📄 zrot_sse.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 2 页
字号:
上一页 12
	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movaps	%xmm0, 0 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	movhps	%xmm2, 2 * SIZE(Y)	MOVSD	4 * SIZE(Y), %xmm1	movhps	6 * SIZE(Y), %xmm1	movaps	4 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movaps	%xmm0, 4 * SIZE(X)	movlps	%xmm2, 4 * SIZE(Y)	movhps	%xmm2, 6 * SIZE(Y)	MOVSD	 8 * SIZE(Y), %xmm1	movhps	10 * SIZE(Y), %xmm1	movaps	 8 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movaps	%xmm0,  8 * SIZE(X)	movlps	%xmm2,  8 * SIZE(Y)	movhps	%xmm2, 10 * SIZE(Y)	MOVSD	12 * SIZE(Y), %xmm1	movhps	14 * SIZE(Y), %xmm1	movaps	12 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movaps	%xmm0, 12 * SIZE(X)	movlps	%xmm2, 12 * SIZE(Y)	movhps	%xmm2, 14 * SIZE(Y)	addl	$16 * SIZE, X	addl	$16 * SIZE, Y	ALIGN_3.L25:	testl	$4, N	jle	.L26	MOVSD	0 * SIZE(Y), %xmm1	movhps	2 * SIZE(Y), %xmm1	movaps	0 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movaps	%xmm0, 0 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	movhps	%xmm2, 2 * SIZE(Y)	MOVSD	4 * SIZE(Y), %xmm1	movhps	6 * SIZE(Y), %xmm1	movaps	4 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movaps	%xmm0, 4 * SIZE(X)	movlps	%xmm2, 4 * SIZE(Y)	movhps	%xmm2, 6 * SIZE(Y)	addl	$8 * SIZE, X	addl	$8 * SIZE, Y	ALIGN_3.L26:	testl	$2, N	jle	.L27	MOVSD	0 * SIZE(Y), %xmm1	movhps	2 * SIZE(Y), %xmm1	movaps	0 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movaps	%xmm0, 0 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	movhps	%xmm2, 2 * SIZE(Y)	addl	$4 * SIZE, X	addl	$4 * SIZE, Y	ALIGN_3.L27:	testl	$1, N	jle	.L999	movsd	0 * SIZE(Y), %xmm1	movsd	0 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 0 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	jmp	.L999	ALIGN_3.L30:	movl	N,  I	sarl	$4, I	jle	.L34	ALIGN_3.L31:#ifdef PENTIUM4	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(X)#endif#ifdef OPTERON	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(X)#endif	MOVSD	0 * SIZE(Y), %xmm1	movhps	2 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhps	2 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 0 * SIZE(X)	movhps	%xmm0, 2 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	movhps	%xmm2, 2 * SIZE(Y)	MOVSD	4 * SIZE(Y), %xmm1	movhps	6 * SIZE(Y), %xmm1	MOVSD	4 * SIZE(X), %xmm0	movhps	6 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 4 * SIZE(X)	movhps	%xmm0, 6 * SIZE(X)	movlps	%xmm2, 4 * SIZE(Y)	movhps	%xmm2, 6 * SIZE(Y)#ifdef OPTERON	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(Y)#endif	MOVSD	 8 * SIZE(Y), %xmm1	movhps	10 * SIZE(Y), %xmm1	MOVSD	 8 * SIZE(X), %xmm0	movhps	10 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0,  8 * SIZE(X)	movhps	%xmm0, 10 * SIZE(X)	movlps	%xmm2,  8 * SIZE(Y)	movhps	%xmm2, 10 * SIZE(Y)	MOVSD	12 * SIZE(Y), %xmm1	movhps	14 * SIZE(Y), %xmm1	MOVSD	12 * SIZE(X), %xmm0	movhps	14 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 12 * SIZE(X)	movhps	%xmm0, 14 * SIZE(X)	movlps	%xmm2, 12 * SIZE(Y)	movhps	%xmm2, 14 * SIZE(Y)#ifdef PENTIUM4	PREFETCH	(PREFETCH_SIZE + 0) * SIZE(Y)#endif#ifdef OPTERON	PREFETCH	(PREFETCH_SIZE + 16) * SIZE(X)#endif	MOVSD	16 * SIZE(Y), %xmm1	movhps	18 * SIZE(Y), %xmm1	MOVSD	16 * SIZE(X), %xmm0	movhps	18 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 16 * SIZE(X)	movhps	%xmm0, 18 * SIZE(X)	movlps	%xmm2, 16 * SIZE(Y)	movhps	%xmm2, 18 * SIZE(Y)	MOVSD	20 * SIZE(Y), %xmm1	movhps	22 * SIZE(Y), %xmm1	MOVSD	20 * SIZE(X), %xmm0	movhps	22 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 20 * SIZE(X)	movhps	%xmm0, 22 * SIZE(X)	movlps	%xmm2, 20 * SIZE(Y)	movhps	%xmm2, 22 * SIZE(Y)#ifdef OPTERON	PREFETCH	(PREFETCH_SIZE + 16) * SIZE(Y)#endif	MOVSD	24 * SIZE(Y), %xmm1	movhps	26 * SIZE(Y), %xmm1	MOVSD	24 * SIZE(X), %xmm0	movhps	26 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 24 * SIZE(X)	movhps	%xmm0, 26 * SIZE(X)	movlps	%xmm2, 24 * SIZE(Y)	movhps	%xmm2, 26 * SIZE(Y)	MOVSD	28 * SIZE(Y), %xmm1	movhps	30 * SIZE(Y), %xmm1	MOVSD	28 * SIZE(X), %xmm0	movhps	30 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 28 * SIZE(X)	movhps	%xmm0, 30 * SIZE(X)	movlps	%xmm2, 28 * SIZE(Y)	movhps	%xmm2, 30 * SIZE(Y)	addl	$32 * SIZE, X	addl	$32 * SIZE, Y	decl	I	jg	.L31	ALIGN_3.L34:	testl	$15, N	jle	.L999	testl	$8, N	jle	.L35	MOVSD	0 * SIZE(Y), %xmm1	movhps	2 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhps	2 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 0 * SIZE(X)	movhps	%xmm0, 2 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	movhps	%xmm2, 2 * SIZE(Y)	MOVSD	4 * SIZE(Y), %xmm1	movhps	6 * SIZE(Y), %xmm1	MOVSD	4 * SIZE(X), %xmm0	movhps	6 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 4 * SIZE(X)	movhps	%xmm0, 6 * SIZE(X)	movlps	%xmm2, 4 * SIZE(Y)	movhps	%xmm2, 6 * SIZE(Y)	MOVSD	 8 * SIZE(Y), %xmm1	movhps	10 * SIZE(Y), %xmm1	MOVSD	 8 * SIZE(X), %xmm0	movhps	10 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0,  8 * SIZE(X)	movhps	%xmm0, 10 * SIZE(X)	movlps	%xmm2,  8 * SIZE(Y)	movhps	%xmm2, 10 * SIZE(Y)	MOVSD	12 * SIZE(Y), %xmm1	movhps	14 * SIZE(Y), %xmm1	MOVSD	12 * SIZE(X), %xmm0	movhps	14 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 12 * SIZE(X)	movhps	%xmm0, 14 * SIZE(X)	movlps	%xmm2, 12 * SIZE(Y)	movhps	%xmm2, 14 * SIZE(Y)	addl	$16 * SIZE, X	addl	$16 * SIZE, Y	ALIGN_3.L35:	testl	$4, N	jle	.L36	MOVSD	0 * SIZE(Y), %xmm1	movhps	2 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhps	2 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 0 * SIZE(X)	movhps	%xmm0, 2 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	movhps	%xmm2, 2 * SIZE(Y)	MOVSD	4 * SIZE(Y), %xmm1	movhps	6 * SIZE(Y), %xmm1	MOVSD	4 * SIZE(X), %xmm0	movhps	6 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 4 * SIZE(X)	movhps	%xmm0, 6 * SIZE(X)	movlps	%xmm2, 4 * SIZE(Y)	movhps	%xmm2, 6 * SIZE(Y)	addl	$8 * SIZE, X	addl	$8 * SIZE, Y	ALIGN_3.L36:	testl	$2, N	jle	.L37	MOVSD	0 * SIZE(Y), %xmm1	movhps	2 * SIZE(Y), %xmm1	MOVSD	0 * SIZE(X), %xmm0	movhps	2 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 0 * SIZE(X)	movhps	%xmm0, 2 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	movhps	%xmm2, 2 * SIZE(Y)	addl	$4 * SIZE, X	addl	$4 * SIZE, Y	ALIGN_3.L37:	testl	$1, N	jle	.L999#ifndef HAVE_SSE2	xorps	%xmm0, %xmm0	xorps	%xmm1, %xmm1#endif	movsd	0 * SIZE(Y), %xmm1	movsd	0 * SIZE(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, 0 * SIZE(X)	movlps	%xmm2, 0 * SIZE(Y)	jmp	.L999	ALIGN_3	ALIGN_3.L50:	movl	N,  I	sarl	$2, I	jle	.L55	ALIGN_3.L53:	MOVSD	(Y), %xmm1	movhps	(Y, INCY), %xmm1	MOVSD	(X), %xmm0	movhps	(X, INCX), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, (X)	movhps	%xmm0, (X, INCX)	movlps	%xmm2, (Y)	movhps	%xmm2, (Y, INCY)	leal	(X, INCX, 2), X	leal	(Y, INCY, 2), Y	MOVSD	(Y), %xmm1	movhps	(Y, INCY), %xmm1	MOVSD	(X), %xmm0	movhps	(X, INCX), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, (X)	movhps	%xmm0, (X, INCX)	movlps	%xmm2, (Y)	movhps	%xmm2, (Y, INCY)	leal	(X, INCX, 2), X	leal	(Y, INCY, 2), Y	decl	I	jg	.L53	ALIGN_3.L55:#ifndef HAVE_SSE2	xorps	%xmm0, %xmm0	xorps	%xmm1, %xmm1#endif	movl	N,  I	andl	$3, I	jle	.L999	ALIGN_3.L56:	movsd	(Y), %xmm1	movsd	(X), %xmm0	movaps	%xmm1, %xmm2	movaps	%xmm0, %xmm3	mulps	C, %xmm0	mulps	S, %xmm1	mulps	C, %xmm2	mulps	S, %xmm3	addps	%xmm1, %xmm0	subps	%xmm3, %xmm2	movlps	%xmm0, (X)	movlps	%xmm2, (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 + -