gemm_kernel_4x4_sse3.s

Optimized GotoBLAS libraries

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

	addq	$2 * SIZE, CO1		# coffset += 4
	addq	$2 * SIZE, CO2		# coffset += 4
	ALIGN_4	

.L30:
	testq	$1, M
	je	.L39
	ALIGN_4

.L31:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(B,  %rax, 4), BO
#endif	

	movddup	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movapd	 0 * SIZE(BO), %xmm9
	pxor	%xmm1, %xmm1
	movddup	 4 * SIZE(AO), %xmm10
	pxor	%xmm2, %xmm2
	movapd	 8 * SIZE(BO), %xmm11
	pxor	%xmm3, %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
	sarq	$3, %rax
	je	.L35
	ALIGN_4

.L32:
	mulpd	%xmm8, %xmm9
	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)
	addpd	%xmm9, %xmm0
	movapd	 2 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movddup	 1 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movapd	 4 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm0
	movapd	 6 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movddup	 2 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movapd	16 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm11
	addpd	%xmm11, %xmm0
	movapd	10 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	movddup	 3 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm1
	movapd	12 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	addpd	%xmm11, %xmm0
	movapd	14 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	movddup	 8 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm1
	movapd	24 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm9
	addpd	%xmm9, %xmm0
	movapd	18 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	movddup	 5 * SIZE(AO), %xmm10
	addpd	%xmm9, %xmm1
	movapd	20 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	addpd	%xmm9, %xmm0
	movapd	22 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	movddup	 6 * SIZE(AO), %xmm10
	addpd	%xmm9, %xmm1
	movapd	32 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm11
	addpd	%xmm11, %xmm0
	movapd	26 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	movddup	 7 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm1
	movapd	28 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	addpd	%xmm11, %xmm0
	movapd	30 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	movddup	12 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm1
	movapd	40 * SIZE(BO), %xmm11

	addq   $ 8 * SIZE, AO
	addq   $32 * SIZE, BO
	decq   %rax
	jne    .L32
	ALIGN_4

.L35:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movddup	ALPHA, %xmm15
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L38
	ALIGN_4

.L36:
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm0
	movapd	 2 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movddup	 1 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movapd	 4 * SIZE(BO), %xmm9

	addq	$1 * SIZE, AO		# aoffset  += 4
	addq	$4 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L36
	ALIGN_4

.L38:
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movsd	0 * SIZE(CO1), %xmm8
	movhpd	0 * SIZE(CO2), %xmm8
	movsd	0 * SIZE(CO1, LDC, 2), %xmm9
	movhpd	0 * SIZE(CO2, LDC, 2), %xmm9
#endif
	mulpd	%xmm15, %xmm0
	mulpd	%xmm15, %xmm1

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addpd	%xmm8,  %xmm0
	addpd	%xmm9,  %xmm1
#endif

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

#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, 1), AO
	leaq	(BO, %rax, 4), BO
#endif

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$1, KK
#endif
	ALIGN_4	
	
.L39:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$4, KK
#endif

	leaq	(C, LDC, 4), C		# c += 4 * ldc
	movq	BO, B
	decq	J			# j --
	jg	.L10
	ALIGN_4

.L40:
	testq	$2, N
	je	.L80
	ALIGN_4

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

	movq	C, CO1			# coffset1 = c
	leaq	(C, LDC, 1), CO2	# coffset2 = c + ldc
	movq	A, AO		# aoffset = a

	movq	K, %rax
	salq	$BASE_SHIFT + 1, %rax
	leaq	(B, %rax), BB

	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))

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(B,  %rax, 2), BO
#endif	

	prefetcht0	  0 * SIZE(BB)
	subq	   $-4 * SIZE, BB

	movapd	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movddup	 0 * SIZE(BO), %xmm9
	pxor	%xmm1, %xmm1
	movapd	 8 * SIZE(AO), %xmm10
	pxor	%xmm4, %xmm4
	movddup	 8 * SIZE(BO), %xmm11
	pxor	%xmm5, %xmm5

#ifdef HAVE_3DNOW
	prefetchw      4 * SIZE(CO1)
	prefetchw      4 * SIZE(CO2)
#else
	prefetchnta     4 * SIZE(CO1)
	prefetchnta     4 * SIZE(CO2)
#endif

#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
	sarq	$3, %rax
	je	.L55
	ALIGN_4

.L52:
	mulpd	%xmm8, %xmm9
	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)
	addpd	%xmm9, %xmm0
	movddup	 1 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 2 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movddup	 0 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm4
	movddup	 1 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 4 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm5
	movddup	 2 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm0
	movddup	 3 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 6 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movddup	 2 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm4
	movddup	 3 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	16 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm5
	movddup	 4 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	addpd	%xmm9, %xmm0
	movddup	 5 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	movapd	10 * SIZE(AO), %xmm10
	addpd	%xmm9, %xmm1
	movddup	 4 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	addpd	%xmm9, %xmm4
	movddup	 5 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	movapd	12 * SIZE(AO), %xmm10
	addpd	%xmm9, %xmm5
	movddup	 6 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	addpd	%xmm9, %xmm0
	movddup	 7 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	movapd	14 * SIZE(AO), %xmm10
	addpd	%xmm9, %xmm1
	movddup	 6 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	addpd	%xmm9, %xmm4
	movddup	 7 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm9
	movapd	40 * SIZE(AO), %xmm10
	addpd	%xmm9, %xmm5
	movddup	16 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm11
	PREFETCH  (PREFETCHSIZE + 16) * SIZE(AO)
	addpd	%xmm11, %xmm0
	movddup	 9 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	movapd	18 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm1
	movddup	 8 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	addpd	%xmm11, %xmm4
	movddup	 9 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	movapd	20 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm5
	movddup	10 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	addpd	%xmm11, %xmm0
	movddup	11 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	movapd	22 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm1
	movddup	10 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	addpd	%xmm11, %xmm4
	movddup	11 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	movapd	24 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm5
	movddup	12 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	addpd	%xmm11, %xmm0
	movddup	13 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	movapd	26 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm1
	movddup	12 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	addpd	%xmm11, %xmm4
	movddup	13 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	movapd	28 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm5
	movddup	14 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	addpd	%xmm11, %xmm0
	movddup	15 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	movapd	30 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm1
	movddup	14 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	addpd	%xmm11, %xmm4
	movddup	15 * SIZE(BO), %xmm11
	mulpd	%xmm8, %xmm11
	movapd	32 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm5
	movddup	24 * SIZE(BO), %xmm11

	addq   $32 * SIZE, AO
	addq   $16 * SIZE, BO
	decq   %rax
	jne    .L52
	ALIGN_4

.L55:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movddup	ALPHA, %xmm15
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L59
	ALIGN_4

.L56:
	mulpd	%xmm8, %xmm9
	movapd	 2 * SIZE(AO), %xmm10
	addpd	%xmm9, %xmm0
	movddup	 1 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movddup	 0 * SIZE(BO), %xmm11
	addpd	%xmm9, %xmm1
	movddup	 2 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm11
	movapd	 4 * SIZE(AO), %xmm8
	addpd	%xmm11, %xmm4
	movddup	 1 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	addpd	%xmm11, %xmm5

	addq	$4 * SIZE, AO		# aoffset  += 4
	addq	$2 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L56
	ALIGN_4

.L59:
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
     movsd	0 * SIZE(CO1), %xmm8
	movhpd	1 * SIZE(CO1), %xmm8
	movsd	2 * SIZE(CO1), %xmm9
	movhpd	3 * SIZE(CO1), %xmm9
	movsd	0 * SIZE(CO2), %xmm10
	movhpd	1 * SIZE(CO2), %xmm10
	movsd	2 * SIZE(CO2), %xmm11
	movhpd	3 * SIZE(CO2), %xmm11
#endif

	mulpd	%xmm15, %xmm0
	mulpd	%xmm15, %xmm1
	mulpd	%xmm15, %xmm4
	mulpd	%xmm15, %xmm5

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addpd	%xmm8,  %xmm0
	addpd	%xmm9,  %xmm4
	addpd	%xmm10, %xmm1
	addpd	%xmm11, %xmm5
#endif

	movsd	%xmm0, 0 * SIZE(CO1)
	movhpd	%xmm0, 1 * SIZE(CO1)
	movsd	%xmm4, 2 * SIZE(CO1)
	movhpd	%xmm4, 3 * SIZE(CO1)
	movsd	%xmm1, 0 * SIZE(CO2)
	movhpd	%xmm1, 1 * SIZE(CO2)
	movsd	%xmm5, 2 * SIZE(CO2)
	movhpd	%xmm5, 3 * 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, 4), AO
	leaq	(BO, %rax, 2), BO
#endif

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

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

	movq	B, BO
#else
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(B,  %rax, 2), BO
#endif	

	movapd	 0 * SIZE(AO), %xmm8
	pxor	%xmm0, %xmm0
	movddup	 0 * SIZE(BO), %xmm9
	pxor	%xmm1, %xmm1
	movapd	 8 * SIZE(AO), %xmm10
	pxor	%xmm2, %xmm2
	movddup	 8 * SIZE(BO), %xmm11
	pxor	%xmm3, %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	$2, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$3, %rax
	je	.L65
	ALIGN_4

.L62:
	mulpd	%xmm8, %xmm9
	PREFETCH  (PREFETCHSIZE + 0) * SIZE(AO)
	addpd	%xmm9, %xmm0
	movddup	 1 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 2 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movddup	 2 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm2
	movddup	 3 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 4 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm3
	movddup	 4 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm0
	movddup	 5 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 6 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movddup	 6 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm2
	movddup	 7 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	16 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm3
	movddup	16 * SIZE(BO), %xmm9
	mulpd	%xmm10, %xmm11
	addpd	%xmm11, %xmm0
	movddup	 9 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	movapd	10 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm1
	movddup	10 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	addpd	%xmm11, %xmm2
	movddup	11 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	movapd	12 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm3
	movddup	12 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	addpd	%xmm11, %xmm0
	movddup	13 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	movapd	14 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm1
	movddup	14 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	addpd	%xmm11, %xmm2
	movddup	15 * SIZE(BO), %xmm11
	mulpd	%xmm10, %xmm11
	movapd	24 * SIZE(AO), %xmm10
	addpd	%xmm11, %xmm3
	movddup	24 * SIZE(BO), %xmm11

	addq   $16 * SIZE, AO
	addq   $16 * SIZE, BO
	decq   %rax
	jne    .L62
	ALIGN_4

.L65:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movddup	ALPHA, %xmm15
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L69
	ALIGN_4

.L66:
	mulpd	%xmm8, %xmm9
	addpd	%xmm9, %xmm0
	movddup	 1 * SIZE(BO), %xmm9
	mulpd	%xmm8, %xmm9
	movapd	 2 * SIZE(AO), %xmm8
	addpd	%xmm9, %xmm1
	movddup	 2 * SIZE(BO), %xmm9

	addq	$2 * SIZE, AO		# aoffset  += 4
	addq	$2 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L66
	ALIGN_4

.L69:
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movsd	0 * SIZE(CO1), %xmm8
	movhpd	1 * SIZE(CO1), %xmm8
	movsd	0 * SIZE(CO2), %xmm10
	movhpd	1 * SIZE(CO2), %xmm10
#endif

	addpd	%xmm2, %xmm0
	addpd	%xmm3, %xmm1

	mulpd	%xmm15, %xmm0
	mulpd	%xmm15, %xmm1

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addpd	%xmm8,  %xmm0
	addpd	%xmm10, %xmm1
#endif

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