⭐ 欢迎来到虫虫下载站! | 📦 资源下载 📁 资源专辑 ℹ️ 关于我们
⭐ 虫虫下载站

📄 trsm_kernel_lt_4x2_sse2.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 3 页
字号:
	movhpd	%xmm2,   3 * SIZE(BB)	movsd	%xmm3,   4 * SIZE(BB)	movsd	%xmm3,   5 * SIZE(BB)	movhpd	%xmm3,   6 * SIZE(BB)	movhpd	%xmm3,   7 * SIZE(BB)	movsd	%xmm5,   8 * SIZE(BB)	movsd	%xmm5,   9 * SIZE(BB)	movhpd	%xmm5,  10 * SIZE(BB)	movhpd	%xmm5,  11 * SIZE(BB)	movsd	%xmm7,  12 * SIZE(BB)	movsd	%xmm7,  13 * SIZE(BB)	movhpd	%xmm7,  14 * SIZE(BB)	movhpd	%xmm7,  15 * SIZE(BB)#else	movapd	%xmm0,   0 * SIZE(AA)	movapd	%xmm1,   2 * SIZE(AA)	movapd	%xmm2,   4 * SIZE(AA)	movapd	%xmm3,   6 * SIZE(AA)#endif#ifdef LN	subl	$4 * SIZE, %esi#endif#if defined(LN) || defined(LT)	movsd	%xmm2,  0 * SIZE(%esi)	movsd	%xmm3,  1 * SIZE(%esi)	movsd	%xmm5,  2 * SIZE(%esi)	movsd	%xmm7,  3 * SIZE(%esi)	movhpd	%xmm2,  0 * SIZE(%esi, LDC)	movhpd	%xmm3,  1 * SIZE(%esi, LDC)	movhpd	%xmm5,  2 * SIZE(%esi, LDC)	movhpd	%xmm7,  3 * SIZE(%esi, LDC)#else	movsd	%xmm0,  0 * SIZE(%esi)	movhpd	%xmm0,  1 * SIZE(%esi)	movsd	%xmm1,  2 * SIZE(%esi)	movhpd	%xmm1,  3 * SIZE(%esi)	movsd	%xmm2,  0 * SIZE(%esi, LDC)	movhpd	%xmm2,  1 * SIZE(%esi, LDC)	movsd	%xmm3,  2 * SIZE(%esi, LDC)	movhpd	%xmm3,  3 * SIZE(%esi, LDC)#endif#ifndef LN	addl	$4 * SIZE, %esi#endif#if defined(LT) || defined(RN)	movl	K,  %eax	subl	KK, %eax	leal	(,%eax, SIZE), %eax	leal	(AA, %eax, 4), AA#ifdef LT	addl	$8 * SIZE, B#endif#endif#ifdef LN	subl	$4, KK	movl	BORIG, B#endif#ifdef LT	addl	$4, KK#endif#ifdef RT	movl	K, %eax	movl	BORIG, B	sall	$2 + BASE_SHIFT, %eax	addl	%eax, AORIG#endif	decl	%ebx			# i --	jg	.L10	ALIGN_2	.L30:	movl	M,  %ebx	testl	$2, %ebx	jle	.L50#ifdef LN       movl	K, %eax       sall	$1 + BASE_SHIFT, %eax       subl	%eax, AORIG#endif#if defined(LN) || defined(RT)	movl	KK, %eax	movl	AORIG, AA	leal	(, %eax, SIZE), %eax	leal	(AA, %eax, 2), AA#endif	leal	BUFFER, BB#if defined(LN) || defined(RT)	movl	KK, %eax	sall	$1 + BASE_SHIFT, %eax	leal	(BB, %eax, 2), BB#endif		movapd	 0 * SIZE(BB), %xmm2	pxor	%xmm4, %xmm4	movapd	 0 * SIZE(AA), %xmm0	pxor	%xmm5, %xmm5	movapd	 8 * SIZE(BB), %xmm3	pxor	%xmm6, %xmm6	movapd	 8 * SIZE(AA), %xmm1	pxor	%xmm7, %xmm7#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K, %eax	subl	KK, %eax#endif	sarl	$3, %eax	je	.L32.L31:		mulpd	%xmm0, %xmm2	mulpd	 2 * SIZE(BB), %xmm0	addpd	%xmm2, %xmm4	movapd	 4 * SIZE(BB), %xmm2	addpd	%xmm0, %xmm5	movapd	 2 * SIZE(AA), %xmm0	mulpd	%xmm0, %xmm2	mulpd	 6 * SIZE(BB), %xmm0	addpd	%xmm2, %xmm6	movapd	16 * SIZE(BB), %xmm2	addpd	%xmm0, %xmm7	movapd	 4 * SIZE(AA), %xmm0	mulpd	%xmm0, %xmm3	mulpd	10 * SIZE(BB), %xmm0	addpd	%xmm3, %xmm4	movapd	12 * SIZE(BB), %xmm3	addpd	%xmm0, %xmm5	movapd	 6 * SIZE(AA), %xmm0	mulpd	%xmm0, %xmm3	mulpd	14 * SIZE(BB), %xmm0	addpd	%xmm3, %xmm6	movapd	24 * SIZE(BB), %xmm3	addpd	%xmm0, %xmm7	movapd	16 * SIZE(AA), %xmm0	mulpd	%xmm1, %xmm2	mulpd	18 * SIZE(BB), %xmm1	addpd	%xmm2, %xmm4	movapd	20 * SIZE(BB), %xmm2	addpd	%xmm1, %xmm5	movapd	10 * SIZE(AA), %xmm1	mulpd	%xmm1, %xmm2	mulpd	22 * SIZE(BB), %xmm1	addpd	%xmm2, %xmm6	movapd	32 * SIZE(BB), %xmm2	addpd	%xmm1, %xmm7	movapd	12 * SIZE(AA), %xmm1	mulpd	%xmm1, %xmm3	mulpd	26 * SIZE(BB), %xmm1	addpd	%xmm3, %xmm4	movapd	28 * SIZE(BB), %xmm3	addpd	%xmm1, %xmm5	movapd	14 * SIZE(AA), %xmm1	mulpd	%xmm1, %xmm3	mulpd	30 * SIZE(BB), %xmm1	addpd	%xmm3, %xmm6	movapd	40 * SIZE(BB), %xmm3	addpd	%xmm1, %xmm7	movapd	24 * SIZE(AA), %xmm1	addl   $16 * SIZE, AA	addl   $32 * SIZE, BB	BRANCH	decl   %eax	jne    .L31.L32:#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K, %eax	subl	KK, %eax#endif	andl	$7, %eax		# if (k & 1)	BRANCH	je .L34.L33:	mulpd	%xmm0, %xmm2	mulpd	 2 * SIZE(BB), %xmm0	addpd	%xmm2, %xmm4	movapd	 4 * SIZE(BB), %xmm2	addpd	%xmm0, %xmm5	movapd	 2 * SIZE(AA), %xmm0	addl	$2 * SIZE, AA		# aoffset  += 8	addl	$4 * SIZE, BB		# boffset1 += 8	decl	%eax	BRANCH	jg	.L33	ALIGN_4.L34:	addpd	%xmm6, %xmm4	addpd	%xmm7, %xmm5#if defined(LN) || defined(RT)	movl	KK, %eax#ifdef LN	subl	$2, %eax#else	subl	$2, %eax#endif	movl	AORIG, AA	movl	BORIG, B	leal	BUFFER, BB	leal	(, %eax, SIZE), %eax	leal	(AA, %eax, 2), AA	leal	(B,  %eax, 2), B	leal	(BB, %eax, 4), BB#endif#if defined(LN) || defined(LT)	movapd	 %xmm4, %xmm0	unpcklpd %xmm5, %xmm4	unpckhpd %xmm5, %xmm0	movapd	 0 * SIZE(B), %xmm2	movapd	 2 * SIZE(B), %xmm3	subpd	%xmm4,  %xmm2	subpd	%xmm0,  %xmm3#else	movapd	 0 * SIZE(AA), %xmm0	movapd	 2 * SIZE(AA), %xmm1	subpd	%xmm4, %xmm0	subpd	%xmm5, %xmm1#endif#ifdef LN	movsd	 3 * SIZE(AA), %xmm0	movhpd	 3 * SIZE(AA), %xmm0	mulpd	 %xmm0, %xmm3	movsd	 2 * SIZE(AA), %xmm0	movhpd	 2 * SIZE(AA), %xmm0	mulpd	 %xmm3, %xmm0	subpd	 %xmm0, %xmm2	movsd	 0 * SIZE(AA), %xmm0	movhpd	 0 * SIZE(AA), %xmm0	mulpd	 %xmm0, %xmm2#endif#ifdef LT	movsd	 0 * SIZE(AA), %xmm0	movhpd	 0 * SIZE(AA), %xmm0	mulpd	 %xmm0, %xmm2	movsd	 1 * SIZE(AA), %xmm0	movhpd	 1 * SIZE(AA), %xmm0	mulpd	 %xmm2, %xmm0	subpd	 %xmm0, %xmm3	movsd	 3 * SIZE(AA), %xmm0	movhpd	 3 * SIZE(AA), %xmm0	mulpd	 %xmm0, %xmm3#endif#ifdef RN	movsd	 0 * SIZE(B), %xmm4	movhpd	 0 * SIZE(B), %xmm4	mulpd	 %xmm4, %xmm0	movsd	 1 * SIZE(B), %xmm4	movhpd	 1 * SIZE(B), %xmm4	mulpd	 %xmm0, %xmm4	subpd	 %xmm4, %xmm1	movsd	 3 * SIZE(B), %xmm4	movhpd	 3 * SIZE(B), %xmm4	mulpd	 %xmm4, %xmm1#endif#ifdef RT	movsd	 3 * SIZE(B), %xmm4	movhpd	 3 * SIZE(B), %xmm4	mulpd	 %xmm4, %xmm1	movsd	 2 * SIZE(B), %xmm4	movhpd	 2 * SIZE(B), %xmm4	mulpd	 %xmm1, %xmm4	subpd	 %xmm4, %xmm0	movsd	 0 * SIZE(B), %xmm4	movhpd	 0 * SIZE(B), %xmm4	mulpd	 %xmm4, %xmm0#endif#if defined(LN) || defined(LT)	movapd	%xmm2,   0 * SIZE(B)	movapd	%xmm3,   2 * SIZE(B)	movsd	%xmm2,   0 * SIZE(BB)	movsd	%xmm2,   1 * SIZE(BB)	movhpd	%xmm2,   2 * SIZE(BB)	movhpd	%xmm2,   3 * SIZE(BB)	movsd	%xmm3,   4 * SIZE(BB)	movsd	%xmm3,   5 * SIZE(BB)	movhpd	%xmm3,   6 * SIZE(BB)	movhpd	%xmm3,   7 * SIZE(BB)#else	movapd	%xmm0,   0 * SIZE(AA)	movapd	%xmm1,   2 * SIZE(AA)#endif#ifdef LN	subl	$2 * SIZE, %esi#endif#if defined(LN) || defined(LT)	movsd	%xmm2,  0 * SIZE(%esi)	movsd	%xmm3,  1 * SIZE(%esi)	movhpd	%xmm2,  0 * SIZE(%esi, LDC)	movhpd	%xmm3,  1 * SIZE(%esi, LDC)#else	movsd	%xmm0,  0 * SIZE(%esi)	movhpd	%xmm0,  1 * SIZE(%esi)	movsd	%xmm1,  0 * SIZE(%esi, LDC)	movhpd	%xmm1,  1 * SIZE(%esi, LDC)#endif#ifndef LN	addl	$2 * SIZE, %esi#endif#if defined(LT) || defined(RN)	movl	K,  %eax	subl	KK, %eax	leal	(,%eax, SIZE), %eax	leal	(AA, %eax, 2), AA#ifdef LT	addl	$4 * SIZE, B#endif#endif#ifdef LN	subl	$2, KK	movl	BORIG, B#endif#ifdef LT	addl	$2, KK#endif#ifdef RT	movl	K, %eax	movl	BORIG, B	sall	$1 + BASE_SHIFT, %eax	addl	%eax, AORIG#endif	ALIGN_2.L50:	movl	M,  %ebx	testl	$1, %ebx	jle	.L99#ifdef LN       movl	K, %eax       sall	$0 + BASE_SHIFT, %eax       subl	%eax, AORIG#endif#if defined(LN) || defined(RT)	movl	KK, %eax	movl	AORIG, AA	leal	(, %eax, SIZE), %eax	leal	(AA, %eax, 1), AA#endif	leal	BUFFER, %ecx#if defined(LN) || defined(RT)	movl	KK, %eax	sall	$1 + BASE_SHIFT, %eax	leal	(BB, %eax, 2), BB#endif		movsd	 0 * SIZE(BB), %xmm2	pxor	%xmm4, %xmm4	movsd	 0 * SIZE(AA), %xmm0	pxor	%xmm5, %xmm5	movsd	 8 * SIZE(BB), %xmm3	pxor	%xmm6, %xmm6	movsd	 4 * SIZE(AA), %xmm1	pxor	%xmm7, %xmm7#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K, %eax	subl	KK, %eax#endif	sarl	$3, %eax	je	.L52.L51:		mulsd	%xmm0, %xmm2	mulsd	 2 * SIZE(BB), %xmm0	addsd	%xmm2, %xmm4	movsd	 4 * SIZE(BB), %xmm2	addsd	%xmm0, %xmm5	movsd	 1 * SIZE(AA), %xmm0	mulsd	%xmm0, %xmm2	mulsd	 6 * SIZE(BB), %xmm0	addsd	%xmm2, %xmm4	movsd	16 * SIZE(BB), %xmm2	addsd	%xmm0, %xmm5	movsd	 2 * SIZE(AA), %xmm0	mulsd	%xmm0, %xmm3	mulsd	10 * SIZE(BB), %xmm0	addsd	%xmm3, %xmm4	movsd	12 * SIZE(BB), %xmm3	addsd	%xmm0, %xmm5	movsd	 3 * SIZE(AA), %xmm0	mulsd	%xmm0, %xmm3	mulsd	14 * SIZE(BB), %xmm0	addsd	%xmm3, %xmm4	movsd	24 * SIZE(BB), %xmm3	addsd	%xmm0, %xmm5	movsd	 8 * SIZE(AA), %xmm0	mulsd	%xmm1, %xmm2	mulsd	18 * SIZE(BB), %xmm1	addsd	%xmm2, %xmm4	movsd	20 * SIZE(BB), %xmm2	addsd	%xmm1, %xmm5	movsd	 5 * SIZE(AA), %xmm1	mulsd	%xmm1, %xmm2	mulsd	22 * SIZE(BB), %xmm1	addsd	%xmm2, %xmm4	movsd	32 * SIZE(BB), %xmm2	addsd	%xmm1, %xmm5	movsd	 6 * SIZE(AA), %xmm1	mulsd	%xmm1, %xmm3	mulsd	26 * SIZE(BB), %xmm1	addsd	%xmm3, %xmm4	movsd	28 * SIZE(BB), %xmm3	addsd	%xmm1, %xmm5	movsd	 7 * SIZE(AA), %xmm1	mulsd	%xmm1, %xmm3	mulsd	30 * SIZE(BB), %xmm1	addsd	%xmm3, %xmm4	movsd	40 * SIZE(BB), %xmm3	addsd	%xmm1, %xmm5	movsd	12 * SIZE(AA), %xmm1	addl   $ 8 * SIZE, AA	addl   $32 * SIZE, BB	BRANCH	decl   %eax	jne    .L51.L52:#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K, %eax	subl	KK, %eax#endif	andl	$7, %eax		# if (k & 1)	BRANCH	je .L54.L53:	mulsd	%xmm0, %xmm2	mulsd	 2 * SIZE(BB), %xmm0	addsd	%xmm2, %xmm4	movsd	 4 * SIZE(BB), %xmm2	addsd	%xmm0, %xmm5	movsd	 1 * SIZE(AA), %xmm0	addl	$1 * SIZE, AA		# aoffset  += 8	addl	$4 * SIZE, BB		# boffset1 += 8	decl	%eax	BRANCH	jg	.L53	ALIGN_4.L54:	addsd	%xmm6, %xmm4	addsd	%xmm7, %xmm5#if defined(LN) || defined(RT)	movl	KK, %eax#ifdef LN	subl	$1, %eax#else	subl	$2, %eax#endif	movl	AORIG, AA	movl	BORIG, B	leal	BUFFER, BB	leal	(, %eax, SIZE), %eax	leal	(AA, %eax, 1), AA	leal	(B,  %eax, 2), B	leal	(BB, %eax, 4), BB#endif#if defined(LN) || defined(LT)	movsd	 0 * SIZE(B), %xmm0	movsd	 1 * SIZE(B), %xmm1#else	movsd	 0 * SIZE(AA), %xmm0	movsd	 1 * SIZE(AA), %xmm1#endif	subsd	%xmm4,  %xmm0	subsd	%xmm5,  %xmm1#if defined(LN) || defined(LT)	movsd	 0 * SIZE(AA), %xmm2	mulsd	 %xmm2, %xmm0	mulsd	 %xmm2, %xmm1#endif#ifdef RN	mulsd	 0 * SIZE(B), %xmm0	movsd	 1 * SIZE(B), %xmm4	mulsd	 %xmm0, %xmm4	subsd	 %xmm4, %xmm1	mulsd	 3 * SIZE(B), %xmm1#endif#ifdef RT	mulsd	 3 * SIZE(B), %xmm1	movsd	 2 * SIZE(B), %xmm4	mulsd	 %xmm1, %xmm4	subsd	 %xmm4, %xmm0	mulsd	 0 * SIZE(B), %xmm0#endif#if defined(LN) || defined(LT)	movsd	%xmm0,   0 * SIZE(B)	movsd	%xmm1,   1 * SIZE(B)	movsd	%xmm0,   0 * SIZE(BB)	movsd	%xmm0,   1 * SIZE(BB)	movsd	%xmm1,   2 * SIZE(BB)	movsd	%xmm1,   3 * SIZE(BB)#else	movsd	%xmm0,   0 * SIZE(AA)	movsd	%xmm1,   1 * SIZE(AA)#endif#ifdef LN	subl	$1 * SIZE, %esi#endif	movsd	%xmm0,  0 * SIZE(%esi)	movsd	%xmm1,  0 * SIZE(%esi, LDC)#ifndef LN	addl	$1 * SIZE, %esi#endif#if defined(LT) || defined(RN)	movl	K,  %eax	subl	KK, %eax	leal	(,%eax, SIZE), %eax	leal	(AA, %eax, 1), AA#ifdef LT	addl	$2 * SIZE, B#endif#endif#ifdef LN	subl	$1, KK	movl	BORIG, B#endif#ifdef LT	addl	$1, KK#endif#ifdef RT	movl	K, %eax	movl	BORIG, B	sall	$0 + BASE_SHIFT, %eax	addl	%eax, AORIG#endif	ALIGN_2.L99:#ifdef LN       movl	K, %eax       leal	(, %eax, SIZE), %eax       leal 	(B, %eax, 2), B#endif#if defined(LT) || defined(RN)	movl	K,  %eax	subl	KK, %eax	leal	(,%eax, SIZE), %eax	leal	(B,  %eax, 2), B#endif#ifdef RN	addl	$2, KK#endif#ifdef RT	subl	$2, KK#endif	decl	J			# j --	jg	.L01	ALIGN_2.L100:	movl	N, %eax	testl	$1, %eax	jle	.L999	ALIGN_2	.L101:/* Copying to Sub Buffer */#ifdef LN	movl	OFFSET, %eax	addl	M, %eax	movl	%eax, KK#endif		leal	BUFFER, %ecx#ifdef RT       movl	K, %eax       sall	$0 + BASE_SHIFT, %eax       subl	%eax, B#endif#if defined(LN) || defined(RT)	movl	KK, %eax	movl	B, BORIG	leal	(, %eax, SIZE), %eax	leal	(B,  %eax, 1), B	leal	(BB, %eax, 2), BB#endif	#ifdef LT	movl	OFFSET, %eax	movl	%eax, KK#endif#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K, %eax	subl	KK, %eax#endif	sarl	$3, %eax	jle	.L103	ALIGN_4	.L102:	movsd	 0 * SIZE(B), %xmm0	movsd	 1 * SIZE(B), %xmm1	movsd	 2 * SIZE(B), %xmm2	movsd	 3 * SIZE(B), %xmm3	movsd	 4 * SIZE(B), %xmm4	movsd	 5 * SIZE(B), %xmm5	movsd	 6 * SIZE(B), %xmm6	movsd	 7 * SIZE(B), %xmm7	unpcklpd  %xmm0, %xmm0	unpcklpd  %xmm1, %xmm1	unpcklpd  %xmm2, %xmm2	unpcklpd  %xmm3, %xmm3	unpcklpd  %xmm4, %xmm4	unpcklpd  %xmm5, %xmm5	unpcklpd  %xmm6, %xmm6	unpcklpd  %xmm7, %xmm7	movapd	%xmm0,  0 * SIZE(%ecx)	movapd	%xmm1,  2 * SIZE(%ecx)	movapd	%xmm2,  4 * SIZE(%ecx)	movapd	%xmm3,  6 * SIZE(%ecx)	movapd	%xmm4,  8 * SIZE(%ecx)	movapd	%xmm5, 10 * SIZE(%ecx)	movapd	%xmm6, 12 * SIZE(%ecx)	movapd	%xmm7, 14 * SIZE(%ecx)	prefetcht0	104 * SIZE(B)	addl	$ 8 * SIZE, B	addl	$16 * SIZE, %ecx	decl	%eax	BRANCH	jne	.L102	ALIGN_2.L103:#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K, %eax	subl	KK, %eax#endif	andl	$7, %eax	BRANCH	jle	.L105	ALIGN_2.L104:	movsd	 0 * SIZE(B), %xmm0	unpcklpd  %xmm0, %xmm0	movapd	%xmm0,  0 * SIZE(%ecx)	addl	$1 * SIZE, B	addl	$2 * SIZE, %ecx	decl	%eax	jne	.L104	ALIGN_4	.L105:#if defined(LT) || defined(RN)	movl	A, AA#else	movl	A, %eax	movl	%eax, AORIG#endif#ifdef RT	subl	LDC, C#endif	movl	C, %esi		# coffset = c#ifndef RT	addl	LDC, C#endif	movl	M,  %ebx	sarl	$2, %ebx	# i = (m >> 2)

⌨️ 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 + -