mlx-native 0.9.0

Pure-Rust Metal GPU compute library for MLX-compatible inference on Apple Silicon
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

//! Quick benchmark: iter-15 per-element vs iter-15b tiled qmm_affine.
//!
//! Measures average per-call wall time over N iterations on a fixed
//! shape, releasing GPU between calls.  Goal: validate the iter-15b
//! "2-5× speedup" claim on a representative shape.
//!
//! Usage: `cargo run --release --bin bench_qmm_affine -- [M] [N] [K]`
//! Default: 64x4096x4096 (close to a real attention out-proj shape).

use std::time::Instant;

use mlx_native::ops::qmm_affine::{
    dispatch_qmm_affine_t_f32, dispatch_qmm_affine_t_f32_simd,
    dispatch_qmm_affine_t_f32_simd4, dispatch_qmm_affine_t_f32_tiled,
};
use mlx_native::{DType, KernelRegistry, MlxDevice};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let mut args = std::env::args().skip(1);
    let m: u32 = args
        .next()
        .map(|s| s.parse().expect("M"))
        .unwrap_or(64);
    let n: u32 = args
        .next()
        .map(|s| s.parse().expect("N"))
        .unwrap_or(4096);
    let k: u32 = args
        .next()
        .map(|s| s.parse().expect("K"))
        .unwrap_or(4096);
    let group_size: u32 = 32;
    let n_iter: usize = args
        .next()
        .map(|s| s.parse().expect("n_iter"))
        .unwrap_or(20);

    println!("[bench] shape M={m} N={n} K={k} group_size={group_size} iter={n_iter}");

    let device = MlxDevice::new()?;
    let mut registry = KernelRegistry::new();

    let m_us = m as usize;
    let n_us = n as usize;
    let k_us = k as usize;
    let groups = k_us / group_size as usize;

    let x: Vec<f32> = (0..(m_us * k_us))
        .map(|i| ((i as f32) * 0.013).sin() * 0.5)
        .collect();
    let q_int: Vec<u8> = (0..(n_us * k_us)).map(|i| ((i * 7) % 16) as u8).collect();
    let scales: Vec<f32> = (0..(n_us * groups))
        .map(|i| 0.05 + (i as f32) * 1e-5)
        .collect();
    let biases: Vec<f32> = (0..(n_us * groups))
        .map(|i| -0.1 + (i as f32) * 1e-5)
        .collect();

    let mut x_buf = device.alloc_buffer(m_us * k_us * 4, DType::F32, vec![m_us, k_us])?;
    x_buf.as_mut_slice::<f32>()?.copy_from_slice(&x);
    let mut q_buf = device.alloc_buffer(n_us * k_us, DType::U8, vec![n_us, k_us])?;
    q_buf.as_mut_slice::<u8>()?.copy_from_slice(&q_int);
    let mut s_buf = device.alloc_buffer(n_us * groups * 4, DType::F32, vec![n_us, groups])?;
    s_buf.as_mut_slice::<f32>()?.copy_from_slice(&scales);
    let mut b_buf = device.alloc_buffer(n_us * groups * 4, DType::F32, vec![n_us, groups])?;
    b_buf.as_mut_slice::<f32>()?.copy_from_slice(&biases);
    let y_buf = device.alloc_buffer(m_us * n_us * 4, DType::F32, vec![m_us, n_us])?;
    let mut meta = device.alloc_buffer(16, DType::U32, vec![4])?;
    meta.as_mut_slice::<u32>()?
        .copy_from_slice(&[m, n, k, group_size]);

    // Warm up both kernels.
    for _ in 0..3 {
        let mut encoder = device.command_encoder()?;
        dispatch_qmm_affine_t_f32(
            &mut encoder, &mut registry, device.metal_device(),
            &x_buf, &q_buf, &s_buf, &b_buf, &y_buf, &meta, m, n, k, group_size,
        )?;
        encoder.commit_and_wait()?;
    }
    for _ in 0..3 {
        let mut encoder = device.command_encoder()?;
        dispatch_qmm_affine_t_f32_tiled(
            &mut encoder, &mut registry, device.metal_device(),
            &x_buf, &q_buf, &s_buf, &b_buf, &y_buf, &meta, m, n, k, group_size,
        )?;
        encoder.commit_and_wait()?;
    }

    // Per-element kernel timing.
    let t0 = Instant::now();
    for _ in 0..n_iter {
        let mut encoder = device.command_encoder()?;
        dispatch_qmm_affine_t_f32(
            &mut encoder, &mut registry, device.metal_device(),
            &x_buf, &q_buf, &s_buf, &b_buf, &y_buf, &meta, m, n, k, group_size,
        )?;
        encoder.commit_and_wait()?;
    }
    let dt_pe = t0.elapsed();
    let avg_pe = dt_pe.as_secs_f64() / n_iter as f64;

    // Tiled kernel timing.
    let t0 = Instant::now();
    for _ in 0..n_iter {
        let mut encoder = device.command_encoder()?;
        dispatch_qmm_affine_t_f32_tiled(
            &mut encoder, &mut registry, device.metal_device(),
            &x_buf, &q_buf, &s_buf, &b_buf, &y_buf, &meta, m, n, k, group_size,
        )?;
        encoder.commit_and_wait()?;
    }
    let dt_tl = t0.elapsed();
    let avg_tl = dt_tl.as_secs_f64() / n_iter as f64;

    // Warm up + time the simdgroup-MMA kernel (iter-15c).
    for _ in 0..3 {
        let mut encoder = device.command_encoder()?;
        dispatch_qmm_affine_t_f32_simd(
            &mut encoder, &mut registry, device.metal_device(),
            &x_buf, &q_buf, &s_buf, &b_buf, &y_buf, &meta, m, n, k, group_size,
        )?;
        encoder.commit_and_wait()?;
    }
    let t0 = Instant::now();
    for _ in 0..n_iter {
        let mut encoder = device.command_encoder()?;
        dispatch_qmm_affine_t_f32_simd(
            &mut encoder, &mut registry, device.metal_device(),
            &x_buf, &q_buf, &s_buf, &b_buf, &y_buf, &meta, m, n, k, group_size,
        )?;
        encoder.commit_and_wait()?;
    }
    let dt_sm = t0.elapsed();
    let avg_sm = dt_sm.as_secs_f64() / n_iter as f64;

    // Warm up + time the 4-simdgroup-MMA kernel (iter-15c-2).
    for _ in 0..3 {
        let mut encoder = device.command_encoder()?;
        dispatch_qmm_affine_t_f32_simd4(
            &mut encoder, &mut registry, device.metal_device(),
            &x_buf, &q_buf, &s_buf, &b_buf, &y_buf, &meta, m, n, k, group_size,
        )?;
        encoder.commit_and_wait()?;
    }
    let t0 = Instant::now();
    for _ in 0..n_iter {
        let mut encoder = device.command_encoder()?;
        dispatch_qmm_affine_t_f32_simd4(
            &mut encoder, &mut registry, device.metal_device(),
            &x_buf, &q_buf, &s_buf, &b_buf, &y_buf, &meta, m, n, k, group_size,
        )?;
        encoder.commit_and_wait()?;
    }
    let dt_s4 = t0.elapsed();
    let avg_s4 = dt_s4.as_secs_f64() / n_iter as f64;

    let speedup_tl = avg_pe / avg_tl;
    let speedup_sm = avg_pe / avg_sm;
    let speedup_s4 = avg_pe / avg_s4;
    let flops_per_call = 2.0 * (m as f64) * (n as f64) * (k as f64);
    let pe_gflops = flops_per_call / avg_pe / 1e9;
    let tl_gflops = flops_per_call / avg_tl / 1e9;
    let sm_gflops = flops_per_call / avg_sm / 1e9;
    let s4_gflops = flops_per_call / avg_s4 / 1e9;

    println!("[bench] per-element: avg {:.3} ms = {:.1} GFLOPS", avg_pe * 1000.0, pe_gflops);
    println!("[bench] tiled:       avg {:.3} ms = {:.1} GFLOPS", avg_tl * 1000.0, tl_gflops);
    println!("[bench] simd-MMA:    avg {:.3} ms = {:.1} GFLOPS", avg_sm * 1000.0, sm_gflops);
    println!("[bench] simd4-MMA:   avg {:.3} ms = {:.1} GFLOPS", avg_s4 * 1000.0, s4_gflops);
    println!("[bench] speedup tiled  / per-element = {speedup_tl:.2}×");
    println!("[bench] speedup simd   / per-element = {speedup_sm:.2}×");
    println!("[bench] speedup simd4  / per-element = {speedup_s4:.2}×");
    Ok(())
}