Function run

python/2_CoreConcepts/simpleZeroCopy/simpleZeroCopy.py:97–222

Zero-copy vector add: map host memory, launch kernel with device pointers, validate on CPU. This function shows how to: 1. Allocate pinned (page-locked) host memory 2. Map host memory into GPU address space (zero-copy) 3. Access host memory directly from GPU kernel 4. V

(num_elements=1048576)

def run(num_elements=1048576):
    """
    Zero-copy vector add: map host memory, launch kernel with device
    pointers, validate on CPU.

    This function shows how to:
    1. Allocate pinned (page-locked) host memory
    2. Map host memory into GPU address space (zero-copy)
    3. Access host memory directly from GPU kernel
    4. Validate results

    Parameters
    ----------
    num_elements : int
        Number of elements in vectors (default: 1048576)
    """
    print("\n" + "=" * 70)
    print("simpleZeroCopy - CUDA Python Sample")
    print("=" * 70)

    # Initialize device
    device = Device()
    device.set_current()
    major, minor = device.compute_capability

    print("\nDevice Information:")
    print(f"  Name: {device.name}")
    print(f"  Compute Capability: {major}.{minor}")

    # Create stream
    stream = device.create_stream()
    mapped_host_ptrs = []

    try:
        print(
            "\n> Memory: mapped pinned host "
            "(cudaHostAlloc + cudaHostGetDevicePointer)"
        )

        print("\nCompiling CUDA kernel...")
        program_options = ProgramOptions(std="c++17", arch=f"sm_{device.arch}")
        prog = Program(VECTOR_ADD_KERNEL, code_type="c++", options=program_options)
        mod = prog.compile("cubin")
        kernel = mod.get_kernel("vectorAddGPU")
        print("  Kernel compiled successfully")

        bytes_total = num_elements * np.dtype(np.float32).itemsize
        print("\nAllocating memory:")
        print(f"  Vector size: {num_elements:,} elements")
        print(f"  Memory per vector: {bytes_total / (1024**2):.2f} MB")
        print(f"  Total memory: {3 * bytes_total / (1024**2):.2f} MB")

        print("\n> Allocating mapped pinned host memory...")
        h_a, d_a = _mapped_host_alloc(num_elements, stream)
        mapped_host_ptrs.append(h_a)
        h_b, d_b = _mapped_host_alloc(num_elements, stream)
        mapped_host_ptrs.append(h_b)
        h_c, d_c = _mapped_host_alloc(num_elements, stream)