gemm_kernel_8x4_core2.s

Optimized GotoBLAS libraries

	pxor	%xmm13, %xmm13

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK	
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$8, %rax
#else
	addq	$2, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle	.L65
	ALIGN_4

.L61:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	movaps	-32 * SIZE(AO), %xmm0
	movaps	-28 * SIZE(AO), %xmm1

	movaps	-32 * SIZE(BO), %xmm2
	movaps	 %xmm2, %xmm3
	movaps	-28 * SIZE(BO), %xmm4
	movaps	 %xmm4, %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm1, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm1, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm12
	addps	%xmm4, %xmm9
	addps	%xmm5, %xmm13

	movaps	-24 * SIZE(AO), %xmm0
	movaps	-20 * SIZE(AO), %xmm1

	movaps	-24 * SIZE(BO), %xmm2
	movaps	 %xmm2, %xmm3
	movaps	-20 * SIZE(BO), %xmm4
	movaps	 %xmm4, %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm1, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm1, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm12
	addps	%xmm4, %xmm9
	addps	%xmm5, %xmm13

	movaps	-16 * SIZE(AO), %xmm0
	movaps	-12 * SIZE(AO), %xmm1

	movaps	-16 * SIZE(BO), %xmm2
	movaps	 %xmm2, %xmm3
	movaps	-12 * SIZE(BO), %xmm4
	movaps	 %xmm4, %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm1, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm1, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm12
	addps	%xmm4, %xmm9
	addps	%xmm5, %xmm13

	movaps	 -8 * SIZE(AO), %xmm0
	movaps	 -4 * SIZE(AO), %xmm1

	movaps	 -8 * SIZE(BO), %xmm2
	movaps	 %xmm2, %xmm3
	movaps	 -4 * SIZE(BO), %xmm4
	movaps	 %xmm4, %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm1, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm1, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm12
	addps	%xmm4, %xmm9
	addps	%xmm5, %xmm13

	subq	$-32 * SIZE, AO
	subq	$-32 * SIZE, BO
	subq	$1, %rax
	jg	.L61
	ALIGN_4

.L65:
	movaps	ALPHA, %xmm7

#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L68
	ALIGN_4

.L66:
	movaps	-32 * SIZE(AO), %xmm0
	movaps	-28 * SIZE(AO), %xmm1

	movaps	-32 * SIZE(BO), %xmm2
	movaps	 %xmm2, %xmm3
	movaps	-28 * SIZE(BO), %xmm4
	movaps	 %xmm4, %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm1, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm1, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm12
	addps	%xmm4, %xmm9
	addps	%xmm5, %xmm13

	addq	$8 * SIZE, AO
	addq	$8 * SIZE, BO
	subq	$1, %rax
	jg	.L66
	ALIGN_4

.L68:
#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhps	2 * SIZE(CO1), %xmm0
	movsd	4 * SIZE(CO1), %xmm1
	movhps	6 * SIZE(CO1), %xmm1
	movsd	0 * SIZE(CO2), %xmm2
	movhps	2 * SIZE(CO2), %xmm2
	movsd	4 * SIZE(CO2), %xmm3
	movhps	6 * SIZE(CO2), %xmm3
#endif

	mulps	%xmm7, %xmm8
	mulps	%xmm7, %xmm9
	mulps	%xmm7, %xmm12
	mulps	%xmm7, %xmm13

#ifndef TRMMKERNEL
	addps	%xmm0,  %xmm8
	addps	%xmm1,  %xmm12
	addps	%xmm2, %xmm9
	addps	%xmm3, %xmm13
#endif

	movlps	%xmm8, 0 * SIZE(CO1)
	movhps	%xmm8, 2 * SIZE(CO1)
	movlps	%xmm12, 4 * SIZE(CO1)
	movhps	%xmm12, 6 * SIZE(CO1)
	movlps	%xmm9, 0 * SIZE(CO2)
	movhps	%xmm9, 2 * SIZE(CO2)
	movlps	%xmm13, 4 * SIZE(CO2)
	movhps	%xmm13, 6 * SIZE(CO2)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax,    8), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 4), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$8, KK
#endif

	addq	$8 * SIZE, CO1
	addq	$8 * SIZE, CO2
	subq	$1, I
	jg	.L60
	ALIGN_4	

.L70:
	testq	$4, M
	jle	.L80

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leaq	32 * SIZE + BUFFER, BO
#else
	leaq	32 * SIZE + BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax,   8), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 4), BO
#endif	

	pxor	%xmm8, %xmm8
	pxor	%xmm9, %xmm9
	pxor	%xmm10, %xmm10
	pxor	%xmm11, %xmm11

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK	
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$4, %rax
#else
	addq	$2, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle	.L75
	ALIGN_4

.L71:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	movaps	-32 * SIZE(AO), %xmm0
	movaps	-28 * SIZE(AO), %xmm1
	movaps	-32 * SIZE(BO), %xmm2
	movaps	-28 * SIZE(BO), %xmm3
	movaps	-24 * SIZE(BO), %xmm4
	movaps	-20 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm1, %xmm4
	mulps	%xmm1, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	movaps	-24 * SIZE(AO), %xmm0
	movaps	-20 * SIZE(AO), %xmm1
	movaps	-16 * SIZE(BO), %xmm2
	movaps	-12 * SIZE(BO), %xmm3
	movaps	 -8 * SIZE(BO), %xmm4
	movaps	 -4 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm1, %xmm4
	mulps	%xmm1, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	subq	$-16 * SIZE, AO
	subq	$-32 * SIZE, BO
	subq	$1, %rax
	jg	.L71
	ALIGN_4

.L75:
	movaps	ALPHA, %xmm7

#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L78
	ALIGN_4

.L76:
	movaps	-32 * SIZE(AO), %xmm0
	movaps	-32 * SIZE(BO), %xmm2
	movaps	-28 * SIZE(BO), %xmm3

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9

	addq	$4 * SIZE, AO
	addq	$8 * SIZE, BO
	subq	$1, %rax
	jg	.L76
	ALIGN_4

.L78:
	addps	%xmm10, %xmm8
	addps	%xmm11, %xmm9

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhps	2 * SIZE(CO1), %xmm0
	movsd	0 * SIZE(CO2), %xmm2
	movhps	2 * SIZE(CO2), %xmm2
#endif

	mulps	%xmm7, %xmm8
	mulps	%xmm7, %xmm9
#ifndef TRMMKERNEL
	addps	%xmm0,  %xmm8
	addps	%xmm2, %xmm9
#endif

	movlps	%xmm8, 0 * SIZE(CO1)
	movhps	%xmm8, 2 * SIZE(CO1)
	movlps	%xmm9, 0 * SIZE(CO2)
	movhps	%xmm9, 2 * SIZE(CO2)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax,    8), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 4), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$4, KK
#endif

	addq	$4 * SIZE, CO1
	addq	$4 * SIZE, CO2
	ALIGN_4
	
.L80:
	testq	$2, M
	jle	.L90

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leaq	32 * SIZE + BUFFER, BO
#else
	leaq	32 * SIZE + BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax,   8), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 4), BO
#endif	

	pxor	%xmm8, %xmm8
	pxor	%xmm9, %xmm9
	pxor	%xmm10, %xmm10
	pxor	%xmm11, %xmm11

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK	
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$2, %rax
#else
	addq	$2, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle	.L85
	ALIGN_4

.L81:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	movsd	-32 * SIZE(AO), %xmm0
	movsd	-30 * SIZE(AO), %xmm1
	movsd	-32 * SIZE(BO), %xmm2
	movsd	-28 * SIZE(BO), %xmm3
	movsd	-24 * SIZE(BO), %xmm4
	movsd	-20 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm1, %xmm4
	mulps	%xmm1, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	movsd	-28 * SIZE(AO), %xmm0
	movsd	-26 * SIZE(AO), %xmm1
	movsd	-16 * SIZE(BO), %xmm2
	movsd	-12 * SIZE(BO), %xmm3
	movsd	 -8 * SIZE(BO), %xmm4
	movsd	 -4 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm1, %xmm4
	mulps	%xmm1, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	subq	$ -8 * SIZE, AO
	subq	$-32 * SIZE, BO
	subq	$1, %rax
	jg	.L81
	ALIGN_4

.L85:
	movsd	ALPHA, %xmm7

#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L88
	ALIGN_4

.L86:
	movsd	-32 * SIZE(AO), %xmm0
	movsd	-32 * SIZE(BO), %xmm2
	movsd	-28 * SIZE(BO), %xmm3

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9

	addq	$2 * SIZE, AO
	addq	$8 * SIZE, BO
	subq	$1, %rax
	jg	.L86
	ALIGN_4

.L88:
	addps	%xmm10, %xmm8
	addps	%xmm11, %xmm9

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movsd	0 * SIZE(CO2), %xmm2
#endif

	mulps	%xmm7, %xmm8
	mulps	%xmm7, %xmm9

#ifndef TRMMKERNEL
	addps	%xmm0,  %xmm8
	addps	%xmm2, %xmm9
#endif

	movlps	%xmm8, 0 * SIZE(CO1)
	movlps	%xmm9, 0 * SIZE(CO2)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax,    8), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 4), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$2, KK
#endif

	addq	$2 * SIZE, CO1
	addq	$2 * SIZE, CO2
	ALIGN_4

.L90:
	testq	$1, M
	jle	.L99

#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leaq	32 * SIZE + BUFFER, BO
#else
	leaq	32 * SIZE + BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax,   4), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 8), BO
#endif	

	pxor	%xmm8, %xmm8
	pxor	%xmm9, %xmm9
	pxor	%xmm10, %xmm10
	pxor	%xmm11, %xmm11

#ifndef TRMMKERNEL
	movq	K, %rax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movq	K, %rax
	subq	KK, %rax
	movq	%rax, KKK	
#else
	movq	KK, %rax
#ifdef LEFT
	addq	$1, %rax
#else
	addq	$2, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle	.L95
	ALIGN_4

.L91:	

	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	movss	-32 * SIZE(AO), %xmm0
	movss	-31 * SIZE(AO), %xmm1
	movss	-32 * SIZE(BO), %xmm2
	movss	-28 * SIZE(BO), %xmm3
	movss	-24 * SIZE(BO), %xmm4
	movss	-20 * SIZE(BO), %xmm5

	mulss	%xmm0, %xmm2
	mulss	%xmm0, %xmm3
	mulss	%xmm1, %xmm4
	mulss	%xmm1, %xmm5

	addss	%xmm2, %xmm8
	addss	%xmm3, %xmm9
	addss	%xmm4, %xmm10
	addss	%xmm5, %xmm11

	movss	-30 * SIZE(AO), %xmm0
	movss	-29 * SIZE(AO), %xmm1
	movss	-16 * SIZE(BO), %xmm2
	movss	-12 * SIZE(BO), %xmm3
	movss	 -8 * SIZE(BO), %xmm4
	movss	 -4 * SIZE(BO), %xmm5

	mulss	%xmm0, %xmm2
	mulss	%xmm0, %xmm3
	mulss	%xmm1, %xmm4
	mulss	%xmm1, %xmm5

	addss	%xmm2, %xmm8
	addss	%xmm3, %xmm9
	addss	%xmm4, %xmm10
	addss	%xmm5, %xmm11

	subq	$ -4 * SIZE, AO
	subq	$-32 * SIZE, BO
	subq	$1, %rax
	jg	.L91
	ALIGN_4

.L95:
	movss	ALPHA, %xmm7

#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L98
	ALIGN_4

.L96:
	movss	-32 * SIZE(AO), %xmm0
	movss	-32 * SIZE(BO), %xmm2
	movss	-28 * SIZE(BO), %xmm3

	mulss	%xmm0, %xmm2
	mulss	%xmm0, %xmm3

	addss	%xmm2, %xmm8
	addss	%xmm3, %xmm9

	addq	$1 * SIZE, AO
	addq	$8 * SIZE, BO
	subq	$1, %rax
	jg	.L96
	ALIGN_4

.L98:
	addss	%xmm10, %xmm8
	addss	%xmm11, %xmm9

#ifndef TRMMKERNEL
	movss	0 * SIZE(CO1), %xmm0
	movss	0 * SIZE(CO2), %xmm2
#endif

	mulss	%xmm7, %xmm8
	mulss	%xmm7, %xmm9

#ifndef TRMMKERNEL
	addss	%xmm0,  %xmm8
	addss	%xmm2, %xmm9
#endif

	movss	%xmm8, 0 * SIZE(CO1)
	movss	%xmm9, 0 * SIZE(CO2)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax,    4), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 8), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$1, KK
#endif
	ALIGN_4

.L99:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$2, KK
#endif

	leaq	(C, LDC, 2), C
	ALIGN_4