numr 0.5.2

High-performance numerical computing with multi-backend GPU acceleration (CPU/CUDA/WebGPU)
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

//! AVX-512 normalization kernels
//!
//! SIMD-optimized RMS norm and layer norm using:
//! - Vertical FMA accumulation for sum of squares
//! - Horizontal reduction intrinsics
//! - Vectorized final normalization pass

pub(super) const F32_LANES: usize = 16;
pub(super) const F64_LANES: usize = 8;

mod fused_add_layer_norm;
mod fused_add_rms_norm;
mod layer_norm;
mod rms_norm;

pub use fused_add_layer_norm::{
    fused_add_layer_norm_bwd_f32, fused_add_layer_norm_bwd_f64, fused_add_layer_norm_f32,
    fused_add_layer_norm_f64,
};
pub use fused_add_rms_norm::{
    fused_add_rms_norm_bwd_f32, fused_add_rms_norm_bwd_f64, fused_add_rms_norm_f32,
    fused_add_rms_norm_f64,
};
pub use layer_norm::{layer_norm_f32, layer_norm_f64};
pub use rms_norm::{rms_norm_f32, rms_norm_f64};