zgemm_kernel_2x2_sse3.s

Optimized GotoBLAS libraries

/*********************************************************************/
/*                                                                   */
/*             Optimized BLAS libraries                              */
/*                     By Kazushige Goto <[e-mail]>     */
/*                                                                   */
/* 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 STACK	16
#define ARGS	 0
	
#define STACK_M	 4 + STACK + ARGS(%esi)
#define STACK_N	 8 + STACK + ARGS(%esi)
#define STACK_K	12 + STACK + ARGS(%esi)
#define STACK_ALPHA_R	16 + STACK + ARGS(%esi)
#define STACK_ALPHA_I	20 + STACK + ARGS(%esi)
#define STACK_A	24 + STACK + ARGS(%esi)
#define STACK_B	28 + STACK + ARGS(%esi)
#define STACK_C	32 + STACK + ARGS(%esi)
#define STACK_LDC	36 + STACK + ARGS(%esi)
#define STACK_OFFT	40 + STACK + ARGS(%esi)

#define POSINV	 0(%esp)
#define ALPHA_R	16(%esp)
#define ALPHA_I	32(%esp)
#define K	48(%esp)
#define N	52(%esp)
#define M	56(%esp)
#define A	60(%esp)
#define C	64(%esp)
#define J	68(%esp)
#define OLD_STACK 72(%esp)
#define OFFSET  76(%esp)
#define KK	80(%esp)
#define KKK	84(%esp)
#define BUFFER 128(%esp)

#define B	%edi
#define	LDC	%ebp
#define AA	%edx
#define BB	%ecx

#define LOCAL_BUFFER_SIZE  GEMM_Q * GEMM_UNROLL_N * COMPSIZE * 8

#ifdef PENTIUM4
#define PREFETCH prefetcht0
#define PREFETCHSIZE 168
#endif

#ifdef PENTIUMM
#define PREFETCH prefetcht0
#define PREFETCHSIZE 168
#endif

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(RN) || defined(RT) || defined(CN) || defined(CT)
#define ADDSUB	addps
#else
#define ADDSUB	subps
#endif

#define KERNEL1(address) \
	mulps	%xmm0, %xmm2; \
	PREFETCH  (PREFETCHSIZE +  0) * SIZE + 1 * (address) * SIZE(AA); \
	addps	%xmm2, %xmm4; \
	movshdup  0 * SIZE + 2 * (address) * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	ADDSUB	%xmm2, %xmm5; \
	movsldup  4 * SIZE + 2 * (address) * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	addps	%xmm2, %xmm6; \
	movshdup  4 * SIZE + 2 * (address) * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	movaps	  4 * SIZE + 1 * (address) * SIZE(AA), %xmm0; \
	ADDSUB	%xmm2, %xmm7; \
	movsldup  8 * SIZE + 2 * (address) * SIZE(BB), %xmm2

#define KERNEL2(address) \
	mulps	%xmm0, %xmm2; \
	addps	%xmm2, %xmm4; \
	movshdup  8 * SIZE + 2 * (address) * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	ADDSUB	%xmm2, %xmm5; \
	movsldup 12 * SIZE + 2 * (address) * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	addps	%xmm2, %xmm6; \
	movshdup 12 * SIZE + 2 * (address) * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	movaps	  8 * SIZE + 1 * (address) * SIZE(AA), %xmm0; \
	ADDSUB	%xmm2, %xmm7; \
	movsldup 32 * SIZE + 2 * (address) * SIZE(BB), %xmm2

#define KERNEL3(address) \
	mulps	%xmm0, %xmm3; \
	addps	%xmm3, %xmm4; \
	movshdup 16 * SIZE + 2 * (address) * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	ADDSUB	%xmm3, %xmm5; \
	movsldup 20 * SIZE + 2 * (address) * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	addps	%xmm3, %xmm6; \
	movshdup 20 * SIZE + 2 * (address) * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	movaps	 12 * SIZE + 1 * (address) * SIZE(AA), %xmm0; \
	ADDSUB	%xmm3, %xmm7; \
	movsldup 24 * SIZE + 2 * (address) * SIZE(BB), %xmm3

#define KERNEL4(address) \
	mulps	%xmm0, %xmm3; \
	addps	%xmm3, %xmm4; \
	movshdup 24 * SIZE + 2 * (address) * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	ADDSUB	%xmm3, %xmm5; \
	movsldup 28 * SIZE + 2 * (address) * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	addps	%xmm3, %xmm6; \
	movshdup 28 * SIZE + 2 * (address) * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	movaps	 32 * SIZE + 1 * (address) * SIZE(AA), %xmm0; \
	ADDSUB	%xmm3, %xmm7; \
	movsldup 48 * SIZE + 2 * (address) * SIZE(BB), %xmm3

#define KERNEL5(address) \
	mulps	%xmm1, %xmm2; \
	addps	%xmm2, %xmm4; \
	movshdup 32 * SIZE + 2 * (address) * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	ADDSUB	%xmm2, %xmm5; \
	movsldup 36 * SIZE + 2 * (address) * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	addps	%xmm2, %xmm6; \
	movshdup 36 * SIZE + 2 * (address) * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	movaps	 20 * SIZE + 1 * (address) * SIZE(AA), %xmm1; \
	ADDSUB	%xmm2, %xmm7; \
	movsldup 40 * SIZE + 2 * (address) * SIZE(BB), %xmm2

#define KERNEL6(address) \
	mulps	%xmm1, %xmm2; \
	addps	%xmm2, %xmm4; \
	movshdup 40 * SIZE + 2 * (address) * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	ADDSUB	%xmm2, %xmm5; \
	movsldup 44 * SIZE + 2 * (address) * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	addps	%xmm2, %xmm6; \
	movshdup 44 * SIZE + 2 * (address) * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	movaps	 24 * SIZE + 1 * (address) * SIZE(AA), %xmm1; \
	ADDSUB	%xmm2, %xmm7; \
	movsldup 64 * SIZE + 2 * (address) * SIZE(BB), %xmm2

#define KERNEL7(address) \
	mulps	%xmm1, %xmm3; \
	addps	%xmm3, %xmm4; \
	movshdup 48 * SIZE + 2 * (address) * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	ADDSUB	%xmm3, %xmm5; \
	movsldup 52 * SIZE + 2 * (address) * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	addps	%xmm3, %xmm6; \
	movshdup 52 * SIZE + 2 * (address) * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	movaps	 28 * SIZE + 1 * (address) * SIZE(AA), %xmm1; \
	ADDSUB	%xmm3, %xmm7; \
	movsldup 56 * SIZE + 2 * (address) * SIZE(BB), %xmm3

#define KERNEL8(address) \
	mulps	%xmm1, %xmm3; \
	addps	%xmm3, %xmm4; \
	movshdup 56 * SIZE + 2 * (address) * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	ADDSUB	%xmm3, %xmm5; \
	movsldup 60 * SIZE + 2 * (address) * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	addps	%xmm3, %xmm6; \
	movshdup 60 * SIZE + 2 * (address) * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	movaps	 48 * SIZE + 1 * (address) * SIZE(AA), %xmm1; \
	ADDSUB	%xmm3, %xmm7; \
	movsldup 80 * SIZE + 2 * (address) * SIZE(BB), %xmm3

	PROLOGUE

	pushl	%ebp
	pushl	%edi
	pushl	%esi
	pushl	%ebx

	PROFCODE

	movl	%esp, %esi	# save old stack

	subl	$128 + LOCAL_BUFFER_SIZE, %esp
	andl	$-1024, %esp	# align stack

#ifdef WINDOWS_ABI
#if LOCAL_BUFFER_SIZE > 12288
	movl	$0,  4096 * 3(%esp)
#endif
#if LOCAL_BUFFER_SIZE >  8192
	movl	$0,  4096 * 2(%esp)
#endif
#if LOCAL_BUFFER_SIZE >  4096
	movl	$0,  4096 * 1(%esp)
#endif
	movl	$0,  4096 * 0(%esp)
#endif

	movl	STACK_M, %ebx
	movl	STACK_N, %eax
	movl	STACK_K, %ecx
	movl	STACK_A, %edx

	movl	%ebx, M
	movl	%eax, N
	movl	%ecx, K
	movl	%edx, A
	movl	%esi, OLD_STACK

	movl	STACK_B, %edi
	movl	STACK_C, %ebx
#ifdef TRMMKERNEL
	movss	STACK_OFFT, %xmm4
#endif

	movss	STACK_ALPHA_R, %xmm0
	movss	STACK_ALPHA_I, %xmm1

	pxor	%xmm7, %xmm7
	cmpeqps	%xmm7, %xmm7
	pslld	$31, %xmm7	# Generate mask

	shufps	$0, %xmm0, %xmm0
	movaps	 %xmm0,  0 + ALPHA_R

	movss	 %xmm1,  4 + ALPHA_I
	movss	 %xmm1, 12 + ALPHA_I
	xorps	 %xmm7, %xmm1
	movss	 %xmm1,  0 + ALPHA_I
	movss	 %xmm1,  8 + ALPHA_I

	movl	%ebx, C
	movl	STACK_LDC, LDC

#ifdef TRMMKERNEL
	movss	%xmm4, OFFSET
	movss	%xmm4, KK
#ifndef LEFT
	negl	KK
#endif	
#endif

	sall	$ZBASE_SHIFT, LDC
	movl	%eax, J			# j = n
	sarl	$1, J
	jle	.L100
	ALIGN_4

.L01:
#if defined(TRMMKERNEL) && defined(LEFT)
	movl	OFFSET, %eax
	movl	%eax, KK
#endif	

/* Copying to Sub Buffer */
	leal	BUFFER, %ecx

	movl	K, %eax
	sarl	$2, %eax
	jle	.L03
	ALIGN_4

.L02:
	movddup	 0 * SIZE(B), %xmm0
	movddup	 2 * SIZE(B), %xmm1
	movddup	 4 * SIZE(B), %xmm2
	movddup	 6 * SIZE(B), %xmm3
	movddup	 8 * SIZE(B), %xmm4
	movddup	10 * SIZE(B), %xmm5
	movddup	12 * SIZE(B), %xmm6
	movddup	14 * SIZE(B), %xmm7

	movaps	%xmm0,  0 * SIZE(BB)
	movaps	%xmm1,  4 * SIZE(BB)
	movaps	%xmm2,  8 * SIZE(BB)
	movaps	%xmm3, 12 * SIZE(BB)
	movaps	%xmm4, 16 * SIZE(BB)
	movaps	%xmm5, 20 * SIZE(BB)
	movaps	%xmm6, 24 * SIZE(BB)
	movaps	%xmm7, 28 * SIZE(BB)

#	prefetcht1	128 * SIZE(%ecx)
	prefetcht0	112 * SIZE(%edi)

	addl	$16 * SIZE, B
	addl	$32 * SIZE, BB

	decl	%eax
	jne	.L02
	ALIGN_4

.L03:
	movl	K, %eax
	andl	$3, %eax
	BRANCH
	jle	.L05
	ALIGN_4

.L04:
	movddup	0 * SIZE(B), %xmm0
	movddup	2 * SIZE(B), %xmm1

	movaps	%xmm0,  0 * SIZE(BB)
	movaps	%xmm1,  4 * SIZE(BB)

	addl	$4 * SIZE, B
	addl	$8 * SIZE, BB
	decl	%eax
	jne	.L04
	ALIGN_4

.L05:
	movl	C,  %esi
	movl	A,  %edx
	movl	M,  %ebx
	sarl	$1, %ebx
	jle	.L30
	ALIGN_4

.L10:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leal	BUFFER, BB	# boffset1 = boffset
#else
	leal	BUFFER, BB	# boffset1 = boffset
	movl	KK, %eax
	leal	(, %eax,   8), %eax
	leal	(AA, %eax, 2), AA
	leal	(BB, %eax, 4), BB
#endif	

	movaps	 0 * SIZE(AA), %xmm0
	pxor	%xmm4, %xmm4
	movaps	16 * SIZE(AA), %xmm1
	pxor	%xmm5, %xmm5
	movsldup  0 * SIZE(BB), %xmm2
	pxor	%xmm6, %xmm6
	movsldup 16 * SIZE(BB), %xmm3
	pxor	%xmm7, %xmm7

	prefetchnta 4 * SIZE(%esi)
	prefetchnta 4 * SIZE(%esi, LDC)

#ifndef TRMMKERNEL
	movl	K, %eax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movl	K, %eax
	subl	KK, %eax
	movl	%eax, KKK	
#else
	movl	KK, %eax
#ifdef LEFT
	addl	$2, %eax
#else
	addl	$2, %eax
#endif
	movl	%eax, KKK
#endif


#if 1
	andl	$-8, %eax
	sall	$4, %eax
	je	.L15
.L1X:	
	KERNEL1(32 *  0)
	KERNEL2(32 *  0)
	KERNEL3(32 *  0)
	KERNEL4(32 *  0)
	KERNEL5(32 *  0)
	KERNEL6(32 *  0)
	KERNEL7(32 *  0)
	KERNEL8(32 *  0)
	cmpl	$128 *  1, %eax
	jle	.L12
	KERNEL1(32 *  1)
	KERNEL2(32 *  1)
	KERNEL3(32 *  1)
	KERNEL4(32 *  1)
	KERNEL5(32 *  1)
	KERNEL6(32 *  1)
	KERNEL7(32 *  1)
	KERNEL8(32 *  1)
	cmpl	$128 *  2, %eax
	jle	.L12
	KERNEL1(32 *  2)
	KERNEL2(32 *  2)
	KERNEL3(32 *  2)
	KERNEL4(32 *  2)
	KERNEL5(32 *  2)
	KERNEL6(32 *  2)
	KERNEL7(32 *  2)
	KERNEL8(32 *  2)
	cmpl	$128 *  3, %eax
	jle	.L12
	KERNEL1(32 *  3)
	KERNEL2(32 *  3)
	KERNEL3(32 *  3)
	KERNEL4(32 *  3)
	KERNEL5(32 *  3)
	KERNEL6(32 *  3)
	KERNEL7(32 *  3)
	KERNEL8(32 *  3)
	cmpl	$128 *  4, %eax
	jle	.L12
	KERNEL1(32 *  4)
	KERNEL2(32 *  4)
	KERNEL3(32 *  4)
	KERNEL4(32 *  4)
	KERNEL5(32 *  4)
	KERNEL6(32 *  4)
	KERNEL7(32 *  4)
	KERNEL8(32 *  4)
	cmpl	$128 *  5, %eax
	jle	.L12
	KERNEL1(32 *  5)
	KERNEL2(32 *  5)
	KERNEL3(32 *  5)
	KERNEL4(32 *  5)
	KERNEL5(32 *  5)
	KERNEL6(32 *  5)
	KERNEL7(32 *  5)
	KERNEL8(32 *  5)
	cmpl	$128 *  6, %eax
	jle	.L12
	KERNEL1(32 *  6)
	KERNEL2(32 *  6)
	KERNEL3(32 *  6)
	KERNEL4(32 *  6)
	KERNEL5(32 *  6)
	KERNEL6(32 *  6)
	KERNEL7(32 *  6)
	KERNEL8(32 *  6)
	cmpl	$128 *  7, %eax
	jle	.L12
	KERNEL1(32 *  7)
	KERNEL2(32 *  7)
	KERNEL3(32 *  7)
	KERNEL4(32 *  7)
	KERNEL5(32 *  7)
	KERNEL6(32 *  7)
	KERNEL7(32 *  7)
	KERNEL8(32 *  7)
#if 1
	cmpl	$128 *  8, %eax
	jle	.L12
	KERNEL1(32 *  8)
	KERNEL2(32 *  8)
	KERNEL3(32 *  8)
	KERNEL4(32 *  8)
	KERNEL5(32 *  8)
	KERNEL6(32 *  8)
	KERNEL7(32 *  8)
	KERNEL8(32 *  8)
	cmpl	$128 *  9, %eax
	jle	.L12
	KERNEL1(32 *  9)
	KERNEL2(32 *  9)
	KERNEL3(32 *  9)
	KERNEL4(32 *  9)
	KERNEL5(32 *  9)
	KERNEL6(32 *  9)
	KERNEL7(32 *  9)
	KERNEL8(32 *  9)
	cmpl	$128 * 10, %eax
	jle	.L12
	KERNEL1(32 * 10)
	KERNEL2(32 * 10)
	KERNEL3(32 * 10)
	KERNEL4(32 * 10)
	KERNEL5(32 * 10)
	KERNEL6(32 * 10)
	KERNEL7(32 * 10)
	KERNEL8(32 * 10)
	cmpl	$128 * 11, %eax
	jle	.L12
	KERNEL1(32 * 11)
	KERNEL2(32 * 11)
	KERNEL3(32 * 11)
	KERNEL4(32 * 11)
	KERNEL5(32 * 11)
	KERNEL6(32 * 11)
	KERNEL7(32 * 11)
	KERNEL8(32 * 11)
	cmpl	$128 * 12, %eax
	jle	.L12
	KERNEL1(32 * 12)
	KERNEL2(32 * 12)
	KERNEL3(32 * 12)
	KERNEL4(32 * 12)
	KERNEL5(32 * 12)
	KERNEL6(32 * 12)
	KERNEL7(32 * 12)
	KERNEL8(32 * 12)
	cmpl	$128 * 13, %eax
	jle	.L12
	KERNEL1(32 * 13)
	KERNEL2(32 * 13)
	KERNEL3(32 * 13)
	KERNEL4(32 * 13)
	KERNEL5(32 * 13)
	KERNEL6(32 * 13)
	KERNEL7(32 * 13)
	KERNEL8(32 * 13)
	cmpl	$128 * 14, %eax
	jle	.L12
	KERNEL1(32 * 14)
	KERNEL2(32 * 14)
	KERNEL3(32 * 14)
	KERNEL4(32 * 14)
	KERNEL5(32 * 14)
	KERNEL6(32 * 14)
	KERNEL7(32 * 14)
	KERNEL8(32 * 14)
	cmpl	$128 * 15, %eax
	jle	.L12
	KERNEL1(32 * 15)
	KERNEL2(32 * 15)
	KERNEL3(32 * 15)
	KERNEL4(32 * 15)
	KERNEL5(32 * 15)
	KERNEL6(32 * 15)
	KERNEL7(32 * 15)
	KERNEL8(32 * 15)
#else
	addl	$128 * 4  * SIZE, BB
	addl	$128 * 2  * SIZE, AA
	subl	$128 * 8, %eax
	jg	.L1X
	jmp	.L15
#endif

.L12:
	leal	(AA, %eax, 1), AA
	leal	(BB, %eax, 2), BB
	ALIGN_4
#else
	sarl	$3, %eax
	je	.L15
	ALIGN_4

.L11:
	KERNEL1(32 *  7)
	KERNEL2(32 *  7)
	KERNEL3(32 *  7)
	KERNEL4(32 *  7)
	KERNEL5(32 *  7)
	KERNEL6(32 *  7)
	KERNEL7(32 *  7)
	KERNEL8(32 *  7)

	addl	$32 * SIZE, AA
	addl	$64 * SIZE, BB
	decl   %eax
	jne    .L11
	ALIGN_4
#endif
	
.L15:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	movaps	ALPHA_R, %xmm1
	movaps	ALPHA_I, %xmm3
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L14
	ALIGN_4

.L13:
	mulps	%xmm0, %xmm2
	addps	%xmm2, %xmm4
	movshdup  0 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	ADDSUB	%xmm2, %xmm5
	movsldup  4 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm2, %xmm6
	movshdup  4 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	movaps	  4 * SIZE(AA), %xmm0
	ADDSUB	%xmm2, %xmm7
	movsldup  8 * SIZE(BB), %xmm2

	addl	$4 * SIZE, AA
	addl	$8 * SIZE, BB
	decl	%eax
	jg	.L13
	ALIGN_4

.L14:
#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(NR) || defined(NC) || defined(TR) || defined(TC)

	shufps	$0xb1, %xmm5, %xmm5
	shufps	$0xb1, %xmm7, %xmm7

	addsubps	%xmm5, %xmm4
	addsubps	%xmm7, %xmm6

	movaps	%xmm4, %xmm5
	movaps	%xmm6, %xmm7

	shufps	$0xb1, %xmm4, %xmm4
	shufps	$0xb1, %xmm6, %xmm6
#else
	shufps	$0xb1, %xmm4, %xmm4
	shufps	$0xb1, %xmm6, %xmm6

	addsubps	%xmm4, %xmm5
	addsubps	%xmm6, %xmm7

	movaps	%xmm5, %xmm4
	movaps	%xmm7, %xmm6

	shufps	$0xb1, %xmm5, %xmm5
	shufps	$0xb1, %xmm7, %xmm7
#endif

	mulps	%xmm1, %xmm5
	mulps	%xmm3, %xmm4
	mulps	%xmm1, %xmm7
	mulps	%xmm3, %xmm6

	addps	%xmm5, %xmm4
	addps	%xmm7, %xmm6

#ifndef TRMMKERNEL
	shufps	$0xe4, %xmm0, %xmm0
	movsd	0 * SIZE(%esi), %xmm0
	movhps	2 * SIZE(%esi), %xmm0

	shufps	$0xe4, %xmm2, %xmm2
	movsd	0 * SIZE(%esi, LDC), %xmm2
	movhps	2 * SIZE(%esi, LDC), %xmm2

	addps	%xmm0, %xmm4
	addps	%xmm2, %xmm6
#endif

	movsd	%xmm4, 0 * SIZE(%esi)
	movhps	%xmm4, 2 * SIZE(%esi)
	movsd	%xmm6, 0 * SIZE(%esi, LDC)
	movhps	%xmm6, 2 * SIZE(%esi, LDC)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movl	K, %eax
	subl	KKK, %eax
	leal	(,%eax, 8), %eax