gemm_kernel_8x2_sse.s

Optimized GotoBLAS libraries

	addl	$2 * SIZE, %esi
	ALIGN_2	

.L70:
	testl	$1, %ebx
	jle	.L99

#if (L1_DATA_LINESIZE == 64)
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leal	BUFFER, BB
	movss	 0 * SIZE + BUFFER, %xmm2
	XORPS	%xmm4, %xmm4
	movss	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movss	 16 * SIZE + BUFFER, %xmm3
	XORPS	%xmm6, %xmm6
	movss	 4 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7
#else
	leal	BUFFER, BB
	movl	KK, %eax
	leal	(, %eax,   4), %eax
	leal	(AA, %eax, 1), AA
	leal	(BB, %eax, 8), BB /* because it's doubled */

	movss	 0 * SIZE(BB), %xmm2
	XORPS	%xmm4, %xmm4
	movss	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movss	 16 * SIZE(BB), %xmm3
	XORPS	%xmm6, %xmm6
	movss	 4 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7
#endif	

#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	$1, %eax
#else
	addl	$2, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L72
	ALIGN_2

.L71:
	mulss	%xmm0, %xmm2
	mulss	 4 * SIZE(BB), %xmm0
	addss	%xmm2, %xmm4
	movss	 8 * SIZE(BB), %xmm2
	addss	%xmm0, %xmm5
	movss	 1 * SIZE(AA), %xmm0
	mulss	%xmm0, %xmm2
	mulss	12 * SIZE(BB), %xmm0
	addss	%xmm2, %xmm6
	movss	32 * SIZE(BB), %xmm2
	addss	%xmm0, %xmm7
	movss	 2 * SIZE(AA), %xmm0
	mulss	%xmm0, %xmm3
	mulss	20 * SIZE(BB), %xmm0
	addss	%xmm3, %xmm4
	movss	24 * SIZE(BB), %xmm3
	addss	%xmm0, %xmm5
	movss	 3 * SIZE(AA), %xmm0
	mulss	%xmm0, %xmm3
	mulss	28 * SIZE(BB), %xmm0
	addss	%xmm3, %xmm6
	movss	48 * SIZE(BB), %xmm3
	addss	%xmm0, %xmm7
	movss	 8 * SIZE(AA), %xmm0
	mulss	%xmm1, %xmm2
	mulss	36 * SIZE(BB), %xmm1
	addss	%xmm2, %xmm4
	movss	40 * SIZE(BB), %xmm2
	addss	%xmm1, %xmm5
	movss	 5 * SIZE(AA), %xmm1
	mulss	%xmm1, %xmm2
	mulss	44 * SIZE(BB), %xmm1
	addss	%xmm2, %xmm6
	movss	64 * SIZE(BB), %xmm2
	addss	%xmm1, %xmm7
	movss	 6 * SIZE(AA), %xmm1
	mulss	%xmm1, %xmm3
	mulss	52 * SIZE(BB), %xmm1
	addss	%xmm3, %xmm4
	movss	56 * SIZE(BB), %xmm3
	addss	%xmm1, %xmm5
	movss	 7 * SIZE(AA), %xmm1
	mulss	%xmm1, %xmm3
	mulss	60 * SIZE(BB), %xmm1
	addss	%xmm3, %xmm6
	movss	80 * SIZE(BB), %xmm3
	addss	%xmm1, %xmm7
	movss	12 * SIZE(AA), %xmm1

	addl   $ 8 * SIZE, AA
	addl   $64 * SIZE, BB
	decl   %eax
	jne    .L71
	ALIGN_2

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

	leal	BUFFER, BB
	movss	 0 * SIZE + BUFFER, %xmm2
	XORPS	%xmm4, %xmm4
	movss	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movss	 8 * SIZE + BUFFER, %xmm3
	XORPS	%xmm6, %xmm6
	movss	 4 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7
#else
	leal	BUFFER, BB
	movl	KK, %eax
	leal	(, %eax,   4), %eax
	leal	(AA, %eax, 1), AA
	leal	(BB, %eax, 8), BB /* because it's doubled */

	movss	 0 * SIZE(BB), %xmm2
	XORPS	%xmm4, %xmm4
	movss	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movss	 8 * SIZE(BB), %xmm3
	XORPS	%xmm6, %xmm6
	movss	 4 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7
#endif	

#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	$1, %eax
#else
	addl	$2, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L72
	ALIGN_2

.L71:
	mulss	%xmm0, %xmm2
	mulss	 4 * SIZE(BB), %xmm0
	addss	%xmm2, %xmm4
	movss	16 * SIZE(BB), %xmm2
	addss	%xmm0, %xmm5
	movss	 1 * SIZE(AA), %xmm0
	mulss	%xmm0, %xmm3
	mulss	12 * SIZE(BB), %xmm0
	addss	%xmm3, %xmm6
	movss	24 * SIZE(BB), %xmm3
	addss	%xmm0, %xmm7
	movss	 2 * SIZE(AA), %xmm0
	mulss	%xmm0, %xmm2
	mulss	20 * SIZE(BB), %xmm0
	addss	%xmm2, %xmm4
	movss	32 * SIZE(BB), %xmm2
	addss	%xmm0, %xmm5
	movss	 3 * SIZE(AA), %xmm0
	mulss	%xmm0, %xmm3
	mulss	28 * SIZE(BB), %xmm0
	addss	%xmm3, %xmm6
	movss	40 * SIZE(BB), %xmm3
	addss	%xmm0, %xmm7
	movss	 8 * SIZE(AA), %xmm0
	mulss	%xmm1, %xmm2
	mulss	36 * SIZE(BB), %xmm1
	addss	%xmm2, %xmm4
	movss	48 * SIZE(BB), %xmm2
	addss	%xmm1, %xmm5
	movss	 5 * SIZE(AA), %xmm1
	mulss	%xmm1, %xmm3
	mulss	44 * SIZE(BB), %xmm1
	addss	%xmm3, %xmm6
	movss	56 * SIZE(BB), %xmm3
	addss	%xmm1, %xmm7
	movss	 6 * SIZE(AA), %xmm1
	mulss	%xmm1, %xmm2
	mulss	52 * SIZE(BB), %xmm1
	addss	%xmm2, %xmm4
	movss	64 * SIZE(BB), %xmm2
	addss	%xmm1, %xmm5
	movss	 7 * SIZE(AA), %xmm1
	mulss	%xmm1, %xmm3
	mulss	60 * SIZE(BB), %xmm1
	addss	%xmm3, %xmm6
	movss	72 * SIZE(BB), %xmm3
	addss	%xmm1, %xmm7
	movss	12 * SIZE(AA), %xmm1

	addl   $ 8 * SIZE, AA
	addl   $64 * SIZE, BB
	decl   %eax
	jne    .L71
	ALIGN_2
#endif

.L72:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	movss	ALPHA,  %xmm3
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L74

.L73:
	mulss	%xmm0, %xmm2
	mulss	 4 * SIZE(BB), %xmm0
	addss	%xmm2, %xmm4
	movss	 8 * SIZE(BB), %xmm2
	addss	%xmm0, %xmm5
	movss	 1 * SIZE(AA), %xmm0

	addl	$1 * SIZE, AA
	addl	$8 * SIZE, BB
	decl	%eax
	jg	.L73
	ALIGN_4

.L74:
	addss	%xmm6, %xmm4
	addss	%xmm7, %xmm5

	mulss	%xmm3, %xmm4
	mulss	%xmm3, %xmm5

#ifndef TRMMKERNEL
	addss	0 * SIZE(%esi), %xmm4
	addss	0 * SIZE(%esi, LDC), %xmm5
#endif

	movss	%xmm4, 0 * SIZE(%esi)
	movss	%xmm5, 0 * SIZE(%esi, LDC)

	addl	$1 * SIZE, %esi

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movl	K, %eax
	subl	KKK, %eax
	leal	(,%eax,    4), %eax
	leal	(AA, %eax, 1), AA
	leal	(BB, %eax, 8), BB
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addl	$1, KK
#endif
	ALIGN_2	

.L99:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$2, KK
#endif

	leal	(, LDC, 2), %eax
	addl	%eax, C			# c += 2 * ldc
	BRANCH
	decl	J			# j --
	jg	.L01
	ALIGN_2

.L100:
	movl	N, %eax
	testl	$1, %eax
	jle	.L999
	ALIGN_2
	
.L101:
#if defined(TRMMKERNEL) && defined(LEFT)
	movl	OFFSET, %eax
	movl	%eax, KK
#endif	

/* Copying to Sub Buffer */
	movl	K, %eax
	leal	BUFFER, %ecx
	sarl	$3, %eax
	jle	.L103
	ALIGN_4
	
.L102:
	prefetchnta	96 * SIZE(B)
	
	movss	0 * SIZE(B), %xmm0
	movss	1 * SIZE(B), %xmm1
	movss	2 * SIZE(B), %xmm2
	movss	3 * SIZE(B), %xmm3
	movss	4 * SIZE(B), %xmm4
	movss	5 * SIZE(B), %xmm5
	movss	6 * SIZE(B), %xmm6
	movss	7 * SIZE(B), %xmm7
	addl	$ 8 * SIZE, B

	shufps	$0, %xmm0, %xmm0
	shufps	$0, %xmm1, %xmm1
	shufps	$0, %xmm2, %xmm2
	shufps	$0, %xmm3, %xmm3
	shufps	$0, %xmm4, %xmm4
	shufps	$0, %xmm5, %xmm5
	shufps	$0, %xmm6, %xmm6
	shufps	$0, %xmm7, %xmm7

	movaps	%xmm0,  0 * SIZE(%ecx)
	movaps	%xmm1,  4 * SIZE(%ecx)
	movaps	%xmm2,  8 * SIZE(%ecx)
	movaps	%xmm3, 12 * SIZE(%ecx)
	movaps	%xmm4, 16 * SIZE(%ecx)
	movaps	%xmm5, 20 * SIZE(%ecx)
	movaps	%xmm6, 24 * SIZE(%ecx)
	movaps	%xmm7, 28 * SIZE(%ecx)
	addl	$32 * SIZE, %ecx

	decl	%eax
	BRANCH
	jne	.L102
	ALIGN_2

.L103:
	movl	K, %eax
	andl	$7, %eax
	BRANCH
	jle	.L105
	ALIGN_2

.L104:
	movss	0 * SIZE(B), %xmm0
	addl	$1 * SIZE, B

	shufps	$0, %xmm0, %xmm0

	movaps	%xmm0,  0 * SIZE(%ecx)
	addl	$4 * SIZE, %ecx
	decl	%eax
	jne	.L104
	ALIGN_4

.L105:
	movl	C, %esi		# coffset = c
	movl	A, AA		# aoffset = a
	movl	M,  %ebx
	sarl	$3, %ebx	# i = (m >> 2)
	jle	.L130
	ALIGN_4

.L110:
#if (L1_DATA_LINESIZE == 64)
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leal	BUFFER, BB
	movaps	 0 * SIZE + BUFFER, %xmm2
	XORPS	%xmm4, %xmm4
	movaps	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movaps	 16 * SIZE + BUFFER, %xmm3
	XORPS	%xmm6, %xmm6
	movaps	 16 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7
#else
	leal	BUFFER, BB
	movl	KK, %eax
	leal	(, %eax,   8), %eax
	leal	(AA, %eax, 4), AA
	leal	(BB, %eax, 2), BB /* because it's doubled */

	movaps	 0 * SIZE(BB), %xmm2
	XORPS	%xmm4, %xmm4
	movaps	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movaps	 16 * SIZE(BB), %xmm3
	XORPS	%xmm6, %xmm6
	movaps	 16 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7
#endif	

#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	$8, %eax
#else
	addl	$1, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L112
	ALIGN_2

.L111:
	mulps	%xmm2, %xmm0
	mulps	 4 * SIZE(AA), %xmm2
	addps	%xmm0, %xmm4
	movaps	 8 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5
	movaps	 4 * SIZE(BB), %xmm2
	mulps	%xmm2, %xmm0
	mulps	12 * SIZE(AA), %xmm2
	addps	%xmm0, %xmm6
	movaps	32 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm7
	movaps	 8 * SIZE(BB), %xmm2
	mulps	%xmm2, %xmm1
	mulps	20 * SIZE(AA), %xmm2
	addps	%xmm1, %xmm4
	movaps	24 * SIZE(AA), %xmm1
	addps	%xmm2, %xmm5
	movaps	12 * SIZE(BB), %xmm2
	mulps	%xmm2, %xmm1
	mulps	28 * SIZE(AA), %xmm2
	addps	%xmm1, %xmm6
	movaps	48 * SIZE(AA), %xmm1
	addps	%xmm2, %xmm7
	movaps	32 * SIZE(BB), %xmm2
	mulps	%xmm3, %xmm0
	mulps	36 * SIZE(AA), %xmm3
	addps	%xmm0, %xmm4
	movaps	40 * SIZE(AA), %xmm0
	addps	%xmm3, %xmm5
	movaps	20 * SIZE(BB), %xmm3
	mulps	%xmm3, %xmm0
	mulps	44 * SIZE(AA), %xmm3
	addps	%xmm0, %xmm6
	movaps	64 * SIZE(AA), %xmm0
	addps	%xmm3, %xmm7
	movaps	24 * SIZE(BB), %xmm3
	mulps	%xmm3, %xmm1
	mulps	52 * SIZE(AA), %xmm3
	addps	%xmm1, %xmm4
	movaps	56 * SIZE(AA), %xmm1
	addps	%xmm3, %xmm5
	movaps	28 * SIZE(BB), %xmm3
	mulps	%xmm3, %xmm1
	mulps	60 * SIZE(AA), %xmm3
	addps	%xmm1, %xmm6
	movaps	80 * SIZE(AA), %xmm1
	addps	%xmm3, %xmm7
	movaps	48 * SIZE(BB), %xmm3

	addl   $64 * SIZE, AA
	addl   $32 * SIZE, BB
	decl   %eax
	jne    .L111
	ALIGN_2

#else

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

	leal	BUFFER, BB
	movaps	 0 * SIZE + BUFFER, %xmm2
	XORPS	%xmm4, %xmm4
	movaps	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movaps	 8 * SIZE + BUFFER, %xmm3
	XORPS	%xmm6, %xmm6
	movaps	 8 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7
#else
	leal	BUFFER, BB
	movl	KK, %eax
	leal	(, %eax,   8), %eax
	leal	(AA, %eax, 4), AA
	leal	(BB, %eax, 2), BB /* because it's doubled */

	movaps	 0 * SIZE(BB), %xmm2
	XORPS	%xmm4, %xmm4
	movaps	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movaps	 8 * SIZE(BB), %xmm3
	XORPS	%xmm6, %xmm6
	movaps	 8 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7
#endif	

#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	$8, %eax
#else
	addl	$1, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L112
	ALIGN_2

.L111:
	mulps	%xmm2, %xmm0
	mulps	 4 * SIZE(AA), %xmm2
	addps	%xmm0, %xmm4
	movaps	16 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5
	movaps	 4 * SIZE(BB), %xmm2
	mulps	%xmm2, %xmm1
	mulps	12 * SIZE(AA), %xmm2
	addps	%xmm1, %xmm6
	movaps	24 * SIZE(AA), %xmm1
	addps	%xmm2, %xmm7
	movaps	16 * SIZE(BB), %xmm2
	mulps	%xmm3, %xmm0
	mulps	20 * SIZE(AA), %xmm3
	addps	%xmm0, %xmm4
	movaps	32 * SIZE(AA), %xmm0
	addps	%xmm3, %xmm5
	movaps	12 * SIZE(BB), %xmm3
	mulps	%xmm3, %xmm1
	mulps	28 * SIZE(AA), %xmm3
	addps	%xmm1, %xmm6
	movaps	40 * SIZE(AA), %xmm1
	addps	%xmm3, %xmm7
	movaps	24 * SIZE(BB), %xmm3
	mulps	%xmm2, %xmm0
	mulps	36 * SIZE(AA), %xmm2
	addps	%xmm0, %xmm4
	movaps	48 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5
	movaps	20 * SIZE(BB), %xmm2
	mulps	%xmm2, %xmm1
	mulps	44 * SIZE(AA), %xmm2
	addps	%xmm1, %xmm6
	movaps	56 * SIZE(AA), %xmm1
	addps	%xmm2, %xmm7
	movaps	32 * SIZE(BB), %xmm2
	mulps	%xmm3, %xmm0
	mulps	52 * SIZE(AA), %xmm3
	addps	%xmm0, %xmm4
	movaps	64 * SIZE(AA), %xmm0
	addps	%xmm3, %xmm5
	movaps	28 * SIZE(BB), %xmm3
	mulps	%xmm3, %xmm1
	mulps	60 * SIZE(AA), %xmm3
	addps	%xmm1, %xmm6
	movaps	72 * SIZE(AA), %xmm1
	addps	%xmm3, %xmm7
	movaps	40 * SIZE(BB), %xmm3

	addl   $64 * SIZE, AA
	addl   $32 * SIZE, BB
	decl   %eax
	jne    .L111
	ALIGN_2
#endif

.L112:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	movaps	ALPHA,  %xmm3
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L114

.L113:
	movaps	 0 * SIZE(BB), %xmm2
	movaps	 0 * SIZE(AA), %xmm0
	mulps	%xmm2, %xmm0
	addps	%xmm0, %xmm4
	mulps	 4 * SIZE(AA), %xmm2
	addps	%xmm2, %xmm5

	addl	$8 * SIZE, AA
	addl	$4 * SIZE, BB
	subl	$1, %eax
	jg	.L113
	ALIGN_4

.L114:
	addps	%xmm6, %xmm4
	addps	%xmm7, %xmm5

	mulps	%xmm3, %xmm4
	mulps	%xmm3, %xmm5

#ifndef TRMMKERNEL
	movlps	0 * SIZE(%esi), %xmm0
	movhps	2 * SIZE(%esi), %xmm0
	addps	%xmm0, %xmm4

	movlps	4 * SIZE(%esi), %xmm0
	movhps	6 * SIZE(%esi), %xmm0
	addps	%xmm0, %xmm5
#endif

#ifdef HAVE_SSE2
	movsd	%xmm4, 0 * SIZE(%esi)
	unpckhpd %xmm4, %xmm4
	movsd	%xmm4, 2 * SIZE(%esi)

	movsd	%xmm5, 4 * SIZE(%esi)
	unpckhpd %xmm5, %xmm5
	movsd	%xmm5, 6 * SIZE(%esi)
#else
	movlps	%xmm4, 0 * SIZE(%esi)
	movhps	%xmm4, 2 * SIZE(%esi)

	movlps	%xmm5, 4 * SIZE(%esi)
	movhps	%xmm5, 6 * SIZE(%esi)
#endif

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movl	K, %eax
	subl	KKK, %eax
	leal	(,%eax,    8), %eax
	leal	(AA, %eax, 4), AA
	leal	(BB, %eax, 2), BB
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addl	$8, KK
#endif

	addl	$8 * SIZE, %esi
	BRANCH
	decl	%ebx			# i --
	jg	.L110
	ALIGN_2	

.L130:
	movl	M,  %ebx
	andl	$7, %ebx
	jle	.L999

	testl	$4, %ebx
	jle	.L150

#if (L1_DATA_LINESIZE == 64)
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leal	BUFFER, BB
	movaps	 0 * SIZE + BUFFER, %xmm2
	XORPS	%xmm4, %xmm4
	movaps	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movaps	 16 * SIZE + BUFFER, %xmm3
	XORPS	%xmm6, %xmm6
	movaps	 16 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7
#else