trsm_kernel_rt_4x4_core2.s

Optimized GotoBLAS libraries

	leaq	(BO, %rax, 2), BO
#endif	

	pxor	%xmm8, %xmm8
	pxor	%xmm9, %xmm9
	pxor	%xmm10, %xmm10
	pxor	%xmm11, %xmm11

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$2, %rax
	je	.L25
	ALIGN_4

.L22:
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	movapd	-16 * SIZE(AO), %xmm0
	movapd	-16 * SIZE(BO), %xmm2
	movapd	-14 * SIZE(BO), %xmm3
	movapd	-12 * SIZE(BO), %xmm4
	movapd	-10 * SIZE(BO), %xmm5

	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm3
	mulpd	%xmm0, %xmm4
	mulpd	%xmm0, %xmm5

	addpd	%xmm2, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm4, %xmm10
	addpd	%xmm5, %xmm11

	movapd	-14 * SIZE(AO), %xmm0
	movapd	 -8 * SIZE(BO), %xmm2
	movapd	 -6 * SIZE(BO), %xmm3
	movapd	 -4 * SIZE(BO), %xmm4
	movapd	 -2 * SIZE(BO), %xmm5

	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm3
	mulpd	%xmm0, %xmm4
	mulpd	%xmm0, %xmm5

	addpd	%xmm2, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm4, %xmm10
	addpd	%xmm5, %xmm11

	movapd	-12 * SIZE(AO), %xmm0
	movapd	  0 * SIZE(BO), %xmm2
	movapd	  2 * SIZE(BO), %xmm3
	movapd	  4 * SIZE(BO), %xmm4
	movapd	  6 * SIZE(BO), %xmm5

	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm3
	mulpd	%xmm0, %xmm4
	mulpd	%xmm0, %xmm5

	addpd	%xmm2, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm4, %xmm10
	addpd	%xmm5, %xmm11

	movapd	-10 * SIZE(AO), %xmm0
	movapd	  8 * SIZE(BO), %xmm2
	movapd	 10 * SIZE(BO), %xmm3
	movapd	 12 * SIZE(BO), %xmm4
	movapd	 14 * SIZE(BO), %xmm5

	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm3
	mulpd	%xmm0, %xmm4
	mulpd	%xmm0, %xmm5

	addpd	%xmm2, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm4, %xmm10
	addpd	%xmm5, %xmm11

	subq	$ -8 * SIZE, AO
	subq	$-32 * SIZE, BO
	subq	$1, %rax
	jne    .L22
	ALIGN_4

.L25:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	BRANCH
	je .L29
	ALIGN_4

.L26:
	movapd	-16 * SIZE(AO), %xmm0
	movapd	-16 * SIZE(BO), %xmm2
	movapd	-14 * SIZE(BO), %xmm3
	movapd	-12 * SIZE(BO), %xmm4
	movapd	-10 * SIZE(BO), %xmm5

	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm3
	mulpd	%xmm0, %xmm4
	mulpd	%xmm0, %xmm5

	addpd	%xmm2, %xmm8
	addpd	%xmm3, %xmm9
	addpd	%xmm4, %xmm10
	addpd	%xmm5, %xmm11

	addq	$2 * SIZE, AO
	addq	$8 * SIZE, BO
	subq	$1, %rax
	jne    .L26
	ALIGN_4

.L29:
#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$2, %rax
#else
	subq	$4, %rax
#endif

	movq	AORIG, AO
	movq	BORIG, B
	leaq	16 * SIZE + BUFFER, BO

	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(B,  %rax, 4), B
	leaq	(BO, %rax, 8), BO
#endif

#if defined(LN) || defined(LT)
	movapd	%xmm8, %xmm0
	unpcklpd %xmm9, %xmm8
	unpckhpd %xmm9, %xmm0

	movapd	%xmm10, %xmm2
	unpcklpd %xmm11, %xmm10
	unpckhpd %xmm11, %xmm2

	movapd	 -16 * SIZE(B), %xmm9
	movapd	 -14 * SIZE(B), %xmm11
	movapd	 -12 * SIZE(B), %xmm13
	movapd	 -10 * SIZE(B), %xmm15

	subpd	%xmm8,  %xmm9
	subpd	%xmm10,  %xmm11
	subpd	%xmm0,  %xmm13
	subpd	%xmm2, %xmm15
#else
	movapd	 -16 * SIZE(AO), %xmm0
	movapd	 -14 * SIZE(AO), %xmm2
	movapd	 -12 * SIZE(AO), %xmm4
	movapd	 -10 * SIZE(AO), %xmm6

	subpd	%xmm8, %xmm0
	subpd	%xmm9, %xmm2
	subpd	%xmm10, %xmm4
	subpd	%xmm11, %xmm6
#endif

#ifdef LN
	movddup	 -13 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm13
	mulpd	 %xmm8, %xmm15

	movddup	 -14 * SIZE(AO), %xmm10
	movapd	 %xmm10, %xmm12
	mulpd	 %xmm13, %xmm10
	subpd	 %xmm10, %xmm9
	mulpd	 %xmm15, %xmm12
	subpd	 %xmm12, %xmm11

	movddup	 -16 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm9
	mulpd	 %xmm8, %xmm11
#endif

#ifdef LT
	movddup	 -16 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm9
	mulpd	 %xmm8, %xmm11

	movddup	 -15 * SIZE(AO), %xmm10
	movapd	 %xmm10, %xmm12
	mulpd	 %xmm9, %xmm10
	subpd	 %xmm10, %xmm13
	mulpd	 %xmm11, %xmm12
	subpd	 %xmm12, %xmm15

	movddup	 -13 * SIZE(AO), %xmm8
	mulpd	 %xmm8, %xmm13
	mulpd	 %xmm8, %xmm15
#endif

#ifdef RN
	movddup	 -16 * SIZE(B), %xmm8
	mulpd	 %xmm8, %xmm0

	movddup	 -15 * SIZE(B), %xmm9
	mulpd	 %xmm0, %xmm9
	subpd	 %xmm9, %xmm2
	movddup	 -14 * SIZE(B), %xmm10
	mulpd	 %xmm0, %xmm10
	subpd	 %xmm10, %xmm4
	movddup	 -13 * SIZE(B), %xmm11
	mulpd	 %xmm0, %xmm11
	subpd	 %xmm11, %xmm6

	movddup	 -11 * SIZE(B), %xmm8
	mulpd	 %xmm8, %xmm2
	movddup	 -10 * SIZE(B), %xmm9
	mulpd	 %xmm2, %xmm9
	subpd	 %xmm9, %xmm4
	movddup	  -9 * SIZE(B), %xmm10
	mulpd	 %xmm2, %xmm10
	subpd	 %xmm10, %xmm6

	movddup	  -6 * SIZE(B), %xmm8
	mulpd	 %xmm8, %xmm4

	movddup	  -5 * SIZE(B), %xmm9
	mulpd	 %xmm4, %xmm9
	subpd	 %xmm9, %xmm6

	movddup	  -1 * SIZE(B), %xmm8
	mulpd	 %xmm8, %xmm6
#endif

#ifdef RT
	movddup	-1 * SIZE(B), %xmm8
	mulpd	 %xmm8, %xmm6

	movddup	-2 * SIZE(B), %xmm9
	mulpd	 %xmm6, %xmm9
	subpd	 %xmm9, %xmm4
	movddup	-3 * SIZE(B), %xmm10
	mulpd	 %xmm6, %xmm10
	subpd	 %xmm10, %xmm2
	movddup	-4 * SIZE(B), %xmm11
	mulpd	 %xmm6, %xmm11
	subpd	 %xmm11, %xmm0

	movddup	-6 * SIZE(B), %xmm8
	mulpd	 %xmm8, %xmm4
	movddup	-7 * SIZE(B), %xmm9
	mulpd	 %xmm4, %xmm9
	subpd	 %xmm9, %xmm2
	movddup	-8 * SIZE(B), %xmm10
	mulpd	 %xmm4, %xmm10
	subpd	 %xmm10, %xmm0

	movddup	-11 * SIZE(B), %xmm8
	mulpd	 %xmm8, %xmm2
	movddup	-12 * SIZE(B), %xmm9
	mulpd	 %xmm2, %xmm9
	subpd	 %xmm9, %xmm0

	movddup	-16 * SIZE(B), %xmm8
	mulpd	 %xmm8, %xmm0
#endif

#ifdef LN
	subq	$2 * SIZE, CO1
	subq	$2 * SIZE, CO2
#endif

#if defined(LN) || defined(LT)
	movsd	%xmm9,  0 * SIZE(CO1)
	movsd	%xmm13,  1 * SIZE(CO1)

	movhpd	%xmm9,  0 * SIZE(CO2)
	movhpd	%xmm13,  1 * SIZE(CO2)

	movsd	%xmm11,  0 * SIZE(CO1, LDC, 2)
	movsd	%xmm15,  1 * SIZE(CO1, LDC, 2)

	movhpd	%xmm11,  0 * SIZE(CO2, LDC, 2)
	movhpd	%xmm15,  1 * SIZE(CO2, LDC, 2)
#else
	movsd	%xmm0,  0 * SIZE(CO1)
	movhpd	%xmm0,  1 * SIZE(CO1)

	movsd	%xmm2,  0 * SIZE(CO2)
	movhpd	%xmm2,  1 * SIZE(CO2)

	movsd	%xmm4,  0 * SIZE(CO1, LDC, 2)
	movhpd	%xmm4,  1 * SIZE(CO1, LDC, 2)

	movsd	%xmm6,  0 * SIZE(CO2, LDC, 2)
	movhpd	%xmm6,  1 * SIZE(CO2, LDC, 2)
#endif

#if defined(LN) || defined(LT)
	movapd	%xmm9,  -16 * SIZE(B)
	movapd	%xmm11,  -14 * SIZE(B)
	movapd	%xmm13,  -12 * SIZE(B)
	movapd	%xmm15,  -10 * SIZE(B)

	movddup	%xmm9,  %xmm8
	SHUFPD_3 %xmm9,  %xmm9
	movddup	%xmm11,  %xmm10
	SHUFPD_3 %xmm11,  %xmm11
	movddup	%xmm13,  %xmm12
	SHUFPD_3 %xmm13,  %xmm13
	movddup	%xmm15,  %xmm14
	SHUFPD_3 %xmm15,  %xmm15

	movapd	%xmm8, -16 * SIZE(BO)
	movapd	%xmm9, -14 * SIZE(BO)
	movapd	%xmm10, -12 * SIZE(BO)
	movapd	%xmm11, -10 * SIZE(BO)
	movapd	%xmm12,  -8 * SIZE(BO)
	movapd	%xmm13,  -6 * SIZE(BO)
	movapd	%xmm14,  -4 * SIZE(BO)
	movapd	%xmm15,  -2 * SIZE(BO)
#else
	movapd	%xmm0,  -16 * SIZE(AO)
	movapd	%xmm2, -14 * SIZE(AO)
	movapd	%xmm4, -12 * SIZE(AO)
	movapd	%xmm6, -10 * SIZE(AO)
#endif

#ifndef LN
	addq	$2 * SIZE, CO1
	addq	$2 * SIZE, CO2
#endif

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
#ifdef LT
	addq	$8 * SIZE, B
#endif
#endif

#ifdef LN
	subq	$2, KK
	movq	BORIG, B
#endif

#ifdef LT
	addq	$2, KK
#endif

#ifdef RT
       movq	K, %rax
       movq	BORIG, B
       salq	$1 + BASE_SHIFT, %rax
       addq	%rax, AORIG
#endif
	ALIGN_4	

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

.L31:
#ifdef LN
       movq	K, %rax
       salq	$0 + BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	movq	AORIG, AO
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
#endif

	leaq	16 * SIZE + BUFFER, BO

#if defined(LN) || defined(RT)
	movq	KK, %rax
	salq	$2 + BASE_SHIFT, %rax
	leaq	(BO, %rax, 2), BO
#endif	

	pxor	%xmm8, %xmm8
	pxor	%xmm9, %xmm9
	pxor	%xmm10, %xmm10
	pxor	%xmm11, %xmm11

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$2, %rax
	je	.L35
	ALIGN_4

.L32:
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	movsd	-16 * SIZE(AO), %xmm0
	movsd	-16 * SIZE(BO), %xmm2
	movsd	-14 * SIZE(BO), %xmm3
	movsd	-12 * SIZE(BO), %xmm4
	movsd	-10 * SIZE(BO), %xmm5

	mulsd	%xmm0, %xmm2
	mulsd	%xmm0, %xmm3
	mulsd	%xmm0, %xmm4
	mulsd	%xmm0, %xmm5

	addsd	%xmm2, %xmm8
	addsd	%xmm3, %xmm9
	addsd	%xmm4, %xmm10
	addsd	%xmm5, %xmm11

	movsd	-15 * SIZE(AO), %xmm0
	movsd	 -8 * SIZE(BO), %xmm2
	movsd	 -6 * SIZE(BO), %xmm3
	movsd	 -4 * SIZE(BO), %xmm4
	movsd	 -2 * SIZE(BO), %xmm5

	mulsd	%xmm0, %xmm2
	mulsd	%xmm0, %xmm3
	mulsd	%xmm0, %xmm4
	mulsd	%xmm0, %xmm5

	addsd	%xmm2, %xmm8
	addsd	%xmm3, %xmm9
	addsd	%xmm4, %xmm10
	addsd	%xmm5, %xmm11

	movsd	-14 * SIZE(AO), %xmm0
	movsd	  0 * SIZE(BO), %xmm2
	movsd	  2 * SIZE(BO), %xmm3
	movsd	  4 * SIZE(BO), %xmm4
	movsd	  6 * SIZE(BO), %xmm5

	mulsd	%xmm0, %xmm2
	mulsd	%xmm0, %xmm3
	mulsd	%xmm0, %xmm4
	mulsd	%xmm0, %xmm5

	addsd	%xmm2, %xmm8
	addsd	%xmm3, %xmm9
	addsd	%xmm4, %xmm10
	addsd	%xmm5, %xmm11

	movsd	-13 * SIZE(AO), %xmm0
	movsd	  8 * SIZE(BO), %xmm2
	movsd	 10 * SIZE(BO), %xmm3
	movsd	 12 * SIZE(BO), %xmm4
	movsd	 14 * SIZE(BO), %xmm5

	mulsd	%xmm0, %xmm2
	mulsd	%xmm0, %xmm3
	mulsd	%xmm0, %xmm4
	mulsd	%xmm0, %xmm5

	addsd	%xmm2, %xmm8
	addsd	%xmm3, %xmm9
	addsd	%xmm4, %xmm10
	addsd	%xmm5, %xmm11

	subq	$ -4 * SIZE, AO
	subq	$-32 * SIZE, BO
	subq	$1, %rax
	jne    .L32
	ALIGN_4

.L35:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	BRANCH
	je .L38
	ALIGN_4

.L36:
	movsd	-16 * SIZE(AO), %xmm0
	movsd	-16 * SIZE(BO), %xmm2
	movsd	-14 * SIZE(BO), %xmm3
	movsd	-12 * SIZE(BO), %xmm4
	movsd	-10 * SIZE(BO), %xmm5

	mulsd	%xmm0, %xmm2
	mulsd	%xmm0, %xmm3
	mulsd	%xmm0, %xmm4
	mulsd	%xmm0, %xmm5

	addsd	%xmm2, %xmm8
	addsd	%xmm3, %xmm9
	addsd	%xmm4, %xmm10
	addsd	%xmm5, %xmm11

	addq	$1 * SIZE, AO
	addq	$8 * SIZE, BO
	subq	$1, %rax
	jg	.L36
	ALIGN_4

.L38:
#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$1, %rax
#else
	subq	$4, %rax
#endif

	movq	AORIG, AO
	movq	BORIG, B
	leaq	16 * SIZE + BUFFER, BO

	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 4), B
	leaq	(BO, %rax, 8), BO
#endif

#if defined(LN) || defined(LT)
	movsd	 -16 * SIZE(B), %xmm12
	movsd	 -15 * SIZE(B), %xmm13
	movsd	 -14 * SIZE(B), %xmm14
	movsd	 -13 * SIZE(B), %xmm15
#else
	movsd	 -16 * SIZE(AO), %xmm12
	movsd	 -15 * SIZE(AO), %xmm13
	movsd	 -14 * SIZE(AO), %xmm14
	movsd	 -13 * SIZE(AO), %xmm15
#endif

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

#ifdef LN
	movsd	-16 * SIZE(AO), %xmm8

	mulsd	 %xmm8, %xmm12
	mulsd	 %xmm8, %xmm13
	mulsd	 %xmm8, %xmm14
	mulsd	 %xmm8, %xmm15
#endif

#ifdef LT
	movsd	-16 * SIZE(AO), %xmm8

	mulsd	 %xmm8, %xmm12
	mulsd	 %xmm8, %xmm13
	mulsd	 %xmm8, %xmm14
	mulsd	 %xmm8, %xmm15
#endif

#ifdef RN
	mulsd	-16 * SIZE(B), %xmm12
	movlpd	-15 * SIZE(B), %xmm9
	mulsd	 %xmm12, %xmm9
	subsd	 %xmm9, %xmm13
	movlpd	-14 * SIZE(B), %xmm10
	mulsd	 %xmm12, %xmm10
	subsd	 %xmm10, %xmm14
	movlpd  -13 * SIZE(B), %xmm11
	mulsd	 %xmm12, %xmm11
	subsd	 %xmm11, %xmm15

	mulsd	-11 * SIZE(B), %xmm13
	movlpd	-10 * SIZE(B), %xmm9
	mulsd	 %xmm13, %xmm9
	subsd	 %xmm9, %xmm14
	movlpd	 -9 * SIZE(B), %xmm10
	mulsd	 %xmm13, %xmm10
	subsd	 %xmm10, %xmm15

	mulsd	 -6 * SIZE(B), %xmm14
	movlpd	 -5 * SIZE(B), %xmm9
	mulsd	 %xmm14, %xmm9
	subsd	 %xmm9, %xmm15

	mulsd	 -1 * SIZE(B), %xmm15
#endif

#ifdef RT
	mulsd	 -1 * SIZE(B), %xmm15

	movlpd	 -2 * SIZE(B), %xmm9
	mulsd	 %xmm15, %xmm9
	subsd	 %xmm9, %xmm14
	movlpd	 -3 * SIZE(B), %xmm10
	mulsd	 %xmm15, %xmm10
	subsd	 %xmm10, %xmm13
	movlpd	 -4 * SIZE(B), %xmm11
	mulsd	 %xmm15, %xmm11
	subsd	 %xmm11, %xmm12

	mulsd	 -6 * SIZE(B), %xmm14

	movlpd	 -7 * SIZE(B), %xmm9
	mulsd	 %xmm14, %xmm9
	subsd	 %xmm9, %xmm13
	movlpd	 -8 * SIZE(B), %xmm10
	mulsd	 %xmm14, %xmm10
	subsd	 %xmm10, %xmm12

	mulsd	-11 * SIZE(B), %xmm13

	movlpd	-12 * SIZE(B), %xmm9
	mulsd	 %xmm13, %xmm9
	subsd	 %xmm9, %xmm12

	mulsd	-16 * SIZE(B), %xmm12
#endif

#ifdef LN
	subq	$1 * SIZE, CO1
	subq	$1 * SIZE, CO2
#endif

	movsd	%xmm12,  0 * SIZE(CO1)
	movsd	%xmm13,  0 * SIZE(CO2)
	movsd	%xmm14,  0 * SIZE(CO1, LDC, 2)
	movsd	%xmm15,  0 * SIZE(CO2, LDC, 2)

#if defined(LN) || defined(LT)
	movsd	%xmm12,  -16 * SIZE(B)
	movsd	%xmm13,  -15 * SIZE(B)
	movsd	%xmm14,  -14 * SIZE(B)
	movsd	%xmm15,  -13 * SIZE(B)

	movsd	%xmm12,  -16 * SIZE(BO)
	movsd	%xmm12,  -15 * SIZE(BO)
	movsd	%xmm13,  -14 * SIZE(BO)
	movsd	%xmm13,  -13 * SIZE(BO)
	movsd	%xmm14,  -12 * SIZE(BO)
	movsd	%xmm14,  -11 * SIZE(BO)
	movsd	%xmm15,  -10 * SIZE(BO)
	movsd	%xmm15,   -9 * SIZE(BO)
#else
	movsd	%xmm12,  -16 * SIZE(AO)
	movsd	%xmm13,  -15 * SIZE(AO)
	movsd	%xmm14,  -14 * SIZE(AO)
	movsd	%xmm15,  -13 * SIZE(AO)
#endif

#ifndef LN
	addq	$1 * SIZE, CO1
	addq	$1 * SIZE, CO2
#endif

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
#ifdef LT
	addq	$4 * SIZE, B
#endif
#endif

#ifdef LN
	subq	$1, KK
	movq	BORIG, B
#endif

#ifdef LT
	addq	$1, KK
#endif

#ifdef RT
       movq	K, %rax
       movq	BORIG, B
       salq	$0 + BASE_SHIFT, %rax
       addq	%rax, AORIG
#endif
	ALIGN_4	
	
.L39:
#ifdef LN
       leaq	(, K, SIZE), %rax
       leaq	(B, %rax, 4), B
#endif

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(B,  %rax, 4), B
#endif

#ifdef RN
	addq	$4, KK
#endif

#ifdef RT
	subq	$4, KK
#endif

	decq	J			# j --
	jg	.L01
	ALIGN_4

.L999:
	movq	%r15, %rsp

	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), %xmm