Blame thirdparty/openblas/xianyi-OpenBLAS-e6e87a2/kernel/x86/zgemm_ncopy_2.S

kusano 2b45e8
/*********************************************************************/
kusano 2b45e8
/* Copyright 2009, 2010 The University of Texas at Austin.           */
kusano 2b45e8
/* All rights reserved.                                              */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/* Redistribution and use in source and binary forms, with or        */
kusano 2b45e8
/* without modification, are permitted provided that the following   */
kusano 2b45e8
/* conditions are met:                                               */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/*   1. Redistributions of source code must retain the above         */
kusano 2b45e8
/*      copyright notice, this list of conditions and the following  */
kusano 2b45e8
/*      disclaimer.                                                  */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/*   2. Redistributions in binary form must reproduce the above      */
kusano 2b45e8
/*      copyright notice, this list of conditions and the following  */
kusano 2b45e8
/*      disclaimer in the documentation and/or other materials       */
kusano 2b45e8
/*      provided with the distribution.                              */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
kusano 2b45e8
/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
kusano 2b45e8
/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
kusano 2b45e8
/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
kusano 2b45e8
/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
kusano 2b45e8
/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
kusano 2b45e8
/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
kusano 2b45e8
/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
kusano 2b45e8
/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
kusano 2b45e8
/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
kusano 2b45e8
/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
kusano 2b45e8
/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
kusano 2b45e8
/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
kusano 2b45e8
/*    POSSIBILITY OF SUCH DAMAGE.                                    */
kusano 2b45e8
/*                                                                   */
kusano 2b45e8
/* The views and conclusions contained in the software and           */
kusano 2b45e8
/* documentation are those of the authors and should not be          */
kusano 2b45e8
/* interpreted as representing official policies, either expressed   */
kusano 2b45e8
/* or implied, of The University of Texas at Austin.                 */
kusano 2b45e8
/*********************************************************************/
kusano 2b45e8
kusano 2b45e8
#define ASSEMBLER
kusano 2b45e8
#include "common.h"
kusano 2b45e8
kusano 2b45e8
#define STACK	16
kusano 2b45e8
#define ARGS	 8
kusano 2b45e8
	
kusano 2b45e8
#define STACK_M	 4 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_N	 8 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_A	12 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_LDA	16 + STACK + ARGS(%esp)
kusano 2b45e8
#define STACK_B	20 + STACK + ARGS(%esp)
kusano 2b45e8
kusano 2b45e8
#define I	%eax
kusano 2b45e8
#define J	%ecx
kusano 2b45e8
#define LDA	%edx
kusano 2b45e8
#define A	%edi
kusano 2b45e8
#define A1	%ebx
kusano 2b45e8
#define A2	%ebp
kusano 2b45e8
#define B	%esi
kusano 2b45e8
kusano 2b45e8
	PROLOGUE
kusano 2b45e8
kusano 2b45e8
	subl	$ARGS, %esp
kusano 2b45e8
	pushl	%ebp
kusano 2b45e8
	pushl	%edi
kusano 2b45e8
	pushl	%esi
kusano 2b45e8
	pushl	%ebx
kusano 2b45e8
kusano 2b45e8
	PROFCODE
kusano 2b45e8
kusano 2b45e8
	EMMS
kusano 2b45e8
kusano 2b45e8
	movl	STACK_A,   A
kusano 2b45e8
	movl	STACK_LDA, LDA
kusano 2b45e8
	movl	STACK_B, B
kusano 2b45e8
	sall	$ZBASE_SHIFT, LDA
kusano 2b45e8
kusano 2b45e8
	movl	STACK_N,  J
kusano 2b45e8
	sarl	$1, J
kusano 2b45e8
	je	.L20
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L21:
kusano 2b45e8
	movl	A, A1
kusano 2b45e8
	leal    (A1, LDA), A2
kusano 2b45e8
	leal	(A,  LDA, 2), A
kusano 2b45e8
kusano 2b45e8
	movl	STACK_M,   I
kusano 2b45e8
	sarl	$1,  I
kusano 2b45e8
	je	.L24
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L25:
kusano 2b45e8
#ifdef HAVE_MMX
kusano 2b45e8
	MMXLOAD		0 * SIZE(A1), %mm0
kusano 2b45e8
	MMXLOAD		1 * SIZE(A1), %mm1
kusano 2b45e8
	MMXLOAD		0 * SIZE(A2), %mm2
kusano 2b45e8
	MMXLOAD		1 * SIZE(A2), %mm3
kusano 2b45e8
kusano 2b45e8
	MMXLOAD		2 * SIZE(A1), %mm4
kusano 2b45e8
	MMXLOAD		3 * SIZE(A1), %mm5
kusano 2b45e8
	MMXLOAD		2 * SIZE(A2), %mm6
kusano 2b45e8
	MMXLOAD		3 * SIZE(A2), %mm7
kusano 2b45e8
kusano 2b45e8
	MMXSTORE	%mm0, 0 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm1, 1 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm2, 2 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm3, 3 * SIZE(B)
kusano 2b45e8
kusano 2b45e8
	MMXSTORE	%mm4, 4 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm5, 5 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm6, 6 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm7, 7 * SIZE(B)
kusano 2b45e8
#else
kusano 2b45e8
	FLD	3 * SIZE(A2)
kusano 2b45e8
	FLD	2 * SIZE(A2)
kusano 2b45e8
	FLD	3 * SIZE(A1)
kusano 2b45e8
	FLD	2 * SIZE(A1)
kusano 2b45e8
	FLD	1 * SIZE(A2)
kusano 2b45e8
	FLD	0 * SIZE(A2)
kusano 2b45e8
	FLD	1 * SIZE(A1)
kusano 2b45e8
	FLD	0 * SIZE(A1)
kusano 2b45e8
kusano 2b45e8
	FST	0 * SIZE(B)
kusano 2b45e8
	FST	1 * SIZE(B)
kusano 2b45e8
	FST	2 * SIZE(B)
kusano 2b45e8
	FST	3 * SIZE(B)
kusano 2b45e8
	FST	4 * SIZE(B)
kusano 2b45e8
	FST	5 * SIZE(B)
kusano 2b45e8
	FST	6 * SIZE(B)
kusano 2b45e8
	FST	7 * SIZE(B)
kusano 2b45e8
#endif
kusano 2b45e8
	addl	$4 * SIZE, A1
kusano 2b45e8
	addl	$4 * SIZE, A2
kusano 2b45e8
	addl	$8 * SIZE, B
kusano 2b45e8
	decl	I
kusano 2b45e8
	jne	.L25
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L24:
kusano 2b45e8
	movl	STACK_M,  I
kusano 2b45e8
	andl	$1, I
kusano 2b45e8
	jle	.L30
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L31:
kusano 2b45e8
#ifdef HAVE_MMX
kusano 2b45e8
	MMXLOAD		0 * SIZE(A1), %mm0
kusano 2b45e8
	MMXLOAD		1 * SIZE(A1), %mm1
kusano 2b45e8
	MMXLOAD		0 * SIZE(A2), %mm2
kusano 2b45e8
	MMXLOAD		1 * SIZE(A2), %mm3
kusano 2b45e8
	MMXSTORE	%mm0, 0 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm1, 1 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm2, 2 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm3, 3 * SIZE(B)
kusano 2b45e8
#else
kusano 2b45e8
	FLD	1 * SIZE(A2)
kusano 2b45e8
	FLD	0 * SIZE(A2)
kusano 2b45e8
	FLD	1 * SIZE(A1)
kusano 2b45e8
	FLD	0 * SIZE(A1)
kusano 2b45e8
	FST	0 * SIZE(B)
kusano 2b45e8
	FST	1 * SIZE(B)
kusano 2b45e8
	FST	2 * SIZE(B)
kusano 2b45e8
	FST	3 * SIZE(B)
kusano 2b45e8
#endif
kusano 2b45e8
	addl	$2 * SIZE, A1
kusano 2b45e8
	addl	$2 * SIZE, A2
kusano 2b45e8
	addl	$4 * SIZE, B
kusano 2b45e8
	decl	I
kusano 2b45e8
	jne	.L31
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L30:
kusano 2b45e8
	decl	J
kusano 2b45e8
	jne	.L21
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L20:
kusano 2b45e8
	movl	A, A1
kusano 2b45e8
	movl	STACK_N, J
kusano 2b45e8
	andl	$1, J
kusano 2b45e8
	jle	.L38
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L39:
kusano 2b45e8
	movl	STACK_M,  I
kusano 2b45e8
	sarl	$2, I
kusano 2b45e8
	je	.L42
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L43:
kusano 2b45e8
#ifdef HAVE_MMX
kusano 2b45e8
	MMXLOAD		0 * SIZE(A1), %mm0
kusano 2b45e8
	MMXLOAD		1 * SIZE(A1), %mm1
kusano 2b45e8
	MMXLOAD		2 * SIZE(A1), %mm2
kusano 2b45e8
	MMXLOAD		3 * SIZE(A1), %mm3
kusano 2b45e8
	MMXLOAD		4 * SIZE(A1), %mm4
kusano 2b45e8
	MMXLOAD		5 * SIZE(A1), %mm5
kusano 2b45e8
	MMXLOAD		6 * SIZE(A1), %mm6
kusano 2b45e8
	MMXLOAD		7 * SIZE(A1), %mm7
kusano 2b45e8
kusano 2b45e8
	MMXSTORE	%mm0,  0 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm1,  1 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm2,  2 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm3,  3 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm4,  4 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm5,  5 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm6,  6 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm7,  7 * SIZE(B)
kusano 2b45e8
#else
kusano 2b45e8
	FLD	 7 * SIZE(A1)
kusano 2b45e8
	FLD	 6 * SIZE(A1)
kusano 2b45e8
	FLD	 5 * SIZE(A1)
kusano 2b45e8
	FLD	 4 * SIZE(A1)
kusano 2b45e8
	FLD	 3 * SIZE(A1)
kusano 2b45e8
	FLD	 2 * SIZE(A1)
kusano 2b45e8
	FLD	 1 * SIZE(A1)
kusano 2b45e8
	FLD	 0 * SIZE(A1)
kusano 2b45e8
kusano 2b45e8
	FST	 0 * SIZE(B)
kusano 2b45e8
	FST	 1 * SIZE(B)
kusano 2b45e8
	FST	 2 * SIZE(B)
kusano 2b45e8
	FST	 3 * SIZE(B)
kusano 2b45e8
	FST	 4 * SIZE(B)
kusano 2b45e8
	FST	 5 * SIZE(B)
kusano 2b45e8
	FST	 6 * SIZE(B)
kusano 2b45e8
	FST	 7 * SIZE(B)
kusano 2b45e8
#endif
kusano 2b45e8
kusano 2b45e8
	addl	$8 * SIZE, A1
kusano 2b45e8
	addl	$8 * SIZE, B
kusano 2b45e8
	decl	I
kusano 2b45e8
	jne	.L43
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L42:
kusano 2b45e8
	movl	STACK_M,  I
kusano 2b45e8
	andl	$3, I
kusano 2b45e8
	jle	.L38
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L49:
kusano 2b45e8
#ifdef HAVE_MMX
kusano 2b45e8
	MMXLOAD		0 * SIZE(A1), %mm0
kusano 2b45e8
	MMXLOAD		1 * SIZE(A1), %mm1
kusano 2b45e8
	MMXSTORE	%mm0,  0 * SIZE(B)
kusano 2b45e8
	MMXSTORE	%mm1,  1 * SIZE(B)
kusano 2b45e8
#else
kusano 2b45e8
	FLD	1 * SIZE(A1)
kusano 2b45e8
	FLD	0 * SIZE(A1)
kusano 2b45e8
	FST	0 * SIZE(B)
kusano 2b45e8
	FST	1 * SIZE(B)
kusano 2b45e8
#endif
kusano 2b45e8
	addl	$2 * SIZE, A1
kusano 2b45e8
	addl	$2 * SIZE, B
kusano 2b45e8
	decl	I
kusano 2b45e8
	jne	.L49
kusano 2b45e8
	ALIGN_3
kusano 2b45e8
kusano 2b45e8
.L38:
kusano 2b45e8
	EMMS
kusano 2b45e8
kusano 2b45e8
	popl	%ebx
kusano 2b45e8
	popl	%esi
kusano 2b45e8
	popl	%edi
kusano 2b45e8
	popl	%ebp
kusano 2b45e8
	addl	$ARGS, %esp
kusano 2b45e8
	ret
kusano 2b45e8
kusano 2b45e8
	EPILOGUE