📄 gemm_kernel_8x4_sse.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 5 页
字号:
	addps	%xmm15, %xmm1	movlps	56 * SIZE(BO), %xmm15	mulps	%xmm8, %xmm15	addps	%xmm15, %xmm2	movlps	60 * SIZE(BO), %xmm15	mulps	%xmm8, %xmm15	movlps	-16 * SIZE(AO), %xmm8	addps	%xmm15, %xmm3	movlps	112 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm0	movlps	68 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm1	movlps	72 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm2	movlps	76 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	movlps	-22 * SIZE(AO), %xmm10	addps	%xmm9, %xmm3	movlps	128 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm0	movlps	84 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm1	movlps	88 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm2	movlps	92 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	movlps	-20 * SIZE(AO), %xmm10	addps	%xmm11, %xmm3	movlps	144 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm13	addps	%xmm13, %xmm0	movlps	100 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	addps	%xmm13, %xmm1	movlps	104 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	addps	%xmm13, %xmm2	movlps	108 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	movlps	-18 * SIZE(AO), %xmm10	addps	%xmm13, %xmm3	movlps	160 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm15	addps	%xmm15, %xmm0	movlps	116 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	addps	%xmm15, %xmm1	movlps	120 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	addps	%xmm15, %xmm2	movlps	124 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	movlps	 -8 * SIZE(AO), %xmm10	addps	%xmm15, %xmm3	movlps	176 * SIZE(BO), %xmm15	addq   $ 16 * SIZE, AO	addq   $128 * SIZE, BO	decq   %rax	jne    .L32	ALIGN_4.L35:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movaps	ALPHA, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je .L38	ALIGN_4.L36:	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movlps	  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm1	movlps	  8 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm2	movlps	 12 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movlps	-30 * SIZE(AO), %xmm8	addps	%xmm9, %xmm3	movlps	 16 * SIZE(BO), %xmm9	addq	$ 2 * SIZE, AO		# aoffset  += 4	addq	$16 * SIZE, BO		# boffset1 += 8	decq	%rax	jg	.L36	ALIGN_4.L38:	mulps	%xmm15, %xmm0	mulps	%xmm15, %xmm1 	mulps	%xmm15, %xmm2	mulps	%xmm15, %xmm3#ifndef TRMMKERNEL	movlps	0 * SIZE(CO1), %xmm8	movlps	0 * SIZE(CO2), %xmm10	movlps	0 * SIZE(CO1, LDC, 2), %xmm12	movlps	0 * SIZE(CO2, LDC, 2), %xmm14	addps	%xmm8,  %xmm0	addps	%xmm10, %xmm1	addps	%xmm12, %xmm2	addps	%xmm14, %xmm3#endif	movlps	%xmm0, 0 * SIZE(CO1)	movlps	%xmm1, 0 * SIZE(CO2)	movlps	%xmm2, 0 * SIZE(CO1, LDC, 2)	movlps	%xmm3, 0 * SIZE(CO2, LDC, 2)#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, 8), 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	.L40:	testq	$1, M	je	.L49#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	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	$4, %rax#endif	movq	%rax, KKK#endif	sarq	$3, %rax	je	.L45	ALIGN_4.L42:	mulss	%xmm8, %xmm9	addss	%xmm9, %xmm0#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)#endif	movss	 4 * SIZE(BO), %xmm9	mulss	%xmm8, %xmm9	addss	%xmm9, %xmm1	movss	 8 * SIZE(BO), %xmm9	mulss	%xmm8, %xmm9	addss	%xmm9, %xmm2	movss	12 * SIZE(BO), %xmm9	mulss	%xmm8, %xmm9	movss	-31 * SIZE(AO), %xmm8	addss	%xmm9, %xmm3	movss	64 * SIZE(BO), %xmm9	mulss	%xmm8, %xmm11	addss	%xmm11, %xmm0	movss	20 * SIZE(BO), %xmm11	mulss	%xmm8, %xmm11	addss	%xmm11, %xmm1	movss	24 * SIZE(BO), %xmm11	mulss	%xmm8, %xmm11	addss	%xmm11, %xmm2	movss	28 * SIZE(BO), %xmm11	mulss	%xmm8, %xmm11	movss	-30 * SIZE(AO), %xmm8	addss	%xmm11, %xmm3	movss	80 * SIZE(BO), %xmm11	mulss	%xmm8, %xmm13	addss	%xmm13, %xmm0	movss	36 * SIZE(BO), %xmm13	mulss	%xmm8, %xmm13	addss	%xmm13, %xmm1	movss	40 * SIZE(BO), %xmm13	mulss	%xmm8, %xmm13	addss	%xmm13, %xmm2	movss	44 * SIZE(BO), %xmm13	mulss	%xmm8, %xmm13	movss	-29 * SIZE(AO), %xmm8	addss	%xmm13, %xmm3	movss	96 * SIZE(BO), %xmm13	mulss	%xmm8, %xmm15	addss	%xmm15, %xmm0	movss	52 * SIZE(BO), %xmm15	mulss	%xmm8, %xmm15	addss	%xmm15, %xmm1	movss	56 * SIZE(BO), %xmm15	mulss	%xmm8, %xmm15	addss	%xmm15, %xmm2	movss	60 * SIZE(BO), %xmm15	mulss	%xmm8, %xmm15	movss	-24 * SIZE(AO), %xmm8	addss	%xmm15, %xmm3	movss	112 * SIZE(BO), %xmm15	mulss	%xmm10, %xmm9	addss	%xmm9, %xmm0	movss	68 * SIZE(BO), %xmm9	mulss	%xmm10, %xmm9	addss	%xmm9, %xmm1	movss	72 * SIZE(BO), %xmm9	mulss	%xmm10, %xmm9	addss	%xmm9, %xmm2	movss	76 * SIZE(BO), %xmm9	mulss	%xmm10, %xmm9	movss	-27 * SIZE(AO), %xmm10	addss	%xmm9, %xmm3	movss	128 * SIZE(BO), %xmm9	mulss	%xmm10, %xmm11	addss	%xmm11, %xmm0	movss	84 * SIZE(BO), %xmm11	mulss	%xmm10, %xmm11	addss	%xmm11, %xmm1	movss	88 * SIZE(BO), %xmm11	mulss	%xmm10, %xmm11	addss	%xmm11, %xmm2	movss	92 * SIZE(BO), %xmm11	mulss	%xmm10, %xmm11	movss	-26 * SIZE(AO), %xmm10	addss	%xmm11, %xmm3	movss	144 * SIZE(BO), %xmm11	mulss	%xmm10, %xmm13	addss	%xmm13, %xmm0	movss	100 * SIZE(BO), %xmm13	mulss	%xmm10, %xmm13	addss	%xmm13, %xmm1	movss	104 * SIZE(BO), %xmm13	mulss	%xmm10, %xmm13	addss	%xmm13, %xmm2	movss	108 * SIZE(BO), %xmm13	mulss	%xmm10, %xmm13	movss	-25 * SIZE(AO), %xmm10	addss	%xmm13, %xmm3	movss	160 * SIZE(BO), %xmm13	mulss	%xmm10, %xmm15	addss	%xmm15, %xmm0	movss	116 * SIZE(BO), %xmm15	mulss	%xmm10, %xmm15	addss	%xmm15, %xmm1	movss	120 * SIZE(BO), %xmm15	mulss	%xmm10, %xmm15	addss	%xmm15, %xmm2	movss	124 * SIZE(BO), %xmm15	mulss	%xmm10, %xmm15	movss	-20 * SIZE(AO), %xmm10	addss	%xmm15, %xmm3	movss	176 * SIZE(BO), %xmm15	addq   $  8 * SIZE, AO	addq   $128 * SIZE, BO	decq   %rax	jne    .L42	ALIGN_4.L45:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movaps	ALPHA, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je .L48	ALIGN_4.L46:	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movss	 4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm1	movss	 8 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm2	movss	12 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movss	-31 * SIZE(AO), %xmm8	addps	%xmm9, %xmm3	movss	16 * SIZE(BO), %xmm9	addq	$ 1 * SIZE, AO		# aoffset  += 4	addq	$16 * SIZE, BO		# boffset1 += 8	decq	%rax	jg	.L46	ALIGN_4.L48:	mulss	%xmm15, %xmm0	mulss	%xmm15, %xmm1 	mulss	%xmm15, %xmm2	mulss	%xmm15, %xmm3#ifndef TRMMKERNEL	movss	0 * SIZE(CO1), %xmm8	movss	0 * SIZE(CO2), %xmm10	movss	0 * SIZE(CO1, LDC, 2), %xmm12	movss	0 * SIZE(CO2, LDC, 2), %xmm14	addss	%xmm8,  %xmm0	addss	%xmm10, %xmm1	addss	%xmm12, %xmm2	addss	%xmm14, %xmm3#endif	movss	%xmm0, 0 * SIZE(CO1)	movss	%xmm1, 0 * SIZE(CO2)	movss	%xmm2, 0 * SIZE(CO1, LDC, 2)	movss	%xmm3, 0 * SIZE(CO2, LDC, 2)#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	movq	K, %rax	subq	KKK, %rax	leaq	(,%rax,    4), %rax	leaq	(AO, %rax, 1), AO	leaq	(BO, %rax, 8), BO	leaq	(BO, %rax, 8), BO#endif#if defined(TRMMKERNEL) && defined(LEFT)	addq	$1, KK#endif	ALIGN_4		.L49:#if defined(TRMMKERNEL) && !defined(LEFT)	addl	$4, KK#endif	leaq	(C, LDC, 4), C		# c += 4 * ldc	decq	J			# j --	jg	.L01.L50:	testq	$2, N	je	.L100.L51:#if defined(TRMMKERNEL) && defined(LEFT)	movq	OFFSET, %rax	movq	%rax, KK#endif	/* Copying to Sub Buffer */	leaq	BUFFER, BO		movq	K, %rax	sarq	$2, %rax	jle	.L53	ALIGN_4	.L52:#if defined(PENTIUM4) || defined(GENERIC)	movss	 0 * SIZE(B), %xmm0	movss	 1 * SIZE(B), %xmm1	movss	 2 * SIZE(B), %xmm2	movss	 3 * SIZE(B), %xmm3	movss	 4 * SIZE(B), %xmm4	movss	 5 * SIZE(B), %xmm5	movss	 6 * SIZE(B), %xmm6	movss	 7 * SIZE(B), %xmm7	PREFETCH	 32 * SIZE(B)	shufps	 $0, %xmm0, %xmm0	shufps	 $0, %xmm1, %xmm1	shufps	 $0, %xmm2, %xmm2	shufps	 $0, %xmm3, %xmm3	shufps	 $0, %xmm4, %xmm4	shufps	 $0, %xmm5, %xmm5	shufps	 $0, %xmm6, %xmm6	shufps	 $0, %xmm7, %xmm7	movaps	%xmm0,  0 * SIZE(BO)	movaps	%xmm1,  4 * SIZE(BO)	movaps	%xmm2,  8 * SIZE(BO)	movaps	%xmm3, 12 * SIZE(BO)	movaps	%xmm4, 16 * SIZE(BO)	movaps	%xmm5, 20 * SIZE(BO)	movaps	%xmm6, 24 * SIZE(BO)	movaps	%xmm7, 28 * SIZE(BO)	addq	$ 8 * SIZE, B	addq	$32 * SIZE, BO#endif#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	 32 * SIZE(B)	movd	 0 * SIZE(B), %mm0	movd	 1 * SIZE(B), %mm1	movd	 2 * SIZE(B), %mm2	movd	 3 * SIZE(B), %mm3	movd	 4 * SIZE(B), %mm4	movd	 5 * SIZE(B), %mm5	movd	 6 * SIZE(B), %mm6	movd	 7 * SIZE(B), %mm7	punpckldq %mm0, %mm0	punpckldq %mm1, %mm1	punpckldq %mm2, %mm2	punpckldq %mm3, %mm3	punpckldq %mm4, %mm4	punpckldq %mm5, %mm5	punpckldq %mm6, %mm6	punpckldq %mm7, %mm7		movq	%mm0,  0 * SIZE(BO)	movq	%mm0,  2 * SIZE(BO)	movq	%mm1,  4 * SIZE(BO)	movq	%mm1,  6 * SIZE(BO)	movq	%mm2,  8 * SIZE(BO)	movq	%mm2, 10 * SIZE(BO)	movq	%mm3, 12 * SIZE(BO)	movq	%mm3, 14 * SIZE(BO)	movq	%mm4, 16 * SIZE(BO)	movq	%mm4, 18 * SIZE(BO)	movq	%mm5, 20 * SIZE(BO)	movq	%mm5, 22 * SIZE(BO)	movq	%mm6, 24 * SIZE(BO)	movq	%mm6, 26 * SIZE(BO)	movq	%mm7, 28 * SIZE(BO)	movq	%mm7, 30 * SIZE(BO)	addq	$ 8 * SIZE, B	addq	$32 * SIZE, BO#endif	decq	%rax	jne	.L52	ALIGN_4.L53:	movq	K, %rax	andq	$3, %rax	BRANCH	jle	.L60	ALIGN_4.L54:#if defined(PENTIUM4) || defined(GENERIC)	movss	 0 * SIZE(B), %xmm0	movss	 1 * SIZE(B), %xmm1	shufps	 $0, %xmm0, %xmm0	shufps	 $0, %xmm1, %xmm1	movaps	%xmm0,  0 * SIZE(BO)	movaps	%xmm1,  4 * SIZE(BO)#endif#if defined(OPTERON) || defined(BARCELONA)	movd	 0 * SIZE(B), %mm0	movd	 1 * SIZE(B), %mm1	punpckldq %mm0, %mm0	punpckldq %mm1, %mm1	movq	%mm0,  0 * SIZE(BO)	movq	%mm0,  2 * SIZE(BO)	movq	%mm1,  4 * SIZE(BO)	movq	%mm1,  6 * SIZE(BO)#endif	addq	$ 2 * SIZE, B	addq	$ 8 * SIZE, BO	decq	%rax	jne	.L54	ALIGN_4	.L60:	movq	C, CO1			# coffset1 = c	leaq	(C, LDC, 1), CO2	# coffset2 = c + ldc	movq	A, AO		# aoffset = a	movq	M,  I	sarq	$3, I	# i = (m >> 3)	jle	.L70	ALIGN_4.L61:#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, 4), AO	leaq	(BO, %rax, 4), BO#endif		movaps	-32 * SIZE(AO), %xmm8	movaps	-16 * SIZE(AO), %xmm10	movaps	  0 * SIZE(AO), %xmm12	movaps	 16 * SIZE(AO), %xmm14	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	PREFETCHW      7 * SIZE(CO1)	pxor	%xmm4, %xmm4	PREFETCHW      7 * SIZE(CO2)	pxor	%xmm5, %xmm5#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	$8, %rax#else	addq	$2, %rax#endif	movq	%rax, KKK#endif	sarq	$3, %rax	je	.L65	ALIGN_4.L62:	mulps	%xmm8, %xmm9#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCHSIZE +  0) * SIZE(AO)#endif	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	mulps	%xmm8, %xmm9	mulps	12 * SIZE(BO), %xmm8	addps	%xmm9, %xmm0	movaps	 8 * SIZE(BO), %xmm9	addps	%xmm8, %xmm1	movaps	-20 * SIZE(AO), %xmm8	mulps	%xmm8, %xmm9	mulps	12 * SIZE(BO), %xmm8	addps	%xmm9, %xmm4	movaps	64 * SIZE(BO), %xmm9	addps	%xmm8, %xmm5	movaps	32 * SIZE(AO), %xmm8#if defined(OPTERON) || defined(BARCELONA)	PREFETCH	(PREFETCHSIZE + 16) * SIZE(AO)#endif	mulps	%xmm10, %xmm11	mulps	20 * SIZE(BO), %xmm10	addps	%xmm11, %xmm0	movaps	16 * SIZE(BO), %xmm11	addps	%xmm10, %xmm1	movaps	-12 * SIZE(AO), %xmm10	mulps	%xmm10, %xmm11
⌨️ 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 + -