zgemm_kernel_4x2_sse.s

Optimized GotoBLAS libraries

	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

.L28:
#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhps	2 * SIZE(CO1), %xmm0

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

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

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

	movaps	%xmm8, %xmm9
	movaps	%xmm10, %xmm11

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

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

	addps	%xmm9, %xmm8
	addps	%xmm11, %xmm10

#ifndef TRMMKERNEL
	addps	%xmm0,  %xmm8
	addps	%xmm1,  %xmm10
#endif

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

	movsd	%xmm10, 0 * SIZE(CO2)
	movhps	%xmm10, 2 * 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, 2), AO
	leaq	(BO, %rax, 8), BO
#endif

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

	addq	$4 * SIZE, CO1		# coffset += 4
	addq	$4 * SIZE, CO2		# coffset += 4
	ALIGN_4	
	
.L30:
	testq	$1, M
	je	.L39

#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, 1), AO
	leaq	(BO, %rax, 8), BO
#endif	

	movaps	-32 * SIZE(AO), %xmm0
	movaps	-24 * SIZE(AO), %xmm2

	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	$1, %rax
#else
	addq	$2, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$3, %rax
	je	.L35
	ALIGN_4

.L32:
	mulps	%xmm0, %xmm1
	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)
	addps	%xmm1, %xmm8
	movsd	-28 * SIZE(BO), %xmm1
	mulps	%xmm0, %xmm1
	addps	%xmm1, %xmm9
	movsd	-24 * SIZE(BO), %xmm1
	mulps	%xmm0, %xmm1
	addps	%xmm1, %xmm10
	movsd	-20 * SIZE(BO), %xmm1
	mulps	%xmm0, %xmm1
	movsd	-30 * SIZE(AO), %xmm0
	addps	%xmm1, %xmm11
	movsd	 32 * SIZE(BO), %xmm1

	mulps	%xmm0, %xmm3
	addps	%xmm3, %xmm8
	movsd	-12 * SIZE(BO), %xmm3
	mulps	%xmm0, %xmm3
	addps	%xmm3, %xmm9
	movsd	 -8 * SIZE(BO), %xmm3
	mulps	%xmm0, %xmm3
	addps	%xmm3, %xmm10
	movsd	 -4 * SIZE(BO), %xmm3
	mulps	%xmm0, %xmm3
	movsd	-28 * SIZE(AO), %xmm0
	addps	%xmm3, %xmm11
	movsd	 48 * SIZE(BO), %xmm3

	mulps	%xmm0, %xmm5
	addps	%xmm5, %xmm8
	movsd	  4 * SIZE(BO), %xmm5
	mulps	%xmm0, %xmm5
	addps	%xmm5, %xmm9
	movsd	  8 * SIZE(BO), %xmm5
	mulps	%xmm0, %xmm5
	addps	%xmm5, %xmm10
	movsd	 12 * SIZE(BO), %xmm5
	mulps	%xmm0, %xmm5
	movsd	-26 * SIZE(AO), %xmm0
	addps	%xmm5, %xmm11
	movsd	 64 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm7
	addps	%xmm7, %xmm8
	movsd	 20 * SIZE(BO), %xmm7
	mulps	%xmm0, %xmm7
	addps	%xmm7, %xmm9
	movsd	 24 * SIZE(BO), %xmm7
	mulps	%xmm0, %xmm7
	addps	%xmm7, %xmm10
	movsd	 28 * SIZE(BO), %xmm7
	mulps	%xmm0, %xmm7
	movsd	-16 * SIZE(AO), %xmm0
	addps	%xmm7, %xmm11
	movsd	 80 * SIZE(BO), %xmm7

	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm8
	movsd	 36 * SIZE(BO), %xmm1
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm9
	movsd	 40 * SIZE(BO), %xmm1
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm10
	movsd	 44 * SIZE(BO), %xmm1
	mulps	%xmm2, %xmm1
	movsd	-22 * SIZE(AO), %xmm2
	addps	%xmm1, %xmm11
	movsd	 96 * SIZE(BO), %xmm1

	mulps	%xmm2, %xmm3
	addps	%xmm3, %xmm8
	movsd	 52 * SIZE(BO), %xmm3
	mulps	%xmm2, %xmm3
	addps	%xmm3, %xmm9
	movsd	 56 * SIZE(BO), %xmm3
	mulps	%xmm2, %xmm3
	addps	%xmm3, %xmm10
	movsd	 60 * SIZE(BO), %xmm3
	mulps	%xmm2, %xmm3
	movsd	-20 * SIZE(AO), %xmm2
	addps	%xmm3, %xmm11
	movsd	112 * SIZE(BO), %xmm3

	mulps	%xmm2, %xmm5
	addps	%xmm5, %xmm8
	movsd	 68 * SIZE(BO), %xmm5
	mulps	%xmm2, %xmm5
	addps	%xmm5, %xmm9
	movsd	 72 * SIZE(BO), %xmm5
	mulps	%xmm2, %xmm5
	addps	%xmm5, %xmm10
	movsd	 76 * SIZE(BO), %xmm5
	mulps	%xmm2, %xmm5
	movsd	-18 * SIZE(AO), %xmm2
	addps	%xmm5, %xmm11
	movsd	128 * SIZE(BO), %xmm5

	mulps	%xmm2, %xmm7
	addps	%xmm7, %xmm8
	movsd	 84 * SIZE(BO), %xmm7
	mulps	%xmm2, %xmm7
	addps	%xmm7, %xmm9
	movsd	 88 * SIZE(BO), %xmm7
	mulps	%xmm2, %xmm7
	addps	%xmm7, %xmm10
	movsd	 92 * SIZE(BO), %xmm7
	mulps	%xmm2, %xmm7
	movsd	 -8 * SIZE(AO), %xmm2
	addps	%xmm7, %xmm11
	movsd	144 * SIZE(BO), %xmm7

	subq   $ -16 * SIZE, AO
	subq   $-128 * SIZE, BO

	decq   %rax
	jne    .L32
	ALIGN_4

.L35:
#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 .L38
	ALIGN_4

.L36:
	mulps	%xmm0, %xmm1
	addps	%xmm1, %xmm8
	movsd	-28 * SIZE(BO), %xmm1
	mulps	%xmm0, %xmm1
	addps	%xmm1, %xmm9
	movsd	-24 * SIZE(BO), %xmm1
	mulps	%xmm0, %xmm1
	addps	%xmm1, %xmm10
	movsd	-20 * SIZE(BO), %xmm1
	mulps	%xmm0, %xmm1
	movsd	-30 * SIZE(AO), %xmm0
	addps	%xmm1, %xmm11
	movsd	-16 * SIZE(BO), %xmm1

	subq	$ -2 * SIZE, AO
	subq	$-16 * SIZE, BO
	decq	%rax
	jg	.L36
	ALIGN_4

.L38:
#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movsd	0 * SIZE(CO2), %xmm1
#endif

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

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

	movaps	%xmm8, %xmm9
	movaps	%xmm10, %xmm11

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

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

	addps	%xmm9, %xmm8
	addps	%xmm11, %xmm10

#ifndef TRMMKERNEL
	addps	%xmm0,  %xmm8
	addps	%xmm1,  %xmm10
#endif

	movsd	%xmm8, 0 * SIZE(CO1)
	movsd	%xmm10, 0 * 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, 1), AO
	leaq	(BO, %rax, 8), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$1, KK
#endif
	ALIGN_4	

.L39:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$2, KK
#endif

	leaq	(C, LDC, 2), C		# c += 2 * ldc
	decq	J			# j --
	jg	.L01
	ALIGN_4

.L40:
	testq	$1, N
	je	.L999
	ALIGN_4

.L41:
#if defined(TRMMKERNEL) && defined(LEFT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif	

/* Copying to Sub Buffer */
	leaq	BUFFER, BO
	movaps	POSINV, %xmm7
	
	movq	K, %rax
	sarq	$2, %rax
	jle	.L43
	ALIGN_4
	
.L42:
	movss	 0 * SIZE(B), %xmm8
	movss	 1 * SIZE(B), %xmm9
	movss	 2 * SIZE(B), %xmm10
	movss	 3 * SIZE(B), %xmm11
	movss	 4 * SIZE(B), %xmm12
	movss	 5 * SIZE(B), %xmm13
	movss	 6 * SIZE(B), %xmm14
	movss	 7 * SIZE(B), %xmm15

	shufps	$0, %xmm8, %xmm8
	shufps	$0, %xmm9, %xmm9
	shufps	$0, %xmm10, %xmm10
	shufps	$0, %xmm11, %xmm11
	shufps	$0, %xmm12, %xmm12
	shufps	$0, %xmm13, %xmm13
	shufps	$0, %xmm14, %xmm14
	shufps	$0, %xmm15, %xmm15

#if defined(NN) || defined(NT) || defined(NR) || defined(NC)  || \
    defined(TN) || defined(TT) || defined(TR) || defined(TC) 
	xorps	 %xmm7, %xmm9
	xorps	 %xmm7, %xmm11
	xorps	 %xmm7, %xmm13
	xorps	 %xmm7, %xmm15
#else
	xorps	 %xmm7, %xmm8
	xorps	 %xmm7, %xmm10
	xorps	 %xmm7, %xmm12
	xorps	 %xmm7, %xmm14
#endif

#if defined(OPTERON) || defined(BARCELONA)
	prefetchnta	 56 * SIZE(B)
#endif

	movaps	%xmm8,  0 * SIZE(BO)
	movaps	%xmm9,  4 * SIZE(BO)
	movaps	%xmm10,  8 * SIZE(BO)
	movaps	%xmm11, 12 * SIZE(BO)
	movaps	%xmm12, 16 * SIZE(BO)
	movaps	%xmm13, 20 * SIZE(BO)
	movaps	%xmm14, 24 * SIZE(BO)
	movaps	%xmm15, 28 * SIZE(BO)

#if defined(PENTIUM4) || defined(GENERIC)
        PREFETCHW	128 * SIZE(BO)
	PREFETCH	112 * SIZE(B)
#endif

	addq	$32 * SIZE, BO
	addq	$ 8 * SIZE, B
	decq	%rax
	jne	.L42
	ALIGN_4

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

.L44:
	movss	 0 * SIZE(B), %xmm8
	movss	 1 * SIZE(B), %xmm9

	shufps	$0, %xmm8, %xmm8
	shufps	$0, %xmm9, %xmm9

#if defined(NN) || defined(NT) || defined(NR) || defined(NC)  || \
    defined(TN) || defined(TT) || defined(TR) || defined(TC) 
	xorps	 %xmm7, %xmm9
#else
	xorps	 %xmm7, %xmm8
#endif

	movaps	%xmm8,  0 * SIZE(BO)
	movaps	%xmm9,  4 * SIZE(BO)

	addq	$2 * SIZE, B
	addq	$8 * SIZE, BO
	decq	%rax
	jne	.L44
	ALIGN_4
	
.L50:
	movq	C, CO1			# coffset1 = c
	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))

	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, 4), BO
#endif	

	movaps	-32 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movaps	-16 * SIZE(AO), %xmm2
	pxor	%xmm9, %xmm9
	movaps	  0 * SIZE(AO), %xmm4
	pxor	%xmm10, %xmm10
	movaps	 16 * SIZE(AO), %xmm6
	pxor	%xmm11, %xmm11

	movaps	-32 * SIZE(BO), %xmm1
	pxor	%xmm12, %xmm12
	movaps	-16 * SIZE(BO), %xmm3
	pxor	%xmm13, %xmm13
	movaps	  0 * SIZE(BO), %xmm5
	pxor	%xmm14, %xmm14
	movaps	 16 * SIZE(BO), %xmm7
	pxor	%xmm15, %xmm15

	PREFETCHW      7 * 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	.L55
	ALIGN_4

.L52:
	mulps	%xmm0, %xmm1
	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)
	mulps	-28 * SIZE(BO), %xmm0
	addps	%xmm1, %xmm8
	movaps	-32 * SIZE(BO), %xmm1
	addps	%xmm0, %xmm9
	movaps	-28 * SIZE(AO), %xmm0
	mulps	%xmm0, %xmm1
	mulps	-28 * SIZE(BO), %xmm0
	addps	%xmm1, %xmm12
	movaps	-24 * SIZE(BO), %xmm1
	addps	%xmm0, %xmm13
	movaps	-24 * SIZE(AO), %xmm0

	mulps	%xmm0, %xmm1
	mulps	-20 * SIZE(BO), %xmm0
	addps	%xmm1, %xmm8
	movaps	-24 * SIZE(BO), %xmm1
	addps	%xmm0, %xmm9
	movaps	-20 * SIZE(AO), %xmm0
	mulps	%xmm0, %xmm1
	mulps	-20 * SIZE(BO), %xmm0
	addps	%xmm1, %xmm12
	movaps	 32 * SIZE(BO), %xmm1
	addps	%xmm0, %xmm13
	movaps	 32 * SIZE(AO), %xmm0

	PREFETCH	(PREFETCHSIZE + 16) * SIZE(AO)

	mulps	%xmm2, %xmm3
	mulps	-12 * SIZE(BO), %xmm2
	addps	%xmm3, %xmm8
	movaps	-16 * SIZE(BO), %xmm3
	addps	%xmm2, %xmm9
	movaps	-12 * SIZE(AO), %xmm2
	mulps	%xmm2, %xmm3
	mulps	-12 * SIZE(BO), %xmm2
	addps	%xmm3, %xmm12
	movaps	 -8 * SIZE(BO), %xmm3
	addps	%xmm2, %xmm13
	movaps	 -8 * SIZE(AO), %xmm2

	mulps	%xmm2, %xmm3
	mulps	 -4 * SIZE(BO), %xmm2
	addps	%xmm3, %xmm8