1 /* wc_avx - Count the number of newlines with avx2 instructions.
2 Copyright (C) 2021-2022 Free Software Foundation, Inc.
4 This program is free software: you can redistribute it and/or modify
5 it under the terms of the GNU General Public License as published by
6 the Free Software Foundation, either version 3 of the License, or
7 (at your option) any later version.
9 This program is distributed in the hope that it will be useful,
10 but WITHOUT ANY WARRANTY; without even the implied warranty of
11 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 GNU General Public License for more details.
14 You should have received a copy of the GNU General Public License
15 along with this program. If not, see <https://www.gnu.org/licenses/>. */
21 #include "safe-read.h"
23 #include <x86intrin.h>
25 /* This must be below 16 KB (16384) or else the accumulators can
26 theoretically overflow, producing wrong result. This is 2*32 bytes below,
27 so there is no single bytes in the optimal case. */
28 #define BUFSIZE (16320)
31 wc_lines_avx2 (char const *file
, int fd
, uintmax_t *lines_out
,
32 uintmax_t *bytes_out
);
35 wc_lines_avx2 (char const *file
, int fd
, uintmax_t *lines_out
,
42 __m256i avx_buf
[BUFSIZE
/ sizeof (__m256i
)];
46 size_t bytes_read
= 0;
49 if (!lines_out
|| !bytes_out
)
52 /* Using two parallel accumulators gave a good performance increase.
53 Adding a third gave no additional benefit, at least on an
54 Intel Xeon E3-1231v3. Maybe on a newer CPU with additional vector
55 execution engines it would be a win. */
56 accumulator
= _mm256_setzero_si256 ();
57 accumulator2
= _mm256_setzero_si256 ();
58 zeroes
= _mm256_setzero_si256 ();
59 endlines
= _mm256_set1_epi8 ('\n');
61 while ((bytes_read
= safe_read (fd
, avx_buf
, sizeof (avx_buf
))) > 0)
68 if (bytes_read
== SAFE_READ_ERROR
)
70 error (0, errno
, "%s", quotef (file
));
77 char *end
= ((char *)avx_buf
) + bytes_read
;
79 while (bytes_read
>= 64)
81 to_match
= _mm256_load_si256 (datap
);
82 to_match2
= _mm256_load_si256 (datap
+ 1);
84 matches
= _mm256_cmpeq_epi8 (to_match
, endlines
);
85 matches2
= _mm256_cmpeq_epi8 (to_match2
, endlines
);
86 /* Compare will set each 8 bit integer in the register to 0xFF
87 on match. When we subtract it the 8 bit accumulators
88 will underflow, so this is equal to adding 1. */
89 accumulator
= _mm256_sub_epi8 (accumulator
, matches
);
90 accumulator2
= _mm256_sub_epi8 (accumulator2
, matches2
);
96 /* Horizontally add all 8 bit integers in the register,
98 accumulator
= _mm256_sad_epu8 (accumulator
, zeroes
);
99 lines
+= _mm256_extract_epi16 (accumulator
, 0)
100 + _mm256_extract_epi16 (accumulator
, 4)
101 + _mm256_extract_epi16 (accumulator
, 8)
102 + _mm256_extract_epi16 (accumulator
, 12);
103 accumulator
= _mm256_setzero_si256 ();
105 accumulator2
= _mm256_sad_epu8 (accumulator2
, zeroes
);
106 lines
+= _mm256_extract_epi16 (accumulator2
, 0)
107 + _mm256_extract_epi16 (accumulator2
, 4)
108 + _mm256_extract_epi16 (accumulator2
, 8)
109 + _mm256_extract_epi16 (accumulator2
, 12);
110 accumulator2
= _mm256_setzero_si256 ();
112 /* Finish up any left over bytes */
113 char *p
= (char *)datap
;
115 lines
+= *p
++ == '\n';