unicode
Case folding benchmark
Times moonbit-community/unicode-casefold on four of the
bench/normperf corpora: English prose (vfear11a)
and personal names in Latin, Cyrillic and Han script.
moon bench ./bench/casefoldperf --target native
| Case | Measures |
|---|---|
case_fold |
Full default case folding of the corpus |
case_fold/folded |
The same on already folded text, which is copied unchanged |
nfkc_casefold |
NFKC_Casefold, including its final NFC |
caseless_match/same |
caseless_match of the corpus with itself |
caseless_match/upper |
caseless_match of the corpus with its full uppercase mapping |
Example results
One run on an Apple M3 Max, native backend, in nanoseconds per code point. "Before" looked every code point up with a binary search over all mappings and folded both strings to compare them; "after" uses the kind trie, skips ASCII without decoding it, copies unchanged runs at once and compares equal code units before decoding anything.
| Corpus | Case | Before | After |
|---|---|---|---|
vfear11a |
case_fold |
10.8 | 0.82 |
case_fold/folded |
10.2 | 0.46 | |
nfkc_casefold |
12.9 | 0.86 | |
caseless_match/same |
21.6 | 0.56 | |
caseless_match/upper |
32.7 | 3.87 | |
TestNames_Latin |
case_fold |
11.9 | 1.83 |
nfkc_casefold |
14.1 | 1.86 | |
caseless_match/upper |
33.5 | 3.15 | |
TestNames_Russian |
case_fold |
13.5 | 3.81 |
nfkc_casefold |
15.3 | 4.16 | |
caseless_match/upper |
33.1 | 4.82 | |
TestNames_Chinese |
case_fold |
10.1 | 2.20 |
nfkc_casefold |
14.7 | 2.50 | |
caseless_match/upper |
21.2 | 0.57 |
1# Case folding benchmark23Times `moonbit-community/unicode-casefold` on four of the4[`bench/normperf`](../normperf/README.md) corpora: English prose (`vfear11a`)5and personal names in Latin, Cyrillic and Han script.67```sh8moon bench ./bench/casefoldperf --target native9```1011| Case | Measures |12|---|---|13| `case_fold` | Full default case folding of the corpus |14| `case_fold/folded` | The same on already folded text, which is copied unchanged |15| `nfkc_casefold` | NFKC_Casefold, including its final NFC |16| `caseless_match/same` | `caseless_match` of the corpus with itself |17| `caseless_match/upper` | `caseless_match` of the corpus with its full uppercase mapping |1819## Example results2021One run on an Apple M3 Max, native backend, in nanoseconds per code point.22"Before" looked every code point up with a binary search over all mappings23and folded both strings to compare them; "after" uses the kind trie, skips24ASCII without decoding it, copies unchanged runs at once and compares equal25code units before decoding anything.2627| Corpus | Case | Before | After |28|---|---|---:|---:|29| `vfear11a` | `case_fold` | 10.8 | 0.82 |30| | `case_fold/folded` | 10.2 | 0.46 |31| | `nfkc_casefold` | 12.9 | 0.86 |32| | `caseless_match/same` | 21.6 | 0.56 |33| | `caseless_match/upper` | 32.7 | 3.87 |34| `TestNames_Latin` | `case_fold` | 11.9 | 1.83 |35| | `nfkc_casefold` | 14.1 | 1.86 |36| | `caseless_match/upper` | 33.5 | 3.15 |37| `TestNames_Russian` | `case_fold` | 13.5 | 3.81 |38| | `nfkc_casefold` | 15.3 | 4.16 |39| | `caseless_match/upper` | 33.1 | 4.82 |40| `TestNames_Chinese` | `case_fold` | 10.1 | 2.20 |41| | `nfkc_casefold` | 14.7 | 2.50 |42| | `caseless_match/upper` | 21.2 | 0.57 |43# Case folding benchmark Times `moonbit-community/unicode-casefold` on four of the [`bench/normperf`](../normperf/README.md) corpora: English prose (`vfear11a`) and personal names in Latin, Cyrillic and Han script. ```sh moon bench ./bench/casefoldperf --target native ``` | Case | Measures | |---|---| | `case_fold` | Full default case folding of the corpus | | `case_fold/folded` | The same on already folded text, which is copied unchanged | | `nfkc_casefold` | NFKC_Casefold, including its final NFC | | `caseless_match/same` | `caseless_match` of the corpus with itself | | `caseless_match/upper` | `caseless_match` of the corpus with its full uppercase mapping | ## Example results One run on an Apple M3 Max, native backend, in nanoseconds per code point. "Before" looked every code point up with a binary search over all mappings and folded both strings to compare them; "after" uses the kind trie, skips ASCII without decoding it, copies unchanged runs at once and compares equal code units before decoding anything. | Corpus | Case | Before | After | |---|---|---:|---:| | `vfear11a` | `case_fold` | 10.8 | 0.82 | | | `case_fold/folded` | 10.2 | 0.46 | | | `nfkc_casefold` | 12.9 | 0.86 | | | `caseless_match/same` | 21.6 | 0.56 | | | `caseless_match/upper` | 32.7 | 3.87 | | `TestNames_Latin` | `case_fold` | 11.9 | 1.83 | | | `nfkc_casefold` | 14.1 | 1.86 | | | `caseless_match/upper` | 33.5 | 3.15 | | `TestNames_Russian` | `case_fold` | 13.5 | 3.81 | | | `nfkc_casefold` | 15.3 | 4.16 | | | `caseless_match/upper` | 33.1 | 4.82 | | `TestNames_Chinese` | `case_fold` | 10.1 | 2.20 | | | `nfkc_casefold` | 14.7 | 2.50 | | | `caseless_match/upper` | 21.2 | 0.57 |