ztrsm_kernel_hummer_ln.s

Optimized GotoBLAS libraries

/*********************************************************************/
/*                                                                   */
/*             Optimized BLAS libraries                              */
/*                     By Kazushige Goto <[e-mail]>     */
/*                                                                   */
/* Copyright (c) The University of Texas, 2005. All rights reserved. */
/* UNIVERSITY EXPRESSLY DISCLAIMS ANY AND ALL WARRANTIES CONCERNING  */
/* THIS SOFTWARE AND DOCUMENTATION, INCLUDING ANY WARRANTIES OF      */
/* MERCHANTABILITY, FITNESS FOR ANY PARTICULAR PURPOSE,              */
/* NON-INFRINGEMENT AND WARRANTIES OF PERFORMANCE, AND ANY WARRANTY  */
/* THAT MIGHT OTHERWISE ARISE FROM COURSE OF DEALING OR USAGE OF     */
/* TRADE. NO WARRANTY IS EITHER EXPRESS OR IMPLIED WITH RESPECT TO   */
/* THE USE OF THE SOFTWARE OR DOCUMENTATION.                         */
/* Under no circumstances shall University be liable for incidental, */
/* special, indirect, direct or consequential damages or loss of     */
/* profits, interruption of business, or related expenses which may  */
/* arise from use of Software or Documentation, including but not    */
/* limited to those resulting from defects in Software and/or        */
/* Documentation, or loss or inaccuracy of data of any kind.         */
/*********************************************************************/

#define ASSEMBLER
#include "common.h"
		
#undef ZERO

#define ALPHA    0
#define FZERO	16

#define	M	r3
#define	N	r4
#define	K	r5

#ifdef linux
#define A	r6
#define	B	r7
#define	C	r8
#define	LDC	r9
#define OFFSET	r10
#endif

#define TEMP	r11
#define AORIG	r12
#define KK	r14
#define INCM1	r15
#define INCM3	r16
#define INCM5	r17
#define INCM7	r18
#define INC2	r19
#define INC	r20
#define INC4	r21

#define	I	r22
#define J	r23
#define AO	r24
#define BO	r25
#define AO2	r26
#define	BO2	r27
	
#define	CO1	r28
#define CO2	r29
#define	ZERO	r31

#ifndef NEEDPARAM

#define A1	f16
#define A2	f17
#define A3	f18
#define A4	f19
#define A5	f20
#define A6	f21
#define A7	f22
#define A8	f23
#define A9	f24
#define A10	f25

#define B1	f26
#define B2	f27
#define B3	f28
#define B4	f29
#define B5	f30
#define B6	f31

#define AP	B6

#ifndef CONJ
#define FXCPMADD	fxcpmadd
#define FXCSMADD	fxcxnpma
#else
#if defined(LN) || defined(LT)
#define FXCPMADD	fxcpnsma
#define FXCSMADD	fxcxma
#else
#define FXCPMADD	fxcpmadd
#define FXCSMADD	fxcxnsma
#endif
#endif

#ifndef CONJ
#define FXCXNPMA	fxcxnpma
#define FXCXNSMA	fxcxnsma
#else
#define FXCXNPMA	fxcxnsma
#define FXCXNSMA	fxcxnpma
#endif


	PROLOGUE
	PROFCODE

	li	r0, -16

	stfpdux	f14, SP, r0
	stfpdux	f15, SP, r0
	stfpdux	f16, SP, r0
	stfpdux	f17, SP, r0
	stfpdux	f18, SP, r0
	stfpdux	f19, SP, r0
	stfpdux	f20, SP, r0
	stfpdux	f21, SP, r0
	stfpdux	f22, SP, r0
	stfpdux	f23, SP, r0
	stfpdux	f24, SP, r0
	stfpdux	f25, SP, r0
	stfpdux	f26, SP, r0
	stfpdux	f27, SP, r0
	stfpdux	f28, SP, r0
	stfpdux	f29, SP, r0
	stfpdux	f30, SP, r0
	stfpdux	f31, SP, r0
	
	stwu	r31,  -4(SP)
	stwu	r30,  -4(SP)
	stwu	r29,  -4(SP)
	stwu	r28,  -4(SP)

	stwu	r27,  -4(SP)
	stwu	r26,  -4(SP)
	stwu	r25,  -4(SP)
	stwu	r24,  -4(SP)

	stwu	r23,  -4(SP)
	stwu	r22,  -4(SP)
	stwu	r21,  -4(SP)
	stwu	r20,  -4(SP)

	stwu	r19,  -4(SP)
	stwu	r18,  -4(SP)
	stwu	r17,  -4(SP)
	stwu	r16,  -4(SP)

	stwu	r15,  -4(SP)
	stwu	r14,  -4(SP)

	li	r0,   0
	stwu	r0,   -4(SP)
	stwu	r0,   -4(SP)

	stfdu	f2,   -8(SP)
	stfdu	f1,   -8(SP)

	slwi	LDC, LDC, ZBASE_SHIFT

	cmpwi	cr0, M, 0
	ble	.L999
	cmpwi	cr0, N, 0
	ble	.L999
	cmpwi	cr0, K, 0
	ble	.L999

	li	INC,    1 * SIZE
	li	INC2,   2 * SIZE
	li	INC4,   4 * SIZE
	li	INCM1, -1 * SIZE
	li	INCM3, -3 * SIZE
	li	INCM5, -5 * SIZE
	li	INCM7, -7 * SIZE

	addi	C, C, - 1 * SIZE
	
#ifdef LN
	mullw	r0, M, K
	slwi	r0, r0, ZBASE_SHIFT
	add	A, A, r0

	slwi	r0, M, ZBASE_SHIFT
	add	C, C, r0
#endif

#ifdef RN
	neg	KK, OFFSET
#endif

#ifdef RT
	mullw	r0, N, K
	slwi	r0, r0, ZBASE_SHIFT
	add	B, B, r0

	mullw	r0, N, LDC
	add	C, C, r0

	sub	KK, N, OFFSET
#endif

	srawi.	J, N,  1
	ble	.L50
	.align 4

.L10:
#ifdef RT
	slwi	r0, K, 1 + ZBASE_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, -4 * SIZE
#else
	addi	AO, A, -4 * SIZE
#endif
#ifndef RT
	add	C,  CO2, LDC
#endif

	li	r0, FZERO
	lfpsx	f0, SP, r0

	
	andi.	I, M,  1
	beq	.L20

#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
	
.L20:
	andi.	I, M,  2
	beq	.L30

#if defined(LT) || defined(RN)
	addi	AO2, AO,   2 * SIZE
	fpmr	f4,  f0
	addi	BO,  B,  - 4 * SIZE
	fpmr	f8,  f0
	addi	BO2, B,  - 2 * SIZE
	fpmr	f12, f0

	srawi.	r0,  KK,  2
 	fpmr	f1,  f0
	fpmr	f5,  f0
	fpmr	f9,  f0
	mtspr	CTR, r0
	fpmr	f13, f0
	ble	.L24
#else
#ifdef LN
	slwi	r0,   K,  1 + ZBASE_SHIFT
	sub	AORIG, AORIG, r0
#endif

	slwi	r0  , KK, 1 + ZBASE_SHIFT
	add	AO, AORIG, r0
	add	BO, B,     r0

	sub	TEMP, K, KK

	addi	AO2, AO,   2 * SIZE
	fpmr	f4,  f0
	addi	BO,  BO,  - 4 * SIZE
	fpmr	f8,  f0
	addi	BO2, BO,    2 * SIZE
	fpmr	f12, f0

 	fpmr	f1,  f0
	fpmr	f5,  f0
	fpmr	f9,  f0
	fpmr	f13, f0
	srawi.	r0,  TEMP,  2
	mtspr	CTR, r0
	ble	.L24
#endif

	LFPDUX	A1,   AO, INC4
	LFPDUX	B1,   BO, INC4
	LFPDUX	A2,  AO2, INC4
	LFPDUX	B2,  BO2, INC4
	LFPDUX	A3,   AO, INC4
	LFPDUX	B3,   BO, INC4
	LFPDUX	A4,  AO2, INC4
	LFPDUX	B4,  BO2, INC4

	LFPDUX	A5,   AO, INC4
	LFPDUX	B5,   BO, INC4
	LFPDUX	A6,  AO2, INC4
	LFPDUX	B6,  BO2, INC4
	LFPDUX	A7,   AO, INC4
	LFPDUX	A9,   BO, INC4
	LFPDUX	A10, BO2, INC4
	bdz-	.L23
	.align 4

.L22:
	FXCPMADD	f0,  B1, A1, f0
	nop
	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