From 77bb4cf7b80d66f750383e4d30232b4fa688ee23 Mon Sep 17 00:00:00 2001
From: xavierrouth <xavierrouth@gmail.com>
Date: Wed, 5 Mar 2025 22:05:06 -0600
Subject: [PATCH] cava gamut fusion debug
---
hercules_opt/src/unforkify.rs | 6 ++
juno_samples/cava/Cargo.toml | 1 +
juno_samples/cava/src/cpu.sch | 8 ++-
juno_samples/cava/src/gpu.sch | 32 ++++++----
juno_samples/edge_detection/Cargo.toml | 1 +
.../benches/edge_detection_bench.rs | 1 +
juno_samples/edge_detection/src/gpu.sch | 64 ++++++++++++++-----
7 files changed, 83 insertions(+), 30 deletions(-)
diff --git a/hercules_opt/src/unforkify.rs b/hercules_opt/src/unforkify.rs
index 2d6cf7b3..2deb7165 100644
--- a/hercules_opt/src/unforkify.rs
+++ b/hercules_opt/src/unforkify.rs
@@ -313,6 +313,12 @@ pub fn unforkify(
edit = edit.delete_node(*reduce)?;
}
+ edit.sub_edit(fork, proj_exit_id);
+ edit.sub_edit(fork, proj_back_id);
+ edit.sub_edit(fork, neq_id);
+ edit.sub_edit(fork, add_id);
+
+
edit = edit.delete_node(fork)?;
edit = edit.delete_node(join)?;
for tid in tids {
diff --git a/juno_samples/cava/Cargo.toml b/juno_samples/cava/Cargo.toml
index bdf144da..279a81f0 100644
--- a/juno_samples/cava/Cargo.toml
+++ b/juno_samples/cava/Cargo.toml
@@ -14,6 +14,7 @@ path = "src/lib.rs"
[features]
cuda = ["juno_build/cuda", "hercules_rt/cuda"]
seq = []
+dont_fuse_gamut = []
[build-dependencies]
juno_build = { path = "../../juno_build" }
diff --git a/juno_samples/cava/src/cpu.sch b/juno_samples/cava/src/cpu.sch
index ada2f552..7d788c2f 100644
--- a/juno_samples/cava/src/cpu.sch
+++ b/juno_samples/cava/src/cpu.sch
@@ -111,11 +111,15 @@ fixpoint {
fork-fusion(fuse4@channel_loop);
}
simpl!(fuse4);
-array-slf(fuse4);
-simpl!(fuse4);
+
+if !feature("dont_fuse_gamut") {
+ array-slf(fuse4);
+ simpl!(fuse4);
+}
if !feature("seq") {
let par = fuse4@image_loop \ fuse4@channel_loop;
+ let par = par \ fuse4@cp_loop;
fork-tile[4, 1, false, false](par);
fork-tile[8, 0, false, false](par);
fork-interchange[1, 2](par);
diff --git a/juno_samples/cava/src/gpu.sch b/juno_samples/cava/src/gpu.sch
index 0ef466c0..daf52339 100644
--- a/juno_samples/cava/src/gpu.sch
+++ b/juno_samples/cava/src/gpu.sch
@@ -115,12 +115,25 @@ fixpoint {
fork-fusion(fuse4@channel_loop);
}
simpl!(fuse4);
-array-slf(fuse4);
-simpl!(fuse4);
-fork-tile[2, 0, false, true](fuse4@channel_loop);
-let out = fork-split(fuse4@channel_loop);
-fork-unroll(out.cava_3.fj1);
-unforkify(fuse4@channel_loop);
+
+if !feature("dont_fuse_gamut") {
+ array-slf(fuse4);
+ simpl!(fuse4);
+ fork-tile[2, 0, false, true](fuse4@channel_loop);
+ let out = fork-split(fuse4@channel_loop);
+ fork-unroll(out.cava_3.fj1);
+ unforkify(fuse4@channel_loop);
+}
+
+let par = fuse4@image_loop \ fuse4@channel_loop;
+let par = par \ fuse4@cp_loop;
+
+fork-tile[4, 1, false, true](par);
+fork-tile[8, 0, false, true](par);
+fork-interchange[1, 2](par);
+let split = fork-split(par);
+fork-coalesce(split.cava_3.fj0 \ split.cava_3.fj2);
+fork-coalesce(split.cava_3.fj2);
no-memset(fuse5@res1);
no-memset(fuse5@res2);
@@ -133,12 +146,7 @@ simpl!(fuse5);
array-slf(fuse5);
simpl!(fuse5);
-fork-tile[4, 1, false, true](fuse4);
-fork-tile[8, 0, false, true](fuse4);
-fork-interchange[1, 2](fuse4);
-let split = fork-split(fuse4);
-fork-coalesce(split.cava_3.fj0 \ split.cava_3.fj2);
-fork-coalesce(split.cava_3.fj2);
+
delete-uncalled(*);
simpl!(*);
diff --git a/juno_samples/edge_detection/Cargo.toml b/juno_samples/edge_detection/Cargo.toml
index 8def7500..6e53500a 100644
--- a/juno_samples/edge_detection/Cargo.toml
+++ b/juno_samples/edge_detection/Cargo.toml
@@ -8,6 +8,7 @@ edition = "2021"
opencv = ["dep:opencv"]
cuda = ["juno_build/cuda", "hercules_rt/cuda"]
seq = []
+warp_tile = []
[[bin]]
name = "juno_edge_detection"
diff --git a/juno_samples/edge_detection/benches/edge_detection_bench.rs b/juno_samples/edge_detection/benches/edge_detection_bench.rs
index 76035275..1fa08506 100644
--- a/juno_samples/edge_detection/benches/edge_detection_bench.rs
+++ b/juno_samples/edge_detection/benches/edge_detection_bench.rs
@@ -17,6 +17,7 @@ juno_build::juno!("edge_detection");
fn edge_detection_bench(c: &mut Criterion) {
let mut group = c.benchmark_group("edge detection bench");
group.sample_size(10);
+ group.measurement_time(std::time::Duration::from_secs(25));
let input = "examples/formula1_scaled.mp4";
diff --git a/juno_samples/edge_detection/src/gpu.sch b/juno_samples/edge_detection/src/gpu.sch
index 666f6cef..3a541dab 100644
--- a/juno_samples/edge_detection/src/gpu.sch
+++ b/juno_samples/edge_detection/src/gpu.sch
@@ -95,22 +95,54 @@ fixpoint {
fork-guard-elim(max_gradient);
fork-coalesce(max_gradient);
}
-simpl!(max_gradient);
-fork-dim-merge(max_gradient);
-simpl!(max_gradient);
-fork-tile[32, 0, false, true](max_gradient);
-let out = fork-split(max_gradient);
-clean-monoid-reduces(max_gradient);
-simpl!(max_gradient);
-let fission = fork-fission[out._4_max_gradient.fj0](max_gradient);
-simpl!(max_gradient);
-fork-tile[32, 0, false, true](fission._4_max_gradient.fj_bottom);
-let out = fork-split(fission._4_max_gradient.fj_bottom);
-clean-monoid-reduces(max_gradient);
-simpl!(max_gradient);
-let top = outline(fission._4_max_gradient.fj_top);
-let bottom = outline(out._4_max_gradient.fj0);
-gpu(top, bottom);
+
+if !feature("seq") {
+ if !feature("warp_tile") {
+ simpl!(max_gradient);
+ fork-dim-merge(max_gradient);
+ simpl!(max_gradient);
+ fork-tile[32, 0, false, true](max_gradient);
+ let out1 = fork-split(max_gradient);
+ clean-monoid-reduces(max_gradient);
+ simpl!(max_gradient);
+ let fission = fork-fission[out1._4_max_gradient.fj0](max_gradient);
+ simpl!(max_gradient);
+ fork-tile[32, 0, false, true](fission._4_max_gradient.fj_bottom);
+ let out2 = fork-split(fission._4_max_gradient.fj_bottom);
+ clean-monoid-reduces(max_gradient);
+ simpl!(max_gradient);
+ unforkify(out1._4_max_gradient.fj1);
+
+ unforkify(out2._4_max_gradient.fj1);
+ simpl!(max_gradient);
+ let top = outline(fission._4_max_gradient.fj_top);
+ let bottom = outline(out2._4_max_gradient.fj0);
+ gpu(top, bottom);
+ } else {
+ simpl!(max_gradient);
+ fork-dim-merge(max_gradient);
+ simpl!(max_gradient);
+ fork-tile[32, 0, false, true](max_gradient);
+ let out = fork-split(max_gradient);
+ clean-monoid-reduces(max_gradient);
+ simpl!(max_gradient);
+ let fission = fork-fission[out._4_max_gradient.fj0](max_gradient);
+ simpl!(max_gradient);
+ fork-tile[32, 0, false, true](fission._4_max_gradient.fj_bottom);
+ let out = fork-split(fission._4_max_gradient.fj_bottom);
+ clean-monoid-reduces(max_gradient);
+ simpl!(max_gradient);
+ let top = outline(fission._4_max_gradient.fj_top);
+ let bottom = outline(out._4_max_gradient.fj0);
+ gpu(top, bottom);
+ }
+} else {
+ simpl!(max_gradient);
+ fork-split(max_gradient);
+ unforkify(max_gradient);
+ gpu(max_gradient);
+}
+
ip-sroa(*);
sroa(*);
simpl!(*);
--
GitLab