ztrsm_kernel_rt_1x2_sse3.s

Optimized GotoBLAS libraries

	addpd	 %xmm4, %xmm5
#endif

#if defined(RN) || defined(RT)
	movddup	 0 * SIZE(BB), %xmm2
	movddup	 1 * SIZE(BB), %xmm3

	movapd	 %xmm5, %xmm4
	SHUFPD_1 %xmm4, %xmm4

	xorpd	 %xmm1, %xmm4

	mulpd	 %xmm2, %xmm5
	mulpd	 %xmm3, %xmm4

	addpd	 %xmm4, %xmm5
#endif

#ifdef LN
	subl	$2 * SIZE, CO1
#endif

	movlpd	%xmm5,   0 * SIZE(CO1)
	movhpd	%xmm5,   1 * SIZE(CO1)

#if defined(LN) || defined(LT)
	movapd	%xmm5,   0 * SIZE(BB)
#else
	movapd	%xmm5,   0 * SIZE(AA)
#endif

#ifndef LN
	addl	$2 * SIZE, CO1
#endif

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	sall	$ZBASE_SHIFT, %eax
	addl	%eax, AA
	addl	%eax, BB
#endif

#ifdef LN
	subl	$1, KK
#endif

#ifdef LT
	addl	$1, KK
#endif

#ifdef RT
	movl	K, %eax
	sall	$ZBASE_SHIFT, %eax
	addl	%eax, AORIG
#endif

	decl	%ebx			# i --
	jg	L110

#ifdef LN
       movl	K, %eax
       sall	$ZBASE_SHIFT, %eax
       addl	%eax, B
#endif

#if defined(LT) || defined(RN)
	movl	BB, B
#endif

#ifdef RN
	addl	$1, KK
#endif

#ifdef RT
	subl	$1, KK
#endif
	ALIGN_4	

.L100:
	movl	N, %eax
	sarl	$1, %eax
	movl	%eax, J			# j = n
	jle	.L500
	ALIGN_4

.L01:
#if defined(LT) || defined(RN)
	movl	A, AA
#else
	movl	A, %eax
	movl	%eax, AORIG
#endif

#ifdef RT
	movl	K, %eax
	sall	$1 + ZBASE_SHIFT, %eax
	subl	%eax, B
#endif

	leal	(, LDC, 2), %eax

#ifdef RT
	subl	%eax, C
#endif
	movl	C, CO1
#ifndef RT
	addl	%eax, C
#endif

#ifdef LN
	movl	OFFSET, %eax
	addl	M, %eax
	movl	%eax, KK
#endif	

#ifdef LT
	movl	OFFSET, %eax
	movl	%eax, KK
#endif

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

.L10:
#ifdef LN
	movl	K, %eax
	sall	$ZBASE_SHIFT, %eax
	subl	%eax, AORIG
#endif

#if defined(LN) || defined(RT)
	movl	KK, %eax
	movl	AORIG, AA
	sall	$ZBASE_SHIFT, %eax
	addl	%eax, AA
#endif

	movl	B, BB

#if defined(LN) || defined(RT)
	movl	KK, %eax
	sall	$1 + ZBASE_SHIFT, %eax
	addl	%eax, 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 LN
	prefetcht0	-2 * SIZE(CO1)
	prefetcht0	-2 * SIZE(CO1, LDC, 1)
#else
	prefetchnta	 2 * SIZE(CO1)
	prefetchnta	 2 * SIZE(CO1, LDC, 1)
#endif

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	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

.L12:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	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:
#if defined(LN) || defined(RT)
	movl	KK, %eax
#ifdef LN
	subl	$1, %eax
#else
	subl	$2, %eax
#endif

	movl	AORIG, AA
	sall	$ZBASE_SHIFT, %eax
	leal	(AA, %eax, 1), AA
	leal	(B,  %eax, 2), BB
#endif

	pcmpeqb	%xmm1, %xmm1
	psllq	$63,   %xmm1

	shufps	$0x40, %xmm1, %xmm1

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

#ifndef CONJ
	xorpd	%xmm1, %xmm5
	xorpd	%xmm1, %xmm7

	subpd	%xmm5, %xmm4
	subpd	%xmm7, %xmm6
#else
#if defined(LN) || defined(LT)
	xorpd	%xmm1, %xmm4
	xorpd	%xmm1, %xmm6
#else
	xorpd	%xmm1, %xmm5
	xorpd	%xmm1, %xmm7
#endif
	addpd	%xmm5, %xmm4
	addpd	%xmm7, %xmm6
#endif

#if defined(LN) || defined(LT)
	movapd	 0 * SIZE(BB), %xmm5
	movapd	 2 * SIZE(BB), %xmm7

	subpd	%xmm4,  %xmm5
	subpd	%xmm6,  %xmm7
#else
	movapd	 0 * SIZE(AA), %xmm5
	movapd	 2 * SIZE(AA), %xmm7

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

#ifndef CONJ
	SHUFPD_1 %xmm1, %xmm1
#endif

#if defined(LN) || defined(LT)
	movddup	 0 * SIZE(AA), %xmm2
	movddup	 1 * SIZE(AA), %xmm3

	movapd	%xmm5, %xmm4
	movapd	%xmm7, %xmm6

	SHUFPD_1 %xmm4, %xmm4
	SHUFPD_1 %xmm6, %xmm6

	xorpd	 %xmm1, %xmm4
	xorpd	 %xmm1, %xmm6

	mulpd	 %xmm2, %xmm5
	mulpd	 %xmm3, %xmm4
	mulpd	 %xmm2, %xmm7
	mulpd	 %xmm3, %xmm6

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

#ifdef RN
	movddup	 0 * SIZE(BB), %xmm2
	movddup	 1 * SIZE(BB), %xmm3

	movapd	%xmm5, %xmm4
	SHUFPD_1 %xmm4, %xmm4

	xorpd	 %xmm1, %xmm4

	mulpd	 %xmm2, %xmm5
	mulpd	 %xmm3, %xmm4

	addpd	 %xmm4, %xmm5

	movddup	 2 * SIZE(BB), %xmm2
	movddup	 3 * SIZE(BB), %xmm3

	movapd	 %xmm5, %xmm4
	movapd	 %xmm5, %xmm6
	SHUFPD_1 %xmm6, %xmm6

	xorpd	 %xmm1, %xmm6

	mulpd	 %xmm2, %xmm4
	mulpd	 %xmm3, %xmm6

	subpd	 %xmm4, %xmm7
	subpd	 %xmm6, %xmm7

	movddup	 6 * SIZE(BB), %xmm2
	movddup	 7 * SIZE(BB), %xmm3

	movapd	 %xmm7, %xmm6
	SHUFPD_1 %xmm6, %xmm6

	xorpd	 %xmm1, %xmm6

	mulpd	 %xmm2, %xmm7
	mulpd	 %xmm3, %xmm6

	addpd	 %xmm6, %xmm7
#endif

#ifdef RT
	movddup	 6 * SIZE(BB), %xmm2
	movddup	 7 * SIZE(BB), %xmm3

	movapd	 %xmm7, %xmm6
	SHUFPD_1 %xmm6, %xmm6

	xorpd	 %xmm1, %xmm6

	mulpd	 %xmm2, %xmm7
	mulpd	 %xmm3, %xmm6

	addpd	 %xmm6, %xmm7

	movddup	 4 * SIZE(BB), %xmm2
	movddup	 5 * SIZE(BB), %xmm3

	movapd	 %xmm7, %xmm4
	movapd	 %xmm7, %xmm6
	SHUFPD_1 %xmm6, %xmm6

	xorpd	 %xmm1, %xmm6

	mulpd	 %xmm2, %xmm4
	mulpd	 %xmm3, %xmm6

	subpd	 %xmm4, %xmm5
	subpd	 %xmm6, %xmm5

	movddup	 0 * SIZE(BB), %xmm2
	movddup	 1 * SIZE(BB), %xmm3

	movapd	 %xmm5, %xmm4
	SHUFPD_1 %xmm4, %xmm4

	xorpd	 %xmm1, %xmm4

	mulpd	 %xmm2, %xmm5
	mulpd	 %xmm3, %xmm4

	addpd	 %xmm4, %xmm5
#endif

#ifdef LN
	subl	$2 * SIZE, CO1
#endif

	movlpd	%xmm5,   0 * SIZE(CO1)
	movhpd	%xmm5,   1 * SIZE(CO1)

	movlpd	%xmm7,   0 * SIZE(CO1, LDC)
	movhpd	%xmm7,   1 * SIZE(CO1, LDC)

#if defined(LN) || defined(LT)
	movapd	%xmm5,   0 * SIZE(BB)
	movapd	%xmm7,   2 * SIZE(BB)
#else
	movapd	%xmm5,   0 * SIZE(AA)
	movapd	%xmm7,   2 * SIZE(AA)

#endif

#ifndef LN
	addl	$2 * SIZE, CO1
#endif

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	sall	$ZBASE_SHIFT, %eax
	addl	%eax, AA
	leal	(BB, %eax, 2), BB
#endif

#ifdef LN
	subl	$1, KK
#endif

#ifdef LT
	addl	$1, KK
#endif

#ifdef RT
	movl	K, %eax
	sall	$ZBASE_SHIFT, %eax
	addl	%eax, AORIG
#endif

	decl	%ebx			# i --
	jg	.L10
	ALIGN_4	

.L99:
#ifdef LN
       movl	K, %eax
       sall	$1 + ZBASE_SHIFT, %eax
       addl	%eax, B
#endif

#if defined(LT) || defined(RN)
	movl	BB, B
#endif

#ifdef RN
	addl	$2, KK
#endif

#ifdef RT
	subl	$2, KK
#endif

	decl	J			# j --
	jg	.L01
	ALIGN_4

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

	addl	$ARGS, %esp
	ret

	EPILOGUE