📄 zgemm_kernel_4x2_penryn.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 3 页
字号:
上一页 1 23
	prefetcht0     7 * SIZE(CO1)	pxor	%xmm8,  %xmm8	pxor	%xmm9,  %xmm9	pxor	%xmm12, %xmm12	pxor	%xmm13, %xmm13#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	$2, %rax	jle,pn	.L45	ALIGN_3.L42:		PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)	addps	%xmm3, %xmm8	pshufd	$0x00, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm12	pshufd	$0x00, %xmm2, %xmm4	mulps	%xmm1, %xmm4	addps	%xmm5, %xmm9	pshufd	$0x55, %xmm2, %xmm5	mulps	%xmm0, %xmm5	movaps	-24 * SIZE(AO), %xmm0	addps	%xmm6, %xmm13	pshufd	$0x55, %xmm2, %xmm6	mulps	%xmm1, %xmm6	movaps	-20 * SIZE(AO), %xmm1	addps	%xmm3, %xmm8	pshufd	$0xaa, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm12	pshufd	$0xaa, %xmm2, %xmm4	mulps	%xmm1, %xmm4	addps	%xmm5, %xmm9	pshufd	$0xff, %xmm2, %xmm5	mulps	%xmm0, %xmm5	movaps	-16 * SIZE(AO), %xmm0	addps	%xmm6, %xmm13	pshufd	$0xff, %xmm2, %xmm6	movaps	-28 * SIZE(BO), %xmm2	mulps	%xmm1, %xmm6	movaps	-12 * SIZE(AO), %xmm1	PREFETCH (PREFETCHSIZE + 16) * SIZE(AO)	addps	%xmm3, %xmm8	pshufd	$0x00, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm12	pshufd	$0x00, %xmm2, %xmm4	mulps	%xmm1, %xmm4	addps	%xmm5, %xmm9	pshufd	$0x55, %xmm2, %xmm5	mulps	%xmm0, %xmm5	movaps	 -8 * SIZE(AO), %xmm0	addps	%xmm6, %xmm13	pshufd	$0x55, %xmm2, %xmm6	mulps	%xmm1, %xmm6	movaps	 -4 * SIZE(AO), %xmm1	addps	%xmm3, %xmm8	pshufd	$0xaa, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm12	pshufd	$0xaa, %xmm2, %xmm4	mulps	%xmm1, %xmm4	addps	%xmm5, %xmm9	pshufd	$0xff, %xmm2, %xmm5	mulps	%xmm0, %xmm5	movaps	  0 * SIZE(AO), %xmm0	addps	%xmm6, %xmm13	pshufd	$0xff, %xmm2, %xmm6	movaps	-24 * SIZE(BO), %xmm2	mulps	%xmm1, %xmm6	movaps	  4 * SIZE(AO), %xmm1	subq	$-32 * SIZE, AO	subq	$ -8 * SIZE, BO	subq	$1, %rax	jg,pt	.L42	ALIGN_3.L45:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	andq	$3, %rax		# if (k & 1)	je,pt	.L48	ALIGN_3.L46:	addps	%xmm3, %xmm8	pshufd	$0x00, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm12	pshufd	$0x00, %xmm2, %xmm4	mulps	%xmm1, %xmm4	addps	%xmm5, %xmm9	pshufd	$0x55, %xmm2, %xmm5	mulps	%xmm0, %xmm5	movaps	-24 * SIZE(AO), %xmm0	addps	%xmm6, %xmm13	pshufd	$0x55, %xmm2, %xmm6	movsd	-30 * SIZE(BO), %xmm2	mulps	%xmm1, %xmm6	movaps	-20 * SIZE(AO), %xmm1	addq	$8 * SIZE, AO	addq	$2 * SIZE, BO	subq	$1, %rax	jg,pt	.L46	ALIGN_3.L48:	addps	%xmm3, %xmm8	addps	%xmm4, %xmm12	addps	%xmm5, %xmm9	addps	%xmm6, %xmm13	pshufd	 $0xb1, %xmm9,  %xmm9	movddup	ALPHA_R, %xmm2	pshufd	 $0xb1, %xmm13, %xmm13	movddup	ALPHA_I, %xmm3	pcmpeqb	%xmm0, %xmm0	psllq	$63,   %xmm0#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)	pxor	%xmm0, %xmm9	pxor	%xmm0, %xmm13	subps	%xmm9,  %xmm8	subps	%xmm13, %xmm12#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)	pxor	%xmm0, %xmm9	pxor	%xmm0, %xmm13	addps	%xmm9,  %xmm8	addps	%xmm13, %xmm12#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)	pxor	%xmm0, %xmm8	pxor	%xmm0, %xmm12	addps	%xmm9,  %xmm8	addps	%xmm13, %xmm12#else	pxor	%xmm0, %xmm8	pxor	%xmm0, %xmm12	subps	%xmm9,  %xmm8	subps	%xmm13, %xmm12#endif	pshufd	$0xb1, %xmm8,  %xmm9	pshufd	$0xb1, %xmm12, %xmm13	mulps	%xmm2, %xmm8	mulps	%xmm3, %xmm9	mulps	%xmm2, %xmm12	mulps	%xmm3, %xmm13	addsubps       %xmm9,  %xmm8	addsubps       %xmm13, %xmm12#ifndef TRMMKERNEL	movsd	0 * SIZE(CO1), %xmm0	movhps	2 * SIZE(CO1), %xmm0	movsd	4 * SIZE(CO1), %xmm1	movhps	6 * SIZE(CO1), %xmm1	addps	%xmm0, %xmm8	addps	%xmm1, %xmm12#endif	movsd	%xmm8,  0 * SIZE(CO1)	movhps	%xmm8,  2 * SIZE(CO1)	movsd	%xmm12, 4 * SIZE(CO1)	movhps	%xmm12, 6 * SIZE(CO1)#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	movq	K, %rax	subq	KKK, %rax	salq	$ZBASE_SHIFT, %rax	leaq	(AO, %rax, 4), AO	leaq	(BO, %rax, 1), BO#endif#if defined(TRMMKERNEL) && defined(LEFT)	addq	$4, KK#endif	addq	$8 * SIZE, CO1		# coffset += 4	decq	I			# i --	jg,pt	.L41	ALIGN_4	.L50:	testq	$2, M	jle,pt	.L60#if !defined(TRMMKERNEL) || \	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	movq	B, BO#else	movq	B, BO	movq	KK, %rax	salq	$ZBASE_SHIFT, %rax	leaq	(AO, %rax, 2), AO	leaq	(BO, %rax, 1), BO#endif		movaps	-32 * SIZE(AO), %xmm0	pxor	%xmm3, %xmm3	pxor	%xmm4, %xmm4	movaps	-32 * SIZE(BO), %xmm2	pxor	%xmm8, %xmm8	pxor	%xmm9, %xmm9#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	$2, %rax	jle,pn	.L55	ALIGN_3.L52:		PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)	addps	%xmm3, %xmm8	pshufd	$0x00, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm9	pshufd	$0x55, %xmm2, %xmm4	mulps	%xmm0, %xmm4	movaps	-28 * SIZE(AO), %xmm0	addps	%xmm3, %xmm8	pshufd	$0xaa, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm9	pshufd	$0xff, %xmm2, %xmm4	movaps	-28 * SIZE(BO), %xmm2	mulps	%xmm0, %xmm4	movaps	-24 * SIZE(AO), %xmm0	addps	%xmm3, %xmm8	pshufd	$0x00, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm9	pshufd	$0x55, %xmm2, %xmm4	mulps	%xmm0, %xmm4	movaps	-20 * SIZE(AO), %xmm0	addps	%xmm3, %xmm8	pshufd	$0xaa, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm9	pshufd	$0xff, %xmm2, %xmm4	movaps	-24 * SIZE(BO), %xmm2	mulps	%xmm0, %xmm4	movaps	-16 * SIZE(AO), %xmm0	subq	$-16 * SIZE, AO	subq	$ -8 * SIZE, BO	subq	$1, %rax	jg,pt	.L52	ALIGN_3.L55:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	andq	$3, %rax		# if (k & 1)	je,pt	.L58	ALIGN_3.L56:	addps	%xmm3, %xmm8	pshufd	$0x00, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm9	pshufd	$0x55, %xmm2, %xmm4	movsd	-30 * SIZE(BO), %xmm2	mulps	%xmm0, %xmm4	movaps	-28 * SIZE(AO), %xmm0	addq	$4 * SIZE, AO	addq	$2 * SIZE, BO	subq	$1, %rax	jg,pt	.L56	ALIGN_3.L58:	addps	%xmm3, %xmm8	movddup	ALPHA_R, %xmm2	addps	%xmm4, %xmm9	movddup	ALPHA_I, %xmm3	pshufd	 $0xb1, %xmm9,  %xmm9	pcmpeqb	%xmm0, %xmm0	psllq	$63,   %xmm0#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)	pxor	%xmm0, %xmm9	subps	%xmm9,  %xmm8#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)	pxor	%xmm0, %xmm9	addps	%xmm9,  %xmm8#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)	pxor	%xmm0, %xmm8	addps	%xmm9,  %xmm8#else	pxor	%xmm0, %xmm8	subps	%xmm9,  %xmm8#endif	pshufd	$0xb1, %xmm8,  %xmm9	mulps	%xmm2, %xmm8	mulps	%xmm3, %xmm9	addsubps       %xmm9,  %xmm8#ifndef TRMMKERNEL	movsd	0 * SIZE(CO1), %xmm0	movhps	2 * SIZE(CO1), %xmm0	addps	%xmm0, %xmm8#endif	movsd	%xmm8,  0 * SIZE(CO1)	movhps	%xmm8,  2 * SIZE(CO1)#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	movq	K, %rax	subq	KKK, %rax	salq	$ZBASE_SHIFT, %rax	leaq	(AO, %rax, 2), AO	leaq	(BO, %rax, 1), BO#endif#if defined(TRMMKERNEL) && defined(LEFT)	addq	$2, KK#endif	addq	$4 * SIZE, CO1	ALIGN_4	.L60:	testq	$1, M	jle,pt	.L999#if !defined(TRMMKERNEL) || \	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	movq	B, BO#else	movq	B, BO	movq	KK, %rax	salq	$ZBASE_SHIFT, %rax	leaq	(AO, %rax, 1), AO	leaq	(BO, %rax, 1), BO#endif		movsd	-32 * SIZE(AO), %xmm0	pxor	%xmm3, %xmm3	pxor	%xmm4, %xmm4	movaps	-32 * SIZE(BO), %xmm2	pxor	%xmm8, %xmm8	pxor	%xmm9, %xmm9#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	$2, %rax	jle,pn	.L65	ALIGN_3.L62:		PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)	addps	%xmm3, %xmm8	pshufd	$0x00, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm9	pshufd	$0x55, %xmm2, %xmm4	mulps	%xmm0, %xmm4	movsd	-30 * SIZE(AO), %xmm0	addps	%xmm3, %xmm8	pshufd	$0xaa, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm9	pshufd	$0xff, %xmm2, %xmm4	movaps	-28 * SIZE(BO), %xmm2	mulps	%xmm0, %xmm4	movsd	-28 * SIZE(AO), %xmm0	addps	%xmm3, %xmm8	pshufd	$0x00, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm9	pshufd	$0x55, %xmm2, %xmm4	mulps	%xmm0, %xmm4	movsd	-26 * SIZE(AO), %xmm0	addps	%xmm3, %xmm8	pshufd	$0xaa, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm9	pshufd	$0xff, %xmm2, %xmm4	movaps	-24 * SIZE(BO), %xmm2	mulps	%xmm0, %xmm4	movsd	-24 * SIZE(AO), %xmm0	subq	$-8 * SIZE, AO	subq	$-8 * SIZE, BO	subq	$1, %rax	jg,pt	.L62	ALIGN_3.L65:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	andq	$3, %rax		# if (k & 1)	je,pt	.L68	ALIGN_3.L66:	addps	%xmm3, %xmm8	pshufd	$0x00, %xmm2, %xmm3	mulps	%xmm0, %xmm3	addps	%xmm4, %xmm9	pshufd	$0x55, %xmm2, %xmm4	movsd	-30 * SIZE(BO), %xmm2	mulps	%xmm0, %xmm4	movsd	-30 * SIZE(AO), %xmm0	addq	$2 * SIZE, AO	addq	$2 * SIZE, BO	subq	$1, %rax	jg,pt	.L66	ALIGN_3.L68:	addps	%xmm3, %xmm8	movddup	ALPHA_R, %xmm2	addps	%xmm4, %xmm9	movddup	ALPHA_I, %xmm3	pshufd	 $0xb1, %xmm9,  %xmm9	pcmpeqb	%xmm0, %xmm0	psllq	$63,   %xmm0#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)	pxor	%xmm0, %xmm9	subps	%xmm9,  %xmm8#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)	pxor	%xmm0, %xmm9	addps	%xmm9,  %xmm8#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)	pxor	%xmm0, %xmm8	addps	%xmm9,  %xmm8#else	pxor	%xmm0, %xmm8	subps	%xmm9,  %xmm8#endif	pshufd	$0xb1, %xmm8,  %xmm9	mulps	%xmm2, %xmm8	mulps	%xmm3, %xmm9	addsubps       %xmm9,  %xmm8#ifndef TRMMKERNEL	movsd	0 * SIZE(CO1), %xmm0	addps	%xmm0, %xmm8#endif	movsd	%xmm8,  0 * SIZE(CO1)	ALIGN_4		.L999:	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 + -