ztrsm_kernel_rt_2x2_sse.s

Optimized GotoBLAS libraries

	pshufd	 $0x44, %xmm5, %xmm6
	pshufd	 $0x11, %xmm5, %xmm7

	pshufd	 $0xa0, %xmm2, %xmm4
	pshufd	 $0xf5, %xmm2, %xmm2

#ifndef CONJ
	xorps	 %xmm0, %xmm2
#else
	xorps	 %xmm0, %xmm4
#endif

	mulps	%xmm6,  %xmm4
	mulps	%xmm7, %xmm2
	addps	%xmm4,  %xmm2

	pshufd	 $0xee, %xmm5, %xmm6
	pshufd	 $0xbb, %xmm5, %xmm7

	pshufd	 $0xa0, %xmm2, %xmm4
	pshufd	 $0xf5, %xmm2, %xmm1

#ifndef CONJ
	xorps	 %xmm0, %xmm1
#else
	xorps	 %xmm0, %xmm4
#endif

	mulps	%xmm6,  %xmm4
	mulps	%xmm7, %xmm1
	subps	%xmm4,  %xmm3
	subps	%xmm1,  %xmm3

	movaps	 4 * SIZE(AA), %xmm5

	pshufd	 $0xee, %xmm5, %xmm6
	pshufd	 $0xbb, %xmm5, %xmm7

	pshufd	 $0xa0, %xmm3, %xmm4
	pshufd	 $0xf5, %xmm3, %xmm3

#ifndef CONJ
	xorps	 %xmm0, %xmm3
#else
	xorps	 %xmm0, %xmm4
#endif

	mulps	%xmm6,  %xmm4
	mulps	%xmm7,  %xmm3
	addps	%xmm4,  %xmm3
#endif

#ifdef RN
	movaps	 0 * SIZE(B), %xmm4

	pshufd	 $0x44, %xmm4, %xmm6
	pshufd	 $0x11, %xmm4, %xmm7

	pshufd	 $0xa0, %xmm1, %xmm3
	pshufd	 $0xf5, %xmm1, %xmm1

#ifndef CONJ
	xorps	 %xmm0, %xmm1
#else
	xorps	 %xmm0, %xmm3
#endif

	mulps	%xmm6,  %xmm3
	mulps	%xmm7,  %xmm1

	addps	%xmm3,  %xmm1

	pshufd	 $0xee, %xmm4, %xmm6
	pshufd	 $0xbb, %xmm4, %xmm7

	pshufd	 $0xa0, %xmm1, %xmm3
	pshufd	 $0xf5, %xmm1, %xmm2

#ifndef CONJ
	xorps	 %xmm0, %xmm2
#else
	xorps	 %xmm0, %xmm3
#endif

	mulps	%xmm6,  %xmm3
	mulps	%xmm7,  %xmm2

	subps	%xmm3,  %xmm5
	subps	%xmm2,  %xmm5

	movaps	 4 * SIZE(B), %xmm4

	pshufd	 $0xee, %xmm4, %xmm6
	pshufd	 $0xbb, %xmm4, %xmm7

	pshufd	 $0xa0, %xmm5, %xmm3
	pshufd	 $0xf5, %xmm5, %xmm5

#ifndef CONJ
	xorps	 %xmm0, %xmm5
#else
	xorps	 %xmm0, %xmm3
#endif

	mulps	%xmm6,  %xmm3
	mulps	%xmm7,  %xmm5

	addps	%xmm3,  %xmm5
#endif

#ifdef RT
	movaps	 4 * SIZE(B), %xmm4

	pshufd	 $0xee, %xmm4, %xmm6
	pshufd	 $0xbb, %xmm4, %xmm7

	pshufd	 $0xa0, %xmm5, %xmm3
	pshufd	 $0xf5, %xmm5, %xmm5

#ifndef CONJ
	xorps	 %xmm0, %xmm5
#else
	xorps	 %xmm0, %xmm3
#endif

	mulps	%xmm6,  %xmm3
	mulps	%xmm7,  %xmm5

	addps	%xmm3,  %xmm5

	pshufd	 $0x44, %xmm4, %xmm6
	pshufd	 $0x11, %xmm4, %xmm7

	pshufd	 $0xa0, %xmm5, %xmm3
	pshufd	 $0xf5, %xmm5, %xmm2

#ifndef CONJ
	xorps	 %xmm0, %xmm2
#else
	xorps	 %xmm0, %xmm3
#endif

	mulps	%xmm6,  %xmm3
	mulps	%xmm7,  %xmm2

	subps	%xmm3,  %xmm1
	subps	%xmm2,  %xmm1

	movaps	 0 * SIZE(B), %xmm4

	pshufd	 $0x44, %xmm4, %xmm6
	pshufd	 $0x11, %xmm4, %xmm7

	pshufd	 $0xa0, %xmm1, %xmm3
	pshufd	 $0xf5, %xmm1, %xmm1

#ifndef CONJ
	xorps	 %xmm0, %xmm1
#else
	xorps	 %xmm0, %xmm3
#endif

	mulps	%xmm6,  %xmm3
	mulps	%xmm7,  %xmm1

	addps	%xmm3,  %xmm1
#endif

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

#if defined(LN) || defined(LT)
	movaps	%xmm2,   0 * SIZE(B)
	movaps	%xmm3,   4 * SIZE(B)

	pshufd	$0x00, %xmm2, %xmm0
	pshufd	$0x55, %xmm2, %xmm1
	pshufd	$0xaa, %xmm2, %xmm4
	pshufd	$0xff, %xmm2, %xmm5

	movaps	%xmm0,   0 * SIZE(BB)
	movaps	%xmm1,   4 * SIZE(BB)
	movaps	%xmm4,   8 * SIZE(BB)
	movaps	%xmm5,  12 * SIZE(BB)

	pshufd	$0x00, %xmm3, %xmm0
	pshufd	$0x55, %xmm3, %xmm1
	pshufd	$0xaa, %xmm3, %xmm4
	pshufd	$0xff, %xmm3, %xmm5

	movaps	%xmm0,  16 * SIZE(BB)
	movaps	%xmm1,  20 * SIZE(BB)
	movaps	%xmm4,  24 * SIZE(BB)
	movaps	%xmm5,  28 * SIZE(BB)

	movlps	%xmm2,   0 * SIZE(CO1)
	movlps	%xmm3,   2 * SIZE(CO1)
	movhps	%xmm2,   0 * SIZE(CO1, LDC)
	movhps	%xmm3,   2 * SIZE(CO1, LDC)
#else
	movaps	%xmm1,   0 * SIZE(AA)
	movaps	%xmm5,   4 * SIZE(AA)

	movlps	%xmm1,   0 * SIZE(CO1)
	movhps	%xmm1,   2 * SIZE(CO1)

	movlps	%xmm5,   0 * SIZE(CO1, LDC)
	movhps	%xmm5,   2 * SIZE(CO1, LDC)
#endif

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

#if defined(LT) || defined(RN)
	movl	K,  %eax
	subl	KK, %eax
	sall	$1 + ZBASE_SHIFT, %eax
	addl	%eax, AA
#ifdef LT
	addl	$8 * 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 + ZBASE_SHIFT, %eax
	addl	%eax, AORIG
#endif

	decl	%ebx
	jg	.L10
	ALIGN_4

.L30:
	movl	M,  %ebx
	andl	$1, %ebx
	jle	.L99
	ALIGN_4

.L40:
#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	# boffset1 = boffset

#if defined(LN) || defined(RT)
	movl	KK, %eax
	sall	$3 + ZBASE_SHIFT, %eax
	addl	%eax, BB
#endif	

	XORPS	%xmm4, %xmm4
	XORPS	%xmm5, %xmm5
	XORPS	%xmm6, %xmm6
	XORPS	%xmm7, %xmm7

	movlps	 0 * SIZE(AA), %xmm0
	movlps	 8 * SIZE(AA), %xmm1
	movaps	 0 * SIZE(BB), %xmm2
	movaps	16 * SIZE(BB), %xmm3

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	sarl	$3, %eax
	je	.L42
	ALIGN_4

.L41:
	mulps	%xmm0, %xmm2
	prefetcht1	(PREFETCHSIZE +  0) * SIZE(AA)
	addps	%xmm2, %xmm4
	movaps	 4 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm2, %xmm5
	movaps	 8 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	mulps	12 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm6
	movaps	32 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm7
	movlps	 2 * SIZE(AA), %xmm0
	mulps	%xmm0, %xmm3
	addps	%xmm3, %xmm4
	movaps	20 * SIZE(BB), %xmm3
	mulps	%xmm0, %xmm3
	addps	%xmm3, %xmm5
	movaps	24 * SIZE(BB), %xmm3
	mulps	%xmm0, %xmm3
	mulps	28 * SIZE(BB), %xmm0
	addps	%xmm3, %xmm6
	movaps	48 * SIZE(BB), %xmm3
	addps	%xmm0, %xmm7
	movlps	 4 * SIZE(AA), %xmm0
	mulps	%xmm0, %xmm2
	addps	%xmm2, %xmm4
	movaps	36 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm2, %xmm5
	movaps	40 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	mulps	44 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm6
	movaps	64 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm7
	movlps	 6 * SIZE(AA), %xmm0
	mulps	%xmm0, %xmm3
	addps	%xmm3, %xmm4
	movaps	52 * SIZE(BB), %xmm3
	mulps	%xmm0, %xmm3
	addps	%xmm3, %xmm5
	movaps	56 * SIZE(BB), %xmm3
	mulps	%xmm0, %xmm3
	mulps	60 * SIZE(BB), %xmm0
	addps	%xmm3, %xmm6
	movaps	80 * SIZE(BB), %xmm3
	addps	%xmm0, %xmm7
	movlps	16 * SIZE(AA), %xmm0
	mulps	%xmm1, %xmm2
#if defined(OPTERON) || defined(BARCELONA)
	prefetcht1     (PREFETCHSIZE + 16) * SIZE(AA)
#endif
	addps	%xmm2, %xmm4
	movaps	68 * SIZE(BB), %xmm2
	mulps	%xmm1, %xmm2
	addps	%xmm2, %xmm5
	movaps	72 * SIZE(BB), %xmm2
	mulps	%xmm1, %xmm2
	mulps	76 * SIZE(BB), %xmm1
	addps	%xmm2, %xmm6
	movaps	96 * SIZE(BB), %xmm2
	addps	%xmm1, %xmm7
	movlps	10 * SIZE(AA), %xmm1
	mulps	%xmm1, %xmm3
	addps	%xmm3, %xmm4
	movaps	84 * SIZE(BB), %xmm3
	mulps	%xmm1, %xmm3
	addps	%xmm3, %xmm5
	movaps	88 * SIZE(BB), %xmm3
	mulps	%xmm1, %xmm3
	mulps	92 * SIZE(BB), %xmm1
	addps	%xmm3, %xmm6
	movaps	112 * SIZE(BB), %xmm3
	addps	%xmm1, %xmm7
	movlps	12 * SIZE(AA), %xmm1
	mulps	%xmm1, %xmm2
	addps	%xmm2, %xmm4
	movaps	100 * SIZE(BB), %xmm2
	mulps	%xmm1, %xmm2
	addps	%xmm2, %xmm5
	movaps	104 * SIZE(BB), %xmm2
	mulps	%xmm1, %xmm2
	mulps	108 * SIZE(BB), %xmm1
	addps	%xmm2, %xmm6
	movaps	128 * SIZE(BB), %xmm2
	addps	%xmm1, %xmm7
	movlps	14 * SIZE(AA), %xmm1
	mulps	%xmm1, %xmm3
	addps	%xmm3, %xmm4
	movaps	116 * SIZE(BB), %xmm3
	mulps	%xmm1, %xmm3
	addps	%xmm3, %xmm5
	movaps	120 * SIZE(BB), %xmm3
	mulps	%xmm1, %xmm3
	mulps	124 * SIZE(BB), %xmm1
	addps	%xmm3, %xmm6
	movaps	144 * SIZE(BB), %xmm3
	addps	%xmm1, %xmm7
	movlps	 24 * SIZE(AA), %xmm1
	addl	$ 16 * SIZE, AA
	addl	$128 * SIZE, BB
	decl	%eax
	jne	.L41
	ALIGN_4
	
.L42:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K,  %eax
	subl	KK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L44
	ALIGN_4

.L43:
	mulps	%xmm0, %xmm2
	addps	%xmm2, %xmm4
	movaps	 4 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	addps	%xmm2, %xmm5
	movaps	 8 * SIZE(BB), %xmm2
	mulps	%xmm0, %xmm2
	mulps	12 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm6
	movaps	16 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm7
	movlps	 2 * SIZE(AA), %xmm0

	addl	$ 2 * SIZE, AA
	addl	$16 * SIZE, BB
	decl	%eax
	jg	.L43
	ALIGN_4

.L44:
	movaps	POSINV,  %xmm0

	shufps	$0xb1, %xmm5, %xmm5
	shufps	$0xb1, %xmm7, %xmm7

#if defined(LN) || defined(LT)
#ifndef CONJ
	xorps	 %xmm0, %xmm5
	xorps	 %xmm0, %xmm7
#else
	xorps	 %xmm0, %xmm4
	xorps	 %xmm0, %xmm6
#endif
#else
	xorps	 %xmm0, %xmm5
	xorps	 %xmm0, %xmm7
#endif

	addps	%xmm5, %xmm4
	addps	%xmm7, %xmm6

#if defined(LN) || defined(RT)
	movl	KK, %eax
#ifdef LN
	subl	$1, %eax
#else
	subl	$2, %eax
#endif

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

	sall	$ZBASE_SHIFT, %eax
	leal	(AA, %eax, 1), AA
	leal	(B,  %eax, 2), B
	leal	(BB, %eax, 8), BB
#endif

#if defined(LN) || defined(LT)
	unpcklpd %xmm6, %xmm4

	movaps	 0 * SIZE(B), %xmm2

	subps	%xmm4,  %xmm2
#else
	movsd	 0 * SIZE(AA), %xmm1
	movsd	 2 * SIZE(AA), %xmm5

	subps	%xmm4,  %xmm1
	subps	%xmm6,  %xmm5
#endif

#if defined(LN) || defined(LT)
	movaps	 0 * SIZE(AA), %xmm5

	pshufd	 $0x44, %xmm5, %xmm6
	pshufd	 $0x11, %xmm5, %xmm7

	pshufd	 $0xa0, %xmm2, %xmm4
	pshufd	 $0xf5, %xmm2, %xmm2

#ifndef CONJ
	xorps	 %xmm0, %xmm2
#else
	xorps	 %xmm0, %xmm4
#endif

	mulps	%xmm6,  %xmm4
	mulps	%xmm7,  %xmm2
	addps	%xmm4,  %xmm2
#endif

#ifdef RN
	movaps	 0 * SIZE(B), %xmm4

	pshufd	 $0x44, %xmm4, %xmm6
	pshufd	 $0x11, %xmm4, %xmm7

	pshufd	 $0xa0, %xmm1, %xmm3
	pshufd	 $0xf5, %xmm1, %xmm1

#ifndef CONJ
	xorps	 %xmm0, %xmm1
#else
	xorps	 %xmm0, %xmm3
#endif

	mulps	%xmm6,  %xmm3
	mulps	%xmm7,  %xmm1

	addps	%xmm3,  %xmm1

	pshufd	 $0xee, %xmm4, %xmm6
	pshufd	 $0xbb, %xmm4, %xmm7

	pshufd	 $0xa0, %xmm1, %xmm3
	pshufd	 $0xf5, %xmm1, %xmm2

#ifndef CONJ
	xorps	 %xmm0, %xmm2
#else
	xorps	 %xmm0, %xmm3
#endif

	mulps	%xmm6,  %xmm3
	mulps	%xmm7,  %xmm2

	subps	%xmm3,  %xmm5
	subps	%xmm2,  %xmm5

	movaps	 4 * SIZE(B), %xmm4

	pshufd	 $0xee, %xmm4, %xmm6
	pshufd	 $0xbb, %xmm4, %xmm7

	pshufd	 $0xa0, %xmm5, %xmm3
	pshufd	 $0xf5, %xmm5, %xmm5

#ifndef CONJ
	xorps	 %xmm0, %xmm5
#else
	xorps	 %xmm0, %xmm3
#endif

	mulps	%xmm6,  %xmm3
	mulps	%xmm7,  %xmm5

	addps	%xmm3,  %xmm5
#endif

#ifdef RT
	movaps	 4 * SIZE(B), %xmm4

	pshufd	 $0xee, %xmm4, %xmm6
	pshufd	 $0xbb, %xmm4, %xmm7

	pshufd	 $0xa0, %xmm5, %xmm3
	pshufd	 $0xf5, %xmm5, %xmm5

#ifndef CONJ
	xorps	 %xmm0, %xmm5
#else
	xorps	 %xmm0, %xmm3
#endif

	mulps	%xmm6,  %xmm3
	mulps	%xmm7,  %xmm5

	addps	%xmm3,  %xmm5

	pshufd	 $0x44, %xmm4, %xmm6
	pshufd	 $0x11, %xmm4, %xmm7

	pshufd	 $0xa0, %xmm5, %xmm3
	pshufd	 $0xf5, %xmm5, %xmm2

#ifndef CONJ
	xorps	 %xmm0, %xmm2
#else
	xorps	 %xmm0, %xmm3
#endif

	mulps	%xmm6,  %xmm3
	mulps	%xmm7,  %xmm2

	subps	%xmm3,  %xmm1
	subps	%xmm2,  %xmm1

	movaps	 0 * SIZE(B), %xmm4

	pshufd	 $0x44, %xmm4, %xmm6
	pshufd	 $0x11, %xmm4, %xmm7

	pshufd	 $0xa0, %xmm1, %xmm3
	pshufd	 $0xf5, %xmm1, %xmm1

#ifndef CONJ
	xorps	 %xmm0, %xmm1
#else
	xorps	 %xmm0, %xmm3
#endif

	mulps	%xmm6,  %xmm3
	mulps	%xmm7,  %xmm1

	addps	%xmm3,  %xmm1
#endif

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

#if defined(LN) || defined(LT)
	movaps	%xmm2,   0 * SIZE(B)

	pshufd	$0x00, %xmm2, %xmm0
	pshufd	$0x55, %xmm2, %xmm1
	pshufd	$0xaa, %xmm2, %xmm4
	pshufd	$0xff, %xmm2, %xmm5

	movaps	%xmm0,   0 * SIZE(BB)
	movaps	%xmm1,   4 * SIZE(BB)
	movaps	%xmm4,   8 * SIZE(BB)
	movaps	%xmm5,  12 * SIZE(BB)

	movlps	%xmm2,   0 * SIZE(CO1)
	movhps	%xmm2,   0 * SIZE(CO1, LDC)
#else
	movlps	%xmm1,   0 * SIZE(AA)
	movlps	%xmm5,   2 * SIZE(AA)

	movlps	%xmm1,   0 * SIZE(CO1)
	movlps	%xmm5,   0 * SIZE(CO1, LDC)
#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
	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


.L999:
	EMMS

	movl	OLD_STACK, %esp
	popl	%ebx
	popl	%esi
	popl	%edi
	popl	%ebp
	ret

	EPILOGUE