krnl

Safe, portable, high performance compute (GPGPU) kernels.

Developed for autograph.

• Similar functionality to CUDA and OpenCL.
• Supports GPU's and other Vulkan 1.2 capable devices.
• MacOS / iOS supported via MoltenVK.
• Kernels are written inline, entirely in Rust.
  • Simple iterator patterns can be implemented without unsafe.
  • Supports inline SPIR-V assembly.
  • DebugPrintf integration, generates backtraces for panics.
• Buffers on the host can be accessed natively as Vecs and slices.

krnlc

Kernel compiler for krnl.

• Built on spirv-builder.
• Supports dependencies defined in Cargo.toml.
• Uses spirv-tools to validate and optimize.
• Compiles to "krnl-cache.rs", so the crate will build on stable Rust.

See the docs for installation and usage instructions.

Installing

For device functionality (kernels), install Vulkan for your platform.

• For development, it's recomended to install the LunarG Vulkan SDK, which includes additional tools:
  • vulkaninfo
  • Validation layers
    • DebugPrintf
  • spirv-tools
    • This is used by krnlc for spirv validation and optimization.
      • krnlc builds by default without needing spirv-tools to be installed.

Test

• Check that vulkaninfo --summary shows your devices.
  • Instance version should be >= 1.2.
• Alternatively, check that cargo test --test integration_tests -- --exact none shows your devices.
  • You can run all the tests with cargo test --all-features.

Getting Started

See the docs or build them locally with cargo doc --all-features.

Example

use krnl::{
    macros::module,
    anyhow::Result,
    device::Device,
    buffer::{Buffer, Slice, SliceMut},
};

#[module]
mod kernels {
    #[cfg(not(target_arch = "spirv"))]
    use krnl::krnl_core;
    use krnl_core::macros::kernel;

    pub fn saxpy_impl(alpha: f32, x: f32, y: &mut f32) {
        *y += alpha * x;
    }

    // Item kernels for iterator patterns.
    #[kernel]
    pub fn saxpy(alpha: f32, #[item] x: f32, #[item] y: &mut f32) {
        saxpy_impl(alpha, x, y);
    }

    // General purpose kernels like CUDA / OpenCL.
    #[kernel]
    pub fn saxpy_global(alpha: f32, #[global] x: Slice<f32>, #[global] y: UnsafeSlice<f32>) {
        use krnl_core::buffer::UnsafeIndex;

        let global_id = kernel.global_id();
        if global_id < x.len().min(y.len()) {
            saxpy_impl(alpha, x[global_id], unsafe { y.unsafe_index_mut(global_id) });
        }
    }
}

fn saxpy(alpha: f32, x: Slice<f32>, mut y: SliceMut<f32>) -> Result<()> {
    if let Some((x, y)) = x.as_host_slice().zip(y.as_host_slice_mut()) {
        x.iter()
            .copied()
            .zip(y.iter_mut())
            .for_each(|(x, y)| kernels::saxpy_impl(alpha, x, y));
        return Ok(());
    }
    if true {
        kernels::saxpy::builder()?
            .build(y.device())?
            .dispatch(alpha, x, y)
    } else {
        // or
        kernels::saxpy_global::builder()?
            .build(y.device())?
            .with_global_threads(y.len() as u32)
            .dispatch(alpha, x, y)
    }
}

fn main() -> Result<()> {
    let x = vec![1f32];
    let alpha = 2f32;
    let y = vec![0f32];
    let device = Device::builder().build().ok().unwrap_or(Device::host());
    let x = Buffer::from(x).into_device(device.clone())?;
    let mut y = Buffer::from(y).into_device(device.clone())?;
    saxpy(alpha, x.as_slice(), y.as_slice_mut())?;
    let y = y.into_vec()?;
    println!("{y:?}");
    Ok(())
}

Performance

NVIDIA GeForce GTX 1060 with Max-Q Design

benches/compute-benches

alloc

	krnl	cuda	ocl
1,000,000	316.90 ns (✅ 1.00x)	112.84 us (❌ 356.06x slower)	495.45 ns (❌ 1.56x slower)
10,000,000	318.15 ns (✅ 1.00x)	1.10 ms (❌ 3454.98x slower)	506.82 ns (❌ 1.59x slower)
64,000,000	317.56 ns (✅ 1.00x)	6.31 ms (❌ 19854.77x slower)	506.15 ns (❌ 1.59x slower)

upload

	krnl	cuda	ocl
1,000,000	332.66 us (✅ 1.00x)	359.18 us (✅ 1.08x slower)	773.51 us (❌ 2.33x slower)
10,000,000	4.83 ms (✅ 1.00x)	3.69 ms (✅ 1.31x faster)	8.76 ms (❌ 1.81x slower)
64,000,000	25.24 ms (✅ 1.00x)	24.34 ms (✅ 1.04x faster)	57.02 ms (❌ 2.26x slower)

download

	krnl	cuda	ocl
1,000,000	584.39 us (✅ 1.00x)	447.38 us (✅ 1.31x faster)	20.17 ms (❌ 34.52x slower)
10,000,000	5.67 ms (✅ 1.00x)	4.03 ms (✅ 1.41x faster)	20.15 ms (❌ 3.55x slower)
64,000,000	28.82 ms (✅ 1.00x)	25.57 ms (✅ 1.13x faster)	37.01 ms (❌ 1.28x slower)

zero

	krnl	cuda	ocl
1,000,000	38.15 us (✅ 1.00x)	25.28 us (✅ 1.51x faster)	34.12 us (✅ 1.12x faster)
10,000,000	250.90 us (✅ 1.00x)	242.95 us (✅ 1.03x faster)	251.86 us (✅ 1.00x slower)
64,000,000	1.53 ms (✅ 1.00x)	1.55 ms (✅ 1.01x slower)	1.56 ms (✅ 1.02x slower)

saxpy

	krnl	cuda	ocl
1,000,000	90.76 us (✅ 1.00x)	81.16 us (✅ 1.12x faster)	88.94 us (✅ 1.02x faster)
10,000,000	746.92 us (✅ 1.00x)	770.03 us (✅ 1.03x slower)	779.90 us (✅ 1.04x slower)
64,000,000	4.71 ms (✅ 1.00x)	4.90 ms (✅ 1.04x slower)	4.91 ms (✅ 1.04x slower)

License

Dual-licensed to be compatible with the Rust project.

Licensed under the Apache License, Version 2.0 http://www.apache.org/licenses/LICENSE-2.0 or the MIT license http://opensource.org/licenses/MIT, at your option. This file may not be copied, modified, or distributed except according to those terms.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.