📄 trsm_kernel_cell_lt.s
字号:
slwi r0, K, 0 + BASE_SHIFT add AORIG, AORIG, r0#endif#if defined(LT) || defined(RN) sub TEMP, K, KK slwi r0, TEMP, 0 + BASE_SHIFT slwi TEMP, TEMP, 2 + BASE_SHIFT add AO, AO, r0 add BO, BO, TEMP#endif#ifdef LN subi KK, KK, 1#endif#ifdef LT addi KK, KK, 1#endif .align 4LL(39):#ifdef LN slwi r0, K, 2 + BASE_SHIFT add B, B, r0#endif#if defined(LT) || defined(RN) mr B, BO#endif#ifdef RN addi KK, KK, 4#endif#ifdef RT subi KK, KK, 4#endif addic. J, J, -1 lfs f0, FZERO bgt LL(10) .align 4LL(40): andi. J, N, 2 ble LL(70)#ifdef RT slwi r0, K, 1 + BASE_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 fmr f1, f0 fmr f2, f0 fmr f3, f0 fmr f4, f0 fmr f5, f0 fmr f6, f0 fmr f7, f0 srawi. I, M, 2#if defined(LN) || defined(RT) mr AORIG, A#else mr AO, A#endif#ifndef RT add C, CO2, LDC#endif ble LL(50) .align 4LL(41):#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) dcbt CO1, PREC dcbt CO2, PREC srawi. r0, KK, 2 mtspr CTR, r0 mr BO, B#else#ifdef LN slwi r0, K, 2 + BASE_SHIFT sub AORIG, AORIG, r0#endif slwi r0, KK, 2 + BASE_SHIFT slwi TEMP, KK, 1 + BASE_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) dcbt CO1, PREC dcbt CO2, PREC srawi. r0, TEMP, 2 mtspr CTR, r0#endif ble LL(45) .align 5LL(42): FMADD f0, f16, f20, f0 FMADD f1, f17, f20, f1 FMADD f2, f18, f20, f2 FMADD f3, f19, f20, f3 FMADD f4, f16, f21, f4 FMADD f5, f17, f21, f5 FMADD f6, f18, f21, f6 FMADD f7, f19, f21, f7 LFD f16, 4 * SIZE(AO) LFD f17, 5 * SIZE(AO) LFD f18, 6 * SIZE(AO) LFD f19, 7 * SIZE(AO) FMADD f0, f16, f22, f0 FMADD f1, f17, f22, f1 FMADD f2, f18, f22, f2 FMADD f3, f19, f22, f3 FMADD f4, f16, f23, f4 FMADD f5, f17, f23, f5 FMADD f6, f18, f23, f6 FMADD f7, f19, f23, f7 LFD f16, 8 * SIZE(AO) LFD f17, 9 * SIZE(AO) LFD f18, 10 * SIZE(AO) LFD f19, 11 * SIZE(AO) LFD f20, 4 * SIZE(BO) LFD f21, 5 * SIZE(BO) LFD f22, 6 * SIZE(BO) LFD f23, 7 * SIZE(BO) FMADD f0, f16, f20, f0 FMADD f1, f17, f20, f1 FMADD f2, f18, f20, f2 FMADD f3, f19, f20, f3 FMADD f4, f16, f21, f4 FMADD f5, f17, f21, f5 FMADD f6, f18, f21, f6 FMADD f7, f19, f21, f7 LFD f16, 12 * SIZE(AO) LFD f17, 13 * SIZE(AO) LFD f18, 14 * SIZE(AO) LFD f19, 15 * SIZE(AO) FMADD f0, f16, f22, f0 FMADD f1, f17, f22, f1 FMADD f2, f18, f22, f2 FMADD f3, f19, f22, f3 FMADD f4, f16, f23, f4 FMADD f5, f17, f23, f5 FMADD f6, f18, f23, f6 FMADD f7, f19, f23, f7 LFD f16, 16 * SIZE(AO) LFD f17, 17 * SIZE(AO) LFD f18, 18 * SIZE(AO) LFD f19, 19 * SIZE(AO) LFD f20, 8 * SIZE(BO) LFD f21, 9 * SIZE(BO) LFD f22, 10 * SIZE(BO) LFD f23, 11 * SIZE(BO) addi AO, AO, 16 * SIZE addi BO, BO, 8 * SIZE PREFETCH_B bdnz LL(42) .align 4LL(45):#if defined(LT) || defined(RN) andi. r0, KK, 3#else andi. r0, TEMP, 3#endif mtspr CTR, r0 ble+ LL(48) .align 4LL(46): FMADD f0, f16, f20, f0 FMADD f1, f17, f20, f1 FMADD f2, f18, f20, f2 FMADD f3, f19, f20, f3 FMADD f4, f16, f21, f4 FMADD f5, f17, f21, f5 FMADD f6, f18, f21, f6 FMADD f7, f19, f21, f7 LFD f16, 4 * SIZE(AO) LFD f17, 5 * SIZE(AO) LFD f18, 6 * SIZE(AO) LFD f19, 7 * SIZE(AO) LFD f20, 2 * SIZE(BO) LFD f21, 3 * SIZE(BO) addi BO, BO, 2 * SIZE addi AO, AO, 4 * SIZE bdnz LL(46) .align 4LL(48):#if defined(LN) || defined(RT)#ifdef LN subi r0, KK, 4#else subi r0, KK, 2#endif slwi TEMP, r0, 2 + BASE_SHIFT slwi r0, r0, 1 + BASE_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 f20, 2 * SIZE(BO) LFD f21, 3 * SIZE(BO) LFD f24, 4 * SIZE(BO) LFD f25, 5 * SIZE(BO) LFD f28, 6 * SIZE(BO) LFD f29, 7 * SIZE(BO) FSUB f0, f16, f0 FSUB f4, f17, f4 FSUB f1, f20, f1 FSUB f5, f21, f5 FSUB f2, f24, f2 FSUB f6, f25, f6 FSUB f3, f28, f3 FSUB f7, f29, f7#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) FSUB f0, f16, f0 FSUB f1, f17, f1 FSUB f2, f18, f2 FSUB f3, f19, f3 FSUB f4, f20, f4 FSUB f5, f21, f5 FSUB f6, f22, f6 FSUB f7, f23, f7#endif#ifdef LN LFD f16, 15 * SIZE(AO) LFD f17, 14 * SIZE(AO) LFD f18, 13 * SIZE(AO) LFD f19, 12 * SIZE(AO) FMUL f3, f16, f3 FMUL f7, f16, f7 FNMSUB f2, f17, f3, f2 FNMSUB f6, f17, f7, f6 FNMSUB f1, f18, f3, f1 FNMSUB f5, f18, f7, f5 FNMSUB f0, f19, f3, f0 FNMSUB f4, f19, f7, f4 LFD f16, 10 * SIZE(AO) LFD f17, 9 * SIZE(AO) LFD f18, 8 * SIZE(AO) LFD f19, 5 * SIZE(AO) LFD f20, 4 * SIZE(AO) LFD f21, 0 * SIZE(AO) FMUL f2, f16, f2 FMUL f6, f16, f6 FNMSUB f1, f17, f2, f1 FNMSUB f5, f17, f6, f5 FNMSUB f0, f18, f2, f0 FNMSUB f4, f18, f6, f4 FMUL f1, f19, f1 FMUL f5, f19, f5 FNMSUB f0, f20, f1, f0 FNMSUB f4, f20, f5, f4 FMUL f0, f21, f0 FMUL f4, f21, f4#endif#ifdef LT LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) LFD f18, 2 * SIZE(AO) LFD f19, 3 * SIZE(AO) FMUL f0, f16, f0 FMUL f4, f16, f4 FNMSUB f1, f17, f0, f1 FNMSUB f5, f17, f4, f5 FNMSUB f2, f18, f0, f2 FNMSUB f6, f18, f4, f6 FNMSUB f3, f19, f0, f3 FNMSUB f7, f19, f4, f7 LFD f17, 5 * SIZE(AO) LFD f18, 6 * SIZE(AO) LFD f19, 7 * SIZE(AO) FMUL f1, f17, f1 FMUL f5, f17, f5 FNMSUB f2, f18, f1, f2 FNMSUB f6, f18, f5, f6 FNMSUB f3, f19, f1, f3 FNMSUB f7, f19, f5, f7 LFD f18, 10 * SIZE(AO) LFD f19, 11 * SIZE(AO) FMUL f2, f18, f2 FMUL f6, f18, f6 FNMSUB f3, f19, f2, f3 FNMSUB f7, f19, f6, f7 LFD f19, 15 * SIZE(AO) FMUL f3, f19, f3 FMUL f7, f19, f7#endif#ifdef RN LFD f16, 0 * SIZE(BO) LFD f17, 1 * SIZE(BO) LFD f18, 3 * SIZE(BO) FMUL f0, f16, f0 FMUL f1, f16, f1 FMUL f2, f16, f2 FMUL f3, f16, f3 FNMSUB f4, f17, f0, f4 FNMSUB f5, f17, f1, f5 FNMSUB f6, f17, f2, f6 FNMSUB f7, f17, f3, f7 FMUL f4, f18, f4 FMUL f5, f18, f5 FMUL f6, f18, f6 FMUL f7, f18, f7#endif#ifdef RT LFD f19, 3 * SIZE(BO) LFD f20, 2 * SIZE(BO) LFD f21, 0 * SIZE(BO) FMUL f4, f19, f4 FMUL f5, f19, f5 FMUL f6, f19, f6 FMUL f7, f19, f7 FNMSUB f0, f20, f4, f0 FNMSUB f1, f20, f5, f1 FNMSUB f2, f20, f6, f2 FNMSUB f3, f20, f7, f3 FMUL f0, f21, f0 FMUL f1, f21, f1 FMUL f2, f21, f2 FMUL f3, f21, f3#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 f4, 1 * SIZE(BO) STFD f1, 2 * SIZE(BO) STFD f5, 3 * SIZE(BO) STFD f2, 4 * SIZE(BO) STFD f6, 5 * SIZE(BO) STFD f3, 6 * SIZE(BO) STFD f7, 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 f4, 4 * SIZE(AO) STFD f5, 5 * SIZE(AO) STFD f6, 6 * SIZE(AO) STFD f7, 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 f4, 0 * SIZE(CO2) STFD f5, 1 * SIZE(CO2) STFD f6, 2 * SIZE(CO2) STFD f7, 3 * SIZE(CO2) lfs f0, FZERO fmr f1, f0 fmr f2, f0 fmr f3, f0 fmr f4, f0 fmr f5, f0 fmr f6, f0 fmr f7, f0#ifndef LN addi CO1, CO1, 4 * SIZE addi CO2, CO2, 4 * SIZE#endif#ifdef RT slwi r0, K, 2 + BASE_SHIFT add AORIG, AORIG, r0#endif#if defined(LT) || defined(RN) sub TEMP, K, KK slwi r0, TEMP, 2 + BASE_SHIFT slwi TEMP, TEMP, 1 + BASE_SHIFT add AO, AO, r0 add BO, BO, TEMP#endif#ifdef LN subi KK, KK, 4#endif#ifdef LT addi KK, KK, 4#endif addic. I, I, -1 bgt+ LL(41) .align 4LL(50): andi. I, M, 2 ble LL(60)#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) srawi. r0, KK, 2 mtspr CTR, r0 mr BO, B#else#ifdef LN slwi r0, K, 1 + BASE_SHIFT sub AORIG, AORIG, r0#endif slwi r0, KK, 1 + BASE_SHIFT slwi TEMP, KK, 1 + BASE_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) srawi. r0, TEMP, 2 mtspr CTR, r0#endif ble LL(55) .align 5LL(52): FMADD f0, f16, f20, f0 FMADD f1, f17, f20, f1 FMADD f2, f16, f21, f2 FMADD f3, f17, f21, f3 FMADD f4, f18, f22, f4 FMADD f5, f19, f22, f5 FMADD f6, f18, f23, f6 FMADD f7, f19, f23, f7 LFD f16, 4 * SIZE(AO) LFD f17, 5 * SIZE(AO) LFD f18, 6 * SIZE(AO) LFD f19, 7 * SIZE(AO) LFD f20, 8 * SIZE(BO) LFD f21, 9 * SIZE(BO) LFD f22, 10 * SIZE(BO) LFD f23, 11 * SIZE(BO) FMADD f0, f16, f24, f0 FMADD f1, f17, f24, f1 FMADD f2, f16, f25, f2 FMADD f3, f17, f25, f3 FMADD f4, f18, f26, f4 FMADD f5, f19, f26, f5 FMADD f6, f18, f27, f6 FMADD f7, f19, f27, f7 LFD f16, 8 * SIZE(AO) LFD f17, 9 * SIZE(AO) LFD f18, 10 * SIZE(AO) LFD f19, 11 * SIZE(AO) LFD f24, 12 * SIZE(BO) LFD f25, 13 * SIZE(BO) LFD f26, 14 * SIZE(BO) LFD f27, 15 * SIZE(BO) addi AO, AO, 8 * SIZE addi BO, BO, 8 * SIZE PREFETCH_B bdnz LL(52) .align 4LL(55):#if defined(LT) || defined(RN) andi. r0, KK, 3#else andi. r0, TEMP, 3#endif mtspr CTR, r0 ble+ LL(58) .align 4LL(56): FMADD f0, f16, f20, f0 FMADD f1, f17, f20, f1 FMADD f2, f16, f21, f2 FMADD f3, f17, f21, f3 LFD f16, 2 * SIZE(AO) LFD f17, 3 * SIZE(AO) LFD f20, 2 * SIZE(BO) LFD f21, 3 * SIZE(BO) addi BO, BO, 2 * SIZE addi AO, AO, 2 * SIZE bdnz LL(56) .align 4LL(58): FADD f0, f4, f0 FADD f1, f5, f1 FADD f2, f6, f2 FADD f3, f7, f3#if defined(LN) || defined(RT)#ifdef LN subi r0, KK, 2#else subi r0, KK, 2#endif slwi TEMP, r0, 1 + BASE_SHIFT slwi r0, r0, 1 + BASE_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 f20, 2 * SIZE(BO) LFD f21, 3 * SIZE(BO) FSUB f0, f16, f0 FSUB f2, f17, f2 FSUB f1, f20, f1 FSUB f3, f21, 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 f19, 3 * SIZE(AO) LFD f20, 2 * SIZE(AO) LFD f21, 0 * SIZE(AO) FMUL f1, f19, f1 FMUL f3, f19, f3 FNMSUB f0, f20, f1, f0 FNMSUB f2, f20, f3, f2 FMUL f0, f21, f0 FMUL f2, f21, f2#endif#ifdef LT LFD f16, 0 * SIZE(AO) LFD f17, 1 * SIZE(AO) FMUL f0, f16, f0 FMUL f2, f16, f2 FNMSUB f1, f17, f0, f1 FNMSUB f3, f17, f2, f3 LFD f17, 3 * SIZE(AO) FMUL f1, f17, f1 FMUL f3, f17, f3#endif#ifdef RN LFD f16, 0 * SIZE(BO) LFD f17, 1 * SIZE(BO) LFD f18, 3 * SIZE(BO) FMUL f0, f16, f0 FMUL f1, f16, f1 FNMSUB f2, f17, f0, f2 FNMSUB f3, f17, f1, f3 FMUL f2, f18, f2 FMUL f3, f18, f3#endif#ifdef RT LFD f19, 3 * SIZE(BO) LFD f20, 2 * SIZE(BO) LFD f21, 0 * SIZE(BO) FMUL f2, f19, f2 FMUL f3, f19, f3 FNMSUB f0, f20, f2, f0 FNMSUB f1, f20, f3, f1 FMUL f0, f21, f0 FMUL f1, f21, f1#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 f2, 1 * SIZE(BO) STFD f1, 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) lfs f0, FZERO fmr f1, f0 fmr f2, f0 fmr f3, f0 fmr f4, f0 fmr f5, f0 fmr f6, f0 fmr f7, f0#ifndef LN addi CO1, CO1, 2 * SIZE addi CO2, CO2, 2 * SIZE#endif#ifdef RT slwi r0, K, 1 + BASE_SHIFT add AORIG, AORIG, r0#endif#if defined(LT) || defined(RN) sub TEMP, K, KK slwi r0, TEMP, 1 + BASE_SHIFT slwi TEMP, TEMP, 1 + BASE_SHIFT add AO, AO, r0 add BO, BO, TEMP#endif#ifdef LN subi KK, KK, 2#endif#ifdef LT addi KK, KK, 2#endif .align 4LL(60): andi. I, M, 1 ble LL(69)#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) srawi. r0, KK, 2 mtspr CTR, r0 mr BO, B#else#ifdef LN slwi r0, K, BASE_SHIFT sub AORIG, AORIG, r0#endif slwi r0, KK, 0 + BASE_SHIFT slwi TEMP, KK, 1 + BASE_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) srawi. r0, TEMP, 2 mtspr CTR, r0#endif ble LL(65) .align 5LL(62): FMADD f0, f16, f20, f0 FMADD f1, f16, f21, f1 FMADD f2, f17, f22, f2 FMADD f3, f17, f23, f3 LFD f20, 8 * SIZE(BO) LFD f21, 9 * SIZE(BO) LFD f22, 10 * SIZE(BO) LFD f23, 11 * SIZE(BO) FMADD f0, f18, f24, f0 FMADD f1, f18, f25, f1 FMADD f2, f19, f26, f2 FMADD f3, f19, f27, f3 LFD f16, 4 * SIZE(AO) LFD f17, 5 * SIZE(AO) LFD f18, 6 * SIZE(AO) LFD f19, 7 * SIZE(AO) LFD f24, 12 * SIZE(BO) LFD f25, 13 * SIZE(BO) LFD f26, 14 * SIZE(BO) LFD f27, 15 * SIZE(BO) addi AO, AO, 4 * SIZE addi BO, BO, 8 * SIZE bdnz LL(62) .align 4LL(65):#if defined(LT) || defined(RN) andi. r0, KK, 3#else andi. r0, TEMP, 3#endif mtspr CTR, r0 ble+ LL(68) .align 4LL(66): FMADD f0, f16, f20, f0 FMADD f1, f16, f21, f1 LFD f16, 1 * SIZE(AO) LFD f20, 2 * SIZE(BO) LFD f21, 3 * SIZE(BO) addi BO, BO, 2 * SIZE addi AO, AO, 1 * SIZE bdnz LL(66) .align 4LL(68): FADD f0, f2, f0 FADD f1, f3, f1#if defined(LN) || defined(RT)#ifdef LN subi r0, KK, 1#else subi r0, KK, 2#endif slwi TEMP, r0, 0 + BASE_SHIFT slwi r0, r0, 1 + BASE_SHIFT add AO, AORIG, TEMP
⌨️ 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 + -