📄 ztrsm_kernel_rt.s
字号:
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 srawi. r0, TEMP, 2 mtspr CTR, r0#endif ble LL(45) .align 4LL(42): fmadd f0, f16, f20, f0 fmadd f1, f17, f21, f1 fmadd f2, f17, f20, f2 fmadd f3, f16, f21, f3 LFD f16, 4 * SIZE(AO) LFD f17, 5 * SIZE(AO) LFD f20, 4 * SIZE(BO) LFD f21, 5 * SIZE(BO) fmadd f4, f18, f22, f4 fmadd f5, f19, f23, f5 fmadd f6, f19, f22, f6 fmadd f7, f18, f23, f7 LFD f18, 6 * SIZE(AO) LFD f19, 7 * SIZE(AO) LFD f22, 6 * SIZE(BO) LFD f23, 7 * SIZE(BO) fmadd f0, f16, f20, f0 fmadd f1, f17, f21, f1 fmadd f2, f17, f20, f2 fmadd f3, f16, f21, f3 LFD f16, 8 * SIZE(AO) LFD f17, 9 * SIZE(AO) LFD f20, 8 * SIZE(BO) LFD f21, 9 * SIZE(BO) fmadd f4, f18, f22, f4 fmadd f5, f19, f23, f5 fmadd f6, f19, f22, f6 fmadd f7, f18, f23, f7 LFD f18, 10 * SIZE(AO) LFD f19, 11 * SIZE(AO) LFD f22, 10 * SIZE(BO) LFD f23, 11 * SIZE(BO) addi AO, AO, 8 * SIZE addi BO, BO, 8 * SIZE bdnz LL(42) .align 4LL(45): fadd f0, f0, f4 fadd f1, f1, f5 fadd f2, f2, f6 fadd f3, f3, f7#if defined(LT) || defined(RN) andi. r0, KK, 3#else andi. r0, TEMP, 3#endif mtspr CTR,r0 ble LL(47) .align 4LL(46): fmadd f0, f16, f20, f0 fmadd f1, f17, f21, f1 fmadd f2, f17, f20, f2 fmadd f3, f16, f21, f3 LFD f16, 2 * SIZE(AO) LFD f17, 3 * SIZE(AO) LFD f20, 2 * SIZE(BO) LFD f21, 3 * SIZE(BO) addi AO, AO, 2 * SIZE addi BO, BO, 2 * SIZE bdnz LL(46) .align 4LL(47):#ifndef CONJ FSUB f0, f0, f1 FADD f1, f2, f3#else FADD f0, f0, f1 FSUB f1, f3, f2#endif#if defined(LN) || defined(RT) subi r0, KK, 1 slwi r0, r0, 0 + 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) FSUB f0, f16, f0 FSUB f1, f17, f1#else LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO)#ifndef CONJ FSUB f0, f16, f0 FSUB f1, f17, f1#else FSUB f0, f16, f0 FADD f1, f17, f1#endif#endif#ifdef LN LFD f20, 0 * SIZE(AO) LFD f21, 1 * SIZE(AO) FMUL f4, f21, f1 FMUL f5, f21, f0#ifndef CONJ FMSUB f0, f20, f0, f4 FMADD f1, f20, f1, f5#else FMADD f0, f20, f0, f4 FMSUB f1, f20, f1, f5#endif#endif#ifdef LT LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) FMUL f4, f17, f1 FMUL f5, f17, f0#ifndef CONJ FMSUB f0, f16, f0, f4 FMADD f1, f16, f1, f5#else FMADD f0, f16, f0, f4 FMSUB f1, f16, f1, f5#endif#endif#ifdef RN LFD f16, 0 * SIZE(BO) LFD f17, 1 * SIZE(BO) FMUL f4, f17, f1 FMUL f5, f17, f0#ifndef CONJ FMSUB f0, f16, f0, f4 FMADD f1, f16, f1, f5#else FMADD f0, f16, f0, f4 FMSUB f1, f16, f1, f5#endif#endif#ifdef RT LFD f20, 0 * SIZE(BO) LFD f21, 1 * SIZE(BO) FMUL f4, f21, f1 FMUL f5, f21, f0#ifndef CONJ FMSUB f0, f20, f0, f4 FMADD f1, f20, f1, f5#else FMADD f0, f20, f0, f4 FMSUB f1, f20, f1, f5#endif#endif#ifdef LN subi CO1, CO1, 2 * SIZE#endif#if defined(LN) || defined(LT) STFD f0, 0 * SIZE(BO) STFD f1, 1 * SIZE(BO)#else STFD f0, 0 * SIZE(AO) STFD f1, 1 * SIZE(AO)#endif STFD f0, 0 * SIZE(CO1) STFD f1, 1 * SIZE(CO1)#ifndef LN addi CO1, CO1, 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 TEMP, TEMP, 0 + ZBASE_SHIFT add AO, AO, TEMP add BO, BO, TEMP#endif#ifdef LT addi KK, KK, 1#endif#ifdef LN subi KK, KK, 1#endif .align 4LL(49):#ifdef LN slwi r0, K, 0 + ZBASE_SHIFT add B, B, r0#endif#if defined(LT) || defined(RN) mr B, BO#endif#ifdef RN addi KK, KK, 1#endif#ifdef RT subi KK, KK, 1#endif .align 4LL(30): srawi. J, N, 1 ble LL(999) .align 4LL(10):#ifdef RT slwi r0, K, 1 + ZBASE_SHIFT sub B, B, r0 slwi r0, LDC, 1 sub C, C, r0#endif mr CO1, C add CO2, C, LDC#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, CO2, LDC#endif ble LL(20) .align 4LL(11):#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 fmr f8, f0 fmr f9, f0 fmr f10, f0 fmr f11, f0 fmr f12, f0 fmr f13, f0 fmr f14, f0 fmr f15, f0 dcbt CO1, PREC dcbt CO2, PREC srawi. r0, KK, 2 mtspr CTR, r0 mr BO, B#else#ifdef LN slwi r0, K, 1 + ZBASE_SHIFT sub AORIG, AORIG, r0#endif slwi TEMP, KK, 1 + ZBASE_SHIFT add AO, AORIG, TEMP 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) 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 dcbt CO1, PREC dcbt CO2, PREC srawi. r0, TEMP, 2 mtspr CTR, r0#endif ble LL(15) .align 4LL(12): fmadd f0, f16, f20, f0 fmadd f5, f17, f21, f5 fmadd f10, f18, f22, f10 fmadd f15, f19, f23, f15 LFD f28, 4 * SIZE(BO) LFD f29, 5 * SIZE(BO) LFD f30, 6 * SIZE(BO) LFD f31, 7 * SIZE(BO) fmadd f1, f17, f20, f1 fmadd f2, f18, f20, f2 fmadd f3, f19, f20, f3 fmadd f4, f16, f21, f4 LFD f24, 4 * SIZE(AO) LFD f25, 5 * SIZE(AO) LFD f26, 6 * SIZE(AO) LFD f27, 7 * SIZE(AO) fmadd f6, f18, f21, f6 fmadd f7, f19, f21, f7 fmadd f8, f16, f22, f8 fmadd f9, f17, f22, f9 fmadd f11, f19, f22, f11 fmadd f12, f16, f23, f12 fmadd f13, f17, f23, f13 fmadd f14, f18, f23, f14 LFD f20, 8 * SIZE(BO) LFD f21, 9 * SIZE(BO) LFD f22, 10 * SIZE(BO) LFD f23, 11 * SIZE(BO) fmadd f0, f24, f28, f0 fmadd f5, f25, f29, f5 fmadd f10, f26, f30, f10 fmadd f15, f27, f31, f15 LFD f16, 8 * SIZE(AO) LFD f17, 9 * SIZE(AO) LFD f18, 10 * SIZE(AO) LFD f19, 11 * SIZE(AO) fmadd f1, f25, f28, f1 fmadd f2, f26, f28, f2 fmadd f3, f27, f28, f3 fmadd f4, f24, f29, f4 fmadd f6, f26, f29, f6 fmadd f7, f27, f29, f7 fmadd f8, f24, f30, f8 fmadd f9, f25, f30, f9 fmadd f11, f27, f30, f11 fmadd f12, f24, f31, f12 fmadd f13, f25, f31, f13 fmadd f14, f26, f31, f14 LFD f28, 12 * SIZE(BO) LFD f29, 13 * SIZE(BO) LFD f30, 14 * SIZE(BO) LFD f31, 15 * SIZE(BO) fmadd f0, f16, f20, f0 fmadd f5, f17, f21, f5 fmadd f10, f18, f22, f10 fmadd f15, f19, f23, f15 LFD f24, 12 * SIZE(AO) LFD f25, 13 * SIZE(AO) LFD f26, 14 * SIZE(AO) LFD f27, 15 * SIZE(AO) fmadd f1, f17, f20, f1 fmadd f2, f18, f20, f2 fmadd f3, f19, f20, f3 fmadd f4, f16, f21, f4 fmadd f6, f18, f21, f6 fmadd f7, f19, f21, f7 fmadd f8, f16, f22, f8 fmadd f9, f17, f22, f9 fmadd f11, f19, f22, f11 fmadd f12, f16, f23, f12 fmadd f13, f17, f23, f13 fmadd f14, f18, f23, f14 LFD f16, 16 * SIZE(AO) LFD f17, 17 * SIZE(AO) LFD f18, 18 * SIZE(AO) LFD f19, 19 * SIZE(AO) fmadd f0, f24, f28, f0 fmadd f5, f25, f29, f5 fmadd f10, f26, f30, f10 fmadd f15, f27, f31, f15 LFD f20, 16 * SIZE(BO) LFD f21, 17 * SIZE(BO) LFD f22, 18 * SIZE(BO) LFD f23, 19 * SIZE(BO) fmadd f1, f25, f28, f1 fmadd f2, f26, f28, f2 fmadd f3, f27, f28, f3 fmadd f4, f24, f29, f4 fmadd f6, f26, f29, f6 fmadd f7, f27, f29, f7 fmadd f8, f24, f30, f8 fmadd f9, f25, f30, f9 fmadd f11, f27, f30, f11 fmadd f12, f24, f31, f12 fmadd f13, f25, f31, f13 fmadd f14, f26, f31, f14 addi AO, AO, 16 * SIZE addi BO, BO, 16 * SIZE#ifdef PPC970#ifndef ALLOC_HUGETLB PREFETCH_A#endif PREFETCH_B#endif#ifdef POWER4#ifndef ALLOC_HUGETLB PREFETCH_A#endif PREFETCH_B#endif#ifdef POWER5 PREFETCH_A PREFETCH_B#endif bdnz LL(12) .align 4LL(15):#if defined(LT) || defined(RN) andi. r0, KK, 3#else andi. r0, TEMP, 3#endif mtspr CTR, r0 ble LL(KERNEL_MainFinish) .align 4LL(16): fmadd f0, f16, f20, f0 fmadd f5, f17, f21, f5 fmadd f10, f18, f22, f10 fmadd f15, f19, f23, f15 fmadd f1, f17, f20, f1 fmadd f2, f18, f20, f2 fmadd f3, f19, f20, f3 fmadd f4, f16, f21, f4 fmadd f6, f18, f21, f6 fmadd f7, f19, f21, f7 fmadd f8, f16, f22, f8 fmadd f9, f17, f22, f9 fmadd f11, f19, f22, f11 fmadd f12, f16, f23, f12 fmadd f13, f17, f23, f13 fmadd f14, f18, f23, f14 LFD f16, 4 * SIZE(AO) LFD f17, 5 * SIZE(AO) LFD f18, 6 * SIZE(AO) LFD f19, 7 * SIZE(AO) LFD f20, 4 * SIZE(BO) LFD f21, 5 * SIZE(BO) LFD f22, 6 * SIZE(BO) LFD f23, 7 * SIZE(BO) addi BO, BO, 4 * SIZE addi AO, AO, 4 * SIZE bdnz LL(16) .align 4LL(KERNEL_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
⌨️ 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 + -