if (TF_GetCode(s) != TF_OK) {
      return;
    }
  }
  int num_outputs = unwrap(o)->expected_num_outputs;
  tsl::Set_TF_Status_from_Status(
      s, unwrap(op)->Execute(
             absl::MakeSpan(reinterpret_cast<AbstractTensorHandle**>(
                                unwrap(o)->outputs.data()),
                            unwrap(o)->outputs.size()),
             &num_outputs));
}

  parallel_device.Execute(context.get(), handle_inputs, "ReadVariableOp",
                          TFE_OpGetAttrs(read_op.get()),
                          /*expected_max_outputs=*/1, status.get());
  ASSERT_FALSE(TF_GetCode(status.get()) == TF_OK);
  TF_SetStatus(status.get(), TF_OK, "");

  // Check that ops still run successfully on the device.
  parallel_device.Execute(context.get(), std::vector<ParallelTensor*>(),

  TFE_DeleteTensorHandle(h0_task0);
  TFE_DeleteTensorHandle(retvals[0]);

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TF_DeleteStatus(status);
}

// Read the value of variable `var` and save it into `out_value`.

      "SetAttrFunctionList has not been "
      "implemented yet.");
}
Status Execute(AbstractOperation* op_, AbstractContext* ctx,
               absl::Span<AbstractTensorHandle*> retvals, int* num_retvals,
               ForwardOperation* forward_op_, Tape* tape,
               const GradientRegistry& registry) {
  TF_RETURN_IF_ERROR(op_->Execute(retvals, num_retvals));
  for (int i = 0; i < *num_retvals; i++) {

  EXPECT_EQ(10, product[1]);
  EXPECT_EQ(15, product[2]);
  EXPECT_EQ(22, product[3]);
  TF_DeleteStatus(status);
}
TEST(CAPI, Execute_MatMul_CPU) { Execute_MatMul_CPU(false); }
TEST(CAPI, Execute_MatMul_CPUAsync) { Execute_MatMul_CPU(true); }

void Execute_MatMul_CPU_Runtime_Error(bool async) {
  TF_Status* status = TF_NewStatus();
  TFE_ContextOptions* opts = TFE_NewContextOptions();

  // Block until the previous `StartExecute` operation has executed. Forwards
  // the status from `TFE_Execute` and returns outputs if the status is OK.
  std::vector<TensorHandlePtr> Join(TF_Status* status);

  // Block until all Ops finished running on the thread.
  void AsyncWait(TF_Status* status);

 private:
  void Run();

  void Execute(TFE_Context* context, const char* operation_name,

#include "tensorflow/c/c_api_experimental.h"
#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

namespace tensorflow {

  TFE_DeleteTensorHandle(h0);
  TFE_DeleteTensorHandle(h1);
  TFE_DeleteTensorHandle(h2);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TFE_ContextRemoveFunction(ctx, "AddVariablesFunction", status);
  TFE_DeleteContext(ctx);

  TF_AbstractTensor* inputs[2] = {at, at};
  TF_OutputList* o = TF_NewOutputList();
  TF_OutputListSetNumOutputs(o, 1, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  // Execute.
  TF_ExecuteOperation(op, 2, inputs, o, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  // Clean up operation and inputs.
  TF_DeleteAbstractOp(op);

  TF_DeleteFunction(fn);

  for (bool async : {false, true, false}) {
    TFE_Executor* old_executor = TFE_ContextGetExecutorForThread(ctx);
    TFE_Executor* executor = TFE_NewExecutor(
        /*is_async=*/async, /*enable_streaming_enqueue=*/true,
        /*in_flight_nodes_limit=*/0);
    TFE_ContextSetExecutorForThread(ctx, executor);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

    TF_Tensor* t =