ztrsm_kernel_hummer_rt.s

Optimized GotoBLAS libraries

	FXCSMADD	f4,  B1, A1, f4
	LFPDUX	A8,  AO2, INC4
	FXCPMADD	f8,  B2, A1, f8
	nop
	FXCSMADD	f12, B2, A1, f12
	LFPDUX	A1,   AO, INC4

	FXCPMADD	f1,  B1, A2, f1
	nop
	FXCSMADD	f5,  B1, A2, f5
	LFPDUX	B1,   BO, INC4
	FXCPMADD	f9,  B2, A2, f9
	nop
	FXCSMADD	f13, B2, A2, f13
	LFPDUX	B2,  BO2, INC4

	FXCPMADD	f0,  B3, A3, f0
	nop
	FXCSMADD	f4,  B3, A3, f4
	LFPDUX	A2,  AO2, INC4
	FXCPMADD	f8,  B4, A3, f8
	nop
	FXCSMADD	f12, B4, A3, f12
	LFPDUX	A3,   AO, INC4

	FXCPMADD	f1,  B3, A4, f1
	nop
	FXCSMADD	f5,  B3, A4, f5
	LFPDUX	B3,   BO, INC4
	FXCPMADD	f9,  B4, A4, f9
	nop
	FXCSMADD	f13, B4, A4, f13
	LFPDUX	B4,  BO2, INC4

	FXCPMADD	f0,  B5, A5, f0
	nop
	FXCSMADD	f4,  B5, A5, f4
	LFPDUX	A4,  AO2, INC4
	FXCPMADD	f8,  B6, A5, f8
	nop
	FXCSMADD	f12, B6, A5, f12
	LFPDUX	A5,   AO, INC4

	FXCPMADD	f1,  B5, A6, f1
	nop
	FXCSMADD	f5,  B5, A6, f5
	LFPDUX	B5,   BO, INC4
	FXCPMADD	f9,  B6, A6, f9
	nop
	FXCSMADD	f13, B6, A6, f13
	LFPDUX	B6,  BO2, INC4

	FXCPMADD	f0,  A9,  A7, f0
	nop
	FXCSMADD	f4,  A9,  A7, f4
	LFPDUX	A6,  AO2, INC4
	FXCPMADD	f8,  A10, A7, f8
	nop
	FXCSMADD	f12, A10, A7, f12
	LFPDUX	A7,   AO, INC4

	FXCPMADD	f1,  A9,  A8, f1
	nop
	FXCSMADD	f5,  A9,  A8, f5
	LFPDUX	A9,   BO, INC4
	FXCPMADD	f9,  A10, A8, f9
	nop
	FXCSMADD	f13, A10, A8, f13
	LFPDUX	A10, BO2, INC4
	bdnz+	.L22
	.align 4

.L23:
	FXCPMADD	f0,  B1, A1, f0
	FXCSMADD	f4,  B1, A1, f4
	LFPDUX	A8,  AO2, INC4
	FXCPMADD	f8,  B2, A1, f8
	FXCSMADD	f12, B2, A1, f12

	FXCPMADD	f1,  B1, A2, f1
	FXCSMADD	f5,  B1, A2, f5
	FXCPMADD	f9,  B2, A2, f9
	FXCSMADD	f13, B2, A2, f13

	FXCPMADD	f0,  B3, A3, f0
	FXCSMADD	f4,  B3, A3, f4
	FXCPMADD	f8,  B4, A3, f8
	FXCSMADD	f12, B4, A3, f12

	FXCPMADD	f1,  B3, A4, f1
	FXCSMADD	f5,  B3, A4, f5
	FXCPMADD	f9,  B4, A4, f9
	FXCSMADD	f13, B4, A4, f13

	FXCPMADD	f0,  B5, A5, f0
	FXCSMADD	f4,  B5, A5, f4
	FXCPMADD	f8,  B6, A5, f8
	FXCSMADD	f12, B6, A5, f12

	FXCPMADD	f1,  B5, A6, f1
	FXCSMADD	f5,  B5, A6, f5
	FXCPMADD	f9,  B6, A6, f9
	FXCSMADD	f13, B6, A6, f13

	FXCPMADD	f0,  A9, A7, f0
	FXCSMADD	f4,  A9, A7, f4
	FXCPMADD	f8,  A10, A7, f8
	FXCSMADD	f12, A10, A7, f12

	FXCPMADD	f1,  A9, A8, f1
	FXCSMADD	f5,  A9, A8, f5
	FXCPMADD	f9,  A10, A8, f9
	FXCSMADD	f13, A10, A8, f13
	.align 4

.L24:
#if defined(LT) || defined(RN)
	andi.	r0,  KK,  3
	mtspr	CTR, r0
	ble+	.L28
#else
	andi.	r0, TEMP, 3
	mtspr	CTR, r0
	ble+	.L28
#endif

	LFPDUX	A1,  AO,  INC4
	LFPDUX	A2,  AO2, INC4
	LFPDUX	B1,  BO,  INC4
	LFPDUX	B2,  BO2, INC4
	bdz-	.L27
	.align 4

.L26:
	FXCPMADD	f0,  B1, A1, f0
	FXCSMADD	f4,  B1, A1, f4
	FXCPMADD	f8,  B2, A1, f8
	FXCSMADD	f12, B2, A1, f12
	LFPDUX	A1,  AO,  INC4

	FXCPMADD	f1,  B1, A2, f1
	FXCSMADD	f5,  B1, A2, f5
	LFPDUX	B1,  BO,  INC4
	FXCPMADD	f9,  B2, A2, f9
	FXCSMADD	f13, B2, A2, f13
	LFPDUX	A2,  AO2, INC4
	LFPDUX	B2,  BO2, INC4
	bdnz+	.L26
	.align 4

.L27:
	FXCPMADD	f0,  B1, A1, f0
	FXCSMADD	f4,  B1, A1, f4
	FXCPMADD	f8,  B2, A1, f8
	FXCSMADD	f12, B2, A1, f12

	FXCPMADD	f1,  B1, A2, f1
	FXCSMADD	f5,  B1, A2, f5
	FXCPMADD	f9,  B2, A2, f9
	FXCSMADD	f13, B2, A2, f13
	.align 4

.L28:
	fpadd	f0, f0, f4
	fpadd	f8, f8, f12
	fpadd	f1, f1, f5
	fpadd	f9, f9, f13

#if defined(LN) || defined(RT)
#ifdef LN
	subi	r0, KK, 2
#else
	subi	r0, KK, 2
#endif
	slwi	r0,   r0, 1 + ZBASE_SHIFT
	add	AO, AORIG, r0
	add	BO, B,     r0
	addi	AO2, AO,   2 * SIZE
	addi	BO,  BO, - 4 * SIZE
	addi	BO2, BO,   2 * SIZE
#endif

#if defined(LN) || defined(LT)
	LFPDUX	f16, BO,  INC4
	LFPDUX	f18, BO2, INC4
	LFPDUX	f17, BO,  INC4
	LFPDUX	f19, BO2, INC4

	subi	BO,  BO,   8 * SIZE
	subi	BO2, BO2,  8 * SIZE
#else
	LFPDUX	f16, AO,  INC4
	LFPDUX	f17, AO2, INC4
	LFPDUX	f18, AO,  INC4
	LFPDUX	f19, AO2, INC4

	subi	AO,  AO,   8 * SIZE
	subi	AO2, AO2,  8 * SIZE
#endif

	fpsub	f0,  f16,  f0
	fpsub	f1,  f17,  f1
	fpsub	f8,  f18,  f8
	fpsub	f9,  f19,  f9

#ifdef LN
	LFPDUX	A1,  AO,  INC4
	add	AO2, AO2, INC4
	LFPDUX	A2,  AO,  INC4
	LFPDUX	A3,  AO2, INC4

	subi	AO,  AO,   8 * SIZE
	subi	AO2, AO2,  8 * SIZE

	fxpmul	  f4,  A3, f1
	fxpmul	  f5,  A3, f9
	FXCXNPMA  f1,  A3, f1,  f4
	FXCXNPMA  f9,  A3, f9,  f5

	fxcpnmsub f0,  A2, f1,  f0
	fxcpnmsub f8,  A2, f9,  f8
	FXCXNSMA  f0,  A2, f1,  f0
	FXCXNSMA  f8,  A2, f9,  f8

	fxpmul	  f4,  A1, f0
	fxpmul	  f5,  A1, f8
	FXCXNPMA  f0,  A1, f0,  f4
	FXCXNPMA  f8,  A1, f8,  f5
#endif

#ifdef LT
	LFPDUX	A1,  AO,  INC4
	LFPDUX	A2,  AO2, INC4
	add	AO,  AO,  INC4
	LFPDUX	A3,  AO2, INC4

	subi	AO,  AO,   8 * SIZE
	subi	AO2, AO2,  8 * SIZE

	fxpmul	  f4,  A1, f0
	fxpmul	  f5,  A1, f8
	FXCXNPMA  f0,  A1, f0, f4
	FXCXNPMA  f8,  A1, f8, f5

	fxcpnmsub f1,  A2, f0, f1
	fxcpnmsub f9,  A2, f8, f9
	FXCXNSMA  f1,  A2, f0, f1
	FXCXNSMA  f9,  A2, f8, f9

	fxpmul	  f6,  A3, f1
	fxpmul	  f7,  A3, f9
	FXCXNPMA  f1,  A3, f1, f6
	FXCXNPMA  f9,  A3, f9, f7
#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
	fxpmul	  f5,  A1, f1

	FXCXNPMA  f0,  A1, f0, f4
	FXCXNPMA  f1,  A1, f1, f5

	fxcpnmsub f8,  A2, f0, f8
	fxcpnmsub f9,  A2, f1, f9

	FXCXNSMA  f8,  A2, f0, f8
	FXCXNSMA  f9,  A2, f1, f9

	fxpmul	  f4,  A3, f8
	fxpmul	  f5,  A3, f9

	FXCXNPMA  f8,  A3, f8,  f4
	FXCXNPMA  f9,  A3, f9,  f5
#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, f8
	fxpmul	  f5,  A3, f9

	FXCXNPMA  f8,  A3, f8,  f4
	FXCXNPMA  f9,  A3, f9,  f5

	fxcpnmsub f0,  A2, f8,  f0
	fxcpnmsub f1,  A2, f9,  f1

	FXCXNSMA  f0,  A2, f8,  f0
	FXCXNSMA  f1,  A2, f9,  f1

	fxpmul	  f4,  A1, f0
	fxpmul	  f5,  A1, f1

	FXCXNPMA  f0,  A1, f0,  f4
	FXCXNPMA  f1,  A1, f1,  f5
#endif

#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

.L999:
	addi	SP, SP, 20

	lwzu	r14,   4(SP)
	lwzu	r15,   4(SP)

	lwzu	r16,   4(SP)
	lwzu	r17,   4(SP)
	lwzu	r18,   4(SP)
	lwzu	r19,   4(SP)

	lwzu	r20,   4(SP)
	lwzu	r21,   4(SP)
	lwzu	r22,   4(SP)
	lwzu	r23,   4(SP)

	lwzu	r24,   4(SP)
	lwzu	r25,   4(SP)
	lwzu	r26,   4(SP)
	lwzu	r27,   4(SP)

	lwzu	r28,   4(SP)
	lwzu	r29,   4(SP)
	lwzu	r30,   4(SP)
	lwzu	r31,   4(SP)

	subi	SP, SP, 12
	li	r0, 16

	lfpdux	f31, SP, r0
	lfpdux	f30, SP, r0
	lfpdux	f29, SP, r0
	lfpdux	f28, SP, r0
	lfpdux	f27, SP, r0
	lfpdux	f26, SP, r0
	lfpdux	f25, SP, r0
	lfpdux	f24, SP, r0
	lfpdux	f23, SP, r0
	lfpdux	f22, SP, r0
	lfpdux	f21, SP, r0
	lfpdux	f20, SP, r0
	lfpdux	f19, SP, r0
	lfpdux	f18, SP, r0
	lfpdux	f17, SP, r0
	lfpdux	f16, SP, r0
	lfpdux	f15, SP, r0
	lfpdux	f14, SP, r0
	addi	SP, SP, 16
	blr
	.align 4


	EPILOGUE
#endif