#ifndef LAYER_CONV_SHARED_INC_CPP

#define LAYER_CONV_SHARED_INC_CPP

#include "layer.conv.inc.cpp"

template<typename iter=""></typename>

void iterateTestConvolutionShared(Layout cl, Layout pl, Kernel k, Neuron *c_neurons, Neuron *p_neurons, Weight *weights) {

  if (!cl) return;

  assert(pl);

  assert(k);

  assert(c_neurons);

  assert(p_neurons);

  assert(weights);

  assert(pl.x0 + k.ox >= 0 && pl.x0 + (cl.getW()-1)*k.dx + k.ox + k.sx <= pl.sx);

  assert(pl.y0 + k.oy >= 0 && pl.y0 + (cl.getH()-1)*k.dy + k.oy + k.sy <= pl.sy);

  for(int cy = cl.y0; cy < cl.y1; ++cy)

  for(int cx = cl.x0; cx < cl.x1; ++cx)

  for(int cz = cl.z0; cz < cl.z1; ++cz) {

    int ci = (cy*cl.sx + cx)*cl.sz + cz;

    Neuron &cn = c_neurons[ci];

    typename Iter::AccumType a = {};

    Iter::init(cn, a);

    for(int ky = 0; ky < k.sy; ++ky)

    for(int kx = 0; kx < k.sx; ++kx)

    for(int pz = pl.z0; pz < pl.z1; ++pz) {

      int wi = (((cz - cl.z0)*k.sy + ky)*k.sx + kx)*pl.getD() + pz - pl.z0;

      Weight &w = weights[wi];

      int px = pl.x0 + (cx - cl.x0)*k.dx + k.ox + kx;

      int py = pl.y0 + (cy - cl.y0)*k.dy + k.oy + ky;

      int pi = (py*pl.sx + px)*pl.sz + pz;

      Neuron &pn = p_neurons[pi];

      Iter::iter(pn, w, a);

    }

    Iter::done(cn, a);

  }

}

template<typename iter=""></typename>

void iterateConvolutionShared(Layout cl, Layout pl, Layout wl, Kernel k, Neuron *c_neurons, Neuron *p_neurons, Weight *weights) {

  if (!cl) return;

  assert(pl);

  assert(wl);

  assert(k);

  assert(c_neurons);

  assert(p_neurons);

  assert(weights);

  assert(cl.isSubLayoutOf(wl));

  assert(pl.x0 + k.ox >= 0 && pl.x0 + (wl.getW()-1)*k.dx + k.ox + k.sx <= pl.sx);

  assert(pl.y0 + k.oy >= 0 && pl.y0 + (wl.getH()-1)*k.dy + k.oy + k.sy <= pl.sy);

  int c_h    = cl.getH();

  int c_w    = cl.getW();

  int c_d    = cl.getD();

  int c_swz  = c_w*cl.sz;

  int c_shxz = c_h*cl.sx*cl.sz;

  int c_dx   = cl.sz - c_d;

  int c_dy   = (cl.sx - c_w)*cl.sz;

  int p_d    = pl.getD();

  int p_dx   = k.dx*pl.sz;

  int p_dy   = k.dy*pl.sx*pl.sz - c_w*p_dx;

  int k_sxd  = k.sx*p_d;

  int k_syxd = k.sy*k_sxd;

  int p_ddy  = (pl.sx - k.sx)*pl.sz;

  int p_ddx  = pl.sz - p_d;

  Neuron *icn = c_neurons + (cl.y0*cl.sx + cl.x0)*cl.sz + cl.z0;

  Neuron *ipn = p_neurons + ((pl.y0 + (cl.y0 - wl.y0)*k.dy + k.oy)*pl.sx + pl.x0 + (cl.x0 - wl.x0)*k.dx + k.ox)*pl.sz + pl.z0;

  weights += (cl.z0 - wl.z0)*k_syxd;

  Weight *iw = weights;

  for(Neuron *e = icn + c_shxz; icn < e; icn += c_dy, ipn += p_dy)

  for(Neuron *e = icn +  c_swz; icn < e; icn += c_dx, ipn += p_dx, iw = weights)

  for(Neuron *e = icn +    c_d; icn < e; ++icn) {

    typename Iter::AccumType a;

    Iter::init(*icn, a);

    Neuron *iipn = ipn;

    for(Weight *e = iw + k_syxd; iw < e; iipn += p_ddy)

    for(Weight *e = iw +  k_sxd; iw < e; iipn += p_ddx)

    for(Weight *e = iw +    p_d; iw < e; ++iw, ++iipn)

      Iter::iter(*iipn, *iw, a);

    Iter::done(*icn, a);

  }

}

template<typename iter=""></typename>

void iterateConvolutionSharedPoint(Layout cl, Layout pl, Layout wl, Kernel k, int kx, int ky, Neuron *c_neurons, Neuron *p_neurons, Weight *weights) {

  if (!cl) return;

  assert(pl);

  assert(wl);

  assert(k);

  assert(c_neurons);

  assert(p_neurons);

  assert(weights);

  assert(cl.isSubLayoutOf(wl));

  assert(kx >= 0 && kx < k.sx);

  assert(ky >= 0 && ky < k.sy);

  assert(pl.x0 + k.ox >= 0 && pl.x0 + (wl.getW()-1)*k.dx + k.ox + k.sx <= pl.sx);

  assert(pl.y0 + k.oy >= 0 && pl.y0 + (wl.getH()-1)*k.dy + k.oy + k.sy <= pl.sy);

  int c_h    = cl.getH();

  int c_w    = cl.getW();

  int c_d    = cl.getD();

  int c_swz  = c_w*cl.sz;

  int c_shxz = c_h*cl.sx*cl.sz;

  int c_dx   = cl.sz - c_d;

  int c_dy   = (cl.sx - c_w)*cl.sz;

  int p_d    = pl.getD();

  int p_dx   = k.dx*pl.sz;

  int p_dy   = k.dy*pl.sx*pl.sz - c_w*p_dx;

  int w_dz   = (k.sy*k.sx - 1)*p_d;

  Neuron *icn = c_neurons + (cl.y0*cl.sx + cl.x0)*cl.sz + cl.z0;

  Neuron *ipn = p_neurons + ((pl.y0 + (cl.y0 - wl.y0)*k.dy + k.oy + ky)*pl.sx + pl.x0 + (cl.x0 - wl.x0)*k.dx + k.ox + kx)*pl.sz + pl.z0;

  weights += (((cl.z0 - wl.z0)*k.sy + ky)*k.sx + kx)*p_d;

  Weight *iw = weights;

  for(Neuron *e = icn + c_shxz; icn < e; icn += c_dy, ipn += p_dy)

  for(Neuron *e = icn +  c_swz; icn < e; icn += c_dx, ipn += p_dx, iw = weights)

  for(Neuron *e = icn +    c_d; icn < e; ++icn,       ipn -= p_d,  iw += w_dz)

  for(Weight *e = iw  +    p_d; iw  < e; ++ipn, ++iw)

    Iter::iter2(*icn, *ipn, *iw);

}

void fillConvolutionWeights(int kx, int ky, int kz, int count, Weight *weights) {

  double kr = 1.5;

  double sum = 0;

  Weight *iw = weights;

  for(int i = 0; i < count; ++i)

  for(int y = 0; y < ky; ++y)

  for(int x = 0; x < kx; ++x)

  for(int z = 0; z < kz; ++z, ++iw) {

    double dx = (2.0*x/(kx-1) - 1)*kr;

    double dy = (2.0*y/(ky-1) - 1)*kr;

    double e = exp( -dx*dx - dy*dy );

    sum += e;

    iw->w = (WeightReal)( (rand()/(double)RAND_MAX*2 - 1)*e );

    //iw->w = (WeightReal)( rand()/(double)RAND_MAX*e );

  }

  WeightReal k = (WeightReal)(10*kz/sum);

  Weight *ew = iw;

  for(iw = weights; iw < ew; ++iw) iw->w *= k;

}

class LayerConvSharedBase: public Layer {

public:

  Kernel kernel;

  std::vector<weight> mtWeights;</weight>

  LayerConvSharedBase(Layer &prev, const Layout &layout, const Kernel &kernel, Weight *weights = nullptr):

    Layer(&prev, layout, kernel.sx * kernel.sy * layout.getD() * prev.back().layout.getD(), weights),

    kernel(kernel)

  {

    assert(kernel);

    stat.links = weightsCount * layout.getW() * layout.getH();

    if (ownWeights) fillWeights(-1, 1);

  }

  void split(int threadsCount) override {

    Layer::split(threadsCount);

    Weight w = {};

    mtWeights.clear();

    mtWeights.resize(threadsCount*weightsCount, w);

  }

  inline void sumWeights(int tid, int threads) {

    int wc = weightsCount;

    Weight *iw = weights + tid;

    Weight *ia = mtWeights.data() + tid;

    Weight *ea = mtWeights.data() + threads*wc;

    for(Weight *ew = weights + wc; iw < ew; iw += threads, ia += threads) {

      WeightReal w = iw->w;

      for(Weight *iia = ia; iia < ea; iia += wc)

        w += iia->w, iia->w = 0;

      iw->w = w;

    }

  }

};

template<func func=""></func>

class LayerConvShared: public LayerConvSharedBase {

public:

  LayerConvShared(Layer &prev, const Layout &layout, const Kernel &kernel, Weight *weights = nullptr):

    LayerConvSharedBase(prev, layout, kernel, weights)

  {

    stat.links = weightsCount * layout.getW() * layout.getH();

    if (ownWeights) fillConvolutionWeights(kernel.sx, kernel.sy, this->prev->layout.getD(), layout.getD(), this->weights);

  }

  void pass(Barrier &barrier) override {

    struct I: public Iter {

      static inline void init(Neuron&, AccumType &a) { a.v = 0; }

      static inline void iter(Neuron &n, Weight &w, AccumType &a) { a.v += n.v * w.w; }

      static inline void done(Neuron &n, AccumType &a) { func(n, a.v); }

    };

    iterateConvolutionShared(mtLayouts[barrier.tid], prev->layout, layout, kernel, neurons, prev->neurons, weights);

  }

  void backpassWeights(Barrier &barrier) override {

    struct I: public Iter {

      static inline void init(Neuron &n, AccumType &a) { a.v = n.d; }

      static inline void iter(Neuron &n, Weight &w, AccumType &a) { w.w += n.v * a.v; }

    };

    iterateConvolutionShared(mtLayouts[barrier.tid], prev->layout, layout, kernel, neurons, prev->neurons, &mtWeights[barrier.tid * weightsCount]);

    barrier.wait();

    sumWeights(barrier.tid, barrier.threads);

  }

  void backpassDeltas(Barrier &barrier) override {

    struct I: public Iter {

      static inline void iter2(Neuron &cn, Neuron &pn, Weight &w) { pn.a.v += cn.d * w.w; }

      static inline void iter3(Neuron &n) { n.d *= n.a.v; n.a.v = 0; }

    };

    int ksx = kernel.sx, ksy = kernel.sy;

    for(int kx = 0; kx < ksx; ++kx)

    for(int ky = 0; ky < ksy; ++ky) {

      iterateConvolutionSharedPoint(mtLayouts[barrier.tid], prev->layout, layout, kernel, kx, ky, neurons, prev->neurons, weights);

      barrier.wait();

    }

    iterateNeurons(mtPrevLayouts[barrier.tid], prev->neurons);

  }

  void testPass() override {

    struct I: public Iter {

      static inline void init(Neuron&, AccumType &a) { a.v = 0; }

      static inline void iter(Neuron &n, Weight &w, AccumType &a) { a.v += n.v * w.w; }

      static inline void done(Neuron &n, AccumType &a) { func(n, a.v); }

    };

    iterateTestConvolutionShared(layout, prev->layout, kernel, neurons, prev->neurons, weights);

  }

  void testBackpass() override {

    struct I: public Iter {

      static inline void init(Neuron &n, AccumType &a) { a.v = n.d; }

      static inline void iter(Neuron &n, Weight &w, AccumType &a) { n.a.v += a.v * w.w; }

      static inline void iter3(Neuron &n) { n.d *= n.a.v; n.a.v = 0; }

    };

    struct IW: public Iter {

      static inline void init(Neuron &n, AccumType &a) { a.v = n.d; }

      static inline void iter(Neuron &n, Weight &w, AccumType &a) { w.w += a.v * n.v; }

    };

    clearAccum();

    iterateTestConvolutionShared(layout, prev->layout, kernel, neurons, prev->neurons, weights);

    iterateTestConvolutionShared<iw>(layout, prev->layout, kernel, neurons, prev->neurons, weights);</iw>

    iterateNeurons(prev->layout, prev->neurons);

    clearAccum();

  }

  bool saveDemo() override

    { return !filename || saveConvDemoImage( filename, layout.getD(), kernel.sx, kernel.sy, prev->layout.getD(), weights ); }

};

template<func func=""></func>

class LayerDeconvShared: public LayerConvSharedBase {

public:

  LayerDeconvShared(Layer &prev, const Layout &layout, const Kernel &kernel, Weight *weights = nullptr):

    LayerConvSharedBase(prev, layout, kernel, weights)

  {

    stat.links = weightsCount * this->prev->layout.getW() * this->prev->layout.getH();

    if (ownWeights) fillConvolutionWeights(kernel.sx, kernel.sy, layout.getD(), this->prev->layout.getD(), this->weights);

  }

  void pass(Barrier &barrier) override {

    struct I: public Iter {

      static inline void iter2(Neuron &cn, Neuron &pn, Weight &w) { pn.a.v += cn.v * w.w; }

      static inline void iter3(Neuron &n) { func(n, n.a.v); n.a.v = 0; }

    };

    int k_sx = kernel.sx, k_sy = kernel.sy;

    for(int kx = 0; kx < k_sx; ++kx)

    for(int ky = 0; ky < k_sy; ++ky) {

      iterateConvolutionSharedPoint(mtPrevLayouts[barrier.tid], layout, prev->layout, kernel, kx, ky, prev->neurons, neurons, weights);

      barrier.wait();

    }

    iterateNeurons(mtLayouts[barrier.tid], neurons);

  }

  void backpassWeights(Barrier &barrier) override {

    struct I: public Iter {

      static inline void init(Neuron &n, AccumType &a) { a.v = n.v; }

      static inline void iter(Neuron &n, Weight &w, AccumType &a) { w.w += n.d * a.v; }

    };

    iterateConvolutionShared(mtPrevLayouts[barrier.tid], layout, prev->layout, kernel, prev->neurons, neurons, &mtWeights[barrier.tid * weightsCount]);

    barrier.wait();

    sumWeights(barrier.tid, barrier.threads);

  }

  void backpassDeltas(Barrier &barrier) override {

    struct I: public Iter {

      static inline void init(Neuron&, AccumType &a) { a.v = 0; }

      static inline void iter(Neuron &n, Weight &w, AccumType &a) { a.v += n.d * w.w; }

      static inline void done(Neuron &n, AccumType &a) { n.d *= a.v; }

    };

    iterateConvolutionShared(mtPrevLayouts[barrier.tid], layout, prev->layout, kernel, prev->neurons, neurons, weights);

  }

  void testPass() override {

    struct I: public Iter {

      static inline void init(Neuron &n, AccumType &a) { a.v = n.v; }

      static inline void iter(Neuron &n, Weight &w, AccumType &a) { n.a.v += a.v * w.w; }

      static inline void iter3(Neuron &n) { func(n, n.a.v); n.a.v = 0; }

    };

    clearAccum();

    iterateTestConvolutionShared(prev->layout, layout, kernel, prev->neurons, neurons, weights);

    iterateNeurons(layout, neurons);

    clearAccum();

  }

  void testBackpass() override {

    struct I: public Iter {

      static inline void init(Neuron &n, AccumType &a) { a.v = 0; }

      static inline void iter(Neuron &n, Weight &w, AccumType &a) { a.v += n.d * w.w; }

      static inline void done(Neuron &n, AccumType &a) { n.d *= a.v; }

    };

    struct IW: public Iter {

      static inline void init(Neuron &n, AccumType &a) { a.v = n.v; }

      static inline void iter(Neuron &n, Weight &w, AccumType &a) { w.w += n.d * a.v; }

    };

    iterateTestConvolutionShared(prev->layout, layout, kernel, prev->neurons, neurons, weights);

    iterateTestConvolutionShared<iw>(prev->layout, layout, kernel, prev->neurons, neurons, weights);</iw>

  }

  bool saveDemo() override

    { return !filename || saveConvDemoImage( filename, prev->layout.getD(), kernel.sx, kernel.sy, layout.getD(), weights ); }

};

#endif