📄 gemv_n_sse2.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 4 页
字号:
上一页 1 2 34
	addpd	 %xmm13, %xmm5	movsd	10 * SIZE(A2), %xmm13	movhpd	11 * SIZE(A2), %xmm13	mulpd	 %xmm0, %xmm9	addpd	 %xmm14, %xmm6	movsd	12 * SIZE(A2), %xmm14	movhpd	13 * SIZE(A2), %xmm14	mulpd	 %xmm0, %xmm10	addpd	 %xmm15, %xmm7	movsd	14 * SIZE(A2), %xmm15	movhpd	15 * SIZE(A2), %xmm15	mulpd	 %xmm0, %xmm11	movapd	 %xmm4, 0 * SIZE(Y1)	movapd	 8 * SIZE(Y1), %xmm4	movapd	 %xmm5, 2 * SIZE(Y1)	movapd	10 * SIZE(Y1), %xmm5	movapd	 %xmm6, 4 * SIZE(Y1)	movapd	12 * SIZE(Y1), %xmm6	movapd	 %xmm7, 6 * SIZE(Y1)	movapd	14 * SIZE(Y1), %xmm7	PREFETCH	PREFETCHSIZE * SIZE(A2)	addpd	 %xmm8,  %xmm4	movapd	 16 * SIZE(A1), %xmm8	mulpd	 %xmm1, %xmm12	addpd	 %xmm9,  %xmm5	movapd	 18 * SIZE(A1), %xmm9	mulpd	 %xmm1, %xmm13	addpd	 %xmm10, %xmm6	movapd	 20 * SIZE(A1), %xmm10	mulpd	 %xmm1, %xmm14	addpd	 %xmm11, %xmm7	movapd	 22 * SIZE(A1), %xmm11	mulpd	 %xmm1, %xmm15	addpd	 %xmm12, %xmm4	movsd	 16 * SIZE(A2), %xmm12	movhpd	 17 * SIZE(A2), %xmm12	mulpd	 %xmm0, %xmm8	addpd	 %xmm13, %xmm5	movsd	 18 * SIZE(A2), %xmm13	movhpd	 19 * SIZE(A2), %xmm13	mulpd	 %xmm0, %xmm9	addpd	 %xmm14, %xmm6	movsd	 20 * SIZE(A2), %xmm14	movhpd	 21 * SIZE(A2), %xmm14	mulpd	 %xmm0, %xmm10	addpd	 %xmm15, %xmm7	movsd	 22 * SIZE(A2), %xmm15	movhpd	 23 * SIZE(A2), %xmm15	mulpd	 %xmm0, %xmm11	movapd	 %xmm4,  8 * SIZE(Y1)	movapd	16 * SIZE(Y1), %xmm4	movapd	 %xmm5, 10 * SIZE(Y1)	movapd	18 * SIZE(Y1), %xmm5	movapd	 %xmm6, 12 * SIZE(Y1)	movapd	20 * SIZE(Y1), %xmm6	movapd	 %xmm7, 14 * SIZE(Y1)	movapd	22 * SIZE(Y1), %xmm7	addq	 $16 * SIZE, A1	addq	 $16 * SIZE, A2	addq	 $16 * SIZE, Y1	decq	 I	jg	.L53	ALIGN_3.L54:	addpd	 %xmm8,  %xmm4	movapd	 8 * SIZE(A1), %xmm8	mulpd	 %xmm1, %xmm12	addpd	 %xmm9,  %xmm5	movapd	10 * SIZE(A1), %xmm9	mulpd	 %xmm1, %xmm13	addpd	 %xmm10, %xmm6	movapd	12 * SIZE(A1), %xmm10	mulpd	 %xmm1, %xmm14	addpd	 %xmm11, %xmm7	movapd	14 * SIZE(A1), %xmm11	mulpd	 %xmm1, %xmm15	addpd	 %xmm12, %xmm4	movsd	 8 * SIZE(A2), %xmm12	movhpd	 9 * SIZE(A2), %xmm12	mulpd	 %xmm0, %xmm8	addpd	 %xmm13, %xmm5	movsd	10 * SIZE(A2), %xmm13	movhpd	11 * SIZE(A2), %xmm13	mulpd	 %xmm0, %xmm9	addpd	 %xmm14, %xmm6	movsd	12 * SIZE(A2), %xmm14	movhpd	13 * SIZE(A2), %xmm14	mulpd	 %xmm0, %xmm10	addpd	 %xmm15, %xmm7	movsd	14 * SIZE(A2), %xmm15	movhpd	15 * SIZE(A2), %xmm15	mulpd	 %xmm0, %xmm11	movapd	 %xmm4, 0 * SIZE(Y1)	movapd	 8 * SIZE(Y1), %xmm4	movapd	 %xmm5, 2 * SIZE(Y1)	movapd	10 * SIZE(Y1), %xmm5	movapd	 %xmm6, 4 * SIZE(Y1)	movapd	12 * SIZE(Y1), %xmm6	movapd	 %xmm7, 6 * SIZE(Y1)	movapd	14 * SIZE(Y1), %xmm7	addpd	 %xmm8,  %xmm4	mulpd	 %xmm1, %xmm12	addpd	 %xmm9,  %xmm5	mulpd	 %xmm1, %xmm13	addpd	 %xmm10, %xmm6	mulpd	 %xmm1, %xmm14	addpd	 %xmm11, %xmm7	mulpd	 %xmm1, %xmm15	addpd	 %xmm12, %xmm4	addpd	 %xmm13, %xmm5	addpd	 %xmm14, %xmm6	addpd	 %xmm15, %xmm7	movapd	 %xmm4,  8 * SIZE(Y1)	movapd	 %xmm5, 10 * SIZE(Y1)	movapd	 %xmm6, 12 * SIZE(Y1)	movapd	 %xmm7, 14 * SIZE(Y1)	addq	 $16 * SIZE, A1	addq	 $16 * SIZE, A2	addq	 $16 * SIZE, Y1	ALIGN_3.L55:	testq	$8, MM	je	.L56	movapd	 0 * SIZE(Y1), %xmm4	movapd	 2 * SIZE(Y1), %xmm5	movapd	 4 * SIZE(Y1), %xmm6	movapd	 6 * SIZE(Y1), %xmm7	movapd	 0 * SIZE(A1), %xmm8	movapd	 2 * SIZE(A1), %xmm9	movapd	 4 * SIZE(A1), %xmm10	movapd	 6 * SIZE(A1), %xmm11	movsd	 0 * SIZE(A2), %xmm12	movhpd	 1 * SIZE(A2), %xmm12	movsd	 2 * SIZE(A2), %xmm13	movhpd	 3 * SIZE(A2), %xmm13	movsd	 4 * SIZE(A2), %xmm14	movhpd	 5 * SIZE(A2), %xmm14	movsd	 6 * SIZE(A2), %xmm15	movhpd	 7 * SIZE(A2), %xmm15	mulpd	 %xmm0, %xmm8	mulpd	 %xmm0, %xmm9	mulpd	 %xmm0, %xmm10	mulpd	 %xmm0, %xmm11	addpd	 %xmm8,  %xmm4	mulpd	 %xmm1, %xmm12	addpd	 %xmm9,  %xmm5	mulpd	 %xmm1, %xmm13	addpd	 %xmm10, %xmm6	mulpd	 %xmm1, %xmm14	addpd	 %xmm11, %xmm7	mulpd	 %xmm1, %xmm15	addpd	 %xmm12, %xmm4	addpd	 %xmm13, %xmm5	addpd	 %xmm14, %xmm6	addpd	 %xmm15, %xmm7	movapd	 %xmm4, 0 * SIZE(Y1)	movapd	 %xmm5, 2 * SIZE(Y1)	movapd	 %xmm6, 4 * SIZE(Y1)	movapd	 %xmm7, 6 * SIZE(Y1)	addq	 $8 * SIZE, A1	addq	 $8 * SIZE, A2	addq	 $8 * SIZE, Y1	ALIGN_3.L56:	testq	$4, MM	je	.L57	movapd	 0 * SIZE(Y1), %xmm4	movapd	 2 * SIZE(Y1), %xmm5	movapd	 0 * SIZE(A1), %xmm8	movapd	 2 * SIZE(A1), %xmm9	movsd	 0 * SIZE(A2), %xmm10	movhpd	 1 * SIZE(A2), %xmm10	movsd	 2 * SIZE(A2), %xmm11	movhpd	 3 * SIZE(A2), %xmm11	mulpd	 %xmm0, %xmm8	mulpd	 %xmm0, %xmm9	mulpd	 %xmm1, %xmm10	mulpd	 %xmm1, %xmm11	addpd	 %xmm8,  %xmm4	addpd	 %xmm9,  %xmm5	addpd	 %xmm10, %xmm4	addpd	 %xmm11, %xmm5	movapd	 %xmm4, 0 * SIZE(Y1)	movapd	 %xmm5, 2 * SIZE(Y1)	addq	 $4 * SIZE, A1	addq	 $4 * SIZE, A2	addq	 $4 * SIZE, Y1	ALIGN_3.L57:	testq	$2, MM	je	.L58	movapd	 0 * SIZE(Y1), %xmm4	movapd	 0 * SIZE(A1), %xmm8	movsd	 0 * SIZE(A2), %xmm10	movhpd	 1 * SIZE(A2), %xmm10	mulpd	 %xmm0, %xmm8	mulpd	 %xmm1, %xmm10	addpd	 %xmm8,  %xmm4	addpd	 %xmm10, %xmm4	movapd	 %xmm4, 0 * SIZE(Y1)	addq	 $2 * SIZE, A1	addq	 $2 * SIZE, A2	addq	 $2 * SIZE, Y1	ALIGN_3.L58:	testq	$1, MM	je	.L30	movsd	 0 * SIZE(Y1), %xmm4	movsd	 0 * SIZE(A1), %xmm8	movsd	 0 * SIZE(A2), %xmm9	mulsd	 %xmm0, %xmm8	mulsd	 %xmm1, %xmm9	addsd	 %xmm8,  %xmm4	addsd	 %xmm9,  %xmm4	movsd	 %xmm4, 0 * SIZE(Y1)	addq	 $1 * SIZE, A1	addq	 $1 * SIZE, A2	addq	 $1 * SIZE, Y1	ALIGN_3.L60:	testq	$1, N	je	.L990	movq	BUFFER, Y1	movq	A,  A1	movsd	(X), %xmm0	mulsd	STACK_ALPHA, %xmm0	unpcklpd %xmm0, %xmm0	testq	$SIZE, A	je	.L62	movsd	 (Y1), %xmm4	movsd	 (A1), %xmm8	mulsd	 %xmm0, %xmm8	addsd	 %xmm8,  %xmm4	movsd	 %xmm4, (Y1)	addq	 $1 * SIZE, A1	addq	 $1 * SIZE, Y1	ALIGN_3.L62:	movq	MM,  I	sarq	$4, I	jle	.L65	movapd	 0 * SIZE(A1), %xmm8	movapd	 2 * SIZE(A1), %xmm9	movapd	 4 * SIZE(A1), %xmm10	movapd	 6 * SIZE(A1), %xmm11	movapd	 8 * SIZE(A1), %xmm12	mulpd	 %xmm0, %xmm8	movapd	10 * SIZE(A1), %xmm13	mulpd	 %xmm0, %xmm9	movapd	12 * SIZE(A1), %xmm14	mulpd	 %xmm0, %xmm10	movapd	14 * SIZE(A1), %xmm15	mulpd	 %xmm0, %xmm11	movapd	 0 * SIZE(Y1), %xmm4	movapd	 2 * SIZE(Y1), %xmm5	movapd	 4 * SIZE(Y1), %xmm6	movapd	 6 * SIZE(Y1), %xmm7	decq	 I	jle	 .L64	ALIGN_3.L63:	PREFETCH	PREFETCHSIZE * SIZE(A1)	addpd	 %xmm8,  %xmm4	movapd	16 * SIZE(A1), %xmm8	mulpd	 %xmm0, %xmm12	addpd	 %xmm9,  %xmm5	movapd	18 * SIZE(A1), %xmm9	mulpd	 %xmm0, %xmm13	addpd	 %xmm10, %xmm6	movapd	20 * SIZE(A1), %xmm10	mulpd	 %xmm0, %xmm14	addpd	 %xmm11, %xmm7	movapd	22 * SIZE(A1), %xmm11	mulpd	 %xmm0, %xmm15	movapd	 %xmm4, 0 * SIZE(Y1)	movapd	 8 * SIZE(Y1), %xmm4	movapd	 %xmm5, 2 * SIZE(Y1)	movapd	10 * SIZE(Y1), %xmm5	movapd	 %xmm6, 4 * SIZE(Y1)	movapd	12 * SIZE(Y1), %xmm6	movapd	 %xmm7, 6 * SIZE(Y1)	movapd	14 * SIZE(Y1), %xmm7	PREFETCH	PREFETCHSIZE * SIZE(Y1)	addpd	 %xmm12, %xmm4	movapd	24 * SIZE(A1), %xmm12	mulpd	 %xmm0, %xmm8	addpd	 %xmm13, %xmm5	movapd	26 * SIZE(A1), %xmm13	mulpd	 %xmm0, %xmm9	addpd	 %xmm14, %xmm6	movapd	28 * SIZE(A1), %xmm14	mulpd	 %xmm0, %xmm10	addpd	 %xmm15, %xmm7	movapd	30 * SIZE(A1), %xmm15	mulpd	 %xmm0, %xmm11	movapd	 %xmm4,  8 * SIZE(Y1)	movapd	16 * SIZE(Y1), %xmm4	movapd	 %xmm5, 10 * SIZE(Y1)	movapd	18 * SIZE(Y1), %xmm5	movapd	 %xmm6, 12 * SIZE(Y1)	movapd	20 * SIZE(Y1), %xmm6	movapd	 %xmm7, 14 * SIZE(Y1)	movapd	22 * SIZE(Y1), %xmm7	addq	 $16 * SIZE, A1	addq	 $16 * SIZE, Y1	decq	 I	jg	.L63	ALIGN_3.L64:	addpd	 %xmm8,  %xmm4	mulpd	 %xmm0, %xmm12	addpd	 %xmm9,  %xmm5	mulpd	 %xmm0, %xmm13	addpd	 %xmm10, %xmm6	mulpd	 %xmm0, %xmm14	addpd	 %xmm11, %xmm7	mulpd	 %xmm0, %xmm15	movapd	 %xmm4, 0 * SIZE(Y1)	movapd	 8 * SIZE(Y1), %xmm4	movapd	 %xmm5, 2 * SIZE(Y1)	movapd	10 * SIZE(Y1), %xmm5	movapd	 %xmm6, 4 * SIZE(Y1)	movapd	12 * SIZE(Y1), %xmm6	movapd	 %xmm7, 6 * SIZE(Y1)	movapd	14 * SIZE(Y1), %xmm7	addpd	 %xmm12, %xmm4	addpd	 %xmm13, %xmm5	addpd	 %xmm14, %xmm6	addpd	 %xmm15, %xmm7	movapd	 %xmm4,  8 * SIZE(Y1)	movapd	 %xmm5, 10 * SIZE(Y1)	movapd	 %xmm6, 12 * SIZE(Y1)	movapd	 %xmm7, 14 * SIZE(Y1)	addq	 $16 * SIZE, A1	addq	 $16 * SIZE, A2	addq	 $16 * SIZE, Y1	ALIGN_3.L65:	testq	$8, MM	je	.L66	movapd	 0 * SIZE(Y1), %xmm4	movapd	 2 * SIZE(Y1), %xmm5	movapd	 4 * SIZE(Y1), %xmm6	movapd	 6 * SIZE(Y1), %xmm7	movapd	 0 * SIZE(A1), %xmm8	movapd	 2 * SIZE(A1), %xmm9	movapd	 4 * SIZE(A1), %xmm10	movapd	 6 * SIZE(A1), %xmm11	mulpd	 %xmm0, %xmm8	mulpd	 %xmm0, %xmm9	mulpd	 %xmm0, %xmm10	mulpd	 %xmm0, %xmm11	addpd	 %xmm8,  %xmm4	addpd	 %xmm9,  %xmm5	addpd	 %xmm10, %xmm6	addpd	 %xmm11, %xmm7	movapd	 %xmm4, 0 * SIZE(Y1)	movapd	 %xmm5, 2 * SIZE(Y1)	movapd	 %xmm6, 4 * SIZE(Y1)	movapd	 %xmm7, 6 * SIZE(Y1)	addq	 $8 * SIZE, A1	addq	 $8 * SIZE, Y1	ALIGN_3.L66:	testq	$4, MM	je	.L67	movapd	 0 * SIZE(Y1), %xmm4	movapd	 2 * SIZE(Y1), %xmm5	movapd	 0 * SIZE(A1), %xmm8	movapd	 2 * SIZE(A1), %xmm9	mulpd	 %xmm0, %xmm8	mulpd	 %xmm0, %xmm9	addpd	 %xmm8,  %xmm4	addpd	 %xmm9,  %xmm5	movapd	 %xmm4, 0 * SIZE(Y1)	movapd	 %xmm5, 2 * SIZE(Y1)	addq	 $4 * SIZE, A1	addq	 $4 * SIZE, Y1	ALIGN_3.L67:	testq	$2, MM	je	.L68	movapd	 0 * SIZE(Y1), %xmm4	movapd	 0 * SIZE(A1), %xmm8	mulpd	 %xmm0, %xmm8	addpd	 %xmm8,  %xmm4	movapd	 %xmm4, 0 * SIZE(Y1)	addq	 $2 * SIZE, A1	addq	 $2 * SIZE, Y1	ALIGN_3.L68:	testq	$1, MM	je	.L990	movsd	 0 * SIZE(Y1), %xmm4	movsd	 0 * SIZE(A1), %xmm8	mulsd	 %xmm0, %xmm8	addsd	 %xmm8,  %xmm4	movsd	 %xmm4, 0 * SIZE(Y1)	ALIGN_3.L990:	movq	Y,  Y1	testq	$SIZE, A	je	.L991	movsd	(Y), %xmm4	addsd	(BUFFER), %xmm4	movsd	%xmm4, (Y1)	addq	INCY, Y	addq	INCY, Y1	addq	$1 * SIZE, BUFFER	ALIGN_3.L991:	movq	MM,  %rax	sarq	$3, %rax	jle	.L994	ALIGN_3.L992:	movsd	(Y), %xmm4	addq	INCY, Y	movhpd	(Y), %xmm4	addq	INCY, Y	movsd	(Y), %xmm5	addq	INCY, Y	movhpd	(Y), %xmm5	addq	INCY, Y	movsd	(Y), %xmm6	addq	INCY, Y	movhpd	(Y), %xmm6	addq	INCY, Y	movsd	(Y), %xmm7	addq	INCY, Y	movhpd	(Y), %xmm7	addq	INCY, Y	movapd	(BUFFER), %xmm0	movapd	2 * SIZE(BUFFER), %xmm1	movapd	4 * SIZE(BUFFER), %xmm2	movapd	6 * SIZE(BUFFER), %xmm3	addpd	%xmm4, %xmm0	addpd	%xmm5, %xmm1	addpd	%xmm6, %xmm2	addpd	%xmm7, %xmm3	movsd	%xmm0, (Y1)	addq	INCY, Y1	movhpd	%xmm0, (Y1)	addq	INCY, Y1	movsd	%xmm1, (Y1)	addq	INCY, Y1	movhpd	%xmm1, (Y1)	addq	INCY, Y1	movsd	%xmm2, (Y1)	addq	INCY, Y1	movhpd	%xmm2, (Y1)	addq	INCY, Y1	movsd	%xmm3, (Y1)	addq	INCY, Y1	movhpd	%xmm3, (Y1)	addq	INCY, Y1	addq	$8 * SIZE, BUFFER	decq	%rax	jg	.L992	ALIGN_3.L994:	testq	$7, MM	jle	.L999	testq	$4, MM	jle	.L995	movsd	(Y), %xmm4	addq	INCY, Y	movhpd	(Y), %xmm4	addq	INCY, Y	movsd	(Y), %xmm5	addq	INCY, Y	movhpd	(Y), %xmm5	addq	INCY, Y	movapd	(BUFFER), %xmm0	movapd	2 * SIZE(BUFFER), %xmm1	addpd	%xmm4, %xmm0	addpd	%xmm5, %xmm1	movsd	%xmm0, (Y1)	addq	INCY, Y1	movhpd	%xmm0, (Y1)	addq	INCY, Y1	movsd	%xmm1, (Y1)	addq	INCY, Y1	movhpd	%xmm1, (Y1)	addq	INCY, Y1	addq	$4 * SIZE, BUFFER	ALIGN_3.L995:	testq	$2, MM	jle	.L996	movsd	(Y), %xmm4	addq	INCY, Y	movhpd	(Y), %xmm4	addq	INCY, Y	movapd	(BUFFER), %xmm0	addpd	%xmm4, %xmm0	movsd	%xmm0, (Y1)	addq	INCY, Y1	movhpd	%xmm0, (Y1)	addq	INCY, Y1	addq	$2 * SIZE, BUFFER	ALIGN_3.L996:	testq	$1, MM	jle	.L999	movsd	(Y), %xmm4	addq	INCY, Y	movsd	(BUFFER), %xmm0	addsd	%xmm4, %xmm0	movsd	%xmm0, (Y1)	ALIGN_3.L999:	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	EPILOGUE
上一页 1 2 34
⌨️ 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 + -