⭐ 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

📄 zdot_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"		#if defined(F_INTERFACE) && defined(F_INTERFACE_F2C)#define RESULT	r3#define N	r4#define X	r5#define INCX	r6	#define Y	r7#define INCY	r8#else#define N	r3#define X	r4#define INCX	r5	#define Y	r6#define INCY	r7#endif#define INCX2	r9#define INCY2	r10#define C1	f1#define C2	f2#define C3	f0#define C4	f3#define C5	f4#define C6	f5#define C7	f6#define C8	f7#define A1	f8#define A2	f9#define A3	f10#define A4	f11#define A5	f12#define A6	f13#define A7	f14#define A8	f15#define B1	f16#define B2	f17#define B3	f18#define B4	f19#define B5	f20#define B6	f21#define B7	f22#define B8	f23#ifndef CONJ#define FXCXNPMA fxcxnpma#else#define FXCXNPMA fxcxnsma#endif	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	stfpdux	f20, SP, r10	stfpdux	f21, SP, r10	stfpdux	f22, SP, r10	stfpdux	f23, SP, r10	li	r10,   0	stwu	r10,   -4(SP)	stwu	r10,   -4(SP)	stwu	r10,   -4(SP)	stwu	r10,   -4(SP)#ifdef F_INTERFACE	LDINT	N,    0(N)	LDINT	INCX, 0(INCX)	LDINT	INCY, 0(INCY)#endif	lfpdx	C1, SP, r10		# Zero clear	slwi	INCX,  INCX, BASE_SHIFT	add	INCX2, INCX, INCX	fpmr	C2, C1	slwi	INCY,  INCY, BASE_SHIFT	fpmr	C3, C1	add	INCY2, INCY, INCY	fpmr	C4, C1	fpmr	C5, C1	fpmr	C6, C1	fpmr	C7, C1	fpmr	C8, C1	cmpwi	cr0, N, 0	ble	LL(99)#ifdef F_INTERFACE	cmpwi	cr0, INCX, 0	bge+	LL(05)	subi	r0, N, 1	mullw	r0, r0, INCX2	sub	X, X, r0	.align 4LL(05):	cmpwi	cr0, INCY, 0	bge+	LL(06)	subi	r0, N, 1	mullw	r0, r0, INCY2	sub	Y, Y, r0	.align 4LL(06):#endif	andi.	r0, X, 2 * SIZE - 1	bne	LL(100)	andi.	r0, Y, 2 * SIZE - 1	bne	LL(100)/* X is aligned, Y is aligned */LL(10):	sub	X, X, INCX2	sub	Y, Y, INCY2	srawi.	r0, N, 3	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	LFPDUX	A5,    X, INCX2	LFPDUX	B5,    Y, INCY2	LFPDUX	A6,    X, INCX2	LFPDUX	B6,    Y, INCY2	LFPDUX	A7,    X, INCX2	LFPDUX	B7,    Y, INCY2	LFPDUX	A8,    X, INCX2	bdz	LL(14)	.align 4LL(13):	fxcpmadd C1, A1, B1, C1	LFPDUX	B8,    Y, INCY2	FXCXNPMA C2, A1, B1, C2	LFPDUX	A1,    X, INCX2	fxcpmadd C3, A2, B2, C3	LFPDUX	B1,    Y, INCY2	FXCXNPMA C4, A2, B2, C4	LFPDUX	A2,    X, INCX2	fxcpmadd C5, A3, B3, C5	LFPDUX	B2,    Y, INCY2	FXCXNPMA C6, A3, B3, C6	LFPDUX	A3,    X, INCX2	fxcpmadd C7, A4, B4, C7	LFPDUX	B3,    Y, INCY2	FXCXNPMA C8, A4, B4, C8	LFPDUX	A4,    X, INCX2	fxcpmadd C1, A5, B5, C1	LFPDUX	B4,    Y, INCY2	FXCXNPMA C2, A5, B5, C2	LFPDUX	A5,    X, INCX2	fxcpmadd C3, A6, B6, C3	LFPDUX	B5,    Y, INCY2	FXCXNPMA C4, A6, B6, C4	LFPDUX	A6,    X, INCX2	fxcpmadd C5, A7, B7, C5	LFPDUX	B6,    Y, INCY2	FXCXNPMA C6, A7, B7, C6	LFPDUX	A7,    X, INCX2	fxcpmadd C7, A8, B8, C7	LFPDUX	B7,    Y, INCY2	FXCXNPMA C8, A8, B8, C8	LFPDUX	A8,    X, INCX2	bdnz	LL(13)	.align 4LL(14):	LFPDUX	B8,    Y, INCY2	fxcpmadd C1, A1, B1, C1	FXCXNPMA C2, A1, B1, C2	fxcpmadd C3, A2, B2, C3	FXCXNPMA C4, A2, B2, C4	fxcpmadd C5, A3, B3, C5	FXCXNPMA C6, A3, B3, C6	fxcpmadd C7, A4, B4, C7	FXCXNPMA C8, A4, B4, C8	fxcpmadd C1, A5, B5, C1	FXCXNPMA C2, A5, B5, C2	fxcpmadd C3, A6, B6, C3	FXCXNPMA C4, A6, B6, C4	fxcpmadd C5, A7, B7, C5	FXCXNPMA C6, A7, B7, C6	fxcpmadd C7, A8, B8, C7	FXCXNPMA C8, A8, B8, C8	.align 4LL(15):	andi.	r0,  N, 7	beq	LL(99)	andi.	r0,  N, 4	beq	LL(16)	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	fxcpmadd C1, A1, B1, C1	FXCXNPMA C2, A1, B1, C2	fxcpmadd C3, A2, B2, C3	FXCXNPMA C4, A2, B2, C4	fxcpmadd C5, A3, B3, C5	FXCXNPMA C6, A3, B3, C6	fxcpmadd C7, A4, B4, C7	FXCXNPMA C8, A4, B4, C8	.align 4LL(16):	andi.	r0,  N, 2	beq	LL(17)	LFPDUX	A1,    X, INCX2	LFPDUX	B1,    Y, INCY2	LFPDUX	A2,    X, INCX2	LFPDUX	B2,    Y, INCY2	fxcpmadd C1, A1, B1, C1	FXCXNPMA C2, A1, B1, C2	fxcpmadd C3, A2, B2, C3	FXCXNPMA C4, A2, B2, C4	.align 4LL(17):	andi.	r0,  N, 1	beq	LL(99)	LFPDUX	A1,    X, INCX2	LFPDUX	B1,    Y, INCY2	fxcpmadd C1, A1, B1, C1	FXCXNPMA C2, A1, B1, C2	.align 4LL(99):	li	r10, 16	fpadd	C1, C1, C5	lfpdux	f23, SP, r10	fpadd	C2, C2, C6	lfpdux	f22, SP, r10	fpadd	C3, C3, C7	lfpdux	f21, SP, r10	fpadd	C4, C4, C8	lfpdux	f20, SP, r10	lfpdux	f19, SP, r10	lfpdux	f18, SP, r10	fpadd	C1, C1, C3	lfpdux	f17, SP, r10	fpadd	C2, C2, C4	lfpdux	f16, SP, r10	fpadd	C1, C1, C2	lfpdux	f15, SP, r10	lfpdux	f14, SP, r10	fsmtp	C2, C1#if defined(F_INTERFACE) && defined(F_INTERFACE_F2C)	STFD	C1,    0 * SIZE(RESULT)	STFD	C2,    1 * SIZE(RESULT)#endif	addi	SP, SP,  16	blr	.align 4/* X is aligned, Y is NOT aligned */LL(100):	subi	INCX2, INCX2, SIZE	subi	INCY2, INCY2, SIZE	li	INCX, SIZE	li	INCY, SIZE	sub	X, X, INCX2	sub	Y, Y, INCY2	srawi.	r0, N, 2	mtspr	CTR,  r0	beq-	LL(105)	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	LFDUX	A5,    X, INCX2	LFDUX	B5,    Y, INCY2	LFDUX	A6,    X, INCX	LFDUX	B6,    Y, INCY	LFDUX	A7,    X, INCX2	LFDUX	B7,    Y, INCY2	LFDUX	A8,    X, INCX	bdz	LL(104)	.align 4LL(103):	fmadd	C1, A1, B1, C1	LFDUX	B8,    Y, INCY	fmadd	C2, A1, B2, C2	LFDUX	A1,    X, INCX2	fmadd	C3, A2, B1, C3	LFDUX	B1,    Y, INCY2	fmadd	C4, A2, B2, C4	LFDUX	A2,    X, INCX	fmadd	C5, A3, B3, C5	LFDUX	B2,    Y, INCY	fmadd	C6, A3, B4, C6	LFDUX	A3,    X, INCX2	fmadd	C7, A4, B3, C7	LFDUX	B3,    Y, INCY2	fmadd	C8, A4, B4, C8	LFDUX	A4,    X, INCX	fmadd	C1, A5, B5, C1	LFDUX	B4,    Y, INCY	fmadd	C2, A5, B6, C2	LFDUX	A5,    X, INCX2	fmadd	C3, A6, B5, C3	LFDUX	B5,    Y, INCY2	fmadd	C4, A6, B6, C4	LFDUX	A6,    X, INCX	fmadd	C5, A7, B7, C5	LFDUX	B6,    Y, INCY	fmadd	C6, A7, B8, C6	LFDUX	A7,    X, INCX2	fmadd	C7, A8, B7, C7	LFDUX	B7,    Y, INCY2	fmadd	C8, A8, B8, C8	LFDUX	A8,    X, INCX	bdnz	LL(103)	.align 4LL(104):	LFDUX	B8,    Y, INCY	fmadd	C1, A1, B1, C1	fmadd	C2, A1, B2, C2	fmadd	C3, A2, B1, C3	fmadd	C4, A2, B2, C4	fmadd	C5, A3, B3, C5	fmadd	C6, A3, B4, C6	fmadd	C7, A4, B3, C7	fmadd	C8, A4, B4, C8	fmadd	C1, A5, B5, C1	fmadd	C2, A5, B6, C2	fmadd	C3, A6, B5, C3	fmadd	C4, A6, B6, C4	fmadd	C5, A7, B7, C5	fmadd	C6, A7, B8, C6	fmadd	C7, A8, B7, C7	fmadd	C8, A8, B8, C8	.align 4LL(105):	andi.	r0,  N, 3	beq	LL(999)	andi.	r0,  N, 2	beq	LL(107)	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	fmadd	C1, A1, B1, C1	fmadd	C2, A1, B2, C2	fmadd	C3, A2, B1, C3	fmadd	C4, A2, B2, C4	fmadd	C5, A3, B3, C5	fmadd	C6, A3, B4, C6	fmadd	C7, A4, B3, C7	fmadd	C8, A4, B4, C8	.align 4LL(107):	andi.	r0,  N, 1	beq	LL(999)	LFDUX	A1,    X, INCX2	LFDUX	B1,    Y, INCY2	LFDUX	A2,    X, INCX	LFDUX	B2,    Y, INCY	fmadd	C1, A1, B1, C1	fmadd	C2, A1, B2, C2	fmadd	C3, A2, B1, C3	fmadd	C4, A2, B2, C4	.align 4LL(999):	li	r10, 16	fadd	C1, C1, C5	lfpdux	f23, SP, r10	fadd	C2, C2, C6	lfpdux	f22, SP, r10	fadd	C3, C3, C7	lfpdux	f21, SP, r10	fadd	C4, C4, C8	lfpdux	f20, SP, r10	lfpdux	f19, SP, r10	lfpdux	f18, SP, r10	lfpdux	f17, SP, r10	lfpdux	f16, SP, r10#ifndef CONJ	FSUB	C1,  C1,  C4	FADD	C2,  C2,  C3#else	FADD	C1,  C1,  C4	FSUB	C2,  C2,  C3#endif	lfpdux	f15, SP, r10	lfpdux	f14, SP, r10#if defined(F_INTERFACE) && defined(F_INTERFACE_F2C)	STFD	C1,    0 * SIZE(RESULT)	STFD	C2,    1 * SIZE(RESULT)#endif	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 + -