⭐ 欢迎来到虫虫下载站! | 📦 资源下载 📁 资源专辑 ℹ️ 关于我们
⭐ 虫虫下载站

📄 zswap_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 Y	r8#define INCY	r9	#define INCX2	r4#define INCY2	r5#define X2	r10#define Y2	r11#define A1	f0#define A2	f1#define A3	f2#define A4	f3#define A5	f4#define B1	f5#define B2	f6#define B3	f7#define B4	f8#define B5	f9#define T1	f10#define T2	f11#define T3	f12#define T4	f13#define T5	f14#define T6	f15#define T7	f16	PROLOGUE	PROFCODE	li	r10, -16	stfpdux	f14, SP, r10	stfpdux	f15, SP, r10	stfpdux	f16, SP, r10		slwi	INCX,  INCX, BASE_SHIFT	slwi	INCY,  INCY, BASE_SHIFT	add	INCX2, INCX, INCX	add	INCY2, INCY, INCY	cmpwi	cr0, N, 0	ble	LL(999)	cmpwi	cr0, INCX, SIZE	bne	LL(100)	cmpwi	cr0, INCY, SIZE	bne	LL(100)	sub	X, X, INCX2	sub	Y, Y, INCY2	mr	X2, X	mr	Y2, Y	andi.	r0, X, 2 * SIZE - 1	bne	LL(30)	andi.	r0, Y, 2 * SIZE - 1	bne	LL(20)	.align 4LL(10):  /* X : aligned     Y : aligned */	srawi.	r0, N, 2	mtspr	CTR,  r0	beq-	LL(15)	LFPDUX	A1,    X,  INCX2	LFPDUX	B1,    Y,  INCY2	LFPDUX	A2,    X,  INCX2	LFPDUX	B2,    Y,  INCY2	LFPDUX	A3,    X,  INCX2	LFPDUX	B3,    Y,  INCY2	LFPDUX	A4,    X,  INCX2	LFPDUX	B4,    Y,  INCY2	bdz	LL(13)	.align 4LL(12):	STFPDUX	B1,    X2, INCY2	LFPDUX	B1,    Y,  INCY2	STFPDUX	A1,    Y2, INCY2	LFPDUX	A1,    X,  INCX2	STFPDUX	B2,    X2, INCY2	LFPDUX	B2,    Y,  INCY2	STFPDUX	A2,    Y2, INCY2	LFPDUX	A2,    X,  INCX2	STFPDUX	B3,    X2, INCY2	LFPDUX	B3,    Y,  INCY2	STFPDUX	A3,    Y2, INCY2	LFPDUX	A3,    X,  INCX2	STFPDUX	B4,    X2, INCY2	LFPDUX	B4,    Y,  INCY2	STFPDUX	A4,    Y2, INCY2	LFPDUX	A4,    X,  INCX2	bdnz	LL(12)	.align 4LL(13):	STFPDUX	B1,    X2, INCY2	STFPDUX	A1,    Y2, INCY2	STFPDUX	B2,    X2, INCY2	STFPDUX	A2,    Y2, INCY2	STFPDUX	B3,    X2, INCY2	STFPDUX	A3,    Y2, INCY2	STFPDUX	B4,    X2, INCY2	STFPDUX	A4,    Y2, INCY2	.align 4LL(15):	andi.	r0,  N, 3	beq	LL(999)	andi.	r0,  N, 2	beq	LL(16)	LFPDUX	A1,    X,  INCX2	LFPDUX	B1,    Y,  INCY2	LFPDUX	A2,    X,  INCX2	LFPDUX	B2,    Y,  INCY2	STFPDUX	B1,    X2, INCY2	STFPDUX	A1,    Y2, INCY2	STFPDUX	B2,    X2, INCY2	STFPDUX	A2,    Y2, INCY2	.align 4LL(16):	andi.	r0,  N, 1	beq	LL(999)	LFPDUX	A1,    X,  INCX2	LFPDUX	B1,    Y,  INCY2	STFPDUX	B1,    X2, INCY2	STFPDUX	A1,    Y2, INCY2	b  LL(999)	.align 4LL(20):  /* X : aligned     Y : unaligned */	LFXDUX	A1, X, INCX2	LFDX	B1, Y, INCY2	STFSDX	A1, Y2, INCY2	add	Y,  Y,  INCY	add	Y2, Y2, INCY	addi	N, N, -1	cmpwi	cr0, N, 0	ble	LL(29)	.align 4	srawi.	r0, N, 2	mtspr	CTR,  r0	beq-	LL(25)	LFXDUX	T1,    X,  INCX2	LFXDUX	T2,    Y,  INCY2	LFXDUX	T3,    X,  INCX2	LFXDUX	T4,    Y,  INCY2	LFPDUX	A4,    X,  INCX2	fsmr	A1, T1	LFPDUX	B4,    Y,  INCY2	fsmr	B1, T2	LFPDUX	A5,    X,  INCX2	fsmr	T1, T3	LFPDUX	B5,    Y,  INCY2	fsmr	T2, T4	bdz	LL(23)	.align 4LL(22):	fxmr	T5, A4	STFPDUX	A1,    Y2, INCY2	fxmr	T6, B4	STFPDUX	B1,    X2, INCX2	fxmr	A1, A5	STFPDUX	T1,    Y2, INCY2	fxmr	B1, B5	STFPDUX	T2,    X2, INCX2	fsmr	T3, T5	LFPDUX	A2,    X,  INCX2	fsmr	T4, T6 	LFPDUX	B2,    Y,  INCY2	fsmr	T5, A1	LFPDUX	A3,    X,  INCX2	fsmr	T6, B1	LFPDUX	B3,    Y,  INCY2	fxmr	T1, A2	STFPDUX	T3,    Y2, INCY2	fxmr	T2, B2	STFPDUX	T4,    X2, INCX2	fxmr	T3, A3	STFPDUX	T5,    Y2, INCY2	fxmr	T4, B3	STFPDUX	T6,    X2, INCX2	fsmr	A1, T1	LFPDUX	A4,    X,  INCX2	fsmr	B1, T2	LFPDUX	B4,    Y,  INCY2	fsmr	T1, T3	LFPDUX	A5,    X,  INCX2	fsmr	T2, T4	LFPDUX	B5,    Y,  INCY2	bdnz	LL(22)	.align 4LL(23):	fxmr	T5, A4	STFPDUX	A1,    Y2, INCY2	fxmr	T6, B4	STFPDUX	B1,    X2, INCX2	fxmr	A1, A5	STFPDUX	T1,    Y2, INCY2	fxmr	B1, B5	STFPDUX	T2,    X2, INCX2	fsmr	T3, T5	fsmr	T4, T6	fsmr	T5, A1	fsmr	T6, B1	STFPDUX	T3,    Y2, INCY2	STFPDUX	T4,    X2, INCX2	STFPDUX	T5,    Y2, INCY2	STFPDUX	T6,    X2, INCX2	.align 4LL(25):	andi.	r0,  N, 3	beq	LL(29)		andi.	r0,  N, 2	beq	LL(27)	LFXDUX	A2,    X,  INCX2	LFXDUX	B2,    Y,  INCY2	LFXDUX	A3,    X,  INCX2	LFXDUX	B3,    Y,  INCY2	fsmr	A1, A2	fsmr	B1, B2	fsmr	A2, A3	fsmr	B2, B3	STFPDUX	A1,    Y2, INCY2	STFPDUX	B1,    X2, INCX2	STFPDUX	A2,    Y2, INCY2	fpmr	A1, A3	STFPDUX	B2,    X2, INCX2	fpmr	B1, B3	.align 4LL(27):	andi.	r0,  N, 1	beq	LL(29)	LFXDUX	A2,    X,  INCX2	LFXDUX	B2,    Y,  INCY2	fsmr	A1, A2	fsmr	B1, B2	STFPDUX	A1,    Y2, INCY2	fpmr	A1, A2	STFPDUX	B1,    X2, INCX2	fpmr	B1, B2	.align 4LL(29):	LFSDX	B1,    Y,  INCY2	STFDX	A1,    Y2, INCY2	STFPDX	B1,    X2, INCX2	b  LL(999)	.align 4LL(30):  /* X : unaligned   Y : aligned */	andi.	r0, Y, 2 * SIZE - 1	bne	LL(40)	LFXDUX	A1, Y, INCY2	LFDX	B1, X, INCX2	STFSDX	A1, X2, INCX2	add	X,  X,  INCX	add	X2, X2, INCX	addi	N, N, -1	cmpwi	cr0, N, 0	ble	LL(39)	.align 4	srawi.	r0, N, 2	mtspr	CTR,  r0	beq-	LL(35)	LFXDUX	T1,    Y,  INCY2	LFXDUX	T2,    X,  INCX2	LFXDUX	T3,    Y,  INCY2	LFXDUX	T4,    X,  INCX2	LFPDUX	A4,    Y,  INCY2	fsmr	A1, T1	LFPDUX	B4,    X,  INCX2	fsmr	B1, T2	LFPDUX	A5,    Y,  INCY2	fsmr	T1, T3	LFPDUX	B5,    X,  INCX2	fsmr	T2, T4	bdz	LL(33)	.align 4LL(32):	fxmr	T5, A4	STFPDUX	A1,    X2, INCX2	fxmr	T6, B4	STFPDUX	B1,    Y2, INCY2	fxmr	A1, A5	STFPDUX	T1,    X2, INCX2	fxmr	B1, B5	STFPDUX	T2,    Y2, INCY2	fsmr	T3, T5	LFPDUX	A2,    Y,  INCY2	fsmr	T4, T6 	LFPDUX	B2,    X,  INCX2	fsmr	T5, A1	LFPDUX	A3,    Y,  INCY2	fsmr	T6, B1	LFPDUX	B3,    X,  INCX2	fxmr	T1, A2	STFPDUX	T3,    X2, INCX2	fxmr	T2, B2	STFPDUX	T4,    Y2, INCY2	fxmr	T3, A3	STFPDUX	T5,    X2, INCX2	fxmr	T4, B3	STFPDUX	T6,    Y2, INCY2	fsmr	A1, T1	LFPDUX	A4,    Y,  INCY2	fsmr	B1, T2	LFPDUX	B4,    X,  INCX2	fsmr	T1, T3	LFPDUX	A5,    Y,  INCY2	fsmr	T2, T4	LFPDUX	B5,    X,  INCX2	bdnz	LL(32)	.align 4LL(33):	fxmr	T5, A4	STFPDUX	A1,    X2, INCX2	fxmr	T6, B4	STFPDUX	B1,    Y2, INCY2	fxmr	A1, A5	STFPDUX	T1,    X2, INCX2	fxmr	B1, B5	STFPDUX	T2,    Y2, INCY2	fsmr	T3, T5	fsmr	T4, T6	fsmr	T5, A1	fsmr	T6, B1	STFPDUX	T3,    X2, INCX2	STFPDUX	T4,    Y2, INCY2	STFPDUX	T5,    X2, INCX2	STFPDUX	T6,    Y2, INCY2	.align 4LL(35):	andi.	r0,  N, 3	beq	LL(39)		andi.	r0,  N, 2	beq	LL(37)	LFXDUX	A2,    Y,  INCY2	LFXDUX	B2,    X,  INCX2	LFXDUX	A3,    Y,  INCY2	LFXDUX	B3,    X,  INCX2	fsmr	A1, A2	fsmr	B1, B2	fsmr	A2, A3	fsmr	B2, B3	STFPDUX	A1,    X2, INCX2	STFPDUX	B1,    Y2, INCY2	STFPDUX	A2,    X2, INCX2	fpmr	A1, A3	STFPDUX	B2,    Y2, INCY2	fpmr	B1, B3	.align 4LL(37):	andi.	r0,  N, 1	beq	LL(39)	LFXDUX	A2,    Y,  INCY2	LFXDUX	B2,    X,  INCX2	fsmr	A1, A2	fsmr	B1, B2	STFPDUX	A1,    X2, INCX2	fpmr	A1, A2	STFPDUX	B1,    Y2, INCY2	fpmr	B1, B2	.align 4LL(39):	LFSDX	B1,    X,  INCX2	STFDX	A1,    X2, INCX2	STFPDX	B1,    Y2, INCY2	b  LL(999)	.align 4LL(40):  /* X : unaligned   Y : unaligned */	LFDX	A1, Y, INCY2	LFDX	B1, X, INCX2	add	X,  X,  INCX	add	Y,  Y,  INCY	addi	N, N, -1	cmpwi	cr0, N, 0	STFDX	A1, X2, INCX2	STFDX	B1, Y2, INCY2	add	X2, X2, INCX	add	Y2, Y2, INCY	ble	LL(49)	srawi.	r0, N, 2	mtspr	CTR,  r0	beq-	LL(45)	LFPDUX	A1,    X,  INCX2	LFPDUX	B1,    Y,  INCY2	LFPDUX	A2,    X,  INCX2	LFPDUX	B2,    Y,  INCY2	LFPDUX	A3,    X,  INCX2	LFPDUX	B3,    Y,  INCY2	LFPDUX	A4,    X,  INCX2	LFPDUX	B4,    Y,  INCY2	bdz	LL(43)	.align 4LL(42):	STFPDUX	B1,    X2, INCY2	LFPDUX	B1,    Y,  INCY2	STFPDUX	A1,    Y2, INCY2	LFPDUX	A1,    X,  INCX2	STFPDUX	B2,    X2, INCY2	LFPDUX	B2,    Y,  INCY2	STFPDUX	A2,    Y2, INCY2	LFPDUX	A2,    X,  INCX2	STFPDUX	B3,    X2, INCY2	LFPDUX	B3,    Y,  INCY2	STFPDUX	A3,    Y2, INCY2	LFPDUX	A3,    X,  INCX2	STFPDUX	B4,    X2, INCY2	LFPDUX	B4,    Y,  INCY2	STFPDUX	A4,    Y2, INCY2	LFPDUX	A4,    X,  INCX2	bdnz	LL(42)	.align 4LL(43):	STFPDUX	B1,    X2, INCY2	STFPDUX	A1,    Y2, INCY2	STFPDUX	B2,    X2, INCY2	STFPDUX	A2,    Y2, INCY2	STFPDUX	B3,    X2, INCY2	STFPDUX	A3,    Y2, INCY2	STFPDUX	B4,    X2, INCY2	STFPDUX	A4,    Y2, INCY2	.align 4LL(45):	andi.	r0,  N, 3	beq	LL(49)	andi.	r0,  N, 2	beq	LL(46)	LFPDUX	A1,    X,  INCX2	LFPDUX	B1,    Y,  INCY2	LFPDUX	A2,    X,  INCX2	LFPDUX	B2,    Y,  INCY2	STFPDUX	B1,    X2, INCY2	STFPDUX	A1,    Y2, INCY2	STFPDUX	B2,    X2, INCY2	STFPDUX	A2,    Y2, INCY2	.align 4LL(46):	andi.	r0,  N, 1	beq	LL(49)	LFPDUX	A1,    X,  INCX2	LFPDUX	B1,    Y,  INCY2	STFPDUX	B1,    X2, INCY2	STFPDUX	A1,    Y2, INCY2	.align 4LL(49):	LFDX	A1, Y, INCY2	LFDX	B1, X, INCX2	STFDX	A1, X2, INCX2	STFDX	B1, Y2, INCY2	b  LL(999)	.align 4LL(100):	subi	INCX2, INCX2, SIZE	subi	INCY2, INCY2, SIZE	li	INCX, SIZE	li	INCY, SIZE	sub	X, X, INCX2	sub	Y, Y, INCY2	mr	X2, X	mr	Y2, Y	srawi.	r0, N, 1	mtspr	CTR,  r0	beq-	LL(115)	LFDUX	A1,   X,  INCX2	LFDUX	B1,   Y,  INCY2	LFDUX	A2,   X,  INCX	LFDUX	B2,   Y,  INCY	LFDUX	A3,   X,  INCX2	LFDUX	B3,   Y,  INCY2	LFDUX	A4,   X,  INCX	LFDUX	B4,   Y,  INCY	bdz	LL(113)	.align 4LL(112):	STFDUX	B1,   X2, INCX2	LFDUX	B1,   Y,  INCY2	STFDUX	A1,   Y2, INCY2	LFDUX	A1,   X,  INCX2	STFDUX	B2,   X2, INCX	LFDUX	B2,   Y,  INCY	STFDUX	A2,   Y2, INCY	LFDUX	A2,   X,  INCX	STFDUX	B3,   X2, INCX2	LFDUX	B3,   Y,  INCY2	STFDUX	A3,   Y2, INCY2	LFDUX	A3,   X,  INCX2	STFDUX	B4,   X2, INCX	LFDUX	B4,   Y,  INCY	STFDUX	A4,   Y2, INCY	LFDUX	A4,   X,  INCX	bdnz	LL(112)	.align 4LL(113):	STFDUX	B1,   X2, INCX2	STFDUX	A1,   Y2, INCY2	STFDUX	B2,   X2, INCX	STFDUX	A2,   Y2, INCY	STFDUX	B3,   X2, INCX2	STFDUX	A3,   Y2, INCY2	STFDUX	B4,   X2, INCX	STFDUX	A4,   Y2, INCY	.align 4LL(115):	andi.	r0,  N, 1	beq	LL(999)	LFDUX	A1,   X,  INCX2	LFDUX	A2,   X,  INCX	LFDUX	B1,   Y,  INCY2	LFDUX	B2,   Y,  INCY	STFDUX	B1,   X2, INCX2	STFDUX	B2,   X2, INCX	STFDUX	A1,   Y2, INCY2	STFDUX	A2,   Y2, INCY	.align 4LL(999):	li	r10, 16	addi	SP, SP,  -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 + -