📄 gemm_kernel_hummer.s
字号:
LFPDUX A5, BO, INC4 fxcpmadd f1, A4, A6, f1 LFPDUX A6, BO2, INC4 fxcsmadd f2, A4, A7, f2 LFPDUX A7, BO, INC4 fxcsmadd f3, A4, A8, f3 LFPDUX A8, BO2, INC4 LFPDUX A4, AO2, INC4 bdnz+ .L42 .align 4.L43: fxcpmadd f0, A1, B1, f0 LFPDUX B1, BO, INC4 fxcpmadd f1, A1, B2, f1 LFPDUX B2, BO2, INC4 fxcsmadd f2, A1, B3, f2 LFPDUX B3, BO, INC4 fxcsmadd f3, A1, B4, f3 LFPDUX B4, BO2, INC4 fxcpmadd f0, A2, A5, f0 LFPDUX A5, BO, INC4 fxcpmadd f1, A2, A6, f1 LFPDUX A6, BO2, INC4 fxcsmadd f2, A2, A7, f2 LFPDUX A7, BO, INC4 fxcsmadd f3, A2, A8, f3 LFPDUX A8, BO2, INC4 fxcpmadd f0, A3, B1, f0 fxcpmadd f1, A3, B2, f1 fxcsmadd f2, A3, B3, f2 fxcsmadd f3, A3, B4, f3 fxcpmadd f0, A4, A5, f0 fxcpmadd f1, A4, A6, f1 fxcsmadd f2, A4, A7, f2 fxcsmadd f3, A4, A8, f3 .align 4.L44: lfd AP, ALPHA(SP)#ifdef TRMMKERNEL fsmfp AP, AP#endif#if defined(TRMMKERNEL)#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) sub TEMP, K, KK#elif defined(LEFT) addi TEMP, KK, 1#else addi TEMP, KK, 4#endif andi. TEMP, TEMP, 7 mtspr CTR, TEMP#else andi. r0, K, 7 mtspr CTR, r0#endif ble+ .L48 LFDX A1, AO, INC4 LFPDUX B1, BO, INC4 LFPDUX B2, BO2, INC4 add AO, AO, INC bdz- .L47 .align 4.L46: fxcpmadd f0, A1, B1, f0 LFPDUX B1, BO, INC4 fxcpmadd f1, A1, B2, f1 LFDX A1, AO, INC4 LFPDUX B2, BO2, INC4 add AO, AO, INC bdnz+ .L46 .align 4.L47: fxcpmadd f0, A1, B1, f0 fxcpmadd f1, A1, B2, f1 .align 4.L48:#ifndef TRMMKERNEL LFDX A1, CO1, INC2 LFDX A2, CO2, INC2 LFDX A3, CO3, INC2 LFDX A4, CO4, INC2 fpadd f0, f0, f2 fpadd f1, f1, f3 fsmfp A1, A2 fsmfp A3, A4 fxcpmadd f0, AP, f0, A1 fxcpmadd f1, AP, f1, A3#else fpadd f0, f0, f2 fpadd f1, f1, f3 fpmul f0, AP, f0 fpmul f1, AP, f1#endif STFDX f0, CO1, INC2 STFSDX f0, CO2, INC2 STFDX f1, CO3, INC2 STFSDX f1, CO4, INC2#ifdef TRMMKERNEL#if ( defined(LEFT) && defined(TRANSA)) || \ (!defined(LEFT) && !defined(TRANSA)) sub TEMP, K, KK#ifdef LEFT addi TEMP, TEMP, -1#else addi TEMP, TEMP, -4#endif slwi r0, TEMP, 0 + BASE_SHIFT slwi TEMP, TEMP, 2 + BASE_SHIFT add AO, AO, r0 add BO, BO, TEMP#endif#ifdef LEFT addi KK, KK, 1#endif#endif .align 4.L49:#if defined(TRMMKERNEL) && !defined(LEFT) addi KK, KK, 4#endif addi B, BO, 4 * SIZE addic. J, J, -1 bgt+ .L10 .align 4.L50: andi. J, N, 2 beq .L90 mr CO1, C add CO2, C, LDC add C, CO2, LDC#if defined(TRMMKERNEL) && defined(LEFT) mr KK, OFFSET#endif addi AO, A, -2 * SIZE li r0, FZERO lfpsx f0, SP, r0 srawi. I, M, 3 ble .L60 .align 4.L51:#if defined(TRMMKERNEL)#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA)) fpmr f4, f0 addi BO, B, - 2 * SIZE fpmr f1, f0 fpmr f5, f0 fpmr f2, f0 fpmr f6, f0#else slwi TEMP, KK, 3 + BASE_SHIFT slwi r0, KK, 1 + BASE_SHIFT add AO, AO, TEMP add BO, B, r0 fpmr f4, f0 addi BO, BO, - 2 * SIZE fpmr f1, f0 fpmr f5, f0 fpmr f2, f0 fpmr f6, f0#endif#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) sub TEMP, K, KK#elif defined(LEFT) addi TEMP, KK, 8#else addi TEMP, KK, 2#endif srawi. r0, TEMP, 2 fpmr f3, f0 mtspr CTR, r0 fpmr f7, f0 ble .L54#else fpmr f4, f0 addi BO, B, - 2 * SIZE fpmr f1, f0 fpmr f5, f0 fpmr f2, f0 fpmr f6, f0 srawi. r0, K, 2 fpmr f3, f0 mtspr CTR, r0 fpmr f7, f0 ble .L54#endif LFPDUX B1, BO, INC2 LFPDUX A1, AO, INC2 LFPDUX A2, AO, INC2 LFPDUX B2, BO, INC2 LFPDUX A3, AO, INC2 LFPDUX A4, AO, INC2 LFPDUX B3, BO, INC2 LFPDUX A5, AO, INC2 LFPDUX A6, AO, INC2 LFPDUX A7, AO, INC2 LFPDUX A8, AO, INC2 bdz- .L53 .align 4.L52: fxcpmadd f0, B1, A1, f0 LFPDUX B4, BO, INC2 fxcsmadd f4, B1, A1, f4 LFPDUX A1, AO, INC2 fxcpmadd f1, B1, A2, f1 nop fxcsmadd f5, B1, A2, f5 LFPDUX A2, AO, INC2 fxcpmadd f2, B1, A3, f2 nop fxcsmadd f6, B1, A3, f6 LFPDUX A3, AO, INC2 fxcpmadd f3, B1, A4, f3 nop fxcsmadd f7, B1, A4, f7 LFPDUX A4, AO, INC2 fxcpmadd f0, B2, A5, f0 LFPDUX B1, BO, INC2 fxcsmadd f4, B2, A5, f4 LFPDUX A5, AO, INC2 fxcpmadd f1, B2, A6, f1 nop fxcsmadd f5, B2, A6, f5 LFPDUX A6, AO, INC2 fxcpmadd f2, B2, A7, f2 nop fxcsmadd f6, B2, A7, f6 LFPDUX A7, AO, INC2 fxcpmadd f3, B2, A8, f3 nop fxcsmadd f7, B2, A8, f7 LFPDUX A8, AO, INC2 fxcpmadd f0, B3, A1, f0 LFPDUX B2, BO, INC2 fxcsmadd f4, B3, A1, f4 LFPDUX A1, AO, INC2 fxcpmadd f1, B3, A2, f1 nop fxcsmadd f5, B3, A2, f5 LFPDUX A2, AO, INC2 fxcpmadd f2, B3, A3, f2 nop fxcsmadd f6, B3, A3, f6 LFPDUX A3, AO, INC2 fxcpmadd f3, B3, A4, f3 nop fxcsmadd f7, B3, A4, f7 LFPDUX A4, AO, INC2 fxcpmadd f0, B4, A5, f0 LFPDUX B3, BO, INC2 fxcsmadd f4, B4, A5, f4 LFPDUX A5, AO, INC2 fxcpmadd f1, B4, A6, f1 nop fxcsmadd f5, B4, A6, f5 LFPDUX A6, AO, INC2 fxcpmadd f2, B4, A7, f2 nop fxcsmadd f6, B4, A7, f6 LFPDUX A7, AO, INC2 fxcpmadd f3, B4, A8, f3 nop fxcsmadd f7, B4, A8, f7 LFPDUX A8, AO, INC2 bdnz+ .L52 .align 4.L53: fxcpmadd f0, B1, A1, f0 LFPDUX B4, BO, INC2 fxcsmadd f4, B1, A1, f4 LFPDUX A1, AO, INC2 fxcpmadd f1, B1, A2, f1 nop fxcsmadd f5, B1, A2, f5 LFPDUX A2, AO, INC2 fxcpmadd f2, B1, A3, f2 nop fxcsmadd f6, B1, A3, f6 LFPDUX A3, AO, INC2 fxcpmadd f3, B1, A4, f3 nop fxcsmadd f7, B1, A4, f7 LFPDUX A4, AO, INC2 fxcpmadd f0, B2, A5, f0 nop fxcsmadd f4, B2, A5, f4 LFPDUX A5, AO, INC2 fxcpmadd f1, B2, A6, f1 nop fxcsmadd f5, B2, A6, f5 LFPDUX A6, AO, INC2 fxcpmadd f2, B2, A7, f2 nop fxcsmadd f6, B2, A7, f6 LFPDUX A7, AO, INC2 fxcpmadd f3, B2, A8, f3 nop fxcsmadd f7, B2, A8, f7 LFPDUX A8, AO, INC2 fxcpmadd f0, B3, A1, f0 fxcsmadd f4, B3, A1, f4 fxcpmadd f1, B3, A2, f1 fxcsmadd f5, B3, A2, f5 fxcpmadd f2, B3, A3, f2 fxcsmadd f6, B3, A3, f6 fxcpmadd f3, B3, A4, f3 fxcsmadd f7, B3, A4, f7 fxcpmadd f0, B4, A5, f0 fxcsmadd f4, B4, A5, f4 fxcpmadd f1, B4, A6, f1 fxcsmadd f5, B4, A6, f5 fxcpmadd f2, B4, A7, f2 fxcsmadd f6, B4, A7, f6 fxcpmadd f3, B4, A8, f3 fxcsmadd f7, B4, A8, f7 .align 4.L54: lfd AP, ALPHA(SP)#ifdef TRMMKERNEL fsmfp AP, AP#endif#if defined(TRMMKERNEL)#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) sub TEMP, K, KK#elif defined(LEFT) addi TEMP, KK, 8#else addi TEMP, KK, 2#endif andi. TEMP, TEMP, 3 mtspr CTR, TEMP#else andi. r0, K, 3 mtspr CTR, r0#endif ble+ .L58 LFPDUX A1, AO, INC2 LFPDUX B1, BO, INC2 LFPDUX A2, AO, INC2 LFPDUX A3, AO, INC2 LFPDUX A4, AO, INC2 bdz- .L57 .align 4.L56: fxcpmadd f0, B1, A1, f0 fxcsmadd f4, B1, A1, f4 LFPDUX A1, AO, INC2 fxcpmadd f1, B1, A2, f1 fxcsmadd f5, B1, A2, f5 LFPDUX A2, AO, INC2 fxcpmadd f2, B1, A3, f2 fxcsmadd f6, B1, A3, f6 LFPDUX A3, AO, INC2 fxcpmadd f3, B1, A4, f3 fxcsmadd f7, B1, A4, f7 LFPDUX A4, AO, INC2 LFPDUX B1, BO, INC2 bdnz+ .L56 .align 4.L57: fxcpmadd f0, B1, A1, f0 fxcsmadd f4, B1, A1, f4 fxcpmadd f1, B1, A2, f1 fxcsmadd f5, B1, A2, f5 fxcpmadd f2, B1, A3, f2 fxcsmadd f6, B1, A3, f6 fxcpmadd f3, B1, A4, f3 fxcsmadd f7, B1, A4, f7 .align 4.L58:#ifndef TRMMKERNEL LFPDUX A1, CO1, INC2 LFPDUX B1, CO1, INC2 LFPDUX A3, CO1, INC2 LFPDUX A5, CO1, INC2 LFPDUX B3, CO2, INC2 LFPDUX A6, CO2, INC2 LFPDUX A7, CO2, INC2 LFPDUX B2, CO2, INC2 fxcpmadd f0, AP, f0, A1 fxcpmadd f1, AP, f1, B1 fxcpmadd f2, AP, f2, A3 fxcpmadd f3, AP, f3, A5 fxcpmadd f4, AP, f4, B3 fxcpmadd f5, AP, f5, A6 STFPDUX f0, CO1, INCM7 fxcpmadd f6, AP, f6, A7 STFPDUX f1, CO1, INC2 fxcpmadd f7, AP, f7, B2 STFPDUX f2, CO1, INC2 STFPDUX f3, CO1, INC2 STFPDUX f4, CO2, INCM7 STFPDUX f5, CO2, INC2 STFPDUX f6, CO2, INC2 STFPDUX f7, CO2, INC2#else fpmul f0, AP, f0 fpmul f1, AP, f1 fpmul f2, AP, f2 fpmul f3, AP, f3 fpmul f4, AP, f4 fpmul f5, AP, f5 STFPDUX f0, CO1, INC2 fpmul f6, AP, f6 STFPDUX f1, CO1, INC2 fpmul f7, AP, f7 STFPDUX f2, CO1, INC2 STFPDUX f3, CO1, INC2 STFPDUX f4, CO2, INC2 STFPDUX f5, CO2, INC2 STFPDUX f6, CO2, INC2 STFPDUX f7, CO2, INC2#endif#ifdef TRMMKERNEL#if ( defined(LEFT) && defined(TRANSA)) || \ (!defined(LEFT) && !defined(TRANSA)) sub TEMP, K, KK#ifdef LEFT addi TEMP, TEMP, -8#else addi TEMP, TEMP, -2#endif slwi r0, TEMP, 3 + BASE_SHIFT slwi TEMP, TEMP, 1 + BASE_SHIFT add AO, AO, r0 add BO, BO, TEMP#endif#ifdef LEFT addi KK, KK, 8#endif#endif addic. I, I, -1 li r0, FZERO lfpsx f0, SP, r0 bgt+ .L51 .align 4.L60: andi. I, M, 4 beq .L70#if defined(TRMMKERNEL)#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA)) addi BO, B, - 2 * SIZE fpmr f1, f0#else slwi TEMP, KK, 2 + BASE_SHIFT slwi r0, KK, 1 + BASE_SHIFT add AO, AO, TEMP add BO, B, r0 addi BO, BO, - 2 * SIZE fpmr f1, 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 fpmr f2, f0 srawi. r0, TEMP, 2 mtspr CTR, r0 fpmr f3, f0 ble .L64#else srawi. r0, K, 2 fpmr f1, f0 addi BO, B, - 2 * SIZE fpmr f2, f0 mtspr CTR, r0 fpmr f3, f0 ble .L64#endif LFPDUX B1, BO, INC2 LFPDUX A1, AO, INC2 LFPDUX A2, AO, INC2 LFPDUX B2, BO, INC2 LFPDUX A3, AO, INC2 LFPDUX A4, AO, INC2 LFPDUX B3, BO, INC2 LFPDUX A5, AO, INC2 LFPDUX A6, AO, INC2 LFPDUX B4, BO, INC2 LFPDUX A7, AO, INC2 LFPDUX A8, AO, INC2 bdz- .L63 .align 4.L62: fxcpmadd f0, B1, A1, f0 fxcsmadd f2, B1, A1, f2 LFPDUX A1, AO, INC2 fxcpmadd f1, B1, A2, f1 fxcsmadd f3, B1, A2, f3 LFPDUX A2, AO, INC2 LFPDUX B1, BO, INC2 fxcpmadd f0, B2, A3, f0 fxcsmadd f2, B2, A3, f2 LFPDUX A3, AO, INC2 fxcpmadd f1, B2, A4, f1 fxcsmadd f3, B2, A4, f3 LFPDUX A4, AO, INC2 LFPDUX B2, BO, INC2 fxcpmadd f0, B3, A5, f0 fxcsmadd f2, B3, A5, f2 LFPDUX A5, AO, INC2 fxcpmadd f1, B3, A6, f1 fxcsmadd f3, B3, A6, f3 LFPDUX A6, AO, INC2 LFPDUX B3, BO, INC2 fxcpmadd f0, B4, A7, f0 fxcsmadd f2, B4, A7, f2 LFPDUX A7, AO, INC2 fxcpmadd f1, B4, A8, f1 fxcsmadd f3, B4, A8, f3 LFPDUX A8, AO, INC2 LFPDUX B4, BO, INC2 bdnz+ .L62 .align 4.L63: fxcpmadd f0, B1, A1, f0 fxcsmadd f2, B1, A1, f2 fxcpmadd f1, B1, A2, f1 fxcsmadd f3, B1, A2, f3 fxcpmadd f0, B2, A3, f0 fxcsmadd f2, B2, A3, f2 fxcpmadd f1, B2, A4, f1 fxcsmadd f3, B2, A4, f3 fxcpmadd f0, B3, A5, f0 fxcsmadd f2, B3, A5, f2 fxcpmadd f1, B3, A6, f1 fxcsmadd f3, B3, A6, f3 fxcpmadd f0, B4, A7, f0 fxcsmadd f2, B4, A7, f2 fxcpmadd f1, B4, A8, f1 fxcsmadd f3, B4, A8, f3 .align 4.L64: lfd AP, ALPHA(SP)#ifdef TRMMKERNEL fsmfp AP, AP#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. TEMP, TEMP, 3 mtspr CTR, TEMP#else andi. r0, K, 3 mtspr CTR, r0#endif ble+ .L68 LFPDUX A1, AO, INC2 LFPDUX B1, BO, INC2 LFPDUX A2, AO, INC2 bdz- .L67 .align 4.L66: fxcpmadd f0, B1, A1, f0 fxcsmadd f2, B1, A1, f2 LFPDUX A1, AO, INC2 fxcpmadd f1, B1, A2, f1 fxcsmadd f3, B1, A2, f3 LFPDUX B1, BO, INC2 LFPDUX A2, AO, INC2 bdnz+ .L66 .align 4.L67: fxcpmadd f0, B1, A1, f0 fxcsmadd f2, B1, A1, f2 fxcpmadd f1, B1, A2, f1 fxcsmadd f3, B1, A2, f3 .align 4.L68:#ifndef TRMMKERNEL LFPDUX A1, CO1, INC2 LFPDUX A2, CO1, INC2 LFPDUX A3, CO2, INC2 LFPDUX A4, CO2, INC2 fxcpmadd f0, AP, f0, A1 fxcpmadd f1, AP, f1, A2 fxcpmadd f2, AP, f2, A3 fxcpmadd f3, AP, f3, A4 STFPDUX f0, CO1, INCM3 STFPDUX f1, CO1, INC2 STFPDUX f2, CO2, INCM3 STFPDUX f3, CO2, INC2#else fpmul f0, AP, f0 fpmul f1, AP, f1 fpmul f2, AP, f2 fpmul f3, AP, f3 STFPDUX f0, CO1, INC2 STFPDUX f1, CO1, INC2 STFPDUX f2, CO2, INC2 STFPDUX f3, CO2, INC2#endif#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 + BASE_SHIFT slwi TEMP, TEMP, 1 + BASE_SHIFT add AO, AO, r0 add BO, BO, TEMP#endif#ifdef LEFT addi KK, KK, 4#endif#endif li r0, FZERO lfpsx f0, SP, r0 .align 4.L70: andi. I, M, 2 beq .L80#if defined(TRMMKERNEL)#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA)) addi BO, B, - 2 * SIZE fpmr f1, f0#else slwi TEMP, KK, 1 + BASE_SHIFT slwi r0, KK, 1 + BASE_SHIFT add AO, AO, TEMP add BO, B, r0 addi BO, BO, - 2 * SIZE fpmr f1, f0#endif#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) sub TEMP, K, KK#elif defined(LEFT) addi TEMP, KK, 2#else addi TEMP, KK, 2#endif srawi. r0, TEMP, 3 fpmr f2, f0 mtspr CTR, r0 fpmr f3, f0 ble .L74#else addi BO, B, - 2 * SIZE fpmr f1, f0 srawi. r0, K, 3 fpmr f2, f0 mtspr CTR, r0 fpmr f3, f0 ble .L74#endif
⌨️ 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 + -