📄 axpy.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" #ifndef NEEDPARAM#ifndef DOUBLE#include "sparam.h"#else#include "dparam.h"#endif#endif #ifdef linux#ifndef __64BIT__#define N r3#define X r6#define INCX r7 #define Y r8#define INCY r9#define PREA r4#define YY r5#else#define N r3#define X r7#define INCX r8 #define Y r9#define INCY r10#define PREA r4#define YY r5#endif#endif#if defined(_AIX) || defined(__APPLE__)#if !defined(__64BIT__) && defined(DOUBLE)#define N r3#define X r8#define INCX r9 #define Y r10#define INCY r4#define PREA r5#define YY r6#else#define N r3#define X r7#define INCX r8 #define Y r9#define INCY r10#define PREA r4#define YY r5#endif#endif#define ALPHA f24#ifndef NEEDPARAM #define STACKSIZE 96 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)#if (defined(_AIX) || defined(__APPLE__)) && !defined(__64BIT__) && defined(DOUBLE) lwz INCY, 56 + STACKSIZE(SP)#endif fmr ALPHA, f1 slwi INCX, INCX, BASE_SHIFT slwi INCY, INCY, BASE_SHIFT#ifdef CELL li PREA, 120 * SIZE#endif#ifdef PPC970 li PREA, 76 * SIZE#endif#ifdef POWER4 li PREA, 26 * SIZE#endif#ifdef POWER5 li PREA, 72 * SIZE#endif cmpwi cr0, N, 0 ble- LL(999) cmpwi cr0, INCX, SIZE bne- cr0, LL(100) cmpwi cr0, INCY, SIZE bne- cr0, LL(100) srawi. r0, N, 4 mtspr CTR, r0 beq- cr0, LL(50) .align 4 LFD f0, 0 * SIZE(X) LFD f1, 1 * SIZE(X) LFD f2, 2 * SIZE(X) LFD f3, 3 * SIZE(X) LFD f8, 0 * SIZE(Y) LFD f9, 1 * SIZE(Y) LFD f10, 2 * SIZE(Y) LFD f11, 3 * SIZE(Y) LFD f4, 4 * SIZE(X) LFD f5, 5 * SIZE(X) LFD f6, 6 * SIZE(X) LFD f7, 7 * SIZE(X) LFD f12, 4 * SIZE(Y) LFD f13, 5 * SIZE(Y) LFD f14, 6 * SIZE(Y) LFD f15, 7 * SIZE(Y) bdz LL(20) .align 4LL(10): FMADD f16, ALPHA, f0, f8 FMADD f17, ALPHA, f1, f9 FMADD f18, ALPHA, f2, f10 FMADD f19, ALPHA, f3, f11 LFD f0, 8 * SIZE(X) LFD f1, 9 * SIZE(X) LFD f2, 10 * SIZE(X) LFD f3, 11 * SIZE(X) LFD f8, 8 * SIZE(Y) LFD f9, 9 * SIZE(Y) LFD f10, 10 * SIZE(Y) LFD f11, 11 * SIZE(Y) STFD f16, 0 * SIZE(Y) STFD f17, 1 * SIZE(Y) STFD f18, 2 * SIZE(Y) STFD f19, 3 * SIZE(Y) FMADD f20, ALPHA, f4, f12 FMADD f21, ALPHA, f5, f13 FMADD f22, ALPHA, f6, f14 FMADD f23, ALPHA, f7, f15 LFD f4, 12 * SIZE(X) LFD f5, 13 * SIZE(X) LFD f6, 14 * SIZE(X) LFD f7, 15 * SIZE(X) LFD f12, 12 * SIZE(Y) LFD f13, 13 * SIZE(Y) LFD f14, 14 * SIZE(Y) LFD f15, 15 * SIZE(Y) STFD f20, 4 * SIZE(Y) STFD f21, 5 * SIZE(Y) STFD f22, 6 * SIZE(Y) STFD f23, 7 * SIZE(Y) FMADD f16, ALPHA, f0, f8 FMADD f17, ALPHA, f1, f9 FMADD f18, ALPHA, f2, f10 FMADD f19, ALPHA, f3, f11 LFD f0, 16 * SIZE(X) LFD f1, 17 * SIZE(X) LFD f2, 18 * SIZE(X) LFD f3, 19 * SIZE(X) LFD f8, 16 * SIZE(Y) LFD f9, 17 * SIZE(Y) LFD f10, 18 * SIZE(Y) LFD f11, 19 * SIZE(Y) STFD f16, 8 * SIZE(Y) STFD f17, 9 * SIZE(Y) STFD f18, 10 * SIZE(Y) STFD f19, 11 * SIZE(Y) FMADD f20, ALPHA, f4, f12 FMADD f21, ALPHA, f5, f13 FMADD f22, ALPHA, f6, f14 FMADD f23, ALPHA, f7, f15 LFD f4, 20 * SIZE(X) LFD f5, 21 * SIZE(X) LFD f6, 22 * SIZE(X) LFD f7, 23 * SIZE(X) LFD f12, 20 * SIZE(Y) LFD f13, 21 * SIZE(Y) LFD f14, 22 * SIZE(Y) LFD f15, 23 * SIZE(Y) STFD f20, 12 * SIZE(Y) STFD f21, 13 * SIZE(Y) STFD f22, 14 * SIZE(Y) STFD f23, 15 * SIZE(Y) addi X, X, 16 * SIZE addi Y, Y, 16 * SIZE PREFETCH_X PREFETCH_Y bdnz LL(10) .align 4LL(20): FMADD f16, ALPHA, f0, f8 FMADD f17, ALPHA, f1, f9 FMADD f18, ALPHA, f2, f10 FMADD f19, ALPHA, f3, f11 LFD f0, 8 * SIZE(X) LFD f1, 9 * SIZE(X) LFD f2, 10 * SIZE(X) LFD f3, 11 * SIZE(X) LFD f8, 8 * SIZE(Y) LFD f9, 9 * SIZE(Y) LFD f10, 10 * SIZE(Y) LFD f11, 11 * SIZE(Y) FMADD f20, ALPHA, f4, f12 FMADD f21, ALPHA, f5, f13 FMADD f22, ALPHA, f6, f14 FMADD f23, ALPHA, f7, f15 LFD f4, 12 * SIZE(X) LFD f5, 13 * SIZE(X) LFD f6, 14 * SIZE(X) LFD f7, 15 * SIZE(X) LFD f12, 12 * SIZE(Y) LFD f13, 13 * SIZE(Y) LFD f14, 14 * SIZE(Y) LFD f15, 15 * SIZE(Y) STFD f16, 0 * SIZE(Y) STFD f17, 1 * SIZE(Y) STFD f18, 2 * SIZE(Y) STFD f19, 3 * SIZE(Y) FMADD f16, ALPHA, f0, f8 FMADD f17, ALPHA, f1, f9 FMADD f18, ALPHA, f2, f10 FMADD f19, ALPHA, f3, f11 STFD f20, 4 * SIZE(Y) STFD f21, 5 * SIZE(Y) STFD f22, 6 * SIZE(Y) STFD f23, 7 * SIZE(Y) FMADD f20, ALPHA, f4, f12 FMADD f21, ALPHA, f5, f13 FMADD f22, ALPHA, f6, f14 FMADD f23, ALPHA, f7, f15 STFD f16, 8 * SIZE(Y) STFD f17, 9 * SIZE(Y) STFD f18, 10 * SIZE(Y) STFD f19, 11 * SIZE(Y) STFD f20, 12 * SIZE(Y) STFD f21, 13 * SIZE(Y) STFD f22, 14 * SIZE(Y) STFD f23, 15 * SIZE(Y) addi X, X, 16 * SIZE addi Y, Y, 16 * SIZE .align 4LL(50): andi. r0, N, 15 mtspr CTR, r0 beq LL(999) .align 4LL(60): LFD f0, 0 * SIZE(X) LFD f8, 0 * SIZE(Y) FMADD f16, ALPHA, f0, f8 STFD f16, 0 * SIZE(Y) addi X, X, 1 * SIZE addi Y, Y, 1 * SIZE bdnz LL(60) b LL(999) .align 4LL(100): sub X, X, INCX sub Y, Y, INCY mr YY, Y srawi. r0, N, 4 mtspr CTR, r0 beq- LL(150) .align 4 LFDUX f0, X, INCX LFDUX f1, X, INCX LFDUX f2, X, INCX LFDUX f3, X, INCX LFDUX f8, Y, INCY LFDUX f9, Y, INCY LFDUX f10, Y, INCY LFDUX f11, Y, INCY LFDUX f4, X, INCX LFDUX f5, X, INCX LFDUX f6, X, INCX LFDUX f7, X, INCX LFDUX f12, Y, INCY LFDUX f13, Y, INCY LFDUX f14, Y, INCY LFDUX f15, Y, INCY bdz LL(120) .align 4LL(110): FMADD f16, ALPHA, f0, f8 FMADD f17, ALPHA, f1, f9 FMADD f18, ALPHA, f2, f10 FMADD f19, ALPHA, f3, f11 LFDUX f0, X, INCX LFDUX f1, X, INCX LFDUX f2, X, INCX LFDUX f3, X, INCX LFDUX f8, Y, INCY LFDUX f9, Y, INCY LFDUX f10, Y, INCY LFDUX f11, Y, INCY FMADD f20, ALPHA, f4, f12 FMADD f21, ALPHA, f5, f13 FMADD f22, ALPHA, f6, f14 FMADD f23, ALPHA, f7, f15 LFDUX f4, X, INCX LFDUX f5, X, INCX LFDUX f6, X, INCX LFDUX f7, X, INCX LFDUX f12, Y, INCY LFDUX f13, Y, INCY LFDUX f14, Y, INCY LFDUX f15, Y, INCY STFDUX f16, YY, INCY STFDUX f17, YY, INCY STFDUX f18, YY, INCY STFDUX f19, YY, INCY FMADD f16, ALPHA, f0, f8 FMADD f17, ALPHA, f1, f9 FMADD f18, ALPHA, f2, f10 FMADD f19, ALPHA, f3, f11 LFDUX f0, X, INCX LFDUX f1, X, INCX LFDUX f2, X, INCX LFDUX f3, X, INCX LFDUX f8, Y, INCY LFDUX f9, Y, INCY LFDUX f10, Y, INCY LFDUX f11, Y, INCY STFDUX f20, YY, INCY STFDUX f21, YY, INCY STFDUX f22, YY, INCY STFDUX f23, YY, INCY FMADD f20, ALPHA, f4, f12 FMADD f21, ALPHA, f5, f13 FMADD f22, ALPHA, f6, f14 FMADD f23, ALPHA, f7, f15 LFDUX f4, X, INCX LFDUX f5, X, INCX LFDUX f6, X, INCX LFDUX f7, X, INCX LFDUX f12, Y, INCY LFDUX f13, Y, INCY LFDUX f14, Y, INCY LFDUX f15, Y, INCY STFDUX f16, YY, INCY STFDUX f17, YY, INCY STFDUX f18, YY, INCY STFDUX f19, YY, INCY STFDUX f20, YY, INCY STFDUX f21, YY, INCY STFDUX f22, YY, INCY STFDUX f23, YY, INCY bdnz LL(110) .align 4LL(120): FMADD f16, ALPHA, f0, f8 FMADD f17, ALPHA, f1, f9 FMADD f18, ALPHA, f2, f10 FMADD f19, ALPHA, f3, f11 LFDUX f0, X, INCX LFDUX f1, X, INCX LFDUX f2, X, INCX LFDUX f3, X, INCX LFDUX f8, Y, INCY LFDUX f9, Y, INCY LFDUX f10, Y, INCY LFDUX f11, Y, INCY FMADD f20, ALPHA, f4, f12 FMADD f21, ALPHA, f5, f13 FMADD f22, ALPHA, f6, f14 FMADD f23, ALPHA, f7, f15 LFDUX f4, X, INCX LFDUX f5, X, INCX LFDUX f6, X, INCX LFDUX f7, X, INCX LFDUX f12, Y, INCY LFDUX f13, Y, INCY LFDUX f14, Y, INCY LFDUX f15, Y, INCY STFDUX f16, YY, INCY STFDUX f17, YY, INCY STFDUX f18, YY, INCY STFDUX f19, YY, INCY FMADD f16, ALPHA, f0, f8 FMADD f17, ALPHA, f1, f9 FMADD f18, ALPHA, f2, f10 FMADD f19, ALPHA, f3, f11 STFDUX f20, YY, INCY STFDUX f21, YY, INCY STFDUX f22, YY, INCY STFDUX f23, YY, INCY FMADD f20, ALPHA, f4, f12 FMADD f21, ALPHA, f5, f13 FMADD f22, ALPHA, f6, f14 FMADD f23, ALPHA, f7, f15 STFDUX f16, YY, INCY STFDUX f17, YY, INCY STFDUX f18, YY, INCY STFDUX f19, YY, INCY STFDUX f20, YY, INCY STFDUX f21, YY, INCY STFDUX f22, YY, INCY STFDUX f23, YY, INCY .align 4LL(150): andi. r0, N, 15 mtspr CTR, r0 beq LL(999) .align 4LL(160): LFDUX f0, X, INCX LFDUX f8, Y, INCY FMADD f16, ALPHA, f0, f8 STFDUX f16, YY, INCY bdnz LL(160) .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) 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 + -