nexar 0.1.2

Distributed runtime with QUIC transport, stream-multiplexed messaging, and built-in collectives
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
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
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246

//! Shared element-wise reduction primitives used by collective algorithms
//! and the CPU device adapter.

use crate::error::Result;
use crate::reduce_types::{Bf16, F16, Identity, LeBytes, Reducible};
use crate::types::{DataType, ReduceOp};

/// Element-wise reduce on byte slices interpreted as `dtype` elements.
///
/// `dst` and `src` must both contain exactly `count * dtype.size_in_bytes()` bytes.
pub(crate) fn reduce_slice(
    dst: &mut [u8],
    src: &[u8],
    count: usize,
    dtype: DataType,
    op: ReduceOp,
) -> Result<()> {
    match dtype {
        DataType::F32 => {
            if unsafe { crate::reduce_simd::reduce_f32_simd(dst, src, count, op) } {
                return Ok(());
            }
            reduce_slice_typed::<f32>(dst, src, count, op);
        }
        DataType::F64 => {
            if unsafe { crate::reduce_simd::reduce_f64_simd(dst, src, count, op) } {
                return Ok(());
            }
            reduce_slice_typed::<f64>(dst, src, count, op);
        }
        DataType::BF16 => {
            if unsafe { crate::reduce_simd::reduce_bf16_simd(dst, src, count, op) } {
                return Ok(());
            }
            reduce_slice_typed::<Bf16>(dst, src, count, op);
        }
        DataType::F16 => {
            if unsafe { crate::reduce_simd::reduce_f16_simd(dst, src, count, op) } {
                return Ok(());
            }
            reduce_slice_typed::<F16>(dst, src, count, op);
        }
        DataType::I32 => reduce_slice_typed::<i32>(dst, src, count, op),
        DataType::I64 => reduce_slice_typed::<i64>(dst, src, count, op),
        DataType::U32 => reduce_slice_typed::<u32>(dst, src, count, op),
        DataType::U64 => reduce_slice_typed::<u64>(dst, src, count, op),
        DataType::I8 => reduce_slice_typed::<i8>(dst, src, count, op),
        DataType::U8 => reduce_slice_typed::<u8>(dst, src, count, op),
    }
    Ok(())
}

/// Element-wise reduce via raw pointer (for device adapter).
///
/// # Safety
/// `dst_ptr` must point to at least `count` elements of type matching `dtype`.
/// `src` must contain exactly `count * dtype.size_in_bytes()` bytes.
pub unsafe fn reduce_ptr(
    dst_ptr: u64,
    src: &[u8],
    count: usize,
    dtype: DataType,
    op: ReduceOp,
) -> Result<()> {
    unsafe {
        match dtype {
            DataType::F32 => reduce_ptr_typed::<f32>(dst_ptr, src, count, op),
            DataType::F64 => reduce_ptr_typed::<f64>(dst_ptr, src, count, op),
            DataType::BF16 => reduce_ptr_typed::<Bf16>(dst_ptr, src, count, op),
            DataType::F16 => reduce_ptr_typed::<F16>(dst_ptr, src, count, op),
            DataType::I32 => reduce_ptr_typed::<i32>(dst_ptr, src, count, op),
            DataType::I64 => reduce_ptr_typed::<i64>(dst_ptr, src, count, op),
            DataType::U32 => reduce_ptr_typed::<u32>(dst_ptr, src, count, op),
            DataType::U64 => reduce_ptr_typed::<u64>(dst_ptr, src, count, op),
            DataType::I8 => reduce_ptr_typed::<i8>(dst_ptr, src, count, op),
            DataType::U8 => reduce_ptr_typed::<u8>(dst_ptr, src, count, op),
        }
    }
    Ok(())
}

fn reduce_slice_typed<T: Reducible + LeBytes>(
    dst: &mut [u8],
    src: &[u8],
    count: usize,
    op: ReduceOp,
) {
    let t_size = std::mem::size_of::<T>();
    for i in 0..count {
        let off = i * t_size;
        let a = T::read_le(&dst[off..off + t_size]);
        let b = T::read_le(&src[off..off + t_size]);
        let r = T::reduce(a, b, op);
        r.write_le(&mut dst[off..off + t_size]);
    }
}

/// Returns a byte buffer of `count * dtype.size_in_bytes()` filled with the
/// identity element for the given op+dtype combination.
///
/// Identity values: Sum→0, Prod→1, Min→type::MAX, Max→type::MIN.
pub(crate) fn identity_slice(count: usize, dtype: DataType, op: ReduceOp) -> Result<Vec<u8>> {
    match dtype {
        DataType::F32 => Ok(identity_slice_typed::<f32>(count, op)),
        DataType::F64 => Ok(identity_slice_typed::<f64>(count, op)),
        DataType::BF16 => Ok(identity_slice_typed::<Bf16>(count, op)),
        DataType::F16 => Ok(identity_slice_typed::<F16>(count, op)),
        DataType::I32 => Ok(identity_slice_typed::<i32>(count, op)),
        DataType::I64 => Ok(identity_slice_typed::<i64>(count, op)),
        DataType::U32 => Ok(identity_slice_typed::<u32>(count, op)),
        DataType::U64 => Ok(identity_slice_typed::<u64>(count, op)),
        DataType::I8 => Ok(identity_slice_typed::<i8>(count, op)),
        DataType::U8 => Ok(identity_slice_typed::<u8>(count, op)),
    }
}

fn identity_slice_typed<T: Identity>(count: usize, op: ReduceOp) -> Vec<u8> {
    let val = T::identity(op);
    let t_size = std::mem::size_of::<T>();
    let mut buf = vec![0u8; count * t_size];
    for i in 0..count {
        val.write_le(&mut buf[i * t_size..(i + 1) * t_size]);
    }
    buf
}

/// # Safety
/// `dst_ptr` must point to at least `count` elements of type `T`.
unsafe fn reduce_ptr_typed<T: Reducible>(dst_ptr: u64, src: &[u8], count: usize, op: ReduceOp) {
    let dst = dst_ptr as *mut T;
    let src_typed = src.as_ptr() as *const T;
    for i in 0..count {
        let d = unsafe { *dst.add(i) };
        let s = unsafe { *src_typed.add(i) };
        unsafe { *dst.add(i) = T::reduce(d, s, op) };
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_reduce_slice_sum_f32() {
        let mut dst = [1.0f32, 2.0, 3.0, 4.0];
        let src = [10.0f32, 20.0, 30.0, 40.0];
        let dst_bytes: &mut [u8] =
            unsafe { std::slice::from_raw_parts_mut(dst.as_mut_ptr() as *mut u8, dst.len() * 4) };
        let src_bytes: &[u8] =
            unsafe { std::slice::from_raw_parts(src.as_ptr() as *const u8, src.len() * 4) };
        reduce_slice(dst_bytes, src_bytes, 4, DataType::F32, ReduceOp::Sum).unwrap();
        assert_eq!(dst, [11.0, 22.0, 33.0, 44.0]);
    }

    #[test]
    fn test_bf16_roundtrip() {
        let values = [0.0f32, 1.0, -1.0, 0.5, 3.125, 65504.0, -0.0];
        for &v in &values {
            let bf = Bf16::from_f32(v);
            let back = bf.to_f32();
            // BF16 has ~7-bit mantissa, so ~1% relative error for non-zero.
            if v == 0.0 {
                assert_eq!(back, 0.0);
            } else {
                assert!(
                    (back - v).abs() / v.abs() < 0.01,
                    "bf16 roundtrip failed for {v}"
                );
            }
        }
    }

    #[test]
    fn test_f16_roundtrip() {
        let values = [0.0f32, 1.0, -1.0, 0.5, 3.125, 65504.0, -0.0];
        for &v in &values {
            let h = F16::from_f32(v);
            let back = h.to_f32();
            if v == 0.0 {
                assert_eq!(back, 0.0);
            } else {
                assert!(
                    (back - v).abs() / v.abs() < 0.002,
                    "f16 roundtrip failed for {v}: got {back}"
                );
            }
        }
    }

    #[test]
    fn test_reduce_slice_bf16_sum() {
        let a = [Bf16::from_f32(1.0), Bf16::from_f32(2.0)];
        let b = [Bf16::from_f32(10.0), Bf16::from_f32(20.0)];
        let mut dst_buf = vec![0u8; 4];
        let mut src_buf = vec![0u8; 4];
        a[0].write_le(&mut dst_buf[0..2]);
        a[1].write_le(&mut dst_buf[2..4]);
        b[0].write_le(&mut src_buf[0..2]);
        b[1].write_le(&mut src_buf[2..4]);

        reduce_slice(&mut dst_buf, &src_buf, 2, DataType::BF16, ReduceOp::Sum).unwrap();

        let r0 = Bf16::read_le(&dst_buf[0..2]).to_f32();
        let r1 = Bf16::read_le(&dst_buf[2..4]).to_f32();
        assert!((r0 - 11.0).abs() < 0.1);
        assert!((r1 - 22.0).abs() < 0.1);
    }

    #[test]
    fn test_reduce_slice_f16_sum() {
        let a = [F16::from_f32(1.0), F16::from_f32(2.0)];
        let b = [F16::from_f32(10.0), F16::from_f32(20.0)];
        let mut dst_buf = vec![0u8; 4];
        let mut src_buf = vec![0u8; 4];
        a[0].write_le(&mut dst_buf[0..2]);
        a[1].write_le(&mut dst_buf[2..4]);
        b[0].write_le(&mut src_buf[0..2]);
        b[1].write_le(&mut src_buf[2..4]);

        reduce_slice(&mut dst_buf, &src_buf, 2, DataType::F16, ReduceOp::Sum).unwrap();

        let r0 = F16::read_le(&dst_buf[0..2]).to_f32();
        let r1 = F16::read_le(&dst_buf[2..4]).to_f32();
        assert!((r0 - 11.0).abs() < 0.01);
        assert!((r1 - 22.0).abs() < 0.01);
    }

    #[test]
    fn test_reduce_ptr_sum_i32() {
        let mut dst: Vec<i32> = vec![1, 2, 3];
        let src: Vec<i32> = vec![10, 20, 30];
        let src_bytes: &[u8] =
            unsafe { std::slice::from_raw_parts(src.as_ptr() as *const u8, src.len() * 4) };
        unsafe {
            reduce_ptr(
                dst.as_mut_ptr() as u64,
                src_bytes,
                3,
                DataType::I32,
                ReduceOp::Sum,
            )
            .unwrap();
        }
        assert_eq!(dst, vec![11, 22, 33]);
    }
}