trsm_kernel_rt_4x2_core2.s

Optimized GotoBLAS libraries

	movddup	-15 * SIZE(AA), %xmm0
	mulpd	 %xmm2, %xmm0
	subpd	 %xmm0, %xmm3
	movddup	-14 * SIZE(AA), %xmm0
	mulpd	 %xmm2, %xmm0
	subpd	 %xmm0, %xmm5
	movddup	-13 * SIZE(AA), %xmm0
	mulpd	 %xmm2, %xmm0
	subpd	 %xmm0, %xmm7

	movddup	-11 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm3
	movddup	-10 * SIZE(AA), %xmm0
	mulpd	 %xmm3, %xmm0
	subpd	 %xmm0, %xmm5
	movddup	 -9 * SIZE(AA), %xmm0
	mulpd	 %xmm3, %xmm0
	subpd	 %xmm0, %xmm7

	movddup	 -6 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm5
	movddup	 -5 * SIZE(AA), %xmm0
	mulpd	 %xmm5, %xmm0
	subpd	 %xmm0, %xmm7

	movddup	 -1 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm7
#endif

#ifdef RN
	movddup	-16 * SIZE(B), %xmm4
	mulpd	 %xmm4, %xmm0
	mulpd	 %xmm4, %xmm1

	movddup	-15 * SIZE(B), %xmm4
	movapd	 %xmm4, %xmm5
	mulpd	 %xmm0, %xmm4
	subpd	 %xmm4, %xmm2
	mulpd	 %xmm1, %xmm5
	subpd	 %xmm5, %xmm3

	movddup	-13 * SIZE(B), %xmm4
	mulpd	 %xmm4, %xmm2
	mulpd	 %xmm4, %xmm3
#endif

#ifdef RT
	movddup	-13 * SIZE(B), %xmm4
	mulpd	 %xmm4, %xmm2
	mulpd	 %xmm4, %xmm3

	movddup	-14 * SIZE(B), %xmm4
	movapd	 %xmm4, %xmm5
	mulpd	 %xmm2, %xmm4
	subpd	 %xmm4, %xmm0
	mulpd	 %xmm3, %xmm5
	subpd	 %xmm5, %xmm1

	movddup	-16 * SIZE(B), %xmm4
	mulpd	 %xmm4, %xmm0
	mulpd	 %xmm4, %xmm1
#endif

#ifdef LN
	subl	$4 * SIZE, CO1
#endif

#if defined(LN) || defined(LT)
	movsd	%xmm2,  0 * SIZE(CO1)
	movsd	%xmm3,  1 * SIZE(CO1)
	movsd	%xmm5,  2 * SIZE(CO1)
	movsd	%xmm7,  3 * SIZE(CO1)

	movhpd	%xmm2,  0 * SIZE(CO1, LDC)
	movhpd	%xmm3,  1 * SIZE(CO1, LDC)
	movhpd	%xmm5,  2 * SIZE(CO1, LDC)
	movhpd	%xmm7,  3 * SIZE(CO1, LDC)
#else
	movsd	%xmm0,  0 * SIZE(CO1)
	movhpd	%xmm0,  1 * SIZE(CO1)
	movsd	%xmm1,  2 * SIZE(CO1)
	movhpd	%xmm1,  3 * SIZE(CO1)

	movsd	%xmm2,  0 * SIZE(CO1, LDC)
	movhpd	%xmm2,  1 * SIZE(CO1, LDC)
	movsd	%xmm3,  2 * SIZE(CO1, LDC)
	movhpd	%xmm3,  3 * SIZE(CO1, LDC)
#endif

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

#if defined(LN) || defined(LT)
	movapd	%xmm2,  -16 * SIZE(B)
	movapd	%xmm3,  -14 * SIZE(B)
	movapd	%xmm5,  -12 * SIZE(B)
	movapd	%xmm7,  -10 * SIZE(B)

	movddup	%xmm2, %xmm0
	movddup	%xmm3, %xmm1
	movddup	%xmm5, %xmm4
	movddup	%xmm7, %xmm6

	unpckhpd %xmm2, %xmm2
	unpckhpd %xmm3, %xmm3
	unpckhpd %xmm5, %xmm5
	unpckhpd %xmm7, %xmm7

	movapd	%xmm0,  -16 * SIZE(BB)
	movapd	%xmm2,  -14 * SIZE(BB)
	movapd	%xmm1,  -12 * SIZE(BB)
	movapd	%xmm3,  -10 * SIZE(BB)
	movapd	%xmm4,   -8 * SIZE(BB)
	movapd	%xmm5,   -6 * SIZE(BB)
	movapd	%xmm6,   -4 * SIZE(BB)
	movapd	%xmm7,   -2 * SIZE(BB)
#else
	movapd	%xmm0,  -16 * SIZE(AA)
	movapd	%xmm1,  -14 * SIZE(AA)
	movapd	%xmm2,  -12 * SIZE(AA)
	movapd	%xmm3,  -10 * SIZE(AA)
#endif

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	leal	(,%eax, SIZE), %eax
	leal	(AA, %eax, 4), AA
#ifdef LT
	addl	$8 * SIZE, B
#endif
#endif

#ifdef LN
	subl	$4, KK
	movl	BORIG, B
#endif

#ifdef LT
	addl	$4, KK
#endif

#ifdef RT
	movl	K, %eax
	movl	BORIG, B
	sall	$2 + BASE_SHIFT, %eax
	addl	%eax, AORIG
#endif

	decl	%ebx			# i --
	jg	.L10
	ALIGN_2	

.L30:
	movl	M,  %ebx
	testl	$2, %ebx
	jle	.L50

#ifdef LN
       movl	K, %eax
       sall	$1 + BASE_SHIFT, %eax
       subl	%eax, AORIG
#endif

#if defined(LN) || defined(RT)
	movl	KK, %eax
	movl	AORIG, AA
	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 2), AA
#endif

	leal	16 * SIZE + BUFFER, BB

#if defined(LN) || defined(RT)
	movl	KK, %eax
	sall	$1 + BASE_SHIFT, %eax
	leal	(BB, %eax, 2), 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
	je	.L32

.L31:	
	mulpd	%xmm0,  %xmm1
	mulpd	-14 * SIZE(BB), %xmm0
	addpd	%xmm1, %xmm4
	movapd	-12 * SIZE(BB), %xmm1
	addpd	%xmm0, %xmm5
	movapd	-14 * SIZE(AA), %xmm0
	mulpd	%xmm0,  %xmm1
	mulpd	-10 * SIZE(BB), %xmm0
	addpd	%xmm1, %xmm6
	movapd	  0 * SIZE(BB), %xmm1
	addpd	%xmm0, %xmm7
	movapd	-12 * SIZE(AA), %xmm0
	mulpd	%xmm0,  %xmm3
	mulpd	 -6 * SIZE(BB), %xmm0
	addpd	%xmm3, %xmm4
	movapd	 -4 * SIZE(BB), %xmm3
	addpd	%xmm0, %xmm5
	movapd	-10 * SIZE(AA), %xmm0
	mulpd	%xmm0,  %xmm3
	mulpd	 -2 * SIZE(BB), %xmm0
	addpd	%xmm3, %xmm6
	movapd	  8 * SIZE(BB), %xmm3
	addpd	%xmm0, %xmm7
	movapd	  0 * SIZE(AA), %xmm0
	mulpd	%xmm2,  %xmm1
	mulpd	  2 * SIZE(BB), %xmm2
	addpd	%xmm1, %xmm4
	movapd	  4 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm5
	movapd	 -6 * SIZE(AA), %xmm2
	mulpd	%xmm2,  %xmm1
	mulpd	  6 * SIZE(BB), %xmm2
	addpd	%xmm1, %xmm6
	movapd	 16 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm7
	movapd	 -4 * SIZE(AA), %xmm2
	mulpd	%xmm2,  %xmm3
	mulpd	 10 * SIZE(BB), %xmm2
	addpd	%xmm3, %xmm4
	movapd	 12 * SIZE(BB), %xmm3
	addpd	%xmm2, %xmm5
	movapd	 -2 * SIZE(AA), %xmm2
	mulpd	%xmm2,  %xmm3
	mulpd	 14 * SIZE(BB), %xmm2
	addpd	%xmm3, %xmm6
	movapd	 24 * SIZE(BB), %xmm3
	addpd	%xmm2, %xmm7
	movapd	  8 * SIZE(AA), %xmm2

	subl   $-16 * SIZE, AA
	addl   $ 32 * SIZE, BB
	decl   %eax
	jne    .L31

.L32:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L34

.L33:
	mulpd	%xmm0,  %xmm1
	mulpd	-14 * SIZE(BB), %xmm0
	addpd	%xmm1, %xmm4
	movapd	-12 * SIZE(BB), %xmm1
	addpd	%xmm0, %xmm5
	movapd	-14 * SIZE(AA), %xmm0

	addl	$2 * SIZE, AA
	addl	$4 * SIZE, BB
	decl	%eax
	jg	.L33
	ALIGN_4

.L34:
	addpd	%xmm6, %xmm4
	addpd	%xmm7, %xmm5

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

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

	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 2), AA
	leal	(B,  %eax, 2), B
	leal	(BB, %eax, 4), BB
#endif

#if defined(LN) || defined(LT)
	movapd	 %xmm4, %xmm0
	unpcklpd %xmm5, %xmm4
	unpckhpd %xmm5, %xmm0

	movapd	-16 * SIZE(B), %xmm2
	movapd	-14 * SIZE(B), %xmm3

	subpd	%xmm4,  %xmm2
	subpd	%xmm0,  %xmm3
#else
	movapd	-16 * SIZE(AA), %xmm0
	movapd	-14 * SIZE(AA), %xmm1

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

#ifdef LN
	movddup	-13 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm3
	movddup	-14 * SIZE(AA), %xmm0
	mulpd	 %xmm3, %xmm0
	subpd	 %xmm0, %xmm2
	movddup	-16 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm2
#endif

#ifdef LT
	movddup	-16 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm2
	movddup	-15 * SIZE(AA), %xmm0
	mulpd	 %xmm2, %xmm0
	subpd	 %xmm0, %xmm3
	movddup	-13 * SIZE(AA), %xmm0
	mulpd	 %xmm0, %xmm3
#endif

#ifdef RN
	movddup	-16 * SIZE(B), %xmm4
	mulpd	 %xmm4, %xmm0
	movddup	-15 * SIZE(B), %xmm4
	mulpd	 %xmm0, %xmm4
	subpd	 %xmm4, %xmm1
	movddup	-13 * SIZE(B), %xmm4
	mulpd	 %xmm4, %xmm1
#endif

#ifdef RT
	movddup	-13 * SIZE(B), %xmm4
	mulpd	 %xmm4, %xmm1
	movddup	-14 * SIZE(B), %xmm4
	mulpd	 %xmm1, %xmm4
	subpd	 %xmm4, %xmm0
	movddup	-16 * SIZE(B), %xmm4
	mulpd	 %xmm4, %xmm0
#endif

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

#if defined(LN) || defined(LT)
	movsd	%xmm2,  0 * SIZE(CO1)
	movsd	%xmm3,  1 * SIZE(CO1)
	movhpd	%xmm2,  0 * SIZE(CO1, LDC)
	movhpd	%xmm3,  1 * SIZE(CO1, LDC)
#else
	movsd	%xmm0,  0 * SIZE(CO1)
	movhpd	%xmm0,  1 * SIZE(CO1)
	movsd	%xmm1,  0 * SIZE(CO1, LDC)
	movhpd	%xmm1,  1 * SIZE(CO1, LDC)
#endif

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

#if defined(LN) || defined(LT)
	movapd	%xmm2,  -16 * SIZE(B)
	movapd	%xmm3,  -14 * SIZE(B)

	movddup	%xmm2, %xmm0
	movddup	%xmm3, %xmm1

	unpckhpd %xmm2, %xmm2
	unpckhpd %xmm3, %xmm3

	movapd	%xmm0,  -16 * SIZE(BB)
	movapd	%xmm2,  -14 * SIZE(BB)
	movapd	%xmm1,  -12 * SIZE(BB)
	movapd	%xmm3,  -10 * SIZE(BB)
#else
	movapd	%xmm0,  -16 * SIZE(AA)
	movapd	%xmm1,  -14 * SIZE(AA)
#endif

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	leal	(,%eax, SIZE), %eax
	leal	(AA, %eax, 2), 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 + BASE_SHIFT, %eax
	addl	%eax, AORIG
#endif
	ALIGN_2

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

#ifdef LN
       movl	K, %eax
       sall	$0 + BASE_SHIFT, %eax
       subl	%eax, AORIG
#endif

#if defined(LN) || defined(RT)
	movl	KK, %eax
	movl	AORIG, AA
	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 1), AA
#endif

	leal	16 * SIZE + BUFFER, BB

#if defined(LN) || defined(RT)
	movl	KK, %eax
	sall	$1 + BASE_SHIFT, %eax
	leal	(BB, %eax, 2), BB
#endif	

	movsd	-16 * SIZE(AA), %xmm0
	pxor	%xmm4, %xmm4
	movsd	-16 * SIZE(BB), %xmm1
	pxor	%xmm5, %xmm5
	movsd	-12 * SIZE(AA), %xmm2
	pxor	%xmm6, %xmm6
	movsd	 -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
	je	.L52

.L51:	
	mulsd	%xmm0,  %xmm1
	mulsd	-14 * SIZE(BB), %xmm0
	addsd	%xmm1, %xmm4
 	movsd	-12 * SIZE(BB), %xmm1
	addsd	%xmm0, %xmm5
	movsd	-15 * SIZE(AA), %xmm0
	mulsd	%xmm0,  %xmm1
	mulsd	-10 * SIZE(BB), %xmm0
	addsd	%xmm1, %xmm6
	movsd	  0 * SIZE(BB), %xmm1
	addsd	%xmm0, %xmm7
	movsd	-14 * SIZE(AA), %xmm0
	mulsd	%xmm0,  %xmm3
	mulsd	 -6 * SIZE(BB), %xmm0
	addsd	%xmm3, %xmm4
	movsd	 -4 * SIZE(BB), %xmm3
	addsd	%xmm0, %xmm5
	movsd	-13 * SIZE(AA), %xmm0
	mulsd	%xmm0,  %xmm3
	mulsd	 -2 * SIZE(BB), %xmm0
	addsd	%xmm3, %xmm6
	movsd	  8 * SIZE(BB), %xmm3
	addsd	%xmm0, %xmm7
	movsd	 -8 * SIZE(AA), %xmm0
	mulsd	%xmm2,  %xmm1
	mulsd	  2 * SIZE(BB), %xmm2
	addsd	%xmm1, %xmm4
	movsd	  4 * SIZE(BB), %xmm1
	addsd	%xmm2, %xmm5
	movsd	-11 * SIZE(AA), %xmm2
	mulsd	%xmm2,  %xmm1
	mulsd	  6 * SIZE(BB), %xmm2
	addsd	%xmm1, %xmm6
	movsd	 16 * SIZE(BB), %xmm1
	addsd	%xmm2, %xmm7
	movsd	-10 * SIZE(AA), %xmm2
	mulsd	%xmm2,  %xmm3
	mulsd	 10 * SIZE(BB), %xmm2
	addsd	%xmm3, %xmm4
	movsd	 12 * SIZE(BB), %xmm3
	addsd	%xmm2, %xmm5
	movsd	 -9 * SIZE(AA), %xmm2
	mulsd	%xmm2,  %xmm3
	mulsd	 14 * SIZE(BB), %xmm2
	addsd	%xmm3, %xmm6
	movsd	 24 * SIZE(BB), %xmm3
	addsd	%xmm2, %xmm7
	movsd	 -4 * SIZE(AA), %xmm2

	subl   $-8 * SIZE, AA
	addl   $32 * SIZE, BB
	decl   %eax
	jne    .L51

.L52:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L54

.L53:
	mulsd	%xmm0,  %xmm1
	mulsd	-14 * SIZE(BB), %xmm0
	addsd	%xmm1, %xmm4
	movsd	-12 * SIZE(BB), %xmm1
	addsd	%xmm0, %xmm5
	movsd	-15 * SIZE(AA), %xmm0

	addl	$1 * SIZE, AA
	addl	$4 * SIZE, BB
	decl	%eax
	BRANCH
	jg	.L53
	ALIGN_4

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

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

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

	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 1), AA
	leal	(B,  %eax, 2), B
	leal	(BB, %eax, 4), BB
#endif

#if defined(LN) || defined(LT)
	movsd	-16 * SIZE(B), %xmm0
	movsd	-15 * SIZE(B), %xmm1
#else
	movsd	-16 * SIZE(AA), %xmm0
	movsd	-15 * SIZE(AA), %xmm1
#endif

	subsd	%xmm4,  %xmm0
	subsd	%xmm5,  %xmm1

#if defined(LN) || defined(LT)
	movsd	-16 * SIZE(AA), %xmm2
	mulsd	 %xmm2, %xmm0
	mulsd	 %xmm2, %xmm1
#endif

#ifdef RN
	mulsd	-16 * SIZE(B), %xmm0
	movsd	-15 * SIZE(B), %xmm4
	mulsd	 %xmm0, %xmm4
	subsd	 %xmm4, %xmm1
	mulsd	-13 * SIZE(B), %xmm1
#endif

#ifdef RT
	mulsd	-13 * SIZE(B), %xmm1
	movsd	-14 * SIZE(B), %xmm4
	mulsd	 %xmm1, %xmm4
	subsd	 %xmm4, %xmm0

	mulsd	-16 * SIZE(B), %xmm0
#endif

#ifdef LN
	subl	$1 * SIZE, CO1
#endif

	movsd	%xmm0,  0 * SIZE(CO1)
	movsd	%xmm1,  0 * SIZE(CO1, LDC)

#ifndef LN
	addl	$1 * SIZE, CO1
#endif

#if defined(LN) || defined(LT)
	movsd	%xmm0,  -16 * SIZE(B)
	movsd	%xmm1,  -15 * SIZE(B)

	movsd	%xmm0,  -16 * SIZE(BB)
	movsd	%xmm0,  -15 * SIZE(BB)
	movsd	%xmm1,  -14 * SIZE(BB)
	movsd	%xmm1,  -13 * SIZE(BB)
#else
	movsd	%xmm0,  -16 * SIZE(AA)
	movsd	%xmm1,  -15 * SIZE(AA)
#endif

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	leal	(,%eax, SIZE), %eax
	leal	(AA, %eax, 1), 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	$0 + BASE_SHIFT, %eax
	addl	%eax, AORIG
#endif
	ALIGN_2

.L99:
#ifdef LN
       movl	K, %eax
       leal	(, %eax, SIZE), %eax
       leal 	(B, %eax, 2), B
#endif

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	leal	(,%eax, SIZE), %eax
	leal	(B,  %eax, 2), B
#endif

#ifdef RN
	addl	$2, KK
#endif

#ifdef RT
	subl	$2, KK
#endif

	decl	J			# j --
	jg	.L01
	ALIGN_2

.L999:
	movl	OLD_STACK, %esp

	EMMS

	popl	%ebx
	popl	%esi
	popl	%edi
	popl	%ebp
	ret
	ALIGN_2


	EPILOGUE