TFE_Op* MatMulOp(TFE_Context* ctx, TFE_TensorHandle* a, TFE_TensorHandle* b);

// Return an identity op.
TFE_Op* IdentityOp(TFE_Context* ctx, TFE_TensorHandle* a);

// Return a shape op fetching the shape of `a`.
TFE_Op* ShapeOp(TFE_Context* ctx, TFE_TensorHandle* a);

// Return an allreduce op adding up input tensor `in` from `group_size` workers.
TFE_Op* AllReduceOp(TFE_Context* ctx, TFE_TensorHandle* in, int group_size);

  TFE_OpAddInput(op, a, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TF_DeleteStatus(status);
  TFE_OpSetAttrType(op, "T", TFE_TensorHandleDataType(a));

  return op;
}

TFE_Op* ShapeOp(TFE_Context* ctx, TFE_TensorHandle* a) {
  TF_Status* status = TF_NewStatus();

  TFE_Op* op = TFE_NewOp(ctx, "Shape", status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

    TFE_TensorHandle* hgpu = TFE_TensorHandleCopyToDevice(
        hcpu, ctx, gpu_device_name.c_str(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

    TFE_Op* shape_op = ShapeOp(ctx, hgpu);
    TFE_OpSetDevice(shape_op, gpu_device_name.c_str(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
    TFE_TensorHandle* retvals[1];
    int num_retvals = 1;

    TFE_TensorHandle* hgpu = TFE_TensorHandleCopyToDevice(
        hcpu, ctx, gpu_device_name.c_str(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

    TFE_Op* shape_op = ShapeOp(ctx, hgpu);
    TFE_OpSetDevice(shape_op, gpu_device_name.c_str(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
    TFE_TensorHandle* retvals[1];
    int num_retvals = 1;