Neon vs cpp

gistfile1.txt

//
//  ViewController.m
//  compare
//
//  Created by Allan on 27/7/2019.
//  Copyright © 2019 Allan. All rights reserved.
//

#import "ViewController.h"
#include <arm_neon.h>
@interface ViewController ()

@end

@implementation ViewController

- (void)viewDidLoad {
    [super viewDidLoad];
//    [self cppTest];
    [self neonTest];
}

-(void)neonTest{
    double tic = CACurrentMediaTime();
    int total = 1024 * 1024 * 100;
#if __ARM_NEON
    
    float *input = new float[total];
    float *output = new float[total];
    float alpha = 1.0f;
    float beta = 2.0f;
    int nn = total >> 2;
    for(int i = 0; i<total; i++){
        input[i] = (float)100;
    }
    float32x4_t _alpha = vdupq_n_f32(alpha);
    float32x4_t _beta = vdupq_n_f32(beta);
#else
    
#endif
    
#if __ARM_NEON
    const float *inputptr = input;
    float *outputptr = input;
#if __aarch64__
    if(nn > 0){
        asm volatile(
                     "0:        \n"
                     "mov    v0.4s,  %5.4s   \n"
                     "mov    v1.4s,  %6.4s   \n"
                     "prfm  pldl1keep,  [%4, #64]   \n"
                     "ld1   {v2.4s},    [%4], #16   \n"
                     "fmul  v3.4s, v2.4s, v0.4s     \n"
                     "fadd  v4.4s, v3.4s, v1.4s     \n"
                     "subs  %w0, %w0, #1            \n"
                     "st1   {v4.4s}, [%1], #16      \n"
                     "bne 0b"
                     :
                     "=r"(nn),      //%0
                     "=r"(outputptr)//%1
                     :
                     "0"(nn),
                     "1"(outputptr),
                     "r"(inputptr), //%4
                     "w"(_alpha),   //%5
                     "w"(_beta)     //%6
                     :"cc","memory","v0","v1","v2","v3","v4"
                     );
    }
#else
#endif// __aarch64
    for (int i=0; i<2; i++){
        int nn = total >> 2;
        outputptr[i] = inputptr [i + nn * 4] *alpha +beta;
    }
#endif// __ARM_NEON
    printf("neon -- cost %.4f ms\n", 1000 * (CACurrentMediaTime() - tic));
}

- (void)cppTest{
    double tic = CACurrentMediaTime();
    int m_total = 1024 * 1024 *100;
    float *input = new float[m_total];
    float *output = new float[m_total];
    
    for(int i = 0; i< m_total; i++){
        input[i] = (float)100;
    }
    
    for(int i=0; i<m_total;i++){
        output[i] = input[i] * 2 +3;
    }
    
    printf("cpp cost %.4f ms\n", 1000 * (CACurrentMediaTime() - tic));
}



@end