📄 zdot_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" #if defined(F_INTERFACE) && defined(F_INTERFACE_F2C)#define RESULT r3#define N r4#define X r5#define INCX r6 #define Y r7#define INCY r8#define PREA r9#else#define N r3#define X r4#define INCX r5 #define Y r6#define INCY r7#define PREA r8#endif#define INCXM1 r10#define INCYM1 r11#define FZERO f0#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) LDINT INCY, 0(INCY)#endif slwi INCX, INCX, ZBASE_SHIFT slwi INCY, INCY, ZBASE_SHIFT subi INCXM1, INCX, SIZE subi INCYM1, INCY, SIZE fmr f1, FZERO fmr f2, FZERO fmr f3, FZERO fmr f4, FZERO fmr f5, FZERO fmr f6, FZERO fmr f7, FZERO fmr f24, FZERO fmr f25, FZERO fmr f26, FZERO fmr f27, FZERO fmr f28, FZERO fmr f29, FZERO fmr f30, FZERO fmr f31, FZERO li PREA, 16 * 10 * 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) .align 4 LFD f8, 0 * SIZE(X) LFD f9, 1 * SIZE(X) LFD f10, 2 * SIZE(X) LFD f11, 3 * SIZE(X) LFD f16, 0 * SIZE(Y) LFD f17, 1 * SIZE(Y) LFD f18, 2 * SIZE(Y) LFD f19, 3 * SIZE(Y) LFD f12, 4 * SIZE(X) LFD f13, 5 * SIZE(X) LFD f14, 6 * SIZE(X) LFD f15, 7 * SIZE(X) LFD f20, 4 * SIZE(Y) LFD f21, 5 * SIZE(Y) LFD f23, 7 * SIZE(Y) bdz LL(20) .align 4LL(10): FMADD f0, f8, f16, f0 LFD f22, 6 * SIZE(Y) FMADD f3, f8, f17, f3 LFD f8, 8 * SIZE(X) FMADD f1, f9, f17, f1 LFD f17, 9 * SIZE(Y) FMADD f2, f9, f16, f2 LFD f9, 9 * SIZE(X) FMADD f4, f10, f18, f4 LFD f16, 8 * SIZE(Y) FMADD f7, f10, f19, f7 LFD f10, 10 * SIZE(X) FMADD f5, f11, f19, f5 LFD f19, 11 * SIZE(Y) FMADD f6, f11, f18, f6 LFD f11, 11 * SIZE(X) FMADD f24, f12, f20, f24 LFD f18, 10 * SIZE(Y) FMADD f27, f12, f21, f27 LFD f12, 12 * SIZE(X) FMADD f25, f13, f21, f25 LFD f21, 13 * SIZE(Y) FMADD f26, f13, f20, f26 LFD f13, 13 * SIZE(X) FMADD f28, f14, f22, f28 LFD f20, 12 * SIZE(Y) FMADD f31, f14, f23, f31 LFD f14, 14 * SIZE(X) FMADD f29, f15, f23, f29 LFD f23, 15 * SIZE(Y) FMADD f30, f15, f22, f30 LFD f15, 15 * SIZE(X) FMADD f0, f8, f16, f0 LFD f22, 14 * SIZE(Y) FMADD f3, f8, f17, f3 LFD f8, 16 * SIZE(X) FMADD f1, f9, f17, f1 LFD f17, 17 * SIZE(Y) FMADD f2, f9, f16, f2 LFD f9, 17 * SIZE(X) FMADD f4, f10, f18, f4 LFD f16, 16 * SIZE(Y) FMADD f7, f10, f19, f7 LFD f10, 18 * SIZE(X) FMADD f5, f11, f19, f5 LFD f19, 19 * SIZE(Y) FMADD f6, f11, f18, f6 LFD f11, 19 * SIZE(X) FMADD f24, f12, f20, f24 LFD f18, 18 * SIZE(Y) FMADD f27, f12, f21, f27 LFD f12, 20 * SIZE(X) FMADD f25, f13, f21, f25 LFD f21, 21 * SIZE(Y) FMADD f26, f13, f20, f26 LFD f13, 21 * SIZE(X) FMADD f28, f14, f22, f28 LFD f20, 20 * SIZE(Y) FMADD f31, f14, f23, f31 LFD f14, 22 * SIZE(X) FMADD f29, f15, f23, f29 LFD f23, 23 * SIZE(Y) FMADD f30, f15, f22, f30 LFD f15, 23 * SIZE(X) dcbt X, PREA addi X, X, 16 * SIZE dcbt Y, PREA addi Y, Y, 16 * SIZE bdnz LL(10) .align 4LL(20): FMADD f0, f8, f16, f0 LFD f22, 6 * SIZE(Y) FMADD f3, f8, f17, f3 LFD f8, 8 * SIZE(X) FMADD f1, f9, f17, f1 LFD f17, 9 * SIZE(Y) FMADD f2, f9, f16, f2 LFD f9, 9 * SIZE(X) FMADD f4, f10, f18, f4 LFD f16, 8 * SIZE(Y) FMADD f7, f10, f19, f7 LFD f10, 10 * SIZE(X) FMADD f5, f11, f19, f5 LFD f19, 11 * SIZE(Y) FMADD f6, f11, f18, f6 LFD f11, 11 * SIZE(X) FMADD f24, f12, f20, f24 LFD f18, 10 * SIZE(Y) FMADD f27, f12, f21, f27 LFD f12, 12 * SIZE(X) FMADD f25, f13, f21, f25 LFD f21, 13 * SIZE(Y) FMADD f26, f13, f20, f26 LFD f13, 13 * SIZE(X) FMADD f28, f14, f22, f28 LFD f20, 12 * SIZE(Y) FMADD f31, f14, f23, f31 LFD f14, 14 * SIZE(X) FMADD f29, f15, f23, f29 LFD f23, 15 * SIZE(Y) FMADD f30, f15, f22, f30 LFD f15, 15 * SIZE(X) FMADD f0, f8, f16, f0 LFD f22, 14 * SIZE(Y) FMADD f3, f8, f17, f3 addi X, X, 16 * SIZE FMADD f1, f9, f17, f1 addi Y, Y, 16 * SIZE FMADD f2, f9, f16, f2 nop FMADD f4, f10, f18, f4 FMADD f7, f10, f19, f7 FMADD f5, f11, f19, f5 FMADD f6, f11, f18, f6 FMADD f24, f12, f20, f24 FMADD f27, f12, f21, f27 FMADD f25, f13, f21, f25 FMADD f26, f13, f20, f26 FMADD f28, f14, f22, f28 FMADD f31, f14, f23, f31 FMADD f29, f15, f23, f29 FMADD f30, f15, f22, f30 .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) LFD f16, 0 * SIZE(Y) LFD f17, 1 * SIZE(Y) addi X, X, 2 * SIZE addi Y, Y, 2 * SIZE FMADD f0, f8, f16, f0 FMADD f3, f8, f17, f3 FMADD f1, f9, f17, f1 FMADD f2, f9, f16, f2 bdnz LL(60) b LL(999) .align 4LL(100):#ifdef F_INTERFACE cmpwi cr0, INCX, 0 bge+ LL(102) subi r0, N, 1 mullw r0, r0, INCX sub X, X, r0 .align 4LL(102): cmpwi cr0, INCY, 0 bge+ LL(104) subi r0, N, 1 mullw r0, r0, INCY sub Y, Y, r0 .align 4LL(104):#endif sub X, X, INCXM1 sub Y, Y, INCYM1 srawi. r0, N, 3 mtspr CTR, r0 beq- LL(150) LFDX f8, X, INCXM1 LFDX f16, Y, INCYM1 LFDUX f9, X, INCX LFDUX f17, Y, INCY LFDX f10, X, INCXM1 LFDX f18, Y, INCYM1 LFDUX f11, X, INCX LFDUX f19, Y, INCY LFDX f12, X, INCXM1 LFDX f20, Y, INCYM1 LFDUX f13, X, INCX LFDUX f21, Y, INCY LFDX f14, X, INCXM1 LFDUX f15, X, INCX bdz LL(120) .align 4LL(110): FMADD f0, f8, f16, f0 LFDX f22, Y, INCYM1 FMADD f3, f8, f17, f3 LFDX f8, X, INCXM1 FMADD f1, f9, f17, f1 LFDUX f23, Y, INCY FMADD f2, f9, f16, f2 LFDUX f9, X, INCX FMADD f4, f10, f18, f4 LFDX f16, Y, INCYM1 FMADD f7, f10, f19, f7 LFDX f10, X, INCXM1 FMADD f5, f11, f19, f5 LFDUX f17, Y, INCY FMADD f6, f11, f18, f6 LFDUX f11, X, INCX FMADD f24, f12, f20, f24 LFDX f18, Y, INCYM1 FMADD f27, f12, f21, f27 LFDX f12, X, INCXM1 FMADD f25, f13, f21, f25 LFDUX f19, Y, INCY FMADD f26, f13, f20, f26 LFDUX f13, X, INCX FMADD f28, f14, f22, f28 LFDX f20, Y, INCYM1 FMADD f31, f14, f23, f31 LFDX f14, X, INCXM1 FMADD f29, f15, f23, f29 LFDUX f21, Y, INCY FMADD f30, f15, f22, f30 LFDUX f15, X, INCX FMADD f0, f8, f16, f0 LFDX f22, Y, INCYM1 FMADD f3, f8, f17, f3 LFDX f8, X, INCXM1 FMADD f1, f9, f17, f1 LFDUX f23, Y, INCY FMADD f2, f9, f16, f2 LFDUX f9, X, INCX FMADD f4, f10, f18, f4 LFDX f16, Y, INCYM1 FMADD f7, f10, f19, f7 LFDX f10, X, INCXM1 FMADD f5, f11, f19, f5 LFDUX f17, Y, INCY FMADD f6, f11, f18, f6 LFDUX f11, X, INCX FMADD f24, f12, f20, f24 LFDX f18, Y, INCYM1 FMADD f27, f12, f21, f27 LFDX f12, X, INCXM1 FMADD f25, f13, f21, f25 LFDUX f19, Y, INCY FMADD f26, f13, f20, f26 LFDUX f13, X, INCX FMADD f28, f14, f22, f28 LFDX f20, Y, INCYM1 FMADD f31, f14, f23, f31 LFDX f14, X, INCXM1 FMADD f29, f15, f23, f29 LFDUX f21, Y, INCY FMADD f30, f15, f22, f30 LFDUX f15, X, INCX bdnz LL(110) .align 4LL(120): FMADD f0, f8, f16, f0 LFDX f22, Y, INCYM1 FMADD f3, f8, f17, f3 LFDX f8, X, INCXM1 FMADD f1, f9, f17, f1 LFDUX f23, Y, INCY FMADD f2, f9, f16, f2 LFDUX f9, X, INCX FMADD f4, f10, f18, f4 LFDX f16, Y, INCYM1 FMADD f7, f10, f19, f7 LFDX f10, X, INCXM1 FMADD f5, f11, f19, f5 LFDUX f17, Y, INCY FMADD f6, f11, f18, f6 LFDUX f11, X, INCX FMADD f24, f12, f20, f24 LFDX f18, Y, INCYM1 FMADD f27, f12, f21, f27 LFDX f12, X, INCXM1 FMADD f25, f13, f21, f25 LFDUX f19, Y, INCY FMADD f26, f13, f20, f26 LFDUX f13, X, INCX FMADD f28, f14, f22, f28 LFDX f20, Y, INCYM1 FMADD f31, f14, f23, f31 LFDX f14, X, INCXM1 FMADD f29, f15, f23, f29 LFDUX f21, Y, INCY FMADD f30, f15, f22, f30 LFDUX f15, X, INCX FMADD f0, f8, f16, f0 LFDX f22, Y, INCYM1 FMADD f3, f8, f17, f3 LFDUX f23, Y, INCY FMADD f1, f9, f17, f1 FMADD f2, f9, f16, f2 FMADD f4, f10, f18, f4 FMADD f7, f10, f19, f7 FMADD f5, f11, f19, f5 FMADD f6, f11, f18, f6 FMADD f24, f12, f20, f24 FMADD f27, f12, f21, f27 FMADD f25, f13, f21, f25 FMADD f26, f13, f20, f26 FMADD f28, f14, f22, f28 FMADD f31, f14, f23, f31 FMADD f29, f15, f23, f29 FMADD f30, f15, f22, f30 .align 4LL(150): andi. r0, N, 7 mtspr CTR, r0 beq LL(999) .align 4LL(160): LFDX f8, X, INCXM1 LFDUX f9, X, INCX LFDX f16, Y, INCYM1 LFDUX f17, Y, INCY FMADD f0, f8, f16, f0 FMADD f3, f8, f17, f3 FMADD f1, f9, f17, f1 FMADD f2, f9, f16, f2 bdnz LL(160) .align 4LL(999): FADD f0, f0, f4 FADD f1, f1, f5 FADD f2, f2, f6 FADD f3, f3, f7 FADD f24, f28, f24 FADD f25, f29, f25 FADD f26, f30, f26 FADD f27, f31, f27 FADD f0, f0, f24 FADD f1, f1, f25 FADD f2, f2, f26 FADD f3, f3, f27#ifndef CONJ FSUB f1, f0, f1 FADD f2, f2, f3#else FADD f1, f0, f1 FSUB f2, f3, f2#endif#if defined(F_INTERFACE) && defined(F_INTERFACE_F2C) STFD f1, 0 * SIZE(RESULT) STFD f2, 1 * SIZE(RESULT)#endif#if defined(F_INTERFACE) && defined(F_INTERFACE_GFORT)#ifndef __64BIT__#ifndef DOUBLE stfs f1, 144(SP) stfs f2, 148(SP) lwz r3, 144(SP) lwz r4, 148(SP)#else stfd f1, 144(SP) stfd f2, 152(SP) lwz r3, 144(SP) lwz r4, 148(SP) lwz r5, 152(SP) lwz r6, 156(SP)#endif#else#ifndef DOUBLE stfs f1, 144(SP) stfs f2, 148(SP) ld r3, 144(SP)#else stfd f1, 144(SP) stfd f2, 152(SP) ld r3, 144(SP) ld r4, 152(SP)#endif#endif#endif 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 + -