📄 scal_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	r6#define INCX	r7	#define INCX2	r4#define X2	r5#define ALPHA	f1#define A1	f0#define A2	f16#define A3	f2#define A4	f3#define A5	f4#define A6	f5#define A7	f6#define A8	f7#define B1	f8#define B2	f9#define B3	f10#define B4	f11#define B5	f12#define B6	f13#define B7	f14#define B8	f15	PROLOGUE	PROFCODE	li	r10, -16	stfpdux	f14, SP, r10	stfpdux	f15, SP, r10	stfpdux	f16, SP, r10		li	r10,   0	stwu	r10,   -4(SP)	stwu	r10,   -4(SP)	stwu	r10,   -4(SP)	stwu	r10,   -4(SP)	lfpdx	A1, SP, r10		# Zero clear	fsmfp	ALPHA, ALPHA	slwi	INCX,  INCX, BASE_SHIFT	add	INCX2, INCX, INCX	cmpwi	cr0, N, 0	ble	LL(999)	cmpwi	cr0, INCX, SIZE	bne	LL(100)	fcmpu	cr7, ALPHA, A1	bne	cr7, LL(50)	sub	X,  X, INCX2	andi.	r0, X, 2 * SIZE - 1	beq	LL(11)	STFDX	A1, X, INCX2	addi	X, X, 1 * SIZE	addi	N, N, -1	cmpwi	cr0, N, 0	ble	LL(999)	.align 4LL(11):	srawi.	r0, N, 4	mtspr	CTR,  r0	beq-	LL(15)	.align 4LL(12):	STFPDUX	A1,   X, INCX2	STFPDUX	A1,   X, INCX2	STFPDUX	A1,   X, INCX2	STFPDUX	A1,   X, INCX2	STFPDUX	A1,   X, INCX2	STFPDUX	A1,   X, INCX2	STFPDUX	A1,   X, INCX2	STFPDUX	A1,   X, INCX2	bdnz	LL(12)	.align 4LL(15):	andi.	r0,  N, 15	beq	LL(999)	andi.	r0,  N, 8	beq	LL(16)	STFPDUX	A1,   X, INCX2	STFPDUX	A1,   X, INCX2	STFPDUX	A1,   X, INCX2	STFPDUX	A1,   X, INCX2	.align 4LL(16):	andi.	r0,  N, 4	beq	LL(17)	STFPDUX	A1,   X, INCX2	STFPDUX	A1,   X, INCX2	.align 4LL(17):	andi.	r0,  N, 2	beq	LL(18)	STFPDUX	A1,   X, INCX2	.align 4LL(18):	andi.	r0,  N, 1	beq	LL(999)	STFDUX	A1,   X, INCX2	b	LL(999)	.align 4LL(50):	sub	X2, X, INCX2	sub	X,  X, INCX2	andi.	r0, X, 2 * SIZE - 1	beq	LL(51)	LFDX	A1, X,  INCX2	addi	X,  X,  1 * SIZE	fmul	B1, ALPHA, A1	addi	N, N, -1	cmpwi	cr0, N, 0	STFDX	B1, X2, INCX2	addi	X2, X2, 1 * SIZE	ble	LL(999)	.align 4LL(51):	srawi.	r0, N, 4	mtspr	CTR,  r0	beq-	LL(55)	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(53)	.align 4LL(52):	fpmul	B1, ALPHA, A1	LFPDUX	A1, X,  INCX2	fpmul	B2, ALPHA, A2	LFPDUX	A2, X,  INCX2	fpmul	B3, ALPHA, A3	LFPDUX	A3, X,  INCX2	fpmul	B4, ALPHA, A4	LFPDUX	A4, X,  INCX2	fpmul	B5, ALPHA, A5	LFPDUX	A5, X,  INCX2	fpmul	B6, ALPHA, A6	LFPDUX	A6, X,  INCX2	fpmul	B7, ALPHA, A7	LFPDUX	A7, X,  INCX2	fpmul	B8, ALPHA, A8	LFPDUX	A8, X,  INCX2	STFPDUX	B1, X2, INCX2	STFPDUX	B2, X2, INCX2	STFPDUX	B3, X2, INCX2	STFPDUX	B4, X2, INCX2	STFPDUX	B5, X2, INCX2	STFPDUX	B6, X2, INCX2	STFPDUX	B7, X2, INCX2	STFPDUX	B8, X2, INCX2	bdnz	LL(52)	.align 4LL(53):	fpmul	B1, ALPHA, A1	fpmul	B2, ALPHA, A2	fpmul	B3, ALPHA, A3	fpmul	B4, ALPHA, A4	fpmul	B5, ALPHA, A5	fpmul	B6, ALPHA, A6	STFPDUX	B1, X2, INCX2	fpmul	B7, ALPHA, A7	STFPDUX	B2, X2, INCX2	fpmul	B8, ALPHA, A8	STFPDUX	B3, X2, INCX2	STFPDUX	B4, X2, INCX2	STFPDUX	B5, X2, INCX2	STFPDUX	B6, X2, INCX2	STFPDUX	B7, X2, INCX2	STFPDUX	B8, X2, INCX2	.align 4LL(55):	andi.	r0,  N, 15	beq	LL(999)	andi.	r0,  N, 8	beq	LL(56)	LFPDUX	A1, X,  INCX2	LFPDUX	A2, X,  INCX2	LFPDUX	A3, X,  INCX2	LFPDUX	A4, X,  INCX2	fpmul	B1, ALPHA, A1	fpmul	B2, ALPHA, A2	fpmul	B3, ALPHA, A3	fpmul	B4, ALPHA, A4	STFPDUX	B1, X2, INCX2	STFPDUX	B2, X2, INCX2	STFPDUX	B3, X2, INCX2	STFPDUX	B4, X2, INCX2	.align 4LL(56):	andi.	r0,  N, 4	beq	LL(57)	LFPDUX	A1, X,  INCX2	LFPDUX	A2, X,  INCX2	fpmul	B1, ALPHA, A1	fpmul	B2, ALPHA, A2	STFPDUX	B1, X2, INCX2	STFPDUX	B2, X2, INCX2	.align 4LL(57):	andi.	r0,  N, 2	beq	LL(58)	LFPDUX	A1, X,  INCX2	fpmul	B1, ALPHA, A1	STFPDUX	B1, X2, INCX2	.align 4LL(58):	andi.	r0,  N, 1	beq	LL(999)	LFDX	A1, X,  INCX2	fmul	B1, ALPHA, A1	STFDX	B1, X2, INCX2	b	LL(999)	.align 4LL(100):	fcmpu	cr7, ALPHA, A1	bne	cr7, LL(200)	sub	X,  X, INCX	srawi.	r0, N, 3	mtspr	CTR,  r0	beq-	LL(115)	.align 4LL(112):	STFDUX	A1,   X, INCX	STFDUX	A1,   X, INCX	STFDUX	A1,   X, INCX	STFDUX	A1,   X, INCX	STFDUX	A1,   X, INCX	STFDUX	A1,   X, INCX	STFDUX	A1,   X, INCX	STFDUX	A1,   X, INCX	bdnz	LL(112)	.align 4LL(115):	andi.	r0,  N, 7	beq	LL(999)	andi.	r0,  N, 4	beq	LL(117)	STFDUX	A1,   X, INCX	STFDUX	A1,   X, INCX	STFDUX	A1,   X, INCX	STFDUX	A1,   X, INCX	.align 4LL(117):	andi.	r0,  N, 2	beq	LL(118)	STFDUX	A1,   X, INCX	STFDUX	A1,   X, INCX	.align 4LL(118):	andi.	r0,  N, 1	beq	LL(999)	STFDUX	A1,   X, INCX	b	LL(999)	.align 4LL(200):	sub	X2, X, INCX	sub	X,  X, INCX	srawi.	r0, N, 3	mtspr	CTR,  r0	beq-	LL(215)	LFDUX	A1,   X, INCX	LFDUX	A2,   X, INCX	LFDUX	A3,   X, INCX	LFDUX	A4,   X, INCX	LFDUX	A5,   X, INCX	LFDUX	A6,   X, INCX	LFDUX	A7,   X, INCX	LFDUX	A8,   X, INCX	bdz	LL(213)	.align 4LL(212):	fmul	B1, ALPHA, A1	LFDUX	A1,   X, INCX	fmul	B2, ALPHA, A2	LFDUX	A2,   X, INCX	fmul	B3, ALPHA, A3	LFDUX	A3,   X, INCX	fmul	B4, ALPHA, A4	LFDUX	A4,   X, INCX	fmul	B5, ALPHA, A5	LFDUX	A5,   X, INCX	fmul	B6, ALPHA, A6	LFDUX	A6,   X, INCX	fmul	B7, ALPHA, A7	LFDUX	A7,   X, INCX	fmul	B8, ALPHA, A8	LFDUX	A8,   X, INCX	STFDUX	B1,   X2, INCX	STFDUX	B2,   X2, INCX	STFDUX	B3,   X2, INCX	STFDUX	B4,   X2, INCX	STFDUX	B5,   X2, INCX	STFDUX	B6,   X2, INCX	STFDUX	B7,   X2, INCX	STFDUX	B8,   X2, INCX	bdnz	LL(212)	.align 4LL(213):	fmul	B1, ALPHA, A1	fmul	B2, ALPHA, A2	fmul	B3, ALPHA, A3	fmul	B4, ALPHA, A4	fmul	B5, ALPHA, A5	fmul	B6, ALPHA, A6	STFDUX	B1,   X2, INCX	fmul	B7, ALPHA, A7	STFDUX	B2,   X2, INCX	fmul	B8, ALPHA, A8	STFDUX	B3,   X2, INCX	STFDUX	B4,   X2, INCX	STFDUX	B5,   X2, INCX	STFDUX	B6,   X2, INCX	STFDUX	B7,   X2, INCX	STFDUX	B8,   X2, INCX	.align 4LL(215):	andi.	r0,  N, 7	beq	LL(999)	andi.	r0,  N, 4	beq	LL(217)	LFDUX	A1,   X, INCX	LFDUX	A2,   X, INCX	LFDUX	A3,   X, INCX	LFDUX	A4,   X, INCX	fmul	B1, ALPHA, A1	fmul	B2, ALPHA, A2	fmul	B3, ALPHA, A3	fmul	B4, ALPHA, A4	STFDUX	B1,   X2, INCX	STFDUX	B2,   X2, INCX	STFDUX	B3,   X2, INCX	STFDUX	B4,   X2, INCX	.align 4LL(217):	andi.	r0,  N, 2	beq	LL(218)	LFDUX	A1,   X, INCX	LFDUX	A2,   X, INCX	fmul	B1, ALPHA, A1	fmul	B2, ALPHA, A2	STFDUX	B1,   X2, INCX	STFDUX	B2,   X2, INCX	.align 4LL(218):	andi.	r0,  N, 1	beq	LL(999)	LFDUX	A1,   X, INCX	fmul	B1, ALPHA, A1	STFDUX	B1,   X2, INCX	.align 4LL(999):	li	r10, 16	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 + -