zgemm_kernel_2x2_penryn.s

Optimized GotoBLAS libraries

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

	subl   $-16 * SIZE, AA
	subl   $-32 * SIZE, BB
	
	decl   %eax
	jne    .L22
	ALIGN_4
	
.L25:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$7, %eax
	BRANCH
	je .L28
	ALIGN_4

.L26:
	addps	%xmm2, %xmm6
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm7
	pshufd	$0x55, %xmm1, %xmm3
	mulps	%xmm0, %xmm3

	addps	%xmm2, %xmm4
	pshufd	$0xaa, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	pshufd	$0xff, %xmm1, %xmm3
	movaps	-28 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm3
	movsd	-30 * SIZE(AA), %xmm0

	addl	$2 * SIZE, AA
	addl	$4 * SIZE, BB

	decl	%eax
	jg	.L26
	ALIGN_4

.L28:
	addps	%xmm2, %xmm6
	addps	%xmm3, %xmm7

	movsd	ALPHA_R, %xmm3

	pshufd	 $0xb1, %xmm5, %xmm5
	pcmpeqb	%xmm0, %xmm0
	pshufd	 $0xb1, %xmm7, %xmm7
	psllq	$63,   %xmm0

#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
	pxor	%xmm0, %xmm5
	pxor	%xmm0, %xmm7

	subps	%xmm5, %xmm4
	subps	%xmm7, %xmm6
#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
	pxor	%xmm0, %xmm5
	pxor	%xmm0, %xmm7

	addps	%xmm5, %xmm4
	addps	%xmm7, %xmm6
#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
	pxor	%xmm0, %xmm4
	pxor	%xmm0, %xmm6

	addps	%xmm5, %xmm4
	addps	%xmm7, %xmm6
#else
	pxor	%xmm0, %xmm4
	pxor	%xmm0, %xmm6

	subps	%xmm5, %xmm4
	subps	%xmm7, %xmm6
#endif

	pshufd	$0x00, %xmm3, %xmm2
	pshufd	$0x55, %xmm3, %xmm3

	pshufd	$0xb1, %xmm4, %xmm5
	pshufd	$0xb1, %xmm6, %xmm7

	mulps	%xmm2, %xmm4
	mulps	%xmm3, %xmm5

	mulps	%xmm2, %xmm6
	mulps	%xmm3, %xmm7

	pxor	%xmm0, %xmm5
	pxor	%xmm0, %xmm7

	subps	%xmm5, %xmm4
	subps	%xmm7, %xmm6

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movsd	0 * SIZE(C1), %xmm2
	movsd	0 * SIZE(C1, LDC), %xmm3

	addps	%xmm2, %xmm4
	addps	%xmm3, %xmm6
#endif

	movsd	%xmm4, 0 * SIZE(C1)
	movsd	%xmm6, 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, 4), BB
#endif

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

	addl	$2 * SIZE, C1
	ALIGN_2

.L29:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$2, KK
#endif

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

	decl	J
	jg	.L01
	ALIGN_4

.L30:
	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,  %ebx
	sarl	$1, %ebx
	jle	.L40
	ALIGN_4

.L31:
#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
	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	$1, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L35
	ALIGN_4

.L32:
	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AA)
	addps	%xmm3, %xmm5
	pshufd	$0x55, %xmm1, %xmm3
	mulps	%xmm0, %xmm3
	movaps	-28 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm4
	pshufd	$0xaa, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	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, %xmm4
	pshufd	$0xaa, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	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, %xmm4
	pshufd	$0xaa, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	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, %xmm4
	pshufd	$0xaa, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	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
	
	decl   %eax
	jne    .L32
	ALIGN_4
	
.L35:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	movsd	-32 * SIZE(BB), %xmm1
	andl	$7, %eax
	BRANCH
	je .L38
	ALIGN_4

.L36:
	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	.L36
	ALIGN_4

.L38:
	addps	%xmm2, %xmm4
	addps	%xmm3, %xmm5

	movsd	ALPHA_R, %xmm3

	pshufd	 $0xb1, %xmm5, %xmm5
	pcmpeqb	%xmm0, %xmm0
	psllq	$63,   %xmm0

#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
	pxor	%xmm0, %xmm5
	subps	%xmm5, %xmm4
#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
	pxor	%xmm0, %xmm5
	addps	%xmm5, %xmm4
#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
	pxor	%xmm0, %xmm4
	addps	%xmm5, %xmm4
#else
	pxor	%xmm0, %xmm4
	subps	%xmm5, %xmm4
#endif

	pshufd	$0x00, %xmm3, %xmm2
	pshufd	$0x55, %xmm3, %xmm3

	pshufd	$0xb1, %xmm4, %xmm5

	mulps	%xmm2, %xmm4
	mulps	%xmm3, %xmm5

	pxor	%xmm0, %xmm5
	subps	%xmm5, %xmm4

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movsd	0 * SIZE(C1), %xmm2
	movhps	2 * SIZE(C1), %xmm2

	addps	%xmm2, %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
	leal	(BB, %eax, 2), BB
#endif

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

	addl	$4 * SIZE, C1
	decl	%ebx
	jg	.L31
	ALIGN_4
	
.L40:
	movl	M,  %ebx
	testl	$1, %ebx
	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
	leal	(AA, %eax, 2), AA
	leal	(BB, %eax, 2), BB
#endif	

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

	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	.L45
	ALIGN_4

.L42:
	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
	movsd	-30 * SIZE(AA), %xmm0

	PREFETCH (PREFETCHSIZE +  0) * SIZE(AA)

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

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	pshufd	$0x55, %xmm1, %xmm3
	movsd	-26 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm3
	movsd	-26 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	pshufd	$0x55, %xmm1, %xmm3
	movsd	-24 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm3
	movsd	-24 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	pshufd	$0x55, %xmm1, %xmm3
	movsd	-22 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm3
	movsd	-22 * SIZE(AA), %xmm0

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

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	pshufd	$0x55, %xmm1, %xmm3
	movsd	-18 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm3
	movsd	-18 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm3, %xmm5
	pshufd	$0x55, %xmm1, %xmm3
	movsd	-16 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm3
	movsd	-16 * SIZE(AA), %xmm0

	subl   $-16 * SIZE, AA
	subl   $-16 * SIZE, BB
	
	decl   %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
	movsd	-30 * SIZE(AA), %xmm0

	addl	$2 * SIZE, AA
	addl	$2 * SIZE, BB

	decl	%eax
	jg	.L46
	ALIGN_4

.L48:
	addps	%xmm2, %xmm4
	addps	%xmm3, %xmm5

	movsd	ALPHA_R, %xmm3

	pshufd	 $0xb1, %xmm5, %xmm5
	pcmpeqb	%xmm0, %xmm0
	psllq	$63,   %xmm0

#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
	pxor	%xmm0, %xmm5
	subps	%xmm5, %xmm4
#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
	pxor	%xmm0, %xmm5
	addps	%xmm5, %xmm4
#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
	pxor	%xmm0, %xmm4
	addps	%xmm5, %xmm4
#else
	pxor	%xmm0, %xmm4
	subps	%xmm5, %xmm4
#endif

	pshufd	$0x00, %xmm3, %xmm2
	pshufd	$0x55, %xmm3, %xmm3

	pshufd	$0xb1, %xmm4, %xmm5

	mulps	%xmm2, %xmm4
	mulps	%xmm3, %xmm5

	pxor	%xmm0, %xmm5
	subps	%xmm5, %xmm4

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movsd	0 * SIZE(C1), %xmm2
	addps	%xmm2, %xmm4
#endif

	movsd	%xmm4, 0 * SIZE(C1)
	ALIGN_4

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

	addl	$ARGS, %esp
	ret

	EPILOGUE