README.md
- ~1.2 clocks per byte
- 100% C (& C++ compatible) without inline assembly
- Both 32 and 64 bits supported
- Portable scalar functions faster than SIMD functions
- More than 3 times faster than naive solution
🆕 more faster, beats even other fast assembler functions
- CPU Sandy Bridge i7-2600k at 4.4GHz, gcc 6.2
- Single thread
- Realistic and practical benchmark with large files.
- No PURE cache benchmark
- Uniform/Skewed distribution:
- Uniform: enwik9
- Skewed: enwik9 bwt generated w. libdivsufsort
- Accurate benchmarking with command "turbohist file -I15"
| Function | Uniform: Time MB/s | Skewed: Time MB/s |
|---|---|---|
| hist_8_32 | 2823 | 2823 |
| hist_4_32 | 2817 | 2679 |
| count2x64 | 2788 | 2788 |
| hist_8_128 | 2780 | 2778 |
| hist_4_128 | 2781 | 2661 |
| hist_8_64 | 2735 | 2734 |
| hist_4_64 | 2731 | 2616 |
| count2x64c | 2433 | 2432 |
| hist_4_8 | 2173 | 2050 |
| hist_8_8 | 2166 | 1897 |
| hist_1_8 | 1840 | 909 |
(bold = pareto) MB=1.000.000
Compile:
makeUsage:
turbohist fileEnvironment:
OS/Compiler (32 + 64 bits):
- Linux: GNU GCC (>=4.6)
- clang (>=3.2)
- Windows: MinGW
- Windows: Visual Studio 2015
- ARM 64 bits/ gcc
References:
Countbench: https://github.com/nkurz/countbench (including "count2x64" with inline assembly)
Last update: 24 DEC 2016