gemm_kernel_2x4_barcelona.s

Optimized GotoBLAS libraries

	leal	(, LDC, 4), %eax
	addl	%eax, C			# c += 4 * ldc
	decl	J			# j --
	jg	.L01
	ALIGN_4

.L30:
	testl	$2, N
	je	.L60
	ALIGN_2

.L31:
#if defined(TRMMKERNEL) && defined(LEFT)
	movl	OFFSET, %eax
	movl	%eax, KK
#endif	

	movl	C, CO		# coffset = c
	movl	A, AO		# aoffset = a

	movl	M,  I
	sarl	$1, I	# 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, BO
#else
	movl	KK, %eax
	leal	(, %eax, SIZE), %eax
	leal	(AO, %eax, 2), AO
	leal	(B,  %eax, 2), BO
#endif	

	movddup	-16 * SIZE(AO), %xmm0
	pxor	%xmm4, %xmm4
	prefetchw 1 * SIZE(CO)
	pxor	%xmm5, %xmm5
	prefetchw 1 * SIZE(CO, LDC)
	pxor	%xmm6, %xmm6
	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	$2, %eax
#else
	addl	$2, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L45
	ALIGN_4

.L42:
	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AO)
	mulpd	-16 * SIZE(BO), %xmm0
	movddup	-15 * SIZE(AO), %xmm1
	addpd	%xmm0, %xmm4
	mulpd	-16 * SIZE(BO), %xmm1
	movddup	-14 * SIZE(AO), %xmm0
	addpd	%xmm1, %xmm5

	mulpd	-14 * SIZE(BO), %xmm0
	movddup	-13 * SIZE(AO), %xmm1
	addpd	%xmm0, %xmm6
	mulpd	-14 * SIZE(BO), %xmm1
	movddup	-12 * SIZE(AO), %xmm0
	addpd	%xmm1, %xmm7

	mulpd	-12 * SIZE(BO), %xmm0
	movddup	-11 * SIZE(AO), %xmm1
	addpd	%xmm0, %xmm4
	mulpd	-12 * SIZE(BO), %xmm1
	movddup	-10 * SIZE(AO), %xmm0
	addpd	%xmm1, %xmm5

	mulpd	-10 * SIZE(BO), %xmm0
	movddup	 -9 * SIZE(AO), %xmm1
	addpd	%xmm0, %xmm6
	mulpd	-10 * SIZE(BO), %xmm1
	movddup	 -8 * SIZE(AO), %xmm0
	addpd	%xmm1, %xmm7

	prefetcht0 (PREFETCHSIZE  + 8) * SIZE(AO)

	mulpd	 -8 * SIZE(BO), %xmm0
	movddup	 -7 * SIZE(AO), %xmm1
	addpd	%xmm0, %xmm4
	mulpd	 -8 * SIZE(BO), %xmm1
	movddup	 -6 * SIZE(AO), %xmm0
	addpd	%xmm1, %xmm5

	mulpd	 -6 * SIZE(BO), %xmm0
	movddup	 -5 * SIZE(AO), %xmm1
	addpd	%xmm0, %xmm6
	mulpd	 -6 * SIZE(BO), %xmm1
	movddup	 -4 * SIZE(AO), %xmm0
	addpd	%xmm1, %xmm7

	mulpd	 -4 * SIZE(BO), %xmm0
	movddup	 -3 * SIZE(AO), %xmm1
	addpd	%xmm0, %xmm4
	mulpd	 -4 * SIZE(BO), %xmm1
	movddup	 -2 * SIZE(AO), %xmm0
	addpd	%xmm1, %xmm5

	mulpd	 -2 * SIZE(BO), %xmm0
	movddup	 -1 * SIZE(AO), %xmm1
	addpd	%xmm0, %xmm6
	mulpd	 -2 * SIZE(BO), %xmm1
	movddup	  0 * SIZE(AO), %xmm0
	addpd	%xmm1, %xmm7

	subl   $-16 * SIZE, AO
	subl   $-16 * SIZE, BO
	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	-16 * SIZE(BO), %xmm0
	movddup	-15 * SIZE(AO), %xmm1
	addpd	%xmm0, %xmm4
	mulpd	-16 * SIZE(BO), %xmm1
	movddup	-14 * SIZE(AO), %xmm0
	addpd	%xmm1, %xmm5

	addl	$2 * SIZE, AO
	addl	$2 * SIZE, BO
	decl	%eax
	jg	.L46
	ALIGN_4

.L48:
#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO), %xmm0
	movhpd	0 * SIZE(CO, LDC), %xmm0

	movsd	1 * SIZE(CO), %xmm1
	movhpd	1 * SIZE(CO, LDC), %xmm1
#endif

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

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

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

	movlpd	%xmm4, 0 * SIZE(CO)
	movlpd	%xmm5, 1 * SIZE(CO)

	movhpd	%xmm4, 0 * SIZE(CO, LDC)
	movhpd	%xmm5, 1 * SIZE(CO, LDC)

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

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

	addl	$2 * SIZE, CO		# coffset += 2
	decl	I			# i --
	jg	.L41
	ALIGN_4

.L50:
	movl	M,  I
	testl	$1, I	# i = (m >> 2)
	jle	.L59

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

	movl	B, BO
#else
	movl	KK, %eax
	leal	(, %eax, SIZE), %eax
	leal	(AO, %eax, 1), AO
	leal	(B,  %eax, 2), BO
#endif	

	movddup -16 * SIZE(AO), %xmm0
	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5
	pxor	%xmm6, %xmm6
	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	$2, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L55
	ALIGN_4

.L52:
	mulpd	-16 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-15 * SIZE(AO), %xmm0

	mulpd	-14 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-14 * SIZE(AO), %xmm0

	mulpd	-12 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-13 * SIZE(AO), %xmm0

	mulpd	-10 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-12 * SIZE(AO), %xmm0

	mulpd	 -8 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-11 * SIZE(AO), %xmm0

	mulpd	 -6 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-10 * SIZE(AO), %xmm0

	mulpd	 -4 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	 -9 * SIZE(AO), %xmm0

	mulpd	 -2 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	 -8 * SIZE(AO), %xmm0

	subl   $ -8 * SIZE, AO
	subl   $-16 * SIZE, BO

	decl   %eax
	jne    .L52
	ALIGN_4

.L55:
	movddup	ALPHA,  %xmm3

#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L58

.L56:
	mulpd	-16 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-15 * SIZE(AO), %xmm0

	subl	$-1 * SIZE, AO
	subl	$-2 * SIZE, BO
	decl	%eax
	jg	.L56
	ALIGN_4

.L58:
	addpd	%xmm6, %xmm4
	addpd	%xmm7, %xmm5
	addpd	%xmm5, %xmm4

	mulpd	%xmm3, %xmm4

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO), %xmm0
	movhpd	0 * SIZE(CO, LDC), %xmm0

	addpd	%xmm0, %xmm4
#endif

	movlpd	%xmm4, 0 * SIZE(CO)
	movhpd	%xmm4, 0 * SIZE(CO, LDC, 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	(AO, %eax, 1), AO
	leal	(BO, %eax, 2), BO
#endif

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

.L59:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$2, KK
#endif

	movl	BO, B

	leal	(, LDC, 2), %eax
	addl	%eax, C			# c += 4 * ldc
	ALIGN_4

.L60:
	testl	$1, N
	je	.L999

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

	movl	C, CO		# coffset = c
	movl	A, AO		# aoffset = a

	movl	M,  I
	sarl	$1, I	# i = (m >> 2)
	jle	.L80
	ALIGN_4

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

	movl	B, BO
#else
	movl	KK, %eax
	leal	(, %eax, SIZE), %eax
	leal	(AO, %eax, 2), AO
	leal	(B,  %eax, 1), BO
#endif	

	movddup	-16 * SIZE(BO), %xmm0
	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5
	pxor	%xmm6, %xmm6
	pxor	%xmm7, %xmm7

	prefetchw   1 * SIZE(CO)

#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	$1, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L75
	ALIGN_4

.L72:
	PREFETCH (PREFETCHSIZE  + 0) * SIZE(AO)

	mulpd	-16 * SIZE(AO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-15 * SIZE(BO), %xmm0

	mulpd	-14 * SIZE(AO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-14 * SIZE(BO), %xmm0

	mulpd	-12 * SIZE(AO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-13 * SIZE(BO), %xmm0

	mulpd	-10 * SIZE(AO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-12 * SIZE(BO), %xmm0

	PREFETCH (PREFETCHSIZE  + 8) * SIZE(AO)

	mulpd	 -8 * SIZE(AO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-11 * SIZE(BO), %xmm0

	mulpd	 -6 * SIZE(AO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-10 * SIZE(BO), %xmm0

	mulpd	 -4 * SIZE(AO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	 -9 * SIZE(BO), %xmm0

	mulpd	 -2 * SIZE(AO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	 -8 * SIZE(BO), %xmm0

	subl   $-16 * SIZE, AO
	subl   $ -8 * SIZE, BO
	decl   %eax
	jne    .L72
	ALIGN_4

.L75:
	movddup	ALPHA,  %xmm3
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L78
	ALIGN_3

.L76:
	mulpd	-16 * SIZE(AO), %xmm0
	addpd	%xmm0, %xmm4
	movddup	-15 * SIZE(BO), %xmm0

	addl	$2 * SIZE, AO
	addl	$1 * SIZE, BO
	decl	%eax
	jg	.L76
	ALIGN_4

.L78:
	mulpd	%xmm3, %xmm4

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO), %xmm0
	movhpd	1 * SIZE(CO), %xmm0

	addpd	%xmm0, %xmm4
#endif

	movsd	%xmm4, 0 * SIZE(CO)
	movhpd	%xmm4, 1 * SIZE(CO)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movl	K, %eax
	subl	KKK, %eax
	leal	(,%eax, SIZE), %eax
	leal	(AO, %eax, 2), AO
	leal	(BO, %eax, 1), BO
#endif

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

	addl	$2 * SIZE, CO		# coffset += 2
	decl	I			# i --
	jg	.L71
	ALIGN_4

.L80:
	movl	M,  I
	testl	$1, I	# i = (m >> 2)
	jle	.L999

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

	movl	B, BO
#else
	movl	KK, %eax
	leal	(, %eax, SIZE), %eax
	leal	(AO, %eax, 1), AO
	leal	(B,  %eax, 1), BO
#endif	

	movaps	-16 * SIZE(AO), %xmm0
	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5
	pxor	%xmm6, %xmm6
	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	$1, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L85
	ALIGN_4

.L82:
	PREFETCH (PREFETCHSIZE  + 0) * SIZE(AO)

	mulpd	-16 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm4
	movapd	-14 * SIZE(AO), %xmm0

	mulpd	-14 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm5
	movapd	-12 * SIZE(AO), %xmm0

	mulpd	-12 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm6
	movapd	-10 * SIZE(AO), %xmm0

	mulpd	-10 * SIZE(BO), %xmm0
	addpd	%xmm0, %xmm7
	movapd	 -8 * SIZE(AO), %xmm0

	subl   $-8 * SIZE, AO
	subl   $-8 * SIZE, BO
	decl   %eax
	jne    .L82
	ALIGN_4

.L85:
	movddup	ALPHA,  %xmm3
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L88

.L86:
	mulsd	-16 * SIZE(BO), %xmm0
	addsd	%xmm0, %xmm4
	movsd	-15 * SIZE(AO), %xmm0

	addl	$1 * SIZE, AO
	addl	$1 * SIZE, BO
	decl	%eax
	jg	.L86
	ALIGN_4

.L88:
	addpd	%xmm5, %xmm4
	addpd	%xmm7, %xmm6
	addpd	%xmm6, %xmm4

	haddpd	%xmm4, %xmm4
	mulsd	%xmm3, %xmm4

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO), %xmm0
	addsd	%xmm0, %xmm4
#endif
	movsd	%xmm4, 0 * SIZE(CO)
	ALIGN_4

.L999:
	popl	%ebx
	popl	%esi
	popl	%edi
	popl	%ebp

	addl	$ARGS, %esp
	ret

	EPILOGUE