basic_array.hpp

Go to the documentation of this file.

// Copyright (c) 2018--present, The Simons Foundation
// This file is part of TRIQS/nda and is licensed under the Apache License, Version 2.0.
// SPDX-License-Identifier: Apache-2.0
// See LICENSE in the root of this distribution for details.

 
#pragma once
 
#include "./accessors.hpp"
#include "./basic_array_view.hpp"
#include "./basic_functions.hpp"
#include "./concepts.hpp"
#include "./iterators.hpp"
#include "./layout/for_each.hpp"
#include "./layout/permutation.hpp"
#include "./layout/range.hpp"
#include "./layout/slice_static.hpp"
#include "./layout_transforms.hpp"
#include "./macros.hpp"
#include "./matrix_functions.hpp"
#include "./mem/address_space.hpp"
#include "./mem/fill.hpp"
#include "./mem/memcpy.hpp"
#include "./mem/policies.hpp"
#include "./stdutil/array.hpp"
#include "./traits.hpp"
 
#include <algorithm>
#include <array>
#include <complex>
#include <concepts>
#include <exception>
#include <initializer_list>
#include <iostream>
#include <random>
#include <ranges>
#include <type_traits>
#include <utility>
 
namespace nda {

  template <typename ValueType, int Rank, typename LayoutPolicy, char Algebra, typename ContainerPolicy>
  class basic_array {
    // Compile-time checks.
    static_assert(!std::is_const_v<ValueType>, "Error in nda::basic_array: ValueType cannot be const");
    static_assert((Algebra != 'N'), "Internal error in nda::basic_array: Algebra 'N' not supported");
    static_assert((Algebra != 'M') or (Rank == 2), "Internal error in nda::basic_array: Algebra 'M' requires a rank 2 array");
    static_assert((Algebra != 'V') or (Rank == 1), "Internal error in nda::basic_array: Algebra 'V' requires a rank 1 array");
 
    public:
    using value_type = ValueType;

    using layout_policy_t = LayoutPolicy;

    using layout_t = typename LayoutPolicy::template mapping<Rank>;

    using container_policy_t = ContainerPolicy;

    using storage_t = typename ContainerPolicy::template handle<ValueType>;

    using regular_type = basic_array;

    static constexpr int rank = Rank;
 
    // Compile-time check.
    static_assert(has_contiguous(layout_t::layout_prop), "Error in nda::basic_array: Memory layout has to be contiguous");
 
    private:
    // Type of the array itself.
    using self_t = basic_array;
 
    // Type of the accessor policy for views (no no_alias_accessor).
    using AccessorPolicy = default_accessor;
 
    // Type of the owning policy for views.
    using OwningPolicy = borrowed<storage_t::address_space>;
 
    // Constexpr variable that is true if the value_type is const (never for basic_array).
    static constexpr bool is_const = false;
 
    // Constexpr variable that is true if the array is a view (never for basic_array).
    static constexpr bool is_view = false;
 
    // Memory layout of the array, i.e. the nda::idx_map.
    layout_t lay;
 
    // Memory handle of the array.
    storage_t sto;
 
    // Construct an array with a given shape and initialize the memory with zeros.
    template <std::integral Int = long>
    basic_array(std::array<Int, Rank> const &shape, mem::init_zero_t) : lay{shape}, sto{lay.size(), mem::init_zero} {}
 
    public:
    auto as_array_view() { return basic_array_view<ValueType, Rank, LayoutPolicy, 'A', AccessorPolicy, OwningPolicy>{*this}; };

    auto as_array_view() const { return basic_array_view<const ValueType, Rank, LayoutPolicy, 'A', AccessorPolicy, OwningPolicy>{*this}; };

    [[deprecated]] auto transpose()
      requires(Rank == 2)
    {
      return permuted_indices_view<encode(std::array<int, 2>{1, 0})>(*this);
    }

    [[deprecated]] auto transpose() const
      requires(Rank == 2)
    {
      return permuted_indices_view<encode(std::array<int, 2>{1, 0})>(*this);
    }

    basic_array() {}; // NOLINT (user-defined constructor to avoid value initialization of the sso buffer)

    basic_array(basic_array &&) = default;

    explicit basic_array(basic_array const &a) = default;

    template <char A, typename CP>
    explicit basic_array(basic_array<ValueType, Rank, LayoutPolicy, A, CP> a) noexcept : lay(a.indexmap()), sto(std::move(a.storage())) {}

    template <std::integral... Ints>
      requires(sizeof...(Ints) == Rank)
    explicit basic_array(Ints... is) {
      // setting the layout and storage in the constructor body improves error messages for wrong # of args
      lay = layout_t{std::array{long(is)...}}; // NOLINT (for better error messages)
      sto = storage_t{lay.size()};             // NOLINT (for better error messages)
    }

    template <std::integral Int, typename RHS>
    explicit basic_array(Int sz, RHS const &val)
      requires((Rank == 1 and is_scalar_for_v<RHS, basic_array>))
       : lay(layout_t{std::array{long(sz)}}), sto{lay.size()} {
      assign_from_scalar(val);
    }

    template <std::integral Int = long>
    explicit basic_array(std::array<Int, Rank> const &shape)
      requires(std::is_default_constructible_v<ValueType>)
       : lay(shape), sto(lay.size()) {}

    explicit basic_array(layout_t const &layout)
      requires(std::is_default_constructible_v<ValueType>)
       : lay{layout}, sto{lay.size()} {}

    explicit basic_array(layout_t const &layout, storage_t &&storage) noexcept : lay{layout}, sto{std::move(storage)} {}

    template <ArrayOfRank<Rank> A>
      requires(HasValueTypeConstructibleFrom<A, ValueType>)
    basic_array(A const &a) : lay(a.shape()), sto{lay.size(), mem::do_not_initialize} {
      static_assert(std::is_constructible_v<ValueType, get_value_t<A>>, "Error in nda::basic_array: Incompatible value types in constructor");
      if constexpr (std::is_trivial_v<ValueType> or is_complex_v<ValueType>) {
        // trivial and complex value types can use the optimized assign_from_ndarray
        if constexpr (std::is_same_v<ValueType, get_value_t<A>>)
          assign_from_ndarray(a);
        else
          assign_from_ndarray(nda::map([](auto const &val) { return ValueType(val); })(a));
      } else {
        // general value types may not be default constructible -> use placement new
        nda::for_each(lay.lengths(), [&](auto const &...is) { new (sto.data() + lay(is...)) ValueType{a(is...)}; });
      }
    }

    template <ArrayInitializer<basic_array> Initializer> // can not be explicit
    basic_array(Initializer const &initializer) : basic_array{initializer.shape()} {
      initializer.invoke(*this);
    }
 
    private:
    // Get the corresponding shape from an initializer list.
    static std::array<long, 1> shape_from_init_list(std::initializer_list<ValueType> const &l) noexcept { return {long(l.size())}; }
 
    // Get the corresponding shape from a nested initializer list.
    template <typename L>
    static auto shape_from_init_list(std::initializer_list<L> const &l) noexcept {
      const auto [min, max] = std::minmax_element(std::begin(l), std::end(l), [](auto &&x, auto &&y) { return x.size() < y.size(); });
      EXPECTS_WITH_MESSAGE(min->size() == max->size(),
                           "Error in nda::basic_array: Arrays can only be initialized with rectangular initializer lists");
      return stdutil::front_append(shape_from_init_list(*max), long(l.size()));
    }
 
    public:
    basic_array(std::initializer_list<ValueType> const &l)
      requires(Rank == 1)
       : lay(std::array<long, 1>{long(l.size())}), sto{lay.size(), mem::do_not_initialize} {
      long i = 0;
      for (auto const &x : l) { new (sto.data() + lay(i++)) ValueType{x}; }
    }

    basic_array(std::initializer_list<std::initializer_list<ValueType>> const &l2)
      requires(Rank == 2)
       : lay(shape_from_init_list(l2)), sto{lay.size(), mem::do_not_initialize} {
      long i = 0, j = 0;
      for (auto const &l1 : l2) {
        for (auto const &x : l1) { new (sto.data() + lay(i, j++)) ValueType{x}; }
        j = 0;
        ++i;
      }
    }

    basic_array(std::initializer_list<std::initializer_list<std::initializer_list<ValueType>>> const &l3)
      requires(Rank == 3)
       : lay(shape_from_init_list(l3)), sto{lay.size(), mem::do_not_initialize} {
      long i = 0, j = 0, k = 0;
      for (auto const &l2 : l3) {
        for (auto const &l1 : l2) {
          for (auto const &x : l1) { new (sto.data() + lay(i, j, k++)) ValueType{x}; }
          k = 0;
          ++j;
        }
        j = 0;
        ++i;
      }
    }

    template <char A2>
    explicit basic_array(basic_array<ValueType, 2, LayoutPolicy, A2, ContainerPolicy> &&a) noexcept
      requires(Rank == 2)
       : basic_array{a.indexmap(), std::move(a).storage()} {}

    template <std::integral Int = long>
    static basic_array zeros(std::array<Int, Rank> const &shape)
      requires(std::is_standard_layout_v<ValueType> && std::is_trivially_copyable_v<ValueType>)
    {
      return basic_array{stdutil::make_std_array<long>(shape), mem::init_zero};
    }

    template <std::integral... Ints>
    static basic_array zeros(Ints... is)
      requires(sizeof...(Ints) == Rank)
    {
      return zeros(std::array<long, Rank>{is...});
    }

    template <std::integral Int = long>
    static basic_array ones(std::array<Int, Rank> const &shape)
      requires(nda::is_scalar_v<ValueType>)
    {
      auto res = basic_array{stdutil::make_std_array<long>(shape)};
      res()    = ValueType{1};
      return res;
    }

    template <std::integral... Ints>
    static basic_array ones(Ints... is)
      requires(sizeof...(Ints) == Rank)
    {
      return ones(std::array<long, Rank>{is...});
    }

    template <std::integral Int = long>
    static basic_array rand(std::array<Int, Rank> const &shape)
      requires(std::is_floating_point_v<ValueType> or nda::is_complex_v<ValueType>)
    {
      auto static gen = std::mt19937_64{};
      auto res        = basic_array{shape};
      if constexpr (nda::is_complex_v<ValueType>) {
        auto static dist = std::uniform_real_distribution<typename ValueType::value_type>(0.0, 1.0);
        using namespace std::complex_literals;
        for (auto &x : res) x = dist(gen) + 1i * dist(gen);
      } else {
        auto static dist = std::uniform_real_distribution<ValueType>(0.0, 1.0);
        for (auto &x : res) x = dist(gen);
      }
      return res;
    }

    template <std::integral... Ints>
    static basic_array rand(Ints... is)
      requires(sizeof...(Ints) == Rank)
    {
      return rand(std::array<long, Rank>{is...});
    }

    basic_array &operator=(basic_array &&) = default;

    basic_array &operator=(basic_array const &) = default;

    template <char A, typename CP>
    basic_array &operator=(basic_array<ValueType, Rank, LayoutPolicy, A, CP> const &rhs) {
      *this = basic_array{rhs};
      return *this;
    }

    template <ArrayOfRank<Rank> RHS>
    basic_array &operator=(RHS const &rhs) {
      resize(rhs.shape());
      assign_from_ndarray(rhs);
      return *this;
    }

    template <typename RHS>
    basic_array &operator=(RHS const &rhs) noexcept
      requires(is_scalar_for_v<RHS, basic_array>)
    {
      assign_from_scalar(rhs);
      return *this;
    }

    template <ArrayInitializer<basic_array> Initializer>
    basic_array &operator=(Initializer const &initializer) {
      resize(initializer.shape());
      initializer.invoke(*this);
      return *this;
    }

    template <std::integral... Ints>
    void resize(Ints const &...is) {
      static_assert(std::is_copy_constructible_v<ValueType>, "Error in nda::basic_array: Resizing requires the value_type to be copy constructible");
      static_assert(sizeof...(is) == Rank, "Error in nda::basic_array: Resizing requires exactly Rank arguments");
      resize(std::array<long, Rank>{long(is)...});
    }

    [[gnu::noinline]] void resize(std::array<long, Rank> const &shape) {
      lay = layout_t(shape);
      if (sto.is_null() or (sto.size() != lay.size())) sto = storage_t{lay.size()};
    }
 
// include common functionality of arrays and views
// Copyright (c) 2019--present, The Simons Foundation
// This file is part of TRIQS/nda and is licensed under the Apache License, Version 2.0.
// SPDX-License-Identifier: Apache-2.0
// See LICENSE in the root of this distribution for details.

[[nodiscard]] constexpr auto const &indexmap() const noexcept { return lay; }

[[nodiscard]] storage_t const &storage() const & noexcept { return sto; }

[[nodiscard]] storage_t &storage() & noexcept { return sto; }

[[nodiscard]] storage_t storage() && noexcept { return std::move(sto); }

[[nodiscard]] constexpr auto stride_order() const noexcept { return lay.stride_order; }

[[nodiscard]] ValueType const *data() const noexcept { return sto.data(); }

[[nodiscard]] ValueType *data() noexcept { return sto.data(); }

[[nodiscard]] auto const &shape() const noexcept { return lay.lengths(); }

[[nodiscard]] auto const &strides() const noexcept { return lay.strides(); }

[[nodiscard]] long size() const noexcept { return lay.size(); }

[[nodiscard]] long is_contiguous() const noexcept { return lay.is_contiguous(); }

[[nodiscard]] long has_positive_strides() const noexcept { return lay.has_positive_strides(); }

[[nodiscard]] bool empty() const { return sto.is_null(); }

[[nodiscard]] bool is_empty() const noexcept { return sto.is_null(); }

[[nodiscard]] long extent(int i) const noexcept {
#ifdef NDA_ENFORCE_BOUNDCHECK
  if (i < 0 || i >= rank) {
    std::cerr << "Error in extent: Dimension " << i << " is incompatible with array of rank " << rank << std::endl;
    std::terminate();
  }
#endif
  return lay.lengths()[i];
}

[[nodiscard]] long shape(int i) const noexcept { return extent(i); }

[[nodiscard]] auto indices() const noexcept { return itertools::product_range(shape()); }

static constexpr bool is_stride_order_C() noexcept { return layout_t::is_stride_order_C(); }

static constexpr bool is_stride_order_Fortran() noexcept { return layout_t::is_stride_order_Fortran(); }

decltype(auto) operator()(_linear_index_t idx) const noexcept {
  if constexpr (layout_t::layout_prop == layout_prop_e::strided_1d)
    return sto[idx.value * lay.min_stride()];
  else if constexpr (layout_t::layout_prop == layout_prop_e::contiguous)
    return sto[idx.value];
  else
    static_assert(always_false<layout_t>, "Internal error in array/view: Calling this type with a _linear_index_t is not allowed");
}

decltype(auto) operator()(_linear_index_t idx) noexcept {
  if constexpr (layout_t::layout_prop == layout_prop_e::strided_1d)
    return sto[idx.value * lay.min_stride()];
  else if constexpr (layout_t::layout_prop == layout_prop_e::contiguous)
    return sto[idx.value];
  else
    static_assert(always_false<layout_t>, "Internal error in array/view: Calling this type with a _linear_index_t is not allowed");
}
 
private:
// Constexpr variable that is true if bounds checking is disabled.
#ifdef NDA_ENFORCE_BOUNDCHECK
static constexpr bool has_no_boundcheck = false;
#else
static constexpr bool has_no_boundcheck = true;
#endif
 
public:
template <char ResultAlgebra, bool SelfIsRvalue, typename Self, typename... Ts>
FORCEINLINE static decltype(auto) call(Self &&self, Ts const &...idxs) noexcept(has_no_boundcheck) {
  // resulting value type
  using r_v_t = std::conditional_t<std::is_const_v<std::remove_reference_t<Self>>, ValueType const, ValueType>;
 
  // behavior depends on the given arguments
  if constexpr (clef::is_any_lazy<Ts...>) {
    // if there are lazy arguments, e.g. as in A(i_) << i_, a lazy expression is returned
    return clef::make_expr_call(std::forward<Self>(self), idxs...);
  } else if constexpr (sizeof...(Ts) == 0) {
    // if no arguments are given, a full view is returned
    return basic_array_view<r_v_t, Rank, LayoutPolicy, Algebra, AccessorPolicy, OwningPolicy>{self.lay, self.sto};
  } else {
    // otherwise we check the arguments and either access a single element or make a slice
    static_assert(((layout_t::template argument_is_allowed_for_call_or_slice<Ts> + ...) > 0),
                  "Error in array/view: Slice arguments must be convertible to range, ellipsis, or long (or string if the layout permits it)");
 
    // number of arguments convertible to long
    static constexpr int n_args_long = (layout_t::template argument_is_allowed_for_call<Ts> + ...);
 
    if constexpr (n_args_long == rank) {
      // access a single element
      long offset = self.lay(idxs...);
      if constexpr (is_view or not SelfIsRvalue) {
        // if the calling object is a view or an lvalue, we return a reference
        return AccessorPolicy::template accessor<r_v_t>::access(self.sto.data(), offset);
      } else {
        // otherwise, we return a copy of the value
        return ValueType{self.sto[offset]};
      }
    } else {
      // access a slice of the view/array
      auto const [offset, idxm]      = self.lay.slice(idxs...);
      static constexpr auto res_rank = decltype(idxm)::rank();
      // resulting algebra
      static constexpr char newAlgebra = (ResultAlgebra == 'M' and (res_rank == 1) ? 'V' : ResultAlgebra);
      // resulting layout policy
      using r_layout_p = typename detail::layout_to_policy<std::decay_t<decltype(idxm)>>::type;
      return basic_array_view<r_v_t, res_rank, r_layout_p, newAlgebra, AccessorPolicy, OwningPolicy>{std::move(idxm), {self.sto, offset}};
    }
  }
}
 
public:
template <typename... Ts>
FORCEINLINE decltype(auto) operator()(Ts const &...idxs) const & noexcept(has_no_boundcheck) {
  static_assert((rank == -1) or (sizeof...(Ts) == rank) or (sizeof...(Ts) == 0) or (ellipsis_is_present<Ts...> and (sizeof...(Ts) <= rank + 1)),
                "Error in array/view: Incorrect number of parameters in call operator");
  return call<Algebra, false>(*this, idxs...);
}

template <typename... Ts>
FORCEINLINE decltype(auto) operator()(Ts const &...idxs) & noexcept(has_no_boundcheck) {
  static_assert((rank == -1) or (sizeof...(Ts) == rank) or (sizeof...(Ts) == 0) or (ellipsis_is_present<Ts...> and (sizeof...(Ts) <= rank + 1)),
                "Error in array/view: Incorrect number of parameters in call operator");
  return call<Algebra, false>(*this, idxs...);
}

template <typename... Ts>
FORCEINLINE decltype(auto) operator()(Ts const &...idxs) && noexcept(has_no_boundcheck) {
  static_assert((rank == -1) or (sizeof...(Ts) == rank) or (sizeof...(Ts) == 0) or (ellipsis_is_present<Ts...> and (sizeof...(Ts) <= rank + 1)),
                "Error in array/view: Incorrect number of parameters in call operator");
  return call<Algebra, true>(*this, idxs...);
}

template <typename T>
decltype(auto) operator[](T const &idx) const & noexcept(has_no_boundcheck) {
  static_assert((rank == 1), "Error in array/view: Subscript operator is only available for rank 1 views/arrays in C++17/20");
  return call<Algebra, false>(*this, idx);
}

template <typename T>
decltype(auto) operator[](T const &x) & noexcept(has_no_boundcheck) {
  static_assert((rank == 1), "Error in array/view: Subscript operator is only available for rank 1 views/arrays in C++17/20");
  return call<Algebra, false>(*this, x);
}

template <typename T>
decltype(auto) operator[](T const &x) && noexcept(has_no_boundcheck) {
  static_assert((rank == 1), "Error in array/view: Subscript operator is only available for rank 1 views/arrays in C++17/20");
  return call<Algebra, true>(*this, x);
}

static constexpr int iterator_rank = (has_strided_1d(layout_t::layout_prop) ? 1 : Rank);

using const_iterator = array_iterator<iterator_rank, ValueType const, typename AccessorPolicy::template accessor<ValueType>::pointer>;

using iterator = array_iterator<iterator_rank, ValueType, typename AccessorPolicy::template accessor<ValueType>::pointer>;
 
private:
// Make an iterator for the view/array depending on its type.
template <typename Iterator>
[[nodiscard]] auto make_iterator(bool at_end) const noexcept {
  if constexpr (iterator_rank == Rank) {
    // multi-dimensional iterator
    if constexpr (layout_t::is_stride_order_C()) {
      // C-order case (array_iterator already traverses the data in C-order)
      return Iterator{indexmap().lengths(), indexmap().strides(), sto.data(), at_end};
    } else {
      // general case (we need to permute the shape and the strides according to the stride order of the layout)
      return Iterator{nda::permutations::apply(layout_t::stride_order, indexmap().lengths()),
                      nda::permutations::apply(layout_t::stride_order, indexmap().strides()), sto.data(), at_end};
    }
  } else {
    // 1-dimensional iterator
    return Iterator{std::array<long, 1>{size()}, std::array<long, 1>{indexmap().min_stride()}, sto.data(), at_end};
  }
}
 
public:
[[nodiscard]] const_iterator begin() const noexcept { return make_iterator<const_iterator>(false); }

[[nodiscard]] const_iterator cbegin() const noexcept { return make_iterator<const_iterator>(false); }

iterator begin() noexcept { return make_iterator<iterator>(false); }

[[nodiscard]] const_iterator end() const noexcept { return make_iterator<const_iterator>(true); }

[[nodiscard]] const_iterator cend() const noexcept { return make_iterator<const_iterator>(true); }

iterator end() noexcept { return make_iterator<iterator>(true); }

template <typename RHS>
auto &operator+=(RHS const &rhs) noexcept {
  static_assert(not is_const, "Error in array/view: Can not assign to a const view");
  return operator=(*this + rhs);
}

template <typename RHS>
auto &operator-=(RHS const &rhs) noexcept {
  static_assert(not is_const, "Error in array/view: Can not assign to a const view");
  return operator=(*this - rhs);
}

template <typename RHS>
auto &operator*=(RHS const &rhs) noexcept {
  static_assert(not is_const, "Error in array/view: Can not assign to a const view");
  return operator=((*this) * rhs);
}

template <typename RHS>
auto &operator/=(RHS const &rhs) noexcept {
  static_assert(not is_const, "Error in array/view: Can not assign to a const view");
  return operator=(*this / rhs);
}

template <std::ranges::contiguous_range R>
auto &operator=(R const &rhs) noexcept
  requires(Rank == 1 and not MemoryArray<R> and not is_scalar_for_v<R, self_t>)
{
  *this = array_const_view<std::ranges::range_value_t<R>, 1>{rhs};
  return *this;
}
 
private:
// Implementation of the assignment from an n-dimensional array type.
template <typename RHS>
void assign_from_ndarray(RHS const &rhs) {
#ifdef NDA_ENFORCE_BOUNDCHECK
  if (this->shape() != rhs.shape()) {
    NDA_RUNTIME_ERROR << "Error in assign_from_ndarray: Size mismatch:"
                      << "\n LHS.shape() = " << this->shape() << "\n RHS.shape() = " << rhs.shape();
  }
#endif
  // compile-time check if assignment is possible
  static_assert(std::is_assignable_v<value_type &, get_value_t<RHS>>, "Error in assign_from_ndarray: Incompatible value types");
 
  // are both operands nda::MemoryArray types?
  static constexpr bool both_in_memory = MemoryArray<self_t> and MemoryArray<RHS>;
 
  // do both operands have the same stride order?
  static constexpr bool same_stride_order = get_layout_info<self_t>.stride_order == get_layout_info<RHS>.stride_order;
 
  // compile-time check for device arrays to avoid runtime errors
  static_assert(!(mem::on_device<self_t> or mem::on_device<RHS>) or (both_in_memory and same_stride_order and have_same_value_type_v<self_t, RHS>),
                "Error in assign_from_ndarray: Assignment to/from device arrays is not supported for the given types.");
 
  // prefer optimized options if possible
  if constexpr (both_in_memory and same_stride_order) {
    if (rhs.empty()) return;
    // are both operands strided in 1d?
    static constexpr bool both_1d_strided = has_layout_strided_1d<self_t> and has_layout_strided_1d<RHS>;
    if constexpr (mem::on_host<self_t, RHS> and both_1d_strided) {
      // vectorizable copy on host
      for (long i = 0; i < size(); ++i) (*this)(_linear_index_t{i}) = rhs(_linear_index_t{i});
      return;
    } else if constexpr (!mem::on_host<self_t, RHS> and have_same_value_type_v<self_t, RHS>) {
      // check for block-layout and use mem::memcpy2D if possible
      auto bl_layout_dst = get_block_layout(*this);
      auto bl_layout_src = get_block_layout(rhs);
      if (bl_layout_dst && bl_layout_src) {
        auto [n_bl_dst, bl_size_dst, bl_str_dst] = *bl_layout_dst;
        auto [n_bl_src, bl_size_src, bl_str_src] = *bl_layout_src;
        // check that the total memory size is the same
        if (n_bl_dst * bl_size_dst != n_bl_src * bl_size_src) NDA_RUNTIME_ERROR << "Error in assign_from_ndarray: Incompatible block sizes";
        // if either destination or source consists of a single block, we can chunk it up to make the layouts compatible
        if (n_bl_dst == 1 && n_bl_src > 1) {
          n_bl_dst = n_bl_src;
          bl_size_dst /= n_bl_src;
          bl_str_dst = bl_size_dst;
        }
        if (n_bl_src == 1 && n_bl_dst > 1) {
          n_bl_src = n_bl_dst;
          bl_size_src /= n_bl_dst;
          bl_str_src = bl_size_src;
        }
        // copy only if block-layouts are compatible, otherwise continue to fallback
        if (n_bl_dst == n_bl_src && bl_size_dst == bl_size_src) {
          mem::memcpy2D<mem::get_addr_space<self_t>, mem::get_addr_space<RHS>>((void *)data(), bl_str_dst * sizeof(value_type), (void *)rhs.data(),
                                                                               bl_str_src * sizeof(value_type), bl_size_src * sizeof(value_type),
                                                                               n_bl_src);
          return;
        }
      }
    }
  }
  // otherwise fallback to elementwise assignment
  if constexpr (mem::on_device<self_t> || mem::on_device<RHS>) {
    NDA_RUNTIME_ERROR << "Error in assign_from_ndarray: Fallback to elementwise assignment not implemented for arrays/views on the GPU";
  }
  nda::for_each(shape(), [this, &rhs](auto const &...args) { (*this)(args...) = rhs(args...); });
}
 
// Implementation to fill a view/array with a constant scalar value.
template <typename Scalar>
void fill_with_scalar(Scalar const &scalar) {
  if constexpr (mem::on_host<self_t>) {
    if constexpr (has_layout_strided_1d<self_t>) {
      const long L             = size();
      auto *__restrict const p = data(); // no alias possible here!
      if constexpr (has_contiguous_layout<self_t>) {
        for (long i = 0; i < L; ++i) p[i] = scalar;
      } else {
        const long stri  = indexmap().min_stride();
        const long Lstri = L * stri;
        for (long i = 0; i != Lstri; i += stri) p[i] = scalar;
      }
    } else {
      for (auto &x : *this) x = scalar;
    }
  } else if constexpr (mem::on_device<self_t> or mem::on_unified<self_t>) {
    if constexpr (has_layout_strided_1d<self_t>) {
      if constexpr (has_contiguous_layout<self_t>) {
        mem::fill_n<mem::get_addr_space<self_t>>(data(), size(), value_type(scalar));
      } else {
        const long stri = indexmap().min_stride();
        mem::fill2D_n<mem::get_addr_space<self_t>>(data(), stri, 1, size(), value_type(scalar));
      }
    } else {
      auto bl_layout = get_block_layout(*this);
      if (bl_layout) {
        auto [n_bl, bl_size, bl_str] = *bl_layout;
        mem::fill2D_n<mem::get_addr_space<self_t>>(data(), bl_str, bl_size, n_bl, value_type(scalar));
      } else {
        // MAM: implement recursive call to fill_with_scalar on (i,nda::ellipsis{})
        NDA_RUNTIME_ERROR << "fill_with_scalar: Not implemented on device for generic (non-blocked) layout.";
      }
    }
  }
}
 
// Implementation of the assignment from a scalar value.
template <typename Scalar>
void assign_from_scalar(Scalar const &scalar) noexcept {
  static_assert(!is_const, "Error in assign_from_ndarray: Cannot assign to a const view");
  if constexpr (Algebra != 'M') {
    // element-wise assignment for non-matrix algebras
    fill_with_scalar(scalar);
  } else {
    // a scalar has to be interpreted as a unit matrix for matrix algebras (the scalar in the shortest diagonal)
    // FIXME : A priori faster to put 0 everywhere and then change the diag to avoid the if.
    // FIXME : Benchmark and confirm.
    if constexpr (is_scalar_or_convertible_v<Scalar>)
      fill_with_scalar(0);
    else
      fill_with_scalar(Scalar{0 * scalar}); // FIXME : improve this
    diagonal(*this).fill_with_scalar(scalar);
  }
}
  };
 
  // Class template argument deduction guides.
  template <MemoryArray A>
  basic_array(A &&a) -> basic_array<get_value_t<A>, get_rank<A>, get_contiguous_layout_policy<get_rank<A>, get_layout_info<A>.stride_order>,
                                    get_algebra<A>, heap<mem::get_addr_space<A>>>;
 
  template <Array A>
  basic_array(A &&a) -> basic_array<get_value_t<A>, get_rank<A>, C_layout, get_algebra<A>, heap<mem::get_addr_space<A>>>;
 
} // namespace nda