qtrsm_kernel_rt_2x2.s

Optimized GotoBLAS libraries

#endif	

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

	movq	M,  I
	sarq	$1, I
	je	.L20
	ALIGN_4

.L11:
#ifdef LN
       movq	K, %rax
       salq	$1 + BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

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

	fldz
	fldz
	fldz
	fldz

#if   defined(HAVE_3DNOW)
	prefetchw	2 * SIZE(CO)
 	prefetchw	2 * SIZE(CO, LDC, 1)
#elif defined(HAVE_SSE)
	prefetchnta	2 * SIZE(CO)
 	prefetchnta	2 * SIZE(CO, LDC, 1)
#endif

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K,  %rax
	subq	KK, %rax
#endif
	sarq	$2, %rax
 	je	.L15
	ALIGN_4

.L12:
	PREFETCH	(PREFETCHSIZE + 0) * SIZE(AO)

	FLD	 -8 * SIZE(AO)

	FLD	 -8 * SIZE(BO)
	fld	 %st(1)
	fmul	 %st(1), %st
	faddp	 %st, %st(3)

	FLD	 -7 * SIZE(BO)
	fmul	 %st, %st(2)
	 
	FLD	 -7 * SIZE(AO)
	fmul	 %st, %st(2)
	fmulp	 %st, %st(1)

	faddp	 %st, %st(6)
	faddp	 %st, %st(4)
	faddp	 %st, %st(2)

	FLD	 -6 * SIZE(AO)

	FLD	 -6 * SIZE(BO)
	fld	 %st(1)
	fmul	 %st(1), %st
	faddp	 %st, %st(3)

	FLD	 -5 * SIZE(BO)
	fmul	 %st, %st(2)
	 
	FLD	 -5 * SIZE(AO)
	fmul	 %st, %st(2)
	fmulp	 %st, %st(1)

	faddp	 %st, %st(6)
	faddp	 %st, %st(4)
	faddp	 %st, %st(2)

	PREFETCH	(PREFETCHSIZE + 4) * SIZE(AO)

	FLD	 -4 * SIZE(AO)

	FLD	 -4 * SIZE(BO)
	fld	 %st(1)
	fmul	 %st(1), %st
	faddp	 %st, %st(3)

	FLD	 -3 * SIZE(BO)
	fmul	 %st, %st(2)
	 
	FLD	 -3 * SIZE(AO)
	fmul	 %st, %st(2)
	fmulp	 %st, %st(1)

	faddp	 %st, %st(6)
	faddp	 %st, %st(4)
	faddp	 %st, %st(2)

	FLD	 -2 * SIZE(AO)

	FLD	 -2 * SIZE(BO)
	fld	 %st(1)
	fmul	 %st(1), %st
	faddp	 %st, %st(3)

	FLD	 -1 * SIZE(BO)
	fmul	 %st, %st(2)
	 
	FLD	 -1 * SIZE(AO)
	fmul	 %st, %st(2)
	fmulp	 %st, %st(1)

	faddp	 %st, %st(6)
	faddp	 %st, %st(4)
	faddp	 %st, %st(2)

	addq	$8 * SIZE,AO
	addq	$8 * SIZE,BO

	decq	%rax
	jne	.L12
	ALIGN_4

.L15:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K,  %rax
	subq	KK, %rax
#endif
	and	$3,  %rax
	je	.L18
	ALIGN_4

.L16:
	FLD	 -8 * SIZE(AO)

	FLD	 -8 * SIZE(BO)
	fld	 %st(1)
	fmul	 %st(1), %st
	faddp	 %st, %st(3)

	FLD	 -7 * SIZE(BO)
	fmul	 %st, %st(2)
	 
	FLD	 -7 * SIZE(AO)
	fmul	 %st, %st(2)
	fmulp	 %st, %st(1)

	faddp	 %st, %st(6)
	faddp	 %st, %st(4)
	faddp	 %st, %st(2)

	addq	$2 * SIZE,AO
	addq	$2 * SIZE,BO

	decq	%rax
	jne	 .L16
	ALIGN_4

.L18:
#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$2, %rax
#else
	subq	$2, %rax
#endif

	salq	$BASE_SHIFT, %rax

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

#if defined(LN) || defined(LT)
	FLD	-8 * SIZE(BO)
	fsubp	%st, %st(1)
	FLD	-7 * SIZE(BO)
	fsubp	%st, %st(2)
	FLD	-6 * SIZE(BO)
	fsubp	%st, %st(3)
	FLD	-5 * SIZE(BO)
	fsubp	%st, %st(4)
#else
	FLD	-8 * SIZE(AO)
	fsubp	%st, %st(1)
	FLD	-7 * SIZE(AO)
	fsubp	%st, %st(3)
	FLD	-6 * SIZE(AO)
	fsubp	%st, %st(2)
	FLD	-5 * SIZE(AO)
	fsubp	%st, %st(4)
#endif

#ifdef LN
       FLD	-5 * SIZE(AO)
       fmul	%st, %st(3)
       fmulp	%st, %st(4)

       FLD	-6 * SIZE(AO)
       fmul	%st(3), %st
       FLD	-6 * SIZE(AO)
       fmul	%st(5), %st

       fsubrp	%st, %st(3)
       fsubrp	%st, %st(1)

       FLD	-8 * SIZE(AO)
       fmul	%st, %st(1)
       fmulp	%st, %st(2)
#endif

#ifdef LT
       FLD	-8 * SIZE(AO)
       fmul	%st, %st(1)
       fmulp	%st, %st(2)

       FLD	-7 * SIZE(AO)
       fmul	%st(1), %st
       FLD	-7 * SIZE(AO)
       fmul	%st(3), %st

       fsubrp	%st, %st(5)
       fsubrp	%st, %st(3)

       FLD	-5 * SIZE(AO)
       fmul	%st, %st(3)
       fmulp	%st, %st(4)
#endif

#ifdef RN
       FLD	-8 * SIZE(BO)
       fmul	%st, %st(1)
       fmulp	%st, %st(3)

       FLD	-7 * SIZE(BO)
       fmul	%st(1), %st
       FLD	-7 * SIZE(BO)
       fmul	%st(4), %st

       fsubrp	%st, %st(5)
       fsubrp	%st, %st(2)

       FLD	-5 * SIZE(BO)
       fmul	%st, %st(2)
       fmulp	%st, %st(4)
#endif

#ifdef RT
       FLD	-5 * SIZE(BO)
       fmul	%st, %st(2)
       fmulp	%st, %st(4)

       FLD	-6 * SIZE(BO)
       fmul	%st(2), %st
       FLD	-6 * SIZE(BO)
       fmul	%st(5), %st

       fsubrp	%st, %st(4)
       fsubrp	%st, %st(1)

       FLD	-8 * SIZE(BO)
       fmul	%st, %st(1)
       fmulp	%st, %st(3)
#endif

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

#if defined(LN) || defined(LT)
	fld	%st
	FST	-8 * SIZE(BO)
	fxch	%st(1)
	fld	%st
	FST	-7 * SIZE(BO)
	fxch	%st(2)
	fld	%st
	FST	-6 * SIZE(BO)
	fxch	%st(3)
	fld	%st
	FST	-5 * SIZE(BO)

	FST	1 * SIZE(CO, LDC)
	FST	0 * SIZE(CO)
	FST	0 * SIZE(CO, LDC)
	FST	1 * SIZE(CO)
#else
	fld	%st
	FST	-8 * SIZE(AO)
	fxch	%st(2)
	fld	%st
	FST	-7 * SIZE(AO)
	fxch	%st(1)
	fld	%st
	FST	-6 * SIZE(AO)
	fxch	%st(3)
	fld	%st
	FST	-5 * SIZE(AO)

	FST	1 * SIZE(CO, LDC)
	FST	1 * SIZE(CO)
	FST	0 * SIZE(CO)
	FST	0 * SIZE(CO, LDC)
#endif

#ifndef LN
	addq	$2 * SIZE, CO
#endif

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	salq	$BASE_SHIFT, %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

	decq	I
	jne	.L11
	ALIGN_4

.L20:
	movq	 M, %rax
	andq	$1, %rax
	je	.L29
	ALIGN_4

.L21:
#ifdef LN
       movq	K, %rax
       salq	$0 + BASE_SHIFT, %rax
       subq	%rax, AORIG
#endif

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

	fldz
	fldz

#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K,  %rax
	subq	KK, %rax
#endif
	sarq	$2, %rax
 	je	.L25
	ALIGN_4

.L22:
	PREFETCH	(PREFETCHSIZE + 0) * SIZE(AO)

	FLD	 -8 * SIZE(AO)

	FLD	 -8 * SIZE(BO)
	fmul	 %st(1), %st
	faddp	 %st, %st(2)

	FLD	 -7 * SIZE(BO)
	fmulp	 %st, %st(1)
	faddp	 %st, %st(2)

	FLD	 -7 * SIZE(AO)

	FLD	 -6 * SIZE(BO)
	fmul	 %st(1), %st
	faddp	 %st, %st(2)

	FLD	 -5 * SIZE(BO)
	fmulp	 %st, %st(1)
	faddp	 %st, %st(2)

	FLD	 -6 * SIZE(AO)

	FLD	 -4 * SIZE(BO)
	fmul	 %st(1), %st
	faddp	 %st, %st(2)

	FLD	 -3 * SIZE(BO)
	fmulp	 %st, %st(1)
	faddp	 %st, %st(2)

	FLD	 -5 * SIZE(AO)

	FLD	 -2 * SIZE(BO)
	fmul	 %st(1), %st
	faddp	 %st, %st(2)

	FLD	 -1 * SIZE(BO)
	fmulp	 %st, %st(1)
	faddp	 %st, %st(2)

	addq	$4 * SIZE,AO
	addq	$8 * SIZE,BO

	decq	%rax
	jne	.L22
	ALIGN_4

.L25:
#if defined(LT) || defined(RN)
	movq	KK, %rax
#else
	movq	K,  %rax
	subq	KK, %rax
#endif
	and	$3,  %rax
	je	.L28
	ALIGN_4

.L26:
	FLD	 -8 * SIZE(AO)

	FLD	 -8 * SIZE(BO)
	fmul	 %st(1), %st
	faddp	 %st, %st(2)

	FLD	 -7 * SIZE(BO)
	fmulp	 %st, %st(1)
	faddp	 %st, %st(2)

	addq	$1 * SIZE,AO
	addq	$2 * SIZE,BO

	decq	%rax
	jne	 .L26
	ALIGN_4

.L28:
#if defined(LN) || defined(RT)
	movq	KK, %rax
#ifdef LN
	subq	$1, %rax
#else
	subq	$2, %rax
#endif

	salq	$BASE_SHIFT, %rax

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

#if defined(LN) || defined(LT)
	FLD	-8 * SIZE(BO)
	fsubp	%st, %st(1)
	FLD	-7 * SIZE(BO)
	fsubp	%st, %st(2)
#else
	FLD	-8 * SIZE(AO)
	fsubp	%st, %st(1)
	FLD	-7 * SIZE(AO)
	fsubp	%st, %st(2)
#endif

#if defined(LN) || defined(LT)
       FLD	-8 * SIZE(AO)
       fmul	%st, %st(1)
       fmulp	%st, %st(2)
#endif

#ifdef RN
       FLD	-8 * SIZE(BO)
       fmulp	%st, %st(1)

       FLD	-7 * SIZE(BO)
       fmul	%st(1), %st

       fsubrp	%st, %st(2)

       FLD	-5 * SIZE(BO)
       fmulp	%st, %st(2)
#endif

#ifdef RT
       FLD	-5 * SIZE(BO)
       fmulp	%st, %st(2)

       FLD	-6 * SIZE(BO)
       fmul	%st(2), %st

       fsubrp	%st, %st(1)

       FLD	-8 * SIZE(BO)
       fmulp	%st, %st(1)
#endif

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

#if defined(LN) || defined(LT)
	fld	%st
	FST	-8 * SIZE(BO)
	fxch	%st(1)
	fld	%st
	FST	-7 * SIZE(BO)
#else
	fld	%st
	FST	-8 * SIZE(AO)
	fxch	%st(1)
	fld	%st
	FST	-7 * SIZE(AO)
#endif

	FST	0 * SIZE(CO, LDC)
	FST	0 * SIZE(CO)

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

#if defined(LT) || defined(RN)
	movq	K,  %rax
	subq	KK, %rax
	salq	$BASE_SHIFT, %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

.L29:
#ifdef LN
       movq	K, %rax
       salq	$BASE_SHIFT, %rax
       leaq	(B, %rax, 2), B
#endif

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

#ifdef RN
	addq	$2, KK
#endif

#ifdef RT
	subq	$2, KK
#endif

	decq	J
	jne	.L01
	ALIGN_4

.L999:
	movq	  0(%rsp), %rbx
	movq	  8(%rsp), %rbp
	movq	 16(%rsp), %r12
	movq	 24(%rsp), %r13
	movq	 32(%rsp), %r14
	movq	 40(%rsp), %r15
	addq	$STACKSIZE, %rsp
	ret

	EPILOGUE