WildRig 0.10.5 beta

multi-algo version is here

Download for Windows(all versions can be found here)
dependencies: Visual C++ Redistributable 2015 x64
note: don't forget to use 398+ drivers for NVIDIA

Download for Linux(no api version)
dependencies: libcurl, libuv1, libmicrohttpd10
note: compiled under Ubuntu 16.04 LTS, so if you get problem with libmicrohttpd10, use "no api" version
note: don't forget to use 398+ drivers for NVIDIA

Download for HiveOS(custom miner integration)
dependencies: libcurl, libuv1, libmicrohttpd10
note: use wildrig as a name for miner while installation

Download HiveOS custom package from HiveOS dev's

mirror on github

Discord: https://discord.gg/RkywAu5

advantages over minerd and ccminer:
- faster for 4%-10% on AMD in comparison with minerd using boosted kernel
- more solved shares than ccminer can do
- fast scratchpad update(less hashrate drop when new addendum blocks received and pushed to vram)
- auto re-downloading scratchpad if failed(known problem with ccminer), also patching the local file every 10Kb of new data(each ~1hour)
- asynchronous share send to avoid wasting GPU resources when internet connection is slow
- no limit for amount of GPU's as in original minerd
- watchdog(see description for --watchdog parameter)
- new interface
- API to check miner stats via web

known issues:
- watchdog can hang miner while trying to terminate the process(depends on how badly drivers are hanged, happens only for AMD as I know), so use it carefully. Later I will extend some parameters to call custom bat-file with rebooting the system or restart the driver if needed
- getting n/a for one or few cards at start(no "scratchpad updated" messages for them). Can happen on rigs with a lot of GPU's(e.g. 12). Use parameter --safe-scratchpad-update in this case
- statistics doesn't work in HiveOS GUI(working on that)

kernels:
AMD cards can use different kernels via --opencl-kernel parameter. Also you can set different kernel for each card, just use comma separation. Some cards can hang on boosted kernel, so probably better to set default one instead of lowering clocks.
- 0 is default one
- 1 can increase hashrate a bit on some cards like RX550
- 2 should be used with old hardware like Radeon HD 7xxx
- 3 boosted kernel
- 4 boosted kernel with the same changes as in kernel 1

note for AMD Vega's
for mining Boolberry better to set --opencl-launch to 128x32 or 256x32 and get significant hashrate increase

note for RX 570/580
for mining Boolberry better to set --opencl-launch to 240x32 and get additional ~20kH/s

Testing and reporting issues are welcome!

coins:
- Boolberry
- Purk

dev-fee:
currently 2% only, fully compensates on NVIDIA's because of improvements and for AMD's because of better hashrate.

TODO:
- make avg. effort calculation based on some amount of shares, instead of all shares
- implement some experiment parameter --cheat-effort, when miner will try to get new job when effort is not acceptable
- implement CPU mining(?)
- implement executing custom command when watchdog catch the gpu hang
- ADL/NVML for monitoring videocards health
- open source some day once all things will be done good

Example of command-line start for AMD videocards(with loop if miner crash):

Code: (bat)

@echo off

:loop
wildrig.exe --opencl-launch=256x0 --send-stale --url=bbr.luckypool.io:5577 --user=WALLET --pass=IFNEEDED --scratchpad-file=scratchpad.bin --scratchpad-url=http://eu-bbr.luckypool.io/scratchpad.bin
if ERRORLEVEL 1000 goto custom
timeout /t 5
goto loop

:custom
echo Custom command here
timeout /t 5
goto loop

Same example for NVIDIA using CUDA:

Code: (bat)

@echo off

:loop
wildrig.exe --cuda-launch=1024x256 --send-stale --url=bbr.luckypool.io:5577 --user=WALLET --pass=IFNEEDED --scratchpad-file=scratchpad.bin --scratchpad-url=http://eu-bbr.luckypool.io/scratchpad.bin
if ERRORLEVEL 1000 goto custom
timeout /t 5
goto loop

:custom
echo Custom command here
timeout /t 5
goto loop

Linux bash-scripts included into archive.

what is avg. effort:
every videocard do some amount of hashes per iteration, and there should be done around diff(set by pool) hashes before miner will find a share. This will be 100% effort. If videocards do less hashes and find a share - effort will be below 100%, more - it will be above 100%. But because of random nature, effort is always differ. A lot of things have influence on effort, but if you get high values(after at least ~30 minutes statistic from start or dev-fee mining finished) take a look on:
- how good variable diff on pool. Giving you high difficulty jobs, which are not suitable for your hashpower can lead to higher effort
- high intensity, more than needed for best hashrate, can lead to higher effort in cases when miner got a new job, but videocard is still hashing and don't know about it. That last hashes will be wasted, also you can get stale share

optimal intensity:
try to set intesity as low as possible, but not lower than point where hashrate goes down a lot

optimal strategy:
if you use static diff, probably --strategy set to 1 will be better in this case. At least it's true for my rigs with static diff set to hashrate_of_rig * 10(e.g. 4820kH/s rig uses diff 48000000).

Changelog:

Code:

0.10.5:
- fixed submit method, now nonce again 64 bit value

0.10.4:
- added new optimized kernels which can improve hashrate for 4-10% on AMD cards
- fixed AMD videocard numeration(was broken in 0.10.3)
- added possibility to set different kernel for each videocard(use parameter --opencl-kernels)
- added parameter --safe-scratchpad-update for rigs with huge amount of gpu's or just not stable motherboards
- [HiveOS]reverted CUDA to 9.0 for compatibility with older drivers

0.10.3
- added workaround for bad risers(to disable it use parameter --good-riser), this should fix problem with n/a hashrate for some gpu's
- CUDA 9.2 used now for Windows build to, so probably newer drivers will be needed
- fixed incorrect choosing of devices when not all of them are specified using --opencl-devices/--cuda-devices
- some minor changes to kernels, should use less registers now
- correct deinitialization of CUDA on exit

0.10.2
- one more fix to crazy effort values after donation mining
- fixed low difficulty shares for old hardware with default kernel
- added new slightly modified kernel(use parameter --opencl-kernel 1), it can give 1%-2% on some cards(like RX550) and up to 30% on old Radeons(7870 now shows ~580kH/s instead of ~440)

0.10.1
- no more final CPU-stage for share calculation, now solved shares should be send faster
- now --send-stale works even with changed scratchpad and pools can accept such shares
- less locks on scratchpad patch/update to make it work faster

0.10.0
- improved precision for share validation, now more shares should be solved again
- now miner can be started without specifying all devices(and --opencl-platform for AMD), but don't forget to specify opencl-launch or cuda-launch
- disabled OpenCL mode for NVIDIA

0.9.25
- reworked share validation, now calculation are the same as on pool side
- after donation mining miner will reconnect to pool now
- implemented --send-stale parameter to send stale shares when new job arrived without scratchpad update
- removed restart miner when watchdog enabled and miner found not valid share on CPU(in 0.9.24 it was a mistake...)
- extended info for invalid share

0.9.24
- fixed rare case when one of 2+ shares solved at once(that green "yay!" messages) can be ignored
- now miner shows difficulty sent from pool, instead of calculation based on target value
- reverted back validation of share(hash will be checked for "<= target", instead of "< target")

0.9.23
- fixed crazy avg. effort after donation mining
- fixed "Duplicate share" sending in rare cases(was possible on fast rigs, when new job is not arriving too long)
- fixed crash when --opencl-devices was set not like 0,1,2,3 and so on

0.9.22
- fixed job timeout problem, now miner make a correct request for new job

0.9.21
- fixed one more ignoring of odd data from pool(found with official Purk pool)

0.9.20
- implemented CUDA calculations for NVIDIA cards
- improved scratchpad fails handling
- instead of luck miner will calculate avg. effort
- now miner will send getjob to pool before disconnect on "job-timeout" event
- now scratchpad will be re-downloaded if it 1 day old
- disabled watchdog by default

0.9.19
- improved scratchpad management, now scratchpad will be updated on disk every 10Kb of new data from pool(each ~1hour)
- fixed possible crash in rare cases
- fixed infinite rejects if miner was started with fresh new scratchpad and only 1 addendum update from pool
- added logging for http daemon to diagnose why API is not working
- now it's possible to set donate level to zero on rigs with 1-2 videocards connected(currently Windows only)
- added parameter --multiple-instance to allow start multiple instances of one miner with the same --opencl-platform

0.9.18
- fixed --no-watchdog parameter, now it should work
- set timeout for watchdog to 30 seconds now
- now if downloading speed of scratchpad is 512 bytes for more than 30 seconds, miner will restart downloading it

0.9.17
- moved scratchpad update into a better place, this should decrease amount of "Low difficulty" shares before "new job"
- added general uptime value to API

0.9.16
- code much more safe now
- fixed problem with watchdog not working(there was a bug with reading parameter no-watchdog)

0.9.15
- improved watchdog, now it works separately and won't decrease hashrate for AMD's when use low intensity values
- improved scratchpad fail handling, now miner will try to reset it to state saved on disk and login again, only then redownload it if fail again
- added experimental parameter --opencl-force1x to support old hardware like Readon HD 7xxx

0.9.14
- fixed small memory leak
- implemented watchdog, now miner will exit with code 1000 if gpu hangs in OpenCL kernel for more than 5 sec.(can be disabled with parameter --no-watchdog)

0.9.13
- reworked strategy a bit, now nvidias should support 0 and 1 too. Changed default strategy to that used in other miners, should be better with vardiff
- fixed processing error "invalid miner height, re-download scratchpad", now miner definitely will redownload it

0.9.12
- fixed "Duplicate share" error after donation mining finished, same was possible to reproduce with pause/resume commands

0.9.11
- new parameter --print-level
- more accurate luck calculation
- dynamic size for scratchpad buffer(no need in recompilation in future, less RAM used for PURK coin to mine :))

0.9.10
- now videocards sorted by BusID
- added luck calculation(need testing)
- fixed crash when logging and --print-full is used
- few more checks for scratchpad fails(also in some cases miner will reconnect if it failed and only then redownload it if this not solve the problem, need testing)

0.9.9:
- fixed JSON decode failed
- more checks for scratchpad failure

0.9.8:
- added validation scratchpad at start and redownloading it if needed
- now should work on old cards too
- added --job-timeout parameter, now miner will disconnect if there is no new job for some time and videocards found all shares(could help for "beast" rigs on Vega)
- a bit more safe code

0.9.7:
- fixed an issue when miner can't redownload scratchpad because of some accident on pool side and going into infinite loop

0.9.6:
- improved logic, now videocard will continue hash even if it waits for new job, just to avoid situation with lowering clocks. Will make something like powersave option later
- added parameter --max-rejects