mircuda-sys 0.1.0

Private native CUDA integration for mircuda
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

use std::{ffi::c_void, ptr::NonNull, sync::Arc};

use cudarc::driver::CudaStream;

use super::super::{
    driver::{Context, Stream},
    memory::{DeviceBuffer, ensure_stream},
};
use crate::{Error, Result};

unsafe extern "C" {
    fn mircuda_fp4_vector_create(
        n: i32,
        k: i32,
        stream: *mut c_void,
        output: *mut *mut c_void,
    ) -> i32;
    fn mircuda_fp4_vector_execute(
        plan: *mut c_void,
        stream: *mut c_void,
        input: *const c_void,
        input_scales: *const c_void,
        weight: *const c_void,
        weight_scales: *const c_void,
        output: *mut c_void,
        alpha: f32,
    ) -> i32;
    fn mircuda_fp4_vector_destroy(plan: *mut c_void);
}

#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub struct BlockScaledFp4VectorSpec {
    pub n: usize,
    pub k: usize,
}

#[derive(Debug)]
pub struct BlockScaledFp4VectorPlan {
    raw: NonNull<c_void>,
    stream: Arc<CudaStream>,
    spec: BlockScaledFp4VectorSpec,
}

// SAFETY: the plan retains its stream and execute requires exclusive access.
unsafe impl Send for BlockScaledFp4VectorPlan {}

impl Context {
    pub fn create_block_scaled_fp4_vector_plan(
        &self,
        stream: &Stream,
        spec: BlockScaledFp4VectorSpec,
    ) -> Result<BlockScaledFp4VectorPlan> {
        if !Arc::ptr_eq(&self.inner, stream.inner.context()) {
            return Err(Error::ContextMismatch);
        }
        self.inner.bind_to_thread()?;
        let mut raw = std::ptr::null_mut();
        // SAFETY: the output pointer is writable and the safe caller validates geometry.
        check(unsafe {
            mircuda_fp4_vector_create(
                i32::try_from(spec.n)?,
                i32::try_from(spec.k)?,
                stream.inner.cu_stream().cast(),
                &raw mut raw,
            )
        })?;
        Ok(BlockScaledFp4VectorPlan {
            raw: NonNull::new(raw).ok_or(Error::NullAllocation)?,
            stream: stream.inner.clone(),
            spec,
        })
    }
}

impl BlockScaledFp4VectorPlan {
    #[allow(clippy::too_many_arguments)]
    pub fn execute(
        &mut self,
        stream: &Stream,
        input: &DeviceBuffer,
        input_scales: &DeviceBuffer,
        weight: &DeviceBuffer,
        weight_scales: &DeviceBuffer,
        output: &DeviceBuffer,
        alpha: f32,
    ) -> Result<()> {
        if !Arc::ptr_eq(&self.stream, &stream.inner) {
            return Err(Error::StreamMismatch);
        }
        for buffer in [input, input_scales, weight, weight_scales, output] {
            ensure_stream(buffer, stream)?;
        }
        validate(self.spec, input, input_scales, weight, weight_scales, output)?;
        self.stream.context().bind_to_thread()?;
        // SAFETY: sizes, context, retained stream, and exclusive plan access were checked.
        check(unsafe {
            mircuda_fp4_vector_execute(
                self.raw.as_ptr(),
                stream.inner.cu_stream().cast(),
                input.pointer() as *const c_void,
                input_scales.pointer() as *const c_void,
                weight.pointer() as *const c_void,
                weight_scales.pointer() as *const c_void,
                output.pointer() as *mut c_void,
                alpha,
            )
        })
    }
}

impl Drop for BlockScaledFp4VectorPlan {
    fn drop(&mut self) {
        self.stream.context().record_err(self.stream.context().bind_to_thread());
        // SAFETY: this value uniquely owns the native plan.
        unsafe { mircuda_fp4_vector_destroy(self.raw.as_ptr()) };
    }
}

fn validate(
    spec: BlockScaledFp4VectorSpec,
    input: &DeviceBuffer,
    input_scales: &DeviceBuffer,
    weight: &DeviceBuffer,
    weight_scales: &DeviceBuffer,
    output: &DeviceBuffer,
) -> Result<()> {
    let valid = input.bytes() == spec.k / 2
        && input_scales.bytes() == scale_bytes(1, spec.k).unwrap_or(0)
        && weight.bytes() == spec.n.checked_mul(spec.k).unwrap_or(0) / 2
        && weight_scales.bytes() == scale_bytes(spec.n, spec.k).unwrap_or(0)
        && output.bytes() == spec.n * size_of::<u16>();
    if valid {
        Ok(())
    } else {
        Err(Error::InvalidMatmulBuffer)
    }
}

fn scale_bytes(rows: usize, columns: usize) -> Option<usize> {
    rows.div_ceil(128).checked_mul(columns.checked_div(64)?)?.checked_mul(512)
}

const fn check(status: i32) -> Result<()> {
    if status == 0 {
        Ok(())
    } else {
        Err(Error::Cutlass(status))
    }
}