📄 znrm2.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 PREA r8#define INCXM1 r9#define FZERO 144(SP)#define FONE 148(SP)#define FMAX 152(SP)#define STACKSIZE 160 PROLOGUE PROFCODE addi SP, SP, -STACKSIZE li r10, 0 lis r11, 0x3f80 lis r12, 0x5fe0 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 lfs f1, FZERO#ifdef F_INTERFACE LDINT N, 0(N) LDINT INCX, 0(INCX)#endif slwi INCX, INCX, ZBASE_SHIFT subi INCXM1, INCX, SIZE li PREA, 4 * 16 * SIZE cmpwi cr0, N, 0 ble- LL(9999) cmpwi cr0, INCX, 0 ble- LL(9999) mr NN, N mr XX, X LFD f0, 0 * SIZE(X) LFD f1, 1 * SIZE(X) add X, X, INCX 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 cmpwi cr0, INCX, 2 * SIZE bne- cr0, LL(1000) srawi. r0, N, 3 mtspr CTR, r0 beq- cr0, LL(50) LFD f24, 0 * SIZE(X) LFD f25, 1 * SIZE(X) LFD f26, 2 * SIZE(X) LFD f27, 3 * SIZE(X) LFD f28, 4 * SIZE(X) LFD f29, 5 * SIZE(X) LFD f30, 6 * SIZE(X) LFD f31, 7 * SIZE(X) fabs f8, f24 fabs f9, f25 fabs f10, f26 fabs f11, f27 LFD f24, 8 * SIZE(X) LFD f25, 9 * SIZE(X) LFD f26, 10 * SIZE(X) LFD f27, 11 * SIZE(X) fabs f12, f28 fabs f13, f29 fabs f14, f30 fabs f15, f31 LFD f28, 12 * SIZE(X) LFD f29, 13 * SIZE(X) LFD f30, 14 * SIZE(X) LFD f31, 15 * SIZE(X) 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 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 LFD f24, 16 * SIZE(X) LFD f25, 17 * SIZE(X) LFD f26, 18 * SIZE(X) LFD f27, 19 * SIZE(X) 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 LFD f28, 20 * SIZE(X) LFD f29, 21 * SIZE(X) LFD f30, 22 * SIZE(X) LFD f31, 23 * SIZE(X) 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 LFD f24, 24 * SIZE(X) LFD f25, 25 * SIZE(X) LFD f26, 26 * SIZE(X) LFD f27, 27 * SIZE(X) 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 LFD f28, 28 * SIZE(X) LFD f29, 29 * SIZE(X) LFD f30, 30 * SIZE(X) LFD f31, 31 * SIZE(X) dcbt X, PREA addi X, X, 16 * SIZE 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 addi X, X, 16 * SIZE .align 4LL(50): andi. r0, N, 7 mtspr CTR, r0 beq LL(100) .align 4LL(60): LFD f8, 0 * SIZE(X) LFD f9, 1 * SIZE(X) addi X, X, 2 * SIZE 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(100): 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(9999) 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) LFD f8, 0 * SIZE(XX) LFD f9, 1 * SIZE(XX) LFD f10, 2 * SIZE(XX) LFD f11, 3 * SIZE(XX) LFD f12, 4 * SIZE(XX) LFD f13, 5 * SIZE(XX) LFD f14, 6 * SIZE(XX) LFD f15, 7 * SIZE(XX) fmul f16, f30, f8 fmul f17, f30, f9 fmul f18, f30, f10 fmul f19, f30, f11 LFD f8, 8 * SIZE(XX) LFD f9, 9 * SIZE(XX) LFD f10, 10 * SIZE(XX) LFD f11, 11 * SIZE(XX) fmul f20, f30, f12 fmul f21, f30, f13 fmul f22, f30, f14 fmul f23, f30, f15 LFD f12, 12 * SIZE(XX) LFD f13, 13 * SIZE(XX) LFD f14, 14 * SIZE(XX) LFD f15, 15 * SIZE(XX) bdz LL(120) .align 4LL(110): 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 LFD f8, 16 * SIZE(XX) LFD f9, 17 * SIZE(XX) LFD f10, 18 * SIZE(XX) LFD f11, 19 * SIZE(XX) 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 LFD f12, 20 * SIZE(XX) LFD f13, 21 * SIZE(XX) LFD f14, 22 * SIZE(XX) LFD f15, 23 * SIZE(XX) 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 LFD f8, 24 * SIZE(XX) LFD f9, 25 * SIZE(XX) LFD f10, 26 * SIZE(XX) LFD f11, 27 * SIZE(XX) 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 LFD f12, 28 * SIZE(XX) LFD f13, 29 * SIZE(XX) LFD f14, 30 * SIZE(XX) LFD f15, 31 * SIZE(XX) dcbt X, PREA addi XX, XX, 16 * SIZE 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
⌨️ 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 + -