Skip to main content

Module offload

Module offload 

Source
🔬This is a nightly-only experimental API. (gpu_offload #131513)
Expand description

This module provides support for gpu offloading. For technical details regarding the offload_kernel and how to use it, see their respective documentation.

§General usage

The offload_kernel macro can be applied to a function to generate the necessary code to launch a kernel on the target device.

#[offload_kernel]
fn kernel(x: *mut [f64; 256]) {
    // SAFETY:
    // calling our `arch` functions and dereferencing a raw pointer is unsafe
    unsafe {
        let n = (*x).len();
        let i = (thread_idx_x() + block_idx_x() * block_dim_x()) as usize;
        if i < n {
            (*x)[i] = i as f64;
        }
    }
}

To launch an offloaded kernel, the only current way is to use the core::intrinsic::offload intrinsic (note that intrinsics usage is discouraged outside the standard library). This allows you to specify grid and block dimensions and pass the required arguments to the device.

let mut x = [0.0f64; 256];
core::intrinsics::offload::<_, _, ()>(kernel, [256, 1, 1], [1, 1, 1], (&mut x as *mut [f64; 256],));

For precise information on the offload intrinsic, see its respective documentation.

§Current limitations:

  • Usage is restricted to types supported by the current device-mapping implementation.
  • Generics and functions accepting dyn Trait are not supported.
  • Kernel execution is currently restricted to intrinsics usage, which is discouraged outside of the standard library.

Modules§

offloadExperimental
This module provides support for gpu offloading. For technical details regarding the offload_kernel and how to use it, see their respective documentation.

Attribute Macros§

offload_kernelExperimental
The offload_kernel macro is applied to a function to generate two separate definitions: a host-side wrapper for dispatch and a device-side kernel.