hpt 0.1.3

High Performance Tensor (HPT) - A fast, efficient, and user-friendly tensor computation library for Rust
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

use super::conv2d;
use super::microkernel_trait::Conv2dMicroKernel;
use crate::backend::Cpu;
use crate::backends::cpu::kernels::conv2d::conv2d_new_mp;
use crate::backends::cpu::kernels::matmul::microkernel_trait::MatmulMicroKernel;
use crate::backends::cpu::kernels::normalization::batch_norm::batch_norm;
use crate::tensor_base::_Tensor;
use hpt_allocator::traits::{Allocator, AllocatorOutputRetrive};
use hpt_common::error::base::TensorError;
use hpt_traits::tensor::CommonBounds;
use hpt_types::dtype::TypeCommon;
use hpt_types::into_scalar::Cast;
use hpt_types::traits::VecTrait;
use hpt_types::type_promote::FloatOutBinary;
use hpt_types::type_promote::FloatOutUnary;
use hpt_types::type_promote::NormalOutPromote;

#[track_caller]
pub(crate) fn batchnorm_conv2d<T, const DEVICE: usize, A>(
    input: &_Tensor<T, Cpu, DEVICE, A>,
    kernels: &_Tensor<T, Cpu, DEVICE, A>,
    mean: &_Tensor<T, Cpu, DEVICE, A>,
    var: &_Tensor<T, Cpu, DEVICE, A>,
    gamma: &_Tensor<T, Cpu, DEVICE, A>,
    beta: &_Tensor<T, Cpu, DEVICE, A>,
    bias: Option<&_Tensor<T, Cpu, DEVICE, A>>,
    eps: T,
    steps: [i64; 2],
    padding: [(i64, i64); 2],
    dilation: [i64; 2],
    post_scalar: Option<fn(T) -> T>,
    post_vec: Option<fn(<T>::Vec) -> <T>::Vec>,
) -> std::result::Result<_Tensor<T, Cpu, DEVICE, A>, TensorError>
where
    T: CommonBounds + Conv2dMicroKernel + MatmulMicroKernel,
    T::Vec: FloatOutBinary<Output = T::Vec> + FloatOutUnary<Output = T::Vec>,
    T: FloatOutBinary<Output = T> + FloatOutUnary<Output = T>,
    A: Allocator + Send + Sync,
    A::Output: AllocatorOutputRetrive,
{
    let conv_res = if T::STR == "bf16" {
        type F32Vec = <<half::bf16 as NormalOutPromote>::Intermediate as TypeCommon>::Vec;
        type BF16Vec = <half::bf16 as TypeCommon>::Vec;
        let inp = input
            .static_cast::<half::bf16>()
            .expect("static_cast bf16 failed");
        let ker = kernels
            .static_cast::<half::bf16>()
            .expect("static_cast bf16 failed");
        let bias = bias.map(|x| {
            x.static_cast::<half::bf16>()
                .expect("static_cast bf16 failed")
        });
        let res = conv2d_new_mp::conv2d::<half::bf16, DEVICE, A>(
            &inp,
            &ker,
            bias.as_ref(),
            steps,
            padding,
            dilation,
            |x| {
                let vec0 = unsafe { x.read_unaligned() };
                let vec1 = unsafe { x.add(1).read_unaligned() };
                BF16Vec::from_2_f32vec([vec0, vec1])
            },
            |x| unsafe {
                let mut bf16_vec = BF16Vec::splat(half::bf16::from_f32_const(0.0));
                for j in 0..F32Vec::SIZE {
                    bf16_vec[j] = *x.add(j);
                }
                let val_f32 = bf16_vec.high_to_f32vec();
                val_f32
            },
            |x| x.cast(),
            |x| x.cast(),
            None,
            None,
        )?;
        Ok(res.static_cast::<T>()?)
    } else if T::STR == "f16" && !cfg!(target_feature = "neon") {
        type F32Vec = <<half::f16 as NormalOutPromote>::Intermediate as TypeCommon>::Vec;
        type F16Vec = <half::f16 as TypeCommon>::Vec;
        let inp = input
            .static_cast::<half::f16>()
            .expect("static_cast f16 failed");
        let ker = kernels
            .static_cast::<half::f16>()
            .expect("static_cast f16 failed");
        let bias = bias.map(|x| {
            x.static_cast::<half::f16>()
                .expect("static_cast f16 failed")
        });
        let res = conv2d_new_mp::conv2d::<half::f16, DEVICE, A>(
            &inp,
            &ker,
            bias.as_ref(),
            steps,
            padding,
            dilation,
            |x| {
                let vec0 = unsafe { x.read_unaligned() };
                let vec1 = unsafe { x.add(1).read_unaligned() };
                F16Vec::from_2_f32vec([vec0, vec1])
            },
            |x| unsafe {
                let mut f16_vec = F16Vec::splat(half::f16::from_f32_const(0.0));
                for j in 0..F32Vec::SIZE {
                    f16_vec[j] = *x.add(j);
                }
                let val_f32 = f16_vec.high_to_f32vec();
                val_f32
            },
            |x| x.cast(),
            |x| x.cast(),
            None,
            None,
        )?;
        Ok(res.static_cast::<T>()?)
    } else {
        conv2d::conv2d(input, kernels, bias, steps, padding, dilation, None, None)
    }?;

    batch_norm(
        &conv_res,
        mean,
        var,
        gamma,
        beta,
        eps,
        post_scalar,
        post_vec,
        Some(conv_res.clone()),
    )?;

    Ok(conv_res)
}