zgemm_kernel_4x2_barcelona.s

Optimized GotoBLAS libraries

	movaps	%xmm4,  16 * SIZE(BO)
	movaps	%xmm5,  20 * SIZE(BO)
	movaps	%xmm6,  24 * SIZE(BO)
	movaps	%xmm7,  28 * SIZE(BO)

	addq	$ 8 * SIZE, B
	addq	$32 * SIZE, BO

	decq	%rax
	jne	.L02
	ALIGN_4

.L03:
	movq	K, %rax
	andq	$3, %rax
	BRANCH
	jle	.L10
	ALIGN_4

.L04:
	movaps	 0 * SIZE(B), %xmm3

	pshufd	 $0x00, %xmm3, %xmm0
	pshufd	 $0x55, %xmm3, %xmm1
	pshufd	 $0xaa, %xmm3, %xmm2
	pshufd	 $0xff, %xmm3, %xmm3

#if defined(NN) || defined(NT) || defined(NR) || defined(NC)  || \
    defined(TN) || defined(TT) || defined(TR) || defined(TC) 
	xorps	 %xmm15, %xmm1
	xorps	 %xmm15, %xmm3
#else
	xorps	 %xmm15, %xmm0
	xorps	 %xmm15, %xmm2
#endif

	movaps	%xmm0,  0 * SIZE(BO)
	movaps	%xmm1,  4 * SIZE(BO)
	movaps	%xmm2,  8 * SIZE(BO)
	movaps	%xmm3, 12 * SIZE(BO)

	addq	$ 4 * SIZE, B
	addq	$16 * SIZE, BO
	decq	%rax
	jne	.L04
	ALIGN_4
	
.L10:
	movq	C, CO1			# coffset1 = c
	leaq	(C, LDC, 1), CO2	# coffset2 = c + ldc
	movq	A, AO		# aoffset = a

	movq	B, BB

	movq	M,  I
	sarq	$2, I		# i = (m >> 2)
	jle	.L20
	ALIGN_4

.L11:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leaq	32 * SIZE + BUFFER, BO
#else
	leaq	32 * SIZE + BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax,   8), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 8), BO
#endif	

	prefetch	 (RPREFETCHSIZE +  0)  * SIZE(BB)
	prefetch	 (RPREFETCHSIZE +  8)  * SIZE(BB)
	prefetch	 (RPREFETCHSIZE + 16)  * SIZE(BB)
	subq		 $-16 * SIZE, BB

	movaps	-32 * SIZE(AO), %xmm0
	movaps	-32 * SIZE(BO), %xmm1
	pxor	%xmm8, %xmm8
 	movaps	-28 * SIZE(BO), %xmm3
	pxor	%xmm9, %xmm9
	movaps	-16 * SIZE(AO), %xmm4
	pxor	%xmm10, %xmm10
	movaps	  0 * SIZE(BO), %xmm5
	pxor	%xmm11, %xmm11

	prefetchw      7 * SIZE(CO1)
	pxor	%xmm12, %xmm12
	prefetchw      7 * SIZE(CO2)
	pxor	%xmm13, %xmm13
	pxor	%xmm14, %xmm14
	pxor	%xmm15, %xmm15

	movaps	%xmm0, %xmm2

#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	$-8, %rax

	leaq	(, %rax, 8), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 8), BO
	negq	%rax
	je,pn	.L15
	ALIGN_3

.L12:
	KERNEL1(32 *  0)
	KERNEL2(32 *  0)
	KERNEL3(32 *  0)
	KERNEL4(32 *  0)
	KERNEL5(32 *  0)
	KERNEL6(32 *  0)
	KERNEL7(32 *  0)
	KERNEL8(32 *  0)
	je,pn	.L15
	KERNEL1(32 *  0)
	KERNEL2(32 *  0)
	KERNEL3(32 *  0)
	KERNEL4(32 *  0)
	KERNEL5(32 *  0)
	KERNEL6(32 *  0)
	KERNEL7(32 *  0)
	KERNEL8(32 *  0)
	je,pn	.L15
	KERNEL1(32 *  0)
	KERNEL2(32 *  0)
	KERNEL3(32 *  0)
	KERNEL4(32 *  0)
	KERNEL5(32 *  0)
	KERNEL6(32 *  0)
	KERNEL7(32 *  0)
	KERNEL8(32 *  0)
	je,pn	.L15
	KERNEL1(32 *  0)
	KERNEL2(32 *  0)
	KERNEL3(32 *  0)
	KERNEL4(32 *  0)
	KERNEL5(32 *  0)
	KERNEL6(32 *  0)
	KERNEL7(32 *  0)
	KERNEL8(32 *  0)
	je,pn	.L15
	KERNEL1(32 *  0)
	KERNEL2(32 *  0)
	KERNEL3(32 *  0)
	KERNEL4(32 *  0)
	KERNEL5(32 *  0)
	KERNEL6(32 *  0)
	KERNEL7(32 *  0)
	KERNEL8(32 *  0)
	je,pn	.L15
	KERNEL1(32 *  0)
	KERNEL2(32 *  0)
	KERNEL3(32 *  0)
	KERNEL4(32 *  0)
	KERNEL5(32 *  0)
	KERNEL6(32 *  0)
	KERNEL7(32 *  0)
	KERNEL8(32 *  0)
	je,pn	.L15
	KERNEL1(32 *  0)
	KERNEL2(32 *  0)
	KERNEL3(32 *  0)
	KERNEL4(32 *  0)
	KERNEL5(32 *  0)
	KERNEL6(32 *  0)
	KERNEL7(32 *  0)
	KERNEL8(32 *  0)
	je,pn	.L15
	KERNEL1(32 *  0)
	KERNEL2(32 *  0)
	KERNEL3(32 *  0)
	KERNEL4(32 *  0)
	KERNEL5(32 *  0)
	KERNEL6(32 *  0)
	KERNEL7(32 *  0)
	KERNEL8(32 *  0)
	jl,pt	.L12
	ALIGN_4

.L15:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	testq	$4, %rax
	je .L16
	xorq	%rax, %rax
	ALIGN_3

	KERNEL_SUB1(32 *  0)
	KERNEL_SUB2(32 *  0)
	KERNEL_SUB3(32 *  0)
	KERNEL_SUB4(32 *  0)

	addq	$64 * SIZE, BO
	addq	$32 * SIZE, AO
	ALIGN_3

.L16:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movaps	ALPHA_R, %xmm6
	movaps	ALPHA_I, %xmm7
	andq	$3, %rax		# if (k & 1)
	BRANCH
	je .L18

	leaq	(, %rax, 8), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 8), BO
	negq	%rax
	ALIGN_4

.L17:
	mulps	%xmm1, %xmm0
	mulps	-28 * SIZE(AO, %rax, 4), %xmm1
	addps	%xmm0, %xmm8
	movaps	%xmm2, %xmm0
	addps	%xmm1, %xmm12
	movaps	-24 * SIZE(BO, %rax, 8), %xmm1
	mulps	%xmm3, %xmm2
	mulps	-28 * SIZE(AO, %rax, 4), %xmm3
	addps	%xmm2, %xmm9
	movaps	%xmm0, %xmm2
	addps	%xmm3, %xmm13
	movaps	-20 * SIZE(BO, %rax, 8), %xmm3
	mulps	%xmm1, %xmm0
	mulps	-28 * SIZE(AO, %rax, 4), %xmm1
	addps	%xmm0, %xmm10
	movaps	-24 * SIZE(AO, %rax, 4), %xmm0
	addps	%xmm1, %xmm14
	movaps	-16 * SIZE(BO, %rax, 8), %xmm1
	mulps	%xmm3, %xmm2
	mulps	-28 * SIZE(AO, %rax, 4), %xmm3
	addps	%xmm2, %xmm11
	addps	%xmm3, %xmm15
 	movaps	-12 * SIZE(BO, %rax, 8), %xmm3
	movaps	%xmm0, %xmm2

	addq	$SIZE * 2, %rax
	jl	.L17
	ALIGN_4

.L18:
#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhps	2 * SIZE(CO1), %xmm0
	movsd	4 * SIZE(CO1), %xmm2
	movhps	6 * SIZE(CO1), %xmm2

	movsd	0 * SIZE(CO2), %xmm1
	movhps	2 * SIZE(CO2), %xmm1
	movsd	4 * SIZE(CO2), %xmm3
	movhps	6 * SIZE(CO2), %xmm3
#endif

	shufps	$0xb1, %xmm9, %xmm9
	shufps	$0xb1, %xmm11, %xmm11
	shufps	$0xb1, %xmm13, %xmm13
	shufps	$0xb1, %xmm15, %xmm15

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(RR) || defined(RC) || defined(CR) || defined(CC)
	subps	%xmm9, %xmm8
	subps	%xmm11, %xmm10
	subps	%xmm13, %xmm12
	subps	%xmm15, %xmm14
#else
	addps	%xmm9, %xmm8
	addps	%xmm11, %xmm10
	addps	%xmm13, %xmm12
	addps	%xmm15, %xmm14
#endif

	movaps	%xmm8, %xmm9
	movaps	%xmm10, %xmm11
	movaps	%xmm12, %xmm13
 	movaps	%xmm14, %xmm15

	shufps	$0xb1, %xmm8, %xmm8
	shufps	$0xb1, %xmm10, %xmm10
	shufps	$0xb1, %xmm12, %xmm12
	shufps	$0xb1, %xmm14, %xmm14

	mulps	%xmm6, %xmm9
	mulps	%xmm7, %xmm8
	mulps	%xmm6, %xmm11
	mulps	%xmm7, %xmm10

	mulps	%xmm6, %xmm13
	mulps	%xmm7, %xmm12
	mulps	%xmm6, %xmm15
	mulps	%xmm7, %xmm14

	addps	%xmm9, %xmm8
	addps	%xmm11, %xmm10
	addps	%xmm13, %xmm12
	addps	%xmm15, %xmm14

#ifndef TRMMKERNEL
	addps	%xmm0,  %xmm8
	addps	%xmm1,  %xmm10
	addps	%xmm2, %xmm12
	addps	%xmm3, %xmm14
#endif

	movsd	%xmm8, 0 * SIZE(CO1)
	movhps	%xmm8, 2 * SIZE(CO1)
	movsd	%xmm12, 4 * SIZE(CO1)
	movhps	%xmm12, 6 * SIZE(CO1)

	movsd	%xmm10, 0 * SIZE(CO2)
	movhps	%xmm10, 2 * SIZE(CO2)
	movsd	%xmm14, 4 * SIZE(CO2)
	movhps	%xmm14, 6 * SIZE(CO2)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movq	K, %rax
	subq	KKK, %rax
	leaq	(,%rax, 8), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 8), BO
#endif

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

	addq	$8 * SIZE, CO1		# coffset += 4
	addq	$8 * SIZE, CO2		# coffset += 4
	decq	I			# i --
	jg	.L11
	ALIGN_4	

.L20:
	testq	$2, M
	je	.L30

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

	leaq	32 * SIZE + BUFFER, BO
#else
	leaq	32 * SIZE + BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax,   8), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 8), BO
#endif	

	movaps	-32 * SIZE(AO), %xmm0
	movaps	-16 * SIZE(AO), %xmm2
	movaps	  0 * SIZE(AO), %xmm4
	movaps	 16 * SIZE(AO), %xmm6

	movaps	-32 * SIZE(BO), %xmm1
	movaps	-16 * SIZE(BO), %xmm3
	movaps	  0 * SIZE(BO), %xmm5
	movaps	 16 * SIZE(BO), %xmm7

	pxor	%xmm8, %xmm8
	pxor	%xmm9, %xmm9
	pxor	%xmm10, %xmm10
	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
	sarq	$3, %rax
	je	.L25
	ALIGN_4

.L22:
	mulps	%xmm0, %xmm1
	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)
	addps	%xmm1, %xmm8
	movaps	-28 * SIZE(BO), %xmm1
	mulps	%xmm0, %xmm1
	addps	%xmm1, %xmm9
	movaps	-24 * SIZE(BO), %xmm1
	mulps	%xmm0, %xmm1
	mulps	-20 * SIZE(BO), %xmm0
	addps	%xmm1, %xmm10
	movaps	 32 * SIZE(BO), %xmm1
	addps	%xmm0, %xmm11
	movaps	-28 * SIZE(AO), %xmm0

	mulps	%xmm0, %xmm3
	addps	%xmm3, %xmm8
	movaps	-12 * SIZE(BO), %xmm3
	mulps	%xmm0, %xmm3
	addps	%xmm3, %xmm9
	movaps	 -8 * SIZE(BO), %xmm3
	mulps	%xmm0, %xmm3
	mulps	 -4 * SIZE(BO), %xmm0
	addps	%xmm3, %xmm10
	movaps	 48 * SIZE(BO), %xmm3
	addps	%xmm0, %xmm11
	movaps	-24 * SIZE(AO), %xmm0

	mulps	%xmm0, %xmm5
	addps	%xmm5, %xmm8
	movaps	  4 * SIZE(BO), %xmm5
	mulps	%xmm0, %xmm5
	addps	%xmm5, %xmm9
	movaps	  8 * SIZE(BO), %xmm5
	mulps	%xmm0, %xmm5
	mulps	 12 * SIZE(BO), %xmm0
	addps	%xmm5, %xmm10
	movaps	 64 * SIZE(BO), %xmm5
	addps	%xmm0, %xmm11
	movaps	-20 * SIZE(AO), %xmm0

	mulps	%xmm0, %xmm7
	addps	%xmm7, %xmm8
	movaps	 20 * SIZE(BO), %xmm7
	mulps	%xmm0, %xmm7
	addps	%xmm7, %xmm9
	movaps	 24 * SIZE(BO), %xmm7
	mulps	%xmm0, %xmm7
	mulps	 28 * SIZE(BO), %xmm0
	addps	%xmm7, %xmm10
	movaps	 80 * SIZE(BO), %xmm7
	addps	%xmm0, %xmm11
	movaps	  0 * SIZE(AO), %xmm0

	mulps	%xmm2, %xmm1
	PREFETCH	(PREFETCHSIZE + 16) * SIZE(AO)
	addps	%xmm1, %xmm8
	movaps	 36 * SIZE(BO), %xmm1
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm9
	movaps	 40 * SIZE(BO), %xmm1
	mulps	%xmm2, %xmm1
	mulps	 44 * SIZE(BO), %xmm2
	addps	%xmm1, %xmm10
	movaps	 96 * SIZE(BO), %xmm1
	addps	%xmm2, %xmm11
	movaps	-12 * SIZE(AO), %xmm2

	mulps	%xmm2, %xmm3
	addps	%xmm3, %xmm8
	movaps	 52 * SIZE(BO), %xmm3
	mulps	%xmm2, %xmm3
	addps	%xmm3, %xmm9
	movaps	 56 * SIZE(BO), %xmm3
	mulps	%xmm2, %xmm3
	mulps	 60 * SIZE(BO), %xmm2
	addps	%xmm3, %xmm10
	movaps	112 * SIZE(BO), %xmm3
	addps	%xmm2, %xmm11
	movaps	 -8 * SIZE(AO), %xmm2

	mulps	%xmm2, %xmm5
	addps	%xmm5, %xmm8
	movaps	 68 * SIZE(BO), %xmm5
	mulps	%xmm2, %xmm5
	addps	%xmm5, %xmm9
	movaps	 72 * SIZE(BO), %xmm5
	mulps	%xmm2, %xmm5
	mulps	 76 * SIZE(BO), %xmm2
	addps	%xmm5, %xmm10
	movaps	128 * SIZE(BO), %xmm5
	addps	%xmm2, %xmm11
	movaps	 -4 * SIZE(AO), %xmm2

	mulps	%xmm2, %xmm7
	addps	%xmm7, %xmm8
	movaps	 84 * SIZE(BO), %xmm7
	mulps	%xmm2, %xmm7
	addps	%xmm7, %xmm9
	movaps	 88 * SIZE(BO), %xmm7
	mulps	%xmm2, %xmm7
	mulps	 92 * SIZE(BO), %xmm2
	addps	%xmm7, %xmm10
	movaps	144 * SIZE(BO), %xmm7
	addps	%xmm2, %xmm11
	movaps	 16 * SIZE(AO), %xmm2

	subq   $ -32 * SIZE, AO
	subq   $-128 * SIZE, BO

	decq   %rax
	jne    .L22
	ALIGN_4

.L25:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movaps	ALPHA_R, %xmm6
	movaps	ALPHA_I, %xmm7
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L28
	ALIGN_4

.L26:
	mulps	%xmm0, %xmm1
	addps	%xmm1, %xmm8
	movaps	-28 * SIZE(BO), %xmm1
	mulps	%xmm0, %xmm1
	addps	%xmm1, %xmm9
	movaps	-24 * SIZE(BO), %xmm1
	mulps	%xmm0, %xmm1
	mulps	-20 * SIZE(BO), %xmm0
	addps	%xmm1, %xmm10
	movaps	-16 * SIZE(BO), %xmm1
	addps	%xmm0, %xmm11
	movaps	-28 * SIZE(AO), %xmm0

	subq	$- 4 * SIZE, AO
	subq	$-16 * SIZE, BO
	decq	%rax
	jg	.L26
	ALIGN_4