bool CanUseGpuDevice(const RuntimeDevices &devices);

// Returns true if all of the GPUs available at runtime support TensorCores
// (NVIDIA compute capability >= 7.0).
bool CanUseTensorCores(const RuntimeDevices &devices);

// Returns true if operation does not have explicit device placement that would
// prevent it from running on GPU device.
bool CanUseGpuDevice(Operation *op);

}  // namespace TF

  return llvm::any_of(devices.device_names(), IsGpuDevice);
}

// Returns true if all of the GPUs available at runtime support TensorCores
// (NVIDIA compute capability >= 7.0).
bool CanUseTensorCores(const RuntimeDevices &devices) {
  auto has_tensor_cores = [&](const DeviceNameUtils::ParsedName &device) {
    auto md = devices.GetGpuDeviceMetadata(device);
    return md ? md->getCcMajor() >= 7 : false;
  };

  // For f16 data type on devices with Tensor Cores support NHWC data format
  // is up to ~2x faster.
  const bool is_f16 = input_ty.getElementType().isF16();
  if (is_f16 && CanUseTensorCores(devices)) return "NHWC";

  // For f32/f16 data type decision depends on the filter size in spatial
  // dimensions, for other data types we keep current data format.