benchmark_backends/
benchmark_backends.rs

1//! CPU (rayon) vs GPU (wgpu) timing for batch RNS arithmetic across batch sizes.
2//!
3//! Run with: `cargo run --release --example benchmark_backends`
4
5use std::time::Instant;
6
7use adele_ring::backend::ArithmeticBackend;
8use adele_ring::{executor, Channels, RnsBatch, RnsInt};
9
10fn time_backend<F: Fn() -> RnsBatch>(f: F, iters: u32) -> f64 {
11    // Warm up.
12    for _ in 0..10 {
13        let _ = f();
14    }
15    let start = Instant::now();
16    for _ in 0..iters {
17        let _ = f();
18    }
19    start.elapsed().as_secs_f64() * 1e6 / iters as f64 // microseconds per op
20}
21
22fn main() {
23    let exec = executor();
24    let ch = Channels::standard(32);
25    let has_gpu = exec.gpu().is_some();
26
27    println!("== adele-ring :: backend benchmark (32 channels) ==");
28    println!(
29        "GPU available: {}\n",
30        if has_gpu {
31            exec.gpu().map(|g| g.adapter_name().to_string()).unwrap_or_default()
32        } else {
33            "no (CPU-only)".to_string()
34        }
35    );
36
37    println!("{:>10} | {:>14} | {:>12} | winner", "batch_size", "cpu_rayon_us", "gpu_us");
38    println!("{}", "-".repeat(56));
39
40    for &size in &[1usize, 16, 128, 1024, 16_384, 65_536] {
41        let a = RnsBatch::from_rns_ints(&vec![RnsInt::from_i64(123, ch.clone()); size]);
42        let b = RnsBatch::from_rns_ints(&vec![RnsInt::from_i64(456, ch.clone()); size]);
43
44        let iters = if size <= 128 { 2000 } else { 100 };
45        let cpu_us = time_backend(|| exec.cpu().batch_rns_add(&a, &b), iters);
46
47        let (gpu_str, winner) = if let Some(gpu) = exec.gpu() {
48            let gpu_us = time_backend(|| gpu.batch_rns_add(&a, &b), iters);
49            let w = if cpu_us <= gpu_us { "CPU" } else { "GPU" };
50            (format!("{gpu_us:>12.2}"), w)
51        } else {
52            ("         n/a".to_string(), "CPU")
53        };
54
55        println!("{size:>10} | {cpu_us:>14.2} | {gpu_str} | {winner}");
56    }
57
58    println!(
59        "\nNote: CPU wins for small batches (GPU upload/dispatch overhead ~100us);\n\
60         GPU pulls ahead once the batch is large enough to amortize that fixed cost."
61    );
62}
benchmark_backends/benchmark_backends.rs

benchmark_backends/
benchmark_backends.rs
