📄 zasum_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 f0#define STACKSIZE 16 PROLOGUE PROFCODE addi SP, SP, -STACKSIZE li r0, 0 stw r0, 0(SP)#ifdef F_INTERFACE LDINT N, 0(N) LDINT INCX, 0(INCX)#endif lfs FZERO, 0(SP) slwi INCX, INCX, ZBASE_SHIFT fmr f1, FZERO li PREA, 8 * 16 * SIZE fmr f2, FZERO subi INCXM1, INCX, SIZE cmpwi cr0, N, 0 fmr f3, FZERO ble- LL(999) cmpwi cr0, INCX, 0 ble- LL(999) cmpwi cr0, INCX, SIZE * 2 bne- cr0, LL(20) srawi. r0, N, 3 mtspr CTR, r0 beq- cr0, LL(15) .align 4 LFD f8, 0 * SIZE(X) LFD f9, 1 * SIZE(X) fabs f4, f8 LFD f10, 2 * SIZE(X) fabs f5, f9 LFD f11, 3 * SIZE(X) fabs f6, f10 LFD f8, 4 * SIZE(X) fabs f7, f11 bdz LL(13) .align 4LL(12): FADD f0, f0, f4 dcbt X, PREA fabs f4, f8 LFD f9, 5 * SIZE(X) FADD f1, f1, f5 nop fabs f5, f9 LFD f10, 6 * SIZE(X) FADD f2, f2, f6 nop fabs f6, f10 LFD f11, 7 * SIZE(X) FADD f3, f3, f7 nop fabs f7, f11 LFD f8, 8 * SIZE(X) FADD f0, f0, f4 nop fabs f4, f8 LFD f9, 9 * SIZE(X) FADD f1, f1, f5 nop fabs f5, f9 LFD f10, 10 * SIZE(X) FADD f2, f2, f6 nop fabs f6, f10 LFD f11, 11 * SIZE(X) FADD f3, f3, f7 nop fabs f7, f11 LFD f8, 12 * SIZE(X) FADD f0, f0, f4 nop fabs f4, f8 LFD f9, 13 * SIZE(X) FADD f1, f1, f5 nop fabs f5, f9 LFD f10, 14 * SIZE(X) FADD f2, f2, f6 nop fabs f6, f10 LFD f11, 15 * SIZE(X) FADD f3, f3, f7 nop fabs f7, f11 LFD f8, 16 * SIZE(X) FADD f0, f0, f4 nop fabs f4, f8 LFD f9, 17 * SIZE(X) FADD f1, f1, f5 addi X, X, 16 * SIZE fabs f5, f9 LFD f10, 2 * SIZE(X) FADD f2, f2, f6 nop fabs f6, f10 LFD f11, 3 * SIZE(X) FADD f3, f3, f7 LFD f8, 4 * SIZE(X) fabs f7, f11 bdnz LL(12) .align 4LL(13): FADD f0, f0, f4 nop fabs f4, f8 LFD f9, 5 * SIZE(X) FADD f1, f1, f5 nop fabs f5, f9 LFD f10, 6 * SIZE(X) FADD f2, f2, f6 nop fabs f6, f10 LFD f11, 7 * SIZE(X) FADD f3, f3, f7 nop fabs f7, f11 LFD f8, 8 * SIZE(X) FADD f0, f0, f4 nop fabs f4, f8 LFD f9, 9 * SIZE(X) FADD f1, f1, f5 nop fabs f5, f9 LFD f10, 10 * SIZE(X) FADD f2, f2, f6 nop fabs f6, f10 LFD f11, 11 * SIZE(X) FADD f3, f3, f7 nop fabs f7, f11 LFD f8, 12 * SIZE(X) FADD f0, f0, f4 nop fabs f4, f8 LFD f9, 13 * SIZE(X) FADD f1, f1, f5 nop fabs f5, f9 LFD f10, 14 * SIZE(X) FADD f2, f2, f6 addi X, X, 16 * SIZE fabs f6, f10 LFD f11, -1 * SIZE(X) FADD f3, f3, f7 fabs f7, f11 FADD f0, f0, f4 FADD f1, f1, f5 FADD f2, f2, f6 FADD f3, f3, f7 .align 4LL(15): andi. r0, N, 7 beq LL(999) andi. r0, N, 4 beq LL(16) LFD f8, 0 * SIZE(X) LFD f9, 1 * SIZE(X) fabs f4, f8 LFD f10, 2 * SIZE(X) fabs f5, f9 LFD f11, 3 * SIZE(X) fabs f6, f10 LFD f8, 4 * SIZE(X) fabs f7, f11 FADD f0, f0, f4 nop fabs f4, f8 LFD f9, 5 * SIZE(X) FADD f1, f1, f5 nop fabs f5, f9 LFD f10, 6 * SIZE(X) FADD f2, f2, f6 addi X, X, 8 * SIZE fabs f6, f10 LFD f11, -1 * SIZE(X) FADD f3, f3, f7 fabs f7, f11 FADD f0, f0, f4 FADD f1, f1, f5 FADD f2, f2, f6 FADD f3, f3, f7 .align 4LL(16): andi. r0, N, 2 beq LL(17) LFD f8, 0 * SIZE(X) LFD f9, 1 * SIZE(X) fabs f4, f8 LFD f10, 2 * SIZE(X) fabs f5, f9 LFD f11, 3 * SIZE(X) fabs f6, f10 addi X, X, 4 * SIZE fabs f7, f11 nop FADD f0, f0, f4 FADD f1, f1, f5 FADD f2, f2, f6 FADD f3, f3, f7 .align 4LL(17): andi. r0, N, 1 beq LL(999) LFD f8, 0 * SIZE(X) LFD f9, 1 * SIZE(X) fabs f4, f8 fabs f5, f9 FADD f0, f0, f4 addi X, X, 2 * SIZE FADD f1, f1, f5 b LL(999) .align 4LL(20): sub X, X, INCXM1 srawi. r0, N, 3 mtspr CTR, r0 beq- cr0, LL(25) LFDX f8, X, INCXM1 LFDUX f9, X, INCX fabs f4, f8 LFDX f10, X, INCXM1 fabs f5, f9 LFDUX f11, X, INCX fabs f6, f10 LFDX f8, X, INCXM1 fabs f7, f11 bdz LL(23) .align 4LL(22): FADD f0, f0, f4 dcbt X, PREA fabs f4, f8 LFDUX f9, X, INCX FADD f1, f1, f5 nop fabs f5, f9 LFDX f10, X, INCXM1 FADD f2, f2, f6 nop fabs f6, f10 LFDUX f11, X, INCX FADD f3, f3, f7 nop fabs f7, f11 LFDX f8, X, INCXM1 FADD f0, f0, f4 nop fabs f4, f8 LFDUX f9, X, INCX FADD f1, f1, f5 nop fabs f5, f9 LFDX f10, X, INCXM1 FADD f2, f2, f6 nop fabs f6, f10 LFDUX f11, X, INCX FADD f3, f3, f7 nop fabs f7, f11 LFDX f8, X, INCXM1 FADD f0, f0, f4 nop fabs f4, f8 LFDUX f9, X, INCX FADD f1, f1, f5 nop fabs f5, f9 LFDX f10, X, INCXM1 FADD f2, f2, f6 nop fabs f6, f10 LFDUX f11, X, INCX FADD f3, f3, f7 nop fabs f7, f11 LFDX f8, X, INCXM1 FADD f0, f0, f4 nop fabs f4, f8 LFDUX f9, X, INCX FADD f1, f1, f5 nop fabs f5, f9 LFDX f10, X, INCXM1 FADD f2, f2, f6 nop fabs f6, f10 LFDUX f11, X, INCX FADD f3, f3, f7 LFDX f8, X, INCXM1 fabs f7, f11 bdnz LL(22) .align 4LL(23): FADD f0, f0, f4 nop fabs f4, f8 LFDUX f9, X, INCX FADD f1, f1, f5 nop fabs f5, f9 LFDX f10, X, INCXM1 FADD f2, f2, f6 nop fabs f6, f10 LFDUX f11, X, INCX FADD f3, f3, f7 nop fabs f7, f11 LFDX f8, X, INCXM1 FADD f0, f0, f4 nop fabs f4, f8 LFDUX f9, X, INCX FADD f1, f1, f5 nop fabs f5, f9 LFDX f10, X, INCXM1 FADD f2, f2, f6 nop fabs f6, f10 LFDUX f11, X, INCX FADD f3, f3, f7 nop fabs f7, f11 LFDX f8, X, INCXM1 FADD f0, f0, f4 nop fabs f4, f8 LFDUX f9, X, INCX FADD f1, f1, f5 nop fabs f5, f9 LFDX f10, X, INCXM1 FADD f2, f2, f6 nop fabs f6, f10 LFDUX f11, X, INCX FADD f3, f3, f7 fabs f7, f11 FADD f0, f0, f4 FADD f1, f1, f5 FADD f2, f2, f6 FADD f3, f3, f7 .align 4LL(25): andi. r0, N, 7 beq LL(999) andi. r0, N, 4 beq LL(26) LFDX f8, X, INCXM1 LFDUX f9, X, INCX fabs f4, f8 LFDX f10, X, INCXM1 fabs f5, f9 LFDUX f11, X, INCX fabs f6, f10 LFDX f8, X, INCXM1 fabs f7, f11 FADD f0, f0, f4 nop fabs f4, f8 LFDUX f9, X, INCX FADD f1, f1, f5 nop fabs f5, f9 LFDX f10, X, INCXM1 FADD f2, f2, f6 fabs f6, f10 LFDUX f11, X, INCX FADD f3, f3, f7 fabs f7, f11 FADD f0, f0, f4 FADD f1, f1, f5 FADD f2, f2, f6 FADD f3, f3, f7 .align 4LL(26): andi. r0, N, 2 beq LL(27) LFDX f8, X, INCXM1 LFDUX f9, X, INCX fabs f4, f8 LFDX f10, X, INCXM1 fabs f5, f9 LFDUX f11, X, INCX fabs f6, f10 fabs f7, f11 FADD f0, f0, f4 FADD f1, f1, f5 FADD f2, f2, f6 FADD f3, f3, f7 .align 4LL(27): andi. r0, N, 1 beq LL(999) LFDX f8, X, INCXM1 LFDUX f9, X, INCX fabs f4, f8 fabs f5, f9 FADD f0, f0, f4 FADD f1, f1, f5 .align 4LL(999): FADD f0, f0, f1 FADD f2, f2, f3 FADD f1, f0, f2 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 + -