⭐ 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_1x4.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 2 页
字号:
	movl	B, B_ORIG#endif#ifdef RN	addl	$4, KK#endif#ifdef RT	subl	$4, KK#endif	decl	J	jne	.L11	ALIGN_4.L20:	movl	N,   %eax	andl	$2,  %eax	je	.L30#if defined(LT) || defined(RN)	movl	STACK_A, A#else	movl	STACK_A, %eax	movl	%eax, AORIG#endif#ifdef RT	movl	K, %eax	sall	$1 + BASE_SHIFT, %eax	subl	%eax, B_ORIG#endif	leal	(, LDC, 2), %eax#ifdef RT	subl	%eax, C#endif	movl	C, %edi#ifndef RT	addl	%eax, C#endif#ifdef LN	movl	OFFSET, %eax	addl	M, %eax	movl	%eax, KK#endif	#ifdef LT	movl	OFFSET, %eax	movl	%eax, KK#endif	movl	B_ORIG, B#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K,  %eax	subl	KK, %eax#endif	sarl	$4, %eax	jle	.L23	ALIGN_4.L22:		movl	 -16 * SIZE(B), %esi	movl	  -8 * SIZE(B), %esi	movl	   0 * SIZE(B), %esi	movl	   8 * SIZE(B), %esi	subl	 $-32 * SIZE, B	decl	  %eax	jne	  .L22	ALIGN_3.L23:	movl	M, %esi	movl	%esi, I	ALIGN_3.L24:#ifdef LN       movl	K, %eax       sall	$BASE_SHIFT, %eax       subl	%eax, AORIG#endif#if defined(LN) || defined(RT)	movl	KK, %eax	leal	(, %eax, SIZE), %eax	movl	AORIG, A	leal	(A     , %eax, 1), A	leal	(B_ORIG, %eax, 2), B#else	movl	B_ORIG, B#endif		fldz	fldz	fldz	fldz	FLD	 -16 * SIZE(A)	FLD	 -16 * SIZE(B)	prefetchw	1 * SIZE(%edi)	prefetchw	1 * SIZE(%edi, LDC)#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K,  %eax	subl	KK, %eax#endif	sarl	$3, %eax 	je	.L26	ALIGN_3.L25:	fmul	%st(1), %st	faddp	%st, %st(2)	FMUL	-15 * SIZE(B)	faddp	%st, %st(2)	FLD	-15 * SIZE(A)	FLD	-14 * SIZE(B)	fmul	%st(1), %st	faddp	%st, %st(4)	FMUL	-13 * SIZE(B)	faddp	%st, %st(4)	FLD	-14 * SIZE(A)	FLD	-12 * SIZE(B)	fmul	%st(1), %st	faddp	%st, %st(2)	FMUL	-11 * SIZE(B)	faddp	%st, %st(2)	FLD	-13 * SIZE(A)	FLD	-10 * SIZE(B)	fmul	%st(1), %st	faddp	%st, %st(4)	FMUL	 -9 * SIZE(B)	faddp	%st, %st(4)	FLD	-12 * SIZE(A)	FLD	 -8 * SIZE(B)	fmul	%st(1), %st	faddp	%st, %st(2)	FMUL	 -7 * SIZE(B)	faddp	%st, %st(2)	FLD	-11 * SIZE(A)	FLD	 -6 * SIZE(B)	fmul	%st(1), %st	faddp	%st, %st(4)	FMUL	 -5 * SIZE(B)	faddp	%st, %st(4)	FLD	-10 * SIZE(A)	FLD	 -4 * SIZE(B)	fmul	%st(1), %st	faddp	%st, %st(2)	FMUL	 -3 * SIZE(B)	faddp	%st, %st(2)	FLD	 -9 * SIZE(A)	FLD	 -2 * SIZE(B)	fmul	%st(1), %st	faddp	%st, %st(4)	FMUL	 -1 * SIZE(B)	faddp	%st, %st(4)	FLD	 -8 * SIZE(A)	FLD	  0 * SIZE(B)	addl	$  8 * SIZE, A	subl	$-16 * SIZE, B	decl	%eax	jne	.L25	ALIGN_4.L26:#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K,  %eax	subl	KK, %eax#endif	and	$7, %eax	je	.L29	ALIGN_4.L27:	fmul	%st(1), %st	faddp	%st, %st(2)	FMUL	-15 * SIZE(B)	faddp	%st, %st(2)	FLD	 -15 * SIZE(A)	FLD	 -14 * SIZE(B)	addl	$1 * SIZE,A	addl	$2 * SIZE,B	decl	%eax	jne	 .L27	ALIGN_4.L29:	ffreep	%st	ffreep	%st	faddp	%st, %st(2)	faddp	%st, %st(2)#if defined(LN) || defined(RT)	movl	KK, %eax#ifdef LN	subl	$1, %eax#else	subl	$2, %eax#endif	leal	(, %eax, SIZE), %eax	movl	AORIG, A	leal	(A,      %eax, 1), A	leal	(B_ORIG, %eax, 2), B#endif#if defined(LN) || defined(LT)	FLD	0 * SIZE - 16 * SIZE(B)	fsubp	%st, %st(1)	FLD	1 * SIZE - 16 * SIZE(B)	fsubp	%st, %st(2)#else	FLD	0 * SIZE - 16 * SIZE(A)	fsubp	%st, %st(1)	FLD	1 * SIZE - 16 * SIZE(A)	fsubp	%st, %st(2)#endif#ifdef LN       FLD	0 * SIZE - 16 * SIZE(A)       fmul	%st, %st(1)       fmulp	%st, %st(2)#endif#ifdef LT       FLD	0 * SIZE - 16 * SIZE(A)       fmul	%st, %st(1)       fmulp	%st, %st(2)#endif#ifdef RN       FMUL	 0 * SIZE - 16 * SIZE(B)       FLD	 1 * SIZE - 16 * SIZE(B)       fmul	 %st(1), %st       fsubrp	 %st, %st(2)       FLD	 3 * SIZE - 16 * SIZE(B)       fmulp	 %st, %st(2)#endif#ifdef RT       FLD	 3 * SIZE - 16 * SIZE(B)       fmulp	 %st, %st(2)       FLD	 2 * SIZE - 16 * SIZE(B)       fmul	 %st(2), %st       fsubrp	 %st, %st(1)       FLD	 0 * SIZE - 16 * SIZE(B)       fmulp	 %st, %st(1)#endif#ifdef LN	subl	$1 * SIZE, %edi#endif#if defined(LN) || defined(LT)	FSTU	0 * SIZE - 16 * SIZE(B)	fxch	%st(1)	FSTU	1 * SIZE - 16 * SIZE(B)#else	FSTU	0 * SIZE - 16 * SIZE(A)	fxch	%st(1)	FSTU	1 * SIZE - 16 * SIZE(A)#endif	FST	0 * SIZE(%edi, LDC)	FST	0 * SIZE(%edi)#ifndef LN	addl	$1 * SIZE, %edi#endif#if defined(LT) || defined(RN)	movl	K,  %eax	subl	KK, %eax	leal	(,%eax, SIZE), %eax	leal	(A, %eax, 1), A	leal	(B, %eax, 2), B#endif#ifdef LN	subl	$1, KK#endif#ifdef LT	addl	$1, KK#endif#ifdef RT       movl	K, %eax       sall	$BASE_SHIFT, %eax       addl	%eax, AORIG#endif	decl	I	jne	.L24#ifdef LN       movl	K, %eax       leal	(    , %eax, SIZE), %eax       leal	(B_ORIG, %eax, 2), B_ORIG#endif#if defined(LT) || defined(RN)	movl	B, B_ORIG#endif#ifdef RN	addl	$2, KK#endif#ifdef RT	subl	$2, KK#endif	ALIGN_4.L30:	movl	N,   %eax	andl	$1,  %eax	je	.L999	ALIGN_3.L31:#if defined(LT) || defined(RN)	movl	STACK_A, A#else	movl	STACK_A, %eax	movl	%eax, AORIG#endif#ifdef RT	movl	K, %eax	sall	$0 + BASE_SHIFT, %eax	subl	%eax, B_ORIG#endif#ifdef RT	subl	LDC, C#endif	movl	C, %edi#ifndef RT	addl	LDC, C#endif#ifdef LN	movl	OFFSET, %eax	addl	M, %eax	movl	%eax, KK#endif	#ifdef LT	movl	OFFSET, %eax	movl	%eax, KK#endif	movl	B_ORIG, B#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K,  %eax	subl	KK, %eax#endif	sarl	$5, %eax	jle	.L33	ALIGN_4.L32:		movl	 -16 * SIZE(B), %esi	movl	  -8 * SIZE(B), %esi	movl	   0 * SIZE(B), %esi	movl	   8 * SIZE(B), %esi	subl	 $-32 * SIZE, B	decl	  %eax	jne	  .L32	ALIGN_3.L33:	movl	M, %esi	movl	%esi, I	ALIGN_3.L34:#ifdef LN       movl	K, %eax       sall	$BASE_SHIFT, %eax       subl	%eax, AORIG#endif#if defined(LN) || defined(RT)	movl	KK, %eax	leal	(, %eax, SIZE), %eax	movl	AORIG, A	leal	(A     , %eax, 1), A	leal	(B_ORIG, %eax, 1), B#else	movl	B_ORIG, B#endif		fldz	fldz	fldz	fldz	prefetchw	1 * SIZE(%edi)#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K,  %eax	subl	KK, %eax#endif	sarl	$3, %eax 	je	.L36	ALIGN_3.L35:	FLD	-16 * SIZE(A)	FMUL	-16 * SIZE(B)	faddp	%st, %st(1)	FLD	-15 * SIZE(A)	FMUL	-15 * SIZE(B)	faddp	%st, %st(2)	FLD	-14 * SIZE(A)	FMUL	-14 * SIZE(B)	faddp	%st, %st(3)	FLD	-13 * SIZE(A)	FMUL	-13 * SIZE(B)	faddp	%st, %st(4)	FLD	-12 * SIZE(A)	FMUL	-12 * SIZE(B)	faddp	%st, %st(1)	FLD	-11 * SIZE(A)	FMUL	-11 * SIZE(B)	faddp	%st, %st(2)	FLD	-10 * SIZE(A)	FMUL	-10 * SIZE(B)	faddp	%st, %st(3)	FLD	 -9 * SIZE(A)	FMUL	 -9 * SIZE(B)	faddp	%st, %st(4)	addl	$8 * SIZE, A	addl	$8 * SIZE, B	decl	%eax	jne	.L35	ALIGN_4.L36:#if defined(LT) || defined(RN)	movl	KK, %eax#else	movl	K,  %eax	subl	KK, %eax#endif	and	$7, %eax	je	.L39	ALIGN_4.L37:	FLD	-16 * SIZE(A)	FMUL	-16 * SIZE(B)	faddp	%st, %st(1)	addl	$1 * SIZE,A	addl	$1 * SIZE,B	decl	%eax	jne	 .L37	ALIGN_4.L39:	faddp	%st, %st(2)	faddp	%st, %st(2)	faddp	%st, %st(1)#if defined(LN) || defined(RT)	movl	KK, %eax	subl	$1, %eax	movl	AORIG, A	leal	(A,      %eax, SIZE), A	leal	(B_ORIG, %eax, SIZE), B#endif#if defined(LN) || defined(LT)	FLD	0 * SIZE - 16 * SIZE(B)	fsubp	%st, %st(1)#else	FLD	0 * SIZE - 16 * SIZE(A)	fsubp	%st, %st(1)#endif#if defined(LN) || defined(LT)       FLD	0 * SIZE - 16 * SIZE(A)       fmulp	%st, %st(1)#endif#if defined(RN) || defined(RT)       FMUL	 0 * SIZE - 16 * SIZE(B)#endif#ifdef LN	subl	$1 * SIZE, %edi#endif#if defined(LN) || defined(LT)	FSTU	0 * SIZE - 16 * SIZE(B)#else	FSTU	0 * SIZE - 16 * SIZE(A)#endif	FST	0 * SIZE(%edi)#ifndef LN	addl	$1 * SIZE, %edi#endif#if defined(LT) || defined(RN)	movl	K,  %eax	subl	KK, %eax	leal	(A, %eax, SIZE), A	leal	(B, %eax, SIZE), B#endif#ifdef LN	subl	$1, KK#endif#ifdef LT	addl	$1, KK#endif#ifdef RT       movl	K, %eax       sall	$BASE_SHIFT, %eax       addl	%eax, AORIG#endif	decl	I	jne	.L34#ifdef LN       movl	K, %eax       leal	(    , %eax, SIZE), %eax       leal	(B_ORIG, %eax, 1), B_ORIG#endif#if defined(LT) || defined(RN)	movl	B, B_ORIG#endif#ifdef RN	addl	$1, KK#endif#ifdef RT	subl	$1, KK#endif	ALIGN_4.L999:	popl	%ebx	popl	%esi	popl	%edi	popl	%ebp	addl	$ARGS, %esp	ret	EPILOGUE

⌨️ 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 + -