gemm_kernel_4x4_penryn.s

Optimized GotoBLAS libraries

	movddup	ALPHA, %xmm3

	movapd	 %xmm8, %xmm0
	movsd    %xmm9, %xmm8
	mulpd	%xmm3, %xmm8
	movsd    %xmm0, %xmm9
	mulpd	%xmm3, %xmm9

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhpd	1 * SIZE(CO1), %xmm0
	movsd	0 * SIZE(CO2), %xmm2
	movhpd	1 * SIZE(CO2), %xmm2

	addpd	%xmm0, %xmm8
	addpd	%xmm2, %xmm9
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movhpd	%xmm8,  1 * SIZE(CO1)
	movsd	%xmm9,  0 * SIZE(CO2)
	movhpd	%xmm9,  1 * SIZE(CO2)

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

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

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

.L70:
	testq	$1, M
	jle,pt	.L79
	ALIGN_4

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

	movq	B, BO
#else
	movq	B, BO

	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	addq	%rax, AO
	leaq	(BO, %rax, 2), BO
#endif	

	movsd	-16 * SIZE(AO), %xmm0
	movaps	-17 * SIZE(BO), %xmm2

	pxor	%xmm8,  %xmm8
	pxor	%xmm9,  %xmm9

#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,pn	.L75
	ALIGN_4

.L72:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	movsd	-15 * SIZE(AO), %xmm0
	addpd	%xmm2, %xmm8
	movaps	-15 * SIZE(BO), %xmm2

	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	movsd	-14 * SIZE(AO), %xmm0
	addpd	%xmm2, %xmm9
	movaps	-13 * SIZE(BO), %xmm2

	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	movsd	-13 * SIZE(AO), %xmm0
	addpd	%xmm2, %xmm8
	movaps	-11 * SIZE(BO), %xmm2

	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	movsd	-12 * SIZE(AO), %xmm0
	addpd	%xmm2, %xmm9
	movaps	 -9 * SIZE(BO), %xmm2

	subq	$-4 * SIZE, AO
	subq	$-8 * SIZE, BO
	subq	$1, %rax
	jg,pt	.L72
	ALIGN_4

.L75:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je,pt	.L78
	ALIGN_4

.L76:
	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	movsd	-15 * SIZE(AO), %xmm0
	addpd	%xmm2, %xmm8
	movaps	-15 * SIZE(BO), %xmm2

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

	subq	$1, %rax
	jg,pt	.L76
	ALIGN_4

.L78:
	movddup	ALPHA, %xmm3

	addpd	%xmm9, %xmm8

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

	mulpd	%xmm3, %xmm8

#ifndef TRMMKERNEL
	addpd	%xmm0, %xmm8
#endif

	movlpd	%xmm8,  0 * SIZE(CO1)
	movhpd	%xmm8,  0 * SIZE(CO2)

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

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

.L79:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addq	$2, KK
#endif

	leaq	(C, LDC, 2), C
	movq	BO, B
	ALIGN_4

.L80:
	testq	$1, N
	jle,pt	.L999

#if defined(TRMMKERNEL) && defined(LEFT)
        movq    OFFSET, %rax
	movq    %rax, KK
#endif

	movq	C, CO1
	movq	A, AO

	movq	M,  I
	sarq	$2, I	# i = (m >> 2)
	jle,pn	.L100
	ALIGN_4

.L91:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	B, BO

	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	addq	%rax, BO
#endif	

	movaps	-16 * SIZE(AO), %xmm0
	movaps	-14 * SIZE(AO), %xmm1
	movsd	-17 * SIZE(BO), %xmm2

	prefetcht0     3 * SIZE(CO1)
	pxor	%xmm8,  %xmm8
	pxor	%xmm9,  %xmm9
	pxor	%xmm12, %xmm12
	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	$4, %rax
#else
	addq	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle,pn	.L95
	ALIGN_4

.L92:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	pshufd	$0x44, %xmm2, %xmm3
	pshufd	$0x44, %xmm2, %xmm4
	movsd	-16 * SIZE(BO), %xmm2

	mulpd	%xmm0, %xmm3
	movaps	-12 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm4
	movaps	-10 * SIZE(AO), %xmm1

	addpd	%xmm3, %xmm8
	addpd	%xmm4, %xmm12

	pshufd	$0x44, %xmm2, %xmm3
	pshufd	$0x44, %xmm2, %xmm4
	movsd	-15 * SIZE(BO), %xmm2

	mulpd	%xmm0, %xmm3
	movaps	 -8 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm4
	movaps	 -6 * SIZE(AO), %xmm1

	addpd	%xmm3, %xmm8
	addpd	%xmm4, %xmm12

	PREFETCH (PREFETCHSIZE +  8) * SIZE(AO)

	pshufd	$0x44, %xmm2, %xmm3
	pshufd	$0x44, %xmm2, %xmm4
	movsd	-14 * SIZE(BO), %xmm2

	mulpd	%xmm0, %xmm3
	movaps	 -4 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm4
	movaps	 -2 * SIZE(AO), %xmm1

	addpd	%xmm3, %xmm8
	addpd	%xmm4, %xmm12

	pshufd	$0x44, %xmm2, %xmm3
	pshufd	$0x44, %xmm2, %xmm4
	movsd	-13 * SIZE(BO), %xmm2

	mulpd	%xmm0, %xmm3
	movaps	  0 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm4
	movaps	  2 * SIZE(AO), %xmm1

	addpd	%xmm3, %xmm8
	addpd	%xmm4, %xmm12

	subq	$-16 * SIZE, AO
	subq	$ -4 * SIZE, BO
	subq	$1, %rax
	jg,pt	.L92
	ALIGN_4

.L95:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je,pt	.L98
	ALIGN_4

.L96:
	pshufd	$0x44, %xmm2, %xmm3
	pshufd	$0x44, %xmm2, %xmm4
	movsd	-16 * SIZE(BO), %xmm2

	mulpd	%xmm0, %xmm3
	movaps	-12 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm4
	movaps	-10 * SIZE(AO), %xmm1

	addpd	%xmm3, %xmm8
	addpd	%xmm4, %xmm12

	addq	$4 * SIZE, AO
	addq	$1 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L96
	ALIGN_4

.L98:
	movddup	ALPHA, %xmm3

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhpd	1 * SIZE(CO1), %xmm0
	movsd	2 * SIZE(CO1), %xmm1
	movhpd	3 * SIZE(CO1), %xmm1
#endif

	mulpd	%xmm3, %xmm8
	mulpd	%xmm3, %xmm12

#ifndef TRMMKERNEL
	addpd	%xmm0, %xmm8
	addpd	%xmm1, %xmm12
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movhpd	%xmm8,  1 * SIZE(CO1)
	movsd	%xmm12, 2 * SIZE(CO1)
	movhpd	%xmm12, 3 * SIZE(CO1)

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

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

	addq	$4 * SIZE, CO1
	decq	I
	jg,pt	.L91
	ALIGN_4	

.L100:
	testq	$2, M
	jle,pt	.L110
	ALIGN_4

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

	movq	B, BO
#else
	movq	B, BO

	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	addq	%rax, BO
#endif	

	movaps	-16 * SIZE(AO), %xmm0
	pxor	%xmm8,  %xmm8
	movaps	-17 * SIZE(BO), %xmm2
	pxor	%xmm9,  %xmm9

#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	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle,pn	.L105
	ALIGN_4

.L102:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	pshufd	$0x44, %xmm2, %xmm3
	movsd	-16 * SIZE(BO), %xmm2

	mulpd	%xmm0, %xmm3
	movaps	-14 * SIZE(AO), %xmm0
	addpd	%xmm3, %xmm8

	pshufd	$0x44, %xmm2, %xmm3
	movsd	-15 * SIZE(BO), %xmm2

	mulpd	%xmm0, %xmm3
	movaps	-12 * SIZE(AO), %xmm0
	addpd	%xmm3, %xmm9

	pshufd	$0x44, %xmm2, %xmm3
	movsd	-14 * SIZE(BO), %xmm2

	mulpd	%xmm0, %xmm3
	movaps	-10 * SIZE(AO), %xmm0
	addpd	%xmm3, %xmm8

	pshufd	$0x44, %xmm2, %xmm3
	movsd	-13 * SIZE(BO), %xmm2

	mulpd	%xmm0, %xmm3
	movaps	 -8 * SIZE(AO), %xmm0
	addpd	%xmm3, %xmm9

	subq	$-8 * SIZE, AO
	subq	$-4 * SIZE, BO
	subq	$1, %rax
	jg,pt	.L102
	ALIGN_4

.L105:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je,pt	.L108
	ALIGN_4

.L106:
	pshufd	$0x44, %xmm2, %xmm3
	movsd	-16 * SIZE(BO), %xmm2

	mulpd	%xmm0, %xmm3
	movaps	-14 * SIZE(AO), %xmm0
	addpd	%xmm3, %xmm8

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

	subq	$1, %rax
	jg,pt	.L106
	ALIGN_4

.L108:
	addpd	%xmm9, %xmm8

	movddup	ALPHA, %xmm3

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhpd	1 * SIZE(CO1), %xmm0
#endif

	mulpd	%xmm3, %xmm8


#ifndef TRMMKERNEL
	addpd	%xmm0, %xmm8
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movhpd	%xmm8,  1 * SIZE(CO1)

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

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

	addq	$2 * SIZE, CO1
	ALIGN_4	

.L110:
	testq	$1, M
	jle,pt	.L999
	
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	B, BO

	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	addq	%rax, AO
	addq	%rax, BO
#endif	

	movsd	-16 * SIZE(AO), %xmm0
	movsd	-17 * SIZE(BO), %xmm2

	pxor	%xmm8,  %xmm8
	pxor	%xmm9,  %xmm9

#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	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle,pn	.L115
	ALIGN_4

.L112:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	mulsd	%xmm0, %xmm2
	addsd	%xmm2, %xmm8
	movsd	-15 * SIZE(AO), %xmm0
	movsd	-16 * SIZE(BO), %xmm2

	mulsd	%xmm0, %xmm2
	addsd	%xmm2, %xmm8
	movsd	-14 * SIZE(AO), %xmm0
	movsd	-15 * SIZE(BO), %xmm2

	mulsd	%xmm0, %xmm2
	addsd	%xmm2, %xmm8
	movsd	-13 * SIZE(AO), %xmm0
	movsd	-14 * SIZE(BO), %xmm2

	mulsd	%xmm0, %xmm2
	addsd	%xmm2, %xmm8
	movsd	-12 * SIZE(AO), %xmm0
	movsd	-13 * SIZE(BO), %xmm2

	subq	$-4 * SIZE, AO
	subq	$-4 * SIZE, BO
	subq	$1, %rax
	jg,pt	.L112
	ALIGN_4

.L115:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je,pt	.L118
	ALIGN_4

.L116:
	mulsd	%xmm0, %xmm2
	addsd	%xmm2, %xmm8
	movsd	-15 * SIZE(AO), %xmm0
	movsd	-16 * SIZE(BO), %xmm2

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

	subq	$1, %rax
	jg,pt	.L116
	ALIGN_4

.L118:
	movddup	ALPHA, %xmm3

	addpd	%xmm9, %xmm8

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

	mulsd	%xmm3, %xmm8

#ifndef TRMMKERNEL
	addpd	%xmm0, %xmm8
#endif

	movlpd	%xmm8,  0 * SIZE(CO1)
	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

#ifdef WINDOWS_ABI
	movq	 48(%rsp), %rdi
	movq	 56(%rsp), %rsi
	movups	 64(%rsp), %xmm6
	movups	 80(%rsp), %xmm7
	movups	 96(%rsp), %xmm8
	movups	112(%rsp), %xmm9
	movups	128(%rsp), %xmm10
	movups	144(%rsp), %xmm11
	movups	160(%rsp), %xmm12
	movups	176(%rsp), %xmm13
	movups	192(%rsp), %xmm14
	movups	208(%rsp), %xmm15
#endif

	addq	$STACKSIZE, %rsp
	ret

	EPILOGUE