1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
// Copyright (C) 2025 zk4x
// SPDX-License-Identifier: LGPL-3.0-only WITH Classpath-exception-2.0
//! Loop-invariant code motion and reassociation.
//!
//! This module provides optimizations for loop-invariant code motion
//! and reassociation of commutative operations.
//!
//! Optimizations include:
//!
//! - `opt_reassociate_commutative`: Reassociate commutative operations
//! - `swap_commutative`: Swap commutative operands
//! - `loop_invariant_code_motion`: Hoist loop-invariant computations
//!
//! These optimizations reduce redundant computations and improve performance.
use super::autotune::Optimization;
use crate::kernel::{Kernel, Op, OpId};
use crate::{Map, Set};
/// Reassociate commutative operations (addition, multiplication)
/// to group them and reduce instruction count.
#[derive(Debug)]
pub struct ReassociateCommutative;
impl Optimization for ReassociateCommutative {
fn nconfigs(&self) -> u64 {
1
}
fn apply(&self, kernel: &mut Kernel, _config: u64) {
kernel.reassociate_commutative();
}
}
impl Kernel {
/// Make the `ReassociateCommutative` optimization.
pub fn opt_reassociate_commutative(&self) -> Box<dyn Optimization> {
Box::new(ReassociateCommutative)
}
/// Swap commutative operands for better instruction scheduling.
///
/// This method swaps operands of commutative operations (addition,
/// multiplication) to improve instruction scheduling and pipeline
/// utilization.
pub(crate) fn swap_commutative(&mut self) {
// Tracks whether a value depends on a loop index
let mut loop_dep: Map<OpId, usize> = Map::default();
let mut loop_depth = 0;
let mut op_id = self.head;
while !op_id.is_null() {
let depth = match self.ops[op_id].op {
Op::Move { .. } | Op::Reduce { .. } | Op::ReduceTile { .. } => {
unreachable!()
}
Op::MatmulTile { x, y, acc } => loop_dep[&x].max(loop_dep[&y]).max(loop_dep[&acc]),
Op::TransposeTile { x } => loop_dep[&x],
Op::BroadcastTile { x, .. } => loop_dep[&x],
Op::Asm { .. } | Op::Index { .. } | Op::Wmma { .. } | Op::Stack { .. } => loop_depth,
Op::If { .. } | Op::Loop { .. } => {
loop_depth += 1;
loop_depth
}
Op::EndIf | Op::EndLoop => {
loop_depth -= 1;
loop_depth
}
Op::Unary { x, .. } | Op::Cast { x, .. } | Op::Bitcast { x, .. } => loop_dep[&x],
Op::Binary { x, y, bop } => {
if bop.is_commutative()
&& !self.ops[x].op.is_const()
&& (loop_dep[&x] > loop_dep[&y] || self.ops[y].op.is_const() || self.ops[x].op.is_load())
{
//println!("Swapping {x}, {y}, loop dep {} > {}: {:?}, {:?}", loop_dep[&x], loop_dep[&y], self.ops[x].op, self.ops[y].op);
if let Op::Binary { x, y, .. } = &mut self.ops[op_id].op {
std::mem::swap(x, y);
}
}
loop_dep[&x].max(loop_dep[&y])
}
Op::Mad { x, y, z } => loop_dep[&x].max(loop_dep[&y]).max(loop_dep[&z]),
Op::Param { .. }
| Op::Barrier
| Op::Range { .. }
| Op::Load { .. }
| Op::Store { .. }
| Op::Const(_)
| Op::Storage { .. } => loop_depth,
};
loop_dep.insert(op_id, depth);
op_id = self.next_op(op_id);
}
self.verify();
}
/// Reassociate commutative operations to group them.
///
/// This method reassociates commutative operations (addition,
/// multiplication) to group them and reduce instruction count.
///
/// For example, `a + b + c` can be transformed to `(a + b) + c`
/// to enable better instruction scheduling.
pub(crate) fn reassociate_commutative(&mut self) {
#[cfg(feature = "time")]
let _timer = crate::Timer::new("reassociate_commutative");
let mut loop_dep: Map<OpId, usize> = Map::default();
let mut loop_depth = 0;
let mut op_id = self.head;
while !op_id.is_null() {
let depth = match self.at(op_id) {
Op::Move { .. } | Op::Reduce { .. } => {
unreachable!()
}
Op::ReduceTile { x, scaler, acc, .. } => loop_dep[x].max(loop_dep[scaler]).max(loop_dep[acc]),
Op::MatmulTile { x, y, acc } => loop_dep[x].max(loop_dep[y]).max(loop_dep[acc]),
Op::TransposeTile { x } => loop_dep[x],
Op::BroadcastTile { x, .. } => loop_dep[x],
Op::Asm { ops, .. } => {
let mut max = 0;
for op in ops.iter() {
max = max.max(loop_dep[op]);
}
max
}
Op::Stack { ops } => {
let mut max = 0;
for op in ops.iter() {
max = max.max(loop_dep[op]);
}
max
}
Op::Index { vec, .. } => loop_dep[vec],
Op::Mad { x, y, z } => loop_dep[x].max(loop_dep[y]).max(loop_dep[z]),
Op::Loop { .. } | Op::If { .. } => {
loop_depth += 1;
loop_depth
}
Op::EndLoop | Op::EndIf => {
loop_depth -= 1;
loop_depth
}
Op::Unary { x, .. } | Op::Cast { x, .. } | Op::Bitcast { x, .. } => loop_dep[x],
Op::Binary { x, y, .. } => loop_dep[x].max(loop_dep[y]),
Op::Range { .. }
| Op::Barrier
| Op::Load { .. }
| Op::Store { .. }
| Op::Const(_)
| Op::Param { .. }
| Op::Storage { .. }
| Op::Wmma { .. } => loop_depth,
};
loop_dep.insert(op_id, depth);
op_id = self.next_op(op_id);
}
let mut op_id = self.head;
'a: while !op_id.is_null() {
let next = self.next_op(op_id);
if let &Op::Binary { bop, .. } = self.at(op_id) {
if !bop.is_commutative() || !bop.is_associative() {
op_id = next;
continue 'a;
}
// Get all the leafs
let mut params = vec![op_id];
let mut chain = Vec::new();
while let Some(param) = params.pop() {
if let &Op::Binary { x, y, bop: t_bop } = self.at(param)
&& t_bop == bop
{
params.push(x);
params.push(y);
continue;
}
chain.push(param);
// We have to be somewhat reasonabe about those chains
if chain.len() > 20 {
op_id = next;
continue 'a;
}
}
if chain.len() < 2 {
op_id = next;
continue 'a;
}
chain.sort_by_key(|id| loop_dep[id]);
// Rebuild chain
let mut prev_acc = chain[0];
let mut j = 1;
while j < chain.len() - 1 {
let op = Op::Binary { x: chain[j], y: prev_acc, bop };
let new_acc = self.insert_before(op_id, op);
prev_acc = new_acc;
j += 1;
}
self.ops[op_id].op = Op::Binary { x: chain[j], y: prev_acc, bop };
}
op_id = next;
}
self.verify();
}
/// Hoist loop-invariant computations outside loops.
///
/// This method identifies computations that are invariant within
/// loops and moves them outside the loop to reduce redundant
/// computations and improve performance.
pub fn loop_invariant_code_motion(&mut self) {
#[cfg(feature = "time")]
let _timer = crate::Timer::new("loop_invariant_code_motion");
let mut endloop_is = Vec::new();
let mut loop_id = self.tail;
while !loop_id.is_null() {
if *self.at(loop_id) == Op::EndLoop {
endloop_is.push(loop_id);
}
if let Op::Loop { .. } = self.at(loop_id) {
let mut op_ids_in_loop = Set::default();
op_ids_in_loop.insert(loop_id); // Loop op is the primary op that breaks LICM
let mut op_id = loop_id;
let endloop_id = endloop_is.pop().unwrap();
while op_id != endloop_id {
let op = self.at(op_id);
let next_op_id = self.next_op(op_id);
if !matches!(
op,
Op::Store { .. }
| Op::Load { .. }
| Op::Loop { .. }
| Op::EndLoop
| Op::Storage { .. }
| Op::Barrier
| Op::If { .. }
| Op::EndIf
) && op.parameters().all(|op_id| !op_ids_in_loop.contains(&op_id))
{
self.move_op_before(op_id, loop_id);
} else {
op_ids_in_loop.insert(op_id);
}
op_id = next_op_id;
}
}
loop_id = self.prev_op(loop_id);
}
self.verify();
}
}