📄 ztrsm_kernel_hummer_lt.s

📁 Optimized GotoBLAS libraries
💻 S
📖 第 1 页 / 共 4 页
字号:
#ifdef LN	subi	CO1, CO1, 4 * SIZE	subi	CO2, CO2, 4 * SIZE#endif#if defined(LN) || defined(LT)	STFPDUX	f0,  BO,  INC4	STFPDUX	f8,  BO2, INC4	STFPDUX	f1,  BO,  INC4	STFPDUX	f9,  BO2, INC4	subi	BO,  BO,   8 * SIZE	subi	BO2, BO2,  8 * SIZE#else	STFPDUX	f0,  AO,  INC4	STFPDUX	f1,  AO2, INC4	STFPDUX	f8,  AO,  INC4	STFPDUX	f9,  AO2, INC4	subi	AO,  AO,   8 * SIZE	subi	AO2, AO2,  8 * SIZE#endif	STFDUX	f0,  CO1, INC	STFSDUX	f0,  CO1, INC	STFDUX	f1,  CO1, INC	STFSDUX	f1,  CO1, INC	STFDUX	f8,  CO2, INC	STFSDUX	f8,  CO2, INC	STFDUX	f9,  CO2, INC	STFSDUX	f9,  CO2, INC#ifdef LN	subi	CO1, CO1, 4 * SIZE	subi	CO2, CO2, 4 * SIZE#endif#ifdef RT	slwi	r0, K, 1 + ZBASE_SHIFT	add	AORIG, AORIG, r0#endif#if defined(LT) || defined(RN)	sub	TEMP, K, KK	slwi	r0,   TEMP, 1 + ZBASE_SHIFT	add	AO, AO, r0	add	BO, BO, r0#endif#ifdef LT	addi	KK, KK, 2#endif#ifdef LN	subi	KK, KK, 2#endif	li	r0, FZERO	lfpsx	f0, SP, r0	.align 4.L30:	andi.	I, M,  1	beq	.L49#if defined(LT) || defined(RN)	addi	AO2, AO,   2 * SIZE	fpmr	f1,  f0	addi	BO,  B,  - 4 * SIZE	fpmr	f2,  f0	addi	BO2, B,  - 2 * SIZE	fpmr	f3, f0	srawi.	r0,  KK,  2	mtspr	CTR, r0	ble	.L34#else#ifdef LN	slwi	r0,   K,  0 + ZBASE_SHIFT	sub	AORIG, AORIG, r0#endif	slwi	r0  , KK, 0 + ZBASE_SHIFT	slwi	TEMP, KK, 1 + ZBASE_SHIFT	add	AO, AORIG, r0	add	BO, B,     TEMP	sub	TEMP, K, KK	addi	AO2, AO,   2 * SIZE	fpmr	f1,  f0	addi	BO,  BO,  - 4 * SIZE	fpmr	f2,  f0	addi	BO2, BO,    2 * SIZE	fpmr	f3, f0	srawi.	r0,  TEMP,  2	mtspr	CTR, r0	ble	.L34#endif	LFPDUX	A1,  AO, INC4	LFPDUX	B1,  BO, INC4	LFPDUX	B2, BO2, INC4	LFPDUX	A2, AO2, INC4	LFPDUX	B3,  BO, INC4	LFPDUX	B4, BO2, INC4	LFPDUX	A3,  AO, INC4	LFPDUX	A5,  BO, INC4	LFPDUX	A6, BO2, INC4	LFPDUX	A4, AO2, INC4	LFPDUX	A7,  BO, INC4	LFPDUX	A8, BO2, INC4	bdz-	.L33	.align 4.L32:	FXCPMADD	f0,  B1, A1, f0	FXCSMADD	f1,  B1, A1, f1	LFPDUX	B1,  BO, INC4	FXCPMADD	f2,  B2, A1, f2	FXCSMADD	f3,  B2, A1, f3	LFPDUX	B2, BO2, INC4	LFPDUX	A1,  AO, INC4	FXCPMADD	f0,  B3, A2, f0	FXCSMADD	f1,  B3, A2, f1	LFPDUX	B3,  BO, INC4	FXCPMADD	f2,  B4, A2, f2	FXCSMADD	f3,  B4, A2, f3	LFPDUX	B4, BO2, INC4	LFPDUX	A2, AO2, INC4	FXCPMADD	f0,  A5, A3, f0	FXCSMADD	f1,  A5, A3, f1	LFPDUX	A5,  BO, INC4	FXCPMADD	f2,  A6, A3, f2	FXCSMADD	f3,  A6, A3, f3	LFPDUX	A6, BO2, INC4	LFPDUX	A3,  AO, INC4	FXCPMADD	f0,  A7, A4, f0	FXCSMADD	f1,  A7, A4, f1	LFPDUX	A7,  BO, INC4	FXCPMADD	f2,  A8, A4, f2	FXCSMADD	f3,  A8, A4, f3	LFPDUX	A8, BO2, INC4	LFPDUX	A4, AO2, INC4	bdnz+	.L32	.align 4.L33:	FXCPMADD	f0,  B1, A1, f0	FXCSMADD	f1,  B1, A1, f1	FXCPMADD	f2,  B2, A1, f2	FXCSMADD	f3,  B2, A1, f3	FXCPMADD	f0,  B3, A2, f0	FXCSMADD	f1,  B3, A2, f1	FXCPMADD	f2,  B4, A2, f2	FXCSMADD	f3,  B4, A2, f3	FXCPMADD	f0,  A5, A3, f0	FXCSMADD	f1,  A5, A3, f1	FXCPMADD	f2,  A6, A3, f2	FXCSMADD	f3,  A6, A3, f3	FXCPMADD	f0,  A7, A4, f0	FXCSMADD	f1,  A7, A4, f1	FXCPMADD	f2,  A8, A4, f2	FXCSMADD	f3,  A8, A4, f3	.align 4.L34:#if defined(LT) || defined(RN)	andi.	r0,  KK,  3	mtspr	CTR, r0	ble+	.L38#else	andi.	r0, TEMP, 3	mtspr	CTR, r0	ble+	.L38#endif	LFPDX	A1,  AO,  INC4	LFPDUX	B1,  BO,  INC4	LFPDUX	B2,  BO2, INC4	add	AO, AO, INC2	bdz-	.L37	.align 4.L36:	FXCPMADD	f0,  B1, A1, f0	FXCSMADD	f1,  B1, A1, f1	LFPDUX	B1,  BO,  INC4	FXCPMADD	f2,  B2, A1, f2	FXCSMADD	f3,  B2, A1, f3	LFPDX	A1,  AO,  INC4	LFPDUX	B2,  BO2, INC4	add	AO, AO, INC2	bdnz+	.L36	.align 4.L37:	FXCPMADD	f0,  B1, A1, f0	FXCSMADD	f1,  B1, A1, f1	FXCPMADD	f2,  B2, A1, f2	FXCSMADD	f3,  B2, A1, f3	.align 4.L38:	fpadd	f0, f0, f1	fpadd	f2, f2, f3#if defined(LN) || defined(RT)#ifdef LN	subi	r0, KK, 1#else	subi	r0, KK, 2#endif	slwi	TEMP, r0, 0 + ZBASE_SHIFT	slwi	r0,   r0, 1 + ZBASE_SHIFT	add	AO, AORIG, TEMP	add	BO, B,     r0	addi	BO,  BO, - 4 * SIZE#endif	addi	AO2, AO,   2 * SIZE	addi	BO2, BO,   2 * SIZE#if defined(LN) || defined(LT)	LFPDX	f16, BO,  INC4	LFPDX	f17, BO2, INC4#else	LFPDX	f16, AO,  INC4	LFPDX	f17, AO2, INC4#endif	fpsub	f0,  f16,  f0	fpsub	f2,  f17,  f2#ifdef LN	LFPDX	A1,  AO,  INC4	fxpmul	  f4,  A1, f0	fxpmul	  f5,  A1, f2	FXCXNPMA  f0,  A1, f0,  f4	FXCXNPMA  f2,  A1, f2,  f5#endif#ifdef LT	LFPDX	A1,  AO,  INC4	fxpmul	  f4,  A1, f0	fxpmul	  f5,  A1, f2	FXCXNPMA  f0,  A1, f0, f4	FXCXNPMA  f2,  A1, f2, f5#endif#ifdef RN	LFPDUX	A1,  BO,  INC4	LFPDUX	A2,  BO2, INC4	add	BO,  BO,  INC4	LFPDUX	A3,  BO2, INC4	subi	BO,  BO,   8 * SIZE	subi	BO2, BO2,  8 * SIZE	fxpmul	  f4,  A1, f0	FXCXNPMA  f0,  A1, f0, f4	fxcpnmsub f2,  A2, f0, f2	FXCXNSMA  f2,  A2, f0, f2	fxpmul	  f4,  A3, f2	FXCXNPMA  f2,  A3, f2,  f4#endif#ifdef RT	LFPDUX	A1,  BO,  INC4	add	BO2, BO2, INC4	LFPDUX	A2,  BO,  INC4	LFPDUX	A3,  BO2, INC4	subi	BO,  BO,   8 * SIZE	subi	BO2, BO2,  8 * SIZE	fxpmul	  f4,  A3, f2	FXCXNPMA  f2,  A3, f2,  f4	fxcpnmsub f0,  A2, f2,  f0	FXCXNSMA  f0,  A2, f2,  f0	fxpmul	  f4,  A1, f0	FXCXNPMA  f0,  A1, f0,  f4#endif#ifdef LN	subi	CO1, CO1, 2 * SIZE	subi	CO2, CO2, 2 * SIZE#endif#if defined(LN) || defined(LT)	STFPDX	f0,  BO,  INC4	STFPDX	f2,  BO2, INC4#else	STFPDX	f0,  AO,  INC4	STFPDX	f2,  AO2, INC4#endif	STFDUX	f0,  CO1, INC	STFSDUX	f0,  CO1, INC	STFDUX	f2,  CO2, INC	STFSDUX	f2,  CO2, INC#ifdef LN	subi	CO1, CO1, 2 * SIZE	subi	CO2, CO2, 2 * SIZE#endif#ifdef RT	slwi	r0, K, 0 + ZBASE_SHIFT	add	AORIG, AORIG, r0#endif#if defined(LT) || defined(RN)	sub	TEMP, K, KK	slwi	r0,   TEMP, 0 + ZBASE_SHIFT	slwi	TEMP, TEMP, 1 + ZBASE_SHIFT	add	AO, AO, r0	add	BO, BO, TEMP#endif#ifdef LT	addi	KK, KK, 1#endif#ifdef LN	subi	KK, KK, 1#endif	li	r0, FZERO	lfpsx	f0, SP, r0	.align 4.L49:#ifdef LN	slwi	r0, K, 1 + ZBASE_SHIFT	add	B, B, r0#endif#if defined(LT) || defined(RN)	addi	B,  BO, 4 * SIZE#endif#ifdef RN	addi	KK, KK, 2#endif#ifdef RT	subi	KK, KK, 2#endif	addic.	J, J, -1	bgt+	.L10	.align 4.L50:	andi.	J, N,  1	beq	.L999#ifdef RT	slwi	r0, K, 0 + ZBASE_SHIFT	sub	B, B, r0	sub	C, C, LDC#endif	mr	CO1, C#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,  2	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	fpmr	f3,  f0	fpmr	f7,  f0	srawi.	r0,  KK,  2	mtspr	CTR, r0	ble	.L54#else#ifdef LN	slwi	r0,   K,  2 + ZBASE_SHIFT	sub	AORIG, AORIG, r0#endif	slwi	r0  , KK, 2 + ZBASE_SHIFT	slwi	TEMP, KK, 0 + ZBASE_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	fpmr	f3,  f0	fpmr	f7,  f0	srawi.	r0, TEMP,  2	mtspr	CTR, r0	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	bdz-	.L53	.align 4.L52:	FXCPMADD	f0,  B1, A1, f0	LFPDUX	B4,  BO,  INC2	FXCSMADD	f4,  B1, A1, f4	LFPDUX	A1,  AO,  INC2	FXCPMADD	f1,  B1, A2, f1	nop	FXCSMADD	f5,  B1, A2, f5	LFPDUX	A2,  AO,  INC2	FXCPMADD	f2,  B1, A3, f2	nop	FXCSMADD	f6,  B1, A3, f6	LFPDUX	A3,  AO,  INC2	FXCPMADD	f3,  B1, A4, f3	nop	FXCSMADD	f7,  B1, A4, f7	LFPDUX	A4,  AO,  INC2	FXCPMADD	f0,  B2, A5, f0	LFPDUX	B1,  BO,  INC2	FXCSMADD	f4,  B2, A5, f4	LFPDUX	A5,  AO,  INC2	FXCPMADD	f1,  B2, A6, f1	nop	FXCSMADD	f5,  B2, A6, f5	LFPDUX	A6,  AO,  INC2	FXCPMADD	f2,  B2, A7, f2	nop	FXCSMADD	f6,  B2, A7, f6	LFPDUX	A7,  AO,  INC2	FXCPMADD	f3,  B2, A8, f3	nop	FXCSMADD	f7,  B2, A8, f7	LFPDUX	A8,  AO,  INC2	FXCPMADD	f0,  B3, A1, f0	LFPDUX	B2,  BO,  INC2	FXCSMADD	f4,  B3, A1, f4	LFPDUX	A1,  AO,  INC2	FXCPMADD	f1,  B3, A2, f1	nop	FXCSMADD	f5,  B3, A2, f5	LFPDUX	A2,  AO,  INC2	FXCPMADD	f2,  B3, A3, f2	nop	FXCSMADD	f6,  B3, A3, f6	LFPDUX	A3,  AO,  INC2	FXCPMADD	f3,  B3, A4, f3	nop	FXCSMADD	f7,  B3, A4, f7	LFPDUX	A4,  AO,  INC2	FXCPMADD	f0,  B4, A5, f0	LFPDUX	B3,  BO,  INC2	FXCSMADD	f4,  B4, A5, f4	LFPDUX	A5,  AO,  INC2	FXCPMADD	f1,  B4, A6, f1	nop	FXCSMADD	f5,  B4, A6, f5	LFPDUX	A6,  AO,  INC2	FXCPMADD	f2,  B4, A7, f2	nop	FXCSMADD	f6,  B4, A7, f6	LFPDUX	A7,  AO,  INC2	FXCPMADD	f3,  B4, A8, f3	nop	FXCSMADD	f7,  B4, A8, f7	LFPDUX	A8,  AO,  INC2	bdnz+	.L52	.align 4.L53:	FXCPMADD	f0,  B1, A1, f0	LFPDUX	B4,  BO,  INC2	FXCSMADD	f4,  B1, A1, f4	LFPDUX	A1,  AO,  INC2	FXCPMADD	f1,  B1, A2, f1	nop	FXCSMADD	f5,  B1, A2, f5	LFPDUX	A2,  AO,  INC2	FXCPMADD	f2,  B1, A3, f2	nop	FXCSMADD	f6,  B1, A3, f6	LFPDUX	A3,  AO,  INC2	FXCPMADD	f3,  B1, A4, f3	nop	FXCSMADD	f7,  B1, A4, f7	LFPDUX	A4,  AO,  INC2	FXCPMADD	f0,  B2, A5, f0	nop	FXCSMADD	f4,  B2, A5, f4	LFPDUX	A5,  AO,  INC2	FXCPMADD	f1,  B2, A6, f1	nop	FXCSMADD	f5,  B2, A6, f5	LFPDUX	A6,  AO,  INC2	FXCPMADD	f2,  B2, A7, f2	nop	FXCSMADD	f6,  B2, A7, f6	LFPDUX	A7,  AO,  INC2	FXCPMADD	f3,  B2, A8, f3	nop	FXCSMADD	f7,  B2, A8, f7	LFPDUX	A8,  AO,  INC2	FXCPMADD	f0,  B3, A1, f0	FXCSMADD	f4,  B3, A1, f4	FXCPMADD	f1,  B3, A2, f1	FXCSMADD	f5,  B3, A2, f5	FXCPMADD	f2,  B3, A3, f2	FXCSMADD	f6,  B3, A3, f6	FXCPMADD	f3,  B3, A4, f3	FXCSMADD	f7,  B3, A4, f7	FXCPMADD	f0,  B4, A5, f0	FXCSMADD	f4,  B4, A5, f4	FXCPMADD	f1,  B4, A6, f1	FXCSMADD	f5,  B4, A6, f5	FXCPMADD	f2,  B4, A7, f2	FXCSMADD	f6,  B4, A7, f6	FXCPMADD	f3,  B4, A8, f3	FXCSMADD	f7,  B4, A8, f7	.align 4.L54:#if defined(LT) || defined(RN)	andi.	r0,  KK,  3	mtspr	CTR, r0	ble+	.L58#else	andi.	r0, TEMP, 3	mtspr	CTR, r0	ble+	.L58#endif	LFPDUX	A1,  AO,  INC2	LFPDUX	B1,  BO,  INC2	LFPDUX	A2,  AO,  INC2	LFPDUX	A3,  AO,  INC2	LFPDUX	A4,  AO,  INC2	bdz-	.L57	.align 4.L56:	FXCPMADD	f0,  B1, A1, f0	FXCSMADD	f4,  B1, A1, f4	LFPDUX	A1,  AO,  INC2	FXCPMADD	f1,  B1, A2, f1	FXCSMADD	f5,  B1, A2, f5	LFPDUX	A2,  AO,  INC2	FXCPMADD	f2,  B1, A3, f2	FXCSMADD	f6,  B1, A3, f6	LFPDUX	A3,  AO,  INC2	FXCPMADD	f3,  B1, A4, f3	FXCSMADD	f7,  B1, A4, f7	LFPDUX	A4,  AO,  INC2	LFPDUX	B1,  BO,  INC2	bdnz+	.L56	.align 4.L57:	FXCPMADD	f0,  B1, A1, f0	FXCSMADD	f4,  B1, A1, f4	FXCPMADD	f1,  B1, A2, f1	FXCSMADD	f5,  B1, A2, f5	FXCPMADD	f2,  B1, A3, f2	FXCSMADD	f6,  B1, A3, f6	FXCPMADD	f3,  B1, A4, f3	FXCSMADD	f7,  B1, A4, f7	.align 4.L58:	fpadd	f0,  f0,  f4	fpadd	f1,  f1,  f5	fpadd	f2,  f2,  f6	fpadd	f3,  f3,  f7#if defined(LN) || defined(RT)#ifdef LN	subi	r0, KK, 4#else	subi	r0, KK, 1#endif	slwi	TEMP, r0, 2 + ZBASE_SHIFT	slwi	r0,   r0, 0 + ZBASE_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	LFPDUX	f18, BO,  INC2	LFPDUX	f19, BO,  INC2	subi	BO,  BO,   8 * SIZE#else	LFPDUX	f16, AO,  INC2	LFPDUX	f17, AO,  INC2	LFPDUX	f18, AO,  INC2	LFPDUX	f19, AO,  INC2	subi	AO,  AO,   8 * SIZE#endif	fpsub	f0,  f16,  f0	fpsub	f1,  f17,  f1	fpsub	f2,  f18,  f2	fpsub	f3,  f19,  f3#ifdef LN	LFPDUX	A1,  AO,  INC2	add	AO,  AO,  INC2	add	AO,  AO,  INC2	add	AO,  AO,  INC2	LFPDUX	A2,  AO,  INC2	LFPDUX	A3,  AO,  INC2	add	AO,  AO,  INC2	add	AO,  AO,  INC2	LFPDUX	A4,  AO,  INC2	LFPDUX	A5,  AO,  INC2	LFPDUX	A6,  AO,  INC2	add	AO,  AO,  INC2	LFPDUX	A7,  AO,  INC2	LFPDUX	A8,  AO,  INC2	LFPDUX	A9,  AO,  INC2	LFPDUX	A10, AO,  INC2	subi	AO,  AO,  32 * SIZE	fxpmul	  f4,  A10, f3	FXCXNPMA  f3,  A10, f3,  f4	fxcpnmsub f2,  A9, f3,  f2	FXCXNSMA  f2,  A9, f3,  f2	fxcpnmsub f1,  A8, f3,  f1	FXCXNSMA  f1,  A8, f3,  f1	fxcpnmsub f0,  A7, f3,  f0	FXCXNSMA  f0,  A7, f3,  f0	fxpmul	  f4,  A6, f2	FXCXNPMA  f2,  A6, f2,  f4	fxcpnmsub f1,  A5, f2,  f1	FXCXNSMA  f1,  A5, f2,  f1	fxcpnmsub f0,  A4, f2,  f0	FXCXNSMA  f0,  A4, f2,  f0	fxpmul	  f4,  A3, f1	FXCXNPMA  f1,  A3, f1,  f4	fxcpnmsub f0,  A2, f1,  f0	FXCXNSMA  f0,  A2, f1,  f0	fxpmul	  f4,  A1, f0	FXCXNPMA  f0,  A1, f0,  f4#endif#ifdef LT
⌨️ 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 + -