Skip to main content

vibe/
accel.rs

1//! Acceleration detection and fail-closed numeric kernels (P16.3).
2//!
3//! Default Vibe eval stays scalar AST. SIMD is used only by the numeric
4//! slice kernel when the CPU advertises it. WASM stays sequential and
5//! caps a cell cluster at 64.
6
7use crate::bind::AccelerationTier;
8
9/// Native ready-cluster ceiling (P16.2).
10pub const CELL_BATCH_NATIVE: usize = 512;
11/// WASM sequential chunk (P16.3).
12pub const CELL_BATCH_WASM: usize = 64;
13
14/// Cell batch capacity.
15pub fn cell_batch_cap() -> usize {
16    #[cfg(target_arch = "wasm32")]
17    {
18        CELL_BATCH_WASM
19    }
20    #[cfg(not(target_arch = "wasm32"))]
21    {
22        CELL_BATCH_NATIVE
23    }
24}
25
26/// What the CPU/adapter could run. Not what this eval pass is using.
27pub fn detect_available_tier() -> AccelerationTier {
28    #[cfg(target_arch = "wasm32")]
29    {
30        AccelerationTier::ScalarCpu
31    }
32    #[cfg(not(target_arch = "wasm32"))]
33    {
34        #[cfg(target_arch = "x86_64")]
35        if is_x86_feature_detected!("avx512f") {
36            return AccelerationTier::VectorSimd512;
37        }
38        AccelerationTier::ScalarCpu
39    }
40}
41
42/// Elementwise `out[i] = a[i] + b[i]`. AVX-512 when present; else scalar.
43pub fn add_f64_slices(a: &[f64], b: &[f64], out: &mut [f64]) {
44    let n = a.len().min(b.len()).min(out.len());
45    #[cfg(all(target_arch = "x86_64", not(target_arch = "wasm32")))]
46    {
47        if is_x86_feature_detected!("avx512f") {
48            unsafe {
49                add_avx512(a, b, out, n);
50            }
51            return;
52        }
53    }
54    for i in 0..n {
55        out[i] = a[i] + b[i];
56    }
57}
58
59#[cfg(all(target_arch = "x86_64", not(target_arch = "wasm32")))]
60#[target_feature(enable = "avx512f")]
61unsafe fn add_avx512(a: &[f64], b: &[f64], out: &mut [f64], n: usize) {
62    use std::arch::x86_64::*;
63    let mut i = 0;
64    while i + 8 <= n {
65        let va = _mm512_loadu_pd(a.as_ptr().add(i));
66        let vb = _mm512_loadu_pd(b.as_ptr().add(i));
67        _mm512_storeu_pd(out.as_mut_ptr().add(i), _mm512_add_pd(va, vb));
68        i += 8;
69    }
70    while i < n {
71        out[i] = a[i] + b[i];
72        i += 1;
73    }
74}
75
76#[cfg(test)]
77mod tests {
78    use super::*;
79
80    #[test]
81    fn wasm_or_native_cap_is_bounded() {
82        let cap = cell_batch_cap();
83        assert!(cap == CELL_BATCH_NATIVE || cap == CELL_BATCH_WASM);
84        assert!(cap <= CELL_BATCH_NATIVE);
85    }
86
87    #[test]
88    fn scalar_add_matches_manual() {
89        let a = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0];
90        let b = [0.5; 9];
91        let mut out = [0.0; 9];
92        add_f64_slices(&a, &b, &mut out);
93        for i in 0..9 {
94            assert!((out[i] - (a[i] + b[i])).abs() < 1e-12);
95        }
96    }
97
98    #[test]
99    fn wasm_never_advertises_simd() {
100        #[cfg(target_arch = "wasm32")]
101        assert_eq!(detect_available_tier(), AccelerationTier::ScalarCpu);
102    }
103}