📄 zgemv_t.s
字号:
FMADD f5, f18, f31, f5 FMADD f6, f19, f30, f6 FMADD f7, f19, f31, f7 LFD f16, 8 * SIZE(AO1) LFD f17, 9 * SIZE(AO1) LFD f18, 8 * SIZE(AO2) LFD f19, 9 * SIZE(AO2) FMADD f8, f20, f30, f8 FMADD f9, f20, f31, f9 FMADD f10, f21, f30, f10 FMADD f11, f21, f31, f11 FMADD f12, f22, f30, f12 FMADD f13, f22, f31, f13 FMADD f14, f23, f30, f14 FMADD f15, f23, f31, f15 LFD f20, 8 * SIZE(AO3) LFD f21, 9 * SIZE(AO3) LFD f22, 8 * SIZE(AO4) LFD f23, 9 * SIZE(AO4) LFD f24, 9 * SIZE(BO) LFD f25, 10 * SIZE(BO) LFD f26, 11 * SIZE(BO) LFD f27, 12 * SIZE(BO) LFD f28, 13 * SIZE(BO) LFD f29, 14 * SIZE(BO) LFD f30, 15 * SIZE(BO) LFDU f31, 16 * SIZE(BO) FMADD f0, f16, f24, f0 FMADD f1, f16, f25, f1 FMADD f2, f17, f24, f2 FMADD f3, f17, f25, f3 FMADD f4, f18, f24, f4 FMADD f5, f18, f25, f5 FMADD f6, f19, f24, f6 FMADD f7, f19, f25, f7 LFD f16, 10 * SIZE(AO1) LFD f17, 11 * SIZE(AO1) LFD f18, 10 * SIZE(AO2) LFD f19, 11 * SIZE(AO2) FMADD f8, f20, f24, f8 FMADD f9, f20, f25, f9 FMADD f10, f21, f24, f10 FMADD f11, f21, f25, f11 FMADD f12, f22, f24, f12 FMADD f13, f22, f25, f13 FMADD f14, f23, f24, f14 FMADD f15, f23, f25, f15 LFD f20, 10 * SIZE(AO3) LFD f21, 11 * SIZE(AO3) LFD f22, 10 * SIZE(AO4) LFD f23, 11 * SIZE(AO4) FMADD f0, f16, f26, f0 FMADD f1, f16, f27, f1 FMADD f2, f17, f26, f2 FMADD f3, f17, f27, f3 FMADD f4, f18, f26, f4 FMADD f5, f18, f27, f5 FMADD f6, f19, f26, f6 FMADD f7, f19, f27, f7 LFD f16, 12 * SIZE(AO1) LFD f17, 13 * SIZE(AO1) LFD f18, 12 * SIZE(AO2) LFD f19, 13 * SIZE(AO2) FMADD f8, f20, f26, f8 FMADD f9, f20, f27, f9 FMADD f10, f21, f26, f10 FMADD f11, f21, f27, f11 FMADD f12, f22, f26, f12 FMADD f13, f22, f27, f13 FMADD f14, f23, f26, f14 FMADD f15, f23, f27, f15 LFD f20, 12 * SIZE(AO3) LFD f21, 13 * SIZE(AO3) LFD f22, 12 * SIZE(AO4) LFD f23, 13 * SIZE(AO4) FMADD f0, f16, f28, f0 FMADD f1, f16, f29, f1 FMADD f2, f17, f28, f2 FMADD f3, f17, f29, f3 FMADD f4, f18, f28, f4 FMADD f5, f18, f29, f5 FMADD f6, f19, f28, f6 FMADD f7, f19, f29, f7 LFD f16, 14 * SIZE(AO1) LFD f17, 15 * SIZE(AO1) LFD f18, 14 * SIZE(AO2) LFD f19, 15 * SIZE(AO2) FMADD f8, f20, f28, f8 FMADD f9, f20, f29, f9 FMADD f10, f21, f28, f10 FMADD f11, f21, f29, f11 FMADD f12, f22, f28, f12 FMADD f13, f22, f29, f13 FMADD f14, f23, f28, f14 FMADD f15, f23, f29, f15 LFD f20, 14 * SIZE(AO3) LFD f21, 15 * SIZE(AO3) LFD f22, 14 * SIZE(AO4) LFD f23, 15 * SIZE(AO4) addi AO1, AO1, 16 * SIZE addi AO2, AO2, 16 * SIZE addi AO3, AO3, 16 * SIZE addi AO4, AO4, 16 * SIZE FMADD f0, f16, f30, f0 FMADD f1, f16, f31, f1 FMADD f2, f17, f30, f2 FMADD f3, f17, f31, f3 FMADD f4, f18, f30, f4 FMADD f5, f18, f31, f5 FMADD f6, f19, f30, f6 FMADD f7, f19, f31, f7 FMADD f8, f20, f30, f8 FMADD f9, f20, f31, f9 FMADD f10, f21, f30, f10 FMADD f11, f21, f31, f11 FMADD f12, f22, f30, f12 FMADD f13, f22, f31, f13 FMADD f14, f23, f30, f14 FMADD f15, f23, f31, f15 .align 4LL(MainN3): andi. r0, MIN_N, 7 mtspr CTR, r0 ble LL(MainFinish) .align 4 LFD f16, 0 * SIZE(AO1) LFD f17, 1 * SIZE(AO1) LFD f18, 0 * SIZE(AO2) LFD f19, 1 * SIZE(AO2) LFD f20, 0 * SIZE(AO3) LFD f21, 1 * SIZE(AO3) LFD f22, 0 * SIZE(AO4) LFD f23, 1 * SIZE(AO4) LFD f24, 1 * SIZE(BO) LFDU f25, 2 * SIZE(BO) addi AO1, AO1, 2 * SIZE addi AO2, AO2, 2 * SIZE addi AO3, AO3, 2 * SIZE addi AO4, AO4, 2 * SIZE bdz LL(MainN3KernelSkip) .align 4LL(MainN3Kernel): FMADD f0, f16, f24, f0 FMADD f1, f16, f25, f1 FMADD f2, f17, f24, f2 FMADD f3, f17, f25, f3 FMADD f4, f18, f24, f4 FMADD f5, f18, f25, f5 FMADD f6, f19, f24, f6 FMADD f7, f19, f25, f7 LFD f16, 0 * SIZE(AO1) LFD f17, 1 * SIZE(AO1) LFD f18, 0 * SIZE(AO2) LFD f19, 1 * SIZE(AO2) FMADD f8, f20, f24, f8 FMADD f9, f20, f25, f9 FMADD f10, f21, f24, f10 FMADD f11, f21, f25, f11 FMADD f12, f22, f24, f12 FMADD f13, f22, f25, f13 FMADD f14, f23, f24, f14 FMADD f15, f23, f25, f15 LFD f20, 0 * SIZE(AO3) LFD f21, 1 * SIZE(AO3) LFD f22, 0 * SIZE(AO4) LFD f23, 1 * SIZE(AO4) LFD f24, 1 * SIZE(BO) LFDU f25, 2 * SIZE(BO) addi AO1, AO1, 2 * SIZE addi AO2, AO2, 2 * SIZE addi AO3, AO3, 2 * SIZE addi AO4, AO4, 2 * SIZE bdnz LL(MainN3Kernel) .align 4 LL(MainN3KernelSkip): FMADD f0, f16, f24, f0 FMADD f1, f16, f25, f1 FMADD f2, f17, f24, f2 FMADD f3, f17, f25, f3 FMADD f4, f18, f24, f4 FMADD f5, f18, f25, f5 FMADD f6, f19, f24, f6 FMADD f7, f19, f25, f7 FMADD f8, f20, f24, f8 FMADD f9, f20, f25, f9 FMADD f10, f21, f24, f10 FMADD f11, f21, f25, f11 FMADD f12, f22, f24, f12 FMADD f13, f22, f25, f13 FMADD f14, f23, f24, f14 FMADD f15, f23, f25, f15 .align 4LL(MainFinish): lfd f30, ALPHA_R lfd f31, ALPHA_I#ifndef XCONJ#ifndef CONJ FSUB f0, f0, f3 FADD f1, f1, f2 FSUB f4, f4, f7 FADD f5, f5, f6 FSUB f8, f8, f11 FADD f9, f9, f10 FSUB f12, f12, f15 FADD f13, f13, f14#else FADD f0, f0, f3 FSUB f1, f1, f2 FADD f4, f4, f7 FSUB f5, f5, f6 FADD f8, f8, f11 FSUB f9, f9, f10 FADD f12, f12, f15 FSUB f13, f13, f14#endif#else#ifndef CONJ FADD f0, f0, f3 FSUB f1, f2, f1 FADD f4, f4, f7 FSUB f5, f6, f5 FADD f8, f8, f11 FSUB f9, f10, f9 FADD f12, f12, f15 FSUB f13, f14, f13#else FSUB f0, f0, f3 FADD f1, f1, f2 FSUB f4, f4, f7 FADD f5, f5, f6 FSUB f8, f8, f11 FADD f9, f9, f10 FSUB f12, f12, f15 FADD f13, f13, f14#endif#endif mr BO, CO cmpwi cr0, INCY, 2 * SIZE bne LL(FinishN1) LFD f16, 0 * SIZE(CO) LFD f17, 1 * SIZE(CO) LFD f18, 2 * SIZE(CO) LFD f19, 3 * SIZE(CO) LFD f20, 4 * SIZE(CO) LFD f21, 5 * SIZE(CO) LFD f22, 6 * SIZE(CO) LFD f23, 7 * SIZE(CO) FMADD f16, f30, f0, f16 FMADDR f17, f30, f1, f17 FMADD f18, f30, f4, f18 FMADDR f19, f30, f5, f19 FMADD f20, f30, f8, f20 FMADDR f21, f30, f9, f21 FMADD f22, f30, f12, f22 FMADDR f23, f30, f13, f23 FMSUBR f16, f31, f1, f16 FMADD f17, f31, f0, f17 FMSUBR f18, f31, f5, f18 FMADD f19, f31, f4, f19 FMSUBR f20, f31, f9, f20 FMADD f21, f31, f8, f21 FMSUBR f22, f31, f13, f22 FMADD f23, f31, f12, f23 STFD f16, 0 * SIZE(CO) STFD f17, 1 * SIZE(CO) STFD f18, 2 * SIZE(CO) STFD f19, 3 * SIZE(CO) STFD f20, 4 * SIZE(CO) STFD f21, 5 * SIZE(CO) STFD f22, 6 * SIZE(CO) STFD f23, 7 * SIZE(CO) addi CO, CO, 8 * SIZE addi J, J, -1 cmpwi cr0, J, 0 bgt LL(MainHead) b LL(Remain) .align 4LL(FinishN1): LFD f16, 0 * SIZE(CO) LFD f17, 1 * SIZE(CO) add CO, CO, INCY LFD f18, 0 * SIZE(CO) LFD f19, 1 * SIZE(CO) add CO, CO, INCY LFD f20, 0 * SIZE(CO) LFD f21, 1 * SIZE(CO) add CO, CO, INCY LFD f22, 0 * SIZE(CO) LFD f23, 1 * SIZE(CO) add CO, CO, INCY FMADD f16, f30, f0, f16 FMADDR f17, f30, f1, f17 FMADD f18, f30, f4, f18 FMADDR f19, f30, f5, f19 FMADD f20, f30, f8, f20 FMADDR f21, f30, f9, f21 FMADD f22, f30, f12, f22 FMADDR f23, f30, f13, f23 FMSUBR f16, f31, f1, f16 FMADD f17, f31, f0, f17 FMSUBR f18, f31, f5, f18 FMADD f19, f31, f4, f19 FMSUBR f20, f31, f9, f20 FMADD f21, f31, f8, f21 FMSUBR f22, f31, f13, f22 FMADD f23, f31, f12, f23 STFD f16, 0 * SIZE(BO) STFD f17, 1 * SIZE(BO) add BO, BO, INCY STFD f18, 0 * SIZE(BO) STFD f19, 1 * SIZE(BO) add BO, BO, INCY STFD f20, 0 * SIZE(BO) STFD f21, 1 * SIZE(BO) add BO, BO, INCY STFD f22, 0 * SIZE(BO) STFD f23, 1 * SIZE(BO) addi J, J, -1 cmpwi cr0, J, 0 bgt LL(MainHead) .align 4 LL(Remain): andi. J, N, 3 ble LL(ISEnd) .align 4LL(RemainHead): mr AO1, A add A, A, LDA mr BO, XP lfd 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 srawi. r0 , MIN_N, 3 mtspr CTR, r0 ble LL(RemainN3) LFD f16, 0 * SIZE(AO1) LFD f17, 1 * SIZE(AO1) LFD f18, 2 * SIZE(AO1) LFD f19, 3 * SIZE(AO1) LFD f20, 4 * SIZE(AO1) LFD f21, 5 * SIZE(AO1) LFD f22, 6 * SIZE(AO1) LFD f23, 7 * SIZE(AO1) LFD f24, 1 * SIZE(BO) LFD f25, 2 * SIZE(BO) LFD f26, 3 * SIZE(BO) LFD f27, 4 * SIZE(BO) LFD f28, 5 * SIZE(BO) LFD f29, 6 * SIZE(BO) LFD f30, 7 * SIZE(BO) LFD f31, 8 * SIZE(BO) bdz LL(RemainKernelSkip) .align 4LL(RemainKernel): FMADD f0, f16, f24, f0 FMADD f1, f16, f25, f1 FMADD f2, f17, f24, f2 FMADD f3, f17, f25, f3 FMADD f4, f18, f26, f4 FMADD f5, f18, f27, f5 FMADD f6, f19, f26, f6 FMADD f7, f19, f27, f7 LFD f16, 8 * SIZE(AO1) LFD f17, 9 * SIZE(AO1) LFD f18, 10 * SIZE(AO1) LFD f19, 11 * SIZE(AO1) LFD f24, 9 * SIZE(BO) LFD f25, 10 * SIZE(BO) LFD f26, 11 * SIZE(BO) LFD f27, 12 * SIZE(BO) FMADD f8, f20, f28, f8 FMADD f9, f20, f29, f9 FMADD f10, f21, f28, f10 FMADD f11, f21, f29, f11 FMADD f12, f22, f30, f12 FMADD f13, f22, f31, f13 FMADD f14, f23, f30, f14 FMADD f15, f23, f31, f15 LFD f20, 12 * SIZE(AO1) LFD f21, 13 * SIZE(AO1) LFD f22, 14 * SIZE(AO1) LFD f23, 15 * SIZE(AO1) LFD f28, 13 * SIZE(BO) LFD f29, 14 * SIZE(BO) LFD f30, 15 * SIZE(BO) LFD f31, 16 * SIZE(BO) FMADD f0, f16, f24, f0 FMADD f1, f16, f25, f1 FMADD f2, f17, f24, f2 FMADD f3, f17, f25, f3 FMADD f4, f18, f26, f4 FMADD f5, f18, f27, f5 FMADD f6, f19, f26, f6 FMADD f7, f19, f27, f7 LFD f16, 16 * SIZE(AO1) LFD f17, 17 * SIZE(AO1) LFD f18, 18 * SIZE(AO1) LFD f19, 19 * SIZE(AO1) LFD f24, 17 * SIZE(BO) LFD f25, 18 * SIZE(BO) LFD f26, 19 * SIZE(BO) LFD f27, 20 * SIZE(BO) FMADD f8, f20, f28, f8 FMADD f9, f20, f29, f9 FMADD f10, f21, f28, f10 FMADD f11, f21, f29, f11 FMADD f12, f22, f30, f12 FMADD f13, f22, f31, f13 FMADD f14, f23, f30, f14 FMADD f15, f23, f31, f15 LFD f20, 20 * SIZE(AO1) LFD f21, 21 * SIZE(AO1) LFD f22, 22 * SIZE(AO1) LFD f23, 23 * SIZE(AO1) LFD f28, 21 * SIZE(BO) LFD f29, 22 * SIZE(BO) LFD f30, 23 * SIZE(BO) LFD f31, 24 * SIZE(BO) addi AO1, AO1, 16 * SIZE addi BO, BO, 16 * SIZE#ifdef linux DCBT(AO1, PREA, SIZE / 2)#else dcbt PREA, AO1#endif bdnz LL(RemainKernel) .align 4 LL(RemainKernelSkip): FMADD f0, f16, f24, f0 FMADD f1, f16, f25, f1 FMADD f2, f17, f24, f2 FMADD f3, f17, f25, f3 FMADD f4, f18, f26, f4 FMADD f5, f18, f27, f5 FMADD f6, f19, f26, f6 FMADD f7, f19, f27, f7 LFD f16, 8 * SIZE(AO1) LFD f17, 9 * SIZE(AO1) LFD f18, 10 * SIZE(AO1) LFD f19, 11 * SIZE(AO1) LFD f24, 9 * SIZE(BO) LFD f25, 10 * SIZE(BO) LFD f26, 11 * SIZE(BO) LFD f27, 12 * SIZE(BO) FMADD f8, f20, f28, f8 FMADD f9, f20, f29, f9 FMADD f10, f21, f28, f10 FMADD f11, f21, f29, f11 FMADD f12, f22, f30, f12 FMADD f13, f22, f31, f13 FMADD f14, f23, f30, f14 FMADD f15, f23, f31, f15 LFD f20, 12 * SIZE(AO1) LFD f21, 13 * SIZE(AO1) LFD f22, 14 * SIZE(AO1) LFD f23, 15 * SIZE(AO1) LFD f28, 13 * SIZE(BO) LFD f29, 14 * SIZE(BO) LFD f30, 15 * SIZE(BO) LFDU f31, 16 * SIZE(BO) FMADD f0, f16, f24, f0 FMADD f1, f16, f25, f1 FMADD f2, f17, f24, f2 FMADD f3, f17, f25, f3 FMADD f4, f18, f26, f4 FMADD f5, f18, f27, f5 FMADD f6, f19, f26, f6 FMADD f7, f19, f27, f7 FMADD f8, f20, f28, f8 FMADD f9, f20, f29, f9 FMADD f10, f21, f28, f10 FMADD f11, f21, f29, f11 FMADD f12, f22, f30, f12 FMADD f13, f22, f31, f13 FMADD f14, f23, f30, f14 FMADD f15, f23, f31, f15 addi AO1, AO1, 16 * SIZE .align 4LL(RemainN3): andi. r0, MIN_N, 7 mtspr CTR, r0 ble LL(RemainFinish) .align 4 LFD f16, 0 * SIZE(AO1) LFD f17, 1 * SIZE(AO1) LFD f24, 1 * SIZE(BO) LFDU f25, 2 * SIZE(BO) addi AO1, AO1, 2 * SIZE bdz LL(RemainN3KernelSkip) .align 4LL(RemainN3Kernel): FMADD f0, f16, f24, f0 FMADD f1, f16, f25, f1 FMADD f2, f17, f24, f2 FMADD f3, f17, f25, f3 LFD f16, 0 * SIZE(AO1) LFD f17, 1 * SIZE(AO1) LFD f24, 1 * SIZE(BO) LFDU f25, 2 * SIZE(BO) addi AO1, AO1, 2 * SIZE bdnz LL(RemainN3Kernel) .align 4 LL(RemainN3KernelSkip): FMADD f0, f16, f24, f0 FMADD f1, f16, f25, f1 FMADD f2, f17, f24, f2 FMADD f3, f17, f25, f3 .align 4LL(RemainFinish): lfd f30, ALPHA_R lfd f31, ALPHA_I LFD f16, 0 * SIZE(CO) LFD f17, 1 * SIZE(CO) FADD f0, f0, f4 FADD f1, f1, f5 FADD f2, f2, f6 FADD f3, f3, f7 FADD f8, f8, f12 FADD f9, f9, f13 FADD f10, f10, f14 FADD f11, f11, f15 FADD f0, f0, f8 FADD f1, f1, f9 FADD f2, f2, f10 FADD f3, f3, f11#ifndef XCONJ#ifndef CONJ FSUB f0, f0, f3 FADD f1, f1, f2#else FADD f0, f0, f3 FSUB f1, f1, f2#endif#else#ifndef CONJ FADD f0, f0, f3 FSUB f1, f2, f1#else FSUB f0, f0, f3 FADD f1, f1, f2#endif#endif FMADD f16, f30, f0, f16 FMADDR f17, f30, f1, f17 FMSUBR f16, f31, f1, f16 FMADD f17, f31, f0, f17 STFD f16, 0 * SIZE(CO) STFD f17, 1 * SIZE(CO) add CO, CO, INCY addi J, J, -1 cmpi cr0, 0, J, 0 bgt LL(RemainHead) .align 4LL(ISEnd): subf A, PLDA_M, A addi IS, IS, P cmp cr0, 0, IS, M blt LL(ISLoop) .align 4LL(End): li r3, 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 r14, 144(SP) ld r15, 152(SP) ld r16, 160(SP) ld r17, 168(SP) ld r18, 176(SP) ld r19, 184(SP) ld r20, 192(SP) ld r21, 200(SP) ld r22, 208(SP) ld r23, 216(SP) ld r24, 224(SP) ld r25, 232(SP)#else lwz r14, 144(SP) lwz r15, 148(SP) lwz r16, 152(SP) lwz r17, 156(SP) lwz r18, 160(SP) lwz r19, 164(SP) lwz r20, 168(SP) lwz r21, 172(SP) lwz r22, 176(SP) lwz r23, 180(SP) lwz r24, 184(SP) lwz r25, 188(SP)#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 + -