// the operation being timed:
//
//	start := time.Now()
//	... operation that takes 20 milliseconds ...
//	t := time.Now()
//	elapsed := t.Sub(start)
//
// Other idioms, such as [time.Since](start), [time.Until](deadline), and
// time.Now().Before(deadline), are similarly robust against wall clock
// resets.
//
// The rest of this section gives the precise details of how operations

  const int num_dims = 2;
  int64_t* dims = new int64_t[num_dims];
  int64_t num_elements = 1;
  dims[0] = batch_size;
  dims[1] = 1;
  for (int64_t i = 0; i < num_dims; ++i) {
    num_elements *= dims[i];
  }
  TF_Tensor* t = TF_AllocateTensor(TF_STRING, dims, num_dims,
                                   sizeof(TF_TString) * num_elements);
  delete[] dims;

  %5 = stablehlo.dynamic_broadcast_in_dim %1, %4, dims = [3] : (tensor<16xf32>, tensor<4xindex>) -> tensor<?x28x28x16xf32>
  %6 = stablehlo.add %3, %5 : tensor<?x28x28x16xf32>
  %7 = shape.shape_of %6 : tensor<?x28x28x16xf32> -> tensor<4xindex>
  %8 = stablehlo.dynamic_broadcast_in_dim %2, %7, dims = [] : (tensor<f32>, tensor<4xindex>) -> tensor<?x28x28x16xf32>

#else
  const char* device_name_ = tensorflow::DEVICE_CPU;
#endif
};

// Validates that the tensor has shape and type corresponding to
// dims and dtype.
void validate_tensor(TF_Tensor* tensor, int64_t* dims, int64_t num_dims,
                     TF_DataType dtype);

// Copies data of length tensor_size_bytes from values to tensor.
template <typename T>

// Returns the shape of the Tensor referenced by `output` in `graph`
// into `dims`. `dims` must be an array large enough to hold `num_dims`
// entries (e.g., the return value of TF_GraphGetTensorNumDims).
//
// If the number of dimensions in the shape is unknown or the shape is
// a scalar, `dims` will remain untouched. Otherwise, each element of
// `dims` will be set corresponding to the size of the dimension. An

      %2 = stablehlo.reshape %1 : (tensor<i32>) -> tensor<1xi32>
      %3 = stablehlo.concatenate %2, %cst_0, dim = 0 : (tensor<1xi32>, tensor<1xi32>) -> tensor<2xi32>
      %4 = stablehlo.dynamic_broadcast_in_dim %arg2, %3, dims = [1] : (tensor<3xf32>, tensor<2xi32>) -> tensor<?x3xf32>
      %5 = stablehlo.add %0, %4 : tensor<?x3xf32>
      return %5 : tensor<?x3xf32>
    }
}

  TFE_DeleteContextOptions(opts);

  std::vector<int64_t> dims(4, 1);
  TFE_Op* op = TFE_NewOp(ctx, "AvgPool", status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TF_Tensor* tensor =
      TF_AllocateTensor(TF_FLOAT, dims.data(), dims.size(), sizeof(float));
  float tensor_data[] = {1};

	// to a 48-bit address space like every other arm64 platform.
	//
	// WebAssembly currently has a limit of 4GB linear memory.
	heapAddrBits = (_64bit*(1-goarch.IsWasm)*(1-goos.IsIos*goarch.IsArm64))*48 + (1-_64bit+goarch.IsWasm)*(32-(goarch.IsMips+goarch.IsMipsle)) + 40*goos.IsIos*goarch.IsArm64

	// maxAlloc is the maximum size of an allocation. On 64-bit,
	// it's theoretically possible to allocate 1<<heapAddrBits bytes. On

   * building caches whose key or value types are incompatible with the types accepted by the
   * weigher already provided; the {@code CacheBuilder} type cannot do this. For best results,
   * simply use the standard method-chaining idiom, as illustrated in the documentation at top,
   * configuring a {@code CacheBuilder} and building your {@link Cache} all in a single statement.
   *

		block *Block    // current handled block
		state walkState // what's to do
	}
	work := make([]bp, 0, 256)
	work = append(work, bp{
		block: f.Entry,
		state: descend,
	})

	idom := f.Idom()
	sdom := f.Sdom()

	// DFS on the dominator tree.
	//
	// For efficiency, we consider only the dominator tree rather
	// than the entire flow graph. On the way down, we consider