diff --git a/Cargo.lock b/Cargo.lock
index 6642aef7745bf91996cb283de2c20c34953b3e90..0835939abb2314765a7ba7ee4ef90eefcd8e3144 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1366,7 +1366,29 @@ dependencies = [
]
[[package]]
-name = "juno_grape_reduction"
+name = "juno_grape_fft"
+version = "0.1.0"
+dependencies = [
+ "async-std",
+ "grape_sim",
+ "hercules_rt",
+ "juno_build",
+ "with_builtin_macros",
+]
+
+[[package]]
+name = "juno_grape_reduction_host"
+version = "0.1.0"
+dependencies = [
+ "async-std",
+ "grape_sim",
+ "hercules_rt",
+ "juno_build",
+ "with_builtin_macros",
+]
+
+[[package]]
+name = "juno_grape_reduction_tree"
version = "0.1.0"
dependencies = [
"async-std",
diff --git a/Cargo.toml b/Cargo.toml
index fd37ab44b5fe189caa1611962a040e7c836b6bfe..75521f82e1b6bf1f11c4cc6834918a8eea679db2 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -36,9 +36,10 @@ members = [
"juno_samples/schedule_test",
"juno_samples/simple3",
"juno_samples/grape",
- "juno_samples/grape_reduction",
+ "juno_samples/grape_reduction_tree",
"juno_samples/grape_conv",
- "juno_samples/grape_reduction",
+ "juno_samples/grape_fft",
+ "juno_samples/grape_reduction_host",
"juno_scheduler",
"juno_utils",
]
diff --git a/grape_sim/src/lib.rs b/grape_sim/src/lib.rs
index 06541a8a5e70105320517d71e0a35afe5aa590d0..9ca5469ebafd3742659debce7a66e66e9801a338 100644
--- a/grape_sim/src/lib.rs
+++ b/grape_sim/src/lib.rs
@@ -65,7 +65,8 @@ pub fn compute() -> [i16; 8] {
println!("row: {}, col: {}, ins: {:?} op: {:?}", row, col, ins, op);
let result = match op {
- hercules_cg::FuOp::Add => ins.0 + ins.1,
+ hercules_cg::FuOp::Add => ins.0.wrapping_add(*ins.1),
+ hercules_cg::FuOp::Sub => ins.0.wrapping_sub(*ins.1),
hercules_cg::FuOp::Mult => ins.0 * ins.1,
hercules_cg::FuOp::PassA => *ins.0,
hercules_cg::FuOp::Default => {
@@ -96,7 +97,7 @@ pub unsafe extern "C" fn await_valid() {}
pub unsafe extern "C" fn read_data(outputs: *mut i16) {
// This forces the computation.
let out = compute();
- print!("out: {:?}", out);
+ print!("out: {:?}\n", out);
for i in 0..7 {
*outputs.add(i) = out[i];
}
diff --git a/hercules_cg/src/grape.rs b/hercules_cg/src/grape.rs
index a7093f777a034be0a89044360fa04c169e14b6ea..085e1f9d53d461bbebde4a5c6f95e6bf31b1d3e7 100644
--- a/hercules_cg/src/grape.rs
+++ b/hercules_cg/src/grape.rs
@@ -74,6 +74,7 @@ impl FunctionalUnit {
let vec = match self.op_type {
FuOp::Add => vec![0, 0, 0, 0],
+ FuOp::Sub => vec![0, 1, 1, 1],
FuOp::Mult => vec![0, 0, 1, 0],
FuOp::PassA => vec![1, 1, 1, 1],
FuOp::Default => vec![1, 1, 1, 1],
@@ -133,6 +134,7 @@ where
#[derive(Clone, Debug, Copy, PartialEq, Serialize, Deserialize)]
pub enum FuOp {
Add,
+ Sub,
Mult,
PassA,
Default, // ...
@@ -441,6 +443,8 @@ where
// let binary_string: String = "010100101".to_owned();
std::fs::write("debug.txt", debug_str);
+ // todo!();
+
// write!(output, "{}", string);
// todo!();
@@ -1199,7 +1203,7 @@ where
for (i, item) in input_nodes.iter().enumerate() {
param_map.insert(*item, i);
}
- return self.schedule_row_recursive(0, param_map, input_mapping, config);
+ return self.schedule_row_recursive(0, param_map, input_mapping, config, HashSet::new());
}
fn schedule_row_recursive(
@@ -1208,6 +1212,7 @@ where
prev_mapping: HashMap<NodeID, usize>, // A node mapping for the previous row
mut input_mapping: Vec<NodeID>,
mut config: SliceDesc<H, W>,
+ mut computed_nodes: HashSet<NodeID>,
) -> Result<
(SliceDesc<H, W>, HashMap<NodeID, usize>, Vec<NodeID>),
(SliceDesc<H, W>, HashMap<NodeID, usize>, Vec<NodeID>),
@@ -1326,6 +1331,23 @@ where
dedup(&mut choices);
// Remove duplicates without changing order.
+ // Move killed nodes to back.
+ let mut i = 0;
+
+ println!("computed_nodes: {:?}", computed_nodes);
+
+ choices = choices
+ .iter()
+ .filter(|i| {
+ !self
+ .def_use_map
+ .get_users(**i)
+ .iter()
+ .all(|u| computed_nodes.contains(u))
+ })
+ .cloned()
+ .collect();
+
println!("num choices: {:?}", choices.clone().len());
println!("choices: {:?}", choices.clone());
@@ -1333,6 +1355,7 @@ where
for combo in choices.iter().combinations(W.min(choices.len())) {
println!("combo: {:?}", combo);
let mut next_mapping = HashMap::new();
+ let mut next_computed_nodes = computed_nodes.clone();
// Heurestic (Correctness): Just assume the first 8 are good choices.
for (col, node) in combo.iter().take(W).enumerate() {
@@ -1361,6 +1384,7 @@ where
} else {
// compute it
next_mapping.insert(**node, col);
+ next_computed_nodes.insert(**node);
// Collect inputs
let inputs: Vec<usize> = get_uses(&self.function.nodes[node.idx()])
@@ -1388,7 +1412,7 @@ where
let op_type = match self.function.nodes[node.idx()] {
Node::Binary { left, right, op } => match op {
BinaryOperator::Add => FuOp::Add,
- BinaryOperator::Sub => todo!(),
+ BinaryOperator::Sub => FuOp::Sub,
BinaryOperator::Mul => FuOp::Mult,
BinaryOperator::Div => todo!(),
BinaryOperator::Rem => todo!(),
@@ -1410,8 +1434,13 @@ where
}
}
- let schedule_attempt =
- self.schedule_row_recursive(row + 1, next_mapping, input_mapping.clone(), config);
+ let schedule_attempt = self.schedule_row_recursive(
+ row + 1,
+ next_mapping,
+ input_mapping.clone(),
+ config,
+ next_computed_nodes.clone(),
+ );
if schedule_attempt.is_ok() {
return schedule_attempt;
diff --git a/hercules_rt/src/grape_header.rs b/hercules_rt/src/grape_header.rs
index 63467b92a6e0fa8a53e403b671d8600212d204e5..93a93f2cf29108d923f46fcc2d0cb06145f72041 100644
--- a/hercules_rt/src/grape_header.rs
+++ b/hercules_rt/src/grape_header.rs
@@ -1,10 +1,9 @@
extern "C" {
fn program_bitstream(ptr: *const u8, len_bits: usize);
- fn send_data(inputs: *const i16); // len is 16
+ fn send_data(inputs: *const i16); // len is 16
fn await_valid();
fn read_data(outputs: *mut i16); // len is 7
-
-}
\ No newline at end of file
+}
diff --git a/hercules_rt/src/grape_lib.rs b/hercules_rt/src/grape_lib.rs
index b2be97bd4391a15df37fe784f68be30d8f74ef92..5a625ce1083b26eb8f74d9471643069a8b0c43d3 100644
--- a/hercules_rt/src/grape_lib.rs
+++ b/hercules_rt/src/grape_lib.rs
@@ -1,20 +1,7 @@
// Grape stuff
#[no_mangle]
-pub unsafe extern "C" fn program_bitstream(ptr: *const u8, len: usize) {
- // Example: read the bytes from the pointer
- if ptr.is_null() {
- panic!("Null pointer passed to program_bitstream");
- }
-
- let slice = std::slice::from_raw_parts(ptr, len);
-
- println!("Received bitstream of length {}", slice.len());
-
- for (i, byte) in slice.iter().enumerate() {
- println!("Byte {}: {:08b}", i, byte);
- }
-}
+pub unsafe extern "C" fn program_bitstream(ptr: *const u8, len: usize) {}
#[no_mangle]
pub unsafe extern "C" fn send_data(inputs: *const i16) {} // len is 16
diff --git a/juno_samples/grape_conv/src/grape.sch b/juno_samples/grape_conv/src/grape.sch
index 3aef5b9af52356456d080dcdf5ce0bca1d2263ee..664c0725c2776bf1d4fcbd88bc929fcd8caee0ee 100644
--- a/juno_samples/grape_conv/src/grape.sch
+++ b/juno_samples/grape_conv/src/grape.sch
@@ -26,5 +26,5 @@ ccp(*);
simplify-cfg(*);
dce(*);
grape(wrapper);
-xdot[true](*);
+xdot[true](wrapper);
gcm(*);
diff --git a/juno_samples/grape_conv/src/main.rs b/juno_samples/grape_conv/src/main.rs
index bf509d5d118324ef0982cbd1a5ad6cf847c34c6f..cad2a6ec5e96a682370a3238b22f630dccc79ee8 100644
--- a/juno_samples/grape_conv/src/main.rs
+++ b/juno_samples/grape_conv/src/main.rs
@@ -1,5 +1,7 @@
#![feature(concat_idents)]
+use std::time::Instant;
+
#[cfg(feature = "cuda")]
use hercules_rt::CUDABox;
@@ -49,10 +51,14 @@ fn main() {
{
let a = HerculesCPURef::from_slice(&a);
let b = HerculesCPURef::from_slice(&b);
+ let start = Instant::now();
let mut r = runner!(entry);
let c = r.run(a, b).await;
print!("{:?}", c);
+ let duration = start.elapsed();
+ println!("Time elapsed in some_function() is: {:?}", duration);
+ assert!(false);
assert_eq!(c, (result[0], result[1], result[2], result[3]));
}
#[cfg(feature = "cuda")]
diff --git a/juno_samples/grape_reduction/Cargo.toml b/juno_samples/grape_fft/Cargo.toml
similarity index 88%
rename from juno_samples/grape_reduction/Cargo.toml
rename to juno_samples/grape_fft/Cargo.toml
index 683268b6368948bc6986d344641ae9ba02f79cfe..7b479813af19ffd84a6e6e153ebb750c3a28ae01 100644
--- a/juno_samples/grape_reduction/Cargo.toml
+++ b/juno_samples/grape_fft/Cargo.toml
@@ -1,11 +1,11 @@
[package]
-name = "juno_grape_reduction"
+name = "juno_grape_fft"
version = "0.1.0"
authors = ["Xavier Routh <xrouth2@illinois.edu>"]
edition = "2021"
[[bin]]
-name = "juno_grape_reduction"
+name = "juno_grape_fft"
path = "src/main.rs"
[features]
diff --git a/juno_samples/grape_fft/build.rs b/juno_samples/grape_fft/build.rs
new file mode 100644
index 0000000000000000000000000000000000000000..2897d0e019e4b1983d5b3db4b4210d0f2ce80fe9
--- /dev/null
+++ b/juno_samples/grape_fft/build.rs
@@ -0,0 +1,35 @@
+use juno_build::JunoCompiler;
+
+fn main() {
+ #[cfg(not(feature = "cuda"))]
+ {
+ JunoCompiler::new()
+ .file_in_src("fft.jn")
+ .unwrap()
+ .schedule_in_src("cpu.sch")
+ .unwrap()
+ .build()
+ .unwrap();
+ }
+ #[cfg(feature = "grape")]
+ {
+ JunoCompiler::new()
+ .file_in_src("fft.jn")
+ .unwrap()
+ .schedule_in_src("grape.sch")
+ .unwrap()
+ .build()
+ .unwrap();
+ }
+
+ #[cfg(feature = "cuda")]
+ {
+ JunoCompiler::new()
+ .file_in_src("simple.jn")
+ .unwrap()
+ .schedule_in_src("gpu.sch")
+ .unwrap()
+ .build()
+ .unwrap();
+ }
+}
diff --git a/juno_samples/grape_reduction/src/cpu.sch b/juno_samples/grape_fft/src/cpu.sch
similarity index 100%
rename from juno_samples/grape_reduction/src/cpu.sch
rename to juno_samples/grape_fft/src/cpu.sch
diff --git a/juno_samples/grape_fft/src/fft.jn b/juno_samples/grape_fft/src/fft.jn
new file mode 100644
index 0000000000000000000000000000000000000000..f8d428d43154426df96f23e6fee6052463442d56
--- /dev/null
+++ b/juno_samples/grape_fft/src/fft.jn
@@ -0,0 +1,46 @@
+fn fft(a : i16[8]) -> i16, i16, i16, i16, i16, i16, i16, i16 {
+ let out : i16[8];
+ let r0 = a[0];
+ let r1 = a[1];
+ let r2 = a[2];
+ let r3 = a[3];
+ let i0 = a[4];
+ let i1 = a[5];
+ let i2 = a[6];
+ let i3 = a[7];
+
+ let r02 = r0 + r2;
+ let r13 = r1 + r3;
+ let i02 = i0 + i2;
+ let i13 = i1 + i3;
+ let r0m2 = r0 - r2;
+ let r1m3 = r1 - r3;
+ let i0m2 = i0 - i2;
+ let i1m3 = i1 - i3;
+
+ out[0] = r02 + r13;
+ out[4] = i02 + i13;
+ out[2] = r02 - r13;
+ out[6] = i02 - i13;
+ out[1] = r0m2 + i1m3;
+ out[5] = i0m2 - r1m3;
+ out[3] = r0m2 - i1m3;
+ out[7] = i0m2 + r1m3;
+ return out[0], out[1], out[2], out[3], out[4], out[5], out[6], out[7];
+}
+
+#[entry]
+fn entry(a: i16[8]) -> i16[8] {
+ let out: i16[8];
+ let a, b, c, d, e, f, g, h = fft(a);
+ out[0] = a;
+ out[1] = b;
+ out[2] = c;
+ out[3] = d;
+ out[4] = e;
+ out[5] = f;
+ out[6] = g;
+ out[7] = h;
+ return out;
+}
+
diff --git a/juno_samples/grape_fft/src/grape.sch b/juno_samples/grape_fft/src/grape.sch
new file mode 100644
index 0000000000000000000000000000000000000000..49486d91178dce73fc41f9a6ef64162d273926df
--- /dev/null
+++ b/juno_samples/grape_fft/src/grape.sch
@@ -0,0 +1,31 @@
+gvn(*);
+phi-elim(*);
+ccp(*);
+simplify-cfg(*);
+dce(*);
+
+delete-uncalled(*);
+
+fixpoint stop after 10 {
+ forkify(*);
+ fork-guard-elim(*);
+ fork-unroll(*);
+ predication(*);
+ gvn(*);
+ phi-elim(*);
+ ccp(*);
+ simplify-cfg(*);
+ dce(*);
+ lift-dc-math(*);
+}
+
+a2p(*);
+sroa(*);
+gvn(*);
+phi-elim(*);
+ccp(*);
+simplify-cfg(*);
+dce(*);
+grape(fft);
+xdot[true](*);
+gcm(*);
diff --git a/juno_samples/grape_fft/src/main.rs b/juno_samples/grape_fft/src/main.rs
new file mode 100644
index 0000000000000000000000000000000000000000..947be6431397ae1ccbc271abcbe9843ea2e06c9c
--- /dev/null
+++ b/juno_samples/grape_fft/src/main.rs
@@ -0,0 +1,45 @@
+#![feature(concat_idents)]
+
+use std::time::Instant;
+
+#[cfg(feature = "cuda")]
+use hercules_rt::CUDABox;
+
+use hercules_rt::{runner, HerculesCPURef};
+
+juno_build::juno!("fft");
+
+#[cfg(feature = "grape")]
+use grape_sim::*;
+fn main() {
+ async_std::task::block_on(async {
+ let a: Box<[i16]> = (1..=8).collect::<Vec<_>>().into_boxed_slice();
+
+ #[cfg(not(feature = "cuda"))]
+ {
+ let a = HerculesCPURef::from_slice(&a);
+ let start = Instant::now();
+
+ let mut r = runner!(entry);
+ let c = r.run(a).await;
+ let duration = start.elapsed();
+ println!("Time elapsed in some_function() is: {:?}", duration);
+
+ assert!(false);
+ print!("{:?}", c);
+ }
+ #[cfg(feature = "cuda")]
+ {
+ let a = CUDABox::from_cpu_ref(HerculesCPURef::from_slice(&a));
+ let b = CUDABox::from_cpu_ref(HerculesCPURef::from_slice(&b));
+ let mut r = runner!(simple);
+ let c = r.run(8, a.get_ref(), b.get_ref()).await;
+ assert_eq!(c, 120);
+ }
+ });
+}
+
+#[test]
+fn simple3_test() {
+ main();
+}
diff --git a/juno_samples/grape_reduction_host/Cargo.toml b/juno_samples/grape_reduction_host/Cargo.toml
new file mode 100644
index 0000000000000000000000000000000000000000..f38a566c11282b4dc1413f6add670723b4bddf29
--- /dev/null
+++ b/juno_samples/grape_reduction_host/Cargo.toml
@@ -0,0 +1,23 @@
+[package]
+name = "juno_grape_reduction_host"
+version = "0.1.0"
+authors = ["Xavier Routh <xrouth2@illinois.edu>"]
+edition = "2021"
+
+[[bin]]
+name = "juno_grape_reduction_host"
+path = "src/main.rs"
+
+[features]
+cuda = ["juno_build/cuda", "hercules_rt/cuda"]
+grape = []
+
+[build-dependencies]
+juno_build = { path = "../../juno_build" }
+
+[dependencies]
+juno_build = { path = "../../juno_build" }
+hercules_rt = { path = "../../hercules_rt" }
+grape_sim = { path = "../../grape_sim" }
+with_builtin_macros = "0.1.0"
+async-std = "*"
diff --git a/juno_samples/grape_reduction/build.rs b/juno_samples/grape_reduction_host/build.rs
similarity index 100%
rename from juno_samples/grape_reduction/build.rs
rename to juno_samples/grape_reduction_host/build.rs
diff --git a/juno_samples/grape_reduction_host/src/cpu.sch b/juno_samples/grape_reduction_host/src/cpu.sch
new file mode 100644
index 0000000000000000000000000000000000000000..7934b277d2ccc3daa385f95182ad6555dd4040f1
--- /dev/null
+++ b/juno_samples/grape_reduction_host/src/cpu.sch
@@ -0,0 +1,18 @@
+gvn(*);
+phi-elim(*);
+dce(*);
+
+
+ip-sroa(*);
+sroa(*);
+dce(*);
+gvn(*);
+phi-elim(*);
+dce(*);
+
+infer-schedules(*);
+
+
+gcm(*);
+dce(*);
+gcm(*);
diff --git a/juno_samples/grape_reduction/src/grape.sch b/juno_samples/grape_reduction_host/src/grape.sch
similarity index 95%
rename from juno_samples/grape_reduction/src/grape.sch
rename to juno_samples/grape_reduction_host/src/grape.sch
index d098a3327e7790ccbea13a38e3abe7d1f2c77ef0..0265e729f52a2108cc6c07bba7fbe2bb6649f27c 100644
--- a/juno_samples/grape_reduction/src/grape.sch
+++ b/juno_samples/grape_reduction_host/src/grape.sch
@@ -8,7 +8,7 @@ inline(fake_entry);
delete-uncalled(*);
forkify(*);
-fork-tile[4, 0, false, true](*);
+fork-tile[8, 0, false, true](*);
let a = fork-split(*);
print[a._1_fake_entry.fj1]();
let inner = outline(a._1_fake_entry.fj1);
@@ -53,7 +53,7 @@ fork-unroll(a._1_fake_entry.fj0);
// xdot[true](*);
-reassociate(inner);
+// reassociate(inner);
diff --git a/juno_samples/grape_reduction_host/src/grape3.sch b/juno_samples/grape_reduction_host/src/grape3.sch
new file mode 100644
index 0000000000000000000000000000000000000000..248254316f8209ed3a4800c070636892e19ac07a
--- /dev/null
+++ b/juno_samples/grape_reduction_host/src/grape3.sch
@@ -0,0 +1,45 @@
+gvn(*);
+phi-elim(*);
+ccp(*);
+simplify-cfg(*);
+dce(*);
+
+inline(fake_entry);
+delete-uncalled(*);
+
+forkify(*);
+fork-guard-elim(*);
+dce(*);
+
+
+fixpoint stop after 10 {
+ forkify(*);
+ fork-guard-elim(*);
+ fork-unroll(*);
+ predication(*);
+ gvn(*);
+ phi-elim(*);
+ ccp(*);
+ simplify-cfg(*);
+ dce(*);
+ lift-dc-math(*);
+}
+// xdot[true](*);
+
+a2p(*);
+sroa(*);
+xdot[true](*);
+
+// reassociate go brr
+reassociate(*);
+xdot[true](*);
+
+gvn(*);
+phi-elim(*);
+ccp(*);
+simplify-cfg(*);
+dce(*);
+grape(fake_entry);
+xdot[true](*);
+gcm(*);
+xdot[true](*);
diff --git a/juno_samples/grape_reduction_host/src/main.rs b/juno_samples/grape_reduction_host/src/main.rs
new file mode 100644
index 0000000000000000000000000000000000000000..81e5ec0cdf12a2ca3ef365d552c4fd0841f12e33
--- /dev/null
+++ b/juno_samples/grape_reduction_host/src/main.rs
@@ -0,0 +1,50 @@
+#![feature(concat_idents)]
+
+use std::time::Instant;
+
+#[cfg(feature = "cuda")]
+use hercules_rt::CUDABox;
+
+use hercules_rt::{runner, HerculesCPURef};
+
+juno_build::juno!("simple");
+
+#[cfg(feature = "grape")]
+use grape_sim::*;
+
+fn main() {
+ async_std::task::block_on(async {
+ let size = 64;
+ let a: Box<[i16]> = (1..=size).collect::<Vec<_>>().into_boxed_slice();
+ let sum: i16 = a.iter().sum();
+ #[cfg(not(feature = "cuda"))]
+ {
+ let a = HerculesCPURef::from_slice(&a);
+ let start = Instant::now();
+
+ let mut r = runner!(entry);
+ let c = r.run(a).await;
+ let duration = start.elapsed();
+
+ println!("Time elapsed in some_function() is: {:?}", duration);
+
+ assert!(false);
+ print!("{:?}", c);
+
+ assert_eq!(c, sum);
+ }
+ #[cfg(feature = "cuda")]
+ {
+ let a = CUDABox::from_cpu_ref(HerculesCPURef::from_slice(&a));
+ let b = CUDABox::from_cpu_ref(HerculesCPURef::from_slice(&b));
+ let mut r = runner!(simple);
+ let c = r.run(8, a.get_ref(), b.get_ref()).await;
+ assert_eq!(c, 120);
+ }
+ });
+}
+
+#[test]
+fn simple3_test() {
+ main();
+}
diff --git a/juno_samples/grape_reduction_host/src/simple.jn b/juno_samples/grape_reduction_host/src/simple.jn
new file mode 100644
index 0000000000000000000000000000000000000000..bd94baf7c0c581038e85a83930e56b83defd91ed
--- /dev/null
+++ b/juno_samples/grape_reduction_host/src/simple.jn
@@ -0,0 +1,27 @@
+fn reduce<n: usize>(a : i16[n]) -> i16 {
+ let acc: i16 = 0;
+
+ for i = 0 to n {
+ acc += a[i];
+ }
+
+ return acc;
+}
+
+fn fake_entry(a: i16[64]) -> i16 {
+ @this {
+ let r = reduce::<64>(a);
+
+ return r;
+ }
+}
+
+#[entry]
+fn entry(a: i16[64]) -> i16 {
+ return fake_entry(a);
+}
+
+
+
+
+
diff --git a/juno_samples/grape_reduction_tree/Cargo.toml b/juno_samples/grape_reduction_tree/Cargo.toml
new file mode 100644
index 0000000000000000000000000000000000000000..a20c606ba90d814bbd49d9783613d89d558a97c7
--- /dev/null
+++ b/juno_samples/grape_reduction_tree/Cargo.toml
@@ -0,0 +1,23 @@
+[package]
+name = "juno_grape_reduction_tree"
+version = "0.1.0"
+authors = ["Xavier Routh <xrouth2@illinois.edu>"]
+edition = "2021"
+
+[[bin]]
+name = "juno_grape_reduction_tree"
+path = "src/main.rs"
+
+[features]
+cuda = ["juno_build/cuda", "hercules_rt/cuda"]
+grape = []
+
+[build-dependencies]
+juno_build = { path = "../../juno_build" }
+
+[dependencies]
+juno_build = { path = "../../juno_build" }
+hercules_rt = { path = "../../hercules_rt" }
+grape_sim = { path = "../../grape_sim" }
+with_builtin_macros = "0.1.0"
+async-std = "*"
diff --git a/juno_samples/grape_reduction_tree/build.rs b/juno_samples/grape_reduction_tree/build.rs
new file mode 100644
index 0000000000000000000000000000000000000000..5ee28f2ccace52074541d72014082dbe056f1f14
--- /dev/null
+++ b/juno_samples/grape_reduction_tree/build.rs
@@ -0,0 +1,35 @@
+use juno_build::JunoCompiler;
+
+fn main() {
+ #[cfg(not(feature = "cuda"))]
+ {
+ JunoCompiler::new()
+ .file_in_src("simple.jn")
+ .unwrap()
+ .schedule_in_src("cpu.sch")
+ .unwrap()
+ .build()
+ .unwrap();
+ }
+ #[cfg(feature = "grape")]
+ {
+ JunoCompiler::new()
+ .file_in_src("simple.jn")
+ .unwrap()
+ .schedule_in_src("grape.sch")
+ .unwrap()
+ .build()
+ .unwrap();
+ }
+
+ #[cfg(feature = "cuda")]
+ {
+ JunoCompiler::new()
+ .file_in_src("simple.jn")
+ .unwrap()
+ .schedule_in_src("gpu.sch")
+ .unwrap()
+ .build()
+ .unwrap();
+ }
+}
diff --git a/juno_samples/grape_reduction_tree/src/cpu.sch b/juno_samples/grape_reduction_tree/src/cpu.sch
new file mode 100644
index 0000000000000000000000000000000000000000..7934b277d2ccc3daa385f95182ad6555dd4040f1
--- /dev/null
+++ b/juno_samples/grape_reduction_tree/src/cpu.sch
@@ -0,0 +1,18 @@
+gvn(*);
+phi-elim(*);
+dce(*);
+
+
+ip-sroa(*);
+sroa(*);
+dce(*);
+gvn(*);
+phi-elim(*);
+dce(*);
+
+infer-schedules(*);
+
+
+gcm(*);
+dce(*);
+gcm(*);
diff --git a/juno_samples/grape_reduction_tree/src/grape.sch b/juno_samples/grape_reduction_tree/src/grape.sch
new file mode 100644
index 0000000000000000000000000000000000000000..248254316f8209ed3a4800c070636892e19ac07a
--- /dev/null
+++ b/juno_samples/grape_reduction_tree/src/grape.sch
@@ -0,0 +1,45 @@
+gvn(*);
+phi-elim(*);
+ccp(*);
+simplify-cfg(*);
+dce(*);
+
+inline(fake_entry);
+delete-uncalled(*);
+
+forkify(*);
+fork-guard-elim(*);
+dce(*);
+
+
+fixpoint stop after 10 {
+ forkify(*);
+ fork-guard-elim(*);
+ fork-unroll(*);
+ predication(*);
+ gvn(*);
+ phi-elim(*);
+ ccp(*);
+ simplify-cfg(*);
+ dce(*);
+ lift-dc-math(*);
+}
+// xdot[true](*);
+
+a2p(*);
+sroa(*);
+xdot[true](*);
+
+// reassociate go brr
+reassociate(*);
+xdot[true](*);
+
+gvn(*);
+phi-elim(*);
+ccp(*);
+simplify-cfg(*);
+dce(*);
+grape(fake_entry);
+xdot[true](*);
+gcm(*);
+xdot[true](*);
diff --git a/juno_samples/grape_reduction/src/main.rs b/juno_samples/grape_reduction_tree/src/main.rs
similarity index 82%
rename from juno_samples/grape_reduction/src/main.rs
rename to juno_samples/grape_reduction_tree/src/main.rs
index 1c209ca181ad5e258cfe6ead86aceb19d534c9d9..d113aa0abc20ce07200273a019e0774e535c3faa 100644
--- a/juno_samples/grape_reduction/src/main.rs
+++ b/juno_samples/grape_reduction_tree/src/main.rs
@@ -1,5 +1,7 @@
#![feature(concat_idents)]
+use std::time::Instant;
+
#[cfg(feature = "cuda")]
use hercules_rt::CUDABox;
@@ -17,10 +19,15 @@ fn main() {
#[cfg(not(feature = "cuda"))]
{
let a = HerculesCPURef::from_slice(&a);
+ let start = Instant::now();
+
let mut r = runner!(entry);
let c = r.run(a).await;
print!("{:?}", c);
+ let duration = start.elapsed();
+ println!("Time elapsed in some_function() is: {:?}", duration);
+ assert!(false);
assert_eq!(c, sum);
}
#[cfg(feature = "cuda")]
diff --git a/juno_samples/grape_reduction/src/simple.jn b/juno_samples/grape_reduction_tree/src/simple.jn
similarity index 100%
rename from juno_samples/grape_reduction/src/simple.jn
rename to juno_samples/grape_reduction_tree/src/simple.jn