#include <iostream>
#include <iomanip>

#include <Vc/Vc>

#define ET_TIMER_RESULT_VAR vx_
#define ET_TIMER_RESULT_VAR_TYPE float*
//#define ET_NO_TIMING
#include <et/Timer.h>

#ifndef VC_IMPL_AVX
# pragma message("!!! no avx !!!")
#else
# pragma message("*** ok avx *** ")
#endif


// always use Vs's allocator for Real_t to produce aligned data
   VC_DECLARE_ALLOCATOR(float)

int main()
{
    std::cout<<"\n@@@ test0 - begin @@@"<<std::endl;
    ET::Timer::reset();
    size_t const nreps=100;
    size_t const ne=1024;
    size_t const nv=ne/Vc::float_v::Size;

    std::vector<float> x(1024);
    float* vx_ = &x[0];

    for( int i=0; i<ne; ++i ) {
        x[i]=0.0;
    }
    size_t nErrors=0;
    for( int i=0; i<ne; ++i ) {
        if( x[i]!=0.0 ) {
            std::cout<<"x["<<i<<"]="<<x[i]<<std::endl;
            ++nErrors;
        }
    }
    if( nErrors ) std::cout<<"#errors: "<<nErrors<<std::endl;
 //====
    ET_TIME_THIS
    ( "std::vector<float> scalar#1 ",
        for( int i=0; i<ne; i+=1 ) {
            x[i] -= 1.0;
        }
    , nreps
    , true
    )
    ET_TIME_THIS
    ( "std::vector<float> scalar#2 ",
        for( int i=0; i<ne; i+=2 ) {
            x[i  ] -= 1.0;
            x[i+1] -= 1.0;
        }
    , nreps
    , true
    )
    ET_TIME_THIS
    ( "std::vector<float> scalar#4 ",
        for( int i=0; i<ne; i+=4 ) {
            x[i  ] -= 1.0;
            x[i+1] -= 1.0;
            x[i+2] -= 1.0;
            x[i+3] -= 1.0;
        }
    , nreps
    , true
    )
    nErrors=0;
    float result = -3.*nreps;
//    std::cout<<result<<std::endl;
//    std::cout<<nreps<<std::endl;
    for( int i=0; i<ne; ++i ) {
        if( x[i]!=result ) {
            std::cout<<"x["<<i<<"]="<<x[i]<<" diff="<<x[i]-result<<std::endl;
            ++nErrors;
        }
    }
    if( nErrors ) std::cout<<"#errors: "<<nErrors<<std::endl;
 //====
    Vc::float_v one(1.);
    ET_TIME_THIS
    ( "std::vector<float> vector",
        for( int i=0; i<ne; i+=Vc::float_v::Size )
        {
            Vc::float_v vx( &x[i] );
            vx -= one;
            vx.store( &x[i] );
        }
    , nreps
    , true
    )
    nErrors=0;
    result -= nreps;
    for( int i=0; i<ne; ++i ) {
        if( x[i]!=result ) {
            std::cout<<"x["<<i<<"]="<<x[i]<<std::endl;
            ++nErrors;
        }
    }
    if( nErrors ) std::cout<<"#errors: "<<nErrors<<std::endl;
 //====
    Vc::Memory<Vc::float_v,ne> Vx;
    vx_ = &Vx[0];
    for( int i=0; i<ne; ++i ) {
        Vx[i] = 1.0;
    }
    ET_TIME_THIS
    ( "Vc::Memory<Vc::float_v,ne>  vector",
        for( int i=0; i<nv; ++i ) {
            Vx.vector(i) -= one;
        }
    , nreps
    , true
    )

    std::cout<<"@@@ test0 - done @@@"<<std::endl;

}