|
1 | | -use criterion::{criterion_group, criterion_main, Criterion}; |
| 1 | +use criterion::{black_box, criterion_group, criterion_main, BenchmarkId, Criterion}; |
2 | 2 | use memmem::{Searcher, TwoWaySearcher}; |
3 | | -use std::fs::File; |
4 | | -use std::io::{self, BufRead, Read}; |
| 3 | +use std::str; |
5 | 4 | use strstr::avx2::*; |
6 | 5 |
|
7 | 6 | fn criterion_benchmark(c: &mut Criterion) { |
8 | | - let mut f = File::open("./data/i386.txt").unwrap(); |
9 | | - let mut buffer = Vec::new(); |
10 | | - f.read_to_end(&mut buffer).unwrap(); |
11 | | - let content = String::from_utf8_lossy(&buffer); |
12 | | - |
13 | | - let file = File::open("./data/words").unwrap(); |
14 | | - let mut words: Vec<String> = io::BufReader::new(file) |
15 | | - .lines() |
16 | | - .filter_map(|line| line.ok()) |
17 | | - .filter(|word| word.len() > 1) |
18 | | - .collect(); |
19 | | - // Sort all words by length then lexicographically |
20 | | - words.sort_by(|a, b| { |
21 | | - if a.len() != b.len() { |
22 | | - a.len().partial_cmp(&b.len()).unwrap() |
23 | | - } else { |
24 | | - a.partial_cmp(b).unwrap() |
| 7 | + let haystack = include_bytes!("../data/haystack"); |
| 8 | + let needle = include_bytes!("../data/needle"); |
| 9 | + |
| 10 | + let sizes = [1, 5, 10, 20, 50, 100, 1000]; |
| 11 | + |
| 12 | + for (i, &size) in sizes.iter().enumerate() { |
| 13 | + let mut group = c.benchmark_group(format!("{}-byte needle", size)); |
| 14 | + let needle = &needle[..size]; |
| 15 | + |
| 16 | + for &size in &sizes[i..] { |
| 17 | + let parameter = &format!("{}-byte haystack", size); |
| 18 | + let haystack = &haystack[..size]; |
| 19 | + |
| 20 | + group.bench_with_input( |
| 21 | + BenchmarkId::new("String::find", parameter), |
| 22 | + &size, |
| 23 | + |b, _| { |
| 24 | + let haystack = str::from_utf8(haystack).unwrap(); |
| 25 | + let needle = str::from_utf8(needle).unwrap(); |
| 26 | + b.iter(|| haystack.find(needle)); |
| 27 | + }, |
| 28 | + ); |
| 29 | + |
| 30 | + group.bench_with_input( |
| 31 | + BenchmarkId::new("TwoWaySearcher::search_in", parameter), |
| 32 | + &size, |
| 33 | + |b, _| { |
| 34 | + let searcher = TwoWaySearcher::new(needle); |
| 35 | + b.iter(|| black_box(searcher.search_in(haystack))); |
| 36 | + }, |
| 37 | + ); |
| 38 | + |
| 39 | + group.bench_with_input( |
| 40 | + BenchmarkId::new("twoway::find_bytes", parameter), |
| 41 | + &size, |
| 42 | + |b, _| { |
| 43 | + b.iter(|| black_box(twoway::find_bytes(haystack, needle))); |
| 44 | + }, |
| 45 | + ); |
| 46 | + |
| 47 | + group.bench_with_input( |
| 48 | + BenchmarkId::new("strstr_avx2_original", parameter), |
| 49 | + &size, |
| 50 | + |b, _| { |
| 51 | + b.iter(|| black_box(unsafe { strstr_avx2_original(haystack, needle) })); |
| 52 | + }, |
| 53 | + ); |
| 54 | + |
| 55 | + group.bench_with_input( |
| 56 | + BenchmarkId::new("strstr_avx2_rust_simple", parameter), |
| 57 | + &size, |
| 58 | + |b, _| { |
| 59 | + b.iter(|| black_box(unsafe { strstr_avx2_rust_simple(haystack, needle) })); |
| 60 | + }, |
| 61 | + ); |
| 62 | + |
| 63 | + group.bench_with_input( |
| 64 | + BenchmarkId::new("strstr_avx2_rust_simple_2", parameter), |
| 65 | + &size, |
| 66 | + |b, _| { |
| 67 | + b.iter(|| black_box(unsafe { strstr_avx2_rust_simple_2(haystack, needle) })); |
| 68 | + }, |
| 69 | + ); |
| 70 | + |
| 71 | + group.bench_with_input( |
| 72 | + BenchmarkId::new("strstr_avx2_rust_fast", parameter), |
| 73 | + &size, |
| 74 | + |b, _| { |
| 75 | + b.iter(|| black_box(unsafe { strstr_avx2_rust_fast(haystack, needle) })); |
| 76 | + }, |
| 77 | + ); |
| 78 | + |
| 79 | + group.bench_with_input( |
| 80 | + BenchmarkId::new("strstr_avx2_rust_fast_2", parameter), |
| 81 | + &size, |
| 82 | + |b, _| { |
| 83 | + b.iter(|| black_box(strstr_avx2_rust_fast_2(haystack, needle))); |
| 84 | + }, |
| 85 | + ); |
| 86 | + |
| 87 | + group.bench_with_input( |
| 88 | + BenchmarkId::new("strstr_avx2_rust_aligned", parameter), |
| 89 | + &size, |
| 90 | + |b, _| { |
| 91 | + b.iter(|| black_box(unsafe { strstr_avx2_rust_aligned(haystack, needle) })); |
| 92 | + }, |
| 93 | + ); |
| 94 | + |
| 95 | + group.bench_with_input( |
| 96 | + BenchmarkId::new("StrStrAVX2Searcher::search_in", parameter), |
| 97 | + &size, |
| 98 | + |b, _| { |
| 99 | + let searcher = StrStrAVX2Searcher::new(needle); |
| 100 | + b.iter(|| black_box(searcher.search_in(haystack))); |
| 101 | + }, |
| 102 | + ); |
| 103 | + |
| 104 | + group.bench_with_input( |
| 105 | + BenchmarkId::new("DynamicAvx2Searcher::search_in", parameter), |
| 106 | + &size, |
| 107 | + |b, _| { |
| 108 | + let searcher = DynamicAvx2Searcher::new(needle.to_owned().into_boxed_slice()); |
| 109 | + b.iter(|| black_box(searcher.search_in(haystack))); |
| 110 | + }, |
| 111 | + ); |
25 | 112 | } |
26 | | - }); |
27 | | - let words = words; |
28 | | - |
29 | | - let twoway_words: Vec<TwoWaySearcher<'_>> = words |
30 | | - .iter() |
31 | | - .map(|word| TwoWaySearcher::new(word.as_bytes())) |
32 | | - .collect(); |
33 | | - |
34 | | - let avx2_words: Vec<StrStrAVX2Searcher> = words |
35 | | - .iter() |
36 | | - .map(|word| StrStrAVX2Searcher::new(word.as_bytes())) |
37 | | - .collect(); |
38 | | - |
39 | | - let dynamic_avx2_searchers: Vec<DynamicAvx2Searcher> = words |
40 | | - .iter() |
41 | | - .map(|word| DynamicAvx2Searcher::new(word.clone().into_boxed_str().into())) |
42 | | - .collect(); |
43 | | - |
44 | | - // Benchmarks against long haystacks |
45 | | - |
46 | | - c.bench_function("String::find with long haystack", |b| { |
47 | | - b.iter(|| { |
48 | | - for word in &words { |
49 | | - content.find(word); |
50 | | - } |
51 | | - }) |
52 | | - }); |
53 | | - |
54 | | - c.bench_function( |
55 | | - "memmem::TwoWaySearcher::search_in with long haystack", |
56 | | - |b| { |
57 | | - b.iter(|| { |
58 | | - for twoway_word in &twoway_words { |
59 | | - twoway_word.search_in(content.as_bytes()); |
60 | | - } |
61 | | - }) |
62 | | - }, |
63 | | - ); |
64 | | - |
65 | | - c.bench_function("twoway::find_bytes with long haystack", |b| { |
66 | | - b.iter(|| { |
67 | | - for word in &words { |
68 | | - twoway::find_bytes(content.as_bytes(), word.as_bytes()); |
69 | | - } |
70 | | - }) |
71 | | - }); |
72 | | - |
73 | | - c.bench_function("strstr_avx2_original with long haystack", |b| { |
74 | | - b.iter(|| { |
75 | | - for word in &words { |
76 | | - unsafe { |
77 | | - strstr_avx2_original(content.as_bytes(), word.as_bytes()); |
78 | | - } |
79 | | - } |
80 | | - }) |
81 | | - }); |
82 | | - |
83 | | - c.bench_function("strstr_avx2_rust_simple with long haystack", |b| { |
84 | | - b.iter(|| { |
85 | | - for word in &words { |
86 | | - unsafe { |
87 | | - strstr_avx2_rust_simple(content.as_bytes(), word.as_bytes()); |
88 | | - } |
89 | | - } |
90 | | - }) |
91 | | - }); |
92 | | - |
93 | | - c.bench_function("strstr_avx2_rust_simple_2 with long haystack", |b| { |
94 | | - b.iter(|| { |
95 | | - for word in &words { |
96 | | - unsafe { |
97 | | - strstr_avx2_rust_simple_2(content.as_bytes(), word.as_bytes()); |
98 | | - } |
99 | | - } |
100 | | - }) |
101 | | - }); |
102 | | - |
103 | | - c.bench_function("strstr_avx2_rust_fast with long haystack", |b| { |
104 | | - b.iter(|| { |
105 | | - for word in &words { |
106 | | - unsafe { |
107 | | - strstr_avx2_rust_fast(content.as_bytes(), word.as_bytes()); |
108 | | - } |
109 | | - } |
110 | | - }) |
111 | | - }); |
112 | | - |
113 | | - c.bench_function("strstr_avx2_rust_fast_2 with long haystack", |b| { |
114 | | - b.iter(|| { |
115 | | - for word in &words { |
116 | | - strstr_avx2_rust_fast_2(content.as_bytes(), word.as_bytes()); |
117 | | - } |
118 | | - }) |
119 | | - }); |
120 | | - |
121 | | - c.bench_function("strstr_avx2_rust_aligned with long haystack", |b| { |
122 | | - b.iter(|| { |
123 | | - for word in &words { |
124 | | - unsafe { |
125 | | - strstr_avx2_rust_aligned(content.as_bytes(), word.as_bytes()); |
126 | | - } |
127 | | - } |
128 | | - }) |
129 | | - }); |
130 | | - |
131 | | - c.bench_function("StrStrAVX2Searcher::search_in with long haystack", |b| { |
132 | | - b.iter(|| { |
133 | | - for avx2_word in &avx2_words { |
134 | | - avx2_word.search_in(content.as_bytes()); |
135 | | - } |
136 | | - }) |
137 | | - }); |
138 | | - |
139 | | - c.bench_function("DynamicAvx2Searcher::search_in with long haystack", |b| { |
140 | | - b.iter(|| { |
141 | | - for searcher in &dynamic_avx2_searchers { |
142 | | - searcher.search_in(content.as_bytes()); |
143 | | - } |
144 | | - }) |
145 | | - }); |
146 | | - |
147 | | - // Benchmarks against short haystacks |
148 | | - // |
149 | | - // Since words are ordered by length, pick a word as needle |
150 | | - // and use bigger words as haystacks. |
151 | | - |
152 | | - c.bench_function("String::find with short haystack", |b| { |
153 | | - b.iter(|| { |
154 | | - for (i, word) in words.iter().enumerate() { |
155 | | - for content in &words[(i + 1)..] { |
156 | | - content.find(word); |
157 | | - } |
158 | | - } |
159 | | - }) |
160 | | - }); |
161 | | - |
162 | | - c.bench_function( |
163 | | - "memmem::TwoWaySearcher::search_in with short haystack", |
164 | | - |b| { |
165 | | - b.iter(|| { |
166 | | - for (i, word) in twoway_words.iter().enumerate() { |
167 | | - for content in &words[(i + 1)..] { |
168 | | - word.search_in(content.as_bytes()); |
169 | | - } |
170 | | - } |
171 | | - }) |
172 | | - }, |
173 | | - ); |
174 | | - |
175 | | - c.bench_function("twoway::find_bytes with short haystack", |b| { |
176 | | - b.iter(|| { |
177 | | - for (i, word) in words.iter().enumerate() { |
178 | | - for content in &words[(i + 1)..] { |
179 | | - twoway::find_bytes(content.as_bytes(), word.as_bytes()); |
180 | | - } |
181 | | - } |
182 | | - }) |
183 | | - }); |
184 | | - |
185 | | - c.bench_function("strstr_avx2_original with short haystack", |b| { |
186 | | - b.iter(|| { |
187 | | - for (i, word) in words.iter().enumerate() { |
188 | | - for content in &words[(i + 1)..] { |
189 | | - unsafe { |
190 | | - strstr_avx2_original(content.as_bytes(), word.as_bytes()); |
191 | | - } |
192 | | - } |
193 | | - } |
194 | | - }) |
195 | | - }); |
196 | | - |
197 | | - c.bench_function("strstr_avx2_rust_simple with short haystack", |b| { |
198 | | - b.iter(|| { |
199 | | - for (i, word) in words.iter().enumerate() { |
200 | | - for content in &words[(i + 1)..] { |
201 | | - unsafe { |
202 | | - strstr_avx2_rust_simple(content.as_bytes(), word.as_bytes()); |
203 | | - } |
204 | | - } |
205 | | - } |
206 | | - }) |
207 | | - }); |
208 | | - |
209 | | - c.bench_function("strstr_avx2_rust_simple_2 with short haystack", |b| { |
210 | | - b.iter(|| { |
211 | | - for (i, word) in words.iter().enumerate() { |
212 | | - for content in &words[(i + 1)..] { |
213 | | - unsafe { |
214 | | - strstr_avx2_rust_simple_2(content.as_bytes(), word.as_bytes()); |
215 | | - } |
216 | | - } |
217 | | - } |
218 | | - }) |
219 | | - }); |
220 | | - |
221 | | - c.bench_function("strstr_avx2_rust_fast with short haystack", |b| { |
222 | | - b.iter(|| { |
223 | | - for (i, word) in words.iter().enumerate() { |
224 | | - for content in &words[(i + 1)..] { |
225 | | - unsafe { |
226 | | - strstr_avx2_rust_fast(content.as_bytes(), word.as_bytes()); |
227 | | - } |
228 | | - } |
229 | | - } |
230 | | - }) |
231 | | - }); |
232 | | - |
233 | | - c.bench_function("strstr_avx2_rust_fast_2 with short haystack", |b| { |
234 | | - b.iter(|| { |
235 | | - for (i, word) in words.iter().enumerate() { |
236 | | - for content in &words[(i + 1)..] { |
237 | | - strstr_avx2_rust_fast_2(content.as_bytes(), word.as_bytes()); |
238 | | - } |
239 | | - } |
240 | | - }) |
241 | | - }); |
242 | | - |
243 | | - c.bench_function("strstr_avx2_rust_aligned with short haystack", |b| { |
244 | | - b.iter(|| { |
245 | | - for (i, word) in words.iter().enumerate() { |
246 | | - for content in &words[(i + 1)..] { |
247 | | - unsafe { |
248 | | - strstr_avx2_rust_aligned(content.as_bytes(), word.as_bytes()); |
249 | | - } |
250 | | - } |
251 | | - } |
252 | | - }) |
253 | | - }); |
254 | | - |
255 | | - c.bench_function("StrStrAVX2Searcher::search_in with short haystack", |b| { |
256 | | - b.iter(|| { |
257 | | - for (i, word) in avx2_words.iter().enumerate() { |
258 | | - for content in &words[(i + 1)..] { |
259 | | - word.search_in(content.as_bytes()); |
260 | | - } |
261 | | - } |
262 | | - }) |
263 | | - }); |
264 | 113 |
|
265 | | - c.bench_function("DynamicAvx2Searcher::search_in with short haystack", |b| { |
266 | | - b.iter(|| { |
267 | | - for (i, searcher) in dynamic_avx2_searchers.iter().enumerate() { |
268 | | - for content in &words[(i + 1)..] { |
269 | | - searcher.search_in(content.as_bytes()); |
270 | | - } |
271 | | - } |
272 | | - }) |
273 | | - }); |
| 114 | + group.finish(); |
| 115 | + } |
274 | 116 | } |
275 | 117 |
|
276 | 118 | criterion_group!(benches, criterion_benchmark); |
|
0 commit comments