zgemm_kernel_1x2_sse3.s

Optimized GotoBLAS libraries

	jle	.L11
	KERNEL1(16  * 13)
	KERNEL2(16  * 13)
	KERNEL3(16  * 13)
	KERNEL4(16  * 13)
	KERNEL5(16  * 13)
	KERNEL6(16  * 13)
	KERNEL7(16  * 13)
	KERNEL8(16  * 13)
	cmpl	$128 * 14, %eax
	jle	.L11
	KERNEL1(16  * 14)
	KERNEL2(16  * 14)
	KERNEL3(16  * 14)
	KERNEL4(16  * 14)
	KERNEL5(16  * 14)
	KERNEL6(16  * 14)
	KERNEL7(16  * 14)
	KERNEL8(16  * 14)
	cmpl	$128 * 15, %eax
	jle	.L11
	KERNEL1(16  * 15)
	KERNEL2(16  * 15)
	KERNEL3(16  * 15)
	KERNEL4(16  * 15)
	KERNEL5(16  * 15)
	KERNEL6(16  * 15)
	KERNEL7(16  * 15)
	KERNEL8(16  * 15)
#else
	addl	$32 * 4  * SIZE, AA
	addl	$32 * 8  * SIZE, BB
	subl	$128 * 8, %eax
	jg	.L1X
#endif

.L11:
	leal	(AA, %eax, 1), AA	# * 16
	leal	(BB, %eax, 2), BB	# * 64

#else

	sarl	$3, %eax
	je	.L12
	ALIGN_4

.L11:
	KERNEL1(16  *  0)
	KERNEL2(16  *  0)
	KERNEL3(16  *  0)
	KERNEL4(16  *  0)
	KERNEL5(16  *  0)
	KERNEL6(16  *  0)
	KERNEL7(16  *  0)
	KERNEL8(16  *  0)

	addl   $32 * SIZE, BB
	addl   $16 * SIZE, AA
	decl   %eax
	jne    .L11
	ALIGN_4
#endif

.L12:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	movddup	ALPHA_R, %xmm1
	movddup	ALPHA_I, %xmm3
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L14
	ALIGN_4

.L13:
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm4
	movddup	 1 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	ADDSUB	 %xmm2, %xmm5
	movddup	 2 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm6
	movddup	 3 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	 2 * SIZE(AA), %xmm0
	ADDSUB	 %xmm2, %xmm7
	movddup	 4 * SIZE(BB), %xmm2

	addl	$2 * SIZE, AA
	addl	$4 * SIZE, BB
	decl	%eax
	jg	.L13
	ALIGN_4

.L14:
	pcmpeqb	%xmm0, %xmm0
	SHUFPD_1 %xmm5, %xmm5
	psllq	$63,   %xmm0
	SHUFPD_1 %xmm7, %xmm7

#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
	shufps	$0x04, %xmm0, %xmm0

	pxor	%xmm0, %xmm5
	pxor	%xmm0, %xmm7
#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
	shufps	$0x40, %xmm0, %xmm0

	pxor	%xmm0, %xmm5
	pxor	%xmm0, %xmm7
#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
	shufps	$0x40, %xmm0, %xmm0

	pxor	%xmm0, %xmm4
	pxor	%xmm0, %xmm6
#else
	shufps	$0x40, %xmm0, %xmm0

	pxor	%xmm0, %xmm4
	pxor	%xmm0, %xmm6
#endif

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

	movaps	%xmm4, %xmm5
	movaps	%xmm6, %xmm7

	SHUFPD_1 %xmm5, %xmm5
	SHUFPD_1 %xmm7, %xmm7

	mulpd	%xmm1, %xmm4
	mulpd	%xmm1, %xmm6

	mulpd	%xmm3, %xmm5
	mulpd	%xmm3, %xmm7

	addsubpd	%xmm5, %xmm4
	addsubpd	%xmm7, %xmm6
	
#ifndef TRMMKERNEL
	movsd	0 * SIZE(%esi), %xmm0
	movhpd	1 * SIZE(%esi), %xmm0
	movsd	0 * SIZE(%esi, LDC), %xmm2
	movhpd	1 * SIZE(%esi, LDC), %xmm2

	addpd	%xmm0, %xmm4
	addpd	%xmm2, %xmm6
#endif

	movsd	%xmm4, 0 * SIZE(%esi)
	movhpd	%xmm4, 1 * SIZE(%esi)
	movsd	%xmm6, 0 * SIZE(%esi, LDC)
	movhpd	%xmm6, 1 * SIZE(%esi, 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, %esi		# coffset += 4
	decl	%ebx			# i --
	jg	.L10
	ALIGN_4	

.L99:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$2, KK
#endif

	leal	(, LDC, 2), %eax
	movl	BB, B
	addl	%eax, C			# c += ldc
	decl	J			# j --
	jg	.L01
	ALIGN_4

.L100:
	movl	N, %eax
	testl	$1, %eax
	jle	.L500

	movl	C, %esi		# coffset = c
	movl	A, AA		# aoffset = a

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

	movl	M,  %ebx
	testl	%ebx, %ebx	
	jle	.L500
	ALIGN_4

L110:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movl	B, BB
#else
	movl	KK, %eax
	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 2), AA
	leal	(B,  %eax, 2), BB
#endif	

	movapd	 0 * SIZE(AA), %xmm0
	pxor	%xmm4, %xmm4
	movapd	 8 * SIZE(AA), %xmm1
	pxor	%xmm5, %xmm5
	movddup	 0 * SIZE(BB), %xmm2
	pxor	%xmm6, %xmm6
	movddup	 8 * SIZE(BB), %xmm3
	pxor	%xmm7, %xmm7

#ifdef PENTIUM4
	prefetchnta	4 * SIZE(%esi)
#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	$1, %eax
#else
	addl	$1, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	L112
	ALIGN_4

L111:
	PREFETCH  (PREFETCHSIZE + 0) * SIZE(AA)
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm4
	movddup	 1 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	 2 * SIZE(AA), %xmm0
	ADDSUB	 %xmm2, %xmm5
	movddup	 2 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm6
	movddup	 3 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	 4 * SIZE(AA), %xmm0
	ADDSUB	 %xmm2, %xmm7
	movddup	 4 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm4
	movddup	 5 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	 6 * SIZE(AA), %xmm0
	ADDSUB	 %xmm2, %xmm5
	movddup	 6 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm6
	movddup	 7 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	16 * SIZE(AA), %xmm0
	ADDSUB	 %xmm2, %xmm7
	movddup	16 * SIZE(BB), %xmm2
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm4
	movddup	 9 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	movapd	10 * SIZE(AA), %xmm1
	ADDSUB	 %xmm3, %xmm5
	movddup	10 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm6
	movddup	11 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	movapd	12 * SIZE(AA), %xmm1
	ADDSUB	 %xmm3, %xmm7
	movddup	12 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm4
	movddup	13 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	movapd	14 * SIZE(AA), %xmm1
	ADDSUB	 %xmm3, %xmm5
	movddup	14 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	addpd	 %xmm3, %xmm6
	movddup	15 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm3
	movapd	24 * SIZE(AA), %xmm1
	ADDSUB	 %xmm3, %xmm7
	movddup	24 * SIZE(BB), %xmm3

	addl   $16 * SIZE, AA
	addl   $16 * SIZE, BB
	decl   %eax
	jne    L111
	ALIGN_4

L112:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	movddup	ALPHA_R, %xmm1
	movddup	ALPHA_I, %xmm3
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je L114
	ALIGN_4

L113:
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm4
	movddup	 1 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	 2 * SIZE(AA), %xmm0
	ADDSUB	 %xmm2, %xmm5
	movddup	 2 * SIZE(BB), %xmm2

	addl	$2 * SIZE, AA
	addl	$2 * SIZE, BB
	decl	%eax
	jg	L113
	ALIGN_4

L114:
	addpd	%xmm6, %xmm4
	addpd	%xmm7, %xmm5

	pcmpeqb	%xmm0, %xmm0
	SHUFPD_1 %xmm5, %xmm5
	psllq	$63,   %xmm0

#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
	shufps	$0x04, %xmm0, %xmm0

	pxor	%xmm0, %xmm5
#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
	shufps	$0x40, %xmm0, %xmm0

	pxor	%xmm0, %xmm5
#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
	shufps	$0x40, %xmm0, %xmm0

	pxor	%xmm0, %xmm4
#else
	shufps	$0x40, %xmm0, %xmm0

	pxor	%xmm0, %xmm4
#endif

	addpd	%xmm5, %xmm4

	movaps	%xmm4, %xmm5

	SHUFPD_1 %xmm5, %xmm5

	mulpd	%xmm1, %xmm4
	mulpd	%xmm3, %xmm5

	addsubpd	%xmm5, %xmm4

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

	addpd	%xmm0, %xmm4
#endif

	movsd	%xmm4, 0 * SIZE(%esi)
	movhpd	%xmm4, 1 * 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	$1, KK
#endif

	addl	$2 * SIZE, %esi		# coffset += 4
	decl	%ebx			# i --
	jg	L110
	ALIGN_4	

.L500:
	popl	%ebx
	popl	%esi
	popl	%edi
	popl	%ebp

	addl	$ARGS, %esp
	ret

	EPILOGUE