zgemm_kernel_2x2_sse2.s

Optimized GotoBLAS libraries

	movlpd	0 * SIZE(CO2), %xmm1
	movhpd	1 * SIZE(CO2), %xmm1
#endif

	SHUFPD_1 %xmm9, %xmm9
	SHUFPD_1 %xmm11, %xmm11

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(NR) || defined(NC) || defined(TR) || defined(TC)
	xorpd	%xmm5, %xmm9
	xorpd	%xmm5, %xmm11
#else
	xorpd	%xmm5, %xmm8
	xorpd	%xmm5, %xmm10
#endif

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

	pshufd	$0x4e, %xmm8, %xmm9
	pshufd	$0x4e, %xmm10, %xmm11
	
	mulpd	%xmm6, %xmm8
	mulpd	%xmm7, %xmm9
	mulpd	%xmm6, %xmm10
	mulpd	%xmm7, %xmm11

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

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

	movlpd	%xmm8, 0 * SIZE(CO1)
	movhpd	%xmm8, 1 * SIZE(CO1)
	movlpd	%xmm10, 0 * SIZE(CO2)
	movhpd	%xmm10, 1 * 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, 2), AO
	leaq	(BO, %rax, 8), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$1, KK
#endif
	ALIGN_4
	
.L99:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$2, KK
#endif

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

.L100:
	testq	$1, N
	jle	.L999

.L101:
#if defined(TRMMKERNEL) && defined(LEFT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif	

/* Copying to Sub Buffer */
	leaq	BUFFER, BO

	movq	K, %rax
	sarq	$2, %rax
	jle	.L103
	ALIGN_4
	
.L102:
	movlpd	 0 * SIZE(B), %xmm8
	movlpd	 1 * SIZE(B), %xmm9
	movlpd	 2 * SIZE(B), %xmm10
	movlpd	 3 * SIZE(B), %xmm11
	movlpd	 4 * SIZE(B), %xmm12
	movlpd	 5 * SIZE(B), %xmm13
	movlpd	 6 * SIZE(B), %xmm14
	movlpd	 7 * SIZE(B), %xmm15

	movlpd	%xmm8,  0 * SIZE(BO)
	movlpd	%xmm8,  1 * SIZE(BO)
	movlpd	%xmm9,  2 * SIZE(BO)
	movlpd	%xmm9,  3 * SIZE(BO)
	movlpd	%xmm10,  4 * SIZE(BO)
	movlpd	%xmm10,  5 * SIZE(BO)
	movlpd	%xmm11,  6 * SIZE(BO)
	movlpd	%xmm11,  7 * SIZE(BO)
	movlpd	%xmm12,  8 * SIZE(BO)
	movlpd	%xmm12,  9 * SIZE(BO)
	movlpd	%xmm13, 10 * SIZE(BO)
	movlpd	%xmm13, 11 * SIZE(BO)
	movlpd	%xmm14, 12 * SIZE(BO)
	movlpd	%xmm14, 13 * SIZE(BO)
	movlpd	%xmm15, 14 * SIZE(BO)
	movlpd	%xmm15, 15 * SIZE(BO)

	subq	$-16 * SIZE, BO
	addq	$ 8 * SIZE, B
	decq	%rax
	jne	.L102
	ALIGN_4

.L103:
	movq	K, %rax
	andq	$3, %rax
	BRANCH
	jle	.L105
	ALIGN_4

.L104:
	movlpd	 0 * SIZE(B), %xmm8
	movlpd	 1 * SIZE(B), %xmm9

	movlpd	%xmm8,  0 * SIZE(BO)
	movlpd	%xmm8,  1 * SIZE(BO)
	movlpd	%xmm9,  2 * SIZE(BO)
	movlpd	%xmm9,  3 * SIZE(BO)

	addq	$4 * SIZE, BO
	addq	$2 * SIZE, B
	decq	%rax
	jne	.L104
	ALIGN_4
	
.L105:
	movq	C, CO1		# coffset1 = c
	movq	A, AO		# aoffset = a

	movq	M,  I
	sarq	$1, I		# i = (m >> 2)
	jle	.L130
	ALIGN_4

.L110:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

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

	movapd	-16 * SIZE(AO), %xmm0
	pxor	%xmm8, %xmm8
	movapd	-16 * SIZE(BO), %xmm1
	pxor	%xmm9, %xmm9
	movapd	 -8 * SIZE(AO), %xmm2
	pxor	%xmm12, %xmm12
	movapd	 -8 * SIZE(BO), %xmm3
	pxor	%xmm13, %xmm13
	PREFETCHW      3 * 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	$2, %rax
#else
	addq	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	je	.L112

.L111:
	mulpd	%xmm0, %xmm1
	PREFETCH	(PREFETCHSIZE     +  0) * SIZE(AO)
	mulpd	-14 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm8
	movapd	-16 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm9
	movapd	-14 * SIZE(AO), %xmm0
	mulpd	%xmm0, %xmm1
	mulpd	-14 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm12
	movapd	-12 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm13
	movapd	-12 * SIZE(AO), %xmm0
	mulpd	%xmm0, %xmm1
	mulpd	-10 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm8
	movapd	-12 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm9
	movapd	-10 * SIZE(AO), %xmm0
	mulpd	%xmm0, %xmm1
	mulpd	-10 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm12
	movapd	  0 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm13
	movapd	  0 * SIZE(AO), %xmm0
	mulpd	%xmm2, %xmm3
	PREFETCH	(PREFETCHSIZE     +  8) * SIZE(AO)
	mulpd	 -6 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm8
	movapd	 -8 * SIZE(BO), %xmm3
	addpd	%xmm2, %xmm9
	movapd	 -6 * SIZE(AO), %xmm2
	mulpd	%xmm2, %xmm3
	mulpd	 -6 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm12
	movapd	 -4 * SIZE(BO), %xmm3
	addpd	%xmm2, %xmm13
	movapd	 -4 * SIZE(AO), %xmm2
	mulpd	%xmm2, %xmm3
	mulpd	 -2 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm8
	movapd	 -4 * SIZE(BO), %xmm3
	addpd	%xmm2, %xmm9
	movapd	 -2 * SIZE(AO), %xmm2
	mulpd	%xmm2, %xmm3
	mulpd	 -2 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm12
	movapd	  8 * SIZE(BO), %xmm3
	addpd	%xmm2, %xmm13
	movapd	  8 * SIZE(AO), %xmm2

	subq   $-16 * SIZE, AO
	subq   $-16 * SIZE, BO
	decq   %rax
	jne    .L111
	ALIGN_4

.L112:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movapd	POSINV,  %xmm5
	movapd	ALPHA_R, %xmm6
	movapd	ALPHA_I, %xmm7
	andq	$3, %rax		# if (k & 1)
	BRANCH
	jle .L114

.L113:
	mulpd	%xmm0, %xmm1
	mulpd	 -14 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm8
	movapd	 -16 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm9
	movapd	 -14 * SIZE(AO), %xmm0
	mulpd	%xmm0, %xmm1
	mulpd	 -14 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm12
	movapd	 -12 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm13
	movapd	 -12 * SIZE(AO), %xmm0

	addq	$4 * SIZE, AO		# aoffset  += 4
	addq	$4 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L113
	ALIGN_4

.L114:
#ifndef TRMMKERNEL
	movlpd	0 * SIZE(CO1), %xmm0
	movhpd	1 * SIZE(CO1), %xmm0
	movlpd	2 * SIZE(CO1), %xmm2
	movhpd	3 * SIZE(CO1), %xmm2
#endif

	SHUFPD_1 %xmm9, %xmm9
	SHUFPD_1 %xmm13, %xmm13

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(NR) || defined(NC) || defined(TR) || defined(TC)
	xorpd	%xmm5, %xmm9
	xorpd	%xmm5, %xmm13
#else
	xorpd	%xmm5, %xmm8
	xorpd	%xmm5, %xmm12
#endif

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(RR) || defined(RC) || defined(CR) || defined(CC)
	subpd	%xmm9, %xmm8
	subpd	%xmm13, %xmm12
#else
	addpd	%xmm9, %xmm8
	addpd	%xmm13, %xmm12
#endif

	pshufd	$0x4e, %xmm8, %xmm9
	pshufd	$0x4e, %xmm12, %xmm13
	
	mulpd	%xmm6, %xmm8
	mulpd	%xmm7, %xmm9
	mulpd	%xmm6, %xmm12
	mulpd	%xmm7, %xmm13

	addpd	%xmm9, %xmm8
	addpd	%xmm13, %xmm12

#ifndef TRMMKERNEL
	addpd	%xmm0,  %xmm8
	addpd	%xmm2, %xmm12
#endif

	movlpd	%xmm8, 0 * SIZE(CO1)
	movhpd	%xmm8, 1 * SIZE(CO1)
	movlpd	%xmm12, 2 * SIZE(CO1)
	movhpd	%xmm12, 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, 4), BO
#endif

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

	addq	$4 * SIZE, CO1		# coffset += 4
	decq	I			# i --
	jg	.L110
	ALIGN_4	

.L130:
	testq	$1, M
	jle	.L999

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

	leaq	16 * SIZE + BUFFER, BO
#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
	movapd	-16 * SIZE(BO), %xmm1
	movapd	 -8 * SIZE(AO), %xmm2
	movapd	 -8 * SIZE(BO), %xmm3

	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	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$3, %rax
	je	.L144
	ALIGN_4

.L141:
	mulpd	%xmm0, %xmm1
	PREFETCH	(PREFETCHSIZE     +  0) * SIZE(AO)
	mulpd	-14 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm8
	movapd	-12 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm9
	movapd	-14 * SIZE(AO), %xmm0
	mulpd	%xmm0, %xmm1
	mulpd	-10 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm10
	movapd	  0 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm11
	movapd	-12 * SIZE(AO), %xmm0
	mulpd	%xmm0, %xmm3
	mulpd	 -6 * SIZE(BO), %xmm0
	addpd	%xmm3, %xmm8
	movapd	 -4 * SIZE(BO), %xmm3
	addpd	%xmm0, %xmm9
	movapd	-10 * SIZE(AO), %xmm0
	mulpd	%xmm0, %xmm3
	mulpd	 -2 * SIZE(BO), %xmm0
	addpd	%xmm3, %xmm10
	movapd	  8 * SIZE(BO), %xmm3
	addpd	%xmm0, %xmm11
	movapd	  0 * SIZE(AO), %xmm0
	mulpd	%xmm2, %xmm1
	PREFETCH	(PREFETCHSIZE     +  8) * SIZE(AO)
	mulpd	  2 * SIZE(BO), %xmm2
	addpd	%xmm1, %xmm8
	movapd	  4 * SIZE(BO), %xmm1
	addpd	%xmm2, %xmm9
	movapd	 -6 * SIZE(AO), %xmm2
	mulpd	%xmm2, %xmm1
	mulpd	  6 * SIZE(BO), %xmm2
	addpd	%xmm1, %xmm10
	movapd	 16 * SIZE(BO), %xmm1
	addpd	%xmm2, %xmm11
	movapd	 -4 * SIZE(AO), %xmm2
	mulpd	%xmm2, %xmm3
	mulpd	 10 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm8
	movapd	 12 * SIZE(BO), %xmm3
	addpd	%xmm2, %xmm9
	movapd	 -2 * SIZE(AO), %xmm2
	mulpd	%xmm2, %xmm3
	mulpd	 14 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm10
	movapd	 24 * SIZE(BO), %xmm3
	addpd	%xmm2, %xmm11
	movapd	  8 * SIZE(AO), %xmm2

	subq   $-16 * SIZE, AO
	subq   $-32 * SIZE, BO
	decq   %rax
	jne    .L141
	ALIGN_4


.L144:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$4, %rax		# if (k & 1)
	BRANCH
	jle .L145

	mulpd	%xmm0, %xmm1
	mulpd	-14 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm8
	movapd	-12 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm9
	movapd	-14 * SIZE(AO), %xmm0
	mulpd	%xmm0, %xmm1
	mulpd	-10 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm10
	movapd	  0 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm11
	movapd	-12 * SIZE(AO), %xmm0
	mulpd	%xmm0, %xmm3
	mulpd	 -6 * SIZE(BO), %xmm0
	addpd	%xmm3, %xmm8
	movapd	 -4 * SIZE(BO), %xmm3
	addpd	%xmm0, %xmm9
	movapd	-10 * SIZE(AO), %xmm0
	mulpd	%xmm0, %xmm3
	mulpd	 -2 * SIZE(BO), %xmm0
	addpd	%xmm3, %xmm10
	addpd	%xmm0, %xmm11
	movapd	 -8 * SIZE(AO), %xmm0

	addq   $8   * SIZE, AO
	subq   $-16 * SIZE, BO
	ALIGN_4

.L145:
	movapd	POSINV, %xmm5
	movapd	ALPHA_R, %xmm6
	movapd	ALPHA_I, %xmm7

#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	BRANCH
	jle .L148
	ALIGN_4

.L146:
	mulpd	%xmm0, %xmm1
	mulpd	-14 * SIZE(BO), %xmm0
	addpd	%xmm1, %xmm8
	movapd	-12 * SIZE(BO), %xmm1
	addpd	%xmm0, %xmm9
	movapd	-14 * SIZE(AO), %xmm0

	addq	$2 * SIZE, AO		# aoffset  += 4
	addq	$4 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L146
	ALIGN_4

.L148:
	addpd	%xmm10, %xmm8
	addpd	%xmm11, %xmm9

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

	SHUFPD_1 %xmm9, %xmm9

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(NR) || defined(NC) || defined(TR) || defined(TC)
	xorpd	%xmm5, %xmm9
#else
	xorpd	%xmm5, %xmm8
#endif

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(RR) || defined(RC) || defined(CR) || defined(CC)
	subpd	%xmm9, %xmm8
#else
	addpd	%xmm9, %xmm8
#endif

	pshufd	$0x4e, %xmm8, %xmm9
	
	mulpd	%xmm6, %xmm8
	mulpd	%xmm7, %xmm9

	addpd	%xmm9, %xmm8

#ifndef TRMMKERNEL
	addpd	%xmm0,  %xmm8
#endif

	movlpd	%xmm8, 0 * SIZE(CO1)
	movhpd	%xmm8, 1 * SIZE(CO1)
	ALIGN_4
	
.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