📄 trsm_kernel_rt_8x4_sse.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 5 页
字号:
	mulss	 %xmm8, %xmm10	pshufd	 $0xff, %xmm6, %xmm8	mulss	 %xmm10, %xmm8	subss	 %xmm8, %xmm11	movaps	20 * SIZE(AO), %xmm7	pshufd	 $0x00, %xmm7, %xmm8	mulss	 %xmm10, %xmm8	subss	 %xmm8, %xmm12	pshufd	 $0x55, %xmm7, %xmm8	mulss	 %xmm10, %xmm8	subss	 %xmm8, %xmm13	pshufd	 $0xaa, %xmm7, %xmm8	mulss	 %xmm10, %xmm8	subss	 %xmm8, %xmm14	pshufd	 $0xff, %xmm7, %xmm8	mulss	 %xmm10, %xmm8	subss	 %xmm8, %xmm15	movaps	24 * SIZE(AO), %xmm6	pshufd	 $0xff, %xmm6, %xmm8	mulss	 %xmm8, %xmm11	movaps	28 * SIZE(AO), %xmm7	pshufd	 $0x00, %xmm7, %xmm8	mulss	 %xmm11, %xmm8	subss	 %xmm8, %xmm12	pshufd	 $0x55, %xmm7, %xmm8	mulss	 %xmm11, %xmm8	subss	 %xmm8, %xmm13	pshufd	 $0xaa, %xmm7, %xmm8	mulss	 %xmm11, %xmm8	subss	 %xmm8, %xmm14	pshufd	 $0xff, %xmm7, %xmm8	mulss	 %xmm11, %xmm8	subss	 %xmm8, %xmm15	movaps	36 * SIZE(AO), %xmm6	pshufd	 $0x00, %xmm6, %xmm8	mulss	 %xmm8, %xmm12	pshufd	 $0x55, %xmm6, %xmm8	mulss	 %xmm12, %xmm8	subss	 %xmm8, %xmm13	pshufd	 $0xaa, %xmm6, %xmm8	mulss	 %xmm12, %xmm8	subss	 %xmm8, %xmm14	pshufd	 $0xff, %xmm6, %xmm8	mulss	 %xmm12, %xmm8	subss	 %xmm8, %xmm15	movaps	44 * SIZE(AO), %xmm7	pshufd	 $0x55, %xmm7, %xmm8	mulss	 %xmm8, %xmm13	pshufd	 $0xaa, %xmm7, %xmm8	mulss	 %xmm13, %xmm8	subss	 %xmm8, %xmm14	pshufd	 $0xff, %xmm7, %xmm8	mulss	 %xmm13, %xmm8	subss	 %xmm8, %xmm15	movaps	52 * SIZE(AO), %xmm6	pshufd	 $0xaa, %xmm6, %xmm8	mulss	 %xmm8, %xmm14	pshufd	 $0xff, %xmm6, %xmm8	mulss	 %xmm14, %xmm8	subss	 %xmm8, %xmm15	movaps	60 * SIZE(AO), %xmm7	pshufd	 $0xff, %xmm7, %xmm8	mulss	 %xmm8, %xmm15#endif#if defined(RN) || defined(RT)	movss	 0 * SIZE(B), %xmm0	pshufd	 $0x00, %xmm0, %xmm2	mulps	 %xmm2, %xmm8	mulps	 %xmm2, %xmm9#endif#ifdef LN	subq	$8 * SIZE, CO1#endif#if defined(LN) || defined(LT)	movss	%xmm1,   0 * SIZE(B)	movss	%xmm5,   1 * SIZE(B)	movss	%xmm10,  2 * SIZE(B)	movss	%xmm11,  3 * SIZE(B)	movss	%xmm12,  4 * SIZE(B)	movss	%xmm13,  5 * SIZE(B)	movss	%xmm14,  6 * SIZE(B)	movss	%xmm15,  7 * SIZE(B)	pshufd	$0x00, %xmm1, %xmm2	movaps	%xmm2,   0 * SIZE(BO)	pshufd	$0x00, %xmm5, %xmm2	movaps	%xmm2,   4 * SIZE(BO)	pshufd	$0x00, %xmm10, %xmm2	movaps	%xmm2,   8 * SIZE(BO)	pshufd	$0x00, %xmm11, %xmm2	movaps	%xmm2,  12 * SIZE(BO)	pshufd	$0x00, %xmm12, %xmm2	movaps	%xmm2,  16 * SIZE(BO)	pshufd	$0x00, %xmm13, %xmm2	movaps	%xmm2,  20 * SIZE(BO)	pshufd	$0x00, %xmm14, %xmm2	movaps	%xmm2,  24 * SIZE(BO)	pshufd	$0x00, %xmm15, %xmm2	movaps	%xmm2,  28 * SIZE(BO)#else	movaps	%xmm8,   0 * SIZE(AO)	movaps	%xmm9,   4 * SIZE(AO)#endif#if defined(LN) || defined(LT)	unpcklps %xmm10, %xmm1	unpcklps %xmm11, %xmm5	unpcklps %xmm5,  %xmm1	unpcklps %xmm14, %xmm12	unpcklps %xmm15, %xmm13	unpcklps %xmm13, %xmm12	movlps	%xmm1,  0 * SIZE(CO1)	movhps	%xmm1,  2 * SIZE(CO1)	movlps	%xmm12, 4 * SIZE(CO1)	movhps	%xmm12, 6 * SIZE(CO1)#else	movlps	%xmm8,   0 * SIZE(CO1)	movhps	%xmm8,   2 * SIZE(CO1)	movlps	%xmm9,   4 * SIZE(CO1)	movhps	%xmm9,   6 * SIZE(CO1)#endif#ifndef LN	addq	$8 * SIZE, CO1#endif#if defined(LT) || defined(RN)	movq	K,  %rax	subq	KK, %rax	leaq	(,%rax, SIZE), %rax	leaq	(AO, %rax, 8), AO#ifdef LT	addq	$8 * SIZE, B#endif#endif#ifdef LN	subq	$8, KK	movq	BORIG, B#endif#ifdef LT	addq	$8, KK#endif#ifdef RT	movq	K, %rax	movq	BORIG, B	salq	$3 + BASE_SHIFT, %rax	addq	%rax, AORIG#endif	decq	I			# i --	jg	.L111	ALIGN_4	.L120:	testq	$4, M	je	.L130#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	BUFFER, BO#if defined(LN) || defined(RT)	movq	KK, %rax	salq	$BASE_SHIFT, %rax	leaq	(BO, %rax, 4), BO#endif		movaps	 0 * SIZE(AO), %xmm8	movaps	16 * SIZE(AO), %xmm10	movaps	 0 * SIZE(BO), %xmm9	movaps	16 * SIZE(BO), %xmm11	pxor	%xmm0, %xmm0	pxor	%xmm1, %xmm1	pxor	%xmm2, %xmm2	pxor	%xmm3, %xmm3#if defined(LT) || defined(RN)	movq	KK, %rax#else	movq	K, %rax	subq	KK, %rax#endif	sarq	$3, %rax	je	.L125	ALIGN_4.L122:	mulps	%xmm8, %xmm9#if defined(OPTERON) && defined(HAVE_PREFETCH)	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)#endif	movaps	 4 * SIZE(AO), %xmm8	mulps	 4 * SIZE(BO), %xmm8	addps	%xmm9, %xmm0	movaps	32 * SIZE(BO), %xmm9	addps	%xmm8, %xmm1	movaps	 8 * SIZE(AO), %xmm8	mulps	 8 * SIZE(BO), %xmm8	addps	%xmm8, %xmm2	movaps	12 * SIZE(AO), %xmm8	mulps	12 * SIZE(BO), %xmm8	addps	%xmm8, %xmm3	movaps	32 * SIZE(AO), %xmm8#if defined(OPTERON) && defined(HAVE_PREFETCH)	PREFETCH	(PREFETCHSIZE + 16) * SIZE(AO)#endif	mulps	%xmm10, %xmm11	movaps	20 * SIZE(AO), %xmm10	mulps	20 * SIZE(BO), %xmm10	addps	%xmm11, %xmm0	movaps	48 * SIZE(BO), %xmm11	addps	%xmm10, %xmm1	movaps	24 * SIZE(AO), %xmm10	mulps	24 * SIZE(BO), %xmm10	addps	%xmm10, %xmm2	movaps	28 * SIZE(AO), %xmm10	mulps	28 * SIZE(BO), %xmm10	addps	%xmm10, %xmm3	movaps	48 * SIZE(AO), %xmm10	addq   $32 * SIZE, AO	addq   $32 * SIZE, BO	decq   %rax	jne    .L122	ALIGN_4.L125:#if defined(LT) || defined(RN)	movq	KK, %rax#else	movq	K, %rax	subq	KK, %rax#endif	andq	$7, %rax		# if (k & 1)	BRANCH	je .L128	ALIGN_4.L126:	mulps	%xmm8, %xmm9	movaps	 4 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movaps	 4 * SIZE(BO), %xmm9	addq	$4 * SIZE, AO		# aoffset  += 4	addq	$4 * SIZE, BO		# boffset1 += 8	decq	%rax	jg	.L126	ALIGN_4.L128:	addps	%xmm1, %xmm0	addps	%xmm3, %xmm2	addps	%xmm2, %xmm0#if defined(LN) || defined(RT)	movq	KK, %rax#ifdef LN	subq	$4, %rax#else	subq	$1, %rax#endif	movq	AORIG, AO	movq	BORIG, B	leaq	BUFFER, BO	salq	$BASE_SHIFT, %rax	leaq	(AO, %rax, 4), 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	movss	 0 * SIZE(B), %xmm1	movss	 1 * SIZE(B), %xmm5	movss	 2 * SIZE(B), %xmm10	movss	 3 * SIZE(B), %xmm11	subss	%xmm0,  %xmm1	subss	%xmm2,  %xmm5	subss	%xmm8,  %xmm10	subss	%xmm3,  %xmm11#else	movaps	 0 * SIZE(AO), %xmm8	subps	%xmm0, %xmm8#endif#ifdef LN	movaps	 12 * SIZE(AO), %xmm6	pshufd	 $0xff, %xmm6, %xmm8	mulss	 %xmm8, %xmm11	pshufd	 $0xaa, %xmm6, %xmm8	mulss	 %xmm11, %xmm8	subss	 %xmm8, %xmm10	pshufd	 $0x55, %xmm6, %xmm8	mulss	 %xmm11, %xmm8	subss	 %xmm8, %xmm5	pshufd	 $0x00, %xmm6, %xmm8	mulss	 %xmm11, %xmm8	subss	 %xmm8, %xmm1	movaps	  8 * SIZE(AO), %xmm6	pshufd	 $0xaa, %xmm6, %xmm8	mulss	 %xmm8, %xmm10	pshufd	 $0x55, %xmm6, %xmm8	mulss	 %xmm10, %xmm8	subss	 %xmm8, %xmm5	pshufd	 $0x00, %xmm6, %xmm8	mulss	 %xmm10, %xmm8	subss	 %xmm8, %xmm1	movaps	  4 * SIZE(AO), %xmm6	pshufd	 $0x55, %xmm6, %xmm8	mulss	 %xmm8, %xmm5	pshufd	 $0x00, %xmm6, %xmm8	mulss	 %xmm5, %xmm8	subss	 %xmm8, %xmm1	movaps	  0 * SIZE(AO), %xmm6	pshufd	 $0x00, %xmm6, %xmm8	mulss	 %xmm8, %xmm1#endif#ifdef LT	movaps	 0 * SIZE(AO), %xmm6	pshufd	 $0x00, %xmm6, %xmm8	mulss	 %xmm8, %xmm1	pshufd	 $0x55, %xmm6, %xmm8	mulss	 %xmm1, %xmm8	subss	 %xmm8, %xmm5	pshufd	 $0xaa, %xmm6, %xmm8	mulss	 %xmm1, %xmm8	subss	 %xmm8, %xmm10	pshufd	 $0xff, %xmm6, %xmm8	mulss	 %xmm1, %xmm8	subss	 %xmm8, %xmm11	movaps	 4 * SIZE(AO), %xmm6	pshufd	 $0x55, %xmm6, %xmm8	mulss	 %xmm8, %xmm5	pshufd	 $0xaa, %xmm6, %xmm8	mulss	 %xmm5, %xmm8	subss	 %xmm8, %xmm10	pshufd	 $0xff, %xmm6, %xmm8	mulss	 %xmm5, %xmm8	subss	 %xmm8, %xmm11	movaps	 8 * SIZE(AO), %xmm6	pshufd	 $0xaa, %xmm6, %xmm8	mulss	 %xmm8, %xmm10	pshufd	 $0xff, %xmm6, %xmm8	mulss	 %xmm10, %xmm8	subss	 %xmm8, %xmm11	movaps	12 * SIZE(AO), %xmm6	pshufd	 $0xff, %xmm6, %xmm8	mulss	 %xmm8, %xmm11#endif#if defined(RN) || defined(RT)	movss	 0 * SIZE(B), %xmm0	pshufd	 $0x00, %xmm0, %xmm2	mulps	 %xmm2, %xmm8#endif#ifdef LN	subq	$4 * SIZE, CO1#endif#if defined(LN) || defined(LT)	movss	%xmm1,   0 * SIZE(B)	movss	%xmm5,   1 * SIZE(B)	movss	%xmm10,  2 * SIZE(B)	movss	%xmm11,  3 * SIZE(B)	pshufd	$0x00, %xmm1, %xmm2	movaps	%xmm2,   0 * SIZE(BO)	pshufd	$0x00, %xmm5, %xmm2	movaps	%xmm2,   4 * SIZE(BO)	pshufd	$0x00, %xmm10, %xmm2	movaps	%xmm2,   8 * SIZE(BO)	pshufd	$0x00, %xmm11, %xmm2	movaps	%xmm2,  12 * SIZE(BO)#else	movaps	%xmm8,   0 * SIZE(AO)#endif#if defined(LN) || defined(LT)	unpcklps %xmm10, %xmm1	unpcklps %xmm11, %xmm5	unpcklps %xmm5,  %xmm1	movlps	%xmm1,  0 * SIZE(CO1)	movhps	%xmm1,  2 * SIZE(CO1)#else	movlps	%xmm8,   0 * SIZE(CO1)	movhps	%xmm8,   2 * SIZE(CO1)#endif#ifndef LN	addq	$4 * SIZE, CO1#endif#if defined(LT) || defined(RN)	movq	K,  %rax	subq	KK, %rax	leaq	(,%rax, SIZE), %rax	leaq	(AO, %rax, 4), AO#ifdef LT	addq	$4 * 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	.L130:	testq	$2, M	je	.L140#ifdef LN       movq	K, %rax       salq	$1 + 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, 2), AO#endif	leaq	BUFFER, BO#if defined(LN) || defined(RT)	movq	KK, %rax	salq	$BASE_SHIFT, %rax	leaq	(BO, %rax, 4), BO#endif		movlps	 0 * SIZE(AO), %xmm8	movhps	 2 * SIZE(AO), %xmm8	movlps	 8 * SIZE(AO), %xmm10	movhps	10 * SIZE(AO), %xmm10	movaps	 0 * SIZE(BO), %xmm9	movaps	16 * SIZE(BO), %xmm11	pxor	%xmm0, %xmm0	pxor	%xmm1, %xmm1	pxor	%xmm2, %xmm2	pxor	%xmm3, %xmm3#if defined(LT) || defined(RN)	movq	KK, %rax#else	movq	K, %rax	subq	KK, %rax#endif	sarq	$3, %rax	je	.L135	ALIGN_4.L132:	mulps	%xmm8, %xmm9#if defined(OPTERON) && defined(HAVE_PREFETCH)	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)#endif	movlps	 2 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movlps	 4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movlps	 4 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movlps	 8 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movlps	 6 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movlps	12 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movlps	16 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movlps	32 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm11	movlps	10 * SIZE(AO), %xmm10	addps	%xmm11, %xmm0	movlps	20 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	movlps	12 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movlps	24 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	movlps	14 * SIZE(AO), %xmm10	addps	%xmm11, %xmm0	movlps	28 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	movlps	24 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movlps	48 * SIZE(BO), %xmm11	addq   $16 * SIZE, AO	addq   $32 * SIZE, BO	decq   %rax	jne    .L132	ALIGN_4.L135:#if defined(LT) || defined(RN)	movq	KK, %rax#else	movq	K, %rax	subq	KK, %rax#endif	andq	$7, %rax		# if (k & 1)	BRANCH	je .L138	ALIGN_4.L136:	mulps	%xmm8, %xmm9	movlps	 2 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movlps	 4 * SIZE(BO), %xmm9	addq	$2 * SIZE, AO		# aoffset  += 4	addq	$4 * SIZE, BO		# boffset1 += 8	decq	%rax	jg	.L136	ALIGN_4.L138:	addps	%xmm1,  %xmm0#if defined(LN) || defined(RT)	movq	KK, %rax#ifdef LN	subq	$2, %rax#else	subq	$1, %rax#endif	movq	AORIG, AO	movq	BORIG, B	leaq	BUFFER, BO	salq	$BASE_SHIFT, %rax	leaq	(AO, %rax, 2), AO	leaq	(B,  %rax, 1), B	leaq	(BO, %rax, 4), BO#endif#if defined(LN) || defined(LT)	unpcklps %xmm2,  %xmm0	unpcklps %xmm3,  %xmm1	movaps	 %xmm0,  %xmm2	unpcklps %xmm1,  %xmm0	unpckhps %xmm1,  %xmm2	movss	 0 * SIZE(B), %xmm1	movss	 1 * SIZE(B), %xmm5	subss	%xmm0,  %xmm1	subss	%xmm2,  %xmm5#else	movsd	 0 * SIZE(AO), %xmm8	subps	%xmm0, %xmm8#endif#ifdef LN	movaps	  0 * SIZE(AO), %xmm6	pshufd	 $0xff, %xmm6, %xmm8	mulss	 %xmm8, %xmm5	pshufd	 $0xaa, %xmm6, %xmm8	mulss	 %xmm5, %xmm8	subss	 %xmm8, %xmm1	pshufd	 $0x00, %xmm6, %xmm8	mulss	 %xmm8, %xmm1#endif#ifdef LT	movaps	 0 * SIZE(AO), %xmm6	pshufd	 $0x00, %xmm6, %xmm8	mulss	 %xmm8, %xmm1	pshufd	 $0x55, %xmm6, %xmm8	mulss	 %xmm1, %xmm8	subss	 %xmm8, %xmm5	pshufd	 $0xff, %xmm6, %xmm8	mulss	 %xmm8, %xmm5#endif#if defined(RN) || defined(RT)	movss	 0 * SIZE(B), %xmm0	pshufd	 $0x00, %xmm0, %xmm2	mulps	 %xmm2, %xmm8#endif#ifdef LN	subq	$2 * SIZE, CO1#endif#if defined(LN) || defined(LT)	movss	%xmm1,   0 * SIZE(B)	movss	%xmm5,   1 * SIZE(B)	pshufd	$0x00, %xmm1, %xmm2	movaps	%xmm2,   0 * SIZE(BO)	pshufd	$0x00, %xmm5, %xmm2	movaps	%xmm2,   4 * SIZE(BO)#else	movlps	%xmm8,   0 * SIZE(AO)#endif#if defined(LN) || defined(LT)	unpcklps %xmm10, %xmm1	unpcklps %xmm11, %xmm5	unpcklps %xmm5,  %xmm1	movlps	%xmm1,  0 * SIZE(CO1)#else	movlps	%xmm8,  0 * SIZE(CO1)#endif#ifndef LN	addq	$2 * SIZE, CO1#endif#if defined(LT) || defined(RN)	movq	K,  %rax	subq	KK, %rax
⌨️ 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 + -