gemm_kernel_4x4_sse.s

Optimized GotoBLAS libraries

	addl	$1, KK
#endif
	ALIGN_4

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

	leal	(, LDC, 4), %eax
	addl	%eax, C			# c += 4 * ldc
	decl	J			# j --
	jg	.L01
	ALIGN_4

.L40:
	testl	$2, N
	je	.L80

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

	movl	K, %eax
	leal	BUFFER, %ecx
	sarl	$2, %eax
	jle	.L45
	ALIGN_4
	
.L42:
	prefetchnta	80 * SIZE(%edi)

#if defined(OPTERON) || defined(BARCELONA)
	prefetchw	112 * SIZE(%ecx)
	prefetchw	120 * SIZE(%ecx)
#endif

#ifdef PENTIUM4
     	prefetcht1	112 * SIZE(%ecx)
#endif

#ifdef HAVE_SSE2
	movss	 0 * SIZE(%edi), %xmm0
	movss	 1 * SIZE(%edi), %xmm1
	movss	 2 * SIZE(%edi), %xmm2
	movss	 3 * SIZE(%edi), %xmm3
	movss	 4 * SIZE(%edi), %xmm4
	movss	 5 * SIZE(%edi), %xmm5
	movss	 6 * SIZE(%edi), %xmm6
	movss	 7 * SIZE(%edi), %xmm7

	shufps	 $0, %xmm0, %xmm0
	shufps	 $0, %xmm1, %xmm1
	shufps	 $0, %xmm2, %xmm2
	shufps	 $0, %xmm3, %xmm3
	shufps	 $0, %xmm4, %xmm4
	shufps	 $0, %xmm5, %xmm5
	shufps	 $0, %xmm6, %xmm6
	shufps	 $0, %xmm7, %xmm7

	movaps	%xmm0,  0 * SIZE(%ecx)
	movaps	%xmm1,  4 * SIZE(%ecx)
	movaps	%xmm2,  8 * SIZE(%ecx)
	movaps	%xmm3, 12 * SIZE(%ecx)
	movaps	%xmm4, 16 * SIZE(%ecx)
	movaps	%xmm5, 20 * SIZE(%ecx)
	movaps	%xmm6, 24 * SIZE(%ecx)
	movaps	%xmm7, 28 * SIZE(%ecx)
#else
	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

	movd	%mm0,  0 * SIZE(%ecx)
	movd	%mm0,  1 * SIZE(%ecx)
	movd	%mm0,  2 * SIZE(%ecx)
	movd	%mm0,  3 * SIZE(%ecx)
	movd	%mm1,  4 * SIZE(%ecx)
	movd	%mm1,  5 * SIZE(%ecx)
	movd	%mm1,  6 * SIZE(%ecx)
	movd	%mm1,  7 * SIZE(%ecx)
	movd	%mm2,  8 * SIZE(%ecx)
	movd	%mm2,  9 * SIZE(%ecx)
	movd	%mm2, 10 * SIZE(%ecx)
	movd	%mm2, 11 * SIZE(%ecx)
	movd	%mm3, 12 * SIZE(%ecx)
	movd	%mm3, 13 * SIZE(%ecx)
	movd	%mm3, 14 * SIZE(%ecx)
	movd	%mm3, 15 * SIZE(%ecx)
	movd	%mm4, 16 * SIZE(%ecx)
	movd	%mm4, 17 * SIZE(%ecx)
	movd	%mm4, 18 * SIZE(%ecx)
	movd	%mm4, 19 * SIZE(%ecx)
	movd	%mm5, 20 * SIZE(%ecx)
	movd	%mm5, 21 * SIZE(%ecx)
	movd	%mm5, 22 * SIZE(%ecx)
	movd	%mm5, 23 * SIZE(%ecx)
	movd	%mm6, 24 * SIZE(%ecx)
	movd	%mm6, 25 * SIZE(%ecx)
	movd	%mm6, 26 * SIZE(%ecx)
	movd	%mm6, 27 * SIZE(%ecx)
	movd	%mm7, 28 * SIZE(%ecx)
	movd	%mm7, 29 * SIZE(%ecx)
	movd	%mm7, 30 * SIZE(%ecx)
	movd	%mm7, 31 * SIZE(%ecx)
#endif
	addl	$ 8 * SIZE, %edi
	addl	$32 * SIZE, %ecx
	decl	%eax
	jne	.L42
	ALIGN_4

.L45:
	movl	K, %eax
	andl	$3, %eax
	BRANCH
	jle	.L50
	ALIGN_4

.L46:

#ifdef HAVE_SSE2
	movss	 0 * SIZE(%edi), %xmm0
	movss	 1 * SIZE(%edi), %xmm1

	shufps	 $0, %xmm0, %xmm0
	shufps	 $0, %xmm1, %xmm1

	movaps	%xmm0,  0 * SIZE(%ecx)
	movaps	%xmm1,  4 * SIZE(%ecx)
#else
	movd	 0 * SIZE(%edi), %mm0
	movd	 1 * SIZE(%edi), %mm1

	movd	%mm0,  0 * SIZE(%ecx)
	movd	%mm0,  1 * SIZE(%ecx)
	movd	%mm0,  2 * SIZE(%ecx)
	movd	%mm0,  3 * SIZE(%ecx)
	movd	%mm1,  4 * SIZE(%ecx)
	movd	%mm1,  5 * SIZE(%ecx)
	movd	%mm1,  6 * SIZE(%ecx)
	movd	%mm1,  7 * SIZE(%ecx)
#endif
	addl	$2 * SIZE, %edi
	addl	$8 * SIZE, %ecx
	decl	%eax
	jne	.L46
	ALIGN_4
	
.L50:
	movl	C, %esi		# coffset = c
	movl	A, %edx		# aoffset = a
	movl	M,  %ebx
	sarl	$2, %ebx	# i = (m >> 2)
	jle	.L60
	ALIGN_4

.L51:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leal	BUFFER, BB	# boffset1 = boffset
#else
	leal	BUFFER, BB	# boffset1 = boffset
	movl	KK, %eax
	leal	(, %eax,   8), %eax
	leal	(AA, %eax, 2), AA
	leal	(BB, %eax, 4), BB
#endif	

	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5
	pxor	%xmm6, %xmm6
	pxor	%xmm7, %xmm7

	movaps	 0 * SIZE(AA), %xmm0
	movaps	16 * SIZE(AA), %xmm1
	movaps	 0 * SIZE(BB), %xmm2
	movaps	16 * SIZE(BB), %xmm3

#ifdef HAVE_3DNOW
	prefetchw	4 * SIZE(%esi)
	prefetchw	4 * SIZE(%esi, LDC)
#elif defined(HAVE_SSE) || defined(HAVE_SSE2)
	prefetcht2	4 * SIZE(%esi)
	prefetcht2	4 * SIZE(%esi, LDC)
#endif

#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	.L55
	ALIGN_4

.L52:
	mulps	%xmm0, %xmm2
#if defined(OPTERON) || defined(BARCELONA)
	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
#endif
	mulps	 4 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm4
	movaps	 8 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm5
	movaps	 4 * SIZE(AA), %xmm0

	mulps	%xmm0, %xmm2
	mulps	12 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm4
	movaps	32 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm5
	movaps	 8 * SIZE(AA), %xmm0

	mulps	%xmm0, %xmm3
	mulps	20 * SIZE(BB), %xmm0
	addps	%xmm3, %xmm4
	movaps	24 * SIZE(BB), %xmm3
	addps	%xmm0, %xmm5
	movaps	12 * SIZE(AA), %xmm0

	mulps	%xmm0, %xmm3
	mulps	28 * SIZE(BB), %xmm0
	addps	%xmm3, %xmm4
	movaps	48 * SIZE(BB), %xmm3
	addps	%xmm0, %xmm5
	movaps	32 * SIZE(AA), %xmm0

#if defined(OPTERON) || defined(BARCELONA)
	prefetcht0 (PREFETCHSIZE  + 16) * SIZE(AA)
#endif
	mulps	%xmm1, %xmm2
	mulps	36 * SIZE(BB), %xmm1
	addps	%xmm2, %xmm4
	movaps	40 * SIZE(BB), %xmm2
	addps	%xmm1, %xmm5
	movaps	20 * SIZE(AA), %xmm1

	mulps	%xmm1, %xmm2
	mulps	44 * SIZE(BB), %xmm1
	addps	%xmm2, %xmm4
	movaps	64 * SIZE(BB), %xmm2
	addps	%xmm1, %xmm5
	movaps	24 * SIZE(AA), %xmm1

	mulps	%xmm1, %xmm3
	mulps	52 * SIZE(BB), %xmm1
	addps	%xmm3, %xmm4
	movaps	56 * SIZE(BB), %xmm3
	addps	%xmm1, %xmm5
	movaps	28 * SIZE(AA), %xmm1

	mulps	%xmm1, %xmm3
	mulps	60 * SIZE(BB), %xmm1
	addps	%xmm3, %xmm4
	movaps	80 * SIZE(BB), %xmm3
	addps	%xmm1, %xmm5
	movaps	48 * SIZE(AA), %xmm1

	addl	$32 * SIZE, AA
	addl	$64 * SIZE, BB
	decl   %eax
	jne    .L52
	ALIGN_4

.L55:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	movaps	ALPHA,  %xmm3
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L58
	ALIGN_4

.L56:
	mulps	%xmm0, %xmm2
	mulps	 4 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm4
	movaps	 8 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm5
	movaps	 4 * SIZE(AA), %xmm0

	addl	$4 * SIZE, AA
	addl	$8 * SIZE, BB
	decl	%eax
	jg	.L56
	ALIGN_4

.L58:
	mulps	%xmm3, %xmm4
	mulps	%xmm3, %xmm5

#ifndef TRMMKERNEL
	movsd	0 * SIZE(%esi), %xmm0
	movhps	2 * SIZE(%esi), %xmm0
	movsd	0 * SIZE(%esi, LDC, 1), %xmm1
	movhps	2 * SIZE(%esi, LDC, 1), %xmm1

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

	movsd	%xmm4, 0 * SIZE(%esi)
	movhps	%xmm4, 2 * SIZE(%esi)
	movsd	%xmm5, 0 * SIZE(%esi, LDC, 1)
	movhps	%xmm5, 2 * SIZE(%esi, LDC, 1)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movl	K, %eax
	subl	KKK, %eax
	leal	(,%eax, 8), %eax
	leal	(AA, %eax, 2), AA
	leal	(BB, %eax, 4), BB
#endif

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

	addl	$4 * SIZE, %esi		# coffset += 2
	decl	%ebx			# i --
	jg	.L51
	ALIGN_4

.L60:
	testl	$2, M
	je	.L70

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

	leal	BUFFER, BB	# boffset1 = boffset
#else
	leal	BUFFER, BB	# boffset1 = boffset
	movl	KK, %eax
	leal	(, %eax,   8), %eax
	leal	(AA, %eax, 1), AA
	leal	(BB, %eax, 4), BB
#endif	

	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5
	pxor	%xmm6, %xmm6
	pxor	%xmm7, %xmm7

	movsd	 0 * SIZE(AA), %xmm0
	movsd	 8 * SIZE(AA), %xmm1
	movsd	 0 * SIZE(BB), %xmm2
	movsd	16 * SIZE(BB), %xmm3

	leal	(LDC, LDC, 2), %eax

#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	.L65
	ALIGN_4

.L62:
#if defined(OPTERON) || defined(BARCELONA)
	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
#endif

	mulps	%xmm0, %xmm2
	addps	%xmm2, %xmm4
	movsd	 4 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	movsd	 2 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5
	movsd	 8 * SIZE(BB), %xmm2

	mulps	%xmm0, %xmm2
	addps	%xmm2, %xmm6
	movsd	12 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	movsd	 4 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm7
	movsd	32 * SIZE(BB), %xmm2

	mulps	%xmm0, %xmm3
	addps	%xmm3, %xmm4
	movsd	20 * SIZE(BB), %xmm3
	mulps	%xmm0, %xmm3
	movsd	 6 * SIZE(AA), %xmm0
	addps	%xmm3, %xmm5
	movsd	24 * SIZE(BB), %xmm3

	mulps	%xmm0, %xmm3
	addps	%xmm3, %xmm6
	movsd	28 * SIZE(BB), %xmm3
	mulps	%xmm0, %xmm3
	movsd	16 * SIZE(AA), %xmm0
	addps	%xmm3, %xmm7
	movsd	48 * SIZE(BB), %xmm3

	mulps	%xmm1, %xmm2
	addps	%xmm2, %xmm4
	movsd	36 * SIZE(BB), %xmm2
	mulps	%xmm1, %xmm2
	movsd	10 * SIZE(AA), %xmm1
	addps	%xmm2, %xmm5
	movsd	40 * SIZE(BB), %xmm2

	mulps	%xmm1, %xmm2
	addps	%xmm2, %xmm6
	movsd	44 * SIZE(BB), %xmm2
	mulps	%xmm1, %xmm2
	movsd	12 * SIZE(AA), %xmm1
	addps	%xmm2, %xmm7
	movsd	64 * SIZE(BB), %xmm2

	mulps	%xmm1, %xmm3
	addps	%xmm3, %xmm4
	movsd	52 * SIZE(BB), %xmm3
	mulps	%xmm1, %xmm3
	movsd	14 * SIZE(AA), %xmm1
	addps	%xmm3, %xmm5
	movsd	56 * SIZE(BB), %xmm3

	mulps	%xmm1, %xmm3
	addps	%xmm3, %xmm6
	movsd	60 * SIZE(BB), %xmm3
	mulps	%xmm1, %xmm3
	movsd	24 * SIZE(AA), %xmm1
	addps	%xmm3, %xmm7
	movsd	80 * SIZE(BB), %xmm3

	addl	$16 * SIZE, AA
	addl	$64 * SIZE, BB
	decl   %eax
	jne    .L62
	ALIGN_4

.L65:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	movaps	ALPHA,  %xmm3
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L68
	ALIGN_4

.L66:
	mulps	%xmm0, %xmm2
	addps	%xmm2, %xmm4
	movsd	 4 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	movsd	 2 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5
	movsd	 8 * SIZE(BB), %xmm2

	addl	$2 * SIZE, AA
	addl	$8 * SIZE, BB
	decl	%eax
	jg	.L66
	ALIGN_4

.L68:
	addps	%xmm6, %xmm4
	addps	%xmm7, %xmm5

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

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

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

	movsd	%xmm4, 0 * SIZE(%esi)
	movsd	%xmm5, 0 * SIZE(%esi, LDC, 1)
	addl	$2 * SIZE, %esi		# coffset += 2

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movl	K, %eax
	subl	KKK, %eax
	leal	(,%eax, 8), %eax
	leal	(AA, %eax, 1), AA
	leal	(BB, %eax, 4), BB
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addl	$2, KK
#endif
	ALIGN_4

.L70:
	testl	$1, M
	je	.L79

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

	leal	BUFFER, BB	# boffset1 = boffset
#else
	leal	BUFFER, BB	# boffset1 = boffset
	movl	KK, %eax
	leal	(, %eax,   4), %eax
	leal	(AA, %eax, 1), AA
	leal	(BB, %eax, 8), BB
#endif	

	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5
	pxor	%xmm6, %xmm6
	pxor	%xmm7, %xmm7

	movss	 0 * SIZE(AA), %xmm0
	movss	 4 * SIZE(AA), %xmm1
	movss	 0 * SIZE(BB), %xmm2
	movss	16 * SIZE(BB), %xmm3

	leal	(LDC, LDC, 2), %eax

#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	.L75
	ALIGN_4

.L72:
	mulss	%xmm0, %xmm2
#if defined(OPTERON) || defined(BARCELONA)
	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
#endif
	mulss	 4 * SIZE(BB), %xmm0
	addss	%xmm2, %xmm4
	movss	 8 * SIZE(BB), %xmm2
	addss	%xmm0, %xmm5
	movss	 1 * SIZE(AA), %xmm0
	mulss	%xmm0, %xmm2
	mulss	12 * SIZE(BB), %xmm0
	addss	%xmm2, %xmm6
	movss	32 * SIZE(BB), %xmm2
	addss	%xmm0, %xmm7
	movss	 2 * SIZE(AA), %xmm0
	mulss	%xmm0, %xmm3
	mulss	20 * SIZE(BB), %xmm0
	addss	%xmm3, %xmm4
	movss	24 * SIZE(BB), %xmm3
	addss	%xmm0, %xmm5
	movss	 3 * SIZE(AA), %xmm0
	mulss	%xmm0, %xmm3
	mulss	28 * SIZE(BB), %xmm0
	addss	%xmm3, %xmm6
	movss	48 * SIZE(BB), %xmm3
	addss	%xmm0, %xmm7
	movss	 8 * SIZE(AA), %xmm0
	mulss	%xmm1, %xmm2
	mulss	36 * SIZE(BB), %xmm1
	addss	%xmm2, %xmm4
	movss	40 * SIZE(BB), %xmm2
	addss	%xmm1, %xmm5
	movss	 5 * SIZE(AA), %xmm1
	mulss	%xmm1, %xmm2
	mulss	44 * SIZE(BB), %xmm1
	addss	%xmm2, %xmm6
	movss	64 * SIZE(BB), %xmm2
	addss	%xmm1, %xmm7
	movss	 6 * SIZE(AA), %xmm1
	mulss	%xmm1, %xmm3
	mulss	52 * SIZE(BB), %xmm1
	addss	%xmm3, %xmm4
	movss	56 * SIZE(BB), %xmm3
	addss	%xmm1, %xmm5
	movss	 7 * SIZE(AA), %xmm1
	mulss	%xmm1, %xmm3
	mulss	60 * SIZE(BB), %xmm1
	addss	%xmm3, %xmm6
	movss	80 * SIZE(BB), %xmm3
	addss	%xmm1, %xmm7
	movss	12 * SIZE(AA), %xmm1

	addl	$ 8 * SIZE, AA
	addl	$64 * SIZE, BB
	decl   %eax