tensor_algebra

Headless dense tensor algebra: fill a vector, then a tensor GEMV through the same retained scheme. There is no window and no capture clip.

cargo run --example tensor_algebra

CUDA compute-only (no graphics):

cargo run --no-default-features --features cuda,tensor --example tensor_algebra

What it demonstrates

TensorKernels / TensorRecorder recording into an ordinary Scheme
Checked views and semantic matmul (cuBLAS / MPS / stdlib)
Host observation through MemoryExchange

Source

examples/tensor_algebra.rs:

//! Headless dense tensor algebra: fill, add with broadcast, and GEMV.

use goldy::{
    Instance, RequestAdapterOptions, RuntimeDescriptor, Scheme, Tensor, TensorDType, TensorKernels, TensorScalar,
    TensorShape,
};
use std::ops::Shr;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let runtime = Instance::new()?
        .request_adapter(&RequestAdapterOptions::default())?
        .request_runtime(&RuntimeDescriptor::default())?;
    let ctx = runtime.create_context()?;
    let kernels = TensorKernels::new(&runtime)?;
    let x = Tensor::from_f32(&runtime, TensorShape::vector(4), &[1.0, 2.0, 3.0, 4.0])?;
    let w = Tensor::from_f32(
        &runtime,
        TensorShape::matrix(2, 4),
        &[1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0],
    )?;
    let mut scheme = Scheme::new(&ctx);
    let y = {
        let mut rec = kernels.recorder(&mut scheme);
        rec.fill("fill_bias", x.view(), TensorScalar::F32(1.0))?;
        rec.matmul("gemv", w.view(), x.view())?
    };
    drop(kernels);

    let mut sub = scheme.submit()?;
    let bytes = (&mut sub >> y.buffer()).take::<u8>()?;
    let out: &[f32] = bytemuck::cast_slice(&bytes);
    println!("gemv(W, ones) = {out:?}");
    assert_eq!(out, &[1.0, 1.0]);
    let _ = TensorDType::F32;
    Ok(())
}

Goldy

tensor_algebra

What it demonstrates

Source