gemm_kernel_4x4_penryn.s

Optimized GotoBLAS libraries

#endif

	pxor	%xmm4, %xmm4
	movsd	-32 * SIZE(AA), %xmm0
	pxor	%xmm5, %xmm5
	movsd	-32 * SIZE(BB), %xmm1

#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	.L65
	ALIGN_4

.L62:
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AA)

	pshufd	$0x00, %xmm0, %xmm2
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm4
	movsd	-30 * SIZE(BB), %xmm1

	pshufd	$0x55, %xmm0, %xmm2
	movsd	-30 * SIZE(AA), %xmm0
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm5
	movsd	-28 * SIZE(BB), %xmm1

	pshufd	$0x00, %xmm0, %xmm2
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm4
	movsd	-26 * SIZE(BB), %xmm1

	pshufd	$0x55, %xmm0, %xmm2
	movsd	-28 * SIZE(AA), %xmm0
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm5
	movsd	-24 * SIZE(BB), %xmm1

	pshufd	$0x00, %xmm0, %xmm2
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm4
	movsd	-22 * SIZE(BB), %xmm1

	pshufd	$0x55, %xmm0, %xmm2
	movsd	-26 * SIZE(AA), %xmm0
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm5
	movsd	-20 * SIZE(BB), %xmm1

	pshufd	$0x00, %xmm0, %xmm2
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm4
	movsd	-18 * SIZE(BB), %xmm1

	pshufd	$0x55, %xmm0, %xmm2
	movsd	-24 * SIZE(AA), %xmm0
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm5
	movsd	-16 * SIZE(BB), %xmm1

	subl   $ -8 * SIZE, AA
	subl   $-16 * SIZE, BB

	subl   $1, %eax
	jne    .L62
	ALIGN_4

.L65:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$7, %eax
	BRANCH
	je .L68
	ALIGN_4

.L66:
	pshufd	$0x00, %xmm0, %xmm2
	movss	-31 * SIZE(AA), %xmm0
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm4
	movsd	-30 * SIZE(BB), %xmm1

	addl	$1 * SIZE, AA
	addl	$2 * SIZE, BB
	decl	%eax
	jg	.L66
	ALIGN_4

.L68:
	movss	ALPHA,  %xmm1

	addps	%xmm5, %xmm4
	shufps	$0, %xmm1, %xmm1
	mulps	%xmm1, %xmm4

	pshufd	$0x55, %xmm4, %xmm5
	pshufd	$0x00, %xmm4, %xmm4

#ifndef TRMMKERNEL
	movss	0 * SIZE(C1), %xmm0
	movss	0 * SIZE(C1, LDC), %xmm1

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

	movss	%xmm4, 0 * SIZE(C1)
	movss	%xmm5, 0 * SIZE(C1, LDC)

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

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

.L69:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$2, KK
#endif

	movl	BB, B

	leal	(, LDC, 2), %eax
	addl	%eax, C
	ALIGN_4

.L70:
	movl	N,  %eax
	testl	$1, %eax
	jle	.L999

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

	movl	C, C1
	movl	A, AA

	movl	M,  I
	sarl	$2, I
	jle	.L80
	ALIGN_4

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

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

	movaps	-32 * SIZE(AA), %xmm0
	pxor	%xmm2, %xmm2
	movsd	-32 * SIZE(BB), %xmm1

	pxor	%xmm4, %xmm4
	prefetcht0	3 * SIZE(C1)
	pxor	%xmm5, %xmm5

#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	$4, %eax
#else
	addl	$1, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L75
	ALIGN_4

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

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movaps	-28 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm5
	pshufd	$0x55, %xmm1, %xmm2
	movsd	-30 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movaps	-24 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movaps	-20 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm5
	pshufd	$0x55, %xmm1, %xmm2
	movsd	-28 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movaps	-16 * SIZE(AA), %xmm0

	PREFETCH (PREFETCHSIZE + 16) * SIZE(AA)

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movaps	-12 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm5
	pshufd	$0x55, %xmm1, %xmm2
	movsd	-26 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movaps	 -8 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movaps	 -4 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm5
	pshufd	$0x55, %xmm1, %xmm2
	movsd	-24 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movaps	  0 * SIZE(AA), %xmm0

	subl   $-32 * SIZE, AA
	subl   $ -8 * SIZE, BB

	subl   $1, %eax
	jne    .L72
	ALIGN_4

.L75:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$7, %eax
	BRANCH
	je .L78
	ALIGN_4

.L76:
	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	movss	-31 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movaps	-28 * SIZE(AA), %xmm0

	addl	$4 * SIZE, AA
	addl	$1 * SIZE, BB
	decl	%eax
	jg	.L76
	ALIGN_4

.L78:
	movss	ALPHA,  %xmm1

	addps	%xmm2, %xmm4
	addps	%xmm5, %xmm4
	shufps	$0, %xmm1, %xmm1
	mulps	%xmm1, %xmm4

#ifndef TRMMKERNEL
	movsd	0 * SIZE(C1), %xmm0
	movhps	2 * SIZE(C1), %xmm0

	addps	%xmm0, %xmm4
#endif

	movsd	%xmm4, 0 * SIZE(C1)
	movhps	%xmm4, 2 * SIZE(C1)

#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, 4), AA
	addl	%eax, BB
#endif

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

	addl	$4 * SIZE, C1
	decl	I
	jg	.L71
	ALIGN_4

.L80:
	movl	M,  I
	testl	$2, I
	jle	.L90

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

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

	movsd	-32 * SIZE(AA), %xmm0
	pxor	%xmm3, %xmm3
	movsd	-32 * SIZE(BB), %xmm1
	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5

#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	.L85
	ALIGN_4

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

	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movsd	-30 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm4

	pshufd	$0x55, %xmm1, %xmm2
	movsd	-30 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movsd	-28 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5

	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movsd	-26 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm4

	pshufd	$0x55, %xmm1, %xmm2
	movsd	-28 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movsd	-24 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5

	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movsd	-22 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm4

	pshufd	$0x55, %xmm1, %xmm2
	movsd	-26 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movsd	-20 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5

	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movsd	-18 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm4

	pshufd	$0x55, %xmm1, %xmm2
	movsd	-24 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movsd	-16 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5

	subl   $-16 * SIZE, AA
	subl   $ -8 * SIZE, BB

	subl   $1, %eax
	jne    .L82
	ALIGN_4

.L85:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$7, %eax
	BRANCH
	je .L88
	ALIGN_4

.L86:
	pshufd	$0x00, %xmm1, %xmm2
	movss	-31 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movsd	-30 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm4

	addl	$2 * SIZE, AA
	addl	$1 * SIZE, BB
	decl	%eax
	jg	.L86
	ALIGN_4

.L88:
	movss	ALPHA,  %xmm1

	addps	%xmm5, %xmm4
	shufps	$0, %xmm1, %xmm1
	mulps	%xmm1, %xmm4

#ifndef TRMMKERNEL
	movsd	0 * SIZE(C1), %xmm0
	addps	%xmm0, %xmm4
#endif

	movsd	%xmm4, 0 * SIZE(C1)

#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
	addl	%eax, BB
#endif

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

	addl	$2 * SIZE, C1
	ALIGN_4

.L90:
	movl	M,  I
	testl	$1, I
	jle	.L999

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

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

	pxor	%xmm4, %xmm4
	movsd	-32 * SIZE(AA), %xmm0
	pxor	%xmm5, %xmm5
	movsd	-32 * SIZE(BB), %xmm1

#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	.L95
	ALIGN_4

.L92:
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AA)

	mulps	%xmm0, %xmm1
	movsd	-30 * SIZE(AA), %xmm0
	addps	%xmm1, %xmm4
	movsd	-30 * SIZE(BB), %xmm1

	mulps	%xmm0, %xmm1
	movsd	-28 * SIZE(AA), %xmm0
	addps	%xmm1, %xmm4
	movsd	-28 * SIZE(BB), %xmm1

	mulps	%xmm0, %xmm1
	movsd	-26 * SIZE(AA), %xmm0
	addps	%xmm1, %xmm4
	movsd	-26 * SIZE(BB), %xmm1

	mulps	%xmm0, %xmm1
	movsd	-24 * SIZE(AA), %xmm0
	addps	%xmm1, %xmm4
	movsd	-24 * SIZE(BB), %xmm1

	subl   $-8 * SIZE, AA
	subl   $-8 * SIZE, BB

	subl   $1, %eax
	jne    .L92
	ALIGN_4

.L95:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$7, %eax
	BRANCH
	je .L98
	ALIGN_4

.L96:
	mulss	%xmm0, %xmm1
	movss	-31 * SIZE(AA), %xmm0
	addss	%xmm1, %xmm4
	movss	-31 * SIZE(BB), %xmm1

	addl	$1 * SIZE, AA
	addl	$1 * SIZE, BB
	decl	%eax
	jg	.L96
	ALIGN_4

.L98:
	movss	ALPHA,  %xmm1

	haddps	%xmm4, %xmm4
	mulss	%xmm1, %xmm4

#ifndef TRMMKERNEL
	movss	0 * SIZE(C1), %xmm0

	addss	%xmm0, %xmm4
#endif

	movss	%xmm4, 0 * SIZE(C1)
	ALIGN_4

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

	addl	$ARGS, %esp
	ret

	EPILOGUE