OXYGENHUB
◆ v0.6.0 oxygen / vendor / crc-fast / src / arch / x86_64 / avx512.rs
5.6 KiB · Rust · last changed Strata: the cloud stores send through ureq, so every release target links 1 month ago source blame history raw
1 
// Copyright 2025 Don MacAskill. Licensed under MIT or Apache-2.0.
2
3 
//! This module provides x86_64-specific implementations of the ArchOps trait.
4 
//!
5 
//! It uses the AVX-512 instruction sets for SIMD operations.
6
7 
#![cfg(target_arch = "x86_64")]
8
9 
use crate::arch::x86::sse::X86SsePclmulqdqOps;
10 
use crate::traits::ArchOps;
11 
use std::arch::x86_64::*;
12
13 
/// x86_64-only AVX512+PCLMULQDQ tier - delegates to SSE tier and overrides XOR3 operations
14 
/// Uses AVX512 ternary logic for XOR3 operations with PCLMULQDQ
15 
#[derive(Debug, Copy, Clone)]
16 
pub struct X86_64Avx512PclmulqdqOps(X86SsePclmulqdqOps);
17
18 
impl Default for X86_64Avx512PclmulqdqOps {
19 
    fn default() -> Self {
20 
        Self::new()
21 
    }
22 
}
23
24 
impl X86_64Avx512PclmulqdqOps {
25 
    #[inline(always)]
26 
    pub fn new() -> Self {
27 
        Self(X86SsePclmulqdqOps)
28 
    }
29 
}
30
31 
impl ArchOps for X86_64Avx512PclmulqdqOps {
32 
    type Vector = __m128i;
33
34 
    // Delegate all methods to the base SSE implementation
35 
    #[inline(always)]
36 
    unsafe fn create_vector_from_u64_pair(
37 
        &self,
38 
        high: u64,
39 
        low: u64,
40 
        reflected: bool,
41 
    ) -> Self::Vector {
42 
        self.0.create_vector_from_u64_pair(high, low, reflected)
43 
    }
44
45 
    #[inline(always)]
46 
    unsafe fn create_vector_from_u64_pair_non_reflected(
47 
        &self,
48 
        high: u64,
49 
        low: u64,
50 
    ) -> Self::Vector {
51 
        self.0.create_vector_from_u64_pair_non_reflected(high, low)
52 
    }
53
54 
    #[inline(always)]
55 
    unsafe fn create_vector_from_u64(&self, value: u64, high: bool) -> Self::Vector {
56 
        self.0.create_vector_from_u64(value, high)
57 
    }
58
59 
    #[inline(always)]
60 
    unsafe fn extract_u64s(&self, vector: Self::Vector) -> [u64; 2] {
61 
        self.0.extract_u64s(vector)
62 
    }
63
64 
    #[inline(always)]
65 
    unsafe fn extract_poly64s(&self, vector: Self::Vector) -> [u64; 2] {
66 
        self.0.extract_poly64s(vector)
67 
    }
68
69 
    #[inline(always)]
70 
    unsafe fn xor_vectors(&self, a: Self::Vector, b: Self::Vector) -> Self::Vector {
71 
        self.0.xor_vectors(a, b)
72 
    }
73
74 
    #[inline(always)]
75 
    unsafe fn load_bytes(&self, ptr: *const u8) -> Self::Vector {
76 
        self.0.load_bytes(ptr)
77 
    }
78
79 
    #[inline(always)]
80 
    unsafe fn load_aligned(&self, ptr: *const [u64; 2]) -> Self::Vector {
81 
        self.0.load_aligned(ptr)
82 
    }
83
84 
    #[inline(always)]
85 
    unsafe fn shuffle_bytes(&self, data: Self::Vector, mask: Self::Vector) -> Self::Vector {
86 
        self.0.shuffle_bytes(data, mask)
87 
    }
88
89 
    #[inline(always)]
90 
    unsafe fn blend_vectors(
91 
        &self,
92 
        a: Self::Vector,
93 
        b: Self::Vector,
94 
        mask: Self::Vector,
95 
    ) -> Self::Vector {
96 
        self.0.blend_vectors(a, b, mask)
97 
    }
98
99 
    #[inline(always)]
100 
    unsafe fn shift_left_8(&self, vector: Self::Vector) -> Self::Vector {
101 
        self.0.shift_left_8(vector)
102 
    }
103
104 
    #[inline(always)]
105 
    unsafe fn set_all_bytes(&self, value: u8) -> Self::Vector {
106 
        self.0.set_all_bytes(value)
107 
    }
108
109 
    #[inline(always)]
110 
    unsafe fn create_compare_mask(&self, vector: Self::Vector) -> Self::Vector {
111 
        self.0.create_compare_mask(vector)
112 
    }
113
114 
    #[inline(always)]
115 
    unsafe fn and_vectors(&self, a: Self::Vector, b: Self::Vector) -> Self::Vector {
116 
        self.0.and_vectors(a, b)
117 
    }
118
119 
    #[inline(always)]
120 
    unsafe fn shift_right_32(&self, vector: Self::Vector) -> Self::Vector {
121 
        self.0.shift_right_32(vector)
122 
    }
123
124 
    #[inline(always)]
125 
    unsafe fn shift_left_32(&self, vector: Self::Vector) -> Self::Vector {
126 
        self.0.shift_left_32(vector)
127 
    }
128
129 
    #[inline(always)]
130 
    unsafe fn create_vector_from_u32(&self, value: u32, high: bool) -> Self::Vector {
131 
        self.0.create_vector_from_u32(value, high)
132 
    }
133
134 
    #[inline(always)]
135 
    unsafe fn shift_left_4(&self, vector: Self::Vector) -> Self::Vector {
136 
        self.0.shift_left_4(vector)
137 
    }
138
139 
    #[inline(always)]
140 
    unsafe fn shift_right_4(&self, vector: Self::Vector) -> Self::Vector {
141 
        self.0.shift_right_4(vector)
142 
    }
143
144 
    #[inline(always)]
145 
    unsafe fn shift_right_8(&self, vector: Self::Vector) -> Self::Vector {
146 
        self.0.shift_right_8(vector)
147 
    }
148
149 
    #[inline(always)]
150 
    unsafe fn shift_right_5(&self, vector: Self::Vector) -> Self::Vector {
151 
        self.0.shift_right_5(vector)
152 
    }
153
154 
    #[inline(always)]
155 
    unsafe fn shift_right_6(&self, vector: Self::Vector) -> Self::Vector {
156 
        self.0.shift_right_6(vector)
157 
    }
158
159 
    #[inline(always)]
160 
    unsafe fn shift_right_7(&self, vector: Self::Vector) -> Self::Vector {
161 
        self.0.shift_right_7(vector)
162 
    }
163
164 
    #[inline(always)]
165 
    unsafe fn shift_right_12(&self, vector: Self::Vector) -> Self::Vector {
166 
        self.0.shift_right_12(vector)
167 
    }
168
169 
    #[inline(always)]
170 
    unsafe fn shift_left_12(&self, vector: Self::Vector) -> Self::Vector {
171 
        self.0.shift_left_12(vector)
172 
    }
173
174 
    #[inline(always)]
175 
    unsafe fn carryless_mul_00(&self, a: Self::Vector, b: Self::Vector) -> Self::Vector {
176 
        self.0.carryless_mul_00(a, b)
177 
    }
178
179 
    #[inline(always)]
180 
    unsafe fn carryless_mul_01(&self, a: Self::Vector, b: Self::Vector) -> Self::Vector {
181 
        self.0.carryless_mul_01(a, b)
182 
    }
183
184 
    #[inline(always)]
185 
    unsafe fn carryless_mul_10(&self, a: Self::Vector, b: Self::Vector) -> Self::Vector {
186 
        self.0.carryless_mul_10(a, b)
187 
    }
188
189 
    #[inline(always)]
190 
    unsafe fn carryless_mul_11(&self, a: Self::Vector, b: Self::Vector) -> Self::Vector {
191 
        self.0.carryless_mul_11(a, b)
192 
    }
193
194 
    #[inline]
195 
    #[target_feature(enable = "avx512vl")]
196 
    unsafe fn xor3_vectors(
197 
        &self,
198 
        a: Self::Vector,
199 
        b: Self::Vector,
200 
        c: Self::Vector,
201 
    ) -> Self::Vector {
202 
        // AVX512 tier always uses ternary logic
203 
        _mm_ternarylogic_epi64(a, b, c, 0x96) // XOR3 operation
204 
    }
205 
}