Fixing parallel features

Cargo.toml

@@ -15,10 +15,10 @@ exclude = [
 
 [dependencies]
 ark-bn254 = "0.4.0"
-ark-ec = "0.4.2"
-ark-ff = "0.4.1"
+ark-ec = {version = "0.4.2", features = ["parallel"]}
+ark-ff = {version = "0.4.2", features = ["parallel"]}
 ark-serialize = "0.4.2"
-ark-std = "0.4.0"
+ark-std = {version = "0.4.0", features = ["parallel"]}
 directories = "5.0.1"
 hex-literal = "0.4.1"
 rand = "0.8.5"
@@ -28,7 +28,7 @@ num-bigint = "0.4"
 rayon = "^1.5"
 num-traits = "0.2"
 byteorder = "1.4"
-ark-poly = "0.4.2"
+ark-poly = {version = "0.4.2", features = ["parallel"]}
 crossbeam-channel = "0.5"
 num_cpus = "1.13.0"
 
@@ -69,6 +69,11 @@ name = "bench_kzg_commit"
 harness = false
 path = "benches/bench_kzg_commit.rs"
 
+[[bench]]
+name = "bench_kzg_commit_large_blobs"
+harness = false
+path = "benches/bench_kzg_commit_large_blobs.rs"
+
 [[bench]]
 name = "bench_kzg_proof"
 harness = false
@@ -81,7 +86,7 @@ path = "benches/bench_kzg_verify.rs"
 
 [profile.bench]
 opt-level = 3
-debug = true
+debug = false
 strip = "none"
 debug-assertions = false
 overflow-checks = false

benches/bench_kzg_commit_large_blobs.rs

@@ -0,0 +1,56 @@
+use criterion::{criterion_group, criterion_main, Criterion};
+use rand::Rng;
+use rust_kzg_bn254::{blob::Blob, kzg::Kzg, polynomial::PolynomialFormat};
+use std::time::Duration;
+
+fn bench_kzg_commit(c: &mut Criterion) {
+    let mut rng = rand::thread_rng();
+    let mut kzg = Kzg::setup(
+        "tests/test-files/mainnet-data/g1.32mb.point",
+        "",
+        "tests/test-files/mainnet-data/g2.point.powerOf2",
+        268435456,
+        524288,
+    )
+    .unwrap();
+
+    c.bench_function("bench_kzg_commit_8mb", |b| {
+        let random_blob: Vec<u8> = (0..8000000)
+            .map(|_| rng.gen_range(32..=126) as u8)
+            .collect();
+        let input = Blob::from_bytes_and_pad(&random_blob);
+        let input_poly = input
+            .to_polynomial(PolynomialFormat::InCoefficientForm)
+            .unwrap();
+        kzg.data_setup_custom(1, input.len().try_into().unwrap())
+            .unwrap();
+        b.iter(|| kzg.commit(&input_poly).unwrap());
+    });
+
+    c.bench_function("bench_kzg_commit_16mb", |b| {
+        let random_blob: Vec<u8> = (0..16_252_000)
+            .map(|_| rng.gen_range(32..=126) as u8)
+            .collect();
+        let input = Blob::from_bytes_and_pad(&random_blob);
+        let input_poly = input
+            .to_polynomial(PolynomialFormat::InCoefficientForm)
+            .unwrap();
+        kzg.data_setup_custom(1, input.len().try_into().unwrap())
+            .unwrap();
+        b.iter(|| kzg.commit(&input_poly).unwrap());
+    });
+}
+
+fn criterion_config() -> Criterion {
+    Criterion::default()
+        .warm_up_time(Duration::from_secs(5))  // Warm-up time
+        .measurement_time(Duration::from_secs(220))  // Measurement time
+        .sample_size(10) // Number of samples to take
+}
+
+criterion_group!(
+    name = benches;
+    config = criterion_config();
+    targets = bench_kzg_commit
+);
+criterion_main!(benches);

src/kzg.rs

@@ -13,6 +13,7 @@ use ark_serialize::Read;
 use ark_std::{ops::Div, str::FromStr, One, Zero};
 use crossbeam_channel::{bounded, Sender};
 use num_traits::ToPrimitive;
+use rayon::iter::{IntoParallelRefIterator, ParallelIterator};
 use std::{fs::File, io, io::BufReader};
 
 #[derive(Debug, PartialEq, Clone)]
@@ -559,15 +560,15 @@ impl Kzg {
         }
 
         let points_projective: Vec<G1Projective> = self.g1[..length]
-            .iter()
+            .par_iter()
             .map(|&p| G1Projective::from(p))
             .collect();
 
         match GeneralEvaluationDomain::<Fr>::new(length) {
             Some(domain) => {
                 let ifft_result = domain.ifft(&points_projective);
                 let ifft_result_affine: Vec<_> =
-                    ifft_result.iter().map(|p| p.into_affine()).collect();
+                    ifft_result.par_iter().map(|p| p.into_affine()).collect();
                 Ok(ifft_result_affine)
             },
             None => Err(KzgError::FftError(