zgemm_kernel_4x1_sse.s

Optimized GotoBLAS libraries

	KERNEL3(32  *  4)
	KERNEL4(32  *  4)
	KERNEL5(32  *  4)
	KERNEL6(32  *  4)
	KERNEL7(32  *  4)
	KERNEL8(32  *  4)
	cmpl	$64 *  5, %eax
	jle,pn	.L11
	KERNEL1(32  *  5)
	KERNEL2(32  *  5)
	KERNEL3(32  *  5)
	KERNEL4(32  *  5)
	KERNEL5(32  *  5)
	KERNEL6(32  *  5)
	KERNEL7(32  *  5)
	KERNEL8(32  *  5)
	cmpl	$64 *  6, %eax
	jle,pn	.L11
	KERNEL1(32  *  6)
	KERNEL2(32  *  6)
	KERNEL3(32  *  6)
	KERNEL4(32  *  6)
	KERNEL5(32  *  6)
	KERNEL6(32  *  6)
	KERNEL7(32  *  6)
	KERNEL8(32  *  6)
	cmpl	$64 *  7, %eax
	jle,pn	.L11
	KERNEL1(32  *  7)
	KERNEL2(32  *  7)
	KERNEL3(32  *  7)
	KERNEL4(32  *  7)
	KERNEL5(32  *  7)
	KERNEL6(32  *  7)
	KERNEL7(32  *  7)
	KERNEL8(32  *  7)

	addl	$128 * 4  * SIZE, AA
	addl	$128 * 4  * SIZE, BB
	subl	$ 64 * 8, %eax
	jg,pt	.L1X

.L11:
	leal	(AA, %eax, 4), AA
	leal	(BB, %eax, 4), BB

#else

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

	leal	BUFFER, BB
	movaps	 0 * SIZE + BUFFER, %xmm2
	XORPS	%xmm4, %xmm4
	movaps	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movaps	 8 * SIZE + BUFFER, %xmm3
	XORPS	%xmm6, %xmm6
	movaps	 8 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7

#else

	leal	BUFFER, BB
	movl	KK, %eax
	leal	(, %eax,   8), %eax
	leal	(AA, %eax, 4), AA
	leal	(BB, %eax, 4), BB /* because it's doubled */

	movaps	 0 * SIZE(BB), %xmm2
	XORPS	%xmm4, %xmm4
	movaps	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movaps	 8 * SIZE(BB), %xmm3
	XORPS	%xmm6, %xmm6
	movaps	 8 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7

#endif	

#ifndef TRMMKERNEL
	movl	K, %eax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movl	K, %eax
	subl	KK, %eax
	movl	%eax, KKK	
#else
	movl	KK, %eax
#ifdef LEFT
	addl	$4, %eax
#else
	addl	$1, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	prefetcht0   8 * SIZE(%esi)
	je	.L12
	ALIGN_4

#define PREFETCHSIZE 48

.L11:
#ifdef CORE_KATMAI
	prefetcht0	PREFETCHSIZE * SIZE(AA)
#endif

	mulps	%xmm0, %xmm2
	mulps	 4 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm4
	movaps	 0 * SIZE(BB), %xmm2

	addps	%xmm0, %xmm5
	movaps	 4 * SIZE(AA), %xmm0
	mulps	%xmm0, %xmm2
	mulps	 4 * SIZE(BB), %xmm0

	addps	%xmm2, %xmm6
	movaps	16 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm7
	movaps	16 * SIZE(AA), %xmm0

#ifdef CORE_KATMAI
	prefetcht0	(PREFETCHSIZE + 8) * SIZE(AA)
#endif

	mulps	%xmm1, %xmm3
	mulps	12 * SIZE(BB), %xmm1
	addps	%xmm3, %xmm4
	movaps	 8 * SIZE(BB), %xmm3

	addps	%xmm1, %xmm5
	movaps	12 * SIZE(AA), %xmm1
	mulps	%xmm1, %xmm3
	mulps	12 * SIZE(BB), %xmm1

	addps	%xmm3, %xmm6
	movaps	24 * SIZE(BB), %xmm3
	addps	%xmm1, %xmm7
	movaps	24 * SIZE(AA), %xmm1

#ifdef CORE_KATMAI
	prefetcht0	(PREFETCHSIZE + 16) * SIZE(AA)
#endif

	mulps	%xmm0, %xmm2
	mulps	20 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm4
	movaps	16 * SIZE(BB), %xmm2

	addps	%xmm0, %xmm5
	movaps	20 * SIZE(AA), %xmm0
	mulps	%xmm0, %xmm2
	mulps	20 * SIZE(BB), %xmm0

	addps	%xmm2, %xmm6
	movaps	32 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm7
	movaps	32 * SIZE(AA), %xmm0

#ifdef CORE_KATMAI
	prefetcht0	(PREFETCHSIZE + 24) * SIZE(AA)
#endif

	mulps	%xmm1, %xmm3
	mulps	28 * SIZE(BB), %xmm1
	addps	%xmm3, %xmm4
	movaps	24 * SIZE(BB), %xmm3

	addps	%xmm1, %xmm5
	movaps	28 * SIZE(AA), %xmm1
	mulps	%xmm1, %xmm3
	mulps	28 * SIZE(BB), %xmm1

	addps	%xmm3, %xmm6
	movaps	40 * SIZE(BB), %xmm3
	addps	%xmm1, %xmm7
	movaps	40 * SIZE(AA), %xmm1

#ifdef CORE_KATMAI
	prefetcht0	(PREFETCHSIZE + 32) * SIZE(AA)
#endif

	mulps	%xmm0, %xmm2
	mulps	36 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm4
	movaps	32 * SIZE(BB), %xmm2

	addps	%xmm0, %xmm5
	movaps	36 * SIZE(AA), %xmm0
	mulps	%xmm0, %xmm2
	mulps	36 * SIZE(BB), %xmm0

	addps	%xmm2, %xmm6
	movaps	48 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm7
	movaps	48 * SIZE(AA), %xmm0

#ifdef CORE_KATMAI
	prefetcht0	(PREFETCHSIZE + 40) * SIZE(AA)
#endif

	mulps	%xmm1, %xmm3
	mulps	44 * SIZE(BB), %xmm1
	addps	%xmm3, %xmm4
	movaps	40 * SIZE(BB), %xmm3

	addps	%xmm1, %xmm5
	movaps	44 * SIZE(AA), %xmm1
	mulps	%xmm1, %xmm3
	mulps	44 * SIZE(BB), %xmm1

	addps	%xmm3, %xmm6
	movaps	56 * SIZE(BB), %xmm3
	addps	%xmm1, %xmm7
	movaps	56 * SIZE(AA), %xmm1

#ifdef CORE_KATMAI
	prefetcht0	(PREFETCHSIZE + 48) * SIZE(AA)
#endif

	mulps	%xmm0, %xmm2
	mulps	52 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm4
	movaps	48 * SIZE(BB), %xmm2

	addps	%xmm0, %xmm5
	movaps	52 * SIZE(AA), %xmm0
	mulps	%xmm0, %xmm2
	mulps	52 * SIZE(BB), %xmm0

	addps	%xmm2, %xmm6
	movaps	64 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm7
	movaps	64 * SIZE(AA), %xmm0

#ifdef CORE_KATMAI
	prefetcht0	(PREFETCHSIZE + 56) * SIZE(AA)
#endif

	mulps	%xmm1, %xmm3
	mulps	60 * SIZE(BB), %xmm1
	addps	%xmm3, %xmm4
	movaps	56 * SIZE(BB), %xmm3

	addps	%xmm1, %xmm5
	movaps	60 * SIZE(AA), %xmm1
	mulps	%xmm1, %xmm3
	mulps	60 * SIZE(BB), %xmm1

	addps	%xmm3, %xmm6
	movaps	72 * SIZE(BB), %xmm3
	addps	%xmm1, %xmm7
	movaps	72 * SIZE(AA), %xmm1

	addl	$64 * SIZE, BB
	addl	$64 * SIZE, AA
	decl	%eax
	jne	.L11
#endif
	
.L12:
#ifndef TRMMKERNEL
	movl	K, %eax
#else
	movl	KKK, %eax
#endif
	movaps	ALPHA_R, %xmm1
	movaps	ALPHA_I, %xmm3
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L14

.L13:
	mulps	%xmm0, %xmm2
	mulps	 4 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm4
	movaps	 0 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm5
	movaps	 4 * SIZE(AA), %xmm0
	mulps	%xmm0, %xmm2
	mulps	 4 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm6
	movaps	 8 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm7
	movaps	 8 * SIZE(AA), %xmm0

	addl	$8 * SIZE, AA		# aoffset  += 8
	addl	$8 * SIZE, BB		# boffset1 += 8

	decl	%eax
	jg	.L13

.L14:
	shufps	$0xb1, %xmm5, %xmm5
	shufps	$0xb1, %xmm7, %xmm7

#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
    defined(RR) || defined(RC) || defined(CR) || defined(CC)
	subps	%xmm5, %xmm4
	subps	%xmm7, %xmm6
#else
	addps	%xmm5, %xmm4
	addps	%xmm7, %xmm6
#endif

	movaps	%xmm4, %xmm5
	movaps	%xmm6, %xmm7

	shufps	$0xb1, %xmm4, %xmm4
	shufps	$0xb1, %xmm6, %xmm6

	mulps	%xmm1, %xmm5
	mulps	%xmm3, %xmm4
	mulps	%xmm1, %xmm7
	mulps	%xmm3, %xmm6

	addps	%xmm5, %xmm4
	addps	%xmm7, %xmm6

	shufps	$0xe4, %xmm4, %xmm4
	shufps	$0xe4, %xmm6, %xmm6

#ifndef TRMMKERNEL
	MOVSD	0 * SIZE(%esi), %xmm0
	movhps	2 * SIZE(%esi), %xmm0
	MOVSD	4 * SIZE(%esi), %xmm2
	movhps	6 * SIZE(%esi), %xmm2

	addps	%xmm0, %xmm4
	addps	%xmm2, %xmm6
#endif

	MOVSD	%xmm4, 0 * SIZE(%esi)
	movhps	%xmm4, 2 * SIZE(%esi)
	MOVSD	%xmm6, 4 * SIZE(%esi)
	movhps	%xmm6, 6 * SIZE(%esi)

#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
	movl	K, %eax
	subl	KKK, %eax
	leal	(,%eax,    8), %eax
	leal	(AA, %eax, 4), AA
	leal	(BB, %eax, 4), BB
#endif

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

	addl	$8 * SIZE, %esi		# coffset += 4
	decl	%ebx			# i --
	jg	.L10
	ALIGN_2
	
.L50:
	movl	M,  %ebx
	testl	$2, %ebx
	jle	.L70


#if (L1_DATA_LINESIZE == 64)

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

	leal	BUFFER, BB
	movaps	 0 * SIZE + BUFFER, %xmm2
	XORPS	%xmm4, %xmm4
	movaps	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movaps	 16 * SIZE + BUFFER, %xmm3
	XORPS	%xmm6, %xmm6
	movaps	 16 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7

#else

	leal	BUFFER, BB
	movl	KK, %eax
	leal	(, %eax,   8), %eax
	leal	(AA, %eax, 2), AA
	leal	(BB, %eax, 4), BB /* because it's doubled */

	movaps	 0 * SIZE(BB), %xmm2
	XORPS	%xmm4, %xmm4
	movaps	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movaps	 16 * SIZE(BB), %xmm3
	XORPS	%xmm6, %xmm6
	movaps	 16 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7

#endif	

#ifndef TRMMKERNEL
	movl	K, %eax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movl	K, %eax
	subl	KK, %eax
	movl	%eax, KKK	
#else
	movl	KK, %eax
#ifdef LEFT
	addl	$2, %eax
#else
	addl	$1, %eax
#endif
	movl	%eax, KKK
#endif
	sarl	$3, %eax
	je	.L52
	ALIGN_4

.L51:
	mulps	%xmm0, %xmm2
	mulps	 4 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm4
	movaps	 8 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm5
	movaps	 4 * SIZE(AA), %xmm0
	mulps	%xmm0, %xmm2
	mulps	12 * SIZE(BB), %xmm0
	addps	%xmm2, %xmm6
	movaps	32 * SIZE(BB), %xmm2
	addps	%xmm0, %xmm7
	movaps	 8 * SIZE(AA), %xmm0
	mulps	%xmm0, %xmm3
	mulps	20 * SIZE(BB), %xmm0
	addps	%xmm3, %xmm4
	movaps	24 * SIZE(BB), %xmm3
	addps	%xmm0, %xmm5
	movaps	12 * SIZE(AA), %xmm0
	mulps	%xmm0, %xmm3
	mulps	28 * SIZE(BB), %xmm0
	addps	%xmm3, %xmm6
	movaps	48 * SIZE(BB), %xmm3
	addps	%xmm0, %xmm7
	movaps	32 * SIZE(AA), %xmm0
	mulps	%xmm1, %xmm2
	mulps	36 * SIZE(BB), %xmm1
	addps	%xmm2, %xmm4
	movaps	40 * SIZE(BB), %xmm2
	addps	%xmm1, %xmm5
	movaps	20 * SIZE(AA), %xmm1
	mulps	%xmm1, %xmm2
	mulps	44 * SIZE(BB), %xmm1
	addps	%xmm2, %xmm6
	movaps	64 * SIZE(BB), %xmm2
	addps	%xmm1, %xmm7
	movaps	24 * SIZE(AA), %xmm1
	mulps	%xmm1, %xmm3
	mulps	52 * SIZE(BB), %xmm1
	addps	%xmm3, %xmm4
	movaps	56 * SIZE(BB), %xmm3
	addps	%xmm1, %xmm5
	movaps	28 * SIZE(AA), %xmm1
	mulps	%xmm1, %xmm3
	mulps	60 * SIZE(BB), %xmm1
	addps	%xmm3, %xmm6
	movaps	80 * SIZE(BB), %xmm3
	addps	%xmm1, %xmm7
	movaps	48 * SIZE(AA), %xmm1

	addl	$32 * SIZE, AA
	addl	$64 * SIZE, BB
	decl	%eax
	jne	.L51
	ALIGN_2

#else

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

	leal	BUFFER, BB
	movaps	 0 * SIZE + BUFFER, %xmm2
	XORPS	%xmm4, %xmm4
	movaps	 0 * SIZE(AA), %xmm0
	XORPS	%xmm5, %xmm5
	movaps	 8 * SIZE + BUFFER, %xmm3
	XORPS	%xmm6, %xmm6
	movaps	 8 * SIZE(AA), %xmm1
	XORPS	%xmm7, %xmm7

#else

	leal	BUFFER, BB
	movl	KK, %eax