📄 gemm_kernel_4x4_sse.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 4 页
字号:
	KERNEL8(32  *  4)	cmpl	$128 *  5, %eax	jle	.L12	KERNEL1(32  *  5)	KERNEL2(32  *  5)	KERNEL3(32  *  5)	KERNEL4(32  *  5)	KERNEL5(32  *  5)	KERNEL6(32  *  5)	KERNEL7(32  *  5)	KERNEL8(32  *  5)	cmpl	$128 *  6, %eax	jle	.L12	KERNEL1(32  *  6)	KERNEL2(32  *  6)	KERNEL3(32  *  6)	KERNEL4(32  *  6)	KERNEL5(32  *  6)	KERNEL6(32  *  6)	KERNEL7(32  *  6)	KERNEL8(32  *  6)	cmpl	$128 *  7, %eax	jle	.L12	KERNEL1(32  *  7)	KERNEL2(32  *  7)	KERNEL3(32  *  7)	KERNEL4(32  *  7)	KERNEL5(32  *  7)	KERNEL6(32  *  7)	KERNEL7(32  *  7)	KERNEL8(32  *  7)	addl	$128 * 8  * SIZE, BB	addl	$128 * 2  * SIZE, AA	subl	$128 * 8, %eax	jg	.L1X	jmp	.L15.L12:	leal	(AA, %eax, 1), AA	leal	(BB, %eax, 4), BB	ALIGN_4#else	sarl	$3, %eax	je	.L15	ALIGN_4.L12:	KERNEL1(32  *  0)	KERNEL2(32  *  0)	KERNEL3(32  *  0)	KERNEL4(32  *  0)	KERNEL5(32  *  0)	KERNEL6(32  *  0)	KERNEL7(32  *  0)	KERNEL8(32  *  0)	addl   $128 * SIZE, BB	addl   $32 * SIZE, AA	decl   %eax	jne    .L12	ALIGN_4#endif.L15:#ifndef TRMMKERNEL	movl	K, %eax#else	movl	KKK, %eax#endif	movaps	ALPHA,  %xmm3	andl	$7, %eax		# if (k & 1)	BRANCH	je .L18	ALIGN_4.L16:	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm4	movaps	 4 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm5	movaps	 8 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	mulps	12 * SIZE(BB), %xmm0	addps	%xmm2, %xmm6	movaps	16 * SIZE(BB), %xmm2	addps	%xmm0, %xmm7	movaps	 4 * SIZE(AA), %xmm0	addl	$ 4 * SIZE, AA	addl	$16 * SIZE, BB	decl	%eax	jg	.L16	ALIGN_4.L18:	leal	(LDC, LDC, 2), %eax#ifndef TRMMKERNEL	shufps	$0xe4, %xmm0, %xmm0	shufps	$0xe4, %xmm1, %xmm1	shufps	$0xe4, %xmm2, %xmm2	shufps	$0xe4, %xmm3, %xmm3	mulps	%xmm3, %xmm4	movsd	0 * SIZE(%esi), %xmm0	movhps	2 * SIZE(%esi), %xmm0	mulps	%xmm3, %xmm5	movsd	0 * SIZE(%esi, LDC, 1), %xmm1	movhps	2 * SIZE(%esi, LDC, 1), %xmm1	mulps	%xmm3, %xmm6	movsd	0 * SIZE(%esi, LDC, 2), %xmm2	movhps	2 * SIZE(%esi, LDC, 2), %xmm2	mulps	%xmm3, %xmm7	movsd	0 * SIZE(%esi, %eax, 1), %xmm3	movhps	2 * SIZE(%esi, %eax, 1), %xmm3	addps	%xmm0, %xmm4	addps	%xmm1, %xmm5	addps	%xmm2, %xmm6	addps	%xmm3, %xmm7#else	mulps	%xmm3, %xmm4	mulps	%xmm3, %xmm5	mulps	%xmm3, %xmm6	mulps	%xmm3, %xmm7#endif	movsd	%xmm4, 0 * SIZE(%esi)	movhps	%xmm4, 2 * SIZE(%esi)	movsd	%xmm5, 0 * SIZE(%esi, LDC, 1)	movhps	%xmm5, 2 * SIZE(%esi, LDC, 1)	movsd	%xmm6, 0 * SIZE(%esi, LDC, 2)	movhps	%xmm6, 2 * SIZE(%esi, LDC, 2)	movsd	%xmm7, 0 * SIZE(%esi, %eax, 1)	movhps	%xmm7, 2 * SIZE(%esi, %eax, 1)#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	movl	K, %eax	subl	KKK, %eax	leal	(,%eax, 8), %eax	leal	(AA, %eax, 2), AA	leal	(BB, %eax, 8), BB#endif#if defined(TRMMKERNEL) && defined(LEFT)	addl	$4, KK#endif	addl	$4 * SIZE, %esi		# coffset += 2	decl	%ebx			# i --	jg	.L11	ALIGN_4.L20:	testl	$2, M	je	.L30#if !defined(TRMMKERNEL) || \	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	leal	BUFFER, BB	# boffset1 = boffset#else	leal	BUFFER, BB	# boffset1 = boffset	movl	KK, %eax	leal	(, %eax,   8), %eax	leal	(AA, %eax, 1), AA	leal	(BB, %eax, 8), BB#endif		movsd	 0 * SIZE(AA), %xmm0	pxor	%xmm4, %xmm4	movsd	 8 * SIZE(AA), %xmm1	pxor	%xmm5, %xmm5	movsd	 0 * SIZE(BB), %xmm2	pxor	%xmm6, %xmm6	movsd	16 * SIZE(BB), %xmm3	pxor	%xmm7, %xmm7#ifndef TRMMKERNEL	movl	K, %eax#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	movl	K, %eax	subl	KK, %eax	movl	%eax, KKK	#else	movl	KK, %eax#ifdef LEFT	addl	$2, %eax#else	addl	$4, %eax#endif	movl	%eax, KKK#endif	sarl	$3, %eax	je	.L25	ALIGN_4.L22:	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm4#if defined(OPTERON) || defined(BARCELONA)	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)#endif	movsd	 4 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm5	movsd	 8 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm6	movsd	12 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	movsd	 2 * SIZE(AA), %xmm0	addps	%xmm2, %xmm7	movsd	32 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm3	addps	%xmm3, %xmm4	movsd	20 * SIZE(BB), %xmm3	mulps	%xmm0, %xmm3	addps	%xmm3, %xmm5	movsd	24 * SIZE(BB), %xmm3	mulps	%xmm0, %xmm3	addps	%xmm3, %xmm6	movsd	28 * SIZE(BB), %xmm3	mulps	%xmm0, %xmm3	movsd	 4 * SIZE(AA), %xmm0	addps	%xmm3, %xmm7	movsd	48 * SIZE(BB), %xmm3	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm4	movsd	36 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm5	movsd	40 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm6	movsd	44 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	movsd	 6 * SIZE(AA), %xmm0	addps	%xmm2, %xmm7	movsd	64 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm3	addps	%xmm3, %xmm4	movsd	52 * SIZE(BB), %xmm3	mulps	%xmm0, %xmm3	addps	%xmm3, %xmm5	movsd	56 * SIZE(BB), %xmm3	mulps	%xmm0, %xmm3	addps	%xmm3, %xmm6	movsd	60 * SIZE(BB), %xmm3	mulps	%xmm0, %xmm3	movsd	 16 * SIZE(AA), %xmm0	addps	%xmm3, %xmm7	movsd	80 * SIZE(BB), %xmm3	mulps	%xmm1, %xmm2	addps	%xmm2, %xmm4	movsd	68 * SIZE(BB), %xmm2	mulps	%xmm1, %xmm2	addps	%xmm2, %xmm5	movsd	72 * SIZE(BB), %xmm2	mulps	%xmm1, %xmm2	addps	%xmm2, %xmm6	movsd	76 * SIZE(BB), %xmm2	mulps	%xmm1, %xmm2	movsd	10 * SIZE(AA), %xmm1	addps	%xmm2, %xmm7	movsd	96 * SIZE(BB), %xmm2	mulps	%xmm1, %xmm3	addps	%xmm3, %xmm4	movsd	84 * SIZE(BB), %xmm3	mulps	%xmm1, %xmm3	addps	%xmm3, %xmm5	movsd	88 * SIZE(BB), %xmm3	mulps	%xmm1, %xmm3	addps	%xmm3, %xmm6	movsd	92 * SIZE(BB), %xmm3	mulps	%xmm1, %xmm3	movsd	12 * SIZE(AA), %xmm1	addps	%xmm3, %xmm7	movsd	112 * SIZE(BB), %xmm3	mulps	%xmm1, %xmm2	addps	%xmm2, %xmm4	movsd	100 * SIZE(BB), %xmm2	mulps	%xmm1, %xmm2	addps	%xmm2, %xmm5	movsd	104 * SIZE(BB), %xmm2	mulps	%xmm1, %xmm2	addps	%xmm2, %xmm6	movsd	108 * SIZE(BB), %xmm2	mulps	%xmm1, %xmm2	movsd	 14 * SIZE(AA), %xmm1	addps	%xmm2, %xmm7	movsd	128 * SIZE(BB), %xmm2	mulps	%xmm1, %xmm3	addps	%xmm3, %xmm4	movsd	116 * SIZE(BB), %xmm3	mulps	%xmm1, %xmm3	addps	%xmm3, %xmm5	movsd	120 * SIZE(BB), %xmm3	mulps	%xmm1, %xmm3	addps	%xmm3, %xmm6	movsd	124 * SIZE(BB), %xmm3	mulps	%xmm1, %xmm3	movsd	 24 * SIZE(AA), %xmm1	addps	%xmm3, %xmm7	movsd	144 * SIZE(BB), %xmm3	addl	$ 16 * SIZE, AA	addl	$128 * SIZE, BB	decl   %eax	jne    .L22	ALIGN_4.L25:#ifndef TRMMKERNEL	movl	K, %eax#else	movl	KKK, %eax#endif	movaps	ALPHA,  %xmm3	andl	$7, %eax		# if (k & 1)	BRANCH	je .L28	ALIGN_4.L26:	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm4	movsd	 4 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm5	movsd	 8 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	addps	%xmm2, %xmm6	movsd	12 * SIZE(BB), %xmm2	mulps	%xmm0, %xmm2	movsd	 2 * SIZE(AA), %xmm0	addps	%xmm2, %xmm7	movsd	16 * SIZE(BB), %xmm2	addl	$ 2 * SIZE, AA	addl	$16 * SIZE, BB	decl	%eax	jg	.L26	ALIGN_4.L28:	leal	(LDC, LDC, 2), %eax#ifndef TRMMKERNEL	mulps	%xmm3, %xmm4	movsd	0 * SIZE(%esi), %xmm0	mulps	%xmm3, %xmm5	movsd	0 * SIZE(%esi, LDC, 1), %xmm1	mulps	%xmm3, %xmm6	movsd	0 * SIZE(%esi, LDC, 2), %xmm2	mulps	%xmm3, %xmm7	movsd	0 * SIZE(%esi, %eax, 1), %xmm3	addps	%xmm0, %xmm4	addps	%xmm1, %xmm5	addps	%xmm2, %xmm6	addps	%xmm3, %xmm7#else	mulps	%xmm3, %xmm4	mulps	%xmm3, %xmm5	mulps	%xmm3, %xmm6	mulps	%xmm3, %xmm7#endif	movsd	%xmm4, 0 * SIZE(%esi)	movsd	%xmm5, 0 * SIZE(%esi, LDC, 1)	movsd	%xmm6, 0 * SIZE(%esi, LDC, 2)	movsd	%xmm7, 0 * SIZE(%esi, %eax, 1)#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	movl	K, %eax	subl	KKK, %eax	leal	(,%eax, 8), %eax	leal	(AA, %eax, 1), AA	leal	(BB, %eax, 8), BB#endif#if defined(TRMMKERNEL) && defined(LEFT)	addl	$2, KK#endif	addl	$2 * SIZE, %esi		# coffset += 2	ALIGN_4.L30:	testl	$1, M	je	.L39#if !defined(TRMMKERNEL) || \	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	leal	BUFFER, BB	# boffset1 = boffset#else	leal	BUFFER, BB	# boffset1 = boffset	movl	KK, %eax	leal	(, %eax,   4), %eax	leal	(AA, %eax, 1), AA	leal	(BB, %eax, 8), BB	leal	(BB, %eax, 8), BB#endif		movss	 0 * SIZE(AA), %xmm0	pxor	%xmm4, %xmm4	movss	 4 * SIZE(AA), %xmm1	pxor	%xmm5, %xmm5	movss	 0 * SIZE(BB), %xmm2	pxor	%xmm6, %xmm6	movss	16 * SIZE(BB), %xmm3	pxor	%xmm7, %xmm7#ifndef TRMMKERNEL	movl	K, %eax#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	movl	K, %eax	subl	KK, %eax	movl	%eax, KKK	#else	movl	KK, %eax#ifdef LEFT	addl	$1, %eax#else	addl	$4, %eax#endif	movl	%eax, KKK#endif	sarl	$3, %eax	je	.L35	ALIGN_4.L32:	mulss	%xmm0, %xmm2	addss	%xmm2, %xmm4#if defined(OPTERON) || defined(BARCELONA)	prefetcht0 (PREFETCHSIZE  + 0) * SIZE(AA)#endif	movss	 4 * SIZE(BB), %xmm2	mulss	%xmm0, %xmm2	addss	%xmm2, %xmm5	movss	 8 * SIZE(BB), %xmm2	mulss	%xmm0, %xmm2	mulss	12 * SIZE(BB), %xmm0	addss	%xmm2, %xmm6	movss	32 * SIZE(BB), %xmm2	addss	%xmm0, %xmm7	movss	 1 * SIZE(AA), %xmm0	mulss	%xmm0, %xmm3	addss	%xmm3, %xmm4	movss	20 * SIZE(BB), %xmm3	mulss	%xmm0, %xmm3	addss	%xmm3, %xmm5	movss	24 * SIZE(BB), %xmm3	mulss	%xmm0, %xmm3	mulss	28 * SIZE(BB), %xmm0	addss	%xmm3, %xmm6	movss	48 * SIZE(BB), %xmm3	addss	%xmm0, %xmm7	movss	 2 * SIZE(AA), %xmm0	mulss	%xmm0, %xmm2	addss	%xmm2, %xmm4	movss	36 * SIZE(BB), %xmm2	mulss	%xmm0, %xmm2	addss	%xmm2, %xmm5	movss	40 * SIZE(BB), %xmm2	mulss	%xmm0, %xmm2	mulss	44 * SIZE(BB), %xmm0	addss	%xmm2, %xmm6	movss	64 * SIZE(BB), %xmm2	addss	%xmm0, %xmm7	movss	 3 * SIZE(AA), %xmm0	mulss	%xmm0, %xmm3	addss	%xmm3, %xmm4	movss	52 * SIZE(BB), %xmm3	mulss	%xmm0, %xmm3	addss	%xmm3, %xmm5	movss	56 * SIZE(BB), %xmm3	mulss	%xmm0, %xmm3	mulss	60 * SIZE(BB), %xmm0	addss	%xmm3, %xmm6	movss	80 * SIZE(BB), %xmm3	addss	%xmm0, %xmm7	movss	 8 * SIZE(AA), %xmm0	mulss	%xmm1, %xmm2	addss	%xmm2, %xmm4	movss	68 * SIZE(BB), %xmm2	mulss	%xmm1, %xmm2	addss	%xmm2, %xmm5	movss	72 * SIZE(BB), %xmm2	mulss	%xmm1, %xmm2	mulss	76 * SIZE(BB), %xmm1	addss	%xmm2, %xmm6	movss	96 * SIZE(BB), %xmm2	addss	%xmm1, %xmm7	movss	 5 * SIZE(AA), %xmm1	mulss	%xmm1, %xmm3	addss	%xmm3, %xmm4	movss	84 * SIZE(BB), %xmm3	mulss	%xmm1, %xmm3	addss	%xmm3, %xmm5	movss	88 * SIZE(BB), %xmm3	mulss	%xmm1, %xmm3	mulss	92 * SIZE(BB), %xmm1	addss	%xmm3, %xmm6	movss	112 * SIZE(BB), %xmm3	addss	%xmm1, %xmm7	movss	 6 * SIZE(AA), %xmm1	mulss	%xmm1, %xmm2	addss	%xmm2, %xmm4	movss	100 * SIZE(BB), %xmm2	mulss	%xmm1, %xmm2	addss	%xmm2, %xmm5	movss	104 * SIZE(BB), %xmm2	mulss	%xmm1, %xmm2	mulss	108 * SIZE(BB), %xmm1	addss	%xmm2, %xmm6	movss	128 * SIZE(BB), %xmm2	addss	%xmm1, %xmm7	movss	 7 * SIZE(AA), %xmm1	mulss	%xmm1, %xmm3	addss	%xmm3, %xmm4	movss	116 * SIZE(BB), %xmm3	mulss	%xmm1, %xmm3	addss	%xmm3, %xmm5	movss	120 * SIZE(BB), %xmm3	mulss	%xmm1, %xmm3	mulss	124 * SIZE(BB), %xmm1	addss	%xmm3, %xmm6	movss	144 * SIZE(BB), %xmm3	addss	%xmm1, %xmm7	movss	12 * SIZE(AA), %xmm1	addl	$  8 * SIZE, AA	addl	$128 * SIZE, BB	decl   %eax	jne    .L32	ALIGN_4.L35:#ifndef TRMMKERNEL	movl	K, %eax#else	movl	KKK, %eax#endif	movaps	ALPHA,  %xmm3	andl	$7, %eax		# if (k & 1)	BRANCH	je .L38	ALIGN_4.L36:	mulss	%xmm0, %xmm2	addss	%xmm2, %xmm4	movss	 4 * SIZE(BB), %xmm2	mulss	%xmm0, %xmm2	addss	%xmm2, %xmm5	movss	 8 * SIZE(BB), %xmm2	mulss	%xmm0, %xmm2	mulss	12 * SIZE(BB), %xmm0	addss	%xmm2, %xmm6	movss	16 * SIZE(BB), %xmm2	addss	%xmm0, %xmm7	movss	 1 * SIZE(AA), %xmm0	addl	$ 1 * SIZE, AA	addl	$16 * SIZE, BB	decl	%eax	jg	.L36	ALIGN_4.L38:	leal	(LDC, LDC, 2), %eax#ifndef TRMMKERNEL	mulss	%xmm3, %xmm4	movss	0 * SIZE(%esi), %xmm0	mulss	%xmm3, %xmm5	movss	0 * SIZE(%esi, LDC, 1), %xmm1	mulss	%xmm3, %xmm6	movss	0 * SIZE(%esi, LDC, 2), %xmm2	mulss	%xmm3, %xmm7	movss	0 * SIZE(%esi, %eax, 1), %xmm3	addss	%xmm0, %xmm4	addss	%xmm1, %xmm5	addss	%xmm2, %xmm6	addss	%xmm3, %xmm7#else	mulss	%xmm3, %xmm4	mulss	%xmm3, %xmm5	mulss	%xmm3, %xmm6	mulss	%xmm3, %xmm7#endif	movss	%xmm4, 0 * SIZE(%esi)	movss	%xmm5, 0 * SIZE(%esi, LDC, 1)	movss	%xmm6, 0 * SIZE(%esi, LDC, 2)	movss	%xmm7, 0 * SIZE(%esi, %eax, 1)#if (defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \    (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))	movl	K, %eax	subl	KKK, %eax	leal	(,%eax, 4), %eax	leal	(AA, %eax, 1), AA	leal	(BB, %eax, 8), BB	leal	(BB, %eax, 8), BB#endif#if defined(TRMMKERNEL) && defined(LEFT)
⌨️ 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 + -