OpenBLAS/kernel/x86_64/gemm_kernel_4x2_atom.S

/*********************************************************************/
/* Copyright 2009, 2010 The University of Texas at Austin.           */
/* All rights reserved.                                              */
/*                                                                   */
/* Redistribution and use in source and binary forms, with or        */
/* without modification, are permitted provided that the following   */
/* conditions are met:                                               */
/*                                                                   */
/*   1. Redistributions of source code must retain the above         */
/*      copyright notice, this list of conditions and the following  */
/*      disclaimer.                                                  */
/*                                                                   */
/*   2. Redistributions in binary form must reproduce the above      */
/*      copyright notice, this list of conditions and the following  */
/*      disclaimer in the documentation and/or other materials       */
/*      provided with the distribution.                              */
/*                                                                   */
/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
/*    POSSIBILITY OF SUCH DAMAGE.                                    */
/*                                                                   */
/* The views and conclusions contained in the software and           */
/* documentation are those of the authors and should not be          */
/* interpreted as representing official policies, either expressed   */
/* or implied, of The University of Texas at Austin.                 */
/*********************************************************************/

#define ASSEMBLER
#include "common.h"

#define M	%rdi
#define N	%rsi
#define K	%rdx

#define A	%rcx
#define B	%r8
#define C	%r9
#define LDC	%r10

#define I	%r11
#define J	%r12
#define AO	%r13
#define BO	%r14
#define	CO1	%r15
#define CO2	%rbx
#define BB	%rbp

#ifndef WINDOWS_ABI

#define STACKSIZE 128

#define OLD_LDC		 8 + STACKSIZE(%rsp)
#define OLD_OFFSET	16 + STACKSIZE(%rsp)

#define ALPHA	 48(%rsp)
#define OFFSET	 56(%rsp)
#define KKK	 64(%rsp)
#define KK	 72(%rsp)

#else

#define STACKSIZE 256

#define OLD_A		40 + STACKSIZE(%rsp)
#define OLD_B		48 + STACKSIZE(%rsp)
#define OLD_C		56 + STACKSIZE(%rsp)
#define OLD_LDC		64 + STACKSIZE(%rsp)
#define OLD_OFFSET	72 + STACKSIZE(%rsp)

#define ALPHA	 224(%rsp)
#define OFFSET	 232(%rsp)
#define KK	 240(%rsp)
#define KKK	 248(%rsp)

#endif

#define PREFETCH     prefetcht0
#define PREFETCHSIZE (8 * 8 + 3)

	PROLOGUE
	PROFCODE

	subq	$STACKSIZE, %rsp
	movq	%rbx,  0(%rsp)
	movq	%rbp,  8(%rsp)
	movq	%r12, 16(%rsp)
	movq	%r13, 24(%rsp)
	movq	%r14, 32(%rsp)
	movq	%r15, 40(%rsp)

#ifdef WINDOWS_ABI
	movq	%rdi,    48(%rsp)
	movq	%rsi,    56(%rsp)
	movups	%xmm6,   64(%rsp)
	movups	%xmm7,   80(%rsp)
	movups	%xmm8,   96(%rsp)
	movups	%xmm9,  112(%rsp)
	movups	%xmm10, 128(%rsp)
	movups	%xmm11, 144(%rsp)
	movups	%xmm12, 160(%rsp)
	movups	%xmm13, 176(%rsp)
	movups	%xmm14, 192(%rsp)
	movups	%xmm15, 208(%rsp)

	movq	ARG1,      M
	movq	ARG2,      N
	movq	ARG3,      K
	movq	OLD_A,     A
	movq	OLD_B,     B
	movq	OLD_C,     C
	movq	OLD_LDC,   LDC
#ifdef TRMMKERNEL
	movsd	OLD_OFFSET, %xmm4
#endif
	movaps	%xmm3, %xmm0

#else
	movq	OLD_LDC,   LDC
#ifdef TRMMKERNEL
	movsd	OLD_OFFSET, %xmm4
#endif

#endif

	movsd	%xmm0, ALPHA

#ifdef TRMMKERNEL
	movsd	%xmm4, OFFSET
	movsd	%xmm4, KK
#ifndef LEFT
	negq	KK
#endif
#endif

	leaq	(, LDC, SIZE), LDC

	movq	N,  J
	sarq	$1, J
	jle	.L40
	ALIGN_4

.L10:
#if defined(TRMMKERNEL) && defined(LEFT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif

	movq	C, CO1
	leaq	(C, LDC, 1), CO2
	leaq	(C, LDC, 2), C

	movq	A, AO

	movq	K, %rax
	salq	$BASE_SHIFT + 1, %rax
	leaq	(B, %rax), BB

	movq	M,  I
	sarq	$2, I
	jle	.L20
	ALIGN_4

.L11:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(B,  %rax, 2), BO
#endif

	prefetcht0	  0 * SIZE(BB)
	subq	   $-8 * SIZE, BB

	movsd	 0 * SIZE(AO), %xmm0
	xorps	%xmm2,  %xmm2
	movsd	 1 * SIZE(AO), %xmm4
	xorps	%xmm5,  %xmm5
	movsd	 2 * SIZE(AO), %xmm5
	xorps	%xmm6,  %xmm6
	xorps	%xmm7,  %xmm7

	movsd	 0 * SIZE(BO), %xmm1
	xorps	%xmm8,  %xmm8
	xorps	%xmm9,  %xmm9
	movsd	 1 * SIZE(BO), %xmm3
	xorps	%xmm10, %xmm10
	xorps	%xmm11, %xmm11

	prefetcht0     3 * SIZE(CO1)
	xorps	%xmm12, %xmm12
	xorps	%xmm13, %xmm13
	prefetcht0     3 * SIZE(CO2)
	xorps	%xmm14, %xmm14
	xorps	%xmm15, %xmm15

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$4, %rax
#else
	addq	$2, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	je	.L15
	ALIGN_4

.L12:
	PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
	addsd	 %xmm2, %xmm13
	movaps	 %xmm0, %xmm2
	mulsd	 %xmm1, %xmm0

	addsd	 %xmm7, %xmm14
	movsd	 3 * SIZE(AO), %xmm7
	mulsd	 %xmm3, %xmm2

	addsd	 %xmm6, %xmm15
	movaps	 %xmm4, %xmm6
	mulsd	 %xmm1, %xmm4

	addsd	 %xmm0, %xmm8
	movsd	 4 * SIZE(AO), %xmm0
	mulsd	 %xmm3, %xmm6

	addsd	 %xmm2, %xmm9
	movaps	 %xmm5, %xmm2
	mulsd	 %xmm1, %xmm5

	addsd	 %xmm4, %xmm10
	movsd	 5 * SIZE(AO), %xmm4
	mulsd	 %xmm3, %xmm2

	addsd	 %xmm6, %xmm11
	movaps	 %xmm7, %xmm6
	mulsd	 %xmm1, %xmm7
	movsd	 2 * SIZE(BO), %xmm1

	addsd	 %xmm5, %xmm12
	movsd	 6 * SIZE(AO), %xmm5
	mulsd	 %xmm3, %xmm6
	movsd	 3 * SIZE(BO), %xmm3

	addsd	 %xmm2, %xmm13
	movaps	 %xmm0, %xmm2
	mulsd	 %xmm1, %xmm0

	addsd	 %xmm7, %xmm14
	movsd	 7 * SIZE(AO), %xmm7
	mulsd	 %xmm3, %xmm2

	addsd	 %xmm6, %xmm15
	movaps	 %xmm4, %xmm6
	mulsd	 %xmm1, %xmm4

	addsd	 %xmm0, %xmm8
	movsd	 8 * SIZE(AO), %xmm0
	mulsd	 %xmm3, %xmm6

	addsd	 %xmm2, %xmm9
	movaps	 %xmm5, %xmm2
	mulsd	 %xmm1, %xmm5

	addsd	 %xmm4, %xmm10
	movsd	 9 * SIZE(AO), %xmm4
	mulsd	 %xmm3, %xmm2

	addsd	 %xmm6, %xmm11
	movaps	 %xmm7, %xmm6
	mulsd	 %xmm1, %xmm7
	movsd	 4 * SIZE(BO), %xmm1

	addsd	 %xmm5, %xmm12
	movsd	10 * SIZE(AO), %xmm5
	mulsd	 %xmm3, %xmm6
	movsd	 5 * SIZE(BO), %xmm3

	addsd	 %xmm2, %xmm13
	PREFETCH (PREFETCHSIZE + 8) * SIZE(AO)
	movaps	 %xmm0, %xmm2
	mulsd	 %xmm1, %xmm0

	addsd	 %xmm7, %xmm14
	movsd	11 * SIZE(AO), %xmm7
	mulsd	 %xmm3, %xmm2

	addsd	 %xmm6, %xmm15
	movaps	 %xmm4, %xmm6
	mulsd	 %xmm1, %xmm4

	addsd	 %xmm0, %xmm8
	movsd	12 * SIZE(AO), %xmm0
	mulsd	 %xmm3, %xmm6

	addsd	 %xmm2, %xmm9
	movaps	 %xmm5, %xmm2
	mulsd	 %xmm1, %xmm5

	addsd	 %xmm4, %xmm10
	movsd	13 * SIZE(AO), %xmm4
	mulsd	 %xmm3, %xmm2

	addsd	 %xmm6, %xmm11
	movaps	 %xmm7, %xmm6
	mulsd	 %xmm1, %xmm7
	movsd	 6 * SIZE(BO), %xmm1

	addsd	 %xmm5, %xmm12
	movsd	14 * SIZE(AO), %xmm5
	mulsd	 %xmm3, %xmm6
	movsd	 7 * SIZE(BO), %xmm3

	addsd	 %xmm2, %xmm13
	movaps	 %xmm0, %xmm2
	mulsd	 %xmm1, %xmm0

	addsd	 %xmm7, %xmm14
	movsd	15 * SIZE(AO), %xmm7
	mulsd	 %xmm3, %xmm2
	subq   $-16 * SIZE, AO

	addsd	 %xmm6, %xmm15
	movaps	 %xmm4, %xmm6
	mulsd	 %xmm1, %xmm4

	addsd	 %xmm0, %xmm8
	movsd	 0 * SIZE(AO), %xmm0
	mulsd	 %xmm3, %xmm6

	addsd	 %xmm2, %xmm9
	movaps	 %xmm5, %xmm2
	mulsd	 %xmm1, %xmm5
	addq   $  8 * SIZE, BO

	addsd	 %xmm4, %xmm10
	movsd	 1 * SIZE(AO), %xmm4
	mulsd	 %xmm3, %xmm2
	decq   %rax

	addsd	 %xmm6, %xmm11
	movaps	 %xmm7, %xmm6
	mulsd	 %xmm1, %xmm7
	movsd	 0 * SIZE(BO), %xmm1

	addsd	 %xmm5, %xmm12
	movsd	 2 * SIZE(AO), %xmm5
	mulsd	 %xmm3, %xmm6
	movsd	 1 * SIZE(BO), %xmm3

	jne    .L12
	ALIGN_4

.L15:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax
	BRANCH
	BRANCH
	je	.L19
	ALIGN_4

.L16:
	addsd	 %xmm2, %xmm13
	movaps	 %xmm0, %xmm2
	mulsd	 %xmm1, %xmm0

	addsd	 %xmm7, %xmm14
	movsd	 3 * SIZE(AO), %xmm7
	mulsd	 %xmm3, %xmm2

	addsd	 %xmm6, %xmm15
	movaps	 %xmm4, %xmm6
	mulsd	 %xmm1, %xmm4

	addsd	 %xmm0, %xmm8
	movsd	 4 * SIZE(AO), %xmm0
	mulsd	 %xmm3, %xmm6

	addsd	 %xmm2, %xmm9
	movaps	 %xmm5, %xmm2
	mulsd	 %xmm1, %xmm5

	addsd	 %xmm4, %xmm10
	movsd	 5 * SIZE(AO), %xmm4
	mulsd	 %xmm3, %xmm2

	addsd	 %xmm6, %xmm11
	movaps	 %xmm7, %xmm6
	mulsd	 %xmm1, %xmm7
	movsd	 2 * SIZE(BO), %xmm1

	addsd	 %xmm5, %xmm12
	movsd	 6 * SIZE(AO), %xmm5
	mulsd	 %xmm3, %xmm6
	movsd	 3 * SIZE(BO), %xmm3

	addq	$4 * SIZE, AO
	addq	$2 * SIZE, BO
	decq	%rax
	BRANCH
	jg	.L16
	ALIGN_4

.L19:
	movsd	ALPHA, %xmm5

	addsd	 %xmm2, %xmm13
	mulsd	%xmm5, %xmm8
	addsd	 %xmm7, %xmm14
	mulsd	%xmm5, %xmm10
	addsd	 %xmm6, %xmm15
	mulsd	%xmm5, %xmm12
	mulsd	%xmm5, %xmm14

	mulsd	%xmm5, %xmm9
	mulsd	%xmm5, %xmm11
	mulsd	%xmm5, %xmm13
	mulsd	%xmm5, %xmm15

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addsd	0 * SIZE(CO1), %xmm8
	addsd	1 * SIZE(CO1), %xmm10
	addsd	2 * SIZE(CO1), %xmm12
	addsd	3 * SIZE(CO1), %xmm14

	addsd	0 * SIZE(CO2), %xmm9
	addsd	1 * SIZE(CO2), %xmm11
	addsd	2 * SIZE(CO2), %xmm13
	addsd	3 * SIZE(CO2), %xmm15
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movsd	%xmm10, 1 * SIZE(CO1)
	movsd	%xmm12, 2 * SIZE(CO1)
	movsd	%xmm14, 3 * SIZE(CO1)

	movsd	%xmm9,  0 * SIZE(CO2)
	movsd	%xmm11, 1 * SIZE(CO2)
	movsd	%xmm13, 2 * SIZE(CO2)
	movsd	%xmm15, 3 * SIZE(CO2)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 2), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$4, KK
#endif

	addq	$4 * SIZE, CO1
	addq	$4 * SIZE, CO2

	decq	I			# i --
	jg	.L11
	ALIGN_4

.L20:
	testq	$2, M
	jle	.L30

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(B,  %rax, 2), BO
#endif

	movsd	 0 * SIZE(AO), %xmm0
	xorps	%xmm2,  %xmm2
	movsd	 1 * SIZE(AO), %xmm4
	xorps	%xmm5,  %xmm5
	movsd	 2 * SIZE(AO), %xmm5
	xorps	%xmm6,  %xmm6
	movsd	 3 * SIZE(AO), %xmm7

	movsd	 0 * SIZE(BO), %xmm1
	xorps	%xmm8,  %xmm8
	xorps	%xmm9,  %xmm9
	movsd	 1 * SIZE(BO), %xmm3
	xorps	%xmm10, %xmm10
	xorps	%xmm11, %xmm11

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$2, %rax
#else
	addq	$2, %rax
#endif
	movq	%rax, KKK
#endif

	sarq	$2, %rax
	je	.L25
	ALIGN_4

.L22:
	PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)
	addsd	 %xmm2, %xmm9
	movaps	 %xmm0, %xmm2
	mulsd	 %xmm1, %xmm0

	addsd	 %xmm6, %xmm11
	movaps	 %xmm4, %xmm6
	mulsd	 %xmm1, %xmm4
	movsd	 2 * SIZE(BO), %xmm1

	addsd	 %xmm0, %xmm8
	movsd	 4 * SIZE(AO), %xmm0
	mulsd	 %xmm3, %xmm2

	addsd	 %xmm4, %xmm10
	movsd	 5 * SIZE(AO), %xmm4
	mulsd	 %xmm3, %xmm6
	movsd	 3 * SIZE(BO), %xmm3

	addsd	 %xmm2, %xmm9
	movaps	 %xmm5, %xmm2
	mulsd	 %xmm1, %xmm5

	addsd	 %xmm6, %xmm11
	movaps	 %xmm7, %xmm6
	mulsd	 %xmm1, %xmm7
	movsd	 4 * SIZE(BO), %xmm1

	addsd	 %xmm5, %xmm8
	movsd	 6 * SIZE(AO), %xmm5
	mulsd	 %xmm3, %xmm2

	addsd	 %xmm7, %xmm10
	movsd	 7 * SIZE(AO), %xmm7
	mulsd	 %xmm3, %xmm6
	movsd	 5 * SIZE(BO), %xmm3

	addsd	 %xmm2, %xmm9
	movaps	 %xmm0, %xmm2
	mulsd	 %xmm1, %xmm0

	addsd	 %xmm6, %xmm11
	movaps	 %xmm4, %xmm6
	mulsd	 %xmm1, %xmm4
	movsd	 6 * SIZE(BO), %xmm1

	addsd	 %xmm0, %xmm8
	movsd	 8 * SIZE(AO), %xmm0
	mulsd	 %xmm3, %xmm2

	addsd	 %xmm4, %xmm10
	movsd	 9 * SIZE(AO), %xmm4
	mulsd	 %xmm3, %xmm6
	movsd	 7 * SIZE(BO), %xmm3

	addsd	 %xmm2, %xmm9
	movaps	 %xmm5, %xmm2
	mulsd	 %xmm1, %xmm5

	addsd	 %xmm6, %xmm11
	movaps	 %xmm7, %xmm6
	mulsd	 %xmm1, %xmm7
	movsd	 8 * SIZE(BO), %xmm1

	addsd	 %xmm5, %xmm8
	movsd	10 * SIZE(AO), %xmm5
	mulsd	 %xmm3, %xmm2

	addsd	 %xmm7, %xmm10
	movsd	11 * SIZE(AO), %xmm7
	mulsd	 %xmm3, %xmm6
	movsd	 9 * SIZE(BO), %xmm3

	addq	$8 * SIZE, AO
	addq	$8 * SIZE, BO

	decq	%rax
	jne    .L22
	ALIGN_4

.L25:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movsd	ALPHA, %xmm7

	andq	$3, %rax
	BRANCH
	BRANCH
	je	.L29
	ALIGN_4

.L26:
	addsd	 %xmm2, %xmm9
	movaps	 %xmm0, %xmm2
	mulsd	 %xmm1, %xmm0

	addsd	 %xmm6, %xmm11
	movaps	 %xmm4, %xmm6
	mulsd	 %xmm1, %xmm4
	movsd	 2 * SIZE(BO), %xmm1

	mulsd	 %xmm3, %xmm2
	addsd	 %xmm0, %xmm8
	movsd	 2 * SIZE(AO), %xmm0

	mulsd	 %xmm3, %xmm6
	movsd	 3 * SIZE(BO), %xmm3
	addsd	 %xmm4, %xmm10
	movsd	 3 * SIZE(AO), %xmm4

	addq	$2 * SIZE, AO
	addq	$2 * SIZE, BO
	decq	%rax
	BRANCH
	jg	.L26
	ALIGN_4

.L29:
	addsd	 %xmm2, %xmm9
	mulsd	%xmm7, %xmm8
	addsd	 %xmm6, %xmm11
	mulsd	%xmm7, %xmm10
	mulsd	%xmm7, %xmm9
	mulsd	%xmm7, %xmm11

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addsd	0 * SIZE(CO1), %xmm8
	addsd	1 * SIZE(CO1), %xmm10

	addsd	0 * SIZE(CO2), %xmm9
	addsd	1 * SIZE(CO2), %xmm11
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movsd	%xmm10, 1 * SIZE(CO1)

	movsd	%xmm9,  0 * SIZE(CO2)
	movsd	%xmm11, 1 * SIZE(CO2)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 2), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$2, KK
#endif

	addq	$2 * SIZE, CO1
	addq	$2 * SIZE, CO2
	ALIGN_4

.L30:
	testq	$1, M
	je	.L39
	ALIGN_4

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 2), BO
#endif

	movsd	 0 * SIZE(AO), %xmm0
	xorps	%xmm7,  %xmm7
	movsd	 1 * SIZE(AO), %xmm2
	xorps	%xmm5,  %xmm5

	movsd	 0 * SIZE(BO), %xmm1
	xorps	%xmm8,  %xmm8
	xorps	%xmm9,  %xmm9
	movsd	 1 * SIZE(BO), %xmm3

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$1, %rax
#else
	addq	$2, %rax
#endif
	movq	%rax, KKK
#endif

	sarq	$2, %rax
	je	.L35
	ALIGN_4

.L32:
	addsd	 %xmm5, %xmm8
	movsd	 2 * SIZE(BO), %xmm5
	mulsd	 %xmm0, %xmm1
	PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)

	addsd	 %xmm7, %xmm9
	movsd	 3 * SIZE(BO), %xmm7
	mulsd	 %xmm0, %xmm3
	movsd	 2 * SIZE(AO), %xmm0

	addsd	 %xmm1, %xmm8
	movsd	 4 * SIZE(BO), %xmm1
	mulsd	 %xmm2, %xmm5

	addsd	 %xmm3, %xmm9
	movsd	 5 * SIZE(BO), %xmm3
	mulsd	 %xmm2, %xmm7
	movsd	 3 * SIZE(AO), %xmm2

	addsd	 %xmm5, %xmm8
	movsd	 6 * SIZE(BO), %xmm5
	mulsd	 %xmm0, %xmm1

	addsd	 %xmm7, %xmm9
	movsd	 7 * SIZE(BO), %xmm7
	mulsd	 %xmm0, %xmm3
	movsd	 4 * SIZE(AO), %xmm0

	addsd	 %xmm1, %xmm8
	movsd	 8 * SIZE(BO), %xmm1
	mulsd	 %xmm2, %xmm5

	addsd	 %xmm3, %xmm9
	movsd	 9 * SIZE(BO), %xmm3
	mulsd	 %xmm2, %xmm7
	movsd	 5 * SIZE(AO), %xmm2

	addq	$4 * SIZE, AO
	addq	$8 * SIZE, BO

	decq	%rax
	jne    .L32
	ALIGN_4

.L35:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	addsd	 %xmm5, %xmm8
	addsd	 %xmm7, %xmm9
	movsd	ALPHA, %xmm7

	andq	$3, %rax
	BRANCH
	BRANCH
	je	.L38
	ALIGN_4

.L36:
	mulsd	 %xmm0, %xmm1
	addq	$2 * SIZE, BO
	mulsd	 %xmm0, %xmm3
	movsd	 1 * SIZE(AO), %xmm0

	addsd	 %xmm1, %xmm8
	movsd	 0 * SIZE(BO), %xmm1
	addsd	 %xmm3, %xmm9
	movsd	 1 * SIZE(BO), %xmm3

	addq	$1 * SIZE, AO
	decq	%rax
	BRANCH
	jg	.L36
	ALIGN_4

.L38:
	mulsd	%xmm7, %xmm8
	mulsd	%xmm7, %xmm9

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addsd	0 * SIZE(CO1), %xmm8
	addsd	0 * SIZE(CO2), %xmm9
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movsd	%xmm9,  0 * SIZE(CO2)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 2), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$1, KK
#endif
	ALIGN_4

.L39:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$2, KK
#endif

	movq	BO, B
	decq	J			# j --
	jg	.L10
	ALIGN_4

.L40:
	testq	$1, N
	je	.L999

#if defined(TRMMKERNEL) && defined(LEFT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif

	movq	C, CO1
	addq	LDC, C

	movq	A, AO

	movq	M,  I
	sarq	$2, I
	jle	.L50
	ALIGN_4

.L41:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(B,  %rax, 1), BO
#endif

	movsd	 0 * SIZE(AO), %xmm0
	xorps	%xmm9,   %xmm9
	movsd	 1 * SIZE(AO), %xmm1
	xorps	%xmm11,  %xmm11
	movsd	 2 * SIZE(AO), %xmm2
	xorps	%xmm13,  %xmm13
	movsd	 3 * SIZE(AO), %xmm3
	xorps	%xmm15,  %xmm15

	movsd	 0 * SIZE(BO), %xmm4
	xorps	%xmm8,  %xmm8
	movsd	 1 * SIZE(BO), %xmm5
	xorps	%xmm10, %xmm10
	prefetcht0     3 * SIZE(CO1)
	xorps	%xmm12, %xmm12
	xorps	%xmm14, %xmm14

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$4, %rax
#else
	addq	$1, %rax
#endif
	movq	%rax, KKK
#endif

	sarq	$2, %rax
	je	.L45
	ALIGN_4

.L42:
	addsd	 %xmm9,  %xmm8
	movsd	 4 * SIZE(AO), %xmm9
	mulsd	 %xmm4, %xmm0
	PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)

	addsd	 %xmm11, %xmm10
	movsd	 5 * SIZE(AO), %xmm11
	mulsd	 %xmm4, %xmm1

	addsd	 %xmm13, %xmm12
	movsd	 6 * SIZE(AO), %xmm13
	mulsd	 %xmm4, %xmm2

	addsd	 %xmm15, %xmm14
	movsd	 7 * SIZE(AO), %xmm15
	mulsd	 %xmm4, %xmm3
	movsd	 2 * SIZE(BO), %xmm4

	addsd	 %xmm0, %xmm8
	movsd	 8 * SIZE(AO), %xmm0
	mulsd	 %xmm5, %xmm9

	addsd	 %xmm1, %xmm10
	movsd	 9 * SIZE(AO), %xmm1
	mulsd	 %xmm5, %xmm11

	addsd	 %xmm2, %xmm12
	movsd	10 * SIZE(AO), %xmm2
	mulsd	 %xmm5, %xmm13

	addsd	 %xmm3, %xmm14
	movsd	11 * SIZE(AO), %xmm3
	mulsd	 %xmm5, %xmm15
	movsd	 3 * SIZE(BO), %xmm5

	addsd	 %xmm9,  %xmm8
	movsd	12 * SIZE(AO), %xmm9
	mulsd	 %xmm4, %xmm0
	PREFETCH (PREFETCHSIZE + 8) * SIZE(AO)

	addsd	 %xmm11, %xmm10
	movsd	13 * SIZE(AO), %xmm11
	mulsd	 %xmm4, %xmm1

	addsd	 %xmm13, %xmm12
	movsd	14 * SIZE(AO), %xmm13
	mulsd	 %xmm4, %xmm2

	addsd	 %xmm15, %xmm14
	movsd	15 * SIZE(AO), %xmm15
	mulsd	 %xmm4, %xmm3
	movsd	 4 * SIZE(BO), %xmm4
	subq	$-16 * SIZE, AO

	addsd	 %xmm0, %xmm8
	movsd	 0 * SIZE(AO), %xmm0
	mulsd	 %xmm5, %xmm9

	addsd	 %xmm1, %xmm10
	movsd	 1 * SIZE(AO), %xmm1
	mulsd	 %xmm5, %xmm11
	addq	$  4 * SIZE, BO

	addsd	 %xmm2, %xmm12
	movsd	 2 * SIZE(AO), %xmm2
	mulsd	 %xmm5, %xmm13
	decq	%rax

	addsd	 %xmm3, %xmm14
	movsd	 3 * SIZE(AO), %xmm3
	mulsd	 %xmm5, %xmm15
	movsd	 1 * SIZE(BO), %xmm5

	jne    .L42
	ALIGN_4

.L45:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif

	movsd	ALPHA, %xmm7

	addsd	 %xmm9,  %xmm8
	addsd	 %xmm11, %xmm10
	addsd	 %xmm13, %xmm12
	addsd	 %xmm15, %xmm14

	andq	$3, %rax
	BRANCH
	BRANCH
	je	.L49
	ALIGN_4

.L46:
	mulsd	 %xmm4, %xmm0
	mulsd	 %xmm4, %xmm1
	mulsd	 %xmm4, %xmm2
	mulsd	 %xmm4, %xmm3
	movsd	 1 * SIZE(BO), %xmm4

	addsd	 %xmm0, %xmm8
	movsd	 4 * SIZE(AO), %xmm0
	addsd	 %xmm1, %xmm10
	movsd	 5 * SIZE(AO), %xmm1
	addsd	 %xmm2, %xmm12
	movsd	 6 * SIZE(AO), %xmm2
	addsd	 %xmm3, %xmm14
	movsd	 7 * SIZE(AO), %xmm3

	addq	$4 * SIZE, AO
	addq	$1 * SIZE, BO
	decq	%rax
	BRANCH
	jg	.L46
	ALIGN_4

.L49:
	mulsd	%xmm7, %xmm8
	mulsd	%xmm7, %xmm10
	mulsd	%xmm7, %xmm12
	mulsd	%xmm7, %xmm14

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addsd	0 * SIZE(CO1), %xmm8
	addsd	1 * SIZE(CO1), %xmm10
	addsd	2 * SIZE(CO1), %xmm12
	addsd	3 * SIZE(CO1), %xmm14
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movsd	%xmm10, 1 * SIZE(CO1)
	movsd	%xmm12, 2 * SIZE(CO1)
	movsd	%xmm14, 3 * SIZE(CO1)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 1), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$4, KK
#endif

	addq	$4 * SIZE, CO1

	decq	I			# i --
	jg	.L41
	ALIGN_4

.L50:
	testq	$2, M
	jle	.L60

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(B,  %rax, 1), BO
#endif

	movsd	 0 * SIZE(AO), %xmm0
	xorps	%xmm2,   %xmm2
	movsd	 1 * SIZE(AO), %xmm1
	xorps	%xmm3,   %xmm3

	movsd	 0 * SIZE(BO), %xmm4
	xorps	%xmm8,  %xmm8
	movsd	 1 * SIZE(BO), %xmm5
	xorps	%xmm10, %xmm10

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$2, %rax
#else
	addq	$1, %rax
#endif
	movq	%rax, KKK
#endif

	sarq	$2, %rax
	je	.L55
	ALIGN_4

.L52:
	addsd	 %xmm2, %xmm8
	movsd	 2 * SIZE(AO), %xmm2
	mulsd	 %xmm4, %xmm0
	PREFETCH (PREFETCHSIZE + 0) * SIZE(AO)

	addsd	 %xmm3, %xmm10
	movsd	 3 * SIZE(AO), %xmm3
	mulsd	 %xmm4, %xmm1
	movsd	 2 * SIZE(BO), %xmm4

	addsd	 %xmm0, %xmm8
	movsd	 4 * SIZE(AO), %xmm0
	mulsd	 %xmm5, %xmm2
	addq	$8 * SIZE, AO

	addsd	 %xmm1, %xmm10
	movsd	-3 * SIZE(AO), %xmm1
	mulsd	 %xmm5, %xmm3
	movsd	 3 * SIZE(BO), %xmm5

	addsd	 %xmm2, %xmm8
	movsd	-2 * SIZE(AO), %xmm2
	mulsd	 %xmm4, %xmm0
	addq	$4 * SIZE, BO

	addsd	 %xmm3, %xmm10
	movsd	-1 * SIZE(AO), %xmm3
	mulsd	 %xmm4, %xmm1
	movsd	 0 * SIZE(BO), %xmm4

	addsd	 %xmm0, %xmm8
	movsd	 0 * SIZE(AO), %xmm0
	mulsd	 %xmm5, %xmm2
	decq	%rax

	addsd	 %xmm1, %xmm10
	movsd	 1 * SIZE(AO), %xmm1
	mulsd	 %xmm5, %xmm3
	movsd	 1 * SIZE(BO), %xmm5

	jne    .L52
	ALIGN_4

.L55:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movsd	ALPHA, %xmm7

	addsd	 %xmm2, %xmm8
	addsd	 %xmm3, %xmm10

	andq	$3, %rax
	BRANCH
	BRANCH
	je	.L59
	ALIGN_4

.L56:
	mulsd	 %xmm4, %xmm0
	mulsd	 %xmm4, %xmm1
	movsd	 1 * SIZE(BO), %xmm4

	addsd	 %xmm0, %xmm8
	movsd	 2 * SIZE(AO), %xmm0
	addsd	 %xmm1, %xmm10
	movsd	 3 * SIZE(AO), %xmm1

	addq	$2 * SIZE, AO
	addq	$1 * SIZE, BO
	decq	%rax
	BRANCH
	jg	.L56
	ALIGN_4

.L59:
	mulsd	%xmm7, %xmm8
	mulsd	%xmm7, %xmm10

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addsd	0 * SIZE(CO1), %xmm8
	addsd	1 * SIZE(CO1), %xmm10
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movsd	%xmm10, 1 * SIZE(CO1)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 1), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$2, KK
#endif

	addq	$2 * SIZE, CO1
	ALIGN_4

.L60:
	testq	$1, M
	je	.L999
	ALIGN_4

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 1), BO
#endif

	movsd	 0 * SIZE(AO), %xmm0
	xorps	%xmm5,  %xmm5
	movsd	 1 * SIZE(AO), %xmm2
	xorps	%xmm7,  %xmm7

	movsd	 0 * SIZE(BO), %xmm1
	xorps	%xmm8,  %xmm8
	movsd	 1 * SIZE(BO), %xmm3
	xorps	%xmm9,  %xmm9
	movsd	 2 * SIZE(AO), %xmm4
	movsd	 3 * SIZE(AO), %xmm6

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$1, %rax
#else
	addq	$1, %rax
#endif
	movq	%rax, KKK
#endif

	sarq	$2, %rax
	je	.L65
	ALIGN_4

.L62:
	addsd	 %xmm5, %xmm8
	movsd	 2 * SIZE(BO), %xmm5
	mulsd	 %xmm0, %xmm1
	movsd	 4 * SIZE(AO), %xmm0

	addsd	 %xmm7, %xmm9
	movsd	 3 * SIZE(BO), %xmm7
	mulsd	 %xmm2, %xmm3
	movsd	 5 * SIZE(AO), %xmm2

	addsd	 %xmm1, %xmm8
	movsd	 4 * SIZE(BO), %xmm1
	mulsd	 %xmm4, %xmm5
	movsd	 6 * SIZE(AO), %xmm4

	addsd	 %xmm3, %xmm9
	movsd	 5 * SIZE(BO), %xmm3
	mulsd	 %xmm6, %xmm7
	movsd	 7 * SIZE(AO), %xmm6

	addq	$4 * SIZE, AO
	addq	$4 * SIZE, BO

	decq	%rax
	jne    .L62

	addsd	 %xmm5, %xmm8
	addsd	 %xmm7, %xmm9
	ALIGN_4

.L65:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movsd	ALPHA, %xmm7

	andq	$3, %rax
	BRANCH
	BRANCH
	je	.L68
	ALIGN_4

.L66:
	movsd	 0 * SIZE(AO), %xmm0
	movsd	 0 * SIZE(BO), %xmm1

	mulsd	 %xmm0, %xmm1
	addsd	 %xmm1, %xmm8

	addq	$1 * SIZE, AO
	addq	$1 * SIZE, BO

	decq	%rax
	BRANCH
	jg	.L66
	ALIGN_4

.L68:
	addsd	%xmm9, %xmm8

	mulsd	%xmm7, %xmm8

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addsd	0 * SIZE(CO1), %xmm8
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	ALIGN_4

.L999:
	movq	  0(%rsp), %rbx
	movq	  8(%rsp), %rbp
	movq	 16(%rsp), %r12
	movq	 24(%rsp), %r13
	movq	 32(%rsp), %r14
	movq	 40(%rsp), %r15

#ifdef WINDOWS_ABI
	movq	 48(%rsp), %rdi
	movq	 56(%rsp), %rsi
	movups	 64(%rsp), %xmm6
	movups	 80(%rsp), %xmm7
	movups	 96(%rsp), %xmm8
	movups	112(%rsp), %xmm9
	movups	128(%rsp), %xmm10
	movups	144(%rsp), %xmm11
	movups	160(%rsp), %xmm12
	movups	176(%rsp), %xmm13
	movups	192(%rsp), %xmm14
	movups	208(%rsp), %xmm15
#endif

	addq	$STACKSIZE, %rsp
	ret

	EPILOGUE