numr 0.5.2

High-performance numerical computing with multi-backend GPU acceleration (CPU/CUDA/WebGPU)
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
147
148
149
150

//! Matrix decompositions: LU, Cholesky, QR

use wgpu::{Buffer, Queue};

use super::check_dtype_f32;
use crate::dtype::DType;
use crate::error::Result;
use crate::runtime::wgpu::shaders::linalg_shaders::decompositions::DECOMPOSITIONS_SHADER;
use crate::runtime::wgpu::shaders::pipeline::{LayoutKey, PipelineCache};

/// Launch LU decomposition kernel with partial pivoting.
pub fn launch_lu_decompose(
    cache: &PipelineCache,
    queue: &Queue,
    lu_matrix: &Buffer,
    pivots: &Buffer,
    num_swaps: &Buffer,
    singular_flag: &Buffer,
    params_buffer: &Buffer,
    dtype: DType,
) -> Result<()> {
    check_dtype_f32!(dtype, "lu_decompose");

    let module = cache.get_or_create_module("linalg_decompositions", DECOMPOSITIONS_SHADER);
    let layout = cache.get_or_create_layout(LayoutKey {
        num_storage_buffers: 4,
        num_uniform_buffers: 1,
        num_readonly_storage: 0,
    });
    let pipeline = cache.get_or_create_pipeline(
        "linalg_decompositions",
        "lu_decompose_f32",
        &module,
        &layout,
    );

    let bind_group = cache.create_bind_group(
        &layout,
        &[lu_matrix, pivots, num_swaps, singular_flag, params_buffer],
    );

    let mut encoder = cache
        .device()
        .create_command_encoder(&wgpu::CommandEncoderDescriptor {
            label: Some("lu_decompose"),
        });

    {
        let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
            label: Some("lu_decompose"),
            timestamp_writes: None,
        });
        pass.set_pipeline(&pipeline);
        pass.set_bind_group(0, Some(&bind_group), &[]);
        pass.dispatch_workgroups(1, 1, 1); // Single thread for sequential algorithm
    }

    queue.submit(std::iter::once(encoder.finish()));
    Ok(())
}

/// Launch Cholesky decomposition kernel.
pub fn launch_cholesky_decompose(
    cache: &PipelineCache,
    queue: &Queue,
    l_matrix: &Buffer,
    not_pd_flag: &Buffer,
    params_buffer: &Buffer,
    dtype: DType,
) -> Result<()> {
    check_dtype_f32!(dtype, "cholesky_decompose");

    let module = cache.get_or_create_module("linalg_decompositions", DECOMPOSITIONS_SHADER);
    let layout = cache.get_or_create_layout(LayoutKey {
        num_storage_buffers: 2,
        num_uniform_buffers: 1,
        num_readonly_storage: 0,
    });
    let pipeline =
        cache.get_or_create_pipeline("linalg", "cholesky_decompose_f32", &module, &layout);

    let bind_group = cache.create_bind_group(&layout, &[l_matrix, not_pd_flag, params_buffer]);

    let mut encoder = cache
        .device()
        .create_command_encoder(&wgpu::CommandEncoderDescriptor {
            label: Some("cholesky_decompose"),
        });

    {
        let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
            label: Some("cholesky_decompose"),
            timestamp_writes: None,
        });
        pass.set_pipeline(&pipeline);
        pass.set_bind_group(0, Some(&bind_group), &[]);
        pass.dispatch_workgroups(1, 1, 1); // Single thread for sequential algorithm
    }

    queue.submit(std::iter::once(encoder.finish()));
    Ok(())
}

/// Launch QR decomposition kernel using Householder reflections.
pub fn launch_qr_decompose(
    cache: &PipelineCache,
    queue: &Queue,
    q_matrix: &Buffer,
    r_matrix: &Buffer,
    workspace: &Buffer,
    params_buffer: &Buffer,
    dtype: DType,
) -> Result<()> {
    check_dtype_f32!(dtype, "qr_decompose");

    let module = cache.get_or_create_module("linalg_decompositions", DECOMPOSITIONS_SHADER);
    let layout = cache.get_or_create_layout(LayoutKey {
        num_storage_buffers: 3,
        num_uniform_buffers: 1,
        num_readonly_storage: 0,
    });
    let pipeline = cache.get_or_create_pipeline(
        "linalg_decompositions",
        "qr_decompose_f32",
        &module,
        &layout,
    );

    let bind_group =
        cache.create_bind_group(&layout, &[q_matrix, r_matrix, workspace, params_buffer]);

    let mut encoder = cache
        .device()
        .create_command_encoder(&wgpu::CommandEncoderDescriptor {
            label: Some("qr_decompose"),
        });

    {
        let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
            label: Some("qr_decompose"),
            timestamp_writes: None,
        });
        pass.set_pipeline(&pipeline);
        pass.set_bind_group(0, Some(&bind_group), &[]);
        pass.dispatch_workgroups(1, 1, 1); // Single thread for sequential algorithm
    }

    queue.submit(std::iter::once(encoder.finish()));
    Ok(())
}