rlx-sparse 0.2.11

Sparse linear algebra for RLX — CSR LU, mat-vec, Conjugate Gradient. Downstream package; registers against rlx's custom-op scaffold.
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

// 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/>.

//! `pcg` op registration — split from `lib.rs` (see `register()`).

#![cfg_attr(not(feature = "cpu"), allow(dead_code))]
#![allow(unused_imports)]

use std::sync::Arc;

use rlx_ir::{DType, Graph, Node, NodeId, Op, OpExtension, Shape, VjpContext, register_op};

#[cfg(feature = "cpu")]
use rlx_cpu::op_registry::{CpuKernel, CpuTensorMut, CpuTensorRef, register_cpu_kernel};

// ── Op names (stable strings; downstream callers use these to look
//    up the registered op or build `Op::Custom` directly) ─────────

use super::*;

pub(crate) struct SparsePcgExt;

impl OpExtension for SparsePcgExt {
    fn name(&self) -> &str {
        SPARSE_PCG_SOLVE
    }
    fn num_inputs(&self) -> usize {
        4
    } // values, col_idx, row_ptr, b

    fn infer_shape(&self, inputs: &[&Shape], _attrs: &[u8]) -> Shape {
        inputs[3].clone()
    }

    fn vjp(&self, node: &Node, ctx: &mut VjpContext) -> Vec<(usize, NodeId)> {
        // Closed-form adjoint identical to CG (same SPD assumption).
        // Recurse into pcg_solve with the same attrs.
        let vals_b = ctx.fwd_map[&node.inputs[0]];
        let cidx_b = ctx.fwd_map[&node.inputs[1]];
        let rptr_b = ctx.fwd_map[&node.inputs[2]];
        let attrs = match &node.op {
            rlx_ir::Op::Custom { attrs, .. } => attrs.clone(),
            _ => Vec::new(),
        };
        let g_b = ctx.bwd.custom_op(
            SPARSE_PCG_SOLVE,
            attrs,
            vec![vals_b, cidx_b, rptr_b, ctx.upstream],
        );
        let y_fwd = ctx.fwd_map[&node.id];
        let raw_grad = ctx.bwd.custom_op(
            SPARSE_VALUES_GRAD,
            Vec::new(),
            vec![cidx_b, rptr_b, g_b, y_fwd],
        );
        let raw_shape = ctx.bwd.node(raw_grad).shape.clone();
        let g_vals = ctx
            .bwd
            .activation(rlx_ir::op::Activation::Neg, raw_grad, raw_shape);
        vec![(0, g_vals), (3, g_b)]
    }
}

#[cfg(feature = "cpu")]
pub(crate) struct SparsePcgCpu;

#[cfg(feature = "cpu")]
impl CpuKernel for SparsePcgCpu {
    fn name(&self) -> &str {
        SPARSE_PCG_SOLVE
    }
    fn execute(
        &self,
        inputs: &[CpuTensorRef<'_>],
        output: CpuTensorMut<'_>,
        attrs: &[u8],
    ) -> Result<(), String> {
        let values = inputs[0].expect_f64("pcg values")?;
        let col_idx = inputs[1].expect_i32("pcg col_idx")?;
        let row_ptr = inputs[2].expect_i32("pcg row_ptr")?;
        let b = inputs[3].expect_f64("pcg b")?;
        let out = output.expect_f64_mut("pcg x")?;
        let (max_iter, tol) = decode_cg_attrs(attrs)?;
        algos::pcg_solve(values, col_idx, row_ptr, b, out, max_iter, tol)
    }
}

// ── BiCGSTAB (non-symmetric) ──────────────────────────────────────