ztrsm_kernel_ln_2x1_sse2.s

Optimized GotoBLAS libraries

	addpd	 %xmm4, %xmm5
#endif

#ifdef LT
	movsd	 0 * SIZE(AA), %xmm2
	movhpd	 0 * SIZE(AA), %xmm2
	movsd	 1 * SIZE(AA), %xmm3
	movhpd	 1 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4

	xorpd	 %xmm1, %xmm4

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

	addpd	 %xmm4, %xmm5
#endif

#ifdef RN
	movsd	 0 * SIZE(B), %xmm2
	movhpd	 0 * SIZE(B), %xmm2
	movsd	 1 * SIZE(B), %xmm3
	movhpd	 1 * SIZE(B), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4

	xorpd	 %xmm1, %xmm4

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

	addpd	 %xmm4, %xmm5
#endif

#ifdef RT
	movsd	 0 * SIZE(B), %xmm2
	movhpd	 0 * SIZE(B), %xmm2
	movsd	 1 * SIZE(B), %xmm3
	movhpd	 1 * 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,   0 * SIZE(B)

	movsd	%xmm5,   0 * SIZE(BB)
	movsd	%xmm5,   1 * SIZE(BB)
	movhpd	%xmm5,   2 * SIZE(BB)
	movhpd	%xmm5,   3 * 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
#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

.L50:
	movl	M,  %ebx
	sarl	$1, %ebx	# i = (m >> 2)
	jle	.L99
	ALIGN_4

.L10:
#ifdef LN
       movl	K, %eax
       sall	$1 + ZBASE_SHIFT, %eax
       subl	%eax, AORIG
#endif

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

	movl	KK, %eax
	sall	$1 + ZBASE_SHIFT, %eax
	addl	%eax, AA
#endif

	leal	BUFFER, BB

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

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

#ifdef LN
	prefetchnta	-4 * SIZE(CO1)
#else
	prefetchnta	 4 * SIZE(CO1)
#endif

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	andl	$-8, %eax
	je,pn	.L12
	sall	$3, %eax
	
.L1X:
	KERNEL1(32  *  0)
	KERNEL2(32  *  0)
	KERNEL3(32  *  0)
	KERNEL4(32  *  0)
	KERNEL5(32  *  0)
	KERNEL6(32  *  0)
	KERNEL7(32  *  0)
	KERNEL8(32  *  0)
	cmpl	$64 *  1, %eax
	jle,pn	.L11
	KERNEL1(32  *  1)
	KERNEL2(32  *  1)
	KERNEL3(32  *  1)
	KERNEL4(32  *  1)
	KERNEL5(32  *  1)
	KERNEL6(32  *  1)
	KERNEL7(32  *  1)
	KERNEL8(32  *  1)
	cmpl	$64 *  2, %eax
	jle,pn	.L11
	KERNEL1(32  *  2)
	KERNEL2(32  *  2)
	KERNEL3(32  *  2)
	KERNEL4(32  *  2)
	KERNEL5(32  *  2)
	KERNEL6(32  *  2)
	KERNEL7(32  *  2)
	KERNEL8(32  *  2)
	cmpl	$64 *  3, %eax
	jle,pn	.L11
	KERNEL1(32  *  3)
	KERNEL2(32  *  3)
	KERNEL3(32  *  3)
	KERNEL4(32  *  3)
	KERNEL5(32  *  3)
	KERNEL6(32  *  3)
	KERNEL7(32  *  3)
	KERNEL8(32  *  3)
	cmpl	$64 *  4, %eax
	jle,pn	.L11
	KERNEL1(32  *  4)
	KERNEL2(32  *  4)
	KERNEL3(32  *  4)
	KERNEL4(32  *  4)
	KERNEL5(32  *  4)
	KERNEL6(32  *  4)
	KERNEL7(32  *  4)
	KERNEL8(32  *  4)
	cmpl	$64 *  5, %eax
	jle,pn	.L11
	KERNEL1(32  *  5)
	KERNEL2(32  *  5)
	KERNEL3(32  *  5)
	KERNEL4(32  *  5)
	KERNEL5(32  *  5)
	KERNEL6(32  *  5)
	KERNEL7(32  *  5)
	KERNEL8(32  *  5)
	cmpl	$64 *  6, %eax
	jle,pn	.L11
	KERNEL1(32  *  6)
	KERNEL2(32  *  6)
	KERNEL3(32  *  6)
	KERNEL4(32  *  6)
	KERNEL5(32  *  6)
	KERNEL6(32  *  6)
	KERNEL7(32  *  6)
	KERNEL8(32  *  6)
	cmpl	$64 *  7, %eax
	jle,pn	.L11
	KERNEL1(32  *  7)
	KERNEL2(32  *  7)
	KERNEL3(32  *  7)
	KERNEL4(32  *  7)
	KERNEL5(32  *  7)
	KERNEL6(32  *  7)
	KERNEL7(32  *  7)
	KERNEL8(32  *  7)

	addl	$64 * 4  * SIZE, AA
	addl	$64 * 4  * SIZE, BB
	subl	$64 * 8, %eax
	jg,pt	.L1X

.L11:
	leal	(BB, %eax, 4), BB
	leal	(AA, %eax, 4), AA
	
.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

.L13:
	movapd	 2 * SIZE(BB), %xmm1
	mulpd	%xmm0, %xmm2
	addpd	%xmm2, %xmm4
	movapd	 0 * SIZE(BB), %xmm2
	mulpd	%xmm0, %xmm1
	movapd	 2 * SIZE(AA), %xmm0
	addpd	%xmm1, %xmm5
	movapd	 2 * SIZE(BB), %xmm1
	mulpd	%xmm0, %xmm2
	addpd	%xmm2, %xmm6
	movapd	 4 * SIZE(BB), %xmm2
	mulpd	%xmm0, %xmm1
	movapd	 4 * SIZE(AA), %xmm0
	addpd	%xmm1, %xmm7

	addl	$4 * SIZE, AA		# aoffset  += 8
	addl	$4 * SIZE, BB		# boffset1 += 8
	subl	$1, %eax
	jg	.L13

.L14:
#if defined(LN) || defined(RT)
	movl	KK, %eax
#ifdef LN
	subl	$2, %eax
#else
	subl	$1, %eax
#endif

	movl	AORIG, AA
	movl	BORIG, B
	leal	BUFFER, BB

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

	movapd	POSINV,  %xmm1

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

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(NR) || defined(NC) || defined(TR) || defined(TC)
	xorpd	%xmm1, %xmm5
	xorpd	%xmm1, %xmm7
#else
	xorpd	%xmm1, %xmm4
	xorpd	%xmm1, %xmm6
#endif

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(RR) || defined(RC) || defined(CR) || defined(CC)
	subpd	%xmm5, %xmm4
	subpd	%xmm7, %xmm6
#else
	addpd	%xmm5, %xmm4
	addpd	%xmm7, %xmm6
#endif

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

#ifdef LN
	movsd	 6 * SIZE(AA), %xmm2
	movhpd	 6 * SIZE(AA), %xmm2
	movsd	 7 * SIZE(AA), %xmm3
	movhpd	 7 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm7, %xmm6

	xorpd	 %xmm1, %xmm6

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

	addpd	 %xmm6, %xmm7

	movsd	 4 * SIZE(AA), %xmm2
	movhpd	 4 * SIZE(AA), %xmm2
	movsd	 5 * SIZE(AA), %xmm3
	movhpd	 5 * 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

	movsd	 0 * SIZE(AA), %xmm2
	movhpd	 0 * SIZE(AA), %xmm2
	movsd	 1 * SIZE(AA), %xmm3
	movhpd	 1 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4

	xorpd	 %xmm1, %xmm4

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

	addpd	 %xmm4, %xmm5
#endif

#ifdef LT
	movsd	 0 * SIZE(AA), %xmm2
	movhpd	 0 * SIZE(AA), %xmm2
	movsd	 1 * SIZE(AA), %xmm3
	movhpd	 1 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4

	xorpd	 %xmm1, %xmm4

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

	addpd	 %xmm4, %xmm5

	movsd	 2 * SIZE(AA), %xmm2
	movhpd	 2 * SIZE(AA), %xmm2
	movsd	 3 * SIZE(AA), %xmm3
	movhpd	 3 * 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

	movsd	 6 * SIZE(AA), %xmm2
	movhpd	 6 * SIZE(AA), %xmm2
	movsd	 7 * SIZE(AA), %xmm3
	movhpd	 7 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm7, %xmm6

	xorpd	 %xmm1, %xmm6

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

	addpd	 %xmm6, %xmm7
#endif

#ifdef RN
	movsd	 0 * SIZE(B), %xmm2
	movhpd	 0 * SIZE(B), %xmm2
	movsd	 1 * SIZE(B), %xmm3
	movhpd	 1 * 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
	movsd	 0 * SIZE(B), %xmm2
	movhpd	 0 * SIZE(B), %xmm2
	movsd	 1 * SIZE(B), %xmm3
	movhpd	 1 * 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,   0 * SIZE(B)
	movapd	%xmm7,   2 * SIZE(B)

	movsd	%xmm5,   0 * SIZE(BB)
	movsd	%xmm5,   1 * SIZE(BB)
	movhpd	%xmm5,   2 * SIZE(BB)
	movhpd	%xmm5,   3 * SIZE(BB)
	movsd	%xmm7,   4 * SIZE(BB)
	movsd	%xmm7,   5 * SIZE(BB)
	movhpd	%xmm7,   6 * SIZE(BB)
	movhpd	%xmm7,   7 * SIZE(BB)
#else
	movapd	%xmm5,   0 * SIZE(AA)
	movapd	%xmm7,   2 * 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

.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