CCminer(SP-MOD) Modded GPU kernels.

[djm34]

@anyone who wants to poke around in the Ethereum CUDA miner source, i've implemented sp_'s mods and also cleaned it up a bit.

from what I checked, using uint2 was a bit slower (for once) than using uint64_t in keccak

No it isn't, but I think you have to seperate the loops into two and remove all the constant inputs, load from memory differenty, play with the pragma urolls, alter the launch bounds etc..

But keccak was just 13% of the total algorithm so the effort will probobly just gain a few percent.

I tried yesterday it was slower on my 980

[1714693274]

1714693274

[1714693274]

1714693274

[dominuspro]

@anyone who wants to poke around in the Ethereum CUDA miner source, i've implemented sp_'s mods and also cleaned it up a bit.

from what I checked, using uint2 was a bit slower (for once) than using uint64_t in keccak

No it isn't, but I think you have to seperate the loops into two and remove all the constant inputs, load from memory differenty, play with the pragma urolls, alter the launch bounds etc..

But keccak was just 13% of the total algorithm so the effort will probobly just gain a few percent.

I tried yesterday it was slower on my 980

I can confirm it beeing a little bit slower on 970 and 960.

[sp_]

yes, but it will be faster with a small rewrite.

Seperate the loops into two(keccakfirst,. keccak last) and remove all the constant work (less instructions),
With the #pragma unroll and __forceinline__ I hoped that the the compiler would remove the constant work but it doesn't understand that rol(constant,constant) etc can be precalculated.. Tune, compile and study the ptx assembly code. Less instructions is normally bether..

Load from memory differenty(remove the uint2 cast) use maxwells  vector instuctions, play with the pragma unrolls(but make sure that you don't exceed the instructioncache), alter the launch bounds to tune the registers etc..

[hashbrown9000]

--diff 2 never worked for me on nicehash. it always reported 1/5 of my total hash.  which diff should i be using now?

[sp_]

Before I started the sp-mod project 1 year ago, the new gtx 980 hashing slower than the AMD 290x (scrypt, x11,x13,x15 etc..) and almost nobody used NVIDIA cards for mining.

Here is a link to an old article:

Hashrate in september 2014 vs november 2014

In sp-mod release 60(opensource) from august 2015 quark is hashing close to 20MHASH on the gtx 980

http://cryptomining-blog.com/tag/gtx-980-mining-hashrate/



Quark opensource Opencl code is still doing 2MHASH on the 280x...

https://github.com/sgminer-dev/sgminer

compile and run....

[djm34]

--diff 2 never worked for me on nicehash. it always reported 1/5 of my total hash.  which diff should i be using now?

according to them, it should have worked since day one...
with the new release, you should remove the --diff 2 completely and it should work (please report if it doesn't)

[hashbrown9000]

well, i just compiled sp_'s version and nicehash is agreeing with my hashrate now.  I just want to know how to get the increased shares! 

[sp_]

--diff 2 never worked for me on nicehash. it always reported 1/5 of my total hash.  which diff should i be using now?

according to them, it should have worked since day one...
with the new release, you should remove the --diff 2 completely and it should work (please report if it doesn't)

I have instamined vertcoin with a few rigs and have almost reached 1000 VTC 2-3 days.

[pokeytex]

--diff 2 never worked for me on nicehash. it always reported 1/5 of my total hash.  which diff should i be using now?

according to them, it should have worked since day one...
with the new release, you should remove the --diff 2 completely and it should work (please report if it doesn't)

I have instamined vertcoin with a few rigs and have almost reached 1000 VTC 2-3 days.

Is selling VTC at 23k insane or is everyone just dumping daily?  Is there any reason to hold this coin?

[scryptr]

--diff 2 never worked for me on nicehash. it always reported 1/5 of my total hash.  which diff should i be using now?

according to them, it should have worked since day one...
with the new release, you should remove the --diff 2 completely and it should work (please report if it doesn't)

I have instamined vertcoin with a few rigs and have almost reached 1000 VTC 2-3 days.

Is selling VTC at 23k insane or is everyone just dumping daily?  Is there any reason to hold this coin?

VTC VALUE--

Vertcoin was at .002 BTC in early June.       --scryptr

[bensam1231]

--diff 2 never worked for me on nicehash. it always reported 1/5 of my total hash.  which diff should i be using now?

according to them, it should have worked since day one...
with the new release, you should remove the --diff 2 completely and it should work (please report if it doesn't)

I have instamined vertcoin with a few rigs and have almost reached 1000 VTC 2-3 days.

Is selling VTC at 23k insane or is everyone just dumping daily?  Is there any reason to hold this coin?

Up to you. Most important aspect of mining right there. Dump or hold. I've gotten burned a lot from holding, so I usually sell daily unless it's a long term coin and hold BTC until a nice spike in price.

That being said VTC has traditionally been a really solid coin so it wouldn't be unheard of for it to climb again. Sometimes you don't have the ability to hold because you don't make enough money too.

[sambiohazard]

--diff 2 never worked for me on nicehash. it always reported 1/5 of my total hash.  which diff should i be using now?

according to them, it should have worked since day one...
with the new release, you should remove the --diff 2 completely and it should work (please report if it doesn't)

I have instamined vertcoin with a few rigs and have almost reached 1000 VTC 2-3 days.

Is selling VTC at 23k insane or is everyone just dumping daily?  Is there any reason to hold this coin?

VTC VALUE--

Vertcoin was at .002 BTC in early June.       --scryptr

Dump some, keep some, donate some.

[dominuspro]

well, i just compiled sp_'s version and nicehash is agreeing with my hashrate now.  I just want to know how to get the increased shares! 

Go mining vtc on give me coins pool... You'll earn more.

[mamahdedeh]

hashing with nvidia always powerful. to the best of my choosing mining GPU Graphic card Nvida

[tbearhere]

Anybody have the 980ti implemented yet? I don't have mine on the new rig yet, but just wondering.
Thx

[tbearhere]

@djm34  ......Sent you some beers for your Lyra2v2.v4....or is it a 6 pack.      .05btc

TrxID: cdd1ad95ace6de79024b50863a51743c0cead3125c09cda59ea09802176ea90c

4.6 mh 750ti oc 1350.

I forgot to add .... 2x 750ti's on my vista doing quark algo drawing 212 watts at the wall... lyra2v2 v4 196 watts at the wall. So 8 watts less per 750ti doing lyra2v2 v4.
Down loading v5 now.

[ltc_bilic]

@_sp i'm not mining directly to you're adress atm, but will continue when the vtc/eth hype is over
at all devs, what you do for this community is mind blowing, so you definitely deserve a beer or two really appreciate it!

@ djm34:  .05 btc https://blockchain.info/en/tx/9a1522198e383355e5128a9c41eef8f4c3f40d35c61f234defa5ed81f3cc45bd
@ _sp:  .05 btc https://blockchain.info/en/tx/9a1522198e383355e5128a9c41eef8f4c3f40d35c61f234defa5ed81f3cc45bd
@ pallas:  .05 btc https://blockchain.info/en/tx/9a1522198e383355e5128a9c41eef8f4c3f40d35c61f234defa5ed81f3cc45bd

thank you guys again!

[djm34]

@_sp i'm not mining directly to you're adress atm, but will continue when the vtc/eth hype is over
at all devs, what you do for this community is mind blowing, so you definitely deserve a beer or two really appreciate it!

@ djm34:  .05 btc https://blockchain.info/en/tx/9a1522198e383355e5128a9c41eef8f4c3f40d35c61f234defa5ed81f3cc45bd
@ _sp:  .05 btc https://blockchain.info/en/tx/9a1522198e383355e5128a9c41eef8f4c3f40d35c61f234defa5ed81f3cc45bd
@ pallas:  .05 btc https://blockchain.info/en/tx/9a1522198e383355e5128a9c41eef8f4c3f40d35c61f234defa5ed81f3cc45bd

thank you guys again!

thanks a lot

[sambiohazard]

Today's share sent

djm34: 7f3ffa68047d03032fdfda9c4e0dacabec8abd943c7c7530f23d0ef4047386d5
pallas: 1a5a6cfafa6a4a52c8350895c406826a4004c8f9763a15776bb922c02f55fbe8
sp_: 03f50081ae4b1a55bf1af84460169254ab6d130a0a6785de464e2b197a737262

I gave myself a 80% donation

[Genoil]

yes, but it will be faster with a small rewrite.

Seperate the loops into two(keccakfirst,. keccak last) and remove all the constant work (less instructions),
With the #pragma unroll and __forceinline__ I hoped that the the compiler would remove the constant work but it doesn't understand that rol(constant,constant) etc can be precalculated.. Tune, compile and study the ptx assembly code. Less instructions is normally bether..

Load from memory differenty(remove the uint2 cast) use maxwells  vector instuctions, play with the pragma unrolls(but make sure that you don't exceed the instructioncache), alter the launch bounds to tune the registers etc..

It will probably be a little faster on its own, but I think the kernel is currently bottleneck by the memory lookups.

I tried making the whole keccak/dagger state uint2 based to get rid of casting, split out the keccak functions, manually unrolled some parts to get rid of branches, removed a bunch or XORs because they were unneeded..no improvement at all. As far as I could see there are no ROL2(const, const) here..

It's in the cudaminer-sp branch if you want to look.