concat_op.cc

/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

// See docs in ../ops/array_ops.cc.

#include <limits>
#include <vector>

#include "absl/status/status.h"
#include "absl/strings/str_cat.h"
#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
#include "tensorflow/core/framework/bounds_check.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/framework/tensor_types.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/kernels/concat_lib.h"
#include "tensorflow/core/lib/core/status.h"
#include "tensorflow/core/platform/errors.h"
#include "tensorflow/core/platform/types.h"

namespace tensorflow {

typedef Eigen::ThreadPoolDevice CPUDevice;
#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
typedef Eigen::GpuDevice GPUDevice;
#endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM

enum AxisArgumentName { NAME_IS_AXIS, NAME_IS_CONCAT_DIM };

// --------------------------------------------------------------------------
template <typename Device, typename T, AxisArgumentName AxisArgName>
class ConcatBaseOp : public OpKernel {
 public:
  typedef std::vector<std::unique_ptr<typename TTypes<T, 2>::ConstMatrix>>
      ConstMatrixVector;

  explicit ConcatBaseOp(OpKernelConstruction* c)
      : OpKernel(c),
        axis_attribute_name_(AxisArgName == NAME_IS_AXIS ? "axis"
                             : AxisArgName == NAME_IS_CONCAT_DIM
                                 ? "concat_dim"
                                 : "<invalid>") {
    int unused;
    OP_REQUIRES_OK(
        c, InputRange(axis_attribute_name_, &axis_input_index_, &unused));
    OP_REQUIRES_OK(c, InputRange("values", &values_input_start_index_,
                                 &values_input_end_index_));
  }

  void Compute(OpKernelContext* c) override {
    const Tensor& concat_dim_tensor = c->input(axis_input_index_);

    // TODO(rmlarsen): Disallow legacy use of length-1 vectors as scalars.
    OP_REQUIRES(c,
                (TensorShapeUtils::IsScalar(concat_dim_tensor.shape()) ||
                 (TensorShapeUtils::IsVector(concat_dim_tensor.shape()) &&
                  concat_dim_tensor.shape().dim_size(0) == 1)),
                errors::InvalidArgument(
                    axis_attribute_name_,
                    " tensor should be a scalar integer, but got shape ",
                    concat_dim_tensor.shape().DebugString()));
    int64_t concat_dim;
    // In case of ConcatV2, "axis" could be int32 or int64
    if (AxisArgName == NAME_IS_AXIS) {
      OP_REQUIRES(
          c,
          (concat_dim_tensor.dtype() == DT_INT32 ||
           concat_dim_tensor.dtype() == DT_INT64),
          errors::InvalidArgument(axis_attribute_name_,
                                  " tensor should be int32 or int64, but got ",
                                  DataTypeString(concat_dim_tensor.dtype())));
    } else {
      OP_REQUIRES(c, (concat_dim_tensor.dtype() == DT_INT32),
                  errors::InvalidArgument(
                      axis_attribute_name_, " tensor should be int32, but got ",
                      DataTypeString(concat_dim_tensor.dtype())));
    }
    if (concat_dim_tensor.dtype() == DT_INT32) {
      concat_dim =
          internal::SubtleMustCopy(concat_dim_tensor.scalar<int32>()());
    } else {
      concat_dim =
          internal::SubtleMustCopy(concat_dim_tensor.scalar<int64_t>()());
    }

    const int N = values_input_end_index_ - values_input_start_index_;
    const Tensor& first_input = c->input(values_input_start_index_);
    const int input_dims = first_input.dims();
    const TensorShape& input_shape = first_input.shape();

    int32_t axis = concat_dim < 0 ? concat_dim + input_dims : concat_dim;
    // concat_dim==0 allows concatenating a list of scalars into a vector.
    OP_REQUIRES(c, (0 <= axis && axis < input_dims) || concat_dim == 0,
                errors::InvalidArgument(
                    "ConcatOp : Expected concatenating dimensions in the range "
                    "[",
                    -input_dims, ", ", input_dims, "), but got ", concat_dim));
    // Note that we reduce the concat of n-dimensional tensors into a two
    // dimensional concat. Assuming the dimensions of any input/output
    // tensor are {x0, x1,...,xn-1, y0, y1,...,ym-1}, where the concat is along
    // the dimension indicated with size y0, we flatten it to {x, y}, where y =
    // Prod_i(yi) and x = ((n > 0) ? Prod_i(xi) : 1).
    ConstMatrixVector inputs_flat;
    inputs_flat.reserve(N);
    int64_t inputs_flat_dim0 = 1;
    for (int d = 0; d < axis; ++d) {
      inputs_flat_dim0 *= input_shape.dim_size(d);
    }
    int64_t output_concat_dim = 0;
    for (int i = 0; i < N; ++i) {
      const auto& in = c->input(values_input_start_index_ + i);
      OP_REQUIRES(
          c, in.dims() > 0,
          errors::InvalidArgument("ConcatOp : Can't concatenate scalars "
                                  "(use tf.stack instead)"));
      OP_REQUIRES(
          c, in.dims() == input_dims,
          errors::InvalidArgument(
              "ConcatOp : Ranks of all input tensors should match: shape[0] = ",
              input_shape.DebugString(), " vs. shape[", i,
              "] = ", in.shape().DebugString()));
      for (int j = 0; j < input_dims; ++j) {
        if (j == axis) {
          continue;
        }
        OP_REQUIRES(
            c, in.dim_size(j) == input_shape.dim_size(j),
            errors::InvalidArgument("ConcatOp : Dimension ", j,
                                    " in both shapes must be equal: "
                                    "shape[0] = ",
                                    input_shape.DebugString(), " vs. shape[", i,
                                    "] = ", in.shape().DebugString()));
      }
      if (in.NumElements() > 0) {
        int64_t inputs_flat_dim1 = in.NumElements() / inputs_flat_dim0;
        inputs_flat.emplace_back(new typename TTypes<T, 2>::ConstMatrix(
            in.template shaped<T, 2>({inputs_flat_dim0, inputs_flat_dim1})));
      }
      // TODO(rmlarsen): Remove check once !allow_legacy_scalars()?
      output_concat_dim += in.dims() > 0 ? in.dim_size(axis) : 1;
    }

    TensorShape output_shape(input_shape);
    // TODO(rmlarsen): Remove rank 0 case once !allow_legacy_scalars()?
    if (output_shape.dims() == 0) {
      OP_REQUIRES_OK(c, output_shape.AddDimWithStatus(output_concat_dim));
    } else {
      output_shape.set_dim(axis, output_concat_dim);
    }
    Tensor* output = nullptr;
    OP_REQUIRES_OK(c, c->allocate_output(0, output_shape, &output));
    if (output->NumElements() > 0) {
      int64_t output_dim1 = output->NumElements() / inputs_flat_dim0;
      auto output_flat = output->shaped<T, 2>({inputs_flat_dim0, output_dim1});
#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
      if (std::is_same<Device, GPUDevice>::value) {
        ConcatGPU<T>(c, inputs_flat, output, &output_flat);
        return;
      }
#endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
      ConcatCPU<T>(c->device(), inputs_flat, &output_flat);
    }
  }

 private:
  const char* const axis_attribute_name_;
  int axis_input_index_;
  int values_input_start_index_;
  int values_input_end_index_;
};

template <typename Device, typename T>
using ConcatOp = ConcatBaseOp<Device, T, NAME_IS_CONCAT_DIM>;
template <typename Device, typename T>
using ConcatV2Op = ConcatBaseOp<Device, T, NAME_IS_AXIS>;

#define REGISTER_CONCAT(type)                            \
  REGISTER_KERNEL_BUILDER(Name("Concat")                 \
                              .Device(DEVICE_CPU)        \
                              .TypeConstraint<type>("T") \
                              .HostMemory("concat_dim"), \
                          ConcatOp<CPUDevice, type>)     \
  REGISTER_KERNEL_BUILDER(Name("ConcatV2")               \
                              .Device(DEVICE_CPU)        \
                              .TypeConstraint<type>("T") \
                              .HostMemory("axis"),       \
                          ConcatV2Op<CPUDevice, type>)

TF_CALL_POD_STRING_TYPES(REGISTER_CONCAT);
TF_CALL_float8_e5m2(REGISTER_CONCAT);
TF_CALL_float8_e4m3fn(REGISTER_CONCAT);
REGISTER_CONCAT(quint8);
REGISTER_CONCAT(qint8);
REGISTER_CONCAT(quint16);
REGISTER_CONCAT(qint16);
REGISTER_CONCAT(qint32);

#undef REGISTER_CONCAT

#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM

#define REGISTER_GPU(type)                               \
  REGISTER_KERNEL_BUILDER(Name("Concat")                 \
                              .Device(DEVICE_GPU)        \
                              .TypeConstraint<type>("T") \
                              .HostMemory("concat_dim"), \
                          ConcatOp<GPUDevice, type>)     \
  REGISTER_KERNEL_BUILDER(Name("ConcatV2")               \
                              .Device(DEVICE_GPU)        \
                              .TypeConstraint<type>("T") \
                              .HostMemory("axis"),       \
                          ConcatV2Op<GPUDevice, type>)

TF_CALL_INTEGRAL_TYPES_NO_INT32(REGISTER_GPU);
TF_CALL_GPU_ALL_TYPES(REGISTER_GPU);
TF_CALL_float8_e5m2(REGISTER_GPU);
TF_CALL_float8_e4m3fn(REGISTER_GPU);
#undef REGISTER_GPU

#endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM

// A special DEVICE_DEFAULT kernel for int32.
// TODO(b/25387198): Also enable int32 in device memory. This kernel
// registration requires all int32 inputs and outputs to be in host memory.
REGISTER_KERNEL_BUILDER(Name("Concat")
                            .Device(DEVICE_DEFAULT)
                            .TypeConstraint<int32>("T")
                            .HostMemory("concat_dim")
                            .HostMemory("values")
                            .HostMemory("output"),
                        ConcatOp<CPUDevice, int32>);
REGISTER_KERNEL_BUILDER(Name("ConcatV2")
                            .Device(DEVICE_DEFAULT)
                            .TypeConstraint<int32>("T")
                            .HostMemory("values")
                            .HostMemory("axis")
                            .HostMemory("output"),
                        ConcatV2Op<CPUDevice, int32>);

template <typename ShapeType>
class ConcatOffsetOp : public OpKernel {
 public:
  explicit ConcatOffsetOp(OpKernelConstruction* ctx) : OpKernel(ctx) {}

  void Compute(OpKernelContext* ctx) override {
    const Tensor& concat_dim = ctx->input(0);
    OP_REQUIRES(
        ctx, TensorShapeUtils::IsScalar(concat_dim.shape()),
        absl::InvalidArgumentError(absl::StrCat(
            "Concat dim tensor should be a scalar integer, but got shape ",
            concat_dim.shape().DebugString())));
    for (int i = 1; i < ctx->num_inputs(); ++i) {
      const Tensor& inp = ctx->input(i);
      OP_REQUIRES(ctx, TensorShapeUtils::IsVector(inp.shape()),
                  absl::InvalidArgumentError(absl::StrCat(
                      "input ", i, " should be a vector, but got shape ",
                      inp.shape().DebugString())));
    }
    // Suppose a Concat() op needs to Concatenate N tensors, each of
    // which has the same number of dimensions.  Their shapes match
    // except the concat dimension.
    //
    // E.g., say, we want to concatenate 3 tensors in the 2nd
    // dimension, and their shapes are:
    //
    //  [2, 2, 5, 7]
    //  [2, 3, 5, 7]
    //  [2, 4, 5, 7]
    //
    // Here, N=3, cdim=1, dims=4. The concatenated tensor has shape
    // [2,9,5,7]. We will compute the cumulative sum along the 2nd
    // dimension to figure out each input's offset in the concatenated
    // output:
    //  [0, 0, 0, 0]
    //  [0, 2, 0, 0]
    //  [0, 5, 0, 0]
    const int32_t N = ctx->num_inputs() - 1;
    const Tensor& inp0 = ctx->input(1);
    auto inp0_vec = inp0.vec<ShapeType>();
    const int64_t cdim = internal::SubtleMustCopy(concat_dim.scalar<int32>()());
    const int64_t dims = inp0.NumElements();
    int32_t axis = cdim < 0 ? cdim + dims : cdim;
    OP_REQUIRES(ctx, FastBoundsCheck(axis, dims),
                absl::InvalidArgumentError(absl::StrCat(
                    "Concat dim is out of range: ", cdim, " vs. ", dims)));
    int64_t offset = 0;
    for (int i = 0; i < N; ++i) {
      const Tensor& inp = ctx->input(1 + i);
      OP_REQUIRES(ctx, dims == inp.NumElements(),
                  absl::InvalidArgumentError(
                      absl::StrCat("input ", i, " should contain ", dims,
                                   " elements, but got ", inp.NumElements())));
      auto inp_vec = inp.vec<ShapeType>();
      Tensor* out = nullptr;
      OP_REQUIRES_OK(ctx, ctx->allocate_output(i, {dims}, &out));
      auto out_vec = out->vec<ShapeType>();
      for (int64_t j = 0; j < dims; ++j) {
        if (j == axis) {
          out_vec(j) = offset;
          offset += inp_vec(j);
        } else {
          OP_REQUIRES(ctx, (inp0_vec(j) == inp_vec(j)),
                      absl::InvalidArgumentError(absl::StrCat(
                          "All dimensions except ", axis, " must match. Input ",
                          i, " has shape [", inp.SummarizeValue(10),
                          "] and doesn't match input 0 with shape [",
                          inp0.SummarizeValue(10), "].")));
          out_vec(j) = 0;
        }
      }
    }
  }

  bool IsExpensive() override { return false; }
};

REGISTER_KERNEL_BUILDER(
    Name("ConcatOffset").Device(DEVICE_CPU).TypeConstraint<int32>("shape_type"),
    ConcatOffsetOp<int32>);
REGISTER_KERNEL_BUILDER(Name("ConcatOffset")
                            .Device(DEVICE_DEFAULT)
                            .TypeConstraint<int32>("shape_type")
                            .HostMemory("concat_dim")
                            .HostMemory("shape")
                            .HostMemory("offset"),
                        ConcatOffsetOp<int32>);
REGISTER_KERNEL_BUILDER(
    Name("ConcatOffset").Device(DEVICE_CPU).TypeConstraint<int64>("shape_type"),
    ConcatOffsetOp<int64>);
REGISTER_KERNEL_BUILDER(Name("ConcatOffset")
                            .Device(DEVICE_DEFAULT)
                            .TypeConstraint<int64>("shape_type")
                            .HostMemory("concat_dim")
                            .HostMemory("shape")
                            .HostMemory("offset"),
                        ConcatOffsetOp<int64>);

}  // namespace tensorflow