diff --git a/hercules_ir/src/ir.rs b/hercules_ir/src/ir.rs
index 5dfe2915f5f3e30f56b6665dc27d23cd40cca3d4..f6aafa35bd2c2d94324e63b2b91213ad0c2e9c4f 100644
--- a/hercules_ir/src/ir.rs
+++ b/hercules_ir/src/ir.rs
@@ -1048,9 +1048,20 @@ impl Constant {
}
}
- /*
- * Useful for GVN.
- */
+ pub fn is_false(&self) -> bool {
+ match self {
+ Constant::Boolean(false) => true,
+ _ => false,
+ }
+ }
+
+ pub fn is_true(&self) -> bool {
+ match self {
+ Constant::Boolean(true) => true,
+ _ => false,
+ }
+ }
+
pub fn is_zero(&self) -> bool {
match self {
Constant::Integer8(0) => true,
diff --git a/hercules_opt/src/inline.rs b/hercules_opt/src/inline.rs
index 38ed1b22d2d81be971aa867857fd01e3752ecc0a..9b0a9200b6301a8928f5fda2f70b515fc1d45dde 100644
--- a/hercules_opt/src/inline.rs
+++ b/hercules_opt/src/inline.rs
@@ -307,7 +307,11 @@ impl ParameterLattice {
* These functions can have that constant "inlined" - the parameter is removed
* and all uses of the parameter becomes uses of the constant directly.
*/
-pub fn const_inline(editors: &mut [FunctionEditor], callgraph: &CallGraph) {
+pub fn const_inline(
+ editors: &mut [FunctionEditor],
+ callgraph: &CallGraph,
+ inline_collections: bool,
+) {
// Run const inlining on each function, starting at the most shallow
// function first, since we want to propagate constants down the call graph.
for func_id in callgraph.topo().into_iter().rev() {
@@ -361,22 +365,29 @@ pub fn const_inline(editors: &mut [FunctionEditor], callgraph: &CallGraph) {
let mut param_tys = edit.get_param_types().clone();
let mut decrement_index_by = 0;
for idx in 0..param_tys.len() {
- if let Some(node) = match param_lattice[idx] {
- ParameterLattice::Top => Some(Node::Undef { ty: param_tys[idx] }),
- ParameterLattice::Constant(id) => Some(Node::Constant { id }),
- ParameterLattice::DynamicConstant(id, _) => {
- // Rust moment.
- let maybe_cons = edit.get_dynamic_constant(id).try_constant();
- if let Some(val) = maybe_cons {
- Some(Node::DynamicConstant {
- id: edit.add_dynamic_constant(DynamicConstant::Constant(val)),
- })
- } else {
- None
+ if (inline_collections
+ || edit
+ .get_type(param_tys[idx - decrement_index_by])
+ .is_primitive())
+ && let Some(node) = match param_lattice[idx] {
+ ParameterLattice::Top => Some(Node::Undef {
+ ty: param_tys[idx - decrement_index_by],
+ }),
+ ParameterLattice::Constant(id) => Some(Node::Constant { id }),
+ ParameterLattice::DynamicConstant(id, _) => {
+ // Rust moment.
+ let maybe_cons = edit.get_dynamic_constant(id).try_constant();
+ if let Some(val) = maybe_cons {
+ Some(Node::DynamicConstant {
+ id: edit.add_dynamic_constant(DynamicConstant::Constant(val)),
+ })
+ } else {
+ None
+ }
}
+ _ => None,
}
- _ => None,
- } && let Some(ids) = param_idx_to_ids.get(&idx)
+ && let Some(ids) = param_idx_to_ids.get(&idx)
{
let node = edit.add_node(node);
for id in ids {
diff --git a/hercules_opt/src/pred.rs b/hercules_opt/src/pred.rs
index ed7c3a855b016608aa194cc9f2cd89f05d836bde..8f1d07454262a621d8f9c753a3c68b01752b3dc7 100644
--- a/hercules_opt/src/pred.rs
+++ b/hercules_opt/src/pred.rs
@@ -136,6 +136,77 @@ pub fn predication(editor: &mut FunctionEditor, typing: &Vec<TypeID>) {
bad_branches.insert(branch);
}
}
+
+ // Do a quick and dirty rewrite to convert select(a, b, false) to a && b and
+ // select(a, b, true) to a || b.
+ for id in editor.node_ids() {
+ let nodes = &editor.func().nodes;
+ if let Node::Ternary {
+ op: TernaryOperator::Select,
+ first,
+ second,
+ third,
+ } = nodes[id.idx()]
+ {
+ if let Some(cons) = nodes[second.idx()].try_constant()
+ && editor.get_constant(cons).is_false()
+ {
+ editor.edit(|mut edit| {
+ let inv = edit.add_node(Node::Unary {
+ op: UnaryOperator::Not,
+ input: first,
+ });
+ let node = edit.add_node(Node::Binary {
+ op: BinaryOperator::And,
+ left: inv,
+ right: third,
+ });
+ edit = edit.replace_all_uses(id, node)?;
+ edit.delete_node(id)
+ });
+ } else if let Some(cons) = nodes[third.idx()].try_constant()
+ && editor.get_constant(cons).is_false()
+ {
+ editor.edit(|mut edit| {
+ let node = edit.add_node(Node::Binary {
+ op: BinaryOperator::And,
+ left: first,
+ right: second,
+ });
+ edit = edit.replace_all_uses(id, node)?;
+ edit.delete_node(id)
+ });
+ } else if let Some(cons) = nodes[second.idx()].try_constant()
+ && editor.get_constant(cons).is_true()
+ {
+ editor.edit(|mut edit| {
+ let node = edit.add_node(Node::Binary {
+ op: BinaryOperator::Or,
+ left: first,
+ right: third,
+ });
+ edit = edit.replace_all_uses(id, node)?;
+ edit.delete_node(id)
+ });
+ } else if let Some(cons) = nodes[third.idx()].try_constant()
+ && editor.get_constant(cons).is_true()
+ {
+ editor.edit(|mut edit| {
+ let inv = edit.add_node(Node::Unary {
+ op: UnaryOperator::Not,
+ input: first,
+ });
+ let node = edit.add_node(Node::Binary {
+ op: BinaryOperator::Or,
+ left: inv,
+ right: second,
+ });
+ edit = edit.replace_all_uses(id, node)?;
+ edit.delete_node(id)
+ });
+ }
+ }
+ }
}
/*
diff --git a/hercules_opt/src/schedule.rs b/hercules_opt/src/schedule.rs
index d7ae40488d75a1da7ef65b8a53a894bc0f62cded..9bc7823ee7f5837cf49387170e548a9174340f42 100644
--- a/hercules_opt/src/schedule.rs
+++ b/hercules_opt/src/schedule.rs
@@ -69,6 +69,26 @@ pub fn infer_parallel_reduce(
chain_id = reduct;
}
+ // If the use is a phi that uses the reduce and a write, then we might
+ // want to parallelize this still. Set the chain ID to the write.
+ if let Node::Phi {
+ control: _,
+ ref data,
+ } = func.nodes[chain_id.idx()]
+ && data.len()
+ == data
+ .into_iter()
+ .filter(|phi_use| **phi_use == last_reduce)
+ .count()
+ + 1
+ {
+ chain_id = *data
+ .into_iter()
+ .filter(|phi_use| **phi_use != last_reduce)
+ .next()
+ .unwrap();
+ }
+
// Check for a Write-Reduce tight cycle.
if let Node::Write {
collect,
@@ -130,12 +150,13 @@ pub fn infer_monoid_reduce(
reduce_cycles: &HashMap<NodeID, HashSet<NodeID>>,
) {
let is_binop_monoid = |op| {
- matches!(
- op,
- BinaryOperator::Add | BinaryOperator::Mul | BinaryOperator::Or | BinaryOperator::And
- )
+ op == BinaryOperator::Add
+ || op == BinaryOperator::Mul
+ || op == BinaryOperator::Or
+ || op == BinaryOperator::And
};
- let is_intrinsic_monoid = |intrinsic| matches!(intrinsic, Intrinsic::Max | Intrinsic::Min);
+ let is_intrinsic_monoid =
+ |intrinsic| intrinsic == Intrinsic::Max || intrinsic == Intrinsic::Min;
for id in editor.node_ids() {
let func = editor.func();
diff --git a/juno_samples/rodinia/bfs/src/bfs.jn b/juno_samples/rodinia/bfs/src/bfs.jn
index 51dcd945429dfde02cb2313afa404e81f8722c84..2534a89c627f137bf4a65a7f3d61879c3d3670e6 100644
--- a/juno_samples/rodinia/bfs/src/bfs.jn
+++ b/juno_samples/rodinia/bfs/src/bfs.jn
@@ -43,10 +43,10 @@ fn bfs<n, m: usize>(graph_nodes: Node[n], source: u32, edges: u32[m]) -> i32[n]
}
@loop2 for i in 0..n {
+ stop = stop && !updated[i];
if updated[i] {
mask[i] = true;
visited[i] = true;
- stop = false;
updated[i] = false;
}
}
diff --git a/juno_samples/rodinia/bfs/src/cpu.sch b/juno_samples/rodinia/bfs/src/cpu.sch
index 44cfa8ad0161fac0afbccc2d383637ec8a2f1aa0..ae67fdd987e961a95311a7d3aaa0f94fe31f1687 100644
--- a/juno_samples/rodinia/bfs/src/cpu.sch
+++ b/juno_samples/rodinia/bfs/src/cpu.sch
@@ -23,7 +23,8 @@ fixpoint {
fork-guard-elim(*);
}
simpl!(*);
+predication(*);
+simpl!(*);
unforkify(*);
-
gcm(*);
diff --git a/juno_samples/rodinia/cfd/src/cpu_euler.sch b/juno_samples/rodinia/cfd/src/cpu_euler.sch
index 5fe48a8395cfb6fada1d668b4f73fa6eb3487f5e..1244f80e54fdad43f58e5c5a5af44646b7a83e89 100644
--- a/juno_samples/rodinia/cfd/src/cpu_euler.sch
+++ b/juno_samples/rodinia/cfd/src/cpu_euler.sch
@@ -24,7 +24,8 @@ fixpoint {
fork-guard-elim(*);
}
simpl!(*);
+no-memset(compute_step_factor@res, compute_flux@res, copy_vars@res);
+parallel-reduce(time_step, copy_vars, compute_flux@outer_loop \ compute_flux@inner_loop);
unforkify(*);
-
gcm(*);
diff --git a/juno_samples/rodinia/cfd/src/euler.jn b/juno_samples/rodinia/cfd/src/euler.jn
index 203cfd96008237f57ec276973d70304e56159682..6966f5ba0887388cf02bafcf80ed66e4059b8b7d 100644
--- a/juno_samples/rodinia/cfd/src/euler.jn
+++ b/juno_samples/rodinia/cfd/src/euler.jn
@@ -47,7 +47,7 @@ fn compute_speed_of_sound(density: f32, pressure: f32) -> f32 {
}
fn compute_step_factor<nelr: usize>(variables: Variables::<nelr>, areas: f32[nelr]) -> f32[nelr] {
- let step_factors : f32[nelr];
+ @res let step_factors : f32[nelr];
for i in 0..nelr {
let density = variables.density[i];
@@ -106,9 +106,9 @@ fn compute_flux<nelr: usize>(
ff_flux_contribution_momentum_z: float3,
) -> Variables::<nelr> {
const smoothing_coefficient : f32 = 0.2;
- let fluxes: Variables::<nelr>;
+ @res let fluxes: Variables::<nelr>;
- for i in 0..nelr {
+ @outer_loop for i in 0..nelr {
let density_i = variables.density[i];
let momentum_i = float3 { x: variables.momentum.x[i],
@@ -131,7 +131,7 @@ fn compute_flux<nelr: usize>(
let flux_i_momentum = float3 { x: 0.0, y: 0.0, z: 0.0 };
let flux_i_density_energy : f32 = 0.0;
- for j in 0..NNB {
+ @inner_loop for j in 0..NNB {
let nb = elements_surrounding_elements[j, i];
let normal = float3 {
x: normals.x[j, i],
@@ -249,7 +249,7 @@ fn time_step<nelr: usize>(
}
fn copy_vars<nelr: usize>(variables: Variables::<nelr>) -> Variables::<nelr> {
- let result : Variables::<nelr>;
+ @res let result : Variables::<nelr>;
for i in 0..nelr {
result.density[i] = variables.density[i];
diff --git a/juno_samples/rodinia/srad/benches/srad_bench.rs b/juno_samples/rodinia/srad/benches/srad_bench.rs
index d327454002a6f9cabe4c40f74098570ea0d22d66..728702d9bcc18405ef291945f81413f49f5715af 100644
--- a/juno_samples/rodinia/srad/benches/srad_bench.rs
+++ b/juno_samples/rodinia/srad/benches/srad_bench.rs
@@ -13,8 +13,8 @@ fn srad_bench(c: &mut Criterion) {
let mut r = runner!(srad);
let niter = 100;
let lambda = 0.5;
- let nrows = 502;
- let ncols = 458;
+ let nrows = 512;
+ let ncols = 512;
let image = "data/image.pgm".to_string();
let Image {
image: image_ori,
diff --git a/juno_samples/rodinia/srad/src/cpu.sch b/juno_samples/rodinia/srad/src/cpu.sch
index 1a81ddad3b55bcf9ffb76660ebdc1069338affd4..2b45e8c956e10cb6af538282df98e32eb35b6b5e 100644
--- a/juno_samples/rodinia/srad/src/cpu.sch
+++ b/juno_samples/rodinia/srad/src/cpu.sch
@@ -28,6 +28,7 @@ fixpoint {
fork-coalesce(*);
}
simpl!(*);
+fork-interchange[0, 1](loop1);
fork-split(*);
unforkify(*);
diff --git a/juno_samples/rodinia/srad/src/gpu.sch b/juno_samples/rodinia/srad/src/gpu.sch
index 149d5cd2fd71005ade5cdbb3461e08b3e65ab34f..289548f9e01cdf402a3e1b1057fa52d4029f6173 100644
--- a/juno_samples/rodinia/srad/src/gpu.sch
+++ b/juno_samples/rodinia/srad/src/gpu.sch
@@ -1,23 +1,57 @@
-gvn(*);
-dce(*);
+macro simpl!(X) {
+ ccp(X);
+ simplify-cfg(X);
+ lift-dc-math(X);
+ gvn(X);
+ phi-elim(X);
+ dce(X);
+ infer-schedules(X);
+}
+
phi-elim(*);
-dce(*);
+let sum_loop = outline(srad@loop1);
+let main_loops = outline(srad@loop2 | srad@loop3);
+gpu(main_loops, extract, compress);
+simpl!(*);
+const-inline[true](*);
crc(*);
-dce(*);
slf(*);
-dce(*);
-
-let auto = auto-outline(srad);
-gpu(auto.srad);
-
-inline(auto.srad);
-inline(auto.srad);
-delete-uncalled(*);
+write-predication(*);
+simpl!(*);
+predication(*);
+simpl!(*);
+predication(*);
+simpl!(*);
+fixpoint {
+ forkify(*);
+ fork-guard-elim(*);
+ fork-coalesce(*);
+}
+simpl!(*);
+reduce-slf(*);
+simpl!(*);
+array-slf(*);
+simpl!(*);
+slf(*);
+simpl!(*);
-sroa[false](auto.srad);
-dce(*);
-float-collections(*);
-dce(*);
+fork-dim-merge(sum_loop);
+simpl!(sum_loop);
+fork-tile[32, 0, false, true](sum_loop);
+let out = fork-split(sum_loop);
+clean-monoid-reduces(sum_loop);
+simpl!(sum_loop);
+let fission = fork-fission[out.srad_0.fj0](sum_loop);
+simpl!(sum_loop);
+fork-tile[32, 0, false, true](fission.srad_0.fj_bottom);
+let out = fork-split(fission.srad_0.fj_bottom);
+clean-monoid-reduces(sum_loop);
+simpl!(sum_loop);
+let top = outline(fission.srad_0.fj_top);
+let bottom = outline(out.srad_0.fj0);
+gpu(top, bottom);
+ip-sroa(*);
+sroa(*);
+simpl!(*);
gcm(*);
-
diff --git a/juno_samples/rodinia/srad/src/lib.rs b/juno_samples/rodinia/srad/src/lib.rs
index d63660070ff0f61d47057ea00b14b3fb31db6e09..a647b94a5ffc8aad3bab91badc1bd58a305e7e75 100644
--- a/juno_samples/rodinia/srad/src/lib.rs
+++ b/juno_samples/rodinia/srad/src/lib.rs
@@ -114,7 +114,7 @@ pub fn srad_harness(args: SRADInputs) {
.max()
.unwrap_or(0);
assert!(
- max_diff <= 1,
+ max_diff <= 2,
"Verification failed: maximum pixel difference of {} exceeds threshold of 1",
max_diff
);
diff --git a/juno_samples/rodinia/srad/src/main.rs b/juno_samples/rodinia/srad/src/main.rs
index 87d1e7e8504584478f51ac2b9dc20dbc04716c81..20da11e73ef8eb90bcf8fde31ca3fa33c734c582 100644
--- a/juno_samples/rodinia/srad/src/main.rs
+++ b/juno_samples/rodinia/srad/src/main.rs
@@ -12,8 +12,8 @@ fn srad_test() {
srad_harness(SRADInputs {
niter: 100,
lambda: 0.5,
- nrows: 502,
- ncols: 458,
+ nrows: 512,
+ ncols: 512,
image: "data/image.pgm".to_string(),
output: None,
verify: true,
diff --git a/juno_samples/rodinia/srad/src/srad.jn b/juno_samples/rodinia/srad/src/srad.jn
index 3e016a99b574c1dcde982e7277a5cbcdc1743c19..6074bf8cb12ccc2ad29c1086d7620b3ef98bcf59 100644
--- a/juno_samples/rodinia/srad/src/srad.jn
+++ b/juno_samples/rodinia/srad/src/srad.jn
@@ -50,10 +50,10 @@ fn srad<nrows, ncols: usize>(
let varROI = (sum2 / nelems as f32) - meanROI * meanROI;
let q0sqr = varROI / (meanROI * meanROI);
- let dN : f32[ncols, nrows];
- let dS : f32[ncols, nrows];
- let dE : f32[ncols, nrows];
- let dW : f32[ncols, nrows];
+ @dirs let dN : f32[ncols, nrows];
+ @dirs let dS : f32[ncols, nrows];
+ @dirs let dE : f32[ncols, nrows];
+ @dirs let dW : f32[ncols, nrows];
let c : f32[ncols, nrows];
diff --git a/juno_scheduler/src/ir.rs b/juno_scheduler/src/ir.rs
index a0db884492120a43d0bb8fff89e689746ef1579e..6aa85fe53689cf015497e56850ef0c197ccbdae0 100644
--- a/juno_scheduler/src/ir.rs
+++ b/juno_scheduler/src/ir.rs
@@ -54,14 +54,15 @@ impl Pass {
pub fn is_valid_num_args(&self, num: usize) -> bool {
match self {
Pass::ArrayToProduct => num == 0 || num == 1,
+ Pass::ConstInline => num == 0 || num == 1,
Pass::ForkChunk => num == 4,
Pass::ForkExtend => num == 1,
Pass::ForkFissionBufferize => num == 2 || num == 1,
Pass::ForkInterchange => num == 2,
+ Pass::InterproceduralSROA => num == 0 || num == 1,
Pass::Print => num == 1,
Pass::Rename => num == 1,
Pass::SROA => num == 0 || num == 1,
- Pass::InterproceduralSROA => num == 0 || num == 1,
Pass::Xdot => num == 0 || num == 1,
_ => num == 0,
}
@@ -70,14 +71,15 @@ impl Pass {
pub fn valid_arg_nums(&self) -> &'static str {
match self {
Pass::ArrayToProduct => "0 or 1",
+ Pass::ConstInline => "0 or 1",
Pass::ForkChunk => "4",
Pass::ForkExtend => "1",
Pass::ForkFissionBufferize => "1 or 2",
Pass::ForkInterchange => "2",
+ Pass::InterproceduralSROA => "0 or 1",
Pass::Print => "1",
Pass::Rename => "1",
Pass::SROA => "0 or 1",
- Pass::InterproceduralSROA => "0 or 1",
Pass::Xdot => "0 or 1",
_ => "0",
}
diff --git a/juno_scheduler/src/pm.rs b/juno_scheduler/src/pm.rs
index e049f985e0db36ae78368b8d33c01d22744fdcc6..70d8e4278169ebdbe9985e00ede161acbe05c24d 100644
--- a/juno_scheduler/src/pm.rs
+++ b/juno_scheduler/src/pm.rs
@@ -1837,7 +1837,17 @@ fn run_pass(
pm.clear_analyses();
}
Pass::ConstInline => {
- assert!(args.is_empty());
+ let inline_collections = match args.get(0) {
+ Some(Value::Boolean { val }) => *val,
+ Some(_) => {
+ return Err(SchedulerError::PassError {
+ pass: "constInline".to_string(),
+ error: "expected boolean argument".to_string(),
+ });
+ }
+ None => true,
+ };
+
pm.make_callgraph();
let callgraph = pm.callgraph.take().unwrap();
@@ -1845,7 +1855,7 @@ fn run_pass(
.into_iter()
.map(|editor| editor.unwrap())
.collect();
- const_inline(&mut editors, &callgraph);
+ const_inline(&mut editors, &callgraph, inline_collections);
for func in editors {
changed |= func.modified();