gemm_kernel_8x4_core2.s

Optimized GotoBLAS libraries

	movaps	-28 * SIZE(AO), %xmm0
	movaps	-16 * SIZE(BO), %xmm2
	movaps	-12 * SIZE(BO), %xmm3
	movaps	 -8 * SIZE(BO), %xmm4
	movaps	 -4 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm0, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	movaps	-24 * SIZE(AO), %xmm0
	movaps	  0 * SIZE(BO), %xmm2
	movaps	  4 * SIZE(BO), %xmm3
	movaps	  8 * SIZE(BO), %xmm4
	movaps	 12 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm0, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	movaps	-20 * SIZE(AO), %xmm0
	movaps	 16 * SIZE(BO), %xmm2
	movaps	 20 * SIZE(BO), %xmm3
	movaps	 24 * SIZE(BO), %xmm4
	movaps	 28 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm0, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	subq	$-16 * SIZE, AO
	subq	$-64 * SIZE, BO
	subq	$1, %rax
	jg	.L21
	ALIGN_4

.L25:
	movaps	ALPHA, %xmm7

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

.L26:
	movaps	-32 * SIZE(AO), %xmm0
	movaps	-32 * SIZE(BO), %xmm2
	movaps	-28 * SIZE(BO), %xmm3
	movaps	-24 * SIZE(BO), %xmm4
	movaps	-20 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm0, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	addq	$ 4 * SIZE, AO
	addq	$16 * SIZE, BO
	subq	$1, %rax
	jg	.L26
	ALIGN_4

.L28:
#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movhps	2 * SIZE(CO1), %xmm0
	movsd	0 * SIZE(CO2), %xmm2
	movhps	2 * SIZE(CO2), %xmm2

	movsd	0 * SIZE(CO1, LDC, 2), %xmm4
	movhps	2 * SIZE(CO1, LDC, 2), %xmm4
	movsd	0 * SIZE(CO2, LDC, 2), %xmm6
	movhps	2 * SIZE(CO2, LDC, 2), %xmm6
#endif

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

#ifndef TRMMKERNEL
	addps	%xmm0,  %xmm8
	addps	%xmm2, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm6, %xmm11
#endif

	movlps	%xmm8, 0 * SIZE(CO1)
	movhps	%xmm8, 2 * SIZE(CO1)
	movlps	%xmm9, 0 * SIZE(CO2)
	movhps	%xmm9, 2 * SIZE(CO2)

	movlps	%xmm10, 0 * SIZE(CO1, LDC, 2)
	movhps	%xmm10, 2 * SIZE(CO1, LDC, 2)
	movlps	%xmm11, 0 * SIZE(CO2, LDC, 2)
	movhps	%xmm11, 2 * 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,    8), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 8), BO
#endif

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

	addq	$4 * SIZE, CO1
	addq	$4 * SIZE, CO2
	subq	$1, I
	ALIGN_4
	
.L30:
	testq	$2, M
	jle	.L40

#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	

	pxor	%xmm8, %xmm8
	pxor	%xmm9, %xmm9
	pxor	%xmm10, %xmm10
	pxor	%xmm11, %xmm11

	pxor	%xmm12, %xmm12
	pxor	%xmm13, %xmm13
	pxor	%xmm14, %xmm14
	pxor	%xmm15, %xmm15

#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	$4, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$2, %rax
	jle	.L35
	ALIGN_4

.L31:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	movsd	-32 * SIZE(AO), %xmm0
	movsd	-32 * SIZE(BO), %xmm2
	movsd	-28 * SIZE(BO), %xmm3
	movsd	-24 * SIZE(BO), %xmm4
	movsd	-20 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm0, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	movsd	-30 * SIZE(AO), %xmm0
	movsd	-16 * SIZE(BO), %xmm2
	movsd	-12 * SIZE(BO), %xmm3
	movsd	 -8 * SIZE(BO), %xmm4
	movsd	 -4 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm0, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	movsd	-28 * SIZE(AO), %xmm0
	movsd	  0 * SIZE(BO), %xmm2
	movsd	  4 * SIZE(BO), %xmm3
	movsd	  8 * SIZE(BO), %xmm4
	movsd	 12 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm0, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	movsd	-26 * SIZE(AO), %xmm0
	movsd	 16 * SIZE(BO), %xmm2
	movsd	 20 * SIZE(BO), %xmm3
	movsd	 24 * SIZE(BO), %xmm4
	movsd	 28 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm0, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	subq	$ -8 * SIZE, AO
	subq	$-64 * SIZE, BO
	subq	$1, %rax
	jg	.L31
	ALIGN_4

.L35:
	movsd	ALPHA, %xmm7

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

.L36:
	movsd	-32 * SIZE(AO), %xmm0
	movsd	-32 * SIZE(BO), %xmm2
	movsd	-28 * SIZE(BO), %xmm3
	movsd	-24 * SIZE(BO), %xmm4
	movsd	-20 * SIZE(BO), %xmm5

	mulps	%xmm0, %xmm2
	mulps	%xmm0, %xmm3
	mulps	%xmm0, %xmm4
	mulps	%xmm0, %xmm5

	addps	%xmm2, %xmm8
	addps	%xmm3, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm11

	addq	$ 2 * SIZE, AO
	addq	$16 * SIZE, BO
	subq	$1, %rax
	jg	.L36
	ALIGN_4

.L38:
#ifndef TRMMKERNEL
	movsd	0 * SIZE(CO1), %xmm0
	movsd	0 * SIZE(CO2), %xmm2
	movsd	0 * SIZE(CO1, LDC, 2), %xmm4
	movsd	0 * SIZE(CO2, LDC, 2), %xmm6
#endif

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

#ifndef TRMMKERNEL
	addps	%xmm0,  %xmm8
	addps	%xmm2, %xmm9
	addps	%xmm4, %xmm10
	addps	%xmm6, %xmm11
#endif

	movlps	%xmm8, 0 * SIZE(CO1)
	movlps	%xmm9, 0 * SIZE(CO2)
	movlps	%xmm10, 0 * SIZE(CO1, LDC, 2)
	movlps	%xmm11, 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,    8), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 8), BO
#endif

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

	addq	$2 * SIZE, CO1
	addq	$2 * SIZE, CO2
	ALIGN_4

.L40:
	testq	$1, M
	jle	.L49

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

	pxor	%xmm8, %xmm8
	pxor	%xmm9, %xmm9
	pxor	%xmm10, %xmm10
	pxor	%xmm11, %xmm11

	pxor	%xmm12, %xmm12
	pxor	%xmm13, %xmm13
	pxor	%xmm14, %xmm14
	pxor	%xmm15, %xmm15

#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	.L45
	ALIGN_4

.L41:	
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)

	movss	-32 * SIZE(AO), %xmm0
	movss	-32 * SIZE(BO), %xmm2
	movss	-28 * SIZE(BO), %xmm3
	movss	-24 * SIZE(BO), %xmm4
	movss	-20 * SIZE(BO), %xmm5

	mulss	%xmm0, %xmm2
	mulss	%xmm0, %xmm3
	mulss	%xmm0, %xmm4
	mulss	%xmm0, %xmm5

	addss	%xmm2, %xmm8
	addss	%xmm3, %xmm9
	addss	%xmm4, %xmm10
	addss	%xmm5, %xmm11

	movss	-31 * SIZE(AO), %xmm0
	movss	-16 * SIZE(BO), %xmm2
	movss	-12 * SIZE(BO), %xmm3
	movss	 -8 * SIZE(BO), %xmm4
	movss	 -4 * SIZE(BO), %xmm5

	mulss	%xmm0, %xmm2
	mulss	%xmm0, %xmm3
	mulss	%xmm0, %xmm4
	mulss	%xmm0, %xmm5

	addss	%xmm2, %xmm8
	addss	%xmm3, %xmm9
	addss	%xmm4, %xmm10
	addss	%xmm5, %xmm11

	movss	-30 * SIZE(AO), %xmm0
	movss	  0 * SIZE(BO), %xmm2
	movss	  4 * SIZE(BO), %xmm3
	movss	  8 * SIZE(BO), %xmm4
	movss	 12 * SIZE(BO), %xmm5

	mulss	%xmm0, %xmm2
	mulss	%xmm0, %xmm3
	mulss	%xmm0, %xmm4
	mulss	%xmm0, %xmm5

	addss	%xmm2, %xmm8
	addss	%xmm3, %xmm9
	addss	%xmm4, %xmm10
	addss	%xmm5, %xmm11

	movss	-29 * SIZE(AO), %xmm0
	movss	 16 * SIZE(BO), %xmm2
	movss	 20 * SIZE(BO), %xmm3
	movss	 24 * SIZE(BO), %xmm4
	movss	 28 * SIZE(BO), %xmm5

	mulss	%xmm0, %xmm2
	mulss	%xmm0, %xmm3
	mulss	%xmm0, %xmm4
	mulss	%xmm0, %xmm5

	addss	%xmm2, %xmm8
	addss	%xmm3, %xmm9
	addss	%xmm4, %xmm10
	addss	%xmm5, %xmm11

	subq	$ -4 * SIZE, AO
	subq	$-64 * SIZE, BO
	subq	$1, %rax
	jg	.L41
	ALIGN_4

.L45:
	movss	ALPHA, %xmm7

#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L48
	ALIGN_4

.L46:
	movss	-32 * SIZE(AO), %xmm0
	movss	-32 * SIZE(BO), %xmm2
	movss	-28 * SIZE(BO), %xmm3
	movss	-24 * SIZE(BO), %xmm4
	movss	-20 * SIZE(BO), %xmm5

	mulss	%xmm0, %xmm2
	mulss	%xmm0, %xmm3
	mulss	%xmm0, %xmm4
	mulss	%xmm0, %xmm5

	addss	%xmm2, %xmm8
	addss	%xmm3, %xmm9
	addss	%xmm4, %xmm10
	addss	%xmm5, %xmm11

	addq	$ 1 * SIZE, AO
	addq	$16 * SIZE, BO
	subq	$1, %rax
	jg	.L46
	ALIGN_4

.L48:
#ifndef TRMMKERNEL
	movss	0 * SIZE(CO1), %xmm0
	movss	0 * SIZE(CO2), %xmm2
	movss	0 * SIZE(CO1, LDC, 2), %xmm4
	movss	0 * SIZE(CO2, LDC, 2), %xmm6
#endif

	mulss	%xmm7, %xmm8
	mulss	%xmm7, %xmm9
	mulss	%xmm7, %xmm10
	mulss	%xmm7, %xmm11

#ifndef TRMMKERNEL
	addss	%xmm0,  %xmm8
	addss	%xmm2, %xmm9
	addss	%xmm4, %xmm10
	addss	%xmm6, %xmm11
#endif

	movss	%xmm8, 0 * SIZE(CO1)
	movss	%xmm9, 0 * SIZE(CO2)
	movss	%xmm10, 0 * SIZE(CO1, LDC, 2)
	movss	%xmm11, 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,    4), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 8), BO
	leaq	(BO, %rax, 8), BO
#endif

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

.L49:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addl	$4, KK
#endif

	leaq	(C, LDC, 4), C
	subq	$1, J
	jg	.L01
	ALIGN_4

.L50:
	testq	$2, N
	jle	.L100
	ALIGN_4

.L51:
/* Copying to Sub Buffer */
	leaq	BUFFER, BO
	
#if defined(TRMMKERNEL) && defined(LEFT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif	

	movq	K, %rax
	sarq	$3, %rax
	jle	.L53

	addq	%rax, %rax
	ALIGN_4
	
.L52:
	movaps	 -32 * SIZE(B), %xmm3
	movaps	 -28 * SIZE(B), %xmm7

	pshufd	$0x00, %xmm3,  %xmm0
	pshufd	$0x55, %xmm3,  %xmm1
	pshufd	$0xaa, %xmm3,  %xmm2
	pshufd	$0xff, %xmm3,  %xmm3
	prefetcht0	(PREFETCH_W +  0) * SIZE(BO)

	pshufd	$0x00, %xmm7,  %xmm4
	pshufd	$0x55, %xmm7,  %xmm5
	pshufd	$0xaa, %xmm7,  %xmm6
	pshufd	$0xff, %xmm7,  %xmm7

	prefetcht0	(PREFETCH_W + 16) * SIZE(BO)

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

	subq	$1, %rax
	jne	.L52
	ALIGN_4

.L53:
	movq	K, %rax
	andq	$7, %rax
	BRANCH
	jle	.L55
	ALIGN_4

.L54:
	movss	 -32 * SIZE(B), %xmm8
	movss	 -31 * SIZE(B), %xmm9

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

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

	addq	$2 * SIZE, B
	addq	$8 * SIZE, BO
	subq	$1, %rax
	jne	.L54
	ALIGN_4
	
.L55:
	movq	C, CO1
	leaq	(C, LDC, 1), CO2
	movq	A, AO		# aoffset = a

	movq	M,  I
	sarq	$3, I
	jle	.L70
	ALIGN_4

.L60:
#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	

	pxor	%xmm8, %xmm8
	pxor	%xmm9, %xmm9

	prefetcht0     7 * SIZE(CO1)
	pxor	%xmm12, %xmm12
	prefetcht0     7 * SIZE(CO2)