为什么我的分散代码的性能优于Vc SIMD？

Question

I'm a rookie to SIMD program

unsigned int Hash(unsigned int f);
uint_v Hash(uint_v vec);

int main()
{
        std::random_device rd;
        unsigned* mem1=new unsigned [_size]();
        for(int i=0;i<_size;++i)
                mem1[i]=rd();

        time_t t1=clock();
        uint_v mem;
        for(int i=0;i<_size;i+=uint_v::size())
        {
                mem.load(mem1+i,Vc::Unaligned);
                uint_v temp=Hash(mem);
        }
        t1=clock()-t1;
        std::cout<<"simd time:"<<(1.0*t1)/CLOCKS_PER_SEC<<"\n";

        time_t t2=clock();
        for(int i=0;i<_size;++i)
                unsigned int temp=Hash(mem1[i]);
        t2=clock()-t2;
        std::cout<<"normal time:"<<(1.0*t2)/CLOCKS_PER_SEC<<"\n";

        return 0;
}
unsigned int Hash(unsigned int f)
{
        return (f>>7)^(f>>13)^(f>>21)^f;
}
uint_v Hash(uint_v vec)
{
        uint_v mem=vec.apply([](unsigned f) ->unsigned{return (f>>7)^(f>>13)^(f>>21)^f;});
        return mem;
}

my code as above,the time result is:
simd time:0.127762 normal time:0.034841
result is similar with comparing the date in mem1 and mem2(Vc uint_v vector)

Answer 1

You are not measuring what you intended to measure. The compiler will do dead code elimination for everything that you calculate but never use (well, everything where the compiler is 100% that it's never used). The compiler should have done DCE on both loops but apparently failed to do it for the Vc case.

Ideas:

• store the result to a global variable
• use inline asm to fake a use of the result