Wrong result on test case, benchmark on 2950X about 1brc HOT 6 CLOSED

• right I think I do rounding to 0 not actual rounding. Not a big issue.
• cities appearing twice is definitely weird though. Will check.
• from these timings I'd infer that your CPU doesn't support PDEP maybe? Does it have the BMI2 feature flag? Maybe it's emulated by the compiler now which could be slow. I'll try to update the code to print a warning if it's not supported.
• you can just run with -j1 to run only 1 thread.
• I'll download your input and use that from now.
• the remaining performance difference is probably just differences in CPU. It does seem that yours has more L1 cache per thread (48kB Vs my 32kB), but that doesn't explain the slower speed.

from 1brc.

Oh right this:
https://en.wikipedia.org/wiki/X86_Bit_manipulation_instruction_set#%3A%7E%3Atext%3Dwithout_affecting_flags-%2CParallel_bit_deposit_and_extract%2Cto_be_packed_or_unpacked.?wprov=sfla1

AMD processors before Zen 3 that implement PDEP and PEXT do so in microcode, with a latency of 18 cycles rather than (Zen 3) 3 cycles. As a result it is often faster to use other instructions on these processors.

from 1brc.

fixed the wrong rounding and the initial bad city was an off-by-one in the initialization.

I don't think I will make non-PDEP code, since it uses a different data storage format.

I'll make a new benchmark at some later point today.

from 1brc.

Thanks! Now the performance numbers make more sense (single thread HT vs no HT not much difference).

It seems Intel's HT is just worse than AMD's.

from 1brc.

Great finding! Also, can you run 6 thread vs 12 thread both with HT on? Your benchmark currently only has 6t HT off vs 12t HT on.

The difference is ~60% for both of ours code on AMD 2950X. I'm curious what's the number on an Intel CPU

from 1brc.

Added latest results here:
https://curiouscoding.nl/posts/1brc/

from 1brc.