zyx 0.17.0

Zyx machine learning library
Documentation 
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
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273

// Copyright (C) 2025 zk4x
// SPDX-License-Identifier: LGPL-3.0-only WITH Classpath-exception-2.0

//! Split loops optimization.
//!
//! This module provides loop splitting optimizations for kernels,
//! which split large loops into smaller iterations for better
//! instruction scheduling and vectorization.
//!
//! Loop splitting is useful for:
//!
//! - Reducing instruction dependencies
//! - Enabling better instruction-level parallelism
//! - Improving vectorization opportunities
//! - Splitting global indices into local factors

use super::autotune::Optimization;
use crate::{
    dtype::Constant,
    kernel::{BOp, Kernel, Op, OpId, RangeKind},
    shape::Dim,
};

/// Split a global index into local factors for parallelization.
#[derive(Debug)]
pub struct SplitGlobalToLocal {
    /// Pairs of (operation_id, split_factor) for each split.
    pub factors: Vec<(OpId, u32)>,
}

impl Optimization for SplitGlobalToLocal {
    fn nconfigs(&self) -> u64 {
        self.factors.len() as u64
    }

    fn apply(&self, kernel: &mut Kernel, config: u64) {
        #[cfg(feature = "time")]
        let _timer = crate::Timer::new("SplitGlobalToLocal");
        let (op_id, factor) = self.factors[config as usize];
        let Op::Range { axis, kind: RangeKind::Group(len) } = kernel.ops[op_id].op else {
            unreachable!()
        };
        // valid factors are checked by opt init
        let len = kernel.resolve_const(len).and_then(crate::dtype::Constant::as_dim).unwrap();
        let group_len = kernel.insert_const_idx_before(op_id, len / Dim::from(factor));
        kernel.split_dim(
            op_id,
            vec![
                Op::Range { axis, kind: RangeKind::Group(group_len) },
                Op::Range { axis, kind: RangeKind::Local(factor) },
            ],
        );
    }
}

/// Split a loop into smaller iterations.
#[derive(Debug)]
pub struct SplitLoop {
    /// Pairs of (loop_id, split_factor) for each split.
    pub factors: Vec<(OpId, u64)>,
}

impl Optimization for SplitLoop {
    fn nconfigs(&self) -> u64 {
        self.factors.len() as u64
    }

    fn apply(&self, kernel: &mut Kernel, config: u64) {
        let (op_id, factor) = self.factors[config as usize];
        let Op::Loop { len: len_id } = kernel.ops[op_id].op else {
            unreachable!()
        };
        let Some(len) = kernel.resolve_const(len_id).and_then(crate::dtype::Constant::as_dim) else {
            return;
        };
        let len1 = kernel.insert_const_idx_before(op_id, len / factor as Dim);
        let len2 = kernel.insert_const_idx_before(op_id, factor);
        kernel.split_dim(op_id, vec![Op::Loop { len: len1 }, Op::Loop { len: len2 }]);
    }
}

impl Kernel {
    /// Make the `SplitGlobalToLocal` optimization: scan the kernel for
    /// global indices that can be split into local factors.
    ///
    /// Config ids are ordered by factor, hardware-aligned factors first
    /// (e.g. 64/32 for warp-sized groups).
    pub fn opt_split_global_to_local(&self) -> Box<dyn Optimization> {
        #[cfg(feature = "time")]
        let _timer = crate::Timer::new("opt_split_global_to_local");

        let dev_info = self.dev_info();

        if self.ops.values().any(|node| matches!(node.op, Op::EndIf)) {
            return Box::new(SplitGlobalToLocal { factors: Vec::new() });
        }
        let mut local_axis_sizes: crate::Map<u32, u32> = crate::Map::default();
        for op in self.ops.values() {
            if let Op::Range { axis, kind: RangeKind::Local(len) } = op.op {
                if let Some(&existing) = local_axis_sizes.get(&axis) {
                    debug_assert_eq!(existing, len);
                } else {
                    local_axis_sizes.insert(axis, len);
                }
            }
        }
        let used_threads: u32 = local_axis_sizes.values().product();
        let remaining_threads = if local_axis_sizes.is_empty() {
            dev_info.max_local_threads
        } else {
            dev_info.max_local_threads / used_threads
        };
        //println!("local_axis_sizes={local_axis_sizes:?}, remaining_threads={remaining_threads}, used_threads={used_threads}");
        let mut op_id = self.head;
        let mut factors = Vec::new();
        while !op_id.is_null() {
            if let Op::Range { axis, kind: RangeKind::Group(len) } = self.ops[op_id].op {
                let mut l_factors: Vec<u32> = vec![64, 32, 16, 8, 4, 2];
                if !local_axis_sizes.contains_key(&axis) {
                    let max_per_axis = dev_info.max_local_work_dims[axis as usize];
                    let Some(len) = self.resolve_const(len).and_then(crate::dtype::Constant::as_dim) else {
                        op_id = self.next_op(op_id);
                        continue;
                    };
                    l_factors.retain(|&f| len % f as Dim == 0 && f <= remaining_threads && f <= max_per_axis);
                    for &f in &l_factors {
                        factors.push((op_id, f));
                    }
                }
            }
            op_id = self.next_op(op_id);
        }
        Box::new(SplitGlobalToLocal { factors })
    }

    /// Make the `SplitLoop` optimization: scan the kernel for large loops
    /// that can be split into smaller iterations.
    pub fn opt_split_loop(&self) -> Box<dyn Optimization> {
        #[cfg(feature = "time")]
        let _timer = crate::Timer::new("opt_split_loop");
        let candidates = vec![8, 16, 4, 2];
        let mut factors = Vec::new();
        let mut op_id = self.head;
        while !op_id.is_null() {
            if let Op::Loop { len: len_id } = self.ops[op_id].op {
                let Some(len) = self.resolve_const(len_id).and_then(crate::dtype::Constant::as_dim) else {
                    op_id = self.next_op(op_id);
                    continue;
                };
                if len >= 16 {
                    for &factor in &candidates {
                        if len % factor as Dim == 0 {
                            factors.push((op_id, factor));
                        }
                    }
                }
            }
            op_id = self.next_op(op_id);
        }
        Box::new(SplitLoop { factors })
    }

    /// Splits dim (index or loop) into multiple indices or loops
    /// Returns the `OpId`s of the created split operations in the order they were provided
    pub(crate) fn split_dim(&mut self, dim_id: OpId, mut splits: Vec<Op>) -> Vec<OpId> {
        #[cfg(feature = "time")]
        let _timer = crate::Timer::new("split_dim");
        let is_loop = matches!(self.ops[dim_id].op, Op::Loop { .. });

        #[cfg(debug_assertions)]
        {
            let mut dim = 1;
            let mut ok = true;
            for op in splits.iter() {
                use crate::shape::Dim;

                match *op {
                    Op::Loop { len, .. } => match self.resolve_const(len).and_then(crate::dtype::Constant::as_dim) {
                        Some(l) => dim *= l,
                        None => {
                            ok = false;
                            break;
                        }
                    },
                    Op::Range { kind, .. } => match kind {
                        RangeKind::Group(len) => match self.resolve_const(len).and_then(crate::dtype::Constant::as_dim) {
                            Some(l) => dim *= l,
                            None => {
                                ok = false;
                                break;
                            }
                        },
                        RangeKind::Local(len) => dim *= len as Dim,
                        // A warp is a view over a local range — it is never split.
                        RangeKind::Warp(_) => {
                            ok = false;
                            break;
                        }
                    },
                    _ => unreachable!("split can be only index or loop"),
                }
            }
            if ok {
                match self.ops[dim_id].op {
                    Op::Range { kind, .. } => {
                        use crate::shape::Dim;

                        match kind {
                            RangeKind::Group(len) => {
                                if let Some(l) = self.resolve_const(len).and_then(crate::dtype::Constant::as_dim) {
                                    debug_assert_eq!(l, dim);
                                }
                            }
                            RangeKind::Local(l) => debug_assert_eq!(l as Dim, dim),
                            RangeKind::Warp(_) => unreachable!("warp dims are never split"),
                        }
                    }
                    Op::Loop { len, .. } => {
                        if let Some(l) = self.resolve_const(len).and_then(crate::dtype::Constant::as_dim) {
                            debug_assert_eq!(l, dim);
                        }
                    }
                    _ => {}
                }
            }
        }

        let last_dim_op = self.get_last_dim_op(dim_id);
        let n_loops = splits.iter().filter(|op| matches!(op, Op::Loop { .. })).count();
        for (i, op) in splits.iter().enumerate() {
            if matches!(op, Op::Loop { .. }) {
                if is_loop && i == n_loops - 1 {
                } else {
                    self.insert_after(last_dim_op, Op::EndLoop);
                }
            }
        }

        let mut strides = Vec::new();
        let mut st = 1;
        for op in splits.iter().rev() {
            strides.push(st);
            match op {
                Op::Loop { len, .. } => st *= self.resolve_const(*len).and_then(crate::dtype::Constant::as_dim).unwrap(),
                Op::Range { kind, .. } => match kind {
                    RangeKind::Group(len) => st *= self.resolve_const(*len).and_then(crate::dtype::Constant::as_dim).unwrap(),
                    RangeKind::Local(len) => st *= i64::from(*len),
                    RangeKind::Warp(_) => unreachable!("warp dims are never split"),
                },
                _ => unreachable!("split can be only index or loop"),
            }
        }
        strides.reverse();
        strides.pop();
        let last_op = splits.pop().unwrap();

        let mut split_ids: Vec<OpId> = Vec::new();
        let mut acc = self.insert_before(dim_id, Op::Const(Constant::idx(0)));
        for (&st, op) in strides.iter().zip(splits) {
            let x = self.insert_before(dim_id, Op::Const(Constant::idx(st)));
            let y = self.insert_before(dim_id, op);
            acc = self.insert_before(dim_id, Op::Mad { x, y, z: acc });
            split_ids.push(y);
        }

        let y = self.insert_before(dim_id, last_op);
        split_ids.push(y);
        self.ops[dim_id].op = Op::Binary { x: acc, y, bop: BOp::Add };

        self.verify();
        split_ids
    }
}