AO) STFD f8, 4 * SIZE(AO) STFD f9, 5 * SIZE(AO) STFD f10, 6 * SIZE(AO) STFD f11, 7 * SIZE(AO)#endif STFD f0, 0 * SIZE(CO1) STFD f1, 1 * SIZE(CO1) STFD f2, 2 * SIZE(CO1) STFD f3, 3 * SIZE(CO1) STFD f8, 0 * SIZE(CO2) STFD f9, 1 * SIZE(CO2) STFD f10, 2 * SIZE(CO2) STFD f11, 3 * SIZE(CO2)#ifndef LN addi CO1, CO1, 4 * SIZE addi CO2, CO2, 4 * SIZE#endif #ifdef RT slwi r0, K, 1 + ZBASE_SHIFT add AORIG, AORIG, r0#endif#if defined(LT) || defined(RN) sub TEMP, K, KK slwi TEMP, TEMP, 1 + ZBASE_SHIFT add AO, AO, TEMP add BO, BO, TEMP#endif#ifdef LT addi KK, KK, 2#endif#ifdef LN subi KK, KK, 2#endif addic. I, I, -1 bgt .L11 .align 4.L29:#ifdef LN slwi r0, K, 1 + ZBASE_SHIFT add B, B, r0#endif#if defined(LT) || defined(RN) mr B, BO#endif#ifdef RN addi KK, KK, 2#endif#ifdef RT subi KK, KK, 2#endif addic. J, J, -1 bgt .L10 .align 4.L30: andi. J, N, 1 ble .L999#ifdef RT slwi r0, K, 0 + ZBASE_SHIFT sub B, B, r0 sub C, C, LDC#endif mr CO1, C#ifdef LN add KK, M, OFFSET#endif#ifdef LT mr KK, OFFSET#endif#if defined(LN) || defined(RT) mr AORIG, A#else mr AO, A#endif#ifndef RT add C, C, LDC#endif andi. I, M, 1 ble .L40#if defined(LT) || defined(RN) LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) LFD f18, 2 * SIZE(AO) LFD f19, 3 * SIZE(AO) LFD f20, 0 * SIZE(B) LFD f21, 1 * SIZE(B) LFD f22, 2 * SIZE(B) LFD f23, 3 * SIZE(B) lfs f0, FZERO fmr f1, f0 fmr f2, f0 fmr f3, f0 fmr f4, f0 fmr f5, f0 fmr f6, f0 fmr f7, f0 srawi. r0, KK, 2 mr BO, B mtspr CTR, r0#else#ifdef LN slwi r0, K, 0 + ZBASE_SHIFT sub AORIG, AORIG, r0#endif slwi r0, KK, 0 + ZBASE_SHIFT add AO, AORIG, r0 add BO, B, r0 sub TEMP, K, KK LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) LFD f18, 2 * SIZE(AO) LFD f19, 3 * SIZE(AO) LFD f20, 0 * SIZE(BO) LFD f21, 1 * SIZE(BO) LFD f22, 2 * SIZE(BO) LFD f23, 3 * SIZE(BO) lfs f0, FZERO fmr f1, f0 fmr f2, f0 fmr f3, f0 fmr f4, f0 fmr f5, f0 fmr f6, f0 fmr f7, f0
⌨️ 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 + -