memcpy optimize copy_cut_border/copy_make_boder

[LittlePeng]

No description provided.

[LittlePeng]

android 6.0 测试脚本 copy_cut_borderv2 为memcpy版本：

static void bench_cut_border(int size, int iteration, char* comment){
    ncnn::Mat src = ncnn::Mat(size, size, 3);
    ncnn::Mat dst;
    bench_start();
    for(int i=0; i<iteration; i++){
        copy_cut_borderv1(src, dst, 1, 1, 1, 1);
    }
    bench_end(comment);
    
    bench_start();
    for(int i=0; i<iteration; i++){
        copy_cut_borderv2(src, dst, 1, 1, 1, 1);
    }
    bench_end(comment);
}

bench_cut_border(5, 10000, "3*5*5");
bench_cut_border(10, 10000, "3*10*10");
bench_cut_border(14, 10000, "3*14*14");
bench_cut_border(20, 10000, "3*20*20");
bench_cut_border(50, 10000, "3*50*50");
bench_cut_border(250, 5000, "3*250*250");
bench_cut_border(500, 1000, "3*500*500");
bench_cut_border(1000, 500, "3*1000*1000");

结果:

4.62ms   3*5*5
5.40ms   3*5*5
9.38ms   3*10*10
11.04ms   3*10*10
20.10ms   3*14*14
16.83ms   3*14*14
30.01ms   3*20*20
23.47ms   3*20*20
163.74ms   3*50*50
90.66ms   3*50*50
5135.47ms   3*250*250
4193.02ms   3*250*250
4396.05ms   3*500*500
3623.97ms   3*500*500
8779.94ms   3*1000*1000
7144.28ms   3*1000*1000

memcpy 从一次copy 12个float 开始速度有优势