Function vector_add_cuda_core

python/1_GettingStarted/vectorAdd/vectorAdd.py:70–155

Perform vector addition using cuda.core API. Parameters ---------- num_elements : int Number of elements in each vector device_id : int CUDA device ID to use verify : bool Whether to verify the result Returns ------- bool True if

(num_elements=50000, device_id=0, verify=True)

def vector_add_cuda_core(num_elements=50000, device_id=0, verify=True):
    """
    Perform vector addition using cuda.core API.

    Parameters
    ----------
    num_elements : int
        Number of elements in each vector
    device_id : int
        CUDA device ID to use
    verify : bool
        Whether to verify the result

    Returns
    -------
    bool
        True if successful, False otherwise
    """
    try:
        # Initialize device
        print("[Vector addition using CUDA Core API]")
        device = Device(device_id)
        device.set_current()

        print(f"Device: {device.name}")
        print(f"Compute Capability: sm_{device.arch}")

        stream = device.create_stream()

        # Compile kernel
        print("Compiling kernel 'vectorAdd<float>'...")
        program_options = ProgramOptions(std="c++17", arch=f"sm_{device.arch}")
        program = Program(VECTOR_ADD_KERNEL, code_type="c++", options=program_options)
        module = program.compile("cubin", name_expressions=("vectorAdd<float>",))
        kernel = module.get_kernel("vectorAdd<float>")
        print("Kernel compiled successfully")

        # Allocate and initialize vectors
        print(f"[Vector addition of {num_elements} elements]")
        dtype = cp.float32

        a = cp.random.rand(num_elements).astype(dtype)
        b = cp.random.rand(num_elements).astype(dtype)
        c = cp.empty(num_elements, dtype=dtype)

        # Synchronize before kernel launch
        device.sync()

        # Configure and launch kernel
        threads_per_block = 256
        blocks_per_grid = (num_elements + threads_per_block - 1) // threads_per_block

        print(
            f"CUDA kernel launch with {blocks_per_grid} blocks "
            f"of {threads_per_block} threads"
        )

        config = LaunchConfig(grid=blocks_per_grid, block=threads_per_block)