gemm_kernel_2x4_sse3.s

Optimized GotoBLAS libraries

	mulpd	 %xmm1, %xmm2
	movapd	10 * SIZE(AA), %xmm1
	addpd	 %xmm2, %xmm7
	movddup	20 * SIZE(BB), %xmm2
	mulpd	 %xmm1, %xmm2
	addpd	 %xmm2, %xmm4
	movddup	21 * SIZE(BB), %xmm2
	mulpd	 %xmm1, %xmm2
	addpd	 %xmm2, %xmm5
	movddup	22 * SIZE(BB), %xmm2
	mulpd	 %xmm1, %xmm2
	addpd	 %xmm2, %xmm6
	movddup	23 * SIZE(BB), %xmm2
	mulpd	 %xmm1, %xmm2
	movapd	12 * SIZE(AA), %xmm1
	addpd	 %xmm2, %xmm7
	movddup	32 * SIZE(BB), %xmm2
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm4
	movddup	25 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm5
	movddup	26 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm6
	movddup	27 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	movapd	14 * SIZE(AA), %xmm1
	addpd	 %xmm3, %xmm7
	movddup	28 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm4
	movddup	29 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm5
	movddup	30 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm6
	movddup	31 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	movapd	24 * SIZE(AA), %xmm1
	addpd	 %xmm3, %xmm7
	movddup	40 * SIZE(BB), %xmm3

	addl   $32 * SIZE, BB
	addl   $16 * SIZE, AA
	decl   %eax
	jne    .L12
	ALIGN_4
#endif

.L15:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	movddup	ALPHA,  %xmm3
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L18
	ALIGN_3

.L16:
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm4
	movddup	 1 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm5
	movddup	 2 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm6
	movddup	 3 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	 2 * SIZE(AA), %xmm0
	addpd	 %xmm2, %xmm7
	movddup	 4 * SIZE(BB), %xmm2

	addl	$2 * SIZE, AA
	addl	$4 * SIZE, BB
	decl	%eax
	jg	.L16
	ALIGN_4

.L18:
	leal	(LDC, LDC, 2), %eax

#ifndef TRMMKERNEL
	SHUFPD_2 %xmm0, %xmm0
	SHUFPD_2 %xmm1, %xmm1
	SHUFPD_2 %xmm2, %xmm2
	SHUFPD_2 %xmm3, %xmm3

	mulpd	%xmm3, %xmm4
	movsd	0 * SIZE(%esi), %xmm0
	movhpd	1 * SIZE(%esi), %xmm0
	mulpd	%xmm3, %xmm5
	movsd	0 * SIZE(%esi, LDC, 1), %xmm1
	movhpd	1 * SIZE(%esi, LDC, 1), %xmm1
	mulpd	%xmm3, %xmm6
	movsd	0 * SIZE(%esi, LDC, 2), %xmm2
	movhpd	1 * SIZE(%esi, LDC, 2), %xmm2
	mulpd	%xmm3, %xmm7
	movsd	0 * SIZE(%esi, %eax, 1), %xmm3
	movhpd	1 * SIZE(%esi, %eax, 1), %xmm3

	addpd	%xmm0, %xmm4
	addpd	%xmm1, %xmm5
	addpd	%xmm2, %xmm6
	addpd	%xmm3, %xmm7
#else
	mulpd	%xmm3, %xmm4
	mulpd	%xmm3, %xmm5
	mulpd	%xmm3, %xmm6
	mulpd	%xmm3, %xmm7
#endif

	movsd	%xmm4, 0 * SIZE(%esi)
	movhpd	%xmm4, 1 * SIZE(%esi)
	movsd	%xmm5, 0 * SIZE(%esi, LDC, 1)
	movhpd	%xmm5, 1 * SIZE(%esi, LDC, 1)
	movsd	%xmm6, 0 * SIZE(%esi, LDC, 2)
	movhpd	%xmm6, 1 * SIZE(%esi, LDC, 2)
	movsd	%xmm7, 0 * SIZE(%esi, %eax, 1)
	movhpd	%xmm7, 1 * SIZE(%esi, %eax, 1)

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

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

	addl	$2 * SIZE, %esi		# coffset += 2
	decl	%ebx			# i --
	jg	.L11
	ALIGN_4

.L20:
	movl	M,  %ebx
	testl	$1, %ebx	# i = (m >> 2)
	jle	.L29


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

	movl	B, BB
#else
	movl	KK, %eax
	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 1), AA
	leal	(B,  %eax, 4), BB
#endif	

	movddup	 0 * SIZE(AA), %xmm0
	pxor	%xmm4, %xmm4
	movddup	 8 * SIZE(AA), %xmm1
	pxor	%xmm5, %xmm5
	movapd	 0 * SIZE(BB), %xmm2
	pxor	%xmm6, %xmm6
	movapd	 8 * SIZE(BB), %xmm3
	pxor	%xmm7, %xmm7

#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	$4, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$4, %eax
	je	.L25
	ALIGN_4

.L22:
	PREFETCH  (PREFETCHSIZE + 0) * SIZE(AA)
	mulpd	 %xmm0, %xmm2
	mulpd	 2 * SIZE(BB), %xmm0
	addpd	 %xmm2, %xmm4
	movapd	 4 * SIZE(BB), %xmm2
	addpd	 %xmm0, %xmm5
	movddup	 1 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm2
	mulpd	 6 * SIZE(BB), %xmm0
	addpd	 %xmm2, %xmm6
	movapd	16 * SIZE(BB), %xmm2
	addpd	 %xmm0, %xmm7
	movddup	 2 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm3
	mulpd	10 * SIZE(BB), %xmm0
	addpd	 %xmm3, %xmm4
	movapd	12 * SIZE(BB), %xmm3
	addpd	 %xmm0, %xmm5
	movddup	 3 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm3
	mulpd	14 * SIZE(BB), %xmm0
	addpd	 %xmm3, %xmm6
	movapd	24 * SIZE(BB), %xmm3
	addpd	 %xmm0, %xmm7
	movddup	 4 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm2
	mulpd	18 * SIZE(BB), %xmm0
	addpd	 %xmm2, %xmm4
	movapd	20 * SIZE(BB), %xmm2
	addpd	 %xmm0, %xmm5
	movddup	 5 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm2
	mulpd	22 * SIZE(BB), %xmm0
	addpd	 %xmm2, %xmm6
	movapd	32 * SIZE(BB), %xmm2
	addpd	 %xmm0, %xmm7
	movddup	 6 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm3
	mulpd	26 * SIZE(BB), %xmm0
	addpd	 %xmm3, %xmm4
	movapd	28 * SIZE(BB), %xmm3
	addpd	 %xmm0, %xmm5
	movddup	 7 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm3
	mulpd	30 * SIZE(BB), %xmm0
	addpd	 %xmm3, %xmm6
	movapd	40 * SIZE(BB), %xmm3
	addpd	 %xmm0, %xmm7
	movddup	16 * SIZE(AA), %xmm0
	mulpd	 %xmm1, %xmm2
	mulpd	34 * SIZE(BB), %xmm1
	addpd	 %xmm2, %xmm4
	movapd	36 * SIZE(BB), %xmm2
	addpd	 %xmm1, %xmm5
	movddup	 9 * SIZE(AA), %xmm1
	mulpd	 %xmm1, %xmm2
	mulpd	38 * SIZE(BB), %xmm1
	addpd	 %xmm2, %xmm6
	movapd	48 * SIZE(BB), %xmm2
	addpd	 %xmm1, %xmm7
	movddup	10 * SIZE(AA), %xmm1
	mulpd	 %xmm1, %xmm3
	mulpd	42 * SIZE(BB), %xmm1
	addpd	 %xmm3, %xmm4
	movapd	44 * SIZE(BB), %xmm3
	addpd	 %xmm1, %xmm5
	movddup	11 * SIZE(AA), %xmm1
	mulpd	 %xmm1, %xmm3
	mulpd	46 * SIZE(BB), %xmm1
	addpd	 %xmm3, %xmm6
	movapd	56 * SIZE(BB), %xmm3
	addpd	 %xmm1, %xmm7
	movddup	12 * SIZE(AA), %xmm1
	mulpd	 %xmm1, %xmm2
	mulpd	50 * SIZE(BB), %xmm1
	addpd	 %xmm2, %xmm4
	movapd	52 * SIZE(BB), %xmm2
	addpd	 %xmm1, %xmm5
	movddup	13 * SIZE(AA), %xmm1
	mulpd	 %xmm1, %xmm2
	mulpd	54 * SIZE(BB), %xmm1
	addpd	 %xmm2, %xmm6
	movapd	64 * SIZE(BB), %xmm2
	addpd	 %xmm1, %xmm7
	movddup	14 * SIZE(AA), %xmm1
	mulpd	 %xmm1, %xmm3
	mulpd	58 * SIZE(BB), %xmm1
	addpd	 %xmm3, %xmm4
	movapd	60 * SIZE(BB), %xmm3
	addpd	 %xmm1, %xmm5
	movddup	15 * SIZE(AA), %xmm1
	mulpd	 %xmm1, %xmm3
	mulpd	62 * SIZE(BB), %xmm1
	addpd	 %xmm3, %xmm6
	movapd	72 * SIZE(BB), %xmm3
	addpd	 %xmm1, %xmm7
	movddup	24 * SIZE(AA), %xmm1

	addl   $16 * SIZE, AA
	addl   $64 * SIZE, BB
	decl   %eax
	jne    .L22
	ALIGN_4

.L25:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	movddup	ALPHA,  %xmm3
	andl	$15, %eax		# if (k & 1)
	BRANCH
	je .L28

.L26:
	mulpd	 %xmm0, %xmm2
	mulpd	 2 * SIZE(BB), %xmm0
	addpd	 %xmm2, %xmm4
	movapd	 4 * SIZE(BB), %xmm2
	addpd	 %xmm0, %xmm5
	movddup	 1 * SIZE(AA), %xmm0

	addl	$1 * SIZE, AA
	addl	$4 * SIZE, BB

	decl	%eax
	jg	.L26
	ALIGN_4

.L28:
	leal	(%esi, LDC, 1), %eax

	addpd	%xmm6, %xmm4
	addpd	%xmm7, %xmm5

	mulpd	%xmm3, %xmm4
	mulpd	%xmm3, %xmm5

#ifndef TRMMKERNEL

#ifdef PENTIUM4
	SHUFPD_2 %xmm0, %xmm0
	SHUFPD_2 %xmm1, %xmm1
#endif

	movsd	0 * SIZE(%esi), %xmm0
	movhpd	0 * SIZE(%eax), %xmm0
	movsd	0 * SIZE(%esi, LDC, 2), %xmm1
	movhpd	0 * SIZE(%eax, LDC, 2), %xmm1

	addpd	%xmm0, %xmm4
	addpd	%xmm1, %xmm5
#endif

	movsd	%xmm4, 0 * SIZE(%esi)
	movhpd	%xmm4, 0 * SIZE(%eax)
	movsd	%xmm5, 0 * SIZE(%esi, LDC, 2)
	movhpd	%xmm5, 0 * SIZE(%eax, LDC, 2)

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

#if defined(TRMMKERNEL) && defined(LEFT)
	addl	$1, KK
#endif
	ALIGN_4

.L29:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$4, KK
#endif

	leal	(, LDC, 4), %eax
	movl	BB, B
	addl	%eax, C			# c += 4 * ldc
	decl	J			# j --
	jg	.L10
	ALIGN_4

.L30:
	testl	$2, N
	je	.L60

	movl	C, %esi		# coffset = c
	movl	A, AA		# aoffset = a

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

	movl	M,  %ebx
	sarl	$1, %ebx	# i = (m >> 2)
	jle	.L50
	ALIGN_4

.L41:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movl	B, BB
#else
	movl	KK, %eax
	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 2), AA
	leal	(B,  %eax, 2), BB
#endif	

	movapd	 0 * SIZE(AA), %xmm0
	pxor	%xmm4, %xmm4
	movapd	 8 * SIZE(AA), %xmm1
	pxor	%xmm5, %xmm5
	movddup	 0 * SIZE(BB), %xmm2
	pxor	%xmm6, %xmm6
	movddup	 8 * SIZE(BB), %xmm3
	pxor	%xmm7, %xmm7

#ifdef HAVE_3DNOW
	prefetchw 2 * SIZE(%esi)
	prefetchw 2 * SIZE(%esi, LDC)
#endif

#ifdef PENTIUM4
	prefetchnta	3 * SIZE(%esi)
	prefetchnta	3 * SIZE(%esi, LDC)
#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	$2, %eax
#else
	addl	$2, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L45
	ALIGN_4

.L42:
	PREFETCH  (PREFETCHSIZE + 0) * SIZE(AA)
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm4
	movddup	 1 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	 2 * SIZE(AA), %xmm0
	addpd	 %xmm2, %xmm5
	movddup	 2 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm6
	movddup	 3 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	 4 * SIZE(AA), %xmm0
	addpd	 %xmm2, %xmm7
	movddup	 4 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm4
	movddup	 5 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	 6 * SIZE(AA), %xmm0
	addpd	 %xmm2, %xmm5
	movddup	 6 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm6
	movddup	 7 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	16 * SIZE(AA), %xmm0
	addpd	 %xmm2, %xmm7
	movddup	16 * SIZE(BB), %xmm2
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm4
	movddup	 9 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	movapd	10 * SIZE(AA), %xmm1
	addpd	 %xmm3, %xmm5
	movddup	10 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm6
	movddup	11 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	movapd	12 * SIZE(AA), %xmm1
	addpd	 %xmm3, %xmm7
	movddup	12 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm4
	movddup	13 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	movapd	14 * SIZE(AA), %xmm1
	addpd	 %xmm3, %xmm5
	movddup	14 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm6
	movddup	15 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	movapd	24 * SIZE(AA), %xmm1
	addpd	 %xmm3, %xmm7
	movddup	24 * SIZE(BB), %xmm3

	addl   $16 * SIZE, AA
	addl   $16 * SIZE, BB
	decl   %eax
	jne    .L42
	ALIGN_4

.L45:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	movddup	ALPHA,  %xmm3
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L48
	ALIGN_3

.L46:
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm4
	movddup	 1 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2