trsm_kernel_lt_4x4_sse.s

Optimized GotoBLAS libraries

       sall	$2 + 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, 4), AA
#endif

	leal	BUFFER, BB

#if defined(LN) || defined(RT)
	movl	KK, %eax
	sall	$BASE_SHIFT, %eax
	leal	(BB, %eax, 4), BB
#endif	

	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5
	pxor	%xmm6, %xmm6
	pxor	%xmm7, %xmm7

	movaps	 0 * SIZE(AA), %xmm0
	movaps	16 * SIZE(AA), %xmm1
	movaps	 0 * SIZE(BB), %xmm2
	movaps	16 * SIZE(BB), %xmm3

	PREFETCHW	3 * SIZE(CO1)

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	sarl	$3, %eax
	je	.L95
	ALIGN_4

.L92:
	mulps	%xmm0, %xmm2
#if defined(OPTERON) || defined(BARCELONA)
	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
#endif
	movaps	 4 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm4
	movaps	32 * SIZE(BB), %xmm2
	mulps	 4 * SIZE(BB), %xmm0
	addps	%xmm0, %xmm5
	movaps	 8 * SIZE(AA), %xmm0
	mulps	 8 * SIZE(BB), %xmm0
	addps	%xmm0, %xmm6
	movaps	12 * SIZE(AA), %xmm0
	mulps	12 * SIZE(BB), %xmm0
	addps	%xmm0, %xmm7
	movaps	32 * SIZE(AA), %xmm0
#if defined(OPTERON) || defined(BARCELONA)
	prefetcht0 (PREFETCHSIZE  + 16) * SIZE(AA)
#endif
	mulps	%xmm1, %xmm3
	movaps	20 * SIZE(AA), %xmm1
	addps	%xmm3, %xmm4
	movaps	48 * SIZE(BB), %xmm3
	mulps	20 * SIZE(BB), %xmm1
	addps	%xmm1, %xmm5
	movaps	24 * SIZE(AA), %xmm1
	mulps	24 * SIZE(BB), %xmm1
	addps	%xmm1, %xmm6
	movaps	28 * SIZE(AA), %xmm1
	mulps	28 * SIZE(BB), %xmm1
	addps	%xmm1, %xmm7
	movaps	48 * SIZE(AA), %xmm1

	addl	$32 * SIZE, AA
	addl	$32 * SIZE, BB
	decl   %eax
	jne    .L92
	ALIGN_4

.L95:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L98
	ALIGN_4

.L96:
	mulps	%xmm0, %xmm2
	addps	%xmm2, %xmm4
	movaps	 4 * SIZE(AA), %xmm0
	movaps	 4 * SIZE(BB), %xmm2

	addl	$4 * SIZE, AA
	addl	$4 * SIZE, BB
	decl	%eax
	jg	.L96
	ALIGN_4

.L98:
	addps	%xmm5, %xmm4
	addps	%xmm7, %xmm6
	addps	%xmm6, %xmm4

#if defined(LN) || defined(RT)
	movl	KK, %eax
#ifdef LN
	subl	$4, %eax
#else
	subl	$1, %eax
#endif

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

	sall	$ BASE_SHIFT, %eax
	leal	(AA, %eax, 4), AA
	leal	(B,  %eax, 1), B
	leal	(BB, %eax, 4), BB
#endif

#if defined(LN) || defined(LT)
	movaps	 %xmm4, %xmm0
	unpcklps %xmm6, %xmm4
	unpckhps %xmm6, %xmm0

	movaps	 %xmm5, %xmm1
	unpcklps %xmm7, %xmm5
	unpckhps %xmm7, %xmm1

	movaps	 %xmm4, %xmm6
	unpcklps %xmm5, %xmm4
	unpckhps %xmm5, %xmm6

	movaps	 %xmm0, %xmm2
	unpcklps %xmm1, %xmm0
	unpckhps %xmm1, %xmm2

	movss	 0 * SIZE(B), %xmm1
	movss	 1 * SIZE(B), %xmm3
	movss	 2 * SIZE(B), %xmm5
	movss	 3 * SIZE(B), %xmm7

	subss	%xmm4,  %xmm1
	subss	%xmm6,  %xmm3
	subss	%xmm0,  %xmm5
	subss	%xmm2,  %xmm7
#else
	movaps	 0 * SIZE(AA), %xmm0

	subps	%xmm4, %xmm0
#endif

#ifdef LN
	movaps	 12 * SIZE(AA), %xmm4
	pshufd	 $0xff, %xmm4, %xmm6
	mulss	 %xmm6, %xmm7
	pshufd	 $0xaa, %xmm4, %xmm6
	mulss	 %xmm7, %xmm6
	subss	 %xmm6, %xmm5
	pshufd	 $0x55, %xmm4, %xmm6
	mulss	 %xmm7, %xmm6
	subss	 %xmm6, %xmm3
	pshufd	 $0x00, %xmm4, %xmm6
	mulss	 %xmm7, %xmm6
	subss	 %xmm6, %xmm1

	movaps	  8 * SIZE(AA), %xmm4
	pshufd	 $0xaa, %xmm4, %xmm6
	mulss	 %xmm6, %xmm5
	pshufd	 $0x55, %xmm4, %xmm6
	mulss	 %xmm5, %xmm6
	subss	 %xmm6, %xmm3
	pshufd	 $0x00, %xmm4, %xmm6
	mulss	 %xmm5, %xmm6
	subss	 %xmm6, %xmm1

	movaps	  4 * SIZE(AA), %xmm4
	pshufd	 $0x55, %xmm4, %xmm6
	mulss	 %xmm6, %xmm3
	pshufd	 $0x00, %xmm4, %xmm6
	mulss	 %xmm3, %xmm6
	subss	 %xmm6, %xmm1

	movaps	  0 * SIZE(AA), %xmm4
	pshufd	 $0x00, %xmm4, %xmm6
	mulss	 %xmm6, %xmm1
#endif

#ifdef LT
	movaps	 0 * SIZE(AA), %xmm4
	pshufd	 $0x00, %xmm4, %xmm6
	mulss	 %xmm6, %xmm1

	pshufd	 $0x55, %xmm4, %xmm6
	mulss	 %xmm1, %xmm6
	subss	 %xmm6, %xmm3
	pshufd	 $0xaa, %xmm4, %xmm6
	mulss	 %xmm1, %xmm6
	subss	 %xmm6, %xmm5
	pshufd	 $0xff, %xmm4, %xmm6
	mulss	 %xmm1, %xmm6
	subss	 %xmm6, %xmm7

	movaps	 4 * SIZE(AA), %xmm4
	pshufd	 $0x55, %xmm4, %xmm6
	mulss	 %xmm6, %xmm3
	pshufd	 $0xaa, %xmm4, %xmm6
	mulss	 %xmm3, %xmm6
	subss	 %xmm6, %xmm5
	pshufd	 $0xff, %xmm4, %xmm6
	mulss	 %xmm3, %xmm6
	subss	 %xmm6, %xmm7

	movaps	 8 * SIZE(AA), %xmm4
	pshufd	 $0xaa, %xmm4, %xmm6
	mulss	 %xmm6, %xmm5
	pshufd	 $0xff, %xmm4, %xmm6
	mulss	 %xmm5, %xmm6
	subss	 %xmm6, %xmm7

	movaps	12 * SIZE(AA), %xmm4
	pshufd	 $0xff, %xmm4, %xmm6
	mulss	 %xmm6, %xmm7
#endif

#if defined(RN) || defined(RT)
	movss	 0 * SIZE(B), %xmm6
	pshufd	 $0x00, %xmm6, %xmm7
	mulps	 %xmm7, %xmm0
#endif

#if defined(LN) || defined(LT)
	movss	%xmm1,   0 * SIZE(B)
	movss	%xmm3,   1 * SIZE(B)
	movss	%xmm5,   2 * SIZE(B)
	movss	%xmm7,   3 * SIZE(B)

	pshufd	$0x00, %xmm1, %xmm0
	movaps	%xmm0,   0 * SIZE(BB)
	pshufd	$0x00, %xmm3, %xmm0
	movaps	%xmm0,   4 * SIZE(BB)

	pshufd	$0x00, %xmm5, %xmm0
	movaps	%xmm0,   8 * SIZE(BB)
	pshufd	$0x00, %xmm7, %xmm0
	movaps	%xmm0,  12 * SIZE(BB)
#else
	movss	%xmm0,   0 * SIZE(AA)
	movss	%xmm1,   1 * SIZE(AA)
	movss	%xmm2,   2 * SIZE(AA)
	movss	%xmm3,   3 * SIZE(AA)
#endif

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

#if defined(LN) || defined(LT)
	unpcklps %xmm5, %xmm1
	unpcklps %xmm7, %xmm3

	unpcklps %xmm3, %xmm1

	movlps	%xmm1, 0 * SIZE(CO1)
	movhps	%xmm1, 2 * SIZE(CO1)
#else
	movlps	%xmm0, 0 * SIZE(CO1)
	movhps	%xmm0, 2 * SIZE(CO1)
#endif

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

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	leal	(,%eax, SIZE), %eax
	leal	(AA, %eax, 4), AA
#ifdef LT
	addl	$4 * 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	.L91
	ALIGN_4

.L100:
	testl	$2, M
	je	.L110

#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	BUFFER, BB

#if defined(LN) || defined(RT)
	movl	KK, %eax
	sall	$BASE_SHIFT, %eax
	leal	(BB, %eax, 4), BB
#endif	

	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5
	pxor	%xmm6, %xmm6
	pxor	%xmm7, %xmm7

	movsd	 0 * SIZE(AA), %xmm0
	movsd	 8 * SIZE(AA), %xmm1
	movsd	 0 * SIZE(BB), %xmm2
	movsd	16 * SIZE(BB), %xmm3

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	sarl	$3, %eax
	je	.L105
	ALIGN_4

.L102:
	mulps	%xmm0, %xmm2
#if defined(OPTERON) || defined(BARCELONA)
	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
#endif
	movsd	 2 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm4
	movsd	 4 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	movsd	 4 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5
	movsd	 8 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	movsd	 6 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm6
	movsd	12 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	movsd	16 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm7
	movsd	32 * SIZE(BB), %xmm2
	mulps	%xmm1, %xmm3
	movsd	10 * SIZE(AA), %xmm1
	addps	%xmm3, %xmm4
	movsd	20 * SIZE(BB), %xmm3
	mulps	%xmm1, %xmm3
	movsd	12 * SIZE(AA), %xmm1
	addps	%xmm3, %xmm5
	movsd	24 * SIZE(BB), %xmm3
	mulps	%xmm1, %xmm3
	movsd	14 * SIZE(AA), %xmm1
	addps	%xmm3, %xmm6
	movsd	28 * SIZE(BB), %xmm3
	mulps	%xmm1, %xmm3
	movsd	24 * SIZE(AA), %xmm1
	addps	%xmm3, %xmm7
	movsd	48 * SIZE(BB), %xmm3

	addl	$16 * SIZE, AA
	addl	$32 * SIZE, BB
	decl   %eax
	jne    .L102
	ALIGN_4

.L105:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L108
	ALIGN_4

.L106:
	mulps	%xmm0, %xmm2
	addps	%xmm2, %xmm4
	movsd	 2 * SIZE(AA), %xmm0
	movsd	 4 * SIZE(BB), %xmm2

	addl	$2 * SIZE, AA
	addl	$4 * SIZE, BB
	decl	%eax
	jg	.L106
	ALIGN_4

.L108:
	addps	%xmm5, %xmm4
	addps	%xmm7, %xmm6
	addps	%xmm6, %xmm4

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

#if defined(LN) || defined(LT)
	pshufd	$1, %xmm4, %xmm6

	movss	 0 * SIZE(B), %xmm1
	movss	 1 * SIZE(B), %xmm3

	subss	%xmm4,  %xmm1
	subss	%xmm6,  %xmm3
#else
	movsd	 0 * SIZE(AA), %xmm0

	subps	%xmm4, %xmm0
#endif

#ifdef LN
	movaps	  0 * SIZE(AA), %xmm4
	pshufd	 $0xff, %xmm4, %xmm6
	mulss	 %xmm6, %xmm3
	pshufd	 $0xaa, %xmm4, %xmm6
	mulss	 %xmm3, %xmm6
	subss	 %xmm6, %xmm1

	pshufd	 $0x00, %xmm4, %xmm6
	mulss	 %xmm6, %xmm1
#endif

#ifdef LT
	movaps	 0 * SIZE(AA), %xmm4
	pshufd	 $0x00, %xmm4, %xmm6
	mulss	 %xmm6, %xmm1
	pshufd	 $0x55, %xmm4, %xmm6
	mulss	 %xmm1, %xmm6
	subss	 %xmm6, %xmm3

	pshufd	 $0xff, %xmm4, %xmm6
	mulss	 %xmm6, %xmm3
#endif

#if defined(RN) || defined(RT)
	movss	 0 * SIZE(B), %xmm6
	pshufd	 $0x00, %xmm6, %xmm7
	mulps	 %xmm7, %xmm0
#endif

#if defined(LN) || defined(LT)
	movss	%xmm1,   0 * SIZE(B)
	movss	%xmm3,   1 * SIZE(B)

	pshufd	$0x00, %xmm1, %xmm0
	movaps	%xmm0,   0 * SIZE(BB)
	pshufd	$0x00, %xmm3, %xmm0
	movaps	%xmm0,   4 * SIZE(BB)
#else
	movlps	%xmm0,   0 * SIZE(AA)
#endif

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

#if defined(LN) || defined(LT)
	movss	%xmm1, 0 * SIZE(CO1)
	movss	%xmm3, 1 * SIZE(CO1)
#else
	movlps	%xmm0, 0 * SIZE(CO1)
#endif

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

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	leal	(,%eax, SIZE), %eax
	leal	(AA, %eax, 2), AA
#ifdef LT
	addl	$2 * 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_4

.L110:
	testl	$1, M
	je	.L119

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

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

	leal	BUFFER, BB

#if defined(LN) || defined(RT)
	movl	KK, %eax
	sall	$BASE_SHIFT, %eax
	leal	(BB, %eax, 4), BB
#endif	

	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5
	pxor	%xmm6, %xmm6
	pxor	%xmm7, %xmm7

	movss	 0 * SIZE(AA), %xmm0
	movss	 4 * SIZE(AA), %xmm1
	movss	 0 * SIZE(BB), %xmm2
	movss	16 * SIZE(BB), %xmm3

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	sarl	$3, %eax
	je	.L115
	ALIGN_4

.L112:
	mulss	%xmm0, %xmm2
#if defined(OPTERON) || defined(BARCELONA)
	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)
#endif
	movss	 1 * SIZE(AA), %xmm0
	addss	%xmm2, %xmm4
	movss	32 * SIZE(BB), %xmm2
	mulss	 4 * SIZE(BB), %xmm0
	addss	%xmm0, %xmm5
	movss	 2 * SIZE(AA), %xmm0
	mulss	 8 * SIZE(BB), %xmm0
	addss	%xmm0, %xmm6
	movss	 3 * SIZE(AA), %xmm0
	mulss	12 * SIZE(BB), %xmm0
	addss	%xmm0, %xmm7
	movss	 8 * SIZE(AA), %xmm0
	mulss	%xmm1, %xmm3
	movss	 5 * SIZE(AA), %xmm1
	addss	%xmm3, %xmm4
	movss	48 * SIZE(BB), %xmm3
	mulss	20 * SIZE(BB), %xmm1
	addss	%xmm1, %xmm5
	movss	 6 * SIZE(AA), %xmm1
	mulss	24 * SIZE(BB), %xmm1
	addss	%xmm1, %xmm6
	movss	 7 * SIZE(AA), %xmm1
	mulss	28 * SIZE(BB), %xmm1
	addss	%xmm1, %xmm7
	movss	12 * SIZE(AA), %xmm1

	addl	$ 8 * SIZE, AA
	addl	$32 * SIZE, BB
	decl   %eax
	jne    .L112
	ALIGN_4

.L115:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L118
	ALIGN_4

.L116:
	mulss	%xmm0, %xmm2
	movss	 1 * SIZE(AA), %xmm0
	addss	%xmm2, %xmm4
	movss	 4 * SIZE(BB), %xmm2

	addl	$ 1 * SIZE, AA
	addl	$ 4 * SIZE, BB
	decl	%eax
	jg	.L116
	ALIGN_4

.L118:
	addss	%xmm5, %xmm4
	addss	%xmm7, %xmm6
	addss	%xmm6, %xmm4

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

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

	sall	$ BASE_SHIFT, %eax
	leal	(AA, %eax, 1), AA
	leal	(B,  %eax, 1), B
	leal	(BB, %eax, 4), BB
#endif

#if defined(LN) || defined(LT)
	movss	 0 * SIZE(B), %xmm1
	subss	%xmm4,  %xmm1
#else
	movss	 0 * SIZE(AA), %xmm0
	subss	%xmm4, %xmm0
#endif

#if defined(LN) || defined(LT)
	mulss	 0 * SIZE(AA), %xmm1
#endif

#if defined(RN) || defined(RT)
	mulss	 0 * SIZE(B), %xmm0
#endif

#if defined(LN) || defined(LT)
	movss	%xmm1,   0 * SIZE(B)

	pshufd	$0x00, %xmm1, %xmm0
	movaps	%xmm0,   0 * SIZE(BB)
#else
	movss	%xmm0,   0 * SIZE(AA)
#endif

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

#if defined(LN) || defined(LT)
	movss	%xmm1, 0 * SIZE(CO1)
#else
	movss	%xmm0, 0 * SIZE(CO1)
#endif

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

#if