📄 zgemv_t_sse2_core2.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 4 页
字号:
	movapd	-12 * SIZE(BO),  %xmm12	movapd	-14 * SIZE(AO1), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm13, %xmm8	addpd	%xmm8,  %xmm2	mulpd	%xmm13, %xmm9	ADD	%xmm9,  %xmm3	movapd	-10 * SIZE(BO),  %xmm13	subq	$-4 * SIZE, AO1	subq	$-4 * SIZE, BO	ALIGN_2.L46:	movq	MIN_N, I	andq	$1,   I	jle	.L47	movapd	-16 * SIZE(AO1), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm12, %xmm8	addpd	%xmm8,  %xmm0	mulpd	%xmm12, %xmm9	ADD	%xmm9,  %xmm1	ALIGN_3.L47:	addpd	%xmm2, %xmm0	addpd	%xmm3, %xmm1#if (!defined(CONJ) && !defined(XCONJ)) || \    ( defined(CONJ) &&  defined(XCONJ))	pxor	BETA, %xmm0#else	pxor	BETA, %xmm1#endif	haddpd	%xmm1, %xmm0	pshufd	  $0x4e, %xmm0, %xmm1	mulpd	  ALPHA, %xmm0	mulpd	  ALPHA, %xmm1	xorpd	  BETA, %xmm0	haddpd	%xmm1, %xmm0	movsd	 0 * SIZE(CO), %xmm8	movhpd	 1 * SIZE(CO), %xmm8	addpd	%xmm8, %xmm0	movsd	 %xmm0, 0 * SIZE(CO)	movhpd	 %xmm0, 1 * SIZE(CO)	ALIGN_3.L99:	addq	NLDA, A	addq	$P, IS	cmpq	M,  IS	jl	.L10	movq	  0(%rsp), %rbx	movq	  8(%rsp), %rbp	movq	 16(%rsp), %r12	movq	 24(%rsp), %r13	movq	 32(%rsp), %r14	movq	 40(%rsp), %r15#ifdef WINDOWS_ABI	movq	 48(%rsp), %rdi	movq	 56(%rsp), %rsi	movups	 64(%rsp), %xmm6	movups	 80(%rsp), %xmm7	movups	 96(%rsp), %xmm8	movups	112(%rsp), %xmm9	movups	128(%rsp), %xmm10	movups	144(%rsp), %xmm11	movups	160(%rsp), %xmm12	movups	176(%rsp), %xmm13	movups	192(%rsp), %xmm14	movups	208(%rsp), %xmm15#endif	addq	$STACKSIZE, %rsp	ret	ALIGN_3.L100:	movq	N,   J	sarq	$2,  J	jle	.L130	ALIGN_3	.L121:	movq	A,  AO1	leaq	(A, LDA, 1), AO2	leaq	(A, LDA, 4), A	movq	BUFFER, BO	addq	$16 * SIZE, BO	movapd	-16 * SIZE(BO),  %xmm12	movapd	-14 * SIZE(BO),  %xmm13	pxor	%xmm0, %xmm0	pxor	%xmm1, %xmm1	pxor	%xmm2, %xmm2	pxor	%xmm3, %xmm3	pxor	%xmm4, %xmm4	pxor	%xmm5, %xmm5	pxor	%xmm6, %xmm6	pxor	%xmm7, %xmm7	movq	MIN_N, I	sarq	$3,   I	jle	.L124	ALIGN_3.L122:	PREFETCH	(PREFETCHSIZE + 0) * SIZE(AO1)	movsd	-16 * SIZE(AO1), %xmm8	movhpd	-15 * SIZE(AO1), %xmm8	movsd	-16 * SIZE(AO2), %xmm10	movhpd	-15 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm8	mulpd	%xmm12, %xmm9	mulpd	%xmm12, %xmm10	mulpd	%xmm12, %xmm11	addpd	%xmm8,  %xmm0	ADD	%xmm9,  %xmm1	addpd	%xmm10, %xmm2	ADD	%xmm11, %xmm3	movsd	-16 * SIZE(AO1, LDA, 2), %xmm8	movhpd	-15 * SIZE(AO1, LDA, 2), %xmm8	movsd	-16 * SIZE(AO2, LDA, 2), %xmm10	movhpd	-15 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm8	mulpd	%xmm12, %xmm9	mulpd	%xmm12, %xmm10	mulpd	%xmm12, %xmm11	movapd	-12 * SIZE(BO),  %xmm12	addpd	%xmm8,  %xmm4	ADD	%xmm9,  %xmm5	addpd	%xmm10, %xmm6	ADD	%xmm11, %xmm7	PREFETCH	(PREFETCHSIZE + 0) * SIZE(AO2)	movsd	-14 * SIZE(AO1), %xmm8	movhpd	-13 * SIZE(AO1), %xmm8	movsd	-14 * SIZE(AO2), %xmm10	movhpd	-13 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm8	mulpd	%xmm13, %xmm9	mulpd	%xmm13, %xmm10	mulpd	%xmm13, %xmm11	addpd	%xmm8,  %xmm0	ADD	%xmm9,  %xmm1	addpd	%xmm10, %xmm2	ADD	%xmm11, %xmm3	movsd	-14 * SIZE(AO1, LDA, 2), %xmm8	movhpd	-13 * SIZE(AO1, LDA, 2), %xmm8	movsd	-14 * SIZE(AO2, LDA, 2), %xmm10	movhpd	-13 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm8	mulpd	%xmm13, %xmm9	mulpd	%xmm13, %xmm10	mulpd	%xmm13, %xmm11	movapd	-10 * SIZE(BO),  %xmm13	addpd	%xmm8,  %xmm4	ADD	%xmm9,  %xmm5	addpd	%xmm10, %xmm6	ADD	%xmm11, %xmm7	PREFETCH	(PREFETCHSIZE + 0) * SIZE(AO1, LDA, 2)	movsd	-12 * SIZE(AO1), %xmm8	movhpd	-11 * SIZE(AO1), %xmm8	movsd	-12 * SIZE(AO2), %xmm10	movhpd	-11 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm8	mulpd	%xmm12, %xmm9	mulpd	%xmm12, %xmm10	mulpd	%xmm12, %xmm11	addpd	%xmm8,  %xmm0	ADD	%xmm9,  %xmm1	addpd	%xmm10, %xmm2	ADD	%xmm11, %xmm3	movsd	-12 * SIZE(AO1, LDA, 2), %xmm8	movhpd	-11 * SIZE(AO1, LDA, 2), %xmm8	movsd	-12 * SIZE(AO2, LDA, 2), %xmm10	movhpd	-11 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm8	mulpd	%xmm12, %xmm9	mulpd	%xmm12, %xmm10	mulpd	%xmm12, %xmm11	movapd	 -8 * SIZE(BO),  %xmm12	addpd	%xmm8,  %xmm4	ADD	%xmm9,  %xmm5	addpd	%xmm10, %xmm6	ADD	%xmm11, %xmm7	PREFETCH	(PREFETCHSIZE + 0) * SIZE(AO2, LDA, 2)	movsd	-10 * SIZE(AO1), %xmm8	movhpd	 -9 * SIZE(AO1), %xmm8	movsd	-10 * SIZE(AO2), %xmm10	movhpd	 -9 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm8	mulpd	%xmm13, %xmm9	mulpd	%xmm13, %xmm10	mulpd	%xmm13, %xmm11	addpd	%xmm8,  %xmm0	ADD	%xmm9,  %xmm1	addpd	%xmm10, %xmm2	ADD	%xmm11, %xmm3	movsd	-10 * SIZE(AO1, LDA, 2), %xmm8	movhpd	 -9 * SIZE(AO1, LDA, 2), %xmm8	movsd	-10 * SIZE(AO2, LDA, 2), %xmm10	movhpd	 -9 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm8	mulpd	%xmm13, %xmm9	mulpd	%xmm13, %xmm10	mulpd	%xmm13, %xmm11	movapd	 -6 * SIZE(BO),  %xmm13	addpd	%xmm8,  %xmm4	ADD	%xmm9,  %xmm5	addpd	%xmm10, %xmm6	ADD	%xmm11, %xmm7	PREFETCH	(PREFETCHSIZE + 8) * SIZE(AO1)	movsd	 -8 * SIZE(AO1), %xmm8	movhpd	 -7 * SIZE(AO1), %xmm8	movsd	 -8 * SIZE(AO2), %xmm10	movhpd	 -7 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm8	mulpd	%xmm12, %xmm9	mulpd	%xmm12, %xmm10	mulpd	%xmm12, %xmm11	addpd	%xmm8,  %xmm0	ADD	%xmm9,  %xmm1	addpd	%xmm10, %xmm2	ADD	%xmm11, %xmm3	movsd	 -8 * SIZE(AO1, LDA, 2), %xmm8	movhpd	 -7 * SIZE(AO1, LDA, 2), %xmm8	movsd	 -8 * SIZE(AO2, LDA, 2), %xmm10	movhpd	 -7 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm8	mulpd	%xmm12, %xmm9	mulpd	%xmm12, %xmm10	mulpd	%xmm12, %xmm11	movapd	 -4 * SIZE(BO),  %xmm12	addpd	%xmm8,  %xmm4	ADD	%xmm9,  %xmm5	addpd	%xmm10, %xmm6	ADD	%xmm11, %xmm7	PREFETCH	(PREFETCHSIZE + 8) * SIZE(AO2)	movsd	 -6 * SIZE(AO1), %xmm8	movhpd	 -5 * SIZE(AO1), %xmm8	movsd	 -6 * SIZE(AO2), %xmm10	movhpd	 -5 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm8	mulpd	%xmm13, %xmm9	mulpd	%xmm13, %xmm10	mulpd	%xmm13, %xmm11	addpd	%xmm8,  %xmm0	ADD	%xmm9,  %xmm1	addpd	%xmm10, %xmm2	ADD	%xmm11, %xmm3	movsd	 -6 * SIZE(AO1, LDA, 2), %xmm8	movhpd	 -5 * SIZE(AO1, LDA, 2), %xmm8	movsd	 -6 * SIZE(AO2, LDA, 2), %xmm10	movhpd	 -5 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm8	mulpd	%xmm13, %xmm9	mulpd	%xmm13, %xmm10	mulpd	%xmm13, %xmm11	movapd	 -2 * SIZE(BO),  %xmm13	addpd	%xmm8,  %xmm4	ADD	%xmm9,  %xmm5	addpd	%xmm10, %xmm6	ADD	%xmm11, %xmm7	PREFETCH	(PREFETCHSIZE + 8) * SIZE(AO1, LDA, 2)	movsd	 -4 * SIZE(AO1), %xmm8	movhpd	 -3 * SIZE(AO1), %xmm8	movsd	 -4 * SIZE(AO2), %xmm10	movhpd	 -3 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm8	mulpd	%xmm12, %xmm9	mulpd	%xmm12, %xmm10	mulpd	%xmm12, %xmm11	addpd	%xmm8,  %xmm0	ADD	%xmm9,  %xmm1	addpd	%xmm10, %xmm2	ADD	%xmm11, %xmm3	movsd	 -4 * SIZE(AO1, LDA, 2), %xmm8	movhpd	 -3 * SIZE(AO1, LDA, 2), %xmm8	movsd	 -4 * SIZE(AO2, LDA, 2), %xmm10	movhpd	 -3 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm8	mulpd	%xmm12, %xmm9	mulpd	%xmm12, %xmm10	mulpd	%xmm12, %xmm11	movapd	  0 * SIZE(BO),  %xmm12	addpd	%xmm8,  %xmm4	ADD	%xmm9,  %xmm5	addpd	%xmm10, %xmm6	ADD	%xmm11, %xmm7	PREFETCH	(PREFETCHSIZE + 8) * SIZE(AO2, LDA, 2)	movsd	 -2 * SIZE(AO1), %xmm8	movhpd	 -1 * SIZE(AO1), %xmm8	movsd	 -2 * SIZE(AO2), %xmm10	movhpd	 -1 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm8	mulpd	%xmm13, %xmm9	mulpd	%xmm13, %xmm10	mulpd	%xmm13, %xmm11	addpd	%xmm8,  %xmm0	ADD	%xmm9,  %xmm1	addpd	%xmm10, %xmm2	ADD	%xmm11, %xmm3	movsd	 -2 * SIZE(AO1, LDA, 2), %xmm8	movhpd	 -1 * SIZE(AO1, LDA, 2), %xmm8	movsd	 -2 * SIZE(AO2, LDA, 2), %xmm10	movhpd	 -1 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm8, %xmm9	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm8	mulpd	%xmm13, %xmm9	mulpd	%xmm13, %xmm10	mulpd	%xmm13, %xmm11	movapd	  2 * SIZE(BO),  %xmm13	addpd	%xmm8,  %xmm4	ADD	%xmm9,  %xmm5	addpd	%xmm10, %xmm6	ADD	%xmm11, %xmm7	subq	$-16 * SIZE, AO1	subq	$-16 * SIZE, AO2	subq	$-16 * SIZE, BO	decq	I	jg	.L122	ALIGN_3.L124:	movq	MIN_N, I	andq	$4,   I	jle	.L125	movsd	-16 * SIZE(AO1), %xmm8	movhpd	-15 * SIZE(AO1), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm12, %xmm8	addpd	%xmm8,  %xmm0	mulpd	%xmm12, %xmm9	ADD	%xmm9,  %xmm1	movsd	-16 * SIZE(AO2), %xmm10	movhpd	-15 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm10	addpd	%xmm10,  %xmm2	mulpd	%xmm12, %xmm11	ADD	%xmm11,  %xmm3	movsd	-16 * SIZE(AO1, LDA, 2), %xmm8	movhpd	-15 * SIZE(AO1, LDA, 2), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm12, %xmm8	addpd	%xmm8,  %xmm4	mulpd	%xmm12, %xmm9	ADD	%xmm9,  %xmm5	movsd	-16 * SIZE(AO2, LDA, 2), %xmm10	movhpd	-15 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm10	addpd	%xmm10, %xmm6	mulpd	%xmm12, %xmm11	ADD	%xmm11, %xmm7	movapd	-12 * SIZE(BO),  %xmm12	movsd	-14 * SIZE(AO1), %xmm8	movhpd	-13 * SIZE(AO1), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm13, %xmm8	addpd	%xmm8,  %xmm0	mulpd	%xmm13, %xmm9	ADD	%xmm9,  %xmm1	movsd	-14 * SIZE(AO2), %xmm10	movhpd	-13 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm10	addpd	%xmm10,  %xmm2	mulpd	%xmm13, %xmm11	ADD	%xmm11,  %xmm3	movsd	-14 * SIZE(AO1, LDA, 2), %xmm8	movhpd	-13 * SIZE(AO1, LDA, 2), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm13, %xmm8	addpd	%xmm8,  %xmm4	mulpd	%xmm13, %xmm9	ADD	%xmm9,  %xmm5	movsd	-14 * SIZE(AO2, LDA, 2), %xmm10	movhpd	-13 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm10	addpd	%xmm10, %xmm6	mulpd	%xmm13, %xmm11	ADD	%xmm11, %xmm7	movapd	-10 * SIZE(BO),  %xmm13	movsd	-12 * SIZE(AO1), %xmm8	movhpd	-11 * SIZE(AO1), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm12, %xmm8	addpd	%xmm8,  %xmm0	mulpd	%xmm12, %xmm9	ADD	%xmm9,  %xmm1	movsd	-12 * SIZE(AO2), %xmm10	movhpd	-11 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm10	addpd	%xmm10,  %xmm2	mulpd	%xmm12, %xmm11	ADD	%xmm11,  %xmm3	movsd	-12 * SIZE(AO1, LDA, 2), %xmm8	movhpd	-11 * SIZE(AO1, LDA, 2), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm12, %xmm8	addpd	%xmm8,  %xmm4	mulpd	%xmm12, %xmm9	ADD	%xmm9,  %xmm5	movsd	-12 * SIZE(AO2, LDA, 2), %xmm10	movhpd	-11 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm10	addpd	%xmm10, %xmm6	mulpd	%xmm12, %xmm11	ADD	%xmm11, %xmm7	movapd	 -8 * SIZE(BO),  %xmm12	movsd	-10 * SIZE(AO1), %xmm8	movhpd	 -9 * SIZE(AO1), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm13, %xmm8	addpd	%xmm8,  %xmm0	mulpd	%xmm13, %xmm9	ADD	%xmm9,  %xmm1	movsd	-10 * SIZE(AO2), %xmm10	movhpd	 -9 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm10	addpd	%xmm10,  %xmm2	mulpd	%xmm13, %xmm11	ADD	%xmm11,  %xmm3	movsd	-10 * SIZE(AO1, LDA, 2), %xmm8	movhpd	 -9 * SIZE(AO1, LDA, 2), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm13, %xmm8	addpd	%xmm8,  %xmm4	mulpd	%xmm13, %xmm9	ADD	%xmm9,  %xmm5	movsd	-10 * SIZE(AO2, LDA, 2), %xmm10	movhpd	 -9 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm10	addpd	%xmm10, %xmm6	mulpd	%xmm13, %xmm11	ADD	%xmm11, %xmm7	movapd	 -6 * SIZE(BO),  %xmm13	subq	$-8 * SIZE, AO1	subq	$-8 * SIZE, AO2	subq	$-8 * SIZE, BO	ALIGN_3.L125:	movq	MIN_N, I	andq	$2,   I	jle	.L126	movsd	-16 * SIZE(AO1), %xmm8	movhpd	-15 * SIZE(AO1), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm12, %xmm8	addpd	%xmm8,  %xmm0	mulpd	%xmm12, %xmm9	ADD	%xmm9,  %xmm1	movsd	-16 * SIZE(AO2), %xmm10	movhpd	-15 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm10	addpd	%xmm10,  %xmm2	mulpd	%xmm12, %xmm11	ADD	%xmm11,  %xmm3	movsd	-16 * SIZE(AO1, LDA, 2), %xmm8	movhpd	-15 * SIZE(AO1, LDA, 2), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm12, %xmm8	addpd	%xmm8,  %xmm4	mulpd	%xmm12, %xmm9	ADD	%xmm9,  %xmm5	movsd	-16 * SIZE(AO2, LDA, 2), %xmm10	movhpd	-15 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm10	addpd	%xmm10, %xmm6	mulpd	%xmm12, %xmm11	ADD	%xmm11, %xmm7	movapd	-12 * SIZE(BO),  %xmm12	movsd	-14 * SIZE(AO1), %xmm8	movhpd	-13 * SIZE(AO1), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm13, %xmm8	addpd	%xmm8,  %xmm0	mulpd	%xmm13, %xmm9	ADD	%xmm9,  %xmm1	movsd	-14 * SIZE(AO2), %xmm10	movhpd	-13 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm10	addpd	%xmm10,  %xmm2	mulpd	%xmm13, %xmm11	ADD	%xmm11,  %xmm3	movsd	-14 * SIZE(AO1, LDA, 2), %xmm8	movhpd	-13 * SIZE(AO1, LDA, 2), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm13, %xmm8	addpd	%xmm8,  %xmm4	mulpd	%xmm13, %xmm9	ADD	%xmm9,  %xmm5	movsd	-14 * SIZE(AO2, LDA, 2), %xmm10	movhpd	-13 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm13, %xmm10	addpd	%xmm10, %xmm6	mulpd	%xmm13, %xmm11	ADD	%xmm11, %xmm7	movapd	-10 * SIZE(BO),  %xmm13	subq	$-4 * SIZE, AO1	subq	$-4 * SIZE, AO2	subq	$-4 * SIZE, BO	ALIGN_2.L126:	movq	MIN_N, I	andq	$1,   I	jle	.L127	movsd	-16 * SIZE(AO1), %xmm8	movhpd	-15 * SIZE(AO1), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm12, %xmm8	addpd	%xmm8,  %xmm0	mulpd	%xmm12, %xmm9	ADD	%xmm9,  %xmm1	movsd	-16 * SIZE(AO2), %xmm10	movhpd	-15 * SIZE(AO2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm10	addpd	%xmm10,  %xmm2	mulpd	%xmm12, %xmm11	ADD	%xmm11,  %xmm3	movsd	-16 * SIZE(AO1, LDA, 2), %xmm8	movhpd	-15 * SIZE(AO1, LDA, 2), %xmm8	pshufd	  $0x4e, %xmm8, %xmm9	mulpd	%xmm12, %xmm8	addpd	%xmm8,  %xmm4	mulpd	%xmm12, %xmm9	ADD	%xmm9,  %xmm5	movsd	-16 * SIZE(AO2, LDA, 2), %xmm10	movhpd	-15 * SIZE(AO2, LDA, 2), %xmm10	pshufd	  $0x4e, %xmm10, %xmm11	mulpd	%xmm12, %xmm10	addpd	%xmm10, %xmm6	mulpd	%xmm12, %xmm11	ADD	%xmm11, %xmm7	ALIGN_3.L127:#if (!defined(CONJ) && !defined(XCONJ)) || \    ( defined(CONJ) &&  defined(XCONJ))	pxor	BETA, %xmm0	pxor	BETA, %xmm2	pxor	BETA, %xmm4	pxor	BETA, %xmm6#else	pxor	BETA, %xmm1	pxor	BETA, %xmm3	pxor	BETA, %xmm5	pxor	BETA, %xmm7#endif	haddpd	%xmm1, %xmm0	haddpd	%xmm3, %xmm2	haddpd	%xmm5, %xmm4	haddpd	%xmm7, %xmm6	pshufd	  $0x4e, %xmm0, %xmm1	pshufd	  $0x4e, %xmm2, %xmm3	pshufd	  $0x4e, %xmm4, %xmm5	pshufd	  $0x4e, %xmm6, %xmm7
⌨️ 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 + -