gemm_kernel_8x4_sse.s

Optimized GotoBLAS libraries

	addps	%xmm13, %xmm2
	movss	44 * SIZE(BO), %xmm13
	mulps	%xmm10, %xmm13
	movss	-26 * SIZE(AO), %xmm10
	addps	%xmm13, %xmm3
	movss	 96 * SIZE(BO), %xmm13

	mulps	%xmm10, %xmm15
	addps	%xmm15, %xmm0
	movss	52 * SIZE(BO), %xmm15
	mulps	%xmm10, %xmm15
	movss	-25 * SIZE(AO), %xmm10
	addps	%xmm15, %xmm1
	movss	56 * SIZE(BO), %xmm15

	mulps	%xmm10, %xmm15
	addps	%xmm15, %xmm2
	movss	60 * SIZE(BO), %xmm15
	mulps	%xmm10, %xmm15
	movss	-20 * SIZE(AO), %xmm10
	addps	%xmm15, %xmm3
	movss	112 * SIZE(BO), %xmm15

	addq   $ 8 * SIZE, AO
	addq   $64 * SIZE, BO
	decq   %rax
	jne    .L92
	ALIGN_4

.L95:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movaps	ALPHA, %xmm15
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L98
	ALIGN_4

.L96:
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm0
	movss	 4 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	movss	-31 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm1
	movss	 8 * SIZE(BO), %xmm9

	addq	$1 * SIZE, AO		# aoffset  += 4
	addq	$8 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L96
	ALIGN_4

.L98:
#ifndef TRMMKERNEL
	movss	0 * SIZE(CO1), %xmm8
	movss	0 * SIZE(CO2), %xmm10
#endif

	addss	%xmm2, %xmm0
	addss	%xmm3, %xmm1
	mulss	%xmm15, %xmm0
	mulss	%xmm15, %xmm1

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

	movss	%xmm0, 0 * SIZE(CO1)
	movss	%xmm1, 0 * SIZE(CO2)

#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
#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 += 4 * ldc
	ALIGN_4


.L100:
	testq	$1, N
	je	.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	$3, %rax
	jle	.L103
	ALIGN_4
	

.L102:
#if defined(PENTIUM4) || defined(GENERIC)
	movss	 0 * SIZE(B), %xmm0
	movss	 1 * SIZE(B), %xmm1
	movss	 2 * SIZE(B), %xmm2
	movss	 3 * SIZE(B), %xmm3
	movss	 4 * SIZE(B), %xmm4
	movss	 5 * SIZE(B), %xmm5
	movss	 6 * SIZE(B), %xmm6
	movss	 7 * SIZE(B), %xmm7

	PREFETCH	 32 * SIZE(B)

	shufps	 $0, %xmm0, %xmm0
	shufps	 $0, %xmm1, %xmm1
	shufps	 $0, %xmm2, %xmm2
	shufps	 $0, %xmm3, %xmm3
	shufps	 $0, %xmm4, %xmm4
	shufps	 $0, %xmm5, %xmm5
	shufps	 $0, %xmm6, %xmm6
	shufps	 $0, %xmm7, %xmm7

	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
#endif

#if defined(OPTERON) || defined(BARCELONA)
	PREFETCH	 32 * SIZE(B)

	movd	 0 * SIZE(B), %mm0
	movd	 1 * SIZE(B), %mm1
	movd	 2 * SIZE(B), %mm2
	movd	 3 * SIZE(B), %mm3
	movd	 4 * SIZE(B), %mm4
	movd	 5 * SIZE(B), %mm5
	movd	 6 * SIZE(B), %mm6
	movd	 7 * SIZE(B), %mm7

	punpckldq %mm0, %mm0
	punpckldq %mm1, %mm1
	punpckldq %mm2, %mm2
	punpckldq %mm3, %mm3
	punpckldq %mm4, %mm4
	punpckldq %mm5, %mm5
	punpckldq %mm6, %mm6
	punpckldq %mm7, %mm7
	
	movq	%mm0,  0 * SIZE(BO)
	movq	%mm0,  2 * SIZE(BO)
	movq	%mm1,  4 * SIZE(BO)
	movq	%mm1,  6 * SIZE(BO)
	movq	%mm2,  8 * SIZE(BO)
	movq	%mm2, 10 * SIZE(BO)
	movq	%mm3, 12 * SIZE(BO)
	movq	%mm3, 14 * SIZE(BO)
	movq	%mm4, 16 * SIZE(BO)
	movq	%mm4, 18 * SIZE(BO)
	movq	%mm5, 20 * SIZE(BO)
	movq	%mm5, 22 * SIZE(BO)
	movq	%mm6, 24 * SIZE(BO)
	movq	%mm6, 26 * SIZE(BO)
	movq	%mm7, 28 * SIZE(BO)
	movq	%mm7, 30 * SIZE(BO)

	addq	$ 8 * SIZE, B
	addq	$32 * SIZE, BO
#endif

	decq	%rax
	jne	.L102
	ALIGN_4

.L103:
	movq	K, %rax
	andq	$7, %rax
	BRANCH
	jle	.L110
	ALIGN_4

.L104:
#if defined(PENTIUM4) || defined(GENERIC)
	movss	 0 * SIZE(B), %xmm0
	shufps	 $0, %xmm0, %xmm0
	movaps	%xmm0,  0 * SIZE(BO)
#endif

#if defined(OPTERON) || defined(BARCELONA)
	movd	 0 * SIZE(B), %mm0
	punpckldq %mm0, %mm0
	movq	%mm0,  0 * SIZE(BO)
	movq	%mm0,  2 * SIZE(BO)
#endif

	addq	$ 1 * SIZE, B
	addq	$ 4 * SIZE, BO
	decq	%rax
	jne	.L104
	ALIGN_4
	
.L110:
	movq	C, CO1			# coffset1 = c
	movq	A, AO		# aoffset = a

	movq	M,  I
	sarq	$3, I	# i = (m >> 3)
	jle	.L120
	ALIGN_4

.L111:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leaq	BUFFER, BO
#else
	leaq	BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax,   8), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 2), BO
#endif	

	movaps	-32 * SIZE(AO), %xmm8
	movaps	-16 * SIZE(AO), %xmm10
	movaps	  0 * SIZE(AO), %xmm12
	movaps	 16 * SIZE(AO), %xmm14

	movaps	 0 * SIZE(BO), %xmm9
	movaps	16 * SIZE(BO), %xmm11
	movaps	32 * SIZE(BO), %xmm13
	movaps	48 * SIZE(BO), %xmm15

	pxor	%xmm0, %xmm0
	pxor	%xmm1, %xmm1

	PREFETCHW      7 * SIZE(CO1)
	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5

#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	$8, %rax
#else
	addq	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$3, %rax
	je	.L115
	ALIGN_4

.L112:
	mulps	%xmm9, %xmm8
#if defined(OPTERON) || defined(BARCELONA)
	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)
#endif

	mulps	-28 * SIZE(AO), %xmm9
	addps	%xmm8, %xmm0
	movaps	-24 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm4
	movaps	 4 * SIZE(BO), %xmm9

	mulps	%xmm9, %xmm8
	mulps	-20 * SIZE(AO), %xmm9
	addps	%xmm8, %xmm0
	movaps	 32 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm4
	movaps	 8 * SIZE(BO), %xmm9

#if defined(OPTERON) || defined(BARCELONA)
	PREFETCH	(PREFETCHSIZE + 16) * SIZE(AO)
#endif
	mulps	%xmm9, %xmm10
	mulps	-12 * SIZE(AO), %xmm9
	addps	%xmm10, %xmm0
	movaps	 -8 * SIZE(AO), %xmm10
	addps	%xmm9, %xmm4
	movaps	12 * SIZE(BO), %xmm9

	mulps	%xmm9, %xmm10
	mulps	 -4 * SIZE(AO), %xmm9
	addps	%xmm10, %xmm0
	movaps	 48 * SIZE(AO), %xmm10
	addps	%xmm9, %xmm4
	movaps	32 * SIZE(BO), %xmm9

#if defined(OPTERON) || defined(BARCELONA)
	PREFETCH	(PREFETCHSIZE + 32) * SIZE(AO)
#endif
	mulps	%xmm11, %xmm12
	mulps	 4 * SIZE(AO), %xmm11
	addps	%xmm12, %xmm0
	movaps	 8 * SIZE(AO), %xmm12
	addps	%xmm11, %xmm4
	movaps	20 * SIZE(BO), %xmm11

	mulps	%xmm11, %xmm12
	mulps	12 * SIZE(AO), %xmm11
	addps	%xmm12, %xmm0
	movaps	64 * SIZE(AO), %xmm12
	addps	%xmm11, %xmm4
	movaps	24 * SIZE(BO), %xmm11

#if defined(OPTERON) || defined(BARCELONA)
	PREFETCH	(PREFETCHSIZE + 48) * SIZE(AO)
#endif
	mulps	%xmm11, %xmm14
	mulps	20 * SIZE(AO), %xmm11
	addps	%xmm14, %xmm0
	movaps	24 * SIZE(AO), %xmm14
	addps	%xmm11, %xmm4
	movaps	28 * SIZE(BO), %xmm11

	mulps	%xmm11, %xmm14
	mulps	28 * SIZE(AO), %xmm11
	addps	%xmm14, %xmm0
	movaps	80 * SIZE(AO), %xmm14
	addps	%xmm11, %xmm4
	movaps	48 * SIZE(BO), %xmm11

	addq   $64 * SIZE, AO
	addq   $32 * SIZE, BO
	decq   %rax
	jne    .L112
	ALIGN_4

.L115:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movaps	ALPHA, %xmm15
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L118
	ALIGN_4

.L116:
	mulps	%xmm9, %xmm8
	mulps	-28 * SIZE(AO), %xmm9
	addps	%xmm8, %xmm0
	movaps	-24 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm4
	movaps	 4 * SIZE(BO), %xmm9

	addq	$8 * SIZE, AO		# aoffset  += 4
	addq	$4 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L116
	ALIGN_4

.L118:
#ifndef TRMMKERNEL
	movlps	0 * SIZE(CO1), %xmm8
	movhps	2 * SIZE(CO1), %xmm8
	movlps	4 * SIZE(CO1), %xmm9
	movhps	6 * SIZE(CO1), %xmm9
#endif

	mulps	%xmm15, %xmm0
	mulps	%xmm15, %xmm4
#ifndef TRMMKERNEL
	addps	%xmm8,  %xmm0
	addps	%xmm9,  %xmm4
#endif

	movlps	%xmm0, 0 * SIZE(CO1)
	movhps	%xmm0, 2 * SIZE(CO1)
	movlps	%xmm4, 4 * SIZE(CO1)
	movhps	%xmm4, 6 * SIZE(CO1)

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

#if defined(TRMMKERNEL) && defined(LEFT)
	addq	$8, KK
#endif

	addq	$8 * SIZE, CO1		# coffset += 4
	decq	I			# i --
	jg	.L111
	ALIGN_4	

.L120:
	testq	$4, M
	je	.L130

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

	leaq	BUFFER, BO
#else
	leaq	BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax,   8), %rax
	leaq	(AO, %rax, 2), AO
	leaq	(BO, %rax, 2), BO
#endif	

	movaps	-32 * SIZE(AO), %xmm8
	movaps	-16 * SIZE(AO), %xmm10

	movaps	 0 * SIZE(BO), %xmm9
	movaps	16 * SIZE(BO), %xmm11

	pxor	%xmm0, %xmm0
	pxor	%xmm1, %xmm1
	pxor	%xmm2, %xmm2
	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	$4, %rax
#else
	addq	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$3, %rax
	je	.L125
	ALIGN_4

.L122:
	mulps	%xmm8, %xmm9
#if defined(OPTERON) || defined(BARCELONA)
	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)
#endif
	movaps	-28 * SIZE(AO), %xmm8
	mulps	 4 * SIZE(BO), %xmm8
	addps	%xmm9, %xmm0
	movaps	32 * SIZE(BO), %xmm9
	addps	%xmm8, %xmm1
	movaps	-24 * SIZE(AO), %xmm8
	mulps	 8 * SIZE(BO), %xmm8
	addps	%xmm8, %xmm2
	movaps	-20 * SIZE(AO), %xmm8
	mulps	12 * SIZE(BO), %xmm8
	addps	%xmm8, %xmm3
	movaps	  0 * SIZE(AO), %xmm8

#if defined(OPTERON) || defined(BARCELONA)
	PREFETCH	(PREFETCHSIZE + 16) * SIZE(AO)
#endif
	mulps	%xmm10, %xmm11
	movaps	-12 * SIZE(AO), %xmm10
	mulps	20 * SIZE(BO), %xmm10
	addps	%xmm11, %xmm0
	movaps	48 * SIZE(BO), %xmm11
	addps	%xmm10, %xmm1
	movaps	 -8 * SIZE(AO), %xmm10
	mulps	24 * SIZE(BO), %xmm10
	addps	%xmm10, %xmm2
	movaps	-4 * SIZE(AO), %xmm10
	mulps	28 * SIZE(BO), %xmm10
	addps	%xmm10, %xmm3
	movaps	16 * SIZE(AO), %xmm10

	addq   $32 * SIZE, AO
	addq   $32 * SIZE, BO
	decq   %rax
	jne    .L122
	ALIGN_4

.L125:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movaps	ALPHA, %xmm15
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L128
	ALIGN_4

.L126:
	mulps	%xmm8, %xmm9
	movaps	-28 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm0
	movaps	 4 * SIZE(BO), %xmm9

	addq	$4 * SIZE, AO		# aoffset  += 4
	addq	$4 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L126
	ALIGN_4

.L128:
#ifndef TRMMKERNEL
	movlps	0 * SIZE(CO1), %xmm8
	movhps	2 * SIZE(CO1), %xmm8
#endif

	addps	%xmm1, %xmm0
	addps	%xmm3, %xmm2
	addps	%xmm2, %xmm0

	mulps	%xmm15, %xmm0
#ifndef TRMMKERNEL
	addps	%xmm8,  %xmm0
#endif

	movlps	%xmm0, 0 * SIZE(CO1)
	movhps	%xmm0, 2 * SIZE(CO1)

#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, 2), BO
#endif

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

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

.L130:
	testq	$2, M
	je	.L140

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

	leaq	BUFFER, BO
#else
	leaq	BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax,   8), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 2), BO
#endif	

	movaps	-32 * SIZE(AO), %xmm8
	movaps	-24 * SIZE(AO), %xmm10

	movaps	 0 * SIZE(BO), %xmm9
	movaps	16 * SIZE(BO), %xmm11

	pxor	%xmm0, %xmm0
	pxor	%xmm1, %xmm1
	pxor	%xmm2, %xmm2
	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	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$3, %rax
	je	.L135
	ALIGN_4

.L132:
	mulps	%xmm8, %xmm9
#if defined(OPTERON) || defined(BARCELONA)
	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)
#endif
	movlps	-30 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm0
	movlps	 4 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	movlps	-28 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm1
	movlps	 8 * SIZE(BO), %xmm9

	mulps	%xmm8, %xmm9
	movlps	-26 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm0
	movlps	12 * SIZE(BO), %xm