gemm_kernel_4x4_sse2.s

Optimized GotoBLAS libraries

	movsd	-15 * SIZE(AO), %xmm0

	mulsd	%xmm0, %xmm1
	mulsd	 6 * SIZE(BO), %xmm0
	addsd	%xmm1, %xmm10
	movsd	32 * SIZE(BO), %xmm1
	addsd	%xmm0, %xmm11
	movsd	-14 * SIZE(AO), %xmm0

	mulsd	%xmm0, %xmm3
	mulsd	10 * SIZE(BO), %xmm0
	addsd	%xmm3, %xmm8
	movsd	12 * SIZE(BO), %xmm3
	addsd	%xmm0, %xmm9
	movsd	-13 * SIZE(AO), %xmm0

	mulsd	%xmm0, %xmm3
	mulsd	14 * SIZE(BO), %xmm0
	addsd	%xmm3, %xmm10
	movsd	40 * SIZE(BO), %xmm3
	addsd	%xmm0, %xmm11
	movsd	-8 * SIZE(AO), %xmm0

	mulsd	%xmm2, %xmm5
	mulsd	18 * SIZE(BO), %xmm2
	addsd	%xmm5, %xmm8
	movsd	20 * SIZE(BO), %xmm5
	addsd	%xmm2, %xmm9
	movsd	-11 * SIZE(AO), %xmm2

	mulsd	%xmm2, %xmm5
	mulsd	22 * SIZE(BO), %xmm2
	addsd	%xmm5, %xmm10
	movsd	48 * SIZE(BO), %xmm5
	addsd	%xmm2, %xmm11
	movsd	-10 * SIZE(AO), %xmm2

	mulsd	%xmm2, %xmm7
	mulsd	26 * SIZE(BO), %xmm2
	addsd	%xmm7, %xmm8
	movsd	28 * SIZE(BO), %xmm7
	addsd	%xmm2, %xmm9
	movsd	-9 * SIZE(AO), %xmm2

	mulsd	%xmm2, %xmm7
	mulsd	30 * SIZE(BO), %xmm2
	addsd	%xmm7, %xmm10
	movsd	56 * SIZE(BO), %xmm7
	addsd	%xmm2, %xmm11
	movsd	-4 * SIZE(AO), %xmm2

	addq   $ 8 * SIZE, AO
	addq   $32 * SIZE, BO
	decq   %rax
	jne    .L72
	ALIGN_3

.L75:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movsd	ALPHA, %xmm7
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L78
	ALIGN_3

.L76:
	mulsd	%xmm0, %xmm1
	mulsd	 2 * SIZE(BO), %xmm0
	addsd	%xmm1, %xmm8
	addsd	%xmm0, %xmm9
	movsd	-15 * SIZE(AO), %xmm0
	movsd	 4 * SIZE(BO), %xmm1

	addq	$1 * SIZE, AO		# aoffset  += 4
	addq	$4 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L76
	ALIGN_3

.L78:
#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movsd	0 * SIZE(CO2), %xmm2
#endif

	addsd	%xmm10, %xmm8
	addsd	%xmm11, %xmm9

	mulsd	%xmm7, %xmm8
	mulsd	%xmm7, %xmm9

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

	movsd	%xmm8, 0 * SIZE(CO1)
	movsd	%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, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 4), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$1, KK
#endif
	ALIGN_3	
	
.L79:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$2, KK
#endif
	leaq	(C, LDC, 2), C
	ALIGN_3

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

.L81:
/* Copying to Sub Buffer */
	leaq	BUFFER, BO
	
#if defined(TRMMKERNEL) && defined(LEFT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif	

	movq	K, %rax
	sarq	$3, %rax
	jle	.L83
	ALIGN_3
	
.L82:
	PREFETCH	 56 * SIZE(B)

	movq	 0 * SIZE(B), %mm0
	movq	 1 * SIZE(B), %mm1
	movq	 2 * SIZE(B), %mm2
	movq	 3 * SIZE(B), %mm3
	movq	 4 * SIZE(B), %mm4
	movq	 5 * SIZE(B), %mm5
	movq	 6 * SIZE(B), %mm6
	movq	 7 * SIZE(B), %mm7

	addq	$ 8 * SIZE, B
	addq	$16 * SIZE, BO

	movq	%mm0, -16 * SIZE(BO)
	movq	%mm0, -15 * SIZE(BO)
	movq	%mm1, -14 * SIZE(BO)
	movq	%mm1, -13 * SIZE(BO)
	movq	%mm2, -12 * SIZE(BO)
	movq	%mm2, -11 * SIZE(BO)
	movq	%mm3, -10 * SIZE(BO)
	movq	%mm3,  -9 * SIZE(BO)
	movq	%mm4,  -8 * SIZE(BO)
	movq	%mm4,  -7 * SIZE(BO)
	movq	%mm5,  -6 * SIZE(BO)
	movq	%mm5,  -5 * SIZE(BO)
	movq	%mm6,  -4 * SIZE(BO)
	movq	%mm6,  -3 * SIZE(BO)
	movq	%mm7,  -2 * SIZE(BO)
	movq	%mm7,  -1 * SIZE(BO)

	decq	%rax
	jne	.L82
	ALIGN_3

.L83:
	movq	K, %rax
	andq	$7, %rax
	BRANCH
	jle	.L90
	ALIGN_3

.L84:
	movq	 0 * SIZE(B), %mm0

	movq	%mm0,  0 * SIZE(BO)
	movq	%mm0,  1 * SIZE(BO)

	addq	$1 * SIZE, B
	addq	$2 * SIZE, BO
	decq	%rax
	jne	.L84
	ALIGN_3
	
.L90:
	movq	C, CO1			# coffset1 = c
	movq	A, AO		# aoffset = a

	movq	M,  I
	sarq	$2, I	# i = (m >> 2)
	jle	.L100
	ALIGN_3

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

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

	movapd	-16 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movapd	 0 * SIZE(BO), %xmm1
	pxor	%xmm9, %xmm9
	movapd	 -8 * SIZE(AO), %xmm2
	pxor	%xmm10, %xmm10
	movapd	 8 * SIZE(BO), %xmm3
	pxor	%xmm11, %xmm11

	movapd	 0 * SIZE(AO), %xmm4
	movapd	 8 * SIZE(AO), %xmm6

	PREFETCHW      4 * SIZE(CO1)

#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_3

.L92:
	mulpd	%xmm1, %xmm0
	PREFETCH	(PREFETCHSIZE     +  0) * SIZE(AO)
	mulpd	-14 * SIZE(AO), %xmm1
	addpd	%xmm0, %xmm8
	movapd	-12 * SIZE(AO), %xmm0
	addpd	%xmm1, %xmm9
	movapd	 2 * SIZE(BO), %xmm1
	mulpd	%xmm1, %xmm0
	mulpd	-10 * SIZE(AO), %xmm1
	addpd	%xmm0, %xmm10
	movapd	 16 * SIZE(AO), %xmm0
	PREFETCH	(PREFETCHSIZE     +  8) * SIZE(AO)
	addpd	%xmm1, %xmm11
	movapd	 4 * SIZE(BO), %xmm1
	mulpd	%xmm1, %xmm2
	mulpd	-6 * SIZE(AO), %xmm1
	addpd	%xmm2, %xmm8
	movapd	-4 * SIZE(AO), %xmm2
	addpd	%xmm1, %xmm9
	movapd	 6 * SIZE(BO), %xmm1
	mulpd	%xmm1, %xmm2
	mulpd	-2 * SIZE(AO), %xmm1
	addpd	%xmm2, %xmm10
	movapd	24 * SIZE(AO), %xmm2
	PREFETCH	(PREFETCHSIZE     + 16) * SIZE(AO)
	addpd	%xmm1, %xmm11
	movapd	16 * SIZE(BO), %xmm1
	mulpd	%xmm3, %xmm4
	mulpd	 2 * SIZE(AO), %xmm3
	addpd	%xmm4, %xmm8
	movapd	 4 * SIZE(AO), %xmm4
	addpd	%xmm3, %xmm9
	movapd	10 * SIZE(BO), %xmm3
	mulpd	%xmm3, %xmm4
	mulpd	 6 * SIZE(AO), %xmm3
	addpd	%xmm4, %xmm10
	movapd	32 * SIZE(AO), %xmm4
	PREFETCH	(PREFETCHSIZE     + 24) * SIZE(AO)
	addpd	%xmm3, %xmm11
	movapd	12 * SIZE(BO), %xmm3
	mulpd	%xmm3, %xmm6
	mulpd	10 * SIZE(AO), %xmm3
	addpd	%xmm6, %xmm8
	movapd	12 * SIZE(AO), %xmm6
	addpd	%xmm3, %xmm9
	movapd	14 * SIZE(BO), %xmm3
	mulpd	%xmm3, %xmm6
	mulpd	14 * SIZE(AO), %xmm3
	addpd	%xmm6, %xmm10
	movapd	40 * SIZE(AO), %xmm6
	addpd	%xmm3, %xmm11
	movapd	24 * SIZE(BO), %xmm3

	addq   $32 * SIZE, AO
	addq   $16 * SIZE, BO
	decq   %rax
	jne    .L92
	ALIGN_3

.L95:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movapd	ALPHA, %xmm7
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L99
	ALIGN_3

.L96:
	mulpd	%xmm1, %xmm0
	mulpd	-14 * SIZE(AO), %xmm1
	addpd	%xmm0, %xmm8
	movapd	-12 * SIZE(AO), %xmm0
	addpd	%xmm1, %xmm9
	movapd	 2 * SIZE(BO), %xmm1

	addq	$4 * SIZE, AO		# aoffset  += 4
	addq	$2 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L96
	ALIGN_3

.L99:
#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhpd	1 * SIZE(CO1), %xmm0
	movsd	2 * SIZE(CO1), %xmm1
	movhpd	3 * SIZE(CO1), %xmm1
#endif

	addpd	%xmm10, %xmm8
	addpd	%xmm11, %xmm9

	mulpd	%xmm7, %xmm8
	mulpd	%xmm7, %xmm9

#ifndef TRMMKERNEL
	addpd	%xmm0,  %xmm8
	addpd	%xmm1,  %xmm9
#endif

	movsd	%xmm8, 0 * SIZE(CO1)
	movhpd	%xmm8, 1 * SIZE(CO1)
	movsd	%xmm9, 2 * SIZE(CO1)
	movhpd	%xmm9, 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, 2), BO
#endif

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

	addq	$4 * SIZE, CO1		# coffset += 4
	decq	I			# i --
	jg	.L91
	ALIGN_3	

.L100:
	testq	$2, M
	je	.L110
	ALIGN_3

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

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

	movapd	-16 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movapd	 0 * SIZE(BO), %xmm1
	pxor	%xmm9, %xmm9
	movapd	-8 * SIZE(AO), %xmm2
	pxor	%xmm10, %xmm10
	movapd	 8 * SIZE(BO), %xmm3
	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	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$3, %rax
	je	.L105
	ALIGN_3

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

	addq   $16 * SIZE, AO
	addq   $16 * SIZE, BO
	decq   %rax
	jne    .L102
	ALIGN_3

.L105:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movapd	ALPHA, %xmm7
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L109
	ALIGN_3

.L106:
	mulpd	%xmm0, %xmm1
	addpd	%xmm1, %xmm8
	movapd	-14 * SIZE(AO), %xmm0
	movapd	  2 * SIZE(BO), %xmm1

	addq	$2 * SIZE, AO		# aoffset  += 4
	addq	$2 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L106
	ALIGN_3

.L109:
	addpd	%xmm9, %xmm8
	addpd	%xmm11, %xmm10
	addpd	%xmm10, %xmm8

	mulpd	%xmm7, %xmm8

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

	addpd	%xmm0,  %xmm8
#endif

	movsd	%xmm8, 0 * SIZE(CO1)
	movhpd	%xmm8, 1 * SIZE(CO1)
	addq	$2 * SIZE, CO1		# coffset += 4

#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
	ALIGN_3	

.L110:
	testq	$1, M
	je	.L999
	ALIGN_3

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

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

	movsd	-16 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movsd	 0 * SIZE(BO), %xmm1
	pxor	%xmm9, %xmm9
	movsd	-12 * SIZE(AO), %xmm2
	pxor	%xmm10, %xmm10
	movsd	 8 * SIZE(BO), %xmm3
	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	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$3, %rax
	je	.L115
	ALIGN_3

.L112:
	mulsd	%xmm0, %xmm1
	PREFETCH	(PREFETCHSIZE     +  0) * SIZE(AO)
	movsd	-15 * SIZE(AO), %xmm0
	addsd	%xmm1, %xmm8
	movsd	16 * SIZE(BO), %xmm1
	mulsd	 2 * SIZE(BO), %xmm0
	addsd	%xmm0, %xmm9
	movsd	-14 * SIZE(AO), %xmm0
	mulsd	 4 * SIZE(BO), %xmm0
	addsd	%xmm0, %xmm10
	movsd	-13 * SIZE(AO), %xmm0
	mulsd	 6 * SIZE(BO), %xmm0
	addsd	%xmm0, %xmm11
	movsd	 -8 * SIZE(AO), %xmm0
	mulsd	%xmm2, %xmm3
	movsd	-11 * SIZE(AO), %xmm2
	addsd	%xmm3, %xmm8
	movsd	24 * SIZE(BO), %xmm3
	mulsd	10 * SIZE(BO), %xmm2
	addsd	%xmm2, %xmm9
	movsd	-10 * SIZE(AO), %xmm2
	mulsd	12 * SIZE(BO), %xmm2
	addsd	%xmm2, %xmm10
	movsd	-9 * SIZE(AO), %xmm2
	mulsd	14 * SIZE(BO), %xmm2
	addsd	%xmm2, %xmm11
	movsd	-4 * SIZE(AO), %xmm2

	addq   $ 8 * SIZE, AO
	addq   $16 * SIZE, BO
	decq   %rax
	jne    .L112
	ALIGN_3

.L115:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movsd	ALPHA, %xmm7
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L118
	ALIGN_3

.L116:
	mulsd	%xmm0, %xmm1
	movsd	-15 * SIZE(AO), %xmm0
	addsd	%xmm1, %xmm8
	movsd	 2 * SIZE(BO), %xmm1

	addq	$1 * SIZE, AO		# aoffset  += 4
	addq	$2 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L116
	ALIGN_3

.L118:
	addsd	%xmm10, %xmm8
	addsd	%xmm11, %xmm9
	addsd	%xmm9, %xmm8

	mulsd	%xmm7, %xmm8
#ifndef TRMMKERNEL
	addsd	0 * SIZE(CO1), %xmm8
#endif
	movsd	%xmm8, 0 * SIZE(CO1)
	ALIGN_3
	
.L999:
	movq	%rbx, %rsp

	EMMS

	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