📄 zgemm3m_kernel_4x4_sse3.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 4 页
字号:
	mulpd	%xmm8, %xmm9	addpd	%xmm9, %xmm5	movddup	 6 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	addpd	%xmm9, %xmm6	movddup	 7 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	movapd	32 * SIZE(AO), %xmm8	addpd	%xmm9, %xmm7	movddup	32 * SIZE(BO), %xmm9	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm0	movddup	 9 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm1	movddup	10 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm2	movddup	11 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	movapd	10 * SIZE(AO), %xmm10	addpd	%xmm11, %xmm3	movddup	 8 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm4	movddup	 9 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm5	movddup	10 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm6	movddup	11 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	movapd	12 * SIZE(AO), %xmm10	addpd	%xmm11, %xmm7	movddup	12 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm0	movddup	13 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm1	movddup	14 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm2	movddup	15 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	movapd	14 * SIZE(AO), %xmm10	addpd	%xmm11, %xmm3	movddup	12 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm4	movddup	13 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm5	movddup	14 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm6	movddup	15 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	movapd	40 * SIZE(AO), %xmm10	addpd	%xmm11, %xmm7	movddup	40 * SIZE(BO), %xmm11	mulpd	%xmm12, %xmm13	PREFETCH  (PREFETCHSIZE + 16) * SIZE(AO)	addpd	%xmm13, %xmm0	movddup	17 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	addpd	%xmm13, %xmm1	movddup	18 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	addpd	%xmm13, %xmm2	movddup	19 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	movapd	18 * SIZE(AO), %xmm12	addpd	%xmm13, %xmm3	movddup	16 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	addpd	%xmm13, %xmm4	movddup	17 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	addpd	%xmm13, %xmm5	movddup	18 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	addpd	%xmm13, %xmm6	movddup	19 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	movapd	20 * SIZE(AO), %xmm12	addpd	%xmm13, %xmm7	movddup	20 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	addpd	%xmm13, %xmm0	movddup	21 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	addpd	%xmm13, %xmm1	movddup	22 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	addpd	%xmm13, %xmm2	movddup	23 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	movapd	22 * SIZE(AO), %xmm12	addpd	%xmm13, %xmm3	movddup	20 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	addpd	%xmm13, %xmm4	movddup 21 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	addpd	%xmm13, %xmm5	movddup	22 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	addpd	%xmm13, %xmm6	movddup	23 * SIZE(BO), %xmm13	mulpd	%xmm12, %xmm13	movapd	48 * SIZE(AO), %xmm12	addpd	%xmm13, %xmm7	movddup	48 * SIZE(BO), %xmm13	mulpd	%xmm14, %xmm15	addpd	%xmm15, %xmm0	movddup	25 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	addpd	%xmm15, %xmm1	movddup	26 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	addpd	%xmm15, %xmm2	movddup	27 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	movapd	26 * SIZE(AO), %xmm14	addpd	%xmm15, %xmm3	movddup	24 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	addpd	%xmm15, %xmm4	movddup	25 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	addpd	%xmm15, %xmm5	movddup	26 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	addpd	%xmm15, %xmm6	movddup	27 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	movapd	28 * SIZE(AO), %xmm14	addpd	%xmm15, %xmm7	movddup	28 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	addpd	%xmm15, %xmm0	movddup	29 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	addpd	%xmm15, %xmm1	movddup	30 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	addpd	%xmm15, %xmm2	movddup	31 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	movapd	30 * SIZE(AO), %xmm14	addpd	%xmm15, %xmm3	movddup	28 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	addpd	%xmm15, %xmm4	movddup	29 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	addpd	%xmm15, %xmm5	movddup	30 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	addpd	%xmm15, %xmm6	movddup	31 * SIZE(BO), %xmm15	mulpd	%xmm14, %xmm15	movapd	56 * SIZE(AO), %xmm14	addpd	%xmm15, %xmm7	movddup	56 * SIZE(BO), %xmm15	addq   $32 * SIZE, BO	addq   $32 * SIZE, AO	decq   %rax	BRANCH	jne    .L12#endif	ALIGN_4.L15:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movsd	ALPHA_R, %xmm15	movhpd	ALPHA_I, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je,pt	.L19	ALIGN_4.L16:	mulpd	%xmm8, %xmm9	movapd	 2 * SIZE(AO), %xmm10	addpd	%xmm9, %xmm0	movddup	 1 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	movddup	 0 * SIZE(BO), %xmm11	addpd	%xmm9, %xmm1	movddup	 2 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	addpd	%xmm9, %xmm2	movddup	 3 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	movapd	 4 * SIZE(AO), %xmm8	addpd	%xmm9, %xmm3	movddup	 4 * SIZE(BO), %xmm9	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm4	movddup	 1 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm5	movddup	 2 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm6	movddup	 3 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm7	addq	$4 * SIZE, AO		# aoffset  += 4	addq	$4 * SIZE, BO		# boffset1 += 8	decq	%rax	jg,pt	.L16	ALIGN_4.L19:	movsd	0 * SIZE(CO1), %xmm8	movhpd	1 * SIZE(CO1), %xmm8	movsd	2 * SIZE(CO1), %xmm9	movhpd	3 * SIZE(CO1), %xmm9	movsd	4 * SIZE(CO1), %xmm10	movhpd	5 * SIZE(CO1), %xmm10	movsd	6 * SIZE(CO1), %xmm11	movhpd	7 * SIZE(CO1), %xmm11	movddup	 %xmm0, %xmm12	unpckhpd %xmm0, %xmm0	movddup	 %xmm4, %xmm13	unpckhpd %xmm4, %xmm4	mulpd	 %xmm15, %xmm12	mulpd	 %xmm15, %xmm0	mulpd	 %xmm15, %xmm13	mulpd	 %xmm15, %xmm4	addpd	 %xmm12, %xmm8	addpd	 %xmm0,  %xmm9	addpd	 %xmm13, %xmm10	addpd	 %xmm4,  %xmm11	movsd	%xmm8,  0 * SIZE(CO1)	movhpd	%xmm8,  1 * SIZE(CO1)	movsd	%xmm9,  2 * SIZE(CO1)	movhpd	%xmm9,  3 * SIZE(CO1)	movsd	%xmm10, 4 * SIZE(CO1)	movhpd	%xmm10, 5 * SIZE(CO1)	movsd	%xmm11, 6 * SIZE(CO1)	movhpd	%xmm11, 7 * SIZE(CO1)		movsd	0 * SIZE(CO2), %xmm8	movhpd	1 * SIZE(CO2), %xmm8	movsd	2 * SIZE(CO2), %xmm9	movhpd	3 * SIZE(CO2), %xmm9	movsd	4 * SIZE(CO2), %xmm10	movhpd	5 * SIZE(CO2), %xmm10	movsd	6 * SIZE(CO2), %xmm11	movhpd	7 * SIZE(CO2), %xmm11	movddup	 %xmm1, %xmm12	unpckhpd %xmm1, %xmm1	movddup	 %xmm5, %xmm13	unpckhpd %xmm5, %xmm5	mulpd	 %xmm15, %xmm12	mulpd	 %xmm15, %xmm1	mulpd	 %xmm15, %xmm13	mulpd	 %xmm15, %xmm5	addpd	 %xmm12, %xmm8	addpd	 %xmm1,  %xmm9	addpd	 %xmm13, %xmm10	addpd	 %xmm5,  %xmm11	movsd	%xmm8,  0 * SIZE(CO2)	movhpd	%xmm8,  1 * SIZE(CO2)	movsd	%xmm9,  2 * SIZE(CO2)	movhpd	%xmm9,  3 * SIZE(CO2)	movsd	%xmm10, 4 * SIZE(CO2)	movhpd	%xmm10, 5 * SIZE(CO2)	movsd	%xmm11, 6 * SIZE(CO2)	movhpd	%xmm11, 7 * SIZE(CO2)	movsd	0 * SIZE(CO1, LDC, 2), %xmm8	movhpd	1 * SIZE(CO1, LDC, 2), %xmm8	movsd	2 * SIZE(CO1, LDC, 2), %xmm9	movhpd	3 * SIZE(CO1, LDC, 2), %xmm9	movsd	4 * SIZE(CO1, LDC, 2), %xmm10	movhpd	5 * SIZE(CO1, LDC, 2), %xmm10	movsd	6 * SIZE(CO1, LDC, 2), %xmm11	movhpd	7 * SIZE(CO1, LDC, 2), %xmm11	movddup	 %xmm2, %xmm12	unpckhpd %xmm2, %xmm2	movddup	 %xmm6, %xmm13	unpckhpd %xmm6, %xmm6	mulpd	 %xmm15, %xmm12	mulpd	 %xmm15, %xmm2	mulpd	 %xmm15, %xmm13	mulpd	 %xmm15, %xmm6	addpd	 %xmm12, %xmm8	addpd	 %xmm2,  %xmm9	addpd	 %xmm13, %xmm10	addpd	 %xmm6,  %xmm11	movsd	%xmm8,  0 * SIZE(CO1, LDC, 2)	movhpd	%xmm8,  1 * SIZE(CO1, LDC, 2)	movsd	%xmm9,  2 * SIZE(CO1, LDC, 2)	movhpd	%xmm9,  3 * SIZE(CO1, LDC, 2)	movsd	%xmm10, 4 * SIZE(CO1, LDC, 2)	movhpd	%xmm10, 5 * SIZE(CO1, LDC, 2)	movsd	%xmm11, 6 * SIZE(CO1, LDC, 2)	movhpd	%xmm11, 7 * SIZE(CO1, LDC, 2)		movsd	0 * SIZE(CO2, LDC, 2), %xmm8	movhpd	1 * SIZE(CO2, LDC, 2), %xmm8	movsd	2 * SIZE(CO2, LDC, 2), %xmm9	movhpd	3 * SIZE(CO2, LDC, 2), %xmm9	movsd	4 * SIZE(CO2, LDC, 2), %xmm10	movhpd	5 * SIZE(CO2, LDC, 2), %xmm10	movsd	6 * SIZE(CO2, LDC, 2), %xmm11	movhpd	7 * SIZE(CO2, LDC, 2), %xmm11	movddup	 %xmm3, %xmm12	unpckhpd %xmm3, %xmm3	movddup	 %xmm7, %xmm13	unpckhpd %xmm7, %xmm7	mulpd	 %xmm15, %xmm12	mulpd	 %xmm15, %xmm3	mulpd	 %xmm15, %xmm13	mulpd	 %xmm15, %xmm7	addpd	 %xmm12, %xmm8	addpd	 %xmm3,  %xmm9	addpd	 %xmm13, %xmm10	addpd	 %xmm7,  %xmm11	movsd	%xmm8,  0 * SIZE(CO2, LDC, 2)	movhpd	%xmm8,  1 * SIZE(CO2, LDC, 2)	movsd	%xmm9,  2 * SIZE(CO2, LDC, 2)	movhpd	%xmm9,  3 * SIZE(CO2, LDC, 2)	movsd	%xmm10, 4 * SIZE(CO2, LDC, 2)	movhpd	%xmm10, 5 * SIZE(CO2, LDC, 2)	movsd	%xmm11, 6 * SIZE(CO2, LDC, 2)	movhpd	%xmm11, 7 * SIZE(CO2, LDC, 2)	addq	$8 * SIZE, CO1		# coffset += 4	addq	$8 * SIZE, CO2		# coffset += 4	decq	I			# i --	jg	.L11	jmp	.L20	ALIGN_4.L20:	testq	$2, M	BRANCH	je	.L30	ALIGN_4.L21:#if !defined(TRMMKERNEL) || \	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	movq	B, BO#else	movq	KK, %rax	leaq	(, %rax, SIZE), %rax	leaq	(AO, %rax, 2), AO	leaq	(B,  %rax, 4), BO#endif		movapd	 0 * SIZE(AO), %xmm8	pxor	%xmm0, %xmm0	movddup	 0 * SIZE(BO), %xmm9	pxor	%xmm1, %xmm1	movapd	 8 * SIZE(AO), %xmm10	pxor	%xmm2, %xmm2	movddup	 8 * SIZE(BO), %xmm11	pxor	%xmm3, %xmm3#ifndef TRMMKERNEL	movq	K, %rax#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	movq	K, %rax	subq	KK, %rax	movq	%rax, KKK	#else	movq	KK, %rax#ifdef LEFT	addq	$2, %rax#else	addq	$4, %rax#endif	movq	%rax, KKK#endif	sarq	$3, %rax	je	.L25	ALIGN_4.L22:	mulpd	%xmm8, %xmm9	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)	addpd	%xmm9, %xmm0	movddup	 1 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	addpd	%xmm9, %xmm1	movddup	 2 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	addpd	%xmm9, %xmm2	movddup	 3 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	movapd	 2 * SIZE(AO), %xmm8	addpd	%xmm9, %xmm3	movddup	 4 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	addpd	%xmm9, %xmm0	movddup	 5 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	addpd	%xmm9, %xmm1	movddup	 6 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	addpd	%xmm9, %xmm2	movddup	 7 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	movapd	 4 * SIZE(AO), %xmm8	addpd	%xmm9, %xmm3	movddup	16 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm11	addpd	%xmm11, %xmm0	movddup	 9 * SIZE(BO), %xmm11	mulpd	%xmm8, %xmm11	addpd	%xmm11, %xmm1	movddup	10 * SIZE(BO), %xmm11	mulpd	%xmm8, %xmm11	addpd	%xmm11, %xmm2	movddup	11 * SIZE(BO), %xmm11	mulpd	%xmm8, %xmm11	movapd	 6 * SIZE(AO), %xmm8	addpd	%xmm11, %xmm3	movddup	12 * SIZE(BO), %xmm11	mulpd	%xmm8, %xmm11	addpd	%xmm11, %xmm0	movddup	13 * SIZE(BO), %xmm11	mulpd	%xmm8, %xmm11	addpd	%xmm11, %xmm1	movddup	14 * SIZE(BO), %xmm11	mulpd	%xmm8, %xmm11	addpd	%xmm11, %xmm2	movddup	15 * SIZE(BO), %xmm11	mulpd	%xmm8, %xmm11	movapd	16 * SIZE(AO), %xmm8	addpd	%xmm11, %xmm3	movddup	24 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm9	addpd	%xmm9, %xmm0	movddup	17 * SIZE(BO), %xmm9	mulpd	%xmm10, %xmm9	addpd	%xmm9, %xmm1	movddup	18 * SIZE(BO), %xmm9	mulpd	%xmm10, %xmm9	addpd	%xmm9, %xmm2	movddup	19 * SIZE(BO), %xmm9	mulpd	%xmm10, %xmm9	movapd	10 * SIZE(AO), %xmm10	addpd	%xmm9, %xmm3	movddup	20 * SIZE(BO), %xmm9	mulpd	%xmm10, %xmm9	addpd	%xmm9, %xmm0	movddup	21 * SIZE(BO), %xmm9	mulpd	%xmm10, %xmm9	addpd	%xmm9, %xmm1	movddup	22 * SIZE(BO), %xmm9	mulpd	%xmm10, %xmm9	addpd	%xmm9, %xmm2	movddup	23 * SIZE(BO), %xmm9	mulpd	%xmm10, %xmm9	movapd	12 * SIZE(AO), %xmm10	addpd	%xmm9, %xmm3	movddup	32 * SIZE(BO), %xmm9	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm0	movddup	25 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm1	movddup	26 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm2	movddup	27 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	movapd	14 * SIZE(AO), %xmm10	addpd	%xmm11, %xmm3	movddup	28 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm0	movddup	29 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm1	movddup	30 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	addpd	%xmm11, %xmm2	movddup	31 * SIZE(BO), %xmm11	mulpd	%xmm10, %xmm11	movapd	24 * SIZE(AO), %xmm10	addpd	%xmm11, %xmm3	movddup	40 * SIZE(BO), %xmm11	addq   $16 * SIZE, AO	addq   $32 * SIZE, BO	decq   %rax	jne    .L22	ALIGN_4.L25:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movsd	ALPHA_R, %xmm15	movhpd	ALPHA_I, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je .L29	ALIGN_4.L26:	mulpd	%xmm8, %xmm9	addpd	%xmm9, %xmm0	movddup	 1 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	addpd	%xmm9, %xmm1	movddup	 2 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	addpd	%xmm9, %xmm2	movddup	 3 * SIZE(BO), %xmm9	mulpd	%xmm8, %xmm9	movapd	 2 * SIZE(AO), %xmm8	addpd	%xmm9, %xmm3	movddup	 4 * SIZE(BO), %xmm9	addq	$2 * SIZE, AO		# aoffset  += 4	addq	$4 * SIZE, BO		# boffset1 += 8	decq	%rax	jg	.L26	ALIGN_4.L29:	movsd	0 * SIZE(CO1), %xmm8	movhpd	1 * SIZE(CO1), %xmm8	movsd	2 * SIZE(CO1), %xmm9	movhpd	3 * SIZE(CO1), %xmm9	movddup	 %xmm0, %xmm12	unpckhpd %xmm0, %xmm0	mulpd	 %xmm15, %xmm12	mulpd	 %xmm15, %xmm0	addpd	 %xmm12, %xmm8	addpd	 %xmm0,  %xmm9	movsd	%xmm8,  0 * SIZE(CO1)	movhpd	%xmm8,  1 * SIZE(CO1)	movsd	%xmm9,  2 * SIZE(CO1)	movhpd	%xmm9,  3 * SIZE(CO1)	movsd	0 * SIZE(CO2), %xmm8	movhpd	1 * SIZE(CO2), %xmm8	movsd	2 * SIZE(CO2), %xmm9	movhpd	3 * SIZE(CO2), %xmm9	movddup	 %xmm1, %xmm12	unpckhpd %xmm1, %xmm1	mulpd	 %xmm15, %xmm12	mulpd	 %xmm15, %xmm1	addpd	 %xmm12, %xmm8	addpd	 %xmm1,  %xmm9	movsd	%xmm8,  0 * SIZE(CO2)	movhpd	%xmm8,  1 * SIZE(CO2)	movsd	%xmm9,  2 * SIZE(CO2)	movhpd	%xmm9,  3 * SIZE(CO2)	movsd	0 * SIZE(CO1, LDC, 2), %xmm8	movhpd	1 * SIZE(CO1, LDC, 2), %xmm8	movsd	2 * SIZE(CO1, LDC, 2), %xmm9	movhpd	3 * SIZE(CO1, LDC, 2), %xmm9	movddup	 %xmm2, %xmm12	unpckhpd %xmm2, %xmm2	mulpd	 %xmm15, %xmm12	mulpd	 %xmm15, %xmm2	addpd	 %xmm12, %xmm8	addpd	 %xmm2,  %xmm9	movsd	%xmm8,  0 * SIZE(CO1, LDC, 2)	movhpd	%xmm8,  1 * SIZE(CO1, LDC, 2)	movsd	%xmm9,  2 * SIZE(CO1, LDC, 2)	movhpd	%xmm9,  3 * SIZE(CO1, LDC, 2)		movsd	0 * SIZE(CO2, LDC, 2), %xmm8	movhpd	1 * SIZE(CO2, LDC, 2), %xmm8	movsd	2 * SIZE(CO2, LDC, 2), %xmm9	movhpd	3 * SIZE(CO2, LDC, 2), %xmm9	movddup	 %xmm3, %xmm12	unpckhpd %xmm3, %xmm3	mulpd	 %xmm15, %xmm12	mulpd	 %xmm15, %xmm3	addpd	 %xmm12, %xmm8	addpd	 %xmm3,  %xmm9	movsd	%xmm8,  0 * SIZE(CO2, LDC, 2)	movhpd	%xmm8,  1 * SIZE(CO2, LDC, 2)	movsd	%xmm9,  2 * SIZE(CO2, LDC, 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 + -