📄 gemm_kernel_8x4_sse.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 5 页
字号:
	mulps	20 * SIZE(BO), %xmm10	addps	%xmm11, %xmm4	movaps	24 * SIZE(BO), %xmm11	addps	%xmm10, %xmm5	movaps	 -8 * SIZE(AO), %xmm10	mulps	%xmm10, %xmm11	mulps	28 * SIZE(BO), %xmm10	addps	%xmm11, %xmm0	movaps	24 * SIZE(BO), %xmm11	addps	%xmm10, %xmm1	movaps	-4 * SIZE(AO), %xmm10	mulps	%xmm10, %xmm11	mulps	28 * SIZE(BO), %xmm10	addps	%xmm11, %xmm4	movaps	80 * SIZE(BO), %xmm11	addps	%xmm10, %xmm5	movaps	48 * SIZE(AO), %xmm10#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCHSIZE + 32) * SIZE(AO)#endif	mulps	%xmm12, %xmm13	mulps	36 * SIZE(BO), %xmm12	addps	%xmm13, %xmm0	movaps	32 * SIZE(BO), %xmm13	addps	%xmm12, %xmm1	movaps	 4 * SIZE(AO), %xmm12	mulps	%xmm12, %xmm13	mulps	36 * SIZE(BO), %xmm12	addps	%xmm13, %xmm4	movaps	40 * SIZE(BO), %xmm13	addps	%xmm12, %xmm5	movaps	 8 * SIZE(AO), %xmm12	mulps	%xmm12, %xmm13	mulps	44 * SIZE(BO), %xmm12	addps	%xmm13, %xmm0	movaps	40 * SIZE(BO), %xmm13	addps	%xmm12, %xmm1	movaps	12 * SIZE(AO), %xmm12	mulps	%xmm12, %xmm13	mulps	44 * SIZE(BO), %xmm12	addps	%xmm13, %xmm4	movaps	96 * SIZE(BO), %xmm13	addps	%xmm12, %xmm5	movaps	64 * SIZE(AO), %xmm12#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCHSIZE + 48) * SIZE(AO)#endif	mulps	%xmm14, %xmm15	mulps	52 * SIZE(BO), %xmm14	addps	%xmm15, %xmm0	movaps	48 * SIZE(BO), %xmm15	addps	%xmm14, %xmm1	movaps	20 * SIZE(AO), %xmm14	mulps	%xmm14, %xmm15	mulps	52 * SIZE(BO), %xmm14	addps	%xmm15, %xmm4	movaps	56 * SIZE(BO), %xmm15	addps	%xmm14, %xmm5	movaps	24 * SIZE(AO), %xmm14	mulps	%xmm14, %xmm15	mulps	60 * SIZE(BO), %xmm14	addps	%xmm15, %xmm0	movaps	56 * SIZE(BO), %xmm15	addps	%xmm14, %xmm1	movaps	28 * SIZE(AO), %xmm14	mulps	%xmm14, %xmm15	mulps	60 * SIZE(BO), %xmm14	addps	%xmm15, %xmm4	movaps	112 * SIZE(BO), %xmm15	addps	%xmm14, %xmm5	movaps	80 * SIZE(AO), %xmm14	addq   $64 * SIZE, AO	addq   $64 * SIZE, BO	decq   %rax	jne    .L62	ALIGN_4.L65:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movaps	ALPHA, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je .L68	ALIGN_4.L66:	mulps	%xmm8, %xmm9	mulps	 4 * SIZE(BO), %xmm8	addps	%xmm9, %xmm0	movaps	 0 * SIZE(BO), %xmm9	addps	%xmm8, %xmm1	movaps	-28 * SIZE(AO), %xmm8	mulps	%xmm8, %xmm9	mulps	 4 * SIZE(BO), %xmm8	addps	%xmm9, %xmm4	movaps	 8 * SIZE(BO), %xmm9	addps	%xmm8, %xmm5	movaps	-24 * SIZE(AO), %xmm8	addq	$8 * SIZE, AO		# aoffset  += 4	addq	$8 * SIZE, BO		# boffset1 += 8	decq	%rax	jg	.L66	ALIGN_4.L68:#ifndef TRMMKERNEL	movlps	0 * SIZE(CO1), %xmm8	movhps	2 * SIZE(CO1), %xmm8	movlps	4 * SIZE(CO1), %xmm9	movhps	6 * SIZE(CO1), %xmm9	movlps	0 * SIZE(CO2), %xmm10	movhps	2 * SIZE(CO2), %xmm10	movlps	4 * SIZE(CO2), %xmm11	movhps	6 * SIZE(CO2), %xmm11#endif	mulps	%xmm15, %xmm0	mulps	%xmm15, %xmm4	mulps	%xmm15, %xmm1	mulps	%xmm15, %xmm5#ifndef TRMMKERNEL	addps	%xmm8,  %xmm0	addps	%xmm9,  %xmm4	addps	%xmm10, %xmm1	addps	%xmm11, %xmm5#endif	movlps	%xmm0, 0 * SIZE(CO1)	movhps	%xmm0, 2 * SIZE(CO1)	movlps	%xmm4, 4 * SIZE(CO1)	movhps	%xmm4, 6 * SIZE(CO1)	movlps	%xmm1, 0 * SIZE(CO2)	movhps	%xmm1, 2 * SIZE(CO2)	movlps	%xmm5, 4 * SIZE(CO2)	movhps	%xmm5, 6 * SIZE(CO2)#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	movq	K, %rax	subq	KKK, %rax	leaq	(,%rax,    8), %rax	leaq	(AO, %rax, 4), AO	leaq	(BO, %rax, 4), BO#endif#if defined(TRMMKERNEL) && defined(LEFT)	addq	$8, KK#endif	addq	$8 * SIZE, CO1		# coffset += 4	addq	$8 * SIZE, CO2		# coffset += 4	decq	I			# i --	jg	.L61	ALIGN_4	.L70:	testq	$4, M	je	.L80#if !defined(TRMMKERNEL) || \	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	leaq	BUFFER, BO#else	leaq	BUFFER, BO	movq	KK, %rax	leaq	(, %rax,   8), %rax	leaq	(AO, %rax, 2), AO	leaq	(BO, %rax, 4), BO#endif		movaps	-32 * SIZE(AO), %xmm8	movaps	-16 * SIZE(AO), %xmm10	movaps	 0 * SIZE(BO), %xmm9	movaps	16 * SIZE(BO), %xmm11	movaps	32 * SIZE(BO), %xmm13	movaps	48 * SIZE(BO), %xmm15	pxor	%xmm0, %xmm0	pxor	%xmm1, %xmm1	pxor	%xmm2, %xmm2	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	$4, %rax#else	addq	$2, %rax#endif	movq	%rax, KKK#endif	sarq	$3, %rax	je	.L75	ALIGN_4.L72:	mulps	%xmm8, %xmm9#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)#endif	mulps	 4 * SIZE(BO), %xmm8	addps	%xmm9, %xmm0	movaps	 8 * SIZE(BO), %xmm9	addps	%xmm8, %xmm1	movaps	-28 * SIZE(AO), %xmm8	mulps	%xmm8, %xmm9	mulps	12 * SIZE(BO), %xmm8	addps	%xmm9, %xmm2	movaps	64 * SIZE(BO), %xmm9	addps	%xmm8, %xmm3	movaps	-24 * SIZE(AO), %xmm8	mulps	%xmm8, %xmm11	mulps	20 * SIZE(BO), %xmm8	addps	%xmm11, %xmm0	movaps	24 * SIZE(BO), %xmm11	addps	%xmm8, %xmm1	movaps	-20 * SIZE(AO), %xmm8	mulps	%xmm8, %xmm11	mulps	28 * SIZE(BO), %xmm8	addps	%xmm11, %xmm2	movaps	80 * SIZE(BO), %xmm11	addps	%xmm8, %xmm3	movaps	 0 * SIZE(AO), %xmm8	mulps	%xmm10, %xmm13	mulps	36 * SIZE(BO), %xmm10	addps	%xmm13, %xmm0	movaps	40 * SIZE(BO), %xmm13	addps	%xmm10, %xmm1	movaps	-12 * SIZE(AO), %xmm10	mulps	%xmm10, %xmm13	mulps	44 * SIZE(BO), %xmm10	addps	%xmm13, %xmm2	movaps	96 * SIZE(BO), %xmm13	addps	%xmm10, %xmm3	movaps	 -8 * SIZE(AO), %xmm10	mulps	%xmm10, %xmm15	mulps	52 * SIZE(BO), %xmm10	addps	%xmm15, %xmm0	movaps	56 * SIZE(BO), %xmm15	addps	%xmm10, %xmm1	movaps	 -4 * SIZE(AO), %xmm10	mulps	%xmm10, %xmm15	mulps	60 * SIZE(BO), %xmm10	addps	%xmm15, %xmm2	movaps	112 * SIZE(BO), %xmm15	addps	%xmm10, %xmm3	movaps	16 * SIZE(AO), %xmm10	addq   $32 * SIZE, AO	addq   $64 * SIZE, BO	decq   %rax	jne    .L72	ALIGN_4.L75:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movaps	ALPHA, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je .L78	ALIGN_4.L76:	mulps	%xmm8, %xmm9	mulps	 4 * SIZE(BO), %xmm8	addps	%xmm9, %xmm0	movaps	 8 * SIZE(BO), %xmm9	addps	%xmm8, %xmm1	movaps	-28 * SIZE(AO), %xmm8	addq	$4 * SIZE, AO		# aoffset  += 4	addq	$8 * SIZE, BO		# boffset1 += 8	decq	%rax	jg	.L76	ALIGN_4.L78:#ifndef TRMMKERNEL	movlps	0 * SIZE(CO1), %xmm8	movhps	2 * SIZE(CO1), %xmm8	movlps	0 * SIZE(CO2), %xmm10	movhps	2 * SIZE(CO2), %xmm10#endif	addps	%xmm2, %xmm0	addps	%xmm3, %xmm1	mulps	%xmm15, %xmm0	mulps	%xmm15, %xmm1#ifndef TRMMKERNEL	addps	%xmm8,  %xmm0	addps	%xmm10, %xmm1#endif	movlps	%xmm0, 0 * SIZE(CO1)	movhps	%xmm0, 2 * SIZE(CO1)	movlps	%xmm1, 0 * SIZE(CO2)	movhps	%xmm1, 2 * SIZE(CO2)#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	movq	K, %rax	subq	KKK, %rax	leaq	(,%rax,    8), %rax	leaq	(AO, %rax, 2), AO	leaq	(BO, %rax, 4), BO#endif#if defined(TRMMKERNEL) && defined(LEFT)	addq	$4, KK#endif	addq	$4 * SIZE, CO1		# coffset += 4	addq	$4 * SIZE, CO2		# coffset += 4	ALIGN_4	.L80:	testq	$2, M	je	.L90#if !defined(TRMMKERNEL) || \	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	leaq	BUFFER, BO#else	leaq	BUFFER, BO	movq	KK, %rax	leaq	(, %rax,   8), %rax	leaq	(AO, %rax, 1), AO	leaq	(BO, %rax, 4), BO#endif		movaps	-32 * SIZE(AO), %xmm8	movaps	-24 * SIZE(AO), %xmm10	movaps	 0 * SIZE(BO), %xmm9	movaps	16 * SIZE(BO), %xmm11	movaps	32 * SIZE(BO), %xmm13	movaps	48 * SIZE(BO), %xmm15	pxor	%xmm0, %xmm0	pxor	%xmm1, %xmm1	pxor	%xmm2, %xmm2	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	$2, %rax#endif	movq	%rax, KKK#endif	sarq	$3, %rax	je	.L85	ALIGN_4.L82:	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)#endif	movlps	 4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movlps	-30 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movlps	 8 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm2	movlps	12 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movlps	-28 * SIZE(AO), %xmm8	addps	%xmm9, %xmm3	movlps	64 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm11	addps	%xmm11, %xmm0	movlps	20 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	movlps	-26 * SIZE(AO), %xmm8	addps	%xmm11, %xmm1	movlps	24 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	addps	%xmm11, %xmm2	movlps	28 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	movlps	-16 * SIZE(AO), %xmm8	addps	%xmm11, %xmm3	movlps	 80 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm13	addps	%xmm13, %xmm0	movlps	36 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	movlps	-22 * SIZE(AO), %xmm10	addps	%xmm13, %xmm1	movlps	40 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	addps	%xmm13, %xmm2	movlps	44 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	movlps	-20 * SIZE(AO), %xmm10	addps	%xmm13, %xmm3	movlps	 96 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm15	addps	%xmm15, %xmm0	movlps	52 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	movlps	-18 * SIZE(AO), %xmm10	addps	%xmm15, %xmm1	movlps	56 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	addps	%xmm15, %xmm2	movlps	60 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	movlps	-8 * SIZE(AO), %xmm10	addps	%xmm15, %xmm3	movlps	112 * SIZE(BO), %xmm15	addq   $16 * SIZE, AO	addq   $64 * SIZE, BO	decq   %rax	jne    .L82	ALIGN_4.L85:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movaps	ALPHA, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je .L88	ALIGN_4.L86:	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movlps	  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movlps	-30 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movlps	  8 * SIZE(BO), %xmm9	addq	$2 * SIZE, AO		# aoffset  += 4	addq	$8 * SIZE, BO		# boffset1 += 8	decq	%rax	jg	.L86	ALIGN_4.L88:#ifndef TRMMKERNEL	movlps	0 * SIZE(CO1), %xmm8	movlps	0 * SIZE(CO2), %xmm10#endif	addps	%xmm2, %xmm0	addps	%xmm3, %xmm1	mulps	%xmm15, %xmm0	mulps	%xmm15, %xmm1#ifndef TRMMKERNEL	addps	%xmm8,  %xmm0	addps	%xmm10, %xmm1#endif	movlps	%xmm0, 0 * SIZE(CO1)	movlps	%xmm1, 0 * SIZE(CO2)#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	movq	K, %rax	subq	KKK, %rax	leaq	(,%rax,    8), %rax	leaq	(AO, %rax, 1), AO	leaq	(BO, %rax, 4), BO#endif#if defined(TRMMKERNEL) && defined(LEFT)	addq	$2, KK#endif	addq	$2 * SIZE, CO1		# coffset += 4	addq	$2 * SIZE, CO2		# coffset += 4	ALIGN_4	.L90:	testq	$1, M	je	.L99#if !defined(TRMMKERNEL) || \	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	leaq	BUFFER, BO#else	leaq	BUFFER, BO	movq	KK, %rax	leaq	(, %rax,   4), %rax	leaq	(AO, %rax, 1), AO	leaq	(BO, %rax, 8), BO#endif		movss	-32 * SIZE(AO), %xmm8	movss	-28 * SIZE(AO), %xmm10	movss	 0 * SIZE(BO), %xmm9	movss	16 * SIZE(BO), %xmm11	movss	32 * SIZE(BO), %xmm13	movss	48 * SIZE(BO), %xmm15	pxor	%xmm0, %xmm0	pxor	%xmm1, %xmm1	pxor	%xmm2, %xmm2	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	$1, %rax#else	addq	$2, %rax#endif	movq	%rax, KKK#endif	sarq	$3, %rax	je	.L95	ALIGN_4.L92:	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)#endif	movss	 4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movss	-31 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movss	 8 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm2	movss	12 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movss	-30 * SIZE(AO), %xmm8	addps	%xmm9, %xmm3	movss	64 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm11	addps	%xmm11, %xmm0	movss	20 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	movss	-29 * SIZE(AO), %xmm8	addps	%xmm11, %xmm1	movss	24 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	addps	%xmm11, %xmm2	movss	28 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	movss	-24 * SIZE(AO), %xmm8	addps	%xmm11, %xmm3	movss	 80 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm13	addps	%xmm13, %xmm0	movss	36 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	movss	-27 * SIZE(AO), %xmm10	addps	%xmm13, %xmm1	movss	40 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13
⌨️ 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 + -