mirror of
https://github.com/hanzoai/ml.git
synced 2026-11-19 23:50:51 +00:00
No crate inherits from the workspace any longer: each manifest holds its own version, package fields, lints and dependency specs, siblings as a path and the exact version they publish at. The root keeps members, exclude, resolver and profiles. Every publishable crate takes a patch bump so the registry carries manifests that stand alone. Claude-Session: https://claude.ai/code/session_01JupC8d38f9Lkfz62hvjG2q
Bindgen Cuda
Similar crate than bindgen in philosophy. It will help create automatic bindgen to cuda kernels source files and make them easier to use directly from Rust.
PTX inclusion
Let's say you have a file
src/cuda.cu
__global__ void cuda_hello(){
printf("Hello World from GPU!\n");
}
You can add bindgen_cuda as a build dependency:
cargo add --build bindgen_cuda
And then create this build.rs
fn main() {
let builder = bindgen_cuda::Builder::default();
let bindings = builder.build_ptx().unwrap();
bindings.write("src/lib.rs");
}
This will create a src file containing the following code:
pub const CUDA: &str = include_str!(concat!(env!("OUT_DIR"), "/cuda.ptx"));
You can then use the PTX directly in your rust code with a library like cudarc.
Raw cuda calls
Alternatively you can build a static library that you can link against in build.rs in order to call cuda directly with the c code.
src/cuda.cu
__global__ void cuda_hello(){
printf("Hello World from GPU!\n");
}
int run() {
cuda_hello<<<1,1>>>();
return 0;
}
Then write the build.rs:
fn main() {
let builder = bindgen_cuda::Builder::default();
builder.build_lib("libcuda.a");
println!("cargo:rustc-link-lib=cuda");
}
Which you can then interface through FFI in src/lib.rs:
extern "C" {
fn cuda_hello();
}
fn main(){
unsafe{ cuda_hello();}
}