gemm_kernel_4x4_sse2.s

Optimized GotoBLAS libraries

	movsd	60 * SIZE(BO), %xmm7
	mulsd	%xmm0, %xmm7
	mulsd	62 * SIZE(BO), %xmm0
	addsd	%xmm7, %xmm10
	movsd	88 * SIZE(BO), %xmm7
	addsd	%xmm0, %xmm11
	movsd	-8 * SIZE(AO), %xmm0

	addq   $ 8 * SIZE, AO
	addq   $64 * SIZE, BO
	decq   %rax
	jne    .L32
	ALIGN_3

.L35:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movsd	ALPHA, %xmm7
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L38
	ALIGN_3

.L36:
	mulsd	%xmm0, %xmm1
	addsd	%xmm1, %xmm8
	movsd	 2 * SIZE(BO), %xmm1
	mulsd	%xmm0, %xmm1
	addsd	%xmm1, %xmm9
	movsd	 4 * SIZE(BO), %xmm1
	mulsd	%xmm0, %xmm1
	mulsd	 6 * SIZE(BO), %xmm0
	addsd	%xmm1, %xmm10
	movsd	 8 * SIZE(BO), %xmm1
	addsd	%xmm0, %xmm11
	movsd	-15 * SIZE(AO), %xmm0

	addq	$1 * SIZE, AO		# aoffset  += 4
	addq	$8 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L36
	ALIGN_3

.L38:
#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movsd	0 * SIZE(CO2), %xmm2
	movsd	0 * SIZE(CO1, LDC, 2), %xmm4
	movsd	0 * SIZE(CO2, LDC, 2), %xmm6
#endif

	mulsd	%xmm7, %xmm8
	mulsd	%xmm7, %xmm9
	mulsd	%xmm7, %xmm10
	mulsd	%xmm7, %xmm11

#ifndef TRMMKERNEL
	addsd	%xmm0,  %xmm8
	addsd	%xmm2, %xmm9
	addsd	%xmm4, %xmm10
	addsd	%xmm6, %xmm11
#endif

	movsd	%xmm8, 0 * SIZE(CO1)
	movsd	%xmm9, 0 * SIZE(CO2)
	movsd	%xmm10, 0 * SIZE(CO1, LDC, 2)
	movsd	%xmm11, 0 * SIZE(CO2, LDC, 2)

#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, 8), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$1, KK
#endif
	ALIGN_3	
	
.L39:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$4, KK
#endif

	leaq	(C, LDC, 4), C		# c += 4 * ldc
	decq	J			# j --
	jg	.L01
	ALIGN_3

.L40:
	testq	$3, N
	je	.L999

	testq	$2, N
	je	.L80
	ALIGN_4

.L41:
/* Copying to Sub Buffer */
	leaq	BUFFER, BO
	
#if defined(TRMMKERNEL) && defined(LEFT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif	

	movq	K, %rax
	sarq	$2, %rax
	jle	.L43
	ALIGN_3
	
.L42:
	PREFETCH	 56 * SIZE(B)

	movq	 0 * SIZE(B), %mm0
	movq	 1 * SIZE(B), %mm1
	movq	 2 * SIZE(B), %mm2
	movq	 3 * SIZE(B), %mm3
	movq	 4 * SIZE(B), %mm4
	movq	 5 * SIZE(B), %mm5
	movq	 6 * SIZE(B), %mm6
	movq	 7 * SIZE(B), %mm7

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

	movq	%mm0, -16 * SIZE(BO)
	movq	%mm0, -15 * SIZE(BO)
	movq	%mm1, -14 * SIZE(BO)
	movq	%mm1, -13 * SIZE(BO)
	movq	%mm2, -12 * SIZE(BO)
	movq	%mm2, -11 * SIZE(BO)
	movq	%mm3, -10 * SIZE(BO)
	movq	%mm3,  -9 * SIZE(BO)
	movq	%mm4,  -8 * SIZE(BO)
	movq	%mm4,  -7 * SIZE(BO)
	movq	%mm5,  -6 * SIZE(BO)
	movq	%mm5,  -5 * SIZE(BO)
	movq	%mm6,  -4 * SIZE(BO)
	movq	%mm6,  -3 * SIZE(BO)
	movq	%mm7,  -2 * SIZE(BO)
	movq	%mm7,  -1 * SIZE(BO)

	decq	%rax
	jne	.L42
	ALIGN_3

.L43:
	movq	K, %rax
	andq	$3, %rax
	BRANCH
	jle	.L50
	ALIGN_3

.L44:
	movq	 0 * SIZE(B), %mm0
	movq	 1 * SIZE(B), %mm1

	movq	%mm0,  0 * SIZE(BO)
	movq	%mm0,  1 * SIZE(BO)
	movq	%mm1,  2 * SIZE(BO)
	movq	%mm1,  3 * SIZE(BO)

	addq	$2 * SIZE, B
	addq	$4 * SIZE, BO
	decq	%rax
	jne	.L44
	ALIGN_3
	
.L50:
	movq	C, CO1			# coffset1 = c
	leaq	(C, LDC, 1), CO2	# coffset2 = c + ldc
	movq	A, AO		# aoffset = a

	movq	M,  I
	sarq	$2, I	# i = (m >> 2)
	jle	.L60
	ALIGN_3

.L51:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

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

	movapd	-16 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movapd	 0 * SIZE(BO), %xmm1
	pxor	%xmm9, %xmm9
	movapd	 -8 * SIZE(AO), %xmm2
	pxor	%xmm12, %xmm12
	movapd	 8 * SIZE(BO), %xmm3
	pxor	%xmm13, %xmm13

	movapd	 0 * SIZE(AO), %xmm4
	movapd	16 * SIZE(BO), %xmm5
	movapd	 8 * SIZE(AO), %xmm6
	movapd	24 * SIZE(BO), %xmm7

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

#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	$3, %rax
	je	.L55
	ALIGN_3

.L52:
	mulpd	%xmm0, %xmm1
	PREFETCH	(PREFETCHSIZE     +  0) * SIZE(AO)
	mulpd	 2 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm8
	movapd	 0 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm9
	movapd	-14 * SIZE(AO), %xmm0
	mulpd	%xmm0, %xmm1
	mulpd	 2 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm12
	movapd	 4 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm13
	movapd	-12 * SIZE(AO), %xmm0

	mulpd	%xmm0, %xmm1
	mulpd	 6 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm8
	movapd	 4 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm9
	movapd	-10 * SIZE(AO), %xmm0
	mulpd	%xmm0, %xmm1
	mulpd	 6 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm12
	movapd	32 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm13
	movapd	16 * SIZE(AO), %xmm0

	PREFETCH	(PREFETCHSIZE     +  8) * SIZE(AO)
	mulpd	%xmm2, %xmm3
	mulpd	10 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm8
	movapd	 8 * SIZE(BO), %xmm3
	addpd	%xmm2, %xmm9
	movapd	-6 * SIZE(AO), %xmm2
	mulpd	%xmm2, %xmm3
	mulpd	10 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm12
	movapd	12 * SIZE(BO), %xmm3
	addpd	%xmm2, %xmm13
	movapd	-4 * SIZE(AO), %xmm2

	mulpd	%xmm2, %xmm3
	mulpd	14 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm8
	movapd	12 * SIZE(BO), %xmm3
	addpd	%xmm2, %xmm9
	movapd	-2 * SIZE(AO), %xmm2
	mulpd	%xmm2, %xmm3
	mulpd	14 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm12
	movapd	40 * SIZE(BO), %xmm3
	addpd	%xmm2, %xmm13
	movapd	24 * SIZE(AO), %xmm2

	PREFETCH	(PREFETCHSIZE     + 16) * SIZE(AO)
	mulpd	%xmm4, %xmm5
	mulpd	18 * SIZE(BO), %xmm4
	addpd	%xmm5, %xmm8
	movapd	16 * SIZE(BO), %xmm5
	addpd	%xmm4, %xmm9
	movapd	 2 * SIZE(AO), %xmm4
	mulpd	%xmm4, %xmm5
	mulpd	18 * SIZE(BO), %xmm4
	addpd	%xmm5, %xmm12
	movapd	20 * SIZE(BO), %xmm5
	addpd	%xmm4, %xmm13
	movapd	 4 * SIZE(AO), %xmm4

	mulpd	%xmm4, %xmm5
	mulpd	22 * SIZE(BO), %xmm4
	addpd	%xmm5, %xmm8
	movapd	20 * SIZE(BO), %xmm5
	addpd	%xmm4, %xmm9
	movapd	 6 * SIZE(AO), %xmm4
	mulpd	%xmm4, %xmm5
	mulpd	22 * SIZE(BO), %xmm4
	addpd	%xmm5, %xmm12
	movapd	48 * SIZE(BO), %xmm5
	addpd	%xmm4, %xmm13
	movapd	32 * SIZE(AO), %xmm4

	PREFETCH	(PREFETCHSIZE     + 24) * SIZE(AO)
	mulpd	%xmm6, %xmm7
	mulpd	26 * SIZE(BO), %xmm6
	addpd	%xmm7, %xmm8
	movapd	24 * SIZE(BO), %xmm7
	addpd	%xmm6, %xmm9
	movapd	10 * SIZE(AO), %xmm6
	mulpd	%xmm6, %xmm7
	mulpd	26 * SIZE(BO), %xmm6
	addpd	%xmm7, %xmm12
	movapd	28 * SIZE(BO), %xmm7
	addpd	%xmm6, %xmm13
	movapd	12 * SIZE(AO), %xmm6

	mulpd	%xmm6, %xmm7
	mulpd	30 * SIZE(BO), %xmm6
	addpd	%xmm7, %xmm8
	movapd	28 * SIZE(BO), %xmm7
	addpd	%xmm6, %xmm9
	movapd	14 * SIZE(AO), %xmm6
	mulpd	%xmm6, %xmm7
	mulpd	30 * SIZE(BO), %xmm6
	addpd	%xmm7, %xmm12
	movapd	56 * SIZE(BO), %xmm7
	addpd	%xmm6, %xmm13
	movapd	40 * SIZE(AO), %xmm6

	addq   $32 * SIZE, AO
	addq   $32 * SIZE, BO
	decq   %rax
	jne    .L52
	ALIGN_3

.L55:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movapd	ALPHA, %xmm7
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L59
	ALIGN_3

.L56:
	movapd	 0 * SIZE(BO), %xmm1
	mulpd	%xmm0, %xmm1
	addpd	%xmm1, %xmm8
	mulpd	 2 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm9
	movapd	-14 * SIZE(AO), %xmm0
	movapd	 0 * SIZE(BO), %xmm1
	mulpd	%xmm0, %xmm1
	addpd	%xmm1, %xmm12
	mulpd	 2 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm13
	movapd	-12 * SIZE(AO), %xmm0

	addq	$4 * SIZE, AO		# aoffset  += 4
	addq	$4 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L56
	ALIGN_3

.L59:
#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhpd	1 * SIZE(CO1), %xmm0
	movsd	2 * SIZE(CO1), %xmm1
	movhpd	3 * SIZE(CO1), %xmm1
	movsd	0 * SIZE(CO2), %xmm2
	movhpd	1 * SIZE(CO2), %xmm2
	movsd	2 * SIZE(CO2), %xmm3
	movhpd	3 * SIZE(CO2), %xmm3
#endif

	mulpd	%xmm7, %xmm8
	mulpd	%xmm7, %xmm9
	mulpd	%xmm7, %xmm12
	mulpd	%xmm7, %xmm13

#ifndef TRMMKERNEL
	addpd	%xmm0,  %xmm8
	addpd	%xmm1,  %xmm12
	addpd	%xmm2, %xmm9
	addpd	%xmm3, %xmm13
#endif

	movsd	%xmm8, 0 * SIZE(CO1)
	movhpd	%xmm8, 1 * SIZE(CO1)
	movsd	%xmm12, 2 * SIZE(CO1)
	movhpd	%xmm12, 3 * SIZE(CO1)
	movsd	%xmm9, 0 * SIZE(CO2)
	movhpd	%xmm9, 1 * SIZE(CO2)
	movsd	%xmm13, 2 * SIZE(CO2)
	movhpd	%xmm13, 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, 4), BO
#endif

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

	addq	$4 * SIZE, CO1		# coffset += 4
	addq	$4 * SIZE, CO2		# coffset += 4
	decq	I			# i --
	jg	.L51
	ALIGN_3	

.L60:
	testq	$2, M
	je	.L70
	ALIGN_3

.L61:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

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

	movapd	-16 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movapd	 0 * SIZE(BO), %xmm1
	pxor	%xmm9, %xmm9
	movapd	 -8 * SIZE(AO), %xmm2
	pxor	%xmm10, %xmm10
	movapd	 8 * SIZE(BO), %xmm3
	pxor	%xmm11, %xmm11

	movapd	16 * SIZE(BO), %xmm5
	movapd	24 * SIZE(BO), %xmm7

#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	$3, %rax
	je	.L65
	ALIGN_3

.L62:
	mulpd	%xmm0, %xmm1
	PREFETCH	(PREFETCHSIZE     +  0) * SIZE(AO)
	mulpd	 2 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm8
	movapd	 4 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm9
	movapd	-14 * SIZE(AO), %xmm0

	mulpd	%xmm0, %xmm1
	mulpd	 6 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm10
	movapd	32 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm11
	movapd	-12 * SIZE(AO), %xmm0

	mulpd	%xmm0, %xmm3
	mulpd	10 * SIZE(BO), %xmm0
	addpd	%xmm3, %xmm8
	movapd	12 * SIZE(BO), %xmm3
	addpd	%xmm0, %xmm9
	movapd	-10 * SIZE(AO), %xmm0

	mulpd	%xmm0, %xmm3
	mulpd	14 * SIZE(BO), %xmm0
	addpd	%xmm3, %xmm10
	movapd	40 * SIZE(BO), %xmm3
	addpd	%xmm0, %xmm11
	movapd	 0 * SIZE(AO), %xmm0

	PREFETCH	(PREFETCHSIZE     +  8) * SIZE(AO)
	mulpd	%xmm2, %xmm5
	mulpd	18 * SIZE(BO), %xmm2
	addpd	%xmm5, %xmm8
	movapd	20 * SIZE(BO), %xmm5
	addpd	%xmm2, %xmm9
	movapd	-6 * SIZE(AO), %xmm2

	mulpd	%xmm2, %xmm5
	mulpd	22 * SIZE(BO), %xmm2
	addpd	%xmm5, %xmm10
	movapd	48 * SIZE(BO), %xmm5
	addpd	%xmm2, %xmm11
	movapd	-4 * SIZE(AO), %xmm2

	mulpd	%xmm2, %xmm7
	mulpd	26 * SIZE(BO), %xmm2
	addpd	%xmm7, %xmm8
	movapd	28 * SIZE(BO), %xmm7
	addpd	%xmm2, %xmm9
	movapd	-2 * SIZE(AO), %xmm2

	mulpd	%xmm2, %xmm7
	mulpd	30 * SIZE(BO), %xmm2
	addpd	%xmm7, %xmm10
	movapd	56 * SIZE(BO), %xmm7
	addpd	%xmm2, %xmm11
	movapd	 8 * SIZE(AO), %xmm2

	addq   $16 * SIZE, AO
	addq   $32 * SIZE, BO
	decq   %rax
	jne    .L62
	ALIGN_3

.L65:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movapd	ALPHA, %xmm7
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L69
	ALIGN_3

.L66:
	mulpd	%xmm0, %xmm1
	mulpd	 2 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm8
	movapd	 4 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm9
	movapd	-14 * SIZE(AO), %xmm0

	addq	$2 * SIZE, AO		# aoffset  += 4
	addq	$4 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L66
	ALIGN_3

.L69:
#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhpd	1 * SIZE(CO1), %xmm0
	movsd	0 * SIZE(CO2), %xmm2
	movhpd	1 * SIZE(CO2), %xmm2
#endif

	addpd	%xmm10, %xmm8
	addpd	%xmm11, %xmm9

	mulpd	%xmm7, %xmm8
	mulpd	%xmm7, %xmm9

#ifndef TRMMKERNEL
	addpd	%xmm0,  %xmm8
	addpd	%xmm2, %xmm9
#endif

	movsd	%xmm8, 0 * SIZE(CO1)
	movhpd	%xmm8, 1 * SIZE(CO1)
	movsd	%xmm9, 0 * SIZE(CO2)
	movhpd	%xmm9, 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, 4), BO
#endif

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

	addq	$2 * SIZE, CO1		# coffset += 4
	addq	$2 * SIZE, CO2		# coffset += 4
	ALIGN_3	

.L70:
	testq	$1, M
	je	.L79
	ALIGN_3

.L71:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

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

	movsd	-16 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movsd	 0 * SIZE(BO), %xmm1
	pxor	%xmm9, %xmm9
	movsd	-12 * SIZE(AO), %xmm2
	pxor	%xmm10, %xmm10
	movsd	 8 * SIZE(BO), %xmm3
	pxor	%xmm11, %xmm11

	movsd	16 * SIZE(BO), %xmm5
	movsd	24 * SIZE(BO), %xmm7

#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	$3, %rax
	je	.L75
	ALIGN_3

.L72:
	mulsd	%xmm0, %xmm1
	PREFETCH	(PREFETCHSIZE     +  0) * SIZE(AO)
	mulsd	 2 * SIZE(BO), %xmm0
	addsd	%xmm1, %xmm8
	movsd	 4 * SIZE(BO), %xmm1
	addsd	%xmm0, %xmm9