gemm_kernel_2x4_penryn.s

Optimized GotoBLAS libraries

#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	.L35
	ALIGN_4

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

	pshufd	$0x4e, %xmm1, %xmm2
	mulpd	%xmm0, %xmm1
	mulpd	%xmm0, %xmm2
	movaps	-14 * SIZE(AA), %xmm0

	addpd	%xmm1, %xmm5
	movaps	-14 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm4

	pshufd	$0x4e, %xmm1, %xmm2
	mulpd	%xmm0, %xmm1
	mulpd	%xmm0, %xmm2
	movaps	-12 * SIZE(AA), %xmm0

	addpd	%xmm1, %xmm7
	movaps	-12 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm6

	pshufd	$0x4e, %xmm1, %xmm2
	mulpd	%xmm0, %xmm1
	mulpd	%xmm0, %xmm2
	movaps	-10 * SIZE(AA), %xmm0

	addpd	%xmm1, %xmm5
	movaps	-10 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm4

	pshufd	$0x4e, %xmm1, %xmm2
	mulpd	%xmm0, %xmm1
	mulpd	%xmm0, %xmm2
	movaps	 -8 * SIZE(AA), %xmm0

	addpd	%xmm1, %xmm7
	movaps	 -8 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm6

	PREFETCH (PREFETCHSIZE +  8) * SIZE(AA)

	pshufd	$0x4e, %xmm1, %xmm2
	mulpd	%xmm0, %xmm1
	mulpd	%xmm0, %xmm2
	movaps	 -6 * SIZE(AA), %xmm0

	addpd	%xmm1, %xmm5
	movaps	 -6 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm4

	pshufd	$0x4e, %xmm1, %xmm2
	mulpd	%xmm0, %xmm1
	mulpd	%xmm0, %xmm2
	movaps	 -4 * SIZE(AA), %xmm0

	addpd	%xmm1, %xmm7
	movaps	 -4 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm6

	pshufd	$0x4e, %xmm1, %xmm2
	mulpd	%xmm0, %xmm1
	mulpd	%xmm0, %xmm2
	movaps	 -2 * SIZE(AA), %xmm0

	addpd	%xmm1, %xmm5
	movaps	 -2 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm4

	pshufd	$0x4e, %xmm1, %xmm2
	mulpd	%xmm0, %xmm1
	mulpd	%xmm0, %xmm2
	movaps	  0 * SIZE(AA), %xmm0

	addpd	%xmm1, %xmm7
	movaps	  0 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm6

	subl   $-16 * SIZE, AA
	subl   $-16 * 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	$0x4e, %xmm1, %xmm2
	mulpd	%xmm0, %xmm1
	mulpd	%xmm0, %xmm2
	movaps	-14 * SIZE(AA), %xmm0

	addpd	%xmm1, %xmm5
	movaps	-14 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm4

	addl	$2 * SIZE, AA
	addl	$2 * SIZE, BB
	decl	%eax
	jg	.L36
	ALIGN_4

.L38:
	movddup	ALPHA,  %xmm3

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

	movaps	 %xmm4, %xmm0
	movsd    %xmm5, %xmm4
	mulpd	%xmm3, %xmm4
	movsd    %xmm0, %xmm5
	mulpd	%xmm3, %xmm5

#ifndef TRMMKERNEL
	movsd	0 * SIZE(C1), %xmm0
	movhpd	1 * SIZE(C1), %xmm0
	movsd	0 * SIZE(C1, LDC), %xmm1
	movhpd	1 * SIZE(C1, LDC), %xmm1

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

	movsd	%xmm4, 0 * SIZE(C1)
	movhpd	%xmm4, 1 * SIZE(C1)
	movsd	%xmm5, 0 * SIZE(C1, LDC)
	movhpd	%xmm5, 1 * 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
	decl	I
	jg	.L31
	ALIGN_4

.L40:
	movl	M,  I
	testl	$1, I
	jle	.L49

#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
#endif

	movaps	-16 * SIZE(AA), %xmm0
	pxor	%xmm4, %xmm4
	movaps	-16 * SIZE(BB), %xmm2
	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	$1, %eax
#else
	addl	$2, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L45
	ALIGN_4

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

	pshufd	$0x44, %xmm0, %xmm1
	mulpd	%xmm1, %xmm2

	addpd	%xmm2, %xmm4
	movaps	-14 * SIZE(BB), %xmm2

	pshufd	$0xee, %xmm0, %xmm1
	movaps	-14 * SIZE(AA), %xmm0
	mulpd	%xmm1, %xmm2

	addpd	%xmm2, %xmm5
	movaps	-12 * SIZE(BB), %xmm2

	pshufd	$0x44, %xmm0, %xmm1
	mulpd	%xmm1, %xmm2

	addpd	%xmm2, %xmm4
	movaps	-10 * SIZE(BB), %xmm2

	pshufd	$0xee, %xmm0, %xmm1
	movaps	-12 * SIZE(AA), %xmm0
	mulpd	%xmm1, %xmm2

	addpd	%xmm2, %xmm5
	movaps	 -8 * SIZE(BB), %xmm2

	pshufd	$0x44, %xmm0, %xmm1
	mulpd	%xmm1, %xmm2

	addpd	%xmm2, %xmm4
	movaps	 -6 * SIZE(BB), %xmm2

	pshufd	$0xee, %xmm0, %xmm1
	movaps	-10 * SIZE(AA), %xmm0
	mulpd	%xmm1, %xmm2

	addpd	%xmm2, %xmm5
	movaps	 -4 * SIZE(BB), %xmm2

	pshufd	$0x44, %xmm0, %xmm1
	mulpd	%xmm1, %xmm2

	addpd	%xmm2, %xmm4
	movaps	 -2 * SIZE(BB), %xmm2

	pshufd	$0xee, %xmm0, %xmm1
	movaps	 -8 * SIZE(AA), %xmm0
	mulpd	%xmm1, %xmm2

	addpd	%xmm2, %xmm5
	movaps	 0 * SIZE(BB), %xmm2

	subl   $ -8 * 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:
	pshufd	$0x44, %xmm0, %xmm1
	movsd	-15 * SIZE(AA), %xmm0
	mulpd	%xmm1, %xmm2

	addpd	%xmm2, %xmm4
	movaps	-14 * SIZE(BB), %xmm2

	addl	$1 * SIZE, AA
	addl	$2 * SIZE, BB
	decl	%eax
	jg	.L46
	ALIGN_4

.L48:
	movddup	ALPHA,  %xmm3

	addpd	%xmm5, %xmm4

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

	mulpd	%xmm3, %xmm4

#ifndef TRMMKERNEL
	addpd	%xmm0, %xmm4
#endif

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

#if defined(TRMMKERNEL) && defined(LEFT)
	addl	$1, KK
#endif
	ALIGN_4
	
.L49:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$2, KK
#endif

	movl	BB, B

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

.L50:
	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	$1, I
	jle	.L60
	ALIGN_4

.L51:
#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

	movaps	-16 * SIZE(AA), %xmm0
	pxor	%xmm4, %xmm4
	movaps	-16 * SIZE(BB), %xmm1
	pxor	%xmm5, %xmm5
	prefetcht0	1 * SIZE(C1)

#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	.L55
	ALIGN_4

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

	pshufd	$0x44, %xmm1, %xmm2
	mulpd	%xmm0, %xmm2
	movaps	-14 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm4

	pshufd	$0xee, %xmm1, %xmm2
	movaps	-14 * SIZE(BB), %xmm1
	mulpd	%xmm0, %xmm2
	movaps	-12 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm5

	pshufd	$0x44, %xmm1, %xmm2
	mulpd	%xmm0, %xmm2
	movaps	-10 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm4

	pshufd	$0xee, %xmm1, %xmm2
	movaps	-12 * SIZE(BB), %xmm1
	mulpd	%xmm0, %xmm2
	movaps	 -8 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm5

	PREFETCH (PREFETCHSIZE +  8) * SIZE(AA)

	pshufd	$0x44, %xmm1, %xmm2
	mulpd	%xmm0, %xmm2
	movaps	 -6 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm4

	pshufd	$0xee, %xmm1, %xmm2
	movaps	-10 * SIZE(BB), %xmm1
	mulpd	%xmm0, %xmm2
	movaps	 -4 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm5

	pshufd	$0x44, %xmm1, %xmm2
	mulpd	%xmm0, %xmm2
	movaps	 -2 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm4

	pshufd	$0xee, %xmm1, %xmm2
	movaps	 -8 * SIZE(BB), %xmm1
	mulpd	%xmm0, %xmm2
	movaps	  0 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm5

	subl   $-16 * SIZE, AA
	subl   $ -8 * 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, %xmm1, %xmm2
	movsd	-15 * SIZE(BB), %xmm1
	mulpd	%xmm0, %xmm2
	movaps	-14 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm4

	addl	$2 * SIZE, AA
	addl	$1 * SIZE, BB
	decl	%eax
	jg	.L56
	ALIGN_4

.L58:
	movddup	ALPHA,  %xmm3

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

#ifndef TRMMKERNEL
	movsd	0 * SIZE(C1), %xmm0
	movhpd	1 * SIZE(C1), %xmm0

	addpd	%xmm0, %xmm4
#endif

	movsd	%xmm4, 0 * SIZE(C1)
	movhpd	%xmm4, 1 * 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
	decl	I
	jg	.L51
	ALIGN_4

.L60:
	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

	movaps	-16 * SIZE(AA), %xmm0
	pxor	%xmm4, %xmm4
	movaps	-16 * SIZE(BB), %xmm2
	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	$1, %eax
#else
	addl	$1, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L65
	ALIGN_4

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

	mulpd	%xmm0, %xmm2
	movaps	-14 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm4
	movaps	-14 * SIZE(BB), %xmm2

	mulpd	%xmm0, %xmm2
	movaps	-12 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm5
	movaps	-12 * SIZE(BB), %xmm2

	mulpd	%xmm0, %xmm2
	movaps	-10 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm4
	movaps	-10 * SIZE(BB), %xmm2

	mulpd	%xmm0, %xmm2
	movaps	 -8 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm5
	movaps	 -8 * SIZE(BB), %xmm2

	subl   $-8 * SIZE, AA
	subl   $-8 * 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:
	mulsd	%xmm0, %xmm2
	movsd	-15 * SIZE(AA), %xmm0
	addsd	%xmm2, %xmm4
	movsd	-15 * SIZE(BB), %xmm2

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

.L68:
	movddup	ALPHA,  %xmm3

	addpd	%xmm5, %xmm4

	haddpd	%xmm4, %xmm4

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

	mulsd	%xmm3, %xmm4

#ifndef TRMMKERNEL
	addsd	%xmm0, %xmm4
#endif

	movsd	%xmm4, 0 * SIZE(C1)
	ALIGN_4
	
.L999:
	popl	%ebx
	popl	%esi
	popl	%edi
	popl	%ebp

	addl	$ARGS, %esp
	ret

	EPILOGUE