📄 trsm_kernel_ln_4x4_core2.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 5 页
字号:
	addpd	%xmm2, %xmm8	addpd	%xmm3, %xmm9	addpd	%xmm4, %xmm10	addpd	%xmm5, %xmm11	subq	$ -8 * SIZE, AO	subq	$-32 * SIZE, BO	subq	$1, %rax	jne    .L22	ALIGN_4.L25:#if defined(LT) || defined(RN)	movq	KK, %rax#else	movq	K, %rax	subq	KK, %rax#endif	andq	$3, %rax		# if (k & 1)	BRANCH	je .L29	ALIGN_4.L26:	movapd	-16 * SIZE(AO), %xmm0	movapd	-16 * SIZE(BO), %xmm2	movapd	-14 * SIZE(BO), %xmm3	movapd	-12 * SIZE(BO), %xmm4	movapd	-10 * SIZE(BO), %xmm5	mulpd	%xmm0, %xmm2	mulpd	%xmm0, %xmm3	mulpd	%xmm0, %xmm4	mulpd	%xmm0, %xmm5	addpd	%xmm2, %xmm8	addpd	%xmm3, %xmm9	addpd	%xmm4, %xmm10	addpd	%xmm5, %xmm11	addq	$2 * SIZE, AO	addq	$8 * SIZE, BO	subq	$1, %rax	jne    .L26	ALIGN_4.L29:#if defined(LN) || defined(RT)	movq	KK, %rax#ifdef LN	subq	$2, %rax#else	subq	$4, %rax#endif	movq	AORIG, AO	movq	BORIG, B	leaq	16 * SIZE + BUFFER, BO	leaq	(, %rax, SIZE), %rax	leaq	(AO, %rax, 2), AO	leaq	(B,  %rax, 4), B	leaq	(BO, %rax, 8), BO#endif#if defined(LN) || defined(LT)	movapd	%xmm8, %xmm0	unpcklpd %xmm9, %xmm8	unpckhpd %xmm9, %xmm0	movapd	%xmm10, %xmm2	unpcklpd %xmm11, %xmm10	unpckhpd %xmm11, %xmm2	movapd	 -16 * SIZE(B), %xmm9	movapd	 -14 * SIZE(B), %xmm11	movapd	 -12 * SIZE(B), %xmm13	movapd	 -10 * SIZE(B), %xmm15	subpd	%xmm8,  %xmm9	subpd	%xmm10,  %xmm11	subpd	%xmm0,  %xmm13	subpd	%xmm2, %xmm15#else	movapd	 -16 * SIZE(AO), %xmm0	movapd	 -14 * SIZE(AO), %xmm2	movapd	 -12 * SIZE(AO), %xmm4	movapd	 -10 * SIZE(AO), %xmm6	subpd	%xmm8, %xmm0	subpd	%xmm9, %xmm2	subpd	%xmm10, %xmm4	subpd	%xmm11, %xmm6#endif#ifdef LN	movddup	 -13 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm13	mulpd	 %xmm8, %xmm15	movddup	 -14 * SIZE(AO), %xmm10	movapd	 %xmm10, %xmm12	mulpd	 %xmm13, %xmm10	subpd	 %xmm10, %xmm9	mulpd	 %xmm15, %xmm12	subpd	 %xmm12, %xmm11	movddup	 -16 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm9	mulpd	 %xmm8, %xmm11#endif#ifdef LT	movddup	 -16 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm9	mulpd	 %xmm8, %xmm11	movddup	 -15 * SIZE(AO), %xmm10	movapd	 %xmm10, %xmm12	mulpd	 %xmm9, %xmm10	subpd	 %xmm10, %xmm13	mulpd	 %xmm11, %xmm12	subpd	 %xmm12, %xmm15	movddup	 -13 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm13	mulpd	 %xmm8, %xmm15#endif#ifdef RN	movddup	 -16 * SIZE(B), %xmm8	mulpd	 %xmm8, %xmm0	movddup	 -15 * SIZE(B), %xmm9	mulpd	 %xmm0, %xmm9	subpd	 %xmm9, %xmm2	movddup	 -14 * SIZE(B), %xmm10	mulpd	 %xmm0, %xmm10	subpd	 %xmm10, %xmm4	movddup	 -13 * SIZE(B), %xmm11	mulpd	 %xmm0, %xmm11	subpd	 %xmm11, %xmm6	movddup	 -11 * SIZE(B), %xmm8	mulpd	 %xmm8, %xmm2	movddup	 -10 * SIZE(B), %xmm9	mulpd	 %xmm2, %xmm9	subpd	 %xmm9, %xmm4	movddup	  -9 * SIZE(B), %xmm10	mulpd	 %xmm2, %xmm10	subpd	 %xmm10, %xmm6	movddup	  -6 * SIZE(B), %xmm8	mulpd	 %xmm8, %xmm4	movddup	  -5 * SIZE(B), %xmm9	mulpd	 %xmm4, %xmm9	subpd	 %xmm9, %xmm6	movddup	  -1 * SIZE(B), %xmm8	mulpd	 %xmm8, %xmm6#endif#ifdef RT	movddup	-1 * SIZE(B), %xmm8	mulpd	 %xmm8, %xmm6	movddup	-2 * SIZE(B), %xmm9	mulpd	 %xmm6, %xmm9	subpd	 %xmm9, %xmm4	movddup	-3 * SIZE(B), %xmm10	mulpd	 %xmm6, %xmm10	subpd	 %xmm10, %xmm2	movddup	-4 * SIZE(B), %xmm11	mulpd	 %xmm6, %xmm11	subpd	 %xmm11, %xmm0	movddup	-6 * SIZE(B), %xmm8	mulpd	 %xmm8, %xmm4	movddup	-7 * SIZE(B), %xmm9	mulpd	 %xmm4, %xmm9	subpd	 %xmm9, %xmm2	movddup	-8 * SIZE(B), %xmm10	mulpd	 %xmm4, %xmm10	subpd	 %xmm10, %xmm0	movddup	-11 * SIZE(B), %xmm8	mulpd	 %xmm8, %xmm2	movddup	-12 * SIZE(B), %xmm9	mulpd	 %xmm2, %xmm9	subpd	 %xmm9, %xmm0	movddup	-16 * SIZE(B), %xmm8	mulpd	 %xmm8, %xmm0#endif#ifdef LN	subq	$2 * SIZE, CO1	subq	$2 * SIZE, CO2#endif#if defined(LN) || defined(LT)	movsd	%xmm9,  0 * SIZE(CO1)	movsd	%xmm13,  1 * SIZE(CO1)	movhpd	%xmm9,  0 * SIZE(CO2)	movhpd	%xmm13,  1 * SIZE(CO2)	movsd	%xmm11,  0 * SIZE(CO1, LDC, 2)	movsd	%xmm15,  1 * SIZE(CO1, LDC, 2)	movhpd	%xmm11,  0 * SIZE(CO2, LDC, 2)	movhpd	%xmm15,  1 * SIZE(CO2, LDC, 2)#else	movsd	%xmm0,  0 * SIZE(CO1)	movhpd	%xmm0,  1 * SIZE(CO1)	movsd	%xmm2,  0 * SIZE(CO2)	movhpd	%xmm2,  1 * SIZE(CO2)	movsd	%xmm4,  0 * SIZE(CO1, LDC, 2)	movhpd	%xmm4,  1 * SIZE(CO1, LDC, 2)	movsd	%xmm6,  0 * SIZE(CO2, LDC, 2)	movhpd	%xmm6,  1 * SIZE(CO2, LDC, 2)#endif#if defined(LN) || defined(LT)	movapd	%xmm9,  -16 * SIZE(B)	movapd	%xmm11,  -14 * SIZE(B)	movapd	%xmm13,  -12 * SIZE(B)	movapd	%xmm15,  -10 * SIZE(B)	movddup	%xmm9,  %xmm8	SHUFPD_3 %xmm9,  %xmm9	movddup	%xmm11,  %xmm10	SHUFPD_3 %xmm11,  %xmm11	movddup	%xmm13,  %xmm12	SHUFPD_3 %xmm13,  %xmm13	movddup	%xmm15,  %xmm14	SHUFPD_3 %xmm15,  %xmm15	movapd	%xmm8, -16 * SIZE(BO)	movapd	%xmm9, -14 * SIZE(BO)	movapd	%xmm10, -12 * SIZE(BO)	movapd	%xmm11, -10 * SIZE(BO)	movapd	%xmm12,  -8 * SIZE(BO)	movapd	%xmm13,  -6 * SIZE(BO)	movapd	%xmm14,  -4 * SIZE(BO)	movapd	%xmm15,  -2 * SIZE(BO)#else	movapd	%xmm0,  -16 * SIZE(AO)	movapd	%xmm2, -14 * SIZE(AO)	movapd	%xmm4, -12 * SIZE(AO)	movapd	%xmm6, -10 * SIZE(AO)#endif#ifndef LN	addq	$2 * SIZE, CO1	addq	$2 * SIZE, CO2#endif#if defined(LT) || defined(RN)	movq	K,  %rax	subq	KK, %rax	leaq	(,%rax, SIZE), %rax	leaq	(AO, %rax, 2), AO#ifdef LT	addq	$8 * SIZE, B#endif#endif#ifdef LN	subq	$2, KK	movq	BORIG, B#endif#ifdef LT	addq	$2, KK#endif#ifdef RT       movq	K, %rax       movq	BORIG, B       salq	$1 + BASE_SHIFT, %rax       addq	%rax, AORIG#endif	ALIGN_4	.L30:	movq	M,  I	sarq	$2, I	# i = (m >> 2)	jle	.L39	ALIGN_4.L11:#ifdef LN       movq	K, %rax       salq	$2 + BASE_SHIFT, %rax       subq	%rax, AORIG#endif#if defined(LN) || defined(RT)	movq	KK, %rax	movq	AORIG, AO	leaq	(, %rax, SIZE), %rax	leaq	(AO, %rax, 4), AO#endif	leaq	16 * SIZE + BUFFER, BO#if defined(LN) || defined(RT)	movq	KK, %rax	salq	$2 + BASE_SHIFT, %rax	leaq	(BO, %rax, 2), BO#endif		pxor	%xmm8, %xmm8	pxor	%xmm9, %xmm9	pxor	%xmm10, %xmm10	pxor	%xmm11, %xmm11#ifdef LN	prefetcht2     -3 * SIZE(CO1)	pxor	%xmm12, %xmm12	prefetcht2     -3 * SIZE(CO2)	pxor	%xmm13, %xmm13	prefetcht2     -3 * SIZE(CO1, LDC, 2)	pxor	%xmm14, %xmm14	prefetcht2     -3 * SIZE(CO2, LDC, 2)	pxor	%xmm15, %xmm15#else	prefetcht2     3 * SIZE(CO1)	pxor	%xmm12, %xmm12	prefetcht2     3 * SIZE(CO2)	pxor	%xmm13, %xmm13	prefetcht2     3 * SIZE(CO1, LDC, 2)	pxor	%xmm14, %xmm14	prefetcht2     3 * SIZE(CO2, LDC, 2)	pxor	%xmm15, %xmm15#endif	pxor	%xmm2, %xmm2	pxor	%xmm3, %xmm3	pxor	%xmm4, %xmm4	pxor	%xmm5, %xmm5#if defined(LT) || defined(RN)	movq	KK, %rax#else	movq	K, %rax	subq	KK, %rax#endif	sarq	$2, %rax	jle	.L15	ALIGN_4.L12:		PREFETCH (PREFETCHSIZE +  0) * SIZE(AO)	addpd	%xmm2, %xmm10	movapd	-16 * SIZE(AO), %xmm0	addpd	%xmm3, %xmm14	movapd	-16 * SIZE(BO), %xmm2	movapd	 %xmm2, %xmm3	movapd	-14 * SIZE(AO), %xmm1	mulpd	%xmm0, %xmm2	mulpd	%xmm1, %xmm3	addpd	%xmm4, %xmm11	movapd	-14 * SIZE(BO), %xmm4	addpd	%xmm5, %xmm15	movapd	 %xmm4, %xmm5	mulpd	%xmm0, %xmm4	mulpd	%xmm1, %xmm5	addpd	%xmm2, %xmm8	addpd	%xmm3, %xmm12	movapd	-12 * SIZE(BO), %xmm2	movapd	 %xmm2, %xmm3	mulpd	%xmm0, %xmm2	mulpd	%xmm1, %xmm3	addpd	%xmm4, %xmm9	movapd	-10 * SIZE(BO), %xmm4	addpd	%xmm5, %xmm13	movapd	 %xmm4, %xmm5	mulpd	%xmm0, %xmm4	movapd	-12 * SIZE(AO), %xmm0	mulpd	%xmm1, %xmm5	addpd	%xmm2, %xmm10	addpd	%xmm3, %xmm14	movapd	 -8 * SIZE(BO), %xmm2	movapd	 %xmm2, %xmm3	movapd	-10 * SIZE(AO), %xmm1	mulpd	%xmm0, %xmm2	mulpd	%xmm1, %xmm3	addpd	%xmm4, %xmm11	addpd	%xmm5, %xmm15	movapd	-6 * SIZE(BO), %xmm4	movapd	 %xmm4, %xmm5	mulpd	%xmm0, %xmm4	mulpd	%xmm1, %xmm5	addpd	%xmm2, %xmm8	addpd	%xmm3, %xmm12	movapd	-4 * SIZE(BO), %xmm2	movapd	 %xmm2, %xmm3	mulpd	%xmm0, %xmm2	mulpd	%xmm1, %xmm3	addpd	%xmm4, %xmm9	addpd	%xmm5, %xmm13	movapd	-2 * SIZE(BO), %xmm4	movapd	 %xmm4, %xmm5	mulpd	%xmm0, %xmm4	PREFETCH (PREFETCHSIZE +  8) * SIZE(AO)	movapd	-8 * SIZE(AO), %xmm0	mulpd	%xmm1, %xmm5	addpd	%xmm2, %xmm10	addpd	%xmm3, %xmm14	movapd	 0 * SIZE(BO), %xmm2	movapd	 %xmm2, %xmm3	movapd	-6 * SIZE(AO), %xmm1	mulpd	%xmm0, %xmm2	mulpd	%xmm1, %xmm3	addpd	%xmm4, %xmm11	addpd	%xmm5, %xmm15	movapd	 2 * SIZE(BO), %xmm4	movapd	 %xmm4, %xmm5	mulpd	%xmm0, %xmm4	mulpd	%xmm1, %xmm5	addpd	%xmm2, %xmm8	addpd	%xmm3, %xmm12	movapd	 4 * SIZE(BO), %xmm2	movapd	 %xmm2, %xmm3	mulpd	%xmm0, %xmm2	mulpd	%xmm1, %xmm3	addpd	%xmm4, %xmm9	addpd	%xmm5, %xmm13	movapd	 6 * SIZE(BO), %xmm4	movapd	 %xmm4, %xmm5	mulpd	%xmm0, %xmm4	movapd	-4 * SIZE(AO), %xmm0	mulpd	%xmm1, %xmm5	addpd	%xmm2, %xmm10	addpd	%xmm3, %xmm14	movapd	 8 * SIZE(BO), %xmm2	movapd	 %xmm2, %xmm3	movapd	-2 * SIZE(AO), %xmm1	mulpd	%xmm0, %xmm2	mulpd	%xmm1, %xmm3	addpd	%xmm4, %xmm11	addpd	%xmm5, %xmm15	movapd	10 * SIZE(BO), %xmm4	movapd	 %xmm4, %xmm5	addq	$32 * SIZE, BO	mulpd	%xmm0, %xmm4	mulpd	%xmm1, %xmm5	addpd	%xmm2, %xmm8	addpd	%xmm3, %xmm12	movapd	-20 * SIZE(BO), %xmm2	movapd	 %xmm2, %xmm3	mulpd	%xmm0, %xmm2	subq	$-16 * SIZE, AO	mulpd	%xmm1, %xmm3	addpd	%xmm4, %xmm9	addpd	%xmm5, %xmm13	movapd	-18 * SIZE(BO), %xmm4	movapd	 %xmm4, %xmm5	mulpd	%xmm0, %xmm4	subq	$1, %rax	mulpd	%xmm1, %xmm5	jg,pt	.L12	ALIGN_4.L15:#if defined(LT) || defined(RN)	movq	KK, %rax#else	movq	K, %rax	subq	KK, %rax#endif	andq	$3, %rax		# if (k & 1)	BRANCH	je .L19	ALIGN_4.L16:	movapd	-16 * SIZE(AO), %xmm0	addpd	%xmm2, %xmm10	movapd	-16 * SIZE(BO), %xmm2	addpd	%xmm3, %xmm14	movapd	 %xmm2, %xmm3	movapd	-14 * SIZE(AO), %xmm1	mulpd	%xmm0, %xmm2	mulpd	%xmm1, %xmm3	addpd	%xmm4, %xmm11	movapd	-14 * SIZE(BO), %xmm4	addpd	%xmm5, %xmm15	movapd	 %xmm4, %xmm5	mulpd	%xmm0, %xmm4	mulpd	%xmm1, %xmm5	addpd	%xmm2, %xmm8	movapd	-12 * SIZE(BO), %xmm2	addpd	%xmm3, %xmm12	movapd	 %xmm2, %xmm3	mulpd	%xmm0, %xmm2	mulpd	%xmm1, %xmm3	addpd	%xmm4, %xmm9	movapd	-10 * SIZE(BO), %xmm4	addpd	%xmm5, %xmm13	movapd	 %xmm4, %xmm5	mulpd	%xmm0, %xmm4	mulpd	%xmm1, %xmm5	addq	$4 * SIZE, AO		# aoffset  += 4	addq	$8 * SIZE, BO		# boffset1 += 8	subq	$1, %rax	jg,pt	.L16	ALIGN_4.L19:	addpd	%xmm2, %xmm10	addpd	%xmm3, %xmm14	addpd	%xmm4, %xmm11	addpd	%xmm5, %xmm15#if defined(LN) || defined(RT)	movq	KK, %rax#ifdef LN	subq	$4, %rax#else	subq	$4, %rax#endif	movq	AORIG, AO	movq	BORIG, B	leaq	16 * SIZE + BUFFER, BO	leaq	(, %rax, SIZE), %rax	leaq	(AO, %rax, 4), AO	leaq	(B,  %rax, 4), B	leaq	(BO, %rax, 8), BO#endif#if defined(LN) || defined(LT)	movapd	%xmm8, %xmm0	unpcklpd %xmm9, %xmm8	unpckhpd %xmm9, %xmm0	movapd	%xmm10, %xmm2	unpcklpd %xmm11, %xmm10	unpckhpd %xmm11, %xmm2	movapd	%xmm12, %xmm4	unpcklpd %xmm13, %xmm12	unpckhpd %xmm13, %xmm4	movapd	%xmm14, %xmm6	unpcklpd %xmm15, %xmm14	unpckhpd %xmm15, %xmm6	movapd	-16 * SIZE(B), %xmm9	movapd	-14 * SIZE(B), %xmm11	movapd	-12 * SIZE(B), %xmm13	movapd	-10 * SIZE(B), %xmm15	movapd	 -8 * SIZE(B), %xmm1	movapd	 -6 * SIZE(B), %xmm3	movapd	 -4 * SIZE(B), %xmm5	movapd	 -2 * SIZE(B), %xmm7	subpd	%xmm8,  %xmm9	subpd	%xmm10,  %xmm11	subpd	%xmm0,  %xmm13	subpd	%xmm2, %xmm15	subpd	%xmm12,  %xmm1	subpd	%xmm14,  %xmm3	subpd	%xmm4, %xmm5	subpd	%xmm6, %xmm7#else	movapd	 -16 * SIZE(AO), %xmm0	movapd	 -14 * SIZE(AO), %xmm1	movapd	 -12 * SIZE(AO), %xmm2	movapd	 -10 * SIZE(AO), %xmm3	movapd	  -8 * SIZE(AO), %xmm4	movapd	  -6 * SIZE(AO), %xmm5	movapd	  -4 * SIZE(AO), %xmm6	movapd	  -2 * SIZE(AO), %xmm7	subpd	%xmm8, %xmm0	subpd	%xmm12, %xmm1	subpd	%xmm9, %xmm2	subpd	%xmm13, %xmm3	subpd	%xmm10, %xmm4	subpd	%xmm14, %xmm5	subpd	%xmm11, %xmm6	subpd	%xmm15, %xmm7#endif#ifdef LN	movddup	-1 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm5	mulpd	 %xmm8, %xmm7	movddup	-2 * SIZE(AO), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm5, %xmm10	subpd	 %xmm10, %xmm1	mulpd	 %xmm7, %xmm12	subpd	 %xmm12, %xmm3	movddup	-3 * SIZE(AO), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm5, %xmm10	subpd	 %xmm10, %xmm13	mulpd	 %xmm7, %xmm12	subpd	 %xmm12, %xmm15	movddup	-4 * SIZE(AO), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm5, %xmm10	subpd	 %xmm10, %xmm9	mulpd	 %xmm7, %xmm12	subpd	 %xmm12, %xmm11	movddup	-6 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm1	mulpd	 %xmm8, %xmm3	movddup	-7 * SIZE(AO), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm1, %xmm10	subpd	 %xmm10, %xmm13	mulpd	 %xmm3, %xmm12	subpd	 %xmm12, %xmm15	movddup	-8 * SIZE(AO), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm1, %xmm10	subpd	 %xmm10, %xmm9	mulpd	 %xmm3, %xmm12	subpd	 %xmm12, %xmm11	movddup	-11 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm13	mulpd	 %xmm8, %xmm15	movddup	-12 * SIZE(AO), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm13, %xmm10	subpd	 %xmm10, %xmm9	mulpd	 %xmm15, %xmm12	subpd	 %xmm12, %xmm11	movddup	-16 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm9	mulpd	 %xmm8, %xmm11#endif#ifdef LT	movddup	-16 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm9	mulpd	 %xmm8, %xmm11	movddup	-15 * SIZE(AO), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm9, %xmm10	subpd	 %xmm10, %xmm13	mulpd	 %xmm11, %xmm12	subpd	 %xmm12, %xmm15	movddup	-14 * SIZE(AO), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm9, %xmm10	subpd	 %xmm10, %xmm1	mulpd	 %xmm11, %xmm12	subpd	 %xmm12, %xmm3	movddup	-13 * SIZE(AO), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm9, %xmm10	subpd	 %xmm10, %xmm5	mulpd	 %xmm11, %xmm12	subpd	 %xmm12, %xmm7	movddup	-11 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm13	mulpd	 %xmm8, %xmm15	movddup	-10 * SIZE(AO), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm13, %xmm10	subpd	 %xmm10, %xmm1	mulpd	 %xmm15, %xmm12	subpd	 %xmm12, %xmm3	movddup	 -9 * SIZE(AO), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm13, %xmm10	subpd	 %xmm10, %xmm5	mulpd	 %xmm15, %xmm12	subpd	 %xmm12, %xmm7	movddup	 -6 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm1	mulpd	 %xmm8, %xmm3	movddup	 -5 * SIZE(AO), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm1, %xmm10	subpd	 %xmm10, %xmm5	mulpd	 %xmm3, %xmm12	subpd	 %xmm12, %xmm7	movddup	 -1 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm5	mulpd	 %xmm8, %xmm7#endif#ifdef RN	movddup	-16 * SIZE(B), %xmm8	mulpd	 %xmm8, %xmm0	mulpd	 %xmm8, %xmm1	movddup	 -15 * SIZE(B), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm0, %xmm10	subpd	 %xmm10, %xmm2	mulpd	 %xmm1, %xmm12	subpd	 %xmm12, %xmm3	movddup	 -14 * SIZE(B), %xmm10	movapd	%xmm10, %xmm12	mulpd	 %xmm0, %xmm10	subpd	 %xmm10, %xmm4	mulpd	 %xmm1, %xmm12
⌨️ 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 + -