📄 zrot.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#define N r3#define X r4#define INCX r5 #define Y r6#define INCY r7#define PREA r8#define XX r9#define YY r10#define INCXM1 r11#define INCYM1 r12#define C f1#define S f2#define STACKSIZE 32#ifndef NEEDPARAM 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) slwi INCX, INCX, ZBASE_SHIFT slwi INCY, INCY, ZBASE_SHIFT subi INCXM1, INCX, SIZE subi INCYM1, INCY, SIZE li PREA, 16 * 8 * SIZE cmpwi cr0, N, 0 ble- LL(999) cmpwi cr0, INCX, 2 * SIZE bne- cr0, LL(100) cmpwi cr0, INCY, 2 * SIZE bne- cr0, LL(100) srawi. r0, N, 3 mtspr CTR, r0 beq- cr0, LL(50) LFD f0, 0 * SIZE(X) LFD f4, 1 * SIZE(X) LFD f6, 2 * SIZE(X) LFD f8, 3 * SIZE(X) LFD f3, 0 * SIZE(Y) LFD f5, 1 * SIZE(Y) LFD f7, 2 * SIZE(Y) LFD f9, 3 * SIZE(Y) bdz LL(12) .align 4LL(10): FMUL f10, C, f0 FMUL f11, C, f3 FMUL f12, C, f4 FMUL f13, C, f5 FMUL f14, C, f6 FMUL f15, C, f7 FMUL f16, C, f8 FMUL f17, C, f9 FMADD f10, S, f3, f10 FNMSUB f11, S, f0, f11 FMADD f12, S, f5, f12 FNMSUB f13, S, f4, f13 FMADD f14, S, f7, f14 FNMSUB f15, S, f6, f15 FMADD f16, S, f9, f16 FNMSUB f17, S, f8, f17 LFD f0, 4 * SIZE(X) LFD f4, 5 * SIZE(X) LFD f6, 6 * SIZE(X) LFD f8, 7 * SIZE(X) LFD f3, 4 * SIZE(Y) LFD f5, 5 * SIZE(Y) LFD f7, 6 * SIZE(Y) LFD f9, 7 * SIZE(Y) STFD f10, 0 * SIZE(X) STFD f12, 1 * SIZE(X) STFD f14, 2 * SIZE(X) STFD f16, 3 * SIZE(X) STFD f11, 0 * SIZE(Y) STFD f13, 1 * SIZE(Y) STFD f15, 2 * SIZE(Y) STFD f17, 3 * SIZE(Y) FMUL f10, C, f0 FMUL f11, C, f3 FMUL f12, C, f4 FMUL f13, C, f5 FMUL f14, C, f6 FMUL f15, C, f7 FMUL f16, C, f8 FMUL f17, C, f9 FMADD f10, S, f3, f10 FNMSUB f11, S, f0, f11 FMADD f12, S, f5, f12 FNMSUB f13, S, f4, f13 FMADD f14, S, f7, f14 FNMSUB f15, S, f6, f15 FMADD f16, S, f9, f16 FNMSUB f17, S, f8, f17 LFD f0, 8 * SIZE(X) LFD f4, 9 * SIZE(X) LFD f6, 10 * SIZE(X) LFD f8, 11 * SIZE(X) LFD f3, 8 * SIZE(Y) LFD f5, 9 * SIZE(Y) LFD f7, 10 * SIZE(Y) LFD f9, 11 * SIZE(Y) STFD f10, 4 * SIZE(X) STFD f12, 5 * SIZE(X) STFD f14, 6 * SIZE(X) STFD f16, 7 * SIZE(X) STFD f11, 4 * SIZE(Y) STFD f13, 5 * SIZE(Y) STFD f15, 6 * SIZE(Y) STFD f17, 7 * SIZE(Y) FMUL f10, C, f0 FMUL f11, C, f3 FMUL f12, C, f4 FMUL f13, C, f5 FMUL f14, C, f6 FMUL f15, C, f7 FMUL f16, C, f8 FMUL f17, C, f9 FMADD f10, S, f3, f10 FNMSUB f11, S, f0, f11 FMADD f12, S, f5, f12 FNMSUB f13, S, f4, f13 FMADD f14, S, f7, f14 FNMSUB f15, S, f6, f15 FMADD f16, S, f9, f16 FNMSUB f17, S, f8, f17 LFD f0, 12 * SIZE(X) LFD f4, 13 * SIZE(X) LFD f6, 14 * SIZE(X) LFD f8, 15 * SIZE(X) LFD f3, 12 * SIZE(Y) LFD f5, 13 * SIZE(Y) LFD f7, 14 * SIZE(Y) LFD f9, 15 * SIZE(Y) STFD f10, 8 * SIZE(X) STFD f12, 9 * SIZE(X) STFD f14, 10 * SIZE(X) STFD f16, 11 * SIZE(X) STFD f11, 8 * SIZE(Y) STFD f13, 9 * SIZE(Y) STFD f15, 10 * SIZE(Y) STFD f17, 11 * SIZE(Y) FMUL f10, C, f0 FMUL f11, C, f3 FMUL f12, C, f4 FMUL f13, C, f5 FMUL f14, C, f6 FMUL f15, C, f7 FMUL f16, C, f8 FMUL f17, C, f9 FMADD f10, S, f3, f10 FNMSUB f11, S, f0, f11 FMADD f12, S, f5, f12 FNMSUB f13, S, f4, f13 FMADD f14, S, f7, f14 FNMSUB f15, S, f6, f15 FMADD f16, S, f9, f16 FNMSUB f17, S, f8, f17 LFD f0, 16 * SIZE(X) LFD f4, 17 * SIZE(X) LFD f6, 18 * SIZE(X) LFD f8, 19 * SIZE(X) LFD f3, 16 * SIZE(Y) LFD f5, 17 * SIZE(Y) LFD f7, 18 * SIZE(Y) LFD f9, 19 * SIZE(Y) STFD f10, 12 * SIZE(X) STFD f12, 13 * SIZE(X) STFD f14, 14 * SIZE(X) STFD f16, 15 * SIZE(X) STFD f11, 12 * SIZE(Y) STFD f13, 13 * SIZE(Y) STFD f15, 14 * SIZE(Y) STFD f17, 15 * SIZE(Y) addi X, X, 16 * SIZE addi Y, Y, 16 * SIZE#if !defined(PPC440) && !defined(PPC440FP2) PREFETCH_X PREFETCH_Y#endif bdnz LL(10) .align 4LL(12): FMUL f10, C, f0 FMUL f11, C, f3 FMUL f12, C, f4 FMUL f13, C, f5 FMUL f14, C, f6 FMUL f15, C, f7 FMUL f16, C, f8 FMUL f17, C, f9 FMADD f10, S, f3, f10 FNMSUB f11, S, f0, f11 FMADD f12, S, f5, f12 FNMSUB f13, S, f4, f13 FMADD f14, S, f7, f14 FNMSUB f15, S, f6, f15 FMADD f16, S, f9, f16 FNMSUB f17, S, f8, f17 STFD f10, 0 * SIZE(X) STFD f12, 1 * SIZE(X) STFD f14, 2 * SIZE(X) STFD f16, 3 * SIZE(X) STFD f11, 0 * SIZE(Y) STFD f13, 1 * SIZE(Y) STFD f15, 2 * SIZE(Y) STFD f17, 3 * SIZE(Y) LFD f0, 4 * SIZE(X) LFD f4, 5 * SIZE(X) LFD f6, 6 * SIZE(X) LFD f8, 7 * SIZE(X) LFD f3, 4 * SIZE(Y) LFD f5, 5 * SIZE(Y) LFD f7, 6 * SIZE(Y) LFD f9, 7 * SIZE(Y) FMUL f10, C, f0 FMUL f11, C, f3 FMUL f12, C, f4 FMUL f13, C, f5 FMUL f14, C, f6 FMUL f15, C, f7 FMUL f16, C, f8 FMUL f17, C, f9 FMADD f10, S, f3, f10 FNMSUB f11, S, f0, f11 FMADD f12, S, f5, f12 FNMSUB f13, S, f4, f13 FMADD f14, S, f7, f14 FNMSUB f15, S, f6, f15 FMADD f16, S, f9, f16 FNMSUB f17, S, f8, f17 STFD f10, 4 * SIZE(X) STFD f12, 5 * SIZE(X) STFD f14, 6 * SIZE(X) STFD f16, 7 * SIZE(X) STFD f11, 4 * SIZE(Y) STFD f13, 5 * SIZE(Y) STFD f15, 6 * SIZE(Y) STFD f17, 7 * SIZE(Y) LFD f0, 8 * SIZE(X) LFD f4, 9 * SIZE(X) LFD f6, 10 * SIZE(X) LFD f8, 11 * SIZE(X) LFD f3, 8 * SIZE(Y) LFD f5, 9 * SIZE(Y) LFD f7, 10 * SIZE(Y) LFD f9, 11 * SIZE(Y) FMUL f10, C, f0 FMUL f11, C, f3 FMUL f12, C, f4 FMUL f13, C, f5 FMUL f14, C, f6 FMUL f15, C, f7 FMUL f16, C, f8 FMUL f17, C, f9 FMADD f10, S, f3, f10 FNMSUB f11, S, f0, f11 FMADD f12, S, f5, f12 FNMSUB f13, S, f4, f13 FMADD f14, S, f7, f14 FNMSUB f15, S, f6, f15 FMADD f16, S, f9, f16 FNMSUB f17, S, f8, f17 STFD f10, 8 * SIZE(X) STFD f12, 9 * SIZE(X) STFD f14, 10 * SIZE(X) STFD f16, 11 * SIZE(X) STFD f11, 8 * SIZE(Y) STFD f13, 9 * SIZE(Y) STFD f15, 10 * SIZE(Y) STFD f17, 11 * SIZE(Y) LFD f0, 12 * SIZE(X) LFD f4, 13 * SIZE(X) LFD f6, 14 * SIZE(X) LFD f8, 15 * SIZE(X) LFD f3, 12 * SIZE(Y) LFD f5, 13 * SIZE(Y) LFD f7, 14 * SIZE(Y) LFD f9, 15 * SIZE(Y) FMUL f10, C, f0 FMUL f11, C, f3 FMUL f12, C, f4 FMUL f13, C, f5 FMUL f14, C, f6 FMUL f15, C, f7 FMUL f16, C, f8 FMUL f17, C, f9 FMADD f10, S, f3, f10 FNMSUB f11, S, f0, f11 FMADD f12, S, f5, f12 FNMSUB f13, S, f4, f13 FMADD f14, S, f7, f14 FNMSUB f15, S, f6, f15 FMADD f16, S, f9, f16 FNMSUB f17, S, f8, f17 STFD f10, 12 * SIZE(X) STFD f12, 13 * SIZE(X) STFD f14, 14 * SIZE(X) STFD f16, 15 * SIZE(X) STFD f11, 12 * SIZE(Y) STFD f13, 13 * SIZE(Y) STFD f15, 14 * SIZE(Y) STFD f17, 15 * SIZE(Y) addi X, X, 16 * SIZE addi Y, Y, 16 * SIZE .align 4LL(50): andi. r0, N, 7 mtspr CTR, r0 beq LL(999) .align 4LL(60): LFD f3, 0 * SIZE(X) LFD f4, 0 * SIZE(Y) LFD f5, 1 * SIZE(X) LFD f6, 1 * SIZE(Y) FMUL f10, C, f3 FMUL f11, C, f4 FMUL f12, C, f5 FMUL f13, C, f6 FMADD f10, S, f4, f10 FNMSUB f11, S, f3, f11 FMADD f12, S, f6, f12 FNMSUB f13, S, f5, f13 STFD f10, 0 * SIZE(X) STFD f11, 0 * SIZE(Y) STFD f12, 1 * SIZE(X) STFD f13, 1 * SIZE(Y) addi X, X, 2 * SIZE addi Y, Y, 2 * SIZE bdnz LL(60) b LL(999) .align 4LL(100): sub X, X, INCXM1 sub Y, Y, INCYM1 mr XX, X mr YY, Y srawi. r0, N, 2 mtspr CTR, r0 beq- LL(150) .align 4LL(110): LFDX f0, X, INCXM1 LFDX f3, Y, INCYM1 LFDUX f4, X, INCX LFDUX f5, Y, INCY LFDX f6, X, INCXM1 LFDX f7, Y, INCYM1 LFDUX f8, X, INCX LFDUX f9, Y, INCY FMUL f10, C, f0 FMUL f11, C, f3 FMUL f12, C, f4 FMUL f13, C, f5 FMUL f14, C, f6 FMUL f15, C, f7 FMUL f16, C, f8 FMUL f17, C, f9 FMADD f10, S, f3, f10 FNMSUB f11, S, f0, f11 FMADD f12, S, f5, f12 FNMSUB f13, S, f4, f13 FMADD f14, S, f7, f14 FNMSUB f15, S, f6, f15 FMADD f16, S, f9, f16 FNMSUB f17, S, f8, f17 STFDX f10, XX, INCXM1 STFDX f11, YY, INCYM1 STFDUX f12, XX, INCX STFDUX f13, YY, INCY STFDX f14, XX, INCXM1 STFDX f15, YY, INCYM1 STFDUX f16, XX, INCX STFDUX f17, YY, INCY LFDX f0, X, INCXM1 LFDX f3, Y, INCYM1 LFDUX f4, X, INCX LFDUX f5, Y, INCY LFDX f6, X, INCXM1 LFDX f7, Y, INCYM1 LFDUX f8, X, INCX LFDUX f9, Y, INCY FMUL f10, C, f0 FMUL f11, C, f3 FMUL f12, C, f4 FMUL f13, C, f5 FMUL f14, C, f6 FMUL f15, C, f7 FMUL f16, C, f8 FMUL f17, C, f9 FMADD f10, S, f3, f10 FNMSUB f11, S, f0, f11 FMADD f12, S, f5, f12 FNMSUB f13, S, f4, f13 FMADD f14, S, f7, f14 FNMSUB f15, S, f6, f15 FMADD f16, S, f9, f16 FNMSUB f17, S, f8, f17 STFDX f10, XX, INCXM1 STFDX f11, YY, INCYM1 STFDUX f12, XX, INCX STFDUX f13, YY, INCY STFDX f14, XX, INCXM1 STFDX f15, YY, INCYM1 STFDUX f16, XX, INCX STFDUX f17, YY, INCY bdnz LL(110) .align 4LL(150): andi. r0, N, 3 mtspr CTR, r0 beq LL(999) .align 4LL(160): LFDX f0, X, INCXM1 LFDX f3, Y, INCYM1 LFDUX f4, X, INCX LFDUX f5, Y, INCY FMUL f10, C, f0 FMUL f11, C, f3 FMUL f12, C, f4 FMUL f13, C, f5 FMADD f10, S, f3, f10 FNMSUB f11, S, f0, f11 FMADD f12, S, f5, f12 FNMSUB f13, S, f4, f13 STFDX f10, XX, INCXM1 STFDX f11, YY, INCYM1 STFDUX f12, XX, INCX STFDUX f13, YY, INCY bdnz LL(160) .align 4LL(999): lfd f14, 0(SP) lfd f15, 8(SP) lfd f16, 16(SP) lfd f17, 24(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 + -