mircuda 0.1.0

Native, explicit-stream Rust gateway to NVIDIA CUDA
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

use std::io::{self, Write};

use mircuda::{
    Context, DenseMatmulPlan, DenseMatmulSpec, DenseVectorPlan, DenseVectorSpec, DeviceBuffer,
    Driver, MemoryPool, Stream, bf16,
};

const CYCLES: u16 = 4;
const TARGET_BANK_BYTES: usize = 768 * 1_024 * 1_024;

#[derive(Clone, Copy)]
struct Case {
    name: &'static str,
    n: usize,
    k: usize,
}

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let driver = Driver::initialize()?;
    let device = driver.devices()?.into_iter().next().ok_or(mircuda::Error::InvalidLaunch)?;
    let context = driver.create_context(device)?;
    let stream = context.create_stream()?;
    let pool = context.default_memory_pool()?;
    let cases = [
        Case { name: "router", n: 128, k: 2_816 },
        Case { name: "dense-down", n: 2_816, k: 2_112 },
        Case {
            name: "attention-output",
            n: 2_816,
            k: 4_096,
        },
        Case {
            name: "dense-gate-up",
            n: 4_224,
            k: 2_816,
        },
        Case { name: "local-qkv", n: 8_192, k: 2_816 },
        Case { name: "global-qkv", n: 10_240, k: 2_816 },
        Case {
            name: "output-head",
            n: 262_144,
            k: 2_816,
        },
    ];
    let mut stdout = io::stdout().lock();
    for case in cases {
        run(&context, &stream, &pool, case, &mut stdout)?;
    }
    Ok(())
}

fn run(
    context: &Context,
    stream: &Stream,
    pool: &MemoryPool,
    case: Case,
    stdout: &mut impl Write,
) -> Result<(), Box<dyn std::error::Error>> {
    let elements = case.n.checked_mul(case.k).ok_or(mircuda::Error::InvalidMatmulShape)?;
    let bytes = elements
        .checked_mul(size_of::<bf16>())
        .ok_or(mircuda::Error::InvalidMatmulShape)?;
    let bank_count = (TARGET_BANK_BYTES / bytes).clamp(1, 64);
    let template = values(elements)?;
    let mut banks = Vec::with_capacity(bank_count);
    for _ in 0..bank_count {
        banks.push(upload(context, stream, pool, &template)?);
    }
    let input = upload(context, stream, pool, &values(case.k)?)?;
    let mut output = pool.allocate::<bf16>(stream, case.n)?;
    let mut matrix =
        DenseMatmulPlan::new(context, stream, DenseMatmulSpec::new(1, case.n, case.k)?)?;
    let mut vector = DenseVectorPlan::new(context, stream, DenseVectorSpec::new(case.n, case.k)?)?;
    let matrix_us = measure(context, stream, &banks, |weight| {
        matrix.execute(stream, &input, weight, &mut output, 1.0, 0.0)
    })?;
    let vector_us = measure(context, stream, &banks, |weight| {
        vector.execute(stream, &input, weight, &mut output, 1.0, 0.0)
    })?;
    writeln!(
        stdout,
        "{}: n={} k={} banks={} matrix={matrix_us:.3} us vector={vector_us:.3} us speedup={:.3}x",
        case.name,
        case.n,
        case.k,
        banks.len(),
        matrix_us / vector_us
    )?;
    Ok(())
}

fn measure(
    context: &Context,
    stream: &Stream,
    banks: &[DeviceBuffer<bf16>],
    mut execute: impl FnMut(&DeviceBuffer<bf16>) -> mircuda::Result<()>,
) -> mircuda::Result<f32> {
    for weight in banks {
        execute(weight)?;
    }
    stream.synchronize()?;
    let started = context.create_event(true)?;
    let completed = context.create_event(true)?;
    started.record(stream)?;
    for _ in 0..CYCLES {
        for weight in banks {
            execute(weight)?;
        }
    }
    completed.record(stream)?;
    completed.synchronize()?;
    let operations = f32::from(CYCLES) * f32::from(u16::try_from(banks.len())?);
    Ok(started.elapsed_ms(&completed)? * 1_000.0 / operations)
}

fn values(count: usize) -> mircuda::Result<Vec<bf16>> {
    (0..count)
        .map(|index| {
            let value = f32::from(u8::try_from(index % 251)?) / 256.0 - 0.5;
            Ok(bf16::from_f32(value))
        })
        .collect()
}

fn upload(
    context: &Context,
    stream: &Stream,
    pool: &MemoryPool,
    values: &[bf16],
) -> mircuda::Result<DeviceBuffer<bf16>> {
    let mut host = context.allocate_pinned::<bf16>(values.len())?;
    host.copy_from_slice(values)?;
    let mut device = pool.allocate::<bf16>(stream, values.len())?;
    stream.copy_to_device(&mut host, &mut device)?;
    Ok(device)
}