⭐ 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

📄 trsm_kernel_hummer_rt.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 5 页
字号:
	fxcpmadd	f1,  B4, A6, f1	LFPDUX	A6,  AO,  INC2	fxcsmadd	f2,  B4, A7, f2	LFPDUX	A7,  AO,  INC2	fxcsmadd	f3,  B4, A8, f3	LFPDUX	A8,  AO,  INC2	LFPDUX	B4,  BO,  INC2	bdnz+	.L102	.align 4.L103:	fxcpmadd	f0,  B1, A1, f0	LFPDUX	A1,  AO,  INC2	fxcpmadd	f1,  B1, A2, f1	LFPDUX	A2,  AO,  INC2	fxcsmadd	f2,  B1, A3, f2	LFPDUX	A3,  AO,  INC2	fxcsmadd	f3,  B1, A4, f3	LFPDUX	A4,  AO,  INC2	fxcpmadd	f0,  B2, A5, f0	LFPDUX	A5,  AO,  INC2	fxcpmadd	f1,  B2, A6, f1	LFPDUX	A6,  AO,  INC2	fxcsmadd	f2,  B2, A7, f2	LFPDUX	A7,  AO,  INC2	fxcsmadd	f3,  B2, A8, f3	LFPDUX	A8,  AO,  INC2	fxcpmadd	f0,  B3, A1, f0	fxcpmadd	f1,  B3, A2, f1	fxcsmadd	f2,  B3, A3, f2	fxcsmadd	f3,  B3, A4, f3	fxcpmadd	f0,  B4, A5, f0	fxcpmadd	f1,  B4, A6, f1	fxcsmadd	f2,  B4, A7, f2	fxcsmadd	f3,  B4, A8, f3	.align 4.L104:#if defined(LT) || defined(RN)	andi.	r0,  KK,  7	mtspr	CTR, r0	ble+	.L108#else	andi.	r0, TEMP, 7	mtspr	CTR, r0	ble+	.L108#endif	LFPDUX	A1,  AO,  INC2	LFDX	B1,  BO,  INC2	LFPDUX	A2,  AO,  INC2	add	BO, BO, INC	bdz-	.L107	.align 4.L106:	fxcpmadd	f0,  B1, A1, f0	LFPDUX	A1,  AO,  INC2	fxcpmadd	f1,  B1, A2, f1	LFDX	B1,  BO,  INC2	LFPDUX	A2,  AO,  INC2	add	BO, BO, INC	bdnz+	.L106	.align 4.L107:	fxcpmadd	f0,  B1, A1, f0	fxcpmadd	f1,  B1, A2, f1	.align 4.L108:	fpadd	f0, f0, f2	fpadd	f1, f1, f3#if defined(LN) || defined(RT)#ifdef LN	subi	r0, KK, 4#else	subi	r0, KK, 1#endif	slwi	TEMP, r0, 2 + BASE_SHIFT	slwi	r0,   r0, 0 + BASE_SHIFT	add	AO, AORIG, TEMP	add	BO,  B,     r0	addi	BO,  BO, - 2 * SIZE#endif#if defined(LN) || defined(LT)	LFPDUX	f16, BO,  INC2	LFPDUX	f17, BO,  INC2	subi	BO,  BO,   4 * SIZE	fpsub	f0,  f16,  f0	fpsub	f1,  f17,  f1#else	LFPDUX	f16, AO,  INC2	LFPDUX	f17, AO,  INC2	subi	AO,  AO,   4 * SIZE	fpsub	f0,  f16,  f0	fpsub	f1,  f17,  f1#endif#ifdef LN	fsmtp	f4, f0	fsmtp	f5, f1	LFD	A1, (2 + 15) * SIZE(AO)	LFD	A2, (2 + 14) * SIZE(AO)	LFD	A3, (2 + 13) * SIZE(AO)	LFD	A4, (2 + 12) * SIZE(AO)	fmul	f5, A1, f5	fnmsub	f1, A2, f5, f1	fnmsub	f4, A3, f5, f4	fnmsub	f0, A4, f5, f0	LFD	A1, (2 + 10) * SIZE(AO)	LFD	A2, (2 +  9) * SIZE(AO)	LFD	A3, (2 +  8) * SIZE(AO)	fmul	f1, A1, f1	fnmsub	f4, A2, f1, f4	fnmsub	f0, A3, f1, f0	LFD	A1, (2 +  5) * SIZE(AO)	LFD	A2, (2 +  4) * SIZE(AO)	fmul	f4, A1, f4	fnmsub	f0, A2, f4, f0	LFD	A1, (2 +  0) * SIZE(AO)	fmul	f0, A1, f0	fsmfp	f0, f4	fsmfp	f1, f5#endif#ifdef LT	fsmtp	f4, f0	fsmtp	f5, f1	LFD	A1, (2 +  0) * SIZE(AO)	LFD	A2, (2 +  1) * SIZE(AO)	LFD	A3, (2 +  2) * SIZE(AO)	LFD	A4, (2 +  3) * SIZE(AO)	fmul	f0, A1, f0	fnmsub	f4, A2, f0, f4	fnmsub	f1, A3, f0, f1	fnmsub	f5, A4, f0, f5	LFD	A1, (2 +  5) * SIZE(AO)	LFD	A2, (2 +  6) * SIZE(AO)	LFD	A3, (2 +  7) * SIZE(AO)	fmul	f4, A1, f4	fnmsub	f1, A2, f4, f1	fnmsub	f5, A3, f4, f5	LFD	A1, (2 + 10) * SIZE(AO)	LFD	A2, (2 + 11) * SIZE(AO)	fmul	f1, A1, f1	fnmsub	f5, A2, f1, f5	LFD	A1, (2 + 15) * SIZE(AO)	fmul	f5, A1, f5	fsmfp	f0, f4	fsmfp	f1, f5#endif#ifdef RN	LFPDX	A1,  BO,  INC2	fxpmul	f0,  A1,  f0	fxpmul	f1,  A1,  f1#endif#ifdef RT	LFPDX	A1,  BO,  INC2	fxpmul	f0,  A1,  f0	fxpmul	f1,  A1,  f1#endif#ifdef LN	subi	CO1, CO1, 4 * SIZE#endif#if defined(LN) || defined(LT)	STFPDUX	f0,  BO,  INC2	STFPDUX	f1,  BO,  INC2	subi	BO,  BO,   4 * SIZE	STFDUX	f0,  CO1, INC	STFSDUX	f0,  CO1, INC	STFDUX	f1,  CO1, INC	STFSDUX	f1,  CO1, INC#else	STFPDUX	f0,  AO,  INC2	STFPDUX	f1,  AO,  INC2	subi	AO,  AO,   4 * SIZE	STFDUX	f0,  CO1, INC	STFSDUX	f0,  CO1, INC	STFDUX	f1,  CO1, INC	STFSDUX	f1,  CO1, INC#endif#ifdef LN	subi	CO1, CO1, 4 * SIZE#endif#ifdef RT	slwi	r0, K, 2 + BASE_SHIFT	add	AORIG, AORIG, r0#endif#if defined(LT) || defined(RN)	sub	TEMP, K, KK	slwi	r0,   TEMP, 2 + BASE_SHIFT	slwi	TEMP, TEMP, 0 + BASE_SHIFT	add	AO, AO, r0	add	BO, BO, TEMP#endif#ifdef LT	addi	KK, KK, 4#endif#ifdef LN	subi	KK, KK, 4#endif	li	r0, FZERO	lfpsx	f0, SP, r0	.align 4.L110:	andi.	I, M,  2	beq	.L120#if defined(LT) || defined(RN)	addi	BO,  B,  - 2 * SIZE	fpmr	f1,  f0	fpmr	f2,  f0	fpmr	f3,  f0	srawi.	r0,  KK,  3	mtspr	CTR, r0	ble	.L114#else#ifdef LN	slwi	r0,   K,  1 + BASE_SHIFT	sub	AORIG, AORIG, r0#endif	slwi	r0  , KK, 1 + BASE_SHIFT	slwi	TEMP, KK, 0 + BASE_SHIFT	add	AO, AORIG, r0	add	BO, B,     TEMP	sub	TEMP, K, KK	addi	BO,  BO,  - 2 * SIZE	fpmr	f1,  f0	fpmr	f2,  f0	fpmr	f3,  f0	srawi.	r0, TEMP,  3	mtspr	CTR, r0	ble	.L114#endif		LFPDUX	A1,  AO,  INC2	LFPDUX	A2,  AO,  INC2	LFPDUX	B1,  BO,  INC2	LFPDUX	A3,  AO,  INC2	LFPDUX	A4,  AO,  INC2	LFPDUX	B2,  BO,  INC2	LFPDUX	A5,  AO,  INC2	LFPDUX	A6,  AO,  INC2	LFPDUX	B3,  BO,  INC2	LFPDUX	A7,  AO,  INC2	LFPDUX	A8,  AO,  INC2	LFPDUX	B4,  BO,  INC2	bdz-	.L113	.align 4.L112:	fxcpmadd	f0,  B1, A1, f0	LFPDUX	A1,  AO,  INC2	fxcsmadd	f1,  B1, A2, f1	LFPDUX	A2,  AO,  INC2	LFPDUX	B1,  BO,  INC2	fxcpmadd	f2,  B2, A3, f2	LFPDUX	A3,  AO,  INC2	fxcsmadd	f3,  B2, A4, f3	LFPDUX	A4,  AO,  INC2	LFPDUX	B2,  BO,  INC2	fxcpmadd	f0,  B3, A5, f0	LFPDUX	A5,  AO,  INC2	fxcsmadd	f1,  B3, A6, f1	LFPDUX	A6,  AO,  INC2	LFPDUX	B3,  BO,  INC2	fxcpmadd	f2,  B4, A7, f2	LFPDUX	A7,  AO,  INC2	fxcsmadd	f3,  B4, A8, f3	LFPDUX	A8,  AO,  INC2	LFPDUX	B4,  BO,  INC2	bdnz+	.L112	.align 4.L113:	fxcpmadd	f0,  B1, A1, f0	fxcsmadd	f1,  B1, A2, f1	fxcpmadd	f2,  B2, A3, f2	fxcsmadd	f3,  B2, A4, f3	fxcpmadd	f0,  B3, A5, f0	fxcsmadd	f1,  B3, A6, f1	fxcpmadd	f2,  B4, A7, f2	fxcsmadd	f3,  B4, A8, f3	.align 4.L114:#if defined(LT) || defined(RN)	andi.	r0,  KK,  7	mtspr	CTR, r0	ble+	.L118#else	andi.	r0, TEMP, 7	mtspr	CTR, r0	ble+	.L118#endif	LFPDUX	A1,  AO,  INC2	LFDX	B1,  BO,  INC2	add	BO, BO, INC	bdz-	.L117	.align 4.L116:	fxcpmadd	f0,  B1, A1, f0	LFPDUX	A1,  AO,  INC2	LFDX	B1,  BO,  INC2	add	BO, BO, INC	bdnz+	.L116	.align 4.L117:	fxcpmadd	f0,  B1, A1, f0	.align 4.L118:	fpadd	f0, f0, f1	fpadd	f2, f3, f2	fpadd	f0, f0, f2#if defined(LN) || defined(RT)#ifdef LN	subi	r0, KK, 2#else	subi	r0, KK, 1#endif	slwi	TEMP, r0, 1 + BASE_SHIFT	slwi	r0,   r0, 0 + BASE_SHIFT	add	AO, AORIG, TEMP	add	BO,  B,     r0	addi	BO,  BO, - 2 * SIZE#endif#if defined(LN) || defined(LT)	LFPDX	f16, BO,  INC2	fpsub	f0,  f16,  f0#else	LFPDX	f16, AO,  INC2	fpsub	f0,  f16,  f0#endif#ifdef LN	fsmtp	f4, f0	LFD	A1, (2 +  3) * SIZE(AO)	LFD	A2, (2 +  2) * SIZE(AO)	LFD	A3, (2 +  0) * SIZE(AO)	fmul	f4, A1, f4	fnmsub	f0, A2, f4, f0	fmul	f0, A3, f0	fsmfp	f0, f4#endif#ifdef LT	fsmtp	f4, f0	LFD	A1, (2 +  0) * SIZE(AO)	LFD	A2, (2 +  1) * SIZE(AO)	LFD	A3, (2 +  3) * SIZE(AO)	fmul	f0, A1, f0	fnmsub	f4, A2, f0, f4	fmul	f4, A3, f4	fsmfp	f0, f4#endif#ifdef RN	LFPDX	A1,  BO,  INC2	fxpmul	f0,  A1,  f0#endif#ifdef RT	LFPDX	A1,  BO,  INC2	fxpmul	f0,  A1,  f0#endif#ifdef LN	subi	CO1, CO1, 2 * SIZE#endif#if defined(LN) || defined(LT)	STFPDX	f0,  BO,  INC2	STFDUX	f0,  CO1, INC	STFSDUX	f0,  CO1, INC#else	STFPDX	f0,  AO,  INC2	STFDUX	f0,  CO1, INC	STFSDUX	f0,  CO1, INC#endif#ifdef LN	subi	CO1, CO1, 2 * SIZE#endif#ifdef RT	slwi	r0, K, 1 + BASE_SHIFT	add	AORIG, AORIG, r0#endif#if defined(LT) || defined(RN)	sub	TEMP, K, KK	slwi	r0,   TEMP, 1 + BASE_SHIFT	slwi	TEMP, TEMP, 0 + BASE_SHIFT	add	AO, AO, r0	add	BO, BO, TEMP#endif#ifdef LT	addi	KK, KK, 2#endif#ifdef LN	subi	KK, KK, 2#endif	li	r0, FZERO	lfpsx	f0, SP, r0	.align 4.L120:	andi.	I, M,  1	beq	.L129#if defined(LT) || defined(RN)	addi	BO,  B,  - 2 * SIZE	fpmr	f1,  f0	fpmr	f2,  f0	fpmr	f3,  f0	srawi.	r0,  KK,  3	mtspr	CTR, r0	ble	.L124#else#ifdef LN	slwi	r0,   K,  0 + BASE_SHIFT	sub	AORIG, AORIG, r0#endif	slwi	r0  , KK, 0 + BASE_SHIFT	slwi	TEMP, KK, 0 + BASE_SHIFT	add	AO, AORIG, r0	add	BO, B,     TEMP	sub	TEMP, K, KK	addi	BO,  BO,  - 2 * SIZE	fpmr	f1,  f0	fpmr	f2,  f0	fpmr	f3,  f0	srawi.	r0,  TEMP,  3	mtspr	CTR, r0	ble	.L124#endif	LFPDUX	A1,  AO,  INC2	LFPDUX	B1,  BO,  INC2	LFPDUX	A2,  AO,  INC2	LFPDUX	B2,  BO,  INC2	LFPDUX	A3,  AO,  INC2	LFPDUX	B3,  BO,  INC2	LFPDUX	A4,  AO,  INC2	LFPDUX	B4,  BO,  INC2	bdz-	.L123	.align 4.L122:	fpmadd	f0,  A1, B1, f0	LFPDUX	A1,  AO,  INC2	LFPDUX	B1,  BO,  INC2	fpmadd	f1,  A2, B2, f1	LFPDUX	A2,  AO,  INC2	LFPDUX	B2,  BO,  INC2	fpmadd	f2,  A3, B3, f2	LFPDUX	A3,  AO,  INC2	LFPDUX	B3,  BO,  INC2	fpmadd	f3,  A4, B4, f3	LFPDUX	A4,  AO,  INC2	LFPDUX	B4,  BO,  INC2	bdnz+	.L122	.align 4.L123:	fpmadd	f0,  A1, B1, f0	fpmadd	f1,  A2, B2, f1	fpmadd	f2,  A3, B3, f2	fpmadd	f3,  A4, B4, f3	.align 4.L124:#if defined(LT) || defined(RN)	andi.	r0,  KK,  7	mtspr	CTR, r0	ble+	.L128#else	andi.	r0, TEMP, 7	mtspr	CTR, r0	ble+	.L128#endif	LFDX	A1,  AO,  INC2	LFDX	B1,  BO,  INC2	add	AO, AO, INC	add	BO, BO, INC	bdz-	.L127	.align 4.L126:	fmadd	f0,  A1, B1, f0	LFDX	A1,  AO,  INC2	LFDX	B1,  BO,  INC2	add	AO, AO, INC	add	BO, BO, INC	bdnz+	.L126	.align 4.L127:	fmadd	f0,  A1, B1, f0	.align 4.L128:	fpadd	f0, f0, f1	fpadd	f2, f2, f3	fpadd	f0, f0, f2	fsmtp	f1, f0	fadd	f0, f0, f1#if defined(LN) || defined(RT)#ifdef LN	subi	r0, KK, 1#else	subi	r0, KK, 1#endif	slwi	TEMP, r0, 0 + BASE_SHIFT	slwi	r0,   r0, 0 + BASE_SHIFT	add	AO, AORIG, TEMP	add	BO,  B,     r0	addi	BO,  BO, - 2 * SIZE#endif#if defined(LN) || defined(LT)	LFDX	f16, BO,  INC2	fsub	f0,  f16,  f0#else	LFDX	f16, AO,  INC2	fsub	f0,  f16,  f0#endif#ifdef LN	LFD	A1, (2 +  0) * SIZE(AO)	fmul	f0, A1, f0#endif#ifdef LT	LFD	A1, (2 +  0) * SIZE(AO)	fmul	f0, A1, f0#endif#ifdef RN	LFDX	A1,  BO,  INC2	fmul	f0,  A1,  f0#endif#ifdef RT	LFDX	A1,  BO,  INC2	fmul	f0,  A1,  f0#endif#ifdef LN	subi	CO1, CO1, 1 * SIZE#endif#if defined(LN) || defined(LT)	STFDX	f0,  BO,  INC2	STFDUX	f0,  CO1, INC#else	STFDX	f0,  AO,  INC2	STFDUX	f0,  CO1, INC#endif#ifdef LN	subi	CO1, CO1, 1 * SIZE#endif#ifdef RT	slwi	r0, K, 0 + BASE_SHIFT	add	AORIG, AORIG, r0#endif#if defined(LT) || defined(RN)	sub	TEMP, K, KK	slwi	r0,   TEMP, 0 + BASE_SHIFT	slwi	TEMP, TEMP, 0 + BASE_SHIFT	add	AO, AO, r0	add	BO, BO, TEMP#endif#ifdef LT	addi	KK, KK, 1#endif#ifdef LN	subi	KK, KK, 1#endif	.align 4.L129:#ifdef LN	slwi	r0, K, 0 + BASE_SHIFT	add	B, B, r0#endif#if defined(LT) || defined(RN)	addi	B,  BO, 2 * SIZE#endif#ifdef RN	addi	KK, KK, 1#endif#ifdef RT	subi	KK, KK, 1#endif	.align 4.L50:	andi.	J, N,  2	beq	.L90#ifdef RT	slwi	r0, K, 1 + BASE_SHIFT	sub	B, B, r0	slwi	r0, LDC, 1	sub	C, C, r0#endif	mr	CO1, C	add	CO2, C,   LDC#ifdef LN	add	KK, M, OFFSET#endif#ifdef LT	mr	KK, OFFSET#endif#if defined(LN) || defined(RT)	addi	AORIG, A, -2 * SIZE#else	addi	AO, A, -2 * SIZE#endif#ifndef RT	add	C,  CO2, LDC#endif	li	r0, FZERO	lfpsx	f0, SP, r0	srawi.	I, M,  3	ble	.L60	.align 4.L51:#if defined(LT) || defined(RN)	fpmr	f4,  f0	addi	BO,   B,  - 2 * SIZE 	fpmr	f1,  f0	fpmr	f5,  f0	fpmr	f2,  f0	fpmr	f6,  f0	srawi.	r0,  KK,  2	fpmr	f3,  f0	mtspr	CTR, r0	fpmr	f7,  f0	ble	.L54#else#ifdef LN	slwi	r0,   K,  3 + BASE_SHIFT	sub	AORIG, AORIG, r0#endif	slwi	r0  , KK, 3 + BASE_SHIFT	slwi	TEMP, KK, 1 + BASE_SHIFT	add	AO, AORIG, r0	add	BO, B,     TEMP	sub	TEMP, K, KK	fpmr	f4,  f0	addi	BO,  BO,  - 2 * SIZE 	fpmr	f1,  f0	fpmr	f5,  f0	fpmr	f2,  f0	fpmr	f6,  f0	srawi.	r0,  TEMP,  2	fpmr	f3,  f0	mtspr	CTR, r0	fpmr	f7,  f0	ble	.L54#endif	LFPDUX	B1,  BO,  INC2	LFPDUX	A1,  AO,  INC2	LFPDUX	A2,  AO,  INC2	LFPDUX	B2,  BO,  INC2	LFPDUX	A3,  AO,  INC2	LFPDUX	A4,  AO,  INC2	LFPDUX	B3,  BO,  INC2	LFPDUX	A5,  AO,  INC2	LFPDUX	A6,  AO,  INC2	LFPDUX	A7,  AO,  INC2	LFPDUX	A8,  AO,  INC2

⌨️ 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 + -