From 3f8acd3bb4d0b7388a74c7dc24e3d1e4ad9771b1 Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Thu, 27 Feb 2025 22:44:17 -0600
Subject: [PATCH 1/9] misc. progress for bfs gpu
---
hercules_cg/src/rt.rs | 2 +-
hercules_opt/src/gcm.rs | 70 ++++++++++++++++------------
juno_samples/rodinia/bfs/src/gpu.sch | 52 ++++++++++++++-------
3 files changed, 76 insertions(+), 48 deletions(-)
diff --git a/hercules_cg/src/rt.rs b/hercules_cg/src/rt.rs
index 6981a3da..d94f0e19 100644
--- a/hercules_cg/src/rt.rs
+++ b/hercules_cg/src/rt.rs
@@ -938,7 +938,7 @@ impl<'a> RTContext<'a> {
let dst_device = self.node_colors.0[&collect];
write!(
block,
- "::hercules_rt::__copy_{}_to_{}({}.byte_add({} as usize).0, {}.0, {});",
+ "::hercules_rt::__copy_{}_to_{}({}.byte_add({} as usize).0, {}.0, {} as usize);",
src_device.name(),
dst_device.name(),
self.get_value(collect, bb, false),
diff --git a/hercules_opt/src/gcm.rs b/hercules_opt/src/gcm.rs
index c612acac..3f326051 100644
--- a/hercules_opt/src/gcm.rs
+++ b/hercules_opt/src/gcm.rs
@@ -1304,37 +1304,49 @@ enum UTerm {
fn unify(
mut equations: VecDeque<(UTerm, UTerm)>,
-) -> Result<BTreeMap<NodeID, Device>, BTreeMap<NodeID, Device>> {
+) -> Result<BTreeMap<NodeID, Device>, (NodeID, NodeID)> {
let mut theta = BTreeMap::new();
+ // First, assign devices to nodes when a rule directly says to.
+ for _ in 0..equations.len() {
+ let (l, r) = equations.pop_front().unwrap();
+ match (l, r) {
+ (UTerm::Node(n), UTerm::Device(d)) | (UTerm::Device(d), UTerm::Node(n)) => {
+ theta.insert(n, d);
+ }
+ _ => equations.push_back((l, r)),
+ }
+ }
+
+ // Second, iterate the rest of the rules until...
+ // 1. The rules are exhausted. All the nodes have device assignments.
+ // 2. No progress is being made. Some nodes may not have device assignments.
+ // 3. An inconsistency has been found. The inconsistency is returned.
let mut no_progress_iters = 0;
while no_progress_iters <= equations.len()
&& let Some((l, r)) = equations.pop_front()
{
- match (l, r) {
- (UTerm::Node(_), UTerm::Node(_)) => {
- if l != r {
- equations.push_back((l, r));
- }
- no_progress_iters += 1;
- }
- (UTerm::Node(n), UTerm::Device(d)) | (UTerm::Device(d), UTerm::Node(n)) => {
- theta.insert(n, d);
- for (l, r) in equations.iter_mut() {
- if *l == UTerm::Node(n) {
- *l = UTerm::Device(d);
- }
- if *r == UTerm::Node(n) {
- *r = UTerm::Device(d);
- }
+ let (UTerm::Node(l), UTerm::Node(r)) = (l, r) else {
+ panic!();
+ };
+
+ match (theta.get(&l), theta.get(&r)) {
+ (Some(ld), Some(rd)) => {
+ if ld != rd {
+ return Err((l, r));
+ } else {
+ no_progress_iters = 0;
}
- no_progress_iters = 0;
}
- (UTerm::Device(d1), UTerm::Device(d2)) if d1 == d2 => {
+ (Some(d), None) | (None, Some(d)) => {
+ let d = *d;
+ theta.insert(l, d);
+ theta.insert(r, d);
no_progress_iters = 0;
}
- _ => {
- return Err(theta);
+ (None, None) => {
+ equations.push_back((UTerm::Node(l), UTerm::Node(r)));
+ no_progress_iters += 1;
}
}
}
@@ -1377,8 +1389,8 @@ fn color_nodes(
} if !editor.get_type(typing[id.idx()]).is_primitive() => {
// Every input to a phi needs to be on the same device. The
// phi itself is also on this device.
- for (l, r) in zip(data.into_iter(), data.into_iter().skip(1).chain(once(&id))) {
- equations.push((UTerm::Node(*l), UTerm::Node(*r)));
+ for data in data {
+ equations.push((UTerm::Node(*data), UTerm::Node(id)));
}
}
Node::Reduce {
@@ -1394,7 +1406,7 @@ fn color_nodes(
} if !editor.get_type(typing[id.idx()]).is_primitive() => {
// Every input to the reduce, and the reduce itself, are on
// the same device.
- equations.push((UTerm::Node(first), UTerm::Node(second)));
+ equations.push((UTerm::Node(first), UTerm::Node(id)));
equations.push((UTerm::Node(second), UTerm::Node(id)));
}
Node::Constant { id: _ }
@@ -1533,12 +1545,12 @@ fn color_nodes(
}
Some(func_colors)
}
- Err(progress) => {
+ Err(inconsistency) => {
// If unification failed, then there's some node using a node in
- // `progress` that's expecting a different type than what it got.
- // Pick one and add potentially inter-device copies on each def-use
- // edge. We'll clean these up later.
- let (id, _) = progress.into_iter().next().unwrap();
+ // that's expecting a different type than what it got. Pick one and
+ // add potentially inter-device copies on each def-use edge. We'll
+ // clean these up later.
+ let id = inconsistency.0;
let users: Vec<_> = editor.get_users(id).collect();
let success = editor.edit(|mut edit| {
let cons = edit.add_zero_constant(typing[id.idx()]);
diff --git a/juno_samples/rodinia/bfs/src/gpu.sch b/juno_samples/rodinia/bfs/src/gpu.sch
index 0a3f4d77..49d44b98 100644
--- a/juno_samples/rodinia/bfs/src/gpu.sch
+++ b/juno_samples/rodinia/bfs/src/gpu.sch
@@ -1,23 +1,39 @@
-gvn(*);
-phi-elim(*);
-dce(*);
+macro simpl!(X) {
+ ccp(X);
+ simplify-cfg(X);
+ lift-dc-math(X);
+ gvn(X);
+ phi-elim(X);
+ dce(X);
+ infer-schedules(X);
+}
-let outline = auto-outline(bfs);
-gpu(outline.bfs);
+phi-elim(bfs);
+no-memset(bfs@cost);
+let cost_init = outline(bfs@cost_init);
+let loop1 = outline(bfs@loop1);
+let loop2 = outline(bfs@loop2);
+gpu(cost_init, loop1, loop2);
-ip-sroa(*);
-sroa(*);
-dce(*);
-gvn(*);
-phi-elim(*);
-dce(*);
+simpl!(*);
+predication(*);
+const-inline(*);
+simpl!(*);
+fixpoint {
+ forkify(*);
+ fork-guard-elim(*);
+}
+simpl!(*);
+predication(*);
+simpl!(*);
-//forkify(*);
-infer-schedules(*);
+parallel-reduce(loop1);
+forkify(*);
+fork-guard-elim(*);
+simpl!(*);
+predication(*);
+reduce-slf(*);
+simpl!(*);
gcm(*);
-fixpoint {
- float-collections(*);
- dce(*);
- gcm(*);
-}
+xdot[true](*);
--
GitLab
From 2654f19895ca35203f8431d46613c5566e194e40 Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Thu, 27 Feb 2025 22:50:22 -0600
Subject: [PATCH 2/9] oh my
---
juno_samples/rodinia/bfs/src/gpu.sch | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/juno_samples/rodinia/bfs/src/gpu.sch b/juno_samples/rodinia/bfs/src/gpu.sch
index 49d44b98..5be2d903 100644
--- a/juno_samples/rodinia/bfs/src/gpu.sch
+++ b/juno_samples/rodinia/bfs/src/gpu.sch
@@ -13,7 +13,7 @@ no-memset(bfs@cost);
let cost_init = outline(bfs@cost_init);
let loop1 = outline(bfs@loop1);
let loop2 = outline(bfs@loop2);
-gpu(cost_init, loop1, loop2);
+gpu(loop1, loop2);
simpl!(*);
predication(*);
@@ -35,5 +35,5 @@ predication(*);
reduce-slf(*);
simpl!(*);
+unforkify(cost_init);
gcm(*);
-xdot[true](*);
--
GitLab
From ab1eafd464dd73b2e9f81633e42ede950766f600 Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Fri, 28 Feb 2025 09:52:12 -0600
Subject: [PATCH 3/9] fix unify
---
hercules_opt/src/gcm.rs | 21 +++++++++++----------
1 file changed, 11 insertions(+), 10 deletions(-)
diff --git a/hercules_opt/src/gcm.rs b/hercules_opt/src/gcm.rs
index 3f326051..1dcedd97 100644
--- a/hercules_opt/src/gcm.rs
+++ b/hercules_opt/src/gcm.rs
@@ -1302,9 +1302,7 @@ enum UTerm {
Device(Device),
}
-fn unify(
- mut equations: VecDeque<(UTerm, UTerm)>,
-) -> Result<BTreeMap<NodeID, Device>, (NodeID, NodeID)> {
+fn unify(mut equations: VecDeque<(UTerm, UTerm)>) -> Result<BTreeMap<NodeID, Device>, NodeID> {
let mut theta = BTreeMap::new();
// First, assign devices to nodes when a rule directly says to.
@@ -1312,7 +1310,11 @@ fn unify(
let (l, r) = equations.pop_front().unwrap();
match (l, r) {
(UTerm::Node(n), UTerm::Device(d)) | (UTerm::Device(d), UTerm::Node(n)) => {
- theta.insert(n, d);
+ if let Some(old_d) = theta.insert(n, d)
+ && old_d != d
+ {
+ return Err(n);
+ }
}
_ => equations.push_back((l, r)),
}
@@ -1333,7 +1335,7 @@ fn unify(
match (theta.get(&l), theta.get(&r)) {
(Some(ld), Some(rd)) => {
if ld != rd {
- return Err((l, r));
+ return Err(l);
} else {
no_progress_iters = 0;
}
@@ -1545,12 +1547,11 @@ fn color_nodes(
}
Some(func_colors)
}
- Err(inconsistency) => {
+ Err(id) => {
// If unification failed, then there's some node using a node in
- // that's expecting a different type than what it got. Pick one and
- // add potentially inter-device copies on each def-use edge. We'll
- // clean these up later.
- let id = inconsistency.0;
+ // that's expecting a different type than what it got. Add
+ // potentially inter-device copies on each def-use edge. We'll clean
+ // these up later.
let users: Vec<_> = editor.get_users(id).collect();
let success = editor.edit(|mut edit| {
let cons = edit.add_zero_constant(typing[id.idx()]);
--
GitLab
From 96c8dc524fe30a271e57a2f9d02f3759411767ad Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Fri, 28 Feb 2025 11:14:12 -0600
Subject: [PATCH 4/9] sequentializing channel loop in gamut in cava is
beneficial
---
juno_samples/cava/src/gpu.sch | 1 +
1 file changed, 1 insertion(+)
diff --git a/juno_samples/cava/src/gpu.sch b/juno_samples/cava/src/gpu.sch
index c8db124e..bacfd3ab 100644
--- a/juno_samples/cava/src/gpu.sch
+++ b/juno_samples/cava/src/gpu.sch
@@ -120,6 +120,7 @@ simpl!(fuse4);
//fork-tile[2, 0, false, true](fuse4@channel_loop);
//fork-split(fuse4@channel_loop);
//clean-monoid-reduces(fuse4);
+unforkify(fuse4@channel_loop);
no-memset(fuse5@res1);
no-memset(fuse5@res2);
--
GitLab
From 28ea01a1afcbb630b570e360693defb3e76c70d9 Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Fri, 28 Feb 2025 13:24:48 -0600
Subject: [PATCH 5/9] Infer more indices as parallel
---
hercules_opt/src/utils.rs | 18 ++++++++++++++++++
juno_samples/rodinia/backprop/src/backprop.jn | 4 ++--
juno_samples/rodinia/backprop/src/cpu.sch | 10 +++-------
3 files changed, 23 insertions(+), 9 deletions(-)
diff --git a/hercules_opt/src/utils.rs b/hercules_opt/src/utils.rs
index e962b81d..b910a128 100644
--- a/hercules_opt/src/utils.rs
+++ b/hercules_opt/src/utils.rs
@@ -532,6 +532,24 @@ where
let fork_thread_id_pairs = node_indices(indices).filter_map(|id| {
if let Node::ThreadID { control, dimension } = nodes[id.idx()] {
Some((control, dimension))
+ } else if let Node::Binary {
+ op: BinaryOperator::Add,
+ left: tid,
+ right: cons,
+ } = nodes[id.idx()]
+ && let Node::ThreadID { control, dimension } = nodes[tid.idx()]
+ && (nodes[cons.idx()].is_constant() || nodes[cons.idx()].is_dynamic_constant())
+ {
+ Some((control, dimension))
+ } else if let Node::Binary {
+ op: BinaryOperator::Add,
+ left: cons,
+ right: tid,
+ } = nodes[id.idx()]
+ && let Node::ThreadID { control, dimension } = nodes[tid.idx()]
+ && (nodes[cons.idx()].is_constant() || nodes[cons.idx()].is_dynamic_constant())
+ {
+ Some((control, dimension))
} else {
None
}
diff --git a/juno_samples/rodinia/backprop/src/backprop.jn b/juno_samples/rodinia/backprop/src/backprop.jn
index 356bb3d9..94c4334c 100644
--- a/juno_samples/rodinia/backprop/src/backprop.jn
+++ b/juno_samples/rodinia/backprop/src/backprop.jn
@@ -7,9 +7,9 @@ fn layer_forward<n, m: usize>(vals: f32[n + 1], weights: f32[n + 1, m + 1]) -> f
@res let result : f32[m + 1];
result[0] = 1.0;
- for j in 1..=m {
+ @outer_loop for j in 1..=m {
let sum = 0.0;
- for k in 0..=n {
+ @inner_loop for k in 0..=n {
sum += weights[k, j] * vals[k];
}
result[j] = squash(sum);
diff --git a/juno_samples/rodinia/backprop/src/cpu.sch b/juno_samples/rodinia/backprop/src/cpu.sch
index d1fe8953..d59fd5f5 100644
--- a/juno_samples/rodinia/backprop/src/cpu.sch
+++ b/juno_samples/rodinia/backprop/src/cpu.sch
@@ -15,20 +15,16 @@ delete-uncalled(*);
no-memset(layer_forward@res);
lift-dc-math(*);
loop-bound-canon(*);
-dce(*);
+simpl!(*);
lift-dc-math(*);
+slf(*);
fixpoint {
forkify(*);
fork-guard-elim(*);
fork-coalesce(*);
}
+simpl!(*);
fork-split(*);
-gvn(*);
-phi-elim(*);
-dce(*);
unforkify(*);
-gvn(*);
-phi-elim(*);
-dce(*);
gcm(*);
--
GitLab
From 3565d96d31fa97f32892548167b359843fafb6dc Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Fri, 28 Feb 2025 14:28:50 -0600
Subject: [PATCH 6/9] Tweak edge multi-core
---
juno_samples/edge_detection/src/cpu.sch | 13 ++++++++++---
1 file changed, 10 insertions(+), 3 deletions(-)
diff --git a/juno_samples/edge_detection/src/cpu.sch b/juno_samples/edge_detection/src/cpu.sch
index 4bd3254b..ec9e423d 100644
--- a/juno_samples/edge_detection/src/cpu.sch
+++ b/juno_samples/edge_detection/src/cpu.sch
@@ -86,7 +86,7 @@ fixpoint {
simpl!(max_gradient);
fork-dim-merge(max_gradient);
simpl!(max_gradient);
-fork-tile[8, 0, false, false](max_gradient);
+fork-tile[16, 0, false, false](max_gradient);
let split = fork-split(max_gradient);
clean-monoid-reduces(max_gradient);
let out = outline(split._4_max_gradient.fj1);
@@ -104,11 +104,18 @@ fixpoint {
}
predication(reject_zero_crossings);
simpl!(reject_zero_crossings);
+fork-tile[4, 1, false, false](reject_zero_crossings);
+fork-tile[4, 0, false, false](reject_zero_crossings);
+fork-interchange[1, 2](reject_zero_crossings);
+let split = fork-split(reject_zero_crossings);
+let reject_zero_crossings_body = outline(split._5_reject_zero_crossings.fj2);
+fork-coalesce(reject_zero_crossings, reject_zero_crossings_body);
+simpl!(reject_zero_crossings, reject_zero_crossings_body);
async-call(edge_detection@le, edge_detection@zc);
-fork-split(gaussian_smoothing_body, laplacian_estimate_body, zero_crossings_body, gradient, reject_zero_crossings);
-unforkify(gaussian_smoothing_body, laplacian_estimate_body, zero_crossings_body, gradient, reject_zero_crossings);
+fork-split(gaussian_smoothing_body, laplacian_estimate_body, zero_crossings_body, gradient, reject_zero_crossings_body);
+unforkify(gaussian_smoothing_body, laplacian_estimate_body, zero_crossings_body, gradient, reject_zero_crossings_body);
simpl!(*);
--
GitLab
From da2cf7557ad1b74556df070c76d16094b9f20256 Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Fri, 28 Feb 2025 14:45:40 -0600
Subject: [PATCH 7/9] gcm moment to allow reduce-slf in CPU code without worry
---
hercules_opt/src/gcm.rs | 7 +++++++
juno_samples/rodinia/srad/src/cpu.sch | 2 ++
2 files changed, 9 insertions(+)
diff --git a/hercules_opt/src/gcm.rs b/hercules_opt/src/gcm.rs
index 1dcedd97..2579e73e 100644
--- a/hercules_opt/src/gcm.rs
+++ b/hercules_opt/src/gcm.rs
@@ -883,6 +883,13 @@ fn spill_clones(
init == *b
&& editor.func().schedules[a.idx()].contains(&Schedule::ParallelReduce)
})
+ .unwrap_or(false)
+ && !editor.func().nodes[a.idx()]
+ .try_phi()
+ .map(|(_, data)| {
+ data.contains(b)
+ && editor.func().schedules[a.idx()].contains(&Schedule::ParallelReduce)
+ })
.unwrap_or(false))
});
diff --git a/juno_samples/rodinia/srad/src/cpu.sch b/juno_samples/rodinia/srad/src/cpu.sch
index 2b45e8c9..a4cd4956 100644
--- a/juno_samples/rodinia/srad/src/cpu.sch
+++ b/juno_samples/rodinia/srad/src/cpu.sch
@@ -29,6 +29,8 @@ fixpoint {
}
simpl!(*);
fork-interchange[0, 1](loop1);
+reduce-slf(*);
+simpl!(*);
fork-split(*);
unforkify(*);
--
GitLab
From 1b78bc2e9d2acce3e9b1eec8aed25489eb9aced5 Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Fri, 28 Feb 2025 15:59:10 -0600
Subject: [PATCH 8/9] Refactor infer parallelreduce
---
hercules_opt/src/schedule.rs | 46 ++++++-----------------
juno_samples/rodinia/backprop/src/cpu.sch | 1 +
2 files changed, 12 insertions(+), 35 deletions(-)
diff --git a/hercules_opt/src/schedule.rs b/hercules_opt/src/schedule.rs
index 9bc7823e..10eca72e 100644
--- a/hercules_opt/src/schedule.rs
+++ b/hercules_opt/src/schedule.rs
@@ -1,6 +1,6 @@
use std::collections::{BTreeSet, HashMap, HashSet};
+use std::iter::once;
-use hercules_ir::def_use::*;
use hercules_ir::ir::*;
use crate::*;
@@ -42,6 +42,10 @@ pub fn infer_parallel_reduce(
fork_join_map: &HashMap<NodeID, NodeID>,
reduce_cycles: &HashMap<NodeID, HashSet<NodeID>>,
) {
+ let join_fork_map: HashMap<_, _> = fork_join_map
+ .into_iter()
+ .map(|(fork, join)| (*join, *fork))
+ .collect();
for id in editor.node_ids() {
let func = editor.func();
if !func.nodes[id.idx()].is_reduce() {
@@ -98,40 +102,11 @@ pub fn infer_parallel_reduce(
&& *collect == last_reduce
&& !reduce_cycles[&last_reduce].contains(data)
{
- // If there is a Write-Reduce tight cycle, get the position indices.
- let positions = indices
- .iter()
- .filter_map(|index| {
- if let Index::Position(indices) = index {
- Some(indices)
- } else {
- None
- }
- })
- .flat_map(|pos| pos.iter());
-
- // Get the Forks corresponding to uses of bare ThreadIDs.
- let fork_thread_id_pairs = positions.filter_map(|id| {
- if let Node::ThreadID { control, dimension } = func.nodes[id.idx()] {
- Some((control, dimension))
- } else {
- None
- }
- });
- let mut forks = HashMap::<NodeID, Vec<usize>>::new();
- for (fork, dim) in fork_thread_id_pairs {
- forks.entry(fork).or_default().push(dim);
- }
-
- // Check if one of the Forks correspond to the Join associated with
- // the Reduce being considered, and has all of its dimensions
- // represented in the indexing.
- let is_parallel = forks.into_iter().any(|(id, mut rep_dims)| {
- rep_dims.sort();
- rep_dims.dedup();
- fork_join_map[&id] == first_control.unwrap()
- && func.nodes[id.idx()].try_fork().unwrap().1.len() == rep_dims.len()
- });
+ let is_parallel = indices_parallel_over_forks(
+ editor,
+ indices,
+ once(join_fork_map[&first_control.unwrap()]),
+ );
if is_parallel {
editor.edit(|edit| edit.add_schedule(id, Schedule::ParallelReduce));
@@ -145,6 +120,7 @@ pub fn infer_parallel_reduce(
* operands must be the Reduce node, and all other operands must not be in the
* Reduce node's cycle.
*/
+#[rustfmt::skip]
pub fn infer_monoid_reduce(
editor: &mut FunctionEditor,
reduce_cycles: &HashMap<NodeID, HashSet<NodeID>>,
diff --git a/juno_samples/rodinia/backprop/src/cpu.sch b/juno_samples/rodinia/backprop/src/cpu.sch
index d59fd5f5..de34d660 100644
--- a/juno_samples/rodinia/backprop/src/cpu.sch
+++ b/juno_samples/rodinia/backprop/src/cpu.sch
@@ -23,6 +23,7 @@ fixpoint {
fork-guard-elim(*);
fork-coalesce(*);
}
+reduce-slf(*);
simpl!(*);
fork-split(*);
--
GitLab
From eaf09eb1b58e1d159839aa0c4b474e7d62d5dd09 Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Fri, 28 Feb 2025 16:31:57 -0600
Subject: [PATCH 9/9] hmm
---
hercules_opt/src/gcm.rs | 4 ++--
juno_samples/rodinia/bfs/src/gpu.sch | 2 +-
2 files changed, 3 insertions(+), 3 deletions(-)
diff --git a/hercules_opt/src/gcm.rs b/hercules_opt/src/gcm.rs
index 2579e73e..d950941a 100644
--- a/hercules_opt/src/gcm.rs
+++ b/hercules_opt/src/gcm.rs
@@ -879,8 +879,8 @@ fn spill_clones(
|| editor.func().nodes[a.idx()].is_reduce())
&& !editor.func().nodes[a.idx()]
.try_reduce()
- .map(|(_, init, _)| {
- init == *b
+ .map(|(_, init, reduct)| {
+ (init == *b || reduct == *b)
&& editor.func().schedules[a.idx()].contains(&Schedule::ParallelReduce)
})
.unwrap_or(false)
diff --git a/juno_samples/rodinia/bfs/src/gpu.sch b/juno_samples/rodinia/bfs/src/gpu.sch
index 5be2d903..6c4d027b 100644
--- a/juno_samples/rodinia/bfs/src/gpu.sch
+++ b/juno_samples/rodinia/bfs/src/gpu.sch
@@ -27,6 +27,7 @@ simpl!(*);
predication(*);
simpl!(*);
+unforkify(cost_init);
parallel-reduce(loop1);
forkify(*);
fork-guard-elim(*);
@@ -35,5 +36,4 @@ predication(*);
reduce-slf(*);
simpl!(*);
-unforkify(cost_init);
gcm(*);
--
GitLab