xref: /external/tensorflow/tensorflow/core/kernels/self_adjoint_eig_v2_op_gpu.cc

/* Copyright 2017 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/linalg_ops.cc.

#if GOOGLE_CUDA

#include <numeric>
#include <type_traits>

#define EIGEN_USE_GPU
#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
#include "tensorflow/core/framework/kernel_def_builder.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/kernels/cast_op.h"
#include "tensorflow/core/kernels/cuda_solvers.h"
#include "tensorflow/core/kernels/cwise_ops.h"
#include "tensorflow/core/kernels/transpose_functor.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/types.h"

namespace tensorflow {

typedef Eigen::GpuDevice GPUDevice;

template <class Scalar>
class SelfAdjointEigV2OpGpu : public AsyncOpKernel {
 public:
  explicit SelfAdjointEigV2OpGpu(OpKernelConstruction* context)
      : AsyncOpKernel(context) {
    OP_REQUIRES_OK(context, context->GetAttr("compute_v", &compute_v_));
  }

  void ComputeAsync(OpKernelContext* context, DoneCallback done) final {
    const Tensor& input = context->input(0);
    const int ndims = input.dims();
    OP_REQUIRES_ASYNC(
        context, ndims >= 2,
        errors::InvalidArgument("Input must have rank >= 2, got ", ndims),
        done);
    const int64 n = input.dim_size(ndims - 1);
    OP_REQUIRES_ASYNC(
        context, input.dim_size(ndims - 2) == n,
        errors::InvalidArgument("Input matrices must be squares, got",
                                input.dim_size(ndims - 2), " != ", n),
        done);
    const int64 batch_size =
        input.template flat_inner_dims<Scalar, 3>().dimension(0);

    // Allocate outputs.
    Tensor* eigenvalues;
    TensorShape eigenvalues_shape = input.shape();
    eigenvalues_shape.RemoveLastDims(1);
    OP_REQUIRES_OK_ASYNC(
        context, context->allocate_output(0, eigenvalues_shape, &eigenvalues),
        done);
    Tensor* eigenvectors;
    TensorShape eigenvectors_shape =
        compute_v_ ? input.shape() : TensorShape({});
    OP_REQUIRES_OK_ASYNC(
        context, context->allocate_output(1, eigenvectors_shape, &eigenvectors),
        done);

    if (input.NumElements() == 0) {
      done();
      return;
    }

    // Allocate workspace.
    // TODO(rmlarsen): Convert to std::make_unique when available.
    std::unique_ptr<CudaSolver> solver(new CudaSolver(context));
    Tensor eigenvalues_real;
    using RealScalar = typename Eigen::NumTraits<Scalar>::Real;
    if (std::is_same<Scalar, RealScalar>::value) {
      eigenvalues_real = *eigenvalues;
    } else {
      OP_REQUIRES_OK_ASYNC(
          context,
          solver->allocate_scoped_tensor(DataTypeToEnum<RealScalar>::value,
                                         eigenvalues_shape, &eigenvalues_real),
          done);
    }

    Tensor input_copy;
    OP_REQUIRES_OK_ASYNC(
        context,
        solver->forward_input_or_allocate_scoped_tensor(
            {0}, DataTypeToEnum<Scalar>::value, input.shape(), &input_copy),
        done);
    // For real symmetric matrices, row-major and column-major are the same. For
    // complex Hermitian, row-major and column-major differ by a conjugation,
    // which is still cheaper than a transpose.
    const GPUDevice& device = context->eigen_device<GPUDevice>();
    if (!input.SharesBufferWith(input_copy)) {
      if (Eigen::NumTraits<Scalar>::IsComplex) {
        functor::UnaryFunctor<GPUDevice, functor::conj<Scalar>> conj;
        conj(device, input_copy.flat<Scalar>() /*out*/,
             input.flat<Scalar>() /*in*/);
      } else {
        device.memcpy(input_copy.flat<Scalar>().data(),
                      input.flat<Scalar>().data(),
                      input.NumElements() * sizeof(Scalar));
      }
    } else if (Eigen::NumTraits<Scalar>::IsComplex) {
      functor::UnaryFunctor<GPUDevice, functor::conj<Scalar>> conj;
      conj(device, const_cast<Tensor*>(&input)->flat<Scalar>() /*out*/,
           input.flat<Scalar>() /*in*/);
    }

    // Compute eigen decomposition in-place in input_copy.
    std::vector<DeviceLapackInfo> dev_info;
    dev_info.push_back(solver->GetDeviceLapackInfo(batch_size, "heevd"));
    auto input_copy_reshaped = input_copy.flat_inner_dims<Scalar, 3>();
    auto eigenvalues_real_reshaped =
        eigenvalues_real.flat_inner_dims<RealScalar, 2>();
    for (int batch = 0; batch < batch_size; ++batch) {
      OP_REQUIRES_OK_ASYNC(
          context,
          solver->Heevd(compute_v_ ? CUSOLVER_EIG_MODE_VECTOR
                                   : CUSOLVER_EIG_MODE_NOVECTOR,
                        CUBLAS_FILL_MODE_UPPER, n,
                        &input_copy_reshaped(batch, 0, 0), n,
                        &eigenvalues_real_reshaped(batch, 0),
                        dev_info.back().mutable_data() + batch),
          done);
    }

    if (!std::is_same<Scalar, RealScalar>::value) {
      functor::CastFunctor<GPUDevice, Scalar, RealScalar> cast;
      cast(device, eigenvalues->flat<Scalar>(),
           const_cast<const Tensor*>(&eigenvalues_real)->flat<RealScalar>());
    }

    if (compute_v_) {
      // Transpose eigenvectors now stored in input_copy in column-major form to
      // output in row-major form.
      OP_REQUIRES_OK_ASYNC(
          context, DoMatrixTranspose(device, input_copy, eigenvectors), done);
    }

    // Asynchronously check return status from cuSolver kernels.
    CudaSolver::CheckLapackInfoAndDeleteSolverAsync(std::move(solver), dev_info,
                                                    std::move(done));
  }

 private:
  bool compute_v_;

  TF_DISALLOW_COPY_AND_ASSIGN(SelfAdjointEigV2OpGpu);
};

#define REGISTER(Scalar)                                                       \
  REGISTER_KERNEL_BUILDER(                                                     \
      Name("SelfAdjointEigV2").Device(DEVICE_GPU).TypeConstraint<Scalar>("T"), \
      (SelfAdjointEigV2OpGpu<Scalar>))

REGISTER(float);
REGISTER(double);
REGISTER(complex64);
REGISTER(complex128);

#undef REGISTER

}  // namespace tensorflow

#endif  // GOOGLE_CUDA