Struct Executor 

pub struct Executor {
    pub gpu_threshold: usize,
    /* private fields */
}

Runtime dispatcher between the CPU and (optional) GPU backends.

Fields

gpu_threshold: usize

Batches smaller than this use the CPU even when a GPU is present, because the GPU’s upload/dispatch/download round-trip (~100µs) dominates for small inputs. Public so callers can tune it for their hardware.

Implementations

impl Executor

pub fn init() -> Self

Probe for a GPU and build the executor. This is the only constructor.

pub fn has_gpu(&self) -> bool

Whether a GPU backend is available.

pub fn cpu(&self) -> &CpuBackend

Borrow the CPU backend directly (used by benchmarks).

Examples found in repository

examples/benchmark_backends.rs (line 45)

fn main() {
    let exec = executor();
    let ch = Channels::standard(32);
    let has_gpu = exec.gpu().is_some();

    println!("== adele-ring :: backend benchmark (32 channels) ==");
    println!(
        "GPU available: {}\n",
        if has_gpu {
            exec.gpu().map(|g| g.adapter_name().to_string()).unwrap_or_default()
        } else {
            "no (CPU-only)".to_string()
        }
    );

    println!("{:>10} | {:>14} | {:>12} | winner", "batch_size", "cpu_rayon_us", "gpu_us");
    println!("{}", "-".repeat(56));

    for &size in &[1usize, 16, 128, 1024, 16_384, 65_536] {
        let a = RnsBatch::from_rns_ints(&vec![RnsInt::from_i64(123, ch.clone()); size]);
        let b = RnsBatch::from_rns_ints(&vec![RnsInt::from_i64(456, ch.clone()); size]);

        let iters = if size <= 128 { 2000 } else { 100 };
        let cpu_us = time_backend(|| exec.cpu().batch_rns_add(&a, &b), iters);

        let (gpu_str, winner) = if let Some(gpu) = exec.gpu() {
            let gpu_us = time_backend(|| gpu.batch_rns_add(&a, &b), iters);
            let w = if cpu_us <= gpu_us { "CPU" } else { "GPU" };
            (format!("{gpu_us:>12.2}"), w)
        } else {
            ("         n/a".to_string(), "CPU")
        };

        println!("{size:>10} | {cpu_us:>14.2} | {gpu_str} | {winner}");
    }

    println!(
        "\nNote: CPU wins for small batches (GPU upload/dispatch overhead ~100us);\n\
         GPU pulls ahead once the batch is large enough to amortize that fixed cost."
    );
}

pub fn gpu(&self) -> Option<&GpuBackend>

Borrow the GPU backend if present (used by benchmarks).

Examples found in repository

examples/benchmark_backends.rs (line 25)

fn main() {
    let exec = executor();
    let ch = Channels::standard(32);
    let has_gpu = exec.gpu().is_some();

    println!("== adele-ring :: backend benchmark (32 channels) ==");
    println!(
        "GPU available: {}\n",
        if has_gpu {
            exec.gpu().map(|g| g.adapter_name().to_string()).unwrap_or_default()
        } else {
            "no (CPU-only)".to_string()
        }
    );

    println!("{:>10} | {:>14} | {:>12} | winner", "batch_size", "cpu_rayon_us", "gpu_us");
    println!("{}", "-".repeat(56));

    for &size in &[1usize, 16, 128, 1024, 16_384, 65_536] {
        let a = RnsBatch::from_rns_ints(&vec![RnsInt::from_i64(123, ch.clone()); size]);
        let b = RnsBatch::from_rns_ints(&vec![RnsInt::from_i64(456, ch.clone()); size]);

        let iters = if size <= 128 { 2000 } else { 100 };
        let cpu_us = time_backend(|| exec.cpu().batch_rns_add(&a, &b), iters);

        let (gpu_str, winner) = if let Some(gpu) = exec.gpu() {
            let gpu_us = time_backend(|| gpu.batch_rns_add(&a, &b), iters);
            let w = if cpu_us <= gpu_us { "CPU" } else { "GPU" };
            (format!("{gpu_us:>12.2}"), w)
        } else {
            ("         n/a".to_string(), "CPU")
        };

        println!("{size:>10} | {cpu_us:>14.2} | {gpu_str} | {winner}");
    }

    println!(
        "\nNote: CPU wins for small batches (GPU upload/dispatch overhead ~100us);\n\
         GPU pulls ahead once the batch is large enough to amortize that fixed cost."
    );
}

pub fn add(&self, a: &RnsBatch, b: &RnsBatch) -> RnsBatch

Elementwise batch addition.

pub fn mul(&self, a: &RnsBatch, b: &RnsBatch) -> RnsBatch

Elementwise batch multiplication.

pub fn crt(&self, batch: &RnsBatch) -> Vec<BigUint>

CRT reconstruction — always CPU-side (Garner’s algorithm is sequential).

Trait Implementations

impl Default for Executor

fn default() -> Self

Returns the “default value” for a type.