📄 zgemm_kernel_hummer.s
字号:
addi SP, SP, 16 blr .align 4.L1000: li INC, 1 * SIZE li INC2, 2 * SIZE li INC4, 4 * SIZE li INCM1, -1 * SIZE li INCM3, -3 * SIZE li INCM5, -5 * SIZE li INCM7, -7 * SIZE addi C, C, - 1 * SIZE srawi. J, N, 1 ble .L1050 .align 4.L1010: mr CO1, C add CO2, C, LDC add C, CO2, LDC#if defined(TRMMKERNEL) && defined(LEFT) mr KK, OFFSET#endif addi AO, A, -4 * SIZE li r0, FZERO lfpsx f0, SP, r0 srawi. I, M, 2 ble .L1020 .align 4.L1011:#if defined(TRMMKERNEL)#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA)) addi AO2, AO, 2 * SIZE fpmr f4, f0 addi BO, B, - 4 * SIZE fpmr f8, f0 addi BO2, B, - 2 * SIZE fpmr f12, f0#else slwi TEMP, KK, 2 + ZBASE_SHIFT slwi r0, KK, 1 + ZBASE_SHIFT add AO, AO, TEMP add BO, B, r0 addi AO2, AO, 2 * SIZE fpmr f4, f0 addi BO, BO, - 4 * SIZE fpmr f8, f0 addi BO2, BO, 2 * SIZE fpmr f12, f0#endif#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) sub TEMP, K, KK#elif defined(LEFT) addi TEMP, KK, 4#else addi TEMP, KK, 2#endif srawi. r0, TEMP, 2 fpmr f1, f0 mtspr CTR, r0 ble .L1014#else addi AO2, AO, 2 * SIZE fpmr f4, f0 addi BO, B, - 4 * SIZE fpmr f8, f0 addi BO2, B, - 2 * SIZE fpmr f12, f0 srawi. r0, K, 2 fpmr f1, f0 mtspr CTR, r0 ble .L1014#endif LFPDUX A1, AO, INC4 fpmr f5, f0 LFPDUX A3, AO, INC4 fpmr f9, f0 LFPDUX B1, BO, INC4 fpmr f13, f0 LFPDUX A5, AO, INC4 fpmr f2, f0 LFPDUX A6, AO, INC4 fpmr f6, f0 LFPDUX B3, BO, INC4 fpmr f10, f0 LFPDUX A7, AO, INC4 fpmr f14, f0 LFPDUX A8, AO, INC4 fpmr f3, f0 LFPDUX B5, BO, INC4 fpmr f7, f0 LFPDUX A9, AO, INC4 fpmr f11, f0 LFPDUX A2, AO2, INC4 fpmr f15, f0 LFPDUX B2, BO2, INC4 bdz- .L1013 .align 4.L1012:## 1 ## FXCPMADD f0, B1, A1, f0 nop FXCSMADD f4, B1, A1, f4 nop FXCPMADD f8, B2, A1, f8 LFPDUX B4, BO2, INC4 FXCSMADD f12, B2, A1, f12 LFPDUX B6, BO, INC4 FXCPMADD f1, B1, A2, f1 nop FXCSMADD f5, B1, A2, f5 LFPDUX A4, AO2, INC4 FXCPMADD f9, B2, A2, f9 LFPDUX A10, AO, INC4 FXCSMADD f13, B2, A2, f13 nop FXCPMADD f2, B1, A3, f2 nop FXCSMADD f6, B1, A3, f6 nop FXCPMADD f10, B2, A3, f10 nop FXCSMADD f14, B2, A3, f14 nop FXCPMADD f3, B1, A4, f3 nop FXCSMADD f7, B1, A4, f7 LFPDUX A2, AO2, INC4 FXCPMADD f11, B2, A4, f11 LFPDUX A1, AO, INC4 FXCSMADD f15, B2, A4, f15 nop## 2 ## FXCPMADD f0, B3, A5, f0 nop FXCSMADD f4, B3, A5, f4 nop FXCPMADD f8, B4, A5, f8 LFPDUX B2, BO2, INC4 FXCSMADD f12, B4, A5, f12 LFPDUX B1, BO, INC4 FXCPMADD f1, B3, A2, f1 nop FXCSMADD f5, B3, A2, f5 LFPDUX A4, AO2, INC4 FXCPMADD f9, B4, A2, f9 LFPDUX A3, AO, INC4 FXCSMADD f13, B4, A2, f13 nop FXCPMADD f2, B3, A6, f2 nop FXCSMADD f6, B3, A6, f6 nop FXCPMADD f10, B4, A6, f10 nop FXCSMADD f14, B4, A6, f14 nop FXCPMADD f3, B3, A4, f3 nop FXCSMADD f7, B3, A4, f7 LFPDUX A2, AO2, INC4 FXCPMADD f11, B4, A4, f11 LFPDUX A5, AO, INC4 FXCSMADD f15, B4, A4, f15 nop## 3 ## FXCPMADD f0, B5, A7, f0 nop FXCSMADD f4, B5, A7, f4 nop FXCPMADD f8, B2, A7, f8 LFPDUX B4, BO2, INC4 FXCSMADD f12, B2, A7, f12 LFPDUX B3, BO, INC4 FXCPMADD f1, B5, A2, f1 nop FXCSMADD f5, B5, A2, f5 LFPDUX A4, AO2, INC4 FXCPMADD f9, B2, A2, f9 LFPDUX A6, AO, INC4 FXCSMADD f13, B2, A2, f13 nop FXCPMADD f2, B5, A8, f2 nop FXCSMADD f6, B5, A8, f6 nop FXCPMADD f10, B2, A8, f10 nop FXCSMADD f14, B2, A8, f14 nop FXCPMADD f3, B5, A4, f3 nop FXCSMADD f7, B5, A4, f7 LFPDUX A2, AO2, INC4 FXCPMADD f11, B2, A4, f11 LFPDUX A7, AO, INC4 FXCSMADD f15, B2, A4, f15 nop## 4 ## FXCPMADD f0, B6, A9, f0 nop FXCSMADD f4, B6, A9, f4 nop FXCPMADD f8, B4, A9, f8 LFPDUX B2, BO2, INC4 FXCSMADD f12, B4, A9, f12 LFPDUX B5, BO, INC4 FXCPMADD f1, B6, A2, f1 nop FXCSMADD f5, B6, A2, f5 LFPDUX A4, AO2, INC4 FXCPMADD f9, B4, A2, f9 LFPDUX A8, AO, INC4 FXCSMADD f13, B4, A2, f13 nop FXCPMADD f2, B6, A10, f2 nop FXCSMADD f6, B6, A10, f6 nop FXCPMADD f10, B4, A10, f10 nop FXCSMADD f14, B4, A10, f14 nop FXCPMADD f3, B6, A4, f3 LFPDUX A2, AO2, INC4 FXCSMADD f7, B6, A4, f7 LFPDUX A9, AO, INC4 FXCPMADD f11, B4, A4, f11 nop FXCSMADD f15, B4, A4, f15 bdnz+ .L1012 .align 4.L1013:## 1 ## FXCPMADD f0, B1, A1, f0 nop FXCSMADD f4, B1, A1, f4 nop FXCPMADD f8, B2, A1, f8 LFPDUX B4, BO2, INC4 FXCSMADD f12, B2, A1, f12 LFPDUX B6, BO, INC4 FXCPMADD f1, B1, A2, f1 nop FXCSMADD f5, B1, A2, f5 LFPDUX A4, AO2, INC4 FXCPMADD f9, B2, A2, f9 LFPDUX A10, AO, INC4 FXCSMADD f13, B2, A2, f13 nop FXCPMADD f2, B1, A3, f2 nop FXCSMADD f6, B1, A3, f6 nop FXCPMADD f10, B2, A3, f10 nop FXCSMADD f14, B2, A3, f14 nop FXCPMADD f3, B1, A4, f3 nop FXCSMADD f7, B1, A4, f7 LFPDUX A2, AO2, INC4 FXCPMADD f11, B2, A4, f11#ifndef TRMMKERNEL LFDUX A1, CO1, INC#else nop#endif FXCSMADD f15, B2, A4, f15 nop## 2 ## FXCPMADD f0, B3, A5, f0 nop FXCSMADD f4, B3, A5, f4 nop FXCPMADD f8, B4, A5, f8 LFPDUX B2, BO2, INC4 FXCSMADD f12, B4, A5, f12#ifndef TRMMKERNEL LFDUX B1, CO1, INC2#else nop#endif FXCPMADD f1, B3, A2, f1 nop FXCSMADD f5, B3, A2, f5 LFPDUX A4, AO2, INC4 FXCPMADD f9, B4, A2, f9#ifndef TRMMKERNEL LFDUX A3, CO1, INC2#else nop#endif FXCSMADD f13, B4, A2, f13 nop FXCPMADD f2, B3, A6, f2 nop FXCSMADD f6, B3, A6, f6 nop FXCPMADD f10, B4, A6, f10 nop FXCSMADD f14, B4, A6, f14 nop FXCPMADD f3, B3, A4, f3 nop FXCSMADD f7, B3, A4, f7 LFPDUX A2, AO2, INC4 FXCPMADD f11, B4, A4, f11#ifndef TRMMKERNEL LFDUX A5, CO1, INC2#else nop#endif FXCSMADD f15, B4, A4, f15 nop## 3 ## FXCPMADD f0, B5, A7, f0 nop FXCSMADD f4, B5, A7, f4 nop FXCPMADD f8, B2, A7, f8 LFPDUX B4, BO2, INC4 FXCSMADD f12, B2, A7, f12#ifndef TRMMKERNEL LFSDUX A1, CO1, INCM5#else nop#endif FXCPMADD f1, B5, A2, f1 nop FXCSMADD f5, B5, A2, f5 LFPDUX A4, AO2, INC4 FXCPMADD f9, B2, A2, f9#ifndef TRMMKERNEL LFSDUX B1, CO1, INC2#else nop#endif FXCSMADD f13, B2, A2, f13 nop FXCPMADD f2, B5, A8, f2 nop FXCSMADD f6, B5, A8, f6 nop FXCPMADD f10, B2, A8, f10 nop FXCSMADD f14, B2, A8, f14 nop FXCPMADD f3, B5, A4, f3 nop FXCSMADD f7, B5, A4, f7 LFPDUX A2, AO2, INC4 FXCPMADD f11, B2, A4, f11#ifndef TRMMKERNEL LFSDUX A3, CO1, INC2#else nop#endif FXCSMADD f15, B2, A4, f15 nop## 4 ## FXCPMADD f0, B6, A9, f0 nop FXCSMADD f4, B6, A9, f4 nop FXCPMADD f8, B4, A9, f8#ifndef TRMMKERNEL LFSDUX A5, CO1, INC2#else nop#endif FXCSMADD f12, B4, A9, f12#ifndef TRMMKERNEL LFDUX B3, CO2, INC#else nop#endif FXCPMADD f1, B6, A2, f1 nop FXCSMADD f5, B6, A2, f5 LFPDUX A4, AO2, INC4 FXCPMADD f9, B4, A2, f9#ifndef TRMMKERNEL LFDUX A6, CO2, INC2#else nop#endif FXCSMADD f13, B4, A2, f13 nop FXCPMADD f2, B6, A10, f2 nop FXCSMADD f6, B6, A10, f6 nop FXCPMADD f10, B4, A10, f10 nop FXCSMADD f14, B4, A10, f14#ifndef TRMMKERNEL LFDUX A7, CO2, INC2#else nop#endif FXCPMADD f3, B6, A4, f3 nop FXCSMADD f7, B6, A4, f7 nop FXCPMADD f11, B4, A4, f11 nop FXCSMADD f15, B4, A4, f15#ifndef TRMMKERNEL LFDUX B2, CO2, INC2#else nop#endif .align 4.L1014: li r0, ALPHA lfpdx AP, SP, r0#ifdef TRMMKERNEL li r0, FZERO lfpsx f30, SP, r0#endif#if defined(TRMMKERNEL)#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) sub TEMP, K, KK#elif defined(LEFT) addi TEMP, KK, 4#else addi TEMP, KK, 2#endif andi. r0, TEMP, 3 mtspr CTR, r0 ble+ .L1018 cmpwi cr0, TEMP, 3 bgt+ .L1015#else andi. r0, K, 3 mtspr CTR, r0 ble+ .L1018 cmpwi cr0, K, 3 bgt+ .L1015#endif#ifndef TRMMKERNEL LFDUX A1, CO1, INC fpmr f5, f0 LFDUX B1, CO1, INC2 fpmr f9, f0 LFDUX A3, CO1, INC2 fpmr f13, f0 LFDUX A5, CO1, INC2 fpmr f2, f0 LFSDUX A1, CO1, INCM5 fpmr f6, f0 LFSDUX B1, CO1, INC2 fpmr f10, f0 LFSDUX A3, CO1, INC2 fpmr f14, f0 LFSDUX A5, CO1, INC2 fpmr f3, f0 LFDUX B3, CO2, INC fpmr f7, f0 LFDUX A6, CO2, INC2 fpmr f11, f0 LFDUX A7, CO2, INC2 fpmr f15, f0 LFDUX B2, CO2, INC2#else fpmr f5, f0 fpmr f9, f0 fpmr f13, f0 fpmr f2, f0 fpmr f6, f0 fpmr f10, f0 fpmr f14, f0 fpmr f3, f0 fpmr f7, f0 fpmr f11, f0 fpmr f15, f0#endif .align 4.L1015: LFPDUX A2, AO, INC4 LFPDUX A4, AO2, INC4 LFPDUX A10, BO, INC4 LFPDUX B4, BO2, INC4 bdz- .L1017 .align 4.L1016: FXCPMADD f0, A10, A2, f0 FXCSMADD f4, A10, A2, f4 FXCPMADD f8, B4, A2, f8 FXCSMADD f12, B4, A2, f12 LFPDUX A2, AO, INC4 FXCPMADD f1, A10, A4, f1 FXCSMADD f5, A10, A4, f5 FXCPMADD f9, B4, A4, f9 FXCSMADD f13, B4, A4, f13 LFPDUX A4, AO2, INC4 FXCPMADD f2, A10, A2, f2 FXCSMADD f6, A10, A2, f6 FXCPMADD f10, B4, A2, f10 FXCSMADD f14, B4, A2, f14 LFPDUX A2, AO, INC4 FXCPMADD f3, A10, A4, f3 FXCSMADD f7, A10, A4, f7 LFPDUX A10, BO, INC4 FXCPMADD f11, B4, A4, f11 FXCSMADD f15, B4, A4, f15 LFPDUX A4, AO2, INC4 LFPDUX B4, BO2, INC4 bdnz+ .L1016 .align 4.L1017: FXCPMADD f0, A10, A2, f0 FXCSMADD f4, A10, A2, f4 FXCPMADD f8, B4, A2, f8 FXCSMADD f12, B4, A2, f12 LFPDUX A2, AO, INC4 FXCPMADD f1, A10, A4, f1 FXCSMADD f5, A10, A4, f5 FXCPMADD f9, B4, A4, f9 FXCSMADD f13, B4, A4, f13 LFPDUX A4, AO2, INC4 FXCPMADD f2, A10, A2, f2 FXCSMADD f6, A10, A2, f6 FXCPMADD f10, B4, A2, f10 FXCSMADD f14, B4, A2, f14 FXCPMADD f3, A10, A4, f3 FXCSMADD f7, A10, A4, f7 FXCPMADD f11, B4, A4, f11 FXCSMADD f15, B4, A4, f15 .align 4.L1018:#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \ defined(RN) || defined(RT) || defined(CN) || defined(CT) fpadd f0, f0, f4 fpadd f8, f8, f12 fpadd f1, f1, f5 fpadd f9, f9, f13 fpadd f2, f2, f6 fpadd f10, f10, f14 fpadd f3, f3, f7 fpadd f11, f11, f15#else fpsub f0, f0, f4 fpsub f8, f8, f12 fpsub f1, f1, f5 fpsub f9, f9, f13 fpsub f2, f2, f6 fpsub f10, f10, f14 fpsub f3, f3, f7 fpsub f11, f11, f15#endif#ifndef TRMMKERNEL fxcpmadd A1, f0, AP, A1 LFSDUX B3, CO2, INCM5 fxcpmadd B1, f1, AP, B1 LFSDUX A6, CO2, INC2 fxcpmadd A3, f2, AP, A3 LFSDUX A7, CO2, INC2 fxcpmadd A5, f3, AP, A5 LFSDUX B2, CO2, INC2 fxcxnpma f0, f0, AP, A1 fxcpmadd B3, f8, AP, B3 fxcxnpma f1, f1, AP, B1 fxcpmadd A6, f9, AP, A6 fxcxnpma f2, f2, AP, A3 fxcpmadd A7, f10, AP, A7 fxcxnpma f3, f3, AP, A5 STFDUX f0, CO1, INCM7 fxcpmadd B2, f11, AP, B2 STFSDUX f0, CO1, INC fxcxnpma f8, f8, AP, B3 STFDUX f1, CO1, INC STFSDUX f1, CO1, INC fxcxnpma f9, f9, AP, A6 STFDUX f2, CO1, INC STFSDUX f2, CO1, INC fxcxnpma f10, f10, AP, A7 STFDUX f3, CO1, INC STFSDUX f3, CO1, INC fxcxnpma f11, f11, AP, B2 STFDUX f8, CO2, INCM7#else fxcpmadd f12, f0, AP, f30 fxcpmadd f13, f1, AP, f30 fxcpmadd f14, f2, AP, f30 fxcpmadd f15, f3, AP, f30 fxcxnpma f0, f0, AP, f12 fxcxnpma f1, f1, AP, f13 fxcxnpma f2, f2, AP, f14 fxcxnpma f3, f3, AP, f15 fxcpmadd f16, f8, AP, f30 fxcpmadd f17, f9, AP, f30 fxcpmadd f18, f10, AP, f30 fxcpmadd f19, f11, AP, f30 fxcxnpma f8, f8, AP, f16 fxcxnpma f9, f9, AP, f17 fxcxnpma f10, f10, AP, f18 fxcxnpma f11, f11, AP, f19 STFDUX f0, CO1, INC STFSDUX f0, CO1, INC STFDUX f1, CO1, INC STFSDUX f1, CO1, INC STFDUX f2, CO1, INC STFSDUX f2, CO1, INC STFDUX f3, CO1, INC STFSDUX f3, CO1, INC STFDUX f8, CO2, INC#endif STFSDUX f8, CO2, INC STFDUX f9, CO2, INC STFSDUX f9, CO2, INC STFDUX f10, CO2, INC STFSDUX f10, CO2, INC STFDUX f11, CO2, INC STFSDUX f11, CO2, INC#ifdef TRMMKERNEL#if ( defined(LEFT) && defined(TRANSA)) || \ (!defined(LEFT) && !defined(TRANSA)) sub TEMP, K, KK#ifdef LEFT addi TEMP, TEMP, -4#else addi TEMP, TEMP, -2#endif slwi r0, TEMP, 2 + ZBASE_SHIFT slwi TEMP, TEMP, 1 + ZBASE_SHIFT add AO, AO, r0 add BO, BO, TEMP#endif#ifdef LEFT addi KK, KK, 4#endif#endif addic. I, I, -1 li r0, FZERO lfpsx f0, SP, r0 bgt+ .L1011 .align 4.L1020: andi. I, M, 2 beq .L1030#if defined(TRMMKERNEL)#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA)) addi AO2, AO, 2 * SIZE fpmr f4, f0 addi BO, B, - 4 * SIZE fpmr f8, f0 addi BO2, B, - 2 * SIZE fpmr f12, f0#else slwi TEMP, KK, 1 + ZBASE_SHIFT slwi r0, KK, 1 + ZBASE_SHIFT
⌨️ 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 + -