gemm_kernel_4x4_sse3.s

Optimized GotoBLAS libraries

	movsd	%xmm1, 0 * SIZE(CO2)
	movhpd	%xmm1, 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		# coffset += 4
	addq	$2 * SIZE, CO2		# coffset += 4
	ALIGN_4	

.L70:
	testq	$1, M
	je	.L79
	ALIGN_4

.L71:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 2), BO
#endif	

	movddup	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movapd	 0 * SIZE(BO), %xmm9
	pxor	%xmm1, %xmm1
	movddup	 4 * SIZE(AO), %xmm10
	pxor	%xmm2, %xmm2
	movapd	 8 * SIZE(BO), %xmm11
	pxor	%xmm3, %xmm3

#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	$3, %rax
	je	.L75
	ALIGN_4

.L72:
	mulpd	%xmm8, %xmm9
	PREFETCH  (PREFETCHSIZE + 0) * SIZE(AO)
	movddup	 1 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm0
	mulpd	 2 * SIZE(BO), %xmm8
	movapd	16 * SIZE(BO), %xmm9
	addpd	%xmm8, %xmm1
	movddup	 2 * SIZE(AO), %xmm8
	mulpd	 4 * SIZE(BO), %xmm8
	addpd	%xmm8, %xmm2
	movddup	 3 * SIZE(AO), %xmm8
	mulpd	 6 * SIZE(BO), %xmm8
	addpd	%xmm8, %xmm3
	movddup	 8 * SIZE(AO), %xmm8
	mulpd	%xmm10, %xmm11
	movddup	 5 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm0
	mulpd	10 * SIZE(BO), %xmm10
	movapd	24 * SIZE(BO), %xmm11
	addpd	%xmm10, %xmm1
	movddup	 6 * SIZE(AO), %xmm10
	mulpd	12 * SIZE(BO), %xmm10
	addpd	%xmm10, %xmm2
	movddup	 7 * SIZE(AO), %xmm10
	mulpd	14 * SIZE(BO), %xmm10
	addpd	%xmm10, %xmm3
	movddup	12 * SIZE(AO), %xmm10

	addq   $ 8 * SIZE, AO
	addq   $16 * SIZE, BO
	decq   %rax
	jne    .L72
	ALIGN_4

.L75:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movddup	ALPHA, %xmm15
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L78
	ALIGN_4

.L76:
	mulpd	%xmm8, %xmm9
	movddup	 1 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm0
	movapd	 2 * SIZE(BO), %xmm9

	addq	$1 * SIZE, AO		# aoffset  += 4
	addq	$2 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L76
	ALIGN_4

.L78:
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movsd	0 * SIZE(CO1), %xmm8
	movhpd	0 * SIZE(CO2), %xmm8
#endif

	addpd	%xmm1, %xmm0
	addpd	%xmm3, %xmm2
	addpd	%xmm2, %xmm0

	mulpd	%xmm15, %xmm0
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addpd	%xmm8,  %xmm0
#endif

	movsd	%xmm0, 0 * SIZE(CO1)
	movhpd	%xmm0, 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
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 2), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$1, KK
#endif
	ALIGN_4	
	
.L79:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$2, KK
#endif
	leaq	(C, LDC, 2), C
	movq	BO, B
	ALIGN_4

.L80:
	testq	$1, N
	je	.L999
	ALIGN_4

#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	.L100
	ALIGN_4

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

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(B,  %rax, 1), BO
#endif	

	movapd	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movddup	 0 * SIZE(BO), %xmm9
	pxor	%xmm1, %xmm1
	movapd	 8 * SIZE(AO), %xmm10
	pxor	%xmm2, %xmm2
	movddup	 4 * SIZE(BO), %xmm11
	pxor	%xmm3, %xmm3

#ifdef HAVE_3DNOW
	prefetchw      4 * SIZE(CO1)
#else
	prefetchnta     4 * SIZE(CO1)
#endif

#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	$3, %rax
	je	.L95
	ALIGN_4

.L92:
	mulpd	%xmm9, %xmm8
	PREFETCH  (PREFETCHSIZE + 0) * SIZE(AO)
	mulpd	 2 * SIZE(AO), %xmm9
	addpd	%xmm8, %xmm0
	movapd	 4 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movddup	 1 * SIZE(BO), %xmm9
	mulpd	%xmm9, %xmm8
	mulpd	 6 * SIZE(AO), %xmm9
	addpd	%xmm8, %xmm2
	movapd	16 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm3
	movddup	 2 * SIZE(BO), %xmm9
	mulpd	%xmm9, %xmm10
	mulpd	10 * SIZE(AO), %xmm9
	addpd	%xmm10, %xmm0
	movapd	12 * SIZE(AO), %xmm10
	addpd	%xmm9, %xmm1
	movddup	 3 * SIZE(BO), %xmm9
	mulpd	%xmm9, %xmm10
	mulpd	14 * SIZE(AO), %xmm9
	addpd	%xmm10, %xmm2
	movapd	24 * SIZE(AO), %xmm10
	PREFETCH  (PREFETCHSIZE + 16) * SIZE(AO)
	addpd	%xmm9, %xmm3
	movddup	 8 * SIZE(BO), %xmm9
	mulpd	%xmm11, %xmm8
	mulpd	18 * SIZE(AO), %xmm11
	addpd	%xmm8, %xmm0
	movapd	20 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm1
	movddup	 5 * SIZE(BO), %xmm11
	mulpd	%xmm11, %xmm8
	mulpd	22 * SIZE(AO), %xmm11
	addpd	%xmm8, %xmm2
	movapd	32 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm3
	movddup	 6 * SIZE(BO), %xmm11
	mulpd	%xmm11, %xmm10
	mulpd	26 * SIZE(AO), %xmm11
	addpd	%xmm10, %xmm0
	movapd	28 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm1
	movddup	 7 * SIZE(BO), %xmm11
	mulpd	%xmm11, %xmm10
	mulpd	30 * SIZE(AO), %xmm11
	addpd	%xmm10, %xmm2
	movapd	40 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm3
	movddup	12 * SIZE(BO), %xmm11

	addq   $32 * SIZE, AO
	addq   $8 * SIZE, BO
	decq   %rax
	jne    .L92
	ALIGN_4

.L95:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movddup	ALPHA, %xmm15
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L99
	ALIGN_4

.L96:
	mulpd	%xmm9, %xmm8
	mulpd	 2 * SIZE(AO), %xmm9
	addpd	%xmm8, %xmm0
	movapd	 4 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movddup	 1 * SIZE(BO), %xmm9

	addq	$4 * SIZE, AO		# aoffset  += 4
	addq	$1 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L96
	ALIGN_4

.L99:
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movsd	0 * SIZE(CO1), %xmm8
	movhpd	1 * SIZE(CO1), %xmm8
	movsd	2 * SIZE(CO1), %xmm9
	movhpd	3 * SIZE(CO1), %xmm9
#endif

	addpd	%xmm2, %xmm0
	addpd	%xmm3, %xmm1

	mulpd	%xmm15, %xmm0
	mulpd	%xmm15, %xmm1

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addpd	%xmm8,  %xmm0
	addpd	%xmm9,  %xmm1
#endif

	movsd	%xmm0, 0 * SIZE(CO1)
	movhpd	%xmm0, 1 * SIZE(CO1)
	movsd	%xmm1, 2 * SIZE(CO1)
	movhpd	%xmm1, 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
	leaq	(BO, %rax, 1), BO
#endif

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

	addq	$4 * SIZE, CO1		# coffset += 4
	decq	I			# i --
	jg	.L91
	ALIGN_4	

.L100:
	testq	$2, M
	je	.L110
	ALIGN_4

.L101:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(B,  %rax, 1), BO
#endif	

	movapd	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movddup	 0 * SIZE(BO), %xmm9
	pxor	%xmm1, %xmm1
	movapd	 8 * SIZE(AO), %xmm10
	pxor	%xmm2, %xmm2
	movddup	 4 * SIZE(BO), %xmm11
	pxor	%xmm3, %xmm3

#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	$3, %rax
	je	.L105
	ALIGN_4

.L102:
	mulpd	%xmm9, %xmm8
	PREFETCH  (PREFETCHSIZE + 0) * SIZE(AO)
	movddup	 1 * SIZE(BO), %xmm9
	addpd	%xmm8, %xmm0
	mulpd	 2 * SIZE(AO), %xmm9
	movapd	16 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movddup	 2 * SIZE(BO), %xmm9
	mulpd	 4 * SIZE(AO), %xmm9
	addpd	%xmm9, %xmm2
	movddup	 3 * SIZE(BO), %xmm9
	mulpd	 6 * SIZE(AO), %xmm9
	addpd	%xmm9, %xmm3
	movddup	 8 * SIZE(BO), %xmm9
	mulpd	%xmm11, %xmm10
	movddup	 5 * SIZE(BO), %xmm11
	addpd	%xmm10, %xmm0
	mulpd	10 * SIZE(AO), %xmm11
	movapd	24 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm1
	movddup	 6 * SIZE(BO), %xmm11
	mulpd	12 * SIZE(AO), %xmm11
	addpd	%xmm11, %xmm2
	movddup	 7 * SIZE(BO), %xmm11
	mulpd	14 * SIZE(AO), %xmm11
	addpd	%xmm11, %xmm3
	movddup	12 * SIZE(BO), %xmm11

	addq   $16 * SIZE, AO
	addq   $ 8 * SIZE, BO
	decq   %rax
	jne    .L102
	ALIGN_4

.L105:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movddup	ALPHA, %xmm15
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L109
	ALIGN_4

.L106:
	mulpd	%xmm9, %xmm8
	movddup	 1 * SIZE(BO), %xmm9
	addpd	%xmm8, %xmm0
	movapd	 2 * SIZE(AO), %xmm8

	addq	$2 * SIZE, AO		# aoffset  += 4
	addq	$1 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L106
	ALIGN_4

.L109:
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movsd	0 * SIZE(CO1), %xmm8
	movhpd	1 * SIZE(CO1), %xmm8
#endif

	addpd	%xmm1, %xmm0
	addpd	%xmm3, %xmm2
	addpd	%xmm2, %xmm0

	mulpd	%xmm15, %xmm0
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
     addpd	%xmm8,  %xmm0
#endif

	movsd	%xmm0, 0 * SIZE(CO1)
	movhpd	%xmm0, 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
	leaq	(BO, %rax, 1), BO
#endif

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

	addq	$2 * SIZE, CO1		# coffset += 4
	ALIGN_4	

.L110:
	testq	$1, M
	je	.L999
	ALIGN_4

.L111:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 1), BO
#endif	

	movsd	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movsd	 0 * SIZE(BO), %xmm9
	pxor	%xmm1, %xmm1
	movsd	 4 * SIZE(AO), %xmm10
	pxor	%xmm2, %xmm2
	movsd	 4 * SIZE(BO), %xmm11
	pxor	%xmm3, %xmm3

	movapd	 0 * SIZE(AO), %xmm9
	movapd	 0 * SIZE(BO), %xmm8
	movapd	 4 * SIZE(AO), %xmm11
	movapd	 4 * SIZE(BO), %xmm10

#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	$3, %rax
	je	.L115
	ALIGN_4

.L112:
	mulpd	%xmm9, %xmm8
	movapd	 2 * SIZE(AO), %xmm9
	addpd	%xmm8, %xmm0
	mulpd	 2 * SIZE(BO), %xmm9
	movapd	 8 * SIZE(BO), %xmm8
	addpd	%xmm9, %xmm1
	movapd	 8 * SIZE(AO), %xmm9
	mulpd	%xmm11, %xmm10
	movapd	 6 * SIZE(AO), %xmm11
	addpd	%xmm10, %xmm0
	mulpd	 6 * SIZE(BO), %xmm11
	movapd	12 * SIZE(BO), %xmm10
	addpd	%xmm11, %xmm1
	movapd	12 * SIZE(AO), %xmm11

	addq   $8 * SIZE, AO
	addq   $8 * SIZE, BO
	decq   %rax
	jne    .L112
	ALIGN_4

.L115:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movddup	ALPHA, %xmm15
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L118
	ALIGN_4

.L116:
	mulsd	 0 * SIZE(BO), %xmm9
	addsd	%xmm9, %xmm0
	movsd	 1 * SIZE(AO), %xmm9

	addq	$1 * SIZE, AO		# aoffset  += 4
	addq	$1 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L116
	ALIGN_4

.L118:
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movsd	0 * SIZE(CO1), %xmm8
#endif

	addpd	%xmm1, %xmm0
	haddpd	%xmm0, %xmm0
	mulsd	%xmm15, %xmm0
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addsd	%xmm8, %xmm0
#endif

	movsd	%xmm0, 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