values[0] = v;
  return TF_NewTensor(TF_FLOAT, nullptr, 0, values, num_bytes,
                      &FloatDeallocator, nullptr);
}

// All the *Helper methods are used as a workaround for the restrictions that
// one cannot call ASSERT_* methods in non-void-returning functions (when
// exceptions are disabled during compilation)
void PlaceholderHelper(TF_Graph* graph, TF_Status* s, const char* name,

  bool valid = true;
  int64_t expected_stride = 1;
  for (int i = ndim - 1; i >= 0; --i) {
    // Empty tensors are always compact regardless of strides.
    if (shape_arr[i] == 0) return true;
    // Note that dimensions with size=1 can have any stride.
    if (shape_arr[i] != 1 && stride_arr[i] != expected_stride) {
      valid = false;
    }
    expected_stride *= shape_arr[i];
  }
  return valid;
}

==============================================================================*/
// This file extends/implements core graph optimizer base classes in terms of
// the C API defined in grappler.h. A class "CSomething" represents a
// "Something" that can be manipulated via calls in the C interface and a C
// struct called "TP_Something".

#include "tensorflow/c/experimental/grappler/grappler.h"

#include <algorithm>
#include <cstddef>
#include <cstring>

}

TEST(TensorPjRtBufferUtilTest, GetPjRtCBufferFromTensorNoBuffer) {
  auto allocator = std::make_unique<AsyncValueAllocator>();
  tensorflow::Tensor tensor(allocator.get(), DT_FLOAT, {1});

  EXPECT_THAT(
      GetPjRtCBufferFromTensor(&tensor),
      StatusIs(error::INTERNAL, HasSubstr(absl::StrCat(
                                    "Input tensor does not have PjRtBuffer"))));
}

      ExecutionState::kIdle;
  // Tells the worker thread that there is new work.
  tensorflow::condition_variable start_execute_;
  // The worker thread notifies that work has finished.
  tensorflow::condition_variable finished_execute_;
  // Notifies a StartExecute that the previous Join has finished.
  tensorflow::condition_variable finished_join_;

  TFE_TensorHandle* h0 = TestVariable(ctx, 1.0, task1_name);
  TFE_TensorHandle* h1 = TestVariable(ctx, 2.0, task2_name);
  TFE_TensorHandle* h2 = TestVariable(ctx, 3.0, task0_name);

  // Add a sync point to make sure that variables have been initialized
  // before the function execution starts.
  TFE_ContextAsyncWait(ctx, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  ExpectScalarEq<float>(result_components[1].get(), 3.);
}

TEST(PARALLEL_DEVICE, TestCollectiveSync) { TestCollective(/*async=*/false); }

// Note that ops on the parallel device currently don't execute
// asynchronously. The test is just that we don't get deadlocks.
TEST(PARALLEL_DEVICE, TestCollectiveAsync) { TestCollective(/*async=*/true); }

void RegisterCollectiveMulFunction(TFE_Context* context,

        //
        // TODO(allenl): There may be smarter ways to do this copy in some
        // cases, i.e. with a collective broadcast. We'll need to be careful
        // about things that are taken as inputs on the host or on their
        // existing device (for multi-device functions).
        std::unique_ptr<ParallelTensor> parallel_tensor(
            parallel_device.CopyToParallelDevice(

    variable->Assign(context, initial_value.get(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  }

  // Read from the variable and verify that we have a parallel tensor.
  {
    TensorHandlePtr read = variable->Read(context, status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
    std::array<TensorHandlePtr, 2> components;

                                    fn_name, "'");
    output_tensors->emplace_back(node, idx);
  }
  return absl::OkStatus();
}

// Populates `body_nodes` with the nodes that will become function's body.
// Performs various checks.
Status ComputeBodyNodes(
    const TF_Graph* fn_body, const char* fn_name, int num_opers,
    const TF_Operation* const* opers,