This repo works to launch CUDA kernels without relying on the CUDA runtime.

## Repository layout

- `src/` – Rust crate that drives the reverse-engineered submission path.
- `cuda_kernels/` – cache of compiled kernels the CLI can load quickly.
- `open-gpu-kernel-modules/` – NVIDIA’s open kernel headers (required for bindgen
  and for compiling new cubins/PTX).
- `orig_cpp/` – the historical C++/CUDA prototype plus helper scripts. Nothing in
  this folder is built by default, but it’s useful reference material and still
  provides optional utilities such as `libsniff`.

## First-time setup

1. Clone NVIDIA’s header repository beside the crate (if you haven’t already):

   ```bash
   git clone https://github.com/NVIDIA/open-gpu-kernel-modules.git
   ```

2. (Optional but recommended) Rebuild the legacy “simple” cubin that the Rust
   loader can fall back to when `TRIVIAL` is unset:

   ```bash
   (cd orig_cpp && ./cudaless.sh)
   # creates orig_cpp/out/simple.o
   ```

   If you prefer to rely exclusively on the built-in trivial program, export
   `TRIVIAL=1` before running any of the binaries and you can skip this step.

3. (Optional) Build the C helper library used by the original sniffing code:

   ```bash
   (cd orig_cpp && ./make_sniff.sh)
   # produces orig_cpp/out/libsniff.so
   ```

## Building and running the Rust tools

- Minimal end-to-end launch (uses built-in “TRIVIAL” kernel automatically):

  ```bash
  cargo run --bin minimal_demo
  ```

- Rich TUI CLI (`gpu_cli`) with optional C shim:

  ```bash
  # Rebuild libsniff if you skipped it earlier: (cd orig_cpp && ./make_sniff.sh)
  cargo run --features with-sniff-c --bin gpu_cli
  # Set LD_LIBRARY_PATH if you enabled the shim:
  LD_LIBRARY_PATH=orig_cpp/out cargo run --features with-sniff-c --bin gpu_cli
  ```

- Compiling new CUDA sources through the CLI:

  ```text
  kernel compile path/to/file.cu arch=sm_86
  kernel demo <name>
  ```

  The command injects include paths for `open-gpu-kernel-modules`, writes the
  resulting PTX/CUBIN into `cuda_kernels/`, and updates the registry so you can
  `kernel use <name>` later.

## Legacy prototype

All of the original C++ sources are under `orig_cpp/`. The noteworthy entry
points are:

- `orig_cpp/gpu_driver.cc` – hand-crafted command submission.
- `orig_cpp/tc_context.cc` – minimal `cuInit` / `cuCtxCreate` replacement.

They continue to build exactly as they did historically once
`open-gpu-kernel-modules/` is present at the repo root.