trsm_kernel_ln_4x4_penryn.s

Optimized GotoBLAS libraries

	je	.L999

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

#ifdef RT
	movl	K, %eax
	sall	$BASE_SHIFT, %eax
	subl	%eax, B
#endif

#ifdef RT
	subl	LDC, C
#endif
	movl	C, CO1
#ifndef RT
	addl	LDC, C
#endif

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

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

	testl	$1, M
	je	.L100

#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

	movl	B, BB

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

	pxor	%xmm4, %xmm4
	movsd	-32 * SIZE(AA), %xmm0
	pxor	%xmm5, %xmm5
	movsd	-32 * SIZE(BB), %xmm1

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

.L112:
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AA)

	mulps	%xmm0, %xmm1
	movsd	-30 * SIZE(AA), %xmm0
	addps	%xmm1, %xmm4
	movsd	-30 * SIZE(BB), %xmm1

	mulps	%xmm0, %xmm1
	movsd	-28 * SIZE(AA), %xmm0
	addps	%xmm1, %xmm4
	movsd	-28 * SIZE(BB), %xmm1

	mulps	%xmm0, %xmm1
	movsd	-26 * SIZE(AA), %xmm0
	addps	%xmm1, %xmm4
	movsd	-26 * SIZE(BB), %xmm1

	mulps	%xmm0, %xmm1
	movsd	-24 * SIZE(AA), %xmm0
	addps	%xmm1, %xmm4
	movsd	-24 * SIZE(BB), %xmm1

	subl   $-8 * SIZE, AA
	subl   $-8 * SIZE, BB

	subl   $1, %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, %xmm1
	movss	-31 * SIZE(AA), %xmm0
	addss	%xmm1, %xmm4
	movss	-31 * SIZE(BB), %xmm1

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

.L118:
#if defined(LN) || defined(RT)
	movl	KK, %eax
	subl	$1, %eax

	movl	AORIG, AA

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

	haddps	%xmm4, %xmm4

#if defined(LN) || defined(LT)
	movss	-32 * SIZE(BB), %xmm1
	subss	%xmm4,  %xmm1
#else
	movss	-32 * SIZE(AA), %xmm0
	subss	%xmm4, %xmm0
#endif

#if defined(LN) || defined(LT)
	mulss	-32 * SIZE(AA), %xmm1
#endif

#if defined(RN) || defined(RT)
	mulss	-32 * SIZE(BB), %xmm0
#endif

#if defined(LN) || defined(LT)
	movss	%xmm1, -32 * SIZE(BB)
#else
	movss	%xmm0, -32 * 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 defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	leal	(AA, %eax, SIZE), AA
	leal	(BB, %eax, SIZE), BB
#endif

#ifdef LN
	subl	$1, KK
#endif

#ifdef LT
	addl	$1, KK
#endif

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

	movl	B, BB

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

	movsd	-32 * SIZE(AA), %xmm0
	pxor	%xmm3, %xmm3
	movsd	-32 * SIZE(BB), %xmm1
	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5

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

.L102:
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AA)

	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movsd	-30 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm4

	pshufd	$0x55, %xmm1, %xmm2
	movsd	-30 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movsd	-28 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5

	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movsd	-26 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm4

	pshufd	$0x55, %xmm1, %xmm2
	movsd	-28 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movsd	-24 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5

	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movsd	-22 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm4

	pshufd	$0x55, %xmm1, %xmm2
	movsd	-26 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movsd	-20 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5

	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movsd	-18 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm4

	pshufd	$0x55, %xmm1, %xmm2
	movsd	-24 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movsd	-16 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm5

	subl   $-16 * SIZE, AA
	subl   $ -8 * SIZE, BB

	subl   $1, %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:
	pshufd	$0x00, %xmm1, %xmm2
	movss	-31 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movsd	-30 * SIZE(AA), %xmm0
	addps	%xmm2, %xmm4

	addl	$2 * SIZE, AA
	addl	$1 * SIZE, BB
	decl	%eax
	jg	.L106
	ALIGN_4

.L108:
#if defined(LN) || defined(RT)
	movl	KK, %eax
#ifdef LN
	subl	$2, %eax
#else
	subl	$1, %eax
#endif

	movl	AORIG, AA

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

	addps	%xmm5, %xmm4

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

	movss	-32 * SIZE(BB), %xmm1
	movss	-31 * SIZE(BB), %xmm3

	subss	%xmm4,  %xmm1
	subss	%xmm6,  %xmm3
#else
	movsd	-32 * SIZE(AA), %xmm0

	subps	%xmm4, %xmm0
#endif

#ifdef LN
	movsd	-32 * SIZE(AA), %xmm4
	movhps	-30 * 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	-32 * 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	-32 * SIZE(BB), %xmm6
	pshufd	 $0x00, %xmm6, %xmm7
	mulps	 %xmm7, %xmm0
#endif

#if defined(LN) || defined(LT)
	movss	%xmm1, -32 * SIZE(BB)
	movss	%xmm3, -31 * SIZE(BB)
#else
	movlps	%xmm0, -32 * 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
	leal	(BB, %eax, 1), BB
#endif

#ifdef LN
	subl	$2, KK
#endif

#ifdef LT
	addl	$2, KK
#endif

#ifdef RT
	movl	K, %eax
	sall	$1 + BASE_SHIFT, %eax
	addl	%eax, AORIG
#endif
	ALIGN_4

.L110:
	movl	M,  %ebx
	sarl	$2, %ebx	# i = (m >> 2)
	jle	.L119
	ALIGN_4

.L91:
#ifdef LN
       movl	K, %eax
       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

	movl	B, BB

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

	movaps	-32 * SIZE(AA), %xmm0
	pxor	%xmm2, %xmm2
	movsd	-32 * SIZE(BB), %xmm1

	pxor	%xmm4, %xmm4
#ifdef LN
	prefetcht0	-4 * SIZE(CO1)
#else
	prefetcht0	 3 * SIZE(CO1)
#endif
	pxor	%xmm5, %xmm5

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

.L92:
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AA)

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movaps	-28 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm5
	pshufd	$0x55, %xmm1, %xmm2
	movsd	-30 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movaps	-24 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movaps	-20 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm5
	pshufd	$0x55, %xmm1, %xmm2
	movsd	-28 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movaps	-16 * SIZE(AA), %xmm0

	PREFETCH (PREFETCHSIZE + 16) * SIZE(AA)

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movaps	-12 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm5
	pshufd	$0x55, %xmm1, %xmm2
	movsd	-26 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movaps	 -8 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	mulps	%xmm0, %xmm2
	movaps	 -4 * SIZE(AA), %xmm0

	addps	%xmm2, %xmm5
	pshufd	$0x55, %xmm1, %xmm2
	movsd	-24 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movaps	  0 * SIZE(AA), %xmm0

	subl   $-32 * SIZE, AA
	subl   $ -8 * SIZE, BB

	subl   $1, %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:
	addps	%xmm2, %xmm4
	pshufd	$0x00, %xmm1, %xmm2
	movss	-31 * SIZE(BB), %xmm1
	mulps	%xmm0, %xmm2
	movaps	-28 * SIZE(AA), %xmm0

	addl	$4 * SIZE, AA
	addl	$1 * SIZE, BB
	decl	%eax
	jg	.L96
	ALIGN_4

.L98:
#if defined(LN) || defined(RT)
	movl	KK, %eax
#ifdef LN
	subl	$4, %eax
#else
	subl	$1, %eax
#endif

	movl	AORIG, AA

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

	addps	%xmm2, %xmm4
	addps	%xmm5, %xmm4

#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	-32 * SIZE(BB), %xmm1
	movss	-31 * SIZE(BB), %xmm3
	movss	-30 * SIZE(BB), %xmm5
	movss	-29 * SIZE(BB), %xmm7

	subss	%xmm4,  %xmm1
	subss	%xmm6,  %xmm3
	subss	%xmm0,  %xmm5
	subss	%xmm2,  %xmm7
#else
	movaps	-32 * SIZE(AA), %xmm0

	subps	%xmm4, %xmm0
#endif

#ifdef LN
	movaps	-20 * 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	 -24 * 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	-28 * SIZE(AA), %xmm4
	pshufd	 $0x55, %xmm4, %xmm6
	mulss	 %xmm6, %xmm3
	pshufd	 $0x00, %xmm4, %xmm6
	mulss	 %xmm3, %xmm6
	subss	 %xmm6, %xmm1

	movaps	-32 * SIZE(AA), %xmm4
	pshufd	 $0x00, %xmm4, %xmm6
	mulss	 %xmm6, %xmm1
#endif

#ifdef LT
	movaps	-32 * 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	-28 * 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	-24 * SIZE(AA), %xmm4
	pshufd	 $0xaa, %xmm4, %xmm6
	mulss	 %xmm6, %xmm5
	pshufd	 $0xff, %xmm4, %xmm6
	mulss	 %xmm5, %xmm6
	subss	 %xmm6, %xmm7

	movaps	-20 * SIZE(AA), %xmm4
	pshufd	 $0xff, %xmm4, %xmm6
	mulss	 %xmm6, %xmm7
#endif

#if defined(RN) || defined(RT)
	movss	-32 * SIZE(BB), %xmm6
	pshufd	 $0x00, %xmm6, %xmm7
	mulps	 %xmm7, %xmm0
#endif

#if defined(LN) || defined(LT)
	movss	%xmm1, -32 * SIZE(BB)
	movss	%xmm3, -31 * SIZE(BB)
	movss	%xmm5, -30 * SIZE(BB)
	movss	%xmm7, -29 * SIZE(BB)
#else
	movaps	%xmm0, -32 * 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
	leal	(BB, %eax, 1), BB
#endif

#ifdef LN
	subl	$4, KK
#endif

#ifdef LT
	addl	$4, KK
#endif

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

	decl	%ebx			# i --
	jg	.L91
	ALIGN_4

.L119:
#ifdef LN
       movl	K, %eax
       leal 	(B, %eax, SIZE), 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


.L999:
	popl	%ebx
	popl	%esi
	popl	%edi
	popl	%ebp

	addl	$ARGS, %esp
	ret

	EPILOGUE