⭐ 欢迎来到虫虫下载站! | 📦 资源下载 📁 资源专辑 ℹ️ 关于我们
⭐ 虫虫下载站

📄 zaxpy_hummer.s

📁 Optimized GotoBLAS libraries
💻 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	r6#define INCX	r7	#define Y	r8#define INCY	r9#define	YY	r4#define INCX2	r5#define INCY2	r10#define X1	r11#define Y1	INCX#define YY1	INCY#define ALPHA	f1#define A1	f0#define A2	f8#define A3	f2#define A4	f3#define A5	f4#define A6	f5#define A7	f6#define A8	f7#define A9	f25#define B1	f9#define B2	f10#define B3	f11#define B4	f12#define B5	f13#define B6	f14#define B7	f15#define B8	f16#define C1	f17#define C2	f18#define C3	f19#define C4	f20#define C5	f21#define C6	f22#define C7	f23#define C8	f24#define ALPHA_R ALPHA#define ALPHA_I A9#ifndef CONJ#define ADD1	FNMSUB#define ADD2	FMADD#else#define ADD1	FMADD#define ADD2	FNMSUB#endif#ifndef CONJ#define FXMADD1	fxcpmadd#define FXMADD2	fxcxnpma#else#define FXMADD1	fxcpnsma#define FXMADD2	fxcxma#endif	PROLOGUE	PROFCODE	li	r10, -16	stfpdux	f14, SP, r10	stfpdux	f15, SP, r10	stfpdux	f16, SP, r10	stfpdux	f17, SP, r10	stfpdux	f18, SP, r10	stfpdux	f19, SP, r10	stfpdux	f20, SP, r10	stfpdux	f21, SP, r10	stfpdux	f22, SP, r10	stfpdux	f23, SP, r10	stfpdux	f24, SP, r10	stfpdux	f25, SP, r10	fsmfp	ALPHA, f2	slwi	INCX,  INCX, BASE_SHIFT	slwi	INCY,  INCY, BASE_SHIFT	add	INCX2, INCX, INCX	add	INCY2, INCY, INCY	cmpwi	cr0, N, 0	ble	LL(999)	andi.	r0, X, 2 * SIZE - 1	bne	LL(100)	andi.	r0, Y, 2 * SIZE - 1	bne	LL(100)	sub	X,  X, INCX2	sub	Y,  Y, INCY2	mr	YY, Y	srawi.	r0, N, 3	mtspr	CTR,  r0	beq-	LL(15)	LFPDUX	A1,   X, INCX2	LFPDUX	B1,   Y, INCY2	LFPDUX	A2,   X, INCX2	LFPDUX	B2,   Y, INCY2	LFPDUX	A3,   X, INCX2	LFPDUX	B3,   Y, INCY2	LFPDUX	A4,   X, INCX2	LFPDUX	B4,   Y, INCY2	LFPDUX	A5,   X, INCX2	LFPDUX	B5,   Y, INCY2	LFPDUX	A6,   X, INCX2	LFPDUX	B6,   Y, INCY2	LFPDUX	A7,   X, INCX2	LFPDUX	B7,   Y, INCY2	LFPDUX	A8,   X, INCX2	LFPDUX	B8,   Y, INCY2	bdz	LL(13)	.align 4LL(12):	FXMADD1 C1, ALPHA, A1, B1	LFPDUX	B1,   Y, INCY2	FXMADD1 C2, ALPHA, A2, B2	LFPDUX	B2,   Y, INCY2	FXMADD1 C3, ALPHA, A3, B3	LFPDUX	B3,   Y, INCY2	FXMADD1 C4, ALPHA, A4, B4	LFPDUX	B4,   Y, INCY2	FXMADD1 C5, ALPHA, A5, B5	LFPDUX	B5,   Y, INCY2	FXMADD1 C6, ALPHA, A6, B6	LFPDUX	B6,   Y, INCY2	FXMADD1 C7, ALPHA, A7, B7	LFPDUX	B7,   Y, INCY2	FXMADD1 C8, ALPHA, A8, B8	LFPDUX	B8,   Y, INCY2	FXMADD2 C1, ALPHA, A1, C1	LFPDUX	A1,   X, INCX2	FXMADD2 C2, ALPHA, A2, C2	LFPDUX	A2,   X, INCX2	FXMADD2 C3, ALPHA, A3, C3	LFPDUX	A3,   X, INCX2	FXMADD2 C4, ALPHA, A4, C4	LFPDUX	A4,   X, INCX2	FXMADD2 C5, ALPHA, A5, C5	LFPDUX	A5,   X, INCX2	FXMADD2 C6, ALPHA, A6, C6	LFPDUX	A6,   X, INCX2	FXMADD2 C7, ALPHA, A7, C7	LFPDUX	A7,   X, INCX2	FXMADD2 C8, ALPHA, A8, C8	LFPDUX	A8,   X, INCX2	STFPDUX	C1,  YY, INCY2	STFPDUX	C2,  YY, INCY2	STFPDUX	C3,  YY, INCY2	STFPDUX	C4,  YY, INCY2	STFPDUX	C5,  YY, INCY2	STFPDUX	C6,  YY, INCY2	STFPDUX	C7,  YY, INCY2	STFPDUX	C8,  YY, INCY2	bdnz	LL(12)	.align 4LL(13):	FXMADD1 C1, ALPHA, A1, B1	FXMADD1 C2, ALPHA, A2, B2	FXMADD1 C3, ALPHA, A3, B3	FXMADD1 C4, ALPHA, A4, B4	FXMADD1 C5, ALPHA, A5, B5	FXMADD1 C6, ALPHA, A6, B6	FXMADD1 C7, ALPHA, A7, B7	FXMADD1 C8, ALPHA, A8, B8	FXMADD2 C1, ALPHA, A1, C1	FXMADD2 C2, ALPHA, A2, C2	FXMADD2 C3, ALPHA, A3, C3	FXMADD2 C4, ALPHA, A4, C4	FXMADD2 C5, ALPHA, A5, C5	FXMADD2 C6, ALPHA, A6, C6	STFPDUX	C1,  YY, INCY2	FXMADD2 C7, ALPHA, A7, C7	STFPDUX	C2,  YY, INCY2	FXMADD2 C8, ALPHA, A8, C8	STFPDUX	C3,  YY, INCY2	STFPDUX	C4,  YY, INCY2	STFPDUX	C5,  YY, INCY2	STFPDUX	C6,  YY, INCY2	STFPDUX	C7,  YY, INCY2	STFPDUX	C8,  YY, INCY2	.align 4LL(15):	andi.	r0,  N, 7	beq	LL(999)	andi.	r0,  N, 4	beq	LL(16)	LFPDUX	A1,   X, INCX2	LFPDUX	B1,   Y, INCY2	LFPDUX	A2,   X, INCX2	LFPDUX	B2,   Y, INCY2	LFPDUX	A3,   X, INCX2	LFPDUX	B3,   Y, INCY2	LFPDUX	A4,   X, INCX2	LFPDUX	B4,   Y, INCY2	FXMADD1 C1, ALPHA, A1, B1	FXMADD1 C2, ALPHA, A2, B2	FXMADD1 C3, ALPHA, A3, B3	FXMADD1 C4, ALPHA, A4, B4	FXMADD2 C1, ALPHA, A1, C1	FXMADD2 C2, ALPHA, A2, C2	FXMADD2 C3, ALPHA, A3, C3	FXMADD2 C4, ALPHA, A4, C4	STFPDUX	C1,  YY, INCY2	STFPDUX	C2,  YY, INCY2	STFPDUX	C3,  YY, INCY2	STFPDUX	C4,  YY, INCY2	.align 4LL(16):	andi.	r0,  N, 2	beq	LL(17)	LFPDUX	A1,   X, INCX2	LFPDUX	B1,   Y, INCY2	LFPDUX	A2,   X, INCX2	LFPDUX	B2,   Y, INCY2	FXMADD1 C1, ALPHA, A1, B1	FXMADD1 C2, ALPHA, A2, B2	FXMADD2 C1, ALPHA, A1, C1	FXMADD2 C2, ALPHA, A2, C2	STFPDUX	C1,  YY, INCY2	STFPDUX	C2,  YY, INCY2	.align 4LL(17):	andi.	r0,  N, 1	beq	LL(999)	LFPDUX	A1,   X, INCX2	LFPDUX	B1,   Y, INCY2	FXMADD1	C1, ALPHA, A1, B1	FXMADD2	C1, ALPHA, A1, C1	STFPDUX	C1,  YY, INCY2	b	LL(999)	.align 4LL(100):	fsmtp	ALPHA_I, ALPHA_R	sub	X,  X, INCX2	sub	Y,  Y, INCY2	addi	X1, X, SIZE	addi	Y1, Y, SIZE	mr	YY,  Y	mr	YY1, Y1	srawi.	r0, N, 2	mtspr	CTR,  r0	beq-	LL(115)	LFDUX	A1,   X,  INCX2	LFDUX	A2,   X1, INCX2	LFDUX	B1,   Y,  INCY2	LFDUX	B2,   Y1, INCY2	LFDUX	A3,   X,  INCX2	LFDUX	A4,   X1, INCX2	LFDUX	B3,   Y,  INCY2	LFDUX	B4,   Y1, INCY2	LFDUX	A5,   X,  INCX2	LFDUX	A6,   X1, INCX2	LFDUX	B5,   Y,  INCY2	LFDUX	B6,   Y1, INCY2	LFDUX	A7,   X,  INCX2	LFDUX	A8,   X1, INCX2	LFDUX	B7,   Y,  INCY2	LFDUX	B8,   Y1, INCY2	bdz	LL(113)	.align 4LL(112):	FMADD	C1,  ALPHA_R, A1, B1	LFDUX	B1,   Y,  INCY2	FMADD	C2,  ALPHA_I, A1, B2	LFDUX	A1,   X,  INCX2	FMADD	C3,  ALPHA_R, A3, B3	LFDUX	B3,   Y,  INCY2	FMADD	C4,  ALPHA_I, A3, B4	LFDUX	A3,   X,  INCX2	FMADD	C5,  ALPHA_R, A5, B5	LFDUX	B5,   Y,  INCY2	FMADD	C6,  ALPHA_I, A5, B6	LFDUX	A5,   X,  INCX2	FMADD	C7,  ALPHA_R, A7, B7	LFDUX	B7,   Y,  INCY2	FMADD	C8,  ALPHA_I, A7, B8	LFDUX	A7,   X,  INCX2	ADD1	C1,  ALPHA_I, A2, C1	LFDUX	B2,   Y1, INCY2	ADD2	C2,  ALPHA_R, A2, C2	LFDUX	A2,   X1, INCX2	ADD1	C3,  ALPHA_I, A4, C3	LFDUX	B4,   Y1, INCY2	ADD2	C4,  ALPHA_R, A4, C4	LFDUX	A4,   X1, INCX2	ADD1	C5,  ALPHA_I, A6, C5	LFDUX	B6,   Y1, INCY2	ADD2	C6,  ALPHA_R, A6, C6	LFDUX	A6,   X1, INCX2	ADD1	C7,  ALPHA_I, A8, C7	LFDUX	B8,   Y1, INCY2	ADD2	C8,  ALPHA_R, A8, C8	LFDUX	A8,   X1, INCX2	STFDUX	C1,  YY,  INCY2	STFDUX	C2,  YY1, INCY2	STFDUX	C3,  YY,  INCY2	STFDUX	C4,  YY1, INCY2	STFDUX	C5,  YY,  INCY2	STFDUX	C6,  YY1, INCY2	STFDUX	C7,  YY,  INCY2	STFDUX	C8,  YY1, INCY2	bdnz	LL(112)	.align 4LL(113):	FMADD	C1,  ALPHA_R, A1, B1	FMADD	C2,  ALPHA_I, A1, B2	FMADD	C3,  ALPHA_R, A3, B3	FMADD	C4,  ALPHA_I, A3, B4	FMADD	C5,  ALPHA_R, A5, B5	FMADD	C6,  ALPHA_I, A5, B6	FMADD	C7,  ALPHA_R, A7, B7	FMADD	C8,  ALPHA_I, A7, B8	ADD1	C1,  ALPHA_I, A2, C1	ADD2	C2,  ALPHA_R, A2, C2	ADD1	C3,  ALPHA_I, A4, C3	ADD2	C4,  ALPHA_R, A4, C4	ADD1	C5,  ALPHA_I, A6, C5	ADD2	C6,  ALPHA_R, A6, C6	STFDUX	C1,  YY,  INCY2	ADD1	C7,  ALPHA_I, A8, C7	STFDUX	C2,  YY1, INCY2	ADD2	C8,  ALPHA_R, A8, C8	STFDUX	C3,  YY,  INCY2	STFDUX	C4,  YY1, INCY2	STFDUX	C5,  YY,  INCY2	STFDUX	C6,  YY1, INCY2	STFDUX	C7,  YY,  INCY2	STFDUX	C8,  YY1, INCY2	.align 4LL(115):	andi.	r0,  N, 3	beq	LL(999)	andi.	r0,  N, 2	beq	LL(117)	LFDUX	A1,   X,  INCX2	LFDUX	A2,   X1, INCX2	LFDUX	B1,   Y,  INCY2	LFDUX	B2,   Y1, INCY2	LFDUX	A3,   X,  INCX2	FMADD	C1,  ALPHA_R, A1, B1	LFDUX	A4,   X1, INCX2	FMADD	C2,  ALPHA_I, A1, B2	LFDUX	B3,   Y,  INCY2	FMADD	C3,  ALPHA_R, A3, B3	LFDUX	B4,   Y1, INCY2	FMADD	C4,  ALPHA_I, A3, B4	ADD1	C1,  ALPHA_I, A2, C1	ADD2	C2,  ALPHA_R, A2, C2	STFDUX	C1,  YY,  INCY2	ADD1	C3,  ALPHA_I, A4, C3	STFDUX	C2,  YY1, INCY2	ADD2	C4,  ALPHA_R, A4, C4	STFDUX	C3,  YY,  INCY2	STFDUX	C4,  YY1, INCY2	.align 4LL(117):	andi.	r0,  N, 1	beq	LL(999)	LFDUX	A1,   X,  INCX2	LFDUX	A2,   X1, INCX2	LFDUX	B1,   Y,  INCY2	LFDUX	B2,   Y1, INCY2	FMADD	C1,  ALPHA_R, A1, B1	FMADD	C2,  ALPHA_I, A1, B2	ADD1	C1,  ALPHA_I, A2, C1	ADD2	C2,  ALPHA_R, A2, C2	STFDUX	C1,  YY,  INCY2	STFDUX	C2,  YY1, INCY2	.align 4LL(999):	li	r10, 16	subi	SP, SP, 16		lfpdux	f25, SP, r10	lfpdux	f24, SP, r10	lfpdux	f23, SP, r10	lfpdux	f22, SP, r10	lfpdux	f21, SP, r10	lfpdux	f20, SP, r10	lfpdux	f19, SP, r10	lfpdux	f18, SP, r10	lfpdux	f17, SP, r10	lfpdux	f16, SP, r10	lfpdux	f15, SP, r10	lfpdux	f14, SP, r10	addi	SP, SP,  16	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 + -