01_scale_numbers

Scale Numbers

Demonstrates how to pass scalar values to a CubeCL kernel using ScalarArg.

Concepts Demonstrated

• Passing scalar parameters to kernels with ScalarArg::new()
• Creating reusable helper functions with #[cube] (not #[cube(launch)])
• Backend-agnostic kernel launching using generics with Runtime trait bounds

Key Code

#[cube(launch)]
fn kernel_scale_numbers(input_data: &Array<u32>, scale: u32, output_data: &mut Array<u32>) {
    let index = CUBE_POS * CUBE_DIM + UNIT_POS;
    output_data[index] = input_data[index] * scale;
}

// Launch with ScalarArg for the scale parameter
kernel_scale_numbers::launch::<R>(
    &client,
    CubeCount::Static(1, 1, 1),
    CubeDim::new(num_elements as u32, 1, 1),
    ArrayArg::from_raw_parts::<u32>(input, num_elements, 1),
    ScalarArg::new(scale),  // <-- Scalar value passed here
    ArrayArg::from_raw_parts::<u32>(output, num_elements, 1),
)

Backend-Agnostic Pattern

The launch_scale_numbers_kernel function uses a generic R: Runtime parameter, allowing the same kernel launching logic to work across different GPU backends (WGPU, CUDA) without code changes.

Running

cargo run --bin 01_scale_numbers