📄 zamax_cell.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 PREA r8#define INCXM1 r9#define FZERO f1#define STACKSIZE 160 PROLOGUE PROFCODE addi SP, SP, -STACKSIZE li r0, 0 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 r0, 144(SP) lfs FZERO,144(SP)#ifdef F_INTERFACE LDINT N, 0(N) LDINT INCX, 0(INCX)#endif slwi INCX, INCX, ZBASE_SHIFT subi INCXM1, INCX, SIZE li PREA, 10 * 16 * SIZE cmpwi cr0, N, 0 ble- LL(9999) cmpwi cr0, INCX, 0 ble- LL(9999) LFD f1, 0 * SIZE(X) LFD f2, 1 * SIZE(X) add X, X, INCX fabs f1, f1 fabs f2, f2 fadd f1, f1, f2 fmr f0, f1 fmr f2, f1 fmr f3, f1 subi N, N, 1 cmpwi cr0, INCX, 2 * SIZE bne- cr0, LL(100) srawi. r0, N, 3 mtspr CTR, r0 beq- cr0, LL(50) .align 4 LFD f24, 0 * SIZE(X) LFD f25, 1 * SIZE(X) fabs f8, f24 LFD f26, 2 * SIZE(X) fabs f9, f25 LFD f27, 3 * SIZE(X) fabs f10, f26 LFD f28, 4 * SIZE(X) fabs f11, f27 LFD f29, 5 * SIZE(X) fabs f12, f28 LFD f30, 6 * SIZE(X) fabs f13, f29 LFD f31, 7 * SIZE(X) fabs f14, f30 nop fabs f15, f31 bdz LL(20) .align 4LL(10): fadd f4, f8, f9 dcbt X, PREA fadd f5, f10, f11 nop fadd f6, f12, f13 LFD f24, 8 * SIZE(X) fadd f7, f14, f15 LFD f25, 9 * SIZE(X) fabs f8, f24 LFD f26, 10 * SIZE(X) fabs f9, f25 LFD f27, 11 * SIZE(X) fabs f10, f26 fabs f11, f27 fsub f16, f0, f4 fsub f17, f1, f5 fsub f18, f2, f6 LFD f28, 12 * SIZE(X) fsub f19, f3, f7 LFD f29, 13 * SIZE(X) fabs f12, f28 LFD f30, 14 * SIZE(X) fabs f13, f29 LFD f31, 15 * SIZE(X) fabs f14, f30 fabs f15, f31 fsel f0, f16, f0, f4 fsel f1, f17, f1, f5 fsel f2, f18, f2, f6 fsel f3, f19, f3, f7 fadd f20, f8, f9 fadd f21, f10, f11 fadd f22, f12, f13 LFD f24, 16 * SIZE(X) fadd f23, f14, f15 LFD f25, 17 * SIZE(X) fabs f8, f24 LFD f26, 18 * SIZE(X) fabs f9, f25 LFD f27, 19 * SIZE(X) fabs f10, f26 fabs f11, f27 fsub f16, f0, f20 fsub f17, f1, f21 fsub f18, f2, f22 LFD f28, 20 * SIZE(X) fsub f19, f3, f23 LFD f29, 21 * SIZE(X) fabs f12, f28 LFD f30, 22 * SIZE(X) fabs f13, f29 LFD f31, 23 * SIZE(X) fabs f14, f30 addi X, X, 16 * SIZE fabs f15, f31 fsel f0, f16, f0, f20 fsel f1, f17, f1, f21 fsel f2, f18, f2, f22 fsel f3, f19, f3, f23 bdnz LL(10) .align 4LL(20): fadd f4, f8, f9 fadd f5, f10, f11 fadd f6, f12, f13 LFD f24, 8 * SIZE(X) fadd f7, f14, f15 LFD f25, 9 * SIZE(X) fabs f8, f24 LFD f26, 10 * SIZE(X) fabs f9, f25 LFD f27, 11 * SIZE(X) fabs f10, f26 fabs f11, f27 fsub f16, f0, f4 fsub f17, f1, f5 fsub f18, f2, f6 LFD f28, 12 * SIZE(X) fsub f19, f3, f7 LFD f29, 13 * SIZE(X) fabs f12, f28 LFD f30, 14 * SIZE(X) fabs f13, f29 LFD f31, 15 * SIZE(X) fabs f14, f30 fabs f15, f31 fsel f0, f16, f0, f4 fsel f1, f17, f1, f5 fsel f2, f18, f2, f6 fsel f3, f19, f3, f7 fadd f20, f8, f9 fadd f21, f10, f11 fadd f22, f12, f13 fadd f23, f14, f15 fsub f16, f0, f20 fsub f17, f1, f21 fsub f18, f2, f22 fsub f19, f3, f23 fsel f0, f16, f0, f20 fsel f1, f17, f1, f21 fsel f2, f18, f2, f22 fsel f3, f19, f3, f23 addi X, X, 16 * SIZE .align 4LL(50): andi. r0, N, 7 mtspr CTR, r0 beq LL(999) .align 4LL(60): LFD f8, 0 * SIZE(X) LFD f9, 1 * SIZE(X) addi X, X, 2 * SIZE fabs f8, f8 fabs f9, f9 fadd f8, f8, f9 fsub f16, f1, f8 fsel f1, f16, f1, f8 bdnz LL(60) b LL(999) .align 4LL(100): sub X, X, INCXM1 srawi. r0, N, 3 mtspr CTR, r0 beq- LL(150) LFDX f24, X, INCXM1 LFDUX f25, X, INCX LFDX f26, X, INCXM1 LFDUX f27, X, INCX LFDX f28, X, INCXM1 LFDUX f29, X, INCX LFDX f30, X, INCXM1 LFDUX f31, X, INCX fabs f8, f24 fabs f9, f25 fabs f10, f26 fabs f11, f27 fabs f12, f28 fabs f13, f29 fabs f14, f30 fabs f15, f31 LFDX f24, X, INCXM1 LFDUX f25, X, INCX LFDX f26, X, INCXM1 LFDUX f27, X, INCX LFDX f28, X, INCXM1 LFDUX f29, X, INCX LFDX f30, X, INCXM1 LFDUX f31, X, INCX bdz LL(120) .align 4LL(110): fadd f4, f8, f9 fadd f5, f10, f11 fadd f6, f12, f13 fadd f7, f14, f15 fabs f8, f24 fabs f9, f25 fabs f10, f26 fabs f11, f27 LFDX f24, X, INCXM1 LFDUX f25, X, INCX LFDX f26, X, INCXM1 LFDUX f27, X, INCX fabs f12, f28 fabs f13, f29 fabs f14, f30 fabs f15, f31 LFDX f28, X, INCXM1 LFDUX f29, X, INCX LFDX f30, X, INCXM1 LFDUX f31, X, INCX fsub f16, f0, f4 fsub f17, f1, f5 fsub f18, f2, f6 fsub f19, f3, f7 fadd f20, f8, f9 fadd f21, f10, f11 fadd f22, f12, f13 fadd f23, f14, f15 fabs f8, f24 fabs f9, f25 fabs f10, f26 fabs f11, f27 LFDX f24, X, INCXM1 LFDUX f25, X, INCX LFDX f26, X, INCXM1 LFDUX f27, X, INCX fsel f0, f16, f0, f4 fsel f1, f17, f1, f5 fsel f2, f18, f2, f6 fsel f3, f19, f3, f7 fabs f12, f28 fabs f13, f29 fabs f14, f30 fabs f15, f31 LFDX f28, X, INCXM1 LFDUX f29, X, INCX LFDX f30, X, INCXM1 LFDUX f31, X, INCX fsub f16, f0, f20 fsub f17, f1, f21 fsub f18, f2, f22 fsub f19, f3, f23 fsel f0, f16, f0, f20 fsel f1, f17, f1, f21 fsel f2, f18, f2, f22 fsel f3, f19, f3, f23 bdnz LL(110) .align 4LL(120): fadd f4, f8, f9 fadd f5, f10, f11 fadd f6, f12, f13 fadd f7, f14, f15 fabs f8, f24 fabs f9, f25 fabs f10, f26 fabs f11, f27 fabs f12, f28 fabs f13, f29 fabs f14, f30 fabs f15, f31 fsub f16, f0, f4 fsub f17, f1, f5 fsub f18, f2, f6 fsub f19, f3, f7 fadd f20, f8, f9 fadd f21, f10, f11 fadd f22, f12, f13 fadd f23, f14, f15 fsel f0, f16, f0, f4 fsel f1, f17, f1, f5 fsel f2, f18, f2, f6 fsel f3, f19, f3, f7 fsub f16, f0, f20 fsub f17, f1, f21 fsub f18, f2, f22 fsub f19, f3, f23 fsel f0, f16, f0, f20 fsel f1, f17, f1, f21 fsel f2, f18, f2, f22 fsel f3, f19, f3, f23 .align 4LL(150): andi. r0, N, 7 mtspr CTR, r0 beq LL(999) .align 4LL(160): LFDX f8, X, INCXM1 LFDUX f9, X, INCX fabs f8, f8 fabs f9, f9 fadd f8, f8, f9 fsub f16, f1, f8 fsel f1, f16, f1, f8 bdnz LL(160) .align 4LL(999): fsub f8, f0, f1 fsub f9, f2, f3 fsel f0, f8, f0, f1 fsel f2, f9, f2, f3 fsub f8, f0, f2 fsel f1, f8, f0, f2 .align 4LL(9999): 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 + -