gemm_kernel_4x4_penryn.s

Optimized GotoBLAS libraries

	addpd	%xmm3, %xmm11
	movaps	 -3 * SIZE(BO), %xmm3
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	addpd	%xmm5, %xmm10
	mulpd	%xmm0, %xmm7
	subq	$ -8 * SIZE, AO

	addpd	%xmm2, %xmm9
	movaps	 -1 * SIZE(BO), %xmm2
	pshufd	$0x4e, %xmm3, %xmm5
	mulpd	%xmm0, %xmm3
	addpd	%xmm7, %xmm8
	mulpd	%xmm0, %xmm5
	movaps	-16 * SIZE(AO), %xmm0

	subq	$-16 * SIZE, BO
	subq	$1, %rax
	jg,pt	.L22
	ALIGN_4

.L25:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je,pt	.L28
	ALIGN_4

.L26:
	addpd	%xmm3, %xmm11
	movaps	-15 * SIZE(BO), %xmm3
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	addpd	%xmm5, %xmm10
	mulpd	%xmm0, %xmm7

	addpd	%xmm2, %xmm9
	movaps	-13 * SIZE(BO), %xmm2
	pshufd	$0x4e, %xmm3, %xmm5
	mulpd	%xmm0, %xmm3
	addpd	%xmm7, %xmm8
	mulpd	%xmm0, %xmm5
	movaps	-14 * SIZE(AO), %xmm0

	addq	$2 * SIZE, AO
	addq	$4 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L26
	ALIGN_4

.L28:
	addpd	%xmm3, %xmm11
	addpd	%xmm5, %xmm10

	movddup	ALPHA, %xmm3

	movapd	 %xmm8, %xmm0
	movsd    %xmm9, %xmm8
	mulpd	%xmm3, %xmm8
	movsd    %xmm0, %xmm9
	mulpd	%xmm3, %xmm9

	movapd	 %xmm10, %xmm0
	movsd    %xmm11, %xmm10
	mulpd	%xmm3, %xmm10
	movsd    %xmm0,  %xmm11
	mulpd	%xmm3, %xmm11

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhpd	1 * SIZE(CO1), %xmm0
	movsd	0 * SIZE(CO2), %xmm2
	movhpd	1 * SIZE(CO2), %xmm2

	movsd	0 * SIZE(CO1, LDC, 2), %xmm4
	movhpd	1 * SIZE(CO1, LDC, 2), %xmm4
	movsd	0 * SIZE(CO2, LDC, 2), %xmm6
	movhpd	1 * SIZE(CO2, LDC, 2), %xmm6

	addpd	%xmm0, %xmm8
	addpd	%xmm2, %xmm9
	addpd	%xmm4, %xmm10
	addpd	%xmm6, %xmm11
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movhpd	%xmm8,  1 * SIZE(CO1)
	movsd	%xmm9,  0 * SIZE(CO2)
	movhpd	%xmm9,  1 * SIZE(CO2)

	movsd	%xmm10, 0 * SIZE(CO1, LDC, 2)
	movhpd	%xmm10, 1 * SIZE(CO1, LDC, 2)
	movsd	%xmm11, 0 * SIZE(CO2, LDC, 2)
	movhpd	%xmm11, 1 * 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, 2), AO
	leaq	(BO, %rax, 4), BO
#endif

#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
	jle,pt	.L39
	ALIGN_4

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

	movq	B, BO
#else
	movq	B, BO

	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	addq	%rax, AO
	leaq	(BO, %rax, 4), BO
#endif	

	movsd	-16 * SIZE(AO), %xmm0
	movaps	-17 * SIZE(BO), %xmm2
	movaps	-15 * 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	$4, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle,pn	.L35
	ALIGN_4

.L32:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm3
	movsd	-15 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm8
	movaps	-13 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm9
	movaps	-11 * SIZE(BO), %xmm3

	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm3
	movsd	-14 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm10
	movaps	 -9 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm11
	movaps	 -7 * SIZE(BO), %xmm3

	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm3
	movsd	-13 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm8
	movaps	 -5 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm9
	movaps	 -3 * SIZE(BO), %xmm3

	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm3
	movsd	-12 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm10
	movaps	 -1 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm11
	movaps	  1 * SIZE(BO), %xmm3

	subq	$ -4 * SIZE, AO
	subq	$-16 * SIZE, BO
	subq	$1, %rax
	jg,pt	.L32
	ALIGN_4

.L35:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je,pt	.L38
	ALIGN_4

.L36:
	shufps	$0x44, %xmm0, %xmm0
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm3
	movsd	-15 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm8
	movaps	-13 * SIZE(BO), %xmm2
	addpd	%xmm3, %xmm9
	movaps	-11 * SIZE(BO), %xmm3

	addq	$1 * SIZE, AO
	addq	$4 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L36
	ALIGN_4

.L38:
	movddup	ALPHA, %xmm3

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

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhpd	0 * SIZE(CO2), %xmm0

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

	mulpd	%xmm3, %xmm8
	mulpd	%xmm3, %xmm9

#ifndef TRMMKERNEL
	addpd	%xmm0, %xmm8
	addpd	%xmm1, %xmm9
#endif

	movlpd	%xmm8,  0 * SIZE(CO1)
	movhpd	%xmm8,  0 * SIZE(CO2)

	movlpd	%xmm9, 0 * SIZE(CO1, LDC, 2)
	movhpd	%xmm9, 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
	addq	%rax, AO
	leaq	(BO, %rax, 4), BO
#endif

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

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

	movq	BO, B

	leaq	(C, LDC, 4), C

	subq	$1, J
	jg,pt	.L01
	ALIGN_4

.L40:
	testq	$2, N
	jle,pt	.L80

	movq	C, CO1
	leaq	(C, LDC, 1), CO2
	movq	A, AO

#if defined(TRMMKERNEL) && defined(LEFT)
        movq    OFFSET, %rax
	movq    %rax, KK
#endif
	
	movq	K, %rax
	salq	$BASE_SHIFT + 1, %rax
	leaq	(B, %rax), BB

	movq	M,  I
	sarq	$2, I	# i = (m >> 2)
	jle,pn	.L60
	ALIGN_4

.L51:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	movq	B, BO
#else
	movq	B, BO

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

	prefetcht2	 -16 * SIZE(BB)
	subq		 $-4 * SIZE, BB

	movaps	-16 * SIZE(AO), %xmm0
	movaps	-14 * SIZE(AO), %xmm1
	movaps	-17 * SIZE(BO), %xmm2

	prefetcht0     3 * SIZE(CO1)
	pxor	%xmm8,  %xmm8
	pxor	%xmm9,  %xmm9
	prefetcht0     3 * SIZE(CO2)
	pxor	%xmm12, %xmm12
	pxor	%xmm13, %xmm13

#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	$2, %rax
	jle,pn	.L55
	ALIGN_4

.L52:	

	movaps	%xmm2, %xmm4
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)
	mulpd	%xmm1, %xmm4

	movaps	%xmm7, %xmm6
	mulpd	%xmm0, %xmm7
	movaps	-12 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm6
	movaps	-10 * SIZE(AO), %xmm1

	addpd	%xmm2, %xmm9
	movaps	-15 * SIZE(BO), %xmm2
	addpd	%xmm4, %xmm13
	addpd	%xmm7, %xmm8
	addpd	%xmm6, %xmm12

	movaps	%xmm2, %xmm4
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	PREFETCH (PREFETCHSIZE +  8) * SIZE(AO)
	mulpd	%xmm1, %xmm4

	movaps	%xmm7, %xmm6
	mulpd	%xmm0, %xmm7
	movaps	 -8 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm6
	movaps	 -6 * SIZE(AO), %xmm1

	addpd	%xmm2, %xmm9
	movaps	-13 * SIZE(BO), %xmm2
	addpd	%xmm4, %xmm13
	addpd	%xmm7, %xmm8
	addpd	%xmm6, %xmm12

	movaps	%xmm2, %xmm4
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm1, %xmm4

	movaps	%xmm7, %xmm6
	mulpd	%xmm0, %xmm7
	movaps	 -4 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm6
	movaps	 -2 * SIZE(AO), %xmm1

	addpd	%xmm2, %xmm9
	movaps	-11 * SIZE(BO), %xmm2
	addpd	%xmm4, %xmm13
	addpd	%xmm7, %xmm8
	addpd	%xmm6, %xmm12

	movaps	%xmm2, %xmm4
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm1, %xmm4

	movaps	%xmm7, %xmm6
	mulpd	%xmm0, %xmm7
	movaps	  0 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm6
	movaps	  2 * SIZE(AO), %xmm1

	addpd	%xmm2, %xmm9
	movaps	 -9 * SIZE(BO), %xmm2
	addpd	%xmm4, %xmm13
	addpd	%xmm7, %xmm8
	addpd	%xmm6, %xmm12

	subq	$-16 * SIZE, AO
	subq	$ -8 * SIZE, BO
	subq	$1, %rax
	jg,pt	.L52
	ALIGN_4

.L55:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je,pt	.L58
	ALIGN_4

.L56:
	movaps	%xmm2, %xmm4
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm1, %xmm4

	movaps	%xmm7, %xmm6
	mulpd	%xmm0, %xmm7
	movaps	-12 * SIZE(AO), %xmm0
	mulpd	%xmm1, %xmm6
	movaps	-10 * SIZE(AO), %xmm1

	addpd	%xmm2, %xmm9
	movaps	-15 * SIZE(BO), %xmm2
	addpd	%xmm4, %xmm13
	addpd	%xmm7, %xmm8
	addpd	%xmm6, %xmm12

	addq	$4 * SIZE, AO
	addq	$2 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L56
	ALIGN_4

.L58:
	movddup	ALPHA, %xmm3

	movapd	 %xmm8, %xmm0
	movsd    %xmm9, %xmm8
	mulpd	%xmm3, %xmm8
	movsd    %xmm0, %xmm9
	mulpd	%xmm3, %xmm9

	movapd	 %xmm12, %xmm0
	movsd    %xmm13, %xmm12
	mulpd	%xmm3, %xmm12
	movsd    %xmm0,  %xmm13
	mulpd	%xmm3, %xmm13

#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhpd	1 * SIZE(CO1), %xmm0
	movsd	2 * SIZE(CO1), %xmm1
	movhpd	3 * SIZE(CO1), %xmm1
	movsd	0 * SIZE(CO2), %xmm2
	movhpd	1 * SIZE(CO2), %xmm2
	movsd	2 * SIZE(CO2), %xmm3
	movhpd	3 * SIZE(CO2), %xmm3

	addpd	%xmm0, %xmm8
	addpd	%xmm1, %xmm12
	addpd	%xmm2, %xmm9
	addpd	%xmm3, %xmm13
#endif

	movsd	%xmm8,  0 * SIZE(CO1)
	movhpd	%xmm8,  1 * SIZE(CO1)
	movsd	%xmm12, 2 * SIZE(CO1)
	movhpd	%xmm12, 3 * SIZE(CO1)
	movsd	%xmm9,  0 * SIZE(CO2)
	movhpd	%xmm9,  1 * SIZE(CO2)
	movsd	%xmm13, 2 * SIZE(CO2)
	movhpd	%xmm13, 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
	addq	$4 * SIZE, CO2
	decq	I
	jg,pt	.L51
	ALIGN_4	

.L60:
	testq	$2, M
	jle,pt	.L70
	ALIGN_4

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

	movq	B, BO
#else
	movq	B, BO

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

	movaps	-16 * SIZE(AO), %xmm0
	pxor	%xmm8,  %xmm8
	pxor	%xmm9,  %xmm9
	movaps	-17 * SIZE(BO), %xmm2
	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	$2, %rax
	jle,pn	.L65
	ALIGN_4

.L62:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm7
	movaps	-14 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm9
	addpd	%xmm7, %xmm8
	movaps	-15 * SIZE(BO), %xmm2

	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm7
	movaps	-12 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm11
	addpd	%xmm7, %xmm10
	movaps	-13 * SIZE(BO), %xmm2

	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm7
	movaps	-10 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm9
	addpd	%xmm7, %xmm8
	movaps	-11 * SIZE(BO), %xmm2

	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm7
	movaps	 -8 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm11
	addpd	%xmm7, %xmm10
	movaps	 -9 * SIZE(BO), %xmm2

	subq	$-8 * SIZE, AO
	subq	$-8 * SIZE, BO
	subq	$1, %rax
	jg,pt	.L62
	ALIGN_4

.L65:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je,pt	.L68
	ALIGN_4

.L66:
	pshufd	$0x4e, %xmm2, %xmm7
	mulpd	%xmm0, %xmm2
	mulpd	%xmm0, %xmm7
	movaps	-14 * SIZE(AO), %xmm0

	addpd	%xmm2, %xmm9
	addpd	%xmm7, %xmm8
	movaps	-15 * SIZE(BO), %xmm2

	addq	$2 * SIZE, AO
	addq	$2 * SIZE, BO

	subq	$1, %rax
	jg,pt	.L66
	ALIGN_4

.L68:
	addpd	%xmm10, %xmm8
	addpd	%xmm11, %xmm9