The project's home page: https://github.com/marknelsonengineer/BenchMark (hosted by GitHub)
The source code is documented at: https://www2.hawaii.edu/~marknels/BenchMark (hosted by The University of Hawaiʻi at Mānoa)
BenchMark is a toolset for analysing the relative performance of various things that I'm interested in.
First, I'm shocked and thrilled at the performance of Release-mode memcpy()
and memset() functions from glibc. Look at these results:
Results of memcpy() in CPU ticks using different compilers & options
| n | Release | MinSizeRel | Clang |
|---|---|---|---|
| 8 | 1 | 13 | 11 |
| 16 | 2 | 13 | 9 |
| 32 | 2 | 13 | 12 |
| 64 | 6 | 13 | 12 |
| 128 | 4 | 19 | 10 |
| 256 | 4 | 22 | 15 |
| 512 | 12 | 27 | 28 |
| 1024 | 20 | 38 | 31 |
| 2048 | 38 | 67 | 56 |
| 4096 | 86 | 141 | 94 |
| 8192 | 172 | 485 | 179 |
| 16384 | 744 | 719 | 840 |
| 32768 | 1836 | 1957 | 1698 |
| 65536 | 3264 | 4240 | 3386 |
- Notice how the speeds are about the same when
n< 256 - Notice how much more efficient the Release build over Clang or MinSizeRel,
especially at low
n. - The results of
memcpy()are so good, it makes me doubt whether I can improve the speed with hand-coded assembly. - Results for
memset()are very similar tomemcpy()(Not shown)
| Build | Compiler | Options |
|---|---|---|
| Release | Statically linked gcc executable | -lstdc++ -fuse-ld=gold -march=native -mtune=native -Ofast -funroll-loops -static -mfma |
| MinSizeRel | Dynamically linked gcc executable | -lstdc++ -fuse-ld=gold -march=native -mtune=native -Oz -mfma |
| Clang | Dynamically linked Clang executable | -march=native |
The tests were performed on a MacBook Pro:
- Architecture: Coffee Lake – 9th Generation Intel Core
- Cache line size: 64 bytes