📄 ztrsm_kernel_ppc440_lt.s
字号:
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.L20: andi. I, M, 1 ble .L29#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) LFD f24, 4 * SIZE(B) LFD f25, 5 * SIZE(B) LFD f26, 6 * SIZE(B) LFD f27, 7 * 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 slwi TEMP, KK, 1 + ZBASE_SHIFT add AO, AORIG, r0 add BO, B, TEMP 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) LFD f24, 4 * SIZE(BO) LFD f25, 5 * SIZE(BO) LFD f26, 6 * SIZE(BO) LFD f27, 7 * 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 srawi. r0, TEMP, 2 mtspr CTR, r0#endif ble .L25 .align 4.L22: fmadd f0, f16, f20, f0 LFD f19, 3 * SIZE(AO) fmadd f1, f16, f21, f1 nop fmadd f2, f16, f22, f2 nop fmadd f3, f16, f23, f3 LFD f16, 4 * SIZE(AO) fmadd f4, f17, f20, f4 LFD f20, 8 * SIZE(BO) fmadd f5, f17, f21, f5 LFD f21, 9 * SIZE(BO) fmadd f6, f17, f22, f6 LFD f22, 10 * SIZE(BO) fmadd f7, f17, f23, f7 LFD f23, 11 * SIZE(BO) fmadd f0, f18, f24, f0 LFD f17, 5 * SIZE(AO) fmadd f1, f18, f25, f1 nop fmadd f2, f18, f26, f2 nop fmadd f3, f18, f27, f3 LFD f18, 6 * SIZE(AO) fmadd f4, f19, f24, f4 LFD f24, 12 * SIZE(BO) fmadd f5, f19, f25, f5 LFD f25, 13 * SIZE(BO) fmadd f6, f19, f26, f6 LFD f26, 14 * SIZE(BO) fmadd f7, f19, f27, f7 LFD f27, 15 * SIZE(BO) fmadd f0, f16, f20, f0 LFD f19, 7 * SIZE(AO) fmadd f1, f16, f21, f1 nop fmadd f2, f16, f22, f2 nop fmadd f3, f16, f23, f3 LFDU f16, 8 * SIZE(AO) fmadd f4, f17, f20, f4 LFDU f20, 16 * SIZE(BO) fmadd f5, f17, f21, f5 LFD f21, 1 * SIZE(BO) fmadd f6, f17, f22, f6 LFD f22, 2 * SIZE(BO) fmadd f7, f17, f23, f7 LFD f23, 3 * SIZE(BO) fmadd f0, f18, f24, f0 LFD f17, 1 * SIZE(AO) fmadd f1, f18, f25, f1 nop fmadd f2, f18, f26, f2 nop fmadd f3, f18, f27, f3 LFD f18, 2 * SIZE(AO) fmadd f4, f19, f24, f4 LFD f24, 4 * SIZE(BO) fmadd f5, f19, f25, f5 LFD f25, 5 * SIZE(BO) fmadd f6, f19, f26, f6 LFD f26, 6 * SIZE(BO) fmadd f7, f19, f27, f7 LFD f27, 7 * SIZE(BO) bdnz .L22 .align 4.L25:#if defined(LT) || defined(RN) andi. r0, KK, 3#else andi. r0, TEMP, 3#endif mtspr CTR, r0 ble .L27 .align 4.L26: fmadd f0, f16, f20, f0 LFD f17, 1 * SIZE(AO) fmadd f1, f16, f21, f1 nop fmadd f2, f16, f22, f2 nop fmadd f3, f16, f23, f3 LFDU f16, 2 * SIZE(AO) fmadd f4, f17, f20, f4 LFDU f20, 4 * SIZE(BO) fmadd f5, f17, f21, f5 LFD f21, 1 * SIZE(BO) fmadd f6, f17, f22, f6 LFD f22, 2 * SIZE(BO) fmadd f7, f17, f23, f7 LFD f23, 3 * SIZE(BO) bdnz .L26 .align 4.L27:#ifndef CONJ FSUB f0, f0, f5 FADD f1, f1, f4 FSUB f2, f2, f7 FADD f3, f3, f6#else#if defined(LN) || defined(LT) FADD f0, f0, f5 FSUB f1, f1, f4 FADD f2, f2, f7 FSUB f3, f3, f6#else FADD f0, f0, f5 FSUB f1, f4, f1 FADD f2, f2, f7 FSUB f3, f6, f3#endif#endif#if defined(LN) || defined(RT)#ifdef LN subi r0, KK, 1#else subi r0, KK, 2#endif slwi TEMP, r0, 0 + ZBASE_SHIFT slwi r0, r0, 1 + ZBASE_SHIFT add AO, AORIG, TEMP 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) FSUB f0, f16, f0 FSUB f1, f17, f1 FSUB f2, f18, f2 FSUB f3, f19, f3#else LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) LFD f20, 2 * SIZE(AO) LFD f21, 3 * SIZE(AO) FSUB f0, f16, f0 FSUB f1, f17, f1 FSUB f2, f20, f2 FSUB f3, f21, f3#endif#ifdef LN LFD f20, 0 * SIZE(AO) LFD f21, 1 * SIZE(AO) FMUL f4, f21, f1 FMUL f5, f21, f0 FMUL f12, f21, f3 FMUL f13, f21, f2#ifndef CONJ FMSUB f0, f20, f0, f4 FMADD f1, f20, f1, f5 FMSUB f2, f20, f2, f12 FMADD f3, f20, f3, f13#else FMADD f0, f20, f0, f4 FMSUB f1, f20, f1, f5 FMADD f2, f20, f2, f12 FMSUB f3, f20, f3, f13#endif#endif#ifdef LT LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) FMUL f4, f17, f1 FMUL f5, f17, f0 FMUL f12, f17, f3 FMUL f13, f17, f2#ifndef CONJ FMSUB f0, f16, f0, f4 FMADD f1, f16, f1, f5 FMSUB f2, f16, f2, f12 FMADD f3, f16, f3, f13#else FMADD f0, f16, f0, f4 FMSUB f1, f16, f1, f5 FMADD f2, f16, f2, f12 FMSUB f3, f16, f3, 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#ifndef CONJ FMSUB f0, f16, f0, f4 FMADD f1, f16, f1, f5 FMADD f2, f19, f1, f2 FNMSUB f3, f19, f0, f3 FNMSUB f2, f18, f0, f2 FNMSUB f3, f18, f1, f3 FMUL f4, f21, f3 FMUL f5, f21, f2 FMSUB f2, f20, f2, f4 FMADD f3, f20, f3, f5#else FMADD f0, f16, f0, f4 FMSUB f1, f16, f1, f5 FMSUB f2, f19, f1, f2 FNMADD f3, f19, f0, f3 FNMADD f2, f18, f0, f2 FNMADD f3, f18, f1, f3 FMUL f4, f21, f3 FMUL f5, f21, f2 FMADD f2, f20, f2, f4 FMSUB f3, f20, f3, f5#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, f3 FMUL f13, f17, f2#ifndef CONJ FMSUB f2, f16, f2, f12 FMADD f3, f16, f3, f13 FMADD f0, f19, f3, f0 FNMSUB f1, f19, f2, f1 FNMSUB f0, f18, f2, f0 FNMSUB f1, f18, f3, f1 FMUL f4, f21, f1 FMUL f5, f21, f0 FMSUB f0, f20, f0, f4 FMADD f1, f20, f1, f5#else FMADD f2, f16, f2, f12 FMSUB f3, f16, f3, f13 FMSUB f0, f19, f3, f0 FNMADD f1, f19, f2, f1 FNMADD f0, f18, f2, f0 FNMADD f1, f18, f3, f1 FMUL f4, f21, f1 FMUL f5, f21, f0 FMADD f0, f20, f0, f4 FMSUB f1, f20, f1, f5#endif#endif#ifdef LN subi CO1, CO1, 2 * SIZE subi CO2, CO2, 2 * SIZE#endif#if defined(LN) || defined(LT) STFD f0, 0 * SIZE(BO) STFD f1, 1 * SIZE(BO) STFD f2, 2 * SIZE(BO) STFD f3, 3 * SIZE(BO)#else STFD f0, 0 * SIZE(AO) STFD f1, 1 * SIZE(AO) STFD f2, 2 * SIZE(AO) STFD f3, 3 * SIZE(AO)#endif STFD f0, 0 * SIZE(CO1) STFD f1, 1 * SIZE(CO1) STFD f2, 0 * SIZE(CO2) STFD f3, 1 * SIZE(CO2)#ifndef LN addi CO1, CO1, 2 * SIZE addi CO2, CO2, 2 * SIZE#endif#ifdef RT slwi r0, K, 0 + ZBASE_SHIFT add AORIG, AORIG, r0#endif#if defined(LT) || defined(RN) sub TEMP, K, KK slwi r0, TEMP, 0 + ZBASE_SHIFT slwi TEMP, TEMP, 1 + ZBASE_SHIFT add AO, AO, r0 add BO, BO, TEMP#endif#ifdef LT addi KK, KK, 1#endif#ifdef LN subi KK, KK, 1#endif .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 srawi. I, M, 1#if defined(LN) || defined(RT) mr AORIG, A#else mr AO, A#endif#ifndef RT add C, C, LDC#endif ble .L40 .align 4.L31:#if defined(LT) || defined(RN) LFD f20, 0 * SIZE(AO) LFD f21, 1 * SIZE(AO) LFD f22, 2 * SIZE(AO) LFD f23, 3 * SIZE(AO) LFD f24, 4 * SIZE(AO) LFD f25, 5 * SIZE(AO) LFD f26, 6 * SIZE(AO) LFD f27, 7 * SIZE(AO) LFD f16, 0 * SIZE(B) LFD f17, 1 * SIZE(B) LFD f18, 2 * SIZE(B) LFD f19, 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
⌨️ 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 + -