RUDA.
Documentation / EnglishView source ↗

Examples and Tutorials

Documentation · Quickstart · 中文 | 日本語 | Deutsch | Русский

1. NVIDIA runtime example

Entry point: ptx-runtime.

The basic case performs FP32 addition twice at each length: 1, 63, 64, 65, and 257. It checks every result and 16 tail sentinel values. It demonstrates device selection, upload, argument binding, tail bounds, readback, and repeated execution.

See the quickstart for building and selecting a backend.

2. Targeted cases

Place these arguments after the command's --. For example, run the tensor case with:

cargo run --locked -p ruda-driver-cuda --features direct-ptx --example ptx-runtime -- --tensor
Argument Behavior Example source
--tensor Tensor metadata and layout checks tensor.rs
--shared Shared memory checks shared.rs
--half FP16/BF16 checks half_precision.rs
--bitwise Bitwise operation checks bitwise.rs
--shared-over-limit Shared memory limit diagnostics shared.rs
--expect-cold Asserts compilation occurred without a disk cache hit Main entry point
--expect-warm Asserts a disk cache hit without recompilation Main entry point

--shared-over-limit takes a separate early-return branch. Do not combine it with cold/warm cache checks.

3. Cold and warm caches

Use a separate new cache directory for each compilation path. Reuse that directory for the cold and warm runs of the same path.

After selecting a compilation path as described in Getting started, run these commands in the same PowerShell session:

$env:RUDA_PTX_TEST_CACHE = 'target/ptx-example-cache-' + [guid]::NewGuid().ToString('N')
cargo run --locked -p ruda-driver-cuda --features direct-ptx --example ptx-runtime -- --expect-cold
cargo run --locked -p ruda-driver-cuda --features direct-ptx --example ptx-runtime -- --expect-warm

Keep the compiler, input arguments, and cache path unchanged between runs.

4. Compute library examples

Prepare the build environment before running examples.

Task Command Output checks
FP32 CSR matrix-vector multiplication cargo run --locked -p ruSPARSE --features cuda --example csrmv Reads back and checks [7.0, 2.0, 18.5]
CUDA Ring AllReduce cargo run --locked -p ruCCL --features cuda --example all_reduce Four logical ranks on GPU 0, 257 elements, Sum/Mean, and input preservation

5. Training and model inference