📄 gemm_kernel_8x4_sse3.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 5 页
字号:
	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, 2), BO#endif	 	movaps	  0 * SIZE(AO), %xmm8	movsldup  0 * SIZE(BO), %xmm9 	movaps	 16 * SIZE(AO), %xmm10	movsldup 16 * SIZE(BO), %xmm11	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	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)	addps	%xmm9, %xmm0	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9 	movaps	  4 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsldup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm2	movshdup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9 	movaps	  8 * SIZE(AO), %xmm8	addps	%xmm9, %xmm3	movsldup  8 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movshdup  8 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9 	movaps	 12 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsldup 12 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm2	movshdup 12 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9 	movaps	 32 * SIZE(AO), %xmm8	addps	%xmm9, %xmm3	movsldup 32 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm0	movshdup 16 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11 	movaps	 20 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movsldup 20 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm2	movshdup 20 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11 	movaps	 24 * SIZE(AO), %xmm10	addps	%xmm11, %xmm3	movsldup 24 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm0	movshdup 24 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11 	movaps	 28 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movsldup 28 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm2	movshdup 28 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11 	movaps	 48 * SIZE(AO), %xmm10	addps	%xmm11, %xmm3	movsldup 48 * SIZE(BO), %xmm11	addq   $32 * SIZE, AO	addq   $32 * 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	addps	%xmm9, %xmm0	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9 	movaps	  4 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsldup  4 * SIZE(BO), %xmm9	addq	$4 * SIZE, AO	addq	$4 * SIZE, BO	decq	%rax	jg	.L76	ALIGN_4.L78:#if! defined(TRMMKERNEL) && !defined(BETAZERO)	movsd	0 * SIZE(CO1), %xmm8	movhps	2 * SIZE(CO1), %xmm8	movsd	0 * SIZE(CO2), %xmm10	movhps	2 * SIZE(CO2), %xmm10#endif	addps	%xmm2, %xmm0	addps	%xmm3, %xmm1	mulps	%xmm15, %xmm0	mulps	%xmm15, %xmm1#if! defined(TRMMKERNEL) && !defined(BETAZERO)	addps	%xmm8,  %xmm0	addps	%xmm10, %xmm1#endif	movsd	%xmm0, 0 * SIZE(CO1)	movhps	%xmm0, 2 * SIZE(CO1)	movsd	%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, 2), 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, 2), BO#endif		movddup	  0 * SIZE(AO), %xmm8	movddup	  8 * SIZE(AO), %xmm10	movsd     0 * SIZE(BO), %xmm9	movsd    16 * SIZE(BO), %xmm11	pxor	%xmm0, %xmm0	pxor	%xmm1, %xmm1#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:	shufps	  $0x50, %xmm9, %xmm9	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)	mulps	%xmm8, %xmm9	movddup	  2 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movsd     4 * SIZE(BO), %xmm9	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	movddup	  4 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsd     8 * SIZE(BO), %xmm9	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	movddup	  6 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movsd    12 * SIZE(BO), %xmm9	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	movddup	 16 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsd    32 * SIZE(BO), %xmm9	shufps	  $0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11	movddup	 10 * SIZE(AO), %xmm10	addps	%xmm11, %xmm0	movsd    20 * SIZE(BO), %xmm11	shufps	  $0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11	movddup	 12 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movsd    24 * SIZE(BO), %xmm11	shufps	  $0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11	movddup	 14 * SIZE(AO), %xmm10	addps	%xmm11, %xmm0	movsd    28 * SIZE(BO), %xmm11	shufps	  $0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11	movddup	 24 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movsd    48 * SIZE(BO), %xmm11	addq   $16 * SIZE, AO	addq   $32 * 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:	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	movddup	  2 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movsd     4 * SIZE(BO), %xmm9	addq	$2 * SIZE, AO	addq	$4 * SIZE, BO	decq	%rax	jg	.L86	ALIGN_4.L88:#if! defined(TRMMKERNEL) && !defined(BETAZERO)	movsd	0 * SIZE(CO1), %xmm8	movhps	0 * SIZE(CO2), %xmm8#endif	addps	%xmm1, %xmm0	mulps	%xmm15, %xmm0#if! defined(TRMMKERNEL) && !defined(BETAZERO)	addps	%xmm8,  %xmm0#endif	movsd	%xmm0, 0 * SIZE(CO1)	movhps	%xmm0, 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, 2), 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, 4), BO#endif		movss	 0 * SIZE(AO), %xmm8	movss	 4 * SIZE(AO), %xmm10	movsd	 0 * SIZE(BO), %xmm9	movsd	16 * SIZE(BO), %xmm11	pxor	%xmm0, %xmm0	pxor	%xmm1, %xmm1#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:	shufps	$0, %xmm8, %xmm8	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)	mulps	%xmm8, %xmm9	movss	 1 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movsd	 4 * SIZE(BO), %xmm9	shufps	$0, %xmm8, %xmm8	mulps	%xmm8, %xmm9	movss	 2 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsd	 8 * SIZE(BO), %xmm9	shufps	$0, %xmm8, %xmm8	mulps	%xmm8, %xmm9	movss	 3 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movsd	12 * SIZE(BO), %xmm9	shufps	$0, %xmm8, %xmm8	mulps	%xmm8, %xmm9	movss	 8 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsd	32 * SIZE(BO), %xmm9	shufps	$0, %xmm10, %xmm10	mulps	%xmm10, %xmm11	movss	 5 * SIZE(AO), %xmm10	addps	%xmm11, %xmm0	movsd	20 * SIZE(BO), %xmm11	shufps	$0, %xmm10, %xmm10	mulps	%xmm10, %xmm11	movss	 6 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movsd	24 * SIZE(BO), %xmm11	shufps	$0, %xmm10, %xmm10	mulps	%xmm10, %xmm11	movss	 7 * SIZE(AO), %xmm10	addps	%xmm11, %xmm0	movsd	28 * SIZE(BO), %xmm11	shufps	$0, %xmm10, %xmm10	mulps	%xmm10, %xmm11	movss	12 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movsd	48 * SIZE(BO), %xmm11	addq   $ 8 * SIZE, AO	addq   $32 * SIZE, BO	decq   %rax	jne    .L92	ALIGN_4.L95:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movaps	ALPHA, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je .L98	ALIGN_4.L96:	shufps	$0, %xmm8, %xmm8	mulps	%xmm8, %xmm9	movss	 1 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movsd	 4 * SIZE(BO), %xmm9	addq	$1 * SIZE, AO	addq	$4 * SIZE, BO	decq	%rax	jg	.L96	ALIGN_4.L98:#if! defined(TRMMKERNEL) && !defined(BETAZERO)	movss	0 * SIZE(CO1), %xmm8	movss	0 * SIZE(CO2), %xmm9	addps	%xmm1, %xmm0	mulps	%xmm15, %xmm0	addss	%xmm0, %xmm8	psrlq	$32, %xmm0	addss	%xmm0, %xmm9	movss	%xmm8, 0 * SIZE(CO1)	movss	%xmm9, 0 * SIZE(CO2)#else	addps	%xmm1, %xmm0	mulps	%xmm15, %xmm0	movss	%xmm0, 0 * SIZE(CO1)	psrlq	$32, %xmm0	movss	%xmm0, 0 * SIZE(CO2)#endif#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, 4), BO#endif#if defined(TRMMKERNEL) && defined(LEFT)	addq	$1, KK#endif	ALIGN_4		.L99:#if defined(TRMMKERNEL) && !defined(LEFT)	addl	$2, KK#endif	leaq	(C, LDC, 2), C		# c += 4 * ldc	ALIGN_4.L100:	testq	$1, N	je	.L999.L101:#if defined(TRMMKERNEL) && defined(LEFT)	movq	OFFSET, %rax	movq	%rax, KK#endif	/* Copying to Sub Buffer */	leaq	BUFFER, BO		movq	K, %rax	sarq	$3, %rax	jle	.L103	ALIGN_4	.L102:	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	movss	%xmm0,  0 * SIZE(BO)	movss	%xmm0,  1 * SIZE(BO)	movss	%xmm1,  2 * SIZE(BO)	movss	%xmm1,  3 * SIZE(BO)	movss	%xmm2,  4 * SIZE(BO)	movss	%xmm2,  5 * SIZE(BO)	movss	%xmm3,  6 * SIZE(BO)	movss	%xmm3,  7 * SIZE(BO)	movss	%xmm4,  8 * SIZE(BO)	movss	%xmm4,  9 * SIZE(BO)	movss	%xmm5, 10 * SIZE(BO)	movss	%xmm5, 11 * SIZE(BO)	movss	%xmm6, 12 * SIZE(BO)	movss	%xmm6, 13 * SIZE(BO)	movss	%xmm7, 14 * SIZE(BO)	movss	%xmm7, 15 * SIZE(BO)	addq	$ 8 * SIZE, B	addq	$16 * SIZE, BO	decq	%rax	jne	.L102	ALIGN_4.L103:	movq	K, %rax	andq	$7, %rax	BRANCH	jle	.L110	ALIGN_4.L104:	movss	 0 * SIZE(B), %xmm0	movss	%xmm0,  0 * SIZE(BO)	movss	%xmm0,  1 * SIZE(BO)	addq	$ 1 * SIZE, B	addq	$ 2 * SIZE, BO	decq	%rax	jne	.L104	ALIGN_4	.L110:	movq	C, CO1			# coffset1 = c	movq	A, AO		# aoffset = a	movq	M,  I	sarq	$3, I	# i = (m >> 3)	jle	.L120	ALIGN_4.L111:#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, 1), BO#endif	 	movaps	  0 * SIZE(AO), %xmm8 	movaps	 16 * SIZE(AO), %xmm10 	movaps	 32 * SIZE(AO), %xmm12 	movaps	 48 * SIZE(AO), %xmm14	movddup   0 * SIZE(BO), %xmm9	movddup   8 * SIZE(BO), %xmm11	pxor	%xmm0, %xmm0	pxor	%xmm1, %xmm1	prefetchnta      8 * SIZE(CO1)	pxor	%xmm4, %xmm4	pxor	%xmm5, %xmm5
⌨️ 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 + -