📄 izamax_sse2.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 M	ARG1	/* rdi */#define X	ARG2	/* rsi */#define INCX	ARG3	/* rdx */#define RET	%rax#define I	ARG4#define XX	%r10#define	MM	%r11	#ifdef MIN#define maxpd	minpd#define maxsd	minsd#endif	PROLOGUE	PROFCODE#ifdef F_INTERFACE#ifndef USE64BITINT	movslq	(M), M	movslq	(INCX), INCX#else	movslq	(M), M	movslq	(INCX), INCX#endif#endif	SAVEREGISTERS	pxor	%xmm0, %xmm0	xor	RET, RET	testq	M, M	jle	.L999	testq	INCX, INCX	jle	.L999	salq	$ZBASE_SHIFT, INCX	movq	M, MM	movq	X, XX	pcmpeqb	%xmm15, %xmm15	psrlq	$1, %xmm15	movsd	0 * SIZE(X), %xmm0	movsd	1 * SIZE(X), %xmm1	addq	INCX, X	decq	M	andpd	 %xmm15, %xmm0	andpd	 %xmm15, %xmm1	addpd	 %xmm1, %xmm0	unpcklpd %xmm0, %xmm0	movapd	 %xmm0, %xmm1	movapd	 %xmm0, %xmm2	movapd	 %xmm0, %xmm3	cmpq	$2 * SIZE, INCX	jne	.L60#define PREFETCHSIZE 256	movq	M,  I	sarq	$3, I	jle	.L25	ALIGN_4.L21:#if defined(OPTERON) || defined(BARCELONA)	prefetch	PREFETCHSIZE * SIZE(X)#endif#ifdef PENTIUM4	prefetchnta	PREFETCHSIZE * SIZE(X)#endif	movsd	0 * SIZE(X), %xmm4	movsd	1 * SIZE(X), %xmm5	movhpd	2 * SIZE(X), %xmm4	movhpd	3 * SIZE(X), %xmm5	movsd	4 * SIZE(X), %xmm6	movsd	5 * SIZE(X), %xmm7	movhpd	6 * SIZE(X), %xmm6	movhpd	7 * SIZE(X), %xmm7	andpd	%xmm15, %xmm4	andpd	%xmm15, %xmm5	andpd	%xmm15, %xmm6	andpd	%xmm15, %xmm7	addpd	%xmm5,  %xmm4	addpd	%xmm7,  %xmm6	maxpd	%xmm4, %xmm0	maxpd	%xmm6, %xmm1#if defined(OPTERON) || defined(BARCELONA)	prefetch	(PREFETCHSIZE + 8) * SIZE(X)#endif	movsd	 8 * SIZE(X), %xmm4	movsd	 9 * SIZE(X), %xmm5	movhpd	10 * SIZE(X), %xmm4	movhpd	11 * SIZE(X), %xmm5	movsd	12 * SIZE(X), %xmm6	movsd	13 * SIZE(X), %xmm7	movhpd	14 * SIZE(X), %xmm6	movhpd	15 * SIZE(X), %xmm7	andpd	%xmm15, %xmm4	andpd	%xmm15, %xmm5	andpd	%xmm15, %xmm6	andpd	%xmm15, %xmm7	addpd	%xmm5,  %xmm4	addpd	%xmm7,  %xmm6	maxpd	%xmm4, %xmm2	maxpd	%xmm6, %xmm3	addq	$16 * SIZE, X	decq	I	jg	.L21	ALIGN_4.L25:	andq	$7,  M	jle	.L30	testq	$4, M	je	.L26	movsd	0 * SIZE(X), %xmm4	movsd	1 * SIZE(X), %xmm5	movhpd	2 * SIZE(X), %xmm4	movhpd	3 * SIZE(X), %xmm5	movsd	4 * SIZE(X), %xmm6	movsd	5 * SIZE(X), %xmm7	movhpd	6 * SIZE(X), %xmm6	movhpd	7 * SIZE(X), %xmm7	andpd	%xmm15, %xmm4	andpd	%xmm15, %xmm5	andpd	%xmm15, %xmm6	andpd	%xmm15, %xmm7	addpd	%xmm5,  %xmm4	addpd	%xmm7,  %xmm6	maxpd	%xmm4, %xmm0	maxpd	%xmm6, %xmm1	addq	$8 * SIZE, X	ALIGN_3.L26:	testq	$2, M	je	.L27	movsd	0 * SIZE(X), %xmm4	movsd	1 * SIZE(X), %xmm5	movhpd	2 * SIZE(X), %xmm4	movhpd	3 * SIZE(X), %xmm5	addq	$4 * SIZE, X	andpd	%xmm15, %xmm4	andpd	%xmm15, %xmm5	addpd	%xmm5,  %xmm4	maxpd	%xmm4, %xmm0	ALIGN_3	.L27:	testq	$1, M	je	.L30	movsd	0 * SIZE(X), %xmm4	movsd	1 * SIZE(X), %xmm5	andpd	%xmm15, %xmm4	andpd	%xmm15, %xmm5	addpd	%xmm5, %xmm4	maxsd	%xmm4, %xmm2	ALIGN_4.L30:	movq	XX, X	movq	MM, M	maxpd	%xmm1, %xmm0	maxpd	%xmm3, %xmm2	maxpd	%xmm2, %xmm0	movapd	%xmm0, %xmm1	unpckhpd %xmm0, %xmm0	maxsd	%xmm1, %xmm0	unpcklpd %xmm0, %xmm0	movq	M,  I	sarq	$2, I	jle	.L35	ALIGN_4.L31:#if defined(OPTERON) || defined(BARCELONA)	prefetch	PREFETCHSIZE * SIZE(X)#endif#ifdef PENTIUM4	prefetchnta	PREFETCHSIZE * SIZE(X)#endif	movsd	0 * SIZE(X), %xmm1	movsd	1 * SIZE(X), %xmm2	movhpd	2 * SIZE(X), %xmm1	movhpd	3 * SIZE(X), %xmm2	movsd	4 * SIZE(X), %xmm3	movsd	5 * SIZE(X), %xmm4	movhpd	6 * SIZE(X), %xmm3	movhpd	7 * SIZE(X), %xmm4	andpd	%xmm15, %xmm1	andpd	%xmm15, %xmm2	andpd	%xmm15, %xmm3	andpd	%xmm15, %xmm4	addpd	%xmm2,  %xmm1	addpd	%xmm4,  %xmm3	cmpeqpd	%xmm0, %xmm1	cmpeqpd	%xmm0, %xmm3	orpd	%xmm3, %xmm1	movmskpd %xmm1, %r11	testq	 $3, %r11	jne	 .L33	addq	$8 * SIZE, X	addq	$4, RET	decq	I	jg	.L31	jmp	.L35	ALIGN_4.L33:	movsd	0 * SIZE(X), %xmm1	movsd	1 * SIZE(X), %xmm2	movsd	2 * SIZE(X), %xmm3	movsd	3 * SIZE(X), %xmm4	movsd	4 * SIZE(X), %xmm5	movsd	5 * SIZE(X), %xmm6	movsd	6 * SIZE(X), %xmm7	movsd	7 * SIZE(X), %xmm8	addq	$8 * SIZE, X	andpd	%xmm15, %xmm1	andpd	%xmm15, %xmm2	andpd	%xmm15, %xmm3	andpd	%xmm15, %xmm4	andpd	%xmm15, %xmm5	andpd	%xmm15, %xmm6	andpd	%xmm15, %xmm7	andpd	%xmm15, %xmm8	addpd	%xmm2,  %xmm1	addpd	%xmm4,  %xmm3	addpd	%xmm6,  %xmm5	addpd	%xmm8,  %xmm7	incq	RET	comisd	%xmm0, %xmm1	je	.L999	incq	RET	comisd	%xmm0, %xmm3	je	.L999	incq	RET	comisd	%xmm0, %xmm5	je	.L999	incq	RET	comisd	%xmm0, %xmm7	je	.L999	ALIGN_3.L35:	testq	$2, M	je	.L36	movsd	0 * SIZE(X), %xmm1	movsd	1 * SIZE(X), %xmm2	movsd	2 * SIZE(X), %xmm3	movsd	3 * SIZE(X), %xmm4	addq	$4 * SIZE, X	andpd	%xmm15, %xmm1	andpd	%xmm15, %xmm2	andpd	%xmm15, %xmm3	andpd	%xmm15, %xmm4	addpd	%xmm2,  %xmm1	addpd	%xmm4,  %xmm3	incq	RET	comisd	%xmm0, %xmm1	je	.L999	incq	RET	comisd	%xmm0, %xmm3	je	.L999	ALIGN_3	.L36:	incq	RET	jmp	.L999	ALIGN_3.L60:	movq	M,  I	sarq	$3, I	jle	.L65	ALIGN_4	.L61:#if defined(OPTERON) || defined(BARCELONA)	prefetch	PREFETCHSIZE * SIZE(X)#endif#ifdef PENTIUM4	prefetchnta	PREFETCHSIZE * SIZE(X)#endif	movsd	0 * SIZE(X), %xmm4	movsd	1 * SIZE(X), %xmm5	addq	INCX, X	movhpd	0 * SIZE(X), %xmm4	movhpd	1 * SIZE(X), %xmm5	addq	INCX, X	movsd	0 * SIZE(X), %xmm6	movsd	1 * SIZE(X), %xmm7	addq	INCX, X	movhpd	0 * SIZE(X), %xmm6	movhpd	1 * SIZE(X), %xmm7	addq	INCX, X	andpd	%xmm15, %xmm4	andpd	%xmm15, %xmm5	andpd	%xmm15, %xmm6	andpd	%xmm15, %xmm7	addpd	%xmm5, %xmm4	addpd	%xmm7, %xmm6	maxpd	%xmm4, %xmm0	maxpd	%xmm6, %xmm1#if defined(OPTERON) || defined(BARCELONA)	prefetch	PREFETCHSIZE * SIZE(X)#endif#ifdef PENTIUM4	prefetchnta	PREFETCHSIZE * SIZE(X)#endif	movsd	0 * SIZE(X), %xmm4	movsd	1 * SIZE(X), %xmm5	addq	INCX, X	movhpd	0 * SIZE(X), %xmm4	movhpd	1 * SIZE(X), %xmm5	addq	INCX, X	movsd	0 * SIZE(X), %xmm6	movsd	1 * SIZE(X), %xmm7	addq	INCX, X	movhpd	0 * SIZE(X), %xmm6	movhpd	1 * SIZE(X), %xmm7	addq	INCX, X	andpd	%xmm15, %xmm4	andpd	%xmm15, %xmm5	andpd	%xmm15, %xmm6	andpd	%xmm15, %xmm7	addpd	%xmm5, %xmm4	addpd	%xmm7, %xmm6	maxpd	%xmm4, %xmm2	maxpd	%xmm6, %xmm3	decq	I	jg	.L61	ALIGN_4.L65:	andq	$7,  M	jle	.L70	testq	$4, M	je	.L66	movsd	0 * SIZE(X), %xmm4	movsd	1 * SIZE(X), %xmm5	addq	INCX, X	movhpd	0 * SIZE(X), %xmm4	movhpd	1 * SIZE(X), %xmm5	addq	INCX, X	movsd	0 * SIZE(X), %xmm6	movsd	1 * SIZE(X), %xmm7	addq	INCX, X	movhpd	0 * SIZE(X), %xmm6	movhpd	1 * SIZE(X), %xmm7	addq	INCX, X	andpd	%xmm15, %xmm4	andpd	%xmm15, %xmm5	andpd	%xmm15, %xmm6	andpd	%xmm15, %xmm7	addpd	%xmm5, %xmm4	addpd	%xmm7, %xmm6	maxpd	%xmm4, %xmm0	maxpd	%xmm6, %xmm1	ALIGN_3.L66:	testq	$2, M	je	.L67	movsd	0 * SIZE(X), %xmm4	movsd	1 * SIZE(X), %xmm5	addq	INCX, X	movhpd	0 * SIZE(X), %xmm4	movhpd	1 * SIZE(X), %xmm5	addq	INCX, X	andpd	%xmm15, %xmm4	andpd	%xmm15, %xmm5	addpd	%xmm5, %xmm4	maxpd	%xmm4, %xmm2	ALIGN_3	.L67:	testq	$1, M	je	.L70	movsd	0 * SIZE(X), %xmm4	movsd	1 * SIZE(X), %xmm5	andpd	%xmm15, %xmm4	andpd	%xmm15, %xmm5	addpd	%xmm5, %xmm4	maxsd	%xmm4, %xmm3	ALIGN_3.L70:	movq	XX, X	movq	MM, M	maxpd	%xmm1, %xmm0	maxpd	%xmm3, %xmm2	maxpd	%xmm2, %xmm0	movapd	%xmm0, %xmm1	unpckhpd %xmm0, %xmm0	maxsd	%xmm1, %xmm0	unpcklpd %xmm0, %xmm0	movq	M,  I	sarq	$2, I	jle	.L75	ALIGN_4.L71:#if defined(OPTERON) || defined(BARCELONA)	prefetch	PREFETCHSIZE * SIZE(X)#endif#ifdef PENTIUM4	prefetchnta	PREFETCHSIZE * SIZE(X)#endif	movsd	0 * SIZE(X), %xmm1	movsd	1 * SIZE(X), %xmm2	addq	INCX, X	movhpd	0 * SIZE(X), %xmm1	movhpd	1 * SIZE(X), %xmm2	addq	INCX, X	movsd	0 * SIZE(X), %xmm3	movsd	1 * SIZE(X), %xmm4	addq	INCX, X	movhpd	0 * SIZE(X), %xmm3	movhpd	1 * SIZE(X), %xmm4	addq	INCX, X	andpd	%xmm15, %xmm1	andpd	%xmm15, %xmm2	andpd	%xmm15, %xmm3	andpd	%xmm15, %xmm4	addpd	%xmm2,  %xmm1	addpd	%xmm4,  %xmm3	cmpeqpd	%xmm0, %xmm1	cmpeqpd	%xmm0, %xmm3	orpd	%xmm3, %xmm1	movmskpd %xmm1, %r11	testq	 $3, %r11	jne	 .L73	addq	$4, RET	decq	I	jg	.L71	jmp	.L75	ALIGN_4.L73:	subq	INCX, X	movsd	0 * SIZE(X), %xmm7	movsd	1 * SIZE(X), %xmm8	subq	INCX, X	movsd	0 * SIZE(X), %xmm5	movsd	1 * SIZE(X), %xmm6	subq	INCX, X	movsd	0 * SIZE(X), %xmm3	movsd	1 * SIZE(X), %xmm4	subq	INCX, X	movsd	0 * SIZE(X), %xmm1	movsd	1 * SIZE(X), %xmm2	andpd	%xmm15, %xmm1	andpd	%xmm15, %xmm2	andpd	%xmm15, %xmm3	andpd	%xmm15, %xmm4	andpd	%xmm15, %xmm5	andpd	%xmm15, %xmm6	andpd	%xmm15, %xmm7	andpd	%xmm15, %xmm8	addpd	%xmm2,  %xmm1	addpd	%xmm4,  %xmm3	addpd	%xmm6,  %xmm5	addpd	%xmm8,  %xmm7	incq	RET	comisd	%xmm0, %xmm1	je	.L999	incq	RET	comisd	%xmm0, %xmm3	je	.L999	incq	RET	comisd	%xmm0, %xmm5	je	.L999	incq	RET	comisd	%xmm0, %xmm7	je	.L999	ALIGN_3.L75:	testq	$2, M	je	.L76	movsd	0 * SIZE(X), %xmm1	movsd	1 * SIZE(X), %xmm2	addq	INCX, X	movsd	0 * SIZE(X), %xmm3	movsd	1 * SIZE(X), %xmm4	addq	INCX, X	andpd	%xmm15, %xmm1	andpd	%xmm15, %xmm2	andpd	%xmm15, %xmm3	andpd	%xmm15, %xmm4	addpd	%xmm2,  %xmm1	addpd	%xmm4,  %xmm3	incq	RET	comisd	%xmm0, %xmm1	je	.L999	incq	RET	comisd	%xmm0, %xmm3	je	.L999	ALIGN_3	.L76:	incq	RET	ALIGN_4.L999:	RESTOREREGISTERS	ret	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 + -