📄 zgemm_kernel_4x2_sse3.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 3 页
字号:
上一页 1 23
	movddup	 4 * SIZE(B), %xmm2	movddup	 6 * SIZE(B), %xmm3	movddup	 8 * SIZE(B), %xmm4	movddup	10 * SIZE(B), %xmm5	movddup	12 * SIZE(B), %xmm6	movddup	14 * SIZE(B), %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)	prefetcht1	128 * SIZE(BO)	prefetcht0	112 * SIZE(B)	addq	$16 * SIZE, B	addq	$32 * SIZE, BO	decq	%rax	jne	.L42	ALIGN_4.L43:	movq	K, %rax	andq	$7, %rax	BRANCH	jle	.L50	ALIGN_4.L44:	movddup	 0 * SIZE(B), %xmm0	movaps	%xmm0,  0 * SIZE(BO)	addq	$2 * SIZE, B	addq	$4 * SIZE, BO	decq	%rax	jne	.L44	ALIGN_4	.L50:	movq	C, CO1			# coffset1 = c	movq	A, AO		# aoffset = a	movq	M,  I	sarq	$2, I		# i = (m >> 2)	jle	.L60	ALIGN_4.L51:#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	  0 * SIZE(AO), %xmm8	pxor	%xmm0, %xmm0	movaps	 16 * SIZE(AO), %xmm10	pxor	%xmm1, %xmm1	movaps	 32 * SIZE(AO), %xmm12	pxor	%xmm4, %xmm4	movaps	 48 * SIZE(AO), %xmm14	pxor	%xmm5, %xmm5	movsldup  0 * SIZE(BO), %xmm9	movsldup 16 * SIZE(BO), %xmm11	prefetchnta     4 * SIZE(CO1)#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	.L55	ALIGN_4.L52:	mulps	%xmm8, %xmm9	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)	addps	%xmm9, %xmm0	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	  4 * SIZE(AO), %xmm8	ADDSUB	%xmm9, %xmm1	movsldup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm4	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	  8 * SIZE(AO), %xmm8	ADDSUB	%xmm9, %xmm5	movsldup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movshdup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	 12 * SIZE(AO), %xmm8	ADDSUB	%xmm9, %xmm1	movsldup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm4	movshdup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	 64 * SIZE(AO), %xmm8	ADDSUB	%xmm9, %xmm5	movsldup  8 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm0	movshdup  8 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	movaps	 20 * SIZE(AO), %xmm10	ADDSUB	%xmm9, %xmm1	movsldup  8 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm4	movshdup  8 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	movaps	 24 * SIZE(AO), %xmm10	ADDSUB	%xmm9, %xmm5	movsldup 12 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm0	movshdup 12 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	movaps	 28 * SIZE(AO), %xmm10	ADDSUB	%xmm9, %xmm1	movsldup 12 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm4	movshdup 12 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	movaps	 80 * SIZE(AO), %xmm10	ADDSUB	%xmm9, %xmm5	movsldup 32 * SIZE(BO), %xmm9	mulps	%xmm12, %xmm11	PREFETCH  (PREFETCHSIZE + 32) * SIZE(AO)	addps	%xmm11, %xmm0	movshdup 16 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11	movaps	 36 * SIZE(AO), %xmm12	ADDSUB	%xmm11, %xmm1	movsldup 16 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11	addps	%xmm11, %xmm4	movshdup 16 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11	movaps	 40 * SIZE(AO), %xmm12	ADDSUB	%xmm11, %xmm5	movsldup 20 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11	addps	%xmm11, %xmm0	movshdup 20 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11	movaps	 44 * SIZE(AO), %xmm12	ADDSUB	%xmm11, %xmm1	movsldup 20 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11	addps	%xmm11, %xmm4	movshdup 20 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11	movaps	 96 * SIZE(AO), %xmm12	ADDSUB	%xmm11, %xmm5	movsldup 24 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	addps	%xmm11, %xmm0	movshdup 24 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	movaps	 52 * SIZE(AO), %xmm14	ADDSUB	%xmm11, %xmm1	movsldup 24 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	addps	%xmm11, %xmm4	movshdup 24 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	movaps	 56 * SIZE(AO), %xmm14	ADDSUB	%xmm11, %xmm5	movsldup 28 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	addps	%xmm11, %xmm0	movshdup 28 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	movaps	 60 * SIZE(AO), %xmm14	ADDSUB	%xmm11, %xmm1	movsldup 28 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	addps	%xmm11, %xmm4	movshdup 28 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	movaps	112 * SIZE(AO), %xmm14	ADDSUB	%xmm11, %xmm5	movsldup 48 * SIZE(BO), %xmm11	addq   $64 * SIZE, AO	addq   $32 * SIZE, BO	decq   %rax	jne    .L52	ALIGN_4.L55:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movaps	ALPHA_R, %xmm14	movaps	ALPHA_I, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je .L58	ALIGN_4.L56:	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	  4 * SIZE(AO), %xmm8	ADDSUB	%xmm9, %xmm1	movsldup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm4	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	  8 * SIZE(AO), %xmm8	ADDSUB	%xmm9, %xmm5	movsldup  4 * SIZE(BO), %xmm9	addq	$ 8 * SIZE, AO	addq	$ 4 * SIZE, BO	decq	%rax	jg	.L56	ALIGN_4.L58:#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \    defined(NR) || defined(NC) || defined(TR) || defined(TC)	shufps	$0xb1, %xmm1, %xmm1	shufps	$0xb1, %xmm5, %xmm5	addsubps	%xmm1, %xmm0	addsubps	%xmm5, %xmm4	movaps	%xmm0, %xmm1	movaps	%xmm4, %xmm5	shufps	$0xb1, %xmm0, %xmm0	shufps	$0xb1, %xmm4, %xmm4#else	shufps	$0xb1, %xmm0, %xmm0	shufps	$0xb1, %xmm4, %xmm4	addsubps	%xmm0, %xmm1	addsubps	%xmm4, %xmm5	movaps	%xmm1, %xmm0	movaps	%xmm5, %xmm4	shufps	$0xb1, %xmm1, %xmm1	shufps	$0xb1, %xmm5, %xmm5#endif	mulps	%xmm14, %xmm1	mulps	%xmm15, %xmm0	mulps	%xmm14, %xmm5	mulps	%xmm15, %xmm4	addps	%xmm1, %xmm0	addps	%xmm5, %xmm4#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	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, 2), BO#endif#if defined(TRMMKERNEL) && defined(LEFT)	addq	$4, KK#endif	addq	$8 * SIZE, CO1		# coffset += 4	decq	I			# i --	jg	.L51	ALIGN_4	.L60:	testq	$2, M	je	.L70#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	  0 * SIZE(AO), %xmm8	pxor	%xmm0, %xmm0	movsldup  0 * SIZE(BO), %xmm9	pxor	%xmm1, %xmm1	movaps	 16 * SIZE(AO), %xmm10	movsldup 16 * SIZE(BO), %xmm11#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	.L65	ALIGN_4.L62:	mulps	%xmm8, %xmm9	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)	addps	%xmm9, %xmm0	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	  4 * SIZE(AO), %xmm8	ADDSUB	%xmm9, %xmm1	movsldup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movshdup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	  8 * SIZE(AO), %xmm8	ADDSUB	%xmm9, %xmm1	movsldup  8 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movshdup  8 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	 12 * SIZE(AO), %xmm8	ADDSUB	%xmm9, %xmm1	movsldup 12 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movshdup 12 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	 32 * SIZE(AO), %xmm8	ADDSUB	%xmm9, %xmm1	movsldup 32 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm0	movshdup 16 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	movaps	 20 * SIZE(AO), %xmm10	ADDSUB	%xmm11, %xmm1	movsldup 20 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm0	movshdup 20 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	movaps	 24 * SIZE(AO), %xmm10	ADDSUB	%xmm11, %xmm1	movsldup 24 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm0	movshdup 24 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	movaps	 28 * SIZE(AO), %xmm10	ADDSUB	%xmm11, %xmm1	movsldup 28 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm0	movshdup 28 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	movaps	 48 * SIZE(AO), %xmm10	ADDSUB	%xmm11, %xmm1	movsldup 48 * SIZE(BO), %xmm11	addq   $32 * SIZE, AO	addq   $32 * SIZE, BO	decq   %rax	jne    .L62	ALIGN_4.L65:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movaps	ALPHA_R, %xmm14	movaps	ALPHA_I, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je .L68	ALIGN_4.L66:	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	  4 * SIZE(AO), %xmm8	ADDSUB	%xmm9, %xmm1	movsldup  4 * SIZE(BO), %xmm9	addq	$4 * SIZE, AO		# aoffset  += 4	addq	$4 * SIZE, BO		# boffset1 += 8	decq	%rax	jg	.L66	ALIGN_4.L68:#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \    defined(NR) || defined(NC) || defined(TR) || defined(TC)	shufps	$0xb1, %xmm1, %xmm1	addsubps	%xmm1, %xmm0	movaps	%xmm0, %xmm1	shufps	$0xb1, %xmm0, %xmm0#else	shufps	$0xb1, %xmm0, %xmm0	addsubps	%xmm0, %xmm1	movaps	%xmm1, %xmm0	shufps	$0xb1, %xmm1, %xmm1#endif	mulps	%xmm14, %xmm1	mulps	%xmm15, %xmm0	addps	%xmm1, %xmm0#if! defined(TRMMKERNEL) && !defined(BETAZERO)	movsd	0 * SIZE(CO1), %xmm8	movhps	2 * SIZE(CO1), %xmm8	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, 2), BO#endif#if defined(TRMMKERNEL) && defined(LEFT)	addq	$2, KK#endif	addq	$4 * SIZE, CO1		# coffset += 4	ALIGN_4		.L70:	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, 8), %rax	leaq	(AO, %rax, 1), AO	leaq	(BO, %rax, 2), BO#endif	 	movddup	  0 * SIZE(AO), %xmm8	pxor	%xmm0, %xmm0	movsd     0 * SIZE(BO), %xmm9	pxor	%xmm1, %xmm1 	movddup	  8 * SIZE(AO), %xmm10	movsd    16 * SIZE(BO), %xmm11#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	.L75	ALIGN_4.L72:	shufps	$0x50, %xmm9, %xmm9	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)	mulps	%xmm8, %xmm9 	movddup	  2 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movsd     4 * SIZE(BO), %xmm9	shufps	$0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9 	movddup	  4 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsd     8 * SIZE(BO), %xmm9	shufps	$0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9 	movddup	  6 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movsd    12 * SIZE(BO), %xmm9	shufps	$0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9 	movddup	 16 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsd    32 * SIZE(BO), %xmm9	shufps	$0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11 	movddup	 10 * SIZE(AO), %xmm10	addps	%xmm11, %xmm0	movsd    20 * SIZE(BO), %xmm11	shufps	$0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11 	movddup	 12 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movsd    24 * SIZE(BO), %xmm11	shufps	$0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11 	movddup	 14 * SIZE(AO), %xmm10	addps	%xmm11, %xmm0	movsd    28 * SIZE(BO), %xmm11	shufps	$0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11 	movddup	 24 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movsd    48 * SIZE(BO), %xmm11	addq   $16 * SIZE, AO	addq   $32 * SIZE, BO	decq   %rax	jne    .L72	ALIGN_4.L75:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movaps	ALPHA_R, %xmm14	movaps	ALPHA_I, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je .L78	ALIGN_4.L76:	shufps	$0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9 	movddup	  2 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movsd     4 * SIZE(BO), %xmm9	addq	$2 * SIZE, AO	addq	$4 * SIZE, BO	decq	%rax	jg	.L76	ALIGN_4.L78:	addps	%xmm1, %xmm0	movhlps	%xmm0, %xmm1#if  defined(NR) || defined(NC) || defined(TR) || defined(TC) || \     defined(RR) || defined(RC) || defined(CR) || defined(CC) 	cmpeqps	%xmm7, %xmm7	pslld	$31,   %xmm7	xorps	%xmm7, %xmm1#endif          #if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \    defined(NR) || defined(NC) || defined(TR) || defined(TC)	shufps	$0xb1, %xmm1, %xmm1	addsubps	%xmm1, %xmm0	movaps	%xmm0, %xmm1	shufps	$0xb1, %xmm0, %xmm0#else	shufps	$0xb1, %xmm0, %xmm0	addsubps	%xmm0, %xmm1	movaps	%xmm1, %xmm0	shufps	$0xb1, %xmm1, %xmm1#endif	mulps	%xmm14, %xmm1	mulps	%xmm15, %xmm0	addps	%xmm1, %xmm0#if! defined(TRMMKERNEL) && !defined(BETAZERO)	movsd	0 * SIZE(CO1), %xmm8	addps	%xmm8, %xmm0#endif	movsd	%xmm0, 0 * SIZE(CO1)	ALIGN_4	.L999:	movq	%rbx, %rsp	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
上一页 1 23
⌨️ 快捷键说明

复制代码 Ctrl + C
搜索代码 Ctrl + F
全屏模式 F11
切换主题 Ctrl + Shift + D
显示快捷键 ?
增大字号 Ctrl + =
减小字号 Ctrl + -