📄 izamax_sse.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 maxps	minps#define maxss	minss#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	psrld	$1, %xmm15	movss	0 * SIZE(X), %xmm0	movss	1 * SIZE(X), %xmm1	addq	INCX, X	decq	M	andps	%xmm15, %xmm0	andps	%xmm15, %xmm1	addps	%xmm1,  %xmm0	shufps	$0, %xmm0, %xmm0	movaps	%xmm0, %xmm1	cmpq	$2 * SIZE, INCX	jne	.L70.L30:	movq	M,  I	sarq	$3, I	jle	.L35	ALIGN_4	#define PREFETCHSIZE 256.L31:#if defined(OPTERON) || defined(BARCELONA)	prefetch	PREFETCHSIZE * SIZE(X)#endif#ifdef PENTIUM4	prefetcht2	PREFETCHSIZE * SIZE(X)#endif	movsd	0 * SIZE(X), %xmm4	movhps	2 * SIZE(X), %xmm4	movsd	4 * SIZE(X), %xmm5	movhps	6 * SIZE(X), %xmm5	movaps	%xmm4, %xmm6	shufps	$0x88, %xmm5, %xmm4	shufps	$0xdd, %xmm5, %xmm6	andps	%xmm15, %xmm4	andps	%xmm15, %xmm6	addps	%xmm6,  %xmm4	maxps	%xmm4,  %xmm0	movsd	 8 * SIZE(X), %xmm7	movhps	10 * SIZE(X), %xmm7	movsd	12 * SIZE(X), %xmm8	movhps	14 * SIZE(X), %xmm8	movaps	%xmm7, %xmm9	shufps	$0x88, %xmm8, %xmm7	shufps	$0xdd, %xmm8, %xmm9	andps	%xmm15, %xmm7	andps	%xmm15, %xmm9	addps	%xmm9,  %xmm7	maxps	%xmm7,  %xmm0	addq	$16 * SIZE, X	decq	I	jg	.L31	ALIGN_4.L35:	andq	$7,  M	jle	.L40	testq	$4, M	je	.L36	movsd	0 * SIZE(X), %xmm4	movhps	2 * SIZE(X), %xmm4	movsd	4 * SIZE(X), %xmm5	movhps	6 * SIZE(X), %xmm5	movaps	%xmm4, %xmm6	shufps	$0x88, %xmm5, %xmm4	shufps	$0xdd, %xmm5, %xmm6	andps	%xmm15, %xmm4	andps	%xmm15, %xmm6	addps	%xmm6,  %xmm4	maxps	%xmm4,  %xmm0	addq	$8 * SIZE, X	ALIGN_3.L36:	testq	$2, M	je	.L37	movss	0 * SIZE(X), %xmm4	movss	1 * SIZE(X), %xmm5	movss	2 * SIZE(X), %xmm6	movss	3 * SIZE(X), %xmm7	andps	%xmm15, %xmm4	andps	%xmm15, %xmm5	andps	%xmm15, %xmm6	andps	%xmm15, %xmm7	addps	%xmm5,  %xmm4	addps	%xmm7,  %xmm6	maxss	%xmm4,  %xmm0	maxss	%xmm6,  %xmm1	addq	$4 * SIZE, X	ALIGN_3	.L37:	testq	$1, M	je	.L40	movss	0 * SIZE(X), %xmm4	movss	1 * SIZE(X), %xmm5	andps	%xmm15, %xmm4	andps	%xmm15, %xmm5	addps	%xmm5,  %xmm4	maxss	%xmm4,  %xmm0	ALIGN_4.L40:	movq	XX, X	movq	MM, M	maxps	%xmm1, %xmm0	movaps	%xmm0, %xmm1	movhlps %xmm0, %xmm0	maxps	%xmm1, %xmm0	movaps	%xmm0, %xmm1	shufps	$1, %xmm0, %xmm0	maxss	%xmm1, %xmm0	shufps	$0, %xmm0, %xmm0	movq	M,  I	sarq	$2, I	jle	.L45	ALIGN_4	.L41:#if defined(OPTERON) || defined(BARCELONA)	prefetch	PREFETCHSIZE * SIZE(X)#endif#ifdef PENTIUM4	prefetcht2	PREFETCHSIZE * SIZE(X)#endif	movsd	0 * SIZE(X), %xmm1	movhps	2 * SIZE(X), %xmm1	movsd	4 * SIZE(X), %xmm2	movhps	6 * SIZE(X), %xmm2	movaps	%xmm1, %xmm3	shufps	$0x88, %xmm2, %xmm1	shufps	$0xdd, %xmm2, %xmm3	andps	%xmm15, %xmm1	andps	%xmm15, %xmm3	addps	%xmm3,  %xmm1	cmpeqps	%xmm0, %xmm1	movmskps %xmm1, %r11	testq	 $15, %r11	jne	 .L43	addq	$8 * SIZE, X	addq	$4, RET	decq	I	jg	.L41	jmp	.L45	ALIGN_4.L43:	movss	0 * SIZE(X), %xmm1	movss	1 * SIZE(X), %xmm2	movss	2 * SIZE(X), %xmm3	movss	3 * SIZE(X), %xmm4	movss	4 * SIZE(X), %xmm5	movss	5 * SIZE(X), %xmm6	movss	6 * SIZE(X), %xmm7	movss	7 * SIZE(X), %xmm8	addq	$8 * SIZE, X	andps	%xmm15, %xmm1	andps	%xmm15, %xmm2	andps	%xmm15, %xmm3	andps	%xmm15, %xmm4	andps	%xmm15, %xmm5	andps	%xmm15, %xmm6	andps	%xmm15, %xmm7	andps	%xmm15, %xmm8	addps	%xmm2,  %xmm1	addps	%xmm4,  %xmm3	addps	%xmm6,  %xmm5	addps	%xmm8,  %xmm7	incq	RET	comiss	%xmm0, %xmm1	je	.L999	incq	RET	comiss	%xmm0, %xmm3	je	.L999	incq	RET	comiss	%xmm0, %xmm5	je	.L999	incq	RET	comiss	%xmm0, %xmm7	je	.L999	ALIGN_3.L45:	testq	$2, M	je	.L47	movss	0 * SIZE(X), %xmm1	movss	1 * SIZE(X), %xmm2	movss	2 * SIZE(X), %xmm3	movss	3 * SIZE(X), %xmm4	addq	$4 * SIZE, X	andps	%xmm15, %xmm1	andps	%xmm15, %xmm2	andps	%xmm15, %xmm3	andps	%xmm15, %xmm4	addps	%xmm2,  %xmm1	addps	%xmm4,  %xmm3	incq	RET	comiss	%xmm0, %xmm1	je	.L999	incq	RET	comiss	%xmm0, %xmm3	je	.L999	ALIGN_3	.L47:	incq	RET	jmp	.L999	ALIGN_3.L70:	movq	M,  I	sarq	$3, I	jle	.L75	ALIGN_4	.L71:#if defined(OPTERON) || defined(BARCELONA)	prefetch	PREFETCHSIZE * SIZE(X)#endif#ifdef PENTIUM4	prefetcht2	PREFETCHSIZE * SIZE(X)#endif	movsd	0 * SIZE(X), %xmm4	addq	INCX, X	movhps	0 * SIZE(X), %xmm4	addq	INCX, X	movsd	0 * SIZE(X), %xmm5	addq	INCX, X	movhps	0 * SIZE(X), %xmm5	addq	INCX, X	movaps	%xmm4, %xmm6	shufps	$0x88, %xmm5, %xmm4	shufps	$0xdd, %xmm5, %xmm6	andps	%xmm15, %xmm4	andps	%xmm15, %xmm6	addps	%xmm6,  %xmm4	maxps	%xmm4,  %xmm0	movsd	 0 * SIZE(X), %xmm7	addq	INCX, X	movhps	 0 * SIZE(X), %xmm7	addq	INCX, X	movsd	 0 * SIZE(X), %xmm8	addq	INCX, X	movhps	 0 * SIZE(X), %xmm8	addq	INCX, X	movaps	%xmm7, %xmm9	shufps	$0x88, %xmm8, %xmm7	shufps	$0xdd, %xmm8, %xmm9	andps	%xmm15, %xmm7	andps	%xmm15, %xmm9	addps	%xmm9,  %xmm7	maxps	%xmm7,  %xmm0	decq	I	jg	.L71	ALIGN_4.L75:	andq	$7,  M	jle	.L80	testq	$4, M	je	.L76	movsd	0 * SIZE(X), %xmm4	addq	INCX, X	movhps	0 * SIZE(X), %xmm4	addq	INCX, X	movsd	0 * SIZE(X), %xmm5	addq	INCX, X	movhps	0 * SIZE(X), %xmm5	addq	INCX, X	movaps	%xmm4, %xmm6	shufps	$0x88, %xmm5, %xmm4	shufps	$0xdd, %xmm5, %xmm6	andps	%xmm15, %xmm4	andps	%xmm15, %xmm6	addps	%xmm6,  %xmm4	maxps	%xmm4,  %xmm0	ALIGN_3	.L76:	testq	$2, M	je	.L77	movss	0 * SIZE(X), %xmm4	movss	1 * SIZE(X), %xmm5	addq	INCX, X	movss	0 * SIZE(X), %xmm6	movss	1 * SIZE(X), %xmm7	addq	INCX, X	andps	%xmm15, %xmm4	andps	%xmm15, %xmm5	andps	%xmm15, %xmm6	andps	%xmm15, %xmm7	addps	%xmm5,  %xmm4	addps	%xmm7,  %xmm6	maxss	%xmm4,  %xmm0	maxss	%xmm6,  %xmm1	ALIGN_3	.L77:	testq	$1, M	je	.L80	movss	0 * SIZE(X), %xmm4	movss	1 * SIZE(X), %xmm5	andps	%xmm15, %xmm4	andps	%xmm15, %xmm5	addps	%xmm5,  %xmm4	maxss	%xmm4,  %xmm0	ALIGN_4.L80:	movq	XX, X	movq	MM, M	maxps	%xmm1, %xmm0	movaps	%xmm0, %xmm1	movhlps %xmm0, %xmm0	maxps	%xmm1, %xmm0	movaps	%xmm0, %xmm1	shufps	$1, %xmm0, %xmm0	maxss	%xmm1, %xmm0	shufps	$0, %xmm0, %xmm0	movq	M,  I	sarq	$2, I	jle	.L85	ALIGN_4	.L81:#if defined(OPTERON) || defined(BARCELONA)	prefetch	PREFETCHSIZE * SIZE(X)#endif#ifdef PENTIUM4	prefetcht2	PREFETCHSIZE * SIZE(X)#endif	movsd	0 * SIZE(X), %xmm1	addq	INCX, X	movhps	0 * SIZE(X), %xmm1	addq	INCX, X	movsd	0 * SIZE(X), %xmm2	addq	INCX, X	movhps	0 * SIZE(X), %xmm2	addq	INCX, X	movaps	%xmm1, %xmm3	shufps	$0x88, %xmm2, %xmm1	shufps	$0xdd, %xmm2, %xmm3	andps	%xmm15, %xmm1	andps	%xmm15, %xmm3	addps	%xmm3,  %xmm1	cmpeqps	%xmm0, %xmm1	movmskps %xmm1, %r11	testq	 $15, %r11	jne	 .L83	addq	$4, RET	decq	I	jg	.L81	jmp	.L85	ALIGN_4.L83:	subq	INCX, X	movss	0 * SIZE(X), %xmm7	movss	1 * SIZE(X), %xmm8	subq	INCX, X	movss	0 * SIZE(X), %xmm5	movss	1 * SIZE(X), %xmm6	subq	INCX, X	movss	0 * SIZE(X), %xmm3	movss	1 * SIZE(X), %xmm4	subq	INCX, X	movss	0 * SIZE(X), %xmm1	movss	1 * SIZE(X), %xmm2	andps	%xmm15, %xmm1	andps	%xmm15, %xmm2	andps	%xmm15, %xmm3	andps	%xmm15, %xmm4	andps	%xmm15, %xmm5	andps	%xmm15, %xmm6	andps	%xmm15, %xmm7	andps	%xmm15, %xmm8	addps	%xmm2,  %xmm1	addps	%xmm4,  %xmm3	addps	%xmm6,  %xmm5	addps	%xmm8,  %xmm7	incq	RET	comiss	%xmm0, %xmm1	je	.L999	incq	RET	comiss	%xmm0, %xmm3	je	.L999	incq	RET	comiss	%xmm0, %xmm5	je	.L999	incq	RET	comiss	%xmm0, %xmm7	je	.L999	ALIGN_3.L85:	testq	$2, M	je	.L87	movss	0 * SIZE(X), %xmm1	movss	1 * SIZE(X), %xmm2	addq	INCX, X	movss	0 * SIZE(X), %xmm3	movss	1 * SIZE(X), %xmm4	addq	INCX, X	andps	%xmm15, %xmm1	andps	%xmm15, %xmm2	andps	%xmm15, %xmm3	andps	%xmm15, %xmm4	addps	%xmm2,  %xmm1	addps	%xmm4,  %xmm3	incq	RET	comiss	%xmm0, %xmm1	je	.L999	incq	RET	comiss	%xmm0, %xmm3	je	.L999	ALIGN_3	.L87:	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 + -