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
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
163
164
165

#![cfg(all(target_os = "linux", feature = "cutlass"))]

use mircuda::{
    Context, DenseMatmulPlan, DenseMatmulSpec, DenseVectorPlan, DenseVectorSpec, DeviceBuffer,
    DeviceElement, Driver, MemoryPool, Stream, bf16, f16,
};

const N: usize = 128;
const K: usize = 128;

#[test]
fn dense_bf16_supports_decode_and_prefill() -> mircuda::Result<()> {
    run_bf16(1)?;
    run_bf16(128)
}

#[test]
fn dense_bf16_supports_profiled_wide_decode() -> mircuda::Result<()> {
    const OUTPUTS: usize = 8_192;
    let (context, stream, pool) = environment()?;
    let input = copy_device(&context, &stream, &pool, &[bf16::ONE; K])?;
    let weight = copy_device(&context, &stream, &pool, &vec![bf16::ONE; OUTPUTS * K])?;
    let mut output = copy_device(&context, &stream, &pool, &[bf16::ZERO; OUTPUTS])?;
    DenseMatmulPlan::new(&context, &stream, DenseMatmulSpec::new(1, OUTPUTS, K)?)?
        .execute(&stream, &input, &weight, &mut output, 1.0, 0.0)?;
    let actual = read_device(&context, &stream, &output)?;
    assert!(actual.iter().all(|value| *value == bf16::from_f32(128.0)));
    Ok(())
}

#[test]
fn dense_bf16_retains_fp32_output() -> mircuda::Result<()> {
    let (context, stream, pool) = environment()?;
    let input = copy_device(&context, &stream, &pool, &[bf16::from_f32(0.1); K])?;
    let weight = copy_device(&context, &stream, &pool, &vec![bf16::from_f32(0.1); N * K])?;
    let mut output = copy_device(&context, &stream, &pool, &[f32::NAN; N])?;
    DenseMatmulPlan::<bf16, f32>::new(&context, &stream, DenseMatmulSpec::new(1, N, K)?)?
        .execute(&stream, &input, &weight, &mut output, 1.0, 0.0)?;
    let actual = read_device(&context, &stream, &output)?;
    let rounded = bf16::from_f32(actual[0]).to_f32();
    assert!(
        actual
            .iter()
            .all(|value| value.is_finite() && (*value - actual[0]).abs() < f32::EPSILON)
    );
    assert!((actual[0] - rounded).abs() > f32::EPSILON, "FP32 result was rounded to BF16");
    Ok(())
}

#[test]
fn dense_bf16_supports_unaligned_shapes() -> mircuda::Result<()> {
    let (context, stream, pool) = environment()?;
    let a = copy_device(&context, &stream, &pool, &[bf16::ONE; 6])?;
    let b = copy_device(&context, &stream, &pool, &[bf16::ONE; 12])?;
    let mut c = copy_device(&context, &stream, &pool, &[bf16::ONE; 8])?;
    let mut plan = DenseMatmulPlan::<bf16>::new(&context, &stream, DenseMatmulSpec::new(2, 4, 3)?)?;
    plan.execute(&stream, &a, &b, &mut c, 1.0, 1.0)?;
    drop(plan);
    let actual = read_device(&context, &stream, &c)?;
    assert!(actual.iter().all(|value| *value == bf16::from_f32(4.0)));
    Ok(())
}

#[test]
fn dense_f16_supports_decode() -> mircuda::Result<()> {
    let (context, stream, pool) = environment()?;
    let a = copy_device(&context, &stream, &pool, &vec![f16::ONE; K])?;
    let b = copy_device(&context, &stream, &pool, &vec![f16::ONE; N * K])?;
    let mut c = copy_device(&context, &stream, &pool, &vec![f16::ONE; N])?;
    let mut plan = DenseMatmulPlan::<f16>::new(&context, &stream, DenseMatmulSpec::new(1, N, K)?)?;
    plan.execute(&stream, &a, &b, &mut c, 1.0, 1.0)?;
    drop(plan);
    let actual = read_device(&context, &stream, &c)?;
    let expected = f16::from_f32(129.0);
    assert!(actual.iter().all(|value| *value == expected));
    Ok(())
}

#[test]
fn dense_bf16_vector_matches_tensor_core_gemm() -> mircuda::Result<()> {
    const OUTPUTS: usize = 257;
    const FEATURES: usize = 2_816;
    let (context, stream, pool) = environment()?;
    let input = (0..FEATURES)
        .map(|index| Ok(bf16::from_f32(f32::from(u8::try_from(index % 31)?) / 32.0 - 0.5)))
        .collect::<mircuda::Result<Vec<_>>>()?;
    let weight = (0..OUTPUTS * FEATURES)
        .map(|index| Ok(bf16::from_f32(f32::from(u8::try_from(index % 17)?) / 64.0 - 0.125)))
        .collect::<mircuda::Result<Vec<_>>>()?;
    let input = copy_device(&context, &stream, &pool, &input)?;
    let weight = copy_device(&context, &stream, &pool, &weight)?;
    let mut expected = copy_device(&context, &stream, &pool, &[bf16::ZERO; OUTPUTS])?;
    let mut actual = copy_device(&context, &stream, &pool, &[bf16::NAN; OUTPUTS])?;
    DenseMatmulPlan::new(&context, &stream, DenseMatmulSpec::new(1, OUTPUTS, FEATURES)?)?
        .execute(&stream, &input, &weight, &mut expected, 1.0, 0.0)?;
    DenseVectorPlan::new(&context, &stream, DenseVectorSpec::new(OUTPUTS, FEATURES)?)?
        .execute(&stream, &input, &weight, &mut actual, 1.0, 0.0)?;
    let expected = read_device(&context, &stream, &expected)?;
    let actual = read_device(&context, &stream, &actual)?;
    assert!(actual.iter().all(|value| value.to_f32().is_finite()));
    let maximum = |values: &[bf16]| {
        values
            .iter()
            .enumerate()
            .max_by(|left, right| left.1.to_f32().total_cmp(&right.1.to_f32()))
            .map(|(index, _)| index)
    };
    assert_eq!(maximum(&actual), maximum(&expected));
    let error = expected
        .iter()
        .zip(&actual)
        .map(|(left, right)| (left.to_f32() - right.to_f32()).abs())
        .fold(0.0_f32, f32::max);
    assert!(error <= 0.125, "maximum BF16 GEMV difference: {error}");
    Ok(())
}

fn run_bf16(m: usize) -> mircuda::Result<()> {
    let (context, stream, pool) = environment()?;
    let a = copy_device(&context, &stream, &pool, &vec![bf16::ONE; m * K])?;
    let b = copy_device(&context, &stream, &pool, &vec![bf16::ONE; N * K])?;
    let mut c = copy_device(&context, &stream, &pool, &vec![bf16::ONE; m * N])?;
    let mut plan = DenseMatmulPlan::<bf16>::new(&context, &stream, DenseMatmulSpec::new(m, N, K)?)?;
    plan.execute(&stream, &a, &b, &mut c, 0.5, 1.0)?;
    assert_eq!(plan.spec(), DenseMatmulSpec::new(m, N, K)?);
    drop(plan);
    let actual = read_device(&context, &stream, &c)?;
    let expected = bf16::from_f32(65.0);
    assert!(actual.iter().all(|value| *value == expected));
    Ok(())
}

fn environment() -> mircuda::Result<(Context, Stream, MemoryPool)> {
    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()?;
    Ok((context, stream, pool))
}

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

fn read_device<T: DeviceElement>(
    context: &Context,
    stream: &Stream,
    device: &DeviceBuffer<T>,
) -> mircuda::Result<Vec<T>> {
    let mut host = context.allocate_pinned::<T>(device.len())?;
    stream.copy_to_host(device, &mut host)?;
    stream.synchronize()?;
    host.to_vec()
}