ztrsm_kernel_lt_1x2_sse2.s

Optimized GotoBLAS libraries

#ifndef CONJ
	SHUFPD_1 %xmm1, %xmm1
#endif

#if defined(LN) || defined(LT)
	movlpd	 0 * SIZE(AA), %xmm2
	movhpd	 0 * SIZE(AA), %xmm2
	movlpd	 1 * SIZE(AA), %xmm3
	movhpd	 1 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4
	pshufd	$0x4e, %xmm7, %xmm6

	xorpd	 %xmm1, %xmm4
	xorpd	 %xmm1, %xmm6

	mulpd	 %xmm2, %xmm5
	mulpd	 %xmm3, %xmm4
	mulpd	 %xmm2, %xmm7
	mulpd	 %xmm3, %xmm6

	addpd	 %xmm4, %xmm5
	addpd	 %xmm6, %xmm7
#endif

#ifdef RN
	movlpd	 0 * SIZE(B), %xmm2
	movhpd	 0 * SIZE(B), %xmm2
	movlpd	 1 * SIZE(B), %xmm3
	movhpd	 1 * SIZE(B), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4

	xorpd	 %xmm1, %xmm4

	mulpd	 %xmm2, %xmm5
	mulpd	 %xmm3, %xmm4

	addpd	 %xmm4, %xmm5

	movlpd	 2 * SIZE(B), %xmm2
	movhpd	 2 * SIZE(B), %xmm2
	movlpd	 3 * SIZE(B), %xmm3
	movhpd	 3 * SIZE(B), %xmm3

	movapd	 %xmm5, %xmm4
	pshufd	 $0x4e, %xmm5, %xmm6

	xorpd	 %xmm1, %xmm6

	mulpd	 %xmm2, %xmm4
	mulpd	 %xmm3, %xmm6

	subpd	 %xmm4, %xmm7
	subpd	 %xmm6, %xmm7

	movlpd	 6 * SIZE(B), %xmm2
	movhpd	 6 * SIZE(B), %xmm2
	movlpd	 7 * SIZE(B), %xmm3
	movhpd	 7 * SIZE(B), %xmm3

	pshufd	$0x4e, %xmm7, %xmm6

	xorpd	 %xmm1, %xmm6

	mulpd	 %xmm2, %xmm7
	mulpd	 %xmm3, %xmm6

	addpd	 %xmm6, %xmm7
#endif

#ifdef RT
	movlpd	 6 * SIZE(B), %xmm2
	movhpd	 6 * SIZE(B), %xmm2
	movlpd	 7 * SIZE(B), %xmm3
	movhpd	 7 * SIZE(B), %xmm3

	pshufd	$0x4e, %xmm7, %xmm6

	xorpd	 %xmm1, %xmm6

	mulpd	 %xmm2, %xmm7
	mulpd	 %xmm3, %xmm6

	addpd	 %xmm6, %xmm7

	movlpd	 4 * SIZE(B), %xmm2
	movhpd	 4 * SIZE(B), %xmm2
	movlpd	 5 * SIZE(B), %xmm3
	movhpd	 5 * SIZE(B), %xmm3

	movapd	 %xmm7, %xmm4
	pshufd	 $0x4e, %xmm7, %xmm6

	xorpd	 %xmm1, %xmm6

	mulpd	 %xmm2, %xmm4
	mulpd	 %xmm3, %xmm6

	subpd	 %xmm4, %xmm5
	subpd	 %xmm6, %xmm5

	movlpd	 0 * SIZE(B), %xmm2
	movhpd	 0 * SIZE(B), %xmm2
	movlpd	 1 * SIZE(B), %xmm3
	movhpd	 1 * SIZE(B), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4

	xorpd	 %xmm1, %xmm4

	mulpd	 %xmm2, %xmm5
	mulpd	 %xmm3, %xmm4

	addpd	 %xmm4, %xmm5
#endif

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

	movlpd	%xmm5,   0 * SIZE(CO1)
	movhpd	%xmm5,   1 * SIZE(CO1)

	movlpd	%xmm7,   0 * SIZE(CO1, LDC)
	movhpd	%xmm7,   1 * SIZE(CO1, LDC)

#if defined(LN) || defined(LT)
	movapd	%xmm5,   0 * SIZE(B)
	movapd	%xmm7,   2 * SIZE(B)

	movlpd	%xmm5,   0 * SIZE(BB)
	movlpd	%xmm5,   1 * SIZE(BB)
	movhpd	%xmm5,   2 * SIZE(BB)
	movhpd	%xmm5,   3 * SIZE(BB)
	movlpd	%xmm7,   4 * SIZE(BB)
	movlpd	%xmm7,   5 * SIZE(BB)
	movhpd	%xmm7,   6 * SIZE(BB)
	movhpd	%xmm7,   7 * SIZE(BB)
#else
	movapd	%xmm5,   0 * SIZE(AA)
	movapd	%xmm7,   2 * SIZE(AA)

#endif

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

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	sall	$ZBASE_SHIFT, %eax
	addl	%eax, AA
#ifdef LT
	addl	$4 * 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	$ZBASE_SHIFT, %eax
	addl	%eax, AORIG
#endif

	decl	%ebx			# i --
	jg	.L10
	ALIGN_4	

.L99:
#ifdef LN
       movl	K, %eax
       sall	$1 + ZBASE_SHIFT, %eax
       addl	%eax, B
#endif

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	sall	$1 + ZBASE_SHIFT, %eax
	addl	%eax, B
#endif

#ifdef RN
	addl	$2, KK
#endif

#ifdef RT
	subl	$2, KK
#endif

	decl	J			# j --
	jg	.L01
	ALIGN_4

.L100:
	movl	N, %eax
	andl	$1, %eax
	jle	.L500
	ALIGN_4

.L101:
#ifdef LN
	movl	OFFSET, %eax
	addl	M, %eax
	movl	%eax, KK
#endif	

	leal	BUFFER, BB

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

#if defined(LN) || defined(RT)
	movl	KK, %eax
	movl	B, BORIG
	sall	$ZBASE_SHIFT, %eax
	addl	%eax, B
	leal	(BB, %eax, 2), BB
#endif	

#if defined(LT)
	movl	OFFSET, %eax
	movl	%eax, KK
#endif

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K,  %eax
	subl	KK, %eax
#endif
	sarl	$2, %eax
	jle	.L103
	ALIGN_4

.L102:
	prefetchnta	 56 * SIZE(B)

	movlpd	 0 * SIZE(B), %xmm0
	movlpd	 1 * SIZE(B), %xmm1
	movlpd	 2 * SIZE(B), %xmm2
	movlpd	 3 * SIZE(B), %xmm3
	movlpd	 4 * SIZE(B), %xmm4
	movlpd	 5 * SIZE(B), %xmm5
	movlpd	 6 * SIZE(B), %xmm6
	movlpd	 7 * SIZE(B), %xmm7

	movlpd	%xmm0,  0 * SIZE(BB)
	movlpd	%xmm0,  1 * SIZE(BB)
	movlpd	%xmm1,  2 * SIZE(BB)
	movlpd	%xmm1,  3 * SIZE(BB)
	movlpd	%xmm2,  4 * SIZE(BB)
	movlpd	%xmm2,  5 * SIZE(BB)
	movlpd	%xmm3,  6 * SIZE(BB)
	movlpd	%xmm3,  7 * SIZE(BB)
	movlpd	%xmm4,  8 * SIZE(BB)
	movlpd	%xmm4,  9 * SIZE(BB)
	movlpd	%xmm5, 10 * SIZE(BB)
	movlpd	%xmm5, 11 * SIZE(BB)
	movlpd	%xmm6, 12 * SIZE(BB)
	movlpd	%xmm6, 13 * SIZE(BB)
	movlpd	%xmm7, 14 * SIZE(BB)
	movlpd	%xmm7, 15 * SIZE(BB)

	addl	$  8 * SIZE, B
	subl	$-16 * SIZE, BB
	decl	%eax
	jne	.L102
	ALIGN_4

.L103:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	andl	$3, %eax
	BRANCH
	jle	.L105
	ALIGN_4

.L104:
	movlpd	 0 * SIZE(B), %xmm0
	movlpd	 1 * SIZE(B), %xmm1

	movlpd	%xmm0,  0 * SIZE(BB)
	movlpd	%xmm0,  1 * SIZE(BB)
	movlpd	%xmm1,  2 * SIZE(BB)
	movlpd	%xmm1,  3 * SIZE(BB)

	addl	$2 * SIZE, B
	addl	$4 * SIZE, BB
	decl	%eax
	jne	.L104
	ALIGN_4

.L105:
#if defined(LT) || defined(RN)
	movl	A, %eax
	movl	%eax, AA
#else
	movl	A, %eax
	movl	%eax, AORIG
#endif

#ifdef RT
       subl	LDC, C
#endif

	movl	C, CO1

#ifndef RT
	addl	LDC, C
#endif

	movl	M,  %ebx
	testl	%ebx, %ebx	
	jle	.L199
	ALIGN_4

.L110:
#ifdef LN
       movl	K, %eax
       sall	$ZBASE_SHIFT, %eax
       subl	%eax, AORIG
#endif

#if defined(LN) || defined(RT)
	movl	AORIG, %eax
	movl	%eax, AA

	movl	KK, %eax
	sall	$ZBASE_SHIFT, %eax
	addl	%eax, AA
#endif

	leal	BUFFER, BB

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

	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5
	pxor	%xmm6, %xmm6
	pxor	%xmm7, %xmm7

	movapd	 0 * SIZE(AA), %xmm0
	movapd	 8 * SIZE(AA), %xmm1
	movapd	 0 * SIZE(BB), %xmm2
	movapd	 8 * SIZE(BB), %xmm3

#ifdef LN
	prefetchw -2 * SIZE(CO1)
#else
	prefetchw  2 * SIZE(CO1)
#endif

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	sarl	$3, %eax
	je	.L112
	ALIGN_4

.L111:
	PREFETCH  (PREFETCHSIZE + 0) * SIZE(AA)
	mulpd	%xmm0, %xmm2
	mulpd	 2 * SIZE(BB), %xmm0
	addpd	%xmm2, %xmm4
	movapd	 4 * SIZE(BB), %xmm2
	addpd	%xmm0, %xmm5
	movapd	 2 * SIZE(AA), %xmm0
	mulpd	%xmm0, %xmm2
	mulpd	 6 * SIZE(BB), %xmm0
	addpd	%xmm2, %xmm6
	movapd	16 * SIZE(BB), %xmm2
	addpd	%xmm0, %xmm7
	movapd	 4 * SIZE(AA), %xmm0
	mulpd	%xmm0, %xmm3
	mulpd	10 * SIZE(BB), %xmm0
	addpd	%xmm3, %xmm4
	movapd	12 * SIZE(BB), %xmm3
	addpd	%xmm0, %xmm5
	movapd	 6 * SIZE(AA), %xmm0
	mulpd	%xmm0, %xmm3
	mulpd	14 * SIZE(BB), %xmm0
	addpd	%xmm3, %xmm6
	movapd	24 * SIZE(BB), %xmm3
	addpd	%xmm0, %xmm7
	movapd	16 * SIZE(AA), %xmm0
	mulpd	%xmm1, %xmm2
	mulpd	18 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm4
	movapd	20 * SIZE(BB), %xmm2
	addpd	%xmm1, %xmm5
	movapd	10 * SIZE(AA), %xmm1
	mulpd	%xmm1, %xmm2
	mulpd	22 * SIZE(BB), %xmm1
	addpd	%xmm2, %xmm6
	movapd	32 * SIZE(BB), %xmm2
	addpd	%xmm1, %xmm7
	movapd	12 * SIZE(AA), %xmm1
	mulpd	%xmm1, %xmm3
	mulpd	26 * SIZE(BB), %xmm1
	addpd	%xmm3, %xmm4
	movapd	28 * SIZE(BB), %xmm3
	addpd	%xmm1, %xmm5
	movapd	14 * SIZE(AA), %xmm1
	mulpd	%xmm1, %xmm3
	mulpd	30 * SIZE(BB), %xmm1
	addpd	%xmm3, %xmm6
	movapd	40 * SIZE(BB), %xmm3
	addpd	%xmm1, %xmm7
	movapd	24 * SIZE(AA), %xmm1

	addl   $16 * SIZE, AA
	addl   $32 * SIZE, BB
	decl   %eax
	jne    .L111
	ALIGN_4

.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
	ALIGN_4

.L113:
	mulpd	%xmm0, %xmm2
	mulpd	 2 * SIZE(BB), %xmm0
	addpd	%xmm2, %xmm4
	movapd	 4 * SIZE(BB), %xmm2
	addpd	%xmm0, %xmm5
	movapd	 2 * SIZE(AA), %xmm0

	addl	$2 * SIZE, AA
	addl	$4 * SIZE, BB
	decl	%eax
	jg	.L113
	ALIGN_4

.L114:
#if defined(LN) || defined(RT)
	movl	KK, %eax
#ifdef LN
	subl	$1, %eax
#else
	subl	$1, %eax
#endif

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

	sall	$ZBASE_SHIFT, %eax
	addl	%eax, AA
	addl	%eax, B
	leal	(BB, %eax, 2), BB
#endif

	movapd	POSINV,  %xmm1

	addpd	%xmm6, %xmm4
	addpd	%xmm7, %xmm5

	SHUFPD_1 %xmm5, %xmm5

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(NR) || defined(NC) || defined(TR) || defined(TC)
	xorpd	%xmm1, %xmm5
#else
	xorpd	%xmm1, %xmm4
#endif

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(RR) || defined(RC) || defined(CR) || defined(CC)
	subpd	%xmm5, %xmm4
#else
	addpd	%xmm5, %xmm4
#endif

#if defined(LN) || defined(LT)
	movapd	 0 * SIZE(B), %xmm5
	subpd	%xmm4,  %xmm5
#else
	movapd	 0 * SIZE(AA), %xmm5
	subpd	%xmm4,  %xmm5
#endif

#ifndef CONJ
	SHUFPD_1 %xmm1, %xmm1
#endif

#if defined(LN) || defined(LT)
	movlpd	 0 * SIZE(AA), %xmm2
	movhpd	 0 * SIZE(AA), %xmm2
	movlpd	 1 * SIZE(AA), %xmm3
	movhpd	 1 * SIZE(AA), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4
	xorpd	 %xmm1, %xmm4

	mulpd	 %xmm2, %xmm5
	mulpd	 %xmm3, %xmm4

	addpd	 %xmm4, %xmm5
#endif

#ifdef RN
	movlpd	 0 * SIZE(B), %xmm2
	movhpd	 0 * SIZE(B), %xmm2
	movlpd	 1 * SIZE(B), %xmm3
	movhpd	 1 * SIZE(B), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4
	xorpd	 %xmm1, %xmm4

	mulpd	 %xmm2, %xmm5
	mulpd	 %xmm3, %xmm4

	addpd	 %xmm4, %xmm5
#endif

#ifdef RT
	movlpd	 0 * SIZE(B), %xmm2
	movhpd	 0 * SIZE(B), %xmm2
	movlpd	 1 * SIZE(B), %xmm3
	movhpd	 1 * SIZE(B), %xmm3

	pshufd	$0x4e, %xmm5, %xmm4
	xorpd	 %xmm1, %xmm4

	mulpd	 %xmm2, %xmm5
	mulpd	 %xmm3, %xmm4

	addpd	 %xmm4, %xmm5
#endif

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

	movlpd	%xmm5,   0 * SIZE(CO1)
	movhpd	%xmm5,   1 * SIZE(CO1)

#if defined(LN) || defined(LT)
	movapd	%xmm5,   0 * SIZE(B)

	movlpd	%xmm5,   0 * SIZE(BB)
	movlpd	%xmm5,   1 * SIZE(BB)
	movhpd	%xmm5,   2 * SIZE(BB)
	movhpd	%xmm5,   3 * SIZE(BB)
#else
	movapd	%xmm5,   0 * SIZE(AA)
#endif

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

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	sall	$ZBASE_SHIFT, %eax
	addl	%eax, AA
#ifdef LT
	addl	$2 * 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	$ZBASE_SHIFT, %eax
	addl	%eax, AORIG
#endif

	decl	%ebx			# i --
	jg	.L110
	ALIGN_4	

.L199:
#ifdef LN
       movl	K, %eax
       sall	$ZBASE_SHIFT, %eax
       addl	%eax, B
#endif

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	sall	$ZBASE_SHIFT, %eax
	addl	%eax, B
#endif

#ifdef RN
	addl	$1, KK
#endif

#ifdef RT
	subl	$1, KK
#endif
	ALIGN_4

.L500:
	movl	OLD_STACK, %esp

	EMMS
	
	popl	%ebx
	popl	%esi
	popl	%edi
	popl	%ebp
	ret

	EPILOGUE