trsm_kernel_lt_4x4_sse2.s

Optimized GotoBLAS libraries

	leaq	BUFFER, BO

#ifdef RT
       movq	K, %rax
       salq	$1 + BASE_SHIFT, %rax
       subq	%rax, B
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	movq	B, BORIG
	leaq	(, %rax, SIZE), %rax
	leaq	(B,  %rax, 2), B
	leaq	(BO, %rax, 4), BO
#endif	

#ifdef LT
	movq	OFFSET, %rax
	movq	%rax, KK
#endif

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$2, %rax
	jle	.L43
	ALIGN_4
	
.L42:
	PREFETCH	 56 * SIZE(B)

	movsd	 0 * SIZE(B), %xmm0
	movsd	 1 * SIZE(B), %xmm1
	movsd	 2 * SIZE(B), %xmm2
	movsd	 3 * SIZE(B), %xmm3
	movsd	 4 * SIZE(B), %xmm4
	movsd	 5 * SIZE(B), %xmm5
	movsd	 6 * SIZE(B), %xmm6
	movsd	 7 * SIZE(B), %xmm7

	addq	$ 8 * SIZE, B
	addq	$16 * SIZE, BO

	movsd	%xmm0, -16 * SIZE(BO)
	movsd	%xmm0, -15 * SIZE(BO)
	movsd	%xmm1, -14 * SIZE(BO)
	movsd	%xmm1, -13 * SIZE(BO)
	movsd	%xmm2, -12 * SIZE(BO)
	movsd	%xmm2, -11 * SIZE(BO)
	movsd	%xmm3, -10 * SIZE(BO)
	movsd	%xmm3,  -9 * SIZE(BO)
	movsd	%xmm4,  -8 * SIZE(BO)
	movsd	%xmm4,  -7 * SIZE(BO)
	movsd	%xmm5,  -6 * SIZE(BO)
	movsd	%xmm5,  -5 * SIZE(BO)
	movsd	%xmm6,  -4 * SIZE(BO)
	movsd	%xmm6,  -3 * SIZE(BO)
	movsd	%xmm7,  -2 * SIZE(BO)
	movsd	%xmm7,  -1 * SIZE(BO)

	decq	%rax
	jne	.L42
	ALIGN_4

.L43:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$3, %rax
	BRANCH
	jle	.L50
	ALIGN_4

.L44:
	movsd	 0 * SIZE(B), %xmm0
	movsd	 1 * SIZE(B), %xmm1

	movsd	%xmm0,  0 * SIZE(BO)
	movsd	%xmm0,  1 * SIZE(BO)
	movsd	%xmm1,  2 * SIZE(BO)
	movsd	%xmm1,  3 * SIZE(BO)

	addq	$2 * SIZE, B
	addq	$4 * SIZE, BO
	decq	%rax
	jne	.L44
	ALIGN_4
	
.L50:
#if defined(LT) || defined(RN)
	movq	A, AO
#else
	movq	A, AORIG
#endif

#ifdef RT
       leaq	(, LDC, 2), %rax
       subq	%rax, C
#endif

	movq	C, CO1			# coffset1 = c
	leaq	(C, LDC, 1), CO2	# coffset2 = c + ldc
#ifndef RT
	leaq	(C, LDC, 2), C
#endif

	movq	M,  I
	sarq	$2, I	# i = (m >> 2)
	jle	.L60
	ALIGN_4

.L51:
#ifdef LN
       movq	K, %rax
       salq	$2 + 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, 4), AO
#endif

	leaq	BUFFER, BO

#if defined(LN) || defined(RT)
	movq	KK, %rax
	salq	$1 + BASE_SHIFT, %rax
	leaq	(BO, %rax, 2), BO
#endif	

	movapd	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movapd	 0 * SIZE(BO), %xmm9
	pxor	%xmm1, %xmm1
	movapd	 8 * SIZE(AO), %xmm10
	pxor	%xmm4, %xmm4
	movapd	 8 * SIZE(BO), %xmm11
	pxor	%xmm5, %xmm5

	movapd	16 * SIZE(AO), %xmm12
	movapd	16 * SIZE(BO), %xmm13
	movapd	24 * SIZE(AO), %xmm14
	movapd	24 * SIZE(BO), %xmm15

	PREFETCHW      4 * SIZE(CO1)
	PREFETCHW      4 * SIZE(CO2)

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$3, %rax
	je	.L55
	ALIGN_4

.L52:
	mulpd	%xmm8, %xmm9
	PREFETCH	(PREFETCHSIZE     +  0) * SIZE(AO)
	mulpd	 2 * SIZE(BO), %xmm8
	addpd	%xmm9, %xmm0
	movapd	 0 * SIZE(BO), %xmm9
	addpd	%xmm8, %xmm1
	movapd	 2 * SIZE(AO), %xmm8
	mulpd	%xmm8, %xmm9
	mulpd	 2 * SIZE(BO), %xmm8
	addpd	%xmm9, %xmm4
	movapd	 4 * SIZE(BO), %xmm9
	addpd	%xmm8, %xmm5
	movapd	 4 * SIZE(AO), %xmm8

	mulpd	%xmm8, %xmm9
	mulpd	 6 * SIZE(BO), %xmm8
	addpd	%xmm9, %xmm0
	movapd	 4 * SIZE(BO), %xmm9
	addpd	%xmm8, %xmm1
	movapd	 6 * SIZE(AO), %xmm8
	mulpd	%xmm8, %xmm9
	mulpd	 6 * SIZE(BO), %xmm8
	addpd	%xmm9, %xmm4
	movapd	32 * SIZE(BO), %xmm9
	addpd	%xmm8, %xmm5
	movapd	32 * SIZE(AO), %xmm8

	PREFETCH	(PREFETCHSIZE     +  8) * SIZE(AO)
	mulpd	%xmm10, %xmm11
	mulpd	10 * SIZE(BO), %xmm10
	addpd	%xmm11, %xmm0
	movapd	 8 * SIZE(BO), %xmm11
	addpd	%xmm10, %xmm1
	movapd	10 * SIZE(AO), %xmm10
	mulpd	%xmm10, %xmm11
	mulpd	10 * SIZE(BO), %xmm10
	addpd	%xmm11, %xmm4
	movapd	12 * SIZE(BO), %xmm11
	addpd	%xmm10, %xmm5
	movapd	12 * SIZE(AO), %xmm10

	mulpd	%xmm10, %xmm11
	mulpd	14 * SIZE(BO), %xmm10
	addpd	%xmm11, %xmm0
	movapd	12 * SIZE(BO), %xmm11
	addpd	%xmm10, %xmm1
	movapd	14 * SIZE(AO), %xmm10
	mulpd	%xmm10, %xmm11
	mulpd	14 * SIZE(BO), %xmm10
	addpd	%xmm11, %xmm4
	movapd	40 * SIZE(BO), %xmm11
	addpd	%xmm10, %xmm5
	movapd	40 * SIZE(AO), %xmm10

	PREFETCH	(PREFETCHSIZE     + 16) * SIZE(AO)
	mulpd	%xmm12, %xmm13
	mulpd	18 * SIZE(BO), %xmm12
	addpd	%xmm13, %xmm0
	movapd	16 * SIZE(BO), %xmm13
	addpd	%xmm12, %xmm1
	movapd	18 * SIZE(AO), %xmm12
	mulpd	%xmm12, %xmm13
	mulpd	18 * SIZE(BO), %xmm12
	addpd	%xmm13, %xmm4
	movapd	20 * SIZE(BO), %xmm13
	addpd	%xmm12, %xmm5
	movapd	20 * SIZE(AO), %xmm12

	mulpd	%xmm12, %xmm13
	mulpd	22 * SIZE(BO), %xmm12
	addpd	%xmm13, %xmm0
	movapd	20 * SIZE(BO), %xmm13
	addpd	%xmm12, %xmm1
	movapd	22 * SIZE(AO), %xmm12
	mulpd	%xmm12, %xmm13
	mulpd	22 * SIZE(BO), %xmm12
	addpd	%xmm13, %xmm4
	movapd	48 * SIZE(BO), %xmm13
	addpd	%xmm12, %xmm5
	movapd	48 * SIZE(AO), %xmm12

	PREFETCH	(PREFETCHSIZE     + 24) * SIZE(AO)
	mulpd	%xmm14, %xmm15
	mulpd	26 * SIZE(BO), %xmm14
	addpd	%xmm15, %xmm0
	movapd	24 * SIZE(BO), %xmm15
	addpd	%xmm14, %xmm1
	movapd	26 * SIZE(AO), %xmm14
	mulpd	%xmm14, %xmm15
	mulpd	26 * SIZE(BO), %xmm14
	addpd	%xmm15, %xmm4
	movapd	28 * SIZE(BO), %xmm15
	addpd	%xmm14, %xmm5
	movapd	28 * SIZE(AO), %xmm14

	mulpd	%xmm14, %xmm15
	mulpd	30 * SIZE(BO), %xmm14
	addpd	%xmm15, %xmm0
	movapd	28 * SIZE(BO), %xmm15
	addpd	%xmm14, %xmm1
	movapd	30 * SIZE(AO), %xmm14
	mulpd	%xmm14, %xmm15
	mulpd	30 * SIZE(BO), %xmm14
	addpd	%xmm15, %xmm4
	movapd	56 * SIZE(BO), %xmm15
	addpd	%xmm14, %xmm5
	movapd	56 * SIZE(AO), %xmm14

	addq   $32 * SIZE, AO
	addq   $32 * SIZE, BO
	decq   %rax
	jne    .L52
	ALIGN_4

.L55:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L59
	ALIGN_4

.L56:
	movapd	 0 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm0
	mulpd	 2 * SIZE(BO), %xmm8
	addpd	%xmm8, %xmm1
	movapd	 2 * SIZE(AO), %xmm8
	movapd	 0 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm4
	mulpd	 2 * SIZE(BO), %xmm8
	addpd	%xmm8, %xmm5
	movapd	 4 * SIZE(AO), %xmm8

	addq	$4 * SIZE, AO		# aoffset  += 4
	addq	$4 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L56
	ALIGN_4

.L59:
#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$4, %rax
#else
	subq	$2, %rax
#endif

	movq	AORIG, AO
	movq	BORIG, B
	leaq	BUFFER, BO

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

#if defined(LN) || defined(LT)
	movapd	%xmm0, %xmm8
	unpcklpd %xmm1, %xmm0
	unpckhpd %xmm1, %xmm8

	movapd	%xmm4, %xmm12
	unpcklpd %xmm5, %xmm4
	unpckhpd %xmm5, %xmm12

	movapd	 0 * SIZE(B), %xmm1
	movapd	 2 * SIZE(B), %xmm5
	movapd	 4 * SIZE(B), %xmm9
	movapd	 6 * SIZE(B), %xmm13

	subpd	%xmm0,  %xmm1
	subpd	%xmm8,  %xmm5
	subpd	%xmm4,  %xmm9
	subpd	%xmm12, %xmm13
#else
	movapd	 0 * SIZE(AO), %xmm8
	movapd	 2 * SIZE(AO), %xmm9
	movapd	 4 * SIZE(AO), %xmm10
	movapd	 6 * SIZE(AO), %xmm11

	subpd	%xmm0, %xmm8
	subpd	%xmm4, %xmm9
	subpd	%xmm1, %xmm10
	subpd	%xmm5, %xmm11
#endif

#ifdef LN
	movlpd	15 * SIZE(AO), %xmm0
	movhpd	15 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm13
	movlpd	14 * SIZE(AO), %xmm2
	movhpd	14 * SIZE(AO), %xmm2
	mulpd	 %xmm13, %xmm2
	subpd	 %xmm2, %xmm9
	movlpd	13 * SIZE(AO), %xmm4
	movhpd	13 * SIZE(AO), %xmm4
	mulpd	 %xmm13, %xmm4
	subpd	 %xmm4, %xmm5
	movlpd	12 * SIZE(AO), %xmm6
	movhpd	12 * SIZE(AO), %xmm6
	mulpd	 %xmm13, %xmm6
	subpd	 %xmm6, %xmm1

	movlpd	10 * SIZE(AO), %xmm0
	movhpd	10 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm9
	movlpd	 9 * SIZE(AO), %xmm2
	movhpd	 9 * SIZE(AO), %xmm2
	mulpd	 %xmm9, %xmm2
	subpd	 %xmm2, %xmm5
	movlpd	 8 * SIZE(AO), %xmm4
	movhpd	 8 * SIZE(AO), %xmm4
	mulpd	 %xmm9, %xmm4
	subpd	 %xmm4, %xmm1

	movlpd	 5 * SIZE(AO), %xmm0
	movhpd	 5 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm5
	movlpd	 4 * SIZE(AO), %xmm2
	movhpd	 4 * SIZE(AO), %xmm2
	mulpd	 %xmm5, %xmm2
	subpd	 %xmm2, %xmm1

	movlpd	 0 * SIZE(AO), %xmm0
	movhpd	 0 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm1
#endif

#ifdef LT
	movlpd	 0 * SIZE(AO), %xmm0
	movhpd	 0 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm1
	movlpd	 1 * SIZE(AO), %xmm2
	movhpd	 1 * SIZE(AO), %xmm2
	mulpd	 %xmm1, %xmm2
	subpd	 %xmm2, %xmm5
	movlpd	 2 * SIZE(AO), %xmm4
	movhpd	 2 * SIZE(AO), %xmm4
	mulpd	 %xmm1, %xmm4
	subpd	 %xmm4, %xmm9
	movlpd	 3 * SIZE(AO), %xmm6
	movhpd	 3 * SIZE(AO), %xmm6
	mulpd	 %xmm1, %xmm6
	subpd	 %xmm6, %xmm13


	movlpd	 5 * SIZE(AO), %xmm0
	movhpd	 5 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm5

	movlpd	 6 * SIZE(AO), %xmm2
	movhpd	 6 * SIZE(AO), %xmm2
	mulpd	 %xmm5, %xmm2
	subpd	 %xmm2, %xmm9
	movlpd	 7 * SIZE(AO), %xmm4
	movhpd	 7 * SIZE(AO), %xmm4
	mulpd	 %xmm5, %xmm4
	subpd	 %xmm4, %xmm13

	movlpd	10 * SIZE(AO), %xmm0
	movhpd	10 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm9
	movlpd	11 * SIZE(AO), %xmm2
	movhpd	11 * SIZE(AO), %xmm2
	mulpd	 %xmm9, %xmm2
	subpd	 %xmm2, %xmm13

	movlpd	15 * SIZE(AO), %xmm0
	movhpd	15 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm13
#endif

#ifdef RN
	movlpd	 0 * SIZE(B), %xmm0
	movhpd	 0 * SIZE(B), %xmm0
	mulpd	 %xmm0, %xmm8
	mulpd	 %xmm0, %xmm9

	movlpd	 1 * SIZE(B), %xmm1
	movhpd	 1 * SIZE(B), %xmm1
	mulpd	 %xmm8, %xmm1
	subpd	 %xmm1, %xmm10
	movlpd	 1 * SIZE(B), %xmm1
	movhpd	 1 * SIZE(B), %xmm1
	mulpd	 %xmm9, %xmm1
	subpd	 %xmm1, %xmm11

	movlpd	 3 * SIZE(B), %xmm0
	movhpd	 3 * SIZE(B), %xmm0
	mulpd	 %xmm0, %xmm10
	mulpd	 %xmm0, %xmm11
#endif

#ifdef RT
	movlpd	 3 * SIZE(B), %xmm0
	movhpd	 3 * SIZE(B), %xmm0
	mulpd	 %xmm0, %xmm10
	mulpd	 %xmm0, %xmm11

	movlpd	 2 * SIZE(B), %xmm1
	movhpd	 2 * SIZE(B), %xmm1
	mulpd	 %xmm10, %xmm1
	subpd	 %xmm1, %xmm8
	movlpd	 2 * SIZE(B), %xmm1
	movhpd	 2 * SIZE(B), %xmm1
	mulpd	 %xmm11, %xmm1
	subpd	 %xmm1, %xmm9

	movlpd	 0 * SIZE(B), %xmm0
	movhpd	 0 * SIZE(B), %xmm0
	mulpd	 %xmm0, %xmm8
	mulpd	 %xmm0, %xmm9
#endif

#ifdef LN
	subq	$4 * SIZE, CO1
	subq	$4 * SIZE, CO2
#endif

#if defined(LN) || defined(LT)
	movsd	%xmm1,  0 * SIZE(CO1)
	movsd	%xmm5,  1 * SIZE(CO1)
	movsd	%xmm9,  2 * SIZE(CO1)
	movsd	%xmm13, 3 * SIZE(CO1)

	movhpd	%xmm1,  0 * SIZE(CO2)
	movhpd	%xmm5,  1 * SIZE(CO2)
	movhpd	%xmm9,  2 * SIZE(CO2)
	movhpd	%xmm13, 3 * SIZE(CO2)
#else
	movsd	%xmm8,  0 * SIZE(CO1)
	movhpd	%xmm8,  1 * SIZE(CO1)
	movsd	%xmm9,  2 * SIZE(CO1)
	movhpd	%xmm9,  3 * SIZE(CO1)

	movsd	%xmm10,  0 * SIZE(CO2)
	movhpd	%xmm10,  1 * SIZE(CO2)
	movsd	%xmm11,  2 * SIZE(CO2)
	movhpd	%xmm11,  3 * SIZE(CO2)
#endif

#if defined(LN) || defined(LT)
	movapd	%xmm1,   0 * SIZE(B)
	movapd	%xmm5,   2 * SIZE(B)
	movapd	%xmm9,   4 * SIZE(B)
	movapd	%xmm13,  6 * SIZE(B)

	movlpd	%xmm1,   0 * SIZE(BO)
	movlpd	%xmm1,   1 * SIZE(BO)
	movhpd	%xmm1,   2 * SIZE(BO)
	movhpd	%xmm1,   3 * SIZE(BO)
	movlpd	%xmm5,   4 * SIZE(BO)
	movlpd	%xmm5,   5 * SIZE(BO)
	movhpd	%xmm5,   6 * SIZE(BO)
	movhpd	%xmm5,   7 * SIZE(BO)
	movlpd	%xmm9,   8 * SIZE(BO)
	movlpd	%xmm9,   9 * SIZE(BO)
	movhpd	%xmm9,  10 * SIZE(BO)
	movhpd	%xmm9,  11 * SIZE(BO)
	movlpd	%xmm13, 12 * SIZE(BO)
	movlpd	%xmm13, 13 * SIZE(BO)
	movhpd	%xmm13, 14 * SIZE(BO)
	movhpd	%xmm13, 15 * SIZE(BO)
#else
	movapd	%xmm8,   0 * SIZE(AO)
	movapd	%xmm9,   2 * SIZE(AO)
	movapd	%xmm10,  4 * SIZE(AO)
	movapd	%xmm11,  6 * SIZE(AO)
#endif

#ifndef LN
	addq	$4 * SIZE, CO1
	addq	$4 * SIZE, CO2
#endif

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
#ifdef LT
	addq	$8 * SIZE, B
#endif
#endif

#ifdef LN
	subq	$4, KK
	movq	BORIG, B
#endif

#ifdef LT
	addq	$4, KK
#endif

#ifdef RT
       movq	K, %rax
       movq	BORIG, B
       salq	$2 + BASE_SHIFT, %rax
       addq	%rax, AORIG
#endif

	decq	I			# i --
	jg	.L51
	ALIGN_4	

.L60:
	testq	$2, M
	je	.L70
	ALIGN_4

.L61:
#ifdef LN
       movq	K, %rax
       salq	$1 + 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, 2), AO
#endif

	leaq	BUFFER, BO

#if defined(LN) || defined(RT)
	movq	KK, %rax
	salq	$1 + BASE_SHIFT, %rax
	leaq	(BO, %rax, 2), BO
#endif	

	movapd	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movapd	 0 * SIZE(BO), %xmm9
	pxor	%xmm1, %xmm1
	movapd	 8 * SIZE(AO), %xmm10
	pxor	%xmm2, %xmm2
	movapd	 8 * SIZE(BO), %xmm11
	pxor	%xmm3, %xmm3

	movapd	16 * SIZE(BO), %xmm13
	movapd	24 * SIZE(BO), %xmm15

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$3, %rax
	je	.L65
	ALIGN_4

.L62:
	mulpd	%xmm8, %xmm9
	PREFETCH	(PREFETCHSIZE     +  0) * SIZE(AO)
	mulpd	 2 * SIZE(BO), %xmm8
	addpd	%xmm9, %xmm0
	movapd	 4 * SIZE(BO), %xmm9
	addpd	%xmm8, %xmm1
	movapd	 2 * SIZE(AO), %xmm8

	mulpd	%xmm8, %xmm9
	mulpd	 6 * SIZE(BO), %xmm8
	addpd	%xmm9, %xmm2
	movapd	32 * SIZE(BO), %xmm9
	addpd	%xmm8, %xmm3
	movapd	 4 * SIZE(AO), %xmm8

	mulpd	%xmm8, %xmm11
	mulpd	10 * SIZE(BO), %xmm8
	addpd	%xmm11, %xmm0
	movapd	12 * SIZE(BO), %xmm11
	addpd	%xmm8, %xmm1
	movapd	 6 * SIZE(AO), %xmm8

	mulpd	%xmm8, %xmm11
	mulpd	14 * SIZE(BO), %xmm8
	addpd	%xmm11, %xmm2
	movapd	40 * SIZE(BO), %xmm11
	addpd	%xmm8, %xmm3
	movapd	16 * SIZE(AO), %xmm8

	PREFETCH	(PREFETCHSIZE     +  8) * SIZE(AO)
	mulpd	%xmm10, %xmm13
	mulpd	18 * SIZE(BO), %xmm10
	addpd	%xmm13, %xmm0
	movapd	20 * SIZE(BO), %xmm13
	addpd	%xmm10, %xmm1
	movapd	10 * SIZE(AO), %xmm10

	mulpd	%xmm10, %xmm13
	mulpd	22 * SIZE(BO), %xmm10
	addpd	%xmm13, %xmm2
	movapd	48 * SIZE(BO), %xmm13
	addpd	%xmm10, %xmm3
	movapd	12 * SIZE(AO), %xmm10

	mulpd	%xmm10, %xmm15
	mulpd	26 * SIZE(BO), %xmm10