trsm_kernel_ln_4x2_sse2.s

Optimized GotoBLAS libraries

#ifdef LN
       movl	K, %eax
       sall	$0 + 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, 1), AA
#endif

	leal	BUFFER, BB

	movsd	 0 * SIZE(BB), %xmm2
	pxor	%xmm4, %xmm4
	movsd	 0 * SIZE(AA), %xmm0
	pxor	%xmm5, %xmm5
	movsd	 8 * SIZE(BB), %xmm3
	pxor	%xmm6, %xmm6
	movsd	 4 * SIZE(AA), %xmm1
	pxor	%xmm7, %xmm7

#ifdef LN
	prefetcht2	-4 * SIZE(%esi)
#else
	prefetcht2	4 * SIZE(%esi)
#endif

#if defined(LN) || defined(RT)
	movl	KK, %eax
	sall	$0 + BASE_SHIFT, %eax
	leal	(BB, %eax, 2), BB
#endif	

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	sarl	$3, %eax
	je	.L152

.L151:	
	mulsd	%xmm0, %xmm2
	movsd	 1 * SIZE(AA), %xmm0
	addsd	%xmm2, %xmm4
	mulsd	 2 * SIZE(BB), %xmm0
	movsd	16 * SIZE(BB), %xmm2
	addsd	%xmm0, %xmm4
	movsd	 2 * SIZE(AA), %xmm0
	mulsd	 4 * SIZE(BB), %xmm0
	addsd	%xmm0, %xmm4
	movsd	 3 * SIZE(AA), %xmm0
	mulsd	 6 * SIZE(BB), %xmm0
	addsd	%xmm0, %xmm4
	movsd	 8 * SIZE(AA), %xmm0
	mulsd	%xmm1, %xmm3
	movsd	 5 * SIZE(AA), %xmm1
	addsd	%xmm3, %xmm4
	mulsd	10 * SIZE(BB), %xmm1
	movsd	24 * SIZE(BB), %xmm3
	addsd	%xmm1, %xmm4
	movsd	 6 * SIZE(AA), %xmm1
	mulsd	12 * SIZE(BB), %xmm1
	addsd	%xmm1, %xmm4
	movsd	 7 * SIZE(AA), %xmm1
	mulsd	14 * SIZE(BB), %xmm1
	addsd	%xmm1, %xmm4
	movsd	12 * SIZE(AA), %xmm1

	addl   $ 8 * SIZE, AA
	addl   $16 * SIZE, BB
	BRANCH
	decl   %eax
	jne    .L151

.L152:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L154

.L153:
	movsd	 0 * SIZE(AA), %xmm0
	mulsd	 0 * SIZE(BB), %xmm0
	addsd	%xmm0, %xmm4

	addl	$1 * SIZE, AA		# aoffset  += 8
	addl	$2 * SIZE, BB		# boffset1 += 8
	decl	%eax
	BRANCH
	jg	.L153
	ALIGN_4

.L154:
	addsd	%xmm6, %xmm4
	addsd	%xmm7, %xmm5

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

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

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

#if defined(LN) || defined(LT)
	movsd	 0 * SIZE(B), %xmm0
#else
	movsd	 0 * SIZE(AA), %xmm0
#endif

	subsd	%xmm4, %xmm0

#if defined(LN) || defined(LT)
	mulsd	 0 * SIZE(AA), %xmm0
#endif

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

#if defined(LN) || defined(LT)
	movsd	%xmm0,   0 * SIZE(B)

	movsd	%xmm0,   0 * SIZE(BB)
	movsd	%xmm0,   1 * SIZE(BB)
#else
	movsd	%xmm0,   0 * SIZE(AA)
#endif

#ifdef LN
	subl	$1 * SIZE, %esi
#endif

	movsd	%xmm0,  0 * SIZE(%esi)

#ifndef LN
	addl	$1 * SIZE, %esi
#endif

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	leal	(,%eax, SIZE), %eax
	leal	(AA, %eax, 1), AA
#ifdef LT
	addl	$1 * SIZE, B
#endif
#endif

#ifdef LN
	subl	$1, KK
	movl	BORIG, B
#endif

#ifdef LT
	addl	$1, KK
#endif

#ifdef RT
	movl	K, %eax
	movl	BORIG, B
	sall	$0 + BASE_SHIFT, %eax
	addl	%eax, AORIG
#endif
	ALIGN_2

.L130:
	movl	M,  %ebx
	testl	$2, %ebx
	jle	.L150

#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

	movapd	 0 * SIZE(BB), %xmm2
	pxor	%xmm4, %xmm4
	movapd	 0 * SIZE(AA), %xmm0
	pxor	%xmm5, %xmm5
	movapd	 8 * SIZE(BB), %xmm3
	pxor	%xmm6, %xmm6
	movapd	 8 * SIZE(AA), %xmm1
	pxor	%xmm7, %xmm7

#if defined(LN) || defined(RT)
	movl	KK, %eax
	sall	$0 + BASE_SHIFT, %eax
	leal	(BB, %eax, 2), BB
#endif	

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	sarl	$3, %eax
	je	.L132

.L131:	
	mulpd	%xmm0, %xmm2
	movapd	 2 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm4
	mulpd	 2 * SIZE(BB), %xmm0
	movapd	16 * SIZE(BB), %xmm2
	addpd	%xmm0, %xmm5
	movapd	 4 * SIZE(AA), %xmm0
	mulpd	 4 * SIZE(BB), %xmm0
	addpd	%xmm0, %xmm6
	movapd	 6 * SIZE(AA), %xmm0
	mulpd	 6 * SIZE(BB), %xmm0
	addpd	%xmm0, %xmm7
	movapd	16 * SIZE(AA), %xmm0
	mulpd	%xmm1, %xmm3
	movapd	10 * SIZE(AA), %xmm1
	addpd	%xmm3, %xmm4
	mulpd	10 * SIZE(BB), %xmm1
	movapd	24 * SIZE(BB), %xmm3
	addpd	%xmm1, %xmm5
	movapd	12 * SIZE(AA), %xmm1
	mulpd	12 * SIZE(BB), %xmm1
	addpd	%xmm1, %xmm6
	movapd	14 * SIZE(AA), %xmm1
	mulpd	14 * SIZE(BB), %xmm1
	addpd	%xmm1, %xmm7
	movapd	24 * SIZE(AA), %xmm1

	addl   $16 * SIZE, AA
	addl   $16 * SIZE, BB
	BRANCH
	decl   %eax
	jne    .L131

.L132:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L134

.L133:
	movapd	 0 * SIZE(AA), %xmm0
	mulpd	 0 * SIZE(BB), %xmm0
	addpd	%xmm0, %xmm4

	addl	$2 * SIZE, AA		# aoffset  += 8
	addl	$2 * SIZE, BB		# boffset1 += 8
	decl	%eax
	BRANCH
	jg	.L133
	ALIGN_4

.L134:
	addpd	%xmm5, %xmm4
	addpd	%xmm7, %xmm6
	addpd	%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

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

#if defined(LN) || defined(LT)
	movapd	 0 * SIZE(B), %xmm0
#else
	movapd	 0 * SIZE(AA), %xmm0
#endif

	subpd	%xmm4, %xmm0

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

	movsd	 3 * SIZE(AA), %xmm4
	mulsd	 %xmm4, %xmm2

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

	movsd	 0 * SIZE(AA), %xmm4
	mulsd	 %xmm4, %xmm0

	unpcklpd %xmm2, %xmm0
#endif

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

	movsd	 0 * SIZE(AA), %xmm4
	mulsd	 %xmm4, %xmm0

	movsd	 1 * SIZE(AA), %xmm5
	mulsd	 %xmm0, %xmm5
	subsd	 %xmm5, %xmm2
	movsd	 3 * SIZE(AA), %xmm4
	mulsd	 %xmm4, %xmm2

	unpcklpd %xmm2, %xmm0
#endif

#if defined(RN) || defined(RT)
	movsd	 0 * SIZE(B), %xmm4
	movhpd	 0 * SIZE(B), %xmm4
	mulpd	 %xmm4, %xmm0
#endif

#if defined(LN) || defined(LT)
	movapd	%xmm0,   0 * SIZE(B)

	movsd	%xmm0,   0 * SIZE(BB)
	movsd	%xmm0,   1 * SIZE(BB)
	movhpd	%xmm0,   2 * SIZE(BB)
	movhpd	%xmm0,   3 * SIZE(BB)
#else
	movapd	%xmm0,   0 * SIZE(AA)
#endif

#ifdef LN
	subl	$2 * SIZE, %esi
#endif

	movsd	%xmm0,  0 * SIZE(%esi)
	movhpd	%xmm0,  1 * SIZE(%esi)

#ifndef LN
	addl	$2 * SIZE, %esi
#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_2

.L150:
	movl	M,  %ebx
	sarl	$2, %ebx	# i = (m >> 2)
	jle	.L159
	ALIGN_4

.L110:
#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

	leal	BUFFER, BB

#if defined(LN) || defined(RT)
	movl	KK, %eax
	sall	$0 + BASE_SHIFT, %eax
	leal	(BB, %eax, 2), BB
#endif	

	movapd	 0 * SIZE(BB), %xmm2
	pxor	%xmm4, %xmm4
	movapd	 0 * SIZE(AA), %xmm0
	pxor	%xmm5, %xmm5
	movapd	 8 * SIZE(BB), %xmm3
	pxor	%xmm6, %xmm6
	movapd	 8 * SIZE(AA), %xmm1
	pxor	%xmm7, %xmm7

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	sarl	$3, %eax
	je	.L112

.L111:	
	mulpd	%xmm2, %xmm0
	mulpd	 2 * SIZE(AA), %xmm2
	addpd	%xmm0, %xmm4
	movapd	 4 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm6
	movapd	 2 * SIZE(BB), %xmm2
	mulpd	%xmm2, %xmm0
	mulpd	 6 * SIZE(AA), %xmm2
	addpd	%xmm0, %xmm5
	movapd	16 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm7
	movapd	 4 * SIZE(BB), %xmm2
	mulpd	%xmm2, %xmm1
	mulpd	10 * SIZE(AA), %xmm2
	addpd	%xmm1, %xmm4
	movapd	12 * SIZE(AA), %xmm1
	addpd	%xmm2, %xmm6
	movapd	 6 * SIZE(BB), %xmm2
	mulpd	%xmm2, %xmm1
	mulpd	14 * SIZE(AA), %xmm2
	addpd	%xmm1, %xmm5
	movapd	24 * SIZE(AA), %xmm1
	addpd	%xmm2, %xmm7
	movapd	16 * SIZE(BB), %xmm2
	mulpd	%xmm3, %xmm0
	mulpd	18 * SIZE(AA), %xmm3
	addpd	%xmm0, %xmm4
	movapd	20 * SIZE(AA), %xmm0
	addpd	%xmm3, %xmm6
	movapd	10 * SIZE(BB), %xmm3
	mulpd	%xmm3, %xmm0
	mulpd	22 * SIZE(AA), %xmm3
	addpd	%xmm0, %xmm5
	movapd	32 * SIZE(AA), %xmm0
	addpd	%xmm3, %xmm7
	movapd	12 * SIZE(BB), %xmm3
	mulpd	%xmm3, %xmm1
	mulpd	26 * SIZE(AA), %xmm3
	addpd	%xmm1, %xmm4
	movapd	28 * SIZE(AA), %xmm1
	addpd	%xmm3, %xmm6
	movapd	14 * SIZE(BB), %xmm3
	mulpd	%xmm3, %xmm1
	mulpd	30 * SIZE(AA), %xmm3
	addpd	%xmm1, %xmm5
	movapd	40 * SIZE(AA), %xmm1
	addpd	%xmm3, %xmm7
	movapd	24 * SIZE(BB), %xmm3

	addl   $32 * SIZE, AA
	addl   $16 * SIZE, BB
	decl   %eax
	jne    .L111

.L112:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L114

.L113:
	mulpd	%xmm2, %xmm0
	mulpd	 2 * SIZE(AA), %xmm2
	addpd	%xmm0, %xmm4
	movapd	 4 * SIZE(AA), %xmm0
	addpd	%xmm2, %xmm6
	movapd	 2 * SIZE(BB), %xmm2

	addl	$4 * SIZE, AA		# aoffset  += 8
	addl	$2 * SIZE, BB		# boffset1 += 8
	subl	$1, %eax
	jg	.L113
	ALIGN_4

.L114:
	addpd	%xmm5, %xmm4
	addpd	%xmm7, %xmm6

#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

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

#if defined(LN) || defined(LT)
	movapd	 0 * SIZE(B), %xmm0
	movapd	 2 * SIZE(B), %xmm1
#else
	movapd	 0 * SIZE(AA), %xmm0
	movapd	 2 * SIZE(AA), %xmm1
#endif

	subpd	%xmm4, %xmm0
	subpd	%xmm6, %xmm1

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

	movapd	%xmm1, %xmm3
        unpckhpd %xmm3, %xmm3

	movsd	15 * SIZE(AA), %xmm4
	mulsd	 %xmm4, %xmm3

	movsd	14 * SIZE(AA), %xmm5
	mulsd	 %xmm3, %xmm5
	subsd	 %xmm5, %xmm1
	movsd	13 * SIZE(AA), %xmm6
	mulsd	 %xmm3, %xmm6
	subsd	 %xmm6, %xmm2
	movsd	12 * SIZE(AA), %xmm7
	mulsd	 %xmm3, %xmm7
	subsd	 %xmm7, %xmm0

	movsd	10 * SIZE(AA), %xmm4
	mulsd	 %xmm4, %xmm1

	movsd	 9 * SIZE(AA), %xmm5
	mulsd	 %xmm1, %xmm5
	subsd	 %xmm5, %xmm2
	movsd	 8 * SIZE(AA), %xmm6
	mulsd	 %xmm1, %xmm6
	subsd	 %xmm6, %xmm0

	movsd	 5 * SIZE(AA), %xmm4
	mulsd	 %xmm4, %xmm2

	movsd	 4 * SIZE(AA), %xmm5
	mulsd	 %xmm2, %xmm5
	subsd	 %xmm5, %xmm0

	movsd	 0 * SIZE(AA), %xmm4
	mulsd	 %xmm4, %xmm0

	unpcklpd %xmm2, %xmm0
	unpcklpd %xmm3, %xmm1
#endif

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

	movapd	%xmm1, %xmm3
        unpckhpd %xmm3, %xmm3

	movsd	 0 * SIZE(AA), %xmm4
	mulsd	 %xmm4, %xmm0

	movsd	 1 * SIZE(AA), %xmm5
	mulsd	 %xmm0, %xmm5
	subsd	 %xmm5, %xmm2
	movsd	 2 * SIZE(AA), %xmm6
	mulsd	 %xmm0, %xmm6
	subsd	 %xmm6, %xmm1
	movsd	 3 * SIZE(AA), %xmm7
	mulsd	 %xmm0, %xmm7
	subsd	 %xmm7, %xmm3

	movsd	 5 * SIZE(AA), %xmm4
	mulsd	 %xmm4, %xmm2

	movsd	 6 * SIZE(AA), %xmm5
	mulsd	 %xmm2, %xmm5
	subsd	 %xmm5, %xmm1
	movsd	 7 * SIZE(AA), %xmm6
	mulsd	 %xmm2, %xmm6
	subsd	 %xmm6, %xmm3

	movsd	10 * SIZE(AA), %xmm4
	mulsd	 %xmm4, %xmm1

	movsd	11 * SIZE(AA), %xmm5
	mulsd	 %xmm1, %xmm5
	subsd	 %xmm5, %xmm3

	movsd	15 * SIZE(AA), %xmm4
	mulsd	 %xmm4, %xmm3

	unpcklpd %xmm2, %xmm0
	unpcklpd %xmm3, %xmm1
#endif

#if defined(RN) || defined(RT)
	movsd	 0 * SIZE(B), %xmm4
	movhpd	 0 * SIZE(B), %xmm4
	mulpd	 %xmm4, %xmm0
	mulpd	 %xmm4, %xmm1
#endif

#if defined(LN) || defined(LT)
	movapd	%xmm0,   0 * SIZE(B)
	movapd	%xmm1,   2 * SIZE(B)

	movsd	%xmm0,   0 * SIZE(BB)
	movsd	%xmm0,   1 * SIZE(BB)
	movhpd	%xmm0,   2 * SIZE(BB)
	movhpd	%xmm0,   3 * SIZE(BB)
	movsd	%xmm1,   4 * SIZE(BB)
	movsd	%xmm1,   5 * SIZE(BB)
	movhpd	%xmm1,   6 * SIZE(BB)
	movhpd	%xmm1,   7 * SIZE(BB)
#else
	movapd	%xmm0,   0 * SIZE(AA)
	movapd	%xmm1,   2 * SIZE(AA)
#endif

#ifdef LN
	subl	$4 * SIZE, %esi
#endif

	movsd	%xmm0,  0 * SIZE(%esi)
	movhpd	%xmm0,  1 * SIZE(%esi)
	movsd	%xmm1,  2 * SIZE(%esi)
	movhpd	%xmm1,  3 * SIZE(%esi)

#ifndef LN
	addl	$4 * SIZE, %esi
#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

	BRANCH
	decl	%ebx			# i --
	jg	.L110
	ALIGN_2	

.L159:
#ifdef LN
       movl	K, %eax
       leal	(, %eax, SIZE), %eax
       leal 	(B, %eax, 1), B
#endif

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	leal	(,%eax, SIZE), %eax
	leal	(B,  %eax, 1), B
#endif

#ifdef RN
	addl	$1, KK
#endif

#ifdef RT
	subl	$1, KK
#endif
	ALIGN_2

.L999:
	movl	OLD_STACK, %esp

	EMMS

	popl	%ebx
	popl	%esi
	popl	%edi
	popl	%ebp
	ret
	ALIGN_2


	EPILOGUE