deep_network/src/layers/convolution.cpp

#include "convolution.hpp"
#include "avx.hpp"

namespace Layer{

  ConvolutionLayer::ConvolutionLayer(size_t nf_,size_t ni_,size_t nj_,size_t p_,size_t q_,size_t mf_):Layer(nf_*ni_*nj_,mf_*(ni_-p_+1)*(nj_-q_+1)){
    nf=nf_;
    ni=ni_;
    nj=nj_;
    p=p_;
    q=q_;
    mf=mf_;
    mi=ni-p+1;
    mj=nj-q+1;
    K=init_vector(mf*nf*p*q);
    b=init_vector(mf*nf);
    nabla_K=init_vector(mf*nf*p*q);
    nabla_b=init_vector(mf*nf);
    size_t n_vindex1=8*AVX_SIZE(p*q);
    vindex1=new v8i[AVX_SIZE(p*q)];
    for(size_t i=0;i<AVX_SIZE(p*q);++i){
      for(size_t j=0;j<8;++j){
        vindex1[i].i[j]=0;
      }
    }
    for(size_t j=0;j<q;++j){
      for(size_t i=0;i<p;++i){
        size_t ind=j*p+i;
        vindex1[ind/8].i[ind%8]=j*nj+i;
      }
    }
  }

  ConvolutionLayer::~ConvolutionLayer(){
    delete_vector(K);
    delete_vector(b);
    delete_vector(nabla_K);
    delete_vector(nabla_b);
  }
  void
  ConvolutionLayer::init(Real m,Real d){
    default_random_engine generator;
    normal_distribution<Real> distribution(m,d);
    for(size_t i=0;i<mf*nf*i*q;++i){
      K[i]=distribution(generator);
    }
    for(size_t i=0;i<mf*nf;++i){
      b[i]=distribution(generator);
    }
  }

  Vector
  ConvolutionLayer::feed_forward(Vector x_){
    return avx_feed_forward(x_);
    x=x_;
    for(size_t g=0;g<mf;++g){
      for(size_t k=0;k<mi;++k){
        for(size_t l=0;l<mj;++l){
          Real temp=0;
          size_t offset_x=k*nj+l;
          for(size_t f=0;f<nf;++f){
            for(size_t r=0;r<p;++r){
              for(size_t s=0;s<q;++s){
                temp+=x[indice3(f,k+r,l+s,ni,nj)]*K[indice4(g,f,r,s,nf,p,q)];
              }
            }
            temp+=b[indice2(g,f,nf)];
          }
          y[indice3(g,k,l,mi,mj)]=temp;
        }
      }
    }
    return y;
  }

  Vector
  ConvolutionLayer::avx_feed_forward(Vector x_){

    x=x_;
    for(size_t g=0;g<mf;++g){
      for(size_t k=0;k<mi;++k){
        for(size_t l=0;l<mj;++l){
          Real temp=0;
          Real temp_avx=0;
          for(size_t f=0;f<nf;++f){
            size_t n=AVX_SIZE(p*q);
            Real* ptrK=&K[(g*nf+f)*p*q];
            Real* ptrx=&x[indice3(f,k,l,ni,nj)];

            for(size_t t=0;t<n;++t){
              cout<<t<<" > "<<endl;
              //__m256 x_data=_mm256_i32gather_ps(&x[indice3(f,k,l,ni,nj)],*(__m256i*)(&vindex1[t]),4);
              v8f u={_mm256_loadu_ps(ptrK)}; //Read 8 values of K
              __m256i vi=*(__m256i*)(&vindex1[t]);
              v8f v={_mm256_i32gather_ps(ptrx,vi,4)};
              u.avx=_mm256_mul_ps(u.avx,v.avx);
              temp_avx+=hadd(u);
              ptrK+=8; //There is 8 floats in m256
            }
            cout<<p<<" "<<q<<endl;
          //  exit(0);
            for(size_t r=0;r<p;++r){
              for(size_t s=0;s<q;++s){
                temp+=x[indice3(f,k+r,l+s,ni,nj)]*K[indice4(g,f,r,s,nf,p,q)];
              }
            }
            cout<<temp<<" vs "<<temp_avx<<endl;
            temp+=b[indice2(g,f,nf)];
            temp_avx+=b[indice2(g,f,nf)];
          }
          cout<<endl;
          char c;
          //cin>>c;
          //exit(0);
          y[indice3(g,k,l,mi,mj)]=temp;
        }
      }
    }
    return y;
  }

  void
  ConvolutionLayer::init_nabla(){
    for(size_t i=0;i<mf*nf;++i){
      nabla_b[i]=0;
    }
    for(size_t i=0;i<mf*nf*p*q;++i){
      nabla_K[i]=0;
    }
  }

  Vector
  ConvolutionLayer::back_propagation(Vector e){
    for(size_t f=0;f<nf;++f){
      for(size_t i=0;i<ni;++i){
        for(size_t j=0;j<nj;++j){
          Real temp=0;
          for(size_t g=0;g<mf;++g){
            size_t r=(i>=mi-1)?i-mi+1:0;
            for(;r<min(i,p);++r){
              size_t s=(j>=mj-1)?j-mj+1:0;
              for(;s<min(j,q);++s){
                temp+=K[indice4(g,f,r,s,nf,p,q)]*e[indice3(g,i-r,j-s,mi,mj)];
              }//s
            }//r
          }//g
          d[indice3(f,i,j,ni,nj)]=temp;
        }//j
      }//i
    }//f
    //display(delta,nf*ni*nj);
    //char a;cin>>a;
    //cout<<"  - Update nabla_b"<<endl;
    //Update nabla_b<<
    for(size_t g=0;g<mf;++g){
      for(size_t f=0;f<nf;++f){
        Real temp=0;
        for(size_t k=0;k<mi;++k){
          for(size_t l=0;l<mj;++l){
            temp+=e[indice3(g,k,l,mi,mj)];
          }//l
        }//k
        nabla_b[indice2(g,f,nf)]+=temp;
      }
    }
    //Update nabla_w
    for(size_t g=0;g<mf;++g){
      for(size_t f=0;f<nf;++f){
        for(size_t r=0;r<p;++r){
          for(size_t s=0;s<q;++s){
            Real temp=0;
            for(size_t k=0;k<mi;++k){
              for(size_t l=0;l<mj;++l){
                temp+=e[indice3(g,k,l,mi,mj)]*x[indice3(f,k+r,l+s,ni,nj)];
              }//l
            }//k
            nabla_K[indice4(g,f,r,s,nf,p,q)]+=temp;
          }
        }
      }
    }
    return d;
  }

  void
  ConvolutionLayer::update(Real eta){
    //Update b
    for(size_t i=0;i<mf*nf;++i){
      b[i]-=eta*nabla_b[i];
    }
    //Update K
    for(size_t i=0;i<mf*nf*p*q;++i){
      K[i]-=eta*nabla_K[i];
    }
  }
}