⭐ 欢迎来到虫虫下载站! | 📦 资源下载 📁 资源专辑 ℹ️ 关于我们
⭐ 虫虫下载站

📄 gemm_kernel.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 4 页
字号:
	fmr	f5,  f0	fmr	f6,  f0	fmr	f7,  f0		srawi.	I, M,  2	add	C,  CO2, LDC	mr	AO, A	ble	LL(50)	.align 4LL(41):#if defined(TRMMKERNEL)#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))	LFD	f16,  0 * SIZE(AO)	LFD	f17,  1 * SIZE(AO)	LFD	f18,  2 * SIZE(AO)	LFD	f19,  3 * SIZE(AO)	LFD	f20,  0 * SIZE(B)	LFD	f21,  1 * SIZE(B)	LFD	f22,  2 * SIZE(B)	LFD	f23,  3 * SIZE(B)	mr	BO,  B#else	slwi	r0,   KK, 2 + BASE_SHIFT	slwi	TEMP, KK, 1 + BASE_SHIFT	add	AO, AO, r0	add	BO, B,  TEMP	LFD	f16,  0 * SIZE(AO)	LFD	f17,  1 * SIZE(AO)	LFD	f18,  2 * SIZE(AO)	LFD	f19,  3 * SIZE(AO)	LFD	f20,  0 * SIZE(BO)	LFD	f21,  1 * SIZE(BO)	LFD	f22,  2 * SIZE(BO)	LFD	f23,  3 * SIZE(BO)#endif	dcbt	CO1, PREC	dcbt	CO2, PREC#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	sub	TEMP, K, KK#elif defined(LEFT)	addi	TEMP, KK, 4#else	addi	TEMP, KK, 2#endif	srawi.	TEMP,  TEMP,  2	mtspr	CTR, TEMP#else	LFD	f16,  0 * SIZE(AO)	LFD	f17,  1 * SIZE(AO)	LFD	f18,  2 * SIZE(AO)	LFD	f19,  3 * SIZE(AO)	LFD	f20,  0 * SIZE(B)	LFD	f21,  1 * SIZE(B)	LFD	f22,  2 * SIZE(B)	LFD	f23,  3 * SIZE(B)	dcbt	CO1, PREC	dcbt	CO2, PREC	srawi.	r0,  K,  2	mtspr	CTR, r0	mr	BO,  B#endif	ble	LL(45)	.align 5LL(42):	FMADD	f0,  f16, f20, f0	FMADD	f1,  f17, f20, f1	FMADD	f2,  f18, f20, f2	FMADD	f3,  f19, f20, f3	FMADD	f4,  f16, f21, f4	FMADD	f5,  f17, f21, f5	FMADD	f6,  f18, f21, f6	FMADD	f7,  f19, f21, f7	LFD	f16,  4 * SIZE(AO)	LFD	f17,  5 * SIZE(AO)	LFD	f18,  6 * SIZE(AO)	LFD	f19,  7 * SIZE(AO)	FMADD	f0,  f16, f22, f0	FMADD	f1,  f17, f22, f1	FMADD	f2,  f18, f22, f2	FMADD	f3,  f19, f22, f3	FMADD	f4,  f16, f23, f4	FMADD	f5,  f17, f23, f5	FMADD	f6,  f18, f23, f6	FMADD	f7,  f19, f23, f7	LFD	f16,  8 * SIZE(AO)	LFD	f17,  9 * SIZE(AO)	LFD	f18, 10 * SIZE(AO)	LFD	f19, 11 * SIZE(AO)	LFD	f20,  4 * SIZE(BO)	LFD	f21,  5 * SIZE(BO)	LFD	f22,  6 * SIZE(BO)	LFD	f23,  7 * SIZE(BO)	FMADD	f0,  f16, f20, f0	FMADD	f1,  f17, f20, f1	FMADD	f2,  f18, f20, f2	FMADD	f3,  f19, f20, f3	FMADD	f4,  f16, f21, f4	FMADD	f5,  f17, f21, f5	FMADD	f6,  f18, f21, f6	FMADD	f7,  f19, f21, f7	LFD	f16, 12 * SIZE(AO)	LFD	f17, 13 * SIZE(AO)	LFD	f18, 14 * SIZE(AO)	LFD	f19, 15 * SIZE(AO)	FMADD	f0,  f16, f22, f0	FMADD	f1,  f17, f22, f1	FMADD	f2,  f18, f22, f2	FMADD	f3,  f19, f22, f3	FMADD	f4,  f16, f23, f4	FMADD	f5,  f17, f23, f5	FMADD	f6,  f18, f23, f6	FMADD	f7,  f19, f23, f7	LFD	f16, 16 * SIZE(AO)	LFD	f17, 17 * SIZE(AO)	LFD	f18, 18 * SIZE(AO)	LFD	f19, 19 * SIZE(AO)	LFD	f20,  8 * SIZE(BO)	LFD	f21,  9 * SIZE(BO)	LFD	f22, 10 * SIZE(BO)	LFD	f23, 11 * SIZE(BO)	addi	AO, AO, 16 * SIZE	addi	BO, BO,  8 * SIZE	PREFETCH_B	bdnz	LL(42)	.align 4LL(45):	lfd	f30,  ALPHA#if defined(TRMMKERNEL)#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	sub	TEMP, K, KK#elif defined(LEFT)	addi	TEMP, KK, 4#else	addi	TEMP, KK, 2#endif	andi.	TEMP,  TEMP,  3	mtspr	CTR, TEMP#else	andi.	r0,  K,  3	mtspr	CTR, r0#endif	ble+	LL(48)	.align 4LL(46):	FMADD	f0,  f16, f20, f0	FMADD	f1,  f17, f20, f1	FMADD	f2,  f18, f20, f2	FMADD	f3,  f19, f20, f3	FMADD	f4,  f16, f21, f4	FMADD	f5,  f17, f21, f5	FMADD	f6,  f18, f21, f6	FMADD	f7,  f19, f21, f7	LFD	f16,  4 * SIZE(AO)	LFD	f17,  5 * SIZE(AO)	LFD	f18,  6 * SIZE(AO)	LFD	f19,  7 * SIZE(AO)	LFD	f20,  2 * SIZE(BO)	LFD	f21,  3 * SIZE(BO)	addi	BO, BO,  2 * SIZE	addi	AO, AO,  4 * SIZE	bdnz	LL(46)	.align 4LL(48):#ifndef TRMMKERNEL	LFD	f16, 0 * SIZE(CO1)	LFD	f17, 1 * SIZE(CO1)	LFD	f18, 2 * SIZE(CO1)	LFD	f19, 3 * SIZE(CO1)	LFD	f20, 0 * SIZE(CO2)	LFD	f21, 1 * SIZE(CO2)	LFD	f22, 2 * SIZE(CO2)	LFD	f23, 3 * SIZE(CO2)	FMADD	f0,  f0, f30, f16	FMADD	f1,  f1, f30, f17	FMADD	f2,  f2, f30, f18	FMADD	f3,  f3, f30, f19	FMADD	f4,  f4, f30, f20	FMADD	f5,  f5, f30, f21	FMADD	f6,  f6, f30, f22	FMADD	f7,  f7, f30, f23#else	FMUL	f0,  f0, f30	FMUL	f1,  f1, f30	FMUL	f2,  f2, f30	FMUL	f3,  f3, f30	FMUL	f4,  f4, f30	FMUL	f5,  f5, f30	FMUL	f6,  f6, f30	FMUL	f7,  f7, f30#endif	STFD	f0,  0 * SIZE(CO1)	STFD	f1,  1 * SIZE(CO1)	STFD	f2,  2 * SIZE(CO1)	STFD	f3,  3 * SIZE(CO1)	lfs	f0,  FZERO 	fmr	f1,  f0	fmr	f2,  f0	fmr	f3,  f0	STFD	f4,  0 * SIZE(CO2)	STFD	f5,  1 * SIZE(CO2)	STFD	f6,  2 * SIZE(CO2)	STFD	f7,  3 * SIZE(CO2)	fmr	f4,  f0	fmr	f5,  f0	fmr	f6,  f0	fmr	f7,  f0	addi	CO1, CO1, 4 * SIZE	addi	CO2, CO2, 4 * SIZE#ifdef TRMMKERNEL#if ( defined(LEFT) &&  defined(TRANSA)) || \    (!defined(LEFT) && !defined(TRANSA))	sub	TEMP, K, KK#ifdef LEFT	addi	TEMP, TEMP, -4#else	addi	TEMP, TEMP, -2#endif	slwi	r0,   TEMP, 2 + BASE_SHIFT	slwi	TEMP, TEMP, 1 + BASE_SHIFT	add	AO, AO, r0	add	BO, BO, TEMP#endif#ifdef LEFT	addi	KK, KK, 4#endif#endif	addic.	I, I, -1	bgt+	LL(41)	.align 4LL(50):	andi.	I,  M,  2	ble	LL(60)#if defined(TRMMKERNEL)#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))	LFD	f16,  0 * SIZE(AO)	LFD	f17,  1 * SIZE(AO)	LFD	f18,  2 * SIZE(AO)	LFD	f19,  3 * SIZE(AO)	LFD	f20,  0 * SIZE(B)	LFD	f21,  1 * SIZE(B)	LFD	f22,  2 * SIZE(B)	LFD	f23,  3 * SIZE(B)	LFD	f24,  4 * SIZE(B)	LFD	f25,  5 * SIZE(B)	LFD	f26,  6 * SIZE(B)	LFD	f27,  7 * SIZE(B)	mr	BO,  B#else	slwi	r0,   KK, 1 + BASE_SHIFT	slwi	TEMP, KK, 1 + BASE_SHIFT	add	AO, AO, r0	add	BO, B,  TEMP	LFD	f16,  0 * SIZE(AO)	LFD	f17,  1 * SIZE(AO)	LFD	f18,  2 * SIZE(AO)	LFD	f19,  3 * SIZE(AO)	LFD	f20,  0 * SIZE(BO)	LFD	f21,  1 * SIZE(BO)	LFD	f22,  2 * SIZE(BO)	LFD	f23,  3 * SIZE(BO)	LFD	f24,  4 * SIZE(BO)	LFD	f25,  5 * SIZE(BO)	LFD	f26,  6 * SIZE(BO)	LFD	f27,  7 * SIZE(BO)#endif#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	sub	TEMP, K, KK#elif defined(LEFT)	addi	TEMP, KK, 2#else	addi	TEMP, KK, 2#endif	srawi.	TEMP,  TEMP,  2	mtspr	CTR, TEMP#else	LFD	f16,  0 * SIZE(AO)	LFD	f17,  1 * SIZE(AO)	LFD	f18,  2 * SIZE(AO)	LFD	f19,  3 * SIZE(AO)	LFD	f20,  0 * SIZE(B)	LFD	f21,  1 * SIZE(B)	LFD	f22,  2 * SIZE(B)	LFD	f23,  3 * SIZE(B)	LFD	f24,  4 * SIZE(B)	LFD	f25,  5 * SIZE(B)	LFD	f26,  6 * SIZE(B)	LFD	f27,  7 * SIZE(B)	srawi.	r0,  K,  2	mtspr	CTR, r0	mr	BO,  B#endif	ble	LL(55)	.align 5LL(52):	FMADD	f0,  f16, f20, f0	FMADD	f1,  f17, f20, f1	FMADD	f2,  f16, f21, f2	FMADD	f3,  f17, f21, f3	FMADD	f4,  f18, f22, f4	FMADD	f5,  f19, f22, f5	FMADD	f6,  f18, f23, f6	FMADD	f7,  f19, f23, f7	LFD	f16,  4 * SIZE(AO)	LFD	f17,  5 * SIZE(AO)	LFD	f18,  6 * SIZE(AO)	LFD	f19,  7 * SIZE(AO)	LFD	f20,  8 * SIZE(BO)	LFD	f21,  9 * SIZE(BO)	LFD	f22, 10 * SIZE(BO)	LFD	f23, 11 * SIZE(BO)	FMADD	f0,  f16, f24, f0	FMADD	f1,  f17, f24, f1	FMADD	f2,  f16, f25, f2	FMADD	f3,  f17, f25, f3	FMADD	f4,  f18, f26, f4	FMADD	f5,  f19, f26, f5	FMADD	f6,  f18, f27, f6	FMADD	f7,  f19, f27, f7	LFD	f16,  8 * SIZE(AO)	LFD	f17,  9 * SIZE(AO)	LFD	f18, 10 * SIZE(AO)	LFD	f19, 11 * SIZE(AO)	LFD	f24, 12 * SIZE(BO)	LFD	f25, 13 * SIZE(BO)	LFD	f26, 14 * SIZE(BO)	LFD	f27, 15 * SIZE(BO)	addi	AO, AO,  8 * SIZE	addi	BO, BO,  8 * SIZE	PREFETCH_B	bdnz	LL(52)	.align 4LL(55):	lfd	f30,  ALPHA#if defined(TRMMKERNEL)#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	sub	TEMP, K, KK#elif defined(LEFT)	addi	TEMP, KK, 2#else	addi	TEMP, KK, 2#endif	andi.	TEMP,  TEMP,  3	mtspr	CTR, TEMP#else	andi.	r0,  K,  3	mtspr	CTR, r0#endif	ble+	LL(58)	.align 4LL(56):	FMADD	f0,  f16, f20, f0	FMADD	f1,  f17, f20, f1	FMADD	f2,  f16, f21, f2	FMADD	f3,  f17, f21, f3	LFD	f16,  2 * SIZE(AO)	LFD	f17,  3 * SIZE(AO)	LFD	f20,  2 * SIZE(BO)	LFD	f21,  3 * SIZE(BO)	addi	BO, BO,  2 * SIZE	addi	AO, AO,  2 * SIZE	bdnz	LL(56)	.align 4LL(58):#ifndef TRMMKERNEL	LFD	f16, 0 * SIZE(CO1)	LFD	f17, 1 * SIZE(CO1)	LFD	f18, 0 * SIZE(CO2)	LFD	f19, 1 * SIZE(CO2)	FADD	f0, f4,  f0	FADD	f1, f5,  f1	FADD	f2, f6,  f2	FADD	f3, f7,  f3	FMADD	f0,  f0, f30, f16	FMADD	f1,  f1, f30, f17	FMADD	f2,  f2, f30, f18	FMADD	f3,  f3, f30, f19#else	FADD	f0, f4,  f0	FADD	f1, f5,  f1	FADD	f2, f6,  f2	FADD	f3, f7,  f3	FMUL	f0,  f0, f30	FMUL	f1,  f1, f30	FMUL	f2,  f2, f30	FMUL	f3,  f3, f30#endif	STFD	f0,  0 * SIZE(CO1)	STFD	f1,  1 * SIZE(CO1)	STFD	f2,  0 * SIZE(CO2)	STFD	f3,  1 * SIZE(CO2)	lfs	f0,  FZERO 	fmr	f1,  f0	fmr	f2,  f0	fmr	f3,  f0	fmr	f4,  f0	fmr	f5,  f0	fmr	f6,  f0	fmr	f7,  f0	addi	CO1, CO1, 2 * SIZE	addi	CO2, CO2, 2 * SIZE#ifdef TRMMKERNEL#if ( defined(LEFT) &&  defined(TRANSA)) || \    (!defined(LEFT) && !defined(TRANSA))	sub	TEMP, K, KK#ifdef LEFT	addi	TEMP, TEMP, -2#else	addi	TEMP, TEMP, -2#endif	slwi	r0,   TEMP, 1 + BASE_SHIFT	slwi	TEMP, TEMP, 1 + BASE_SHIFT	add	AO, AO, r0	add	BO, BO, TEMP#endif#ifdef LEFT	addi	KK, KK, 2#endif#endif	.align 4LL(60):	andi.	I,  M,  1	ble	LL(69)#if defined(TRMMKERNEL)#if (defined(LEFT) &&  defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))	LFD	f16,  0 * SIZE(AO)	LFD	f17,  1 * SIZE(AO)	LFD	f18,  2 * SIZE(AO)	LFD	f19,  3 * SIZE(AO)	LFD	f20,  0 * SIZE(B)	LFD	f21,  1 * SIZE(B)	LFD	f22,  2 * SIZE(B)	LFD	f23,  3 * SIZE(B)	LFD	f24,  4 * SIZE(B)	LFD	f25,  5 * SIZE(B)	LFD	f26,  6 * SIZE(B)	LFD	f27,  7 * SIZE(B)	mr	BO,  B#else	slwi	r0,   KK, 0 + BASE_SHIFT	slwi	TEMP, KK, 1 + BASE_SHIFT	add	AO, AO, r0	add	BO, B,  TEMP	LFD	f16,  0 * SIZE(AO)	LFD	f17,  1 * SIZE(AO)	LFD	f18,  2 * SIZE(AO)	LFD	f19,  3 * SIZE(AO)	LFD	f20,  0 * SIZE(BO)	LFD	f21,  1 * SIZE(BO)	LFD	f22,  2 * SIZE(BO)	LFD	f23,  3 * SIZE(BO)	LFD	f24,  4 * SIZE(BO)	LFD	f25,  5 * SIZE(BO)	LFD	f26,  6 * SIZE(BO)	LFD	f27,  7 * SIZE(BO)#endif#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	sub	TEMP, K, KK#elif defined(LEFT)	addi	TEMP, KK, 1#else	addi	TEMP, KK, 2#endif	srawi.	TEMP,  TEMP,  2	mtspr	CTR, TEMP#else	LFD	f16,  0 * SIZE(AO)	LFD	f17,  1 * SIZE(AO)	LFD	f18,  2 * SIZE(AO)	LFD	f19,  3 * SIZE(AO)	LFD	f20,  0 * SIZE(B)	LFD	f21,  1 * SIZE(B)	LFD	f22,  2 * SIZE(B)	LFD	f23,  3 * SIZE(B)	LFD	f24,  4 * SIZE(B)	LFD	f25,  5 * SIZE(B)	LFD	f26,  6 * SIZE(B)	LFD	f27,  7 * SIZE(B)	srawi.	r0,  K,  2	mtspr	CTR, r0	mr	BO,  B#endif	ble	LL(65)	.align 5LL(62):	FMADD	f0,  f16, f20, f0	FMADD	f1,  f16, f21, f1	FMADD	f2,  f17, f22, f2	FMADD	f3,  f17, f23, f3	LFD	f20,  8 * SIZE(BO)	LFD	f21,  9 * SIZE(BO)	LFD	f22, 10 * SIZE(BO)	LFD	f23, 11 * SIZE(BO)	FMADD	f0,  f18, f24, f0	FMADD	f1,  f18, f25, f1	FMADD	f2,  f19, f26, f2	FMADD	f3,  f19, f27, f3	LFD	f16,  4 * SIZE(AO)	LFD	f17,  5 * SIZE(AO)	LFD	f18,  6 * SIZE(AO)	LFD	f19,  7 * SIZE(AO)	LFD	f24, 12 * SIZE(BO)	LFD	f25, 13 * SIZE(BO)	LFD	f26, 14 * SIZE(BO)	LFD	f27, 15 * SIZE(BO)	addi	AO, AO,  4 * SIZE	addi	BO, BO,  8 * SIZE	bdnz	LL(62)	.align 4LL(65):	lfd	f30,  ALPHA#if defined(TRMMKERNEL)#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))	sub	TEMP, K, KK#elif defined(LEFT)	addi	TEMP, KK, 1#else	addi	TEMP, KK, 2#endif	andi.	TEMP,  TEMP,  3	mtspr	CTR, TEMP#else	andi.	r0,  K,  3	mtspr	CTR, r0#endif	ble+	LL(68)	.align 4LL(66):	FMADD	f0,  f16, f20, f0	FMADD	f1,  f16, f21, f1	LFD	f16,  1 * SIZE(AO)	LFD	f20,  2 * SIZE(BO)	LFD	f21,  3 * SIZE(BO)	addi	BO, BO,  2 * SIZE	addi	AO, AO,  1 * SIZE	bdnz	LL(66)	.align 4LL(68):#ifndef TRMMKERNEL	LFD	f16, 0 * SIZE(CO1)	LFD	f18, 0 * SIZE(CO2)	FADD	f0, f2, f0	FADD	f1, f3, f1	FMADD	f0,  f0,  f30, f16	FMADD	f1,  f1,  f30, f18#else	FADD	f0, f2, f0	FADD	f1, f3, f1	FMUL	f0,  f0,  f30	FMUL	f1,  f1,  f30#endif

⌨️ 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 + -