device_query/
device_query.rs

1//! Enumerate CUDA devices via the Runtime API, print compute capability
2//! and total global memory for each.
3//!
4//! Mirrors `nvidia-smi -L` (minus the rich properties), demonstrating
5//! the Runtime-API view: a device is just an ordinal, the primary
6//! context is implicit.
7//!
8//! Run with:
9//!
10//! ```text
11//! cargo run --example device_query -p baracuda-runtime
12//! ```
13
14use baracuda_runtime::{
15    device_synchronize, driver_version, query::device_properties, runtime_version, Device,
16};
17
18fn main() -> Result<(), Box<dyn std::error::Error>> {
19    println!("CUDA driver version: {}", driver_version()?);
20    println!("CUDA runtime version: {}", runtime_version()?);
21
22    let devices = Device::all()?;
23    println!("found {} CUDA device(s)", devices.len());
24
25    for dev in &devices {
26        // Switch current-thread device before querying properties:
27        // `device_properties` reads via `cudaGetDeviceProperties`, which
28        // is keyed off the ordinal we pass — but other Runtime calls
29        // (memory queries, kernel launches) operate on the current
30        // device, so it's good practice to bind first.
31        dev.set_current()?;
32        let props = device_properties(dev)?;
33        let (cc_major, cc_minor) = dev.compute_capability()?;
34        let sm_count = dev.multiprocessor_count()?;
35        let warp = dev.warp_size()?;
36
37        let gib = (props.total_global_memory_bytes as f64) / (1024.0 * 1024.0 * 1024.0);
38        println!(
39            "  [{}] {:30}  cc {}.{}  {} SMs  warp {}  {:.2} GiB",
40            dev.ordinal(),
41            props.name,
42            cc_major,
43            cc_minor,
44            sm_count,
45            warp,
46            gib,
47        );
48    }
49
50    // Make sure all pending work on every device (none here) is done.
51    device_synchronize()?;
52    println!("OK");
53    Ok(())
54}
device_query/device_query.rs

device_query/
device_query.rs
