tract-linalg 0.23.5

Tiny, no-nonsense, self contained, TensorFlow and ONNX inference
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

use crate::Ops;
use crate::frame::element_wise::ElementWiseKer;
use crate::frame::reduce::{MapReduceKer, ReduceKer};
use crate::x86_64_fma::softmax::x86_64_avx512_softmax2_fastcompact_f16_64n;
use crate::x86_64_fma::softmax::x86_64_fma_softmax2_fastcompact_f32_32n;

pub mod mmm;

mod amd_avx512_linear;
mod amd_fma_linear;
mod intel_avx512_linear;
mod intel_avx512_mmv_linear;
mod intel_fma_linear;

/// CPU vendor, the axis (with the AVX-512-vs-FMA tier, decided by which plug runs)
/// that selects a per-target `LinearCostModel`. `TRACT_X86_KIND=intel|amd|other`
/// overrides the CPUID probe (for forcing a cohort under emulation or in CI).
#[derive(PartialEq, Clone, Copy)]
pub(crate) enum Vendor {
    Intel,
    Amd,
    Other,
}

pub(crate) fn vendor() -> Vendor {
    if let Ok(k) = std::env::var("TRACT_X86_KIND") {
        return match k.as_str() {
            "intel" => Vendor::Intel,
            "amd" => Vendor::Amd,
            _ => Vendor::Other,
        };
    }
    // `unsafe` is required on the MSRV (1.91); newer rustc deems it redundant.
    #[allow(unused_unsafe)]
    let id = unsafe { std::arch::x86_64::__cpuid(0) };
    let mut s = [0u8; 12];
    s[0..4].copy_from_slice(&id.ebx.to_le_bytes());
    s[4..8].copy_from_slice(&id.edx.to_le_bytes());
    s[8..12].copy_from_slice(&id.ecx.to_le_bytes());
    match &s {
        b"GenuineIntel" => Vendor::Intel,
        b"AuthenticAMD" => Vendor::Amd,
        _ => Vendor::Other,
    }
}

pub mod act;
pub mod act_f16;
pub mod act_f16_fp16;

pub mod amx;
pub mod amx_bf16;
pub mod avxvnni;
pub mod by_scalar;
pub mod erf;
#[cfg(tract_avx512vnni)]
pub mod fma_width;
pub mod max;
pub mod min;
pub mod panel_extract;
pub mod rms_norm;
pub mod softmax;

const AVX: fn() -> bool = || is_x86_feature_detected!("avx");
const AVX2: fn() -> bool = || is_x86_feature_detected!("avx2");
const FMA: fn() -> bool = || is_x86_feature_detected!("fma");
const AVX512F: fn() -> bool = || is_x86_feature_detected!("avx512f");
#[cfg(tract_avx512vnni)]
const AVX512VNNI: fn() -> bool = || is_x86_feature_detected!("avx512vnni");

tanh_impl!(f32, fma_tanh_f32, 8, 8, is_x86_feature_detected!("fma"));
sigmoid_impl!(f32, fma_sigmoid_f32, 8, 8, is_x86_feature_detected!("fma"));
silu_impl!(f32, fma_silu_f32, 8, 8, is_x86_feature_detected!("fma"));

// AVX-without-FMA ports of the fma kernels above (each vfmadd132ps expanded
// to an in-place vmulps+vaddps pair) for CPUs outside the fma tier.
tanh_impl!(f32, avx_tanh_f32, 8, 8, is_x86_feature_detected!("avx"));
sigmoid_impl!(f32, avx_sigmoid_f32, 8, 8, is_x86_feature_detected!("avx"));

// AVX-512 (zmm, 16-wide) variants. The assembly lives in x86_64/avx512/; the
// main loop handles 64 lanes (4 zmm) per iteration with a 16-lane tail, so
// nr()=16 (any multiple of 16 is safe).
tanh_impl!(f32, avx512_tanh_f32, 16, 16, is_x86_feature_detected!("avx512f"));
sigmoid_impl!(f32, avx512_sigmoid_f32, 16, 16, is_x86_feature_detected!("avx512f"));
silu_impl!(f32, avx512_silu_f32, 16, 16, is_x86_feature_detected!("avx512f"));

fn plug_avx2(_ops: &mut Ops) {}

/// Element-wise kernels for AVX-capable CPUs outside the fma tier: the
/// mul_by_scalar / max / min asm is plain AVX, and sigmoid / tanh have
/// dedicated mul+add ports. softmax uses fma asm and keeps its generic
/// fallback on this tier.
fn plug_avx(ops: &mut Ops) {
    ops.sigmoid_f32 = Box::new(|| avx_sigmoid_f32::ew());
    ops.tanh_f32 = Box::new(|| avx_tanh_f32::ew());

    ops.mul_by_scalar_f32 = Box::new(|| by_scalar::x86_64_avx_f32_mul_by_scalar_32n::ew());
    ops.max_f32 = Box::new(|| max::x86_64_fma_max_f32_32n::red());
    ops.min_f32 = Box::new(|| min::x86_64_fma_min_f32_32n::red());

    log::info!("sigmoid_f32, tanh_f32, mul_by_scalar_f32, max_f32, min_f32: x86_64/avx activated");
}

fn plug_fma(ops: &mut Ops) {
    panel_extract::plug(ops);

    ops.sigmoid_f32 = Box::new(|| fma_sigmoid_f32::ew());
    ops.tanh_f32 = Box::new(|| fma_tanh_f32::ew());
    ops.silu_f32 = Box::new(|| fma_silu_f32::ew());

    ops.mul_by_scalar_f32 = Box::new(|| by_scalar::x86_64_avx_f32_mul_by_scalar_32n::ew());
    ops.max_f32 = Box::new(|| max::x86_64_fma_max_f32_32n::red());
    ops.min_f32 = Box::new(|| min::x86_64_fma_min_f32_32n::red());
    ops.softmax2_fastcompact_f32 = Box::new(|| x86_64_fma_softmax2_fastcompact_f32_32n::red());

    log::info!("sigmoid_f32, tanh_f32, silu_f32: x86_64/fma activated");
}

/// On hosts that also support AVX-512_FP16 (Sapphire Rapids / Granite Rapids /
/// later, and recent Xeon-D / consumer parts), upgrade the f16 element-wise
/// kernels from the f32-roundtrip implementations in `act_f16.rs` to the
/// native f16 implementations in `act_f16_fp16.rs` where the native path is
/// actually faster on this uarch. We benched each op against its f32-roundtrip
/// equivalent on Sapphire Rapids and only plug in the ones that win:
///
///   hardswish_f16:  8.71 → 31.6 Gelem/s  (3.62× native) — plug in
///   leaky_relu_f16: 9.44 →  5.85 Gelem/s (0.62× native — regression) — keep
///                   the f32-roundtrip version from act_f16.rs. The native
///                   kernel exists in act_f16_fp16.rs for future revisits but
///                   is not wired here.
fn plug_avx512fp16(ops: &mut Ops) {
    ops.hardswish_f16 = Box::new(|| act_f16_fp16::x86_64_avx512fp16_hardswish_f16_128n::ew());

    log::info!("hardswish_f16: x86_64/avx512fp16 native activated");
}

fn plug_avx512f(ops: &mut Ops) {
    ops.sigmoid_f32 = Box::new(|| avx512_sigmoid_f32::ew());
    ops.tanh_f32 = Box::new(|| avx512_tanh_f32::ew());
    ops.hardswish_f32 = Box::new(|| act::x86_64_avx512_hardswish_f32_64n::ew());
    ops.leaky_relu_f32 = Box::new(|| act::x86_64_avx512_leaky_relu_f32_64n::ew());
    ops.silu_f32 = Box::new(|| avx512_silu_f32::ew());
    ops.gelu_f32 = Box::new(|| act::x86_64_avx512_gelu_f32_16n::ew());

    ops.sigmoid_f16 = Box::new(|| act_f16::x86_64_avx512_sigmoid_f16_16n::ew());
    ops.tanh_f16 = Box::new(|| act_f16::x86_64_avx512_tanh_f16_16n::ew());
    ops.hardswish_f16 = Box::new(|| act_f16::x86_64_avx512_hardswish_f16_64n::ew());
    ops.leaky_relu_f16 = Box::new(|| act_f16::x86_64_avx512_leaky_relu_f16_64n::ew());
    ops.silu_f16 = Box::new(|| act_f16::x86_64_avx512_silu_f16_16n::ew());
    ops.gelu_f16 = Box::new(|| act_f16::x86_64_avx512_gelu_f16_16n::ew());

    ops.max_f32 = Box::new(|| max::x86_64_avx512_max_f32_64n::red());
    ops.softmax2_fastcompact_f32 =
        Box::new(|| softmax::x86_64_avx512_softmax2_fastcompact_f32_64n::red());
    ops.softmax2_fastcompact_f16 = Box::new(|| x86_64_avx512_softmax2_fastcompact_f16_64n::red());

    ops.erf_f32 = Box::new(|| erf::x86_64_avx512_erf_f32_64n::ew());

    ops.rms_norm_f32 = Box::new(rms_norm::rms_norm_f32);

    log::info!(
        "sigmoid_f32, tanh_f32, hardswish_f32, leaky_relu_f32, \
         silu_f32, gelu_f32, \
         sigmoid_f16, tanh_f16, hardswish_f16, leaky_relu_f16, \
         silu_f16, gelu_f16, \
         max_f32, softmax2_fastcompact_f32, softmax2_fastcompact_f16, erf_f32, \
         rms_norm_f32: x86_64/avx512f activated"
    );
}

pub fn plug(ops: &mut Ops) {
    mmm::plug(ops);
    if is_x86_feature_detected!("avx")
        && !(is_x86_feature_detected!("avx2") && is_x86_feature_detected!("fma"))
    {
        plug_avx(ops);
    }
    if is_x86_feature_detected!("avx2") {
        plug_avx2(ops);
        if is_x86_feature_detected!("fma") {
            plug_fma(ops);
            if is_x86_feature_detected!("avx512f") {
                plug_avx512f(ops);
                if is_x86_feature_detected!("avx512fp16") {
                    plug_avx512fp16(ops);
                }
            }
        }
    }
}