xref: /external/tensorflow/tensorflow/compiler/jit/xla_launch_util.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/compiler/jit/xla_launch_util.h"

#include <memory>

#include "absl/algorithm/container.h"
#include "absl/memory/memory.h"
#include "tensorflow/compiler/jit/defs.h"
#include "tensorflow/compiler/tf2xla/shape_util.h"
#include "tensorflow/compiler/tf2xla/xla_compiler.h"
#include "tensorflow/compiler/xla/client/client_library.h"
#include "tensorflow/compiler/xla/client/local_client.h"
#include "tensorflow/compiler/xla/statusor.h"
#include "tensorflow/core/common_runtime/dma_helper.h"
#include "tensorflow/core/common_runtime/function.h"
#include "tensorflow/core/common_runtime/gpu_device_context.h"
#include "tensorflow/core/framework/allocator.h"
#include "tensorflow/core/framework/node_def_util.h"
#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/resource_mgr.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/core/refcount.h"
#include "tensorflow/core/util/stream_executor_util.h"

namespace tensorflow {
namespace {
using xla::ScopedShapedBuffer;
using xla::ShapedBuffer;

}  // anonymous namespace

VariableInfo::VariableInfo(int index, absl::string_view name, Var* var)
    : index_(index), name_(name), var_(var) {}
VariableInfo::VariableInfo(VariableInfo&& other)
    : index_(other.index_), var_(other.var_), lock_held_(other.lock_held_) {
  other.index_ = -1;
  other.var_ = nullptr;
}

VariableInfo& VariableInfo::operator=(VariableInfo&& other) {
  index_ = other.index_;
  var_ = other.var_;
  lock_held_ = other.lock_held_;

  other.index_ = -1;
  other.var_ = nullptr;

  return *this;
}

VariableInfo::~VariableInfo() {
  // Release the variable's lock if we hold it. Ensures that the lock is
  // released even on error.  It does not matter in what order we release the
  // locks.
  if (var()) {
    if (lock_held()) {
      var()->mu()->unlock();
    }

    // Unref the variable so it can be released by ResourceManager.
    var()->Unref();
  }
}

Status GetVariableInfosFromInputs(ResourceMgr* rm, DeviceBase* dev,
                                  absl::Span<const Tensor* const> inputs,
                                  absl::Span<const int> variable_indices,
                                  std::vector<VariableInfo>* result) {
  result->clear();
  result->reserve(variable_indices.size());
  for (int var_idx : variable_indices) {
    Var* variable = nullptr;
    ResourceHandle handle = inputs[var_idx]->flat<ResourceHandle>()(0);
    if (handle.device() != dev->attributes().name()) {
      std::string definition_location = [&]() -> std::string {
        if (handle.definition_stack_trace()) {
          std::vector<StackFrame> stack_frames =
              handle.definition_stack_trace()->ToStackFrames(
                  {}, IsInternalFrameForFilename,
                  /*reverse_traversal=*/true,
                  /*limit=*/1);
          if (!stack_frames.empty()) {
            const StackFrame& last_frame = stack_frames[0];
            return absl::StrCat(" (defined @ ", last_frame.file_name, ":",
                                last_frame.line_number, ")");
          }
        }
        return "";
      }();
      return errors::InvalidArgument("Trying to access resource ",
                                     handle.name(), definition_location,
                                     " located in device ", handle.device(),
                                     " from device ", dev->attributes().name());
    }
    TF_RETURN_IF_ERROR(rm->LookupOrCreate<Var>(
        handle.container(), handle.name(), &variable, [](Var** ptr) {
          // This var is uninitialized for now.
          *ptr = new Var(DT_INVALID);
          return Status::OK();
        }));
    result->emplace_back(var_idx, handle.name(), variable);
  }
  return Status::OK();
}

std::vector<const Tensor*> InputsFromContext(OpKernelContext* ctx) {
  std::vector<const Tensor*> inputs;
  inputs.reserve(ctx->num_inputs());
  for (int input_idx = 0; input_idx < ctx->num_inputs(); input_idx++) {
    inputs.push_back(&ctx->input(input_idx));
  }
  return inputs;
}

Status LockVariables(absl::Span<VariableInfo> variables) {
  std::vector<int> lock_order(variables.size());
  std::iota(lock_order.begin(), lock_order.end(), 0);

  // VariableInfoComparator orders all empty VariableInfo instances as
  // equivalent so it looks like we may want to stable sort these to maintain a
  // deterministic order between the empty VariableInfo instances.  However
  // since we're sorting by pointer value the sort is pretty non-deterministic
  // anyway so we don't bother using std::stable_sort for now.
  absl::c_sort(lock_order, [&](int a, int b) {
    if (variables[a].var() && variables[b].var()) {
      return variables[a].var()->mu() < variables[b].var()->mu();
    }

    // Move all the empty VariableInfo instances to the end.
    return variables[a].var() != nullptr;
  });

  mutex* prev = nullptr;
  for (int i : lock_order) {
    Var* variable = variables[i].var();
    if (variable == nullptr) {
      // All empty VariableInfo instances are at the end of the order
      // so we're done.
      break;
    }
    mutex* mu = variable->mu();
    if (prev == mu) {
      // It is an error to pass the same variable handle twice to the same XLA
      // cluster because we would not handle variable updates correctly.  Any
      // locks we have already acquired will be released when the VariableInfo
      // objects are destroyed.
      // TODO(b/128495870) Add support for passing aliased resource variables.
      return errors::Unimplemented("Duplicate variable passed to XLA cluster");
    }
    VLOG(4) << "Acquiring lock for variable "
            << reinterpret_cast<void*>(variable);
    mu->lock();
    variables[i].set_lock_held();
    prev = mu;
  }
  VLOG(4) << "Finished acquiring variable locks.";
  return Status::OK();
}

Status SnapshotResourceVariables(OpKernelContext* ctx,
                                 absl::Span<const int> variable_indices,
                                 absl::Span<VariableInfo const> variable_infos,
                                 ResourceVarsSnapshot* result) {
  for (int i = 0, end = variable_indices.size(); i < end; i++) {
    Var* var = variable_infos[i].var();
    (*result)[variable_indices[i]] =
        var ? absl::make_optional(*var->tensor()) : absl::nullopt;
  }
  return Status::OK();
}

XlaComputationLaunchContext::XlaComputationLaunchContext(
    xla::LocalClient* client, se::DeviceMemoryAllocator* xla_allocator,
    int device_ordinal, bool allocate_xla_tensors, bool use_multiple_streams)
    : client_(client),
      xla_allocator_(xla_allocator),
      allocate_xla_tensors_(allocate_xla_tensors),
      use_multiple_streams_(use_multiple_streams),
      device_ordinal_(device_ordinal) {
  if (use_multiple_streams_) {
    CHECK(allocate_xla_tensors_) << "To use multiple streams correctly we must "
                                    "be allocating XLA tensors!";
  }
}

// Fills in `execution_input` with `buffer` for `index`.
static void PopulateExecutionInputBuffer(xla::ExecutionInput& execution_input,
                                         xla::ShapeIndex index,
                                         se::DeviceMemoryBase& buffer,
                                         bool donate_buffer, int device_ordinal,
                                         se::DeviceMemoryAllocator* allocator) {
  xla::MaybeOwningDeviceMemory* in_buffer =
      execution_input.MutableBuffer(index);
  if (donate_buffer) {
    *in_buffer = se::OwningDeviceMemory(buffer, device_ordinal, allocator);
    buffer = se::DeviceMemoryBase();
  } else {
    *in_buffer = buffer;
  }
}

xla::StatusOr<std::vector<xla::ExecutionInput>>
XlaComputationLaunchContext::PopulateInputs(
    OpKernelContext* ctx,
    const XlaCompiler::CompilationResult* compilation_result,
    const std::map<int, const Tensor*>& resource_vars,
    int missing_ctx_input_prefix,
    const xla::HloInputOutputAliasConfig& input_output_alias) {
  std::vector<xla::ExecutionInput> arguments;
  arguments.reserve(compilation_result->xla_input_shapes.size());

  xla::TransferManager* transfer_manager =
      client_->backend().transfer_manager();
  for (int i = 0, end = compilation_result->xla_input_shapes.size(); i < end;
       ++i) {
    int arg_num = compilation_result->input_mapping[i];
    CHECK_GE(arg_num, missing_ctx_input_prefix);
    const xla::Shape& shape = compilation_result->xla_input_shapes[i];
    const xla::Shape& device_shape =
        transfer_manager->HostShapeToDeviceShape(shape);

    bool is_resource_variable = resource_vars.count(arg_num);
    bool is_updated_resource_variable =
        is_resource_variable &&
        absl::c_any_of(compilation_result->resource_updates,
                       [&](const XlaCompiler::ResourceUpdate& update) {
                         return update.input_index == i && update.modified;
                       });

    const Tensor* t = is_resource_variable
                          ? resource_vars.at(arg_num)
                          : &(ctx->input(arg_num - missing_ctx_input_prefix));
    CHECK(t);
    bool donate_buffer =
        t->RefCountIsOne() && is_updated_resource_variable &&
        input_output_alias.ParameterHasAlias(i, xla::ShapeIndex{});
    VLOG(3) << "Processing input: " << i
            << "; is_resource_variable=" << is_resource_variable
            << "; is_updated_resource_variable=" << is_updated_resource_variable
            << "; donate_buffer=" << donate_buffer;

    if (use_multiple_streams_) {
      CHECK(ctx->op_device_context() && ctx->op_device_context()->stream())
          << "Must have a stream available when using XLA tensors!";
      XlaTensor* xla_tensor = XlaTensor::FromTensor(t);
      CHECK(xla_tensor);
      xla_tensor->WaitForDefinitionEventOnStream(
          ctx->op_device_context()->stream());
    }

    arguments.emplace_back(device_shape, shape);
    xla::ExecutionInput& execution_input = arguments.back();
    if (xla::Shape::Equal().MinorToMajorOnlyInLayout()(shape, device_shape)) {
      se::DeviceMemoryBase dmem = XlaTensor::DeviceMemoryFromTensor(*t);
      PopulateExecutionInputBuffer(execution_input, xla::ShapeIndex{}, dmem,
                                   donate_buffer, device_ordinal_,
                                   xla_allocator_);
    } else {
      XlaTensor* xla_tensor = XlaTensor::FromTensor(t);
      CHECK(xla_tensor && xla_tensor->has_shaped_buffer());
      xla_tensor->shaped_buffer().buffers().ForEachMutableElement(
          [&](const xla::ShapeIndex& index, se::DeviceMemoryBase* buffer) {
            PopulateExecutionInputBuffer(execution_input, index, *buffer,
                                         donate_buffer, device_ordinal_,
                                         xla_allocator_);
          });
    }
  }
  return std::move(arguments);
}

// Construct the tensor for the given type and buffer.
static Tensor MakeTensor(DataType dtype, const TensorShape& shape,
                         se::DeviceMemoryBase buffer, Allocator* allocator) {
  size_t expected_size = shape.num_elements() * DataTypeSize(dtype);
  auto* tensor_buffer = new XlaTensorBuffer(buffer.opaque(), expected_size,
                                            buffer.size(), allocator);
  Tensor t(dtype, shape, tensor_buffer);
  tensor_buffer->Unref();
  return t;
}

// Get aliased tensor, or make a new one for the corresponding output operation.
static Tensor GetOrCreateTensorForOutput(
    int output_num, OpKernelContext* ctx, int missing_ctx_input_prefix,
    const xla::HloInputOutputAliasConfig& input_output_alias,
    absl::Span<const int> input_mapping,
    const std::map<int, const Tensor*>& resource_vars_snapshots,
    DataType output_dtype, const TensorShape& output_shape,
    se::DeviceMemoryBase output_buffer, Allocator* output_allocator) {
  xla::ShapeIndex output_index = input_output_alias.shape().IsTuple()
                                     ? xla::ShapeIndex({output_num})
                                     : xla::ShapeIndex({});

  CHECK(input_output_alias.shape().IsTuple() || output_num == 0);
  if (absl::optional<xla::HloInputOutputAliasConfig::Alias> alias =
          input_output_alias.GetAliasedParameter(output_index)) {
    VLOG(3) << "Found alias: " << alias->ToString();
    int tf_param =
        input_mapping[alias->parameter_number] - missing_ctx_input_prefix;
    const Tensor input_tensor =
        ctx->input(tf_param).dtype() != DT_RESOURCE
            ? ctx->input(tf_param)
            : *resource_vars_snapshots.at(missing_ctx_input_prefix + tf_param);
    if (output_buffer.opaque() == input_tensor.data()) {
      return input_tensor;
    }
  }
  return MakeTensor(output_dtype, output_shape, output_buffer,
                    output_allocator);
}

static void PopulateXlaTensor(Tensor* output_tensor,
                              xla::ScopedShapedBuffer* output, int output_num,
                              se::Stream* stream, bool use_multiple_streams,
                              std::shared_ptr<se::Event> definition_event) {
  XlaTensor* xla_tensor = XlaTensor::FromTensor(output_tensor);
  CHECK(xla_tensor);
  xla_tensor->set_shaped_buffer(output->TakeSubTree({output_num}));
  if (use_multiple_streams) {
    xla_tensor->ResetDefinitionEvent(definition_event, stream);
  }
}

// Sets output `output_num` for `ctx` provided it is known at a compile time.
static Status SetOutputForConstant(
    OpKernelContext* ctx, se::Stream* stream,
    const XlaCompiler::CompilationResult* compilation_result, int output_num) {
  CHECK(compilation_result->outputs[output_num].is_constant);
  const Tensor& const_tensor =
      compilation_result->outputs[output_num].constant_value;
  Tensor* output_tensor;
  if (stream && const_tensor.TotalBytes() > 0) {
    // Copy host -> device. (Empty tensors don't have backing buffers.)
    // Manually allocate memory using an XlaTensorBuffer so we can allocate
    // as much memory as the device requires (as given by
    // GetByteSizeRequirement). This avoids XlaTransferManager having to
    // reallocate the device buffer later.
    VLOG(1) << "Constant output tensor on device";

    TF_RETURN_IF_ERROR(
        ctx->allocate_output(output_num, const_tensor.shape(), &output_tensor));
    Device* device = dynamic_cast<Device*>(ctx->device());
    if (device == nullptr) {
      return errors::Internal("DeviceBase was not a Device.");
    }
    ctx->op_device_context()->CopyCPUTensorToDevice(
        &const_tensor, device, output_tensor,
        [&](Status status) { TF_CHECK_OK(status); });

    if (device->device_type() == DEVICE_GPU) {
      // The GPUDeviceContext enqueues the host->device transfer in a
      // separate stream from the main compute stream. We must ensure the
      // compute stream is synchronized with the host->device transfer
      // stream now otherwise we will create a race condition.
      auto* gpu_device_context =
          static_cast<GPUDeviceContext*>(ctx->op_device_context());
      gpu_device_context->stream()->ThenWaitFor(
          gpu_device_context->host_to_device_stream());
    }
  } else {
    // No copy required.
    ctx->set_output(output_num, const_tensor);
    output_tensor = ctx->mutable_output(output_num);
  }
  return Status::OK();
}

static xla::StatusOr<Var*> GetOrCreateResourceVar(
    OpKernelContext* ctx, const ResourceHandle& handle,
    const XlaCompiler::ResourceUpdate& write) {
  Var* variable = nullptr;
  TF_RETURN_IF_ERROR(
      LookupOrCreateResource<Var>(ctx, handle, &variable, [&write](Var** ptr) {
        *ptr = new Var(write.type);
        return Status::OK();
      }));
  return variable;
}

xla::StatusOr<std::vector<VariableInfo>> GatherVariableInfo(
    OpKernelContext* ctx,
    const XlaCompiler::CompilationResult& compilation_result,
    int missing_ctx_input_prefix) {
  std::vector<VariableInfo> out;
  out.reserve(compilation_result.resource_updates.size());
  for (int i = 0; i < compilation_result.resource_updates.size(); ++i) {
    const XlaCompiler::ResourceUpdate& write =
        compilation_result.resource_updates[i];
    int actual_input_index = write.input_index - missing_ctx_input_prefix;
    if (actual_input_index < 0 || actual_input_index >= ctx->num_inputs()) {
      return errors::Internal("Invalid input index for variable write.");
    }

    const ResourceHandle handle = HandleFromInput(ctx, actual_input_index);
    TF_ASSIGN_OR_RETURN(Var * variable,
                        GetOrCreateResourceVar(ctx, handle, write));
    out.emplace_back(actual_input_index, handle.name(), variable);
  }
  return std::move(out);
}

Status XlaComputationLaunchContext::PopulateOutputs(
    OpKernelContext* ctx,
    const XlaCompiler::CompilationResult* compilation_result,
    ScopedShapedBuffer output, int missing_ctx_input_prefix,
    absl::Span<VariableInfo> variable_infos,
    const xla::HloInputOutputAliasConfig& input_output_alias,
    const std::map<int, const Tensor*>& resource_vars) {
  se::Stream* stream =
      ctx->op_device_context() ? ctx->op_device_context()->stream() : nullptr;
  Allocator* allocator = ctx->device()->GetAllocator({});

  // Computation output should always be a tuple.
  VLOG(2) << "Result tuple shape: " << output.on_host_shape().DebugString();
  VLOG(2) << "Result tuple shape (on device): "
          << output.on_device_shape().DebugString();
  CHECK_EQ(ctx->num_outputs(), compilation_result->outputs.size());

  // If the on-host-shape isn't a tuple, create a new single-element tuple
  // buffer with a nullptr root index table. This allows the code below to treat
  // output as a tuple unconditionally.
  if (!output.on_host_shape().IsTuple()) {
    ShapedBuffer nontuple_buffer = output.release();
    ShapedBuffer buffer(
        xla::ShapeUtil::MakeTupleShape({nontuple_buffer.on_host_shape()}),
        xla::ShapeUtil::MakeTupleShape({nontuple_buffer.on_device_shape()}),
        output.device_ordinal());
    buffer.buffers().CopySubtreeFrom(nontuple_buffer.buffers(),
                                     /*source_base_index=*/{},
                                     /*target_base_index=*/{0});
    output = ScopedShapedBuffer(std::move(buffer), output.memory_allocator());
  }

  std::shared_ptr<se::Event> definition_event;
  if (use_multiple_streams_) {
    definition_event = std::make_shared<se::Event>(stream->parent());
    if (!definition_event->Init()) {
      return errors::Internal("Failed to initialize tensor definition event.");
    }
    stream->ThenRecordEvent(definition_event.get());
  }

  std::vector<TensorShape> output_tensor_shapes;
  output_tensor_shapes.reserve(ctx->num_outputs());
  if (output.on_host_shape().is_dynamic()) {
    TF_ASSIGN_OR_RETURN(
        auto transfer_manager,
        xla::TransferManager::GetForPlatform(stream->parent()->platform()));

    xla::Shape output_device_shape = output.on_device_shape();
    TF_RETURN_IF_ERROR(transfer_manager->ReadDynamicShapes(
        stream, &output, &output_device_shape));

    output.set_shapes(output_device_shape, output_device_shape);
    for (int i = 0; i < ctx->num_outputs(); ++i) {
      const xla::Shape& subshape =
          xla::ShapeUtil::GetSubshape(output_device_shape, {i});
      TensorShape shape;
      TF_RETURN_IF_ERROR(XLAShapeToTensorShape(subshape, &shape));
      output_tensor_shapes.push_back(shape);
    }
  } else {
    for (int i = 0; i < ctx->num_outputs(); ++i) {
      output_tensor_shapes.push_back(compilation_result->outputs[i].shape);
    }
  }

  // Copy XLA results to the OpOutputList.
  int output_num = 0;
  for (int i = 0, end = ctx->num_outputs(); i < end; ++i) {
    const TensorShape& shape = output_tensor_shapes[i];
    const DataType& type = compilation_result->outputs[i].type;
    VLOG(2) << "Populating output for retval " << i << " shape "
            << shape.DebugString() << " type " << DataTypeString(type);
    if (type == DT_VARIANT) {
      return errors::Unimplemented(
          "Support for TensorList crossing the XLA/TF boundary "
          "is not implemented");
    }

    if (compilation_result->outputs[i].is_constant) {
      TF_RETURN_IF_ERROR(
          SetOutputForConstant(ctx, stream, compilation_result, i));
    } else if (type == DT_RESOURCE) {
      int input_index =
          compilation_result->outputs[i].input_index - missing_ctx_input_prefix;
      TF_RET_CHECK(input_index >= 0 && input_index < ctx->num_inputs())
          << "Invalid input for outputs " << i << ": " << input_index;
      ctx->set_output(i, ctx->input(input_index));
    } else {
      if (allocate_xla_tensors_) {
        Tensor* output_tensor;
        TF_RETURN_IF_ERROR(ctx->allocate_output(i, shape, &output_tensor));
        if (output_tensor->TotalBytes() > 0) {
          PopulateXlaTensor(output_tensor, &output, output_num, stream,
                            use_multiple_streams_, definition_event);
        }
      } else {
        se::DeviceMemoryBase buffer = output.buffer({output_num});
        Tensor output_tensor = GetOrCreateTensorForOutput(
            output_num, ctx, missing_ctx_input_prefix, input_output_alias,
            compilation_result->input_mapping, resource_vars,
            ctx->expected_output_dtype(i), shape, buffer, allocator);
        ctx->set_output(i, output_tensor);
      }
      output.set_buffer(se::OwningDeviceMemory(), {output_num});
      ++output_num;
    }
  }

  // input_index -> index into variable_infos.
  absl::flat_hash_map<int, int> variable_info_lookup;
  for (int i = 0; i < variable_infos.size(); i++) {
    variable_info_lookup.emplace(variable_infos[i].index(), i);
  }

  // Apply variable updates, if any.
  for (int i = 0, end = compilation_result->resource_updates.size(); i < end;
       ++i) {
    const XlaCompiler::ResourceUpdate& write =
        compilation_result->resource_updates[i];
    int actual_input_index = write.input_index - missing_ctx_input_prefix;
    CHECK_GE(actual_input_index, 0);
    CHECK_LT(actual_input_index, ctx->num_inputs());
    Var* var = variable_infos[variable_info_lookup[actual_input_index]].var();
    CHECK(var);

    VLOG(2) << "Updating variable #" << i
            << " at input index: " << actual_input_index << " with shape "
            << write.shape.DebugString() << "; variable tensor has shape: "
            << var->tensor()->shape().DebugString();

    if (var->is_initialized && var->tensor()->dtype() != write.type) {
      return errors::Internal("Mismatched type in variable write");
    }

    Tensor output_tensor;
    if (allocate_xla_tensors_) {
      TF_RETURN_IF_ERROR(
          ctx->allocate_temp(write.type, write.shape, &output_tensor));
      if (write.shape.num_elements() > 0) {
        PopulateXlaTensor(&output_tensor, &output, output_num, stream,
                          use_multiple_streams_, definition_event);
      }
    } else {
      se::DeviceMemoryBase buffer = output.buffer({output_num});
      output_tensor = GetOrCreateTensorForOutput(
          output_num, ctx, missing_ctx_input_prefix, input_output_alias,
          compilation_result->input_mapping, resource_vars, write.type,
          write.shape, buffer, allocator);
    }
    output.set_buffer(se::OwningDeviceMemory(), {output_num});
    var->is_initialized |= write.modified;
    *var->tensor() = output_tensor;
    ++output_num;
  }
  return Status::OK();
}

xla::StatusOr<std::vector<XlaCompiler::Argument>>
XlaComputationLaunchContext::BuildXlaCompilerArguments(
    absl::Span<int const> must_be_constant_idxs,
    absl::Span<const Tensor* const> inputs,
    absl::Span<VariableInfo const> variable_args, Device* device) {
  CHECK(absl::c_is_sorted(must_be_constant_idxs));
  std::vector<XlaCompiler::Argument> out;
  out.resize(inputs.size());

  // TODO(cheshire): Avoid duplication with framework/op_kernel.h
  DeviceContext* device_context = nullptr;
  TF_RETURN_IF_ERROR(device->TryGetDeviceContext(&device_context));
  bool using_default_context = false;
  auto cleanup = xla::MakeCleanup([&] {
    if (device_context != nullptr && !using_default_context) {
      device_context->Unref();
    }
  });
  if (device_context == nullptr) {
    using_default_context = true;
    auto* dev_info = device->tensorflow_gpu_device_info();
    if (dev_info) device_context = dev_info->default_context;
  }

  absl::flat_hash_map<int, const VariableInfo*> variable_info_lookup;
  for (const VariableInfo& info : variable_args) {
    CHECK(!info.var() || info.lock_held())
        << "Need to hold the lock on resource variables "
           "before calling BuildXlaCompilerArguments";
    variable_info_lookup.emplace(info.index(), &info);
  }

  for (int64 input_num = 0; input_num < inputs.size(); ++input_num) {
    const Tensor* input = inputs[input_num];

    XlaCompiler::Argument& arg = out[input_num];
    if (variable_info_lookup.count(input_num)) {
      // Handles resource variables.
      TF_RET_CHECK(input->dtype() == DT_RESOURCE);
      const VariableInfo& variable = *variable_info_lookup[input_num];
      arg.name = std::string(variable.name());
      arg.kind = XlaCompiler::Argument::kResource;
      arg.resource_kind = XlaResource::kVariable;
      if (variable.var() && variable.var()->is_initialized) {
        const Tensor* value = variable.var()->tensor();
        arg.type = value->dtype();
        arg.shape = value->shape();
        arg.initialized = true;
      } else {
        // The values of uninitialized variables are not passed as inputs, since
        // they are meaningless. However, it is legal to assign to a resource
        // variable for the first time inside the XLA computation, so we do
        // permit uninitialized variables.
        arg.initialized = false;
        arg.type = DT_INVALID;
        arg.shape = TensorShape();
      }

      if (absl::c_binary_search(must_be_constant_idxs, input_num)) {
        TF_RET_CHECK(variable.var() && variable.var()->is_initialized);
        const Tensor* value = variable.var()->tensor();
        Tensor value_on_host(value->dtype(), value->shape());
        if (!device_context) {
          value_on_host = *value;
        } else {
          TF_RETURN_IF_ERROR(device_context->CopyDeviceTensorToCPUSync(
              value, "", device, &value_on_host));
        }
        arg.kind = XlaCompiler::Argument::kConstantResource;
        arg.constant_value = value_on_host;
      }
    } else if (absl::c_binary_search(must_be_constant_idxs, input_num)) {
      arg.kind = XlaCompiler::Argument::kConstant;
      arg.type = input->dtype();
      arg.shape = input->shape();
      arg.constant_value = *input;
    } else {
      // Normal inputs.
      TF_RET_CHECK(input->dtype() != DT_RESOURCE);
      if (input->NumElements() > 0) {
        arg.kind = XlaCompiler::Argument::kParameter;
      } else {
        arg.kind = XlaCompiler::Argument::kConstant;
        arg.constant_value = *input;
      }
      arg.type = input->dtype();
      arg.shape = input->shape();
    }
  }

  return out;
}

}  // namespace tensorflow