When doing interviews, one of the theoritical exercise we gave to potential employees was to find a list of letter pairs in a long string (e.g. a book).
After a few times, I had a few ideas myself of how such a (very useful indeed) algorithm could be written, and then I could not resist anymore and I had to write one.
This implementation ignores disk reading speed and JIT, and focuses on the algorithm itself.
- Socolin for
MemoryArraySearchDigraphsLookupV2 - christianrondeau for
BitShiftingBinarySearchDigraphsLookup
This runs in Release mode, searching for 5 digraphs in Frankenstein, 1k warmup runs and 1k measured runs. Not perfect but good enough.
This is the "naive" implementation (SubstringPairsDigraphsLookup). We check every pairs of letters using substring and use a dictionary for lookup.
Average time: 8457 µs Percentiles (in microseconds): | Min | 25% | 50% | 75% | 95% | Max | | 8089 | 8247 | 8505 | 8619 | 8723 | 13298 |
This is my implementation (BitShiftingBinarySearchDigraphsLookup), which reads one character as a byte, and use bit shifting and bit masking to create an integer lookup key. I also use binary search instead of a dictionary (note that a linear search was pretty much the same speed, but with more keys it would have been faster. For a huge amount of digraphs, a dictionary would have been better).
Average time: 2316 µs Percentiles (in microseconds): | Min | 25% | 50% | 75% | 95% | Max | | 2273 | 2293 | 2300 | 2310 | 2340 | 3389 |
This is @Socolin's solution, which creates an array large enough to cover all possible ASCII digraphs. This replaces a few arithmetic operations by a larger memory space, making this the fastest solution!
Average time: 1668 µs
Percentiles (in microseconds):
| Min | 25% | 50% | 75% | 95% | Max |
| 1627 | 1648 | 1653 | 1660 | 1678 | 2389 |
Other variation using SIMD / Parallelism give similar result. But with GPU it's faster:
Average time: 0469 µs
Percentiles (in microseconds):
| Min | 25% | 50% | 75% | 95% | Max |
| 360 | 472 | 475 | 481 | 502 | 1694 |