string fn_name = "two_adds";
  TF_ExecutionContext* graph_ctx = TF_CreateFunction(fn_name.c_str(), s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  auto* arg0 = TF_AddFunctionParameter(graph_ctx, TF_FLOAT, {-1, nullptr}, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
  auto* arg1 = TF_AddFunctionParameter(graph_ctx, TF_FLOAT, {-1, nullptr}, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

      // But in gcd(a - b, b), a - b is even and b is odd, so we can divide out powers of two.

      // We bend over backwards to avoid branching, adapting a technique from
      // http://graphics.stanford.edu/~seander/bithacks.html#IntegerMinOrMax

      long delta = a - b; // can't overflow, since a and b are nonnegative

      long minDeltaOrZero = delta & (delta >> (Long.SIZE - 1));

build:cuda_clang --config=cuda
build:cuda_clang --@local_config_cuda//:cuda_compiler=clang
build:cuda_clang --copt=-Qunused-arguments
# Select supported compute capabilities (supported graphics cards).
# This is the same as the official TensorFlow builds.
# See https://developer.nvidia.com/cuda-gpus#compute
# `compute_XY` enables PTX embedding in addition to SASS. PTX

      // But in gcd(a - b, b), a - b is even and b is odd, so we can divide out powers of two.

      // We bend over backwards to avoid branching, adapting a technique from
      // http://graphics.stanford.edu/~seander/bithacks.html#IntegerMinOrMax

      long delta = a - b; // can't overflow, since a and b are nonnegative

      long minDeltaOrZero = delta & (delta >> (Long.SIZE - 1));