rscopulas 0.2.3

Core Rust library for fitting, evaluating, and sampling copula models and vine copulas
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

use crate::{
    accel,
    domain::{Device, ExecPolicy},
    errors::{BackendError, CopulaError},
};

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum Operation {
    DensityEval,
    VineLogPdf,
    PairBatchEval,
    PairFitScoring,
    KendallTauMatrix,
    Sample,
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum ExecutionStrategy {
    CpuSerial,
    CpuParallel,
    Cuda(u32),
    Metal,
}

pub(crate) fn resolve_strategy(
    policy: ExecPolicy,
    operation: Operation,
    batch_len: usize,
) -> Result<ExecutionStrategy, CopulaError> {
    match policy {
        ExecPolicy::Force(Device::Cpu) => Ok(cpu_strategy(operation, batch_len)),
        // GPU force requests only become device strategies for operations that
        // have an actual GPU-aware bridge. Everything else remains explicit:
        // either a deliberate CPU fallback at the call site or an
        // `Unsupported` backend error here.
        ExecPolicy::Force(Device::Cuda(device)) => {
            resolve_accelerated_device(Device::Cuda(device), operation, batch_len)
        }
        ExecPolicy::Force(Device::Metal) => {
            resolve_accelerated_device(Device::Metal, operation, batch_len)
        }
        // `Auto` stays conservative for now: it can pick CPU parallelism, but it
        // will not opportunistically jump to CUDA or Metal until the GPU parity
        // story is broader than the current narrow kernel set.
        ExecPolicy::Auto => Ok(cpu_strategy(operation, batch_len)),
    }
}

pub(crate) fn parallel_try_map_range_collect<T, F>(
    len: usize,
    strategy: ExecutionStrategy,
    f: F,
) -> Result<Vec<T>, CopulaError>
where
    T: Send,
    F: Fn(usize) -> Result<T, CopulaError> + Sync + Send,
{
    match strategy {
        ExecutionStrategy::CpuSerial => (0..len).map(f).collect(),
        ExecutionStrategy::CpuParallel => accel::parallel_try_map_range_collect(len, f),
        ExecutionStrategy::Cuda(_) | ExecutionStrategy::Metal => {
            unreachable!("GPU strategies must be handled before CPU range mapping")
        }
    }
}

fn resolve_accelerated_device(
    device: Device,
    operation: Operation,
    batch_len: usize,
) -> Result<ExecutionStrategy, CopulaError> {
    match operation {
        // Pair-fit uses CPU control flow around pair-batch evaluation. The outer
        // candidate loop remains a CPU strategy, while the pair-batch helper can
        // still use the forced GPU backend for supported Gaussian kernels.
        Operation::PairFitScoring => return Ok(cpu_strategy(operation, batch_len)),
        Operation::PairBatchEval | Operation::VineLogPdf => {}
        Operation::DensityEval | Operation::KendallTauMatrix | Operation::Sample => {
            let backend = backend_name(device);
            return Err(BackendError::Unsupported { backend }.into());
        }
    }

    let accel_device = accel_device(device);
    let backend = backend_name(device);
    if !accel::is_device_available(accel_device) {
        return Err(BackendError::Unavailable { backend }.into());
    }

    Ok(match device {
        Device::Cpu => ExecutionStrategy::CpuSerial,
        Device::Cuda(ordinal) => ExecutionStrategy::Cuda(ordinal),
        Device::Metal => ExecutionStrategy::Metal,
    })
}

fn cpu_parallel_supported(operation: Operation) -> bool {
    matches!(
        operation,
        Operation::DensityEval
            | Operation::VineLogPdf
            | Operation::PairBatchEval
            | Operation::PairFitScoring
            | Operation::KendallTauMatrix
            | Operation::Sample
    )
}

fn cpu_strategy(operation: Operation, batch_len: usize) -> ExecutionStrategy {
    if cpu_parallel_supported(operation) && batch_len >= parallel_threshold(operation) {
        ExecutionStrategy::CpuParallel
    } else {
        ExecutionStrategy::CpuSerial
    }
}

fn parallel_threshold(operation: Operation) -> usize {
    match operation {
        Operation::DensityEval | Operation::VineLogPdf | Operation::PairBatchEval => 128,
        Operation::PairFitScoring => 2,
        Operation::KendallTauMatrix => 4,
        Operation::Sample => 128,
    }
}

fn backend_name(device: Device) -> &'static str {
    match device {
        Device::Cpu => "cpu",
        Device::Cuda(_) => "cuda",
        Device::Metal => "metal",
    }
}

fn accel_device(device: Device) -> accel::Device {
    match device {
        Device::Cpu => accel::Device::Cpu,
        Device::Cuda(ordinal) => accel::Device::Cuda(ordinal),
        Device::Metal => accel::Device::Metal,
    }
}