Repository navigation
Expand file tree
/
Copy pathsimd-checksum-x86_64.cpp
More file actions
749 lines (637 loc) · 30.6 KB
/
Copy pathsimd-checksum-x86_64.cpp
File metadata and controls
749 lines (637 loc) · 30.6 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
/*
* SSE2/SSSE3/AVX2-optimized routines to support checksumming of bytes.
*
* Copyright (C) 1996 Andrew Tridgell
* Copyright (C) 1996 Paul Mackerras
* Copyright (C) 2004-2020 Wayne Davison
* Copyright (C) 2020 Jorrit Jongma
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 3 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License along
* with this program; if not, visit the http://fsf.org website.
*/
/*
* Optimization target for get_checksum1() was the Intel Atom D2700, the
* slowest CPU in the test set and the most likely to be CPU limited during
* transfers. The combination of intrinsics was chosen specifically for the
* most gain on that CPU, other combinations were occasionally slightly
* faster on the others.
*
* While on more modern CPUs transfers are less likely to be CPU limited
* (at least by this specific function), lower CPU usage is always better.
* Improvements may still be seen when matching chunks from NVMe storage
* even on newer CPUs.
*
* Benchmarks (in MB/s) C SSE2 SSSE3 AVX2
* - Intel Atom D2700 550 750 1000 N/A
* - Intel i7-7700hq 1850 2550 4050 6200
* - AMD ThreadRipper 2950x 2900 5600 8950 8100
*
* Curiously the AMD is slower with AVX2 than SSSE3, while the Intel is
* significantly faster. AVX2 is kept because it's more likely to relieve
* the bottleneck on the slower CPU.
*
* This optimization for get_checksum1() is intentionally limited to x86-64
* as no 32-bit CPU was available for testing. As 32-bit CPUs only have half
* the available xmm registers, this optimized version may not be faster than
* the pure C version anyway. Note that all x86-64 CPUs support at least SSE2.
*
* This file is compiled using GCC 4.8+/clang 6+'s C++ front end to allow the
* use of the target attribute, selecting the fastest code path based on
* dispatch priority (GCC 5) or runtime detection of CPU capabilities (GCC 6+).
* GCC 4.x are not supported to ease configure.ac logic.
*/
#ifdef __x86_64__ /* { */
#ifdef __cplusplus /* { */
#include "rsync.h"
#ifdef USE_ROLL_SIMD /* { */
#include <immintrin.h>
/* Some clang versions don't like it when you use static with multi-versioned functions: linker errors */
#ifdef __clang__
#define MVSTATIC
#else
#define MVSTATIC static
#endif
// Missing from the headers on gcc 6 and older, clang 8 and older
typedef long long __m128i_u __attribute__((__vector_size__(16), __may_alias__, __aligned__(16)));
typedef long long __m256i_u __attribute__((__vector_size__(32), __may_alias__, __aligned__(16)));
/* Compatibility macros to let our SSSE3 algorithm run with only SSE2.
These used to be neat individual functions with target attributes switching between SSE2 and SSSE3 implementations
as needed, but though this works perfectly with GCC, clang fails to inline those properly leading to a near 50%
performance drop - combined with static and inline modifiers gets you linker errors and even compiler crashes...
*/
#define SSE2_INTERLEAVE_ODD_EPI16(a, b) _mm_packs_epi32(_mm_srai_epi32(a, 16), _mm_srai_epi32(b, 16))
#define SSE2_INTERLEAVE_EVEN_EPI16(a, b) SSE2_INTERLEAVE_ODD_EPI16(_mm_slli_si128(a, 2), _mm_slli_si128(b, 2))
#define SSE2_MULU_ODD_EPI8(a, b) _mm_mullo_epi16(_mm_srli_epi16(a, 8), _mm_srai_epi16(b, 8))
#define SSE2_MULU_EVEN_EPI8(a, b) _mm_mullo_epi16(_mm_and_si128(a, _mm_set1_epi16(0xFF)), _mm_srai_epi16(_mm_slli_si128(b, 1), 8))
#define SSE2_HADDS_EPI16(a, b) _mm_adds_epi16(SSE2_INTERLEAVE_EVEN_EPI16(a, b), SSE2_INTERLEAVE_ODD_EPI16(a, b))
#define SSE2_MADDUBS_EPI16(a, b) _mm_adds_epi16(SSE2_MULU_EVEN_EPI8(a, b), SSE2_MULU_ODD_EPI8(a, b))
#ifndef USE_ROLL_ASM
__attribute__ ((target("default"))) MVSTATIC int32 get_checksum1_avx2_64(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2) { return i; }
#endif
__attribute__ ((target("default"))) MVSTATIC int32 get_checksum1_ssse3_32(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2) { return i; }
__attribute__ ((target("default"))) MVSTATIC int32 get_checksum1_sse2_32(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2) { return i; }
/*
Original loop per 4 bytes:
s2 += 4*(s1 + buf[i]) + 3*buf[i+1] + 2*buf[i+2] + buf[i+3] + 10*CHAR_OFFSET;
s1 += buf[i] + buf[i+1] + buf[i+2] + buf[i+3] + 4*CHAR_OFFSET;
SSE2/SSSE3 loop per 32 bytes:
int16 t1[8];
int16 t2[8];
for (int j = 0; j < 8; j++) {
t1[j] = buf[j*4 + i] + buf[j*4 + i+1] + buf[j*4 + i+2] + buf[j*4 + i+3];
t2[j] = 4*buf[j*4 + i] + 3*buf[j*4 + i+1] + 2*buf[j*4 + i+2] + buf[j*4 + i+3];
}
s2 += 32*s1 + (uint32)(
28*t1[0] + 24*t1[1] + 20*t1[2] + 16*t1[3] + 12*t1[4] + 8*t1[5] + 4*t1[6] +
t2[0] + t2[1] + t2[2] + t2[3] + t2[4] + t2[5] + t2[6] + t2[7]
) + 528*CHAR_OFFSET;
s1 += (uint32)(t1[0] + t1[1] + t1[2] + t1[3] + t1[4] + t1[5] + t1[6] + t1[7]) +
32*CHAR_OFFSET;
*/
__attribute__ ((target("ssse3"))) MVSTATIC int32 get_checksum1_ssse3_32(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2)
{
if (len > 32) {
int aligned = ((uintptr_t)buf & 15) == 0;
uint32 x[4] = {0};
x[0] = *ps1;
__m128i ss1 = _mm_loadu_si128((__m128i_u*)x);
x[0] = *ps2;
__m128i ss2 = _mm_loadu_si128((__m128i_u*)x);
const int16 mul_t1_buf[8] = {28, 24, 20, 16, 12, 8, 4, 0};
__m128i mul_t1 = _mm_loadu_si128((__m128i_u*)mul_t1_buf);
for (; i < (len-32); i+=32) {
// Load ... 2*[int8*16]
__m128i in8_1, in8_2;
if (!aligned) {
// Synonymous with _mm_loadu_si128 on all but a handful of old CPUs
in8_1 = _mm_lddqu_si128((__m128i_u*)&buf[i]);
in8_2 = _mm_lddqu_si128((__m128i_u*)&buf[i + 16]);
} else {
in8_1 = _mm_load_si128((__m128i_u*)&buf[i]);
in8_2 = _mm_load_si128((__m128i_u*)&buf[i + 16]);
}
// (1*buf[i] + 1*buf[i+1]), (1*buf[i+2], 1*buf[i+3]), ... 2*[int16*8]
// Fastest, even though multiply by 1
__m128i mul_one = _mm_set1_epi8(1);
__m128i add16_1 = _mm_maddubs_epi16(mul_one, in8_1);
__m128i add16_2 = _mm_maddubs_epi16(mul_one, in8_2);
// (4*buf[i] + 3*buf[i+1]), (2*buf[i+2], buf[i+3]), ... 2*[int16*8]
__m128i mul_const = _mm_set1_epi32(4 + (3 << 8) + (2 << 16) + (1 << 24));
__m128i mul_add16_1 = _mm_maddubs_epi16(mul_const, in8_1);
__m128i mul_add16_2 = _mm_maddubs_epi16(mul_const, in8_2);
// s2 += 32*s1
ss2 = _mm_add_epi32(ss2, _mm_slli_epi32(ss1, 5));
// [sum(t1[0]..t1[7]), X, X, X] [int32*4]; faster than multiple _mm_hadds_epi16
// Shifting left, then shifting right again and shuffling (rather than just
// shifting right as with mul32 below) to cheaply end up with the correct sign
// extension as we go from int16 to int32.
__m128i sum_add32 = _mm_add_epi16(add16_1, add16_2);
sum_add32 = _mm_add_epi16(sum_add32, _mm_slli_si128(sum_add32, 2));
sum_add32 = _mm_add_epi16(sum_add32, _mm_slli_si128(sum_add32, 4));
sum_add32 = _mm_add_epi16(sum_add32, _mm_slli_si128(sum_add32, 8));
sum_add32 = _mm_srai_epi32(sum_add32, 16);
sum_add32 = _mm_shuffle_epi32(sum_add32, 3);
// [sum(t2[0]..t2[7]), X, X, X] [int32*4]; faster than multiple _mm_hadds_epi16
__m128i sum_mul_add32 = _mm_add_epi16(mul_add16_1, mul_add16_2);
sum_mul_add32 = _mm_add_epi16(sum_mul_add32, _mm_slli_si128(sum_mul_add32, 2));
sum_mul_add32 = _mm_add_epi16(sum_mul_add32, _mm_slli_si128(sum_mul_add32, 4));
sum_mul_add32 = _mm_add_epi16(sum_mul_add32, _mm_slli_si128(sum_mul_add32, 8));
sum_mul_add32 = _mm_srai_epi32(sum_mul_add32, 16);
sum_mul_add32 = _mm_shuffle_epi32(sum_mul_add32, 3);
// s1 += t1[0] + t1[1] + t1[2] + t1[3] + t1[4] + t1[5] + t1[6] + t1[7]
ss1 = _mm_add_epi32(ss1, sum_add32);
// s2 += t2[0] + t2[1] + t2[2] + t2[3] + t2[4] + t2[5] + t2[6] + t2[7]
ss2 = _mm_add_epi32(ss2, sum_mul_add32);
// [t1[0] + t1[1], t1[2] + t1[3] ...] [int16*8]
// We could've combined this with generating sum_add32 above and
// save an instruction but benchmarking shows that as being slower
__m128i add16 = _mm_hadds_epi16(add16_1, add16_2);
// [t1[0], t1[1], ...] -> [t1[0]*28 + t1[1]*24, ...] [int32*4]
__m128i mul32 = _mm_madd_epi16(add16, mul_t1);
// [sum(mul32), X, X, X] [int32*4]; faster than multiple _mm_hadd_epi32
mul32 = _mm_add_epi32(mul32, _mm_srli_si128(mul32, 4));
mul32 = _mm_add_epi32(mul32, _mm_srli_si128(mul32, 8));
// s2 += 28*t1[0] + 24*t1[1] + 20*t1[2] + 16*t1[3] + 12*t1[4] + 8*t1[5] + 4*t1[6]
ss2 = _mm_add_epi32(ss2, mul32);
#if CHAR_OFFSET != 0
// s1 += 32*CHAR_OFFSET
__m128i char_offset_multiplier = _mm_set1_epi32(32 * CHAR_OFFSET);
ss1 = _mm_add_epi32(ss1, char_offset_multiplier);
// s2 += 528*CHAR_OFFSET
char_offset_multiplier = _mm_set1_epi32(528 * CHAR_OFFSET);
ss2 = _mm_add_epi32(ss2, char_offset_multiplier);
#endif
}
_mm_store_si128((__m128i_u*)x, ss1);
*ps1 = x[0];
_mm_store_si128((__m128i_u*)x, ss2);
*ps2 = x[0];
}
return i;
}
/*
Same as SSSE3 version, but using macros defined above to emulate SSSE3 calls that are not available with SSE2.
For GCC-only the SSE2 and SSSE3 versions could be a single function calling other functions with the right
target attributes to emulate SSSE3 calls on SSE2 if needed, but clang doesn't inline those properly leading
to a near 50% performance drop.
*/
__attribute__ ((target("sse2"))) MVSTATIC int32 get_checksum1_sse2_32(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2)
{
if (len > 32) {
int aligned = ((uintptr_t)buf & 15) == 0;
uint32 x[4] = {0};
x[0] = *ps1;
__m128i ss1 = _mm_loadu_si128((__m128i_u*)x);
x[0] = *ps2;
__m128i ss2 = _mm_loadu_si128((__m128i_u*)x);
const int16 mul_t1_buf[8] = {28, 24, 20, 16, 12, 8, 4, 0};
__m128i mul_t1 = _mm_loadu_si128((__m128i_u*)mul_t1_buf);
for (; i < (len-32); i+=32) {
// Load ... 2*[int8*16]
__m128i in8_1, in8_2;
if (!aligned) {
in8_1 = _mm_loadu_si128((__m128i_u*)&buf[i]);
in8_2 = _mm_loadu_si128((__m128i_u*)&buf[i + 16]);
} else {
in8_1 = _mm_load_si128((__m128i_u*)&buf[i]);
in8_2 = _mm_load_si128((__m128i_u*)&buf[i + 16]);
}
// (1*buf[i] + 1*buf[i+1]), (1*buf[i+2], 1*buf[i+3]), ... 2*[int16*8]
// Fastest, even though multiply by 1
__m128i mul_one = _mm_set1_epi8(1);
__m128i add16_1 = SSE2_MADDUBS_EPI16(mul_one, in8_1);
__m128i add16_2 = SSE2_MADDUBS_EPI16(mul_one, in8_2);
// (4*buf[i] + 3*buf[i+1]), (2*buf[i+2], buf[i+3]), ... 2*[int16*8]
__m128i mul_const = _mm_set1_epi32(4 + (3 << 8) + (2 << 16) + (1 << 24));
__m128i mul_add16_1 = SSE2_MADDUBS_EPI16(mul_const, in8_1);
__m128i mul_add16_2 = SSE2_MADDUBS_EPI16(mul_const, in8_2);
// s2 += 32*s1
ss2 = _mm_add_epi32(ss2, _mm_slli_epi32(ss1, 5));
// [sum(t1[0]..t1[7]), X, X, X] [int32*4]; faster than multiple _mm_hadds_epi16
// Shifting left, then shifting right again and shuffling (rather than just
// shifting right as with mul32 below) to cheaply end up with the correct sign
// extension as we go from int16 to int32.
__m128i sum_add32 = _mm_add_epi16(add16_1, add16_2);
sum_add32 = _mm_add_epi16(sum_add32, _mm_slli_si128(sum_add32, 2));
sum_add32 = _mm_add_epi16(sum_add32, _mm_slli_si128(sum_add32, 4));
sum_add32 = _mm_add_epi16(sum_add32, _mm_slli_si128(sum_add32, 8));
sum_add32 = _mm_srai_epi32(sum_add32, 16);
sum_add32 = _mm_shuffle_epi32(sum_add32, 3);
// [sum(t2[0]..t2[7]), X, X, X] [int32*4]; faster than multiple _mm_hadds_epi16
__m128i sum_mul_add32 = _mm_add_epi16(mul_add16_1, mul_add16_2);
sum_mul_add32 = _mm_add_epi16(sum_mul_add32, _mm_slli_si128(sum_mul_add32, 2));
sum_mul_add32 = _mm_add_epi16(sum_mul_add32, _mm_slli_si128(sum_mul_add32, 4));
sum_mul_add32 = _mm_add_epi16(sum_mul_add32, _mm_slli_si128(sum_mul_add32, 8));
sum_mul_add32 = _mm_srai_epi32(sum_mul_add32, 16);
sum_mul_add32 = _mm_shuffle_epi32(sum_mul_add32, 3);
// s1 += t1[0] + t1[1] + t1[2] + t1[3] + t1[4] + t1[5] + t1[6] + t1[7]
ss1 = _mm_add_epi32(ss1, sum_add32);
// s2 += t2[0] + t2[1] + t2[2] + t2[3] + t2[4] + t2[5] + t2[6] + t2[7]
ss2 = _mm_add_epi32(ss2, sum_mul_add32);
// [t1[0] + t1[1], t1[2] + t1[3] ...] [int16*8]
// We could've combined this with generating sum_add32 above and
// save an instruction but benchmarking shows that as being slower
__m128i add16 = SSE2_HADDS_EPI16(add16_1, add16_2);
// [t1[0], t1[1], ...] -> [t1[0]*28 + t1[1]*24, ...] [int32*4]
__m128i mul32 = _mm_madd_epi16(add16, mul_t1);
// [sum(mul32), X, X, X] [int32*4]; faster than multiple _mm_hadd_epi32
mul32 = _mm_add_epi32(mul32, _mm_srli_si128(mul32, 4));
mul32 = _mm_add_epi32(mul32, _mm_srli_si128(mul32, 8));
// s2 += 28*t1[0] + 24*t1[1] + 20*t1[2] + 16*t1[3] + 12*t1[4] + 8*t1[5] + 4*t1[6]
ss2 = _mm_add_epi32(ss2, mul32);
#if CHAR_OFFSET != 0
// s1 += 32*CHAR_OFFSET
__m128i char_offset_multiplier = _mm_set1_epi32(32 * CHAR_OFFSET);
ss1 = _mm_add_epi32(ss1, char_offset_multiplier);
// s2 += 528*CHAR_OFFSET
char_offset_multiplier = _mm_set1_epi32(528 * CHAR_OFFSET);
ss2 = _mm_add_epi32(ss2, char_offset_multiplier);
#endif
}
_mm_store_si128((__m128i_u*)x, ss1);
*ps1 = x[0];
_mm_store_si128((__m128i_u*)x, ss2);
*ps2 = x[0];
}
return i;
}
#ifdef USE_ROLL_ASM /* { */
extern "C" __attribute__ ((target("avx2"))) int32 get_checksum1_avx2_asm(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2);
/* The asm routine is AVX2-only and, unlike the multi-versioned intrinsic
* paths, has no compiler-generated fallback, so it must not be called on a
* CPU without AVX2 (it would fault with SIGILL). Gate it on a cached runtime
* check; when AVX2 is absent we skip it and the SSSE3/SSE2/scalar steps,
* which are safe everywhere, do all the work. */
static int roll_asm_have_avx2(void)
{
static int have = -1;
if (have < 0) {
__builtin_cpu_init();
have = __builtin_cpu_supports("avx2") ? 1 : 0;
}
return have;
}
#else /* } { */
/*
AVX2 loop per 64 bytes:
int16 t1[16];
int16 t2[16];
for (int j = 0; j < 16; j++) {
t1[j] = buf[j*4 + i] + buf[j*4 + i+1] + buf[j*4 + i+2] + buf[j*4 + i+3];
t2[j] = 4*buf[j*4 + i] + 3*buf[j*4 + i+1] + 2*buf[j*4 + i+2] + buf[j*4 + i+3];
}
s2 += 64*s1 + (uint32)(
60*t1[0] + 56*t1[1] + 52*t1[2] + 48*t1[3] + 44*t1[4] + 40*t1[5] + 36*t1[6] + 32*t1[7] + 28*t1[8] + 24*t1[9] + 20*t1[10] + 16*t1[11] + 12*t1[12] + 8*t1[13] + 4*t1[14] +
t2[0] + t2[1] + t2[2] + t2[3] + t2[4] + t2[5] + t2[6] + t2[7] + t2[8] + t2[9] + t2[10] + t2[11] + t2[12] + t2[13] + t2[14] + t2[15]
) + 2080*CHAR_OFFSET;
s1 += (uint32)(t1[0] + t1[1] + t1[2] + t1[3] + t1[4] + t1[5] + t1[6] + t1[7] + t1[8] + t1[9] + t1[10] + t1[11] + t1[12] + t1[13] + t1[14] + t1[15]) +
64*CHAR_OFFSET;
*/
__attribute__ ((target("avx2"))) MVSTATIC int32 get_checksum1_avx2_64(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2)
{
if (len > 64) {
uint32 x[4] = {0};
__m128i ss1 = _mm_cvtsi32_si128(*ps1);
__m128i ss2 = _mm_cvtsi32_si128(*ps2);
const char mul_t1_buf[16] = {60, 56, 52, 48, 44, 40, 36, 32, 28, 24, 20, 16, 12, 8, 4, 0};
__m128i tmp = _mm_load_si128((__m128i*) mul_t1_buf);
__m256i mul_t1 = _mm256_cvtepu8_epi16(tmp);
__m256i mul_const = _mm256_broadcastd_epi32(_mm_cvtsi32_si128(4 | (3 << 8) | (2 << 16) | (1 << 24)));
__m256i mul_one = _mm256_set1_epi8(1);
for (; i < (len-64); i+=64) {
// Load ... 4*[int8*16]
__m256i in8_1, in8_2;
__m128i in8_1_low, in8_2_low, in8_1_high, in8_2_high;
in8_1_low = _mm_loadu_si128((__m128i_u*)&buf[i]);
in8_2_low = _mm_loadu_si128((__m128i_u*)&buf[i+16]);
in8_1_high = _mm_loadu_si128((__m128i_u*)&buf[i+32]);
in8_2_high = _mm_loadu_si128((__m128i_u*)&buf[i+48]);
in8_1 = _mm256_inserti128_si256(_mm256_castsi128_si256(in8_1_low), in8_1_high,1);
in8_2 = _mm256_inserti128_si256(_mm256_castsi128_si256(in8_2_low), in8_2_high,1);
// (1*buf[i] + 1*buf[i+1]), (1*buf[i+2], 1*buf[i+3]), ... 2*[int16*8]
// Fastest, even though multiply by 1
__m256i add16_1 = _mm256_maddubs_epi16(mul_one, in8_1);
__m256i add16_2 = _mm256_maddubs_epi16(mul_one, in8_2);
// (4*buf[i] + 3*buf[i+1]), (2*buf[i+2], buf[i+3]), ... 2*[int16*8]
__m256i mul_add16_1 = _mm256_maddubs_epi16(mul_const, in8_1);
__m256i mul_add16_2 = _mm256_maddubs_epi16(mul_const, in8_2);
// s2 += 64*s1
ss2 = _mm_add_epi32(ss2, _mm_slli_epi32(ss1, 6));
// [sum(t1[0]..t1[7]), X, X, X] [int32*4]; faster than multiple _mm_hadds_epi16
__m256i sum_add32 = _mm256_add_epi16(add16_1, add16_2);
sum_add32 = _mm256_add_epi16(sum_add32, _mm256_srli_epi32(sum_add32, 16));
sum_add32 = _mm256_add_epi16(sum_add32, _mm256_srli_si256(sum_add32, 4));
sum_add32 = _mm256_add_epi16(sum_add32, _mm256_srli_si256(sum_add32, 8));
// [sum(t2[0]..t2[7]), X, X, X] [int32*4]; faster than multiple _mm_hadds_epi16
__m256i sum_mul_add32 = _mm256_add_epi16(mul_add16_1, mul_add16_2);
sum_mul_add32 = _mm256_add_epi16(sum_mul_add32, _mm256_srli_epi32(sum_mul_add32, 16));
sum_mul_add32 = _mm256_add_epi16(sum_mul_add32, _mm256_srli_si256(sum_mul_add32, 4));
sum_mul_add32 = _mm256_add_epi16(sum_mul_add32, _mm256_srli_si256(sum_mul_add32, 8));
// s1 += t1[0] + t1[1] + t1[2] + t1[3] + t1[4] + t1[5] + t1[6] + t1[7]
__m128i sum_add32_hi = _mm256_extracti128_si256(sum_add32, 0x1);
ss1 = _mm_add_epi32(ss1, _mm256_castsi256_si128(sum_add32));
ss1 = _mm_add_epi32(ss1, sum_add32_hi);
// s2 += t2[0] + t2[1] + t2[2] + t2[3] + t2[4] + t2[5] + t2[6] + t2[7]
__m128i sum_mul_add32_hi = _mm256_extracti128_si256(sum_mul_add32, 0x1);
ss2 = _mm_add_epi32(ss2, _mm256_castsi256_si128(sum_mul_add32));
ss2 = _mm_add_epi32(ss2, sum_mul_add32_hi);
// [t1[0] + t1[1], t1[2] + t1[3] ...] [int16*8]
// We could've combined this with generating sum_add32 above and
// save an instruction but benchmarking shows that as being slower
__m256i add16 = _mm256_hadds_epi16(add16_1, add16_2);
// [t1[0], t1[1], ...] -> [t1[0]*28 + t1[1]*24, ...] [int32*4]
__m256i mul32 = _mm256_madd_epi16(add16, mul_t1);
// [sum(mul32), X, X, X] [int32*4]; faster than multiple _mm_hadd_epi32
mul32 = _mm256_add_epi32(mul32, _mm256_srli_si256(mul32, 4));
mul32 = _mm256_add_epi32(mul32, _mm256_srli_si256(mul32, 8));
// prefetch 2 cacheline ahead
_mm_prefetch(&buf[i + 160], _MM_HINT_T0);
// s2 += 28*t1[0] + 24*t1[1] + 20*t1[2] + 16*t1[3] + 12*t1[4] + 8*t1[5] + 4*t1[6]
__m128i mul32_hi = _mm256_extracti128_si256(mul32, 0x1);
ss2 = _mm_add_epi32(ss2, _mm256_castsi256_si128(mul32));
ss2 = _mm_add_epi32(ss2, mul32_hi);
#if CHAR_OFFSET != 0
// s1 += 32*CHAR_OFFSET
__m128i char_offset_multiplier = _mm_set1_epi32(32 * CHAR_OFFSET);
ss1 = _mm_add_epi32(ss1, char_offset_multiplier);
// s2 += 528*CHAR_OFFSET
char_offset_multiplier = _mm_set1_epi32(528 * CHAR_OFFSET);
ss2 = _mm_add_epi32(ss2, char_offset_multiplier);
#endif
}
_mm_store_si128((__m128i_u*)x, ss1);
*ps1 = x[0];
_mm_store_si128((__m128i_u*)x, ss2);
*ps2 = x[0];
}
return i;
}
#endif /* } !USE_ROLL_ASM */
static int32 get_checksum1_default_1(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2)
{
uint32 s1 = *ps1;
uint32 s2 = *ps2;
for (; i < (len-4); i+=4) {
s2 += 4*(s1 + buf[i]) + 3*buf[i+1] + 2*buf[i+2] + buf[i+3] + 10*CHAR_OFFSET;
s1 += (buf[i+0] + buf[i+1] + buf[i+2] + buf[i+3] + 4*CHAR_OFFSET);
}
for (; i < len; i++) {
s1 += (buf[i]+CHAR_OFFSET); s2 += s1;
}
*ps1 = s1;
*ps2 = s2;
return i;
}
/* With GCC 10 putting this implementation inside 'extern "C"' causes an
assembler error. That worked fine on GCC 5-9 and clang 6-10...
*/
static inline uint32 get_checksum1_cpp(char *buf1, int32 len)
{
int32 i = 0;
uint32 s1 = 0;
uint32 s2 = 0;
// multiples of 64 bytes using AVX2 (if available)
#ifdef USE_ROLL_ASM
if (roll_asm_have_avx2())
i = get_checksum1_avx2_asm((schar*)buf1, len, i, &s1, &s2);
#else
i = get_checksum1_avx2_64((schar*)buf1, len, i, &s1, &s2);
#endif
// multiples of 32 bytes using SSSE3 (if available)
i = get_checksum1_ssse3_32((schar*)buf1, len, i, &s1, &s2);
// multiples of 32 bytes using SSE2 (if available)
i = get_checksum1_sse2_32((schar*)buf1, len, i, &s1, &s2);
// whatever is left (updates s1/s2; the returned offset is unused here)
get_checksum1_default_1((schar*)buf1, len, i, &s1, &s2);
return (s1 & 0xffff) + (s2 << 16);
}
extern "C" {
uint32 get_checksum1(char *buf1, int32 len)
{
return get_checksum1_cpp(buf1, len);
}
} // extern "C"
#ifdef BENCHMARK_SIMD_CHECKSUM1
#pragma clang optimize off
#pragma GCC push_options
#pragma GCC optimize ("O0")
#define ROUNDS 1024
#define BLOCK_LEN 1024*1024
#ifndef CLOCK_MONOTONIC_RAW
#define CLOCK_MONOTONIC_RAW CLOCK_MONOTONIC
#endif
static void benchmark(const char* desc, int32 (*func)(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2), schar* buf, int32 len) {
struct timespec start, end;
uint64_t us;
uint32_t cs, s1, s2;
int i, next;
clock_gettime(CLOCK_MONOTONIC_RAW, &start);
for (i = 0; i < ROUNDS; i++) {
s1 = s2 = 0;
next = func((schar*)buf, len, 0, &s1, &s2);
get_checksum1_default_1((schar*)buf, len, next, &s1, &s2);
}
clock_gettime(CLOCK_MONOTONIC_RAW, &end);
us = next == 0 ? 0 : (end.tv_sec - start.tv_sec) * 1000000 + (end.tv_nsec - start.tv_nsec) / 1000;
cs = next == 0 ? 0 : (s1 & 0xffff) + (s2 << 16);
printf("%-5s :: %5.0f MB/s :: %08x\n", desc, us ? (float)(len / (1024 * 1024) * ROUNDS) / ((float)us / 1000000.0f) : 0, cs);
}
static int32 get_checksum1_auto(schar* buf, int32 len, int32 i, uint32* ps1, uint32* ps2) {
uint32 cs = get_checksum1((char*)buf, len);
*ps1 = cs & 0xffff;
*ps2 = cs >> 16;
return len;
}
int main() {
int i;
unsigned char* buf = (unsigned char*)aligned_alloc(64,BLOCK_LEN);
for (i = 0; i < BLOCK_LEN; i++) buf[i] = (i + (i % 3) + (i % 11)) % 256;
benchmark("Auto", get_checksum1_auto, (schar*)buf, BLOCK_LEN);
benchmark("Raw-C", get_checksum1_default_1, (schar*)buf, BLOCK_LEN);
benchmark("SSE2", get_checksum1_sse2_32, (schar*)buf, BLOCK_LEN);
benchmark("SSSE3", get_checksum1_ssse3_32, (schar*)buf, BLOCK_LEN);
#ifdef USE_ROLL_ASM
benchmark("AVX2-ASM", get_checksum1_avx2_asm, (schar*)buf, BLOCK_LEN);
#else
benchmark("AVX2", get_checksum1_avx2_64, (schar*)buf, BLOCK_LEN);
#endif
free(buf);
return 0;
}
#pragma GCC pop_options
#pragma clang optimize on
#endif /* BENCHMARK_SIMD_CHECKSUM1 */
#ifdef TEST_SIMD_CHECKSUM1
#include <sys/mman.h>
#include <unistd.h>
static uint32 checksum_via_default(char *buf, int32 len)
{
uint32 s1 = 0, s2 = 0;
get_checksum1_default_1((schar*)buf, len, 0, &s1, &s2);
return (s1 & 0xffff) + (s2 << 16);
}
static uint32 checksum_via_sse2(char *buf, int32 len)
{
int32 i;
uint32 s1 = 0, s2 = 0;
i = get_checksum1_sse2_32((schar*)buf, len, 0, &s1, &s2);
get_checksum1_default_1((schar*)buf, len, i, &s1, &s2);
return (s1 & 0xffff) + (s2 << 16);
}
static uint32 checksum_via_ssse3(char *buf, int32 len)
{
int32 i;
uint32 s1 = 0, s2 = 0;
i = get_checksum1_ssse3_32((schar*)buf, len, 0, &s1, &s2);
get_checksum1_default_1((schar*)buf, len, i, &s1, &s2);
return (s1 & 0xffff) + (s2 << 16);
}
static uint32 checksum_via_avx2(char *buf, int32 len)
{
int32 i;
uint32 s1 = 0, s2 = 0;
#ifdef USE_ROLL_ASM
if (roll_asm_have_avx2())
i = get_checksum1_avx2_asm((schar*)buf, len, 0, &s1, &s2);
else
i = 0;
#else
i = get_checksum1_avx2_64((schar*)buf, len, 0, &s1, &s2);
#endif
get_checksum1_default_1((schar*)buf, len, i, &s1, &s2);
return (s1 & 0xffff) + (s2 << 16);
}
/* Run every implementation on a buffer placed flush against an unreadable page,
* so any read past buf+len faults here instead of in a user's transfer. The
* buffers above deliberately carry 64 spare bytes for the unaligned case, which
* is exactly what let a 64-byte over-read in the AVX2 assembly go unnoticed:
* it only crashed when an allocation happened to end at a page boundary. */
static int test_no_overread(void)
{
#if defined HAVE_SYS_MMAN_H || defined __linux__ || defined __APPLE__ || defined BSD
/* Setting the guard up is not optional on the platforms this is compiled
* for: if it fails we are not testing what the caller thinks, so say so and
* fail rather than return a pass that looks identical to a real one. */
long pagesz = sysconf(_SC_PAGESIZE);
if (pagesz <= 0) {
printf("FAIL guard-page: sysconf(_SC_PAGESIZE) gave %ld\n", pagesz);
return 1;
}
size_t region = (size_t)pagesz * 4;
char *base = (char *)mmap(NULL, region, PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
if (base == MAP_FAILED) {
printf("FAIL guard-page: mmap failed\n");
return 1;
}
if (mprotect(base + region - pagesz, pagesz, PROT_NONE) != 0) {
printf("FAIL guard-page: mprotect failed\n");
munmap(base, region);
return 1;
}
/* Whether the ASSEMBLY was reached is a separate question from whether the
* checks passed. Without AVX2 the dispatcher falls back and the guard loop
* proves nothing about it, which must not read as coverage. */
__builtin_cpu_init();
printf("guard-page: AVX2 %s\n",
__builtin_cpu_supports("avx2")
? "present, assembly and intrinsics exercised"
: "ABSENT -- fallback only, the AVX2 paths were NOT exercised");
int failures = 0;
/* 128 is the shortest the AVX2 paths will touch; step by 1 so every
* remainder mod 64, and both alignments, get covered. */
for (int32 len = 128; len <= 4096; len++) {
char *buf = base + region - pagesz - len; /* last byte abuts the guard */
for (int32 i = 0; i < len; i++)
buf[i] = (char)((i + (i % 3) + (i % 11)) % 256);
uint32 ref = checksum_via_default(buf, len);
if (checksum_via_sse2(buf, len) != ref
|| checksum_via_ssse3(buf, len) != ref
|| checksum_via_avx2(buf, len) != ref
|| get_checksum1(buf, len) != ref) {
printf("FAIL guard-page size=%5d: mismatch at the page boundary\n", len);
if (++failures > 4)
break;
}
}
munmap(base, region);
return failures;
#else
printf("guard-page: no mmap on this platform, over-read check NOT RUN\n");
return 0;
#endif
}
int main()
{
static const int sizes[] = {1, 4, 31, 32, 33, 63, 64, 65, 128, 129, 256, 700, 1024, 4096, 65536};
int num_sizes = sizeof(sizes) / sizeof(sizes[0]);
int max_size = sizes[num_sizes - 1];
int failures = 0;
/* Allocate with extra bytes for unaligned test */
unsigned char *raw = (unsigned char *)malloc(max_size + 64 + 1);
if (!raw) {
fprintf(stderr, "malloc failed\n");
return 1;
}
/* Fill with deterministic data */
for (int i = 0; i < max_size + 64 + 1; i++)
raw[i] = (i + (i % 3) + (i % 11)) % 256;
/* Test with aligned buffer (64-byte aligned) */
unsigned char *aligned = raw + (64 - ((uintptr_t)raw % 64));
/* Test with unaligned buffer (+1 byte offset) */
unsigned char *unaligned = aligned + 1;
struct { const char *name; unsigned char *buf; } buffers[] = {
{"aligned", aligned},
{"unaligned", unaligned},
};
for (int b = 0; b < 2; b++) {
char *buf = (char *)buffers[b].buf;
const char *bname = buffers[b].name;
for (int s = 0; s < num_sizes; s++) {
int32 len = sizes[s];
uint32 ref = checksum_via_default(buf, len);
uint32 cs_sse2 = checksum_via_sse2(buf, len);
uint32 cs_ssse3 = checksum_via_ssse3(buf, len);
uint32 cs_avx2 = checksum_via_avx2(buf, len);
uint32 cs_auto = get_checksum1(buf, len);
if (cs_sse2 != ref) {
printf("FAIL %-9s size=%5d: SSE2=%08x ref=%08x\n", bname, len, cs_sse2, ref);
failures++;
}
if (cs_ssse3 != ref) {
printf("FAIL %-9s size=%5d: SSSE3=%08x ref=%08x\n", bname, len, cs_ssse3, ref);
failures++;
}
if (cs_avx2 != ref) {
printf("FAIL %-9s size=%5d: AVX2=%08x ref=%08x\n", bname, len, cs_avx2, ref);
failures++;
}
if (cs_auto != ref) {
printf("FAIL %-9s size=%5d: auto=%08x ref=%08x\n", bname, len, cs_auto, ref);
failures++;
}
}
}
free(raw);
failures += test_no_overread();
if (failures) {
printf("%d checksum mismatches!\n", failures);
return 1;
}
printf("All SIMD checksum tests passed.\n");
return 0;
}
#endif /* TEST_SIMD_CHECKSUM1 */
#endif /* } USE_ROLL_SIMD */
#endif /* } __cplusplus */
#endif /* } __x86_64__ */