📄 znrm2_ppc440.s
字号:
/*********************************************************************//* *//* Optimized BLAS libraries *//* By Kazushige Goto <kgoto@tacc.utexas.edu> *//* *//* Copyright (c) The University of Texas, 2005. All rights reserved. *//* UNIVERSITY EXPRESSLY DISCLAIMS ANY AND ALL WARRANTIES CONCERNING *//* THIS SOFTWARE AND DOCUMENTATION, INCLUDING ANY WARRANTIES OF *//* MERCHANTABILITY, FITNESS FOR ANY PARTICULAR PURPOSE, *//* NON-INFRINGEMENT AND WARRANTIES OF PERFORMANCE, AND ANY WARRANTY *//* THAT MIGHT OTHERWISE ARISE FROM COURSE OF DEALING OR USAGE OF *//* TRADE. NO WARRANTY IS EITHER EXPRESS OR IMPLIED WITH RESPECT TO *//* THE USE OF THE SOFTWARE OR DOCUMENTATION. *//* Under no circumstances shall University be liable for incidental, *//* special, indirect, direct or consequential damages or loss of *//* profits, interruption of business, or related expenses which may *//* arise from use of Software or Documentation, including but not *//* limited to those resulting from defects in Software and/or *//* Documentation, or loss or inaccuracy of data of any kind. *//*********************************************************************/#define ASSEMBLER#include "common.h" #define N r3#define X r4#define INCX r5 #define NN r6#define XX r7#define INC1 r9#define PRE r10#define FZERO 144(SP)#define FONE 148(SP)#define FMAX 152(SP)#define C1 156(SP)#define C2 160(SP)#define STACKSIZE 168 PROLOGUE PROFCODE addi SP, SP, -STACKSIZE li r10, 0 lis r11, 0x3f80 lis r12, 0x5fe0 lis r6, 0x3f00 lis r7, 0x4040 stfd f14, 0(SP) stfd f15, 8(SP) stfd f16, 16(SP) stfd f17, 24(SP) stfd f18, 32(SP) stfd f19, 40(SP) stfd f20, 48(SP) stfd f21, 56(SP) stfd f22, 64(SP) stfd f23, 72(SP) stfd f24, 80(SP) stfd f25, 88(SP) stfd f26, 96(SP) stfd f27, 104(SP) stfd f28, 112(SP) stfd f29, 120(SP) stfd f30, 128(SP) stfd f31, 136(SP) stw r10, FZERO stw r11, FONE stw r12, FMAX stw r10, 4 + FMAX stw r6, C1 stw r7, C2 lfs f1, FZERO#ifdef F_INTERFACE LDINT N, 0(N) LDINT INCX, 0(INCX)#endif slwi INCX, INCX, ZBASE_SHIFT sub X, X, INCX li INC1, SIZE li PRE, 3 * 16 * SIZE cmpwi cr0, N, 0 ble- LL(999) cmpwi cr0, INCX, 0 ble- LL(999) mr NN, N mr XX, X LFDUX f0, X, INCX LFDX f1, X, INC1 fabs f2, f0 fabs f3, f1 fabs f4, f0 fabs f5, f1 fabs f6, f0 fabs f7, f1 fabs f0, f0 fabs f1, f1 subi N, N, 1 srawi. r0, N, 3 mtspr CTR, r0 beq- LL(50) LFDUX f24, X, INCX LFDX f25, X, INC1 LFDUX f26, X, INCX LFDX f27, X, INC1 LFDUX f28, X, INCX LFDX f29, X, INC1 LFDUX f30, X, INCX LFDX f31, X, INC1 fabs f8, f24 LFDUX f24, X, INCX fabs f9, f25 LFDX f25, X, INC1 fabs f10, f26 LFDUX f26, X, INCX fabs f11, f27 LFDX f27, X, INC1 fabs f12, f28 LFDUX f28, X, INCX fabs f13, f29 LFDX f29, X, INC1 fabs f14, f30 LFDUX f30, X, INCX fabs f15, f31 LFDX f31, X, INC1 bdz LL(20) .align 4LL(10): fsub f16, f0, f8 fsub f17, f1, f9 fsub f18, f2, f10 fsub f19, f3, f11 fsub f20, f4, f12 fsub f21, f5, f13 fsub f22, f6, f14 fsub f23, f7, f15 fsel f0, f16, f0, f8#ifdef PPCG4 dcbt X, PRE#endif fabs f8, f24 LFDUX f24, X, INCX fsel f1, f17, f1, f9 fabs f9, f25 LFDX f25, X, INC1 fsel f2, f18, f2, f10 fabs f10, f26 LFDUX f26, X, INCX fsel f3, f19, f3, f11 fabs f11, f27 LFDX f27, X, INC1 fsel f4, f20, f4, f12#ifdef PPCG4 dcbt X, PRE#endif fabs f12, f28 LFDUX f28, X, INCX fsel f5, f21, f5, f13 fabs f13, f29 LFDX f29, X, INC1 fsel f6, f22, f6, f14 fabs f14, f30 LFDUX f30, X, INCX fsel f7, f23, f7, f15 fabs f15, f31 LFDX f31, X, INC1 fsub f16, f0, f8 fsub f17, f1, f9 fsub f18, f2, f10 fsub f19, f3, f11 fsub f20, f4, f12 fsub f21, f5, f13 fsub f22, f6, f14 fsub f23, f7, f15 fsel f0, f16, f0, f8#ifdef PPCG4 dcbt X, PRE#endif fabs f8, f24 LFDUX f24, X, INCX fsel f1, f17, f1, f9 fabs f9, f25 LFDX f25, X, INC1 fsel f2, f18, f2, f10 fabs f10, f26 LFDUX f26, X, INCX fsel f3, f19, f3, f11 fabs f11, f27 LFDX f27, X, INC1 fsel f4, f20, f4, f12#ifdef PPCG4 dcbt X, PRE#endif fabs f12, f28 LFDUX f28, X, INCX fsel f5, f21, f5, f13 fabs f13, f29 LFDX f29, X, INC1 fsel f6, f22, f6, f14 fabs f14, f30 LFDUX f30, X, INCX fsel f7, f23, f7, f15 fabs f15, f31 LFDX f31, X, INC1 bdnz LL(10) .align 4LL(20): fsub f16, f0, f8 fsub f17, f1, f9 fsub f18, f2, f10 fsub f19, f3, f11 fsub f20, f4, f12 fsub f21, f5, f13 fsub f22, f6, f14 fsub f23, f7, f15 fsel f0, f16, f0, f8 fabs f8, f24 fsel f1, f17, f1, f9 fabs f9, f25 fsel f2, f18, f2, f10 fabs f10, f26 fsel f3, f19, f3, f11 fabs f11, f27 fsel f4, f20, f4, f12 fabs f12, f28 fsel f5, f21, f5, f13 fabs f13, f29 fsel f6, f22, f6, f14 fabs f14, f30 fsel f7, f23, f7, f15 fabs f15, f31 fsub f16, f0, f8 fsub f17, f1, f9 fsub f18, f2, f10 fsub f19, f3, f11 fsub f20, f4, f12 fsub f21, f5, f13 fsub f22, f6, f14 fsub f23, f7, f15 fsel f0, f16, f0, f8 fsel f1, f17, f1, f9 fsel f2, f18, f2, f10 fsel f3, f19, f3, f11 fsel f4, f20, f4, f12 fsel f5, f21, f5, f13 fsel f6, f22, f6, f14 fsel f7, f23, f7, f15 .align 4LL(50): andi. r0, N, 7 mtspr CTR, r0 beq LL(99) .align 4LL(60): LFDUX f8, X, INCX LFDX f9, X, INC1 fabs f8, f8 fabs f9, f9 fsub f16, f0, f8 fsub f17, f1, f9 fsel f0, f16, f0, f8 fsel f1, f17, f1, f9 bdnz LL(60) .align 4LL(99): fsub f8, f0, f1 fsub f9, f2, f3 fsub f10, f4, f5 fsub f11, f6, f7 fsel f0, f8, f0, f1 fsel f2, f9, f2, f3 fsel f4, f10, f4, f5 fsel f6, f11, f6, f7 fsub f8, f0, f2 fsub f9, f4, f6 fsel f0, f8, f0, f2 fsel f4, f9, f4, f6 fsub f8, f0, f4 fsel f31, f8, f0, f4 lfs f1, FZERO lfs f0, FONE fcmpu cr0, f1, f31 beq- cr0, LL(999) fdiv f30, f0, f31 fmr f0, f1 fmr f2, f1 fmr f3, f1 fmr f4, f1 fmr f5, f1 fmr f6, f1 fmr f7, f1 srawi. r0, NN, 3 mtspr CTR, r0 beq- cr0, LL(150) LFDUX f8, XX, INCX LFDX f9, XX, INC1 LFDUX f10, XX, INCX LFDX f11, XX, INC1 LFDUX f12, XX, INCX LFDX f13, XX, INC1 LFDUX f14, XX, INCX LFDX f15, XX, INC1 fmul f16, f30, f8 LFDUX f8, XX, INCX fmul f17, f30, f9 LFDX f9, XX, INC1 fmul f18, f30, f10 LFDUX f10, XX, INCX fmul f19, f30, f11 LFDX f11, XX, INC1 fmul f20, f30, f12 LFDUX f12, XX, INCX fmul f21, f30, f13 LFDX f13, XX, INC1 fmul f22, f30, f14 LFDUX f14, XX, INCX fmul f23, f30, f15 LFDX f15, XX, INC1 bdz LL(120) .align 4LL(110): fmadd f0, f16, f16, f0#ifdef PPCG4 dcbt XX, PRE#endif fmul f16, f30, f8 LFDUX f8, XX, INCX fmadd f1, f17, f17, f1 fmul f17, f30, f9 LFDX f9, XX, INC1 fmadd f2, f18, f18, f2 fmul f18, f30, f10 LFDUX f10, XX, INCX fmadd f3, f19, f19, f3 fmul f19, f30, f11 LFDX f11, XX, INC1 fmadd f4, f20, f20, f4#ifdef PPCG4 dcbt XX, PRE#endif fmul f20, f30, f12 LFDUX f12, XX, INCX fmadd f5, f21, f21, f5 fmul f21, f30, f13 LFDX f13, XX, INC1 fmadd f6, f22, f22, f6 fmul f22, f30, f14 LFDUX f14, XX, INCX fmadd f7, f23, f23, f7 fmul f23, f30, f15 LFDX f15, XX, INC1 fmadd f0, f16, f16, f0#ifdef PPCG4 dcbt XX, PRE#endif fmul f16, f30, f8 LFDUX f8, XX, INCX fmadd f1, f17, f17, f1 fmul f17, f30, f9 LFDX f9, XX, INC1 fmadd f2, f18, f18, f2 fmul f18, f30, f10 LFDUX f10, XX, INCX fmadd f3, f19, f19, f3 fmul f19, f30, f11 LFDX f11, XX, INC1 fmadd f4, f20, f20, f4#ifdef PPCG4 dcbt XX, PRE#endif fmul f20, f30, f12 LFDUX f12, XX, INCX fmadd f5, f21, f21, f5 fmul f21, f30, f13 LFDX f13, XX, INC1 fmadd f6, f22, f22, f6 fmul f22, f30, f14 LFDUX f14, XX, INCX fmadd f7, f23, f23, f7 fmul f23, f30, f15 LFDX f15, XX, INC1 bdnz LL(110) .align 4LL(120): fmadd f0, f16, f16, f0 fmul f16, f30, f8 fmadd f1, f17, f17, f1 fmul f17, f30, f9 fmadd f2, f18, f18, f2 fmul f18, f30, f10 fmadd f3, f19, f19, f3 fmul f19, f30, f11 fmadd f4, f20, f20, f4 fmul f20, f30, f12 fmadd f5, f21, f21, f5 fmul f21, f30, f13 fmadd f6, f22, f22, f6 fmul f22, f30, f14 fmadd f7, f23, f23, f7 fmul f23, f30, f15 fmadd f0, f16, f16, f0 fmadd f1, f17, f17, f1 fmadd f2, f18, f18, f2 fmadd f3, f19, f19, f3 fmadd f4, f20, f20, f4 fmadd f5, f21, f21, f5 fmadd f6, f22, f22, f6 fmadd f7, f23, f23, f7 .align 4LL(150): andi. r0, NN, 7 mtspr CTR, r0 beq- cr0, LL(170) .align 4LL(160): LFDUX f8, XX, INCX LFDX f9, XX, INC1 fmul f16, f30, f8 fmul f17, f30, f9 fmadd f0, f16, f16, f0 fmadd f1, f17, f17, f1 bdnz LL(160) .align 4LL(170): fadd f0, f0, f1 fadd f2, f2, f3 fadd f4, f4, f5 fadd f6, f6, f7 fadd f0, f0, f2 fadd f4, f4, f6 fadd f1, f0, f4 frsqrte f0, f1 lfs f8, C1 lfs f9, C2 fmul f2, f1, f0 fadd f7, f8, f8 fmul f3, f0, f8 fnmsub f4, f2, f0, f9 fmul f0, f3, f4 fmul f2, f1, f0 fmul f3, f0, f8 fnmsub f4, f2, f0, f9 fmul f0, f3, f4 fmul f2, f1, f0 fmul f3, f0, f8 fnmsub f4, f2, f0, f9 fmul f0, f3, f4 fmul f5, f1, f0 fmul f2, f5, f8 fnmsub f3, f5, f0, f7 fmadd f1, f2, f3, f5 fmul f1, f31, f1 .align 4LL(999): 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) addi SP, SP, STACKSIZE blr EPILOGUE
⌨️ 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 + -