tenflowers-core 0.1.1

Core tensor operations and execution engine for TenfloweRS
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

//! Common Pattern Optimizations for Einstein Summation
//!
//! This module contains optimizations for frequently used einsum patterns
//! including batch operations, transpose operations, and trace computations.

use super::utils::{batch_transpose, cache_friendly_trace, compute_outer_product};
use crate::{Result, Tensor};
use scirs2_core::numeric::{One, Zero};

#[cfg(feature = "gpu")]
use super::gpu::{
    gpu_einsum_batched_matmul, gpu_einsum_diagonal, gpu_einsum_matmul, gpu_einsum_outer_product,
    gpu_einsum_trace, gpu_einsum_transpose, gpu_einsum_vector_dot,
};

/// Try to optimize common einsum patterns
pub fn try_optimize_common_patterns<T>(
    equation: &str,
    operands: &[&Tensor<T>],
) -> Option<Result<Tensor<T>>>
where
    T: Clone
        + Default
        + Zero
        + One
        + std::ops::Add<Output = T>
        + std::ops::Mul<Output = T>
        + Send
        + Sync
        + 'static
        + bytemuck::Pod
        + bytemuck::Zeroable,
{
    // Check if this is a GPU operation that we can optimize
    #[cfg(feature = "gpu")]
    let is_gpu = operands
        .iter()
        .any(|op| matches!(&op.storage, crate::tensor::TensorStorage::Gpu(_)));
    #[cfg(not(feature = "gpu"))]
    let is_gpu = false;

    if is_gpu {
        return try_optimize_gpu_patterns(equation, operands);
    }

    match equation {
        // Batch matrix multiplication patterns
        "bij,bjk->bik" | "bik,bkj->bij" if operands.len() == 2 => {
            Some(crate::ops::matmul(operands[0], operands[1]))
        }
        // Trace operation with cache-friendly access
        "ii->" if operands.len() == 1 => Some(cache_friendly_trace(operands[0])),
        // Inner product (vectorized)
        "i,i->" if operands.len() == 2 => Some(crate::ops::dot(operands[0], operands[1])),
        // Outer product
        "i,j->ij" if operands.len() == 2 => Some(compute_outer_product(operands[0], operands[1])),
        // Batch transpose
        "bij->bji" if operands.len() == 1 => Some(batch_transpose(operands[0])),
        _ => None,
    }
}

/// Try to optimize GPU einsum patterns
#[cfg(feature = "gpu")]
pub fn try_optimize_gpu_patterns<T>(
    equation: &str,
    operands: &[&Tensor<T>],
) -> Option<Result<Tensor<T>>>
where
    T: Clone
        + Default
        + Zero
        + One
        + std::ops::Add<Output = T>
        + std::ops::Mul<Output = T>
        + Send
        + Sync
        + 'static
        + bytemuck::Pod
        + bytemuck::Zeroable,
{
    match equation {
        // Matrix multiplication: "ij,jk->ik"
        "ij,jk->ik" if operands.len() == 2 => Some(gpu_einsum_matmul(operands[0], operands[1])),
        // Batched matrix multiplication: "bij,bjk->bik"
        "bij,bjk->bik" if operands.len() == 2 => {
            Some(gpu_einsum_batched_matmul(operands[0], operands[1]))
        }
        // Transpose: "ij->ji"
        "ij->ji" if operands.len() == 1 => Some(gpu_einsum_transpose(operands[0])),
        // Diagonal extraction: "ii->i"
        "ii->i" if operands.len() == 1 => Some(gpu_einsum_diagonal(operands[0])),
        // Element-wise multiplication: "ij,ij->ij"
        eq if operands.len() == 2 && eq.starts_with("ij,ij->ij") => {
            Some(operands[0].mul(operands[1]))
        }
        // Sum of element-wise multiplication: "ij,ij->"
        eq if operands.len() == 2 && eq.starts_with("ij,ij->") && eq.ends_with("->") => {
            let elementwise = operands[0].mul(operands[1]);
            Some(elementwise.and_then(|t| crate::ops::sum(&t, None, false)))
        }
        // Outer product: "i,j->ij"
        "i,j->ij" if operands.len() == 2 => {
            Some(gpu_einsum_outer_product(operands[0], operands[1]))
        }
        // Vector dot product: "i,i->"
        "i,i->" if operands.len() == 2 => Some(gpu_einsum_vector_dot(operands[0], operands[1])),
        // Trace: "ii->"
        "ii->" if operands.len() == 1 => Some(gpu_einsum_trace(operands[0])),
        _ => None,
    }
}

#[cfg(not(feature = "gpu"))]
pub fn try_optimize_gpu_patterns<T>(
    _equation: &str,
    _operands: &[&Tensor<T>],
) -> Option<Result<Tensor<T>>>
where
    T: Clone
        + Default
        + Zero
        + One
        + std::ops::Add<Output = T>
        + std::ops::Mul<Output = T>
        + Send
        + Sync
        + 'static
        + bytemuck::Pod
        + bytemuck::Zeroable,
{
    None
}