search

xmrig / xmrig

RandomX, KawPow, CryptoNight and GhostRider unified CPU/GPU miner and RandomX benchmark
https://xmrig.com/
GNU General Public License v3.0
8.71k stars 3.45k forks source link

Hashrate dropped 30% for CNv2 on ARMv7 compared to CNv1 #865

Closed bladeFury closed 6 years ago

bladeFury commented 6 years ago

I've done some test, found out that the hashrate loss is due to VARIANT2_INTEGER_MATH macro, this line : https://github.com/xmrig/xmrig/blob/2b0b71b9f695466f8b434fbbbcbfecfb3f9ecd60/src/crypto/CryptoNight_monero.h#L116 Seems "load" sqrt_result_0 is very slow. I did the following tests:

change line 116 to cl ^= sqrt_result_##part; \ -> slow change line 116 to cl ^= 12345;\ -> fast

Could you do some optimization with this macro or give me some hint ? @SChernykh @xmrig

SChernykh commented 6 years ago

When you do cl ^= 12345; you basically remove integer math entirely. Compiler sees that you don't use any results and removes it from the loop.

bladeFury commented 6 years ago

@SChernykh cl ^= 12345 still calculate cl = cl XOR 12345, so compiler can't remove this from loop because it's changing cl every loop, right? And I checked the assembly code , movw r2, #12345 and eors r1, r2 is still there.

SChernykh commented 6 years ago

Compiler removes div and sqrt code completely if it sees that results are not used. This is why it gets faster.

bladeFury commented 6 years ago

@SChernykh ohh I got it, Compilers are really doing a lot work, Thanks for pointing it out.

Is there any optimization possibilities for this snippet? 30% hr drop seems a lot compared to other cpus.