const int num_dims = 2;
  int64_t* dims = new int64_t[num_dims];
  int64_t num_elements = 1;
  dims[0] = batch_size;
  dims[1] = 1;
  for (int64_t i = 0; i < num_dims; ++i) {
    num_elements *= dims[i];
  }
  TF_Tensor* t = TF_AllocateTensor(TF_STRING, dims, num_dims,
                                   sizeof(TF_TString) * num_elements);
  delete[] dims;

    left_padding = split_outputs[0]
    core = split_outputs[1]
    right_padding = split_outputs[2]
    reversed_left_padding = tf.raw_ops.Reverse(tensor=left_padding, dims=rdims)
    reversed_right_padding = tf.raw_ops.Reverse(
        tensor=right_padding, dims=rdims)
    zero_like = tf.raw_ops.ZerosLike(x=core)
    left_offset, _ = tf.raw_ops.SplitV(
        value=zero_like,
        size_splits=[-1, left_padding_size],

  dlm_tensor->dl_tensor.device = tf_dlm_context;
  int ndim = tensor->dims();
  dlm_tensor->dl_tensor.ndim = ndim;
  dlm_tensor->dl_tensor.data = tf_dlm_data;
  dlm_tensor->dl_tensor.dtype = tf_dlm_type;

  std::vector<int64_t>* shape_arr = &tf_dlm_tensor_ctx->shape;
  std::vector<int64_t>* stride_arr = &tf_dlm_tensor_ctx->strides;
  shape_arr->resize(ndim);
  stride_arr->resize(ndim, 1);
  for (int i = 0; i < ndim; i++) {

# tensor 't' shape is [1, 2, 3, 4]

# 'dims' is [3] or 'dims' is [-1]
reverse(t, dims) ==> [[[[ 3,  2,  1,  0],
                        [ 7,  6,  5,  4],
                        [ 11, 10, 9, 8]],
                       [[15, 14, 13, 12],
                        [19, 18, 17, 16],
                        [23, 22, 21, 20]]]]

# 'dims' is '[1]' (or 'dims' is '[-3]')
reverse(t, dims) ==> [[[[12, 13, 14, 15],

  EXPECT_EQ(dltensor_out->device.device_id, dltensor_in->device.device_id);
  EXPECT_EQ(dltensor_out->ndim, dltensor_in->ndim);
  EXPECT_EQ(dltensor_out->dtype.code, dltensor_in->dtype.code);
  EXPECT_EQ(dltensor_out->dtype.bits, dltensor_in->dtype.bits);
  EXPECT_EQ(dltensor_out->dtype.lanes, dltensor_in->dtype.lanes);
  for (int i = 0; i < dltensor_in->ndim; ++i) {
    EXPECT_EQ(dltensor_out->shape[i], dltensor_in->shape[i]);
    if (dltensor_out->strides) {

  constexpr int64_t dims[] = {100, 100};
  constexpr int num_elements = dims[0] * dims[1];
  float data[num_elements];
  for (int i = 0; i < num_elements; ++i) {
    data[i] = 1.0f;
  }
  TF_Status* status = TF_NewStatus();
  TF_Tensor* t = TFE_AllocateHostTensor(ctx, TF_FLOAT, &dims[0],
                                        sizeof(dims) / sizeof(int64_t), status);

  %1 = stablehlo.transpose %0, dims = [1, 0] : (tensor<2x3xi32>) -> tensor<3x2xi32>
  return %1 : tensor<3x2xi32>
}
// CHECK: transpose

// -----

// Tests that transposing an argument cannot be folded.

// CHECK-LABEL: transpose_arg
func.func @transpose_arg(%arg0: tensor<2x3xf32>) -> tensor<3x2xf32> {
  %0 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<2x3xf32>) -> tensor<3x2xf32>

    - mikedanese
  sig-auth-authorizers-reviewers:
    - deads2k
    - dims
    - enj
    - liggitt
    - mikedanese
    - ncdc
    - smarterclayton
    - sttts
    - thockin
    - wojtek-t
  sig-auth-certificates-approvers:
    - liggitt
    - mikedanese
    - smarterclayton
  sig-auth-certificates-reviewers:
    - deads2k
    - dims
    - enj
    - liggitt
    - mikedanese

  static char empty;
  int64_t nelems = 1;
  std::vector<int64_t> dims;
  auto shape_dims = shape.dims();
  dims.reserve(shape_dims);
  for (int i = 0; i < shape_dims; ++i) {
    dims.push_back(shape.dim_size(i));
    nelems *= shape.dim_size(i);
  }
  CHECK_EQ(nelems, 0);
  return TF_NewTensor(
      dtype, reinterpret_cast<const int64_t*>(dims.data()), shape.dims(),

  %1 = stablehlo.transpose %arg0, dims = [0, 3, 1, 2] : (tensor<1x3x3x4xf32>) -> tensor<1x4x3x3xf32>
  %2 = stablehlo.add %1, %0 : tensor<1x4x3x3xf32>
  return %2 : tensor<1x4x3x3xf32>
}
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x3x3x4xf32>) -> tensor<1x4x3x3xf32>
// CHECK-DAG: %[[CONST_0:.+]] = stablehlo.constant