rlx-fusion 0.2.6

Fusion passes and fused-op decomposition for RLX MIR
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

// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
//
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, version 3.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program. If not, see <https://www.gnu.org/licenses/>.

//! Per-backend caps for fused IR (elementwise region chains, etc.).

use std::cell::Cell;

/// Hardware / encoder limits for fusion passes.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct FusionLimits {
    /// Max steps in one `Op::ElementwiseRegion` chain (Metal/wgpu/CUDA: 32).
    pub max_elementwise_steps: u32,
    /// Max distinct external inputs in one region (Metal/wgpu/CUDA: 16).
    pub max_elementwise_inputs: u32,
}

impl FusionLimits {
    /// Caps shared by native elementwise-region kernels today.
    pub const GPU_NATIVE: Self = Self {
        max_elementwise_steps: 32,
        max_elementwise_inputs: 16,
    };

    /// No practical cap — used when regions are unfused to primitives (CPU).
    pub const UNBOUNDED: Self = Self {
        max_elementwise_steps: u32::MAX,
        max_elementwise_inputs: u32::MAX,
    };
}

impl Default for FusionLimits {
    fn default() -> Self {
        Self::GPU_NATIVE
    }
}

thread_local! {
    static ACTIVE_LIMITS: Cell<FusionLimits> = Cell::new(FusionLimits::default());
}

/// Limits used by [`crate::fusion::MarkElementwiseRegions`] during this compile.
pub fn active_fusion_limits() -> FusionLimits {
    ACTIVE_LIMITS.with(|c| c.get())
}

/// Run `f` with `limits` installed for mark/clip passes (single-threaded compile).
pub fn with_fusion_limits<T>(limits: FusionLimits, f: impl FnOnce() -> T) -> T {
    ACTIVE_LIMITS.with(|c| {
        let prev = c.get();
        c.set(limits);
        let out = f();
        c.set(prev);
        out
    })
}