📄 trsm_kernel_rt_4x4_barcelona.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 5 页
字号:
#ifdef LT	addq	$1, KK#endif#ifdef RT	movq	K, %rax	salq	$0 + BASE_SHIFT, %rax	addq	%rax, AORIG#endif	ALIGN_4		.L79:#ifdef LN       leaq	(, K, SIZE), %rax       leaq	(B, %rax, 2), B#endif#if defined(LT) || defined(RN)	movq	BO, B#endif#ifdef RN	addq	$2, KK#endif#ifdef RT	subq	$2, KK#endif	ALIGN_4.L80:	movq	N,  J	sarq	$2, J		# j = (n >> 2)	jle	.L999.L01:#if defined(LT) || defined(RN)	movq	A, AO#else	movq	A, AORIG#endif#ifdef RT       movq	K, %rax       salq	$2 + BASE_SHIFT, %rax       subq	%rax, B       leaq	(, LDC, 4), %rax       subq	%rax, C#endif	movq	C, CO1			# coffset1 = c	leaq	(C, LDC, 1), CO2	# coffset2 = c + ldc#ifndef RT	leaq	(C, LDC, 4), C#endif#ifdef LN	movq	OFFSET, %rax	addq	M, %rax	movq	%rax, KK#endif		movq	K, %rax	salq	$BASE_SHIFT + 2, %rax	movq	B, BB	subq	%rax, BB#if defined(LT)	movq	OFFSET, %rax	movq	%rax, KK#endif	movq	M,  I	sarq	$2, I	# i = (m >> 2)	jle	.L20	ALIGN_4.L11:#ifdef LN       movq	K, %rax       salq	$2 + BASE_SHIFT, %rax       subq	%rax, AORIG#endif#if defined(LN) || defined(RT)	movq	KK, %rax	movq	AORIG, AO	leaq	(, %rax, SIZE), %rax	leaq	(AO, %rax, 4), AO#endif	movq	B, BO#if defined(LN) || defined(RT)	movq	KK, %rax	leaq	(, %rax, SIZE), %rax	leaq	(BO, %rax, 4), BO#endif		movapd	-16 * SIZE(AO), %xmm0	movddup	-16 * SIZE(BO), %xmm1	pxor	%xmm8, %xmm8 	movddup	-15 * SIZE(BO), %xmm3	pxor	%xmm9, %xmm9	movapd	 -8 * SIZE(AO), %xmm4	pxor	%xmm10, %xmm10	movddup	 -8 * SIZE(BO), %xmm5	pxor	%xmm11, %xmm11#ifndef LN	prefetchw      3 * SIZE(CO1)	pxor	%xmm12, %xmm12	prefetchw      3 * SIZE(CO2)	pxor	%xmm13, %xmm13	prefetchw      3 * SIZE(CO1, LDC, 2)	pxor	%xmm14, %xmm14	prefetchw      3 * SIZE(CO2, LDC, 2)	pxor	%xmm15, %xmm15	movapd	%xmm0, %xmm2#else	prefetchw     -8 * SIZE(CO1)	pxor	%xmm12, %xmm12	prefetchw     -8 * SIZE(CO2)	pxor	%xmm13, %xmm13	prefetchw     -8 * SIZE(CO1, LDC, 2)	pxor	%xmm14, %xmm14	prefetchw     -8 * SIZE(CO2, LDC, 2)	pxor	%xmm15, %xmm15	movapd	%xmm0, %xmm2#endif	prefetch	 -16 * SIZE(BB)	prefetch	  -8 * SIZE(BB)	subq		 $-16 * SIZE, BB#if defined(LT) || defined(RN)	movq	KK, %rax#else	movq	K, %rax	subq	KK, %rax#endif	andq	$-8, %rax	leaq	(, %rax, SIZE), %rax	leaq	(AO, %rax, 4), AO	leaq	(BO, %rax, 4), BO	negq	%rax	je,pn	.L15	ALIGN_4.L12:	KERNEL1(16 *  0)	KERNEL2(16 *  0)	KERNEL3(16 *  0)	KERNEL4(16 *  0)	KERNEL5(16 *  0)	KERNEL6(16 *  0)	KERNEL7(16 *  0)	KERNEL8(16 *  0)	jl,pt	.L12	ALIGN_4.L15:#if defined(LT) || defined(RN)	movq	KK, %rax#else	movq	K, %rax	subq	KK, %rax#endif	testq	$4, %rax	je .L16	xorq	%rax, %rax	ALIGN_4	KERNEL_SUB1(16 *  0)	KERNEL_SUB2(16 *  0)	KERNEL_SUB3(16 *  0)	KERNEL_SUB4(16 *  0)	subq	$-16 * SIZE, BO	subq	$-16 * SIZE, AO	ALIGN_4.L16:#if defined(LT) || defined(RN)	movq	KK, %rax#else	movq	K, %rax	subq	KK, %rax#endif	andq	$3, %rax		# if (k & 1)	je .L19	leaq	(, %rax, SIZE), %rax	leaq	(AO, %rax, 4), AO	leaq	(BO, %rax, 4), BO	negq	%rax	ALIGN_4.L17:	mulpd	%xmm1, %xmm0	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1	addpd	%xmm0, %xmm8	movapd	%xmm2, %xmm0	addpd	%xmm1, %xmm12	movddup	-14 * SIZE(BO, %rax, 4), %xmm1	mulpd	%xmm3, %xmm2	mulpd	-14 * SIZE(AO, %rax, 4), %xmm3	addpd	%xmm2, %xmm9	movapd	%xmm0, %xmm2	addpd	%xmm3, %xmm13	movddup	-13 * SIZE(BO, %rax, 4), %xmm3	mulpd	%xmm1, %xmm0	mulpd	-14 * SIZE(AO, %rax, 4), %xmm1	addpd	%xmm0, %xmm10	movapd	-12 * SIZE(AO, %rax, 4), %xmm0	addpd	%xmm1, %xmm14	movddup	-12 * SIZE(BO, %rax, 4), %xmm1	mulpd	%xmm3, %xmm2	mulpd	-14 * SIZE(AO, %rax, 4), %xmm3	addpd	%xmm2, %xmm11	addpd	%xmm3, %xmm15 	movddup	-11 * SIZE(BO, %rax, 4), %xmm3	movapd	%xmm0, %xmm2	addq	$SIZE, %rax	jl	.L17	ALIGN_4.L19:#if defined(LN) || defined(RT)	movq	KK, %rax#ifdef LN	subq	$4, %rax#else	subq	$4, %rax#endif	leaq	(, %rax, SIZE), %rax	movq	AORIG, AO	leaq	(AO, %rax, 4), AO	leaq	(B,  %rax, 4), BO#endif#if defined(LN) || defined(LT)	movapd	%xmm8, %xmm0	unpcklpd %xmm9, %xmm8	unpckhpd %xmm9, %xmm0	movapd	%xmm10, %xmm2	unpcklpd %xmm11, %xmm10	unpckhpd %xmm11, %xmm2	movapd	%xmm12, %xmm4	unpcklpd %xmm13, %xmm12	unpckhpd %xmm13, %xmm4	movapd	%xmm14, %xmm6	unpcklpd %xmm15, %xmm14	unpckhpd %xmm15, %xmm6	movapd	-16 * SIZE(BO), %xmm9	movapd	-14 * SIZE(BO), %xmm11	movapd	-12 * SIZE(BO), %xmm13	movapd	-10 * SIZE(BO), %xmm15	movapd	 -8 * SIZE(BO), %xmm1	movapd	 -6 * SIZE(BO), %xmm3	movapd	 -4 * SIZE(BO), %xmm5	movapd	 -2 * SIZE(BO), %xmm7	subpd	%xmm8,  %xmm9	subpd	%xmm10, %xmm11	subpd	%xmm0,  %xmm13	subpd	%xmm2,  %xmm15	subpd	%xmm12, %xmm1	subpd	%xmm14, %xmm3	subpd	%xmm4,  %xmm5	subpd	%xmm6,  %xmm7#else	movapd	-16 * SIZE(AO), %xmm0	movapd	-14 * SIZE(AO), %xmm1	movapd	-12 * SIZE(AO), %xmm2	movapd	-10 * SIZE(AO), %xmm3	movapd	 -8 * SIZE(AO), %xmm4	movapd	 -6 * SIZE(AO), %xmm5	movapd	 -4 * SIZE(AO), %xmm6	movapd	 -2 * SIZE(AO), %xmm7	subpd	%xmm8,  %xmm0	subpd	%xmm12, %xmm1	subpd	%xmm9,  %xmm2	subpd	%xmm13, %xmm3	subpd	%xmm10, %xmm4	subpd	%xmm14, %xmm5	subpd	%xmm11, %xmm6	subpd	%xmm15, %xmm7#endif#ifdef LN	movddup	 -1 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm5	mulpd	 %xmm8, %xmm7	movddup	 -2 * SIZE(AO), %xmm10	mulpd	 %xmm5, %xmm10	subpd	 %xmm10, %xmm1	movddup	 -2 * SIZE(AO), %xmm10	mulpd	 %xmm7, %xmm10	subpd	 %xmm10, %xmm3	movddup	 -3 * SIZE(AO), %xmm12	mulpd	 %xmm5, %xmm12	subpd	 %xmm12, %xmm13	movddup	 -3 * SIZE(AO), %xmm12	mulpd	 %xmm7, %xmm12	subpd	 %xmm12, %xmm15	movddup	 -4 * SIZE(AO), %xmm14	mulpd	 %xmm5, %xmm14	subpd	 %xmm14, %xmm9	movddup	 -4 * SIZE(AO), %xmm14	mulpd	 %xmm7, %xmm14	subpd	 %xmm14, %xmm11	movddup	 -6 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm1	mulpd	 %xmm8, %xmm3	movddup	 -7 * SIZE(AO), %xmm10	mulpd	 %xmm1, %xmm10	subpd	 %xmm10, %xmm13	movddup	 -7 * SIZE(AO), %xmm10	mulpd	 %xmm3, %xmm10	subpd	 %xmm10, %xmm15	movddup	 -8 * SIZE(AO), %xmm12	mulpd	 %xmm1, %xmm12	subpd	 %xmm12, %xmm9	movddup	 -8 * SIZE(AO), %xmm12	mulpd	 %xmm3, %xmm12	subpd	 %xmm12, %xmm11	movddup	-11 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm13	mulpd	 %xmm8, %xmm15	movddup	-12 * SIZE(AO), %xmm10	mulpd	 %xmm13, %xmm10	subpd	 %xmm10, %xmm9	movddup	-12 * SIZE(AO), %xmm10	mulpd	 %xmm15, %xmm10	subpd	 %xmm10, %xmm11	movddup	-16 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm9	mulpd	 %xmm8, %xmm11#endif#ifdef LT	movddup -16 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm9	mulpd	 %xmm8, %xmm11	movddup	-15 * SIZE(AO), %xmm10	mulpd	 %xmm9, %xmm10	subpd	 %xmm10, %xmm13	movddup	-15 * SIZE(AO), %xmm10	mulpd	 %xmm11, %xmm10	subpd	 %xmm10, %xmm15	movddup	-14 * SIZE(AO), %xmm12	mulpd	 %xmm9, %xmm12	subpd	 %xmm12, %xmm1	movddup	-14 * SIZE(AO), %xmm12	mulpd	 %xmm11, %xmm12	subpd	 %xmm12, %xmm3	movddup	-13 * SIZE(AO), %xmm14	mulpd	 %xmm9, %xmm14	subpd	 %xmm14, %xmm5	movddup	-13 * SIZE(AO), %xmm14	mulpd	 %xmm11, %xmm14	subpd	 %xmm14, %xmm7	movddup	-11 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm13	mulpd	 %xmm8, %xmm15	movddup	-10 * SIZE(AO), %xmm10	mulpd	 %xmm13, %xmm10	subpd	 %xmm10, %xmm1	movddup	-10 * SIZE(AO), %xmm10	mulpd	 %xmm15, %xmm10	subpd	 %xmm10, %xmm3	movddup	 -9 * SIZE(AO), %xmm12	mulpd	 %xmm13, %xmm12	subpd	 %xmm12, %xmm5	movddup	 -9 * SIZE(AO), %xmm12	mulpd	 %xmm15, %xmm12	subpd	 %xmm12, %xmm7	movddup	 -6 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm1	mulpd	 %xmm8, %xmm3	movddup	 -5 * SIZE(AO), %xmm10	mulpd	 %xmm1, %xmm10	subpd	 %xmm10, %xmm5	movddup	 -5 * SIZE(AO), %xmm10	mulpd	 %xmm3, %xmm10	subpd	 %xmm10, %xmm7	movddup	 -1 * SIZE(AO), %xmm8	mulpd	 %xmm8, %xmm5	mulpd	 %xmm8, %xmm7#endif#ifdef RN	movddup	-16 * SIZE(BO), %xmm8	mulpd	 %xmm8, %xmm0	mulpd	 %xmm8, %xmm1	movddup	-15 * SIZE(BO), %xmm9	mulpd	 %xmm0, %xmm9	subpd	 %xmm9, %xmm2	movddup	-15 * SIZE(BO), %xmm9	mulpd	 %xmm1, %xmm9	subpd	 %xmm9, %xmm3	movddup	-14 * SIZE(BO), %xmm10	mulpd	 %xmm0, %xmm10	subpd	 %xmm10, %xmm4	movddup	-14 * SIZE(BO), %xmm10	mulpd	 %xmm1, %xmm10	subpd	 %xmm10, %xmm5	movddup	 -13 * SIZE(BO), %xmm11	mulpd	 %xmm0, %xmm11	subpd	 %xmm11, %xmm6	movddup	 -13 * SIZE(BO), %xmm11	mulpd	 %xmm1, %xmm11	subpd	 %xmm11, %xmm7	movddup	 -11 * SIZE(BO), %xmm8	mulpd	 %xmm8, %xmm2	mulpd	 %xmm8, %xmm3	movddup	 -10 * SIZE(BO), %xmm9	mulpd	 %xmm2, %xmm9	subpd	 %xmm9, %xmm4	movddup	 -10 * SIZE(BO), %xmm9	mulpd	 %xmm3, %xmm9	subpd	 %xmm9, %xmm5	movddup	  -9 * SIZE(BO), %xmm10	mulpd	 %xmm2, %xmm10	subpd	 %xmm10, %xmm6	movddup	  -9 * SIZE(BO), %xmm10	mulpd	 %xmm3, %xmm10	subpd	 %xmm10, %xmm7	movddup	 -6 * SIZE(BO), %xmm8	mulpd	 %xmm8, %xmm4	mulpd	 %xmm8, %xmm5	movddup	 -5 * SIZE(BO), %xmm9	mulpd	 %xmm4, %xmm9	subpd	 %xmm9, %xmm6	movddup	 -5 * SIZE(BO), %xmm9	mulpd	 %xmm5, %xmm9	subpd	 %xmm9, %xmm7	movddup	 -1 * SIZE(BO), %xmm8	mulpd	 %xmm8, %xmm6	mulpd	 %xmm8, %xmm7#endif#ifdef RT	movddup	 -1 * SIZE(BO), %xmm8	mulpd	 %xmm8, %xmm6	mulpd	 %xmm8, %xmm7	movddup	 -2 * SIZE(BO), %xmm9	mulpd	 %xmm6, %xmm9	subpd	 %xmm9, %xmm4	movddup	 -2 * SIZE(BO), %xmm9	mulpd	 %xmm7, %xmm9	subpd	 %xmm9, %xmm5	movddup	 -3 * SIZE(BO), %xmm10	mulpd	 %xmm6, %xmm10	subpd	 %xmm10, %xmm2	movddup	 -3 * SIZE(BO), %xmm10	mulpd	 %xmm7, %xmm10	subpd	 %xmm10, %xmm3	movddup	 -4 * SIZE(BO), %xmm11	mulpd	 %xmm6, %xmm11	subpd	 %xmm11, %xmm0	movddup	 -4 * SIZE(BO), %xmm11	mulpd	 %xmm7, %xmm11	subpd	 %xmm11, %xmm1	movddup	 -6 * SIZE(BO), %xmm8	mulpd	 %xmm8, %xmm4	mulpd	 %xmm8, %xmm5	movddup	 -7 * SIZE(BO), %xmm9	mulpd	 %xmm4, %xmm9	subpd	 %xmm9, %xmm2	movddup	 -7 * SIZE(BO), %xmm9	mulpd	 %xmm5, %xmm9	subpd	 %xmm9, %xmm3	movddup	 -8 * SIZE(BO), %xmm10	mulpd	 %xmm4, %xmm10	subpd	 %xmm10, %xmm0	movddup	 -8 * SIZE(BO), %xmm10	mulpd	 %xmm5, %xmm10	subpd	 %xmm10, %xmm1	movddup	-11 * SIZE(BO), %xmm8	mulpd	 %xmm8, %xmm2	mulpd	 %xmm8, %xmm3	movddup	-12 * SIZE(BO), %xmm9	mulpd	 %xmm2, %xmm9	subpd	 %xmm9, %xmm0	movddup	-12 * SIZE(BO), %xmm9	mulpd	 %xmm3, %xmm9	subpd	 %xmm9, %xmm1	movddup	-16 * SIZE(BO), %xmm8	mulpd	 %xmm8, %xmm0	mulpd	 %xmm8, %xmm1#endif#ifdef LN	subq	$4 * SIZE, CO1	subq	$4 * SIZE, CO2#endif#if defined(LN) || defined(LT)	movlpd	%xmm9,  0 * SIZE(CO1)	movlpd	%xmm13, 1 * SIZE(CO1)	movlpd	%xmm1,  2 * SIZE(CO1)	movlpd	%xmm5,  3 * SIZE(CO1)	movhpd	%xmm9,  0 * SIZE(CO2)	movhpd	%xmm13, 1 * SIZE(CO2)	movhpd	%xmm1,  2 * SIZE(CO2)	movhpd	%xmm5,  3 * SIZE(CO2)	movlpd	%xmm11, 0 * SIZE(CO1, LDC, 2)	movlpd	%xmm15, 1 * SIZE(CO1, LDC, 2)	movlpd	%xmm3,  2 * SIZE(CO1, LDC, 2)	movlpd	%xmm7,  3 * SIZE(CO1, LDC, 2)	movhpd	%xmm11, 0 * SIZE(CO2, LDC, 2)	movhpd	%xmm15, 1 * SIZE(CO2, LDC, 2)	movhpd	%xmm3,  2 * SIZE(CO2, LDC, 2)	movhpd	%xmm7,  3 * SIZE(CO2, LDC, 2)#else	movlpd	%xmm0,  0 * SIZE(CO1)	movhpd	%xmm0,  1 * SIZE(CO1)	movlpd	%xmm1,  2 * SIZE(CO1)	movhpd	%xmm1,  3 * SIZE(CO1)	movlpd	%xmm2,  0 * SIZE(CO2)	movhpd	%xmm2,  1 * SIZE(CO2)	movlpd	%xmm3,  2 * SIZE(CO2)	movhpd	%xmm3,  3 * SIZE(CO2)	movlpd	%xmm4,  0 * SIZE(CO1, LDC, 2)	movhpd	%xmm4,  1 * SIZE(CO1, LDC, 2)	movlpd	%xmm5,  2 * SIZE(CO1, LDC, 2)	movhpd	%xmm5,  3 * SIZE(CO1, LDC, 2)	movlpd	%xmm6,  0 * SIZE(CO2, LDC, 2)	movhpd	%xmm6,  1 * SIZE(CO2, LDC, 2)	movlpd	%xmm7,  2 * SIZE(CO2, LDC, 2)	movhpd	%xmm7,  3 * SIZE(CO2, LDC, 2)#endif#if defined(LN) || defined(LT)	movaps	%xmm9,  -16 * SIZE(BO)	movaps	%xmm11, -14 * SIZE(BO)	movaps	%xmm13, -12 * SIZE(BO)	movaps	%xmm15, -10 * SIZE(BO)	movaps	%xmm1,   -8 * SIZE(BO)	movaps	%xmm3,   -6 * SIZE(BO)	movaps	%xmm5,   -4 * SIZE(BO)	movaps	%xmm7,   -2 * SIZE(BO)#else	movaps	%xmm0,  -16 * SIZE(AO)	movaps	%xmm1,  -14 * SIZE(AO)	movaps	%xmm2,  -12 * SIZE(AO)	movaps	%xmm3,  -10 * SIZE(AO)	movaps	%xmm4,   -8 * SIZE(AO)	movaps	%xmm5,   -6 * SIZE(AO)	movaps	%xmm6,   -4 * SIZE(AO)	movaps	%xmm7,   -2 * SIZE(AO)#endif#ifndef LN	addq	$4 * SIZE, CO1	addq	$4 * SIZE, CO2#endif#if defined(LT) || defined(RN)	movq	K,  %rax	subq	KK, %rax	leaq	(,%rax, SIZE), %rax	leaq	(AO, %rax, 4), AO	leaq	(BO, %rax, 4), BO#endif#ifdef LN	subq	$4, KK#endif#ifdef LT	addq	$4, KK#endif#ifdef RT	movq	K, %rax	salq	$2 + BASE_SHIFT, %rax	addq	%rax, AORIG#endif	decq	I			# i --	jg	.L11	ALIGN_4	.L20:	testq	$3, M	je	.L39	testq	$2, M	je	.L30	ALIGN_4.L21:#ifdef LN       movq	K, %rax       salq	$1 + BASE_SHIFT, %rax       subq	%rax, AORIG#endif#if defined(LN) || defined(RT)	movq	KK, %rax	movq	AORIG, AO	leaq	(, %rax, SIZE), %rax	leaq	(AO, %rax, 2), AO#endif	movq	B, BO#if defined(LN) || defined(RT)	movq	KK, %rax	leaq	(, %rax, SIZE), %rax	leaq	(BO, %rax, 4), BO#endif		movapd	-16 * SIZE(AO), %xmm0	pxor	%xmm8, %xmm8
⌨️ 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 + -