Adapt CUDA kernel raw pointers

You are given a CUDA kernel file that currently uses custom OpenCV CUDA types such as PtrStepSz for image data representation on the device. Your task is to refactor this CUDA kernel code to remove any dependencies on these custom OpenCV types and instead use only raw CUDA pointers for image and label data. Specifically, you should: - Replace all uses of OpenCV CUDA types (e.g., PtrStep, PtrStepSz) with raw pointer variables and explicitly manage parameters such as image width, height, step (pitch in bytes), and element size. - Adjust all indexing and pointer arithmetic accordingly, based on raw pointers and the explicitly passed dimensions and steps. - Implement a helper function that accepts a host std::vector<uint8_t> representing the input image data, uploads this image data to the GPU device memory, launches the adapted CUDA kernel(s) to process the image entirely on the device, and then downloads the resulting label matrix back to host memory as a std::vector<uint32_t> or an equivalent host-side container. - Make sure the entire workflow correctly handles memory allocation, copying to device, kernel launches, synchronization, and copying results back to host. - Preserve all original CUDA kernel logic but adapt only the interface and data management so that no OpenCV-specific CUDA types are used. Be thorough in your reasoning about the needed strides, pitches, and element sizes tracked previously by PtrStepSz and adjust kernel indexing to maintain correctness. # Steps 1. Identify all uses of OpenCV CUDA pointer types and their associated properties (step, elem_size). 2. Replace them with raw pointers (e.g., unsigned char* for input image, unsigned int* for output labels) alongside explicit parameters for width, height, and pitch (step). 3. Adjust all kernel indexing and pointer calculations to use the raw pointers and these parameters. 4. Implement a host-side helper function that: - Receives a host std::vector<uint8_t> containing the input image data. - Allocates device memory for the input image and output labels. - Copies the input image data to device memory. - Launches the kernel with the new pointer-based interface. - Copies the output labels back to a host vector. - Frees all device memory. 5. Ensure proper synchronization and error checking. # Output Format Provide the complete adapted CUDA kernel code and the helper function code in C++ that demonstrates the upload, kernel execution, and download of data. Include comments explaining the key changes and adaptations for clarity. # Notes - Assume the input image data is single-channel and stored as uint8_t (grayscale). - The output label matrix should be stored as a 2D array with one label value (such as uint32_t) per pixel. - All kernel calls and memory operations must use standard CUDA runtime API calls (no OpenCV CUDA functions). - Ensure that the pitch or step is handled correctly for alignment if necessary. # Response Formats Respond only with the adapted code, formatted for readability, including the helper function and any necessary declarations. No additional explanations outside code comments are needed.