/src/duckdb/third_party/fastpforlib/bitpackinghelpers.h
Line | Count | Source (jump to first uncovered line) |
1 | | /** |
2 | | * This code is released under the |
3 | | * Apache License Version 2.0 http://www.apache.org/licenses/. |
4 | | * |
5 | | * (c) Daniel Lemire, http://lemire.me/en/ |
6 | | */ |
7 | | #pragma once |
8 | | #include "bitpacking.h" |
9 | | |
10 | | |
11 | | #include <stdexcept> |
12 | | |
13 | | namespace duckdb_fastpforlib { |
14 | | |
15 | | namespace internal { |
16 | | |
17 | | // Note that this only packs 8 values |
18 | 0 | inline void fastunpack_quarter(const uint8_t *__restrict in, uint8_t *__restrict out, const uint32_t bit) { |
19 | | // Could have used function pointers instead of switch. |
20 | | // Switch calls do offer the compiler more opportunities for optimization in |
21 | | // theory. In this case, it makes no difference with a good compiler. |
22 | 0 | switch (bit) { |
23 | 0 | case 0: |
24 | 0 | internal::__fastunpack0(in, out); |
25 | 0 | break; |
26 | 0 | case 1: |
27 | 0 | internal::__fastunpack1(in, out); |
28 | 0 | break; |
29 | 0 | case 2: |
30 | 0 | internal::__fastunpack2(in, out); |
31 | 0 | break; |
32 | 0 | case 3: |
33 | 0 | internal::__fastunpack3(in, out); |
34 | 0 | break; |
35 | 0 | case 4: |
36 | 0 | internal::__fastunpack4(in, out); |
37 | 0 | break; |
38 | 0 | case 5: |
39 | 0 | internal::__fastunpack5(in, out); |
40 | 0 | break; |
41 | 0 | case 6: |
42 | 0 | internal::__fastunpack6(in, out); |
43 | 0 | break; |
44 | 0 | case 7: |
45 | 0 | internal::__fastunpack7(in, out); |
46 | 0 | break; |
47 | 0 | case 8: |
48 | 0 | internal::__fastunpack8(in, out); |
49 | 0 | break; |
50 | 0 | default: |
51 | 0 | throw std::logic_error("Invalid bit width for bitpacking"); |
52 | 0 | } |
53 | 0 | } |
54 | | |
55 | | // Note that this only packs 8 values |
56 | 0 | inline void fastpack_quarter(const uint8_t *__restrict in, uint8_t *__restrict out, const uint32_t bit) { |
57 | | // Could have used function pointers instead of switch. |
58 | | // Switch calls do offer the compiler more opportunities for optimization in |
59 | | // theory. In this case, it makes no difference with a good compiler. |
60 | 0 | switch (bit) { |
61 | 0 | case 0: |
62 | 0 | internal::__fastpack0(in, out); |
63 | 0 | break; |
64 | 0 | case 1: |
65 | 0 | internal::__fastpack1(in, out); |
66 | 0 | break; |
67 | 0 | case 2: |
68 | 0 | internal::__fastpack2(in, out); |
69 | 0 | break; |
70 | 0 | case 3: |
71 | 0 | internal::__fastpack3(in, out); |
72 | 0 | break; |
73 | 0 | case 4: |
74 | 0 | internal::__fastpack4(in, out); |
75 | 0 | break; |
76 | 0 | case 5: |
77 | 0 | internal::__fastpack5(in, out); |
78 | 0 | break; |
79 | 0 | case 6: |
80 | 0 | internal::__fastpack6(in, out); |
81 | 0 | break; |
82 | 0 | case 7: |
83 | 0 | internal::__fastpack7(in, out); |
84 | 0 | break; |
85 | 0 | case 8: |
86 | 0 | internal::__fastpack8(in, out); |
87 | 0 | break; |
88 | 0 | default: |
89 | 0 | throw std::logic_error("Invalid bit width for bitpacking"); |
90 | 0 | } |
91 | 0 | } |
92 | | |
93 | | // Note that this only packs 16 values |
94 | 0 | inline void fastunpack_half(const uint16_t *__restrict in, uint16_t *__restrict out, const uint32_t bit) { |
95 | | // Could have used function pointers instead of switch. |
96 | | // Switch calls do offer the compiler more opportunities for optimization in |
97 | | // theory. In this case, it makes no difference with a good compiler. |
98 | 0 | switch (bit) { |
99 | 0 | case 0: |
100 | 0 | internal::__fastunpack0(in, out); |
101 | 0 | break; |
102 | 0 | case 1: |
103 | 0 | internal::__fastunpack1(in, out); |
104 | 0 | break; |
105 | 0 | case 2: |
106 | 0 | internal::__fastunpack2(in, out); |
107 | 0 | break; |
108 | 0 | case 3: |
109 | 0 | internal::__fastunpack3(in, out); |
110 | 0 | break; |
111 | 0 | case 4: |
112 | 0 | internal::__fastunpack4(in, out); |
113 | 0 | break; |
114 | 0 | case 5: |
115 | 0 | internal::__fastunpack5(in, out); |
116 | 0 | break; |
117 | 0 | case 6: |
118 | 0 | internal::__fastunpack6(in, out); |
119 | 0 | break; |
120 | 0 | case 7: |
121 | 0 | internal::__fastunpack7(in, out); |
122 | 0 | break; |
123 | 0 | case 8: |
124 | 0 | internal::__fastunpack8(in, out); |
125 | 0 | break; |
126 | 0 | case 9: |
127 | 0 | internal::__fastunpack9(in, out); |
128 | 0 | break; |
129 | 0 | case 10: |
130 | 0 | internal::__fastunpack10(in, out); |
131 | 0 | break; |
132 | 0 | case 11: |
133 | 0 | internal::__fastunpack11(in, out); |
134 | 0 | break; |
135 | 0 | case 12: |
136 | 0 | internal::__fastunpack12(in, out); |
137 | 0 | break; |
138 | 0 | case 13: |
139 | 0 | internal::__fastunpack13(in, out); |
140 | 0 | break; |
141 | 0 | case 14: |
142 | 0 | internal::__fastunpack14(in, out); |
143 | 0 | break; |
144 | 0 | case 15: |
145 | 0 | internal::__fastunpack15(in, out); |
146 | 0 | break; |
147 | 0 | case 16: |
148 | 0 | internal::__fastunpack16(in, out); |
149 | 0 | break; |
150 | 0 | default: |
151 | 0 | throw std::logic_error("Invalid bit width for bitpacking"); |
152 | 0 | } |
153 | 0 | } |
154 | | |
155 | | // Note that this only packs 16 values |
156 | 0 | inline void fastpack_half(const uint16_t *__restrict in, uint16_t *__restrict out, const uint32_t bit) { |
157 | | // Could have used function pointers instead of switch. |
158 | | // Switch calls do offer the compiler more opportunities for optimization in |
159 | | // theory. In this case, it makes no difference with a good compiler. |
160 | 0 | switch (bit) { |
161 | 0 | case 0: |
162 | 0 | internal::__fastpack0(in, out); |
163 | 0 | break; |
164 | 0 | case 1: |
165 | 0 | internal::__fastpack1(in, out); |
166 | 0 | break; |
167 | 0 | case 2: |
168 | 0 | internal::__fastpack2(in, out); |
169 | 0 | break; |
170 | 0 | case 3: |
171 | 0 | internal::__fastpack3(in, out); |
172 | 0 | break; |
173 | 0 | case 4: |
174 | 0 | internal::__fastpack4(in, out); |
175 | 0 | break; |
176 | 0 | case 5: |
177 | 0 | internal::__fastpack5(in, out); |
178 | 0 | break; |
179 | 0 | case 6: |
180 | 0 | internal::__fastpack6(in, out); |
181 | 0 | break; |
182 | 0 | case 7: |
183 | 0 | internal::__fastpack7(in, out); |
184 | 0 | break; |
185 | 0 | case 8: |
186 | 0 | internal::__fastpack8(in, out); |
187 | 0 | break; |
188 | 0 | case 9: |
189 | 0 | internal::__fastpack9(in, out); |
190 | 0 | break; |
191 | 0 | case 10: |
192 | 0 | internal::__fastpack10(in, out); |
193 | 0 | break; |
194 | 0 | case 11: |
195 | 0 | internal::__fastpack11(in, out); |
196 | 0 | break; |
197 | 0 | case 12: |
198 | 0 | internal::__fastpack12(in, out); |
199 | 0 | break; |
200 | 0 | case 13: |
201 | 0 | internal::__fastpack13(in, out); |
202 | 0 | break; |
203 | 0 | case 14: |
204 | 0 | internal::__fastpack14(in, out); |
205 | 0 | break; |
206 | 0 | case 15: |
207 | 0 | internal::__fastpack15(in, out); |
208 | 0 | break; |
209 | 0 | case 16: |
210 | 0 | internal::__fastpack16(in, out); |
211 | 0 | break; |
212 | 0 | default: |
213 | 0 | throw std::logic_error("Invalid bit width for bitpacking"); |
214 | 0 | } |
215 | 0 | } |
216 | | } |
217 | | |
218 | 0 | inline void fastunpack(const uint8_t *__restrict in, uint8_t *__restrict out, const uint32_t bit) { |
219 | 0 | for (uint8_t i = 0; i < 4; i++) { |
220 | 0 | internal::fastunpack_quarter(in + (i*bit), out+(i*8), bit); |
221 | 0 | } |
222 | 0 | } |
223 | | |
224 | 0 | inline void fastunpack(const uint16_t *__restrict in, uint16_t *__restrict out, const uint32_t bit) { |
225 | 0 | internal::fastunpack_half(in, out, bit); |
226 | 0 | internal::fastunpack_half(in + bit, out+16, bit); |
227 | 0 | } |
228 | | |
229 | | inline void fastunpack(const uint32_t *__restrict in, |
230 | 0 | uint32_t *__restrict out, const uint32_t bit) { |
231 | | // Could have used function pointers instead of switch. |
232 | | // Switch calls do offer the compiler more opportunities for optimization in |
233 | | // theory. In this case, it makes no difference with a good compiler. |
234 | 0 | switch (bit) { |
235 | 0 | case 0: |
236 | 0 | internal::__fastunpack0(in, out); |
237 | 0 | break; |
238 | 0 | case 1: |
239 | 0 | internal::__fastunpack1(in, out); |
240 | 0 | break; |
241 | 0 | case 2: |
242 | 0 | internal::__fastunpack2(in, out); |
243 | 0 | break; |
244 | 0 | case 3: |
245 | 0 | internal::__fastunpack3(in, out); |
246 | 0 | break; |
247 | 0 | case 4: |
248 | 0 | internal::__fastunpack4(in, out); |
249 | 0 | break; |
250 | 0 | case 5: |
251 | 0 | internal::__fastunpack5(in, out); |
252 | 0 | break; |
253 | 0 | case 6: |
254 | 0 | internal::__fastunpack6(in, out); |
255 | 0 | break; |
256 | 0 | case 7: |
257 | 0 | internal::__fastunpack7(in, out); |
258 | 0 | break; |
259 | 0 | case 8: |
260 | 0 | internal::__fastunpack8(in, out); |
261 | 0 | break; |
262 | 0 | case 9: |
263 | 0 | internal::__fastunpack9(in, out); |
264 | 0 | break; |
265 | 0 | case 10: |
266 | 0 | internal::__fastunpack10(in, out); |
267 | 0 | break; |
268 | 0 | case 11: |
269 | 0 | internal::__fastunpack11(in, out); |
270 | 0 | break; |
271 | 0 | case 12: |
272 | 0 | internal::__fastunpack12(in, out); |
273 | 0 | break; |
274 | 0 | case 13: |
275 | 0 | internal::__fastunpack13(in, out); |
276 | 0 | break; |
277 | 0 | case 14: |
278 | 0 | internal::__fastunpack14(in, out); |
279 | 0 | break; |
280 | 0 | case 15: |
281 | 0 | internal::__fastunpack15(in, out); |
282 | 0 | break; |
283 | 0 | case 16: |
284 | 0 | internal::__fastunpack16(in, out); |
285 | 0 | break; |
286 | 0 | case 17: |
287 | 0 | internal::__fastunpack17(in, out); |
288 | 0 | break; |
289 | 0 | case 18: |
290 | 0 | internal::__fastunpack18(in, out); |
291 | 0 | break; |
292 | 0 | case 19: |
293 | 0 | internal::__fastunpack19(in, out); |
294 | 0 | break; |
295 | 0 | case 20: |
296 | 0 | internal::__fastunpack20(in, out); |
297 | 0 | break; |
298 | 0 | case 21: |
299 | 0 | internal::__fastunpack21(in, out); |
300 | 0 | break; |
301 | 0 | case 22: |
302 | 0 | internal::__fastunpack22(in, out); |
303 | 0 | break; |
304 | 0 | case 23: |
305 | 0 | internal::__fastunpack23(in, out); |
306 | 0 | break; |
307 | 0 | case 24: |
308 | 0 | internal::__fastunpack24(in, out); |
309 | 0 | break; |
310 | 0 | case 25: |
311 | 0 | internal::__fastunpack25(in, out); |
312 | 0 | break; |
313 | 0 | case 26: |
314 | 0 | internal::__fastunpack26(in, out); |
315 | 0 | break; |
316 | 0 | case 27: |
317 | 0 | internal::__fastunpack27(in, out); |
318 | 0 | break; |
319 | 0 | case 28: |
320 | 0 | internal::__fastunpack28(in, out); |
321 | 0 | break; |
322 | 0 | case 29: |
323 | 0 | internal::__fastunpack29(in, out); |
324 | 0 | break; |
325 | 0 | case 30: |
326 | 0 | internal::__fastunpack30(in, out); |
327 | 0 | break; |
328 | 0 | case 31: |
329 | 0 | internal::__fastunpack31(in, out); |
330 | 0 | break; |
331 | 0 | case 32: |
332 | 0 | internal::__fastunpack32(in, out); |
333 | 0 | break; |
334 | 0 | default: |
335 | 0 | throw std::logic_error("Invalid bit width for bitpacking"); |
336 | 0 | } |
337 | 0 | } |
338 | | |
339 | | inline void fastunpack(const uint32_t *__restrict in, |
340 | 0 | uint64_t *__restrict out, const uint32_t bit) { |
341 | | // Could have used function pointers instead of switch. |
342 | | // Switch calls do offer the compiler more opportunities for optimization in |
343 | | // theory. In this case, it makes no difference with a good compiler. |
344 | 0 | switch (bit) { |
345 | 0 | case 0: |
346 | 0 | internal::__fastunpack0(in, out); |
347 | 0 | break; |
348 | 0 | case 1: |
349 | 0 | internal::__fastunpack1(in, out); |
350 | 0 | break; |
351 | 0 | case 2: |
352 | 0 | internal::__fastunpack2(in, out); |
353 | 0 | break; |
354 | 0 | case 3: |
355 | 0 | internal::__fastunpack3(in, out); |
356 | 0 | break; |
357 | 0 | case 4: |
358 | 0 | internal::__fastunpack4(in, out); |
359 | 0 | break; |
360 | 0 | case 5: |
361 | 0 | internal::__fastunpack5(in, out); |
362 | 0 | break; |
363 | 0 | case 6: |
364 | 0 | internal::__fastunpack6(in, out); |
365 | 0 | break; |
366 | 0 | case 7: |
367 | 0 | internal::__fastunpack7(in, out); |
368 | 0 | break; |
369 | 0 | case 8: |
370 | 0 | internal::__fastunpack8(in, out); |
371 | 0 | break; |
372 | 0 | case 9: |
373 | 0 | internal::__fastunpack9(in, out); |
374 | 0 | break; |
375 | 0 | case 10: |
376 | 0 | internal::__fastunpack10(in, out); |
377 | 0 | break; |
378 | 0 | case 11: |
379 | 0 | internal::__fastunpack11(in, out); |
380 | 0 | break; |
381 | 0 | case 12: |
382 | 0 | internal::__fastunpack12(in, out); |
383 | 0 | break; |
384 | 0 | case 13: |
385 | 0 | internal::__fastunpack13(in, out); |
386 | 0 | break; |
387 | 0 | case 14: |
388 | 0 | internal::__fastunpack14(in, out); |
389 | 0 | break; |
390 | 0 | case 15: |
391 | 0 | internal::__fastunpack15(in, out); |
392 | 0 | break; |
393 | 0 | case 16: |
394 | 0 | internal::__fastunpack16(in, out); |
395 | 0 | break; |
396 | 0 | case 17: |
397 | 0 | internal::__fastunpack17(in, out); |
398 | 0 | break; |
399 | 0 | case 18: |
400 | 0 | internal::__fastunpack18(in, out); |
401 | 0 | break; |
402 | 0 | case 19: |
403 | 0 | internal::__fastunpack19(in, out); |
404 | 0 | break; |
405 | 0 | case 20: |
406 | 0 | internal::__fastunpack20(in, out); |
407 | 0 | break; |
408 | 0 | case 21: |
409 | 0 | internal::__fastunpack21(in, out); |
410 | 0 | break; |
411 | 0 | case 22: |
412 | 0 | internal::__fastunpack22(in, out); |
413 | 0 | break; |
414 | 0 | case 23: |
415 | 0 | internal::__fastunpack23(in, out); |
416 | 0 | break; |
417 | 0 | case 24: |
418 | 0 | internal::__fastunpack24(in, out); |
419 | 0 | break; |
420 | 0 | case 25: |
421 | 0 | internal::__fastunpack25(in, out); |
422 | 0 | break; |
423 | 0 | case 26: |
424 | 0 | internal::__fastunpack26(in, out); |
425 | 0 | break; |
426 | 0 | case 27: |
427 | 0 | internal::__fastunpack27(in, out); |
428 | 0 | break; |
429 | 0 | case 28: |
430 | 0 | internal::__fastunpack28(in, out); |
431 | 0 | break; |
432 | 0 | case 29: |
433 | 0 | internal::__fastunpack29(in, out); |
434 | 0 | break; |
435 | 0 | case 30: |
436 | 0 | internal::__fastunpack30(in, out); |
437 | 0 | break; |
438 | 0 | case 31: |
439 | 0 | internal::__fastunpack31(in, out); |
440 | 0 | break; |
441 | 0 | case 32: |
442 | 0 | internal::__fastunpack32(in, out); |
443 | 0 | break; |
444 | 0 | case 33: |
445 | 0 | internal::__fastunpack33(in, out); |
446 | 0 | break; |
447 | 0 | case 34: |
448 | 0 | internal::__fastunpack34(in, out); |
449 | 0 | break; |
450 | 0 | case 35: |
451 | 0 | internal::__fastunpack35(in, out); |
452 | 0 | break; |
453 | 0 | case 36: |
454 | 0 | internal::__fastunpack36(in, out); |
455 | 0 | break; |
456 | 0 | case 37: |
457 | 0 | internal::__fastunpack37(in, out); |
458 | 0 | break; |
459 | 0 | case 38: |
460 | 0 | internal::__fastunpack38(in, out); |
461 | 0 | break; |
462 | 0 | case 39: |
463 | 0 | internal::__fastunpack39(in, out); |
464 | 0 | break; |
465 | 0 | case 40: |
466 | 0 | internal::__fastunpack40(in, out); |
467 | 0 | break; |
468 | 0 | case 41: |
469 | 0 | internal::__fastunpack41(in, out); |
470 | 0 | break; |
471 | 0 | case 42: |
472 | 0 | internal::__fastunpack42(in, out); |
473 | 0 | break; |
474 | 0 | case 43: |
475 | 0 | internal::__fastunpack43(in, out); |
476 | 0 | break; |
477 | 0 | case 44: |
478 | 0 | internal::__fastunpack44(in, out); |
479 | 0 | break; |
480 | 0 | case 45: |
481 | 0 | internal::__fastunpack45(in, out); |
482 | 0 | break; |
483 | 0 | case 46: |
484 | 0 | internal::__fastunpack46(in, out); |
485 | 0 | break; |
486 | 0 | case 47: |
487 | 0 | internal::__fastunpack47(in, out); |
488 | 0 | break; |
489 | 0 | case 48: |
490 | 0 | internal::__fastunpack48(in, out); |
491 | 0 | break; |
492 | 0 | case 49: |
493 | 0 | internal::__fastunpack49(in, out); |
494 | 0 | break; |
495 | 0 | case 50: |
496 | 0 | internal::__fastunpack50(in, out); |
497 | 0 | break; |
498 | 0 | case 51: |
499 | 0 | internal::__fastunpack51(in, out); |
500 | 0 | break; |
501 | 0 | case 52: |
502 | 0 | internal::__fastunpack52(in, out); |
503 | 0 | break; |
504 | 0 | case 53: |
505 | 0 | internal::__fastunpack53(in, out); |
506 | 0 | break; |
507 | 0 | case 54: |
508 | 0 | internal::__fastunpack54(in, out); |
509 | 0 | break; |
510 | 0 | case 55: |
511 | 0 | internal::__fastunpack55(in, out); |
512 | 0 | break; |
513 | 0 | case 56: |
514 | 0 | internal::__fastunpack56(in, out); |
515 | 0 | break; |
516 | 0 | case 57: |
517 | 0 | internal::__fastunpack57(in, out); |
518 | 0 | break; |
519 | 0 | case 58: |
520 | 0 | internal::__fastunpack58(in, out); |
521 | 0 | break; |
522 | 0 | case 59: |
523 | 0 | internal::__fastunpack59(in, out); |
524 | 0 | break; |
525 | 0 | case 60: |
526 | 0 | internal::__fastunpack60(in, out); |
527 | 0 | break; |
528 | 0 | case 61: |
529 | 0 | internal::__fastunpack61(in, out); |
530 | 0 | break; |
531 | 0 | case 62: |
532 | 0 | internal::__fastunpack62(in, out); |
533 | 0 | break; |
534 | 0 | case 63: |
535 | 0 | internal::__fastunpack63(in, out); |
536 | 0 | break; |
537 | 0 | case 64: |
538 | 0 | internal::__fastunpack64(in, out); |
539 | 0 | break; |
540 | 0 | default: |
541 | 0 | throw std::logic_error("Invalid bit width for bitpacking"); |
542 | 0 | } |
543 | 0 | } |
544 | | |
545 | 0 | inline void fastpack(const uint8_t *__restrict in, uint8_t *__restrict out, const uint32_t bit) { |
546 | |
|
547 | 0 | for (uint8_t i = 0; i < 4; i++) { |
548 | 0 | internal::fastpack_quarter(in+(i*8), out + (i*bit), bit); |
549 | 0 | } |
550 | 0 | } |
551 | | |
552 | 0 | inline void fastpack(const uint16_t *__restrict in, uint16_t *__restrict out, const uint32_t bit) { |
553 | 0 | internal::fastpack_half(in, out, bit); |
554 | 0 | internal::fastpack_half(in+16, out + bit, bit); |
555 | 0 | } |
556 | | |
557 | | inline void fastpack(const uint32_t *__restrict in, |
558 | 0 | uint32_t *__restrict out, const uint32_t bit) { |
559 | | // Could have used function pointers instead of switch. |
560 | | // Switch calls do offer the compiler more opportunities for optimization in |
561 | | // theory. In this case, it makes no difference with a good compiler. |
562 | 0 | switch (bit) { |
563 | 0 | case 0: |
564 | 0 | internal::__fastpack0(in, out); |
565 | 0 | break; |
566 | 0 | case 1: |
567 | 0 | internal::__fastpack1(in, out); |
568 | 0 | break; |
569 | 0 | case 2: |
570 | 0 | internal::__fastpack2(in, out); |
571 | 0 | break; |
572 | 0 | case 3: |
573 | 0 | internal::__fastpack3(in, out); |
574 | 0 | break; |
575 | 0 | case 4: |
576 | 0 | internal::__fastpack4(in, out); |
577 | 0 | break; |
578 | 0 | case 5: |
579 | 0 | internal::__fastpack5(in, out); |
580 | 0 | break; |
581 | 0 | case 6: |
582 | 0 | internal::__fastpack6(in, out); |
583 | 0 | break; |
584 | 0 | case 7: |
585 | 0 | internal::__fastpack7(in, out); |
586 | 0 | break; |
587 | 0 | case 8: |
588 | 0 | internal::__fastpack8(in, out); |
589 | 0 | break; |
590 | 0 | case 9: |
591 | 0 | internal::__fastpack9(in, out); |
592 | 0 | break; |
593 | 0 | case 10: |
594 | 0 | internal::__fastpack10(in, out); |
595 | 0 | break; |
596 | 0 | case 11: |
597 | 0 | internal::__fastpack11(in, out); |
598 | 0 | break; |
599 | 0 | case 12: |
600 | 0 | internal::__fastpack12(in, out); |
601 | 0 | break; |
602 | 0 | case 13: |
603 | 0 | internal::__fastpack13(in, out); |
604 | 0 | break; |
605 | 0 | case 14: |
606 | 0 | internal::__fastpack14(in, out); |
607 | 0 | break; |
608 | 0 | case 15: |
609 | 0 | internal::__fastpack15(in, out); |
610 | 0 | break; |
611 | 0 | case 16: |
612 | 0 | internal::__fastpack16(in, out); |
613 | 0 | break; |
614 | 0 | case 17: |
615 | 0 | internal::__fastpack17(in, out); |
616 | 0 | break; |
617 | 0 | case 18: |
618 | 0 | internal::__fastpack18(in, out); |
619 | 0 | break; |
620 | 0 | case 19: |
621 | 0 | internal::__fastpack19(in, out); |
622 | 0 | break; |
623 | 0 | case 20: |
624 | 0 | internal::__fastpack20(in, out); |
625 | 0 | break; |
626 | 0 | case 21: |
627 | 0 | internal::__fastpack21(in, out); |
628 | 0 | break; |
629 | 0 | case 22: |
630 | 0 | internal::__fastpack22(in, out); |
631 | 0 | break; |
632 | 0 | case 23: |
633 | 0 | internal::__fastpack23(in, out); |
634 | 0 | break; |
635 | 0 | case 24: |
636 | 0 | internal::__fastpack24(in, out); |
637 | 0 | break; |
638 | 0 | case 25: |
639 | 0 | internal::__fastpack25(in, out); |
640 | 0 | break; |
641 | 0 | case 26: |
642 | 0 | internal::__fastpack26(in, out); |
643 | 0 | break; |
644 | 0 | case 27: |
645 | 0 | internal::__fastpack27(in, out); |
646 | 0 | break; |
647 | 0 | case 28: |
648 | 0 | internal::__fastpack28(in, out); |
649 | 0 | break; |
650 | 0 | case 29: |
651 | 0 | internal::__fastpack29(in, out); |
652 | 0 | break; |
653 | 0 | case 30: |
654 | 0 | internal::__fastpack30(in, out); |
655 | 0 | break; |
656 | 0 | case 31: |
657 | 0 | internal::__fastpack31(in, out); |
658 | 0 | break; |
659 | 0 | case 32: |
660 | 0 | internal::__fastpack32(in, out); |
661 | 0 | break; |
662 | 0 | default: |
663 | 0 | throw std::logic_error("Invalid bit width for bitpacking"); |
664 | 0 | } |
665 | 0 | } |
666 | | |
667 | | inline void fastpack(const uint64_t *__restrict in, |
668 | 0 | uint32_t *__restrict out, const uint32_t bit) { |
669 | 0 | switch (bit) { |
670 | 0 | case 0: |
671 | 0 | internal::__fastpack0(in, out); |
672 | 0 | break; |
673 | 0 | case 1: |
674 | 0 | internal::__fastpack1(in, out); |
675 | 0 | break; |
676 | 0 | case 2: |
677 | 0 | internal::__fastpack2(in, out); |
678 | 0 | break; |
679 | 0 | case 3: |
680 | 0 | internal::__fastpack3(in, out); |
681 | 0 | break; |
682 | 0 | case 4: |
683 | 0 | internal::__fastpack4(in, out); |
684 | 0 | break; |
685 | 0 | case 5: |
686 | 0 | internal::__fastpack5(in, out); |
687 | 0 | break; |
688 | 0 | case 6: |
689 | 0 | internal::__fastpack6(in, out); |
690 | 0 | break; |
691 | 0 | case 7: |
692 | 0 | internal::__fastpack7(in, out); |
693 | 0 | break; |
694 | 0 | case 8: |
695 | 0 | internal::__fastpack8(in, out); |
696 | 0 | break; |
697 | 0 | case 9: |
698 | 0 | internal::__fastpack9(in, out); |
699 | 0 | break; |
700 | 0 | case 10: |
701 | 0 | internal::__fastpack10(in, out); |
702 | 0 | break; |
703 | 0 | case 11: |
704 | 0 | internal::__fastpack11(in, out); |
705 | 0 | break; |
706 | 0 | case 12: |
707 | 0 | internal::__fastpack12(in, out); |
708 | 0 | break; |
709 | 0 | case 13: |
710 | 0 | internal::__fastpack13(in, out); |
711 | 0 | break; |
712 | 0 | case 14: |
713 | 0 | internal::__fastpack14(in, out); |
714 | 0 | break; |
715 | 0 | case 15: |
716 | 0 | internal::__fastpack15(in, out); |
717 | 0 | break; |
718 | 0 | case 16: |
719 | 0 | internal::__fastpack16(in, out); |
720 | 0 | break; |
721 | 0 | case 17: |
722 | 0 | internal::__fastpack17(in, out); |
723 | 0 | break; |
724 | 0 | case 18: |
725 | 0 | internal::__fastpack18(in, out); |
726 | 0 | break; |
727 | 0 | case 19: |
728 | 0 | internal::__fastpack19(in, out); |
729 | 0 | break; |
730 | 0 | case 20: |
731 | 0 | internal::__fastpack20(in, out); |
732 | 0 | break; |
733 | 0 | case 21: |
734 | 0 | internal::__fastpack21(in, out); |
735 | 0 | break; |
736 | 0 | case 22: |
737 | 0 | internal::__fastpack22(in, out); |
738 | 0 | break; |
739 | 0 | case 23: |
740 | 0 | internal::__fastpack23(in, out); |
741 | 0 | break; |
742 | 0 | case 24: |
743 | 0 | internal::__fastpack24(in, out); |
744 | 0 | break; |
745 | 0 | case 25: |
746 | 0 | internal::__fastpack25(in, out); |
747 | 0 | break; |
748 | 0 | case 26: |
749 | 0 | internal::__fastpack26(in, out); |
750 | 0 | break; |
751 | 0 | case 27: |
752 | 0 | internal::__fastpack27(in, out); |
753 | 0 | break; |
754 | 0 | case 28: |
755 | 0 | internal::__fastpack28(in, out); |
756 | 0 | break; |
757 | 0 | case 29: |
758 | 0 | internal::__fastpack29(in, out); |
759 | 0 | break; |
760 | 0 | case 30: |
761 | 0 | internal::__fastpack30(in, out); |
762 | 0 | break; |
763 | 0 | case 31: |
764 | 0 | internal::__fastpack31(in, out); |
765 | 0 | break; |
766 | 0 | case 32: |
767 | 0 | internal::__fastpack32(in, out); |
768 | 0 | break; |
769 | 0 | case 33: |
770 | 0 | internal::__fastpack33(in, out); |
771 | 0 | break; |
772 | 0 | case 34: |
773 | 0 | internal::__fastpack34(in, out); |
774 | 0 | break; |
775 | 0 | case 35: |
776 | 0 | internal::__fastpack35(in, out); |
777 | 0 | break; |
778 | 0 | case 36: |
779 | 0 | internal::__fastpack36(in, out); |
780 | 0 | break; |
781 | 0 | case 37: |
782 | 0 | internal::__fastpack37(in, out); |
783 | 0 | break; |
784 | 0 | case 38: |
785 | 0 | internal::__fastpack38(in, out); |
786 | 0 | break; |
787 | 0 | case 39: |
788 | 0 | internal::__fastpack39(in, out); |
789 | 0 | break; |
790 | 0 | case 40: |
791 | 0 | internal::__fastpack40(in, out); |
792 | 0 | break; |
793 | 0 | case 41: |
794 | 0 | internal::__fastpack41(in, out); |
795 | 0 | break; |
796 | 0 | case 42: |
797 | 0 | internal::__fastpack42(in, out); |
798 | 0 | break; |
799 | 0 | case 43: |
800 | 0 | internal::__fastpack43(in, out); |
801 | 0 | break; |
802 | 0 | case 44: |
803 | 0 | internal::__fastpack44(in, out); |
804 | 0 | break; |
805 | 0 | case 45: |
806 | 0 | internal::__fastpack45(in, out); |
807 | 0 | break; |
808 | 0 | case 46: |
809 | 0 | internal::__fastpack46(in, out); |
810 | 0 | break; |
811 | 0 | case 47: |
812 | 0 | internal::__fastpack47(in, out); |
813 | 0 | break; |
814 | 0 | case 48: |
815 | 0 | internal::__fastpack48(in, out); |
816 | 0 | break; |
817 | 0 | case 49: |
818 | 0 | internal::__fastpack49(in, out); |
819 | 0 | break; |
820 | 0 | case 50: |
821 | 0 | internal::__fastpack50(in, out); |
822 | 0 | break; |
823 | 0 | case 51: |
824 | 0 | internal::__fastpack51(in, out); |
825 | 0 | break; |
826 | 0 | case 52: |
827 | 0 | internal::__fastpack52(in, out); |
828 | 0 | break; |
829 | 0 | case 53: |
830 | 0 | internal::__fastpack53(in, out); |
831 | 0 | break; |
832 | 0 | case 54: |
833 | 0 | internal::__fastpack54(in, out); |
834 | 0 | break; |
835 | 0 | case 55: |
836 | 0 | internal::__fastpack55(in, out); |
837 | 0 | break; |
838 | 0 | case 56: |
839 | 0 | internal::__fastpack56(in, out); |
840 | 0 | break; |
841 | 0 | case 57: |
842 | 0 | internal::__fastpack57(in, out); |
843 | 0 | break; |
844 | 0 | case 58: |
845 | 0 | internal::__fastpack58(in, out); |
846 | 0 | break; |
847 | 0 | case 59: |
848 | 0 | internal::__fastpack59(in, out); |
849 | 0 | break; |
850 | 0 | case 60: |
851 | 0 | internal::__fastpack60(in, out); |
852 | 0 | break; |
853 | 0 | case 61: |
854 | 0 | internal::__fastpack61(in, out); |
855 | 0 | break; |
856 | 0 | case 62: |
857 | 0 | internal::__fastpack62(in, out); |
858 | 0 | break; |
859 | 0 | case 63: |
860 | 0 | internal::__fastpack63(in, out); |
861 | 0 | break; |
862 | 0 | case 64: |
863 | 0 | internal::__fastpack64(in, out); |
864 | 0 | break; |
865 | 0 | default: |
866 | 0 | throw std::logic_error("Invalid bit width for bitpacking"); |
867 | 0 | } |
868 | 0 | } |
869 | | } // namespace fastpfor_lib |