| 1 |
#![allow(non_camel_case_types)] |
| 2 |
use core::ops::{Add, AddAssign, BitAnd, BitOr, BitXor, BitXorAssign, Not}; |
| 3 |
|
| 4 |
pub trait AndNot { |
| 5 |
type Output; |
| 6 |
fn andnot(self, rhs: Self) -> Self::Output; |
| 7 |
} |
| 8 |
pub trait BSwap { |
| 9 |
fn bswap(self) -> Self; |
| 10 |
} |
| 11 |
/// Ops that depend on word size |
| 12 |
pub trait ArithOps: Add<Output = Self> + AddAssign + Sized + Copy + Clone + BSwap {} |
| 13 |
/// Ops that are independent of word size and endian |
| 14 |
pub trait BitOps0: |
| 15 |
BitAnd<Output = Self> |
| 16 |
+ BitOr<Output = Self> |
| 17 |
+ BitXor<Output = Self> |
| 18 |
+ BitXorAssign |
| 19 |
+ Not<Output = Self> |
| 20 |
+ AndNot<Output = Self> |
| 21 |
+ Sized |
| 22 |
+ Copy |
| 23 |
+ Clone |
| 24 |
{ |
| 25 |
} |
| 26 |
|
| 27 |
pub trait BitOps32: BitOps0 + RotateEachWord32 {} |
| 28 |
pub trait BitOps64: BitOps32 + RotateEachWord64 {} |
| 29 |
pub trait BitOps128: BitOps64 + RotateEachWord128 {} |
| 30 |
|
| 31 |
pub trait RotateEachWord32 { |
| 32 |
fn rotate_each_word_right7(self) -> Self; |
| 33 |
fn rotate_each_word_right8(self) -> Self; |
| 34 |
fn rotate_each_word_right11(self) -> Self; |
| 35 |
fn rotate_each_word_right12(self) -> Self; |
| 36 |
fn rotate_each_word_right16(self) -> Self; |
| 37 |
fn rotate_each_word_right20(self) -> Self; |
| 38 |
fn rotate_each_word_right24(self) -> Self; |
| 39 |
fn rotate_each_word_right25(self) -> Self; |
| 40 |
} |
| 41 |
|
| 42 |
pub trait RotateEachWord64 { |
| 43 |
fn rotate_each_word_right32(self) -> Self; |
| 44 |
} |
| 45 |
|
| 46 |
pub trait RotateEachWord128 {} |
| 47 |
|
| 48 |
// Vector type naming scheme: |
| 49 |
// uN[xP]xL |
| 50 |
// Unsigned; N-bit words * P bits per lane * L lanes |
| 51 |
// |
| 52 |
// A lane is always 128-bits, chosen because common SIMD architectures treat 128-bit units of |
| 53 |
// wide vectors specially (supporting e.g. intra-lane shuffles), and tend to have limited and |
| 54 |
// slow inter-lane operations. |
| 55 |
|
| 56 |
use crate::arch::{vec128_storage, vec256_storage, vec512_storage}; |
| 57 |
|
| 58 |
#[allow(clippy::missing_safety_doc)] |
| 59 |
pub trait UnsafeFrom<T> { |
| 60 |
unsafe fn unsafe_from(t: T) -> Self; |
| 61 |
} |
| 62 |
|
| 63 |
/// A vector composed of two elements, which may be words or themselves vectors. |
| 64 |
pub trait Vec2<W> { |
| 65 |
fn extract(self, i: u32) -> W; |
| 66 |
fn insert(self, w: W, i: u32) -> Self; |
| 67 |
} |
| 68 |
|
| 69 |
/// A vector composed of four elements, which may be words or themselves vectors. |
| 70 |
pub trait Vec4<W> { |
| 71 |
fn extract(self, i: u32) -> W; |
| 72 |
fn insert(self, w: W, i: u32) -> Self; |
| 73 |
} |
| 74 |
/// Vec4 functions which may not be implemented yet for all Vec4 types. |
| 75 |
/// NOTE: functions in this trait may be moved to Vec4 in any patch release. To avoid breakage, |
| 76 |
/// import Vec4Ext only together with Vec4, and don't qualify its methods. |
| 77 |
pub trait Vec4Ext<W> { |
| 78 |
fn transpose4(a: Self, b: Self, c: Self, d: Self) -> (Self, Self, Self, Self) |
| 79 |
where |
| 80 |
Self: Sized; |
| 81 |
} |
| 82 |
pub trait Vector<T> { |
| 83 |
fn to_scalars(self) -> T; |
| 84 |
} |
| 85 |
|
| 86 |
// TODO: multiples of 4 should inherit this |
| 87 |
/// A vector composed of four words; depending on their size, operations may cross lanes. |
| 88 |
pub trait Words4 { |
| 89 |
fn shuffle1230(self) -> Self; |
| 90 |
fn shuffle2301(self) -> Self; |
| 91 |
fn shuffle3012(self) -> Self; |
| 92 |
} |
| 93 |
|
| 94 |
/// A vector composed one or more lanes each composed of four words. |
| 95 |
pub trait LaneWords4 { |
| 96 |
fn shuffle_lane_words1230(self) -> Self; |
| 97 |
fn shuffle_lane_words2301(self) -> Self; |
| 98 |
fn shuffle_lane_words3012(self) -> Self; |
| 99 |
} |
| 100 |
|
| 101 |
// TODO: make this a part of BitOps |
| 102 |
/// Exchange neigboring ranges of bits of the specified size |
| 103 |
pub trait Swap64 { |
| 104 |
fn swap1(self) -> Self; |
| 105 |
fn swap2(self) -> Self; |
| 106 |
fn swap4(self) -> Self; |
| 107 |
fn swap8(self) -> Self; |
| 108 |
fn swap16(self) -> Self; |
| 109 |
fn swap32(self) -> Self; |
| 110 |
fn swap64(self) -> Self; |
| 111 |
} |
| 112 |
|
| 113 |
pub trait u32x4<M: Machine>: |
| 114 |
BitOps32 |
| 115 |
+ Store<vec128_storage> |
| 116 |
+ ArithOps |
| 117 |
+ Vec4<u32> |
| 118 |
+ Words4 |
| 119 |
+ LaneWords4 |
| 120 |
+ StoreBytes |
| 121 |
+ MultiLane<[u32; 4]> |
| 122 |
+ Into<vec128_storage> |
| 123 |
{ |
| 124 |
} |
| 125 |
pub trait u64x2<M: Machine>: |
| 126 |
BitOps64 + Store<vec128_storage> + ArithOps + Vec2<u64> + MultiLane<[u64; 2]> + Into<vec128_storage> |
| 127 |
{ |
| 128 |
} |
| 129 |
pub trait u128x1<M: Machine>: |
| 130 |
BitOps128 + Store<vec128_storage> + Swap64 + MultiLane<[u128; 1]> + Into<vec128_storage> |
| 131 |
{ |
| 132 |
} |
| 133 |
|
| 134 |
pub trait u32x4x2<M: Machine>: |
| 135 |
BitOps32 |
| 136 |
+ Store<vec256_storage> |
| 137 |
+ Vec2<M::u32x4> |
| 138 |
+ MultiLane<[M::u32x4; 2]> |
| 139 |
+ ArithOps |
| 140 |
+ Into<vec256_storage> |
| 141 |
+ StoreBytes |
| 142 |
{ |
| 143 |
} |
| 144 |
pub trait u64x2x2<M: Machine>: |
| 145 |
BitOps64 |
| 146 |
+ Store<vec256_storage> |
| 147 |
+ Vec2<M::u64x2> |
| 148 |
+ MultiLane<[M::u64x2; 2]> |
| 149 |
+ ArithOps |
| 150 |
+ StoreBytes |
| 151 |
+ Into<vec256_storage> |
| 152 |
{ |
| 153 |
} |
| 154 |
pub trait u64x4<M: Machine>: |
| 155 |
BitOps64 |
| 156 |
+ Store<vec256_storage> |
| 157 |
+ Vec4<u64> |
| 158 |
+ MultiLane<[u64; 4]> |
| 159 |
+ ArithOps |
| 160 |
+ Words4 |
| 161 |
+ StoreBytes |
| 162 |
+ Into<vec256_storage> |
| 163 |
{ |
| 164 |
} |
| 165 |
pub trait u128x2<M: Machine>: |
| 166 |
BitOps128 |
| 167 |
+ Store<vec256_storage> |
| 168 |
+ Vec2<M::u128x1> |
| 169 |
+ MultiLane<[M::u128x1; 2]> |
| 170 |
+ Swap64 |
| 171 |
+ Into<vec256_storage> |
| 172 |
{ |
| 173 |
} |
| 174 |
|
| 175 |
pub trait u32x4x4<M: Machine>: |
| 176 |
BitOps32 |
| 177 |
+ Store<vec512_storage> |
| 178 |
+ Vec4<M::u32x4> |
| 179 |
+ Vec4Ext<M::u32x4> |
| 180 |
+ Vector<[u32; 16]> |
| 181 |
+ MultiLane<[M::u32x4; 4]> |
| 182 |
+ ArithOps |
| 183 |
+ LaneWords4 |
| 184 |
+ Into<vec512_storage> |
| 185 |
+ StoreBytes |
| 186 |
{ |
| 187 |
} |
| 188 |
pub trait u64x2x4<M: Machine>: |
| 189 |
BitOps64 |
| 190 |
+ Store<vec512_storage> |
| 191 |
+ Vec4<M::u64x2> |
| 192 |
+ MultiLane<[M::u64x2; 4]> |
| 193 |
+ ArithOps |
| 194 |
+ Into<vec512_storage> |
| 195 |
{ |
| 196 |
} |
| 197 |
// TODO: Words4 |
| 198 |
pub trait u128x4<M: Machine>: |
| 199 |
BitOps128 |
| 200 |
+ Store<vec512_storage> |
| 201 |
+ Vec4<M::u128x1> |
| 202 |
+ MultiLane<[M::u128x1; 4]> |
| 203 |
+ Swap64 |
| 204 |
+ Into<vec512_storage> |
| 205 |
{ |
| 206 |
} |
| 207 |
|
| 208 |
/// A vector composed of multiple 128-bit lanes. |
| 209 |
pub trait MultiLane<Lanes> { |
| 210 |
/// Split a multi-lane vector into single-lane vectors. |
| 211 |
fn to_lanes(self) -> Lanes; |
| 212 |
/// Build a multi-lane vector from individual lanes. |
| 213 |
fn from_lanes(lanes: Lanes) -> Self; |
| 214 |
} |
| 215 |
|
| 216 |
/// Combine single vectors into a multi-lane vector. |
| 217 |
pub trait VZip<V> { |
| 218 |
fn vzip(self) -> V; |
| 219 |
} |
| 220 |
|
| 221 |
impl<V, T> VZip<V> for T |
| 222 |
where |
| 223 |
V: MultiLane<T>, |
| 224 |
{ |
| 225 |
#[inline(always)] |
| 226 |
fn vzip(self) -> V { |
| 227 |
V::from_lanes(self) |
| 228 |
} |
| 229 |
} |
| 230 |
|
| 231 |
pub trait Machine: Sized + Copy { |
| 232 |
type u32x4: u32x4<Self>; |
| 233 |
type u64x2: u64x2<Self>; |
| 234 |
type u128x1: u128x1<Self>; |
| 235 |
|
| 236 |
type u32x4x2: u32x4x2<Self>; |
| 237 |
type u64x2x2: u64x2x2<Self>; |
| 238 |
type u64x4: u64x4<Self>; |
| 239 |
type u128x2: u128x2<Self>; |
| 240 |
|
| 241 |
type u32x4x4: u32x4x4<Self>; |
| 242 |
type u64x2x4: u64x2x4<Self>; |
| 243 |
type u128x4: u128x4<Self>; |
| 244 |
|
| 245 |
#[inline(always)] |
| 246 |
fn unpack<S, V: Store<S>>(self, s: S) -> V { |
| 247 |
unsafe { V::unpack(s) } |
| 248 |
} |
| 249 |
|
| 250 |
#[inline(always)] |
| 251 |
fn vec<V, A>(self, a: A) -> V |
| 252 |
where |
| 253 |
V: MultiLane<A>, |
| 254 |
{ |
| 255 |
V::from_lanes(a) |
| 256 |
} |
| 257 |
|
| 258 |
#[inline(always)] |
| 259 |
fn read_le<V>(self, input: &[u8]) -> V |
| 260 |
where |
| 261 |
V: StoreBytes, |
| 262 |
{ |
| 263 |
unsafe { V::unsafe_read_le(input) } |
| 264 |
} |
| 265 |
|
| 266 |
#[inline(always)] |
| 267 |
fn read_be<V>(self, input: &[u8]) -> V |
| 268 |
where |
| 269 |
V: StoreBytes, |
| 270 |
{ |
| 271 |
unsafe { V::unsafe_read_be(input) } |
| 272 |
} |
| 273 |
|
| 274 |
/// # Safety |
| 275 |
/// Caller must ensure the type of Self is appropriate for the hardware of the execution |
| 276 |
/// environment. |
| 277 |
unsafe fn instance() -> Self; |
| 278 |
} |
| 279 |
|
| 280 |
pub trait Store<S> { |
| 281 |
/// # Safety |
| 282 |
/// Caller must ensure the type of Self is appropriate for the hardware of the execution |
| 283 |
/// environment. |
| 284 |
unsafe fn unpack(p: S) -> Self; |
| 285 |
} |
| 286 |
|
| 287 |
pub trait StoreBytes { |
| 288 |
/// # Safety |
| 289 |
/// Caller must ensure the type of Self is appropriate for the hardware of the execution |
| 290 |
/// environment. |
| 291 |
unsafe fn unsafe_read_le(input: &[u8]) -> Self; |
| 292 |
/// # Safety |
| 293 |
/// Caller must ensure the type of Self is appropriate for the hardware of the execution |
| 294 |
/// environment. |
| 295 |
unsafe fn unsafe_read_be(input: &[u8]) -> Self; |
| 296 |
fn write_le(self, out: &mut [u8]); |
| 297 |
fn write_be(self, out: &mut [u8]); |
| 298 |
} |