zgemm3m_kernel_4x4_barcelona.s

Optimized GotoBLAS libraries

#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	$4, %rax
#endif
	movq	%rax, KKK
#endif

	andq	$-8, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 4), BO
	negq	%rax
	je,pn	.L15
	ALIGN_4

.L12:
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)
	je,pn	.L15
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)
	je,pn	.L15
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)
	je,pn	.L15
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)
	je,pn	.L15
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)
	je,pn	.L15
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)
	je,pn	.L15
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)
	je,pn	.L15
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)
	jl,pt	.L12
	ALIGN_4

.L15:
	movapd	ALPHA, %xmm7

#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	testq	$4, %rax
	je .L16
	xorq	%rax, %rax
	ALIGN_4

	KERNEL_SUB1(16 *  0)
	KERNEL_SUB2(16 *  0)
	KERNEL_SUB3(16 *  0)
	KERNEL_SUB4(16 *  0)

	subq	$-16 * SIZE, BO
	subq	$-16 * SIZE, AO
	ALIGN_4

.L16:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L19

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

.L17:
	mulpd	%xmm1, %xmm0
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm8
	movapd	%xmm2, %xmm0
	addpd	%xmm1, %xmm12
	movddup	-14 * SIZE(BO, %rax, 4), %xmm1
	mulpd	%xmm3, %xmm2
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm3
	addpd	%xmm2, %xmm9
	movapd	%xmm0, %xmm2
	addpd	%xmm3, %xmm13
	movddup	-13 * SIZE(BO, %rax, 4), %xmm3
	mulpd	%xmm1, %xmm0
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm10
	movapd	-12 * SIZE(AO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm14
	movddup	-12 * SIZE(BO, %rax, 4), %xmm1
	mulpd	%xmm3, %xmm2
	mulpd	-14 * SIZE(AO, %rax, 4), %xmm3
	addpd	%xmm2, %xmm11
	addpd	%xmm3, %xmm15
 	movddup	-11 * SIZE(BO, %rax, 4), %xmm3
	movapd	%xmm0, %xmm2

	addq	$SIZE, %rax
	jl	.L17
	ALIGN_4

.L19:
	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)
	
	movsd	0 * SIZE(CO1, LDC, 2), %xmm0
	movhpd	1 * SIZE(CO1, LDC, 2), %xmm0
	movsd	2 * SIZE(CO1, LDC, 2), %xmm1
	movhpd	3 * SIZE(CO1, LDC, 2), %xmm1

	movsd	4 * SIZE(CO1, LDC, 2), %xmm2
	movhpd	5 * SIZE(CO1, LDC, 2), %xmm2
	movsd	6 * SIZE(CO1, LDC, 2), %xmm3
	movhpd	7 * SIZE(CO1, LDC, 2), %xmm3

	movddup	 %xmm10, %xmm4
	unpckhpd %xmm10, %xmm10
	movddup	 %xmm14, %xmm5
	unpckhpd %xmm14, %xmm14

	mulpd	 %xmm7,  %xmm4
	mulpd	 %xmm7,  %xmm10
	mulpd	 %xmm7,  %xmm5
	mulpd	 %xmm7,  %xmm14

	addpd	 %xmm4,  %xmm0
	addpd	 %xmm10, %xmm1
	addpd	 %xmm5,  %xmm2
	addpd	 %xmm14, %xmm3

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

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

	movsd	4 * SIZE(CO2, LDC, 2), %xmm2
	movhpd	5 * SIZE(CO2, LDC, 2), %xmm2
	movsd	6 * SIZE(CO2, LDC, 2), %xmm3
	movhpd	7 * SIZE(CO2, LDC, 2), %xmm3

	movddup	 %xmm11, %xmm4
	unpckhpd %xmm11, %xmm11
	movddup	 %xmm15, %xmm5
	unpckhpd %xmm15, %xmm15

	mulpd	 %xmm7,  %xmm4
	mulpd	 %xmm7,  %xmm11
	mulpd	 %xmm7,  %xmm5
	mulpd	 %xmm7,  %xmm15

	addpd	 %xmm4,  %xmm0
	addpd	 %xmm11, %xmm1
	addpd	 %xmm5,  %xmm2
	addpd	 %xmm15, %xmm3

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

	movsd	%xmm2, 4 * SIZE(CO2, LDC, 2)
	movhpd	%xmm2, 5 * SIZE(CO2, LDC, 2)
	movsd	%xmm3, 6 * SIZE(CO2, LDC, 2)
	movhpd	%xmm3, 7 * SIZE(CO2, LDC, 2)

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

	decq	I			# i --
	jg,pt	.L11
	ALIGN_4	

.L20:
	testq	$3, M
	je	.L39

	testq	$2, M
	je	.L30
	ALIGN_4

.L21:
#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, 4), 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), %xmm5
	pxor	%xmm11, %xmm11
	movddup	 -8 * SIZE(BO), %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	$4, %rax
#endif
	movq	%rax, KKK
#endif

	andq	$-4, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 4), BO
	negq	%rax
	je,pn	.L26
	ALIGN_4

.L22:
	mulpd	%xmm0, %xmm1
	addpd	%xmm1, %xmm8
	movddup	-14 * SIZE(BO, %rax, 4), %xmm1
	mulpd	%xmm0, %xmm5
	addpd	%xmm5, %xmm9
	movddup	-13 * SIZE(BO, %rax, 4), %xmm5
	mulpd	%xmm0, %xmm1
	addpd	%xmm1, %xmm10
	movddup	-12 * SIZE(BO, %rax, 4), %xmm1
	mulpd	%xmm0, %xmm5
	movapd	-14 * SIZE(AO, %rax, 2), %xmm0
	addpd	%xmm5, %xmm11
	movddup	-11 * SIZE(BO, %rax, 4), %xmm5
	mulpd	%xmm0, %xmm1
	addpd	%xmm1, %xmm8
	movddup	-10 * SIZE(BO, %rax, 4), %xmm1
	mulpd	%xmm0, %xmm5
	addpd	%xmm5, %xmm9
	movddup	 -9 * SIZE(BO, %rax, 4), %xmm5
	mulpd	%xmm0, %xmm1
	addpd	%xmm1, %xmm10
	movddup	  (BO, %rax, 4), %xmm1
	mulpd	%xmm0, %xmm5
	movapd	 -8 * SIZE(AO, %rax, 2), %xmm0
	addpd	%xmm5, %xmm11
	movddup	 -7 * SIZE(BO, %rax, 4), %xmm5
	mulpd	%xmm2, %xmm3
	addpd	%xmm3, %xmm8
	movddup	 -6 * SIZE(BO, %rax, 4), %xmm3
	mulpd	%xmm2, %xmm5
	addpd	%xmm5, %xmm9
	movddup	 -5 * SIZE(BO, %rax, 4), %xmm5
	mulpd	%xmm2, %xmm3
	addpd	%xmm3, %xmm10
	movddup	 -4 * SIZE(BO, %rax, 4), %xmm3
	mulpd	%xmm2, %xmm5
	movapd	-10 * SIZE(AO, %rax, 2), %xmm2
	addpd	%xmm5, %xmm11
	movddup	 -3 * SIZE(BO, %rax, 4), %xmm5
	mulpd	%xmm2, %xmm3
	addpd	%xmm3, %xmm8
	movddup	 -2 * SIZE(BO, %rax, 4), %xmm3
	mulpd	%xmm2, %xmm5
	addpd	%xmm5, %xmm9
	movddup	 -1 * SIZE(BO, %rax, 4), %xmm5
	mulpd	%xmm2, %xmm3
	addpd	%xmm3, %xmm10
	movddup	  8 * SIZE(BO, %rax, 4), %xmm3
	mulpd	%xmm2, %xmm5
	movapd	 -4 * SIZE(AO, %rax, 2), %xmm2
	addpd	%xmm5, %xmm11
	movddup	  1 * SIZE(BO, %rax, 4), %xmm5

	addq	$4 * SIZE, %rax
	jl,pt	.L22
	ALIGN_4

.L26:
	movapd	ALPHA, %xmm7

#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L29

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

.L27:
	mulpd	%xmm0, %xmm1
	addpd	%xmm1, %xmm8
	movddup	-14 * SIZE(BO, %rax, 4), %xmm1
	mulpd	%xmm0, %xmm5
	addpd	%xmm5, %xmm9
	movddup	-13 * SIZE(BO, %rax, 4), %xmm5
	mulpd	%xmm0, %xmm1
	addpd	%xmm1, %xmm10
	movddup	-12 * SIZE(BO, %rax, 4), %xmm1
	mulpd	%xmm0, %xmm5
	movapd	-14 * SIZE(AO, %rax, 2), %xmm0
	addpd	%xmm5, %xmm11
	movddup	-11 * SIZE(BO, %rax, 4), %xmm5

	addq	$SIZE, %rax
	jl	.L27
	ALIGN_4

.L29:
	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)

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

	movddup	 %xmm10, %xmm4
	unpckhpd %xmm10, %xmm10

	mulpd	 %xmm7,  %xmm4
	mulpd	 %xmm7,  %xmm10
	addpd	 %xmm4,  %xmm0
	addpd	 %xmm10, %xmm1

	movsd	%xmm0, 0 * SIZE(CO1, LDC, 2)
	movhpd	%xmm0, 1 * SIZE(CO1, LDC, 2)
	movsd	%xmm1, 2 * SIZE(CO1, LDC, 2)
	movhpd	%xmm1, 3 * SIZE(CO1, LDC, 2)
	
	movsd	0 * SIZE(CO2, LDC, 2), %xmm0
	movhpd	1 * SIZE(CO2, LDC, 2), %xmm0
	movsd	2 * SIZE(CO2, LDC, 2), %xmm1
	movhpd	3 * SIZE(CO2, LDC, 2), %xmm1

	movddup	 %xmm11, %xmm4
	unpckhpd %xmm11, %xmm11

	mulpd	 %xmm7,  %xmm4
	mulpd	 %xmm7,  %xmm11
	addpd	 %xmm4,  %xmm0
	addpd	 %xmm11, %xmm1

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

	addq	$4 * SIZE, CO1
	addq	$4 * SIZE, CO2
	ALIGN_4	

.L30:
	testq	$1, M
	je	.L39

#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, 4), BO
#endif	

	movddup	-16 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movddup	-14 * SIZE(AO), %xmm2
	pxor	%xmm9, %xmm9
	movddup	-15 * SIZE(AO), %xmm4
	pxor	%xmm10, %xmm10
	movapd	-16 * SIZE(BO), %xmm1
	pxor	%xmm11, %xmm11
	movapd	 -8 * SIZE(BO), %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	$4, %rax
#endif
	movq	%rax, KKK
#endif

	andq	$-4, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 4), BO
	negq	%rax
	je,pn	.L36
	ALIGN_4

.L32:
	mulpd	%xmm0, %xmm1
	mulpd	-14 * SIZE(BO, %rax, 4), %xmm0
	addpd	%xmm1, %xmm8
	movapd	-12 * SIZE(BO, %rax, 4), %xmm1
	addpd	%xmm0, %xmm9
	movddup	-12 * SIZE(AO, %rax, 1), %xmm0
	mulpd	%xmm4, %xmm1
	mulpd	-10 * SIZE(BO, %rax, 4), %xmm4