ruda-driver-cpu 0.1.2

CPU driver backend for Ruda
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

use crate::{
    compilation::{MlirCompiler, register_supported_types},
    execution::server::CpuServer,
    device::CpuDevice,
};
use ruda_core::{device::DeviceService, profile::TimingMethod};
use ruda_kernel::dsl::{
    MemoryConfiguration, Runtime,
    client::ComputeClient,
    device::{DeviceId, ServerUtilitiesHandle},
    ir::{
        DeviceProperties, HardwareProperties, MemoryDeviceProperties, TargetProperties, VectorSize,
        features::Features,
    },
    server::ServerUtilities,
    zspace::{Shape, Strides},
};
use ruda::runtime::{allocator::ContiguousMemoryLayoutPolicy, logging::ServerLogger};
use ruda_kernel::library::tensor::is_contiguous;
use std::sync::Arc;
use sysinfo::System;

#[derive(Default)]
pub struct RuntimeOptions {
    /// Configures the memory management.
    pub memory_config: MemoryConfiguration,
}

#[derive(Debug, Clone)]
pub struct CpuRuntime;

pub type CpuCompiler = MlirCompiler;

mod device_service;

impl Runtime for CpuRuntime {
    type Compiler = CpuCompiler;
    type Server = CpuServer;
    type Device = CpuDevice;

    fn client(device: &Self::Device) -> ComputeClient<Self> {
        ComputeClient::load(device)
    }

    fn name(_client: &ComputeClient<Self>) -> &'static str {
        "cpu"
    }

    fn autotune_driver_fingerprint(_client: &ComputeClient<Self>) -> Option<String> {
        static ID: std::sync::OnceLock<Option<String>> = std::sync::OnceLock::new();
        ID.get_or_init(|| {
            let info = std::fs::read_to_string("/proc/cpuinfo").ok()?;
            let mut features: Vec<_> = info.lines().filter(|line| line.starts_with("model name") || line.starts_with("flags") || line.starts_with("Features") || line.starts_with("CPU implementer") || line.starts_with("CPU part")).collect();
            features.sort_unstable(); features.dedup();
            if features.is_empty() { None } else { Some(format!("host={}/{};{}", std::env::consts::ARCH, std::env::consts::OS, features.join(";"))) }
        }).clone()
    }

    fn max_ruda_count() -> (u32, u32, u32) {
        (u32::MAX, u32::MAX, u32::MAX)
    }

    fn can_read_tensor(shape: &Shape, strides: &Strides) -> bool {
        is_contiguous(shape, strides)
    }

    fn target_properties() -> TargetProperties {
        TargetProperties {
            // Values are irrelevant, since no wgsl backends currently support manual mma
            mma: Default::default(),
        }
    }

    fn enumerate_devices(
        _: u16,
        _: &<Self::Server as ruda_kernel::dsl::server::ComputeServer>::Info,
    ) -> Vec<DeviceId> {
        vec![DeviceId {
            type_id: 0,
            index_id: 0,
        }]
    }
}