// 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

#define EIGEN_USE_THREADS

#include "main.h"
#include <atomic>
#include <iostream>
#include <Eigen/Tensor>

using Eigen::Tensor;

class TestAllocator : public Allocator {
 public:
  ~TestAllocator() EIGEN_OVERRIDE {}
  EIGEN_DEVICE_FUNC void* allocate(size_t num_bytes) const EIGEN_OVERRIDE {
    const_cast<TestAllocator*>(this)->alloc_count_++;
    return internal::aligned_malloc(num_bytes);
  }
  EIGEN_DEVICE_FUNC void deallocate(void* buffer) const EIGEN_OVERRIDE {
    const_cast<TestAllocator*>(this)->dealloc_count_++;
    internal::aligned_free(buffer);
  }

  int alloc_count() const { return alloc_count_; }
  int dealloc_count() const { return dealloc_count_; }

 private:
  int alloc_count_ = 0;
  int dealloc_count_ = 0;
};

template <typename = void>
void test_multithread_elementwise() {
  Tensor<float, 3> in1(200, 30, 70);
  Tensor<float, 3> in2(200, 30, 70);
  Tensor<double, 3> out(200, 30, 70);

  in1.setRandom();
  in2.setRandom();

  Eigen::ThreadPool tp(internal::random<int>(3, 11));
  Eigen::ThreadPoolDevice thread_pool_device(&tp, internal::random<int>(3, 11));
  out.device(thread_pool_device) = (in1 + in2 * 3.14f).cast<double>();

  for (int i = 0; i < 200; ++i) {
    for (int j = 0; j < 30; ++j) {
      for (int k = 0; k < 70; ++k) {
        VERIFY_IS_APPROX(out(i, j, k), static_cast<double>(in1(i, j, k) + in2(i, j, k) * 3.14f));
      }
    }
  }
}

template <typename = void>
void test_async_multithread_elementwise() {
  Tensor<float, 3> in1(200, 30, 70);
  Tensor<float, 3> in2(200, 30, 70);
  Tensor<double, 3> out(200, 30, 70);

  in1.setRandom();
  in2.setRandom();

  Eigen::ThreadPool tp(internal::random<int>(3, 11));
  Eigen::ThreadPoolDevice thread_pool_device(&tp, internal::random<int>(3, 11));

  Eigen::Barrier b(1);
  out.device(thread_pool_device, [&b]() { b.Notify(); }) = (in1 + in2 * 3.14f).cast<double>();
  b.Wait();

  for (int i = 0; i < 200; ++i) {
    for (int j = 0; j < 30; ++j) {
      for (int k = 0; k < 70; ++k) {
        VERIFY_IS_APPROX(out(i, j, k), static_cast<double>(in1(i, j, k) + in2(i, j, k) * 3.14f));
      }
    }
  }
}

template <typename = void>
void test_multithread_chip() {
  Tensor<float, 5> in(2, 3, 5, 7, 11);
  Tensor<float, 4> out(3, 5, 7, 11);

  in.setRandom();

  Eigen::ThreadPool tp(internal::random<int>(3, 11));
  Eigen::ThreadPoolDevice thread_pool_device(&tp, internal::random<int>(3, 11));

  out.device(thread_pool_device) = in.chip(1, 0);

  for (int i = 0; i < 3; ++i) {
    for (int j = 0; j < 5; ++j) {
      for (int k = 0; k < 7; ++k) {
        for (int l = 0; l < 11; ++l) {
          VERIFY_IS_EQUAL(out(i, j, k, l), in(1, i, j, k, l));
        }
      }
    }
  }
}

template <typename = void>
void test_async_multithread_chip() {
  Tensor<float, 5> in(2, 3, 5, 7, 11);
  Tensor<float, 4> out(3, 5, 7, 11);

  in.setRandom();

  Eigen::ThreadPool tp(internal::random<int>(3, 11));
  Eigen::ThreadPoolDevice thread_pool_device(&tp, internal::random<int>(3, 11));

  Eigen::Barrier b(1);
  out.device(thread_pool_device, [&b]() { b.Notify(); }) = in.chip(1, 0);
  b.Wait();

  for (int i = 0; i < 3; ++i) {
    for (int j = 0; j < 5; ++j) {
      for (int k = 0; k < 7; ++k) {
        for (int l = 0; l < 11; ++l) {
          VERIFY_IS_EQUAL(out(i, j, k, l), in(1, i, j, k, l));
        }
      }
    }
  }
}

template <typename = void>
void test_multithread_volume_patch() {
  Tensor<float, 5> in(4, 2, 3, 5, 7);
  Tensor<float, 6> out(4, 1, 1, 1, 2 * 3 * 5, 7);

  in.setRandom();

  Eigen::ThreadPool tp(internal::random<int>(3, 11));
  Eigen::ThreadPoolDevice thread_pool_device(&tp, internal::random<int>(3, 11));

  out.device(thread_pool_device) = in.extract_volume_patches(1, 1, 1);

  for (int i = 0; i < in.size(); ++i) {
    VERIFY_IS_EQUAL(in.data()[i], out.data()[i]);
  }
}

template <typename = void>
void test_async_multithread_volume_patch() {
  Tensor<float, 5> in(4, 2, 3, 5, 7);
  Tensor<float, 6> out(4, 1, 1, 1, 2 * 3 * 5, 7);

  in.setRandom();

  Eigen::ThreadPool tp(internal::random<int>(3, 11));
  Eigen::ThreadPoolDevice thread_pool_device(&tp, internal::random<int>(3, 11));

  Eigen::Barrier b(1);
  out.device(thread_pool_device, [&b]() { b.Notify(); }) = in.extract_volume_patches(1, 1, 1);
  b.Wait();

  for (int i = 0; i < in.size(); ++i) {
    VERIFY_IS_EQUAL(in.data()[i], out.data()[i]);
  }
}

// Stateful-but-copyable unary functor: flags an overlap when two threads call
// the same instance concurrently. The executors must copy the evaluator (and
// with it the functor) into each task, so no instance is ever called
// concurrently even though the expression is evaluated by multiple threads.
struct OverlapDetectingOp {
  explicit OverlapDetectingOp(std::atomic<int>* overlaps) : overlaps_(overlaps) {}
  OverlapDetectingOp(const OverlapDetectingOp& other) : overlaps_(other.overlaps_) {}

  float operator()(const float& x) const {
    if (busy_.exchange(true)) overlaps_->fetch_add(1);
    const float result = x + 1.0f;
    busy_.store(false);
    return result;
  }

  std::atomic<int>* overlaps_;
  mutable std::atomic<bool> busy_{false};
};

// The driver expression uses reverse() because its block() has always read
// the argument through coeff() with BlockAccess independent of the argument,
// so these tests exercise the tiled executor's per-task evaluator copy
// regardless of how any one evaluator's traits evolve.
template <typename = void>
void test_multithread_tiled_stateful_functor() {
  Tensor<float, 4> in(64, 64, 64, 8);
  in.setRandom();

  const array<bool, 4> rev = {{false, true, true, false}};
  Tensor<float, 4> expected(64, 64, 64, 8);
  expected = (in + 1.0f).reverse(rev);

  std::atomic<int> overlaps(0);
  Eigen::ThreadPool tp(4);
  Eigen::ThreadPoolDevice thread_pool_device(&tp, 4);

  Tensor<float, 4> out(64, 64, 64, 8);
  out.device(thread_pool_device) = in.unaryExpr(OverlapDetectingOp(&overlaps)).reverse(rev);

  VERIFY_IS_EQUAL(overlaps.load(), 0);
  for (int i = 0; i < out.size(); ++i) {
    VERIFY_IS_EQUAL(out.data()[i], expected.data()[i]);
  }
}

template <typename = void>
void test_async_multithread_tiled_stateful_functor() {
  Tensor<float, 4> in(64, 64, 64, 8);
  in.setRandom();

  const array<bool, 4> rev = {{false, true, true, false}};
  Tensor<float, 4> expected(64, 64, 64, 8);
  expected = (in + 1.0f).reverse(rev);

  std::atomic<int> overlaps(0);
  Eigen::ThreadPool tp(4);
  Eigen::ThreadPoolDevice thread_pool_device(&tp, 4);

  Tensor<float, 4> out(64, 64, 64, 8);
  Eigen::Barrier b(1);
  out.device(thread_pool_device, [&b]() { b.Notify(); }) = in.unaryExpr(OverlapDetectingOp(&overlaps)).reverse(rev);
  b.Wait();

  VERIFY_IS_EQUAL(overlaps.load(), 0);
  for (int i = 0; i < out.size(); ++i) {
    VERIFY_IS_EQUAL(out.data()[i], expected.data()[i]);
  }
}

template <typename = void>
void test_multithread_compound_assignment() {
  Tensor<float, 3> in1(2, 3, 7);
  Tensor<float, 3> in2(2, 3, 7);
  Tensor<float, 3> out(2, 3, 7);

  in1.setRandom();
  in2.setRandom();

  Eigen::ThreadPool tp(internal::random<int>(3, 11));
  Eigen::ThreadPoolDevice thread_pool_device(&tp, internal::random<int>(3, 11));
  out.device(thread_pool_device) = in1;
  out.device(thread_pool_device) += in2 * 3.14f;

  for (int i = 0; i < 2; ++i) {
    for (int j = 0; j < 3; ++j) {
      for (int k = 0; k < 7; ++k) {
        VERIFY_IS_APPROX(out(i, j, k), in1(i, j, k) + in2(i, j, k) * 3.14f);
      }
    }
  }
}

template <int DataLayout>
void test_multithread_contraction() {
  Tensor<float, 4, DataLayout> t_left(30, 50, 37, 31);
  Tensor<float, 5, DataLayout> t_right(37, 31, 70, 2, 10);
  Tensor<float, 5, DataLayout> t_result(30, 50, 70, 2, 10);

  t_left.setRandom();
  t_right.setRandom();

  // this contraction should be equivalent to a single matrix multiplication
  typedef Tensor<float, 1>::DimensionPair DimPair;
  Eigen::array<DimPair, 2> dims{{DimPair(2, 0), DimPair(3, 1)}};

  typedef Map<Matrix<float, Dynamic, Dynamic, DataLayout>> MapXf;
  MapXf m_left(t_left.data(), 1500, 1147);
  MapXf m_right(t_right.data(), 1147, 1400);
  Matrix<float, Dynamic, Dynamic, DataLayout> m_result(1500, 1400);

  Eigen::ThreadPool tp(4);
  Eigen::ThreadPoolDevice thread_pool_device(&tp, 4);

  // compute results by separate methods
  t_result.device(thread_pool_device) = t_left.contract(t_right, dims);
  m_result = m_left * m_right;

  // The tensor contraction and the matrix product sum the k=1147 terms in different orders; their results differ
  // by a small multiple of eps, covered here as absolute slack before the relative check.
  for (ptrdiff_t i = 0; i < t_result.size(); i++) {
    VERIFY(&t_result.data()[i] != &m_result.data()[i]);
    if (fabsf(t_result(i) - m_result(i)) < 128 * NumTraits<float>::epsilon()) {
      continue;
    }
    if (Eigen::internal::isApprox(t_result(i), m_result(i), 128 * NumTraits<float>::epsilon())) {
      continue;
    }
    std::cout << "mismatch detected at index " << i << ": " << t_result(i) << " vs " << m_result(i) << std::endl;
    assert(false);
  }
}

// Issue #1648: lock the threaded dispatch of gemv-shape contractions
// (TensorContractionThreadPool forces num_threads = 1 for n == 1 and routes
// through evalProductSequential, which now selects one of four direct GEMV
// fast paths). Covers y = A^T * x where the LHS contracted dim is the
// contiguous one — the path that motivated the fix.
template <int DataLayout>
void test_multithread_gemv_transpose() {
  for (int M : {17, 64, 129}) {
    for (int N : {16, 33, 100}) {
      Tensor<float, 2, DataLayout> t_left(M, N);
      Tensor<float, 1, DataLayout> t_right(M);
      t_left.setRandom();
      t_right.setRandom();

      typedef Tensor<float, 1>::DimensionPair DimPair;
      Eigen::array<DimPair, 1> dims{{DimPair(0, 0)}};

      Eigen::ThreadPool tp(4);
      Eigen::ThreadPoolDevice thread_pool_device(&tp, 4);
      Tensor<float, 1, DataLayout> t_result(N);
      t_result.device(thread_pool_device) = t_left.contract(t_right, dims);

      typedef Map<const Matrix<float, Dynamic, Dynamic, DataLayout>> MapMat;
      typedef Map<const Matrix<float, Dynamic, 1>> MapVec;
      MapMat m_left(t_left.data(), M, N);
      MapVec m_right(t_right.data(), M);
      Matrix<float, Dynamic, 1> m_result = m_left.transpose() * m_right;

      for (int j = 0; j < N; ++j) {
        VERIFY(internal::isApprox(t_result(j), m_result(j), 16 * NumTraits<float>::epsilon()));
      }
    }
  }
}

template <int DataLayout>
void test_contraction_corner_cases() {
  Tensor<float, 2, DataLayout> t_left(32, 500);
  Tensor<float, 2, DataLayout> t_right(32, 28 * 28);
  Tensor<float, 2, DataLayout> t_result(500, 28 * 28);

  t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f;
  t_right = (t_right.constant(-0.6f) + t_right.random()) * 2.0f;
  t_result = t_result.constant(NAN);

  // this contraction should be equivalent to a single matrix multiplication
  typedef Tensor<float, 1>::DimensionPair DimPair;
  Eigen::array<DimPair, 1> dims{{DimPair(0, 0)}};

  typedef Map<Matrix<float, Dynamic, Dynamic, DataLayout>> MapXf;
  MapXf m_left(t_left.data(), 32, 500);
  MapXf m_right(t_right.data(), 32, 28 * 28);
  Matrix<float, Dynamic, Dynamic, DataLayout> m_result(500, 28 * 28);

  Eigen::ThreadPool tp(12);
  Eigen::ThreadPoolDevice thread_pool_device(&tp, 12);

  // compute results by separate methods
  t_result.device(thread_pool_device) = t_left.contract(t_right, dims);
  m_result = m_left.transpose() * m_right;

  for (ptrdiff_t i = 0; i < t_result.size(); i++) {
    assert(!(numext::isnan)(t_result.data()[i]));
    if (fabsf(t_result.data()[i] - m_result.data()[i]) >= 256 * NumTraits<float>::epsilon()) {
      std::cout << "mismatch detected at index " << i << " : " << t_result.data()[i] << " vs " << m_result.data()[i]
                << std::endl;
      assert(false);
    }
  }

  t_left.resize(32, 1);
  t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f;
  t_result.resize(1, 28 * 28);
  t_result = t_result.constant(NAN);
  t_result.device(thread_pool_device) = t_left.contract(t_right, dims);
  new (&m_left) MapXf(t_left.data(), 32, 1);
  m_result = m_left.transpose() * m_right;
  for (ptrdiff_t i = 0; i < t_result.size(); i++) {
    assert(!(numext::isnan)(t_result.data()[i]));
    if (fabsf(t_result.data()[i] - m_result.data()[i]) >= 256 * NumTraits<float>::epsilon()) {
      std::cout << "mismatch detected: " << t_result.data()[i] << " vs " << m_result.data()[i] << std::endl;
      assert(false);
    }
  }

  t_left.resize(32, 500);
  t_right.resize(32, 4);
  t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f;
  t_right = (t_right.constant(-0.6f) + t_right.random()) * 2.0f;
  t_result.resize(500, 4);
  t_result = t_result.constant(NAN);
  t_result.device(thread_pool_device) = t_left.contract(t_right, dims);
  new (&m_left) MapXf(t_left.data(), 32, 500);
  new (&m_right) MapXf(t_right.data(), 32, 4);
  m_result = m_left.transpose() * m_right;
  for (ptrdiff_t i = 0; i < t_result.size(); i++) {
    assert(!(numext::isnan)(t_result.data()[i]));
    if (fabsf(t_result.data()[i] - m_result.data()[i]) >= 256 * NumTraits<float>::epsilon()) {
      std::cout << "mismatch detected: " << t_result.data()[i] << " vs " << m_result.data()[i] << std::endl;
      assert(false);
    }
  }

  t_left.resize(32, 1);
  t_right.resize(32, 4);
  t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f;
  t_right = (t_right.constant(-0.6f) + t_right.random()) * 2.0f;
  t_result.resize(1, 4);
  t_result = t_result.constant(NAN);
  t_result.device(thread_pool_device) = t_left.contract(t_right, dims);
  new (&m_left) MapXf(t_left.data(), 32, 1);
  new (&m_right) MapXf(t_right.data(), 32, 4);
  m_result = m_left.transpose() * m_right;
  for (ptrdiff_t i = 0; i < t_result.size(); i++) {
    assert(!(numext::isnan)(t_result.data()[i]));
    if (fabsf(t_result.data()[i] - m_result.data()[i]) >= 256 * NumTraits<float>::epsilon()) {
      std::cout << "mismatch detected: " << t_result.data()[i] << " vs " << m_result.data()[i] << std::endl;
      assert(false);
    }
  }
}

template <int DataLayout>
void test_multithread_contraction_agrees_with_singlethread() {
  int contract_size = internal::random<int>(1, 5000);

  Tensor<float, 3, DataLayout> left(internal::random<int>(1, 80), contract_size, internal::random<int>(1, 100));

  Tensor<float, 4, DataLayout> right(internal::random<int>(1, 25), internal::random<int>(1, 37), contract_size,
                                     internal::random<int>(1, 51));

  left.setRandom();
  right.setRandom();

  // add constants to shift values away from 0 for more precision
  left += left.constant(1.5f);
  right += right.constant(1.5f);

  typedef Tensor<float, 1>::DimensionPair DimPair;
  Eigen::array<DimPair, 1> dims{{DimPair(1, 2)}};

  Eigen::ThreadPool tp(internal::random<int>(2, 11));
  Eigen::ThreadPoolDevice thread_pool_device(&tp, internal::random<int>(2, 11));

  Tensor<float, 5, DataLayout> st_result;
  st_result = left.contract(right, dims);

  Tensor<float, 5, DataLayout> tp_result(st_result.dimensions());
  tp_result.device(thread_pool_device) = left.contract(right, dims);

  VERIFY(dimensions_match(st_result.dimensions(), tp_result.dimensions()));
  for (ptrdiff_t i = 0; i < st_result.size(); i++) {
    // if both of the values are very small, then do nothing (because the test will fail
    // due to numerical precision issues when values are small)
    if (numext::abs(st_result.data()[i] - tp_result.data()[i]) >= 1024 * NumTraits<float>::epsilon()) {
      VERIFY_IS_APPROX(st_result.data()[i], tp_result.data()[i]);
    }
  }
}

// Apply Sqrt to all output elements.
struct SqrtOutputKernel {
  template <typename Index, typename Scalar>
  EIGEN_ALWAYS_INLINE void operator()(const internal::blas_data_mapper<Scalar, Index, ColMajor>& output_mapper,
                                      const TensorContractionParams&, Index, Index, Index num_rows,
                                      Index num_cols) const {
    for (int i = 0; i < num_rows; ++i) {
      for (int j = 0; j < num_cols; ++j) {
        output_mapper(i, j) = std::sqrt(output_mapper(i, j));
      }
    }
  }
};

template <int DataLayout>
static void test_multithread_contraction_with_output_kernel() {
  typedef Tensor<float, 1>::DimensionPair DimPair;

  const int num_threads = internal::random<int>(2, 11);
  ThreadPool threads(num_threads);
  Eigen::ThreadPoolDevice device(&threads, num_threads);

  Tensor<float, 4, DataLayout> t_left(30, 50, 8, 31);
  Tensor<float, 5, DataLayout> t_right(8, 31, 7, 20, 10);
  Tensor<float, 5, DataLayout> t_result(30, 50, 7, 20, 10);

  t_left.setRandom();
  t_right.setRandom();
  // Put trash in mat4 to verify contraction clears output memory.
  t_result.setRandom();

  // Add a little offset so that the results won't be close to zero.
  t_left += t_left.constant(1.0f);
  t_right += t_right.constant(1.0f);

  typedef Map<Eigen::Matrix<float, Dynamic, Dynamic, DataLayout>> MapXf;
  MapXf m_left(t_left.data(), 1500, 248);
  MapXf m_right(t_right.data(), 248, 1400);
  Eigen::Matrix<float, Dynamic, Dynamic, DataLayout> m_result(1500, 1400);

  // this contraction should be equivalent to a single matrix multiplication
  Eigen::array<DimPair, 2> dims{{DimPair(2, 0), DimPair(3, 1)}};

  // compute results by separate methods
  t_result.device(device) = t_left.contract(t_right, dims, SqrtOutputKernel());

  m_result = m_left * m_right;

  for (Index i = 0; i < t_result.dimensions().TotalSize(); i++) {
    VERIFY(&t_result.data()[i] != &m_result.data()[i]);
    VERIFY_IS_APPROX(t_result.data()[i], std::sqrt(m_result.data()[i]));
  }
}

template <int DataLayout>
void test_async_multithread_contraction_agrees_with_singlethread() {
  int contract_size = internal::random<int>(100, 500);

  Tensor<float, 3, DataLayout> left(internal::random<int>(10, 40), contract_size, internal::random<int>(10, 40));

  Tensor<float, 4, DataLayout> right(internal::random<int>(1, 20), internal::random<int>(1, 20), contract_size,
                                     internal::random<int>(1, 20));

  left.setRandom();
  right.setRandom();

  // add constants to shift values away from 0 for more precision
  left += left.constant(1.5f);
  right += right.constant(1.5f);

  typedef Tensor<float, 1>::DimensionPair DimPair;
  Eigen::array<DimPair, 1> dims{{DimPair(1, 2)}};

  Eigen::ThreadPool tp(internal::random<int>(2, 11));
  Eigen::ThreadPoolDevice thread_pool_device(&tp, internal::random<int>(8, 32));

  Tensor<float, 5, DataLayout> st_result;
  st_result = left.contract(right, dims);

  Tensor<float, 5, DataLayout> tp_result(st_result.dimensions());

  Eigen::Barrier barrier(1);
  tp_result.device(thread_pool_device, [&barrier]() { barrier.Notify(); }) = left.contract(right, dims);
  barrier.Wait();

  VERIFY(dimensions_match(st_result.dimensions(), tp_result.dimensions()));
  for (ptrdiff_t i = 0; i < st_result.size(); i++) {
    // if both of the values are very small, then do nothing (because the test
    // will fail due to numerical precision issues when values are small)
    if (numext::abs(st_result.data()[i] - tp_result.data()[i]) >= 1024 * NumTraits<float>::epsilon()) {
      VERIFY_IS_APPROX(st_result.data()[i], tp_result.data()[i]);
    }
  }
}

template <typename Contraction>
struct AsyncContractionEvalToProbe;

namespace Eigen {
namespace internal {
template <typename Contraction>
struct traits<AsyncContractionEvalToProbe<Contraction>> : traits<TensorEvaluator<const Contraction, ThreadPoolDevice>> {
};
}  // namespace internal
}  // namespace Eigen

template <typename Contraction>
struct AsyncContractionEvalToProbe : Eigen::TensorContractionEvaluatorBase<AsyncContractionEvalToProbe<Contraction>> {
  using Base = Eigen::TensorContractionEvaluatorBase<AsyncContractionEvalToProbe<Contraction>>;
  using Scalar = typename Base::Scalar;

  AsyncContractionEvalToProbe(const Contraction& op, const Eigen::ThreadPoolDevice& device) : Base(op, device) {}

  template <int Alignment>
  void evalProduct(Scalar*) const {}

  template <typename EvalToCallback, int Alignment>
  void evalProductAsync(Scalar* buffer, EvalToCallback done) const {
    static int shared_instantiation_calls = 0;
    last_instantiation_calls = ++shared_instantiation_calls;
    last_buffer = buffer;
    done();
  }

  mutable Scalar* last_buffer = nullptr;
  mutable int last_instantiation_calls = 0;
};

template <int DataLayout>
void test_async_contraction_eval_subexprs_dedup() {
  using DimPair = typename Tensor<float, 1>::DimensionPair;
  Eigen::ThreadPool tp(4);
  Eigen::ThreadPoolDevice device(&tp, 4);

  Tensor<float, 2, DataLayout> left(4, 4);
  Tensor<float, 2, DataLayout> right(4, 4);
  Tensor<float, 2, DataLayout> dest(4, 4);
  left.setRandom();
  right.setRandom();

  Eigen::array<DimPair, 1> dims{{DimPair(1, 0)}};
  auto expr = left.contract(right, dims);
  AsyncContractionEvalToProbe<decltype(expr)> probe(expr, device);

  bool need_assign = true;
  auto on_done = [&need_assign](bool na) { need_assign = na; };

  probe.evalSubExprsIfNeededAsync(dest.data(), on_done);
  VERIFY_IS_EQUAL(need_assign, false);
  VERIFY_IS_EQUAL(probe.last_buffer, dest.data());
  VERIFY_IS_EQUAL(probe.data(), static_cast<float*>(nullptr));
  VERIFY_IS_EQUAL(probe.last_instantiation_calls, 1);

  probe.evalSubExprsIfNeededAsync(nullptr, on_done);
  VERIFY_IS_EQUAL(need_assign, true);
  VERIFY(probe.data() != nullptr);
  VERIFY_IS_EQUAL(probe.last_buffer, probe.data());
  VERIFY_IS_EQUAL(probe.last_instantiation_calls, 2);
  probe.cleanup();
  VERIFY_IS_EQUAL(probe.data(), static_cast<float*>(nullptr));

  // Verify real ThreadPoolDevice evaluation along the dest == nullptr branch
  // (nested contraction subexpression) across parallel GEMM, inner-dim sharded,
  // and k = 0 shapes.
  const int shapes[][3] = {{64, 64, 64}, {2, 2048, 10}, {10, 0, 20}};
  for (const auto& shape : shapes) {
    const int m = shape[0];
    const int k = shape[1];
    const int n = shape[2];
    Tensor<float, 2, DataLayout> t_left(m, k);
    Tensor<float, 2, DataLayout> t_right(k, n);
    Tensor<float, 2, DataLayout> t_bias(m, n);
    t_left.setRandom();
    t_right.setRandom();
    t_bias.setRandom();
    t_left += t_left.constant(1.5f);
    t_right += t_right.constant(1.5f);
    t_bias += t_bias.constant(1.5f);

    Tensor<float, 2, DataLayout> st_result(m, n);
    Tensor<float, 2, DataLayout> tp_result(m, n);
    tp_result.setZero();

    st_result = t_left.contract(t_right, dims) + t_bias;

    Eigen::Barrier barrier(1);
    tp_result.device(device, [&barrier]() { barrier.Notify(); }) = t_left.contract(t_right, dims) + t_bias;
    barrier.Wait();

    for (Index i = 0; i < st_result.size(); ++i) {
      VERIFY_IS_APPROX(tp_result.data()[i], st_result.data()[i]);
    }
  }
}

// We are triggering 'evalShardedByInnerDim' optimization.
template <int DataLayout>
static void test_sharded_by_inner_dim_contraction() {
  using DimPair = typename Tensor<float, 1>::DimensionPair;

  const int num_threads = internal::random<int>(4, 16);
  ThreadPool threads(num_threads);
  Eigen::ThreadPoolDevice device(&threads, num_threads);

  const Index k = 30000 * PacketType<float, Eigen::ThreadPoolDevice>::size;
  Tensor<float, 2, DataLayout> t_left(2, k);
  Tensor<float, 2, DataLayout> t_right(k, 10);
  Tensor<float, 2, DataLayout> t_result(2, 10);

  t_left.setRandom();
  t_right.setRandom();
  // Put trash in t_result to verify contraction clears output memory.
  t_result.setRandom();

  // Add a little offset so that the results won't be close to zero.
  t_left += t_left.constant(1.0f);
  t_right += t_right.constant(1.0f);

  using MapXf = Map<Eigen::Matrix<float, Dynamic, Dynamic, DataLayout>>;
  MapXf m_left(t_left.data(), 2, k);
  MapXf m_right(t_right.data(), k, 10);
  Eigen::Matrix<float, Dynamic, Dynamic, DataLayout> m_result(2, 10);

  // this contraction should be equivalent to a single matrix multiplication
  Eigen::array<DimPair, 1> dims{{DimPair(1, 0)}};

  auto contraction = t_left.contract(t_right, dims);
  using ContractionEval = TensorEvaluator<const decltype(contraction), Eigen::ThreadPoolDevice>;
  ContractionEval eval(contraction, device);
  const int nt_k = eval.numThreadsInnerDim(2, 10, k);
  VERIFY(nt_k > 1);
  VERIFY(ContractionEval::shardByInnerDim(2, 10, k, /*num_threads=*/1, nt_k));

  // compute results by separate methods
  t_result.device(device) = contraction;
  m_result = m_left * m_right;

  for (Index i = 0; i < t_result.dimensions().TotalSize(); i++) {
    VERIFY_IS_APPROX(t_result.data()[i], m_result.data()[i]);
  }
}

// We are triggering 'evalShardedByInnerDim' optimization with output kernel.
template <int DataLayout>
static void test_sharded_by_inner_dim_contraction_with_output_kernel() {
  using DimPair = typename Tensor<float, 1>::DimensionPair;

  const int num_threads = internal::random<int>(4, 16);
  ThreadPool threads(num_threads);
  Eigen::ThreadPoolDevice device(&threads, num_threads);

  const Index k = 30000 * PacketType<float, Eigen::ThreadPoolDevice>::size;
  Tensor<float, 2, DataLayout> t_left(2, k);
  Tensor<float, 2, DataLayout> t_right(k, 10);
  Tensor<float, 2, DataLayout> t_result(2, 10);

  t_left.setRandom();
  t_right.setRandom();
  // Put trash in t_result to verify contraction clears output memory.
  t_result.setRandom();

  // Add a little offset so that the results won't be close to zero.
  t_left += t_left.constant(1.0f);
  t_right += t_right.constant(1.0f);

  using MapXf = Map<Eigen::Matrix<float, Dynamic, Dynamic, DataLayout>>;
  MapXf m_left(t_left.data(), 2, k);
  MapXf m_right(t_right.data(), k, 10);
  Eigen::Matrix<float, Dynamic, Dynamic, DataLayout> m_result(2, 10);

  // this contraction should be equivalent to a single matrix multiplication
  Eigen::array<DimPair, 1> dims{{DimPair(1, 0)}};

  auto contraction = t_left.contract(t_right, dims, SqrtOutputKernel());
  using ContractionEval = TensorEvaluator<const decltype(contraction), Eigen::ThreadPoolDevice>;
  ContractionEval eval(contraction, device);
  const int nt_k = eval.numThreadsInnerDim(2, 10, k);
  VERIFY(nt_k > 1);
  VERIFY(ContractionEval::shardByInnerDim(2, 10, k, /*num_threads=*/1, nt_k));

  // compute results by separate methods
  t_result.device(device) = contraction;
  m_result = m_left * m_right;

  for (Index i = 0; i < t_result.dimensions().TotalSize(); i++) {
    VERIFY_IS_APPROX(t_result.data()[i], std::sqrt(m_result.data()[i]));
  }
}

// We are triggering 'evalShardedByInnerDim' optimization.
template <int DataLayout>
static void test_async_sharded_by_inner_dim_contraction() {
  using DimPair = typename Tensor<float, 1>::DimensionPair;

  const int num_threads = internal::random<int>(4, 16);
  ThreadPool threads(num_threads);
  Eigen::ThreadPoolDevice device(&threads, num_threads);

  const Index k = 30000 * PacketType<float, Eigen::ThreadPoolDevice>::size;
  Tensor<float, 2, DataLayout> t_left(2, k);
  Tensor<float, 2, DataLayout> t_right(k, 10);
  Tensor<float, 2, DataLayout> t_result(2, 10);

  t_left.setRandom();
  t_right.setRandom();
  // Put trash in t_result to verify contraction clears output memory.
  t_result.setRandom();

  // Add a little offset so that the results won't be close to zero.
  t_left += t_left.constant(1.0f);
  t_right += t_right.constant(1.0f);

  using MapXf = Map<Eigen::Matrix<float, Dynamic, Dynamic, DataLayout>>;
  MapXf m_left(t_left.data(), 2, k);
  MapXf m_right(t_right.data(), k, 10);
  Eigen::Matrix<float, Dynamic, Dynamic, DataLayout> m_result(2, 10);

  // this contraction should be equivalent to a single matrix multiplication
  Eigen::array<DimPair, 1> dims{{DimPair(1, 0)}};

  auto contraction = t_left.contract(t_right, dims);
  using ContractionEval = TensorEvaluator<const decltype(contraction), Eigen::ThreadPoolDevice>;
  ContractionEval eval(contraction, device);
  const int nt_k = eval.numThreadsInnerDim(2, 10, k);
  VERIFY(nt_k > 1);
  VERIFY(ContractionEval::shardByInnerDim(2, 10, k, /*num_threads=*/1, nt_k));

  // compute results by separate methods
  Eigen::Barrier barrier(1);
  t_result.device(device, [&barrier]() { barrier.Notify(); }) = contraction;
  barrier.Wait();

  m_result = m_left * m_right;

  for (Index i = 0; i < t_result.dimensions().TotalSize(); i++) {
    VERIFY_IS_APPROX(t_result.data()[i], m_result.data()[i]);
  }
}

// We are triggering 'evalShardedByInnerDim' optimization with output kernel.
template <int DataLayout>
static void test_async_sharded_by_inner_dim_contraction_with_output_kernel() {
  using DimPair = typename Tensor<float, 1>::DimensionPair;

  const int num_threads = internal::random<int>(4, 16);
  ThreadPool threads(num_threads);
  Eigen::ThreadPoolDevice device(&threads, num_threads);

  const Index k = 30000 * PacketType<float, Eigen::ThreadPoolDevice>::size;
  Tensor<float, 2, DataLayout> t_left(2, k);
  Tensor<float, 2, DataLayout> t_right(k, 10);
  Tensor<float, 2, DataLayout> t_result(2, 10);

  t_left.setRandom();
  t_right.setRandom();
  // Put trash in t_result to verify contraction clears output memory.
  t_result.setRandom();

  // Add a little offset so that the results won't be close to zero.
  t_left += t_left.constant(1.0f);
  t_right += t_right.constant(1.0f);

  using MapXf = Map<Eigen::Matrix<float, Dynamic, Dynamic, DataLayout>>;
  MapXf m_left(t_left.data(), 2, k);
  MapXf m_right(t_right.data(), k, 10);
  Eigen::Matrix<float, Dynamic, Dynamic, DataLayout> m_result(2, 10);

  // this contraction should be equivalent to a single matrix multiplication
  Eigen::array<DimPair, 1> dims{{DimPair(1, 0)}};

  auto contraction = t_left.contract(t_right, dims, SqrtOutputKernel());
  using ContractionEval = TensorEvaluator<const decltype(contraction), Eigen::ThreadPoolDevice>;
  ContractionEval eval(contraction, device);
  const int nt_k = eval.numThreadsInnerDim(2, 10, k);
  VERIFY(nt_k > 1);
  VERIFY(ContractionEval::shardByInnerDim(2, 10, k, /*num_threads=*/1, nt_k));

  // compute results by separate methods
  Eigen::Barrier barrier(1);
  t_result.device(device, [&barrier]() { barrier.Notify(); }) = contraction;
  barrier.Wait();
  m_result = m_left * m_right;

  for (Index i = 0; i < t_result.dimensions().TotalSize(); i++) {
    VERIFY_IS_APPROX(t_result.data()[i], std::sqrt(m_result.data()[i]));
  }
}

template <int DataLayout>
static void test_sharded_by_inner_dim_cost_model() {
  using DimPair = typename Tensor<float, 1>::DimensionPair;
  Eigen::array<DimPair, 1> dims{{DimPair(1, 0)}};

  ThreadPool threads(8);
  Eigen::ThreadPoolDevice device(&threads, 8);

  Tensor<float, 2, DataLayout> t_left(2, 2);
  Tensor<float, 2, DataLayout> t_right(2, 2);
  auto expr = t_left.contract(t_right, dims);
  using ContractionEval = TensorEvaluator<const decltype(expr), Eigen::ThreadPoolDevice>;
  ContractionEval eval(expr, device);

  // Scale k by the packet width so the parallel cost (6.6e4..2.6e5 cycles across FMA and panel-width regimes) sits
  // between the old hardcoded threshold (~5.7e4 cycles) and the TensorCostModel threshold (~4.0e5 cycles) under every
  // SIMD width.
  const Index f32_packet_size = PacketType<float, Eigen::ThreadPoolDevice>::size;
  const Index moderate_f32_k = 64 * f32_packet_size;
  VERIFY_IS_EQUAL(eval.numThreadsInnerDim(32, 64, moderate_f32_k), 1);
  VERIFY(!ContractionEval::shardByInnerDim(32, 64, moderate_f32_k, /*num_threads=*/1,
                                           eval.numThreadsInnerDim(32, 64, moderate_f32_k)));

  // Cover the bfloat16 contraction shape (16 x 128 x k) that motivated the fix.
  Tensor<Eigen::bfloat16, 2, DataLayout> bf16_left(2, 2);
  Tensor<Eigen::bfloat16, 2, DataLayout> bf16_right(2, 2);
  auto bf16_expr = bf16_left.contract(bf16_right, dims);
  using Bf16ContractionEval = TensorEvaluator<const decltype(bf16_expr), Eigen::ThreadPoolDevice>;
  Bf16ContractionEval bf16_eval(bf16_expr, device);
  const Index bf16_packet_size = PacketType<Eigen::bfloat16, Eigen::ThreadPoolDevice>::size;
  const Index moderate_bf16_k = 64 * bf16_packet_size;
  VERIFY_IS_EQUAL(bf16_eval.numThreadsInnerDim(16, 128, moderate_bf16_k), 1);
  VERIFY(!Bf16ContractionEval::shardByInnerDim(16, 128, moderate_bf16_k, /*num_threads=*/1,
                                               bf16_eval.numThreadsInnerDim(16, 128, moderate_bf16_k)));

  // Deep inner dimension with tiny outer dimensions should still shard by k.
  const Index deep_f32_k = 30000 * f32_packet_size;
  const int nt_k = eval.numThreadsInnerDim(2, 10, deep_f32_k);
  VERIFY(nt_k > 1);
  VERIFY(ContractionEval::shardByInnerDim(2, 10, deep_f32_k, /*num_threads=*/1, nt_k));
}

template <int DataLayout>
void test_full_contraction() {
  int contract_size1 = internal::random<int>(1, 500);
  int contract_size2 = internal::random<int>(1, 500);

  Tensor<float, 2, DataLayout> left(contract_size1, contract_size2);
  Tensor<float, 2, DataLayout> right(contract_size1, contract_size2);
  left.setRandom();
  right.setRandom();

  // add constants to shift values away from 0 for more precision
  left += left.constant(1.5f);
  right += right.constant(1.5f);

  typedef Tensor<float, 2>::DimensionPair DimPair;
  Eigen::array<DimPair, 2> dims{{DimPair(0, 0), DimPair(1, 1)}};

  Eigen::ThreadPool tp(internal::random<int>(2, 11));
  Eigen::ThreadPoolDevice thread_pool_device(&tp, internal::random<int>(2, 11));

  Tensor<float, 0, DataLayout> st_result;
  st_result = left.contract(right, dims);

  Tensor<float, 0, DataLayout> tp_result;
  tp_result.device(thread_pool_device) = left.contract(right, dims);

  VERIFY(dimensions_match(st_result.dimensions(), tp_result.dimensions()));
  // if both of the values are very small, then do nothing (because the test will fail
  // due to numerical precision issues when values are small)
  if (numext::abs(st_result() - tp_result()) >= 1024 * NumTraits<float>::epsilon()) {
    VERIFY_IS_APPROX(st_result(), tp_result());
  }
}

template <int DataLayout>
void test_multithreaded_reductions() {
  const int num_threads = internal::random<int>(3, 11);
  ThreadPool thread_pool(num_threads);
  Eigen::ThreadPoolDevice thread_pool_device(&thread_pool, num_threads);

  const int num_rows = internal::random<int>(13, 732);
  const int num_cols = internal::random<int>(13, 732);
  Tensor<float, 2, DataLayout> t1(num_rows, num_cols);
  t1.setRandom();

  Tensor<float, 0, DataLayout> full_redux;
  full_redux = t1.sum();

  Tensor<float, 0, DataLayout> full_redux_tp;
  full_redux_tp.device(thread_pool_device) = t1.sum();

  // Check that the single threaded and the multi threaded reductions return
  // the same result.
  VERIFY_IS_APPROX(full_redux(), full_redux_tp());
}

// reduce() squares each value, so partial results cannot be merged through it.
struct SquaredSumReducer {
  void reduce(const double t, double* accum) const { *accum += t * t; }
  double initialize() const { return 0; }
  double finalize(const double accum) const { return accum; }
};

template <typename = void>
void test_multithreaded_value_transforming_full_reduction() {
  ThreadPool thread_pool(4);
  Eigen::ThreadPoolDevice thread_pool_device(&thread_pool, 4);
  Tensor<double, 1> t(1 << 20);
  t.setConstant(1.0);
  Tensor<double, 0> result;
  result.device(thread_pool_device) = t.reduce(array<Index, 1>{{0}}, SquaredSumReducer());
  VERIFY_IS_EQUAL(result(), static_cast<double>(t.size()));
}

// argmax()/argmin() take the sharded full reduction; each extremum recurs in every shard, so the
// merge must keep the first occurrence.
template <typename = void>
void test_multithreaded_argmax() {
  STATIC_CHECK((internal::reducer_can_reorder_accumulators<internal::ArgMaxPairReducer<Pair<Index, float>>>::value));
  STATIC_CHECK((internal::reducer_can_reorder_accumulators<internal::ArgMinPairReducer<Pair<Index, float>>>::value));
  ThreadPool thread_pool(4);
  Eigen::ThreadPoolDevice thread_pool_device(&thread_pool, 4);
  Tensor<float, 1> t(1 << 20);
  for (Index i = 0; i < t.size(); ++i) t(i) = static_cast<float>((i + 500) % 1000);
  Tensor<Index, 0> result;
  result.device(thread_pool_device) = t.argmax();
  VERIFY_IS_EQUAL(result(), Index(499));
  result.device(thread_pool_device) = t.argmin();
  VERIFY_IS_EQUAL(result(), Index(500));
}

template <typename = void>
void test_multithreaded_complex_reduction() {
  using Scalar = std::complex<float>;
  constexpr Index size = 4096;

  // Preserve the mapped source and fixed-size mapped destination from the issue #1647 backtrace: this exact
  // instantiation triggers the old GCC compiler bug.
  Tensor<Scalar, 1, RowMajor> storage(size);
  storage.setConstant(Scalar(1.0f, 2.0f));
  TensorMap<Tensor<Scalar, 1, RowMajor>> input(storage.data(), size);
  EIGEN_ALIGN_MAX Scalar output_storage;
  TensorMap<TensorFixedSize<Scalar, Sizes<>, RowMajor>, Aligned> output(&output_storage);

  ThreadPool thread_pool(2);
  ThreadPoolDevice thread_pool_device(&thread_pool, 2);
  output.device(thread_pool_device) = input.sum();

  VERIFY_IS_EQUAL(output(), Scalar(static_cast<float>(size), static_cast<float>(2 * size)));
}

template <typename = void>
void test_multithreaded_complex_partial_reductions() {
  using Scalar = std::complex<float>;
  constexpr Index outer_size = 32;
  constexpr Index reduced_size = 257;

  Tensor<Scalar, 2, RowMajor> input(outer_size, reduced_size);
  for (Index i = 0; i < outer_size; ++i) {
    for (Index j = 0; j < reduced_size; ++j) {
      input(i, j) = Scalar(static_cast<float>(i + 1), static_cast<float>(2 * i + 1));
    }
  }

  ThreadPool thread_pool(2);
  ThreadPoolDevice thread_pool_device(&thread_pool, 2);

  IndexList<type2index<1>> static_reduction_dim;
  Tensor<Scalar, 1, RowMajor> static_result(outer_size);
  static_result.device(thread_pool_device) = input.sum(static_reduction_dim);

  array<Index, 1> runtime_reduction_dim{1};
  Tensor<Scalar, 1, RowMajor> runtime_result(outer_size);
  runtime_result.device(thread_pool_device) = input.sum(runtime_reduction_dim);

  for (Index i = 0; i < outer_size; ++i) {
    const Scalar expected = static_cast<float>(reduced_size) * input(i, 0);
    VERIFY_IS_EQUAL(static_result(i), expected);
    VERIFY_IS_EQUAL(runtime_result(i), expected);
  }
}

template <typename = void>
void test_memcpy() {
  for (int i = 0; i < 5; ++i) {
    const int num_threads = internal::random<int>(3, 11);
    Eigen::ThreadPool tp(num_threads);
    Eigen::ThreadPoolDevice thread_pool_device(&tp, num_threads);

    const int size = internal::random<int>(13, 7632);
    Tensor<float, 1> t1(size);
    t1.setRandom();
    std::vector<float> result(size);
    thread_pool_device.memcpy(&result[0], t1.data(), size * sizeof(float));
    for (int j = 0; j < size; j++) {
      VERIFY_IS_EQUAL(t1(j), result[j]);
    }
  }
}

// The executors copy the evaluator per task; a copy must not destroy the forced-eval buffer's elements.
template <typename = void>
void test_multithread_forced_eval_non_pod() {
  Eigen::ThreadPool tp(4);
  Eigen::ThreadPoolDevice thread_pool_device(&tp, 4);
  Tensor<std::string, 1> in(64), out(64);
  for (int i = 0; i < 64; ++i) in(i) = std::string(40, static_cast<char>('a' + i % 26));
  out.device(thread_pool_device) = in.eval();
  for (int i = 0; i < 64; ++i) VERIFY_IS_EQUAL(out(i), in(i));
}

template <typename = void>
void test_enqueue_with_args() {
  Eigen::ThreadPool tp(2);
  Eigen::ThreadPoolDevice thread_pool_device(&tp, 2);
  Eigen::Barrier barrier(2);
  std::atomic<int> sum(0);
  thread_pool_device.enqueue([&sum](Eigen::Barrier* b, int x) { sum += x, b->Notify(); }, &barrier, 3);
  thread_pool_device.enqueueNoNotification([&sum](Eigen::Barrier* b) { sum += 4, b->Notify(); }, &barrier);
  barrier.Wait();
  VERIFY_IS_EQUAL(sum.load(), 7);
}

// The async path constructs the forced-eval buffer's non-POD elements before evaluating into them.
template <typename = void>
void test_async_forced_eval_non_pod() {
  Eigen::ThreadPool tp(4);
  Eigen::ThreadPoolDevice thread_pool_device(&tp, 4);
  Tensor<std::string, 1> in(64), out(64);
  for (int i = 0; i < 64; ++i) in(i) = std::string(40, static_cast<char>('a' + i % 26));
  Eigen::Barrier done(1);
  out.device(thread_pool_device, [&done]() { done.Notify(); }) = in.eval();
  done.Wait();
  for (int i = 0; i < 64; ++i) VERIFY_IS_EQUAL(out(i), in(i));
}

// The single-block async path frees its scratch before the done callback, after which the device may be destroyed.
template <typename = void>
void test_async_single_block_scratch_freed_before_done() {
  TestAllocator allocator;
  Eigen::ThreadPool tp(4);
  Eigen::ThreadPoolDevice thread_pool_device(&tp, 4, &allocator);
  Tensor<float, 3> src(4, 5, 6), other(6, 5, 4), dst(6, 5, 4);
  src.setRandom();
  other.setRandom();
  Eigen::array<Index, 3> perm{{2, 1, 0}};
  // The binary op does not hand the destination buffer to the shuffle, so its block is materialized in scratch.
  const auto expr = src.shuffle(perm) + other;
  int deallocs_at_done = -1;
  Eigen::Barrier done(1);
  auto on_done = [&]() {
    deallocs_at_done = allocator.dealloc_count();
    done.Notify();
  };
  using Assign = TensorAssignOp<decltype(dst), const decltype(expr)>;
  using Executor = internal::TensorAsyncExecutor<const Assign, ThreadPoolDevice, decltype(on_done),
                                                 /*Vectorizable=*/true, internal::TiledEvaluation::On>;
  // With no async subexpressions and a single block, runAsync evaluates and signals on this thread.
  Executor::runAsync(Assign(dst, expr), thread_pool_device, on_done);
  done.Wait();
  VERIFY(allocator.alloc_count() > 0);
  VERIFY_IS_EQUAL(allocator.dealloc_count(), deallocs_at_done);
}

template <typename = void>
void test_multithread_random() {
  Eigen::ThreadPool tp(2);
  Eigen::ThreadPoolDevice device(&tp, 2);
  Tensor<float, 1> t(1 << 20);
  t.device(device) = t.random<Eigen::internal::NormalRandomGenerator<float>>();
}

template <int DataLayout>
void test_multithread_shuffle(Allocator* allocator) {
  Tensor<float, 4, DataLayout> tensor(17, 5, 7, 11);
  tensor.setRandom();

  const int num_threads = internal::random<int>(2, 11);
  ThreadPool threads(num_threads);
  Eigen::ThreadPoolDevice device(&threads, num_threads, allocator);

  Tensor<float, 4, DataLayout> shuffle(7, 5, 11, 17);
  array<ptrdiff_t, 4> shuffles = {{2, 1, 3, 0}};
  shuffle.device(device) = tensor.shuffle(shuffles);

  for (int i = 0; i < 17; ++i) {
    for (int j = 0; j < 5; ++j) {
      for (int k = 0; k < 7; ++k) {
        for (int l = 0; l < 11; ++l) {
          VERIFY_IS_EQUAL(tensor(i, j, k, l), shuffle(k, j, l, i));
        }
      }
    }
  }
}

template <typename = void>
void test_threadpool_allocate(TestAllocator* allocator) {
  const int num_threads = internal::random<int>(2, 11);
  const int num_allocs = internal::random<int>(2, 11);
  ThreadPool threads(num_threads);
  Eigen::ThreadPoolDevice device(&threads, num_threads, allocator);

  for (int a = 0; a < num_allocs; ++a) {
    void* ptr = device.allocate(512);
    device.deallocate(ptr);
  }
  VERIFY(allocator != nullptr);
  VERIFY_IS_EQUAL(allocator->alloc_count(), num_allocs);
  VERIFY_IS_EQUAL(allocator->dealloc_count(), num_allocs);
}

template <int DataLayout>
void test_multithread_zero_dim_contraction() {
  const int num_threads = internal::random<int>(2, 11);
  ThreadPool threads(num_threads);
  Eigen::ThreadPoolDevice device(&threads, num_threads);

  // 2D contraction: {10, 0} * {0, 20} -> {10, 20}
  {
    Tensor<float, 2, DataLayout> t_left(10, 0);
    Tensor<float, 2, DataLayout> t_right(0, 20);
    Tensor<float, 2, DataLayout> t_result(10, 20);
    t_result.setConstant(123.0f);

    typedef Tensor<float, 1>::DimensionPair DimPair;
    Eigen::array<DimPair, 1> dims{{DimPair(1, 0)}};

    t_result.device(device) = t_left.contract(t_right, dims);

    VERIFY_IS_EQUAL(t_result.dimension(0), 10);
    VERIFY_IS_EQUAL(t_result.dimension(1), 20);
    for (Index i = 0; i < t_result.size(); ++i) {
      VERIFY_IS_EQUAL(t_result.data()[i], 0.0f);
    }
  }

  // 2D contraction with non-empty dim >= 48: {64, 0} * {0, 64} -> {64, 64}
  {
    Tensor<float, 2, DataLayout> t_left(64, 0);
    Tensor<float, 2, DataLayout> t_right(0, 64);
    Tensor<float, 2, DataLayout> t_result(64, 64);
    t_result.setConstant(123.0f);

    typedef Tensor<float, 1>::DimensionPair DimPair;
    Eigen::array<DimPair, 1> dims{{DimPair(1, 0)}};

    t_result.device(device) = t_left.contract(t_right, dims);

    VERIFY_IS_EQUAL(t_result.dimension(0), 64);
    VERIFY_IS_EQUAL(t_result.dimension(1), 64);
    for (Index i = 0; i < t_result.size(); ++i) {
      VERIFY_IS_EQUAL(t_result.data()[i], 0.0f);
    }
  }

  // Multidim contraction: {5, 0, 4} * {0, 3, 4} with dims {1, 0} and {2, 2} -> {5, 3}
  {
    Tensor<float, 3, DataLayout> t_left(5, 0, 4);
    Tensor<float, 3, DataLayout> t_right(0, 3, 4);
    Tensor<float, 2, DataLayout> t_result(5, 3);
    t_result.setConstant(123.0f);

    typedef Tensor<float, 1>::DimensionPair DimPair;
    Eigen::array<DimPair, 2> dims{{DimPair(1, 0), DimPair(2, 2)}};

    t_result.device(device) = t_left.contract(t_right, dims);

    VERIFY_IS_EQUAL(t_result.dimension(0), 5);
    VERIFY_IS_EQUAL(t_result.dimension(1), 3);
    for (Index i = 0; i < t_result.size(); ++i) {
      VERIFY_IS_EQUAL(t_result.data()[i], 0.0f);
    }
  }
}

template <int DataLayout>
void test_async_multithread_zero_dim_contraction() {
  const int num_threads = internal::random<int>(2, 11);
  ThreadPool threads(num_threads);
  Eigen::ThreadPoolDevice device(&threads, num_threads);

  Tensor<float, 2, DataLayout> t_left(10, 0);
  Tensor<float, 2, DataLayout> t_right(0, 20);
  Tensor<float, 2, DataLayout> t_result(10, 20);
  t_result.setConstant(123.0f);

  typedef Tensor<float, 1>::DimensionPair DimPair;
  Eigen::array<DimPair, 1> dims{{DimPair(1, 0)}};

  Eigen::Barrier barrier(1);
  t_result.device(device, [&barrier]() { barrier.Notify(); }) = t_left.contract(t_right, dims);
  barrier.Wait();

  VERIFY_IS_EQUAL(t_result.dimension(0), 10);
  VERIFY_IS_EQUAL(t_result.dimension(1), 20);
  for (Index i = 0; i < t_result.size(); ++i) {
    VERIFY_IS_EQUAL(t_result.data()[i], 0.0f);
  }
}

template <typename Scalar>
struct AddBiasOutputKernel {
  Scalar bias;
  explicit AddBiasOutputKernel(Scalar b) : bias(b) {}

  template <typename Index, typename ResScalar>
  EIGEN_ALWAYS_INLINE void operator()(const internal::blas_data_mapper<ResScalar, Index, ColMajor>& output_mapper,
                                      const TensorContractionParams&, Index, Index, Index num_rows,
                                      Index num_cols) const {
    for (Index i = 0; i < num_rows; ++i) {
      for (Index j = 0; j < num_cols; ++j) {
        output_mapper(i, j) += bias;
      }
    }
  }
};

template <int DataLayout>
void test_multithread_zero_dim_contraction_with_output_kernel() {
  const int num_threads = internal::random<int>(2, 11);
  ThreadPool threads(num_threads);
  Eigen::ThreadPoolDevice device(&threads, num_threads);

  const float bias = 42.0f;
  typedef Tensor<float, 1>::DimensionPair DimPair;
  Eigen::array<DimPair, 1> dims{{DimPair(1, 0)}};

  // {10, 0} * {0, 20} -> {10, 20}
  {
    Tensor<float, 2, DataLayout> t_left(10, 0);
    Tensor<float, 2, DataLayout> t_right(0, 20);
    Tensor<float, 2, DataLayout> t_result(10, 20);
    t_result.setConstant(123.0f);

    t_result.device(device) = t_left.contract(t_right, dims, AddBiasOutputKernel<float>(bias));

    VERIFY_IS_EQUAL(t_result.dimension(0), 10);
    VERIFY_IS_EQUAL(t_result.dimension(1), 20);
    for (Index i = 0; i < t_result.size(); ++i) {
      VERIFY_IS_EQUAL(t_result.data()[i], bias);
    }
  }

  // {64, 0} * {0, 64} -> {64, 64}
  {
    Tensor<float, 2, DataLayout> t_left(64, 0);
    Tensor<float, 2, DataLayout> t_right(0, 64);
    Tensor<float, 2, DataLayout> t_result(64, 64);
    t_result.setConstant(123.0f);

    t_result.device(device) = t_left.contract(t_right, dims, AddBiasOutputKernel<float>(bias));

    VERIFY_IS_EQUAL(t_result.dimension(0), 64);
    VERIFY_IS_EQUAL(t_result.dimension(1), 64);
    for (Index i = 0; i < t_result.size(); ++i) {
      VERIFY_IS_EQUAL(t_result.data()[i], bias);
    }
  }
}

template <int DataLayout>
void test_async_multithread_zero_dim_contraction_with_output_kernel() {
  const int num_threads = internal::random<int>(2, 11);
  ThreadPool threads(num_threads);
  Eigen::ThreadPoolDevice device(&threads, num_threads);

  const float bias = 42.0f;
  typedef Tensor<float, 1>::DimensionPair DimPair;
  Eigen::array<DimPair, 1> dims{{DimPair(1, 0)}};

  // {10, 0} * {0, 20} -> {10, 20}
  {
    Tensor<float, 2, DataLayout> t_left(10, 0);
    Tensor<float, 2, DataLayout> t_right(0, 20);
    Tensor<float, 2, DataLayout> t_result(10, 20);
    t_result.setConstant(123.0f);

    Eigen::Barrier barrier(1);
    t_result.device(device, [&barrier]() { barrier.Notify(); }) =
        t_left.contract(t_right, dims, AddBiasOutputKernel<float>(bias));
    barrier.Wait();

    VERIFY_IS_EQUAL(t_result.dimension(0), 10);
    VERIFY_IS_EQUAL(t_result.dimension(1), 20);
    for (Index i = 0; i < t_result.size(); ++i) {
      VERIFY_IS_EQUAL(t_result.data()[i], bias);
    }
  }

  // {64, 0} * {0, 64} -> {64, 64}
  {
    Tensor<float, 2, DataLayout> t_left(64, 0);
    Tensor<float, 2, DataLayout> t_right(0, 64);
    Tensor<float, 2, DataLayout> t_result(64, 64);
    t_result.setConstant(123.0f);

    Eigen::Barrier barrier(1);
    t_result.device(device, [&barrier]() { barrier.Notify(); }) =
        t_left.contract(t_right, dims, AddBiasOutputKernel<float>(bias));
    barrier.Wait();

    VERIFY_IS_EQUAL(t_result.dimension(0), 64);
    VERIFY_IS_EQUAL(t_result.dimension(1), 64);
    for (Index i = 0; i < t_result.size(); ++i) {
      VERIFY_IS_EQUAL(t_result.data()[i], bias);
    }
  }
}

template <int DataLayout>
void test_multithread_zero_dim_outer_zeros() {
  const int num_threads = internal::random<int>(2, 11);
  ThreadPool threads(num_threads);
  Eigen::ThreadPoolDevice device(&threads, num_threads);

  typedef Tensor<float, 1>::DimensionPair DimPair;
  Eigen::array<DimPair, 1> dims{{DimPair(1, 0)}};

  // Outer zero on LHS: {0, 5} * {5, 10} -> {0, 10}
  {
    Tensor<float, 2, DataLayout> t_left(0, 5);
    Tensor<float, 2, DataLayout> t_right(5, 10);
    t_right.setRandom();
    Tensor<float, 2, DataLayout> t_result(0, 10);
    t_result.device(device) = t_left.contract(t_right, dims);
    VERIFY_IS_EQUAL(t_result.dimension(0), 0);
    VERIFY_IS_EQUAL(t_result.dimension(1), 10);
    VERIFY_IS_EQUAL(t_result.size(), 0);
  }

  // Outer zero on RHS: {10, 5} * {5, 0} -> {10, 0}
  {
    Tensor<float, 2, DataLayout> t_left(10, 5);
    Tensor<float, 2, DataLayout> t_right(5, 0);
    t_left.setRandom();
    Tensor<float, 2, DataLayout> t_result(10, 0);
    t_result.device(device) = t_left.contract(t_right, dims);
    VERIFY_IS_EQUAL(t_result.dimension(0), 10);
    VERIFY_IS_EQUAL(t_result.dimension(1), 0);
    VERIFY_IS_EQUAL(t_result.size(), 0);
  }

  // Outer zero on both: {0, 5} * {5, 0} -> {0, 0}
  {
    Tensor<float, 2, DataLayout> t_left(0, 5);
    Tensor<float, 2, DataLayout> t_right(5, 0);
    Tensor<float, 2, DataLayout> t_result(0, 0);
    t_result.device(device) = t_left.contract(t_right, dims);
    VERIFY_IS_EQUAL(t_result.dimension(0), 0);
    VERIFY_IS_EQUAL(t_result.dimension(1), 0);
    VERIFY_IS_EQUAL(t_result.size(), 0);
  }

  // Async zero outer dim
  {
    Tensor<float, 2, DataLayout> t_left(0, 5);
    Tensor<float, 2, DataLayout> t_right(5, 10);
    t_right.setRandom();
    Tensor<float, 2, DataLayout> t_result(0, 10);
    Eigen::Barrier barrier(1);
    t_result.device(device, [&barrier]() { barrier.Notify(); }) = t_left.contract(t_right, dims);
    barrier.Wait();
    VERIFY_IS_EQUAL(t_result.dimension(0), 0);
    VERIFY_IS_EQUAL(t_result.dimension(1), 10);
    VERIFY_IS_EQUAL(t_result.size(), 0);
  }
}

EIGEN_DECLARE_TEST(tensor_thread_pool) {
  CALL_SUBTEST_1(test_multithread_elementwise<>());
  CALL_SUBTEST_1(test_async_multithread_elementwise<>());
  CALL_SUBTEST_1(test_multithread_compound_assignment<>());

  CALL_SUBTEST_2(test_multithread_contraction<ColMajor>());
  CALL_SUBTEST_2(test_multithread_contraction<RowMajor>());
  CALL_SUBTEST_2(test_multithread_gemv_transpose<ColMajor>());
  CALL_SUBTEST_2(test_multithread_gemv_transpose<RowMajor>());

  CALL_SUBTEST_3(test_multithread_chip<>());
  CALL_SUBTEST_3(test_async_multithread_chip<>());

  CALL_SUBTEST_4(test_multithread_volume_patch<>());
  CALL_SUBTEST_4(test_async_multithread_volume_patch<>());
  CALL_SUBTEST_4(test_multithread_tiled_stateful_functor<>());
  CALL_SUBTEST_4(test_async_multithread_tiled_stateful_functor<>());

  CALL_SUBTEST_5(test_multithread_contraction_agrees_with_singlethread<ColMajor>());
  CALL_SUBTEST_5(test_multithread_contraction_agrees_with_singlethread<RowMajor>());
  CALL_SUBTEST_5(test_multithread_contraction_with_output_kernel<ColMajor>());
  CALL_SUBTEST_5(test_multithread_contraction_with_output_kernel<RowMajor>());

  CALL_SUBTEST_6(test_async_multithread_contraction_agrees_with_singlethread<ColMajor>());
  CALL_SUBTEST_6(test_async_multithread_contraction_agrees_with_singlethread<RowMajor>());
  CALL_SUBTEST_6(test_async_contraction_eval_subexprs_dedup<ColMajor>());
  CALL_SUBTEST_6(test_async_contraction_eval_subexprs_dedup<RowMajor>());

  // Test EvalShardedByInnerDimContext parallelization strategy.
  CALL_SUBTEST_7(test_sharded_by_inner_dim_contraction<ColMajor>());
  CALL_SUBTEST_7(test_sharded_by_inner_dim_contraction<RowMajor>());
  CALL_SUBTEST_7(test_sharded_by_inner_dim_contraction_with_output_kernel<ColMajor>());
  CALL_SUBTEST_7(test_sharded_by_inner_dim_contraction_with_output_kernel<RowMajor>());
  CALL_SUBTEST_7(test_sharded_by_inner_dim_cost_model<ColMajor>());
  CALL_SUBTEST_7(test_sharded_by_inner_dim_cost_model<RowMajor>());

  CALL_SUBTEST_8(test_async_sharded_by_inner_dim_contraction<ColMajor>());
  CALL_SUBTEST_8(test_async_sharded_by_inner_dim_contraction<RowMajor>());
  CALL_SUBTEST_8(test_async_sharded_by_inner_dim_contraction_with_output_kernel<ColMajor>());
  CALL_SUBTEST_8(test_async_sharded_by_inner_dim_contraction_with_output_kernel<RowMajor>());

  // Exercise various cases that have been problematic in the past.
  CALL_SUBTEST_9(test_contraction_corner_cases<ColMajor>());
  CALL_SUBTEST_9(test_contraction_corner_cases<RowMajor>());

  CALL_SUBTEST_10(test_full_contraction<ColMajor>());
  CALL_SUBTEST_10(test_full_contraction<RowMajor>());

  CALL_SUBTEST_11(test_multithreaded_reductions<ColMajor>());
  CALL_SUBTEST_11(test_multithreaded_reductions<RowMajor>());
  CALL_SUBTEST_11(test_multithreaded_complex_reduction<>());
  CALL_SUBTEST_11(test_multithreaded_value_transforming_full_reduction<>());
  CALL_SUBTEST_11(test_multithreaded_argmax<>());
  CALL_SUBTEST_11(test_multithreaded_complex_partial_reductions<>());

  CALL_SUBTEST_12(test_memcpy<>());
  CALL_SUBTEST_12(test_multithread_random<>());
  CALL_SUBTEST_12(test_multithread_forced_eval_non_pod<>());
  CALL_SUBTEST_12(test_enqueue_with_args<>());
  CALL_SUBTEST_12(test_async_forced_eval_non_pod<>());
  CALL_SUBTEST_12(test_async_single_block_scratch_freed_before_done<>());

  TestAllocator test_allocator;
  CALL_SUBTEST_13(test_multithread_shuffle<ColMajor>(nullptr));
  CALL_SUBTEST_13(test_multithread_shuffle<RowMajor>(&test_allocator));
  CALL_SUBTEST_13(test_threadpool_allocate<>(&test_allocator));

  CALL_SUBTEST_14(test_multithread_zero_dim_contraction<ColMajor>());
  CALL_SUBTEST_14(test_multithread_zero_dim_contraction<RowMajor>());
  CALL_SUBTEST_14(test_async_multithread_zero_dim_contraction<ColMajor>());
  CALL_SUBTEST_14(test_async_multithread_zero_dim_contraction<RowMajor>());
  CALL_SUBTEST_14(test_multithread_zero_dim_contraction_with_output_kernel<ColMajor>());
  CALL_SUBTEST_14(test_multithread_zero_dim_contraction_with_output_kernel<RowMajor>());
  CALL_SUBTEST_14(test_async_multithread_zero_dim_contraction_with_output_kernel<ColMajor>());
  CALL_SUBTEST_14(test_async_multithread_zero_dim_contraction_with_output_kernel<RowMajor>());
  CALL_SUBTEST_14(test_multithread_zero_dim_outer_zeros<ColMajor>());
  CALL_SUBTEST_14(test_multithread_zero_dim_outer_zeros<RowMajor>());

  // Force CMake to split this test.
  // EIGEN_SUFFIXES;1;2;3;4;5;6;7;8;9;10;11;12;13;14
}