solvr 0.2.0-beta.2

Advanced computing library for real-world problem solving - optimization, differential equations, interpolation, statistics, and more
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
151
152
153
154
155
156
157
158
159
160
161
162

//! CUDA implementations of multivariate minimization algorithms.

mod newton_cg;
mod trust_exact;
mod trust_krylov;
mod trust_ncg;

use numr::error::Result;
use numr::runtime::cuda::{CudaClient, CudaRuntime};
use numr::tensor::Tensor;

use crate::optimize::error::OptimizeResult;
use crate::optimize::minimize::MinimizeOptions;

use super::impl_generic::{
    LbfgsOptions, TensorMinimizeResult, bfgs_impl, conjugate_gradient_impl, lbfgs_impl,
    nelder_mead_impl, powell_impl,
};
use crate::optimize::impl_generic::scalar::{
    bisect_impl, brentq_impl, minimize_scalar_brent_impl, newton_impl,
};
use crate::optimize::scalar::{MinimizeResult, RootResult, ScalarOptions};

impl crate::optimize::OptimizationAlgorithms<CudaRuntime> for CudaClient {
    // Scalar root finding methods
    fn bisect<F>(&self, f: F, a: f64, b: f64, options: &ScalarOptions) -> OptimizeResult<RootResult>
    where
        F: Fn(f64) -> f64,
    {
        bisect_impl(f, a, b, options)
    }

    fn brentq<F>(&self, f: F, a: f64, b: f64, options: &ScalarOptions) -> OptimizeResult<RootResult>
    where
        F: Fn(f64) -> f64,
    {
        brentq_impl(f, a, b, options)
    }

    fn newton<F, DF>(
        &self,
        f: F,
        df: DF,
        x0: f64,
        options: &ScalarOptions,
    ) -> OptimizeResult<RootResult>
    where
        F: Fn(f64) -> f64,
        DF: Fn(f64) -> f64,
    {
        newton_impl(f, df, x0, options)
    }

    fn minimize_scalar_brent<F>(
        &self,
        f: F,
        bracket: Option<(f64, f64, f64)>,
        options: &ScalarOptions,
    ) -> OptimizeResult<MinimizeResult>
    where
        F: Fn(f64) -> f64,
    {
        minimize_scalar_brent_impl(f, bracket, options)
    }

    // Multivariate minimization methods
    fn bfgs<F>(
        &self,
        f: F,
        x0: &Tensor<CudaRuntime>,
        options: &MinimizeOptions,
    ) -> OptimizeResult<TensorMinimizeResult<CudaRuntime>>
    where
        F: Fn(&Tensor<CudaRuntime>) -> Result<f64>,
    {
        bfgs_impl(self, f, x0, options)
    }

    fn lbfgs<F>(
        &self,
        f: F,
        x0: &Tensor<CudaRuntime>,
        options: &LbfgsOptions,
    ) -> OptimizeResult<TensorMinimizeResult<CudaRuntime>>
    where
        F: Fn(&Tensor<CudaRuntime>) -> Result<f64>,
    {
        lbfgs_impl(self, f, x0, options)
    }

    fn nelder_mead<F>(
        &self,
        f: F,
        x0: &Tensor<CudaRuntime>,
        options: &MinimizeOptions,
    ) -> OptimizeResult<TensorMinimizeResult<CudaRuntime>>
    where
        F: Fn(&Tensor<CudaRuntime>) -> Result<f64>,
    {
        nelder_mead_impl(self, f, x0, options)
    }

    fn powell<F>(
        &self,
        f: F,
        x0: &Tensor<CudaRuntime>,
        options: &MinimizeOptions,
    ) -> OptimizeResult<TensorMinimizeResult<CudaRuntime>>
    where
        F: Fn(&Tensor<CudaRuntime>) -> Result<f64>,
    {
        powell_impl(self, f, x0, options)
    }

    fn conjugate_gradient<F>(
        &self,
        f: F,
        x0: &Tensor<CudaRuntime>,
        options: &MinimizeOptions,
    ) -> OptimizeResult<TensorMinimizeResult<CudaRuntime>>
    where
        F: Fn(&Tensor<CudaRuntime>) -> Result<f64>,
    {
        conjugate_gradient_impl(self, f, x0, options)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::optimize::OptimizationAlgorithms;
    use numr::runtime::cuda::CudaDevice;

    fn setup() -> Option<(CudaDevice, CudaClient)> {
        let device = CudaDevice::new(0);
        let client = CudaClient::new(device.clone()).ok()?;
        Some((device, client))
    }

    #[test]
    fn test_bfgs_cuda() {
        let Some((device, client)) = setup() else {
            eprintln!("Skipping CUDA test: no device");
            return;
        };
        let x0 = Tensor::<CudaRuntime>::from_slice(&[1.0, 1.0], &[2], &device);

        let result = client
            .bfgs(
                |x| {
                    let data: Vec<f64> = x.to_vec();
                    Ok(data.iter().map(|xi| xi * xi).sum())
                },
                &x0,
                &MinimizeOptions::default(),
            )
            .unwrap();

        assert!(result.converged);
        assert!(result.fun < 1e-6);
    }
}