📄 imax_hummer.s

📁 Optimized GotoBLAS libraries
💻 S
字号:
/*********************************************************************//*                                                                   *//*             Optimized BLAS libraries                              *//*                     By Kazushige Goto <kgoto@tacc.utexas.edu>     *//*                                                                   *//* 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"		#define N	r3#define X	r4#define INCX	r5	#define INCX2	r6#define X2	r7#define	XX	r8#define RET	r9#define NN	r10#define C1	f1#define C2	f0#define C3	f2#define C4	f3#define A1	f4#define A2	f5#define A3	f6#define A4	f7#define A5	f8#define A6	f9#define A7	f10#define A8	f11#define F1	f12#define F2	f13#define F3	f14#define F4	f15#define F5	f16#define F6	f17#define F7	f18#define F8	f19	PROLOGUE	PROFCODE	li	r10, -16	stfpdux	f14, SP, r10	stfpdux	f15, SP, r10		stfpdux	f16, SP, r10	stfpdux	f17, SP, r10	stfpdux	f18, SP, r10	stfpdux	f19, SP, r10#ifdef F_INTERFACE	LDINT	N,    0(N)	LDINT	INCX, 0(INCX)#endif	slwi	INCX,  INCX, BASE_SHIFT	add	INCX2, INCX, INCX	li	RET, 0	cmpwi	cr0, N, 0	ble	LL(999)	mr	NN, N	cmpwi	cr0, INCX, 0	ble	LL(999)	LFD	C1, 0 * SIZE(X)	addi	N, N, -1	cmpwi	cr0, N, 0	li	RET, 1	ble	LL(999)	fsmfp	C1, C1	mr	XX, X	fpmr	C2, C1	add	X, X, INCX	fpmr	C3, C1	fpmr	C4, C1	cmpwi	cr0, INCX, SIZE	bne	LL(100)	andi.	r0, X, 2 * SIZE - 1	beq	LL(05)	LFD	C2, 0 * SIZE(X)	add	X, X, INCX	addi	N, N, -1	cmpwi	cr0, N, 0	ble	LL(20)	.align 4LL(05):	sub	X, X, INCX2	srawi.	r0, N, 4	mtspr	CTR,  r0	beq-	LL(15)	LFPDUX	A1,    X, INCX2	LFPDUX	A2,    X, INCX2	LFPDUX	A3,    X, INCX2	LFPDUX	A4,    X, INCX2	LFPDUX	A5,    X, INCX2	LFPDUX	A6,    X, INCX2	LFPDUX	A7,    X, INCX2	LFPDUX	A8,    X, INCX2	bdz	LL(13)	.align 4LL(12):	fpsub	F1, C1, A1	fpsub	F2, C2, A2	fpsub	F3, C3, A3	fpsub	F4, C4, A4	fpsel	C1, F1, C1, A1	LFPDUX	A1,    X, INCX2	fpsel	C2, F2, C2, A2	LFPDUX	A2,    X, INCX2	fpsel	C3, F3, C3, A3	LFPDUX	A3,    X, INCX2	fpsel	C4, F4, C4, A4	LFPDUX	A4,    X, INCX2	fpsub	F5, C1, A5	fpsub	F6, C2, A6	fpsub	F7, C3, A7	fpsub	F8, C4, A8	fpsel	C1, F5, C1, A5	LFPDUX	A5,    X, INCX2	fpsel	C2, F6, C2, A6	LFPDUX	A6,    X, INCX2	fpsel	C3, F7, C3, A7	LFPDUX	A7,    X, INCX2	fpsel	C4, F8, C4, A8	LFPDUX	A8,    X, INCX2	bdnz	LL(12)	.align 4LL(13):	fpsub	F1, C1, A1	fpsub	F2, C2, A2	fpsub	F3, C3, A3	fpsub	F4, C4, A4	fpsel	C1, F1, C1, A1	fpsel	C2, F2, C2, A2	fpsel	C3, F3, C3, A3	fpsel	C4, F4, C4, A4	fpsub	F5, C1, A5	fpsub	F6, C2, A6	fpsub	F7, C3, A7	fpsub	F8, C4, A8	fpsel	C1, F5, C1, A5	fpsel	C2, F6, C2, A6	fpsel	C3, F7, C3, A7	fpsel	C4, F8, C4, A8	.align 4LL(15):	andi.	r0,  N, 15	beq	LL(20)	andi.	r0,  N, 8	beq	LL(16)	LFPDUX	A1,    X, INCX2	LFPDUX	A2,    X, INCX2	LFPDUX	A3,    X, INCX2	LFPDUX	A4,    X, INCX2	fpsub	F1, C1, A1	fpsub	F2, C2, A2	fpsub	F3, C3, A3	fpsub	F4, C4, A4	fpsel	C1, F1, C1, A1	fpsel	C2, F2, C2, A2	fpsel	C3, F3, C3, A3	fpsel	C4, F4, C4, A4	.align 4LL(16):	andi.	r0,  N, 4	beq	LL(17)	LFPDUX	A1,    X, INCX2	LFPDUX	A2,    X, INCX2	fpsub	F1, C1, A1	fpsub	F2, C2, A2	fpsel	C1, F1, C1, A1	fpsel	C2, F2, C2, A2	.align 4LL(17):	andi.	r0,  N, 2	beq	LL(18)	LFPDUX	A1,    X, INCX2	fpsub	F1, C1, A1	fpsel	C1, F1, C1, A1	.align 4LL(18):	andi.	r0,  N, 1	beq	LL(20)	LFDUX	A1,    X, INCX2	fsub	F1, C1, A1	fsel	C1, F1, C1, A1	.align 4LL(20):	fpsub	F1,  C1,  C2	fpsub	F2,  C3,  C4	fpsel	C1,  F1,  C1,  C2	fpsel	C3,  F2,  C3,  C4	fpsub	F1,  C1,  C3	fpsel	C1,  F1,  C1,  C3	fsmtp	C2, C1	fsub	F1,  C1,  C2	fsel	C1,  F1,  C1,  C2 	li	RET, 0	fsmfp	C1, C1	andi.	r0, XX, 2 * SIZE - 1	beq	LL(21)	LFD	A1, 0 * SIZE(XX)	add	XX, XX, INCX	addi	NN, NN, -1 	addi	RET, RET, 1	fcmpu	cr0, C1, A1	beq	cr0, LL(999)	.align 4LL(21):	sub	XX, XX, INCX2	srawi.	r0, NN, 4	mtspr	CTR,  r0	beq-	LL(25)	LFPDUX	A1,    XX, INCX2	LFPDUX	A2,    XX, INCX2	LFPDUX	A3,    XX, INCX2	LFPDUX	A4,    XX, INCX2	LFPDUX	A5,    XX, INCX2	LFPDUX	A6,    XX, INCX2	LFPDUX	A7,    XX, INCX2	LFPDUX	A8,    XX, INCX2	bdz	LL(23)	.align 4LL(22):	addi	RET, RET, 1	fcmpu	cr0, C1, A1	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A1	LFPDUX	A1,    XX, INCX2	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A2	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A2	LFPDUX	A2,    XX, INCX2	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A3	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A3	LFPDUX	A3,    XX, INCX2	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A4	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A4	LFPDUX	A4,    XX, INCX2	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A5	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A5	LFPDUX	A5,    XX, INCX2	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A6	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A6	LFPDUX	A6,    XX, INCX2	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A7	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A7	LFPDUX	A7,    XX, INCX2	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A8	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A8	LFPDUX	A8,    XX, INCX2	beq	cr0, LL(999)	bdnz	LL(22)	.align 4LL(23):	addi	RET, RET, 1	fcmpu	cr0, C1, A1	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A1	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A2	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A2	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A3	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A3	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A4	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A4	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A5	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A5	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A6	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A6	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A7	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A7	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A8	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A8	beq	cr0, LL(999)	.align 4LL(25):	andi.	r0,  NN, 8	beq	LL(26)	LFPDUX	A1,    XX, INCX2	LFPDUX	A2,    XX, INCX2	LFPDUX	A3,    XX, INCX2	LFPDUX	A4,    XX, INCX2	addi	RET, RET, 1	fcmpu	cr0, C1, A1	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A1	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A2	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A2	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A3	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A3	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A4	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A4	beq	cr0, LL(999)	.align 4LL(26):	andi.	r0,  NN, 4	beq	LL(27)	LFPDUX	A1,    XX, INCX2	LFPDUX	A2,    XX, INCX2	addi	RET, RET, 1	fcmpu	cr0, C1, A1	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A1	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A2	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A2	beq	cr0, LL(999)	.align 4LL(27):	andi.	r0,  NN, 2	beq	LL(28)	LFPDUX	A1,    XX, INCX2	addi	RET, RET, 1	fcmpu	cr0, C1, A1	beq	cr0, LL(999)	addi	RET, RET, 1	fscmp	cr0, C1, A1	beq	cr0, LL(999)	.align 4LL(28):	addi	RET, RET, 1	b	LL(999)	.align 4LL(100):	sub	X, X, INCX	srawi.	r0, N, 4	mtspr	CTR,  r0	beq-	LL(105)	LFDUX	A1,    X, INCX	LFDUX	A2,    X, INCX	LFDUX	A3,    X, INCX	LFDUX	A4,    X, INCX	LFSDUX	A1,    X, INCX	LFSDUX	A2,    X, INCX	LFSDUX	A3,    X, INCX	LFSDUX	A4,    X, INCX	LFDUX	A5,    X, INCX	LFDUX	A6,    X, INCX	LFDUX	A7,    X, INCX	LFDUX	A8,    X, INCX	bdz	LL(103)	.align 4LL(102):	fpsub	F1, C1, A1	LFSDUX	A5,    X, INCX	fpsub	F2, C2, A2	LFSDUX	A6,    X, INCX	fpsub	F3, C3, A3	LFSDUX	A7,    X, INCX	fpsub	F4, C4, A4	LFSDUX	A8,    X, INCX	fpsel	C1, F1, C1, A1	LFDUX	A1,    X, INCX	fpsel	C2, F2, C2, A2	LFDUX	A2,    X, INCX	fpsel	C3, F3, C3, A3	LFDUX	A3,    X, INCX	fpsel	C4, F4, C4, A4	LFDUX	A4,    X, INCX	fpsub	F5, C1, A5	LFSDUX	A1,    X, INCX	fpsub	F6, C2, A6	LFSDUX	A2,    X, INCX	fpsub	F7, C3, A7	LFSDUX	A3,    X, INCX	fpsub	F8, C4, A8	LFSDUX	A4,    X, INCX	fpsel	C1, F5, C1, A5	LFDUX	A5,    X, INCX	fpsel	C2, F6, C2, A6	LFDUX	A6,    X, INCX	fpsel	C3, F7, C3, A7	LFDUX	A7,    X, INCX	fpsel	C4, F8, C4, A8	LFDUX	A8,    X, INCX	bdnz	LL(102)	.align 4LL(103):	fpsub	F1, C1, A1	LFSDUX	A5,    X, INCX	fpsub	F2, C2, A2	LFSDUX	A6,    X, INCX	fpsub	F3, C3, A3	LFSDUX	A7,    X, INCX	fpsub	F4, C4, A4	LFSDUX	A8,    X, INCX	fpsel	C1, F1, C1, A1	fpsel	C2, F2, C2, A2	fpsel	C3, F3, C3, A3	fpsel	C4, F4, C4, A4	fpsub	F5, C1, A5	fpsub	F6, C2, A6	fpsub	F7, C3, A7	fpsub	F8, C4, A8	fpsel	C1, F5, C1, A5	fpsel	C2, F6, C2, A6	fpsel	C3, F7, C3, A7	fpsel	C4, F8, C4, A8	.align 4LL(105):	andi.	r0,  N, 15	beq	LL(120)	andi.	r0,  N, 8	beq	LL(106)	LFDUX	A1,    X, INCX	LFDUX	A2,    X, INCX	LFDUX	A3,    X, INCX	LFDUX	A4,    X, INCX	LFSDUX	A1,    X, INCX	LFSDUX	A2,    X, INCX	LFSDUX	A3,    X, INCX	LFSDUX	A4,    X, INCX	fpsub	F1, C1, A1	fpsub	F2, C2, A2	fpsub	F3, C3, A3	fpsub	F4, C4, A4	fpsel	C1, F1, C1, A1	fpsel	C2, F2, C2, A2	fpsel	C3, F3, C3, A3	fpsel	C4, F4, C4, A4	.align 4LL(106):	andi.	r0,  N, 4	beq	LL(107)	LFDUX	A1,    X, INCX	LFDUX	A2,    X, INCX	LFDUX	A3,    X, INCX	LFDUX	A4,    X, INCX	fsub	F1, C1, A1	fsub	F2, C2, A2	fsub	F3, C3, A3	fsub	F4, C4, A4	fsel	C1, F1, C1, A1	fsel	C2, F2, C2, A2	fsel	C3, F3, C3, A3	fsel	C4, F4, C4, A4	.align 4LL(107):	andi.	r0,  N, 2	beq	LL(108)	LFDUX	A1,    X, INCX	LFDUX	A2,    X, INCX	fsub	F1, C1, A1	fsub	F2, C2, A2	fsel	C1, F1, C1, A1	fsel	C2, F2, C2, A2	.align 4LL(108):	andi.	r0,  N, 1	beq	LL(120)	LFDUX	A1,    X, INCX	fsub	F1, C1, A1	fsel	C1, F1, C1, A1	.align 4LL(120):	fpsub	F1,  C1,  C2	fpsub	F2,  C3,  C4	fpsel	C1,  F1,  C1,  C2	fpsel	C3,  F2,  C3,  C4	fpsub	F1,  C1,  C3	fpsel	C1,  F1,  C1,  C3	fsmtp	C2, C1	fsub	F1,  C1,  C2	fsel	C1,  F1,  C1,  C2 	li	RET, 0	sub	XX, XX, INCX	srawi.	r0, NN, 3	mtspr	CTR,  r0	beq-	LL(126)	LFDUX	A1,    XX, INCX	LFDUX	A2,    XX, INCX	LFDUX	A3,    XX, INCX	LFDUX	A4,    XX, INCX	LFDUX	A5,    XX, INCX	LFDUX	A6,    XX, INCX	LFDUX	A7,    XX, INCX	LFDUX	A8,    XX, INCX	bdz	LL(123)	.align 4LL(122):	addi	RET, RET, 1	fcmpu	cr0, C1, A1	LFDUX	A1,    XX, INCX	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A2	LFDUX	A2,    XX, INCX	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A3	LFDUX	A3,    XX, INCX	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A4	LFDUX	A4,    XX, INCX	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A5	LFDUX	A5,    XX, INCX	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A6	LFDUX	A6,    XX, INCX	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A7	LFDUX	A7,    XX, INCX	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A8	LFDUX	A8,    XX, INCX	beq	cr0, LL(999)	bdnz	LL(122)	.align 4LL(123):	addi	RET, RET, 1	fcmpu	cr0, C1, A1	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A2	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A3	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A4	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A5	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A6	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A7	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A8	beq	cr0, LL(999)	.align 4LL(126):	andi.	r0,  NN, 4	beq	LL(127)	LFDUX	A1,    XX, INCX	LFDUX	A2,    XX, INCX	LFDUX	A3,    XX, INCX	LFDUX	A4,    XX, INCX	addi	RET, RET, 1	fcmpu	cr0, C1, A1	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A2	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A3	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A4	beq	cr0, LL(999)	.align 4LL(127):	andi.	r0,  NN, 2	beq	LL(128)	LFDUX	A1,    XX, INCX	LFDUX	A2,    XX, INCX	addi	RET, RET, 1	fcmpu	cr0, C1, A1	beq	cr0, LL(999)	addi	RET, RET, 1	fcmpu	cr0, C1, A2	beq	cr0, LL(999)	.align 4LL(128):	addi	RET, RET, 1	.align 4LL(999):	li	r10, 16	addi	SP, SP, -16	mr	r3, RET	lfpdux	f19, SP, r10	lfpdux	f18, SP, r10	lfpdux	f17, SP, r10	lfpdux	f16, SP, r10	lfpdux	f15, SP, r10	lfpdux	f14, SP, r10	addi	SP, SP,  16	blr	EPILOGUE
⌨️ 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 + -