TalkChess.com

Henk, Just tried what happens when I disable the hash-table and all other pruning, just plain a-b with 2 killers and history for move-ordering, and in quiescence at depth zero besides captures I generate direct checking moves, nothing else. On a single core with these conditions it takes about 3 min...

This is not difficult at all. Nightmare on 10 cores needs 12 ply and 11.4 msec to solve this. The IO probably takes the largest part of the time, console IO is rather slow on Windows 10.

You can also use:

_lzcnt_u64() and _tzcnt_u64() which are declared in <intrin.h>

uint64_t _lzcnt_u64 (uint64_t b);
uint64_t _tzcnt_u64 (uint64_t b);

Back in 1977 I started working on my first chess engine for the Intel 8080 and 8086, because compilers didn't exist for the Intel architecture there was no other option than to program in assembler. My memory of that time is not so good, but I still remember that my move-generator ran at 15 knps, ro...

When I had first written qperft (I was using a Pentium M at the time), I tried to speed it up by using assembler. To not waste time on non-time-ritical code, I started with the assembler output of the C compiler, and 'hand-optimized' the code of the time-critical loops. It turned out I could delete...

thought I would reply out of interest. On my 2 year old MacBook, 2.2ghz i7 circa 2017 I see this number for perf 6 from start position: White(1): perft 6 total moves=119060324 time=3.11 White(1): Here's the next one: White(1): perft 7 total moves=3195901860 time=82.65 My perft was never optimized f...

AFAIK, movegeneration speed is a good chunk of the runtime in modern engines. Atleast in FabChess, it is about 25-30% of the time used, being on par with the evaluation function (IIRC). 25% to 30% of the time used by the move generation is quite a lot. Maybe it depends upon staged move generation o...

On my 6950X at 4GHz. I get on average 52 mnps without bulk counting and including updating/downdating hash, pawn-hash, material-hash, material-score and phase. The latest processors from Intel have a 15 to 20% higher IPC and a somewhat higher clock-frequency, so my guess is that I would get between...

On my 6950X at 4GHz. I get on average 52 mnps without bulk counting and including updating/downdating hash, pawn-hash, material-hash, material-score and phase. The latest processors from Intel have a 15 to 20% higher IPC and a somewhat higher clock-frequency, so my guess is that I would get between...

Well, I don't think it is impossible to beat the compiler with assembler code, but a programmer who knows all the ins and outs of the compiler he uses is capable of producing a program in a short amount of time that is very hard to beat with assembler code. And then there is the issue of portability...