// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
// SPDX-License-Identifier: MPL-2.0

#include "main.h"

#include <Eigen/Tensor>

using Eigen::array;
using Eigen::Tensor;

template <int DataLayout>
static void test_simple_shuffling() {
  Tensor<float, 4, DataLayout> tensor(2, 3, 5, 7);
  tensor.setRandom();
  array<ptrdiff_t, 4> shuffles;
  shuffles[0] = 0;
  shuffles[1] = 1;
  shuffles[2] = 2;
  shuffles[3] = 3;

  Tensor<float, 4, DataLayout> no_shuffle;
  no_shuffle = tensor.shuffle(shuffles);

  VERIFY_IS_EQUAL(no_shuffle.dimension(0), 2);
  VERIFY_IS_EQUAL(no_shuffle.dimension(1), 3);
  VERIFY_IS_EQUAL(no_shuffle.dimension(2), 5);
  VERIFY_IS_EQUAL(no_shuffle.dimension(3), 7);

  for (int i = 0; i < 2; ++i) {
    for (int j = 0; j < 3; ++j) {
      for (int k = 0; k < 5; ++k) {
        for (int l = 0; l < 7; ++l) {
          VERIFY_IS_EQUAL(tensor(i, j, k, l), no_shuffle(i, j, k, l));
        }
      }
    }
  }

  shuffles[0] = 2;
  shuffles[1] = 3;
  shuffles[2] = 1;
  shuffles[3] = 0;
  Tensor<float, 4, DataLayout> shuffle;
  shuffle = tensor.shuffle(shuffles);

  VERIFY_IS_EQUAL(shuffle.dimension(0), 5);
  VERIFY_IS_EQUAL(shuffle.dimension(1), 7);
  VERIFY_IS_EQUAL(shuffle.dimension(2), 3);
  VERIFY_IS_EQUAL(shuffle.dimension(3), 2);

  for (int i = 0; i < 2; ++i) {
    for (int j = 0; j < 3; ++j) {
      for (int k = 0; k < 5; ++k) {
        for (int l = 0; l < 7; ++l) {
          VERIFY_IS_EQUAL(tensor(i, j, k, l), shuffle(k, l, j, i));
        }
      }
    }
  }
}

template <int DataLayout>
static void test_expr_shuffling() {
  Tensor<float, 4, DataLayout> tensor(2, 3, 5, 7);
  tensor.setRandom();

  array<ptrdiff_t, 4> shuffles;
  shuffles[0] = 2;
  shuffles[1] = 3;
  shuffles[2] = 1;
  shuffles[3] = 0;
  Tensor<float, 4, DataLayout> expected;
  expected = tensor.shuffle(shuffles);

  Tensor<float, 4, DataLayout> result(5, 7, 3, 2);

  array<ptrdiff_t, 4> src_slice_dim{{2, 3, 1, 7}};
  array<ptrdiff_t, 4> src_slice_start{{0, 0, 0, 0}};
  array<ptrdiff_t, 4> dst_slice_dim{{1, 7, 3, 2}};
  array<ptrdiff_t, 4> dst_slice_start{{0, 0, 0, 0}};

  for (int i = 0; i < 5; ++i) {
    result.slice(dst_slice_start, dst_slice_dim) = tensor.slice(src_slice_start, src_slice_dim).shuffle(shuffles);
    src_slice_start[2] += 1;
    dst_slice_start[0] += 1;
  }

  VERIFY_IS_EQUAL(result.dimension(0), 5);
  VERIFY_IS_EQUAL(result.dimension(1), 7);
  VERIFY_IS_EQUAL(result.dimension(2), 3);
  VERIFY_IS_EQUAL(result.dimension(3), 2);

  for (int i = 0; i < expected.dimension(0); ++i) {
    for (int j = 0; j < expected.dimension(1); ++j) {
      for (int k = 0; k < expected.dimension(2); ++k) {
        for (int l = 0; l < expected.dimension(3); ++l) {
          VERIFY_IS_EQUAL(result(i, j, k, l), expected(i, j, k, l));
        }
      }
    }
  }

  dst_slice_start[0] = 0;
  result.setRandom();
  for (int i = 0; i < 5; ++i) {
    result.slice(dst_slice_start, dst_slice_dim) = tensor.shuffle(shuffles).slice(dst_slice_start, dst_slice_dim);
    dst_slice_start[0] += 1;
  }

  for (int i = 0; i < expected.dimension(0); ++i) {
    for (int j = 0; j < expected.dimension(1); ++j) {
      for (int k = 0; k < expected.dimension(2); ++k) {
        for (int l = 0; l < expected.dimension(3); ++l) {
          VERIFY_IS_EQUAL(result(i, j, k, l), expected(i, j, k, l));
        }
      }
    }
  }
}

template <int DataLayout>
static void test_shuffling_as_value() {
  Tensor<float, 4, DataLayout> tensor(2, 3, 5, 7);
  tensor.setRandom();
  array<ptrdiff_t, 4> shuffles;
  shuffles[2] = 0;
  shuffles[3] = 1;
  shuffles[1] = 2;
  shuffles[0] = 3;
  Tensor<float, 4, DataLayout> shuffle(5, 7, 3, 2);
  shuffle.shuffle(shuffles) = tensor;

  VERIFY_IS_EQUAL(shuffle.dimension(0), 5);
  VERIFY_IS_EQUAL(shuffle.dimension(1), 7);
  VERIFY_IS_EQUAL(shuffle.dimension(2), 3);
  VERIFY_IS_EQUAL(shuffle.dimension(3), 2);

  for (int i = 0; i < 2; ++i) {
    for (int j = 0; j < 3; ++j) {
      for (int k = 0; k < 5; ++k) {
        for (int l = 0; l < 7; ++l) {
          VERIFY_IS_EQUAL(tensor(i, j, k, l), shuffle(k, l, j, i));
        }
      }
    }
  }

  array<ptrdiff_t, 4> no_shuffle;
  no_shuffle[0] = 0;
  no_shuffle[1] = 1;
  no_shuffle[2] = 2;
  no_shuffle[3] = 3;
  Tensor<float, 4, DataLayout> shuffle2(5, 7, 3, 2);
  shuffle2.shuffle(shuffles) = tensor.shuffle(no_shuffle);
  for (int i = 0; i < 5; ++i) {
    for (int j = 0; j < 7; ++j) {
      for (int k = 0; k < 3; ++k) {
        for (int l = 0; l < 2; ++l) {
          VERIFY_IS_EQUAL(shuffle2(i, j, k, l), shuffle(i, j, k, l));
        }
      }
    }
  }
}

template <int DataLayout>
static void test_shuffle_unshuffle() {
  Tensor<float, 4, DataLayout> tensor(2, 3, 5, 7);
  tensor.setRandom();

  // Choose a random permutation.
  array<ptrdiff_t, 4> shuffles;
  for (int i = 0; i < 4; ++i) {
    shuffles[i] = i;
  }
  array<ptrdiff_t, 4> shuffles_inverse;
  for (int i = 0; i < 4; ++i) {
    const ptrdiff_t index = internal::random<ptrdiff_t>(i, 3);
    shuffles_inverse[shuffles[index]] = i;
    std::swap(shuffles[i], shuffles[index]);
  }

  Tensor<float, 4, DataLayout> shuffle;
  shuffle = tensor.shuffle(shuffles).shuffle(shuffles_inverse);

  VERIFY_IS_EQUAL(shuffle.dimension(0), 2);
  VERIFY_IS_EQUAL(shuffle.dimension(1), 3);
  VERIFY_IS_EQUAL(shuffle.dimension(2), 5);
  VERIFY_IS_EQUAL(shuffle.dimension(3), 7);

  for (int i = 0; i < 2; ++i) {
    for (int j = 0; j < 3; ++j) {
      for (int k = 0; k < 5; ++k) {
        for (int l = 0; l < 7; ++l) {
          VERIFY_IS_EQUAL(tensor(i, j, k, l), shuffle(i, j, k, l));
        }
      }
    }
  }
}

template <int DataLayout>
static void test_empty_shuffling() {
  Tensor<float, 4, DataLayout> tensor(2, 3, 0, 7);
  tensor.setRandom();
  array<ptrdiff_t, 4> shuffles;
  shuffles[0] = 0;
  shuffles[1] = 1;
  shuffles[2] = 2;
  shuffles[3] = 3;

  Tensor<float, 4, DataLayout> no_shuffle;
  no_shuffle = tensor.shuffle(shuffles);

  VERIFY_IS_EQUAL(no_shuffle.dimension(0), 2);
  VERIFY_IS_EQUAL(no_shuffle.dimension(1), 3);
  VERIFY_IS_EQUAL(no_shuffle.dimension(2), 0);
  VERIFY_IS_EQUAL(no_shuffle.dimension(3), 7);

  for (int i = 0; i < 2; ++i) {
    for (int j = 0; j < 3; ++j) {
      for (int k = 0; k < 0; ++k) {
        for (int l = 0; l < 7; ++l) {
          VERIFY_IS_EQUAL(tensor(i, j, k, l), no_shuffle(i, j, k, l));
        }
      }
    }
  }

  shuffles[0] = 2;
  shuffles[1] = 3;
  shuffles[2] = 1;
  shuffles[3] = 0;
  Tensor<float, 4, DataLayout> shuffle;
  shuffle = tensor.shuffle(shuffles);

  VERIFY_IS_EQUAL(shuffle.dimension(0), 0);
  VERIFY_IS_EQUAL(shuffle.dimension(1), 7);
  VERIFY_IS_EQUAL(shuffle.dimension(2), 3);
  VERIFY_IS_EQUAL(shuffle.dimension(3), 2);

  for (int i = 0; i < 2; ++i) {
    for (int j = 0; j < 3; ++j) {
      for (int k = 0; k < 0; ++k) {
        for (int l = 0; l < 7; ++l) {
          VERIFY_IS_EQUAL(tensor(i, j, k, l), shuffle(k, l, j, i));
        }
      }
    }
  }
}

// Regression: a scalar-changing consumer (here `.cast<double>()`) sits above
// shuffle in the assign. Before TensorConversionOp::block dropped the
// forwarded destination on a non-degenerate cast, shuffle's prepareStorage
// would reuse a double-sized buffer as int storage (assert in debug,
// corruption in release). Mirrors test_concatenation_through_cast.
template <int DataLayout>
static void test_shuffling_through_cast() {
  Tensor<int, 2, DataLayout> src(2, 3);
  for (int j = 0; j < 3; ++j) {
    for (int i = 0; i < 2; ++i) src(i, j) = i + 1 + 10 * j;
  }
  array<ptrdiff_t, 2> shuffles{1, 0};
  Tensor<double, 2, DataLayout> out(3, 2);
  out = src.shuffle(shuffles).template cast<double>();
  for (int j = 0; j < 3; ++j) {
    for (int i = 0; i < 2; ++i) {
      VERIFY_IS_APPROX(out(j, i), static_cast<double>(src(i, j)));
    }
  }
}

template <typename T, int DataLayout>
static void test_shuffling_packet_paths() {
  // Sizes with partial-packet tails; expression-sourced reads so the packet
  // path (not the block path) is exercised, sweeping permutations that
  // preserve and that shuffle the inner dimension.
  Tensor<T, 3, DataLayout> tensor(17, 5, 7);
  tensor.setRandom();

  const ptrdiff_t perms[][3] = {{0, 1, 2}, {0, 2, 1}, {1, 0, 2}, {2, 1, 0}, {1, 2, 0}, {2, 0, 1}};
  for (const auto& p : perms) {
    array<ptrdiff_t, 3> shuffle{{p[0], p[1], p[2]}};

    // Expression-sourced: the argument has no raw buffer, so the shuffle
    // cannot use its block path and must serve packets.
    Tensor<T, 3, DataLayout> result = (tensor * tensor.constant(T(2))).shuffle(shuffle);
    for (Index i = 0; i < result.dimension(0); ++i) {
      for (Index j = 0; j < result.dimension(1); ++j) {
        for (Index k = 0; k < result.dimension(2); ++k) {
          array<Index, 3> out_idx{{i, j, k}};
          array<Index, 3> in_idx;
          in_idx[p[0]] = i;
          in_idx[p[1]] = j;
          in_idx[p[2]] = k;
          VERIFY_IS_EQUAL(result(out_idx), T(2) * tensor(in_idx[0], in_idx[1], in_idx[2]));
        }
      }
    }

    // Lvalue: writePacket through the shuffled destination.
    Tensor<T, 3, DataLayout> dst(17, 5, 7);
    dst.setZero();
    array<Index, 3> dst_dims{{17, 5, 7}};
    Tensor<T, 3, DataLayout> src(dst_dims[p[0]], dst_dims[p[1]], dst_dims[p[2]]);
    src.setRandom();
    array<Index, 3> no_strides{{1, 1, 1}};
    // Unit striding keeps packet access but disables block evaluation.
    dst.shuffle(shuffle) = src.stride(no_strides);
    for (Index i = 0; i < src.dimension(0); ++i) {
      for (Index j = 0; j < src.dimension(1); ++j) {
        for (Index k = 0; k < src.dimension(2); ++k) {
          array<Index, 3> in_idx;
          in_idx[p[0]] = i;
          in_idx[p[1]] = j;
          in_idx[p[2]] = k;
          VERIFY_IS_EQUAL(src(i, j, k), dst(in_idx[0], in_idx[1], in_idx[2]));
        }
      }
    }
  }
}

template <typename T, int DataLayout>
static void test_identity_shuffling_without_impl_packet_access() {
  Tensor<T, 3, DataLayout> tensor(17, 5, 7);
  tensor.setRandom();
  const Eigen::TensorRef<const Tensor<T, 3, DataLayout>> ref(tensor);
  array<ptrdiff_t, 3> identity{{0, 1, 2}};

  Tensor<T, 3, DataLayout> result = ref.shuffle(identity);
  for (Index i = 0; i < result.size(); ++i) {
    VERIFY_IS_EQUAL(result.coeff(i), tensor.coeff(i));
  }
}

EIGEN_DECLARE_TEST(tensor_shuffling) {
  CALL_SUBTEST(test_simple_shuffling<ColMajor>());
  CALL_SUBTEST(test_simple_shuffling<RowMajor>());
  CALL_SUBTEST(test_expr_shuffling<ColMajor>());
  CALL_SUBTEST(test_expr_shuffling<RowMajor>());
  CALL_SUBTEST(test_shuffling_as_value<ColMajor>());
  CALL_SUBTEST(test_shuffling_as_value<RowMajor>());
  CALL_SUBTEST(test_shuffle_unshuffle<ColMajor>());
  CALL_SUBTEST(test_shuffle_unshuffle<RowMajor>());
  CALL_SUBTEST(test_empty_shuffling<ColMajor>());
  CALL_SUBTEST(test_empty_shuffling<RowMajor>());
  CALL_SUBTEST(test_shuffling_through_cast<ColMajor>());
  CALL_SUBTEST(test_shuffling_through_cast<RowMajor>());
  CALL_SUBTEST((test_shuffling_packet_paths<float, ColMajor>()));
  CALL_SUBTEST((test_shuffling_packet_paths<float, RowMajor>()));
  CALL_SUBTEST((test_shuffling_packet_paths<double, ColMajor>()));
  CALL_SUBTEST((test_shuffling_packet_paths<double, RowMajor>()));
  CALL_SUBTEST((test_identity_shuffling_without_impl_packet_access<float, ColMajor>()));
  CALL_SUBTEST((test_identity_shuffling_without_impl_packet_access<float, RowMajor>()));
}