mobench-sdk

Rust SDK for mobench mobile benchmarking. It provides the benchmark timing harness, #[benchmark] registry integration, generated runner support, Android and iOS builders, UniFFI compatibility, native JSON C ABI exports, and semantic profiling helpers used by the mobench CLI.

Current release: 0.1.42.

Features

• #[benchmark] re-export from mobench-macros.
• Inventory-backed benchmark discovery and runtime execution.
• Lightweight timing harness with warmup and measured iterations.
• Setup, teardown, and per-iteration benchmark inputs.
• Android and iOS build automation used by cargo mobench build/run/ci run.
• Generated mobile runner templates for:
  • uniffi (default compatibility backend)
  • native-c-abi (direct mobench JSON C ABI backend)
• profile_phase(...) semantic phase annotations for native profile summaries.
• Stable serializable timing/report types for mobile and CI integrations.

Install

Full SDK, including builders and code generation:

[dependencies]
mobench-sdk = "0.1.42"
inventory = "0.3"

Runtime-only benchmark crates that do not call SDK builders directly can use the narrower registry feature:

[dependencies]
mobench-sdk = { version = "0.1.42", default-features = false, features = ["registry"] }
inventory = "0.3"

For generated mobile libraries, configure crate types:

[lib]
crate-type = ["cdylib", "staticlib", "lib"]

Basic Benchmark

use mobench_sdk::benchmark;

#[benchmark]
pub fn fibonacci_30() {
    let result = fibonacci(30);
    std::hint::black_box(result);
}

fn fibonacci(n: u32) -> u64 {
    match n {
        0 => 0,
        1 => 1,
        _ => fibonacci(n - 1) + fibonacci(n - 2),
    }
}

Run the benchmark through the CLI:

cargo install mobench
cargo mobench list
cargo mobench run --target android --function fibonacci_30 --local-only

Programmatic Runtime

use mobench_sdk::{run_benchmark, BenchSpec};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let spec = BenchSpec::new("fibonacci_30", 100, 10)?;
    let report = run_benchmark(spec)?;

    println!("mean: {} ns", report.mean_ns());
    println!("median: {} ns", report.median_ns());
    Ok(())
}

Setup and Teardown

use mobench_sdk::benchmark;

fn setup_data() -> Vec<u8> {
    vec![42; 1024 * 1024]
}

#[benchmark(setup = setup_data)]
pub fn checksum(data: &Vec<u8>) {
    let sum: u64 = data.iter().map(|b| *b as u64).sum();
    std::hint::black_box(sum);
}

Use per_iteration for mutable or consumed input:

use mobench_sdk::benchmark;

fn unsorted_vec() -> Vec<i32> {
    (0..1000).rev().collect()
}

#[benchmark(setup = unsorted_vec, per_iteration)]
pub fn sort_vec(mut data: Vec<i32>) {
    data.sort();
    std::hint::black_box(data);
}

Generated Runner Backends

Generated mobile runners read backend configuration from mobench.toml:

[project]
crate = "my-bench-crate"
library_name = "my_bench_crate"
ffi_backend = "uniffi" # default; also supports "native-c-abi"

Use uniffi for the historical generated Kotlin/Swift binding path.

Use native-c-abi when you want generated Android and iOS runners to call mobench's direct JSON C ABI and avoid UniFFI binding-generation overhead in the measured path. Native C ABI benchmark crates should export the ABI from the crate root:

mobench_sdk::export_native_c_abi!();

This exports:

• mobench_run_benchmark_json
• mobench_free_buf
• mobench_last_error_message
• MobenchBuf

Semantic Profiling Phases

Native profile runs can include semantic phase timings:

use mobench_sdk::{benchmark, profile_phase};

#[benchmark]
pub fn prove_and_verify() {
    let proof = profile_phase("prove", || prove());
    profile_phase("verify", || verify(&proof));
}

cargo mobench profile run merges these phases into the profile manifest and summary while keeping native stack flamegraphs separate from semantic timing.

Build Through The CLI

The SDK builders are primarily driven by the CLI:

cargo mobench check --target android
cargo mobench check --target ios
cargo mobench build --target android --progress
cargo mobench build --target ios --progress
cargo mobench ci run --target android --function fibonacci_30 --local-only

Build/run/list/verify/package commands resolve the benchmark crate and project root from explicit flags, mobench.toml, Cargo workspace metadata, git root, or the legacy bench-mobile/ layout.

Important Re-Exports

• benchmark
• debug_benchmarks
• BenchmarkBuilder
• run_benchmark
• discover_benchmarks
• find_benchmark
• list_benchmark_names
• BenchSpec, BenchSample, RunnerReport, BenchSummary
• SemanticPhase, HarnessTimelineSpan, TimingError
• profile_phase, run_closure
• Target, FfiBackend, BuildConfig, BuildProfile, BuildResult
• MobenchBuf

License

MIT licensed, World Foundation 2026.