gemm_kernel_2x4_3dnow.s

Optimized GotoBLAS libraries

	movl	%eax, KKK	
#else
	movl	KK, %eax
#ifdef LEFT
	addl	$1, %eax
#else
	addl	$2, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$4, %eax
	je	.L55
	ALIGN_4

.L52:
	pfmul	%mm0, %mm2
	pfadd	%mm2, %mm4
	PADDING movd	(  4 + BOFFSET) * SIZE(BB), %mm2
	pfmul	(  2 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm5
	movd	        (  1 + AOFFSET) * SIZE(AA), %mm0

	PADDING prefetch	(PREFETCHSIZE +  0) * SIZE(AA)

	pfmul	%mm0, %mm2
	pfadd	%mm2, %mm6
	PADDING movd	(  8 + BOFFSET) * SIZE(BB), %mm2
	pfmul	(  6 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm7
	movd	        (  2 + AOFFSET) * SIZE(AA), %mm0

	pfmul	%mm0, %mm2
	pfadd	%mm2, %mm4
	PADDING movd	( 12 + BOFFSET) * SIZE(BB), %mm2
	pfmul	( 10 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm5
	movd	        (  3 + AOFFSET) * SIZE(AA), %mm0

	pfmul	%mm0, %mm2
	pfadd	%mm2, %mm6
	PADDING movd	( 32 + BOFFSET) * SIZE(BB), %mm2
	pfmul	( 14 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm7
	movd	        (  4 + AOFFSET) * SIZE(AA), %mm0

	pfmul	%mm0, %mm3
	pfadd	%mm3, %mm4
	PADDING movd	( 20 + BOFFSET) * SIZE(BB), %mm3
	pfmul	( 18 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm5
	movd	        (  5 + AOFFSET) * SIZE(AA), %mm0

	pfmul	%mm0, %mm3
	pfadd	%mm3, %mm6
	PADDING movd	( 24 + BOFFSET) * SIZE(BB), %mm3
	pfmul	( 22 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm7
	movd	        (  6 + AOFFSET) * SIZE(AA), %mm0

	pfmul	%mm0, %mm3
	pfadd	%mm3, %mm4
	PADDING movd	( 28 + BOFFSET) * SIZE(BB), %mm3
	pfmul	( 26 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm5
	movd	        (  7 + AOFFSET) * SIZE(AA), %mm0

	pfmul	%mm0, %mm3
	pfadd	%mm3, %mm6
	PADDING movd	( 48 + BOFFSET) * SIZE(BB), %mm3
	pfmul	( 30 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm7
	movd	        ( 16 + AOFFSET) * SIZE(AA), %mm0

	pfmul	%mm1, %mm2
	pfadd	%mm2, %mm4
	PADDING movd	( 36 + BOFFSET) * SIZE(BB), %mm2
	pfmul	( 34 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm5
	movd	        (  9 + AOFFSET) * SIZE(AA), %mm1

	pfmul	%mm1, %mm2
	pfadd	%mm2, %mm6
	PADDING movd	( 40 + BOFFSET) * SIZE(BB), %mm2
	pfmul	( 38 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm7
	movd	        ( 10 + AOFFSET) * SIZE(AA), %mm1

	pfmul	%mm1, %mm2
	pfadd	%mm2, %mm4
	PADDING movd	( 44 + BOFFSET) * SIZE(BB), %mm2
	pfmul	( 42 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm5
	movd	        ( 11 + AOFFSET) * SIZE(AA), %mm1

	pfmul	%mm1, %mm2
	pfadd	%mm2, %mm6
	PADDING movd	( 64 + BOFFSET) * SIZE(BB), %mm2
	pfmul	( 46 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm7
	movd	        ( 12 + AOFFSET) * SIZE(AA), %mm1

	pfmul	%mm1, %mm3
	pfadd	%mm3, %mm4
	PADDING movd	( 52 + BOFFSET) * SIZE(BB), %mm3
	pfmul	( 50 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm5
	movd	        ( 13 + AOFFSET) * SIZE(AA), %mm1

	pfmul	%mm1, %mm3
	pfadd	%mm3, %mm6
	PADDING movd	( 56 + BOFFSET) * SIZE(BB), %mm3
	pfmul	( 54 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm7
	movd	        ( 14 + AOFFSET) * SIZE(AA), %mm1

	pfmul	%mm1, %mm3
	pfadd	%mm3, %mm4
	PADDING movd	( 60 + BOFFSET) * SIZE(BB), %mm3
	pfmul	( 58 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm5
	movd	        ( 15 + AOFFSET) * SIZE(AA), %mm1

	pfmul	%mm1, %mm3
	pfadd	%mm3, %mm6
	PADDING movd	( 80 + BOFFSET) * SIZE(BB), %mm3
	pfmul	( 62 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm7
	movd	        ( 24 + AOFFSET) * SIZE(AA), %mm1

	subl	$-16 * SIZE, AA
	addl	$ 64 * SIZE, BB
	decl	%eax
	jne    .L52
	ALIGN_3

.L55:
	movd	ALPHA,  %mm3
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$15, %eax
	BRANCH
	je .L58
	ALIGN_3

.L56:
	pfmul	%mm0, %mm2
	pfadd	%mm2, %mm4
	PADDING movd	(  4 + BOFFSET) * SIZE(BB), %mm2

	pfmul	(  2 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm5
	movd	        (  1 + AOFFSET) * SIZE(AA), %mm0

	addl	$1 * SIZE, AA
	addl	$4 * SIZE, BB
	decl	%eax
	jg	.L56
	ALIGN_3

.L58:
	pfadd	%mm6, %mm4
	pfadd	%mm7, %mm5

	pfmul	%mm3, %mm4
	pfmul	%mm3, %mm5

#ifndef TRMMKERNEL
	movd	0 * SIZE(%esi)         , %mm0
	movd	0 * SIZE(%esi, %ebp, 1), %mm1

	pfadd	%mm0, %mm4
	pfadd	%mm1, %mm5
#endif

	movd	%mm4, 0 * SIZE(%esi)
	movd	%mm5, 0 * SIZE(%esi, %ebp, 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	(AA, %eax, 1), AA
	leal	(BB, %eax, 4), BB
#endif

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

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

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

.L60:
	movl	N, %eax
	testl	$1, %eax
	jle	.L999
	ALIGN_3

.L61:
/* Copying to Sub Buffer */
	leal	BUFFER, %ecx

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

	movl	K, %eax
	sarl	$3, %eax
	jle	.L63
	ALIGN_3
	
.L62:
	movd	 0 * SIZE(%edi), %mm0
	movd	 1 * SIZE(%edi), %mm1
	movd	 2 * SIZE(%edi), %mm2
	movd	 3 * SIZE(%edi), %mm3
	movd	 4 * SIZE(%edi), %mm4
	movd	 5 * SIZE(%edi), %mm5
	movd	 6 * SIZE(%edi), %mm6
	movd	 7 * SIZE(%edi), %mm7

	prefetchnta	72 * SIZE(%edi)

	punpckldq %mm0, %mm0
	punpckldq %mm1, %mm1
	punpckldq %mm2, %mm2
	punpckldq %mm3, %mm3
	punpckldq %mm4, %mm4
	punpckldq %mm5, %mm5
	punpckldq %mm6, %mm6
	punpckldq %mm7, %mm7

	movq	%mm0,  0 * SIZE(%ecx)
	movq	%mm1,  2 * SIZE(%ecx)
	movq	%mm2,  4 * SIZE(%ecx)
	movq	%mm3,  6 * SIZE(%ecx)
	movq	%mm4,  8 * SIZE(%ecx)
	movq	%mm5, 10 * SIZE(%ecx)
	movq	%mm6, 12 * SIZE(%ecx)
	movq	%mm7, 14 * SIZE(%ecx)

	addl	$ 8 * SIZE, %edi
	addl	$16 * SIZE, %ecx
	decl	%eax
	jne	.L62

.L63:
	movl	K, %eax
	andl	$7, %eax
	BRANCH
	jle	.L70
	ALIGN_2

.L64:
	movd	 0 * SIZE(%edi), %mm0
	punpckldq %mm0, %mm0
	movq	%mm0,  0 * SIZE(%ecx)

	addl	$1 * SIZE, %edi
	addl	$2 * SIZE, %ecx
	decl	%eax
	jne	.L64
	ALIGN_4
	
.L70:
	movl	C, %esi		# coffset = c
	movl	A, %edx		# aoffset = a
	movl	M,  %ebx
	sarl	$1, %ebx	# i = (m >> 2)
	jle	.L90
	ALIGN_4

.L71:
	leal	- BOFFSET * SIZE + BUFFER, BB

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
#else
	movl	KK,   %eax
	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 2), AA
	leal	(BB, %eax, 2), BB
#endif

	movq	        (  0 + AOFFSET) * SIZE(AA), %mm0
	pxor	%mm4, %mm4
	movq	        ( 16 + AOFFSET) * SIZE(AA), %mm1
	pxor	%mm5, %mm5
	PADDING movq	(  0 + BOFFSET) * SIZE(BB), %mm2
	pxor	%mm6, %mm6
	pxor	%mm7, %mm7

	prefetchw 2 * SIZE(%esi)
	prefetchw 2 * SIZE(%esi, %ebp)

#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	$4, %eax
	je	.L75
	ALIGN_4

.L72:
	pfmul		(  0 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm4
	movq	        (  2 + AOFFSET) * SIZE(AA), %mm0

	PADDING prefetch	(PREFETCHSIZE +  0) * SIZE(AA)

	pfmul		(  2 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm5
	movq	        (  4 + AOFFSET) * SIZE(AA), %mm0

	pfmul		(  4 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm6
	movq	        (  6 + AOFFSET) * SIZE(AA), %mm0

	pfmul		(  6 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm7
	movq	        (  8 + AOFFSET) * SIZE(AA), %mm0

	pfmul		(  8 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm4
	movq	        ( 10 + AOFFSET) * SIZE(AA), %mm0

	pfmul		( 10 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm5
	movq	        ( 12 + AOFFSET) * SIZE(AA), %mm0

	pfmul		( 12 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm6
	movq	        ( 14 + AOFFSET) * SIZE(AA), %mm0

	pfmul		( 14 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm7
	movq	        ( 32 + AOFFSET) * SIZE(AA), %mm0

	pfmul		( 16 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm4
	movq	        ( 18 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 18 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm5
	movq	        ( 20 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 20 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm6
	movq	        ( 22 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 22 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm7
	movq	        ( 24 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 24 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm4
	movq	        ( 26 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 26 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm5
	movq	        ( 28 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 28 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm6
	movq	        ( 30 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 30 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm7
	movq	        ( 48 + AOFFSET) * SIZE(AA), %mm1

	subl	$-32 * SIZE, AA
	addl	$ 32 * SIZE, BB
	decl	%eax
	jne    .L72
	ALIGN_3

.L75:
	movq	ALPHA,  %mm3
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$15, %eax
	BRANCH
	je .L78
	ALIGN_3

.L76:
	pfmul		(  0 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm4
	movq	        (  2 + AOFFSET) * SIZE(AA), %mm0

	addl	$2 * SIZE, AA
	addl	$2 * SIZE, BB
	decl	%eax
	jg	.L76
	ALIGN_3

.L78:
	pfadd	%mm5, %mm4
	pfadd	%mm7, %mm6
	pfadd	%mm6, %mm4

	pfmul	%mm3, %mm4
#ifndef TRMMKERNEL
	pfadd	0 * SIZE(%esi), %mm4
#endif
	movq	%mm4, 0 * SIZE(%esi)

#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, %esi		# coffset += 2
	decl	%ebx			# i --
	jg	.L71
	ALIGN_4

.L90:
	movl	M,  %ebx
	testl	$1, %ebx	# i = (m >> 2)
	jle	.L999
	ALIGN_4

.L91:
	leal	- BOFFSET * SIZE + BUFFER, BB

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
#else
	movl	KK,   %eax
	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 1), AA
	leal	(BB, %eax, 2), BB
#endif

	movq	        (  0 + AOFFSET) * SIZE(AA), %mm0
	pxor	%mm4, %mm4
	movq	        (  8 + AOFFSET) * SIZE(AA), %mm1
	pxor	%mm5, %mm5
	PADDING movq	(  0 + BOFFSET) * SIZE(BB), %mm2
	pxor	%mm6, %mm6
	pxor	%mm7, %mm7

#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	$4, %eax
	je	.L95
	ALIGN_4

.L92:
	PADDING prefetch	(PREFETCHSIZE +  0) * SIZE(AA)

	pfmul		(  0 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm4
	movd	        (  1 + AOFFSET) * SIZE(AA), %mm0

	pfmul		(  2 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm5
	movd	        (  2 + AOFFSET) * SIZE(AA), %mm0

	pfmul		(  4 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm6
	movd	        (  3 + AOFFSET) * SIZE(AA), %mm0

	pfmul		(  6 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm7
	movd	        (  4 + AOFFSET) * SIZE(AA), %mm0

	pfmul		(  8 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm4
	movd	        (  5 + AOFFSET) * SIZE(AA), %mm0

	pfmul		( 10 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm5
	movd	        (  6 + AOFFSET) * SIZE(AA), %mm0

	pfmul		( 12 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm6
	movd	        (  7 + AOFFSET) * SIZE(AA), %mm0

	pfmul		( 14 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm7
	movd	        ( 16 + AOFFSET) * SIZE(AA), %mm0

	pfmul		( 16 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm4
	movd	        (  9 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 18 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm5
	movd	        ( 10 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 20 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm6
	movd	        ( 11 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 22 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm7
	movd	        ( 12 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 24 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm4
	movd	        ( 13 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 26 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm5
	movd	        ( 14 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 28 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm6
	movd	        ( 15 + AOFFSET) * SIZE(AA), %mm1

	pfmul		( 30 + BOFFSET) * SIZE(BB), %mm1
	pfadd	%mm1, %mm7
	movd	        ( 24 + AOFFSET) * SIZE(AA), %mm1

	subl	$-16 * SIZE, AA
	addl	$ 32 * SIZE, BB
	decl	%eax
	jne    .L92
	ALIGN_3

.L95:
	movd	ALPHA,  %mm3
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	andl	$15, %eax
	BRANCH
	je .L98
	ALIGN_3

.L96:
	pfmul		(  0 + BOFFSET) * SIZE(BB), %mm0
	pfadd	%mm0, %mm4
	movd	        (  1 + AOFFSET) * SIZE(AA), %mm0

	addl	$1 * SIZE, AA
	addl	$2 * SIZE, BB
	decl	%eax
	jg	.L96
	ALIGN_3

.L98:
#ifndef TRMMKERNEL
	movd	0 * SIZE(%esi), %mm0
#endif

	pfadd	%mm5, %mm4
	pfadd	%mm7, %mm6
	pfadd	%mm6, %mm4

	pfmul	%mm3, %mm4
	pfmul	%mm3, %mm5

#ifndef TRMMKERNEL
	pfadd	%mm0, %mm4
#endif
	movd	%mm4, 0 * SIZE(%esi)

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

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

.L999:
	EMMS

	movl	OLD_STACK, %esp
	popl	%ebx
	popl	%esi
	popl	%edi
	popl	%ebp
	ret

	EPILOGUE