meuron 0.4.0

Meuron is a modular neural network library written in rust for training simple neural networks.
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

use super::context::GpuContext;
use super::params::TENSOR_USAGE;
use bytemuck::cast_slice;
use ndarray::Dimension;
use std::sync::Arc;
use wgpu;
use wgpu::util::DeviceExt;

pub struct GpuTensor<D: Dimension> {
    pub buffer: Arc<wgpu::Buffer>,
    pub shape: D,
    pub size: usize,
    pub(super) ctx: Arc<GpuContext>,
}

impl<D: Dimension> GpuTensor<D> {
    pub fn upload(arr: ndarray::Array<f32, D>, ctx: Arc<GpuContext>) -> Self {
        let contig = arr.as_standard_layout();
        let buffer = ctx
            .device
            .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: None,
                contents: cast_slice(contig.as_slice().unwrap()),
                usage: TENSOR_USAGE,
            });
        let shape = contig.raw_dim();
        let size = contig.len();
        GpuTensor {
            buffer: Arc::new(buffer),
            shape,
            size,
            ctx,
        }
    }

    pub fn download(&self) -> ndarray::Array<f32, D> {
        self.ctx.flush();

        let byte_size = (self.size * 4) as u64;
        let staging = self.ctx.device.create_buffer(&wgpu::BufferDescriptor {
            label: None,
            size: byte_size,
            usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::MAP_READ,
            mapped_at_creation: false,
        });
        let mut enc = self.ctx.device.create_command_encoder(&Default::default());
        enc.copy_buffer_to_buffer(&self.buffer, 0, &staging, 0, byte_size);
        let si = self.ctx.queue.submit([enc.finish()]);

        let slice = staging.slice(..);
        slice.map_async(wgpu::MapMode::Read, |_| {});
        let _ = self.ctx.device.poll(wgpu::PollType::Wait {
            submission_index: Some(si),
            timeout: None,
        });

        let data = cast_slice::<u8, f32>(&slice.get_mapped_range()).to_vec();
        staging.unmap();
        ndarray::Array::from_shape_vec(self.shape.clone(), data).unwrap()
    }

    pub fn alloc_like(&self) -> wgpu::Buffer {
        self.ctx.device.create_buffer(&wgpu::BufferDescriptor {
            label: None,
            size: (self.size * 4) as u64,
            usage: TENSOR_USAGE,
            mapped_at_creation: false,
        })
    }

    pub fn download_raw(ctx: &GpuContext, buf: &wgpu::Buffer, count: usize) -> Vec<f32> {
        let byte_size = (count * 4) as u64;
        let staging = ctx.device.create_buffer(&wgpu::BufferDescriptor {
            label: None,
            size: byte_size,
            usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::MAP_READ,
            mapped_at_creation: false,
        });
        let mut enc = ctx.device.create_command_encoder(&Default::default());
        enc.copy_buffer_to_buffer(buf, 0, &staging, 0, byte_size);
        let si = ctx.queue.submit([enc.finish()]);

        let slice = staging.slice(..);
        slice.map_async(wgpu::MapMode::Read, |_| {});
        let _ = ctx.device.poll(wgpu::PollType::Wait {
            submission_index: Some(si),
            timeout: None,
        });

        let data = cast_slice::<u8, f32>(&slice.get_mapped_range()).to_vec();
        staging.unmap();
        data
    }
}

impl<D: Dimension> Clone for GpuTensor<D> {
    fn clone(&self) -> Self {
        self.ctx.flush();

        let new_buf = self.ctx.device.create_buffer(&wgpu::BufferDescriptor {
            label: None,
            size: self.buffer.size(),
            usage: TENSOR_USAGE,
            mapped_at_creation: false,
        });
        let mut enc = self.ctx.device.create_command_encoder(&Default::default());
        enc.copy_buffer_to_buffer(&self.buffer, 0, &new_buf, 0, self.buffer.size());
        self.ctx.queue.submit([enc.finish()]);
        GpuTensor {
            buffer: Arc::new(new_buf),
            shape: self.shape.clone(),
            size: self.size,
            ctx: self.ctx.clone(),
        }
    }
}

impl<D: Dimension + serde::Serialize> serde::Serialize for GpuTensor<D> {
    fn serialize<S: serde::Serializer>(&self, s: S) -> Result<S::Ok, S::Error> {
        self.download().serialize(s)
    }
}
impl<'de, D: Dimension + serde::Deserialize<'de>> serde::Deserialize<'de> for GpuTensor<D> {
    fn deserialize<De: serde::Deserializer<'de>>(d: De) -> Result<Self, De::Error> {
        let arr = ndarray::Array::<f32, D>::deserialize(d)?;
        Ok(GpuTensor::upload(arr, super::context::GpuContext::global()))
    }
}

impl<D: Dimension> From<ndarray::Array<f32, D>> for GpuTensor<D> {
    fn from(arr: ndarray::Array<f32, D>) -> Self {
        GpuTensor::upload(arr, GpuContext::global())
    }
}