| 1 |
// NOTE: The descriptions for each of the vector methods on the traits below |
| 2 |
// are pretty inscrutable. For this reason, there are tests for every method |
| 3 |
// on for every trait impl below. If you're confused about what an op does, |
| 4 |
// consult its test. (They probably should be doc tests, but I couldn't figure |
| 5 |
// out how to write them in a non-annoying way.) |
| 6 |
|
| 7 |
use core::{ |
| 8 |
fmt::Debug, |
| 9 |
panic::{RefUnwindSafe, UnwindSafe}, |
| 10 |
}; |
| 11 |
|
| 12 |
/// A trait for describing vector operations used by vectorized searchers. |
| 13 |
/// |
| 14 |
/// The trait is highly constrained to low level vector operations needed for |
| 15 |
/// the specific algorithms used in this crate. In general, it was invented |
| 16 |
/// mostly to be generic over x86's __m128i and __m256i types. At time of |
| 17 |
/// writing, it also supports wasm and aarch64 128-bit vector types as well. |
| 18 |
/// |
| 19 |
/// # Safety |
| 20 |
/// |
| 21 |
/// All methods are not safe since they are intended to be implemented using |
| 22 |
/// vendor intrinsics, which are also not safe. Callers must ensure that |
| 23 |
/// the appropriate target features are enabled in the calling function, |
| 24 |
/// and that the current CPU supports them. All implementations should |
| 25 |
/// avoid marking the routines with `#[target_feature]` and instead mark |
| 26 |
/// them as `#[inline(always)]` to ensure they get appropriately inlined. |
| 27 |
/// (`inline(always)` cannot be used with target_feature.) |
| 28 |
pub(crate) trait Vector: |
| 29 |
Copy + Debug + Send + Sync + UnwindSafe + RefUnwindSafe |
| 30 |
{ |
| 31 |
/// The number of bits in the vector. |
| 32 |
const BITS: usize; |
| 33 |
/// The number of bytes in the vector. That is, this is the size of the |
| 34 |
/// vector in memory. |
| 35 |
const BYTES: usize; |
| 36 |
|
| 37 |
/// Create a vector with 8-bit lanes with the given byte repeated into each |
| 38 |
/// lane. |
| 39 |
/// |
| 40 |
/// # Safety |
| 41 |
/// |
| 42 |
/// Callers must ensure that this is okay to call in the current target for |
| 43 |
/// the current CPU. |
| 44 |
unsafe fn splat(byte: u8) -> Self; |
| 45 |
|
| 46 |
/// Read a vector-size number of bytes from the given pointer. The pointer |
| 47 |
/// does not need to be aligned. |
| 48 |
/// |
| 49 |
/// # Safety |
| 50 |
/// |
| 51 |
/// Callers must ensure that this is okay to call in the current target for |
| 52 |
/// the current CPU. |
| 53 |
/// |
| 54 |
/// Callers must guarantee that at least `BYTES` bytes are readable from |
| 55 |
/// `data`. |
| 56 |
unsafe fn load_unaligned(data: *const u8) -> Self; |
| 57 |
|
| 58 |
/// Returns true if and only if this vector has zero in all of its lanes. |
| 59 |
/// |
| 60 |
/// # Safety |
| 61 |
/// |
| 62 |
/// Callers must ensure that this is okay to call in the current target for |
| 63 |
/// the current CPU. |
| 64 |
unsafe fn is_zero(self) -> bool; |
| 65 |
|
| 66 |
/// Do an 8-bit pairwise equality check. If lane `i` is equal in this |
| 67 |
/// vector and the one given, then lane `i` in the resulting vector is set |
| 68 |
/// to `0xFF`. Otherwise, it is set to `0x00`. |
| 69 |
/// |
| 70 |
/// # Safety |
| 71 |
/// |
| 72 |
/// Callers must ensure that this is okay to call in the current target for |
| 73 |
/// the current CPU. |
| 74 |
unsafe fn cmpeq(self, vector2: Self) -> Self; |
| 75 |
|
| 76 |
/// Perform a bitwise 'and' of this vector and the one given and return |
| 77 |
/// the result. |
| 78 |
/// |
| 79 |
/// # Safety |
| 80 |
/// |
| 81 |
/// Callers must ensure that this is okay to call in the current target for |
| 82 |
/// the current CPU. |
| 83 |
unsafe fn and(self, vector2: Self) -> Self; |
| 84 |
|
| 85 |
/// Perform a bitwise 'or' of this vector and the one given and return |
| 86 |
/// the result. |
| 87 |
/// |
| 88 |
/// # Safety |
| 89 |
/// |
| 90 |
/// Callers must ensure that this is okay to call in the current target for |
| 91 |
/// the current CPU. |
| 92 |
#[allow(dead_code)] // unused, but useful enough to keep around? |
| 93 |
unsafe fn or(self, vector2: Self) -> Self; |
| 94 |
|
| 95 |
/// Shift each 8-bit lane in this vector to the right by the number of |
| 96 |
/// bits indictated by the `BITS` type parameter. |
| 97 |
/// |
| 98 |
/// # Safety |
| 99 |
/// |
| 100 |
/// Callers must ensure that this is okay to call in the current target for |
| 101 |
/// the current CPU. |
| 102 |
unsafe fn shift_8bit_lane_right<const BITS: i32>(self) -> Self; |
| 103 |
|
| 104 |
/// Shift this vector to the left by one byte and shift the most |
| 105 |
/// significant byte of `vector2` into the least significant position of |
| 106 |
/// this vector. |
| 107 |
/// |
| 108 |
/// Stated differently, this behaves as if `self` and `vector2` were |
| 109 |
/// concatenated into a `2 * Self::BITS` temporary buffer and then shifted |
| 110 |
/// right by `Self::BYTES - 1` bytes. |
| 111 |
/// |
| 112 |
/// With respect to the Teddy algorithm, `vector2` is usually a previous |
| 113 |
/// `Self::BYTES` chunk from the haystack and `self` is the chunk |
| 114 |
/// immediately following it. This permits combining the last two bytes |
| 115 |
/// from the previous chunk (`vector2`) with the first `Self::BYTES - 1` |
| 116 |
/// bytes from the current chunk. This permits aligning the result of |
| 117 |
/// various shuffles so that they can be and-ed together and a possible |
| 118 |
/// candidate discovered. |
| 119 |
/// |
| 120 |
/// # Safety |
| 121 |
/// |
| 122 |
/// Callers must ensure that this is okay to call in the current target for |
| 123 |
/// the current CPU. |
| 124 |
unsafe fn shift_in_one_byte(self, vector2: Self) -> Self; |
| 125 |
|
| 126 |
/// Shift this vector to the left by two bytes and shift the two most |
| 127 |
/// significant bytes of `vector2` into the least significant position of |
| 128 |
/// this vector. |
| 129 |
/// |
| 130 |
/// Stated differently, this behaves as if `self` and `vector2` were |
| 131 |
/// concatenated into a `2 * Self::BITS` temporary buffer and then shifted |
| 132 |
/// right by `Self::BYTES - 2` bytes. |
| 133 |
/// |
| 134 |
/// With respect to the Teddy algorithm, `vector2` is usually a previous |
| 135 |
/// `Self::BYTES` chunk from the haystack and `self` is the chunk |
| 136 |
/// immediately following it. This permits combining the last two bytes |
| 137 |
/// from the previous chunk (`vector2`) with the first `Self::BYTES - 2` |
| 138 |
/// bytes from the current chunk. This permits aligning the result of |
| 139 |
/// various shuffles so that they can be and-ed together and a possible |
| 140 |
/// candidate discovered. |
| 141 |
/// |
| 142 |
/// # Safety |
| 143 |
/// |
| 144 |
/// Callers must ensure that this is okay to call in the current target for |
| 145 |
/// the current CPU. |
| 146 |
unsafe fn shift_in_two_bytes(self, vector2: Self) -> Self; |
| 147 |
|
| 148 |
/// Shift this vector to the left by three bytes and shift the three most |
| 149 |
/// significant bytes of `vector2` into the least significant position of |
| 150 |
/// this vector. |
| 151 |
/// |
| 152 |
/// Stated differently, this behaves as if `self` and `vector2` were |
| 153 |
/// concatenated into a `2 * Self::BITS` temporary buffer and then shifted |
| 154 |
/// right by `Self::BYTES - 3` bytes. |
| 155 |
/// |
| 156 |
/// With respect to the Teddy algorithm, `vector2` is usually a previous |
| 157 |
/// `Self::BYTES` chunk from the haystack and `self` is the chunk |
| 158 |
/// immediately following it. This permits combining the last three bytes |
| 159 |
/// from the previous chunk (`vector2`) with the first `Self::BYTES - 3` |
| 160 |
/// bytes from the current chunk. This permits aligning the result of |
| 161 |
/// various shuffles so that they can be and-ed together and a possible |
| 162 |
/// candidate discovered. |
| 163 |
/// |
| 164 |
/// # Safety |
| 165 |
/// |
| 166 |
/// Callers must ensure that this is okay to call in the current target for |
| 167 |
/// the current CPU. |
| 168 |
unsafe fn shift_in_three_bytes(self, vector2: Self) -> Self; |
| 169 |
|
| 170 |
/// Shuffles the bytes in this vector according to the indices in each of |
| 171 |
/// the corresponding lanes in `indices`. |
| 172 |
/// |
| 173 |
/// If `i` is the index of corresponding lanes, `A` is this vector, `B` is |
| 174 |
/// indices and `C` is the resulting vector, then `C = A[B[i]]`. |
| 175 |
/// |
| 176 |
/// # Safety |
| 177 |
/// |
| 178 |
/// Callers must ensure that this is okay to call in the current target for |
| 179 |
/// the current CPU. |
| 180 |
unsafe fn shuffle_bytes(self, indices: Self) -> Self; |
| 181 |
|
| 182 |
/// Call the provided function for each 64-bit lane in this vector. The |
| 183 |
/// given function is provided the lane index and lane value as a `u64`. |
| 184 |
/// |
| 185 |
/// If `f` returns `Some`, then iteration over the lanes is stopped and the |
| 186 |
/// value is returned. Otherwise, this returns `None`. |
| 187 |
/// |
| 188 |
/// # Notes |
| 189 |
/// |
| 190 |
/// Conceptually it would be nice if we could have a |
| 191 |
/// `unpack64(self) -> [u64; BITS / 64]` method, but defining that is |
| 192 |
/// tricky given Rust's [current support for const generics][support]. |
| 193 |
/// And even if we could, it would be tricky to write generic code over |
| 194 |
/// it. (Not impossible. We could introduce another layer that requires |
| 195 |
/// `AsRef<[u64]>` or something.) |
| 196 |
/// |
| 197 |
/// [support]: https://github.com/rust-lang/rust/issues/60551 |
| 198 |
/// |
| 199 |
/// # Safety |
| 200 |
/// |
| 201 |
/// Callers must ensure that this is okay to call in the current target for |
| 202 |
/// the current CPU. |
| 203 |
unsafe fn for_each_64bit_lane<T>( |
| 204 |
self, |
| 205 |
f: impl FnMut(usize, u64) -> Option<T>, |
| 206 |
) -> Option<T>; |
| 207 |
} |
| 208 |
|
| 209 |
/// This trait extends the `Vector` trait with additional operations to support |
| 210 |
/// Fat Teddy. |
| 211 |
/// |
| 212 |
/// Fat Teddy uses 16 buckets instead of 8, but reads half as many bytes (as |
| 213 |
/// the vector size) instead of the full size of a vector per iteration. For |
| 214 |
/// example, when using a 256-bit vector, Slim Teddy reads 32 bytes at a timr |
| 215 |
/// but Fat Teddy reads 16 bytes at a time. |
| 216 |
/// |
| 217 |
/// Fat Teddy is useful when searching for a large number of literals. |
| 218 |
/// The extra number of buckets spreads the literals out more and reduces |
| 219 |
/// verification time. |
| 220 |
/// |
| 221 |
/// Currently we only implement this for AVX on x86_64. It would be nice to |
| 222 |
/// implement this for SSE on x86_64 and NEON on aarch64, with the latter two |
| 223 |
/// only reading 8 bytes at a time. It's not clear how well it would work, but |
| 224 |
/// there are some tricky things to figure out in terms of implementation. The |
| 225 |
/// `half_shift_in_{one,two,three}_bytes` methods in particular are probably |
| 226 |
/// the trickiest of the bunch. For AVX2, these are implemented by taking |
| 227 |
/// advantage of the fact that `_mm256_alignr_epi8` operates on each 128-bit |
| 228 |
/// half instead of the full 256-bit vector. (Where as `_mm_alignr_epi8` |
| 229 |
/// operates on the full 128-bit vector and not on each 64-bit half.) I didn't |
| 230 |
/// do a careful survey of NEON to see if it could easily support these |
| 231 |
/// operations. |
| 232 |
pub(crate) trait FatVector: Vector { |
| 233 |
type Half: Vector; |
| 234 |
|
| 235 |
/// Read a half-vector-size number of bytes from the given pointer, and |
| 236 |
/// broadcast it across both halfs of a full vector. The pointer does not |
| 237 |
/// need to be aligned. |
| 238 |
/// |
| 239 |
/// # Safety |
| 240 |
/// |
| 241 |
/// Callers must ensure that this is okay to call in the current target for |
| 242 |
/// the current CPU. |
| 243 |
/// |
| 244 |
/// Callers must guarantee that at least `Self::HALF::BYTES` bytes are |
| 245 |
/// readable from `data`. |
| 246 |
unsafe fn load_half_unaligned(data: *const u8) -> Self; |
| 247 |
|
| 248 |
/// Like `Vector::shift_in_one_byte`, except this is done for each half |
| 249 |
/// of the vector instead. |
| 250 |
/// |
| 251 |
/// # Safety |
| 252 |
/// |
| 253 |
/// Callers must ensure that this is okay to call in the current target for |
| 254 |
/// the current CPU. |
| 255 |
unsafe fn half_shift_in_one_byte(self, vector2: Self) -> Self; |
| 256 |
|
| 257 |
/// Like `Vector::shift_in_two_bytes`, except this is done for each half |
| 258 |
/// of the vector instead. |
| 259 |
/// |
| 260 |
/// # Safety |
| 261 |
/// |
| 262 |
/// Callers must ensure that this is okay to call in the current target for |
| 263 |
/// the current CPU. |
| 264 |
unsafe fn half_shift_in_two_bytes(self, vector2: Self) -> Self; |
| 265 |
|
| 266 |
/// Like `Vector::shift_in_two_bytes`, except this is done for each half |
| 267 |
/// of the vector instead. |
| 268 |
/// |
| 269 |
/// # Safety |
| 270 |
/// |
| 271 |
/// Callers must ensure that this is okay to call in the current target for |
| 272 |
/// the current CPU. |
| 273 |
unsafe fn half_shift_in_three_bytes(self, vector2: Self) -> Self; |
| 274 |
|
| 275 |
/// Swap the 128-bit lanes in this vector. |
| 276 |
/// |
| 277 |
/// # Safety |
| 278 |
/// |
| 279 |
/// Callers must ensure that this is okay to call in the current target for |
| 280 |
/// the current CPU. |
| 281 |
unsafe fn swap_halves(self) -> Self; |
| 282 |
|
| 283 |
/// Unpack and interleave the 8-bit lanes from the low 128 bits of each |
| 284 |
/// vector and return the result. |
| 285 |
/// |
| 286 |
/// # Safety |
| 287 |
/// |
| 288 |
/// Callers must ensure that this is okay to call in the current target for |
| 289 |
/// the current CPU. |
| 290 |
unsafe fn interleave_low_8bit_lanes(self, vector2: Self) -> Self; |
| 291 |
|
| 292 |
/// Unpack and interleave the 8-bit lanes from the high 128 bits of each |
| 293 |
/// vector and return the result. |
| 294 |
/// |
| 295 |
/// # Safety |
| 296 |
/// |
| 297 |
/// Callers must ensure that this is okay to call in the current target for |
| 298 |
/// the current CPU. |
| 299 |
unsafe fn interleave_high_8bit_lanes(self, vector2: Self) -> Self; |
| 300 |
|
| 301 |
/// Call the provided function for each 64-bit lane in the lower half |
| 302 |
/// of this vector and then in the other vector. The given function is |
| 303 |
/// provided the lane index and lane value as a `u64`. (The high 128-bits |
| 304 |
/// of each vector are ignored.) |
| 305 |
/// |
| 306 |
/// If `f` returns `Some`, then iteration over the lanes is stopped and the |
| 307 |
/// value is returned. Otherwise, this returns `None`. |
| 308 |
/// |
| 309 |
/// # Safety |
| 310 |
/// |
| 311 |
/// Callers must ensure that this is okay to call in the current target for |
| 312 |
/// the current CPU. |
| 313 |
unsafe fn for_each_low_64bit_lane<T>( |
| 314 |
self, |
| 315 |
vector2: Self, |
| 316 |
f: impl FnMut(usize, u64) -> Option<T>, |
| 317 |
) -> Option<T>; |
| 318 |
} |
| 319 |
|
| 320 |
#[cfg(all(target_arch = "x86_64", target_feature = "sse2"))] |
| 321 |
mod x86_64_ssse3 { |
| 322 |
use core::arch::x86_64::*; |
| 323 |
|
| 324 |
use crate::util::int::{I32, I8}; |
| 325 |
|
| 326 |
use super::Vector; |
| 327 |
|
| 328 |
impl Vector for __m128i { |
| 329 |
const BITS: usize = 128; |
| 330 |
const BYTES: usize = 16; |
| 331 |
|
| 332 |
#[inline(always)] |
| 333 |
unsafe fn splat(byte: u8) -> __m128i { |
| 334 |
_mm_set1_epi8(i8::from_bits(byte)) |
| 335 |
} |
| 336 |
|
| 337 |
#[inline(always)] |
| 338 |
unsafe fn load_unaligned(data: *const u8) -> __m128i { |
| 339 |
_mm_loadu_si128(data.cast::<__m128i>()) |
| 340 |
} |
| 341 |
|
| 342 |
#[inline(always)] |
| 343 |
unsafe fn is_zero(self) -> bool { |
| 344 |
let cmp = self.cmpeq(Self::splat(0)); |
| 345 |
_mm_movemask_epi8(cmp).to_bits() == 0xFFFF |
| 346 |
} |
| 347 |
|
| 348 |
#[inline(always)] |
| 349 |
unsafe fn cmpeq(self, vector2: Self) -> __m128i { |
| 350 |
_mm_cmpeq_epi8(self, vector2) |
| 351 |
} |
| 352 |
|
| 353 |
#[inline(always)] |
| 354 |
unsafe fn and(self, vector2: Self) -> __m128i { |
| 355 |
_mm_and_si128(self, vector2) |
| 356 |
} |
| 357 |
|
| 358 |
#[inline(always)] |
| 359 |
unsafe fn or(self, vector2: Self) -> __m128i { |
| 360 |
_mm_or_si128(self, vector2) |
| 361 |
} |
| 362 |
|
| 363 |
#[inline(always)] |
| 364 |
unsafe fn shift_8bit_lane_right<const BITS: i32>(self) -> Self { |
| 365 |
// Apparently there is no _mm_srli_epi8, so we emulate it by |
| 366 |
// shifting 16-bit integers and masking out the high nybble of each |
| 367 |
// 8-bit lane (since that nybble will contain bits from the low |
| 368 |
// nybble of the previous lane). |
| 369 |
let lomask = Self::splat(0xF); |
| 370 |
_mm_srli_epi16(self, BITS).and(lomask) |
| 371 |
} |
| 372 |
|
| 373 |
#[inline(always)] |
| 374 |
unsafe fn shift_in_one_byte(self, vector2: Self) -> Self { |
| 375 |
_mm_alignr_epi8(self, vector2, 15) |
| 376 |
} |
| 377 |
|
| 378 |
#[inline(always)] |
| 379 |
unsafe fn shift_in_two_bytes(self, vector2: Self) -> Self { |
| 380 |
_mm_alignr_epi8(self, vector2, 14) |
| 381 |
} |
| 382 |
|
| 383 |
#[inline(always)] |
| 384 |
unsafe fn shift_in_three_bytes(self, vector2: Self) -> Self { |
| 385 |
_mm_alignr_epi8(self, vector2, 13) |
| 386 |
} |
| 387 |
|
| 388 |
#[inline(always)] |
| 389 |
unsafe fn shuffle_bytes(self, indices: Self) -> Self { |
| 390 |
_mm_shuffle_epi8(self, indices) |
| 391 |
} |
| 392 |
|
| 393 |
#[inline(always)] |
| 394 |
unsafe fn for_each_64bit_lane<T>( |
| 395 |
self, |
| 396 |
mut f: impl FnMut(usize, u64) -> Option<T>, |
| 397 |
) -> Option<T> { |
| 398 |
// We could just use _mm_extract_epi64 here, but that requires |
| 399 |
// SSE 4.1. It isn't necessarily a problem to just require SSE 4.1, |
| 400 |
// but everything else works with SSSE3 so we stick to that subset. |
| 401 |
let lanes: [u64; 2] = core::mem::transmute(self); |
| 402 |
if let Some(t) = f(0, lanes[0]) { |
| 403 |
return Some(t); |
| 404 |
} |
| 405 |
if let Some(t) = f(1, lanes[1]) { |
| 406 |
return Some(t); |
| 407 |
} |
| 408 |
None |
| 409 |
} |
| 410 |
} |
| 411 |
} |
| 412 |
|
| 413 |
#[cfg(all(target_arch = "x86_64", target_feature = "sse2"))] |
| 414 |
mod x86_64_avx2 { |
| 415 |
use core::arch::x86_64::*; |
| 416 |
|
| 417 |
use crate::util::int::{I32, I64, I8}; |
| 418 |
|
| 419 |
use super::{FatVector, Vector}; |
| 420 |
|
| 421 |
impl Vector for __m256i { |
| 422 |
const BITS: usize = 256; |
| 423 |
const BYTES: usize = 32; |
| 424 |
|
| 425 |
#[inline(always)] |
| 426 |
unsafe fn splat(byte: u8) -> __m256i { |
| 427 |
_mm256_set1_epi8(i8::from_bits(byte)) |
| 428 |
} |
| 429 |
|
| 430 |
#[inline(always)] |
| 431 |
unsafe fn load_unaligned(data: *const u8) -> __m256i { |
| 432 |
_mm256_loadu_si256(data.cast::<__m256i>()) |
| 433 |
} |
| 434 |
|
| 435 |
#[inline(always)] |
| 436 |
unsafe fn is_zero(self) -> bool { |
| 437 |
let cmp = self.cmpeq(Self::splat(0)); |
| 438 |
_mm256_movemask_epi8(cmp).to_bits() == 0xFFFFFFFF |
| 439 |
} |
| 440 |
|
| 441 |
#[inline(always)] |
| 442 |
unsafe fn cmpeq(self, vector2: Self) -> __m256i { |
| 443 |
_mm256_cmpeq_epi8(self, vector2) |
| 444 |
} |
| 445 |
|
| 446 |
#[inline(always)] |
| 447 |
unsafe fn and(self, vector2: Self) -> __m256i { |
| 448 |
_mm256_and_si256(self, vector2) |
| 449 |
} |
| 450 |
|
| 451 |
#[inline(always)] |
| 452 |
unsafe fn or(self, vector2: Self) -> __m256i { |
| 453 |
_mm256_or_si256(self, vector2) |
| 454 |
} |
| 455 |
|
| 456 |
#[inline(always)] |
| 457 |
unsafe fn shift_8bit_lane_right<const BITS: i32>(self) -> Self { |
| 458 |
let lomask = Self::splat(0xF); |
| 459 |
_mm256_srli_epi16(self, BITS).and(lomask) |
| 460 |
} |
| 461 |
|
| 462 |
#[inline(always)] |
| 463 |
unsafe fn shift_in_one_byte(self, vector2: Self) -> Self { |
| 464 |
// Credit goes to jneem for figuring this out: |
| 465 |
// https://github.com/jneem/teddy/blob/9ab5e899ad6ef6911aecd3cf1033f1abe6e1f66c/src/x86/teddy_simd.rs#L145-L184 |
| 466 |
// |
| 467 |
// TL;DR avx2's PALIGNR instruction is actually just two 128-bit |
| 468 |
// PALIGNR instructions, which is not what we want, so we need to |
| 469 |
// do some extra shuffling. |
| 470 |
let v = _mm256_permute2x128_si256(vector2, self, 0x21); |
| 471 |
_mm256_alignr_epi8(self, v, 15) |
| 472 |
} |
| 473 |
|
| 474 |
#[inline(always)] |
| 475 |
unsafe fn shift_in_two_bytes(self, vector2: Self) -> Self { |
| 476 |
// Credit goes to jneem for figuring this out: |
| 477 |
// https://github.com/jneem/teddy/blob/9ab5e899ad6ef6911aecd3cf1033f1abe6e1f66c/src/x86/teddy_simd.rs#L145-L184 |
| 478 |
// |
| 479 |
// TL;DR avx2's PALIGNR instruction is actually just two 128-bit |
| 480 |
// PALIGNR instructions, which is not what we want, so we need to |
| 481 |
// do some extra shuffling. |
| 482 |
let v = _mm256_permute2x128_si256(vector2, self, 0x21); |
| 483 |
_mm256_alignr_epi8(self, v, 14) |
| 484 |
} |
| 485 |
|
| 486 |
#[inline(always)] |
| 487 |
unsafe fn shift_in_three_bytes(self, vector2: Self) -> Self { |
| 488 |
// Credit goes to jneem for figuring this out: |
| 489 |
// https://github.com/jneem/teddy/blob/9ab5e899ad6ef6911aecd3cf1033f1abe6e1f66c/src/x86/teddy_simd.rs#L145-L184 |
| 490 |
// |
| 491 |
// TL;DR avx2's PALIGNR instruction is actually just two 128-bit |
| 492 |
// PALIGNR instructions, which is not what we want, so we need to |
| 493 |
// do some extra shuffling. |
| 494 |
let v = _mm256_permute2x128_si256(vector2, self, 0x21); |
| 495 |
_mm256_alignr_epi8(self, v, 13) |
| 496 |
} |
| 497 |
|
| 498 |
#[inline(always)] |
| 499 |
unsafe fn shuffle_bytes(self, indices: Self) -> Self { |
| 500 |
_mm256_shuffle_epi8(self, indices) |
| 501 |
} |
| 502 |
|
| 503 |
#[inline(always)] |
| 504 |
unsafe fn for_each_64bit_lane<T>( |
| 505 |
self, |
| 506 |
mut f: impl FnMut(usize, u64) -> Option<T>, |
| 507 |
) -> Option<T> { |
| 508 |
// NOTE: At one point in the past, I used transmute to this to |
| 509 |
// get a [u64; 4], but it turned out to lead to worse codegen IIRC. |
| 510 |
// I've tried it more recently, and it looks like that's no longer |
| 511 |
// the case. But since there's no difference, we stick with the |
| 512 |
// slightly more complicated but transmute-free version. |
| 513 |
let lane = _mm256_extract_epi64(self, 0).to_bits(); |
| 514 |
if let Some(t) = f(0, lane) { |
| 515 |
return Some(t); |
| 516 |
} |
| 517 |
let lane = _mm256_extract_epi64(self, 1).to_bits(); |
| 518 |
if let Some(t) = f(1, lane) { |
| 519 |
return Some(t); |
| 520 |
} |
| 521 |
let lane = _mm256_extract_epi64(self, 2).to_bits(); |
| 522 |
if let Some(t) = f(2, lane) { |
| 523 |
return Some(t); |
| 524 |
} |
| 525 |
let lane = _mm256_extract_epi64(self, 3).to_bits(); |
| 526 |
if let Some(t) = f(3, lane) { |
| 527 |
return Some(t); |
| 528 |
} |
| 529 |
None |
| 530 |
} |
| 531 |
} |
| 532 |
|
| 533 |
impl FatVector for __m256i { |
| 534 |
type Half = __m128i; |
| 535 |
|
| 536 |
#[inline(always)] |
| 537 |
unsafe fn load_half_unaligned(data: *const u8) -> Self { |
| 538 |
let half = Self::Half::load_unaligned(data); |
| 539 |
_mm256_broadcastsi128_si256(half) |
| 540 |
} |
| 541 |
|
| 542 |
#[inline(always)] |
| 543 |
unsafe fn half_shift_in_one_byte(self, vector2: Self) -> Self { |
| 544 |
_mm256_alignr_epi8(self, vector2, 15) |
| 545 |
} |
| 546 |
|
| 547 |
#[inline(always)] |
| 548 |
unsafe fn half_shift_in_two_bytes(self, vector2: Self) -> Self { |
| 549 |
_mm256_alignr_epi8(self, vector2, 14) |
| 550 |
} |
| 551 |
|
| 552 |
#[inline(always)] |
| 553 |
unsafe fn half_shift_in_three_bytes(self, vector2: Self) -> Self { |
| 554 |
_mm256_alignr_epi8(self, vector2, 13) |
| 555 |
} |
| 556 |
|
| 557 |
#[inline(always)] |
| 558 |
unsafe fn swap_halves(self) -> Self { |
| 559 |
_mm256_permute4x64_epi64(self, 0x4E) |
| 560 |
} |
| 561 |
|
| 562 |
#[inline(always)] |
| 563 |
unsafe fn interleave_low_8bit_lanes(self, vector2: Self) -> Self { |
| 564 |
_mm256_unpacklo_epi8(self, vector2) |
| 565 |
} |
| 566 |
|
| 567 |
#[inline(always)] |
| 568 |
unsafe fn interleave_high_8bit_lanes(self, vector2: Self) -> Self { |
| 569 |
_mm256_unpackhi_epi8(self, vector2) |
| 570 |
} |
| 571 |
|
| 572 |
#[inline(always)] |
| 573 |
unsafe fn for_each_low_64bit_lane<T>( |
| 574 |
self, |
| 575 |
vector2: Self, |
| 576 |
mut f: impl FnMut(usize, u64) -> Option<T>, |
| 577 |
) -> Option<T> { |
| 578 |
let lane = _mm256_extract_epi64(self, 0).to_bits(); |
| 579 |
if let Some(t) = f(0, lane) { |
| 580 |
return Some(t); |
| 581 |
} |
| 582 |
let lane = _mm256_extract_epi64(self, 1).to_bits(); |
| 583 |
if let Some(t) = f(1, lane) { |
| 584 |
return Some(t); |
| 585 |
} |
| 586 |
let lane = _mm256_extract_epi64(vector2, 0).to_bits(); |
| 587 |
if let Some(t) = f(2, lane) { |
| 588 |
return Some(t); |
| 589 |
} |
| 590 |
let lane = _mm256_extract_epi64(vector2, 1).to_bits(); |
| 591 |
if let Some(t) = f(3, lane) { |
| 592 |
return Some(t); |
| 593 |
} |
| 594 |
None |
| 595 |
} |
| 596 |
} |
| 597 |
} |
| 598 |
|
| 599 |
#[cfg(all( |
| 600 |
target_arch = "aarch64", |
| 601 |
target_feature = "neon", |
| 602 |
target_endian = "little" |
| 603 |
))] |
| 604 |
mod aarch64_neon { |
| 605 |
use core::arch::aarch64::*; |
| 606 |
|
| 607 |
use super::Vector; |
| 608 |
|
| 609 |
impl Vector for uint8x16_t { |
| 610 |
const BITS: usize = 128; |
| 611 |
const BYTES: usize = 16; |
| 612 |
|
| 613 |
#[inline(always)] |
| 614 |
unsafe fn splat(byte: u8) -> uint8x16_t { |
| 615 |
vdupq_n_u8(byte) |
| 616 |
} |
| 617 |
|
| 618 |
#[inline(always)] |
| 619 |
unsafe fn load_unaligned(data: *const u8) -> uint8x16_t { |
| 620 |
vld1q_u8(data) |
| 621 |
} |
| 622 |
|
| 623 |
#[inline(always)] |
| 624 |
unsafe fn is_zero(self) -> bool { |
| 625 |
// Could also use vmaxvq_u8. |
| 626 |
// ... I tried that and couldn't observe any meaningful difference |
| 627 |
// in benchmarks. |
| 628 |
let maxes = vreinterpretq_u64_u8(vpmaxq_u8(self, self)); |
| 629 |
vgetq_lane_u64(maxes, 0) == 0 |
| 630 |
} |
| 631 |
|
| 632 |
#[inline(always)] |
| 633 |
unsafe fn cmpeq(self, vector2: Self) -> uint8x16_t { |
| 634 |
vceqq_u8(self, vector2) |
| 635 |
} |
| 636 |
|
| 637 |
#[inline(always)] |
| 638 |
unsafe fn and(self, vector2: Self) -> uint8x16_t { |
| 639 |
vandq_u8(self, vector2) |
| 640 |
} |
| 641 |
|
| 642 |
#[inline(always)] |
| 643 |
unsafe fn or(self, vector2: Self) -> uint8x16_t { |
| 644 |
vorrq_u8(self, vector2) |
| 645 |
} |
| 646 |
|
| 647 |
#[inline(always)] |
| 648 |
unsafe fn shift_8bit_lane_right<const BITS: i32>(self) -> Self { |
| 649 |
debug_assert!(BITS <= 7); |
| 650 |
vshrq_n_u8(self, BITS) |
| 651 |
} |
| 652 |
|
| 653 |
#[inline(always)] |
| 654 |
unsafe fn shift_in_one_byte(self, vector2: Self) -> Self { |
| 655 |
vextq_u8(vector2, self, 15) |
| 656 |
} |
| 657 |
|
| 658 |
#[inline(always)] |
| 659 |
unsafe fn shift_in_two_bytes(self, vector2: Self) -> Self { |
| 660 |
vextq_u8(vector2, self, 14) |
| 661 |
} |
| 662 |
|
| 663 |
#[inline(always)] |
| 664 |
unsafe fn shift_in_three_bytes(self, vector2: Self) -> Self { |
| 665 |
vextq_u8(vector2, self, 13) |
| 666 |
} |
| 667 |
|
| 668 |
#[inline(always)] |
| 669 |
unsafe fn shuffle_bytes(self, indices: Self) -> Self { |
| 670 |
vqtbl1q_u8(self, indices) |
| 671 |
} |
| 672 |
|
| 673 |
#[inline(always)] |
| 674 |
unsafe fn for_each_64bit_lane<T>( |
| 675 |
self, |
| 676 |
mut f: impl FnMut(usize, u64) -> Option<T>, |
| 677 |
) -> Option<T> { |
| 678 |
let this = vreinterpretq_u64_u8(self); |
| 679 |
let lane = vgetq_lane_u64(this, 0); |
| 680 |
if let Some(t) = f(0, lane) { |
| 681 |
return Some(t); |
| 682 |
} |
| 683 |
let lane = vgetq_lane_u64(this, 1); |
| 684 |
if let Some(t) = f(1, lane) { |
| 685 |
return Some(t); |
| 686 |
} |
| 687 |
None |
| 688 |
} |
| 689 |
} |
| 690 |
} |
| 691 |
|
| 692 |
#[cfg(all(test, target_arch = "x86_64", target_feature = "sse2"))] |
| 693 |
mod tests_x86_64_ssse3 { |
| 694 |
use core::arch::x86_64::*; |
| 695 |
|
| 696 |
use crate::util::int::{I32, U32}; |
| 697 |
|
| 698 |
use super::*; |
| 699 |
|
| 700 |
fn is_runnable() -> bool { |
| 701 |
std::is_x86_feature_detected!("ssse3") |
| 702 |
} |
| 703 |
|
| 704 |
#[target_feature(enable = "ssse3")] |
| 705 |
unsafe fn load(lanes: [u8; 16]) -> __m128i { |
| 706 |
__m128i::load_unaligned(&lanes as *const u8) |
| 707 |
} |
| 708 |
|
| 709 |
#[target_feature(enable = "ssse3")] |
| 710 |
unsafe fn unload(v: __m128i) -> [u8; 16] { |
| 711 |
[ |
| 712 |
_mm_extract_epi8(v, 0).to_bits().low_u8(), |
| 713 |
_mm_extract_epi8(v, 1).to_bits().low_u8(), |
| 714 |
_mm_extract_epi8(v, 2).to_bits().low_u8(), |
| 715 |
_mm_extract_epi8(v, 3).to_bits().low_u8(), |
| 716 |
_mm_extract_epi8(v, 4).to_bits().low_u8(), |
| 717 |
_mm_extract_epi8(v, 5).to_bits().low_u8(), |
| 718 |
_mm_extract_epi8(v, 6).to_bits().low_u8(), |
| 719 |
_mm_extract_epi8(v, 7).to_bits().low_u8(), |
| 720 |
_mm_extract_epi8(v, 8).to_bits().low_u8(), |
| 721 |
_mm_extract_epi8(v, 9).to_bits().low_u8(), |
| 722 |
_mm_extract_epi8(v, 10).to_bits().low_u8(), |
| 723 |
_mm_extract_epi8(v, 11).to_bits().low_u8(), |
| 724 |
_mm_extract_epi8(v, 12).to_bits().low_u8(), |
| 725 |
_mm_extract_epi8(v, 13).to_bits().low_u8(), |
| 726 |
_mm_extract_epi8(v, 14).to_bits().low_u8(), |
| 727 |
_mm_extract_epi8(v, 15).to_bits().low_u8(), |
| 728 |
] |
| 729 |
} |
| 730 |
|
| 731 |
#[test] |
| 732 |
fn vector_splat() { |
| 733 |
#[target_feature(enable = "ssse3")] |
| 734 |
unsafe fn test() { |
| 735 |
let v = __m128i::splat(0xAF); |
| 736 |
assert_eq!( |
| 737 |
unload(v), |
| 738 |
[ |
| 739 |
0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, |
| 740 |
0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF |
| 741 |
] |
| 742 |
); |
| 743 |
} |
| 744 |
if !is_runnable() { |
| 745 |
return; |
| 746 |
} |
| 747 |
unsafe { test() } |
| 748 |
} |
| 749 |
|
| 750 |
#[test] |
| 751 |
fn vector_is_zero() { |
| 752 |
#[target_feature(enable = "ssse3")] |
| 753 |
unsafe fn test() { |
| 754 |
let v = load([0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 755 |
assert!(!v.is_zero()); |
| 756 |
let v = load([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 757 |
assert!(v.is_zero()); |
| 758 |
} |
| 759 |
if !is_runnable() { |
| 760 |
return; |
| 761 |
} |
| 762 |
unsafe { test() } |
| 763 |
} |
| 764 |
|
| 765 |
#[test] |
| 766 |
fn vector_cmpeq() { |
| 767 |
#[target_feature(enable = "ssse3")] |
| 768 |
unsafe fn test() { |
| 769 |
let v1 = |
| 770 |
load([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 1]); |
| 771 |
let v2 = |
| 772 |
load([16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1]); |
| 773 |
assert_eq!( |
| 774 |
unload(v1.cmpeq(v2)), |
| 775 |
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0xFF] |
| 776 |
); |
| 777 |
} |
| 778 |
if !is_runnable() { |
| 779 |
return; |
| 780 |
} |
| 781 |
unsafe { test() } |
| 782 |
} |
| 783 |
|
| 784 |
#[test] |
| 785 |
fn vector_and() { |
| 786 |
#[target_feature(enable = "ssse3")] |
| 787 |
unsafe fn test() { |
| 788 |
let v1 = |
| 789 |
load([0, 0, 0, 0, 0, 0b1001, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 790 |
let v2 = |
| 791 |
load([0, 0, 0, 0, 0, 0b1010, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 792 |
assert_eq!( |
| 793 |
unload(v1.and(v2)), |
| 794 |
[0, 0, 0, 0, 0, 0b1000, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] |
| 795 |
); |
| 796 |
} |
| 797 |
if !is_runnable() { |
| 798 |
return; |
| 799 |
} |
| 800 |
unsafe { test() } |
| 801 |
} |
| 802 |
|
| 803 |
#[test] |
| 804 |
fn vector_or() { |
| 805 |
#[target_feature(enable = "ssse3")] |
| 806 |
unsafe fn test() { |
| 807 |
let v1 = |
| 808 |
load([0, 0, 0, 0, 0, 0b1001, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 809 |
let v2 = |
| 810 |
load([0, 0, 0, 0, 0, 0b1010, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 811 |
assert_eq!( |
| 812 |
unload(v1.or(v2)), |
| 813 |
[0, 0, 0, 0, 0, 0b1011, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] |
| 814 |
); |
| 815 |
} |
| 816 |
if !is_runnable() { |
| 817 |
return; |
| 818 |
} |
| 819 |
unsafe { test() } |
| 820 |
} |
| 821 |
|
| 822 |
#[test] |
| 823 |
fn vector_shift_8bit_lane_right() { |
| 824 |
#[target_feature(enable = "ssse3")] |
| 825 |
unsafe fn test() { |
| 826 |
let v = load([ |
| 827 |
0, 0, 0, 0, 0b1011, 0b0101, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 828 |
]); |
| 829 |
assert_eq!( |
| 830 |
unload(v.shift_8bit_lane_right::<2>()), |
| 831 |
[0, 0, 0, 0, 0b0010, 0b0001, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] |
| 832 |
); |
| 833 |
} |
| 834 |
if !is_runnable() { |
| 835 |
return; |
| 836 |
} |
| 837 |
unsafe { test() } |
| 838 |
} |
| 839 |
|
| 840 |
#[test] |
| 841 |
fn vector_shift_in_one_byte() { |
| 842 |
#[target_feature(enable = "ssse3")] |
| 843 |
unsafe fn test() { |
| 844 |
let v1 = |
| 845 |
load([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]); |
| 846 |
let v2 = load([ |
| 847 |
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 848 |
]); |
| 849 |
assert_eq!( |
| 850 |
unload(v1.shift_in_one_byte(v2)), |
| 851 |
[32, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15], |
| 852 |
); |
| 853 |
} |
| 854 |
if !is_runnable() { |
| 855 |
return; |
| 856 |
} |
| 857 |
unsafe { test() } |
| 858 |
} |
| 859 |
|
| 860 |
#[test] |
| 861 |
fn vector_shift_in_two_bytes() { |
| 862 |
#[target_feature(enable = "ssse3")] |
| 863 |
unsafe fn test() { |
| 864 |
let v1 = |
| 865 |
load([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]); |
| 866 |
let v2 = load([ |
| 867 |
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 868 |
]); |
| 869 |
assert_eq!( |
| 870 |
unload(v1.shift_in_two_bytes(v2)), |
| 871 |
[31, 32, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14], |
| 872 |
); |
| 873 |
} |
| 874 |
if !is_runnable() { |
| 875 |
return; |
| 876 |
} |
| 877 |
unsafe { test() } |
| 878 |
} |
| 879 |
|
| 880 |
#[test] |
| 881 |
fn vector_shift_in_three_bytes() { |
| 882 |
#[target_feature(enable = "ssse3")] |
| 883 |
unsafe fn test() { |
| 884 |
let v1 = |
| 885 |
load([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]); |
| 886 |
let v2 = load([ |
| 887 |
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 888 |
]); |
| 889 |
assert_eq!( |
| 890 |
unload(v1.shift_in_three_bytes(v2)), |
| 891 |
[30, 31, 32, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13], |
| 892 |
); |
| 893 |
} |
| 894 |
if !is_runnable() { |
| 895 |
return; |
| 896 |
} |
| 897 |
unsafe { test() } |
| 898 |
} |
| 899 |
|
| 900 |
#[test] |
| 901 |
fn vector_shuffle_bytes() { |
| 902 |
#[target_feature(enable = "ssse3")] |
| 903 |
unsafe fn test() { |
| 904 |
let v1 = |
| 905 |
load([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]); |
| 906 |
let v2 = |
| 907 |
load([0, 0, 0, 0, 4, 4, 4, 4, 8, 8, 8, 8, 12, 12, 12, 12]); |
| 908 |
assert_eq!( |
| 909 |
unload(v1.shuffle_bytes(v2)), |
| 910 |
[1, 1, 1, 1, 5, 5, 5, 5, 9, 9, 9, 9, 13, 13, 13, 13], |
| 911 |
); |
| 912 |
} |
| 913 |
if !is_runnable() { |
| 914 |
return; |
| 915 |
} |
| 916 |
unsafe { test() } |
| 917 |
} |
| 918 |
|
| 919 |
#[test] |
| 920 |
fn vector_for_each_64bit_lane() { |
| 921 |
#[target_feature(enable = "ssse3")] |
| 922 |
unsafe fn test() { |
| 923 |
let v = load([ |
| 924 |
0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0x0A, |
| 925 |
0x0B, 0x0C, 0x0D, 0x0E, 0x0F, 0x10, |
| 926 |
]); |
| 927 |
let mut lanes = [0u64; 2]; |
| 928 |
v.for_each_64bit_lane(|i, lane| { |
| 929 |
lanes[i] = lane; |
| 930 |
None::<()> |
| 931 |
}); |
| 932 |
assert_eq!(lanes, [0x0807060504030201, 0x100F0E0D0C0B0A09],); |
| 933 |
} |
| 934 |
if !is_runnable() { |
| 935 |
return; |
| 936 |
} |
| 937 |
unsafe { test() } |
| 938 |
} |
| 939 |
} |
| 940 |
|
| 941 |
#[cfg(all(test, target_arch = "x86_64", target_feature = "sse2"))] |
| 942 |
mod tests_x86_64_avx2 { |
| 943 |
use core::arch::x86_64::*; |
| 944 |
|
| 945 |
use crate::util::int::{I32, U32}; |
| 946 |
|
| 947 |
use super::*; |
| 948 |
|
| 949 |
fn is_runnable() -> bool { |
| 950 |
std::is_x86_feature_detected!("avx2") |
| 951 |
} |
| 952 |
|
| 953 |
#[target_feature(enable = "avx2")] |
| 954 |
unsafe fn load(lanes: [u8; 32]) -> __m256i { |
| 955 |
__m256i::load_unaligned(&lanes as *const u8) |
| 956 |
} |
| 957 |
|
| 958 |
#[target_feature(enable = "avx2")] |
| 959 |
unsafe fn load_half(lanes: [u8; 16]) -> __m256i { |
| 960 |
__m256i::load_half_unaligned(&lanes as *const u8) |
| 961 |
} |
| 962 |
|
| 963 |
#[target_feature(enable = "avx2")] |
| 964 |
unsafe fn unload(v: __m256i) -> [u8; 32] { |
| 965 |
[ |
| 966 |
_mm256_extract_epi8(v, 0).to_bits().low_u8(), |
| 967 |
_mm256_extract_epi8(v, 1).to_bits().low_u8(), |
| 968 |
_mm256_extract_epi8(v, 2).to_bits().low_u8(), |
| 969 |
_mm256_extract_epi8(v, 3).to_bits().low_u8(), |
| 970 |
_mm256_extract_epi8(v, 4).to_bits().low_u8(), |
| 971 |
_mm256_extract_epi8(v, 5).to_bits().low_u8(), |
| 972 |
_mm256_extract_epi8(v, 6).to_bits().low_u8(), |
| 973 |
_mm256_extract_epi8(v, 7).to_bits().low_u8(), |
| 974 |
_mm256_extract_epi8(v, 8).to_bits().low_u8(), |
| 975 |
_mm256_extract_epi8(v, 9).to_bits().low_u8(), |
| 976 |
_mm256_extract_epi8(v, 10).to_bits().low_u8(), |
| 977 |
_mm256_extract_epi8(v, 11).to_bits().low_u8(), |
| 978 |
_mm256_extract_epi8(v, 12).to_bits().low_u8(), |
| 979 |
_mm256_extract_epi8(v, 13).to_bits().low_u8(), |
| 980 |
_mm256_extract_epi8(v, 14).to_bits().low_u8(), |
| 981 |
_mm256_extract_epi8(v, 15).to_bits().low_u8(), |
| 982 |
_mm256_extract_epi8(v, 16).to_bits().low_u8(), |
| 983 |
_mm256_extract_epi8(v, 17).to_bits().low_u8(), |
| 984 |
_mm256_extract_epi8(v, 18).to_bits().low_u8(), |
| 985 |
_mm256_extract_epi8(v, 19).to_bits().low_u8(), |
| 986 |
_mm256_extract_epi8(v, 20).to_bits().low_u8(), |
| 987 |
_mm256_extract_epi8(v, 21).to_bits().low_u8(), |
| 988 |
_mm256_extract_epi8(v, 22).to_bits().low_u8(), |
| 989 |
_mm256_extract_epi8(v, 23).to_bits().low_u8(), |
| 990 |
_mm256_extract_epi8(v, 24).to_bits().low_u8(), |
| 991 |
_mm256_extract_epi8(v, 25).to_bits().low_u8(), |
| 992 |
_mm256_extract_epi8(v, 26).to_bits().low_u8(), |
| 993 |
_mm256_extract_epi8(v, 27).to_bits().low_u8(), |
| 994 |
_mm256_extract_epi8(v, 28).to_bits().low_u8(), |
| 995 |
_mm256_extract_epi8(v, 29).to_bits().low_u8(), |
| 996 |
_mm256_extract_epi8(v, 30).to_bits().low_u8(), |
| 997 |
_mm256_extract_epi8(v, 31).to_bits().low_u8(), |
| 998 |
] |
| 999 |
} |
| 1000 |
|
| 1001 |
#[test] |
| 1002 |
fn vector_splat() { |
| 1003 |
#[target_feature(enable = "avx2")] |
| 1004 |
unsafe fn test() { |
| 1005 |
let v = __m256i::splat(0xAF); |
| 1006 |
assert_eq!( |
| 1007 |
unload(v), |
| 1008 |
[ |
| 1009 |
0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, |
| 1010 |
0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, |
| 1011 |
0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, |
| 1012 |
0xAF, 0xAF, 0xAF, 0xAF, 0xAF, |
| 1013 |
] |
| 1014 |
); |
| 1015 |
} |
| 1016 |
if !is_runnable() { |
| 1017 |
return; |
| 1018 |
} |
| 1019 |
unsafe { test() } |
| 1020 |
} |
| 1021 |
|
| 1022 |
#[test] |
| 1023 |
fn vector_is_zero() { |
| 1024 |
#[target_feature(enable = "avx2")] |
| 1025 |
unsafe fn test() { |
| 1026 |
let v = load([ |
| 1027 |
0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1028 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1029 |
]); |
| 1030 |
assert!(!v.is_zero()); |
| 1031 |
let v = load([ |
| 1032 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1033 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1034 |
]); |
| 1035 |
assert!(v.is_zero()); |
| 1036 |
} |
| 1037 |
if !is_runnable() { |
| 1038 |
return; |
| 1039 |
} |
| 1040 |
unsafe { test() } |
| 1041 |
} |
| 1042 |
|
| 1043 |
#[test] |
| 1044 |
fn vector_cmpeq() { |
| 1045 |
#[target_feature(enable = "avx2")] |
| 1046 |
unsafe fn test() { |
| 1047 |
let v1 = load([ |
| 1048 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, |
| 1049 |
19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 1, |
| 1050 |
]); |
| 1051 |
let v2 = load([ |
| 1052 |
32, 31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, |
| 1053 |
17, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, |
| 1054 |
]); |
| 1055 |
assert_eq!( |
| 1056 |
unload(v1.cmpeq(v2)), |
| 1057 |
[ |
| 1058 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1059 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0xFF |
| 1060 |
] |
| 1061 |
); |
| 1062 |
} |
| 1063 |
if !is_runnable() { |
| 1064 |
return; |
| 1065 |
} |
| 1066 |
unsafe { test() } |
| 1067 |
} |
| 1068 |
|
| 1069 |
#[test] |
| 1070 |
fn vector_and() { |
| 1071 |
#[target_feature(enable = "avx2")] |
| 1072 |
unsafe fn test() { |
| 1073 |
let v1 = load([ |
| 1074 |
0, 0, 0, 0, 0, 0b1001, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1075 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1076 |
]); |
| 1077 |
let v2 = load([ |
| 1078 |
0, 0, 0, 0, 0, 0b1010, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1079 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1080 |
]); |
| 1081 |
assert_eq!( |
| 1082 |
unload(v1.and(v2)), |
| 1083 |
[ |
| 1084 |
0, 0, 0, 0, 0, 0b1000, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1085 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1086 |
] |
| 1087 |
); |
| 1088 |
} |
| 1089 |
if !is_runnable() { |
| 1090 |
return; |
| 1091 |
} |
| 1092 |
unsafe { test() } |
| 1093 |
} |
| 1094 |
|
| 1095 |
#[test] |
| 1096 |
fn vector_or() { |
| 1097 |
#[target_feature(enable = "avx2")] |
| 1098 |
unsafe fn test() { |
| 1099 |
let v1 = load([ |
| 1100 |
0, 0, 0, 0, 0, 0b1001, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1101 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1102 |
]); |
| 1103 |
let v2 = load([ |
| 1104 |
0, 0, 0, 0, 0, 0b1010, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1105 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1106 |
]); |
| 1107 |
assert_eq!( |
| 1108 |
unload(v1.or(v2)), |
| 1109 |
[ |
| 1110 |
0, 0, 0, 0, 0, 0b1011, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1111 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1112 |
] |
| 1113 |
); |
| 1114 |
} |
| 1115 |
if !is_runnable() { |
| 1116 |
return; |
| 1117 |
} |
| 1118 |
unsafe { test() } |
| 1119 |
} |
| 1120 |
|
| 1121 |
#[test] |
| 1122 |
fn vector_shift_8bit_lane_right() { |
| 1123 |
#[target_feature(enable = "avx2")] |
| 1124 |
unsafe fn test() { |
| 1125 |
let v = load([ |
| 1126 |
0, 0, 0, 0, 0b1011, 0b0101, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1127 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1128 |
]); |
| 1129 |
assert_eq!( |
| 1130 |
unload(v.shift_8bit_lane_right::<2>()), |
| 1131 |
[ |
| 1132 |
0, 0, 0, 0, 0b0010, 0b0001, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1133 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1134 |
] |
| 1135 |
); |
| 1136 |
} |
| 1137 |
if !is_runnable() { |
| 1138 |
return; |
| 1139 |
} |
| 1140 |
unsafe { test() } |
| 1141 |
} |
| 1142 |
|
| 1143 |
#[test] |
| 1144 |
fn vector_shift_in_one_byte() { |
| 1145 |
#[target_feature(enable = "avx2")] |
| 1146 |
unsafe fn test() { |
| 1147 |
let v1 = load([ |
| 1148 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, |
| 1149 |
19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1150 |
]); |
| 1151 |
let v2 = load([ |
| 1152 |
33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, |
| 1153 |
48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, |
| 1154 |
63, 64, |
| 1155 |
]); |
| 1156 |
assert_eq!( |
| 1157 |
unload(v1.shift_in_one_byte(v2)), |
| 1158 |
[ |
| 1159 |
64, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, |
| 1160 |
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, |
| 1161 |
31, |
| 1162 |
], |
| 1163 |
); |
| 1164 |
} |
| 1165 |
if !is_runnable() { |
| 1166 |
return; |
| 1167 |
} |
| 1168 |
unsafe { test() } |
| 1169 |
} |
| 1170 |
|
| 1171 |
#[test] |
| 1172 |
fn vector_shift_in_two_bytes() { |
| 1173 |
#[target_feature(enable = "avx2")] |
| 1174 |
unsafe fn test() { |
| 1175 |
let v1 = load([ |
| 1176 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, |
| 1177 |
19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1178 |
]); |
| 1179 |
let v2 = load([ |
| 1180 |
33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, |
| 1181 |
48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, |
| 1182 |
63, 64, |
| 1183 |
]); |
| 1184 |
assert_eq!( |
| 1185 |
unload(v1.shift_in_two_bytes(v2)), |
| 1186 |
[ |
| 1187 |
63, 64, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, |
| 1188 |
16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, |
| 1189 |
30, |
| 1190 |
], |
| 1191 |
); |
| 1192 |
} |
| 1193 |
if !is_runnable() { |
| 1194 |
return; |
| 1195 |
} |
| 1196 |
unsafe { test() } |
| 1197 |
} |
| 1198 |
|
| 1199 |
#[test] |
| 1200 |
fn vector_shift_in_three_bytes() { |
| 1201 |
#[target_feature(enable = "avx2")] |
| 1202 |
unsafe fn test() { |
| 1203 |
let v1 = load([ |
| 1204 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, |
| 1205 |
19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1206 |
]); |
| 1207 |
let v2 = load([ |
| 1208 |
33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, |
| 1209 |
48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, |
| 1210 |
63, 64, |
| 1211 |
]); |
| 1212 |
assert_eq!( |
| 1213 |
unload(v1.shift_in_three_bytes(v2)), |
| 1214 |
[ |
| 1215 |
62, 63, 64, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, |
| 1216 |
15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, |
| 1217 |
29, |
| 1218 |
], |
| 1219 |
); |
| 1220 |
} |
| 1221 |
if !is_runnable() { |
| 1222 |
return; |
| 1223 |
} |
| 1224 |
unsafe { test() } |
| 1225 |
} |
| 1226 |
|
| 1227 |
#[test] |
| 1228 |
fn vector_shuffle_bytes() { |
| 1229 |
#[target_feature(enable = "avx2")] |
| 1230 |
unsafe fn test() { |
| 1231 |
let v1 = load([ |
| 1232 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, |
| 1233 |
19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1234 |
]); |
| 1235 |
let v2 = load([ |
| 1236 |
0, 0, 0, 0, 4, 4, 4, 4, 8, 8, 8, 8, 12, 12, 12, 12, 16, 16, |
| 1237 |
16, 16, 20, 20, 20, 20, 24, 24, 24, 24, 28, 28, 28, 28, |
| 1238 |
]); |
| 1239 |
assert_eq!( |
| 1240 |
unload(v1.shuffle_bytes(v2)), |
| 1241 |
[ |
| 1242 |
1, 1, 1, 1, 5, 5, 5, 5, 9, 9, 9, 9, 13, 13, 13, 13, 17, |
| 1243 |
17, 17, 17, 21, 21, 21, 21, 25, 25, 25, 25, 29, 29, 29, |
| 1244 |
29 |
| 1245 |
], |
| 1246 |
); |
| 1247 |
} |
| 1248 |
if !is_runnable() { |
| 1249 |
return; |
| 1250 |
} |
| 1251 |
unsafe { test() } |
| 1252 |
} |
| 1253 |
|
| 1254 |
#[test] |
| 1255 |
fn vector_for_each_64bit_lane() { |
| 1256 |
#[target_feature(enable = "avx2")] |
| 1257 |
unsafe fn test() { |
| 1258 |
let v = load([ |
| 1259 |
0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0x0A, |
| 1260 |
0x0B, 0x0C, 0x0D, 0x0E, 0x0F, 0x10, 0x11, 0x12, 0x13, 0x14, |
| 1261 |
0x15, 0x16, 0x17, 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, |
| 1262 |
0x1F, 0x20, |
| 1263 |
]); |
| 1264 |
let mut lanes = [0u64; 4]; |
| 1265 |
v.for_each_64bit_lane(|i, lane| { |
| 1266 |
lanes[i] = lane; |
| 1267 |
None::<()> |
| 1268 |
}); |
| 1269 |
assert_eq!( |
| 1270 |
lanes, |
| 1271 |
[ |
| 1272 |
0x0807060504030201, |
| 1273 |
0x100F0E0D0C0B0A09, |
| 1274 |
0x1817161514131211, |
| 1275 |
0x201F1E1D1C1B1A19 |
| 1276 |
] |
| 1277 |
); |
| 1278 |
} |
| 1279 |
if !is_runnable() { |
| 1280 |
return; |
| 1281 |
} |
| 1282 |
unsafe { test() } |
| 1283 |
} |
| 1284 |
|
| 1285 |
#[test] |
| 1286 |
fn fat_vector_half_shift_in_one_byte() { |
| 1287 |
#[target_feature(enable = "avx2")] |
| 1288 |
unsafe fn test() { |
| 1289 |
let v1 = load_half([ |
| 1290 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, |
| 1291 |
]); |
| 1292 |
let v2 = load_half([ |
| 1293 |
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1294 |
]); |
| 1295 |
assert_eq!( |
| 1296 |
unload(v1.half_shift_in_one_byte(v2)), |
| 1297 |
[ |
| 1298 |
32, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 32, |
| 1299 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 |
| 1300 |
], |
| 1301 |
); |
| 1302 |
} |
| 1303 |
if !is_runnable() { |
| 1304 |
return; |
| 1305 |
} |
| 1306 |
unsafe { test() } |
| 1307 |
} |
| 1308 |
|
| 1309 |
#[test] |
| 1310 |
fn fat_vector_half_shift_in_two_bytes() { |
| 1311 |
#[target_feature(enable = "avx2")] |
| 1312 |
unsafe fn test() { |
| 1313 |
let v1 = load_half([ |
| 1314 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, |
| 1315 |
]); |
| 1316 |
let v2 = load_half([ |
| 1317 |
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1318 |
]); |
| 1319 |
assert_eq!( |
| 1320 |
unload(v1.half_shift_in_two_bytes(v2)), |
| 1321 |
[ |
| 1322 |
31, 32, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 31, |
| 1323 |
32, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, |
| 1324 |
], |
| 1325 |
); |
| 1326 |
} |
| 1327 |
if !is_runnable() { |
| 1328 |
return; |
| 1329 |
} |
| 1330 |
unsafe { test() } |
| 1331 |
} |
| 1332 |
|
| 1333 |
#[test] |
| 1334 |
fn fat_vector_half_shift_in_three_bytes() { |
| 1335 |
#[target_feature(enable = "avx2")] |
| 1336 |
unsafe fn test() { |
| 1337 |
let v1 = load_half([ |
| 1338 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, |
| 1339 |
]); |
| 1340 |
let v2 = load_half([ |
| 1341 |
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1342 |
]); |
| 1343 |
assert_eq!( |
| 1344 |
unload(v1.half_shift_in_three_bytes(v2)), |
| 1345 |
[ |
| 1346 |
30, 31, 32, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 30, |
| 1347 |
31, 32, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, |
| 1348 |
], |
| 1349 |
); |
| 1350 |
} |
| 1351 |
if !is_runnable() { |
| 1352 |
return; |
| 1353 |
} |
| 1354 |
unsafe { test() } |
| 1355 |
} |
| 1356 |
|
| 1357 |
#[test] |
| 1358 |
fn fat_vector_swap_halves() { |
| 1359 |
#[target_feature(enable = "avx2")] |
| 1360 |
unsafe fn test() { |
| 1361 |
let v = load([ |
| 1362 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, |
| 1363 |
19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1364 |
]); |
| 1365 |
assert_eq!( |
| 1366 |
unload(v.swap_halves()), |
| 1367 |
[ |
| 1368 |
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, |
| 1369 |
31, 32, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, |
| 1370 |
16, |
| 1371 |
], |
| 1372 |
); |
| 1373 |
} |
| 1374 |
if !is_runnable() { |
| 1375 |
return; |
| 1376 |
} |
| 1377 |
unsafe { test() } |
| 1378 |
} |
| 1379 |
|
| 1380 |
#[test] |
| 1381 |
fn fat_vector_interleave_low_8bit_lanes() { |
| 1382 |
#[target_feature(enable = "avx2")] |
| 1383 |
unsafe fn test() { |
| 1384 |
let v1 = load([ |
| 1385 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, |
| 1386 |
19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1387 |
]); |
| 1388 |
let v2 = load([ |
| 1389 |
33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, |
| 1390 |
48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, |
| 1391 |
63, 64, |
| 1392 |
]); |
| 1393 |
assert_eq!( |
| 1394 |
unload(v1.interleave_low_8bit_lanes(v2)), |
| 1395 |
[ |
| 1396 |
1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39, 8, 40, |
| 1397 |
17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55, |
| 1398 |
24, 56, |
| 1399 |
], |
| 1400 |
); |
| 1401 |
} |
| 1402 |
if !is_runnable() { |
| 1403 |
return; |
| 1404 |
} |
| 1405 |
unsafe { test() } |
| 1406 |
} |
| 1407 |
|
| 1408 |
#[test] |
| 1409 |
fn fat_vector_interleave_high_8bit_lanes() { |
| 1410 |
#[target_feature(enable = "avx2")] |
| 1411 |
unsafe fn test() { |
| 1412 |
let v1 = load([ |
| 1413 |
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, |
| 1414 |
19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1415 |
]); |
| 1416 |
let v2 = load([ |
| 1417 |
33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, |
| 1418 |
48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, |
| 1419 |
63, 64, |
| 1420 |
]); |
| 1421 |
assert_eq!( |
| 1422 |
unload(v1.interleave_high_8bit_lanes(v2)), |
| 1423 |
[ |
| 1424 |
9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47, 16, |
| 1425 |
48, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, |
| 1426 |
63, 32, 64, |
| 1427 |
], |
| 1428 |
); |
| 1429 |
} |
| 1430 |
if !is_runnable() { |
| 1431 |
return; |
| 1432 |
} |
| 1433 |
unsafe { test() } |
| 1434 |
} |
| 1435 |
|
| 1436 |
#[test] |
| 1437 |
fn fat_vector_for_each_low_64bit_lane() { |
| 1438 |
#[target_feature(enable = "avx2")] |
| 1439 |
unsafe fn test() { |
| 1440 |
let v1 = load([ |
| 1441 |
0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0x0A, |
| 1442 |
0x0B, 0x0C, 0x0D, 0x0E, 0x0F, 0x10, 0x11, 0x12, 0x13, 0x14, |
| 1443 |
0x15, 0x16, 0x17, 0x18, 0x19, 0x1A, 0x1B, 0x1C, 0x1D, 0x1E, |
| 1444 |
0x1F, 0x20, |
| 1445 |
]); |
| 1446 |
let v2 = load([ |
| 1447 |
0x21, 0x22, 0x23, 0x24, 0x25, 0x26, 0x27, 0x28, 0x29, 0x2A, |
| 1448 |
0x2B, 0x2C, 0x2D, 0x2E, 0x2F, 0x30, 0x31, 0x32, 0x33, 0x34, |
| 1449 |
0x35, 0x36, 0x37, 0x38, 0x39, 0x3A, 0x3B, 0x3C, 0x3D, 0x3E, |
| 1450 |
0x3F, 0x40, |
| 1451 |
]); |
| 1452 |
let mut lanes = [0u64; 4]; |
| 1453 |
v1.for_each_low_64bit_lane(v2, |i, lane| { |
| 1454 |
lanes[i] = lane; |
| 1455 |
None::<()> |
| 1456 |
}); |
| 1457 |
assert_eq!( |
| 1458 |
lanes, |
| 1459 |
[ |
| 1460 |
0x0807060504030201, |
| 1461 |
0x100F0E0D0C0B0A09, |
| 1462 |
0x2827262524232221, |
| 1463 |
0x302F2E2D2C2B2A29 |
| 1464 |
] |
| 1465 |
); |
| 1466 |
} |
| 1467 |
if !is_runnable() { |
| 1468 |
return; |
| 1469 |
} |
| 1470 |
unsafe { test() } |
| 1471 |
} |
| 1472 |
} |
| 1473 |
|
| 1474 |
#[cfg(all(test, target_arch = "aarch64", target_feature = "neon"))] |
| 1475 |
mod tests_aarch64_neon { |
| 1476 |
use core::arch::aarch64::*; |
| 1477 |
|
| 1478 |
use super::*; |
| 1479 |
|
| 1480 |
#[target_feature(enable = "neon")] |
| 1481 |
unsafe fn load(lanes: [u8; 16]) -> uint8x16_t { |
| 1482 |
uint8x16_t::load_unaligned(&lanes as *const u8) |
| 1483 |
} |
| 1484 |
|
| 1485 |
#[target_feature(enable = "neon")] |
| 1486 |
unsafe fn unload(v: uint8x16_t) -> [u8; 16] { |
| 1487 |
[ |
| 1488 |
vgetq_lane_u8(v, 0), |
| 1489 |
vgetq_lane_u8(v, 1), |
| 1490 |
vgetq_lane_u8(v, 2), |
| 1491 |
vgetq_lane_u8(v, 3), |
| 1492 |
vgetq_lane_u8(v, 4), |
| 1493 |
vgetq_lane_u8(v, 5), |
| 1494 |
vgetq_lane_u8(v, 6), |
| 1495 |
vgetq_lane_u8(v, 7), |
| 1496 |
vgetq_lane_u8(v, 8), |
| 1497 |
vgetq_lane_u8(v, 9), |
| 1498 |
vgetq_lane_u8(v, 10), |
| 1499 |
vgetq_lane_u8(v, 11), |
| 1500 |
vgetq_lane_u8(v, 12), |
| 1501 |
vgetq_lane_u8(v, 13), |
| 1502 |
vgetq_lane_u8(v, 14), |
| 1503 |
vgetq_lane_u8(v, 15), |
| 1504 |
] |
| 1505 |
} |
| 1506 |
|
| 1507 |
// Example functions. These don't test the Vector traits, but rather, |
| 1508 |
// specific NEON instructions. They are basically little experiments I |
| 1509 |
// wrote to figure out what an instruction does since their descriptions |
| 1510 |
// are so dense. I decided to keep the experiments around as example tests |
| 1511 |
// in case there' useful. |
| 1512 |
|
| 1513 |
#[test] |
| 1514 |
fn example_vmaxvq_u8_non_zero() { |
| 1515 |
#[target_feature(enable = "neon")] |
| 1516 |
unsafe fn example() { |
| 1517 |
let v = load([0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 1518 |
assert_eq!(vmaxvq_u8(v), 1); |
| 1519 |
} |
| 1520 |
unsafe { example() } |
| 1521 |
} |
| 1522 |
|
| 1523 |
#[test] |
| 1524 |
fn example_vmaxvq_u8_zero() { |
| 1525 |
#[target_feature(enable = "neon")] |
| 1526 |
unsafe fn example() { |
| 1527 |
let v = load([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 1528 |
assert_eq!(vmaxvq_u8(v), 0); |
| 1529 |
} |
| 1530 |
unsafe { example() } |
| 1531 |
} |
| 1532 |
|
| 1533 |
#[test] |
| 1534 |
fn example_vpmaxq_u8_non_zero() { |
| 1535 |
#[target_feature(enable = "neon")] |
| 1536 |
unsafe fn example() { |
| 1537 |
let v = load([0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 1538 |
let r = vpmaxq_u8(v, v); |
| 1539 |
assert_eq!( |
| 1540 |
unload(r), |
| 1541 |
[0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0] |
| 1542 |
); |
| 1543 |
} |
| 1544 |
unsafe { example() } |
| 1545 |
} |
| 1546 |
|
| 1547 |
#[test] |
| 1548 |
fn example_vpmaxq_u8_self() { |
| 1549 |
#[target_feature(enable = "neon")] |
| 1550 |
unsafe fn example() { |
| 1551 |
let v = |
| 1552 |
load([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]); |
| 1553 |
let r = vpmaxq_u8(v, v); |
| 1554 |
assert_eq!( |
| 1555 |
unload(r), |
| 1556 |
[2, 4, 6, 8, 10, 12, 14, 16, 2, 4, 6, 8, 10, 12, 14, 16] |
| 1557 |
); |
| 1558 |
} |
| 1559 |
unsafe { example() } |
| 1560 |
} |
| 1561 |
|
| 1562 |
#[test] |
| 1563 |
fn example_vpmaxq_u8_other() { |
| 1564 |
#[target_feature(enable = "neon")] |
| 1565 |
unsafe fn example() { |
| 1566 |
let v1 = |
| 1567 |
load([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]); |
| 1568 |
let v2 = load([ |
| 1569 |
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1570 |
]); |
| 1571 |
let r = vpmaxq_u8(v1, v2); |
| 1572 |
assert_eq!( |
| 1573 |
unload(r), |
| 1574 |
[2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30, 32] |
| 1575 |
); |
| 1576 |
} |
| 1577 |
unsafe { example() } |
| 1578 |
} |
| 1579 |
|
| 1580 |
// Now we test the actual methods on the Vector trait. |
| 1581 |
|
| 1582 |
#[test] |
| 1583 |
fn vector_splat() { |
| 1584 |
#[target_feature(enable = "neon")] |
| 1585 |
unsafe fn test() { |
| 1586 |
let v = uint8x16_t::splat(0xAF); |
| 1587 |
assert_eq!( |
| 1588 |
unload(v), |
| 1589 |
[ |
| 1590 |
0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, |
| 1591 |
0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF, 0xAF |
| 1592 |
] |
| 1593 |
); |
| 1594 |
} |
| 1595 |
unsafe { test() } |
| 1596 |
} |
| 1597 |
|
| 1598 |
#[test] |
| 1599 |
fn vector_is_zero() { |
| 1600 |
#[target_feature(enable = "neon")] |
| 1601 |
unsafe fn test() { |
| 1602 |
let v = load([0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 1603 |
assert!(!v.is_zero()); |
| 1604 |
let v = load([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 1605 |
assert!(v.is_zero()); |
| 1606 |
} |
| 1607 |
unsafe { test() } |
| 1608 |
} |
| 1609 |
|
| 1610 |
#[test] |
| 1611 |
fn vector_cmpeq() { |
| 1612 |
#[target_feature(enable = "neon")] |
| 1613 |
unsafe fn test() { |
| 1614 |
let v1 = |
| 1615 |
load([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 1]); |
| 1616 |
let v2 = |
| 1617 |
load([16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1]); |
| 1618 |
assert_eq!( |
| 1619 |
unload(v1.cmpeq(v2)), |
| 1620 |
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0xFF] |
| 1621 |
); |
| 1622 |
} |
| 1623 |
unsafe { test() } |
| 1624 |
} |
| 1625 |
|
| 1626 |
#[test] |
| 1627 |
fn vector_and() { |
| 1628 |
#[target_feature(enable = "neon")] |
| 1629 |
unsafe fn test() { |
| 1630 |
let v1 = |
| 1631 |
load([0, 0, 0, 0, 0, 0b1001, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 1632 |
let v2 = |
| 1633 |
load([0, 0, 0, 0, 0, 0b1010, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 1634 |
assert_eq!( |
| 1635 |
unload(v1.and(v2)), |
| 1636 |
[0, 0, 0, 0, 0, 0b1000, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] |
| 1637 |
); |
| 1638 |
} |
| 1639 |
unsafe { test() } |
| 1640 |
} |
| 1641 |
|
| 1642 |
#[test] |
| 1643 |
fn vector_or() { |
| 1644 |
#[target_feature(enable = "neon")] |
| 1645 |
unsafe fn test() { |
| 1646 |
let v1 = |
| 1647 |
load([0, 0, 0, 0, 0, 0b1001, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 1648 |
let v2 = |
| 1649 |
load([0, 0, 0, 0, 0, 0b1010, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]); |
| 1650 |
assert_eq!( |
| 1651 |
unload(v1.or(v2)), |
| 1652 |
[0, 0, 0, 0, 0, 0b1011, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] |
| 1653 |
); |
| 1654 |
} |
| 1655 |
unsafe { test() } |
| 1656 |
} |
| 1657 |
|
| 1658 |
#[test] |
| 1659 |
fn vector_shift_8bit_lane_right() { |
| 1660 |
#[target_feature(enable = "neon")] |
| 1661 |
unsafe fn test() { |
| 1662 |
let v = load([ |
| 1663 |
0, 0, 0, 0, 0b1011, 0b0101, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1664 |
]); |
| 1665 |
assert_eq!( |
| 1666 |
unload(v.shift_8bit_lane_right::<2>()), |
| 1667 |
[0, 0, 0, 0, 0b0010, 0b0001, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] |
| 1668 |
); |
| 1669 |
} |
| 1670 |
unsafe { test() } |
| 1671 |
} |
| 1672 |
|
| 1673 |
#[test] |
| 1674 |
fn vector_shift_in_one_byte() { |
| 1675 |
#[target_feature(enable = "neon")] |
| 1676 |
unsafe fn test() { |
| 1677 |
let v1 = |
| 1678 |
load([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]); |
| 1679 |
let v2 = load([ |
| 1680 |
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1681 |
]); |
| 1682 |
assert_eq!( |
| 1683 |
unload(v1.shift_in_one_byte(v2)), |
| 1684 |
[32, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15], |
| 1685 |
); |
| 1686 |
} |
| 1687 |
unsafe { test() } |
| 1688 |
} |
| 1689 |
|
| 1690 |
#[test] |
| 1691 |
fn vector_shift_in_two_bytes() { |
| 1692 |
#[target_feature(enable = "neon")] |
| 1693 |
unsafe fn test() { |
| 1694 |
let v1 = |
| 1695 |
load([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]); |
| 1696 |
let v2 = load([ |
| 1697 |
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1698 |
]); |
| 1699 |
assert_eq!( |
| 1700 |
unload(v1.shift_in_two_bytes(v2)), |
| 1701 |
[31, 32, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14], |
| 1702 |
); |
| 1703 |
} |
| 1704 |
unsafe { test() } |
| 1705 |
} |
| 1706 |
|
| 1707 |
#[test] |
| 1708 |
fn vector_shift_in_three_bytes() { |
| 1709 |
#[target_feature(enable = "neon")] |
| 1710 |
unsafe fn test() { |
| 1711 |
let v1 = |
| 1712 |
load([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]); |
| 1713 |
let v2 = load([ |
| 1714 |
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1715 |
]); |
| 1716 |
assert_eq!( |
| 1717 |
unload(v1.shift_in_three_bytes(v2)), |
| 1718 |
[30, 31, 32, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13], |
| 1719 |
); |
| 1720 |
} |
| 1721 |
unsafe { test() } |
| 1722 |
} |
| 1723 |
|
| 1724 |
#[test] |
| 1725 |
fn vector_shuffle_bytes() { |
| 1726 |
#[target_feature(enable = "neon")] |
| 1727 |
unsafe fn test() { |
| 1728 |
let v1 = |
| 1729 |
load([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]); |
| 1730 |
let v2 = |
| 1731 |
load([0, 0, 0, 0, 4, 4, 4, 4, 8, 8, 8, 8, 12, 12, 12, 12]); |
| 1732 |
assert_eq!( |
| 1733 |
unload(v1.shuffle_bytes(v2)), |
| 1734 |
[1, 1, 1, 1, 5, 5, 5, 5, 9, 9, 9, 9, 13, 13, 13, 13], |
| 1735 |
); |
| 1736 |
} |
| 1737 |
unsafe { test() } |
| 1738 |
} |
| 1739 |
|
| 1740 |
#[test] |
| 1741 |
fn vector_for_each_64bit_lane() { |
| 1742 |
#[target_feature(enable = "neon")] |
| 1743 |
unsafe fn test() { |
| 1744 |
let v = load([ |
| 1745 |
0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0x0A, |
| 1746 |
0x0B, 0x0C, 0x0D, 0x0E, 0x0F, 0x10, |
| 1747 |
]); |
| 1748 |
let mut lanes = [0u64; 2]; |
| 1749 |
v.for_each_64bit_lane(|i, lane| { |
| 1750 |
lanes[i] = lane; |
| 1751 |
None::<()> |
| 1752 |
}); |
| 1753 |
assert_eq!(lanes, [0x0807060504030201, 0x100F0E0D0C0B0A09],); |
| 1754 |
} |
| 1755 |
unsafe { test() } |
| 1756 |
} |
| 1757 |
} |