Add new TensorBlock api implementation + tests

2 files changed, 777 insertions, 0 deletions

diff --git a/unsupported/test/cxx11_tensor_block_eval.cpp b/unsupported/test/cxx11_tensor_block_eval.cpp
new file mode 100644
index 000000000..e85b81141
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_block_eval.cpp
@@ -0,0 +1,339 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// 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/.
+
+// clang-format off
+#include "main.h"
+#include <Eigen/CXX11/Tensor>
+// clang-format on
+
+using Eigen::internal::TensorBlockDescriptor;
+using Eigen::internal::TensorExecutor;
+
+// -------------------------------------------------------------------------- //
+// Utility functions to generate random tensors, blocks, and evaluate them.
+
+template <int NumDims>
+static DSizes<Index, NumDims> RandomDims(Index min, Index max) {
+  DSizes<Index, NumDims> dims;
+  for (int i = 0; i < NumDims; ++i) {
+    dims[i] = internal::random<Index>(min, max);
+  }
+  return DSizes<Index, NumDims>(dims);
+}
+
+// Block offsets and extents allows to construct a TensorSlicingOp corresponding
+// to a TensorBlockDescriptor.
+template <int NumDims>
+struct TensorBlockParams {
+  DSizes<Index, NumDims> offsets;
+  DSizes<Index, NumDims> sizes;
+  TensorBlockDescriptor<NumDims, Index> desc;
+};
+
+template <int Layout, int NumDims>
+static TensorBlockParams<NumDims> RandomBlock(DSizes<Index, NumDims> dims,
+                                              Index min, Index max) {
+  // Choose random offsets and sizes along all tensor dimensions.
+  DSizes<Index, NumDims> offsets(RandomDims<NumDims>(min, max));
+  DSizes<Index, NumDims> sizes(RandomDims<NumDims>(min, max));
+
+  // Make sure that offset + size do not overflow dims.
+  for (int i = 0; i < NumDims; ++i) {
+    offsets[i] = numext::mini(dims[i] - 1, offsets[i]);
+    sizes[i] = numext::mini(sizes[i], dims[i] - offsets[i]);
+  }
+
+  Index offset = 0;
+  DSizes<Index, NumDims> strides = Eigen::internal::strides<Layout>(dims);
+  for (int i = 0; i < NumDims; ++i) {
+    offset += strides[i] * offsets[i];
+  }
+
+  return {offsets, sizes, TensorBlockDescriptor<NumDims, Index>(offset, sizes)};
+}
+
+// Generate block with block sizes skewed towards inner dimensions. This type of
+// block is required for evaluating broadcast expressions.
+template <int Layout, int NumDims>
+static TensorBlockParams<NumDims> SkewedInnerBlock(
+    DSizes<Index, NumDims> dims) {
+  using BlockMapper = internal::TensorBlockMapper<int, Index, NumDims, Layout>;
+  BlockMapper block_mapper(dims,
+                           internal::TensorBlockShapeType::kSkewedInnerDims,
+                           internal::random<Index>(1, dims.TotalSize()));
+
+  Index total_blocks = block_mapper.total_block_count();
+  Index block_index = internal::random<Index>(0, total_blocks - 1);
+  auto block = block_mapper.GetBlockForIndex(block_index, nullptr);
+  DSizes<Index, NumDims> sizes = block.block_sizes();
+
+  auto strides = internal::strides<Layout>(dims);
+  DSizes<Index, NumDims> offsets;
+
+  // Compute offsets for the first block coefficient.
+  Index index = block.first_coeff_index();
+  if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+    for (int i = NumDims - 1; i > 0; --i) {
+      const Index idx = index / strides[i];
+      index -= idx * strides[i];
+      offsets[i] = idx;
+    }
+    offsets[0] = index;
+  } else {
+    for (int i = 0; i < NumDims - 1; ++i) {
+      const Index idx = index / strides[i];
+      index -= idx * strides[i];
+      offsets[i] = idx;
+    }
+    offsets[NumDims - 1] = index;
+  }
+
+  auto desc = TensorBlockDescriptor<NumDims>(block.first_coeff_index(), sizes);
+  return {offsets, sizes, desc};
+}
+
+template <int NumDims>
+static TensorBlockParams<NumDims> FixedSizeBlock(DSizes<Index, NumDims> dims) {
+  DSizes<Index, NumDims> offsets;
+  for (int i = 0; i < NumDims; ++i) offsets[i] = 0;
+
+  return {offsets, dims, TensorBlockDescriptor<NumDims, Index>(0, dims)};
+}
+
+// -------------------------------------------------------------------------- //
+// Verify that block expression evaluation produces the same result as a
+// TensorSliceOp (reading a tensor block is same to taking a tensor slice).
+
+template <typename T, int NumDims, int Layout, typename Expression,
+          typename GenBlockParams>
+static void VerifyBlockEvaluator(Expression expr, GenBlockParams gen_block) {
+  using Device = DefaultDevice;
+  auto d = Device();
+
+  // Scratch memory allocator for block evaluation.
+  typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
+  TensorBlockScratch scratch(d);
+
+  // TensorEvaluator is needed to produce tensor blocks of the expression.
+  auto eval = TensorEvaluator<const decltype(expr), Device>(expr, d);
+
+  // Choose a random offsets, sizes and TensorBlockDescriptor.
+  TensorBlockParams<NumDims> block_params = gen_block();
+
+  // Evaluate TensorBlock expression into a tensor.
+  Tensor<T, NumDims, Layout> block(block_params.desc.dimensions());
+
+  // Maybe use this tensor as a block desc destination.
+  Tensor<T, NumDims, Layout> dst(block_params.desc.dimensions());
+  if (internal::random<bool>()) {
+    block_params.desc.template AddDestinationBuffer(
+        dst.data(), internal::strides<Layout>(dst.dimensions()),
+        dst.dimensions().TotalSize() * sizeof(T));
+  }
+
+  auto tensor_block = eval.blockV2(block_params.desc, scratch);
+  auto b_expr = tensor_block.expr();
+
+  // We explicitly disable vectorization and tiling, to run a simple coefficient
+  // wise assignment loop, because it's very simple and should be correct.
+  using BlockAssign = TensorAssignOp<decltype(block), const decltype(b_expr)>;
+  using BlockExecutor = TensorExecutor<const BlockAssign, Device, false,
+                                       internal::TiledEvaluation::Off>;
+  BlockExecutor::run(BlockAssign(block, b_expr), d);
+
+  // Cleanup temporary buffers owned by a tensor block.
+  tensor_block.cleanup();
+
+  // Compute a Tensor slice corresponding to a Tensor block.
+  Tensor<T, NumDims, Layout> slice(block_params.desc.dimensions());
+  auto s_expr = expr.slice(block_params.offsets, block_params.sizes);
+
+  // Explicitly use coefficient assignment to evaluate slice expression.
+  using SliceAssign = TensorAssignOp<decltype(slice), const decltype(s_expr)>;
+  using SliceExecutor = TensorExecutor<const SliceAssign, Device, false,
+                                       internal::TiledEvaluation::Off>;
+  SliceExecutor::run(SliceAssign(slice, s_expr), d);
+
+  // Tensor block and tensor slice must be the same.
+  for (Index i = 0; i < block.dimensions().TotalSize(); ++i) {
+    VERIFY_IS_EQUAL(block.coeff(i), slice.coeff(i));
+  }
+}
+
+// -------------------------------------------------------------------------- //
+
+template <typename T, int NumDims, int Layout>
+static void test_eval_tensor_block() {
+  DSizes<Index, NumDims> dims = RandomDims<NumDims>(10, 20);
+  Tensor<T, NumDims, Layout> input(dims);
+  input.setRandom();
+
+  // Identity tensor expression transformation.
+  VerifyBlockEvaluator<T, NumDims, Layout>(
+      input, [&dims]() { return RandomBlock<Layout>(dims, 10, 20); });
+}
+
+template <typename T, int NumDims, int Layout>
+static void test_eval_tensor_unary_expr_block() {
+  DSizes<Index, NumDims> dims = RandomDims<NumDims>(10, 20);
+  Tensor<T, NumDims, Layout> input(dims);
+  input.setRandom();
+
+  VerifyBlockEvaluator<T, NumDims, Layout>(
+      input.square(), [&dims]() { return RandomBlock<Layout>(dims, 10, 20); });
+}
+
+template <typename T, int NumDims, int Layout>
+static void test_eval_tensor_binary_expr_block() {
+  DSizes<Index, NumDims> dims = RandomDims<NumDims>(10, 20);
+  Tensor<T, NumDims, Layout> lhs(dims), rhs(dims);
+  lhs.setRandom();
+  rhs.setRandom();
+
+  VerifyBlockEvaluator<T, NumDims, Layout>(
+      lhs + rhs, [&dims]() { return RandomBlock<Layout>(dims, 10, 20); });
+}
+
+template <typename T, int NumDims, int Layout>
+static void test_eval_tensor_binary_with_unary_expr_block() {
+  DSizes<Index, NumDims> dims = RandomDims<NumDims>(10, 20);
+  Tensor<T, NumDims, Layout> lhs(dims), rhs(dims);
+  lhs.setRandom();
+  rhs.setRandom();
+
+  VerifyBlockEvaluator<T, NumDims, Layout>(
+      (lhs.square() + rhs.square()).sqrt(),
+      [&dims]() { return RandomBlock<Layout>(dims, 10, 20); });
+}
+
+template <typename T, int NumDims, int Layout>
+static void test_eval_tensor_broadcast() {
+  DSizes<Index, NumDims> dims = RandomDims<NumDims>(1, 10);
+  Tensor<T, NumDims, Layout> input(dims);
+  input.setRandom();
+
+  DSizes<Index, NumDims> bcast = RandomDims<NumDims>(1, 5);
+
+  DSizes<Index, NumDims> bcasted_dims;
+  for (int i = 0; i < NumDims; ++i) bcasted_dims[i] = dims[i] * bcast[i];
+
+  VerifyBlockEvaluator<T, NumDims, Layout>(
+      input.broadcast(bcast),
+      [&bcasted_dims]() { return SkewedInnerBlock<Layout>(bcasted_dims); });
+
+  VerifyBlockEvaluator<T, NumDims, Layout>(
+      input.broadcast(bcast),
+      [&bcasted_dims]() { return FixedSizeBlock(bcasted_dims); });
+
+  // Check that desc.destination() memory is not shared between two broadcast
+  // materializations.
+  VerifyBlockEvaluator<T, NumDims, Layout>(
+      input.broadcast(bcast) + input.square().broadcast(bcast),
+      [&bcasted_dims]() { return SkewedInnerBlock<Layout>(bcasted_dims); });
+}
+
+// -------------------------------------------------------------------------- //
+// Verify that assigning block to a Tensor expression produces the same result
+// as an assignment to TensorSliceOp (writing a block is is identical to
+// assigning one tensor to a slice of another tensor).
+
+template <typename T, int NumDims, int Layout, typename Expression,
+          typename GenBlockParams>
+static void VerifyBlockAssignment(Tensor<T, NumDims, Layout>& tensor,
+                                  Expression expr, GenBlockParams gen_block) {
+  using Device = DefaultDevice;
+  auto d = Device();
+
+  // We use tensor evaluator as a target for block and slice assignments.
+  auto eval = TensorEvaluator<decltype(expr), Device>(expr, d);
+
+  // Generate a random block, or choose a block that fits in full expression.
+  TensorBlockParams<NumDims> block_params = gen_block();
+
+  // Generate random data of the selected block size.
+  Tensor<T, NumDims, Layout> block(block_params.desc.dimensions());
+  block.setRandom();
+
+  // ************************************************************************ //
+  // (1) Assignment from a block.
+
+  // Construct a materialize block from a random generated block tensor.
+  internal::TensorMaterializedBlock<T, NumDims, Layout> blk(
+      internal::TensorBlockKind::kView, block.data(), block.dimensions());
+
+  // Reset all underlying tensor values to zero.
+  tensor.setZero();
+
+  // Use evaluator to write block into a tensor.
+  eval.writeBlockV2(block_params.desc, blk);
+
+  // Make a copy of the result after assignment.
+  Tensor<T, NumDims, Layout> block_assigned = tensor;
+
+  // ************************************************************************ //
+  // (2) Assignment to a slice
+
+  // Reset all underlying tensor values to zero.
+  tensor.setZero();
+
+  // Assign block to a slice of original expression
+  auto s_expr = expr.slice(block_params.offsets, block_params.sizes);
+
+  // Explicitly use coefficient assignment to evaluate slice expression.
+  using SliceAssign = TensorAssignOp<decltype(s_expr), const decltype(block)>;
+  using SliceExecutor = TensorExecutor<const SliceAssign, Device, false,
+                                       internal::TiledEvaluation::Off>;
+  SliceExecutor::run(SliceAssign(s_expr, block), d);
+
+  // Make a copy of the result after assignment.
+  Tensor<T, NumDims, Layout> slice_assigned = tensor;
+
+  for (Index i = 0; i < tensor.dimensions().TotalSize(); ++i) {
+    VERIFY_IS_EQUAL(block_assigned.coeff(i), slice_assigned.coeff(i));
+  }
+}
+
+// -------------------------------------------------------------------------- //
+
+template <typename T, int NumDims, int Layout>
+static void test_assign_tensor_block() {
+  DSizes<Index, NumDims> dims = RandomDims<NumDims>(10, 20);
+  Tensor<T, NumDims, Layout> tensor(dims);
+
+  TensorMap<Tensor<T, NumDims, Layout>> map(tensor.data(), dims);
+
+  VerifyBlockAssignment<T, NumDims, Layout>(
+      tensor, map, [&dims]() { return RandomBlock<Layout>(dims, 10, 20); });
+  VerifyBlockAssignment<T, NumDims, Layout>(
+      tensor, map, [&dims]() { return FixedSizeBlock(dims); });
+}
+
+// -------------------------------------------------------------------------- //
+
+//#define CALL_SUBTESTS(NAME) CALL_SUBTEST((NAME<float, 2, RowMajor>()))
+
+#define CALL_SUBTESTS(NAME)                   \
+  CALL_SUBTEST((NAME<float, 1, RowMajor>())); \
+  CALL_SUBTEST((NAME<float, 2, RowMajor>())); \
+  CALL_SUBTEST((NAME<float, 4, RowMajor>())); \
+  CALL_SUBTEST((NAME<float, 5, RowMajor>())); \
+  CALL_SUBTEST((NAME<float, 1, ColMajor>())); \
+  CALL_SUBTEST((NAME<float, 2, ColMajor>())); \
+  CALL_SUBTEST((NAME<float, 4, ColMajor>())); \
+  CALL_SUBTEST((NAME<float, 5, ColMajor>()))
+
+EIGEN_DECLARE_TEST(cxx11_tensor_block_eval) {
+  // clang-format off
+  CALL_SUBTESTS(test_eval_tensor_block);
+  CALL_SUBTESTS(test_eval_tensor_unary_expr_block);
+  CALL_SUBTESTS(test_eval_tensor_binary_expr_block);
+  CALL_SUBTESTS(test_eval_tensor_binary_with_unary_expr_block);
+  CALL_SUBTESTS(test_eval_tensor_broadcast);
+
+  CALL_SUBTESTS(test_assign_tensor_block);
+  // clang-format on
+}
diff --git a/unsupported/test/cxx11_tensor_block_io.cpp b/unsupported/test/cxx11_tensor_block_io.cpp
new file mode 100644
index 000000000..8a03c7dd4
--- /dev/null
+++ b/unsupported/test/cxx11_tensor_block_io.cpp
@@ -0,0 +1,438 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// 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/.
+
+// clang-format off
+#include "main.h"
+#include <Eigen/CXX11/Tensor>
+// clang-format on
+
+// -------------------------------------------------------------------------- //
+// A set of tests for TensorBlockIO: copying data between tensor blocks.
+
+template <int NumDims>
+static DSizes<Index, NumDims> RandomDims(Index min, Index max) {
+  DSizes<Index, NumDims> dims;
+  for (int i = 0; i < NumDims; ++i) {
+    dims[i] = internal::random<Index>(min, max);
+  }
+  return DSizes<Index, NumDims>(dims);
+}
+
+static internal::TensorBlockShapeType RandomBlockShape() {
+  return internal::random<bool>() ? internal::kUniformAllDims
+                                  : internal::kSkewedInnerDims;
+}
+
+template <int NumDims>
+static Index RandomTargetBlockSize(const DSizes<Index, NumDims>& dims) {
+  return internal::random<Index>(1, dims.TotalSize());
+}
+
+template <int Layout, int NumDims>
+static Index GetInputIndex(Index output_index,
+                           const array<Index, NumDims>& output_to_input_dim_map,
+                           const array<Index, NumDims>& input_strides,
+                           const array<Index, NumDims>& output_strides) {
+  int input_index = 0;
+  if (Layout == ColMajor) {
+    for (int i = NumDims - 1; i > 0; --i) {
+      const Index idx = output_index / output_strides[i];
+      input_index += idx * input_strides[output_to_input_dim_map[i]];
+      output_index -= idx * output_strides[i];
+    }
+    return input_index +
+           output_index * input_strides[output_to_input_dim_map[0]];
+  } else {
+    for (int i = 0; i < NumDims - 1; ++i) {
+      const Index idx = output_index / output_strides[i];
+      input_index += idx * input_strides[output_to_input_dim_map[i]];
+      output_index -= idx * output_strides[i];
+    }
+    return input_index +
+           output_index * input_strides[output_to_input_dim_map[NumDims - 1]];
+  }
+}
+
+template <typename T, int NumDims, int Layout>
+static void test_block_io_copy_data_from_source_to_target() {
+  using TensorBlockIO = internal::TensorBlockIOV2<T, Index, NumDims, Layout>;
+  using IODst = typename TensorBlockIO::Dst;
+  using IOSrc = typename TensorBlockIO::Src;
+
+  // Generate a random input Tensor.
+  DSizes<Index, NumDims> dims = RandomDims<NumDims>(1, 30);
+  Tensor<T, NumDims, Layout> input(dims);
+  input.setRandom();
+
+  // Write data to an output Tensor.
+  Tensor<T, NumDims, Layout> output(dims);
+
+  // Construct a tensor block mapper.
+  using TensorBlockMapper =
+      internal::TensorBlockMapper<T, Index, NumDims, Layout>;
+  TensorBlockMapper block_mapper(dims, RandomBlockShape(),
+                                 RandomTargetBlockSize(dims));
+
+  // We will copy data from input to output through this buffer.
+  Tensor<T, NumDims, Layout> block(block_mapper.block_dim_sizes());
+
+  // Precompute strides for TensorBlockIO::Copy.
+  auto input_strides = internal::strides<Layout>(dims);
+  auto output_strides = internal::strides<Layout>(dims);
+
+  const T* input_data = input.data();
+  T* output_data = output.data();
+  T* block_data = block.data();
+
+  for (int i = 0; i < block_mapper.total_block_count(); ++i) {
+    using TensorBlock = internal::TensorBlock<T, Index, NumDims, Layout>;
+    TensorBlock blk = block_mapper.GetBlockForIndex(i, block_data);
+
+    auto blk_dims = blk.block_sizes();
+    auto blk_strides = internal::strides<Layout>(blk_dims);
+
+    {
+      // Read from input into a block buffer.
+      IODst dst(blk_dims, blk_strides, block_data, 0);
+      IOSrc src(input_strides, input_data, blk.first_coeff_index());
+
+      TensorBlockIO::Copy(dst, src);
+    }
+
+    {
+      // Write from block buffer to output.
+      IODst dst(blk_dims, output_strides, output_data, blk.first_coeff_index());
+      IOSrc src(blk_strides, block_data, 0);
+
+      TensorBlockIO::Copy(dst, src);
+    }
+  }
+
+  for (int i = 0; i < dims.TotalSize(); ++i) {
+    VERIFY_IS_EQUAL(input_data[i], output_data[i]);
+  }
+}
+
+template <typename T, int NumDims, int Layout>
+static void test_block_io_copy_using_reordered_dimensions() {
+  // Generate a random input Tensor.
+  DSizes<Index, NumDims> dims = RandomDims<NumDims>(1, 30);
+  Tensor<T, NumDims, Layout> input(dims);
+  input.setRandom();
+
+  // Create a random dimension re-ordering/shuffle.
+  std::vector<int> shuffle;
+
+  for (int i = 0; i < NumDims; ++i) shuffle.push_back(i);
+  std::shuffle(shuffle.begin(), shuffle.end(), std::mt19937(g_seed));
+
+  DSizes<Index, NumDims> output_tensor_dims;
+  DSizes<Index, NumDims> input_to_output_dim_map;
+  DSizes<Index, NumDims> output_to_input_dim_map;
+  for (Index i = 0; i < NumDims; ++i) {
+    output_tensor_dims[shuffle[i]] = dims[i];
+    input_to_output_dim_map[i] = shuffle[i];
+    output_to_input_dim_map[shuffle[i]] = i;
+  }
+
+  // Write data to an output Tensor.
+  Tensor<T, NumDims, Layout> output(output_tensor_dims);
+
+  // Construct a tensor block mapper.
+  // NOTE: Tensor block mapper works with shuffled dimensions.
+  using TensorBlockMapper =
+      internal::TensorBlockMapper<T, Index, NumDims, Layout>;
+  TensorBlockMapper block_mapper(output_tensor_dims, RandomBlockShape(),
+                                 RandomTargetBlockSize(output_tensor_dims));
+
+  // We will copy data from input to output through this buffer.
+  Tensor<T, NumDims, Layout> block(block_mapper.block_dim_sizes());
+
+  // Precompute strides for TensorBlockIO::Copy.
+  auto input_strides = internal::strides<Layout>(dims);
+  auto output_strides = internal::strides<Layout>(output_tensor_dims);
+
+  const T* input_data = input.data();
+  T* output_data = output.data();
+  T* block_data = block.data();
+
+  for (Index i = 0; i < block_mapper.total_block_count(); ++i) {
+    using TensorBlock = internal::TensorBlock<T, Index, NumDims, Layout>;
+    TensorBlock blk = block_mapper.GetBlockForIndex(i, block_data);
+
+    const Index first_coeff_index = GetInputIndex<Layout, NumDims>(
+        blk.first_coeff_index(), output_to_input_dim_map, input_strides,
+        output_strides);
+
+    // NOTE: Block dimensions are in the same order as output dimensions.
+
+    using TensorBlockIO = internal::TensorBlockIOV2<T, Index, NumDims, Layout>;
+    using IODst = typename TensorBlockIO::Dst;
+    using IOSrc = typename TensorBlockIO::Src;
+
+    auto blk_dims = blk.block_sizes();
+    auto blk_strides = internal::strides<Layout>(blk_dims);
+
+    {
+      // Read from input into a block buffer.
+      IODst dst(blk_dims, blk_strides, block_data, 0);
+      IOSrc src(input_strides, input_data, first_coeff_index);
+
+      TensorBlockIO::Copy(dst, src,
+                          /*dst_to_src_dim_map=*/output_to_input_dim_map);
+    }
+
+    {
+      // We need to convert block dimensions from output to input order.
+      auto dst_dims = blk_dims;
+      for (int out_dim = 0; out_dim < NumDims; ++out_dim) {
+        dst_dims[output_to_input_dim_map[out_dim]] = blk_dims[out_dim];
+      }
+
+      // Write from block buffer to output.
+      IODst dst(dst_dims, input_strides, output_data, first_coeff_index);
+      IOSrc src(blk_strides, block_data, 0);
+
+      TensorBlockIO::Copy(dst, src,
+                          /*dst_to_src_dim_map=*/input_to_output_dim_map);
+    }
+  }
+
+  for (Index i = 0; i < dims.TotalSize(); ++i) {
+    VERIFY_IS_EQUAL(input_data[i], output_data[i]);
+  }
+}
+
+// This is the special case for reading data with reordering, when dimensions
+// before/after reordering are the same. Squeezing reads along inner dimensions
+// in this case is illegal, because we reorder innermost dimension.
+template <int Layout>
+static void test_block_io_copy_using_reordered_dimensions_do_not_squeeze() {
+  DSizes<Index, 3> tensor_dims(7, 9, 7);
+  DSizes<Index, 3> block_dims = tensor_dims;
+
+  DSizes<Index, 3> block_to_tensor_dim;
+  block_to_tensor_dim[0] = 2;
+  block_to_tensor_dim[1] = 1;
+  block_to_tensor_dim[2] = 0;
+
+  auto tensor_strides = internal::strides<Layout>(tensor_dims);
+  auto block_strides = internal::strides<Layout>(block_dims);
+
+  Tensor<float, 3, Layout> block(block_dims);
+  Tensor<float, 3, Layout> tensor(tensor_dims);
+  tensor.setRandom();
+
+  float* tensor_data = tensor.data();
+  float* block_data = block.data();
+
+  typedef internal::TensorBlock<float, Index, 3, Layout> TensorBlock;
+  TensorBlock blk(0, block_dims, block_strides, tensor_strides, block_data);
+
+  using TensorBlockIO = internal::TensorBlockIOV2<float, Index, 3, Layout>;
+  using IODst = typename TensorBlockIO::Dst;
+  using IOSrc = typename TensorBlockIO::Src;
+
+  // Read from a tensor into a block.
+  IODst dst(blk.block_sizes(), block_strides, block_data, 0);
+  IOSrc src(tensor_strides, tensor_data, blk.first_coeff_index());
+
+  TensorBlockIO::Copy(dst, src,
+                      /*dst_to_src_dim_map=*/block_to_tensor_dim);
+
+  TensorMap<Tensor<float, 3, Layout> > block_tensor(block_data, block_dims);
+  TensorMap<Tensor<float, 3, Layout> > tensor_tensor(tensor_data, tensor_dims);
+
+  for (Index d0 = 0; d0 < tensor_dims[0]; ++d0) {
+    for (Index d1 = 0; d1 < tensor_dims[1]; ++d1) {
+      for (Index d2 = 0; d2 < tensor_dims[2]; ++d2) {
+        float block_value = block_tensor(d2, d1, d0);
+        float tensor_value = tensor_tensor(d0, d1, d2);
+        VERIFY_IS_EQUAL(block_value, tensor_value);
+      }
+    }
+  }
+}
+
+// This is the special case for reading data with reordering, when dimensions
+// before/after reordering are the same. Squeezing reads in this case is allowed
+// because we reorder outer dimensions.
+template <int Layout>
+static void test_block_io_copy_using_reordered_dimensions_squeeze() {
+  DSizes<Index, 4> tensor_dims(7, 5, 9, 9);
+  DSizes<Index, 4> block_dims = tensor_dims;
+
+  DSizes<Index, 4> block_to_tensor_dim;
+  block_to_tensor_dim[0] = 0;
+  block_to_tensor_dim[1] = 1;
+  block_to_tensor_dim[2] = 3;
+  block_to_tensor_dim[3] = 2;
+
+  auto tensor_strides = internal::strides<Layout>(tensor_dims);
+  auto block_strides = internal::strides<Layout>(block_dims);
+
+  Tensor<float, 4, Layout> block(block_dims);
+  Tensor<float, 4, Layout> tensor(tensor_dims);
+  tensor.setRandom();
+
+  float* tensor_data = tensor.data();
+  float* block_data = block.data();
+
+  typedef internal::TensorBlock<float, Index, 4, Layout> TensorBlock;
+  TensorBlock blk(0, block_dims, block_strides, tensor_strides, block_data);
+
+  using TensorBlockIO = internal::TensorBlockIOV2<float, Index, 4, Layout>;
+  using IODst = typename TensorBlockIO::Dst;
+  using IOSrc = typename TensorBlockIO::Src;
+
+  // Read from a tensor into a block.
+  IODst dst(blk.block_sizes(), block_strides, block_data, 0);
+  IOSrc src(tensor_strides, tensor_data, blk.first_coeff_index());
+
+  TensorBlockIO::Copy(dst, src,
+                      /*dst_to_src_dim_map=*/block_to_tensor_dim);
+
+  TensorMap<Tensor<float, 4, Layout> > block_tensor(block_data, block_dims);
+  TensorMap<Tensor<float, 4, Layout> > tensor_tensor(tensor_data, tensor_dims);
+
+  for (Index d0 = 0; d0 < tensor_dims[0]; ++d0) {
+    for (Index d1 = 0; d1 < tensor_dims[1]; ++d1) {
+      for (Index d2 = 0; d2 < tensor_dims[2]; ++d2) {
+        for (Index d3 = 0; d3 < tensor_dims[3]; ++d3) {
+          float block_value = block_tensor(d0, d1, d3, d2);
+          float tensor_value = tensor_tensor(d0, d1, d2, d3);
+          VERIFY_IS_EQUAL(block_value, tensor_value);
+        }
+      }
+    }
+  }
+}
+
+template <int Layout>
+static void test_block_io_zero_stride() {
+  DSizes<Index, 5> rnd_dims = RandomDims<5>(1, 30);
+
+  DSizes<Index, 5> input_tensor_dims = rnd_dims;
+  input_tensor_dims[0] = 1;
+  input_tensor_dims[2] = 1;
+  input_tensor_dims[4] = 1;
+
+  Tensor<float, 5, Layout> input(input_tensor_dims);
+  input.setRandom();
+
+  DSizes<Index, 5> output_tensor_dims = rnd_dims;
+
+  auto input_tensor_strides = internal::strides<Layout>(input_tensor_dims);
+  auto output_tensor_strides = internal::strides<Layout>(output_tensor_dims);
+
+  auto input_tensor_strides_with_zeros = input_tensor_strides;
+  input_tensor_strides_with_zeros[0] = 0;
+  input_tensor_strides_with_zeros[2] = 0;
+  input_tensor_strides_with_zeros[4] = 0;
+
+  Tensor<float, 5, Layout> output(output_tensor_dims);
+  output.setRandom();
+
+  using TensorBlockIO = internal::TensorBlockIOV2<float, Index, 5, Layout>;
+  using IODst = typename TensorBlockIO::Dst;
+  using IOSrc = typename TensorBlockIO::Src;
+
+  // Write data from input to output with broadcasting in dims [0, 2, 4].
+  IODst dst(output_tensor_dims, output_tensor_strides, output.data(), 0);
+  IOSrc src(input_tensor_strides_with_zeros, input.data(), 0);
+  TensorBlockIO::Copy(dst, src);
+
+  for (int i = 0; i < output_tensor_dims[0]; ++i) {
+    for (int j = 0; j < output_tensor_dims[1]; ++j) {
+      for (int k = 0; k < output_tensor_dims[2]; ++k) {
+        for (int l = 0; l < output_tensor_dims[3]; ++l) {
+          for (int m = 0; m < output_tensor_dims[4]; ++m) {
+            float input_value = input(0, j, 0, l, 0);
+            float output_value = output(i, j, k, l, m);
+            VERIFY_IS_EQUAL(input_value, output_value);
+          }
+        }
+      }
+    }
+  }
+}
+
+template <int Layout>
+static void test_block_io_squeeze_ones() {
+  using TensorBlockIO = internal::TensorBlockIOV2<float, Index, 5, Layout>;
+  using IODst = typename TensorBlockIO::Dst;
+  using IOSrc = typename TensorBlockIO::Src;
+
+  // Total size > 1.
+  {
+    DSizes<Index, 5> block_sizes(1, 2, 1, 2, 1);
+    auto strides = internal::strides<Layout>(block_sizes);
+
+    // Create a random input tensor.
+    Tensor<float, 5> input(block_sizes);
+    input.setRandom();
+
+    Tensor<float, 5> output(block_sizes);
+
+    IODst dst(block_sizes, strides, output.data(), 0);
+    IOSrc src(strides, input.data());
+    TensorBlockIO::Copy(dst, src);
+
+    for (Index i = 0; i < block_sizes.TotalSize(); ++i) {
+      VERIFY_IS_EQUAL(output.data()[i], input.data()[i]);
+    }
+  }
+
+  // Total size == 1.
+  {
+    DSizes<Index, 5> block_sizes(1, 1, 1, 1, 1);
+    auto strides = internal::strides<Layout>(block_sizes);
+
+    // Create a random input tensor.
+    Tensor<float, 5> input(block_sizes);
+    input.setRandom();
+
+    Tensor<float, 5> output(block_sizes);
+
+    IODst dst(block_sizes, strides, output.data(), 0);
+    IOSrc src(strides, input.data());
+    TensorBlockIO::Copy(dst, src);
+
+    for (Index i = 0; i < block_sizes.TotalSize(); ++i) {
+      VERIFY_IS_EQUAL(output.data()[i], input.data()[i]);
+    }
+  }
+}
+
+#define CALL_SUBTESTS(NAME)                   \
+  CALL_SUBTEST((NAME<float, 1, RowMajor>())); \
+  CALL_SUBTEST((NAME<float, 2, RowMajor>())); \
+  CALL_SUBTEST((NAME<float, 4, RowMajor>())); \
+  CALL_SUBTEST((NAME<float, 5, RowMajor>())); \
+  CALL_SUBTEST((NAME<float, 1, ColMajor>())); \
+  CALL_SUBTEST((NAME<float, 2, ColMajor>())); \
+  CALL_SUBTEST((NAME<float, 4, ColMajor>())); \
+  CALL_SUBTEST((NAME<float, 5, ColMajor>()))
+
+EIGEN_DECLARE_TEST(cxx11_tensor_block_io) {
+  // clang-format off
+  CALL_SUBTESTS(test_block_io_copy_data_from_source_to_target);
+  CALL_SUBTESTS(test_block_io_copy_using_reordered_dimensions);
+
+  CALL_SUBTEST(test_block_io_copy_using_reordered_dimensions_do_not_squeeze<RowMajor>());
+  CALL_SUBTEST(test_block_io_copy_using_reordered_dimensions_do_not_squeeze<ColMajor>());
+
+  CALL_SUBTEST(test_block_io_copy_using_reordered_dimensions_squeeze<RowMajor>());
+  CALL_SUBTEST(test_block_io_copy_using_reordered_dimensions_squeeze<ColMajor>());
+
+  CALL_SUBTEST(test_block_io_zero_stride<RowMajor>());
+  CALL_SUBTEST(test_block_io_zero_stride<ColMajor>());
+
+  CALL_SUBTEST(test_block_io_squeeze_ones<RowMajor>());
+  CALL_SUBTEST(test_block_io_squeeze_ones<ColMajor>());
+  // clang-format on
+}
\ No newline at end of file