trsm_kernel_ln_8x4_sse.s

Optimized GotoBLAS libraries

	ALIGN_4
	
.L60:
#if defined(LT) || defined(RN)
	movq	A, AO
#else
	movq	A, AORIG
#endif

#ifdef RT
       leaq	(, LDC, 2), %rax
       subq	%rax, C
#endif

	movq	C, CO1			# coffset1 = c
#ifndef RT
	leaq	(C, LDC, 2), C
#endif

	testq	$1, M
	je	.L70

#ifdef LN
       movq	K, %rax
       salq	$BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	movq	AORIG, AO
	leaq	(AO, %rax, SIZE), AO
#endif

	leaq	BUFFER, BO

#if defined(LN) || defined(RT)
	movq	KK, %rax
	salq	$1 + BASE_SHIFT, %rax
	leaq	(BO, %rax, 4), BO
#endif	

	movss	 0 * SIZE(AO), %xmm8
	movss	 4 * SIZE(AO), %xmm10

	movss	 0 * SIZE(BO), %xmm9
	movss	16 * SIZE(BO), %xmm11
	movss	32 * SIZE(BO), %xmm13
	movss	48 * SIZE(BO), %xmm15

	pxor	%xmm0, %xmm0
	pxor	%xmm1, %xmm1
	pxor	%xmm2, %xmm2
	pxor	%xmm3, %xmm3

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$3, %rax
	je	.L95
	ALIGN_4

.L92:
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm0
#if defined(OPTERON) && defined(HAVE_PREFETCH)
	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)
#endif
	movss	 4 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	movss	 1 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm1
	movss	 8 * SIZE(BO), %xmm9

	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm2
	movss	12 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	movss	 2 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm3
	movss	64 * SIZE(BO), %xmm9

	mulps	%xmm8, %xmm11
	addps	%xmm11, %xmm0
	movss	20 * SIZE(BO), %xmm11
	mulps	%xmm8, %xmm11
	movss	 3 * SIZE(AO), %xmm8
	addps	%xmm11, %xmm1
	movss	24 * SIZE(BO), %xmm11

	mulps	%xmm8, %xmm11
	addps	%xmm11, %xmm2
	movss	28 * SIZE(BO), %xmm11
	mulps	%xmm8, %xmm11
	movss	 8 * SIZE(AO), %xmm8
	addps	%xmm11, %xmm3
	movss	 80 * SIZE(BO), %xmm11

	mulps	%xmm10, %xmm13
	addps	%xmm13, %xmm0
	movss	36 * SIZE(BO), %xmm13
	mulps	%xmm10, %xmm13
	movss	 5 * SIZE(AO), %xmm10
	addps	%xmm13, %xmm1
	movss	40 * SIZE(BO), %xmm13

	mulps	%xmm10, %xmm13
	addps	%xmm13, %xmm2
	movss	44 * SIZE(BO), %xmm13
	mulps	%xmm10, %xmm13
	movss	 6 * SIZE(AO), %xmm10
	addps	%xmm13, %xmm3
	movss	 96 * SIZE(BO), %xmm13

	mulps	%xmm10, %xmm15
	addps	%xmm15, %xmm0
	movss	52 * SIZE(BO), %xmm15
	mulps	%xmm10, %xmm15
	movss	 7 * SIZE(AO), %xmm10
	addps	%xmm15, %xmm1
	movss	56 * SIZE(BO), %xmm15

	mulps	%xmm10, %xmm15
	addps	%xmm15, %xmm2
	movss	60 * SIZE(BO), %xmm15
	mulps	%xmm10, %xmm15
	movss	12 * SIZE(AO), %xmm10
	addps	%xmm15, %xmm3
	movss	112 * SIZE(BO), %xmm15

	addq   $ 8 * SIZE, AO
	addq   $64 * SIZE, BO
	decq   %rax
	jne    .L92
	ALIGN_4

.L95:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L98
	ALIGN_4

.L96:
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm0
	movss	 4 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	movss	 1 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm1
	movss	 8 * SIZE(BO), %xmm9

	addq	$1 * SIZE, AO		# aoffset  += 4
	addq	$8 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L96
	ALIGN_4

.L98:
	addss	%xmm2, %xmm0
	addss	%xmm3, %xmm1

#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$1, %rax
#else
	subq	$2, %rax
#endif

	movq	AORIG, AO
	movq	BORIG, B
	leaq	BUFFER, BO

	salq	$ BASE_SHIFT, %rax
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 2), B
	leaq	(BO, %rax, 8), BO
#endif

#if defined(LN) || defined(LT)
	unpcklps %xmm1,  %xmm0

	movsd	 0 * SIZE(B), %xmm1
	subps	%xmm0,  %xmm1
#else
	movss	 0 * SIZE(AO), %xmm8
	movss	 1 * SIZE(AO), %xmm10
	subss	%xmm0, %xmm8
	subss	%xmm1, %xmm10
#endif

#if defined(LN) || defined(LT)
	movaps	  0 * SIZE(AO), %xmm6
	pshufd	 $0x00, %xmm6, %xmm8
	mulps	 %xmm8, %xmm1
#endif

#ifdef RN
	movaps	 0 * SIZE(B), %xmm0
	pshufd	 $0x00, %xmm0, %xmm2
	mulss	 %xmm2, %xmm8
	pshufd	 $0x55, %xmm0, %xmm2
	mulss	 %xmm8, %xmm2
	subss	 %xmm2, %xmm10

	pshufd	 $0xff, %xmm0, %xmm2
	mulss	 %xmm2, %xmm10
#endif

#ifdef RT
	movaps	  0 * SIZE(B), %xmm0
	pshufd	 $0xff, %xmm0, %xmm2
	mulss	 %xmm2, %xmm10
	pshufd	 $0xaa, %xmm0, %xmm2
	mulss	 %xmm10, %xmm2
	subss	 %xmm2, %xmm8
	pshufd	 $0x00, %xmm0, %xmm2
	mulss	 %xmm2, %xmm8
#endif

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

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

	pshufd	$0x00, %xmm1, %xmm2
	pshufd	$0x55, %xmm1, %xmm3
	movaps	%xmm2,   0 * SIZE(BO)
	movaps	%xmm3,   4 * SIZE(BO)
#else
	movss	%xmm8,   0 * SIZE(AO)
	movss	%xmm10,  1 * SIZE(AO)
#endif

#if defined(LN) || defined(LT)
	unpcklps %xmm10, %xmm1
	unpcklps %xmm11, %xmm5

	movaps	 %xmm1,  %xmm10
	unpcklps %xmm5,  %xmm1
	unpckhps %xmm5,  %xmm10

	movss	%xmm1,   0 * SIZE(CO1)
	movss	%xmm10,  0 * SIZE(CO1, LDC, 1)
#else
	movss	%xmm8,   0 * SIZE(CO1)
	movss	%xmm10,  0 * SIZE(CO1, LDC, 1)
#endif

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

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(AO, %rax, SIZE), AO
#ifdef LT
	addq	$ 2 * SIZE, B
#endif
#endif

#ifdef LN
	subq	$1, KK
	movq	BORIG, B
#endif

#ifdef LT
	addq	$1, KK
#endif

#ifdef RT
	movq	K, %rax
	movq	BORIG, B
	salq	$BASE_SHIFT, %rax
	addq	%rax, AORIG
#endif
	ALIGN_4	

.L70:
	testq	$2, M
	je	.L80

#ifdef LN
       movq	K, %rax
       salq	$1 + BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	movq	AORIG, AO
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
#endif

	leaq	BUFFER, BO

#if defined(LN) || defined(RT)
	movq	KK, %rax
	salq	$1 + BASE_SHIFT, %rax
	leaq	(BO, %rax, 4), BO
#endif	

	movaps	 0 * SIZE(AO), %xmm8
	movaps	 8 * SIZE(AO), %xmm10

	movaps	 0 * SIZE(BO), %xmm9
	movaps	16 * SIZE(BO), %xmm11
	movaps	32 * SIZE(BO), %xmm13
	movaps	48 * SIZE(BO), %xmm15

	pxor	%xmm0, %xmm0
	pxor	%xmm1, %xmm1
	pxor	%xmm2, %xmm2
	pxor	%xmm3, %xmm3

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$3, %rax
	je	.L85
	ALIGN_4

.L82:
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm0
#if defined(OPTERON) && defined(HAVE_PREFETCH)
	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)
#endif
	movlps	 4 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	movlps	 2 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm1
	movlps	 8 * SIZE(BO), %xmm9

	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm2
	movlps	12 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	movlps	 4 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm3
	movlps	64 * SIZE(BO), %xmm9

	mulps	%xmm8, %xmm11
	addps	%xmm11, %xmm0
	movlps	20 * SIZE(BO), %xmm11
	mulps	%xmm8, %xmm11
	movlps	 6 * SIZE(AO), %xmm8
	addps	%xmm11, %xmm1
	movlps	24 * SIZE(BO), %xmm11

	mulps	%xmm8, %xmm11
	addps	%xmm11, %xmm2
	movlps	28 * SIZE(BO), %xmm11
	mulps	%xmm8, %xmm11
	movlps	16 * SIZE(AO), %xmm8
	addps	%xmm11, %xmm3
	movlps	 80 * SIZE(BO), %xmm11

	mulps	%xmm10, %xmm13
	addps	%xmm13, %xmm0
	movlps	36 * SIZE(BO), %xmm13
	mulps	%xmm10, %xmm13
	movlps	10 * SIZE(AO), %xmm10
	addps	%xmm13, %xmm1
	movlps	40 * SIZE(BO), %xmm13

	mulps	%xmm10, %xmm13
	addps	%xmm13, %xmm2
	movlps	44 * SIZE(BO), %xmm13
	mulps	%xmm10, %xmm13
	movlps	12 * SIZE(AO), %xmm10
	addps	%xmm13, %xmm3
	movlps	 96 * SIZE(BO), %xmm13

	mulps	%xmm10, %xmm15
	addps	%xmm15, %xmm0
	movlps	52 * SIZE(BO), %xmm15
	mulps	%xmm10, %xmm15
	movlps	14 * SIZE(AO), %xmm10
	addps	%xmm15, %xmm1
	movlps	56 * SIZE(BO), %xmm15

	mulps	%xmm10, %xmm15
	addps	%xmm15, %xmm2
	movlps	60 * SIZE(BO), %xmm15
	mulps	%xmm10, %xmm15
	movlps	24 * SIZE(AO), %xmm10
	addps	%xmm15, %xmm3
	movlps	112 * SIZE(BO), %xmm15

	addq   $16 * SIZE, AO
	addq   $64 * SIZE, BO
	decq   %rax
	jne    .L82
	ALIGN_4

.L85:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L88
	ALIGN_4

.L86:
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm0
	movlps	 4 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	movlps	 2 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm1
	movlps	 8 * SIZE(BO), %xmm9

	addq	$2 * SIZE, AO		# aoffset  += 4
	addq	$8 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L86
	ALIGN_4

.L88:
	addps	%xmm2, %xmm0
	addps	%xmm3, %xmm1

#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$2, %rax
#else
	subq	$2, %rax
#endif

	movq	AORIG, AO
	movq	BORIG, B
	leaq	BUFFER, BO

	salq	$1 + BASE_SHIFT, %rax
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 1), B
	leaq	(BO, %rax, 4), BO
#endif

#if defined(LN) || defined(LT)
	unpcklps %xmm2,  %xmm0
	unpcklps %xmm3,  %xmm1

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

	movsd	 0 * SIZE(B), %xmm1
	movsd	 2 * SIZE(B), %xmm5

	subps	%xmm0,  %xmm1
	subps	%xmm2,  %xmm5
#else
	movsd	 0 * SIZE(AO), %xmm8
	movsd	 2 * SIZE(AO), %xmm10

	subps	%xmm0, %xmm8
	subps	%xmm1, %xmm10
#endif

#ifdef LN
	movaps	  0 * SIZE(AO), %xmm6
	pshufd	 $0xff, %xmm6, %xmm8
	mulps	 %xmm8, %xmm5
	pshufd	 $0xaa, %xmm6, %xmm8
	mulps	 %xmm5, %xmm8
	subps	 %xmm8, %xmm1
	pshufd	 $0x00, %xmm6, %xmm8
	mulps	 %xmm8, %xmm1
#endif

#ifdef LT
	movaps	 0 * SIZE(AO), %xmm6

	pshufd	 $0x00, %xmm6, %xmm8
	mulps	 %xmm8, %xmm1
	pshufd	 $0x55, %xmm6, %xmm8
	mulps	 %xmm1, %xmm8
	subps	 %xmm8, %xmm5

	pshufd	 $0xff, %xmm6, %xmm8
	mulps	 %xmm8, %xmm5
#endif

#ifdef RN
	movaps	 0 * SIZE(B), %xmm0
	pshufd	 $0x00, %xmm0, %xmm2
	mulps	 %xmm2, %xmm8
	pshufd	 $0x55, %xmm0, %xmm2
	mulps	 %xmm8, %xmm2
	subps	 %xmm2, %xmm10

	pshufd	 $0xff, %xmm0, %xmm2
	mulps	 %xmm2, %xmm10
#endif

#ifdef RT
	movaps	  0 * SIZE(B), %xmm0
	pshufd	 $0xff, %xmm0, %xmm2
	mulps	 %xmm2, %xmm10
	pshufd	 $0xaa, %xmm0, %xmm2
	mulps	 %xmm10, %xmm2
	subps	 %xmm2, %xmm8
	pshufd	 $0x00, %xmm0, %xmm2
	mulps	 %xmm2, %xmm8
#endif

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

#if defined(LN) || defined(LT)
	movlps	%xmm1,   0 * SIZE(B)
	movlps	%xmm5,   2 * SIZE(B)

	pshufd	$0x00, %xmm1, %xmm2
	pshufd	$0x55, %xmm1, %xmm3
	movaps	%xmm2,   0 * SIZE(BO)
	movaps	%xmm3,   4 * SIZE(BO)

	pshufd	$0x00, %xmm5, %xmm2
	pshufd	$0x55, %xmm5, %xmm3
	movaps	%xmm2,   8 * SIZE(BO)
	movaps	%xmm3,  12 * SIZE(BO)
#else
	movlps	%xmm8,   0 * SIZE(AO)
	movlps	%xmm10,  2 * SIZE(AO)
#endif

#if defined(LN) || defined(LT)
	unpcklps %xmm10, %xmm1
	unpcklps %xmm11, %xmm5

	movaps	 %xmm1,  %xmm10
	unpcklps %xmm5,  %xmm1
	unpckhps %xmm5,  %xmm10

	movlps	%xmm1,   0 * SIZE(CO1)
	movlps	%xmm10,  0 * SIZE(CO1, LDC, 1)
#else
	movlps	%xmm8,   0 * SIZE(CO1)
	movlps	%xmm10,  0 * SIZE(CO1, LDC, 1)
#endif

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

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
#ifdef LT
	addq	$ 4 * SIZE, B
#endif
#endif

#ifdef LN
	subq	$2, KK
	movq	BORIG, B
#endif

#ifdef LT
	addq	$2, KK
#endif

#ifdef RT
	movq	K, %rax
	movq	BORIG, B
	salq	$1 + BASE_SHIFT, %rax
	addq	%rax, AORIG
#endif
	ALIGN_4	

.L80:
	testq	$4, M
	je	.L90

#ifdef LN
       movq	K, %rax
       salq	$2 + BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	movq	AORIG, AO
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
#endif

	leaq	BUFFER, BO

#if defined(LN) || defined(RT)
	movq	KK, %rax
	salq	$1 + BASE_SHIFT, %rax
	leaq	(BO, %rax, 4), BO
#endif	

	movaps	 0 * SIZE(AO), %xmm8
	movaps	16 * SIZE(AO), %xmm10

	movaps	 0 * SIZE(BO), %xmm9
	movaps	16 * SIZE(BO), %xmm11
	movaps	32 * SIZE(BO), %xmm13
	movaps	48 * SIZE(BO), %xmm15

	pxor	%xmm0, %xmm0
	pxor	%xmm1, %xmm1
	pxor	%xmm2, %xmm2
	pxor	%xmm3, %xmm3

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$3, %rax
	je	.L75
	ALIGN_4

.L72:
	mulps	%xmm8, %xmm9
#if defined(OPTERON) && defined(HAVE_PREFETCH)
	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)
#endif

	mulps	 4 * SIZE(BO), %xmm8
	addps	%xmm9, %xmm0
	movaps	 8 * SIZE(BO), %xmm9
	addps	%xmm8, %xmm1
	movaps	 4 * SIZE(AO), %xmm8

	mulps	%xmm8, %xmm9
	mulps	12 * SIZE(BO), %xmm8
	addps	%xmm9, %xmm2
	movaps	64 * SIZE(BO), %xmm9
	addps	%xmm8, %xmm3
	movaps	 8 * SIZE(AO), %xmm8

	mulps	%xmm8, %xmm11
	mulps	20 * SIZE(BO), %xmm8
	addp