trsm_kernel_ln_4x4_sse3.s

Optimized GotoBLAS libraries

	movsd	%xmm3,  0 * SIZE(CO1, LDC, 2)
	movsd	%xmm7,  1 * SIZE(CO1, LDC, 2)
	movsd	%xmm11, 2 * SIZE(CO1, LDC, 2)
	movsd	%xmm15, 3 * SIZE(CO1, LDC, 2)

	movhpd	%xmm3,  0 * SIZE(CO2, LDC, 2)
	movhpd	%xmm7,  1 * SIZE(CO2, LDC, 2)
	movhpd	%xmm11, 2 * SIZE(CO2, LDC, 2)
	movhpd	%xmm15, 3 * SIZE(CO2, LDC, 2)
#else
	movsd	%xmm8,  0 * SIZE(CO1)
	movhpd	%xmm8,  1 * SIZE(CO1)
	movsd	%xmm9,  2 * SIZE(CO1)
	movhpd	%xmm9,  3 * SIZE(CO1)

	movsd	%xmm10,  0 * SIZE(CO2)
	movhpd	%xmm10,  1 * SIZE(CO2)
	movsd	%xmm11,  2 * SIZE(CO2)
	movhpd	%xmm11,  3 * SIZE(CO2)

	movsd	%xmm12,  0 * SIZE(CO1, LDC, 2)
	movhpd	%xmm12,  1 * SIZE(CO1, LDC, 2)
	movsd	%xmm13,  2 * SIZE(CO1, LDC, 2)
	movhpd	%xmm13,  3 * SIZE(CO1, LDC, 2)

	movsd	%xmm14,  0 * SIZE(CO2, LDC, 2)
	movhpd	%xmm14,  1 * SIZE(CO2, LDC, 2)
	movsd	%xmm15,  2 * SIZE(CO2, LDC, 2)
	movhpd	%xmm15,  3 * SIZE(CO2, LDC, 2)
#endif

#if defined(LN) || defined(LT)
	movapd	%xmm1,   0 * SIZE(BO)
	movapd	%xmm3,   2 * SIZE(BO)
	movapd	%xmm5,   4 * SIZE(BO)
	movapd	%xmm7,   6 * SIZE(BO)
	movapd	%xmm9,   8 * SIZE(BO)
	movapd	%xmm11, 10 * SIZE(BO)
	movapd	%xmm13, 12 * SIZE(BO)
	movapd	%xmm15, 14 * SIZE(BO)
#else
	movapd	%xmm8,   0 * SIZE(AO)
	movapd	%xmm9,   2 * SIZE(AO)
	movapd	%xmm10,  4 * SIZE(AO)
	movapd	%xmm11,  6 * SIZE(AO)
	movapd	%xmm12,  8 * SIZE(AO)
	movapd	%xmm13, 10 * SIZE(AO)
	movapd	%xmm14, 12 * SIZE(AO)
	movapd	%xmm15, 14 * SIZE(AO)
#endif

#ifndef LN
	addq	$4 * SIZE, CO1
	addq	$4 * SIZE, CO2
#endif

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 4), BO
#endif

#ifdef LN
	subq	$4, KK
#endif

#ifdef LT
	addq	$4, KK
#endif

#ifdef RT
       movq	K, %rax
       salq	$2 + BASE_SHIFT, %rax
       addq	%rax, AORIG
#endif

	decq	I			# i --
	jg	.L11
	ALIGN_4


.L39:
#ifdef LN
       leaq	(, K, SIZE), %rax
       leaq	(B, %rax, 4), B
#endif
#if defined(LT) || defined(RN)
	movq	BO, B
#endif

#ifdef RN
	addq	$4, KK
#endif

#ifdef RT
	subq	$4, KK
#endif

	decq	J			# j --
	jg	.L10
	ALIGN_4

.L40:
	testq	$2, N
	je	.L80
	ALIGN_4

#if defined(LT) || defined(RN)
	movq	A, AO
#else
	movq	A, AORIG
#endif

#ifdef RT
       movq	K, %rax
       salq	$1 + BASE_SHIFT, %rax
       subq	%rax, B
       
       leaq	(, LDC, 2), %rax
       subq	%rax, C
#endif

	movq	C, CO1
	leaq	(C, LDC, 1), CO2
#ifndef RT
	leaq	(C, LDC, 2), C
#endif

#ifdef LN
	movq	OFFSET, %rax
	addq	M, %rax
	movq	%rax, KK
#endif	

	movq	K, %rax
	salq	$BASE_SHIFT + 1, %rax
	leaq	(B, %rax), BB

#ifdef LT
	movq	OFFSET, %rax
	movq	%rax, KK
#endif


	testq	$1, M
	je	.L70
	ALIGN_4

.L71:
#ifdef LN
       movq	K, %rax
       salq	$0 + BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	movq	AORIG, AO
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 2), BO
#else
	movq	B, BO
#endif	

	movddup	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movapd	 0 * SIZE(BO), %xmm9
	pxor	%xmm1, %xmm1
	movddup	 4 * SIZE(AO), %xmm10
	pxor	%xmm2, %xmm2
	movapd	 8 * SIZE(BO), %xmm11
	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:
	mulpd	%xmm8, %xmm9
	PREFETCH  (PREFETCHSIZE + 0) * SIZE(AO)
	movddup	 1 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm0
	mulpd	 2 * SIZE(BO), %xmm8
	movapd	16 * SIZE(BO), %xmm9
	addpd	%xmm8, %xmm1
	movddup	 2 * SIZE(AO), %xmm8
	mulpd	 4 * SIZE(BO), %xmm8
	addpd	%xmm8, %xmm2
	movddup	 3 * SIZE(AO), %xmm8
	mulpd	 6 * SIZE(BO), %xmm8
	addpd	%xmm8, %xmm3
	movddup	 8 * SIZE(AO), %xmm8
	mulpd	%xmm10, %xmm11
	movddup	 5 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm0
	mulpd	10 * SIZE(BO), %xmm10
	movapd	24 * SIZE(BO), %xmm11
	addpd	%xmm10, %xmm1
	movddup	 6 * SIZE(AO), %xmm10
	mulpd	12 * SIZE(BO), %xmm10
	addpd	%xmm10, %xmm2
	movddup	 7 * SIZE(AO), %xmm10
	mulpd	14 * SIZE(BO), %xmm10
	addpd	%xmm10, %xmm3
	movddup	12 * SIZE(AO), %xmm10

	addq   $ 8 * SIZE, AO
	addq   $16 * SIZE, BO
	decq   %rax
	jne    .L72
	ALIGN_4

.L75:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L78
	ALIGN_4

.L76:
	mulpd	%xmm8, %xmm9
	movddup	 1 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm0
	movapd	 2 * SIZE(BO), %xmm9

	addq	$1 * SIZE, AO		# aoffset  += 4
	addq	$2 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L76
	ALIGN_4

.L78:
	addpd	%xmm1, %xmm0
	addpd	%xmm3, %xmm2
	addpd	%xmm2, %xmm0

#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$1, %rax
#else
	subq	$2, %rax
#endif
	leaq	(, %rax, SIZE), %rax
	movq	AORIG, AO
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 2), BO
#endif

#if defined(LN) || defined(LT)
	movapd	 0 * SIZE(BO), %xmm2
	subpd	%xmm0, %xmm2
#else
	movapd	 0 * SIZE(AO), %xmm2
	subpd	%xmm0, %xmm2
#endif

#ifdef LN
	movddup	 0 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm2
#endif

#ifdef LT
	movddup	 0 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm2
#endif

#ifdef RN
	movapd	%xmm2, %xmm0
        unpckhpd %xmm0, %xmm0

	movsd	 0 * SIZE(BO), %xmm4
	mulsd	 %xmm4, %xmm2

	movsd	 1 * SIZE(BO), %xmm5
	mulsd	 %xmm2, %xmm5
	subsd	 %xmm5, %xmm0

	movsd	 3 * SIZE(BO), %xmm4
	mulsd	 %xmm4, %xmm0

	unpcklpd %xmm0, %xmm2
#endif

#ifdef RT
	movapd	%xmm2, %xmm0
        unpckhpd %xmm0, %xmm0

	movsd	 3 * SIZE(BO), %xmm4
	mulsd	 %xmm4, %xmm0

	movsd	 2 * SIZE(BO), %xmm5
	mulsd	 %xmm0, %xmm5
	subsd	 %xmm5, %xmm2

	movsd	 0 * SIZE(BO), %xmm4
	mulsd	 %xmm4, %xmm2

	unpcklpd %xmm0, %xmm2
#endif

#ifdef LN
	subq	$1 * SIZE, CO1
	subq	$1 * SIZE, CO2
#endif

#if defined(LN) || defined(LT)
	movsd	%xmm2,  0 * SIZE(CO1)
	movhpd	%xmm2,  0 * SIZE(CO2)
#else
	movsd	%xmm2,  0 * SIZE(CO1)
	movhpd	%xmm2,  0 * SIZE(CO2)
#endif

#if defined(LN) || defined(LT)
	movapd	%xmm2,   0 * SIZE(BO)
#else
	movapd	%xmm2,   0 * SIZE(AO)
#endif

#ifndef LN
	addq	$1 * SIZE, CO1
	addq	$1 * SIZE, CO2
#endif

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 2), BO
#endif

#ifdef LN
	subq	$1, KK
#endif

#ifdef LT
	addq	$1, KK
#endif

#ifdef RT
       movq	K, %rax
       salq	$0 + BASE_SHIFT, %rax
       addq	%rax, AORIG
#endif
	ALIGN_4	

.L70:	
	testq	$2, M
	je	.L60
	ALIGN_4

.L61:
#ifdef LN
       movq	K, %rax
       salq	$1 + BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

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

	movapd	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movddup	 0 * SIZE(BO), %xmm9
	pxor	%xmm1, %xmm1
	movapd	 8 * SIZE(AO), %xmm10
	pxor	%xmm2, %xmm2
	movddup	 8 * SIZE(BO), %xmm11
	pxor	%xmm3, %xmm3

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	sarq	$3, %rax
	je	.L65
	ALIGN_4

.L62:
	mulpd	%xmm8, %xmm9
	PREFETCH  (PREFETCHSIZE + 0) * SIZE(AO)
	addpd	%xmm9, %xmm0
	movddup	 1 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 2 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movddup	 2 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm2
	movddup	 3 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 4 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm3
	movddup	 4 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm0
	movddup	 5 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 6 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movddup	 6 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm2
	movddup	 7 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	16 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm3
	movddup	16 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm11
	addpd	%xmm11, %xmm0
	movddup	 9 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	movapd	10 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm1
	movddup	10 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	addpd	%xmm11, %xmm2
	movddup	11 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	movapd	12 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm3
	movddup	12 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	addpd	%xmm11, %xmm0
	movddup	13 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	movapd	14 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm1
	movddup	14 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	addpd	%xmm11, %xmm2
	movddup	15 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	movapd	24 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm3
	movddup	24 * SIZE(BO), %xmm11

	addq   $16 * SIZE, AO
	addq   $16 * SIZE, BO
	decq   %rax
	jne    .L62
	ALIGN_4

.L65:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L69
	ALIGN_4

.L66:
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm0
	movddup	 1 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 2 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movddup	 2 * SIZE(BO), %xmm9

	addq	$2 * SIZE, AO		# aoffset  += 4
	addq	$2 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L66
	ALIGN_4

.L69:
	addpd	%xmm2, %xmm0
	addpd	%xmm3, %xmm1

#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$2, %rax
#else
	subq	$2, %rax
#endif
	leaq	(, %rax, SIZE), %rax
	movq	AORIG, AO
	leaq	(AO, %rax, 2), AO
	leaq	(B,  %rax, 2), BO
#endif

#if defined(LN) || defined(LT)
	movapd	%xmm0, %xmm8
	unpcklpd %xmm1, %xmm0
	unpckhpd %xmm1, %xmm8

	movapd	 0 * SIZE(BO), %xmm1
	movapd	 2 * SIZE(BO), %xmm5

	subpd	%xmm0,  %xmm1
	subpd	%xmm8,  %xmm5
#else

	movapd	 0 * SIZE(AO), %xmm8
	movapd	 2 * SIZE(AO), %xmm10

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

#ifdef LN
	movddup	 3 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm5
	movddup	 2 * SIZE(AO), %xmm2
	mulpd	 %xmm5, %xmm2
	subpd	 %xmm2, %xmm1

	movddup	 0 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm1
#endif

#ifdef LT
	movddup	 0 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm1

	movddup	 1 * SIZE(AO), %xmm2
	mulpd	 %xmm1, %xmm2
	subpd	 %xmm2, %xmm5

	movddup	 3 * SIZE(AO), %xmm0
	mulpd	 %xmm0, %xmm5
#endif

#ifdef RN
	movddup	 0 * SIZE(BO), %xmm0
	mulpd	 %xmm0, %xmm8

	movddup	 1 * SIZE(BO), %xmm1
	mulpd	 %xmm8, %xmm1
	subpd	 %xmm1, %xmm10

	movddup	 3 * SIZE(BO), %xmm0
	mulpd	 %xmm0, %xmm10
#endif

#ifdef RT
	movddup	 3 * SIZE(BO), %xmm0
	mulpd	 %xmm0, %xmm10

	movddup	 2 * SIZE(BO), %xmm1
	mulpd	 %xmm10, %xmm1
	subpd	 %xmm1, %xmm8

	movddup	 0 * SIZE(BO), %xmm0
	mulpd	 %xmm0, %xmm8
#endif

#ifdef LN
	subq	$2 * SIZE, CO1
	subq	$2 * SIZE, CO2
#endif

#if defined(LN) || defined(LT)
	movsd	%xmm1,  0 * SIZE(CO1)
	movsd	%xmm5,  1 * SIZE(CO1)
	movhpd	%xmm1,  0 * SIZE(CO2)
	movhpd	%xmm5,  1 * SIZE(CO2)
#else
	movsd	%xmm8,  0 * SIZE(CO1)
	movhpd	%xmm8,  1 * SIZE(CO1)
	movsd	%xmm10, 0 * SIZE(CO2)
	movhpd	%xmm10, 1 * SIZE(CO2)
#endif

#if defined(LN) || defined(LT)
	movapd	%xmm1,   0 * SIZE(BO)
	movapd	%xmm5,   2 * SIZE(BO)
#else
	movapd	%xmm8,   0 * SIZE(AO)
	movapd	%xmm10,  2 * SIZE(AO)
#endif

#ifndef LN
	addq	$2 * SIZE, CO1
	addq	$2 * SIZE, CO2
#endif

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 2), BO
#endif

#ifdef LN
	subq	$2, KK
#endif

#ifdef LT
	addq	$2, KK
#endif

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

.L60:
	movq	M,  I
	sarq	$2, I	# i = (m >> 2)
	jle	.L79
	ALIGN_4

.L51:
#ifdef LN
       movq	K, %rax
       salq	$2 + BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

#if defined(LN) || defined(RT)
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	movq	AORIG, AO
	leaq	(AO, %rax, 4), AO
	leaq	(B,  %rax, 2), BO
#else
	movq	B, BO
#endif	

	prefetcht0	  0 * SIZE(BB)
	subq	   $-4 * SIZE, BB

	movapd	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movddup	 0 * SIZE(BO), %xmm9