📄 gemm_kernel_8x4_sse3.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 5 页
字号:
	KERNEL4 (64  *  4)	KERNEL5 (64  *  4)	KERNEL6 (64  *  4)	KERNEL7 (64  *  4)	KERNEL8 (64  *  4)	KERNEL9 (64  *  4)	KERNEL10(64  *  4)	KERNEL11(64  *  4)	KERNEL12(64  *  4)	KERNEL13(64  *  4)	KERNEL14(64  *  4)	KERNEL15(64  *  4)	KERNEL16(64  *  4)	cmpq	$128 *  5, %rax	jle,pn	.L12	KERNEL1 (64  *  5)	KERNEL2 (64  *  5)	KERNEL3 (64  *  5)	KERNEL4 (64  *  5)	KERNEL5 (64  *  5)	KERNEL6 (64  *  5)	KERNEL7 (64  *  5)	KERNEL8 (64  *  5)	KERNEL9 (64  *  5)	KERNEL10(64  *  5)	KERNEL11(64  *  5)	KERNEL12(64  *  5)	KERNEL13(64  *  5)	KERNEL14(64  *  5)	KERNEL15(64  *  5)	KERNEL16(64  *  5)	cmpq	$128 *  6, %rax	jle,pn	.L12	KERNEL1 (64  *  6)	KERNEL2 (64  *  6)	KERNEL3 (64  *  6)	KERNEL4 (64  *  6)	KERNEL5 (64  *  6)	KERNEL6 (64  *  6)	KERNEL7 (64  *  6)	KERNEL8 (64  *  6)	KERNEL9 (64  *  6)	KERNEL10(64  *  6)	KERNEL11(64  *  6)	KERNEL12(64  *  6)	KERNEL13(64  *  6)	KERNEL14(64  *  6)	KERNEL15(64  *  6)	KERNEL16(64  *  6)	cmpq	$128 *  7, %rax	jle,pn	.L12	KERNEL1 (64  *  7)	KERNEL2 (64  *  7)	KERNEL3 (64  *  7)	KERNEL4 (64  *  7)	KERNEL5 (64  *  7)	KERNEL6 (64  *  7)	KERNEL7 (64  *  7)	KERNEL8 (64  *  7)	KERNEL9 (64  *  7)	KERNEL10(64  *  7)	KERNEL11(64  *  7)	KERNEL12(64  *  7)	KERNEL13(64  *  7)	KERNEL14(64  *  7)	KERNEL15(64  *  7)	KERNEL16(64  *  7)	addq	$64 * 8  * SIZE, AO	addq	$64 * 8  * SIZE, BO	subq	$128 * 8, %rax	jg	.L1X.L12:	leaq	(AO, %rax, 2), AO	leaq	(BO, %rax, 2), BO#else	sarq	$3, %rax	je	.L15	ALIGN_4.L12:	KERNEL1 (64  *  0)	KERNEL2 (64  *  0)	KERNEL3 (64  *  0)	KERNEL4 (64  *  0)	KERNEL5 (64  *  0)	KERNEL6 (64  *  0)	KERNEL7 (64  *  0)	KERNEL8 (64  *  0)	KERNEL9 (64  *  0)	KERNEL10(64  *  0)	KERNEL11(64  *  0)	KERNEL12(64  *  0)	KERNEL13(64  *  0)	KERNEL14(64  *  0)	KERNEL15(64  *  0)	KERNEL16(64  *  0)	addq   $64 * SIZE, AO	addq   $64 * SIZE, BO	decq   %rax	jne    .L12#endif	ALIGN_4.L15:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movaps	ALPHA, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je .L18	ALIGN_4.L16:	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm1	movsldup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm2	movshdup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	  4 * SIZE(AO), %xmm8	addps	%xmm9, %xmm3	movsldup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm4	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm5	movsldup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm6	movshdup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9 	movaps	  8 * SIZE(AO), %xmm8	addps	%xmm9, %xmm7	movsldup  8 * SIZE(BO), %xmm9	addq	$8 * SIZE, AO	addq	$8 * SIZE, BO	decq	%rax	jg	.L16	ALIGN_4.L18:#if! defined(TRMMKERNEL) && !defined(BETAZERO)	movsd	0 * SIZE(CO1), %xmm8	mulps	%xmm15, %xmm0	movhps	2 * SIZE(CO1), %xmm8	mulps	%xmm15, %xmm1	movsd	4 * SIZE(CO1), %xmm9	mulps	%xmm15, %xmm2	movhps	6 * SIZE(CO1), %xmm9	mulps	%xmm15, %xmm3	movsd	0 * SIZE(CO2), %xmm10	mulps	%xmm15, %xmm4	movhps	2 * SIZE(CO2), %xmm10	mulps	%xmm15, %xmm5	movsd	4 * SIZE(CO2), %xmm11	mulps	%xmm15, %xmm6	movhps	6 * SIZE(CO2), %xmm11	mulps	%xmm15, %xmm7	movsd	0 * SIZE(CO1, LDC, 2), %xmm12	movhps	2 * SIZE(CO1, LDC, 2), %xmm12	movsd	4 * SIZE(CO1, LDC, 2), %xmm13	movhps	6 * SIZE(CO1, LDC, 2), %xmm13	movsd	0 * SIZE(CO2, LDC, 2), %xmm14	movhps	2 * SIZE(CO2, LDC, 2), %xmm14	movsd	4 * SIZE(CO2, LDC, 2), %xmm15	movhps	6 * SIZE(CO2, LDC, 2), %xmm15	addps	%xmm8,  %xmm0	addps	%xmm9,  %xmm4	addps	%xmm10, %xmm1	addps	%xmm11, %xmm5	addps	%xmm12, %xmm2	movsd	%xmm0, 0 * SIZE(CO1)	movhps	%xmm0, 2 * SIZE(CO1)	addps	%xmm13, %xmm6	movsd	%xmm4, 4 * SIZE(CO1)	movhps	%xmm4, 6 * SIZE(CO1)	addps	%xmm14, %xmm3	movsd	%xmm1, 0 * SIZE(CO2)	movhps	%xmm1, 2 * SIZE(CO2)	addps	%xmm15, %xmm7	movsd	%xmm5, 4 * SIZE(CO2)	movhps	%xmm5, 6 * SIZE(CO2)#else	mulps	%xmm15, %xmm0	mulps	%xmm15, %xmm1	mulps	%xmm15, %xmm2	mulps	%xmm15, %xmm3	mulps	%xmm15, %xmm4	mulps	%xmm15, %xmm5	mulps	%xmm15, %xmm6	mulps	%xmm15, %xmm7	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)#endif	movsd	%xmm2, 0 * SIZE(CO1, LDC, 2)	movhps	%xmm2, 2 * SIZE(CO1, LDC, 2)	movsd	%xmm6, 4 * SIZE(CO1, LDC, 2)	movhps	%xmm6, 6 * SIZE(CO1, LDC, 2)	movsd	%xmm3, 0 * SIZE(CO2, LDC, 2)	movhps	%xmm3, 2 * SIZE(CO2, LDC, 2)	movsd	%xmm7, 4 * SIZE(CO2, LDC, 2)	movhps	%xmm7, 6 * 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, 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	.L11	ALIGN_4	.L20:	testq	$4, M	je	.L30#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	  0 * SIZE(AO), %xmm8	movaps	 16 * SIZE(AO), %xmm10	movsldup  0 * SIZE(BO), %xmm9	movsldup 16 * SIZE(BO), %xmm11	movsldup 32 * SIZE(BO), %xmm13	movsldup 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	$4, %rax#endif	movq	%rax, KKK#endif	sarq	$3, %rax	je	.L25	ALIGN_4.L22:	mulps	%xmm8, %xmm9	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)	addps	%xmm9, %xmm0	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm1	movsldup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm2	movshdup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	  4 * 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	addps	%xmm9, %xmm1	movsldup 12 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm2	movshdup 12 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	  8 * SIZE(AO), %xmm8	addps	%xmm9, %xmm3	movsldup 64 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm11	addps	%xmm11, %xmm0	movshdup 16 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	addps	%xmm11, %xmm1	movsldup 20 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	addps	%xmm11, %xmm2	movshdup 20 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	movaps	 12 * SIZE(AO), %xmm8	addps	%xmm11, %xmm3	movsldup 24 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	addps	%xmm11, %xmm0	movshdup 24 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	addps	%xmm11, %xmm1	movsldup 28 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	addps	%xmm11, %xmm2	movshdup 28 * SIZE(BO), %xmm11	mulps	%xmm8, %xmm11	movaps	 32 * SIZE(AO), %xmm8	addps	%xmm11, %xmm3	movsldup 80 * SIZE(BO), %xmm11	mulps	%xmm10, %xmm13	addps	%xmm13, %xmm0	movshdup 32 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	addps	%xmm13, %xmm1	movsldup 36 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	addps	%xmm13, %xmm2	movshdup 36 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	movaps	 20 * SIZE(AO), %xmm10	addps	%xmm13, %xmm3	movsldup 40 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	addps	%xmm13, %xmm0	movshdup 40 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	addps	%xmm13, %xmm1	movsldup 44 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	addps	%xmm13, %xmm2	movshdup 44 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm13	movaps	 24 * SIZE(AO), %xmm10	addps	%xmm13, %xmm3	movsldup 96 * SIZE(BO), %xmm13	mulps	%xmm10, %xmm15	addps	%xmm15, %xmm0	movshdup 48 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	addps	%xmm15, %xmm1	movsldup 52 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	addps	%xmm15, %xmm2	movshdup 52 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	movaps	 28 * SIZE(AO), %xmm10	addps	%xmm15, %xmm3	movsldup 56 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	addps	%xmm15, %xmm0	movshdup 56 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	addps	%xmm15, %xmm1	movsldup 60 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	addps	%xmm15, %xmm2	movshdup 60 * SIZE(BO), %xmm15	mulps	%xmm10, %xmm15	movaps	 48 * SIZE(AO), %xmm10	addps	%xmm15, %xmm3	movsldup 112 * SIZE(BO), %xmm15	addq   $32 * SIZE, AO	addq   $64 * SIZE, BO	decq   %rax	jne    .L22	ALIGN_4.L25:#ifndef TRMMKERNEL	movq	K, %rax#else	movq	KKK, %rax#endif	movaps	ALPHA, %xmm15	andq	$7, %rax		# if (k & 1)	BRANCH	je .L28	ALIGN_4.L26:	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movshdup  0 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm1	movsldup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm2	movshdup  4 * SIZE(BO), %xmm9	mulps	%xmm8, %xmm9	movaps	  4 * SIZE(AO), %xmm8	addps	%xmm9, %xmm3	movsldup  8 * SIZE(BO), %xmm9	addq	$4 * SIZE, AO	addq	$8 * SIZE, BO	decq	%rax	jg	.L26	ALIGN_4.L28:#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	mulps	%xmm15, %xmm0	mulps	%xmm15, %xmm1 	mulps	%xmm15, %xmm2	mulps	%xmm15, %xmm3	movsd	0 * SIZE(CO1, LDC, 2), %xmm12	movhps	2 * SIZE(CO1, LDC, 2), %xmm12	movsd	0 * SIZE(CO2, LDC, 2), %xmm14	movhps	2 * SIZE(CO2, LDC, 2), %xmm14	addps	%xmm8,  %xmm0	addps	%xmm10, %xmm1	addps	%xmm12, %xmm2	addps	%xmm14, %xmm3#else	mulps	%xmm15, %xmm0	mulps	%xmm15, %xmm1 	mulps	%xmm15, %xmm2	mulps	%xmm15, %xmm3#endif	movsd	%xmm0, 0 * SIZE(CO1)	movhps	%xmm0, 2 * SIZE(CO1)	movsd	%xmm1, 0 * SIZE(CO2)	movhps	%xmm1, 2 * SIZE(CO2)	movsd	%xmm2, 0 * SIZE(CO1, LDC, 2)	movhps	%xmm2, 2 * SIZE(CO1, LDC, 2)	movsd	%xmm3, 0 * SIZE(CO2, LDC, 2)	movhps	%xmm3, 2 * 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, 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	.L30:	testq	$2, M	je	.L40#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		movddup	  0 * SIZE(AO), %xmm8	movddup	  8 * SIZE(AO), %xmm10	movsd     0 * SIZE(BO), %xmm9	movsd    32 * 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	$2, %rax#else	addq	$4, %rax#endif	movq	%rax, KKK#endif	sarq	$3, %rax	je	.L35	ALIGN_4.L32:	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	PREFETCH  (PREFETCHSIZE +  0) * SIZE(AO)	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	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm2	movsd    12 * SIZE(BO), %xmm9	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	movddup	  4 * SIZE(AO), %xmm8	addps	%xmm9, %xmm3	movsd    16 * SIZE(BO), %xmm9	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm0	movsd    20 * SIZE(BO), %xmm9	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	movddup	  6 * SIZE(AO), %xmm8	addps	%xmm9, %xmm1	movsd    24 * SIZE(BO), %xmm9	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	addps	%xmm9, %xmm2	movsd    28 * SIZE(BO), %xmm9	shufps	  $0x50, %xmm9, %xmm9	mulps	%xmm8, %xmm9	movddup	 16 * SIZE(AO), %xmm8	addps	%xmm9, %xmm3	shufps	  $0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11	movsd    64 * SIZE(BO), %xmm9	addps	%xmm11, %xmm0	movsd    36 * SIZE(BO), %xmm11	shufps	  $0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11	movddup	 10 * SIZE(AO), %xmm10	addps	%xmm11, %xmm1	movsd    40 * SIZE(BO), %xmm11	shufps	  $0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm2	movsd    44 * SIZE(BO), %xmm11	shufps	  $0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11	movddup	 12 * SIZE(AO), %xmm10	addps	%xmm11, %xmm3	movsd    48 * SIZE(BO), %xmm11	shufps	  $0x50, %xmm11, %xmm11	mulps	%xmm10, %xmm11	addps	%xmm11, %xmm0	movsd    52 * SIZE(BO), %xmm11	shufps	  $0x50, %xmm11, %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 + -