📄 gemm_kernel.s
字号:
STFD f0, 0 * SIZE(CO1) STFD f1, 0 * SIZE(CO2) lfs f0, FZERO fmr f1, f0 fmr f4, f0 fmr f5, f0#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, -2#endif slwi r0, TEMP, 0 + BASE_SHIFT slwi TEMP, TEMP, 1 + BASE_SHIFT add AO, AO, r0 add BO, BO, TEMP#endif#ifdef LEFT addi KK, KK, 1#endif#endif .align 4LL(69):#if defined(TRMMKERNEL) && !defined(LEFT) addi KK, KK, 2#endif mr B, BO .align 4LL(70): mr CO1, C andi. J, N, 1 ble LL(999)#if defined(TRMMKERNEL) && defined(LEFT) mr KK, OFFSET#endif lfs f0, FZERO fmr f1, f0 fmr f2, f0 fmr f3, f0 srawi. I, M, 2 mr AO, A ble LL(80) .align 4LL(71):#if defined(TRMMKERNEL)#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA)) 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) mr BO, B#else slwi r0, KK, 2 + BASE_SHIFT slwi TEMP, KK, 0 + BASE_SHIFT add AO, AO, r0 add BO, B, TEMP 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)#endif dcbt CO1, PREC#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) sub TEMP, K, KK#elif defined(LEFT) addi TEMP, KK, 4#else addi TEMP, KK, 1#endif srawi. TEMP, TEMP, 2 mtspr CTR, TEMP#else 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 srawi. r0, K, 2 mtspr CTR, r0 mr BO, B#endif ble LL(75) .align 5LL(72): FMADD f0, f16, f20, f0 FMADD f1, f17, f20, f1 FMADD f2, f18, f20, f2 FMADD f3, f19, f20, f3 LFD f16, 4 * SIZE(AO) LFD f17, 5 * SIZE(AO) LFD f18, 6 * SIZE(AO) LFD f19, 7 * SIZE(AO) FMADD f0, f16, f21, f0 FMADD f1, f17, f21, f1 FMADD f2, f18, f21, f2 FMADD f3, f19, f21, f3 LFD f16, 8 * SIZE(AO) LFD f17, 9 * SIZE(AO) LFD f18, 10 * SIZE(AO) LFD f19, 11 * SIZE(AO) FMADD f0, f16, f22, f0 FMADD f1, f17, f22, f1 FMADD f2, f18, f22, f2 FMADD f3, f19, f22, f3 LFD f16, 12 * SIZE(AO) LFD f17, 13 * SIZE(AO) LFD f18, 14 * SIZE(AO) LFD f19, 15 * SIZE(AO) FMADD f0, f16, f23, f0 FMADD f1, f17, f23, f1 FMADD f2, f18, f23, f2 FMADD f3, f19, f23, f3 LFD f16, 16 * SIZE(AO) LFD f17, 17 * SIZE(AO) LFD f18, 18 * SIZE(AO) LFD f19, 19 * SIZE(AO) LFD f20, 4 * SIZE(BO) LFD f21, 5 * SIZE(BO) LFD f22, 6 * SIZE(BO) LFD f23, 7 * SIZE(BO) addi AO, AO, 16 * SIZE addi BO, BO, 4 * SIZE PREFETCH_B bdnz LL(72) .align 4LL(75): lfd f30, ALPHA#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, 1#endif andi. TEMP, TEMP, 3 mtspr CTR, TEMP#else andi. r0, K, 3 mtspr CTR, r0#endif ble+ LL(78) .align 4LL(76): FMADD f0, f16, f20, f0 FMADD f1, f17, f20, f1 FMADD f2, f18, f20, f2 FMADD f3, f19, f20, f3 LFD f16, 4 * SIZE(AO) LFD f17, 5 * SIZE(AO) LFD f18, 6 * SIZE(AO) LFD f19, 7 * SIZE(AO) LFD f20, 1 * SIZE(BO) addi BO, BO, 1 * SIZE addi AO, AO, 4 * SIZE bdnz LL(76) .align 4LL(78):#ifndef TRMMKERNEL LFD f16, 0 * SIZE(CO1) LFD f17, 1 * SIZE(CO1) LFD f18, 2 * SIZE(CO1) LFD f19, 3 * SIZE(CO1) FMADD f0, f0, f30, f16 FMADD f1, f1, f30, f17 FMADD f2, f2, f30, f18 FMADD f3, f3, f30, f19#else FMUL f0, f0, f30 FMUL f1, f1, f30 FMUL f2, f2, f30 FMUL f3, f3, f30#endif STFD f0, 0 * SIZE(CO1) STFD f1, 1 * SIZE(CO1) STFD f2, 2 * SIZE(CO1) STFD f3, 3 * SIZE(CO1) lfs f0, FZERO fmr f1, f0 fmr f2, f0 fmr f3, f0 #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, -1#endif slwi r0 , TEMP, 2 + BASE_SHIFT slwi TEMP, TEMP, 0 + BASE_SHIFT add AO, AO, r0 add BO, BO, TEMP#endif#ifdef LEFT addi KK, KK, 4#endif#endif addi CO1, CO1, 4 * SIZE addic. I, I, -1 bgt+ LL(71) .align 4LL(80): andi. I, M, 2 ble LL(90)#if defined(TRMMKERNEL)#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA)) 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) mr BO, B#else slwi r0, KK, 1 + BASE_SHIFT slwi TEMP, KK, 0 + BASE_SHIFT add AO, AO, r0 add BO, B, TEMP 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)#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, 1#endif srawi. TEMP, TEMP, 2 mtspr CTR, TEMP#else 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) srawi. r0, K, 2 mtspr CTR, r0 mr BO, B#endif ble LL(85) .align 5LL(82): FMADD f0, f16, f20, f0 FMADD f1, f17, f20, f1 FMADD f2, f18, f21, f2 FMADD f3, f19, f21, f3 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, f23, f2 FMADD f3, f19, f23, f3 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) addi AO, AO, 8 * SIZE addi BO, BO, 4 * SIZE PREFETCH_B bdnz LL(82) .align 4LL(85): lfd f30, ALPHA#if defined(TRMMKERNEL)#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) sub TEMP, K, KK#elif defined(LEFT) addi TEMP, KK, 2#else addi TEMP, KK, 1#endif andi. TEMP, TEMP, 3 mtspr CTR, TEMP#else andi. r0, K, 3 mtspr CTR, r0#endif ble+ LL(88) .align 4LL(86): FMADD f0, f16, f20, f0 FMADD f1, f17, f20, f1 LFD f16, 2 * SIZE(AO) LFD f17, 3 * SIZE(AO) LFD f20, 1 * SIZE(BO) addi BO, BO, 1 * SIZE addi AO, AO, 2 * SIZE bdnz LL(86) .align 4LL(88):#ifndef TRMMKERNEL LFD f16, 0 * SIZE(CO1) LFD f17, 1 * SIZE(CO1) FADD f0, f2, f0 FADD f1, f3, f1 FMADD f0, f0, f30, f16 FMADD f1, f1, f30, f17#else FADD f0, f2, f0 FADD f1, f3, f1 FMUL f0, f0, f30 FMUL f1, f1, f30#endif STFD f0, 0 * SIZE(CO1) STFD f1, 1 * SIZE(CO1) lfs f0, FZERO fmr f1, f0 fmr f2, f0 fmr f3, f0 addi CO1, CO1, 2 * SIZE#ifdef TRMMKERNEL#if ( defined(LEFT) && defined(TRANSA)) || \ (!defined(LEFT) && !defined(TRANSA)) sub TEMP, K, KK#ifdef LEFT addi TEMP, TEMP, -2#else addi TEMP, TEMP, -1#endif slwi r0 , TEMP, 1 + BASE_SHIFT slwi TEMP, TEMP, 0 + BASE_SHIFT add AO, AO, r0 add BO, BO, TEMP#endif#ifdef LEFT addi KK, KK, 2#endif#endif .align 4LL(90): andi. I, M, 1 ble LL(999)#if defined(TRMMKERNEL)#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA)) 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) mr BO, B#else slwi r0, KK, 0 + BASE_SHIFT slwi TEMP, KK, 0 + BASE_SHIFT add AO, AO, r0 add BO, B, TEMP 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)#endif#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA)) sub TEMP, K, KK#elif defined(LEFT) addi TEMP, KK, 1#else addi TEMP, KK, 1#endif srawi. TEMP, TEMP, 3 mtspr CTR, TEMP#else 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) srawi. r0, K, 3 mtspr CTR, r0 mr BO, B#endif ble LL(95) .align 5LL(92): FMADD f0, f16, f20, f0 FMADD f1, f17, f21, f1 FMADD f2, f18, f22, f2 FMADD f3, f19, f23, f3 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) FMADD f0, f16, f20, f0 FMADD f1, f17, f21, f1 FMADD f2, f18, f22, f2 FMADD f3, f19, f23, f3 LFD f16, 8 * SIZE(AO) LFD f17, 9 * SIZE(AO) LFD f18, 10 * SIZE(AO) LFD f19, 11 * 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, 8 * SIZE addi BO, BO, 8 * SIZE bdnz LL(92) .align 4LL(95): lfd f30, ALPHA#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, 1#endif andi. TEMP, TEMP, 7 mtspr CTR, TEMP#else andi. r0, K, 7 mtspr CTR, r0#endif ble+ LL(98) .align 4LL(96): FMADD f0, f16, f20, f0 LFD f16, 1 * SIZE(AO) LFD f20, 1 * SIZE(BO) addi BO, BO, 1 * SIZE addi AO, AO, 1 * SIZE bdnz LL(96) .align 4LL(98):#ifndef TRMMKERNEL LFD f16, 0 * SIZE(CO1) FADD f0, f1, f0 FADD f2, f3, f2 FADD f0, f2, f0 FMADD f0, f0, f30, f16#else FADD f0, f1, f0 FADD f2, f3, f2 FADD f0, f2, f0 FMUL f0, f0, f30#endif STFD f0, 0 * SIZE(CO1) .align 4LL(999): addi r3, 0, 0 lfd f14, 0(SP) lfd f15, 8(SP) lfd f16, 16(SP) lfd f17, 24(SP) lfd f18, 32(SP) lfd f19, 40(SP) lfd f20, 48(SP) lfd f21, 56(SP) lfd f22, 64(SP) lfd f23, 72(SP) lfd f24, 80(SP) lfd f25, 88(SP) lfd f26, 96(SP) lfd f27, 104(SP) lfd f28, 112(SP) lfd f29, 120(SP) lfd f30, 128(SP) lfd f31, 136(SP)#ifdef __64BIT__ ld r31, 144(SP) ld r30, 152(SP) ld r29, 160(SP) ld r28, 168(SP) ld r27, 176(SP) ld r26, 184(SP) ld r25, 192(SP) ld r24, 200(SP) ld r23, 208(SP) ld r22, 216(SP) ld r21, 224(SP) ld r20, 232(SP)#if defined(TRMMKERNEL) || defined(TRSMKERNEL) ld r19, 240(SP) ld r18, 248(SP)#endif#else lwz r31, 144(SP) lwz r30, 148(SP) lwz r29, 152(SP) lwz r28, 156(SP) lwz r27, 160(SP) lwz r26, 164(SP) lwz r25, 168(SP) lwz r24, 172(SP) lwz r23, 176(SP) lwz r22, 180(SP) lwz r21, 184(SP) lwz r20, 188(SP)#if defined(TRMMKERNEL) || defined(TRSMKERNEL) lwz r19, 192(SP) lwz r18, 196(SP)#endif#endif addi SP, SP, STACKSIZE blr EPILOGUE#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 + -