Open achimste opened 5 months ago
Matthies
commented
5 months ago Not sure if it is worth to switch to C++20 but I will give it a try. First step (and not last hopefully) will be to check if all platforms that I want to support allow switching to new standard.
Matthies
commented
5 months ago Speed comparison of master vs. 0935b7ae19ed0282925456cb2568875910b9ffce of branch c++20 on BMI2 Intel CPU running Linux.
Every version was measured three times using Stockfish perf test:
sudo perf stat -r 5 -a -B -e cycles:u,instructions:u ./RubiChess bench > /dev/null
| Version | GCC-11.4 | Clang-14.0 | ICX-2023.0.0 |
|---|---|---|---|
| Master | 4.85s / 4.85s / 4.86s | 4.84s / 4.86s / 4.86s | 4.62s / 4.62s / 4.62s |
| c++20 | 4.87s / 4.88s / 4.87s | 4.92s / 4.93s / 4.92s | 4.66s / 4.66s / 4.67s |
Seems that the compilers are not as smart as expected. Or my generic pullXsb implementation can be improved.
Matthies
commented
5 months ago After fixing pullLsb in bc22873a0d484a4adbe8773490abebacfb18f3ad this looks better now. Well, not for Clang. This testing is on AVX2 AMD Ryzen 3700x running Linux and using more recent comnpilers.
| Version | GCC-13.2 | Clang-16.0 | ICX-2023.2.0 |
|---|---|---|---|
| Master | 3.33s / 3.36s / 3.33s | 3.36s / 3.36s / 3.35s | 3.35s / 3.34s / 3.35s |
| c++20 | 3.32s / 3.32s / 3.33s | 3.42s / 3.41s / 3.41s | 3.33s / 3.33s / 3.34s |
Matthies
commented
4 months ago
Hi!
Sorry, this is not an issue but a suggestion and I just would like to hear your opinion.
In my chess engines, I replaced all that CPU-specialised code (popcount, msb, lsb, etc) with C++20 std commands, where the compiler automatically uses the corresponding intrinsics or other codes depending on the code generation flags.
So my question is: Would you like to switch to C++20 or remain at a lower version? At least with C++20 your code would simplify (a lot?).
I downloaded your latest code and I am indeed replacing your CPU specific code with std-lib commands. I am compiling for Windows only.
Regards Achim