1 | //! SIMD compiler intrinsics. |
2 | //! |
3 | //! In this module, a "vector" is any `repr(simd)` type. |
4 | |
5 | extern "platform-intrinsic" { |
6 | /// Add two simd vectors elementwise. |
7 | /// |
8 | /// `T` must be a vector of integer or floating point primitive types. |
9 | pub fn simd_add<T>(x: T, y: T) -> T; |
10 | |
11 | /// Subtract `rhs` from `lhs` elementwise. |
12 | /// |
13 | /// `T` must be a vector of integer or floating point primitive types. |
14 | pub fn simd_sub<T>(lhs: T, rhs: T) -> T; |
15 | |
16 | /// Multiply two simd vectors elementwise. |
17 | /// |
18 | /// `T` must be a vector of integer or floating point primitive types. |
19 | pub fn simd_mul<T>(x: T, y: T) -> T; |
20 | |
21 | /// Divide `lhs` by `rhs` elementwise. |
22 | /// |
23 | /// `T` must be a vector of integer or floating point primitive types. |
24 | /// |
25 | /// # Safety |
26 | /// For integers, `rhs` must not contain any zero elements. |
27 | /// Additionally for signed integers, `<int>::MIN / -1` is undefined behavior. |
28 | pub fn simd_div<T>(lhs: T, rhs: T) -> T; |
29 | |
30 | /// Remainder of two vectors elementwise |
31 | /// |
32 | /// `T` must be a vector of integer or floating point primitive types. |
33 | /// |
34 | /// # Safety |
35 | /// For integers, `rhs` must not contain any zero elements. |
36 | /// Additionally for signed integers, `<int>::MIN / -1` is undefined behavior. |
37 | pub fn simd_rem<T>(lhs: T, rhs: T) -> T; |
38 | |
39 | /// Elementwise vector left shift, with UB on overflow. |
40 | /// |
41 | /// Shift `lhs` left by `rhs`, shifting in sign bits for signed types. |
42 | /// |
43 | /// `T` must be a vector of integer primitive types. |
44 | /// |
45 | /// # Safety |
46 | /// |
47 | /// Each element of `rhs` must be less than `<int>::BITS`. |
48 | pub fn simd_shl<T>(lhs: T, rhs: T) -> T; |
49 | |
50 | /// Elementwise vector right shift, with UB on overflow. |
51 | /// |
52 | /// `T` must be a vector of integer primitive types. |
53 | /// |
54 | /// Shift `lhs` right by `rhs`, shifting in sign bits for signed types. |
55 | /// |
56 | /// # Safety |
57 | /// |
58 | /// Each element of `rhs` must be less than `<int>::BITS`. |
59 | pub fn simd_shr<T>(lhs: T, rhs: T) -> T; |
60 | |
61 | /// Elementwise vector "and". |
62 | /// |
63 | /// `T` must be a vector of integer primitive types. |
64 | pub fn simd_and<T>(x: T, y: T) -> T; |
65 | |
66 | /// Elementwise vector "or". |
67 | /// |
68 | /// `T` must be a vector of integer primitive types. |
69 | pub fn simd_or<T>(x: T, y: T) -> T; |
70 | |
71 | /// Elementwise vector "exclusive or". |
72 | /// |
73 | /// `T` must be a vector of integer primitive types. |
74 | pub fn simd_xor<T>(x: T, y: T) -> T; |
75 | |
76 | /// Numerically cast a vector, elementwise. |
77 | /// |
78 | /// `T` and `U` must be vectors of integer or floating point primitive types, and must have the |
79 | /// same length. |
80 | /// |
81 | /// When casting floats to integers, the result is truncated. Out-of-bounds result lead to UB. |
82 | /// When casting integers to floats, the result is rounded. |
83 | /// Otherwise, truncates or extends the value, maintaining the sign for signed integers. |
84 | /// |
85 | /// # Safety |
86 | /// Casting from integer types is always safe. |
87 | /// Casting between two float types is also always safe. |
88 | /// |
89 | /// Casting floats to integers truncates, following the same rules as `to_int_unchecked`. |
90 | /// Specifically, each element must: |
91 | /// * Not be `NaN` |
92 | /// * Not be infinite |
93 | /// * Be representable in the return type, after truncating off its fractional part |
94 | pub fn simd_cast<T, U>(x: T) -> U; |
95 | |
96 | /// Numerically cast a vector, elementwise. |
97 | /// |
98 | /// `T` and `U` be a vectors of integer or floating point primitive types, and must have the |
99 | /// same length. |
100 | /// |
101 | /// Like `simd_cast`, but saturates float-to-integer conversions (NaN becomes 0). |
102 | /// This matches regular `as` and is always safe. |
103 | /// |
104 | /// When casting floats to integers, the result is truncated. |
105 | /// When casting integers to floats, the result is rounded. |
106 | /// Otherwise, truncates or extends the value, maintaining the sign for signed integers. |
107 | pub fn simd_as<T, U>(x: T) -> U; |
108 | |
109 | /// Elementwise negation of a vector. |
110 | /// |
111 | /// `T` must be a vector of integer or floating-point primitive types. |
112 | /// |
113 | /// Rust panics for `-<int>::Min` due to overflow, but it is not UB with this intrinsic. |
114 | pub fn simd_neg<T>(x: T) -> T; |
115 | |
116 | /// Elementwise absolute value of a vector. |
117 | /// |
118 | /// `T` must be a vector of floating-point primitive types. |
119 | pub fn simd_fabs<T>(x: T) -> T; |
120 | |
121 | /// Elementwise minimum of a vector. |
122 | /// |
123 | /// `T` must be a vector of floating-point primitive types. |
124 | /// |
125 | /// Follows IEEE-754 `minNum` semantics. |
126 | pub fn simd_fmin<T>(x: T, y: T) -> T; |
127 | |
128 | /// Elementwise maximum of a vector. |
129 | /// |
130 | /// `T` must be a vector of floating-point primitive types. |
131 | /// |
132 | /// Follows IEEE-754 `maxNum` semantics. |
133 | pub fn simd_fmax<T>(x: T, y: T) -> T; |
134 | |
135 | /// Tests elementwise equality of two vectors. |
136 | /// |
137 | /// `T` must be a vector of floating-point primitive types. |
138 | /// |
139 | /// `U` must be a vector of integers with the same number of elements and element size as `T`. |
140 | /// |
141 | /// Returns `0` for false and `!0` for true. |
142 | pub fn simd_eq<T, U>(x: T, y: T) -> U; |
143 | |
144 | /// Tests elementwise inequality equality of two vectors. |
145 | /// |
146 | /// `T` must be a vector of floating-point primitive types. |
147 | /// |
148 | /// `U` must be a vector of integers with the same number of elements and element size as `T`. |
149 | /// |
150 | /// Returns `0` for false and `!0` for true. |
151 | pub fn simd_ne<T, U>(x: T, y: T) -> U; |
152 | |
153 | /// Tests if `x` is less than `y`, elementwise. |
154 | /// |
155 | /// `T` must be a vector of floating-point primitive types. |
156 | /// |
157 | /// `U` must be a vector of integers with the same number of elements and element size as `T`. |
158 | /// |
159 | /// Returns `0` for false and `!0` for true. |
160 | pub fn simd_lt<T, U>(x: T, y: T) -> U; |
161 | |
162 | /// Tests if `x` is less than or equal to `y`, elementwise. |
163 | /// |
164 | /// `T` must be a vector of floating-point primitive types. |
165 | /// |
166 | /// `U` must be a vector of integers with the same number of elements and element size as `T`. |
167 | /// |
168 | /// Returns `0` for false and `!0` for true. |
169 | pub fn simd_le<T, U>(x: T, y: T) -> U; |
170 | |
171 | /// Tests if `x` is greater than `y`, elementwise. |
172 | /// |
173 | /// `T` must be a vector of floating-point primitive types. |
174 | /// |
175 | /// `U` must be a vector of integers with the same number of elements and element size as `T`. |
176 | /// |
177 | /// Returns `0` for false and `!0` for true. |
178 | pub fn simd_gt<T, U>(x: T, y: T) -> U; |
179 | |
180 | /// Tests if `x` is greater than or equal to `y`, elementwise. |
181 | /// |
182 | /// `T` must be a vector of floating-point primitive types. |
183 | /// |
184 | /// `U` must be a vector of integers with the same number of elements and element size as `T`. |
185 | /// |
186 | /// Returns `0` for false and `!0` for true. |
187 | pub fn simd_ge<T, U>(x: T, y: T) -> U; |
188 | |
189 | /// Shuffle two vectors by const indices. |
190 | /// |
191 | /// `T` must be a vector. |
192 | /// |
193 | /// `U` must be a const array of `i32`s. |
194 | /// |
195 | /// `V` must be a vector with the same element type as `T` and the same length as `U`. |
196 | /// |
197 | /// Concatenates `x` and `y`, then returns a new vector such that each element is selected from |
198 | /// the concatenation by the matching index in `idx`. |
199 | pub fn simd_shuffle<T, U, V>(x: T, y: T, idx: U) -> V; |
200 | |
201 | /// Read a vector of pointers. |
202 | /// |
203 | /// `T` must be a vector. |
204 | /// |
205 | /// `U` must be a vector of pointers to the element type of `T`, with the same length as `T`. |
206 | /// |
207 | /// `V` must be a vector of integers with the same length as `T` (but any element size). |
208 | /// |
209 | /// `idx` must be a constant: either naming a constant item, or an inline |
210 | /// `const {}` expression. |
211 | /// |
212 | /// For each pointer in `ptr`, if the corresponding value in `mask` is `!0`, read the pointer. |
213 | /// Otherwise if the corresponding value in `mask` is `0`, return the corresponding value from |
214 | /// `val`. |
215 | /// |
216 | /// # Safety |
217 | /// Unmasked values in `T` must be readable as if by `<ptr>::read` (e.g. aligned to the element |
218 | /// type). |
219 | /// |
220 | /// `mask` must only contain `0` or `!0` values. |
221 | pub fn simd_gather<T, U, V>(val: T, ptr: U, mask: V) -> T; |
222 | |
223 | /// Write to a vector of pointers. |
224 | /// |
225 | /// `T` must be a vector. |
226 | /// |
227 | /// `U` must be a vector of pointers to the element type of `T`, with the same length as `T`. |
228 | /// |
229 | /// `V` must be a vector of integers with the same length as `T` (but any element size). |
230 | /// |
231 | /// For each pointer in `ptr`, if the corresponding value in `mask` is `!0`, write the |
232 | /// corresponding value in `val` to the pointer. |
233 | /// Otherwise if the corresponding value in `mask` is `0`, do nothing. |
234 | /// |
235 | /// # Safety |
236 | /// Unmasked values in `T` must be writeable as if by `<ptr>::write` (e.g. aligned to the element |
237 | /// type). |
238 | /// |
239 | /// `mask` must only contain `0` or `!0` values. |
240 | pub fn simd_scatter<T, U, V>(val: T, ptr: U, mask: V); |
241 | |
242 | /// Read a vector of pointers. |
243 | /// |
244 | /// `T` must be a vector. |
245 | /// |
246 | /// `U` must be a vector of pointers to the element type of `T`, with the same length as `T`. |
247 | /// |
248 | /// `V` must be a vector of integers with the same length as `T` (but any element size). |
249 | /// |
250 | /// For each element, if the corresponding value in `mask` is `!0`, read the corresponding |
251 | /// pointer from `ptr`. |
252 | /// Otherwise if the corresponding value in `mask` is `0`, return the corresponding value from |
253 | /// `val`. |
254 | /// |
255 | /// # Safety |
256 | /// Unmasked values in `T` must be readable as if by `<ptr>::read` (e.g. aligned to the element |
257 | /// type). |
258 | /// |
259 | /// `mask` must only contain `0` or `!0` values. |
260 | pub fn simd_masked_load<V, U, T>(mask: V, ptr: U, val: T) -> T; |
261 | |
262 | /// Write to a vector of pointers. |
263 | /// |
264 | /// `T` must be a vector. |
265 | /// |
266 | /// `U` must be a vector of pointers to the element type of `T`, with the same length as `T`. |
267 | /// |
268 | /// `V` must be a vector of integers with the same length as `T` (but any element size). |
269 | /// |
270 | /// For each element, if the corresponding value in `mask` is `!0`, write the corresponding |
271 | /// value in `val` to the pointer. |
272 | /// Otherwise if the corresponding value in `mask` is `0`, do nothing. |
273 | /// |
274 | /// # Safety |
275 | /// Unmasked values in `T` must be writeable as if by `<ptr>::write` (e.g. aligned to the element |
276 | /// type). |
277 | /// |
278 | /// `mask` must only contain `0` or `!0` values. |
279 | pub fn simd_masked_store<V, U, T>(mask: V, ptr: U, val: T); |
280 | |
281 | /// Add two simd vectors elementwise, with saturation. |
282 | /// |
283 | /// `T` must be a vector of integer primitive types. |
284 | pub fn simd_saturating_add<T>(x: T, y: T) -> T; |
285 | |
286 | /// Subtract two simd vectors elementwise, with saturation. |
287 | /// |
288 | /// `T` must be a vector of integer primitive types. |
289 | /// |
290 | /// Subtract `rhs` from `lhs`. |
291 | pub fn simd_saturating_sub<T>(lhs: T, rhs: T) -> T; |
292 | |
293 | /// Add elements within a vector from left to right. |
294 | /// |
295 | /// `T` must be a vector of integer or floating-point primitive types. |
296 | /// |
297 | /// `U` must be the element type of `T`. |
298 | /// |
299 | /// Starting with the value `y`, add the elements of `x` and accumulate. |
300 | pub fn simd_reduce_add_ordered<T, U>(x: T, y: U) -> U; |
301 | |
302 | /// Multiply elements within a vector from left to right. |
303 | /// |
304 | /// `T` must be a vector of integer or floating-point primitive types. |
305 | /// |
306 | /// `U` must be the element type of `T`. |
307 | /// |
308 | /// Starting with the value `y`, multiply the elements of `x` and accumulate. |
309 | pub fn simd_reduce_mul_ordered<T, U>(x: T, y: U) -> U; |
310 | |
311 | /// Check if all mask values are true. |
312 | /// |
313 | /// `T` must be a vector of integer primitive types. |
314 | /// |
315 | /// # Safety |
316 | /// `x` must contain only `0` or `!0`. |
317 | pub fn simd_reduce_all<T>(x: T) -> bool; |
318 | |
319 | /// Check if all mask values are true. |
320 | /// |
321 | /// `T` must be a vector of integer primitive types. |
322 | /// |
323 | /// # Safety |
324 | /// `x` must contain only `0` or `!0`. |
325 | pub fn simd_reduce_any<T>(x: T) -> bool; |
326 | |
327 | /// Return the maximum element of a vector. |
328 | /// |
329 | /// `T` must be a vector of integer or floating-point primitive types. |
330 | /// |
331 | /// `U` must be the element type of `T`. |
332 | /// |
333 | /// For floating-point values, uses IEEE-754 `maxNum`. |
334 | pub fn simd_reduce_max<T, U>(x: T) -> U; |
335 | |
336 | /// Return the minimum element of a vector. |
337 | /// |
338 | /// `T` must be a vector of integer or floating-point primitive types. |
339 | /// |
340 | /// `U` must be the element type of `T`. |
341 | /// |
342 | /// For floating-point values, uses IEEE-754 `minNum`. |
343 | pub fn simd_reduce_min<T, U>(x: T) -> U; |
344 | |
345 | /// Logical "and" all elements together. |
346 | /// |
347 | /// `T` must be a vector of integer or floating-point primitive types. |
348 | /// |
349 | /// `U` must be the element type of `T`. |
350 | pub fn simd_reduce_and<T, U>(x: T) -> U; |
351 | |
352 | /// Logical "or" all elements together. |
353 | /// |
354 | /// `T` must be a vector of integer or floating-point primitive types. |
355 | /// |
356 | /// `U` must be the element type of `T`. |
357 | pub fn simd_reduce_or<T, U>(x: T) -> U; |
358 | |
359 | /// Logical "exclusive or" all elements together. |
360 | /// |
361 | /// `T` must be a vector of integer or floating-point primitive types. |
362 | /// |
363 | /// `U` must be the element type of `T`. |
364 | pub fn simd_reduce_xor<T, U>(x: T) -> U; |
365 | |
366 | /// Truncate an integer vector to a bitmask. |
367 | /// |
368 | /// `T` must be an integer vector. |
369 | /// |
370 | /// `U` must be either the smallest unsigned integer with at least as many bits as the length |
371 | /// of `T`, or the smallest array of `u8` with as many bits as the length of `T`. |
372 | /// |
373 | /// Each element is truncated to a single bit and packed into the result. |
374 | /// |
375 | /// No matter whether the output is an array or an unsigned integer, it is treated as a single |
376 | /// contiguous list of bits. The bitmask is always packed on the least-significant side of the |
377 | /// output, and padded with 0s in the most-significant bits. The order of the bits depends on |
378 | /// endianess: |
379 | /// |
380 | /// * On little endian, the least significant bit corresponds to the first vector element. |
381 | /// * On big endian, the least significant bit corresponds to the last vector element. |
382 | /// |
383 | /// For example, `[!0, 0, !0, !0]` packs to `0b1101` on little endian and `0b1011` on big |
384 | /// endian. |
385 | /// |
386 | /// To consider a larger example, `[!0, 0, 0, 0, 0, 0, 0, 0, !0, !0, 0, 0, 0, 0, !0, 0]` packs |
387 | /// to `[0b00000001, 0b01000011]` or `0b0100001100000001` on little endian, and `[0b10000000, |
388 | /// 0b11000010]` or `0b1000000011000010` on big endian. |
389 | /// |
390 | /// # Safety |
391 | /// `x` must contain only `0` and `!0`. |
392 | pub fn simd_bitmask<T, U>(x: T) -> U; |
393 | |
394 | /// Select elements from a mask. |
395 | /// |
396 | /// `M` must be an integer vector. |
397 | /// |
398 | /// `T` must be a vector with the same number of elements as `M`. |
399 | /// |
400 | /// For each element, if the corresponding value in `mask` is `!0`, select the element from |
401 | /// `if_true`. If the corresponding value in `mask` is `0`, select the element from |
402 | /// `if_false`. |
403 | /// |
404 | /// # Safety |
405 | /// `mask` must only contain `0` and `!0`. |
406 | pub fn simd_select<M, T>(mask: M, if_true: T, if_false: T) -> T; |
407 | |
408 | /// Select elements from a bitmask. |
409 | /// |
410 | /// `M` must be an unsigned integer or array of `u8`, matching `simd_bitmask`. |
411 | /// |
412 | /// `T` must be a vector. |
413 | /// |
414 | /// For each element, if the bit in `mask` is `1`, select the element from |
415 | /// `if_true`. If the corresponding bit in `mask` is `0`, select the element from |
416 | /// `if_false`. |
417 | /// |
418 | /// The bitmask bit order matches `simd_bitmask`. |
419 | /// |
420 | /// # Safety |
421 | /// Padding bits must be all zero. |
422 | pub fn simd_select_bitmask<M, T>(m: M, yes: T, no: T) -> T; |
423 | |
424 | /// Elementwise calculates the offset from a pointer vector, potentially wrapping. |
425 | /// |
426 | /// `T` must be a vector of pointers. |
427 | /// |
428 | /// `U` must be a vector of `isize` or `usize` with the same number of elements as `T`. |
429 | /// |
430 | /// Operates as if by `<ptr>::wrapping_offset`. |
431 | pub fn simd_arith_offset<T, U>(ptr: T, offset: U) -> T; |
432 | |
433 | /// Cast a vector of pointers. |
434 | /// |
435 | /// `T` and `U` must be vectors of pointers with the same number of elements. |
436 | pub fn simd_cast_ptr<T, U>(ptr: T) -> U; |
437 | |
438 | /// Expose a vector of pointers as a vector of addresses. |
439 | /// |
440 | /// `T` must be a vector of pointers. |
441 | /// |
442 | /// `U` must be a vector of `usize` with the same length as `T`. |
443 | pub fn simd_expose_addr<T, U>(ptr: T) -> U; |
444 | |
445 | /// Create a vector of pointers from a vector of addresses. |
446 | /// |
447 | /// `T` must be a vector of `usize`. |
448 | /// |
449 | /// `U` must be a vector of pointers, with the same length as `T`. |
450 | pub fn simd_from_exposed_addr<T, U>(addr: T) -> U; |
451 | |
452 | /// Swap bytes of each element. |
453 | /// |
454 | /// `T` must be a vector of integers. |
455 | pub fn simd_bswap<T>(x: T) -> T; |
456 | |
457 | /// Reverse bits of each element. |
458 | /// |
459 | /// `T` must be a vector of integers. |
460 | pub fn simd_bitreverse<T>(x: T) -> T; |
461 | |
462 | /// Count the leading zeros of each element. |
463 | /// |
464 | /// `T` must be a vector of integers. |
465 | pub fn simd_ctlz<T>(x: T) -> T; |
466 | |
467 | /// Count the trailing zeros of each element. |
468 | /// |
469 | /// `T` must be a vector of integers. |
470 | pub fn simd_cttz<T>(x: T) -> T; |
471 | } |
472 | |