📄 gemm_kernel_8x4_sse3.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 5 页
字号:
	mulps	%xmm10, %xmm11	movddup	 14 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movsd    56 * SIZE(BO), %xmm11	shufps	  $0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm2	movsd    60 * SIZE(BO), %xmm11	shufps	  $0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11	movddup	 24 * SIZE(AO), %xmm10	addps	%xmm11, %xmm3	movsd    96 * SIZE(BO), %xmm11	addq   $16 * SIZE, AO	addq   $64 * 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:	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movsd     4 * SIZE(BO), %xmm9	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	movddup	  2 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsd     8 * SIZE(BO), %xmm9	addq	$2 * SIZE, AO	addq	$8 * SIZE, BO	decq	%rax	jg	.L36	ALIGN_4.L38:#if! defined(TRMMKERNEL) && !defined(BETAZERO)	movsd	0 * SIZE(CO1), %xmm8	movhps	0 * SIZE(CO2), %xmm8	movsd	0 * SIZE(CO1, LDC, 2), %xmm9	movhps	0 * SIZE(CO2, LDC, 2), %xmm9#endif	addps	%xmm2, %xmm0	addps	%xmm3, %xmm1	mulps	%xmm15, %xmm0	mulps	%xmm15, %xmm1#if! defined(TRMMKERNEL) && !defined(BETAZERO)	addps	%xmm8, %xmm0	addps	%xmm9, %xmm1#endif	movsd	%xmm0, 0 * SIZE(CO1)	movhps	%xmm0, 0 * SIZE(CO2)	movsd	%xmm1, 0 * SIZE(CO1, LDC, 2)	movhps	%xmm1, 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, 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	.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#endif		movss	 0 * SIZE(AO), %xmm8	movss	 4 * SIZE(AO), %xmm10	movsd	 0 * SIZE(BO), %xmm9	movsd	32 * 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	$4, %rax#endif	movq	%rax, KKK#endif	sarq	$3, %rax	je	.L45	ALIGN_4.L42:	shufps	$0, %xmm8, %xmm8	movhps	 4 * SIZE(BO), %xmm9	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)	mulps	%xmm8, %xmm9	movss	 1 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movsd	 8 * SIZE(BO), %xmm9	shufps	$0, %xmm8, %xmm8	movhps	12 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movss	 2 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsd	16 * SIZE(BO), %xmm9	shufps	$0, %xmm8, %xmm8	movhps	20 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movss	 3 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movsd	24 * SIZE(BO), %xmm9	shufps	$0, %xmm8, %xmm8	movhps	28 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movss	 8 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsd	64 * SIZE(BO), %xmm9	shufps	$0, %xmm10, %xmm10	movhps	36 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	movss	 5 * SIZE(AO), %xmm10	addps	%xmm11, %xmm0	movsd	40 * SIZE(BO), %xmm11	shufps	$0, %xmm10, %xmm10	movhps	44 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	movss	 6 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movsd	48 * SIZE(BO), %xmm11	shufps	$0, %xmm10, %xmm10	movhps	52 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	movss	 7 * SIZE(AO), %xmm10	addps	%xmm11, %xmm0	movsd	56 * SIZE(BO), %xmm11	shufps	$0, %xmm10, %xmm10	movhps	60 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm11	movss	12 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movsd	96 * SIZE(BO), %xmm11	addq   $ 8 * SIZE, AO	addq   $64 * 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:	shufps	$0, %xmm8, %xmm8	movhps	 4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movss	 1 * SIZE(AO), %xmm8	addps	%xmm9, %xmm0	movsd	 8 * SIZE(BO), %xmm9	addq	$1 * SIZE, AO		# aoffset  += 4	addq	$8 * SIZE, BO		# boffset1 += 8	decq	%rax	jg	.L46	ALIGN_4.L48:#if! defined(TRMMKERNEL) && !defined(BETAZERO)	movss	0 * SIZE(CO1), %xmm8	movss	0 * SIZE(CO2), %xmm9	movss	0 * SIZE(CO1, LDC, 2), %xmm10	movss	0 * SIZE(CO2, LDC, 2), %xmm11#endif	addps	%xmm1, %xmm0	mulps	%xmm15, %xmm0	movhlps %xmm0, %xmm1#if! defined(TRMMKERNEL) && !defined(BETAZERO)	addss	%xmm0, %xmm8	psrlq	$32, %xmm0	addss	%xmm0, %xmm9	addss	%xmm1, %xmm10	psrlq	$32, %xmm1	addss	%xmm1, %xmm11	movss	%xmm8,  0 * SIZE(CO1)	movss	%xmm9,  0 * SIZE(CO2)	movss	%xmm10, 0 * SIZE(CO1, LDC, 2)	movss	%xmm11, 0 * SIZE(CO2, LDC, 2)#else	movss	%xmm0,  0 * SIZE(CO1)	psrlq	$32, %xmm0	movss	%xmm0,  0 * SIZE(CO2)	movss	%xmm1,  0 * SIZE(CO1, LDC, 2)	psrlq	$32, %xmm1	movss	%xmm1,  0 * SIZE(CO2, LDC, 2)#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, 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	$3, %rax	jle	.L53	ALIGN_4	.L52:	movddup	 0 * SIZE(B), %xmm0	movddup	 2 * SIZE(B), %xmm1	movddup	 4 * SIZE(B), %xmm2	movddup	 6 * SIZE(B), %xmm3	movddup	 8 * SIZE(B), %xmm4	movddup	10 * SIZE(B), %xmm5	movddup	12 * SIZE(B), %xmm6	movddup	14 * SIZE(B), %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)	prefetcht1	128 * SIZE(BO)	prefetcht0	112 * SIZE(B)	addq	$16 * SIZE, B	addq	$32 * SIZE, BO	decq	%rax	jne	.L52	ALIGN_4.L53:	movq	K, %rax	andq	$7, %rax	BRANCH	jle	.L60	ALIGN_4.L54:	movddup	 0 * SIZE(B), %xmm0	movaps	%xmm0,  0 * SIZE(BO)	addq	$ 2 * SIZE, B	addq	$ 4 * 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, 2), BO#endif	 	movaps	  0 * SIZE(AO), %xmm8 	movaps	 16 * SIZE(AO), %xmm10 	movaps	 32 * SIZE(AO), %xmm12 	movaps	 48 * SIZE(AO), %xmm14	movsldup  0 * SIZE(BO), %xmm9	movsldup 16 * SIZE(BO), %xmm11	pxor	%xmm0, %xmm0	pxor	%xmm1, %xmm1	prefetcht2     4 * SIZE(CO1)	pxor	%xmm4, %xmm4	prefetcht2     4 * 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	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  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm4	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9 	movaps	  8 * SIZE(AO), %xmm8	addps	%xmm9, %xmm5	movsldup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movshdup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	 12 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsldup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm4	movshdup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9 	movaps	 64 * SIZE(AO), %xmm8	addps	%xmm9, %xmm5	movsldup  8 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm0	movshdup  8 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	movaps	 20 * SIZE(AO), %xmm10	addps	%xmm9, %xmm1	movsldup  8 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm4	movshdup  8 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9 	movaps	 24 * SIZE(AO), %xmm10	addps	%xmm9, %xmm5	movsldup 12 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm0	movshdup 12 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	movaps	 28 * SIZE(AO), %xmm10	addps	%xmm9, %xmm1	movsldup 12 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9	addps	%xmm9, %xmm4	movshdup 12 * SIZE(BO), %xmm9	mulps	%xmm10, %xmm9 	movaps	 80 * SIZE(AO), %xmm10	addps	%xmm9, %xmm5	movsldup 32 * SIZE(BO), %xmm9	mulps	%xmm12, %xmm11	PREFETCH  (PREFETCHSIZE + 32) * SIZE(AO)	addps	%xmm11, %xmm0	movshdup 16 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11	movaps	 36 * SIZE(AO), %xmm12	addps	%xmm11, %xmm1	movsldup 16 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11	addps	%xmm11, %xmm4	movshdup 16 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11 	movaps	 40 * SIZE(AO), %xmm12	addps	%xmm11, %xmm5	movsldup 20 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11	addps	%xmm11, %xmm0	movshdup 20 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11	movaps	 44 * SIZE(AO), %xmm12	addps	%xmm11, %xmm1	movsldup 20 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11	addps	%xmm11, %xmm4	movshdup 20 * SIZE(BO), %xmm11	mulps	%xmm12, %xmm11 	movaps	 96 * SIZE(AO), %xmm12	addps	%xmm11, %xmm5	movsldup 24 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	addps	%xmm11, %xmm0	movshdup 24 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	movaps	 52 * SIZE(AO), %xmm14	addps	%xmm11, %xmm1	movsldup 24 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	addps	%xmm11, %xmm4	movshdup 24 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11 	movaps	 56 * SIZE(AO), %xmm14	addps	%xmm11, %xmm5	movsldup 28 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	addps	%xmm11, %xmm0	movshdup 28 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	movaps	 60 * SIZE(AO), %xmm14	addps	%xmm11, %xmm1	movsldup 28 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11	addps	%xmm11, %xmm4	movshdup 28 * SIZE(BO), %xmm11	mulps	%xmm14, %xmm11 	movaps	112 * SIZE(AO), %xmm14	addps	%xmm11, %xmm5	movsldup 48 * SIZE(BO), %xmm11	addq   $64 * SIZE, AO	addq   $32 * 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	addps	%xmm9, %xmm0	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	  4 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsldup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm4	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9 	movaps	  8 * SIZE(AO), %xmm8	addps	%xmm9, %xmm5	movsldup  4 * SIZE(BO), %xmm9	addq	$8 * SIZE, AO	addq	$4 * SIZE, BO	decq	%rax	jg	.L66	ALIGN_4.L68:#if! defined(TRMMKERNEL) && !defined(BETAZERO)	movsd	0 * SIZE(CO1), %xmm8	movhps	2 * SIZE(CO1), %xmm8	movsd	4 * SIZE(CO1), %xmm9	movhps	6 * SIZE(CO1), %xmm9	movsd	0 * SIZE(CO2), %xmm10	movhps	2 * SIZE(CO2), %xmm10	movsd	4 * SIZE(CO2), %xmm11	movhps	6 * SIZE(CO2), %xmm11#endif	mulps	%xmm15, %xmm0	mulps	%xmm15, %xmm4	mulps	%xmm15, %xmm1	mulps	%xmm15, %xmm5#if! defined(TRMMKERNEL) && !defined(BETAZERO)	addps	%xmm8,  %xmm0	addps	%xmm9,  %xmm4	addps	%xmm10, %xmm1	addps	%xmm11, %xmm5#endif	movsd	%xmm0, 0 * SIZE(CO1)	movhps	%xmm0, 2 * SIZE(CO1)	movsd	%xmm4, 4 * SIZE(CO1)	movhps	%xmm4, 6 * SIZE(CO1)	movsd	%xmm1, 0 * SIZE(CO2)	movhps	%xmm1, 2 * SIZE(CO2)	movsd	%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, 2), BO#endif#if defined(TRMMKERNEL) && defined(LEFT)	addq	$8, KK#endif	addq	$8 * SIZE, CO1		# coffset += 4	addq	$8 * SIZE, CO2		# coffset += 4
⌨️ 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 + -