Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * simd.h
4 : * Support for platform-specific vector operations.
5 : *
6 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : * src/include/port/simd.h
10 : *
11 : * NOTES
12 : * - VectorN in this file refers to a register where the element operands
13 : * are N bits wide. The vector width is platform-specific, so users that care
14 : * about that will need to inspect "sizeof(VectorN)".
15 : *
16 : *-------------------------------------------------------------------------
17 : */
18 : #ifndef SIMD_H
19 : #define SIMD_H
20 :
21 : #if defined(USE_SSE2)
22 : /*
23 : * We use emmintrin.h rather than the comprehensive header immintrin.h in
24 : * order to exclude extensions beyond SSE2. This is because MSVC, at least,
25 : * will allow the use of intrinsics that haven't been enabled at compile
26 : * time.
27 : */
28 : #include <emmintrin.h>
29 : typedef __m128i Vector8;
30 : typedef __m128i Vector32;
31 :
32 : #elif defined(USE_NEON)
33 : #include <arm_neon.h>
34 : typedef uint8x16_t Vector8;
35 : typedef uint32x4_t Vector32;
36 :
37 : #else
38 : /*
39 : * If no SIMD instructions are available, we can in some cases emulate vector
40 : * operations using bitwise operations on unsigned integers. Note that many
41 : * of the functions in this file presently do not have non-SIMD
42 : * implementations. In particular, none of the functions involving Vector32
43 : * are implemented without SIMD since it's likely not worthwhile to represent
44 : * two 32-bit integers using a uint64.
45 : */
46 : #define USE_NO_SIMD
47 : typedef uint64 Vector8;
48 : #endif
49 :
50 : /* load/store operations */
51 : static inline void vector8_load(Vector8 *v, const uint8 *s);
52 : #ifndef USE_NO_SIMD
53 : static inline void vector32_load(Vector32 *v, const uint32 *s);
54 : #endif
55 :
56 : /* assignment operations */
57 : static inline Vector8 vector8_broadcast(const uint8 c);
58 : #ifndef USE_NO_SIMD
59 : static inline Vector32 vector32_broadcast(const uint32 c);
60 : #endif
61 :
62 : /* element-wise comparisons to a scalar */
63 : static inline bool vector8_has(const Vector8 v, const uint8 c);
64 : static inline bool vector8_has_zero(const Vector8 v);
65 : static inline bool vector8_has_le(const Vector8 v, const uint8 c);
66 : static inline bool vector8_is_highbit_set(const Vector8 v);
67 : #ifndef USE_NO_SIMD
68 : static inline bool vector32_is_highbit_set(const Vector32 v);
69 : static inline uint32 vector8_highbit_mask(const Vector8 v);
70 : #endif
71 :
72 : /* arithmetic operations */
73 : static inline Vector8 vector8_or(const Vector8 v1, const Vector8 v2);
74 : #ifndef USE_NO_SIMD
75 : static inline Vector32 vector32_or(const Vector32 v1, const Vector32 v2);
76 : #endif
77 :
78 : /*
79 : * comparisons between vectors
80 : *
81 : * Note: These return a vector rather than boolean, which is why we don't
82 : * have non-SIMD implementations.
83 : */
84 : #ifndef USE_NO_SIMD
85 : static inline Vector8 vector8_eq(const Vector8 v1, const Vector8 v2);
86 : static inline Vector8 vector8_min(const Vector8 v1, const Vector8 v2);
87 : static inline Vector32 vector32_eq(const Vector32 v1, const Vector32 v2);
88 : #endif
89 :
90 : /*
91 : * Load a chunk of memory into the given vector.
92 : */
93 : static inline void
94 15773401 : vector8_load(Vector8 *v, const uint8 *s)
95 : {
96 : #if defined(USE_SSE2)
97 15773401 : *v = _mm_loadu_si128((const __m128i *) s);
98 : #elif defined(USE_NEON)
99 : *v = vld1q_u8(s);
100 : #else
101 : memcpy(v, s, sizeof(Vector8));
102 : #endif
103 15773401 : }
104 :
105 : #ifndef USE_NO_SIMD
106 : static inline void
107 708 : vector32_load(Vector32 *v, const uint32 *s)
108 : {
109 : #ifdef USE_SSE2
110 708 : *v = _mm_loadu_si128((const __m128i *) s);
111 : #elif defined(USE_NEON)
112 : *v = vld1q_u32(s);
113 : #endif
114 708 : }
115 : #endif /* ! USE_NO_SIMD */
116 :
117 : /*
118 : * Store a vector into the given memory address.
119 : */
120 : #ifndef USE_NO_SIMD
121 : static inline void
122 2317855 : vector8_store(uint8 *s, Vector8 v)
123 : {
124 : #ifdef USE_SSE2
125 : _mm_storeu_si128((Vector8 *) s, v);
126 : #elif defined(USE_NEON)
127 : vst1q_u8(s, v);
128 : #endif
129 2317855 : }
130 : #endif /* ! USE_NO_SIMD */
131 :
132 : /*
133 : * Create a vector with all elements set to the same value.
134 : */
135 : static inline Vector8
136 24553434 : vector8_broadcast(const uint8 c)
137 : {
138 : #if defined(USE_SSE2)
139 49106868 : return _mm_set1_epi8(c);
140 : #elif defined(USE_NEON)
141 : return vdupq_n_u8(c);
142 : #else
143 : return ~UINT64CONST(0) / 0xFF * c;
144 : #endif
145 : }
146 :
147 : #ifndef USE_NO_SIMD
148 : static inline Vector32
149 8284776 : vector32_broadcast(const uint32 c)
150 : {
151 : #ifdef USE_SSE2
152 16569552 : return _mm_set1_epi32(c);
153 : #elif defined(USE_NEON)
154 : return vdupq_n_u32(c);
155 : #endif
156 : }
157 : #endif /* ! USE_NO_SIMD */
158 :
159 : /*
160 : * Return true if any elements in the vector are equal to the given scalar.
161 : */
162 : static inline bool
163 3388387 : vector8_has(const Vector8 v, const uint8 c)
164 : {
165 : bool result;
166 :
167 : /* pre-compute the result for assert checking */
168 : #ifdef USE_ASSERT_CHECKING
169 : bool assert_result = false;
170 :
171 : for (Size i = 0; i < sizeof(Vector8); i++)
172 : {
173 : if (((const uint8 *) &v)[i] == c)
174 : {
175 : assert_result = true;
176 : break;
177 : }
178 : }
179 : #endif /* USE_ASSERT_CHECKING */
180 :
181 : #if defined(USE_NO_SIMD)
182 : /* any bytes in v equal to c will evaluate to zero via XOR */
183 : result = vector8_has_zero(v ^ vector8_broadcast(c));
184 : #else
185 3388387 : result = vector8_is_highbit_set(vector8_eq(v, vector8_broadcast(c)));
186 : #endif
187 :
188 : Assert(assert_result == result);
189 3388387 : return result;
190 : }
191 :
192 : /*
193 : * Convenience function equivalent to vector8_has(v, 0)
194 : */
195 : static inline bool
196 : vector8_has_zero(const Vector8 v)
197 : {
198 : #if defined(USE_NO_SIMD)
199 : /*
200 : * We cannot call vector8_has() here, because that would lead to a
201 : * circular definition.
202 : */
203 : return vector8_has_le(v, 0);
204 : #else
205 : return vector8_has(v, 0);
206 : #endif
207 : }
208 :
209 : /*
210 : * Return true if any elements in the vector are less than or equal to the
211 : * given scalar.
212 : */
213 : static inline bool
214 418842 : vector8_has_le(const Vector8 v, const uint8 c)
215 : {
216 418842 : bool result = false;
217 : #ifdef USE_SSE2
218 : Vector8 umin;
219 : Vector8 cmpe;
220 : #endif
221 :
222 : /* pre-compute the result for assert checking */
223 : #ifdef USE_ASSERT_CHECKING
224 : bool assert_result = false;
225 :
226 : for (Size i = 0; i < sizeof(Vector8); i++)
227 : {
228 : if (((const uint8 *) &v)[i] <= c)
229 : {
230 : assert_result = true;
231 : break;
232 : }
233 : }
234 : #endif /* USE_ASSERT_CHECKING */
235 :
236 : #if defined(USE_NO_SIMD)
237 :
238 : /*
239 : * To find bytes <= c, we can use bitwise operations to find bytes < c+1,
240 : * but it only works if c+1 <= 128 and if the highest bit in v is not set.
241 : * Adapted from
242 : * https://graphics.stanford.edu/~seander/bithacks.html#HasLessInWord
243 : */
244 : if ((int64) v >= 0 && c < 0x80)
245 : result = (v - vector8_broadcast(c + 1)) & ~v & vector8_broadcast(0x80);
246 : else
247 : {
248 : /* one byte at a time */
249 : for (Size i = 0; i < sizeof(Vector8); i++)
250 : {
251 : if (((const uint8 *) &v)[i] <= c)
252 : {
253 : result = true;
254 : break;
255 : }
256 : }
257 : }
258 : #elif defined(USE_SSE2)
259 418842 : umin = vector8_min(v, vector8_broadcast(c));
260 418842 : cmpe = vector8_eq(umin, v);
261 418842 : result = vector8_is_highbit_set(cmpe);
262 : #elif defined(USE_NEON)
263 : result = vminvq_u8(v) <= c;
264 : #endif
265 :
266 : Assert(assert_result == result);
267 418842 : return result;
268 : }
269 :
270 : /*
271 : * Returns true if any elements in the vector are greater than or equal to the
272 : * given scalar.
273 : */
274 : #ifndef USE_NO_SIMD
275 : static inline bool
276 245642 : vector8_has_ge(const Vector8 v, const uint8 c)
277 : {
278 : #ifdef USE_SSE2
279 245642 : Vector8 umax = _mm_max_epu8(v, vector8_broadcast(c));
280 245642 : Vector8 cmpe = vector8_eq(umax, v);
281 :
282 245642 : return vector8_is_highbit_set(cmpe);
283 : #elif defined(USE_NEON)
284 : return vmaxvq_u8(v) >= c;
285 : #endif
286 : }
287 : #endif /* ! USE_NO_SIMD */
288 :
289 : /*
290 : * Return true if the high bit of any element is set
291 : */
292 : static inline bool
293 5898441 : vector8_is_highbit_set(const Vector8 v)
294 : {
295 : #ifdef USE_SSE2
296 5898441 : return _mm_movemask_epi8(v) != 0;
297 : #elif defined(USE_NEON)
298 : return vmaxvq_u8(v) > 0x7F;
299 : #else
300 : return v & vector8_broadcast(0x80);
301 : #endif
302 : }
303 :
304 : /*
305 : * Exactly like vector8_is_highbit_set except for the input type, so it
306 : * looks at each byte separately.
307 : *
308 : * XXX x86 uses the same underlying type for 8-bit, 16-bit, and 32-bit
309 : * integer elements, but Arm does not, hence the need for a separate
310 : * function. We could instead adopt the behavior of Arm's vmaxvq_u32(), i.e.
311 : * check each 32-bit element, but that would require an additional mask
312 : * operation on x86.
313 : */
314 : #ifndef USE_NO_SIMD
315 : static inline bool
316 177 : vector32_is_highbit_set(const Vector32 v)
317 : {
318 : #if defined(USE_NEON)
319 : return vector8_is_highbit_set((Vector8) v);
320 : #else
321 177 : return vector8_is_highbit_set(v);
322 : #endif
323 : }
324 : #endif /* ! USE_NO_SIMD */
325 :
326 : /*
327 : * Return a bitmask formed from the high-bit of each element.
328 : */
329 : #ifndef USE_NO_SIMD
330 : static inline uint32
331 6964372 : vector8_highbit_mask(const Vector8 v)
332 : {
333 : #ifdef USE_SSE2
334 6964372 : return (uint32) _mm_movemask_epi8(v);
335 : #elif defined(USE_NEON)
336 : /*
337 : * Note: It would be faster to use vget_lane_u64 and vshrn_n_u16, but that
338 : * returns a uint64, making it inconvenient to combine mask values from
339 : * multiple vectors.
340 : */
341 : static const uint8 mask[16] = {
342 : 1 << 0, 1 << 1, 1 << 2, 1 << 3,
343 : 1 << 4, 1 << 5, 1 << 6, 1 << 7,
344 : 1 << 0, 1 << 1, 1 << 2, 1 << 3,
345 : 1 << 4, 1 << 5, 1 << 6, 1 << 7,
346 : };
347 :
348 : uint8x16_t masked = vandq_u8(vld1q_u8(mask), (uint8x16_t) vshrq_n_s8((int8x16_t) v, 7));
349 : uint8x16_t maskedhi = vextq_u8(masked, masked, 8);
350 :
351 : return (uint32) vaddvq_u16((uint16x8_t) vzip1q_u8(masked, maskedhi));
352 : #endif
353 : }
354 : #endif /* ! USE_NO_SIMD */
355 :
356 : /*
357 : * Return the bitwise OR of the inputs
358 : */
359 : static inline Vector8
360 7627214 : vector8_or(const Vector8 v1, const Vector8 v2)
361 : {
362 : #ifdef USE_SSE2
363 7627214 : return _mm_or_si128(v1, v2);
364 : #elif defined(USE_NEON)
365 : return vorrq_u8(v1, v2);
366 : #else
367 : return v1 | v2;
368 : #endif
369 : }
370 :
371 : #ifndef USE_NO_SIMD
372 : static inline Vector32
373 531 : vector32_or(const Vector32 v1, const Vector32 v2)
374 : {
375 : #ifdef USE_SSE2
376 531 : return _mm_or_si128(v1, v2);
377 : #elif defined(USE_NEON)
378 : return vorrq_u32(v1, v2);
379 : #endif
380 : }
381 : #endif /* ! USE_NO_SIMD */
382 :
383 : /*
384 : * Return the bitwise AND of the inputs.
385 : */
386 : #ifndef USE_NO_SIMD
387 : static inline Vector8
388 5618278 : vector8_and(const Vector8 v1, const Vector8 v2)
389 : {
390 : #ifdef USE_SSE2
391 5618278 : return _mm_and_si128(v1, v2);
392 : #elif defined(USE_NEON)
393 : return vandq_u8(v1, v2);
394 : #endif
395 : }
396 : #endif /* ! USE_NO_SIMD */
397 :
398 : /*
399 : * Return the result of adding the respective elements of the input vectors.
400 : */
401 : #ifndef USE_NO_SIMD
402 : static inline Vector8
403 4881352 : vector8_add(const Vector8 v1, const Vector8 v2)
404 : {
405 : #ifdef USE_SSE2
406 4881352 : return _mm_add_epi8(v1, v2);
407 : #elif defined(USE_NEON)
408 : return vaddq_u8(v1, v2);
409 : #endif
410 : }
411 : #endif /* ! USE_NO_SIMD */
412 :
413 : /*
414 : * Return the result of subtracting the respective elements of the input
415 : * vectors using signed saturation (i.e., if the operation would yield a value
416 : * less than -128, -128 is returned instead). For more information on
417 : * saturation arithmetic, see
418 : * https://en.wikipedia.org/wiki/Saturation_arithmetic
419 : */
420 : #ifndef USE_NO_SIMD
421 : static inline Vector8
422 245642 : vector8_issub(const Vector8 v1, const Vector8 v2)
423 : {
424 : #ifdef USE_SSE2
425 245642 : return _mm_subs_epi8(v1, v2);
426 : #elif defined(USE_NEON)
427 : return (Vector8) vqsubq_s8((int8x16_t) v1, (int8x16_t) v2);
428 : #endif
429 : }
430 : #endif /* ! USE_NO_SIMD */
431 :
432 : /*
433 : * Return a vector with all bits set in each lane where the corresponding
434 : * lanes in the inputs are equal.
435 : */
436 : #ifndef USE_NO_SIMD
437 : static inline Vector8
438 14708029 : vector8_eq(const Vector8 v1, const Vector8 v2)
439 : {
440 : #ifdef USE_SSE2
441 14708029 : return _mm_cmpeq_epi8(v1, v2);
442 : #elif defined(USE_NEON)
443 : return vceqq_u8(v1, v2);
444 : #endif
445 : }
446 : #endif /* ! USE_NO_SIMD */
447 :
448 : #ifndef USE_NO_SIMD
449 : static inline Vector32
450 708 : vector32_eq(const Vector32 v1, const Vector32 v2)
451 : {
452 : #ifdef USE_SSE2
453 708 : return _mm_cmpeq_epi32(v1, v2);
454 : #elif defined(USE_NEON)
455 : return vceqq_u32(v1, v2);
456 : #endif
457 : }
458 : #endif /* ! USE_NO_SIMD */
459 :
460 : /*
461 : * Return a vector with all bits set for each lane of v1 that is greater than
462 : * the corresponding lane of v2. NB: The comparison treats the elements as
463 : * signed.
464 : */
465 : #ifndef USE_NO_SIMD
466 : static inline Vector8
467 2931960 : vector8_gt(const Vector8 v1, const Vector8 v2)
468 : {
469 : #ifdef USE_SSE2
470 2931960 : return _mm_cmpgt_epi8(v1, v2);
471 : #elif defined(USE_NEON)
472 : return vcgtq_s8((int8x16_t) v1, (int8x16_t) v2);
473 : #endif
474 : }
475 : #endif /* ! USE_NO_SIMD */
476 :
477 : /*
478 : * Given two vectors, return a vector with the minimum element of each.
479 : */
480 : #ifndef USE_NO_SIMD
481 : static inline Vector8
482 523526 : vector8_min(const Vector8 v1, const Vector8 v2)
483 : {
484 : #ifdef USE_SSE2
485 523526 : return _mm_min_epu8(v1, v2);
486 : #elif defined(USE_NEON)
487 : return vminq_u8(v1, v2);
488 : #endif
489 : }
490 : #endif /* ! USE_NO_SIMD */
491 :
492 : /*
493 : * Interleave elements of low halves (e.g., for SSE2, bits 0-63) of given
494 : * vectors. Bytes 0, 2, 4, etc. use v1, and bytes 1, 3, 5, etc. use v2.
495 : */
496 : #ifndef USE_NO_SIMD
497 : static inline Vector8
498 1711622 : vector8_interleave_low(const Vector8 v1, const Vector8 v2)
499 : {
500 : #ifdef USE_SSE2
501 1711622 : return _mm_unpacklo_epi8(v1, v2);
502 : #elif defined(USE_NEON)
503 : return vzip1q_u8(v1, v2);
504 : #endif
505 : }
506 : #endif /* ! USE_NO_SIMD */
507 :
508 : /*
509 : * Interleave elements of high halves (e.g., for SSE2, bits 64-127) of given
510 : * vectors. Bytes 0, 2, 4, etc. use v1, and bytes 1, 3, 5, etc. use v2.
511 : */
512 : #ifndef USE_NO_SIMD
513 : static inline Vector8
514 1097517 : vector8_interleave_high(const Vector8 v1, const Vector8 v2)
515 : {
516 : #ifdef USE_SSE2
517 1097517 : return _mm_unpackhi_epi8(v1, v2);
518 : #elif defined(USE_NEON)
519 : return vzip2q_u8(v1, v2);
520 : #endif
521 : }
522 : #endif /* ! USE_NO_SIMD */
523 :
524 : /*
525 : * Pack 16-bit elements in the given vectors into a single vector of 8-bit
526 : * elements. The first half of the return vector (e.g., for SSE2, bits 0-63)
527 : * uses v1, and the second half (e.g., for SSE2, bits 64-127) uses v2.
528 : *
529 : * NB: The upper 8-bits of each 16-bit element must be zeros, else this will
530 : * produce different results on different architectures.
531 : */
532 : #ifndef USE_NO_SIMD
533 : static inline Vector8
534 122821 : vector8_pack_16(const Vector8 v1, const Vector8 v2)
535 : {
536 : Vector8 mask PG_USED_FOR_ASSERTS_ONLY;
537 :
538 122821 : mask = vector8_interleave_low(vector8_broadcast(0), vector8_broadcast(0xff));
539 : Assert(!vector8_has_ge(vector8_and(v1, mask), 1));
540 : Assert(!vector8_has_ge(vector8_and(v2, mask), 1));
541 : #ifdef USE_SSE2
542 122821 : return _mm_packus_epi16(v1, v2);
543 : #elif defined(USE_NEON)
544 : return vuzp1q_u8(v1, v2);
545 : #endif
546 : }
547 : #endif /* ! USE_NO_SIMD */
548 :
549 : /*
550 : * Unsigned shift left of each 32-bit element in the vector by "i" bits.
551 : *
552 : * XXX AArch64 requires an integer literal, so we have to list all expected
553 : * values of "i" from all callers in a switch statement. If you add a new
554 : * caller, be sure your expected values of "i" are handled.
555 : */
556 : #ifndef USE_NO_SIMD
557 : static inline Vector8
558 245642 : vector8_shift_left(const Vector8 v1, int i)
559 : {
560 : #ifdef USE_SSE2
561 245642 : return _mm_slli_epi32(v1, i);
562 : #elif defined(USE_NEON)
563 : switch (i)
564 : {
565 : case 4:
566 : return (Vector8) vshlq_n_u32((Vector32) v1, 4);
567 : default:
568 : Assert(false);
569 : return vector8_broadcast(0);
570 : }
571 : #endif
572 : }
573 : #endif /* ! USE_NO_SIMD */
574 :
575 : /*
576 : * Unsigned shift right of each 32-bit element in the vector by "i" bits.
577 : *
578 : * XXX AArch64 requires an integer literal, so we have to list all expected
579 : * values of "i" from all callers in a switch statement. If you add a new
580 : * caller, be sure your expected values of "i" are handled.
581 : */
582 : #ifndef USE_NO_SIMD
583 : static inline Vector8
584 1343159 : vector8_shift_right(const Vector8 v1, int i)
585 : {
586 : #ifdef USE_SSE2
587 1343159 : return _mm_srli_epi32(v1, i);
588 : #elif defined(USE_NEON)
589 : switch (i)
590 : {
591 : case 4:
592 : return (Vector8) vshrq_n_u32((Vector32) v1, 4);
593 : case 8:
594 : return (Vector8) vshrq_n_u32((Vector32) v1, 8);
595 : default:
596 : Assert(false);
597 : return vector8_broadcast(0);
598 : }
599 : #endif
600 : }
601 : #endif /* ! USE_NO_SIMD */
602 :
603 : #endif /* SIMD_H */
|
