numrs/llo/
mod.rs

1//! Low-level operations (LLO) — target-aware operations ready for codegen
2use serde::{Serialize, Deserialize};
3
4pub mod elementwise;
5pub mod reduction;
6pub mod matmul;
7pub mod linalg;
8pub mod shape;
9pub mod stats;
10pub mod random;
11pub mod model;
12
13pub use elementwise::*;
14pub use reduction::*;
15pub use matmul::*;
16pub use linalg::*;
17pub use shape::*;
18pub use stats::*;
19pub use random::*;
20pub use model::*;
21
22/// Top-level LLO program container
23#[derive(Debug, Clone, Serialize, Deserialize)]
24pub struct LLOProgram {
25    pub ops: Vec<LloOp>,
26}
27
28impl LLOProgram {
29    pub fn new() -> Self { Self { ops: vec![] } }
30    pub fn add_op(&mut self, op: LloOp) { self.ops.push(op); }
31}
32
33/// Enumeration of LLO operations. Each variant holds a backend-ready shape and input references
34#[derive(Debug, Clone, Serialize, Deserialize)]
35pub enum LloOp {
36    Elementwise { kind: ElementwiseKind, inputs: Vec<usize>, output_shape: Vec<usize>, strategy: ElementwiseStrategy },
37    Reduction { axis: Option<usize>, inputs: Vec<usize>, output_shape: Vec<usize> },
38    MatMul { a: usize, b: usize, output_shape: Vec<usize> },
39    /// Shape-related operations (reshape, transpose, concat, etc.)
40    Shape { kind: ShapeKind, inputs: Vec<usize>, output_shape: Vec<usize> },
41    /// Higher-level stats ops (softmax, norm, cross-entropy, ...)
42    Stats { kind: StatsKind, inputs: Vec<usize>, output_shape: Vec<usize> },
43    /// Random number generation ops (rand, randn, randint, seed)
44    Random { kind: RandomKind, inputs: Vec<usize>, output_shape: Vec<usize> },
45    /// Model-related operations (save, load, training)
46    Model { kind: ModelKind, inputs: Vec<usize>, output_shape: Vec<usize> },
47    /// Training operations (forward, backward, update)
48    Training { kind: TrainingKind, inputs: Vec<usize>, output_shape: Vec<usize> },
49}
50
51/// Convert Array to OnnxTensor
52pub fn array_to_onnx_tensor(name: &str, array: &crate::array::Array) -> anyhow::Result<OnnxTensor> {
53    // Flatten data to bytes (F32)
54    let mut data_bytes = Vec::with_capacity(array.data.len() * 4);
55    for &val in &array.data {
56        data_bytes.extend_from_slice(&val.to_le_bytes());
57    }
58    
59    Ok(OnnxTensor {
60        name: name.to_string(),
61        dtype: 1, // FLOAT
62        shape: array.shape.clone(),
63        data: data_bytes,
64    })
65}
numrs/llo/mod.rs

numrs/llo/
mod.rs
