gemm_kernel_8x4_sse3.s

Optimized GotoBLAS libraries

#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	%xmm8, %xmm9
	movaps	  4 * SIZE(AO), %xmm8
	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)
	addps	%xmm9, %xmm0
	movddup   0 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
 	movaps	  8 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm4
	movddup   2 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	movaps	 12 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm1
	movddup   2 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
 	movaps	 64 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm5
	movddup   4 * SIZE(BO), %xmm9
	mulps	%xmm10, %xmm9
	movaps	 20 * SIZE(AO), %xmm10
	addps	%xmm9, %xmm0
	movddup   4 * SIZE(BO), %xmm9
	mulps	%xmm10, %xmm9
 	movaps	 24 * SIZE(AO), %xmm10
	addps	%xmm9, %xmm4
	movddup   6 * SIZE(BO), %xmm9
	mulps	%xmm10, %xmm9
	movaps	 28 * SIZE(AO), %xmm10
	addps	%xmm9, %xmm1
	movddup   6 * SIZE(BO), %xmm9
	mulps	%xmm10, %xmm9
 	movaps	 80 * SIZE(AO), %xmm10
	addps	%xmm9, %xmm5
	PREFETCH  (PREFETCHSIZE + 32) * SIZE(AO)
	movddup   8 * SIZE(BO), %xmm9
	mulps	%xmm12, %xmm9
	movaps	 36 * SIZE(AO), %xmm12
	addps	%xmm9, %xmm0
	movddup  16 * SIZE(BO), %xmm9
	mulps	%xmm12, %xmm11
 	movaps	 40 * SIZE(AO), %xmm12
	addps	%xmm11, %xmm4
	movddup  10 * SIZE(BO), %xmm11
	mulps	%xmm12, %xmm11
	movaps	 44 * SIZE(AO), %xmm12
	addps	%xmm11, %xmm1
	movddup  10 * SIZE(BO), %xmm11
	mulps	%xmm12, %xmm11
 	movaps	 96 * SIZE(AO), %xmm12
	addps	%xmm11, %xmm5
	movddup  12 * SIZE(BO), %xmm11
	mulps	%xmm14, %xmm11
	movaps	 52 * SIZE(AO), %xmm14
	addps	%xmm11, %xmm0
	movddup  12 * SIZE(BO), %xmm11
	mulps	%xmm14, %xmm11
 	movaps	 56 * SIZE(AO), %xmm14
	addps	%xmm11, %xmm4
	movddup  14 * SIZE(BO), %xmm11
	mulps	%xmm14, %xmm11
	movaps	 60 * SIZE(AO), %xmm14
	addps	%xmm11, %xmm1
	movddup  14 * SIZE(BO), %xmm11
	mulps	%xmm14, %xmm11
 	movaps	112 * SIZE(AO), %xmm14
	addps	%xmm11, %xmm5
	movddup  24 * SIZE(BO), %xmm11

	addq   $64 * SIZE, AO
	addq   $16 * 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	%xmm8, %xmm9
	movaps	  4 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm0
	movddup   0 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
 	movaps	  8 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm4
	movddup   2 * SIZE(BO), %xmm9

	addq	$8 * SIZE, AO
	addq	$2 * SIZE, BO
	decq	%rax
	jg	.L116
	ALIGN_4

.L118:
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movsd	0 * SIZE(CO1), %xmm8
	movhps	2 * SIZE(CO1), %xmm8
	movsd	4 * SIZE(CO1), %xmm9
	movhps	6 * SIZE(CO1), %xmm9
#endif

	addps	%xmm1, %xmm0
	addps	%xmm5, %xmm4

	mulps	%xmm15, %xmm0
	mulps	%xmm15, %xmm4
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addps	%xmm8,  %xmm0
	addps	%xmm9,  %xmm4
#endif

	movsd	%xmm0, 0 * SIZE(CO1)
	movhps	%xmm0, 2 * SIZE(CO1)
	movsd	%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, 1), 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, 1), BO
#endif	

	movaps	 0 * SIZE(AO), %xmm8
	movaps	16 * SIZE(AO), %xmm10

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

 	movaps	  0 * SIZE(AO), %xmm8
	movddup   0 * SIZE(BO), %xmm9
 	movaps	 16 * SIZE(AO), %xmm10
	movddup   8 * SIZE(BO), %xmm11

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

#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
 	movaps	  4 * SIZE(AO), %xmm8
	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)
	addps	%xmm9, %xmm0
	movddup   2 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
 	movaps	  8 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm1
	movddup   4 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
 	movaps	 12 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm0
	movddup   6 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
 	movaps	 32 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm1
	movddup  16 * SIZE(BO), %xmm9
	mulps	%xmm10, %xmm11
 	movaps	 20 * SIZE(AO), %xmm10
	addps	%xmm11, %xmm0
	movddup  10 * SIZE(BO), %xmm11
	mulps	%xmm10, %xmm11
 	movaps	 24 * SIZE(AO), %xmm10
	addps	%xmm11, %xmm1
	movddup  12 * SIZE(BO), %xmm11
	mulps	%xmm10, %xmm11
 	movaps	 28 * SIZE(AO), %xmm10
	addps	%xmm11, %xmm0
	movddup  14 * SIZE(BO), %xmm11
	mulps	%xmm10, %xmm11
 	movaps	 48 * SIZE(AO), %xmm10
	addps	%xmm11, %xmm1
	movddup  24 * SIZE(BO), %xmm11

	addq   $32 * SIZE, AO
	addq   $16 * 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	  4 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm0
	movddup   2 * SIZE(BO), %xmm9

	addq	$4 * SIZE, AO
	addq	$2 * SIZE, BO
	decq	%rax
	jg	.L126
	ALIGN_4

.L128:
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movsd	0 * SIZE(CO1), %xmm8
	movhps	2 * SIZE(CO1), %xmm8
#endif

	addps	%xmm1, %xmm0
	mulps	%xmm15, %xmm0

#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addps	%xmm8,  %xmm0
#endif

	movsd	%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, 1), 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, 1), BO
#endif	

 	movaps	  0 * SIZE(AO), %xmm8
	movaps    0 * SIZE(BO), %xmm9
 	movaps	 16 * SIZE(AO), %xmm10
	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	$4, %rax
	je	.L135
	ALIGN_4

.L132:
	mulps	%xmm8, %xmm9
	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)
 	movaps	  4 * SIZE(AO), %xmm8
	addps	%xmm9, %xmm0
	mulps     4 * SIZE(BO), %xmm8
	addps	%xmm8, %xmm1
 	movaps	  8 * SIZE(AO), %xmm8
	mulps     8 * SIZE(BO), %xmm8
	addps	%xmm8, %xmm2
 	movaps	 12 * SIZE(AO), %xmm8
	mulps    12 * SIZE(BO), %xmm8
	addps	%xmm8, %xmm3
 	movaps	 32 * SIZE(AO), %xmm8
	movaps   32 * SIZE(BO), %xmm9
	mulps	%xmm10, %xmm11
 	movaps	 20 * SIZE(AO), %xmm10
	addps	%xmm11, %xmm0
	movaps   48 * SIZE(BO), %xmm11
	mulps    20 * SIZE(BO), %xmm10
	addps	%xmm10, %xmm1
 	movaps	 24 * SIZE(AO), %xmm10
	mulps    24 * SIZE(BO), %xmm10
	addps	%xmm10, %xmm2
 	movaps	 28 * SIZE(AO), %xmm10
	mulps    28 * SIZE(BO), %xmm10
	addps	%xmm10, %xmm3
 	movaps	 48 * SIZE(AO), %xmm10

	addq   $32 * SIZE, AO
	addq   $32 * SIZE, BO
	decq   %rax
	jne    .L132
	ALIGN_4

.L135:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movaps	ALPHA, %xmm15
	andq	$15, %rax		# if (k & 1)
	BRANCH
	je .L138
	ALIGN_4

.L136:
 	movsd	  0 * SIZE(AO), %xmm8
	movsd     0 * SIZE(BO), %xmm9

	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm0

	addq	$2 * SIZE, AO
	addq	$2 * SIZE, BO
	decq	%rax
	jg	.L136
	ALIGN_4

.L138:
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movsd	0 * SIZE(CO1), %xmm8
#endif
	addps	%xmm1, %xmm0
	addps	%xmm3, %xmm2
	addps	%xmm2, %xmm0

	movhlps	%xmm0, %xmm1
	addps	%xmm1, %xmm0

	mulps	%xmm15, %xmm0
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addps	%xmm8,  %xmm0
#endif
	movsd	%xmm0, 0 * 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, 1), AO
	leaq	(BO, %rax, 1), BO
#endif

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

	addq	$2 * SIZE, CO1		# coffset += 4
	ALIGN_4	

.L140:
	testq	$1, M
	je	.L999

#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,   4), %rax
	leaq	(AO, %rax, 1), AO
	leaq	(BO, %rax, 2), BO
#endif	

	movss	 0 * SIZE(AO), %xmm8
	movss	 4 * SIZE(AO), %xmm10
	movss	 0 * SIZE(BO), %xmm9
	movss	 8 * 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	$1, %rax
#else
	addq	$1, %rax
#endif
	movq	%rax, KKK
#endif
	sarq	$3, %rax
	je	.L145
	ALIGN_4

.L142:
	mulss	%xmm8, %xmm9
	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)
	movss	 1 * SIZE(AO), %xmm8
	mulss	 2 * SIZE(BO), %xmm8
	addss	%xmm9, %xmm0
	movss	16 * SIZE(BO), %xmm9
	addss	%xmm8, %xmm1
	movss	 2 * SIZE(AO), %xmm8
	mulss	 4 * SIZE(BO), %xmm8
	addss	%xmm8, %xmm2
	movss	 3 * SIZE(AO), %xmm8
	mulss	 6 * SIZE(BO), %xmm8
	addss	%xmm8, %xmm3
	movss	 8 * SIZE(AO), %xmm8
	mulss	%xmm10, %xmm11
	movss	 5 * SIZE(AO), %xmm10
	mulss	10 * SIZE(BO), %xmm10
	addss	%xmm11, %xmm0
	movss	24 * SIZE(BO), %xmm11
	addss	%xmm10, %xmm1
	movss	 6 * SIZE(AO), %xmm10
	mulss	12 * SIZE(BO), %xmm10
	addss	%xmm10, %xmm2
	movss	 7 * SIZE(AO), %xmm10
	mulss	14 * SIZE(BO), %xmm10
	addss	%xmm10, %xmm3
	movss	12 * SIZE(AO), %xmm10

	addq   $ 8 * SIZE, AO
	addq   $16 * SIZE, BO
	decq   %rax
	jne    .L142
	ALIGN_4

.L145:
#ifndef TRMMKERNEL
	movq	K, %rax
#else
	movq	KKK, %rax
#endif
	movss	ALPHA, %xmm15
	andq	$7, %rax		# if (k & 1)
	BRANCH
	je .L148
	ALIGN_4

.L146:
 	movss	  0 * SIZE(AO), %xmm8
	movss     0 * SIZE(BO), %xmm9
	mulps	%xmm8, %xmm9
	addps	%xmm9, %xmm0

	addq	$1 * SIZE, AO
	addq	$2 * SIZE, BO
	decq	%rax
	jg	.L146
	ALIGN_4

.L148:
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	movss	0 * SIZE(CO1), %xmm8
#endif
	addss	%xmm1, %xmm0
	addss	%xmm3, %xmm2
	addss	%xmm2, %xmm0

	mulss	%xmm15, %xmm0
#if! defined(TRMMKERNEL) && !defined(BETAZERO)
	addss	%xmm8,  %xmm0
#endif
	movss	%xmm0, 0 * SIZE(CO1)
	ALIGN_4	
	
.L999:
	movq	%rbx, %rsp
	EMMS
	movq	  0(%rsp), %rbx
	movq	  8(%rsp), %rbp
	movq	 16(%rsp), %r12
	movq	 24(%rsp), %r13
	movq	 32(%rsp), %r14
	movq	 40(%rsp), %r15

#ifdef WINDOWS_ABI
	movq	 48(%rsp), %rdi
	movq	 56(%rsp), %rsi
	movups	 64(%rsp), %xmm6
	movups	 80(%rsp), %xmm7
	movups	 96(%rsp), %xmm8
	movups	112(%rsp), %xmm9
	movups	128(%rsp), %xmm10
	movups	144(%rsp), %xmm11
	movups	160(%rsp), %xmm12
	movups	176(%rsp), %xmm13
	movups	192(%rsp), %xmm14
	movups	208(%rsp), %xmm15
#endif

	addq	$STACKSIZE, %rsp
	ret

	EPILOGUE