diff --git a/Cargo.lock b/Cargo.lock
index 61cde7f161b7c4177cb781addeb2f484af3b7477..ccd42c296c953d24a081c13a4725f6a4449aa15a 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1301,6 +1301,7 @@ dependencies = [
"hercules_rt",
"juno_build",
"nom 8.0.0",
+ "rayon",
"with_builtin_macros",
]
@@ -1497,6 +1498,7 @@ dependencies = [
"hercules_rt",
"juno_build",
"nom 8.0.0",
+ "rayon",
"with_builtin_macros",
]
diff --git a/hercules_cg/src/rt.rs b/hercules_cg/src/rt.rs
index d94f0e19191a028dfaded785c460164513b712a4..5c5d35629a42bb2465b1558f5aed37fc20a1237a 100644
--- a/hercules_cg/src/rt.rs
+++ b/hercules_cg/src/rt.rs
@@ -160,7 +160,7 @@ impl<'a> RTContext<'a> {
// Dump the function signature.
write!(
w,
- "#[allow(unused_assignments,unused_variables,unused_mut,unused_parens,unused_unsafe,non_snake_case)]async unsafe fn {}_{}(",
+ "#[allow(unused_assignments,unused_variables,unused_mut,unused_parens,unused_unsafe,non_snake_case)]unsafe fn {}_{}(",
self.module_name,
func.name
)?;
@@ -198,7 +198,7 @@ impl<'a> RTContext<'a> {
}
write!(w, ") -> ")?;
self.write_rust_return_type(w, &func.return_types)?;
- write!(w, " {{")?;
+ write!(w, " {{use rayon::prelude::*;")?;
// Dump signatures for called device functions.
// For single-return functions we directly expose the device function
@@ -407,9 +407,14 @@ impl<'a> RTContext<'a> {
// Emit loops for the thread IDs.
for (idx, factor) in factors.into_iter().enumerate() {
- write!(prologue, "for tid_{}_{} in 0..", id.idx(), idx)?;
+ write!(prologue, "(0..")?;
self.codegen_dynamic_constant(*factor, prologue)?;
- write!(prologue, " {{")?;
+ write!(
+ prologue,
+ ").into_par_iter().for_each(|tid_{}_{}| {{",
+ id.idx(),
+ idx
+ )?;
}
// Emit clones of arcs used inside the fork-join.
@@ -422,13 +427,6 @@ impl<'a> RTContext<'a> {
}
}
- // Spawn an async closure and push its future to a Vec.
- write!(
- prologue,
- "fork_{}.push(::async_std::task::spawn(async move {{",
- id.idx()
- )?;
-
// Open a new environment.
self.codegen_open_environment(id, prologue)?;
@@ -454,22 +452,15 @@ impl<'a> RTContext<'a> {
// Close the branch inside the async closure.
let epilogue = &mut blocks.get_mut(&id).unwrap().epilogue;
- write!(
- epilogue,
- "::std::sync::atomic::fence(::std::sync::atomic::Ordering::Release);return;}}"
- )?;
+ write!(epilogue, "return;}}")?;
// Close the fork's environment.
self.codegen_close_environment(epilogue)?;
- // Close the async closure and push statement from the
- // fork.
- write!(epilogue, "}}));")?;
-
// Close the loops emitted by the fork node.
let fork = self.join_fork_map[&id];
for _ in 0..func.nodes[fork.idx()].try_fork().unwrap().1.len() {
- write!(epilogue, "}}")?;
+ write!(epilogue, "}});")?;
}
// Close the branch for the fork outside the async closure.
@@ -479,14 +470,6 @@ impl<'a> RTContext<'a> {
let succ = self.control_subgraph.succs(id).next().unwrap();
write!(epilogue, "{} => {{", id.idx())?;
- // Await the empty futures for the fork-joins, waiting for them
- // to complete.
- write!(
- epilogue,
- "for fut in fork_{}.drain(..) {{ fut.await; }}; ::std::sync::atomic::fence(::std::sync::atomic::Ordering::Acquire);",
- fork.idx(),
- )?;
-
// Emit the assignments to the reduce variables in the
// surrounding context. It's very unfortunate that we have to do
// it while lowering the join node (rather than the reduce nodes
@@ -669,9 +652,9 @@ impl<'a> RTContext<'a> {
}
};
let postfix = match (device, is_async) {
- (Device::AsyncRust, false) => ".await",
+ (Device::AsyncRust, false) => "",
(_, false) => "",
- (Device::AsyncRust, true) => ".await}))",
+ (Device::AsyncRust, true) => "}))",
(_, true) => "}))",
};
write!(
@@ -1264,11 +1247,6 @@ impl<'a> RTContext<'a> {
}
}
- // Declare Vecs for storing futures of fork-joins.
- for fork in self.fork_tree[&root].iter() {
- write!(w, "let mut fork_{} = vec![];", fork.idx())?;
- }
-
// The core executor is a Rust loop. We literally run a "control token"
// as described in the original sea of nodes paper through the basic
// blocks to drive execution.
@@ -1358,7 +1336,7 @@ impl<'a> RTContext<'a> {
})
.collect();
- write!(w, "async fn run<'runner:")?;
+ write!(w, "fn run<'runner:")?;
for (ret_idx, origins) in returned_origins.iter().enumerate() {
if origins.iter().any(|origin| !origin.is_parameter()) {
write!(w, " 'r{} +", ret_idx)?;
@@ -1482,7 +1460,7 @@ impl<'a> RTContext<'a> {
for idx in 0..func.param_types.len() {
write!(w, "p{}, ", idx)?;
}
- write!(w, ").await;")?;
+ write!(w, ");")?;
// Return the result, appropriately wrapping pointers
if num_returns == 1 {
if self.module.types[func.return_types[0].idx()].is_primitive() {
diff --git a/juno_samples/rodinia/cfd/Cargo.toml b/juno_samples/rodinia/cfd/Cargo.toml
index 6720b5275381594a63f31571ccf6266ebe4e46f4..e39d16ab522a208bab816ca7c70a8ef64f856499 100644
--- a/juno_samples/rodinia/cfd/Cargo.toml
+++ b/juno_samples/rodinia/cfd/Cargo.toml
@@ -21,6 +21,7 @@ juno_build = { path = "../../../juno_build" }
juno_build = { path = "../../../juno_build" }
hercules_rt = { path = "../../../hercules_rt" }
async-std = "*"
+rayon = "*"
clap = { version = "*", features = ["derive"] }
with_builtin_macros = "0.1.0"
nom = "*"
diff --git a/juno_samples/rodinia/cfd/benches/cfd_bench.rs b/juno_samples/rodinia/cfd/benches/cfd_bench.rs
index 5fc73db9d369be5ef9695f37ad1f39122facf991..8be321d17e18e1c146082787c2ae873460a536ba 100644
--- a/juno_samples/rodinia/cfd/benches/cfd_bench.rs
+++ b/juno_samples/rodinia/cfd/benches/cfd_bench.rs
@@ -46,40 +46,37 @@ fn cfd_bench(c: &mut Criterion) {
group.bench_function("cfd bench euler", |b| {
b.iter(|| {
- async_std::task::block_on(async {
- r.run(
- nelr as u64,
- iterations as u64,
- v_density.to(),
- v_momentum_x.to(),
- v_momentum_y.to(),
- v_momentum_z.to(),
- v_energy.to(),
- areas.to(),
- elements_surrounding_elements.to(),
- normals_x.to(),
- normals_y.to(),
- normals_z.to(),
- ff_variable.density,
- ff_variable.momentum.x,
- ff_variable.momentum.y,
- ff_variable.momentum.z,
- ff_variable.energy,
- ff_fc_density_energy.x,
- ff_fc_density_energy.y,
- ff_fc_density_energy.z,
- ff_fc_momentum_x.x,
- ff_fc_momentum_x.y,
- ff_fc_momentum_x.z,
- ff_fc_momentum_y.x,
- ff_fc_momentum_y.y,
- ff_fc_momentum_y.z,
- ff_fc_momentum_z.x,
- ff_fc_momentum_z.y,
- ff_fc_momentum_z.z,
- )
- .await
- });
+ r.run(
+ nelr as u64,
+ iterations as u64,
+ v_density.to(),
+ v_momentum_x.to(),
+ v_momentum_y.to(),
+ v_momentum_z.to(),
+ v_energy.to(),
+ areas.to(),
+ elements_surrounding_elements.to(),
+ normals_x.to(),
+ normals_y.to(),
+ normals_z.to(),
+ ff_variable.density,
+ ff_variable.momentum.x,
+ ff_variable.momentum.y,
+ ff_variable.momentum.z,
+ ff_variable.energy,
+ ff_fc_density_energy.x,
+ ff_fc_density_energy.y,
+ ff_fc_density_energy.z,
+ ff_fc_momentum_x.x,
+ ff_fc_momentum_x.y,
+ ff_fc_momentum_x.z,
+ ff_fc_momentum_y.x,
+ ff_fc_momentum_y.y,
+ ff_fc_momentum_y.z,
+ ff_fc_momentum_z.x,
+ ff_fc_momentum_z.y,
+ ff_fc_momentum_z.z,
+ );
})
});
@@ -118,40 +115,37 @@ fn cfd_bench(c: &mut Criterion) {
group.bench_function("cfd bench pre-euler", |b| {
b.iter(|| {
- async_std::task::block_on(async {
- r.run(
- nelr as u64,
- iterations as u64,
- v_density.to(),
- v_momentum_x.to(),
- v_momentum_y.to(),
- v_momentum_z.to(),
- v_energy.to(),
- areas.to(),
- elements_surrounding_elements.to(),
- normals_x.to(),
- normals_y.to(),
- normals_z.to(),
- ff_variable.density,
- ff_variable.momentum.x,
- ff_variable.momentum.y,
- ff_variable.momentum.z,
- ff_variable.energy,
- ff_fc_density_energy.x,
- ff_fc_density_energy.y,
- ff_fc_density_energy.z,
- ff_fc_momentum_x.x,
- ff_fc_momentum_x.y,
- ff_fc_momentum_x.z,
- ff_fc_momentum_y.x,
- ff_fc_momentum_y.y,
- ff_fc_momentum_y.z,
- ff_fc_momentum_z.x,
- ff_fc_momentum_z.y,
- ff_fc_momentum_z.z,
- )
- .await
- });
+ r.run(
+ nelr as u64,
+ iterations as u64,
+ v_density.to(),
+ v_momentum_x.to(),
+ v_momentum_y.to(),
+ v_momentum_z.to(),
+ v_energy.to(),
+ areas.to(),
+ elements_surrounding_elements.to(),
+ normals_x.to(),
+ normals_y.to(),
+ normals_z.to(),
+ ff_variable.density,
+ ff_variable.momentum.x,
+ ff_variable.momentum.y,
+ ff_variable.momentum.z,
+ ff_variable.energy,
+ ff_fc_density_energy.x,
+ ff_fc_density_energy.y,
+ ff_fc_density_energy.z,
+ ff_fc_momentum_x.x,
+ ff_fc_momentum_x.y,
+ ff_fc_momentum_x.z,
+ ff_fc_momentum_y.x,
+ ff_fc_momentum_y.y,
+ ff_fc_momentum_y.z,
+ ff_fc_momentum_z.x,
+ ff_fc_momentum_z.y,
+ ff_fc_momentum_z.z,
+ );
})
});
}
diff --git a/juno_samples/rodinia/cfd/src/lib.rs b/juno_samples/rodinia/cfd/src/lib.rs
index d61df4c5b7bb0c7915d8605debd72e375a5482b0..bb8bf079c67f26c75ea0d83cbfc7037577754ee8 100644
--- a/juno_samples/rodinia/cfd/src/lib.rs
+++ b/juno_samples/rodinia/cfd/src/lib.rs
@@ -48,41 +48,37 @@ fn run_euler(
let normals_z = HerculesImmBox::from(normals.z.as_slice());
let mut runner = runner!(euler);
- let (density, momentum_x, momentum_y, momentum_z, energy) = async_std::task::block_on(async {
- runner
- .run(
- nelr as u64,
- iterations as u64,
- v_density.to(),
- v_momentum_x.to(),
- v_momentum_y.to(),
- v_momentum_z.to(),
- v_energy.to(),
- areas.to(),
- elements_surrounding_elements.to(),
- normals_x.to(),
- normals_y.to(),
- normals_z.to(),
- ff_variable.density,
- ff_variable.momentum.x,
- ff_variable.momentum.y,
- ff_variable.momentum.z,
- ff_variable.energy,
- ff_fc_density_energy.x,
- ff_fc_density_energy.y,
- ff_fc_density_energy.z,
- ff_fc_momentum_x.x,
- ff_fc_momentum_x.y,
- ff_fc_momentum_x.z,
- ff_fc_momentum_y.x,
- ff_fc_momentum_y.y,
- ff_fc_momentum_y.z,
- ff_fc_momentum_z.x,
- ff_fc_momentum_z.y,
- ff_fc_momentum_z.z,
- )
- .await
- });
+ let (density, momentum_x, momentum_y, momentum_z, energy) = runner.run(
+ nelr as u64,
+ iterations as u64,
+ v_density.to(),
+ v_momentum_x.to(),
+ v_momentum_y.to(),
+ v_momentum_z.to(),
+ v_energy.to(),
+ areas.to(),
+ elements_surrounding_elements.to(),
+ normals_x.to(),
+ normals_y.to(),
+ normals_z.to(),
+ ff_variable.density,
+ ff_variable.momentum.x,
+ ff_variable.momentum.y,
+ ff_variable.momentum.z,
+ ff_variable.energy,
+ ff_fc_density_energy.x,
+ ff_fc_density_energy.y,
+ ff_fc_density_energy.z,
+ ff_fc_momentum_x.x,
+ ff_fc_momentum_x.y,
+ ff_fc_momentum_x.z,
+ ff_fc_momentum_y.x,
+ ff_fc_momentum_y.y,
+ ff_fc_momentum_y.z,
+ ff_fc_momentum_z.x,
+ ff_fc_momentum_z.y,
+ ff_fc_momentum_z.z,
+ );
Variables {
density: AlignedSlice::from_slice(HerculesMutBox::from(density).as_slice()),
@@ -122,41 +118,37 @@ fn run_pre_euler(
let normals_z = HerculesImmBox::from(normals.z.as_slice());
let mut runner = runner!(pre_euler);
- let (density, momentum_x, momentum_y, momentum_z, energy) = async_std::task::block_on(async {
- runner
- .run(
- nelr as u64,
- iterations as u64,
- v_density.to(),
- v_momentum_x.to(),
- v_momentum_y.to(),
- v_momentum_z.to(),
- v_energy.to(),
- areas.to(),
- elements_surrounding_elements.to(),
- normals_x.to(),
- normals_y.to(),
- normals_z.to(),
- ff_variable.density,
- ff_variable.momentum.x,
- ff_variable.momentum.y,
- ff_variable.momentum.z,
- ff_variable.energy,
- ff_fc_density_energy.x,
- ff_fc_density_energy.y,
- ff_fc_density_energy.z,
- ff_fc_momentum_x.x,
- ff_fc_momentum_x.y,
- ff_fc_momentum_x.z,
- ff_fc_momentum_y.x,
- ff_fc_momentum_y.y,
- ff_fc_momentum_y.z,
- ff_fc_momentum_z.x,
- ff_fc_momentum_z.y,
- ff_fc_momentum_z.z,
- )
- .await
- });
+ let (density, momentum_x, momentum_y, momentum_z, energy) = runner.run(
+ nelr as u64,
+ iterations as u64,
+ v_density.to(),
+ v_momentum_x.to(),
+ v_momentum_y.to(),
+ v_momentum_z.to(),
+ v_energy.to(),
+ areas.to(),
+ elements_surrounding_elements.to(),
+ normals_x.to(),
+ normals_y.to(),
+ normals_z.to(),
+ ff_variable.density,
+ ff_variable.momentum.x,
+ ff_variable.momentum.y,
+ ff_variable.momentum.z,
+ ff_variable.energy,
+ ff_fc_density_energy.x,
+ ff_fc_density_energy.y,
+ ff_fc_density_energy.z,
+ ff_fc_momentum_x.x,
+ ff_fc_momentum_x.y,
+ ff_fc_momentum_x.z,
+ ff_fc_momentum_y.x,
+ ff_fc_momentum_y.y,
+ ff_fc_momentum_y.z,
+ ff_fc_momentum_z.x,
+ ff_fc_momentum_z.y,
+ ff_fc_momentum_z.z,
+ );
Variables {
density: AlignedSlice::from_slice(HerculesMutBox::from(density).as_slice()),
diff --git a/juno_samples/rodinia/srad/Cargo.toml b/juno_samples/rodinia/srad/Cargo.toml
index facf8c3bc7c92fe0b77dd85900c3e53307d358e5..46c2ae8e316011c9d563797ea1eedb88bd8d54d9 100644
--- a/juno_samples/rodinia/srad/Cargo.toml
+++ b/juno_samples/rodinia/srad/Cargo.toml
@@ -21,6 +21,7 @@ juno_build = { path = "../../../juno_build" }
juno_build = { path = "../../../juno_build" }
hercules_rt = { path = "../../../hercules_rt" }
async-std = "*"
+rayon = "*"
clap = { version = "*", features = ["derive"] }
with_builtin_macros = "0.1.0"
nom = "*"
diff --git a/juno_samples/rodinia/srad/benches/srad_bench.rs b/juno_samples/rodinia/srad/benches/srad_bench.rs
index 6af13aae5d9093bd59cf6299dfa64ac3e73209a6..2ea230ab25c6cf254d8a7914bcff629f9ae00ea7 100644
--- a/juno_samples/rodinia/srad/benches/srad_bench.rs
+++ b/juno_samples/rodinia/srad/benches/srad_bench.rs
@@ -26,17 +26,8 @@ fn srad_bench(c: &mut Criterion) {
let mut image_h = HerculesMutBox::from(image.clone());
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(),
- max,
- lambda,
- )
- .await
- });
+ let to = image_h.to();
+ r.run(nrows as u64, ncols as u64, niter as u64, to, max, lambda);
})
});
}
diff --git a/juno_samples/rodinia/srad/src/cpu.sch b/juno_samples/rodinia/srad/src/cpu.sch
index 7b7a6c9e2203b8e280f323ed9e6589ab149537c0..186ef8d53f4a411cabc00f9c9294878ebf03a961 100644
--- a/juno_samples/rodinia/srad/src/cpu.sch
+++ b/juno_samples/rodinia/srad/src/cpu.sch
@@ -39,7 +39,6 @@ fork-tile[32, 0, false, false](loop2);
let split = fork-split(loop2);
let loop2_body = outline(split.srad_1.fj1);
simpl!(loop2, loop2_body);
-
inline(srad@loop2);
delete-uncalled(*);
diff --git a/juno_samples/rodinia/srad/src/lib.rs b/juno_samples/rodinia/srad/src/lib.rs
index cb156d9db712155ad9e5b3f9775e3c6c29dbf22a..5d5576e48948ee5f4db858a854aa230109fc5019 100644
--- a/juno_samples/rodinia/srad/src/lib.rs
+++ b/juno_samples/rodinia/srad/src/lib.rs
@@ -27,66 +27,60 @@ pub struct SRADInputs {
}
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 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 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 runner = runner!(srad);
- let result: Vec<f32> = HerculesMutBox::from(
- runner
- .run(
- nrows as u64,
- ncols as u64,
- niter as u64,
- image_h.to(),
- max,
- lambda,
- )
- .await,
- )
- .as_slice()
- .to_vec();
+ 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(),
+ max,
+ lambda,
+ ))
+ .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, max, lambda);
+ if let Some(output) = output {
+ write_graphics(output, &result, nrows, ncols, max);
+ }
- if let Some(output) = output_verify {
- write_graphics(output, &rust_result, nrows, ncols, max);
- }
+ if verify {
+ let mut rust_result = image;
+ rust_srad::srad(nrows, ncols, niter, &mut rust_result, max, lambda);
- 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 <= 2,
- "Verification failed: maximum pixel difference of {} exceeds threshold of 1",
- max_diff
- );
+ 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 <= 2,
+ "Verification failed: maximum pixel difference of {} exceeds threshold of 1",
+ max_diff
+ );
+ }
}