📄 asum_hummer.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 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 T1 f12#define T2 f13#define T3 f14#define T4 f15 PROLOGUE PROFCODE li r10, -16 stfpdux f14, SP, r10 stfpdux f15, 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)#endif lfpdx C1, SP, r10 # Zero clear slwi INCX, INCX, BASE_SHIFT add INCX2, INCX, INCX fpmr C2, C1 fpmr C3, C1 fpmr C4, C1 cmpwi cr0, N, 0 ble LL(999) cmpwi cr0, INCX, 0 ble LL(999) cmpwi cr0, INCX, SIZE bne LL(100) andi. r0, X, 2 * SIZE - 1 beq LL(05) LFD C1, 0(X) addi X, X, 1 * SIZE addi N, N, -1 cmpwi cr0, N, 0 fabs C1, C1 ble LL(999) .align 4LL(05): srawi. r0, N, 4 sub X, X, INCX2 mtspr CTR, r0 beq- LL(15) LFPDUX A1, X, INCX2 fpmr T1, C2 LFPDUX A2, X, INCX2 fpmr T2, C2 LFPDUX A3, X, INCX2 fpmr T3, C2 LFPDUX A4, X, INCX2 fpmr T4, C2 LFPDUX A5, X, INCX2 LFPDUX A6, X, INCX2 LFPDUX A7, X, INCX2 LFPDUX A8, X, INCX2 bdz LL(13) .align 4LL(12): fpadd C1, C1, T1 nop fpabs T1, A1 LFPDUX A1, X, INCX2 fpadd C2, C2, T2 nop fpabs T2, A2 LFPDUX A2, X, INCX2 fpadd C3, C3, T3 nop fpabs T3, A3 LFPDUX A3, X, INCX2 fpadd C4, C4, T4 nop fpabs T4, A4 LFPDUX A4, X, INCX2 fpadd C1, C1, T1 nop fpabs T1, A5 LFPDUX A5, X, INCX2 fpadd C2, C2, T2 nop fpabs T2, A6 LFPDUX A6, X, INCX2 fpadd C3, C3, T3 nop fpabs T3, A7 LFPDUX A7, X, INCX2 fpadd C4, C4, T4 fpabs T4, A8 LFPDUX A8, X, INCX2 bdnz LL(12) .align 4LL(13): fpadd C1, C1, T1 fpabs T1, A1 fpadd C2, C2, T2 fpabs T2, A2 fpadd C3, C3, T3 fpabs T3, A3 fpadd C4, C4, T4 fpabs T4, A4 fpadd C1, C1, T1 fpabs T1, A5 fpadd C2, C2, T2 fpabs T2, A6 fpadd C3, C3, T3 fpabs T3, A7 fpadd C4, C4, T4 fpabs T4, A8 fpadd C1, C1, T1 fpadd C2, C2, T2 fpadd C3, C3, T3 fpadd C4, C4, T4 .align 4LL(15): andi. r0, N, 15 beq LL(999) andi. r0, N, 8 beq LL(16) LFPDUX A1, X, INCX2 LFPDUX A2, X, INCX2 LFPDUX A3, X, INCX2 LFPDUX A4, X, INCX2 fpabs T1, A1 fpabs T2, A2 fpabs T3, A3 fpabs T4, A4 fpadd C1, C1, T1 fpadd C2, C2, T2 fpadd C3, C3, T3 fpadd C4, C4, T4 .align 4LL(16): andi. r0, N, 4 beq LL(17) LFPDUX A1, X, INCX2 LFPDUX A2, X, INCX2 fpabs T1, A1 fpabs T2, A2 fpadd C1, C1, T1 fpadd C2, C2, T2 .align 4LL(17): andi. r0, N, 2 beq LL(18) LFPDUX A1, X, INCX2 fpabs T1, A1 fpadd C1, C1, T1 .align 4LL(18): andi. r0, N, 1 beq LL(999) LFDX A1, X, INCX2 fabs T1, A1 fadd C1, C1, T1 b LL(999) .align 4LL(100): sub X2, X, INCX sub X, X, INCX2 srawi. r0, N, 4 mtspr CTR, r0 beq- LL(115) LFDUX A1, X, INCX2 fpmr T1, C2 LFDUX A2, X, INCX2 fpmr T2, C2 LFDUX A3, X, INCX2 fpmr T3, C2 LFDUX A4, X, INCX2 fpmr T4, C2 LFDUX A5, X, INCX2 LFSDUX A1, X2, INCX2 LFDUX A6, X, INCX2 LFSDUX A2, X2, INCX2 LFDUX A7, X, INCX2 LFSDUX A3, X2, INCX2 LFDUX A8, X, INCX2 LFSDUX A4, X2, INCX2 bdz LL(113) .align 4LL(112): fpadd C1, C1, T1 LFSDUX A5, X2, INCX2 fpabs T1, A1 LFDUX A1, X, INCX2 fpadd C2, C2, T2 LFSDUX A6, X2, INCX2 fpabs T2, A2 LFDUX A2, X, INCX2 fpadd C3, C3, T3 LFSDUX A7, X2, INCX2 fpabs T3, A3 LFDUX A3, X, INCX2 fpadd C4, C4, T4 LFSDUX A8, X2, INCX2 fpabs T4, A4 LFDUX A4, X, INCX2 fpadd C1, C1, T1 LFSDUX A1, X2, INCX2 fpabs T1, A5 LFDUX A5, X, INCX2 fpadd C2, C2, T2 LFSDUX A2, X2, INCX2 fpabs T2, A6 LFDUX A6, X, INCX2 fpadd C3, C3, T3 LFSDUX A3, X2, INCX2 fpabs T3, A7 LFDUX A7, X, INCX2 fpadd C4, C4, T4 LFSDUX A4, X2, INCX2 fpabs T4, A8 LFDUX A8, X, INCX2 bdnz LL(112) .align 4LL(113): fpadd C1, C1, T1 nop fpabs T1, A1 LFSDUX A5, X2, INCX2 fpadd C2, C2, T2 nop fpabs T2, A2 LFSDUX A6, X2, INCX2 fpadd C3, C3, T3 nop fpabs T3, A3 LFSDUX A7, X2, INCX2 fpadd C4, C4, T4 nop fpabs T4, A4 LFSDUX A8, X2, INCX2 fpadd C1, C1, T1 fpabs T1, A5 fpadd C2, C2, T2 fpabs T2, A6 fpadd C3, C3, T3 fpabs T3, A7 fpadd C4, C4, T4 fpabs T4, A8 fpadd C1, C1, T1 fpadd C2, C2, T2 fpadd C3, C3, T3 fpadd C4, C4, T4 .align 4LL(115): andi. r0, N, 15 beq LL(999) andi. r0, N, 8 beq LL(116) LFDUX A1, X, INCX2 LFDUX A2, X2, INCX2 LFDUX A3, X, INCX2 LFDUX A4, X2, INCX2 fabs T1, A1 LFDUX A5, X, INCX2 fabs T2, A2 LFDUX A6, X2, INCX2 fabs T3, A3 LFDUX A7, X, INCX2 fabs T4, A4 LFDUX A8, X2, INCX2 fadd C1, C1, T1 fabs T1, A5 fadd C2, C2, T2 fabs T2, A6 fadd C3, C3, T3 fabs T3, A7 fadd C4, C4, T4 fabs T4, A8 fadd C1, C1, T1 fadd C2, C2, T2 fadd C3, C3, T3 fadd C4, C4, T4 .align 4LL(116): andi. r0, N, 4 beq LL(117) LFDUX A1, X, INCX2 LFDUX A2, X2, INCX2 LFDUX A3, X, INCX2 LFDUX A4, X2, INCX2 fabs T1, A1 fabs T2, A2 fabs T3, A3 fabs T4, A4 fadd C1, C1, T1 fadd C2, C2, T2 fadd C3, C3, T3 fadd C4, C4, T4 .align 4LL(117): andi. r0, N, 2 beq LL(118) LFDUX A1, X, INCX2 LFDUX A2, X2, INCX2 fabs T1, A1 fabs T2, A2 fadd C1, C1, T1 fadd C2, C2, T2 .align 4LL(118): andi. r0, N, 1 beq LL(999) LFDX A1, X, INCX2 fabs T1, A1 fadd C1, C1, T1 .align 4LL(999): fpadd C1, C1, C2 li r10, 16 fpadd C3, C3, C4 fpadd C1, C1, C3 lfpdux f15, SP, r10 fsmtp C2, C1 lfpdux f14, SP, r10 addi SP, SP, 16 fadd C1, C2, C1 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 + -