gemm_kernel_4x4_penryn.s

Optimized GotoBLAS libraries

	movsd	%xmm5, 0 * SIZE(%eax)
	movhps	%xmm5, 0 * SIZE(%eax, 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	(AA, %eax, 2), AA
	leal	(BB, %eax, 4), BB
#endif

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

	addl	$2 * SIZE, C1
	ALIGN_4

.L30:
	movl	M,  I
	testl	$1, I
	jle	.L39

#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
	leal	(BB, %eax, 4), BB
#endif

	pxor	%xmm4, %xmm4
	movsd	-32 * SIZE(AA), %xmm0
	pxor	%xmm5, %xmm5
	movaps	-32 * SIZE(BB), %xmm1
	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	$4, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L35
	ALIGN_4

.L32:
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AA)

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

	pshufd	$0x55, %xmm0, %xmm2
	movsd	-30 * SIZE(AA), %xmm0

	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm4
	movaps	-24 * SIZE(BB), %xmm1

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

	pshufd	$0x55, %xmm0, %xmm2
	movsd	-28 * SIZE(AA), %xmm0

	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm4
	movaps	-16 * SIZE(BB), %xmm1

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

	pshufd	$0x55, %xmm0, %xmm2
	movsd	-26 * SIZE(AA), %xmm0

	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm4
	movaps	 -8 * SIZE(BB), %xmm1

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

	pshufd	$0x55, %xmm0, %xmm2
	movsd	-24 * SIZE(AA), %xmm0

	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm4
	movaps	  0 * SIZE(BB), %xmm1

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

	subl   $1, %eax
	jne    .L32
	ALIGN_4

.L35:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$7, %eax
	BRANCH
	je .L38
	ALIGN_4

.L36:
	pshufd	$0x00, %xmm0, %xmm2
	movss	-31 * SIZE(AA), %xmm0

	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm4
	movaps	-28 * SIZE(BB), %xmm1

	addl	$1 * SIZE, AA
	addl	$4 * SIZE, BB
	decl	%eax
	jg	.L36
	ALIGN_4

.L38:
	movss	ALPHA,  %xmm1

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

	pshufd	$0xff, %xmm4, %xmm7
	pshufd	$0xaa, %xmm4, %xmm6
	pshufd	$0x55, %xmm4, %xmm5
	pshufd	$0x00, %xmm4, %xmm4

	leal		(C1, LDC, 2), %eax

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

	movss	0 * SIZE(%eax), %xmm2
	movss	0 * SIZE(%eax, LDC), %xmm3

	addss	%xmm0, %xmm4
	addss	%xmm1, %xmm5
	addss	%xmm2, %xmm6
	addss	%xmm3, %xmm7
#endif

	movss	%xmm4, 0 * SIZE(C1)
	movss	%xmm5, 0 * SIZE(C1, LDC)
	movss	%xmm6, 0 * SIZE(%eax)
	movss	%xmm7, 0 * SIZE(%eax, 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, 4), BB
#endif

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

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

	movl	BB, B

	leal	(, LDC, 4), %eax
	addl	%eax, C
	decl	J
	jg	.L01
	ALIGN_4

.L40:
	movl	N,  %eax
	testl	$2, %eax
	jle	.L70

#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	.L50
	ALIGN_4

.L41:
#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
	leal	(BB, %eax, 2), BB
#endif

	movaps	-32 * SIZE(AA), %xmm0
	pxor	%xmm2, %xmm2
	movaps	-32 * SIZE(BB), %xmm1
	pxor	%xmm3, %xmm3

	pxor	%xmm4, %xmm4
	prefetcht0	3 * SIZE(C1)
	pxor	%xmm5, %xmm5
	prefetcht0	3 * SIZE(C1, 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	$4, %eax
#else
	addl	$2, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L45
	ALIGN_4

.L42:
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AA)

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	pshufd	$0x55, %xmm1, %xmm3
	mulps	%xmm0, %xmm3
	movaps	-28 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm6
	pshufd	$0xaa, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm7
	pshufd	$0xff, %xmm1, %xmm3
	movaps	-28 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm3
	movaps	-24 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	pshufd	$0x55, %xmm1, %xmm3
	mulps	%xmm0, %xmm3
	movaps	-20 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm6
	pshufd	$0xaa, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm7
	pshufd	$0xff, %xmm1, %xmm3
	movaps	-24 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm3
	movaps	-16 * SIZE(AA), %xmm0

	PREFETCH (PREFETCHSIZE + 16) * SIZE(AA)

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	pshufd	$0x55, %xmm1, %xmm3
	mulps	%xmm0, %xmm3
	movaps	-12 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm6
	pshufd	$0xaa, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm7
	pshufd	$0xff, %xmm1, %xmm3
	movaps	-20 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm3
	movaps	 -8 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	pshufd	$0x55, %xmm1, %xmm3
	mulps	%xmm0, %xmm3
	movaps	 -4 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm6
	pshufd	$0xaa, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm7
	pshufd	$0xff, %xmm1, %xmm3
	movaps	-16 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm3
	movaps	  0 * SIZE(AA), %xmm0

	subl   $-32 * SIZE, AA
	subl   $-16 * SIZE, BB

	subl   $1, %eax
	jne    .L42
	ALIGN_4

.L45:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$7, %eax
	BRANCH
	je .L48
	ALIGN_4

.L46:
	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	pshufd	$0x55, %xmm1, %xmm3
	movsd	-30 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm3
	movaps	-28 * SIZE(AA), %xmm0

	addl	$4 * SIZE, AA
	addl	$2 * SIZE, BB
	decl	%eax
	jg	.L46
	ALIGN_4

.L48:
	movss	ALPHA,  %xmm1

	addps	%xmm6, %xmm4
	addps	%xmm7, %xmm5

	addps	%xmm2, %xmm4
	addps	%xmm3, %xmm5

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

#ifndef TRMMKERNEL
	movsd	0 * SIZE(C1), %xmm0
	movhps	2 * SIZE(C1), %xmm0
	movsd	0 * SIZE(C1, LDC), %xmm1
	movhps	2 * SIZE(C1, LDC), %xmm1

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

	movsd	%xmm4, 0 * SIZE(C1)
	movhps	%xmm4, 2 * SIZE(C1)
	movsd	%xmm5, 0 * SIZE(C1, LDC)
	movhps	%xmm5, 2 * 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
	leal	(AA, %eax, 4), AA
	leal	(BB, %eax, 2), BB
#endif

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

	addl	$4 * SIZE, C1
	decl	I
	jg	.L41
	ALIGN_4

.L50:
	movl	M,  I
	testl	$2, I
	jle	.L60

#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
	leal	(BB, %eax, 2), BB
#endif

	movaps	-32 * SIZE(AA), %xmm0
	pxor	%xmm3, %xmm3
	movaps	-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	$2, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L55
	ALIGN_4

.L52:
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AA)

	pshufd	$0x44, %xmm0, %xmm2
	addps	%xmm3, %xmm4
	pshufd	$0x50, %xmm1, %xmm3
	mulps	%xmm2, %xmm3

	pshufd	$0xee, %xmm0, %xmm2
	movaps	-28 * SIZE(AA), %xmm0
	addps	%xmm3, %xmm5
	pshufd	$0xfa, %xmm1, %xmm3
	movaps	-28 * SIZE(BB), %xmm1
	mulps	%xmm2, %xmm3

	pshufd	$0x44, %xmm0, %xmm2
	addps	%xmm3, %xmm4
	pshufd	$0x50, %xmm1, %xmm3
	mulps	%xmm2, %xmm3

	pshufd	$0xee, %xmm0, %xmm2
	movaps	-24 * SIZE(AA), %xmm0
	addps	%xmm3, %xmm5
	pshufd	$0xfa, %xmm1, %xmm3
	movaps	-24 * SIZE(BB), %xmm1
	mulps	%xmm2, %xmm3

	pshufd	$0x44, %xmm0, %xmm2
	addps	%xmm3, %xmm4
	pshufd	$0x50, %xmm1, %xmm3
	mulps	%xmm2, %xmm3

	pshufd	$0xee, %xmm0, %xmm2
	movaps	-20 * SIZE(AA), %xmm0
	addps	%xmm3, %xmm5
	pshufd	$0xfa, %xmm1, %xmm3
	movaps	-20 * SIZE(BB), %xmm1
	mulps	%xmm2, %xmm3

	pshufd	$0x44, %xmm0, %xmm2
	addps	%xmm3, %xmm4
	pshufd	$0x50, %xmm1, %xmm3
	mulps	%xmm2, %xmm3

	pshufd	$0xee, %xmm0, %xmm2
	movaps	-16 * SIZE(AA), %xmm0
	addps	%xmm3, %xmm5
	pshufd	$0xfa, %xmm1, %xmm3
	movaps	-16 * SIZE(BB), %xmm1
	mulps	%xmm2, %xmm3

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

	subl   $1, %eax
	jne    .L52
	ALIGN_4

.L55:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$7, %eax
	BRANCH
	je .L58
	ALIGN_4

.L56:
	pshufd	$0x44, %xmm0, %xmm2
	movsd	-30 * SIZE(AA), %xmm0
	addps	%xmm3, %xmm4
	pshufd	$0x50, %xmm1, %xmm3
	movsd	-30 * SIZE(BB), %xmm1
	mulps	%xmm2, %xmm3

	addl	$2 * SIZE, AA
	addl	$2 * SIZE, BB
	decl	%eax
	jg	.L56
	ALIGN_4

.L58:
	movss	ALPHA,  %xmm1

	addps	%xmm3, %xmm4
	addps	%xmm5, %xmm4

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

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

	addps	%xmm0, %xmm4
#endif

	movsd	%xmm4, 0 * SIZE(C1)
	movhps	%xmm4, 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
	leal	(AA, %eax, 2), AA
	leal	(BB, %eax, 2), BB
#endif

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

	addl	$2 * SIZE, C1
	ALIGN_4

.L60:
	movl	M,  I
	testl	$1, I
	jle	.L69

#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
	leal	(BB, %eax, 2), BB