tenflowers-dataset 0.1.1

Data pipeline and dataset utilities 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
134
135
136
137
138
139
140
141
142
143
144
145
146

//! SIMD-accelerated element-wise operations
//!
//! This module provides vectorized implementations of element-wise operations
//! using SIMD instructions for enhanced performance.

#![allow(unsafe_code)]

use crate::Transform;
use std::marker::PhantomData;
use tenflowers_core::{Result, Tensor, TensorError};

#[cfg(target_arch = "x86_64")]
use std::arch::x86_64::*;

/// Types of SIMD operations supported
#[derive(Debug, Clone, Copy)]
pub enum SimdOperation {
    Add,
    Multiply,
    Subtract,
    Divide,
}

/// SIMD-accelerated element-wise operations
pub struct SimdElementWise<T> {
    operation: SimdOperation,
    value: T,
    _phantom: PhantomData<T>,
}

impl<T> SimdElementWise<T>
where
    T: Clone + Default + scirs2_core::numeric::Float + Send + Sync + 'static,
{
    /// Create a new SIMD element-wise transform
    pub fn new(operation: SimdOperation, value: T) -> Self {
        Self {
            operation,
            value,
            _phantom: PhantomData,
        }
    }

    /// SIMD-accelerated element-wise operation for f32 data
    #[cfg(target_arch = "x86_64")]
    unsafe fn apply_f32_simd(&self, data: &mut [f32], value: f32) {
        if data.len() < 8 {
            self.apply_scalar_f32(data, value);
            return;
        }

        let value_vec = _mm256_set1_ps(value);
        let chunks = data.len() / 8;
        let remainder = data.len() % 8;

        for i in 0..chunks {
            let offset = i * 8;
            let values = _mm256_loadu_ps(data.as_ptr().add(offset));

            let result = match self.operation {
                SimdOperation::Add => _mm256_add_ps(values, value_vec),
                SimdOperation::Multiply => _mm256_mul_ps(values, value_vec),
                SimdOperation::Subtract => _mm256_sub_ps(values, value_vec),
                SimdOperation::Divide => _mm256_div_ps(values, value_vec),
            };

            _mm256_storeu_ps(data.as_mut_ptr().add(offset), result);
        }

        if remainder > 0 {
            let start = chunks * 8;
            self.apply_scalar_f32(&mut data[start..], value);
        }
    }

    /// Scalar fallback for element-wise operations
    fn apply_scalar(&self, data: &mut [T], value: T)
    where
        T: scirs2_core::numeric::Float,
    {
        for element in data.iter_mut() {
            *element = match self.operation {
                SimdOperation::Add => *element + value,
                SimdOperation::Multiply => *element * value,
                SimdOperation::Subtract => *element - value,
                SimdOperation::Divide => *element / value,
            };
        }
    }

    /// Scalar fallback for f32 element-wise operations
    #[allow(dead_code)]
    fn apply_scalar_f32(&self, data: &mut [f32], value: f32) {
        for element in data.iter_mut() {
            *element = match self.operation {
                SimdOperation::Add => *element + value,
                SimdOperation::Multiply => *element * value,
                SimdOperation::Subtract => *element - value,
                SimdOperation::Divide => *element / value,
            };
        }
    }
}

impl<T> Transform<T> for SimdElementWise<T>
where
    T: Clone + Default + scirs2_core::numeric::Float + Send + Sync + 'static,
{
    fn apply(&self, sample: (Tensor<T>, Tensor<T>)) -> Result<(Tensor<T>, Tensor<T>)> {
        let (features, labels) = sample;

        if let Some(data) = features.as_slice() {
            let mut mutable_data = data.to_vec();

            #[cfg(target_arch = "x86_64")]
            {
                if is_x86_feature_detected!("avx2") && std::mem::size_of::<T>() == 4 {
                    let value_f32 = unsafe { std::mem::transmute_copy::<T, f32>(&self.value) };
                    let data_f32 = unsafe {
                        std::slice::from_raw_parts_mut(
                            mutable_data.as_mut_ptr() as *mut f32,
                            mutable_data.len(),
                        )
                    };

                    unsafe {
                        self.apply_f32_simd(data_f32, value_f32);
                    }
                } else {
                    self.apply_scalar(&mut mutable_data, self.value);
                }
            }
            #[cfg(not(target_arch = "x86_64"))]
            {
                self.apply_scalar(&mut mutable_data, self.value);
            }

            let new_features = Tensor::from_vec(mutable_data, features.shape().dims())?;
            Ok((new_features, labels))
        } else {
            Err(TensorError::invalid_argument(
                "Cannot access tensor data for element-wise operation".to_string(),
            ))
        }
    }
}