zgemm3m_kernel_4x4_barcelona.s

Optimized GotoBLAS libraries

	addpd	%xmm1, %xmm10
	movapd	  (BO, %rax, 4), %xmm1
	addpd	%xmm4, %xmm11
	movddup	-11 * SIZE(AO, %rax, 1), %xmm4
	mulpd	%xmm2, %xmm3
	mulpd	 -6 * SIZE(BO, %rax, 4), %xmm2
	addpd	%xmm3, %xmm8
	movapd	 -4 * SIZE(BO, %rax, 4), %xmm3
	addpd	%xmm2, %xmm9
	movddup	-13 * SIZE(AO, %rax, 1), %xmm2
	mulpd	%xmm2, %xmm3
	mulpd	 -2 * SIZE(BO, %rax, 4), %xmm2
	addpd	%xmm3, %xmm10
	movapd	  8 * SIZE(BO, %rax, 4), %xmm3
	addpd	%xmm2, %xmm11
	movddup	-10 * SIZE(AO, %rax, 1), %xmm2

	addq	$4 * SIZE, %rax
	jl,pt	.L32
	ALIGN_4

.L36:
	movapd	ALPHA, %xmm7

#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L38

	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 4), BO
	negq	%rax
	ALIGN_4

.L37:
	mulpd	%xmm0, %xmm1
	mulpd	-14 * SIZE(BO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm8
	movapd	-12 * SIZE(BO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm9
	movddup	-15 * SIZE(AO, %rax, 1), %xmm0

	addq	$SIZE, %rax
	jl	.L37
	ALIGN_4

.L38:
	addpd	%xmm10, %xmm8
	addpd	%xmm11, %xmm9


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

	movddup	 %xmm8,  %xmm4
	mulpd	 %xmm7,  %xmm4
	addpd	 %xmm4,  %xmm0

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

	movsd	0 * SIZE(CO2), %xmm0
	movhpd	1 * SIZE(CO2), %xmm0

	unpckhpd %xmm8,  %xmm8
	mulpd	 %xmm7,  %xmm8
	addpd	 %xmm8,  %xmm0

	movsd	%xmm0, 0 * SIZE(CO2)
	movhpd	%xmm0, 1 * SIZE(CO2)

	movsd	0 * SIZE(CO1, LDC, 2), %xmm0
	movhpd	1 * SIZE(CO1, LDC, 2), %xmm0

	movddup	 %xmm9,  %xmm4
	mulpd	 %xmm7,  %xmm4
	addpd	 %xmm4,  %xmm0

	movsd	%xmm0, 0 * SIZE(CO1, LDC, 2)
	movhpd	%xmm0, 1 * SIZE(CO1, LDC, 2)
	
	movsd	0 * SIZE(CO2, LDC, 2), %xmm0
	movhpd	1 * SIZE(CO2, LDC, 2), %xmm0

	unpckhpd %xmm9,  %xmm9
	mulpd	 %xmm7,  %xmm9
	addpd	 %xmm9,  %xmm0

	movsd	%xmm0, 0 * SIZE(CO2, LDC, 2)
	movhpd	%xmm0, 1 * SIZE(CO2, LDC, 2)

	ALIGN_4	
	
.L39:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$4, KK
#endif

#ifndef BUFFERED
	movq	BO, B
#endif

	leaq	(C, LDC, 4), C		# c += 4 * ldc
	decq	J			# j --
	jg	.L01
	ALIGN_4

.L40:
	testq	$3, N
	je	.L999

	testq	$2, N
	je	.L80
	ALIGN_4

.L41:
#ifdef BUFFERED
	leaq	16 * SIZE + BUFFER, BO
#endif
		
#if defined(TRMMKERNEL) && defined(LEFT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif	

#ifdef BUFFERED
	movq	K, %rax
	sarq	$2, %rax
	jle	.L43
	ALIGN_4
	
.L42:
	prefetchnta	 (RPREFETCHSIZE +  0)  * SIZE(B)

	movaps	 (B), %xmm0
	movaps	%xmm0, -16 * SIZE(BO)

	movaps	 2 * SIZE(B), %xmm1
	movaps	%xmm1, -14 * SIZE(BO)

	prefetchw	 (WPREFETCHSIZE +  0)  * SIZE(BO)

	movaps	 4 * SIZE(B), %xmm2
	movaps	%xmm2, -12 * SIZE(BO)

	movaps	 6 * SIZE(B), %xmm3
	movaps	%xmm3, -10 * SIZE(BO)

	subq	$-8 * SIZE, BO
	subq	$-8 * SIZE, B

	subq	$1, %rax
	jne	.L42
	ALIGN_4

.L43:
	movq	K, %rax
	andq	$3, %rax
	BRANCH
	jle	.L50
	ALIGN_4

.L44:
	movaps	 (B), %xmm0
	movaps	%xmm0, -16 * SIZE(BO)

	addq	$2 * SIZE, B
	addq	$2 * SIZE, BO
	subq	$1, %rax
	jne	.L44
	ALIGN_4
	
.L50:
#endif
	movq	C, CO1			# coffset1 = c
	leaq	(C, LDC, 1), CO2	# coffset2 = c + ldc
	movq	A, AO		# aoffset = a

	movq	M,  I
	sarq	$2, I	# i = (m >> 2)
	jle	.L60
	ALIGN_4

.L51:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

#ifdef BUFFERED
	leaq	16 * SIZE + BUFFER, BO
#else
	movq	B, BO
#endif
#else
	leaq	16 * SIZE + BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 2), BO
#endif	

	movddup	-16 * SIZE(BO), %xmm1
	movddup	-15 * SIZE(BO), %xmm5
	pxor	%xmm8, %xmm8
	movddup	-12 * SIZE(BO), %xmm3
	pxor	%xmm9, %xmm9
	movapd	-16 * SIZE(AO), %xmm0
	pxor	%xmm12, %xmm12
	movapd	 -8 * SIZE(AO), %xmm4
	pxor	%xmm13, %xmm13
	prefetchw      7 * SIZE(CO1)
	movapd	%xmm0, %xmm2
	prefetchw      7 * SIZE(CO2)

#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

	andq	$-4, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 2), BO
	negq	%rax
	je,pn	.L56
	ALIGN_4

.L52:
	mulpd	%xmm1, %xmm0
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm8
	movapd	-12 * SIZE(AO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm12
	movddup	-14 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm5, %xmm2
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	-13 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm0, %xmm2
	mulpd	%xmm1, %xmm0
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm8
	movapd	  (AO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm12
	movddup	 -8 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm5, %xmm2
	mulpd	-10 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	-11 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm4, %xmm2
	mulpd	%xmm3, %xmm4
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm3
	addpd	%xmm4, %xmm8
	movapd	 -4 * SIZE(AO, %rax, 4), %xmm4
	addpd	%xmm3, %xmm12
	movddup	-10 * SIZE(BO, %rax, 2), %xmm3
	mulpd	%xmm5, %xmm2
	mulpd	 -6 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	 -9 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm4, %xmm2
	mulpd	%xmm3, %xmm4
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm3
	addpd	%xmm4, %xmm8
	movapd	  8 * SIZE(AO, %rax, 4), %xmm4
	addpd	%xmm3, %xmm12
	movddup	 -4 * SIZE(BO, %rax, 2), %xmm3
	mulpd	%xmm5, %xmm2
	mulpd	 -2 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	 -7 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm0, %xmm2

	addq	$4 * SIZE, %rax
	jl,pt	.L52
	ALIGN_4

.L56:
	movapd	ALPHA, %xmm7

#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L59

	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 2), BO
	negq	%rax
	ALIGN_4

.L57:
	mulpd	%xmm1, %xmm0
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm8
	movapd	-12 * SIZE(AO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm12
	movddup	-14 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm5, %xmm2
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm5
	addpd	%xmm2, %xmm9
	addpd	%xmm5, %xmm13
	movddup	-13 * SIZE(BO, %rax, 2), %xmm5
	movapd	%xmm0, %xmm2

	addq	$SIZE, %rax
	jl	.L57
	ALIGN_4

.L59:
	movsd	0 * SIZE(CO1), %xmm0
	movhpd	1 * SIZE(CO1), %xmm0
	movsd	2 * SIZE(CO1), %xmm1
	movhpd	3 * SIZE(CO1), %xmm1

	movsd	4 * SIZE(CO1), %xmm2
	movhpd	5 * SIZE(CO1), %xmm2
	movsd	6 * SIZE(CO1), %xmm3
	movhpd	7 * SIZE(CO1), %xmm3

	movddup	 %xmm8,  %xmm4
	unpckhpd %xmm8,  %xmm8
	movddup	 %xmm12, %xmm5
	unpckhpd %xmm12, %xmm12

	mulpd	 %xmm7,  %xmm4
	mulpd	 %xmm7,  %xmm8
	mulpd	 %xmm7,  %xmm5
	mulpd	 %xmm7,  %xmm12

	addpd	 %xmm4,  %xmm0
	addpd	 %xmm8,  %xmm1
	addpd	 %xmm5,  %xmm2
	addpd	 %xmm12, %xmm3

	movsd	%xmm0, 0 * SIZE(CO1)
	movhpd	%xmm0, 1 * SIZE(CO1)
	movsd	%xmm1, 2 * SIZE(CO1)
	movhpd	%xmm1, 3 * SIZE(CO1)

	movsd	%xmm2, 4 * SIZE(CO1)
	movhpd	%xmm2, 5 * SIZE(CO1)
	movsd	%xmm3, 6 * SIZE(CO1)
	movhpd	%xmm3, 7 * SIZE(CO1)
	
	movsd	0 * SIZE(CO2), %xmm0
	movhpd	1 * SIZE(CO2), %xmm0
	movsd	2 * SIZE(CO2), %xmm1
	movhpd	3 * SIZE(CO2), %xmm1

	movsd	4 * SIZE(CO2), %xmm2
	movhpd	5 * SIZE(CO2), %xmm2
	movsd	6 * SIZE(CO2), %xmm3
	movhpd	7 * SIZE(CO2), %xmm3

	movddup	 %xmm9,  %xmm4
	unpckhpd %xmm9,  %xmm9
	movddup	 %xmm13, %xmm5
	unpckhpd %xmm13, %xmm13

	mulpd	 %xmm7,  %xmm4
	mulpd	 %xmm7,  %xmm9
	mulpd	 %xmm7,  %xmm5
	mulpd	 %xmm7,  %xmm13

	addpd	 %xmm4,  %xmm0
	addpd	 %xmm9,  %xmm1
	addpd	 %xmm5,  %xmm2
	addpd	 %xmm13, %xmm3

	movsd	%xmm0, 0 * SIZE(CO2)
	movhpd	%xmm0, 1 * SIZE(CO2)
	movsd	%xmm1, 2 * SIZE(CO2)
	movhpd	%xmm1, 3 * SIZE(CO2)

	movsd	%xmm2, 4 * SIZE(CO2)
	movhpd	%xmm2, 5 * SIZE(CO2)
	movsd	%xmm3, 6 * SIZE(CO2)
	movhpd	%xmm3, 7 * SIZE(CO2)

	addq	$8 * SIZE, CO1		# coffset += 4
	addq	$8 * SIZE, CO2		# coffset += 4

	decq	I			# i --
	jg	.L51
	ALIGN_4	

.L60:
	testq	$2, M
	je	.L70
	ALIGN_4

.L61:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

#ifdef BUFFERED
	leaq	16 * SIZE + BUFFER, BO
#else
	movq	B, BO
#endif
#else
	leaq	16 * SIZE + 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	-12 * SIZE(AO), %xmm2
	pxor	%xmm9, %xmm9
	movddup	-16 * SIZE(BO), %xmm1
	pxor	%xmm10, %xmm10
	movddup	-15 * 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	$2, %rax
#endif
	movq	%rax, KKK
#endif

	andq	$-4, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 2), BO
	negq	%rax
	je,pn	.L66
	ALIGN_4

.L62:
	mulpd	%xmm0, %xmm1
	addpd	%xmm1, %xmm8
	movddup	-14 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm0, %xmm3
	movapd	-14 * SIZE(AO, %rax, 2), %xmm0
	addpd	%xmm3, %xmm9
	movddup	-13 * SIZE(BO, %rax, 2), %xmm3
	mulpd	%xmm0, %xmm1
	addpd	%xmm1, %xmm10
	movddup	-12 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm0, %xmm3
	movapd	 -8 * SIZE(AO, %rax, 2), %xmm0
	addpd	%xmm3, %xmm11
	movddup	-11 * SIZE(BO, %rax, 2), %xmm3
	mulpd	%xmm2, %xmm1
	addpd	%xmm1, %xmm8
	movddup	-10 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm2, %xmm3
	movapd	-10 * SIZE(AO, %rax, 2), %xmm2
	addpd	%xmm3, %xmm9
	movddup	 -9 * SIZE(BO, %rax, 2), %xmm3
	mulpd	%xmm2, %xmm1
	addpd	%xmm1, %xmm10
	movddup	 -8 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm2, %xmm3
	movapd	 -4 * SIZE(AO, %rax, 2), %xmm2
	addpd	%xmm3, %xmm11
	movddup	 -7 * SIZE(BO, %rax, 2), %xmm3

	addq	$4 * SIZE, %rax
	jl,pt	.L62
	ALIGN_4

.L66:
	movapd	ALPHA, %xmm7

#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L69

	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 2), BO
	negq	%rax
	ALIGN_4

.L67:
	mulpd	%xmm0, %xmm1
	addpd	%xmm1, %xmm8
	movddup	-14 * SIZE(BO, %rax, 2), %xmm1
	mulpd	%xmm0, %xmm3
	movapd	-14 * SIZE(AO, %rax, 2), %xmm0
	addpd	%xmm3, %xmm9
	movddup	-13 * SIZE(BO, %rax, 2), %xmm3

	addq	$SIZE, %rax
	jl	.L67
	ALIGN_4

.L69:
	addpd	%xmm10, %xmm8
	addpd	%xmm11, %xmm9
	
	movsd	0 * SIZE(CO1), %xmm0
	movhpd	1 * SIZE(CO1), %xmm0
	movsd	2 * SIZE(CO1), %xmm1
	movhpd	3 * SIZE(CO1), %xmm1

	movddup	 %xmm8,  %xmm4
	unpckhpd %xmm8,  %xmm8

	mulpd	 %xmm7,  %xmm4
	mulpd	 %xmm7,  %xmm8
	addpd	 %xmm4,  %xmm0
	addpd	 %xmm8,  %xmm1

	movsd	%xmm0, 0 * SIZE(CO1)
	movhpd	%xmm0, 1 * SIZE(CO1)
	movsd	%xmm1, 2 * SIZE(CO1)
	movhpd	%xmm1, 3 * SIZE(CO1)

	movsd	0 * SIZE(CO2), %xmm0
	movhpd	1 * SIZE(CO2), %xmm0
	movsd	2 * SIZE(CO2), %xmm1
	movhpd	3 * SIZE(CO2), %xmm1

	movddup	 %xmm9,  %xmm4
	unpckhpd %xmm9,  %xmm9

	mulpd	 %xmm7,  %xmm4
	mulpd	 %xmm7,  %xmm9
	addpd	 %xmm4,  %xmm0
	addpd	 %xmm9,  %xmm1

	movsd	%xmm0, 0 * SIZE(CO2)
	movhpd	%xmm0, 1 * SIZE(CO2)
	movsd	%xmm1, 2 * SIZE(CO2)
	movhpd	%xmm1, 3 * SIZE(CO2)

	addq	$4 * SIZE, CO1		# coffset += 4
	addq	$4 * 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))

#ifdef BUFFERED
	leaq	16 * SIZE + BUFFER, BO
#else
	movq	B, BO
#endif
#else
	leaq	16 * SIZE + BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 2), BO
#endif	

	movddup	-16 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movddup	-15 * SIZE(AO), %xmm1
	pxor	%xmm9, %xmm9
	movddup	-14 * SIZE(AO), %xmm2
	pxor	%xmm10, %xmm10
	movddup	-13 * SIZE(AO), %xmm3
	pxor	%xmm11, %xmm11

#ifndef TRMMKERNEL
	movq	K, %rax