trsm_kernel_rt_4x4_barcelona.s

Optimized GotoBLAS libraries

	movsd	-11 * SIZE(AO), %xmm12
	mulsd	 %xmm12, %xmm8

	movsd	-12 * SIZE(AO), %xmm13
	mulsd	 %xmm8, %xmm13
	subsd	 %xmm13, %xmm10

	movsd	-16 * SIZE(AO), %xmm12
	mulsd	 %xmm12, %xmm10

	unpcklpd %xmm8, %xmm10
	unpcklpd %xmm9, %xmm11
#endif

#ifdef LT
	movapd	%xmm10, %xmm8
        unpckhpd %xmm8, %xmm8

	movapd	%xmm11, %xmm9
        unpckhpd %xmm9, %xmm9

	movsd	-16 * SIZE(AO), %xmm12
	mulsd	 %xmm12, %xmm10

	movsd	-15 * SIZE(AO), %xmm13
	mulsd	 %xmm10, %xmm13
	subsd	 %xmm13, %xmm8
	movsd	-14 * SIZE(AO), %xmm14
	mulsd	 %xmm10, %xmm14
	subsd	 %xmm14, %xmm11
	movsd	-13 * SIZE(AO), %xmm15
	mulsd	 %xmm10, %xmm15
	subsd	 %xmm15, %xmm9

	movsd	-11 * SIZE(AO), %xmm12
	mulsd	 %xmm12, %xmm8

	movsd	-10 * SIZE(AO), %xmm13
	mulsd	 %xmm8, %xmm13
	subsd	 %xmm13, %xmm11
	movsd	 -9 * SIZE(AO), %xmm14
	mulsd	 %xmm8, %xmm14
	subsd	 %xmm14, %xmm9

	movsd	 -6 * SIZE(AO), %xmm12
	mulsd	 %xmm12, %xmm11

	movsd	 -5 * SIZE(AO), %xmm13
	mulsd	 %xmm11, %xmm13
	subsd	 %xmm13, %xmm9

	movsd	 -1 * SIZE(AO), %xmm12
	mulsd	 %xmm12, %xmm9

	unpcklpd %xmm8, %xmm10
	unpcklpd %xmm9, %xmm11
#endif

#ifdef RN
	movddup	-16 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm10
	mulpd	 %xmm8, %xmm11
#endif

#ifdef RT
	movddup	-16 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm10
	mulpd	 %xmm8, %xmm11
#endif

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

	movlpd	%xmm10,  0 * SIZE(CO1)
	movhpd	%xmm10,  1 * SIZE(CO1)
	movlpd	%xmm11,  2 * SIZE(CO1)
	movhpd	%xmm11,  3 * SIZE(CO1)

#if defined(LN) || defined(LT)
	movaps	%xmm10, -16 * SIZE(BO)
	movaps	%xmm11, -14 * SIZE(BO)
#else
	movaps	%xmm10, -16 * SIZE(AO)
	movaps	%xmm11, -14 * SIZE(AO)
#endif

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

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	leaq	(,%rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	addq	%rax, 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	.L91
	ALIGN_4	

.L100:
	testq	$2, M
	je	.L110

#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

	movq	B, BO

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

	movddup	-16 * SIZE(BO), %xmm0
	pxor	%xmm8, %xmm8
	movddup	-15 * SIZE(BO), %xmm1
	pxor	%xmm9, %xmm9
	movddup	-14 * SIZE(BO), %xmm2
	pxor	%xmm10, %xmm10
	movddup	-13 * SIZE(BO), %xmm3
	pxor	%xmm11, %xmm11

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$-4, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 1), BO
	negq	%rax
	je,pn	.L106
	ALIGN_4

.L102:
	mulpd	-16 * SIZE(AO, %rax, 2), %xmm0
	addpd	%xmm0, %xmm8
	movddup	-12 * SIZE(BO, %rax, 1), %xmm0

	mulpd	-14 * SIZE(AO, %rax, 2), %xmm1
	addpd	%xmm1, %xmm9
	movddup	-11 * SIZE(BO, %rax, 1), %xmm1

	mulpd	-12 * SIZE(AO, %rax, 2), %xmm2
	addpd	%xmm2, %xmm10
	movddup	-10 * SIZE(BO, %rax, 1), %xmm2

	mulpd	-10 * SIZE(AO, %rax, 2), %xmm3
	addpd	%xmm3, %xmm11
	movddup	 -9 * SIZE(BO, %rax, 1), %xmm3

	addq	$4 * SIZE, %rax
	jl,pt	.L102
	ALIGN_4

.L106:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L109

	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 1), BO
	negq	%rax
	ALIGN_4

.L107:
	movddup	-16 * SIZE(BO, %rax, 1), %xmm0
	mulpd	-16 * SIZE(AO, %rax, 2), %xmm0
	addpd	%xmm0, %xmm8

	addq	$SIZE, %rax
	jl	.L107
	ALIGN_4

.L109:
	addpd	%xmm9, %xmm8
	addpd	%xmm11, %xmm10
	addpd	%xmm10, %xmm8

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

	leaq	(, %rax, SIZE), %rax

	movq	AORIG, AO
	leaq	(AO, %rax, 2), AO
	leaq	(B,  %rax, 1), BO
#endif

#if defined(LN) || defined(LT)
	movapd	-16 * SIZE(BO), %xmm10
	subpd	%xmm8,  %xmm10
#else
	movapd	-16 * SIZE(AO), %xmm10
	subpd	%xmm8, %xmm10
#endif

#ifdef LN
	movapd	%xmm10, %xmm8
        unpckhpd %xmm8, %xmm8

	movsd	-13 * SIZE(AO), %xmm12
	mulsd	 %xmm12, %xmm8

	movsd	-14 * SIZE(AO), %xmm13
	mulsd	 %xmm8, %xmm13
	subsd	 %xmm13, %xmm10

	movsd	-16 * SIZE(AO), %xmm12
	mulsd	 %xmm12, %xmm10

	unpcklpd %xmm8, %xmm10
#endif

#ifdef LT
	movapd	%xmm10, %xmm8
        unpckhpd %xmm8, %xmm8

	movsd	-16 * SIZE(AO), %xmm12
	mulsd	 %xmm12, %xmm10

	movsd	-15 * SIZE(AO), %xmm13
	mulsd	 %xmm10, %xmm13
	subsd	 %xmm13, %xmm8

	movsd	-13 * SIZE(AO), %xmm12
	mulsd	 %xmm12, %xmm8

	unpcklpd %xmm8, %xmm10
#endif

#ifdef RN
	movddup	-16 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm10
#endif

#ifdef RT
	movddup	-16 * SIZE(BO), %xmm8
	mulpd	 %xmm8, %xmm10
#endif

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

#if defined(LN) || defined(LT)
	movlpd	%xmm10,  0 * SIZE(CO1)
	movhpd	%xmm10,  1 * SIZE(CO1)
#else
	movlpd	%xmm10,  0 * SIZE(CO1)
	movhpd	%xmm10,  1 * SIZE(CO1)
#endif

#if defined(LN) || defined(LT)
	movaps	%xmm10, -16 * SIZE(BO)
#else
	movaps	%xmm10, -16 * SIZE(AO)
#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
	addq	%rax, 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	

.L110:
	testq	$1, M
	je	.L119

#ifdef LN
       movq	K, %rax
       salq	$0 + 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, 1), AO
#endif

	movq	B, BO

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

	movapd	-16 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movapd	-14 * SIZE(AO), %xmm1
	pxor	%xmm9, %xmm9

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$-4, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 1), BO
	negq	%rax
	je,pn	.L116
	ALIGN_4

.L112:
	mulpd	-16 * SIZE(BO, %rax, 1), %xmm0
	addpd	%xmm0, %xmm8
	movapd	-12 * SIZE(AO, %rax, 1), %xmm0

	mulpd	-14 * SIZE(BO, %rax, 1), %xmm1
	addpd	%xmm1, %xmm9
	movapd	-10 * SIZE(AO, %rax, 1), %xmm1

	addq	$4 * SIZE, %rax
	jl,pt	.L112
	ALIGN_4

.L116:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L118

	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 1), BO
	negq	%rax
	ALIGN_4

.L117:
	mulsd	-16 * SIZE(BO, %rax, 1), %xmm0
	addsd	%xmm0, %xmm8
	movsd	-15 * SIZE(AO, %rax, 1), %xmm0

	addq	$SIZE, %rax
	jl	.L117
	ALIGN_4

.L118:
	addpd	%xmm9, %xmm8
	haddpd	%xmm8, %xmm8

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

	leaq	(, %rax, SIZE), %rax

	movq	AORIG, AO
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 1), BO
#endif

#if defined(LN) || defined(LT)
	movsd	-16 * SIZE(BO), %xmm10
	subsd	%xmm8,  %xmm10
#else
	movsd	-16 * SIZE(AO), %xmm10
	subsd	%xmm8, %xmm10
#endif

#if defined(LN) || defined(LT)
	movsd	-16 * SIZE(AO), %xmm12
	mulsd	 %xmm12, %xmm10
#endif

#if defined(RN) || defined(RT)
	movsd	-16 * SIZE(BO), %xmm8
	mulsd	 %xmm8, %xmm10
#endif

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

	movsd	%xmm10,  0 * SIZE(CO1)

#if defined(LN) || defined(LT)
	movlpd	%xmm10, -16 * SIZE(BO)
#else
	movlpd	%xmm10, -16 * SIZE(AO)
#endif

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

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

.L119:
#ifdef LN
       leaq	(B, K, SIZE), B
#endif

#if defined(LT) || defined(RN)
	movq	BO, B
#endif

#ifdef RN
	addq	$1, KK
#endif

#ifdef RT
	subq	$1, KK
#endif
	ALIGN_4

.L40:	
	testq	$2, N
	je	.L80

#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			# coffset1 = c
	leaq	(C, LDC, 1), CO2	# coffset2 = c + ldc
#ifndef RT
	leaq	(C, LDC, 2), C
#endif

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

#if defined(LT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif

	movq	M,  I
	sarq	$2, I	# i = (m >> 2)
	jle	.L60
	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
	movq	AORIG, AO
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
#endif

	movq	B, BO

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

	movddup	-16 * SIZE(BO), %xmm1
	movddup	-15 * SIZE(BO), %xmm5
	pxor	%xmm8, %xmm8
	movddup	-12 * SIZE(BO), %xmm3
	pxor	%xmm9, %xmm9
	movapd	-16 * SIZE(AO), %xmm0
	pxor	%xmm12, %xmm12
	movapd	 -8 * SIZE(AO), %xmm4
	pxor	%xmm13, %xmm13

#ifndef LN
	prefetchw      3 * SIZE(CO1)
	movapd	%xmm0, %xmm2
	prefetchw      3 * SIZE(CO2)
#else
	prefetchw     -8 * SIZE(CO1)
	movapd	%xmm0, %xmm2
	prefetchw     -8 * SIZE(CO2)
#endif


#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$-4, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 2), BO
	negq	%rax
	je,pn	.L56
	ALIGN_4

.L52:
	mulpd	%xmm1, %xmm0
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm8
	movapd	-12 * SIZE(AO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm12
	movddup	-14 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm5, %xmm2
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	-13 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm0, %xmm2
	mulpd	%xmm1, %xmm0
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm8
	movapd	  (AO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm12
	movddup	 -8 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm5, %xmm2
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	-11 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm4, %xmm2
	mulpd	%xmm3, %xmm4
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm3
	addpd	%xmm4, %xmm8
	movapd	 -4 * SIZE(AO, %rax, 4), %xmm4
	addpd	%xmm3, %xmm12
	movddup	-10 * SIZE(BO, %rax, 2), %xmm3
	mulpd	%xmm5, %xmm2
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	 -9 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm4, %xmm2
	mulpd	%xmm3, %xmm4
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm3
	addpd	%xmm4, %xmm8
	movapd	  8 * SIZE(AO, %rax, 4), %xmm4
	addpd	%xmm3, %xmm12
	movddup	 -4 * SIZE(BO, %rax, 2), %xmm3
	mulpd	%xmm5, %xmm2
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	 -7 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm0, %xmm2

	addq	$4 * SIZE, %rax
	jl,pt	.L52
	ALIGN_4

.L56:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K, %rax
	subq	KK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L59

	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 2), BO
	negq	%rax
	ALIGN_4

.L57:
	mulpd	%xmm1, %xmm0