sparse_slice_grad_op.cc

/* Copyright 2018 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.
==============================================================================*/

#include "tensorflow/core/kernels/sparse_slice_grad_op.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_util.h"
#include "tensorflow/core/framework/types.h"

namespace tensorflow {

typedef Eigen::ThreadPoolDevice CPUDevice;

namespace functor {

template <typename T>
struct SparseSliceGradFunctor<CPUDevice, T> {
  void operator()(OpKernelContext *ctx,
                  typename TTypes<T>::ConstFlat backprop_val_grad,
                  typename TTypes<int64_t>::ConstMatrix input_indices_mat,
                  typename TTypes<int64_t>::ConstFlat input_start_flat,
                  typename TTypes<int64_t>::ConstMatrix output_indices_mat,
                  typename TTypes<T>::Flat val_grad) const {
    const int64_t input_nnz = input_indices_mat.dimension(0);
    const int num_dims = input_indices_mat.dimension(1);

    T *val_grad_flat = val_grad.data();
    const T *backprop_val_grad_flat = backprop_val_grad.data();
    memset(val_grad_flat, 0, sizeof(T) * input_nnz);

    // Fill gradients for position where indices of input and output are same.
    int64_t j = 0;
    for (int64_t i = 0; i < input_nnz && j < backprop_val_grad.dimension(0);
         ++i) {
      bool is_same = true;
      for (int d = 0; d < num_dims; ++d) {
        const int64_t a = input_indices_mat(i, d);
        const int64_t b = output_indices_mat(j, d);
        const int64_t offset = input_start_flat(d);
        if (a != b + offset) {
          is_same = false;
          break;
        }
      }
      if (is_same) {
        val_grad_flat[i] = backprop_val_grad_flat[j];
        ++j;
      }
    }
    OP_REQUIRES(
        ctx, backprop_val_grad.dimension(0) == j,
        errors::Internal("Elements of backprop_val_grad aren't all propagated. "
                         "Num elements:",
                         backprop_val_grad.dimension(0), ", used: ", j));
  }
};

}  // namespace functor

template <typename Device, typename T>
class SparseSliceGradOp : public OpKernel {
 public:
  explicit SparseSliceGradOp(OpKernelConstruction *ctx) : OpKernel(ctx) {}

  void Compute(OpKernelContext *ctx) override {
    const Tensor *backprop_val_grad, *input_indices, *output_indices, *input_start;
    OP_REQUIRES_OK(ctx, ctx->input("backprop_val_grad", &backprop_val_grad));
    OP_REQUIRES_OK(ctx, ctx->input("input_indices", &input_indices));
    OP_REQUIRES_OK(ctx, ctx->input("input_start", &input_start));
    OP_REQUIRES_OK(ctx, ctx->input("output_indices", &output_indices));

    OP_REQUIRES(ctx,
                TensorShapeUtils::IsMatrix(input_indices->shape()) &&
                    TensorShapeUtils::IsMatrix(output_indices->shape()),
                errors::InvalidArgument(
                    "Input and output indices should be matrices "
                    "but received shapes: ",
                    input_indices->shape().DebugString(), " and ",
                    output_indices->shape().DebugString()));
    OP_REQUIRES(
        ctx, TensorShapeUtils::IsVector(backprop_val_grad->shape()),
        errors::InvalidArgument(
            "Input backprop_val_grad should be a vector but received shape: ",
            backprop_val_grad->shape().DebugString()));
    OP_REQUIRES(
        ctx,
        input_indices->dim_size(1) == output_indices->dim_size(1),
        errors::InvalidArgument("The input and output should have the same "
                                "ndims: got: ", input_indices->dim_size(1), " and ",
                                output_indices->dim_size(1)));
    OP_REQUIRES(
        ctx, output_indices->dim_size(0) <= input_indices->dim_size(0),
        errors::InvalidArgument("# rows of output_indices should be not greater "
                                "than of input_indices, got ",
                                output_indices->dim_size(0), " and ",
                                input_indices->dim_size(0)));
    OP_REQUIRES(
        ctx, backprop_val_grad->NumElements() == output_indices->dim_size(0),
        errors::InvalidArgument("# elements of backprop_val_grad and # rows of "
                                "output_indices should match (#nnz of sum): got ",
                                backprop_val_grad->NumElements(), " and ",
                                output_indices->dim_size(0)));
    OP_REQUIRES(ctx, TensorShapeUtils::IsVector(input_start->shape()),
                errors::InvalidArgument(
                    "The input_start should be a vector but received shape ",
                    input_start->shape().DebugString()));

    const int num_dims = input_indices->dim_size(1);
    OP_REQUIRES(ctx, num_dims == input_start->NumElements(),
                errors::InvalidArgument(
                    "Expected input_start to be a vector of length ", num_dims,
                    " but got length ", input_start->NumElements()));

    const int64_t input_nnz = input_indices->dim_size(0);

    Tensor *val_grad;
    OP_REQUIRES_OK(ctx,
                   ctx->allocate_output(0, TensorShape({input_nnz}), &val_grad));

    if (input_nnz == 0) return;

    functor::SparseSliceGradFunctor<Device, T>()(
        ctx, backprop_val_grad->flat<T>(), input_indices->matrix<int64_t>(),
        input_start->flat<int64_t>(), output_indices->matrix<int64_t>(),
        val_grad->flat<T>());
  }
};

#define REGISTER_KERNELS(type)                                              \
  REGISTER_KERNEL_BUILDER(                                                  \
      Name("SparseSliceGrad").Device(DEVICE_CPU).TypeConstraint<type>("T"), \
      SparseSliceGradOp<CPUDevice, type>)

TF_CALL_NUMBER_TYPES(REGISTER_KERNELS);
#undef REGISTER_KERNELS

#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM

typedef Eigen::GpuDevice GPUDevice;

#define REGISTER_KERNELS(type)                                              \
  REGISTER_KERNEL_BUILDER(                                                  \
      Name("SparseSliceGrad").Device(DEVICE_GPU).TypeConstraint<type>("T"), \
      SparseSliceGradOp<GPUDevice, type>)
TF_CALL_NUMBER_TYPES(REGISTER_KERNELS);
#undef REGISTER_KERNELS

#endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM

}  // namespace tensorflow