📄 trsm_kernel_ln_8x4_sse.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 5 页
字号:
	addps	%xmm9, %xmm0	movaps	 4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm1	movaps	 8 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	mulps	12 * SIZE(BO), %xmm8	addps	%xmm9, %xmm2	movaps	16 * SIZE(BO), %xmm9	addps	%xmm8, %xmm3	movaps	 4 * SIZE(AO), %xmm8	addq	$ 4 * SIZE, AO		# aoffset  += 4	addq	$16 * SIZE, BO		# boffset1 += 8	decq	%rax	jg	.L26	ALIGN_4.L28:#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	BUFFER, BO	salq	$2 + BASE_SHIFT, %rax	leaq	(AO, %rax, 1), AO	leaq	(B,  %rax, 1), B	leaq	(BO, %rax, 4), BO#endif#if defined(LN) || defined(LT)	movaps	 %xmm0,  %xmm8	unpcklps %xmm2,  %xmm0	unpckhps %xmm2,  %xmm8	movaps	 %xmm1,  %xmm14	unpcklps %xmm3,  %xmm1	unpckhps %xmm3,  %xmm14	movaps	 %xmm0,  %xmm2	unpcklps %xmm1,  %xmm0	unpckhps %xmm1,  %xmm2	movaps	 %xmm8,  %xmm3	unpcklps %xmm14, %xmm8	unpckhps %xmm14, %xmm3	movaps	 0 * SIZE(B), %xmm1	movaps	 4 * SIZE(B), %xmm5	movaps	 8 * SIZE(B), %xmm10	movaps	12 * SIZE(B), %xmm11	subps	%xmm0,  %xmm1	subps	%xmm2,  %xmm5	subps	%xmm8,  %xmm10	subps	%xmm3,  %xmm11#else	movaps	 0 * SIZE(AO), %xmm8	movaps	 4 * SIZE(AO), %xmm10	movaps	 8 * SIZE(AO), %xmm12	movaps	12 * SIZE(AO), %xmm14	subps	%xmm0, %xmm8	subps	%xmm1, %xmm10	subps	%xmm2, %xmm12	subps	%xmm3, %xmm14#endif#ifdef LN	movaps	 12 * SIZE(AO), %xmm6	pshufd	 $0xff, %xmm6, %xmm8	mulps	 %xmm8, %xmm11	pshufd	 $0xaa, %xmm6, %xmm8	mulps	 %xmm11, %xmm8	subps	 %xmm8, %xmm10	pshufd	 $0x55, %xmm6, %xmm8	mulps	 %xmm11, %xmm8	subps	 %xmm8, %xmm5	pshufd	 $0x00, %xmm6, %xmm8	mulps	 %xmm11, %xmm8	subps	 %xmm8, %xmm1	movaps	  8 * SIZE(AO), %xmm6	pshufd	 $0xaa, %xmm6, %xmm8	mulps	 %xmm8, %xmm10	pshufd	 $0x55, %xmm6, %xmm8	mulps	 %xmm10, %xmm8	subps	 %xmm8, %xmm5	pshufd	 $0x00, %xmm6, %xmm8	mulps	 %xmm10, %xmm8	subps	 %xmm8, %xmm1	movaps	  4 * SIZE(AO), %xmm6	pshufd	 $0x55, %xmm6, %xmm8	mulps	 %xmm8, %xmm5	pshufd	 $0x00, %xmm6, %xmm8	mulps	 %xmm5, %xmm8	subps	 %xmm8, %xmm1	movaps	  0 * SIZE(AO), %xmm6	pshufd	 $0x00, %xmm6, %xmm8	mulps	 %xmm8, %xmm1#endif#ifdef LT	movaps	 0 * SIZE(AO), %xmm6	pshufd	 $0x00, %xmm6, %xmm8	mulps	 %xmm8, %xmm1	pshufd	 $0x55, %xmm6, %xmm8	mulps	 %xmm1, %xmm8	subps	 %xmm8, %xmm5	pshufd	 $0xaa, %xmm6, %xmm8	mulps	 %xmm1, %xmm8	subps	 %xmm8, %xmm10	pshufd	 $0xff, %xmm6, %xmm8	mulps	 %xmm1, %xmm8	subps	 %xmm8, %xmm11	movaps	 4 * SIZE(AO), %xmm6	pshufd	 $0x55, %xmm6, %xmm8	mulps	 %xmm8, %xmm5	pshufd	 $0xaa, %xmm6, %xmm8	mulps	 %xmm5, %xmm8	subps	 %xmm8, %xmm10	pshufd	 $0xff, %xmm6, %xmm8	mulps	 %xmm5, %xmm8	subps	 %xmm8, %xmm11	movaps	 8 * SIZE(AO), %xmm6	pshufd	 $0xaa, %xmm6, %xmm8	mulps	 %xmm8, %xmm10	pshufd	 $0xff, %xmm6, %xmm8	mulps	 %xmm10, %xmm8	subps	 %xmm8, %xmm11	movaps	12 * SIZE(AO), %xmm6	pshufd	 $0xff, %xmm6, %xmm8	mulps	 %xmm8, %xmm11#endif#ifdef RN	movaps	 0 * SIZE(B), %xmm0	pshufd	 $0x00, %xmm0, %xmm2	mulps	 %xmm2, %xmm8	pshufd	 $0x55, %xmm0, %xmm2	mulps	 %xmm8, %xmm2	subps	 %xmm2, %xmm10	pshufd	 $0xaa, %xmm0, %xmm2	mulps	 %xmm8, %xmm2	subps	 %xmm2, %xmm12	pshufd	 $0xff, %xmm0, %xmm2	mulps	 %xmm8, %xmm2	subps	 %xmm2, %xmm14	movaps	 4 * SIZE(B), %xmm0	pshufd	 $0x55, %xmm0, %xmm2	mulps	 %xmm2, %xmm10	pshufd	 $0xaa, %xmm0, %xmm2	mulps	 %xmm10, %xmm2	subps	 %xmm2, %xmm12	pshufd	 $0xff, %xmm0, %xmm2	mulps	 %xmm10, %xmm2	subps	 %xmm2, %xmm14	movaps	 8 * SIZE(B), %xmm0	pshufd	 $0xaa, %xmm0, %xmm2	mulps	 %xmm2, %xmm12	pshufd	 $0xff, %xmm0, %xmm2	mulps	 %xmm12, %xmm2	subps	 %xmm2, %xmm14	movaps	 12 * SIZE(B), %xmm0	pshufd	 $0xff, %xmm0, %xmm2	mulps	 %xmm2, %xmm14#endif#ifdef RT	movaps	 12 * SIZE(B), %xmm0	pshufd	 $0xff, %xmm0, %xmm2	mulps	 %xmm2, %xmm14	pshufd	 $0xaa, %xmm0, %xmm2	mulps	 %xmm14, %xmm2	subps	 %xmm2, %xmm12	pshufd	 $0x55, %xmm0, %xmm2	mulps	 %xmm14, %xmm2	subps	 %xmm2, %xmm10	pshufd	 $0x00, %xmm0, %xmm2	mulps	 %xmm14, %xmm2	subps	 %xmm2, %xmm8	movaps	  8 * SIZE(B), %xmm0	pshufd	 $0xaa, %xmm0, %xmm2	mulps	 %xmm2, %xmm12	pshufd	 $0x55, %xmm0, %xmm2	mulps	 %xmm12, %xmm2	subps	 %xmm2, %xmm10	pshufd	 $0x00, %xmm0, %xmm2	mulps	 %xmm12, %xmm2	subps	 %xmm2, %xmm8	movaps	  4 * SIZE(B), %xmm0	pshufd	 $0x55, %xmm0, %xmm2	mulps	 %xmm2, %xmm10	pshufd	 $0x00, %xmm0, %xmm2	mulps	 %xmm10, %xmm2	subps	 %xmm2, %xmm8	movaps	  0 * SIZE(B), %xmm0	pshufd	 $0x00, %xmm0, %xmm2	mulps	 %xmm2, %xmm8#endif#ifdef LN	subq	$4 * SIZE, CO1	subq	$4 * SIZE, CO2#endif#if defined(LN) || defined(LT)	movaps	%xmm1,   0 * SIZE(B)	movaps	%xmm5,   4 * SIZE(B)	movaps	%xmm10,  8 * SIZE(B)	movaps	%xmm11, 12 * SIZE(B)	pshufd	$0x00, %xmm1, %xmm2	pshufd	$0x55, %xmm1, %xmm3	pshufd	$0xaa, %xmm1, %xmm4	pshufd	$0xff, %xmm1, %xmm6	movaps	%xmm2,   0 * SIZE(BO)	movaps	%xmm3,   4 * SIZE(BO)	movaps	%xmm4,   8 * SIZE(BO)	movaps	%xmm6,  12 * SIZE(BO)	pshufd	$0x00, %xmm5, %xmm2	pshufd	$0x55, %xmm5, %xmm3	pshufd	$0xaa, %xmm5, %xmm4	pshufd	$0xff, %xmm5, %xmm6	movaps	%xmm2,  16 * SIZE(BO)	movaps	%xmm3,  20 * SIZE(BO)	movaps	%xmm4,  24 * SIZE(BO)	movaps	%xmm6,  28 * SIZE(BO)	pshufd	$0x00, %xmm10, %xmm2	pshufd	$0x55, %xmm10, %xmm3	pshufd	$0xaa, %xmm10, %xmm4	pshufd	$0xff, %xmm10, %xmm6	movaps	%xmm2,  32 * SIZE(BO)	movaps	%xmm3,  36 * SIZE(BO)	movaps	%xmm4,  40 * SIZE(BO)	movaps	%xmm6,  44 * SIZE(BO)	pshufd	$0x00, %xmm11, %xmm2	pshufd	$0x55, %xmm11, %xmm3	pshufd	$0xaa, %xmm11, %xmm4	pshufd	$0xff, %xmm11, %xmm6	movaps	%xmm2,  48 * SIZE(BO)	movaps	%xmm3,  52 * SIZE(BO)	movaps	%xmm4,  56 * SIZE(BO)	movaps	%xmm6,  60 * SIZE(BO)#else	movaps	%xmm8,   0 * SIZE(AO)	movaps	%xmm10,  4 * SIZE(AO)	movaps	%xmm12,  8 * SIZE(AO)	movaps	%xmm14, 12 * SIZE(AO)#endif#if defined(LN) || defined(LT)	movaps	 %xmm1,  %xmm0	unpcklps %xmm10, %xmm1	unpckhps %xmm10, %xmm0	movaps	 %xmm5,  %xmm7	unpcklps %xmm11, %xmm5	unpckhps %xmm11, %xmm7	movaps	 %xmm1,  %xmm10	unpcklps %xmm5,  %xmm1	unpckhps %xmm5,  %xmm10	movaps	 %xmm0,  %xmm11	unpcklps %xmm7,  %xmm0	unpckhps %xmm7,  %xmm11	movlps	%xmm1,  0 * SIZE(CO1)	movhps	%xmm1,  2 * SIZE(CO1)	movlps	%xmm10, 0 * SIZE(CO2)	movhps	%xmm10, 2 * SIZE(CO2)	movlps	%xmm0,  0 * SIZE(CO1, LDC, 2)	movhps	%xmm0,  2 * SIZE(CO1, LDC, 2)	movlps	%xmm11, 0 * SIZE(CO2, LDC, 2)	movhps	%xmm11, 2 * SIZE(CO2, LDC, 2)#else	movlps	%xmm8,   0 * SIZE(CO1)	movhps	%xmm8,   2 * SIZE(CO1)	movlps	%xmm10,  0 * SIZE(CO2)	movhps	%xmm10,  2 * SIZE(CO2)	movlps	%xmm12,  0 * SIZE(CO1, LDC, 2)	movhps	%xmm12,  2 * SIZE(CO1, LDC, 2)	movlps	%xmm14,  0 * SIZE(CO2, LDC, 2)	movhps	%xmm14,  2 * SIZE(CO2, LDC, 2)#endif#ifndef LN	addq	$4 * SIZE, CO1	addq	$4 * SIZE, CO2#endif#if defined(LT) || defined(RN)	movq	K,  %rax	subq	KK, %rax	leaq	(,%rax, SIZE), %rax	leaq	(AO, %rax, 4), AO#ifdef LT	addq	$16 * SIZE, B#endif#endif#ifdef LN	subq	$4, KK	movq	BORIG, B#endif#ifdef LT	addq	$4, KK#endif#ifdef RT	movq	K, %rax	movq	BORIG, B	salq	$2 + BASE_SHIFT, %rax	addq	%rax, AORIG#endif	ALIGN_4	.L40:	movq	M,  I	sarq	$3, I	# i = (m >> 3)	jle	.L49	ALIGN_4.L11:#ifdef LN       movq	K, %rax       salq	$3 + 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, 8), AO#endif	leaq	BUFFER, BO#if defined(LN) || defined(RT)	movq	KK, %rax	salq	$2 + BASE_SHIFT, %rax	leaq	(BO, %rax, 4), BO#endif		movaps	 0 * SIZE(BO), %xmm9	movaps	 4 * SIZE(BO), %xmm11	movaps	 8 * SIZE(BO), %xmm13	movaps	16 * SIZE(BO), %xmm15	movaps	 0 * SIZE(AO), %xmm8	pxor	%xmm0, %xmm0	movaps	 4 * SIZE(AO), %xmm10	pxor	%xmm1, %xmm1	movaps	 8 * SIZE(AO), %xmm12	pxor	%xmm2, %xmm2	movaps	12 * SIZE(AO), %xmm14	pxor	%xmm3, %xmm3	PREFETCHW     -8 * SIZE(CO1)	pxor	%xmm4, %xmm4	PREFETCHW     -8 * SIZE(CO2)	pxor	%xmm5, %xmm5	PREFETCHW     -8 * SIZE(CO1, LDC, 2)	pxor	%xmm6, %xmm6	PREFETCHW     -8 * SIZE(CO2, LDC, 2)	pxor	%xmm7, %xmm7#if defined(LT) || defined(RN)	movq	KK, %rax#else	movq	K, %rax	subq	KK, %rax#endif	sarq	$2, %rax	je	.L15	ALIGN_4	.L12:	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movaps	 0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm11	addps	%xmm11, %xmm1	movaps	 4 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm13	mulps	12 * SIZE(BO), %xmm8	addps	%xmm13, %xmm2	movaps	 8 * SIZE(BO), %xmm13	addps	%xmm8, %xmm3	movaps	16 * SIZE(AO), %xmm8	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm4	movaps	32 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm5	movaps	20 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm13	mulps	12 * SIZE(BO), %xmm10	addps	%xmm13, %xmm6	movaps	24 * SIZE(BO), %xmm13	addps	%xmm10, %xmm7	movaps	20 * SIZE(AO), %xmm10	mulps	%xmm12, %xmm15	addps	%xmm15, %xmm0	movaps	16 * SIZE(BO), %xmm15	mulps	%xmm12, %xmm11	addps	%xmm11, %xmm1	movaps	20 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm13	mulps	28 * SIZE(BO), %xmm12	addps	%xmm13, %xmm2	movaps	24 * SIZE(BO), %xmm13	addps	%xmm12, %xmm3	movaps	24 * SIZE(AO), %xmm12	mulps	%xmm14, %xmm15	addps	%xmm15, %xmm4	movaps	48 * SIZE(BO), %xmm15	mulps	%xmm14, %xmm11	addps	%xmm11, %xmm5	movaps	36 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm13	mulps	28 * SIZE(BO), %xmm14	addps	%xmm13, %xmm6	movaps	40 * SIZE(BO), %xmm13	addps	%xmm14, %xmm7	movaps	28 * SIZE(AO), %xmm14	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movaps	32 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm11	addps	%xmm11, %xmm1	movaps	36 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm13	mulps	44 * SIZE(BO), %xmm8	addps	%xmm13, %xmm2	movaps	40 * SIZE(BO), %xmm13	addps	%xmm8, %xmm3	movaps	32 * SIZE(AO), %xmm8	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm4	movaps	64 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm5	movaps	52 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm13	mulps	44 * SIZE(BO), %xmm10	addps	%xmm13, %xmm6	movaps	56 * SIZE(BO), %xmm13	addps	%xmm10, %xmm7	movaps	36 * SIZE(AO), %xmm10	mulps	%xmm12, %xmm15	addps	%xmm15, %xmm0	movaps	48 * SIZE(BO), %xmm15	mulps	%xmm12, %xmm11	addps	%xmm11, %xmm1	movaps	52 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm13	mulps	60 * SIZE(BO), %xmm12	addps	%xmm13, %xmm2	movaps	56 * SIZE(BO), %xmm13	addps	%xmm12, %xmm3	movaps	40 * SIZE(AO), %xmm12	mulps	%xmm14, %xmm15	addps	%xmm15, %xmm4	movaps	80 * SIZE(BO), %xmm15	mulps	%xmm14, %xmm11	addps	%xmm11, %xmm5	movaps	68 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm13	mulps	60 * SIZE(BO), %xmm14	addps	%xmm13, %xmm6	movaps	72 * SIZE(BO), %xmm13	addps	%xmm14, %xmm7	movaps	44 * SIZE(AO), %xmm14	addq	$32 * SIZE, AO	addq	$64 * SIZE, BO	decq	%rax	jg	.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 .L18	ALIGN_4.L16:	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movaps	 4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm1	movaps	 8 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	mulps	12 * SIZE(BO), %xmm8	addps	%xmm9, %xmm2	movaps	 0 * SIZE(BO), %xmm9	addps	%xmm8, %xmm3	movaps	 8 * SIZE(AO), %xmm8	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm4	movaps	 4 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm5	movaps	 8 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	mulps	12 * SIZE(BO), %xmm10	addps	%xmm9, %xmm6	movaps	16 * SIZE(BO), %xmm9	addps	%xmm10, %xmm7	movaps	12 * SIZE(AO), %xmm10	addq	$8  * SIZE, AO	addq	$16 * SIZE, BO	decq	%rax	jg	.L16	ALIGN_4.L18:#if defined(LN) || defined(RT)	movq	KK, %rax#ifdef LN	subq	$8, %rax#else	subq	$4, %rax#endif	movq	AORIG, AO	movq	BORIG, B	leaq	BUFFER, BO	salq	$2 + BASE_SHIFT, %rax	leaq	(AO, %rax, 2), AO	leaq	(B,  %rax, 1), B	leaq	(BO, %rax, 4), BO#endif#if defined(LN) || defined(LT)	movaps	 %xmm0,  %xmm8	unpcklps %xmm2,  %xmm0	unpckhps %xmm2,  %xmm8	movaps	 %xmm1,  %xmm14	unpcklps %xmm3,  %xmm1	unpckhps %xmm3,  %xmm14	movaps	 %xmm0,  %xmm2	unpcklps %xmm1,  %xmm0	unpckhps %xmm1,  %xmm2	movaps	 %xmm8,  %xmm3	unpcklps %xmm14, %xmm8	unpckhps %xmm14, %xmm3	movaps	 %xmm4,  %xmm9	unpcklps %xmm6,  %xmm4	unpckhps %xmm6,  %xmm9	movaps	 %xmm5,  %xmm14	unpcklps %xmm7,  %xmm5	unpckhps %xmm7,  %xmm14	movaps	 %xmm4,  %xmm6	unpcklps %xmm5,  %xmm4	unpckhps %xmm5,  %xmm6	movaps	 %xmm9,  %xmm7	unpcklps %xmm14, %xmm9	unpckhps %xmm14, %xmm7	movaps	 0 * SIZE(B), %xmm1	movaps	 4 * SIZE(B), %xmm5	movaps	 8 * SIZE(B), %xmm10	movaps	12 * SIZE(B), %xmm11	movaps	16 * SIZE(B), %xmm12	movaps	20 * SIZE(B), %xmm13	movaps	24 * SIZE(B), %xmm14	movaps	28 * SIZE(B), %xmm15	subps	%xmm0,  %xmm1	subps	%xmm2,  %xmm5	subps	%xmm8,  %xmm10	subps	%xmm3,  %xmm11	subps	%xmm4,  %xmm12	subps	%xmm6,  %xmm13	subps	%xmm9,  %xmm14	subps	%xmm7,  %xmm15#else	movaps	 0 * SIZE(AO), %xmm8	movaps	 4 * SIZE(AO), %xmm9	movaps	 8 * SIZE(AO), %xmm10	movaps	12 * SIZE(AO), %xmm11	movaps	16 * SIZE(AO), %xmm12	movaps	20 * SIZE(AO), %xmm13	movaps	24 * SIZE(AO), %xmm14	movaps	28 * SIZE(AO), %xmm15	subps	%xmm0, %xmm8	subps	%xmm4, %xmm9	subps	%xmm1, %xmm10	subps	%xmm5, %xmm11	subps	%xmm2, %xmm12	subps	%xmm6, %xmm13	subps	%xmm3, %xmm14	subps	%xmm7, %xmm15#endif#ifdef LN	movaps	 60 * SIZE(AO), %xmm6	pshufd	 $0xff, %xmm6, %xmm8	mulps	 %xmm8, %xmm15	pshufd	 $0xaa, %xmm6, %xmm8	mulps	 %xmm15, %xmm8	subps	 %xmm8, %xmm14	pshufd	 $0x55, %xmm6, %xmm8	mulps	 %xmm15, %xmm8	subps	 %xmm8, %xmm13	pshufd	 $0x00, %xmm6, %xmm8	mulps	 %xmm15, %xmm8	subps	 %xmm8, %xmm12	movaps	 56 * SIZE(AO), %xmm6	pshufd	 $0xff, %xmm6, %xmm8	mulps	 %xmm15, %xmm8	subps	 %xmm8, %xmm11	pshufd	 $0xaa, %xmm6, %xmm8	mulps	 %xmm15, %xmm8	subps	 %xmm8, %xmm10	pshufd	 $0x55, %xmm6, %xmm8	mulps	 %xmm15, %xmm8	subps	 %xmm8, %xmm5	pshufd	 $0x00, %xmm6, %xmm8	mulps	 %xmm15, %xmm8	subps	 %xmm8, %xmm1
⌨️ 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 + -