📄 zrot_sse2.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 2 页
字号:
上一页 12
	movlpd	%xmm4, 2 * SIZE(X)	movhpd	%xmm4, 3 * SIZE(X)	movapd	%xmm2, 2 * SIZE(Y)#ifdef OPTERON	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(Y)#endif	movapd	 5 * SIZE(X), %xmm4	movapd	 4 * SIZE(Y), %xmm1	SHUFPD_1 %xmm4, %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 4 * SIZE(X)	movhpd	%xmm0, 5 * SIZE(X)	movapd	%xmm2, 4 * SIZE(Y)	movapd	 7 * SIZE(X), %xmm0	movapd	 6 * SIZE(Y), %xmm1	SHUFPD_1 %xmm0, %xmm4	movapd	%xmm1, %xmm2	movapd	%xmm4, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm4	subpd	%xmm3, %xmm2	movlpd	%xmm4, 6 * SIZE(X)	movhpd	%xmm4, 7 * SIZE(X)	movapd	%xmm2, 6 * SIZE(Y)#ifdef PENTIUM4	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(Y)#endif#ifdef OPTERON	PREFETCH	(PREFETCH_SIZE + 8) * SIZE(X)#endif	movapd	 9 * SIZE(X), %xmm4	movapd	 8 * SIZE(Y), %xmm1	SHUFPD_1 %xmm4, %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 8 * SIZE(X)	movhpd	%xmm0, 9 * SIZE(X)	movapd	%xmm2, 8 * SIZE(Y)	movapd	 11 * SIZE(X), %xmm0	movapd	 10 * SIZE(Y), %xmm1	SHUFPD_1 %xmm0, %xmm4	movapd	%xmm1, %xmm2	movapd	%xmm4, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm4	subpd	%xmm3, %xmm2	movlpd	%xmm4, 10 * SIZE(X)	movhpd	%xmm4, 11 * SIZE(X)	movapd	%xmm2, 10 * SIZE(Y)#ifdef OPTERON	PREFETCH	(PREFETCH_SIZE + 8) * SIZE(Y)#endif	movapd	13 * SIZE(X), %xmm4	movapd	12 * SIZE(Y), %xmm1	SHUFPD_1 %xmm4, %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 12 * SIZE(X)	movhpd	%xmm0, 13 * SIZE(X)	movapd	%xmm2, 12 * SIZE(Y)	movapd	 15 * SIZE(X), %xmm0	movapd	 14 * SIZE(Y), %xmm1	SHUFPD_1 %xmm0, %xmm4	movapd	%xmm1, %xmm2	movapd	%xmm4, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm4	subpd	%xmm3, %xmm2	movlpd	%xmm4, 14 * SIZE(X)	movhpd	%xmm4, 15 * SIZE(X)	movapd	%xmm2, 14 * SIZE(Y)	addl	$16 * SIZE, Y	addl	$16 * SIZE, X	decl	I	jg	.L31	ALIGN_3.L34:	testl	$7, N	jle	.L999	testl	$4, N	jle	.L35	movapd	 1 * SIZE(X), %xmm4	movapd	 0 * SIZE(Y), %xmm1	SHUFPD_1 %xmm4, %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movapd	%xmm2, 0 * SIZE(Y)	movapd	 3 * SIZE(X), %xmm0	movapd	 2 * SIZE(Y), %xmm1	SHUFPD_1 %xmm0, %xmm4	movapd	%xmm1, %xmm2	movapd	%xmm4, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm4	subpd	%xmm3, %xmm2	movlpd	%xmm4, 2 * SIZE(X)	movhpd	%xmm4, 3 * SIZE(X)	movapd	%xmm2, 2 * SIZE(Y)	movapd	 5 * SIZE(X), %xmm4	movapd	 4 * SIZE(Y), %xmm1	SHUFPD_1 %xmm4, %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 4 * SIZE(X)	movhpd	%xmm0, 5 * SIZE(X)	movapd	%xmm2, 4 * SIZE(Y)	movapd	 7 * SIZE(X), %xmm0	movapd	 6 * SIZE(Y), %xmm1	SHUFPD_1 %xmm0, %xmm4	movapd	%xmm1, %xmm2	movapd	%xmm4, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm4	subpd	%xmm3, %xmm2	movlpd	%xmm4, 6 * SIZE(X)	movhpd	%xmm4, 7 * SIZE(X)	movapd	%xmm2, 6 * SIZE(Y)	addl	$8 * SIZE, Y	addl	$8 * SIZE, X	ALIGN_3.L35:	testl	$2, N	jle	.L36	movapd	 1 * SIZE(X), %xmm4	movapd	 0 * SIZE(Y), %xmm1	SHUFPD_1 %xmm4, %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movapd	%xmm2, 0 * SIZE(Y)	movapd	 3 * SIZE(X), %xmm0	movapd	 2 * SIZE(Y), %xmm1	SHUFPD_1 %xmm0, %xmm4	movapd	%xmm1, %xmm2	movapd	%xmm4, %xmm3	mulpd	C, %xmm4	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm4	subpd	%xmm3, %xmm2	movlpd	%xmm4, 2 * SIZE(X)	movhpd	%xmm4, 3 * SIZE(X)	movapd	%xmm2, 2 * SIZE(Y)	addl	$4 * SIZE, Y	addl	$4 * SIZE, X	ALIGN_3.L36:	testl	$1, N	jle	.L999	movapd	 1 * SIZE(X), %xmm4	movapd	 0 * SIZE(Y), %xmm1	SHUFPD_1 %xmm4, %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movapd	%xmm2, 0 * SIZE(Y)	jmp	.L999	ALIGN_3.L40:	movsd	0 * SIZE(Y), %xmm1	movsd	0 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulsd	C, %xmm0	mulsd	S, %xmm1	mulsd	C, %xmm2	mulsd	S, %xmm3	addsd	%xmm1, %xmm0	subsd	%xmm3, %xmm2	movsd	%xmm0, 0 * SIZE(X)	movsd	%xmm2, 0 * SIZE(Y)	addl	$1 * SIZE, Y	addl	$1 * SIZE, X	decl	N	jle	.L47	movl	N,  I	sarl	$3, I	jle	.L44	ALIGN_3.L41:#ifdef PENTIUM4	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(X)#endif#ifdef OPTERON	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(X)#endif	movapd	0 * SIZE(Y), %xmm1	movapd	0 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 0 * SIZE(X)	movapd	%xmm2, 0 * SIZE(Y)	movapd	2 * SIZE(Y), %xmm1	movapd	2 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 2 * SIZE(X)	movapd	%xmm2, 2 * SIZE(Y)#ifdef OPTERON	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(Y)#endif	movapd	4 * SIZE(Y), %xmm1	movapd	4 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 4 * SIZE(X)	movapd	%xmm2, 4 * SIZE(Y)	movapd	6 * SIZE(Y), %xmm1	movapd	6 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 6 * SIZE(X)	movapd	%xmm2, 6 * SIZE(Y)#ifdef PENTIUM4	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(Y)#endif#ifdef OPTERON	PREFETCH	(PREFETCH_SIZE + 8) * SIZE(X)#endif	movapd	8 * SIZE(Y), %xmm1	movapd	8 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 8 * SIZE(X)	movapd	%xmm2, 8 * SIZE(Y)	movapd	10 * SIZE(Y), %xmm1	movapd	10 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 10 * SIZE(X)	movapd	%xmm2, 10 * SIZE(Y)#ifdef OPTERON	PREFETCH	(PREFETCH_SIZE + 8) * SIZE(Y)#endif	movapd	12 * SIZE(Y), %xmm1	movapd	12 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 12 * SIZE(X)	movapd	%xmm2, 12 * SIZE(Y)	movapd	14 * SIZE(Y), %xmm1	movapd	14 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 14 * SIZE(X)	movapd	%xmm2, 14 * SIZE(Y)	addl	$16 * SIZE, X	addl	$16 * SIZE, Y	decl	I	jg	.L41	ALIGN_3.L44:	testl	$4, N	jle	.L45	movapd	0 * SIZE(Y), %xmm1	movapd	0 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 0 * SIZE(X)	movapd	%xmm2, 0 * SIZE(Y)	movapd	2 * SIZE(Y), %xmm1	movapd	2 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 2 * SIZE(X)	movapd	%xmm2, 2 * SIZE(Y)	movapd	4 * SIZE(Y), %xmm1	movapd	4 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 4 * SIZE(X)	movapd	%xmm2, 4 * SIZE(Y)	movapd	6 * SIZE(Y), %xmm1	movapd	6 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 6 * SIZE(X)	movapd	%xmm2, 6 * SIZE(Y)	addl	$8 * SIZE, X	addl	$8 * SIZE, Y	ALIGN_3.L45:	testl	$2, N	jle	.L46	movapd	0 * SIZE(Y), %xmm1	movapd	0 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 0 * SIZE(X)	movapd	%xmm2, 0 * SIZE(Y)	movapd	2 * SIZE(Y), %xmm1	movapd	2 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 2 * SIZE(X)	movapd	%xmm2, 2 * SIZE(Y)	addl	$4 * SIZE, X	addl	$4 * SIZE, Y	ALIGN_3.L46:	testl	$1, N	jle	.L47	movapd	0 * SIZE(Y), %xmm1	movapd	0 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movapd	%xmm0, 0 * SIZE(X)	movapd	%xmm2, 0 * SIZE(Y)	addl	$2 * SIZE, Y	addl	$2 * SIZE, X	ALIGN_3.L47:	movsd	0 * SIZE(Y), %xmm1	movsd	0 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulsd	C, %xmm0	mulsd	S, %xmm1	mulsd	C, %xmm2	mulsd	S, %xmm3	addsd	%xmm1, %xmm0	subsd	%xmm3, %xmm2	movsd	%xmm0, 0 * SIZE(X)	movsd	%xmm2, 0 * SIZE(Y)	jmp	.L999	ALIGN_3.L50:	movl	N,  I	sarl	$2, I	jle	.L55	ALIGN_3.L53:	MOVSD	0 * SIZE(Y), %xmm1	movhpd	1 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhpd	1 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movlpd	%xmm2, 0 * SIZE(Y)	movhpd	%xmm2, 1 * SIZE(Y)	addl	INCX, X	addl	INCY, Y	MOVSD	0 * SIZE(Y), %xmm1	movhpd	1 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhpd	1 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movlpd	%xmm2, 0 * SIZE(Y)	movhpd	%xmm2, 1 * SIZE(Y)	addl	INCX, X	addl	INCY, Y	MOVSD	0 * SIZE(Y), %xmm1	movhpd	1 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhpd	1 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movlpd	%xmm2, 0 * SIZE(Y)	movhpd	%xmm2, 1 * SIZE(Y)	addl	INCX, X	addl	INCY, Y	MOVSD	0 * SIZE(Y), %xmm1	movhpd	1 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhpd	1 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movlpd	%xmm2, 0 * SIZE(Y)	movhpd	%xmm2, 1 * SIZE(Y)	addl	INCX, X	addl	INCY, Y	decl	I	jg	.L53	ALIGN_3.L55:	movl	N,  I	andl	$3, I	jle	.L999	ALIGN_3.L56:	MOVSD	0 * SIZE(Y), %xmm1	movhpd	1 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhpd	1 * SIZE(X), %xmm0	movapd	%xmm1, %xmm2	movapd	%xmm0, %xmm3	mulpd	C, %xmm0	mulpd	S, %xmm1	mulpd	C, %xmm2	mulpd	S, %xmm3	addpd	%xmm1, %xmm0	subpd	%xmm3, %xmm2	movlpd	%xmm0, 0 * SIZE(X)	movhpd	%xmm0, 1 * SIZE(X)	movlpd	%xmm2, 0 * SIZE(Y)	movhpd	%xmm2, 1 * SIZE(Y)	addl	INCX, X	addl	INCY, Y	decl	I	jg	.L56	ALIGN_3.L999:	popl	%ebx	popl	%esi	popl	%edi	ret	EPILOGUE
上一页 12
⌨️ 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 + -