gemm_kernel_8x4_sse.s

Optimized GotoBLAS libraries

	KERNEL4(16 *  2)
	KERNEL5(16 *  2)
	KERNEL6(16 *  2)
	KERNEL7(16 *  2)
	KERNEL8(16 *  2)

	addq	$16 * SIZE, %rax
	je,pn	.L15
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)

	KERNEL1(16 *  2)
	KERNEL2(16 *  2)
	KERNEL3(16 *  2)
	KERNEL4(16 *  2)
	KERNEL5(16 *  2)
	KERNEL6(16 *  2)
	KERNEL7(16 *  2)
	KERNEL8(16 *  2)

	addq	$16 * SIZE, %rax
	je,pn	.L15
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)

	KERNEL1(16 *  2)
	KERNEL2(16 *  2)
	KERNEL3(16 *  2)
	KERNEL4(16 *  2)
	KERNEL5(16 *  2)
	KERNEL6(16 *  2)
	KERNEL7(16 *  2)
	KERNEL8(16 *  2)

	addq	$16 * SIZE, %rax
	je,pn	.L15
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)

	KERNEL1(16 *  2)
	KERNEL2(16 *  2)
	KERNEL3(16 *  2)
	KERNEL4(16 *  2)
	KERNEL5(16 *  2)
	KERNEL6(16 *  2)
	KERNEL7(16 *  2)
	KERNEL8(16 *  2)

	addq	$16 * SIZE, %rax
	je,pn	.L15
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)

	KERNEL1(16 *  2)
	KERNEL2(16 *  2)
	KERNEL3(16 *  2)
	KERNEL4(16 *  2)
	KERNEL5(16 *  2)
	KERNEL6(16 *  2)
	KERNEL7(16 *  2)
	KERNEL8(16 *  2)

	addq	$16 * SIZE, %rax
	je,pn	.L15
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)

	KERNEL1(16 *  2)
	KERNEL2(16 *  2)
	KERNEL3(16 *  2)
	KERNEL4(16 *  2)
	KERNEL5(16 *  2)
	KERNEL6(16 *  2)
	KERNEL7(16 *  2)
	KERNEL8(16 *  2)

	addq	$16 * SIZE, %rax
	jl,pt	.L12
	ALIGN_3

.L15:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	testq	$4, %rax
	je .L16
	xorq	%rax, %rax
	ALIGN_3

	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)

	addq	$64 * SIZE, BO
	addq	$32 * SIZE, AO
	ALIGN_3
#else
	sarq	$2, %rax
	jle,pn	.L16
	ALIGN_3

.L12:
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)

	addq	$ 64 * SIZE, BO
	subq	$-32 * SIZE, AO
	decq	%rax
	jg,pt	.L12
#endif

.L16:
	movaps	ALPHA, %xmm7

#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	andq	$3, %rax		# if (k & 1)
	je .L18

	leaq	(, %rax, 8), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 8), BO
	negq	%rax
	ALIGN_4

.L17:
	mulps	%xmm0, %xmm1
	addps	%xmm1, %xmm8
	movaps	-28 * SIZE(BO, %rax, 8), %xmm1
	mulps	%xmm0, %xmm1
	addps	%xmm1, %xmm9
	movaps	-24 * SIZE(BO, %rax, 8), %xmm1
	mulps	%xmm0, %xmm1
	mulps	-20 * SIZE(BO, %rax, 8), %xmm0
	addps	%xmm1, %xmm10
	movaps	-32 * SIZE(BO, %rax, 8), %xmm1
	addps	%xmm0, %xmm11
	movaps	-24 * SIZE(AO, %rax, 4), %xmm0
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm12
	movaps	-28 * SIZE(BO, %rax, 8), %xmm1
	mulps	%xmm2, %xmm1
	addps	%xmm1, %xmm13
	movaps	-24 * SIZE(BO, %rax, 8), %xmm1
	mulps	%xmm2, %xmm1
	mulps	-20 * SIZE(BO, %rax, 8), %xmm2
	addps	%xmm1, %xmm14
	movaps	-16 * SIZE(BO, %rax, 8), %xmm1
	addps	%xmm2, %xmm15
	movaps	-20 * SIZE(AO, %rax, 4), %xmm2

	addq	$SIZE * 2, %rax
	jl	.L17
	ALIGN_4

.L18:
#ifndef TRMMKERNEL
	movlps	0 * SIZE(CO1), %xmm0
	movhps	2 * SIZE(CO1), %xmm0
	movlps	4 * SIZE(CO1), %xmm1
	movhps	6 * SIZE(CO1), %xmm1

	movlps	0 * SIZE(CO2), %xmm2
	movhps	2 * SIZE(CO2), %xmm2
	movlps	4 * SIZE(CO2), %xmm3
	movhps	6 * SIZE(CO2), %xmm3
#endif

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

	mulps	%xmm7, %xmm12
	mulps	%xmm7, %xmm13
	mulps	%xmm7, %xmm14
	mulps	%xmm7, %xmm15

#ifndef TRMMKERNEL
	movlps	0 * SIZE(CO1, LDC, 2), %xmm4
	movhps	2 * SIZE(CO1, LDC, 2), %xmm4
	movlps	4 * SIZE(CO1, LDC, 2), %xmm5
	movhps	6 * SIZE(CO1, LDC, 2), %xmm5

	movlps	0 * SIZE(CO2, LDC, 2), %xmm6
	movhps	2 * SIZE(CO2, LDC, 2), %xmm6
	movlps	4 * SIZE(CO2, LDC, 2), %xmm7
	movhps	6 * SIZE(CO2, LDC, 2), %xmm7

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

	movlps	%xmm8,  0 * SIZE(CO1)
	movhps	%xmm8,  2 * SIZE(CO1)
	movlps	%xmm12, 4 * SIZE(CO1)
	movhps	%xmm12, 6 * SIZE(CO1)

	movlps	%xmm9,  0 * SIZE(CO2)
	movhps	%xmm9,  2 * SIZE(CO2)
	movlps	%xmm13, 4 * SIZE(CO2)
	movhps	%xmm13, 6 * SIZE(CO2)

#ifndef TRMMKERNEL
	addps	%xmm4, %xmm10
	addps	%xmm5, %xmm14
	addps	%xmm6, %xmm11
	addps	%xmm7, %xmm15
#endif

	movlps	%xmm10, 0 * SIZE(CO1, LDC, 2)
	movhps	%xmm10, 2 * SIZE(CO1, LDC, 2)
	movlps	%xmm14, 4 * SIZE(CO1, LDC, 2)
	movhps	%xmm14, 6 * SIZE(CO1, LDC, 2)

	movlps	%xmm11, 0 * SIZE(CO2, LDC, 2)
	movhps	%xmm11, 2 * SIZE(CO2, LDC, 2)
	movlps	%xmm15, 4 * SIZE(CO2, LDC, 2)
	movhps	%xmm15, 6 * 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, 4), AO
	leaq	(BO, %rax, 8), BO
#endif

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

	addq	$8 * SIZE, CO1		# coffset += 4
	addq	$8 * SIZE, CO2		# coffset += 4
	decq	I			# i --
	jg	.L11
	ALIGN_4	

.L20:
	testq	$4, M
	je	.L30

#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, 8), BO
#endif	

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

	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
	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	$4, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$3, %rax
	je	.L25
	ALIGN_4

.L22:
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm0
#if defined(OPTERON) || defined(BARCELONA)
	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)
#endif
	movaps	 4 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm1
	movaps	 8 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	mulps	12 * SIZE(BO), %xmm8
	addps	%xmm9, %xmm2
	movaps	64 * SIZE(BO), %xmm9
	addps	%xmm8, %xmm3
	movaps	-28 * SIZE(AO), %xmm8

	mulps	%xmm8, %xmm11
	addps	%xmm11, %xmm0
	movaps	20 * SIZE(BO), %xmm11
	mulps	%xmm8, %xmm11
	addps	%xmm11, %xmm1
	movaps	24 * SIZE(BO), %xmm11
	mulps	%xmm8, %xmm11
	mulps	28 * SIZE(BO), %xmm8
	addps	%xmm11, %xmm2
	movaps	80 * SIZE(BO), %xmm11
	addps	%xmm8, %xmm3
	movaps	-24 * SIZE(AO), %xmm8

	mulps	%xmm8, %xmm13
	addps	%xmm13, %xmm0
	movaps	36 * SIZE(BO), %xmm13
	mulps	%xmm8, %xmm13
	addps	%xmm13, %xmm1
	movaps	40 * SIZE(BO), %xmm13
	mulps	%xmm8, %xmm13
	mulps	44 * SIZE(BO), %xmm8
	addps	%xmm13, %xmm2
	movaps	96 * SIZE(BO), %xmm13
	addps	%xmm8, %xmm3
	movaps	-20 * SIZE(AO), %xmm8

	mulps	%xmm8, %xmm15
	addps	%xmm15, %xmm0
	movaps	52 * SIZE(BO), %xmm15
	mulps	%xmm8, %xmm15
	addps	%xmm15, %xmm1
	movaps	56 * SIZE(BO), %xmm15
	mulps	%xmm8, %xmm15
	mulps	60 * SIZE(BO), %xmm8
	addps	%xmm15, %xmm2
	movaps	112 * SIZE(BO), %xmm15
	addps	%xmm8, %xmm3
	movaps	 0 * SIZE(AO), %xmm8

#if defined(OPTERON) || defined(BARCELONA)
	PREFETCH	(PREFETCHSIZE + 16) * SIZE(AO)
#endif
	mulps	%xmm10, %xmm9
	addps	%xmm9, %xmm0
	movaps	68 * SIZE(BO), %xmm9
	mulps	%xmm10, %xmm9
	addps	%xmm9, %xmm1
	movaps	72 * SIZE(BO), %xmm9
	mulps	%xmm10, %xmm9
	mulps	76 * SIZE(BO), %xmm10
	addps	%xmm9, %xmm2
	movaps	128 * SIZE(BO), %xmm9
	addps	%xmm10, %xmm3
	movaps	-12 * SIZE(AO), %xmm10

	mulps	%xmm10, %xmm11
	addps	%xmm11, %xmm0
	movaps	84 * SIZE(BO), %xmm11
	mulps	%xmm10, %xmm11
	addps	%xmm11, %xmm1
	movaps	88 * SIZE(BO), %xmm11
	mulps	%xmm10, %xmm11
	mulps	92 * SIZE(BO), %xmm10
	addps	%xmm11, %xmm2
	movaps	144 * SIZE(BO), %xmm11
	addps	%xmm10, %xmm3
	movaps	-8 * SIZE(AO), %xmm10

	mulps	%xmm10, %xmm13
	addps	%xmm13, %xmm0
	movaps	100 * SIZE(BO), %xmm13
	mulps	%xmm10, %xmm13
	addps	%xmm13, %xmm1
	movaps	104 * SIZE(BO), %xmm13
	mulps	%xmm10, %xmm13
	mulps	108 * SIZE(BO), %xmm10
	addps	%xmm13, %xmm2
	movaps	160 * SIZE(BO), %xmm13
	addps	%xmm10, %xmm3
	movaps	-4 * SIZE(AO), %xmm10

	mulps	%xmm10, %xmm15
	addps	%xmm15, %xmm0
	movaps	116 * SIZE(BO), %xmm15
	mulps	%xmm10, %xmm15
	addps	%xmm15, %xmm1
	movaps	120 * SIZE(BO), %xmm15
	mulps	%xmm10, %xmm15
	mulps	124 * SIZE(BO), %xmm10
	addps	%xmm15, %xmm2
	movaps	176 * SIZE(BO), %xmm15
	addps	%xmm10, %xmm3
	movaps	16 * SIZE(AO), %xmm10

	addq   $ 32 * SIZE, AO
	addq   $128 * SIZE, BO
	decq   %rax
	jne    .L22
	ALIGN_4

.L25:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movaps	ALPHA, %xmm15
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L28
	ALIGN_4

.L26:
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm0
	movaps	 4 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm1
	movaps	 8 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	mulps	12 * SIZE(BO), %xmm8
	addps	%xmm9, %xmm2
	movaps	16 * SIZE(BO), %xmm9
	addps	%xmm8, %xmm3
	movaps	-28 * SIZE(AO), %xmm8

	addq	$ 4 * SIZE, AO		# aoffset  += 4
	addq	$16 * SIZE, BO		# boffset1 += 8
	decq	%rax
	jg	.L26
	ALIGN_4

.L28:
	mulps	%xmm15, %xmm0
	mulps	%xmm15, %xmm1
 	mulps	%xmm15, %xmm2
	mulps	%xmm15, %xmm3

#ifndef TRMMKERNEL
	movlps	0 * SIZE(CO1), %xmm8
	movhps	2 * SIZE(CO1), %xmm8
	movlps	0 * SIZE(CO2), %xmm10
	movhps	2 * SIZE(CO2), %xmm10

	movlps	0 * SIZE(CO1, LDC, 2), %xmm12
	movhps	2 * SIZE(CO1, LDC, 2), %xmm12
	movlps	0 * SIZE(CO2, LDC, 2), %xmm14
	movhps	2 * SIZE(CO2, LDC, 2), %xmm14

	addps	%xmm8,  %xmm0
	addps	%xmm10, %xmm1
	addps	%xmm12, %xmm2
	addps	%xmm14, %xmm3
#endif

	movlps	%xmm0, 0 * SIZE(CO1)
	movhps	%xmm0, 2 * SIZE(CO1)
	movlps	%xmm1, 0 * SIZE(CO2)
	movhps	%xmm1, 2 * SIZE(CO2)

	movlps	%xmm2, 0 * SIZE(CO1, LDC, 2)
	movhps	%xmm2, 2 * SIZE(CO1, LDC, 2)
	movlps	%xmm3, 0 * SIZE(CO2, LDC, 2)
	movhps	%xmm3, 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		# coffset += 4
	addq	$4 * SIZE, CO2		# coffset += 4
	ALIGN_4	

.L30:
	testq	$2, M
	je	.L40

#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, 8), BO
#endif	

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

	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
	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	$4, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$3, %rax
	je	.L35
	ALIGN_4

.L32:
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm0
#if defined(OPTERON) || defined(BARCELONA)
	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)
#endif
	movlps	 4 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm1
	movlps	 8 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm2
	movlps	12 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	movlps	-30 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm3
	movlps	64 * SIZE(BO), %xmm9

	mulps	%xmm8, %xmm11
	addps	%xmm11, %xmm0
	movlps	20 * SIZE(BO), %xmm11
	mulps	%xmm8, %xmm11
	addps	%xmm11, %xmm1
	movlps	24 * SIZE(BO), %xmm11
	mulps	%xmm8, %xmm11
	addps	%xmm11, %xmm2
	movlps	28 * SIZE(BO), %xmm11
	mulps	%xmm8, %xmm11
	movlps	-28 * SIZE(AO), %xmm8
	addps	%xmm11, %xmm3
	movlps	80 * SIZE(BO), %xmm11

	mulps	%xmm8, %xmm13
	addps	%xmm13, %xmm0
	movlps	36 * SIZE(BO), %xmm13
	mulps	%xmm8, %xmm13
	addps	%xmm13, %xmm1
	movlps	40 * SIZE(BO), %xmm13
	mulps	%xmm8, %xmm13
	addps	%xmm13, %xmm2
	movlps	44 * SIZE(BO), %xmm13
	mulps	%xmm8, %xmm13
	movlps	-26 * SIZE(AO), %xmm8
	addps	%xmm13, %xmm3
	movlps	96 * SIZE(BO), %xmm13

	mulps	%xmm8, %xmm15
	addps	%xmm15, %xmm0
	movlps	52 * SIZE(BO), %xmm15
	mulps	%xmm8, %xmm15