From 00154685ccaee4c128231c2f52c931845e2adc71 Mon Sep 17 00:00:00 2001
From: Russel Arbore <rarbore2@illinois.edu>
Date: Wed, 5 Mar 2025 22:08:40 -0600
Subject: [PATCH 1/2] make bfs faster on cpu
---
hercules_opt/src/fork_transforms.rs | 119 +++++++++++++++++++++++++++
juno_samples/rodinia/bfs/src/cpu.sch | 6 +-
juno_samples/rodinia/bfs/src/gpu.sch | 5 +-
3 files changed, 128 insertions(+), 2 deletions(-)
diff --git a/hercules_opt/src/fork_transforms.rs b/hercules_opt/src/fork_transforms.rs
index c46e4e98..bebb8c6c 100644
--- a/hercules_opt/src/fork_transforms.rs
+++ b/hercules_opt/src/fork_transforms.rs
@@ -1515,6 +1515,10 @@ fn fork_fusion(
* Looks for would-be monoid reduces, if not for a gate on the reduction.
* Partially predicate the gated reduction to allow for a proper monoid
* reduction.
+ *
+ * Looks for monoid reduces that occur through a gated write to a single
+ * location, and lift them to a proper monoid reduction with a single gated
+ * write afterwards.
*/
pub fn clean_monoid_reduces(editor: &mut FunctionEditor, typing: &Vec<TypeID>) {
for id in editor.node_ids() {
@@ -1676,6 +1680,121 @@ pub fn clean_monoid_reduces(editor: &mut FunctionEditor, typing: &Vec<TypeID>) {
});
}
}
+
+ for id in editor.node_ids() {
+ // Identify reduce/write/phi cycle through which a sparse AND reduction
+ // is occurring.
+ let nodes = &editor.func().nodes;
+ let Some((join, init, reduct)) = nodes[id.idx()].try_reduce() else {
+ continue;
+ };
+ let join_pred = nodes[join.idx()].try_join().unwrap();
+ let join_succ = editor
+ .get_users(join)
+ .filter(|id| nodes[id.idx()].is_control())
+ .next()
+ .unwrap();
+ let Some((_, phi_data)) = nodes[reduct.idx()].try_phi() else {
+ continue;
+ };
+ if phi_data.len() != 2 {
+ continue;
+ }
+ let phi_other_use = if phi_data[0] == id {
+ phi_data[1]
+ } else if phi_data[1] == id {
+ phi_data[0]
+ } else {
+ continue;
+ };
+ let Some((collect, data, indices)) = nodes[phi_other_use.idx()].try_write() else {
+ continue;
+ };
+ if collect != id {
+ continue;
+ }
+ if indices.into_iter().any(|idx| idx.is_position()) {
+ continue;
+ }
+ if !is_false(editor, data) {
+ continue;
+ }
+ let Some(preds) = nodes[join_pred.idx()].try_region() else {
+ continue;
+ };
+ if preds.len() != 2 {
+ continue;
+ }
+ let Some((if1, _)) = nodes[preds[0].idx()].try_control_proj() else {
+ continue;
+ };
+ let Some((if2, sel)) = nodes[preds[1].idx()].try_control_proj() else {
+ continue;
+ };
+ if if1 != if2 {
+ continue;
+ }
+ let Some((_, mut cond)) = nodes[if1.idx()].try_if() else {
+ continue;
+ };
+
+ // Transform to a monoid reduction and a single gated write.
+ let negated = phi_other_use == phi_data[sel];
+ let indices = indices.to_vec().into_boxed_slice();
+ editor.edit(|mut edit| {
+ let t = edit.add_constant(Constant::Boolean(true));
+ let t = edit.add_node(Node::Constant { id: t });
+ let f = edit.add_constant(Constant::Boolean(false));
+ let f = edit.add_node(Node::Constant { id: f });
+ if negated {
+ cond = edit.add_node(Node::Unary {
+ op: UnaryOperator::Not,
+ input: cond,
+ });
+ }
+ let reduce_id = NodeID::new(edit.num_node_ids());
+ let and_id = NodeID::new(edit.num_node_ids() + 1);
+ edit.add_node(Node::Reduce {
+ control: join,
+ init: t,
+ reduct: and_id,
+ });
+ edit.add_node(Node::Binary {
+ op: BinaryOperator::And,
+ left: cond,
+ right: reduce_id,
+ });
+
+ let new_if = edit.add_node(Node::If {
+ control: join,
+ cond: reduce_id,
+ });
+ let cpj1 = edit.add_node(Node::ControlProjection {
+ control: new_if,
+ selection: 0,
+ });
+ let cpj2 = edit.add_node(Node::ControlProjection {
+ control: new_if,
+ selection: 1,
+ });
+ let region = edit.add_node(Node::Region {
+ preds: Box::new([cpj1, cpj2]),
+ });
+ let write = edit.add_node(Node::Write {
+ collect: init,
+ data: f,
+ indices,
+ });
+ let phi = edit.add_node(Node::Phi {
+ control: region,
+ data: Box::new([write, init]),
+ });
+ edit = edit.replace_all_uses_where(id, phi, |other_id| {
+ *other_id != phi_other_use && *other_id != reduct
+ })?;
+ edit.replace_all_uses_where(join, region, |id| *id == join_succ)
+ });
+ }
}
/*
diff --git a/juno_samples/rodinia/bfs/src/cpu.sch b/juno_samples/rodinia/bfs/src/cpu.sch
index f564cd36..07006edb 100644
--- a/juno_samples/rodinia/bfs/src/cpu.sch
+++ b/juno_samples/rodinia/bfs/src/cpu.sch
@@ -54,9 +54,13 @@ if !feature("seq") {
inline(bfs@cost_init, bfs@loop1, bfs@loop2);
init = init_body;
}
-fork-tile[8, 0, false, true](init, traverse, collect);
delete-uncalled(*);
const-inline(*);
+clean-monoid-reduces(collect);
+simpl!(*);
+
+fork-tile[8, 0, false, true](init, traverse, collect);
+clean-monoid-reduces(collect);
simpl!(*);
fork-split(init, traverse, collect);
diff --git a/juno_samples/rodinia/bfs/src/gpu.sch b/juno_samples/rodinia/bfs/src/gpu.sch
index 541d15d7..11a1723f 100644
--- a/juno_samples/rodinia/bfs/src/gpu.sch
+++ b/juno_samples/rodinia/bfs/src/gpu.sch
@@ -31,6 +31,8 @@ predication(*);
simpl!(*);
reduce-slf(*);
simpl!(*);
+slf(*);
+simpl!(*);
fixpoint {
forkify(collect);
@@ -39,6 +41,7 @@ fixpoint {
simpl!(collect);
fork-tile[1024, 0, false, true](init, traverse, collect);
-fork-split(init, traverse, collect);
+let out = fork-split(init, traverse, collect);
+simpl!(*);
gcm(*);
--
GitLab
From 08e67ade68f310063276b91c5dfbbc5552949b80 Mon Sep 17 00:00:00 2001
From: Russel Arbore <russel.jma@gmail.com>
Date: Wed, 5 Mar 2025 22:18:03 -0600
Subject: [PATCH 2/2] sigh
---
juno_samples/rodinia/bfs/src/gpu.sch | 2 --
1 file changed, 2 deletions(-)
diff --git a/juno_samples/rodinia/bfs/src/gpu.sch b/juno_samples/rodinia/bfs/src/gpu.sch
index 11a1723f..ea81f330 100644
--- a/juno_samples/rodinia/bfs/src/gpu.sch
+++ b/juno_samples/rodinia/bfs/src/gpu.sch
@@ -31,8 +31,6 @@ predication(*);
simpl!(*);
reduce-slf(*);
simpl!(*);
-slf(*);
-simpl!(*);
fixpoint {
forkify(collect);
--
GitLab