📄 gemv_n_ppc440.s
字号:
FMADD y01, alpha3, a5, y01 FMADD y02, alpha3, a6, y02 FMADD y01, alpha4, a7, y01 FMADD y02, alpha4, a8, y02 STFDU y01, 1 * SIZE(Y2) STFDU y02, 1 * SIZE(Y2) .align 4LL(28): andi. r0, M, 1 ble LL(29) LFDU a1, 1 * SIZE(AO1) LFDU y01, 1 * SIZE(Y1) LFDU a2, 1 * SIZE(AO2) LFDU a3, 1 * SIZE(AO3) LFDU a4, 1 * SIZE(AO4) FMADD y01, alpha1, a1, y01 FMADD y01, alpha2, a2, y01 FMADD y01, alpha3, a3, y01 FMADD y01, alpha4, a4, y01 STFDU y01, 1 * SIZE(Y2) .align 4LL(29): addi J, J, -1 lfd alpha, ALPHA cmpi cr0, 0, J, 0 bgt LL(21) .align 4LL(30): andi. J, N, 2 ble LL(40) LFDUX alpha1, X, INCX LFDUX alpha2, X, INCX mr AO1, A add AO2, A, LDA add A, AO2, LDA FMUL alpha1, alpha, alpha1 mr Y1, YY FMUL alpha2, alpha, alpha2 mr Y2, YY srawi. r0, M, 3 mtspr CTR, r0 ble LL(35) LFDU y01, 1 * SIZE(Y1) LFDU a1, 1 * SIZE(AO1) LFDU y02, 1 * SIZE(Y1) LFDU a2, 1 * SIZE(AO1) LFDU y03, 1 * SIZE(Y1) LFDU a3, 1 * SIZE(AO1) LFDU y04, 1 * SIZE(Y1) LFDU a4, 1 * SIZE(AO1) LFDU y05, 1 * SIZE(Y1) LFDU a5, 1 * SIZE(AO1) LFDU y06, 1 * SIZE(Y1) LFDU a6, 1 * SIZE(AO1) LFDU y07, 1 * SIZE(Y1) LFDU a7, 1 * SIZE(AO1) LFDU y08, 1 * SIZE(Y1) LFDU a8, 1 * SIZE(AO1) bdz LL(33) .align 4LL(32):#ifdef PPCG4 dcbtst Y1, PREA#endif FMADD y09, alpha1, a1, y01 LFDU a1, 1 * SIZE(AO2) FMADD y10, alpha1, a2, y02 LFDU a2, 1 * SIZE(AO2) FMADD y11, alpha1, a3, y03 LFDU a3, 1 * SIZE(AO2) FMADD y12, alpha1, a4, y04 LFDU a4, 1 * SIZE(AO2) LFDU y01, 1 * SIZE(Y1) LFDU y02, 1 * SIZE(Y1)#ifdef PPCG4 dcbt AO2, PREA#endif FMADD y13, alpha1, a5, y05 LFDU a5, 1 * SIZE(AO2) FMADD y14, alpha1, a6, y06 LFDU a6, 1 * SIZE(AO2) FMADD y15, alpha1, a7, y07 LFDU a7, 1 * SIZE(AO2) FMADD y16, alpha1, a8, y08 LFDU a8, 1 * SIZE(AO2) LFDU y03, 1 * SIZE(Y1) LFDU y04, 1 * SIZE(Y1)#if defined(PPCG4) && defined(DOUBLE) dcbt AO2, PREA#endif FMADD y09, alpha2, a1, y09 LFDU a1, 1 * SIZE(AO1) FMADD y10, alpha2, a2, y10 LFDU a2, 1 * SIZE(AO1) FMADD y11, alpha2, a3, y11 LFDU a3, 1 * SIZE(AO1) FMADD y12, alpha2, a4, y12 LFDU a4, 1 * SIZE(AO1)#if defined(PPCG4) && defined(DOUBLE) dcbtst Y1, PREA#endif LFDU y05, 1 * SIZE(Y1) LFDU y06, 1 * SIZE(Y1)#ifdef PPCG4 dcbt AO1, PREA#endif FMADD y13, alpha2, a5, y13 LFDU a5, 1 * SIZE(AO1) FMADD y14, alpha2, a6, y14 LFDU a6, 1 * SIZE(AO1) FMADD y15, alpha2, a7, y15 LFDU a7, 1 * SIZE(AO1) FMADD y16, alpha2, a8, y16 LFDU a8, 1 * SIZE(AO1) LFDU y07, 1 * SIZE(Y1) LFDU y08, 1 * SIZE(Y1)#if defined(PPCG4) && defined(DOUBLE) dcbt AO1, PREA#endif STFDU y09, 1 * SIZE(Y2) STFDU y10, 1 * SIZE(Y2) STFDU y11, 1 * SIZE(Y2) STFDU y12, 1 * SIZE(Y2) STFDU y13, 1 * SIZE(Y2) STFDU y14, 1 * SIZE(Y2) STFDU y15, 1 * SIZE(Y2) STFDU y16, 1 * SIZE(Y2) bdnz LL(32) .align 4LL(33): FMADD y01, alpha1, a1, y01 LFDU a1, 1 * SIZE(AO2) FMADD y02, alpha1, a2, y02 LFDU a2, 1 * SIZE(AO2) FMADD y03, alpha1, a3, y03 LFDU a3, 1 * SIZE(AO2) FMADD y04, alpha1, a4, y04 LFDU a4, 1 * SIZE(AO2) FMADD y05, alpha1, a5, y05 LFDU a5, 1 * SIZE(AO2) FMADD y06, alpha1, a6, y06 LFDU a6, 1 * SIZE(AO2) FMADD y07, alpha1, a7, y07 LFDU a7, 1 * SIZE(AO2) FMADD y08, alpha1, a8, y08 LFDU a8, 1 * SIZE(AO2) FMADD y01, alpha2, a1, y01 FMADD y02, alpha2, a2, y02 FMADD y03, alpha2, a3, y03 FMADD y04, alpha2, a4, y04 FMADD y05, alpha2, a5, y05 STFDU y01, 1 * SIZE(Y2) FMADD y06, alpha2, a6, y06 STFDU y02, 1 * SIZE(Y2) FMADD y07, alpha2, a7, y07 STFDU y03, 1 * SIZE(Y2) FMADD y08, alpha2, a8, y08 STFDU y04, 1 * SIZE(Y2) STFDU y05, 1 * SIZE(Y2) STFDU y06, 1 * SIZE(Y2) STFDU y07, 1 * SIZE(Y2) STFDU y08, 1 * SIZE(Y2) .align 4LL(35): andi. r0, M, 7 ble LL(40) andi. r0, M, 4 ble LL(37) LFDU a1, 1 * SIZE(AO1) LFDU y01, 1 * SIZE(Y1) LFDU a2, 1 * SIZE(AO1) LFDU y02, 1 * SIZE(Y1) LFDU a3, 1 * SIZE(AO1) LFDU y03, 1 * SIZE(Y1) LFDU a4, 1 * SIZE(AO1) LFDU y04, 1 * SIZE(Y1) FMADD y01, alpha1, a1, y01 LFDU a5, 1 * SIZE(AO2) FMADD y02, alpha1, a2, y02 LFDU a6, 1 * SIZE(AO2) FMADD y03, alpha1, a3, y03 LFDU a7, 1 * SIZE(AO2) FMADD y04, alpha1, a4, y04 LFDU a8, 1 * SIZE(AO2) FMADD y01, alpha2, a5, y01 FMADD y02, alpha2, a6, y02 FMADD y03, alpha2, a7, y03 FMADD y04, alpha2, a8, y04 STFDU y01, 1 * SIZE(Y2) STFDU y02, 1 * SIZE(Y2) STFDU y03, 1 * SIZE(Y2) STFDU y04, 1 * SIZE(Y2) .align 4LL(37): andi. r0, M, 2 ble LL(38) LFDU a1, 1 * SIZE(AO1) LFDU y01, 1 * SIZE(Y1) LFDU a2, 1 * SIZE(AO1) LFDU y02, 1 * SIZE(Y1) LFDU a3, 1 * SIZE(AO2) LFDU a4, 1 * SIZE(AO2) FMADD y01, alpha1, a1, y01 FMADD y02, alpha1, a2, y02 FMADD y01, alpha2, a3, y01 FMADD y02, alpha2, a4, y02 STFDU y01, 1 * SIZE(Y2) STFDU y02, 1 * SIZE(Y2) .align 4LL(38): andi. r0, M, 1 ble LL(40) LFDU a1, 1 * SIZE(AO1) LFDU y01, 1 * SIZE(Y1) LFDU a2, 1 * SIZE(AO2) FMADD y01, alpha1, a1, y01 FMADD y01, alpha2, a2, y01 STFDU y01, 1 * SIZE(Y2) .align 4LL(40): andi. J, N, 1 lfd alpha, ALPHA ble LL(990) LFDUX alpha1, X, INCX mr AO1, A add A, A, LDA FMUL alpha1, alpha, alpha1 mr Y1, YY mr Y2, YY srawi. r0, M, 3 mtspr CTR, r0 ble LL(45) LFDU y01, 1 * SIZE(Y1) LFDU a1, 1 * SIZE(AO1) LFDU y02, 1 * SIZE(Y1) LFDU a2, 1 * SIZE(AO1) LFDU y03, 1 * SIZE(Y1) LFDU a3, 1 * SIZE(AO1) LFDU y04, 1 * SIZE(Y1) LFDU a4, 1 * SIZE(AO1) LFDU y05, 1 * SIZE(Y1) LFDU a5, 1 * SIZE(AO1) LFDU y06, 1 * SIZE(Y1) LFDU a6, 1 * SIZE(AO1) LFDU y07, 1 * SIZE(Y1) LFDU a7, 1 * SIZE(AO1) LFDU y08, 1 * SIZE(Y1) LFDU a8, 1 * SIZE(AO1) bdz LL(43) .align 4LL(42):#ifdef PPCG4 dcbtst Y1, PREA#endif FMADD y09, alpha1, a1, y01 LFDU a1, 1 * SIZE(AO1) FMADD y10, alpha1, a2, y02 LFDU a2, 1 * SIZE(AO1) FMADD y11, alpha1, a3, y03 LFDU a3, 1 * SIZE(AO1) FMADD y12, alpha1, a4, y04 LFDU a4, 1 * SIZE(AO1) LFDU y01, 1 * SIZE(Y1) LFDU y02, 1 * SIZE(Y1) LFDU y03, 1 * SIZE(Y1) LFDU y04, 1 * SIZE(Y1)#ifdef PPCG4 dcbt AO1, PREA#endif FMADD y13, alpha1, a5, y05 LFDU a5, 1 * SIZE(AO1) FMADD y14, alpha1, a6, y06 LFDU a6, 1 * SIZE(AO1) FMADD y15, alpha1, a7, y07 LFDU a7, 1 * SIZE(AO1) FMADD y16, alpha1, a8, y08 LFDU a8, 1 * SIZE(AO1)#if defined(PPCG4) && defined(DOUBLE) dcbtst Y1, PREA#endif LFDU y05, 1 * SIZE(Y1) LFDU y06, 1 * SIZE(Y1) LFDU y07, 1 * SIZE(Y1) LFDU y08, 1 * SIZE(Y1)#if defined(PPCG4) && defined(DOUBLE) dcbt AO1, PREA#endif STFDU y09, 1 * SIZE(Y2) STFDU y10, 1 * SIZE(Y2) STFDU y11, 1 * SIZE(Y2) STFDU y12, 1 * SIZE(Y2) STFDU y13, 1 * SIZE(Y2) STFDU y14, 1 * SIZE(Y2) STFDU y15, 1 * SIZE(Y2) STFDU y16, 1 * SIZE(Y2) bdnz LL(42) .align 4LL(43): FMADD y01, alpha1, a1, y01 FMADD y02, alpha1, a2, y02 FMADD y03, alpha1, a3, y03 FMADD y04, alpha1, a4, y04 FMADD y05, alpha1, a5, y05 STFDU y01, 1 * SIZE(Y2) FMADD y06, alpha1, a6, y06 STFDU y02, 1 * SIZE(Y2) FMADD y07, alpha1, a7, y07 STFDU y03, 1 * SIZE(Y2) FMADD y08, alpha1, a8, y08 STFDU y04, 1 * SIZE(Y2) STFDU y05, 1 * SIZE(Y2) STFDU y06, 1 * SIZE(Y2) STFDU y07, 1 * SIZE(Y2) STFDU y08, 1 * SIZE(Y2) .align 4LL(45): andi. r0, M, 7 ble LL(990) andi. r0, M, 4 ble LL(47) LFDU a1, 1 * SIZE(AO1) LFDU y01, 1 * SIZE(Y1) LFDU a2, 1 * SIZE(AO1) LFDU y02, 1 * SIZE(Y1) LFDU a3, 1 * SIZE(AO1) LFDU y03, 1 * SIZE(Y1) LFDU a4, 1 * SIZE(AO1) LFDU y04, 1 * SIZE(Y1) FMADD y01, alpha1, a1, y01 FMADD y02, alpha1, a2, y02 FMADD y03, alpha1, a3, y03 FMADD y04, alpha1, a4, y04 STFDU y01, 1 * SIZE(Y2) STFDU y02, 1 * SIZE(Y2) STFDU y03, 1 * SIZE(Y2) STFDU y04, 1 * SIZE(Y2) .align 4LL(47): andi. r0, M, 2 ble LL(48) LFDU a1, 1 * SIZE(AO1) LFDU y01, 1 * SIZE(Y1) LFDU a2, 1 * SIZE(AO1) LFDU y02, 1 * SIZE(Y1) FMADD y01, alpha1, a1, y01 FMADD y02, alpha1, a2, y02 STFDU y01, 1 * SIZE(Y2) STFDU y02, 1 * SIZE(Y2) .align 4LL(48): andi. r0, M, 1 ble LL(990) LFDU a1, 1 * SIZE(AO1) LFDU y01, 1 * SIZE(Y1) FMADD y01, alpha1, a1, y01 STFDU y01, 1 * SIZE(Y2) .align 4LL(990): cmpi cr0, 0, INCY, SIZE beq LL(999) addi YY, BUFFER, -SIZE mr Y1, Y srawi. r0, M, 3 mtspr CTR, r0 ble LL(995) .align 4LL(991): LFDUX f0, Y, INCY LFDUX f1, Y, INCY LFDUX f2, Y, INCY LFDUX f3, Y, INCY LFDUX f4, Y, INCY LFDUX f5, Y, INCY LFDUX f6, Y, INCY LFDUX f7, Y, INCY LFDU f8, 1 * SIZE(YY) LFDU f9, 1 * SIZE(YY) LFDU f10, 1 * SIZE(YY) LFDU f11, 1 * SIZE(YY) LFDU f12, 1 * SIZE(YY) LFDU f13, 1 * SIZE(YY) LFDU f14, 1 * SIZE(YY) LFDU f15, 1 * SIZE(YY) FADD f8, f8, f0 FADD f9, f9, f1 FADD f10, f10, f2 FADD f11, f11, f3 FADD f12, f12, f4 FADD f13, f13, f5 FADD f14, f14, f6 FADD f15, f15, f7 STFDUX f8, Y1, INCY STFDUX f9, Y1, INCY STFDUX f10, Y1, INCY STFDUX f11, Y1, INCY STFDUX f12, Y1, INCY STFDUX f13, Y1, INCY STFDUX f14, Y1, INCY STFDUX f15, Y1, INCY bdnz LL(991) .align 4LL(995): andi. J, M, 4 ble LL(996) LFDUX f0, Y, INCY LFDUX f1, Y, INCY LFDUX f2, Y, INCY LFDUX f3, Y, INCY LFDU f8, 1 * SIZE(YY) LFDU f9, 1 * SIZE(YY) LFDU f10, 1 * SIZE(YY) LFDU f11, 1 * SIZE(YY) FADD f8, f8, f0 FADD f9, f9, f1 FADD f10, f10, f2 FADD f11, f11, f3 STFDUX f8, Y1, INCY STFDUX f9, Y1, INCY STFDUX f10, Y1, INCY STFDUX f11, Y1, INCY .align 4LL(996): andi. J, M, 2 ble LL(997) LFDUX f0, Y, INCY LFDUX f1, Y, INCY LFDU f8, 1 * SIZE(YY) LFDU f9, 1 * SIZE(YY) FADD f8, f8, f0 FADD f9, f9, f1 STFDUX f8, Y1, INCY STFDUX f9, Y1, INCY .align 4LL(997): andi. J, M, 1 ble LL(999) LFDUX f0, Y, INCY LFDU f8, 1 * SIZE(YY) FADD f8, f8, f0 STFDUX f8, Y1, INCY .align 4LL(999): 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)#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)#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)#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 + -