diff --git a/Cargo.lock b/Cargo.lock
index 170a7f3e5edd54afa809a2874a2b364d4d9b7d7f..61cde7f161b7c4177cb781addeb2f484af3b7477 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1493,6 +1493,7 @@ version = "0.1.0"
dependencies = [
"async-std",
"clap",
+ "criterion",
"hercules_rt",
"juno_build",
"nom 8.0.0",
diff --git a/juno_samples/rodinia/srad/Cargo.toml b/juno_samples/rodinia/srad/Cargo.toml
index e41a8871d3694045d42206028d56c429698e41fc..facf8c3bc7c92fe0b77dd85900c3e53307d358e5 100644
--- a/juno_samples/rodinia/srad/Cargo.toml
+++ b/juno_samples/rodinia/srad/Cargo.toml
@@ -8,6 +8,9 @@ edition = "2021"
name = "juno_srad"
path = "src/main.rs"
+[lib]
+path = "src/lib.rs"
+
[features]
cuda = ["juno_build/cuda", "hercules_rt/cuda"]
@@ -21,3 +24,10 @@ async-std = "*"
clap = { version = "*", features = ["derive"] }
with_builtin_macros = "0.1.0"
nom = "*"
+
+[dev-dependencies]
+criterion = { version = "0.5", features = ["html_reports"] }
+
+[[bench]]
+name = "srad_bench"
+harness = false
diff --git a/juno_samples/rodinia/srad/benches/srad_bench.rs b/juno_samples/rodinia/srad/benches/srad_bench.rs
new file mode 100644
index 0000000000000000000000000000000000000000..d327454002a6f9cabe4c40f74098570ea0d22d66
--- /dev/null
+++ b/juno_samples/rodinia/srad/benches/srad_bench.rs
@@ -0,0 +1,62 @@
+#![feature(concat_idents)]
+use criterion::{criterion_group, criterion_main, Criterion};
+
+use hercules_rt::{runner, HerculesImmBox, HerculesImmBoxTo, HerculesMutBox, HerculesMutBoxTo};
+
+juno_build::juno!("srad");
+
+use juno_srad::*;
+
+fn srad_bench(c: &mut Criterion) {
+ let mut group = c.benchmark_group("srad bench");
+
+ let mut r = runner!(srad);
+ let niter = 100;
+ let lambda = 0.5;
+ let nrows = 502;
+ let ncols = 458;
+ let image = "data/image.pgm".to_string();
+ let Image {
+ image: image_ori,
+ max,
+ rows: image_ori_rows,
+ cols: image_ori_cols,
+ } = read_graphics(image);
+ let image = resize(&image_ori, image_ori_rows, image_ori_cols, nrows, ncols);
+ let mut image_h = HerculesMutBox::from(image.clone());
+ let mut iN = (0..nrows).map(|i| i as i32 - 1).collect::<Vec<_>>();
+ let mut iS = (0..nrows).map(|i| i as i32 + 1).collect::<Vec<_>>();
+ let mut jW = (0..ncols).map(|j| j as i32 - 1).collect::<Vec<_>>();
+ let mut jE = (0..ncols).map(|j| j as i32 + 1).collect::<Vec<_>>();
+ // Fix boundary conditions
+ iN[0] = 0;
+ iS[nrows - 1] = (nrows - 1) as i32;
+ jW[0] = 0;
+ jE[ncols - 1] = (ncols - 1) as i32;
+ let iN_h = HerculesImmBox::from(iN.as_slice());
+ let iS_h = HerculesImmBox::from(iS.as_slice());
+ let jW_h = HerculesImmBox::from(jW.as_slice());
+ let jE_h = HerculesImmBox::from(jE.as_slice());
+ group.bench_function("srad bench", |b| {
+ b.iter(|| {
+ async_std::task::block_on(async {
+ r.run(
+ nrows as u64,
+ ncols as u64,
+ niter as u64,
+ image_h.to(),
+ iN_h.to(),
+ iS_h.to(),
+ jW_h.to(),
+ jE_h.to(),
+ max,
+ lambda,
+ )
+ .await
+ });
+ })
+ });
+}
+
+criterion_group!(benches, srad_bench);
+criterion_main!(benches);
diff --git a/juno_samples/rodinia/srad/src/lib.rs b/juno_samples/rodinia/srad/src/lib.rs
new file mode 100644
index 0000000000000000000000000000000000000000..d63660070ff0f61d47057ea00b14b3fb31db6e09
--- /dev/null
+++ b/juno_samples/rodinia/srad/src/lib.rs
@@ -0,0 +1,123 @@
+#![feature(concat_idents)]
+mod graphics;
+mod rust_srad;
+
+pub use graphics::*;
+
+use clap::Parser;
+
+use hercules_rt::{runner, HerculesImmBox, HerculesImmBoxTo, HerculesMutBox, HerculesMutBoxTo};
+
+juno_build::juno!("srad");
+
+#[derive(Parser)]
+#[clap(author, version, about, long_about = None)]
+pub struct SRADInputs {
+ pub niter: usize,
+ pub lambda: f32,
+ pub nrows: usize,
+ pub ncols: usize,
+ pub image: String,
+ #[clap(short, long)]
+ pub output: Option<String>,
+ #[clap(short, long)]
+ pub verify: bool,
+ #[clap(long = "output-verify", value_name = "PATH")]
+ pub output_verify: Option<String>,
+}
+
+pub fn srad_harness(args: SRADInputs) {
+ async_std::task::block_on(async {
+ let SRADInputs {
+ niter,
+ lambda,
+ nrows,
+ ncols,
+ image,
+ output,
+ verify,
+ output_verify,
+ } = args;
+
+ let Image {
+ image: image_ori,
+ max,
+ rows: image_ori_rows,
+ cols: image_ori_cols,
+ } = read_graphics(image);
+ let image = resize(&image_ori, image_ori_rows, image_ori_cols, nrows, ncols);
+ let mut image_h = HerculesMutBox::from(image.clone());
+
+ let mut iN = (0..nrows).map(|i| i as i32 - 1).collect::<Vec<_>>();
+ let mut iS = (0..nrows).map(|i| i as i32 + 1).collect::<Vec<_>>();
+ let mut jW = (0..ncols).map(|j| j as i32 - 1).collect::<Vec<_>>();
+ let mut jE = (0..ncols).map(|j| j as i32 + 1).collect::<Vec<_>>();
+
+ // Fix boundary conditions
+ iN[0] = 0;
+ iS[nrows - 1] = (nrows - 1) as i32;
+ jW[0] = 0;
+ jE[ncols - 1] = (ncols - 1) as i32;
+
+ let iN_h = HerculesImmBox::from(iN.as_slice());
+ let iS_h = HerculesImmBox::from(iS.as_slice());
+ let jW_h = HerculesImmBox::from(jW.as_slice());
+ let jE_h = HerculesImmBox::from(jE.as_slice());
+
+ let mut runner = runner!(srad);
+ let result: Vec<f32> = HerculesMutBox::from(
+ runner
+ .run(
+ nrows as u64,
+ ncols as u64,
+ niter as u64,
+ image_h.to(),
+ iN_h.to(),
+ iS_h.to(),
+ jW_h.to(),
+ jE_h.to(),
+ max,
+ lambda,
+ )
+ .await,
+ )
+ .as_slice()
+ .to_vec();
+
+ if let Some(output) = output {
+ write_graphics(output, &result, nrows, ncols, max);
+ }
+
+ if verify {
+ let mut rust_result = image;
+ rust_srad::srad(
+ nrows,
+ ncols,
+ niter,
+ &mut rust_result,
+ &iN,
+ &iS,
+ &jW,
+ &jE,
+ max,
+ lambda,
+ );
+
+ if let Some(output) = output_verify {
+ write_graphics(output, &rust_result, nrows, ncols, max);
+ }
+
+ let max_diff = result
+ .iter()
+ .zip(rust_result.iter())
+ .map(|(a, b)| (*a as i32 - *b as i32).abs())
+ .max()
+ .unwrap_or(0);
+ assert!(
+ max_diff <= 1,
+ "Verification failed: maximum pixel difference of {} exceeds threshold of 1",
+ max_diff
+ );
+ }
+ })
+}
diff --git a/juno_samples/rodinia/srad/src/main.rs b/juno_samples/rodinia/srad/src/main.rs
index 1b99b41aada5341fb3157cfb0d97be3a99e73796..87d1e7e8504584478f51ac2b9dc20dbc04716c81 100644
--- a/juno_samples/rodinia/srad/src/main.rs
+++ b/juno_samples/rodinia/srad/src/main.rs
@@ -1,126 +1,6 @@
-#![feature(concat_idents)]
-mod graphics;
-mod rust_srad;
-
-use graphics::*;
-
use clap::Parser;
-use hercules_rt::{runner, HerculesImmBox, HerculesImmBoxTo, HerculesMutBox, HerculesMutBoxTo};
-
-juno_build::juno!("srad");
-
-#[derive(Parser)]
-#[clap(author, version, about, long_about = None)]
-struct SRADInputs {
- niter: usize,
- lambda: f32,
- nrows: usize,
- ncols: usize,
- image: String,
- #[clap(short, long)]
- output: Option<String>,
- #[clap(short, long)]
- verify: bool,
- #[clap(long = "output-verify", value_name = "PATH")]
- output_verify: Option<String>,
-}
-
-fn srad_harness(args: SRADInputs) {
- async_std::task::block_on(async {
- let SRADInputs {
- niter,
- lambda,
- nrows,
- ncols,
- image,
- output,
- verify,
- output_verify,
- } = args;
-
- let Image {
- image: image_ori,
- max,
- rows: image_ori_rows,
- cols: image_ori_cols,
- } = read_graphics(image);
- let image = resize(&image_ori, image_ori_rows, image_ori_cols, nrows, ncols);
- let mut image_h = HerculesMutBox::from(image.clone());
-
- let mut iN = (0..nrows).map(|i| i as i32 - 1).collect::<Vec<_>>();
- let mut iS = (0..nrows).map(|i| i as i32 + 1).collect::<Vec<_>>();
- let mut jW = (0..ncols).map(|j| j as i32 - 1).collect::<Vec<_>>();
- let mut jE = (0..ncols).map(|j| j as i32 + 1).collect::<Vec<_>>();
-
- // Fix boundary conditions
- iN[0] = 0;
- iS[nrows - 1] = (nrows - 1) as i32;
- jW[0] = 0;
- jE[ncols - 1] = (ncols - 1) as i32;
-
- let iN_h = HerculesImmBox::from(iN.as_slice());
- let iS_h = HerculesImmBox::from(iS.as_slice());
- let jW_h = HerculesImmBox::from(jW.as_slice());
- let jE_h = HerculesImmBox::from(jE.as_slice());
-
- let mut runner = runner!(srad);
- let result: Vec<f32> = HerculesMutBox::from(
- runner
- .run(
- nrows as u64,
- ncols as u64,
- niter as u64,
- image_h.to(),
- iN_h.to(),
- iS_h.to(),
- jW_h.to(),
- jE_h.to(),
- max,
- lambda,
- )
- .await,
- )
- .as_slice()
- .to_vec();
-
- if let Some(output) = output {
- write_graphics(output, &result, nrows, ncols, max);
- }
-
- if verify {
- let mut rust_result = image;
- rust_srad::srad(
- nrows,
- ncols,
- niter,
- &mut rust_result,
- &iN,
- &iS,
- &jW,
- &jE,
- max,
- lambda,
- );
-
- if let Some(output) = output_verify {
- write_graphics(output, &rust_result, nrows, ncols, max);
- }
-
- let max_diff = result
- .iter()
- .zip(rust_result.iter())
- .map(|(a, b)| (*a as i32 - *b as i32).abs())
- .max()
- .unwrap_or(0);
- assert!(
- max_diff <= 1,
- "Verification failed: maximum pixel difference of {} exceeds threshold of 1",
- max_diff
- );
- }
- })
-}
+use juno_srad::*;
fn main() {
let args = SRADInputs::parse();