ztrsm_kernel_lt_2x1_core2.s

Optimized GotoBLAS libraries

#endif

#if defined(LN) || defined(LT)
	movapd	-16 * SIZE(B), %xmm5
	movapd	-14 * SIZE(B), %xmm7

	subpd	%xmm4,  %xmm5
	subpd	%xmm6,  %xmm7
#else
	movapd	-16 * SIZE(AA), %xmm5
	movapd	-14 * SIZE(AA), %xmm7

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

#ifndef CONJ
	SHUFPD_1 %xmm1, %xmm1
#endif

#ifdef LN
	movddup	-10 * SIZE(AA), %xmm2
	movddup	 -9 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm7, %xmm6

	xorpd	 %xmm1, %xmm6

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

	addpd	 %xmm6, %xmm7

	movddup	-12 * SIZE(AA), %xmm2
	movddup	-11 * SIZE(AA), %xmm3

	movapd	 %xmm7, %xmm4
	pshufd	 $0x4e, %xmm7, %xmm6

	xorpd	 %xmm1, %xmm6

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

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

	movddup	-16 * SIZE(AA), %xmm2
	movddup	-15 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4

	xorpd	 %xmm1, %xmm4

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

	addpd	 %xmm4, %xmm5
#endif

#ifdef LT
	movddup	-16 * SIZE(AA), %xmm2
	movddup	-15 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4

	xorpd	 %xmm1, %xmm4

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

	addpd	 %xmm4, %xmm5

	movddup	-14 * SIZE(AA), %xmm2
	movddup	-13 * SIZE(AA), %xmm3

	movapd	 %xmm5, %xmm4
	pshufd	 $0x4e, %xmm5, %xmm6

	xorpd	 %xmm1, %xmm6

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

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

	movddup	-10 * SIZE(AA), %xmm2
	movddup	 -9 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm7, %xmm6

	xorpd	 %xmm1, %xmm6

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

	addpd	 %xmm6, %xmm7
#endif

#ifdef RN
	movddup	-16 * SIZE(B), %xmm2
	movddup	-15 * SIZE(B), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4
	pshufd	$0x4e, %xmm7, %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 RT
	movddup	-16 * SIZE(B), %xmm2
	movddup	-15 * SIZE(B), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4
	pshufd	$0x4e, %xmm7, %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 LN
	subl	$4 * SIZE, CO1
#endif

	movsd	%xmm5,   0 * SIZE(CO1)
	movhpd	%xmm5,   1 * SIZE(CO1)
	movsd	%xmm7,   2 * SIZE(CO1)
	movhpd	%xmm7,   3 * SIZE(CO1)

#if defined(LN) || defined(LT)
	movapd	%xmm5, -16 * SIZE(B)
	movapd	%xmm7, -14 * SIZE(B)

	movddup	%xmm5, %xmm4
	unpckhpd %xmm5, %xmm5
	movddup	%xmm7, %xmm6
	unpckhpd %xmm7, %xmm7

	movapd	%xmm4, -16 * SIZE(BB)
	movapd	%xmm5, -14 * SIZE(BB)
	movapd	%xmm6, -12 * SIZE(BB)
	movapd	%xmm7, -10 * SIZE(BB)
#else
	movapd	%xmm5, -16 * SIZE(AA)
	movapd	%xmm7, -14 * SIZE(AA)

#endif

#ifndef LN
	addl	$4 * SIZE, CO1
#endif

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	sall	$1 + ZBASE_SHIFT, %eax
	addl	%eax, AA
#ifdef LT
	addl	$4 * SIZE, B
#endif
#endif

#ifdef LN
	subl	$2, KK
	movl	BORIG, B
#endif

#ifdef LT
	addl	$2, KK
#endif

#ifdef RT
	movl	K, %eax
	movl	BORIG, B
	sall	$1 + ZBASE_SHIFT, %eax
	addl	%eax, AORIG
#endif

	decl	%ebx			# i --
	jg	.L10

.L50:
	movl	M,  %ebx
	testl	$1, %ebx
	je	.L99

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

#if defined(LN) || defined(RT)
	movl	AORIG, %eax
	movl	%eax, AA

	movl	KK, %eax
	sall	$ZBASE_SHIFT, %eax
	addl	%eax, AA
#endif

	leal	16 * SIZE + BUFFER, BB

#if defined(LN) || defined(RT)
	movl	KK, %eax
	sall	$1 + ZBASE_SHIFT, %eax
	addl	%eax, BB
#endif	

	movapd	-16 * SIZE(AA), %xmm0
	pxor	%xmm4, %xmm4
	movapd	-16 * SIZE(BB), %xmm1
	pxor	%xmm5, %xmm5
	movapd	 -8 * SIZE(AA), %xmm2
	pxor	%xmm6, %xmm6
	movapd	 -8 * SIZE(BB), %xmm3
	pxor	%xmm7, %xmm7

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	sarl	$3, %eax
	jle	.L52

.L51:
	mulpd	%xmm0,  %xmm1
	mulpd	-14 * SIZE(BB), %xmm0
	ADD1	%xmm1, %xmm4
	movapd	-12 * SIZE(BB), %xmm1
	ADD2	%xmm0, %xmm5
	movapd	-14 * SIZE(AA), %xmm0
	mulpd	%xmm0,  %xmm1
	mulpd	-10 * SIZE(BB), %xmm0
	ADD1	%xmm1, %xmm6
	movapd	  0 * SIZE(BB), %xmm1
	ADD2	%xmm0, %xmm7
	movapd	-12 * SIZE(AA), %xmm0
	mulpd	%xmm0,  %xmm3
	mulpd	 -6 * SIZE(BB), %xmm0
	ADD1	%xmm3, %xmm4
	movapd	 -4 * SIZE(BB), %xmm3
	ADD2	%xmm0, %xmm5
	movapd	-10 * SIZE(AA), %xmm0
	mulpd	%xmm0,  %xmm3
	mulpd	 -2 * SIZE(BB), %xmm0
	ADD1	%xmm3, %xmm6
	movapd	  8 * SIZE(BB), %xmm3
	ADD2	%xmm0, %xmm7
	movapd	  0 * SIZE(AA), %xmm0
	mulpd	%xmm2,  %xmm1
	mulpd	  2 * SIZE(BB), %xmm2
	ADD1	%xmm1, %xmm4
	movapd	  4 * SIZE(BB), %xmm1
	ADD2	%xmm2, %xmm5
	movapd	 -6 * SIZE(AA), %xmm2
	mulpd	%xmm2,  %xmm1
	mulpd	  6 * SIZE(BB), %xmm2
	ADD1	%xmm1, %xmm6
	movapd	 16 * SIZE(BB), %xmm1
	ADD2	%xmm2, %xmm7
	movapd	 -4 * SIZE(AA), %xmm2
	mulpd	%xmm2,  %xmm3
	mulpd	 10 * SIZE(BB), %xmm2
	ADD1	%xmm3, %xmm4
	movapd	 12 * SIZE(BB), %xmm3
	ADD2	%xmm2, %xmm5
	movapd	 -2 * SIZE(AA), %xmm2
	mulpd	%xmm2,  %xmm3
	mulpd	 14 * SIZE(BB), %xmm2
	ADD1	%xmm3, %xmm6
	movapd	 24 * SIZE(BB), %xmm3
	ADD2	%xmm2, %xmm7
	movapd	  8 * SIZE(AA), %xmm2

	subl   $-16 * SIZE, AA
	addl   $ 32 * SIZE, BB
	decl	%eax			# l--
	jg	.L51
	ALIGN_2

.L52:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K,  %eax
	subl	KK, %eax
#endif
	andl	$7, %eax		# l = (k & 3)
	jle	.L54
	ALIGN_2

.L53:
	mulpd	%xmm0,  %xmm1
	mulpd	-14 * SIZE(BB), %xmm0
	ADD1	%xmm1, %xmm4
	movapd	-12 * SIZE(BB), %xmm1
	ADD2	%xmm0, %xmm5
	movapd	-14 * SIZE(AA), %xmm0

	addl	$2 * SIZE, AA
	addl	$4 * SIZE, BB
	decl	%eax			# l--
	jg	.L53

.L54:
	addpd	%xmm6, %xmm4
	addpd	%xmm7, %xmm5

#if defined(LN) || defined(RT)
	movl	KK, %eax
#ifdef LN
	subl	$1, %eax
#else
	subl	$1, %eax
#endif

	movl	AORIG, AA
	movl	BORIG, B
	leal	16 * SIZE + BUFFER, BB

	sall	$ZBASE_SHIFT, %eax
	addl	%eax, AA
	addl	%eax, B
	leal	(BB, %eax, 2), BB
#endif

	movapd	POSINV,  %xmm1

	SHUFPD_1 %xmm5, %xmm5

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(NR) || defined(NC) || defined(TR) || defined(TC)
	xorpd	%xmm1, %xmm5
#else
	xorpd	%xmm1, %xmm4
#endif

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(RR) || defined(RC) || defined(CR) || defined(CC)
	subpd	%xmm5, %xmm4
#else
	addpd	%xmm5, %xmm4
#endif

#if defined(LN) || defined(LT)
	movapd	-16 * SIZE(B), %xmm5

	subpd	%xmm4,  %xmm5
#else
	movapd	-16 * SIZE(AA), %xmm5

	subpd	%xmm4,  %xmm5
#endif

#ifndef CONJ
	SHUFPD_1 %xmm1, %xmm1
#endif

#ifdef LN
	movddup	-16 * SIZE(AA), %xmm2
	movddup	-15 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4

	xorpd	 %xmm1, %xmm4

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

	addpd	 %xmm4, %xmm5
#endif

#ifdef LT
	movddup	-16 * SIZE(AA), %xmm2
	movddup	-15 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4

	xorpd	 %xmm1, %xmm4

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

	addpd	 %xmm4, %xmm5
#endif

#ifdef RN
	movddup	-16 * SIZE(B), %xmm2
	movddup	-15 * SIZE(B), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4

	xorpd	 %xmm1, %xmm4

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

	addpd	 %xmm4, %xmm5
#endif

#ifdef RT
	movddup	-16 * SIZE(B), %xmm2
	movddup	-15 * SIZE(B), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4

	xorpd	 %xmm1, %xmm4

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

	addpd	 %xmm4, %xmm5
#endif

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

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

#if defined(LN) || defined(LT)
	movapd	%xmm5, -16 * SIZE(B)

	movddup	%xmm5, %xmm4
	unpckhpd %xmm5, %xmm5

	movapd	%xmm4, -16 * SIZE(BB)
	movapd	%xmm5, -14 * SIZE(BB)
#else
	movapd	%xmm5, -16 * 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
#ifdef LT
	addl	$2 * SIZE, B
#endif
#endif

#ifdef LN
	subl	$1, KK
	movl	BORIG, B
#endif

#ifdef LT
	addl	$1, KK
#endif

#ifdef RT
	movl	K, %eax
	movl	BORIG, B
	sall	$ZBASE_SHIFT, %eax
	addl	%eax, AORIG
#endif
	ALIGN_4

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

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

#ifdef RN
	addl	$1, KK
#endif

#ifdef RT
	subl	$1, KK
#endif

	decl	J			# j --
	jg	.L01

.L999:
	movl	OLD_STACK, %esp

	EMMS

	popl	%ebx
	popl	%esi
	popl	%edi
	popl	%ebp
	ret

	EPILOGUE