multipers 2.3.3b6__cp312-cp312-macosx_10_13_x86_64.whl

multipers/tensor/tensor.h

@@ -0,0 +1,672 @@
+#pragma once
+
+#include <cassert>
+#include <cstddef>
+#include <iostream>
+#include <numeric>
+#include <vector>
+
+// TODO : sparse version, current operator[] is already a hash.
+namespace tensor {
+
+template <typename indices_type>
+inline std::vector<indices_type> compute_backward_cumprod(const std::vector<indices_type> &resolution) {
+  constexpr bool verbose = false;
+  std::vector<indices_type> cum_prod_resolution_(resolution.size());
+  cum_prod_resolution_.back() = 1;
+  for (auto i = resolution.size() - 1; i > 0; i--) {
+    // std::cout << i << " " << cum_prod_resolution_.size() << std::endl;
+    cum_prod_resolution_[i - 1] = resolution[i] * cum_prod_resolution_[i];
+  }
+  if constexpr (verbose) {
+    std::cout << "Cum resolution ";
+    for (auto c : cum_prod_resolution_) std::cout << c << " ";
+    std::cout << std::endl;
+  }
+  return cum_prod_resolution_;
+}
+
+template <typename dtype, typename indices_type>
+class static_tensor_view {  // Python handles the construction - destruction of
+                            // the data,
+ public:
+  using sparse_type = std::pair<std::vector<std::vector<indices_type>>, std::vector<dtype>>;
+  static_tensor_view();
+
+  static_tensor_view(dtype *data_ptr, const std::vector<indices_type> &resolution)
+      : data_ptr_(data_ptr),
+        size_(resolution.size() == 0
+                  ? 0
+                  : std::accumulate(begin(resolution), end(resolution), 1, std::multiplies<indices_type>())),
+        resolution_(resolution)
+  // cum_prod_resolution_(compute_backward_cumprod(resolution))
+  {
+    // cum_prod_resolution_ = std::vector<std::size_t>(resolution.size());
+    // std::size_t last = 1;
+    // for (auto i = resolution.size() -1; i > 0; i--){
+    // 	last *=resolution[i];
+    // 	// std::cout << i << " " << cum_prod_resolution_.size() << std::endl;
+    // 	cum_prod_resolution_[resolution.size()-1 - i] = last;
+    // }
+    // cum_prod_resolution_.back() = 1;
+    cum_prod_resolution_ = std::move(compute_backward_cumprod(resolution));
+  };
+
+  // dtype[]& data_ref(){
+  // 	return *data_ptr;
+  // }
+  inline std::size_t size() const { return size_; }
+
+  inline bool empty() const { return size_ == 0; }
+
+  inline dtype &data_back() const { return *(data_ptr_ + size_ - 1); }
+
+  inline std::size_t ndim() const { return resolution_.size(); }
+
+  template <class oned_array_like = std::initializer_list<indices_type>>
+  inline dtype &operator[](const oned_array_like &coordinates) const {
+    const bool check = false;
+    dtype *data_index = data_ptr_;
+    /* 0; // max is 4*10^9, should be fine. just put an assert in python. */
+
+    if constexpr (check) {
+      if (coordinates.size() != resolution_.size()) {
+        auto it = coordinates.begin();
+        for (size_t i = 0u; i < coordinates.size(); i++) std::cerr << *(it++) << "/" << resolution_[i] << ", ";
+        std::cerr << ")" << std::endl;
+        throw std::invalid_argument("Invalid coordinate dimension.");
+      }
+      // for (auto [ci, cum_res, res] : std::views::zip(coordinates,
+      // cum_prod_resolution_, resolution_)){ // NIK Apple clang for
+      // (indices_type i : std::views::iota(0,coordinates.size())){
+      auto it = coordinates.begin();
+      for (size_t i = 0u; i < coordinates.size(); i++) {
+        auto &ci = *(it++);
+        auto cum_res = cum_prod_resolution_[i];
+        auto res = resolution_[i];
+        if (ci >= res) [[unlikely]] {
+          std::cerr << "Crash log. Coordinates : (";
+          auto it = coordinates.begin();
+          for (auto i = 0u; i < coordinates.size(); i++) std::cerr << *(it++) << "/" << resolution_[i] << ", ";
+          // for (auto [c, r] : std::views::zip(coordinates, resolution_))
+          // std::cerr << c << "/" << r << ", "; // NIK APPLE CLANG
+          std::cerr << ")" << std::endl;
+          throw std::invalid_argument("Illegal coordinates.");
+        }
+        data_index += ci * cum_res;
+      }
+      if (data_index >= this->size()) [[unlikely]] {
+        std::cerr << "Crash log. Coordinates : (";
+        auto it = coordinates.begin();
+        for (size_t i = 0u; i < coordinates.size(); i++) std::cerr << *(it++) << "/" << resolution_[i] << ", ";
+        std::cerr << ")" << std::endl;
+        throw std::invalid_argument("Internal error : asked data " + std::to_string(data_index) + "/" +
+                                    std::to_string(this->size()));
+      }
+      // std::cout << data_index << " " << this->size() << std::endl;
+      // std::cout << data_index << "/" << this->size() << std::endl;
+    } else {
+      // for (auto [ci, cum_res] : std::views::zip(coordinates,
+      // cum_prod_resolution_)){ // small so i'm not sure reduce can be
+      // efficient here // NIK Apple clang 	data_index += ci*cum_res;
+      // }
+
+      auto coord_ptr = coordinates.begin();
+      auto cum_res_ptr = cum_prod_resolution_.begin();
+      for (; coord_ptr != coordinates.end(); ++coord_ptr, ++cum_res_ptr) {
+        data_index += (*coord_ptr) * (*cum_res_ptr);
+      }
+    }
+    /* return *(data_ptr_ + data_index); */
+    return *data_index;
+  }
+
+  template <typename idx_type>
+  inline dtype &data_at_index(idx_type i) {
+    return *(data_ptr_ + i);
+  }
+
+  template <typename indice_type_like>
+  inline std::vector<indices_type> data_index_inverse(indice_type_like data_index,
+                                                      const std::vector<bool> &flip_axes = {}) const {
+    std::vector<indices_type> coordinates(resolution_.size());
+    int data_index_ = data_index;
+    for (int parameter = static_cast<int>(coordinates.size()) - 1; parameter >= 0; parameter--) {
+      auto [q, r] = std::div(data_index_, static_cast<int>(resolution_[parameter]));
+      if (static_cast<int>(flip_axes.size()) > parameter && flip_axes[parameter])
+        coordinates[parameter] = resolution_[parameter] - r;
+      else
+        coordinates[parameter] = r;
+      data_index_ = q;
+    }
+    return coordinates;
+  }
+
+  // friend std::ostream& operator<<(std::ostream& stream, const
+  // static_tensor_view<dtype,indices_type>& truc){
+  //     stream << "[";
+  //     for(indices_type i = 0; i < truc.size()-1; i++){
+  //         stream << *(truc.data_ptr_ + i) << ", ";
+  //     }
+  //     if(!truc.empty()) stream << truc.data_back();
+  //     stream << "]";
+  // 	stream << "\n resolution : ";
+  // 	for(indices_type i = 0; i < truc.resolution_.size(); i++){
+  //         stream << truc.resolution_[i] << ", ";
+  //     }
+  // 	stream << "\n cum resolution : ";
+  // 	for(indices_type i = 0; i < truc.cum_prod_resolution_.size(); i++){
+  //         stream << truc.cum_prod_resolution_[i] << ", ";
+  //     }
+  //     return stream;
+  // }
+
+  friend std::ostream &operator<<(std::ostream &stream, const static_tensor_view<dtype, indices_type> &truc) {
+    // constexpr bool verbose = false;
+    for (auto parameter = 0u; parameter < truc.ndim(); parameter++) stream << "[";
+    // iterate over data, update coordinates in a vector, and print if in free
+    // coords i.e. add one to last coord, modulo if greater, and propagate to
+    // the next
+    std::vector<indices_type> coordinates(truc.ndim());  /// 0,...,0
+    for (auto i = 0u; i < truc.size() - 1; i++) {
+      stream << truc.data_at(i);
+
+      // for (indices_type parameter =0; parameter < coordinates.size();
+      // parameter++){ 	stream << coordinates[parameter];
+      // }
+      // stream << "\n";
+      coordinates[0]++;
+      indices_type parameter = 0;
+      for (; parameter < static_cast<int>(coordinates.size()) - 1; ++parameter) {
+        if (coordinates[parameter] < truc.get_resolution()[parameter]) {
+          // stream << ", ";
+          // if (parameter == 1)
+          // 	stream << "\n";
+          break;
+        }
+        // for (indices_type i =0; i < parameter; i++)
+        // 	stream << ";";
+        // for (indices_type i =0; i < parameter+1; i++)
+        // 	stream << "]";
+        // stream << ", ";
+        // for (indices_type i =0; i < parameter; i++)
+        // 	stream << "[";
+        coordinates[parameter] = 0;  // 1 by 1 so should be fine not doing mods
+        coordinates[parameter + 1]++;
+      }
+      if (parameter == 1)
+        stream << "],\n [";
+      else {
+        for (indices_type i = 0; i < parameter; i++) stream << "]";
+        stream << ", ";
+        for (indices_type i = 0; i < parameter; i++) stream << "[";
+      }
+    }
+
+    stream << truc.data_back();
+    for (auto parameter = 0u; parameter < truc.ndim(); parameter++) stream << "]";
+    return stream;
+  }
+
+  // template<class
+  // twod_array_like=std::initializer_list<std::initializer_list<indices_type>>>
+  // static_tensor_view_view<dtype,indices_type> view(twod_array_like
+  // coordinates){ 	auto out = static_tensor_view_view(data_ptr_,
+  // resolution_); 	out.free_coordinates = coordinates; 	return out;
+  // }
+  inline const std::vector<indices_type> &get_resolution() const { return resolution_; }
+
+  inline const std::vector<indices_type> &get_cum_resolution() const { return cum_prod_resolution_; }
+
+  template <typename indice_type_like>
+  inline dtype &data_at(indice_type_like i) const {
+    return *(data_ptr_ + i);
+  }
+
+  void differentiate(indices_type axis);
+
+  inline sparse_type sparsify(const std::vector<bool> &flip_axes = {}, bool verbose = false) const {
+    std::vector<std::vector<indices_type>> coordinates;
+    std::vector<dtype> values;
+    // for (indices_type i = 0; i < static_cast<indices_type>(this->size());
+    // i++){
+    for (auto i = 0u; i < this->size(); i++) {
+      auto stuff = this->data_at(i);
+      if (stuff == 0) [[likely]]  // as this is sparse
+        continue;
+      coordinates.push_back(this->data_index_inverse(i, flip_axes));
+      values.push_back(stuff);
+    }
+    if (verbose) [[unlikely]] {
+      // for (auto [pt,w] : std::views::zip(coordinates, values)){ NIK apple
+      // clang
+      for (auto i = 0u; i < coordinates.size(); i++) {
+        for (const auto &v : coordinates[i]) std::cout << v << " ";
+        std::cout << "| " << values[i] << std::endl;
+      }
+    }
+    return {coordinates, values};
+  }
+
+  // template<class oned_array_like=std::initializer_list<indices_type>>
+  void _rec_add_cone(const std::vector<indices_type> &basepoint,
+                     dtype value,
+                     std::vector<indices_type> &coordinates,
+                     int _rec_parameter) const {
+    if (_rec_parameter < 0) {
+      (*this)[coordinates] += value;
+      return;
+    }
+    for (indices_type c = basepoint[_rec_parameter]; c < this->get_resolution()[_rec_parameter]; c++) {
+      coordinates[_rec_parameter] = c;
+      this->_rec_add_cone(basepoint, value, coordinates, _rec_parameter - 1);
+    }
+  }
+
+  inline void add_cone(const std::vector<indices_type> &basepoint, dtype value) const {
+    constexpr const bool check = false;
+    constexpr const bool verbose = false;
+    if constexpr (check) {
+      if (basepoint.size() != this->ndim()) throw std::logic_error("Invalid coordinate for cone");
+    }
+    if constexpr (verbose) {
+      std::cout << "Adding cone ";
+      for (auto b : basepoint) std::cout << b << " ,";
+      std::cout << std::endl;
+    }
+    std::vector<indices_type> temp_container(this->ndim());
+    this->_rec_add_cone(basepoint, value, temp_container, static_cast<int>(this->ndim()) - 1);
+  }
+
+  // template<class oned_array_like=std::initializer_list<indices_type>>
+  void _rec_add_cone_boundary(const std::vector<indices_type> &basepoint,
+                              dtype value,
+                              std::vector<indices_type> &coordinates,
+                              int _rec_parameter) const {
+    if (_rec_parameter < 0) {
+      (*this)[coordinates] += value;
+      return;
+    }
+
+    // for (auto c=basepoint[_rec_parameter];
+    // c<this->get_resolution()[_rec_parameter]; c++){
+    // 	coordinates[_rec_parameter] = c;
+    // 	this->_rec_add_cone(basepoint, value, coordinates, _rec_parameter-1);
+    // }
+
+    coordinates[_rec_parameter] = basepoint[_rec_parameter];
+    this->_rec_add_cone_boundary(std::vector<indices_type>(basepoint), value, coordinates, _rec_parameter - 1);
+
+    coordinates[_rec_parameter] = this->get_resolution()[_rec_parameter] - 1;
+    this->_rec_add_cone_boundary(basepoint, -value, coordinates, _rec_parameter - 1);
+  }
+
+  inline void add_cone_boundary(const std::vector<indices_type> &basepoint, dtype value) const {
+    const bool check = false;
+    if constexpr (check) {
+      if (basepoint.size() != this->ndim()) throw std::logic_error("Invalid coordinate for cone boundary");
+    }
+    std::vector<indices_type> temp_container(this->ndim());
+    this->_rec_add_cone_boundary(basepoint, value, temp_container, static_cast<int>(this->ndim()) - 1);
+  }
+
+ public:
+ private:
+  dtype *data_ptr_;
+  std::size_t size_;
+  std::vector<indices_type> resolution_;
+  std::vector<indices_type> cum_prod_resolution_;
+  // std::vector<std::vector<indices_types>> fixed_coordinates; // in child
+};
+
+template <typename dtype, typename indices_type>
+class static_tensor_view_view
+    : public static_tensor_view<dtype, indices_type> {  // i'm not sure this class is very efficient.
+ public:
+  using base = static_tensor_view<dtype, indices_type>;
+
+  static_tensor_view_view(dtype *data_ptr,
+                          const std::vector<indices_type> &resolution,
+                          const std::vector<std::vector<indices_type>> &free_coordinates,
+                          bool use_sparse = true)
+      : base(data_ptr, resolution),
+        resolution_view(this->compute_resolution(free_coordinates))
+  // free_coordinates(free_coordinates)
+  {
+    this->compute_ptrs(free_coordinates, use_sparse);
+  };
+
+  static_tensor_view_view(const static_tensor_view<dtype, indices_type> &parent,
+                          const std::vector<std::vector<indices_type>> &free_coordinates,
+                          bool use_sparse = true)
+      : base(parent),
+        resolution_view(this->compute_resolution(free_coordinates))
+  // free_coordinates(free_coordinates)
+  {
+    this->compute_ptrs(free_coordinates, use_sparse);
+  };
+
+  inline bool is_float(const std::vector<indices_type> &resolution) const {
+    indices_type dim = this->dimension();
+    for (indices_type i = 0; i < dim; i++)
+      if (resolution[i] > 1) return false;
+    return true;
+  }
+
+  inline bool is_float() const { return this->is_float(this->resolution_view); }
+
+  template <class oned_array_like = std::initializer_list<indices_type>>
+  inline bool is_in_view(const oned_array_like &coordinates,
+                         const std::vector<std::vector<indices_type>> &free_coordinates) {
+    assert(coordinates.size() == this->ndim());
+    auto it = coordinates.begin();
+    for (indices_type parameter = 0; parameter < static_cast<indices_type>(this->ndim()); ++parameter) {
+      const auto &x = *it;
+      it++;
+      for (auto stuff : free_coordinates[parameter]) {
+        if (stuff < x)
+          continue;
+        else if (stuff == x)
+          break;
+        else
+          return false;
+      }
+      if (x > free_coordinates[parameter].back()) return false;
+    }
+    return true;
+  }
+
+  std::size_t _size() const {  // for construction
+    std::size_t out = 1;
+    for (const auto &r : resolution_view) out *= r;
+    return out;
+  }
+
+  std::size_t size() const { return ptrs.size(); }
+
+  std::vector<indices_type> compute_resolution(const std::vector<std::vector<indices_type>> &free_coordinates) {
+    std::vector<indices_type> out(free_coordinates.size());
+    // for (auto [s, stuff] : std::views::zip(out, free_coordinates)) s =
+    // stuff.size(); // NIK apple clang
+    for (auto i = 0u; i < free_coordinates.size(); i++) out[i] = free_coordinates[i].size();
+    return out;
+  }
+
+  void compute_ptrs_dense(const std::vector<std::vector<indices_type>> &free_coordinates) {  // todo redo from
+    // DO NOT USE
+    constexpr bool verbose = false;
+    std::vector<dtype *> out(this->_size());
+    std::vector<indices_type> coordinates(this->ndim());  /// 0,...,0
+    std::size_t count = 0;
+
+    for (int i = 0; i < static_cast<int>(static_tensor_view<dtype, indices_type>::size()) - 1; i++) {
+      if constexpr (verbose) {
+        std::cout << "Coordinate : ";
+        for (auto x : coordinates) std::cout << x << " ";
+        if (this->is_in_view(coordinates, free_coordinates))
+          std::cout << " in view";
+        else
+          std::cout << "not in view";
+        std::cout << std::endl;
+      }
+
+      if (this->is_in_view(coordinates, free_coordinates)) {
+        out[count] = &this->data_at(i);
+        count++;
+      }
+      coordinates.back()++;
+      for (indices_type parameter = coordinates.size() - 1; parameter > 0; parameter--) {
+        if (coordinates[parameter] < this->get_resolution()[parameter]) {
+          break;
+        }
+        for (indices_type i = parameter; i < static_cast<indices_type>(coordinates.size()); i++)
+          coordinates[i] = 0;  // 1 by 1 so should be fine not doing mods
+        coordinates[parameter - 1]++;
+      }
+    }
+    if (this->is_in_view(coordinates, free_coordinates)) {
+      out[count] = &this->data_back();
+      count++;
+    }
+    // assert(count == this->size());
+    ptrs.swap(out);
+  }
+
+  inline void compute_ptrs_sparse(const std::vector<std::vector<indices_type>> &free_coordinates,
+                                  std::vector<indices_type> _rec_coordinates_begin = {}) {  // todo redo from
+    constexpr bool verbose = false;
+    if (_rec_coordinates_begin.size() == 0) ptrs.reserve(this->_size());
+    indices_type parameter = _rec_coordinates_begin.size();
+    if (parameter == static_cast<indices_type>(this->ndim())) {
+      auto &value = tensor::static_tensor_view<dtype, indices_type>::operator[](
+          _rec_coordinates_begin);  // calling [] is not efficient, but not
+                                    // bottleneck
+      if constexpr (verbose) {
+        std::cout << "Adding coordinates ";
+        for (auto c : _rec_coordinates_begin) std::cout << c << " ";
+        std::cout << " of value " << value;
+        std::cout << std::endl;
+      }
+      ptrs.push_back(&value);
+      return;
+    }
+    _rec_coordinates_begin.reserve(this->ndim());
+    _rec_coordinates_begin.resize(parameter + 1);
+    for (indices_type coord : free_coordinates[parameter]) {
+      _rec_coordinates_begin.back() = coord;
+      compute_ptrs_sparse(free_coordinates, _rec_coordinates_begin);
+    }
+    return;
+  }
+
+  inline void compute_ptrs(const std::vector<std::vector<indices_type>> &free_coordinates, bool use_sparse = true) {
+    if (use_sparse)
+      compute_ptrs_sparse(free_coordinates);
+    else
+      compute_ptrs_dense(free_coordinates);
+  }
+
+  inline void shift_coordinate(indices_type idx, indices_type shift_value) {
+    // resolution stays the same,
+    auto to_add = this->get_cum_resolution()[idx] * shift_value;
+    for (auto &ptr : this->ptrs) ptr += to_add;
+  }
+
+  // constant additions
+  inline void operator+=(dtype x) {
+    // if (ptrs.empty()) this->compute_ptrs_dense();
+    for (auto stuff : ptrs) *stuff += x;
+    return;
+  }
+
+  inline void operator-=(dtype x) {
+    // if (ptrs.empty()) this->compute_ptrs_dense();
+    for (auto stuff : ptrs) *stuff -= x;
+    return;
+  }
+
+  inline void operator*=(dtype x) {
+    // if (ptrs.empty()) this->compute_ptrs_dense();
+    for (auto stuff : ptrs) *stuff *= x;
+    return;
+  }
+
+  inline void operator/=(dtype x) {
+    // if (ptrs.empty()) this->compute_ptrs_dense();
+    for (auto stuff : ptrs) *stuff /= x;
+    return;
+  }
+
+  inline void operator=(dtype x) {
+    for (auto stuff : ptrs) *stuff = x;
+    return;
+  }
+
+  inline void operator=(const static_tensor_view_view<dtype, indices_type> &x) {
+    assert(this->size() == x.size());
+    this->ptrs = x.ptrs;
+    return;
+  }
+
+  inline void swap(static_tensor_view_view<dtype, indices_type> &x) {
+    this->ptrs.swap(x.ptrs);
+    return;
+  }
+
+  // retrieves data from ptrs
+  inline void operator+=(const static_tensor_view_view<dtype, indices_type> &x) {
+    std::size_t num_data = this->size();
+    assert(num_data == x.size());
+    for (auto idx = 0u; idx < num_data; idx++) *ptrs[idx] += *x[idx];
+    return;
+  }
+
+  inline void operator-=(const static_tensor_view_view<dtype, indices_type> &x) {
+    std::size_t num_data = this->size();
+    assert(num_data == x.size());
+    for (auto idx = 0u; idx < num_data; idx++) *ptrs[idx] -= *x[idx];
+    return;
+  }
+
+  inline void operator*=(const static_tensor_view_view<dtype, indices_type> &x) {
+    std::size_t num_data = this->size();
+    assert(num_data == x.size());
+    for (auto idx = 0u; idx < num_data; idx++) *ptrs[idx] *= *x[idx];
+    return;
+  }
+
+  inline void operator/=(const static_tensor_view_view<dtype, indices_type> &x) {
+    std::size_t num_data = this->size();
+    assert(num_data == x.size());
+    for (auto idx = 0u; idx < num_data; idx++) *ptrs[idx] /= *x[idx];
+    return;
+  }
+
+  // Default array_like template
+  template <typename array_like = std::initializer_list<dtype>>
+  inline void operator+=(const array_like &x) {
+    std::size_t num_data = this->size();
+    assert(num_data == x.size());
+    for (auto idx = 0u; idx < num_data; idx++) *ptrs[idx] += *(x.begin() + idx);
+    return;
+  }
+
+  template <typename array_like = std::initializer_list<dtype>>
+  inline void operator-=(const array_like &x) {
+    std::size_t num_data = this->size();
+    assert(num_data == x.size());
+    for (auto idx = 0u; idx < num_data; idx++) *ptrs[idx] -= *(x.begin() + idx);
+    return;
+  }
+
+  template <typename array_like = std::initializer_list<dtype>>
+  inline void operator*=(const array_like &x) {
+    std::size_t num_data = this->size();
+    assert(num_data == x.size());
+    for (auto idx = 0u; idx < num_data; idx++) *ptrs[idx] *= *(x.begin() + idx);
+    return;
+  }
+
+  template <typename array_like = std::initializer_list<dtype>>
+  inline void operator/=(const array_like &x) {
+    std::size_t num_data = this->size();
+    assert(num_data == x.size());
+    for (auto idx = 0u; idx < num_data; idx++) *ptrs[idx] /= *(x.begin() + idx);
+    return;
+  }
+
+  // void compute_cum_res(){
+  // 	if (cum_resolution_view.size() == 0){
+
+  // 		cum_resolution_view =
+  // compute_backward_cumprod(this->resolution_view);
+  // 	}
+  // }
+  template <typename T = std::initializer_list<indices_type>>
+  inline dtype &operator[]([[maybe_unused]] T coords) {
+    throw std::logic_error("Not yet implemented");
+    // this->compute_cum_res();
+    // assert(this->cum_resolution_view.size() == coords.size());
+    // std::size_t data_index = 0;
+    // // for (indices_type i = 0, auto coords_it = coords.begin(); coords_it !=
+    // coords.end(); coords_it++, i++)
+    // // 	{data_index += (*(coords_it))*cum_resolution_view[i];};
+    // for (auto [c, cr] : std::views::zip(coords, cum_resolution_view))
+    // 	data_index += c*cr;
+    // std::cout << ptrs.size() << " vs " << data_index << std::endl;
+    // return *ptrs[data_index];
+  }
+
+  void print_data() const {
+    std::cout << "[";
+    for (auto stuff : ptrs) std::cout << *stuff << " ";
+    std::cout << "]\n";
+  }
+
+  inline std::vector<dtype> copy_data() {
+    std::vector<dtype> out(ptrs.size());
+    for (auto i = 0u; i < ptrs.size(); i++) out[i] = *ptrs[i];
+    return out;
+  }
+
+ public:
+  // juste besoin de la resolution, avec les ptrs : ok pour l'affichage
+  // const std::vector<std::vector<indices_type>> free_coordinates; // for each
+  // parameter, the fixed indices, TODO:REMOVE
+  const std::vector<indices_type> resolution_view;
+
+ private:
+  std::vector<dtype *> ptrs;
+  // std::vector<std::size_t> cum_resolution_view; // not really useful.
+};
+
+template <typename dtype, typename indices_type>
+void inline static_tensor_view<dtype, indices_type>::differentiate(indices_type axis) {
+  std::vector<std::vector<indices_type>> free_coordinates(this->ndim());
+
+  // initialize free_coordinates of the view, full coordinates on each axis
+  // exept for axis on which we iterate
+  for (auto i = 0u; i < free_coordinates.size(); i++) {
+    if (static_cast<indices_type>(i) == axis) continue;
+    free_coordinates[i] = std::vector<indices_type>(this->get_resolution()[i]);
+    for (auto j = 0u; j < free_coordinates[i].size(); j++) {  // TODO optimize
+      free_coordinates[i][j] = j;
+    }
+  }
+  // iterate over coordinate of this axis with ab -> b-a -> ab=b[newslice]
+  free_coordinates[axis] = {{0}};
+  static_tensor_view_view<dtype, indices_type> x_i(*this, free_coordinates);
+  std::vector<dtype> a, b;
+  a = x_i.copy_data();
+  for (indices_type h = 1; h < this->get_resolution()[axis]; h++) {
+    free_coordinates[axis] = {{h}};
+    // x_i = static_tensor_view_view<dtype,
+    // indices_type>(*this,free_coordinates);
+    x_i.shift_coordinate(axis, 1);
+    b = std::move(x_i.copy_data());
+    x_i -= a;
+    a.swap(b);
+  }
+}
+
+template <typename T>
+std::vector<std::vector<T>> cart_product(const std::vector<std::vector<T>> &v) {
+  std::vector<std::vector<T>> s = {{}};
+  for (const auto &u : v) {
+    std::vector<std::vector<T>> r;
+    for (const auto &x : s) {
+      for (const auto y : u) {
+        r.push_back(x);
+        r.back().push_back(y);
+      }
+    }
+    s = std::move(r);
+  }
+  for (const auto &truc : s) {
+    for (const auto &machin : truc) std::cout << machin << ", ";
+    std::cout << "\n";
+  }
+  return s;
+}
+
+}  // namespace tensor

multipers/tensor.pxd

@@ -0,0 +1,13 @@
+from libc.stdint cimport uint16_t
+from libcpp.vector cimport vector
+from libcpp cimport bool, float
+
+
+ctypedef float dtype
+ctypedef uint16_t index_type
+
+cdef extern from "tensor/tensor.h" namespace "tensor":
+	cdef cppclass static_tensor_view[float, uint16_t]:
+		static_tensor_view()   except + nogil
+		static_tensor_view(dtype*,const  vector[index_type]&) except + nogil
+		const vector[index_type]& get_resolution()