archmage 0.9.15

Safely invoke your intrinsic power, using the tokens granted to you by the CPU. Cast primitive magics faster than any mage alive.
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

//! ASM pattern verification for slice→SIMD load patterns.
//!
//! Each function is `#[unsafe(no_mangle)]` + `#[arcane]` so `cargo asm` can find it.
//! Verified by `scripts/verify-asm.sh`.
//!
//! Run: cargo asm -p archmage --bench asm_patterns --features "std avx512"

#![cfg(target_arch = "x86_64")]

use archmage::{Desktop64, SimdToken, arcane};
use std::arch::x86_64::*;

// ============================================================================
// 256-bit float loads
// ============================================================================

/// Baseline: load from array reference → vmovups
#[unsafe(no_mangle)]
#[arcane(import_intrinsics)]
fn load_array_ref(_t: Desktop64, data: &[f32; 8]) -> __m256 {
    _mm256_loadu_ps(data)
}

/// Slice via .first_chunk() → should produce same vmovups
#[unsafe(no_mangle)]
#[arcane(import_intrinsics)]
fn load_first_chunk(_t: Desktop64, data: &[f32]) -> __m256 {
    let arr: &[f32; 8] = data.first_chunk().unwrap();
    _mm256_loadu_ps(arr)
}

/// Slice via try_into → should produce same vmovups
#[unsafe(no_mangle)]
#[arcane(import_intrinsics)]
fn load_try_into(_t: Desktop64, data: &[f32]) -> __m256 {
    let arr: &[f32; 8] = data[..8].try_into().unwrap();
    _mm256_loadu_ps(arr)
}

// ============================================================================
// 256-bit integer load
// ============================================================================

/// Integer load via first_chunk → vmovdqu
#[unsafe(no_mangle)]
#[arcane(import_intrinsics)]
fn load_first_chunk_i(_t: Desktop64, data: &[u8]) -> __m256i {
    let arr: &[u8; 32] = data.first_chunk().unwrap();
    _mm256_loadu_si256(arr)
}

// ============================================================================
// 128-bit float load
// ============================================================================

/// 128-bit first_chunk → vmovups (128-bit)
#[unsafe(no_mangle)]
#[arcane(import_intrinsics)]
fn load_first_chunk_128(_t: Desktop64, data: &[f32]) -> __m128 {
    let arr: &[f32; 4] = data.first_chunk().unwrap();
    _mm_loadu_ps(arr)
}

// ============================================================================
// Store patterns
// ============================================================================

/// Store via first_chunk_mut → vmovups (store)
#[unsafe(no_mangle)]
#[arcane(import_intrinsics)]
fn store_first_chunk_mut(_t: Desktop64, v: __m256, out: &mut [f32]) {
    let arr: &mut [f32; 8] = out.first_chunk_mut().unwrap();
    _mm256_storeu_ps(arr, v);
}

// ============================================================================
// Magetypes patterns
// ============================================================================

/// magetypes from_slice → should produce vmovups
#[unsafe(no_mangle)]
#[arcane(import_intrinsics)]
fn load_f32x8_from_slice(_t: Desktop64, data: &[f32]) -> __m256 {
    use magetypes::simd::f32x8;
    let v = f32x8::from_slice(_t, data);
    v.raw()
}

/// magetypes load via first_chunk → should produce vmovups
#[unsafe(no_mangle)]
#[arcane(import_intrinsics)]
fn load_f32x8_first_chunk(_t: Desktop64, data: &[f32]) -> __m256 {
    use magetypes::simd::f32x8;
    let arr: &[f32; 8] = data.first_chunk().unwrap();
    let v = f32x8::load(_t, arr);
    v.raw()
}

// ============================================================================
// Criterion benchmark (required for cargo asm --bench to work)
// ============================================================================

use zenbench::criterion_compat::*;
use zenbench::{criterion_group, criterion_main};

fn bench_load_patterns(c: &mut Criterion) {
    let data = [1.0f32, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0];
    let slice: &[f32] = &data;
    let bytes = [0u8; 32];
    let byte_slice: &[u8] = &bytes;

    if let Some(token) = Desktop64::summon() {
        c.bench_function("load_array_ref", |b| {
            b.iter(|| load_array_ref(token, black_box(&data)))
        });

        c.bench_function("load_first_chunk", |b| {
            b.iter(|| load_first_chunk(token, black_box(slice)))
        });

        c.bench_function("load_try_into", |b| {
            b.iter(|| load_try_into(token, black_box(slice)))
        });

        c.bench_function("load_first_chunk_i", |b| {
            b.iter(|| load_first_chunk_i(token, black_box(byte_slice)))
        });

        c.bench_function("load_first_chunk_128", |b| {
            b.iter(|| load_first_chunk_128(token, black_box(slice)))
        });

        c.bench_function("load_f32x8_from_slice", |b| {
            b.iter(|| load_f32x8_from_slice(token, black_box(slice)))
        });

        c.bench_function("load_f32x8_first_chunk", |b| {
            b.iter(|| load_f32x8_first_chunk(token, black_box(slice)))
        });
    } else {
        eprintln!("Desktop64 not available, skipping benchmarks");
    }
}

criterion_group!(benches, bench_load_patterns);
criterion_main!(benches);