CCminer(SP-MOD) Modded NVIDIA Maxwell / Pascal kernels. - page 827.

djm34

legendary

Activity: 1400

Merit: 1050

Quote from: bensam1231 on December 29, 2015, 04:04:09 AM

Quote from: djm34 on December 28, 2015, 09:12:20 PM

Quote from: bensam1231 on December 28, 2015, 02:40:58 PM

Quote from: djm34 on December 27, 2015, 12:36:29 PM

Quote from: bensam1231 on December 27, 2015, 08:40:19 AM

No one has released anything new in close to six months or so it seems. Everything that has been talked about here is old hat or the same epeen wars. Where you talk about how fast your super optimized miner is you only sell to a handful of really rich individuals.

Biggest improvement has come from Nicehash returning dividends with their paid bounty program, but that only applies to AMD.

I joined the Nvidia mining community because it was more open and forthright then AMD mining. That seems to have died out and turned back into the same shit as the AMD community. Nothing improves or changes. I still end up going back and mining Neoscrypt with ancient miners (that are definitely outdated). Donating doesn't do anything.

How would you know ? you never donated for neoscrypt, you just started mining it after I decided to release the source publicly.
Such an hypocrite Grin

You were paid by the feathercoin community. Why would I pay you for something that is already been paid for? You worked for commission. You're not in the same shoes as SP.

I never been paid by the feathercoin community not sure why you are always making things up, I opened an ANN for a crowdfunding campaign...
, but your excuse proves, I was right not dealing with you...

Yuh and who do you think is funding you? People NOT mining with Neoscrypt? What do you mine with Neoscrypt, oh yeah Feathercoin. Just shows you how inept you are at anything other then toodling around with kernel development... and that definitely includes economics, coin development, and, of course, business.

Sure you can drudge up a few other Neoscrypt coins with almost 0 volume, community, and buy support like UFOcoin that multipools hammer all the time, but as we both know (or maybe not) those don't matter.

each time I discuss with you, I have the feeling I am discussing with a complete moron, well so you won't be surprised to learn that my opinion didn't change with that message. Grin

I meant paid by the ftc devs (they didn't contribute in any way to the crowdfunding... ) only people wanting to mine ftc did pay for it
and apparently even 0.02btc (around 5euros at that time) was too much for you. Hence you are probably the most irrelevant person on that whole thread when it comes to business, donation and else.
Also since you mined ftc for quite some time that makes you a member of that community, hence as part of that community you didn't pay it either...

bensam1231

legendary

Activity: 1764

Merit: 1024

Quote from: djm34 on December 28, 2015, 09:12:20 PM

Quote from: bensam1231 on December 28, 2015, 02:40:58 PM

Quote from: djm34 on December 27, 2015, 12:36:29 PM

Quote from: bensam1231 on December 27, 2015, 08:40:19 AM

No one has released anything new in close to six months or so it seems. Everything that has been talked about here is old hat or the same epeen wars. Where you talk about how fast your super optimized miner is you only sell to a handful of really rich individuals.

Biggest improvement has come from Nicehash returning dividends with their paid bounty program, but that only applies to AMD.

I joined the Nvidia mining community because it was more open and forthright then AMD mining. That seems to have died out and turned back into the same shit as the AMD community. Nothing improves or changes. I still end up going back and mining Neoscrypt with ancient miners (that are definitely outdated). Donating doesn't do anything.

How would you know ? you never donated for neoscrypt, you just started mining it after I decided to release the source publicly.
Such an hypocrite Grin

You were paid by the feathercoin community. Why would I pay you for something that is already been paid for? You worked for commission. You're not in the same shoes as SP.

I never been paid by the feathercoin community not sure why you are always making things up, I opened an ANN for a crowdfunding campaign...
, but your excuse proves, I was right not dealing with you...

Yuh and who do you think is funding you? People NOT mining with Neoscrypt? What do you mine with Neoscrypt, oh yeah Feathercoin. Just shows you how inept you are at anything other then toodling around with kernel development... and that definitely includes economics, coin development, and, of course, business.

Sure you can drudge up a few other Neoscrypt coins with almost 0 volume, community, and buy support like UFOcoin that multipools hammer all the time, but as we both know (or maybe not) those don't matter.

djm34

legendary

Activity: 1400

Merit: 1050

Quote from: sp_ on December 28, 2015, 06:22:55 PM

The compiler engineers at NVIDIA is probobly using EESL. Altcoin miners make their own compilers from scratach.

lolwut ? (are the same guy as Bensam ) ? Grin

djm34

legendary

Activity: 1400

Merit: 1050

Quote from: bensam1231 on December 28, 2015, 02:40:58 PM

Quote from: djm34 on December 27, 2015, 12:36:29 PM

Quote from: bensam1231 on December 27, 2015, 08:40:19 AM

No one has released anything new in close to six months or so it seems. Everything that has been talked about here is old hat or the same epeen wars. Where you talk about how fast your super optimized miner is you only sell to a handful of really rich individuals.

Biggest improvement has come from Nicehash returning dividends with their paid bounty program, but that only applies to AMD.

I joined the Nvidia mining community because it was more open and forthright then AMD mining. That seems to have died out and turned back into the same shit as the AMD community. Nothing improves or changes. I still end up going back and mining Neoscrypt with ancient miners (that are definitely outdated). Donating doesn't do anything.

How would you know ? you never donated for neoscrypt, you just started mining it after I decided to release the source publicly.
Such an hypocrite Grin

You were paid by the feathercoin community. Why would I pay you for something that is already been paid for? You worked for commission. You're not in the same shoes as SP.

I never been paid by the feathercoin community not sure why you are always making things up, I opened an ANN for a crowdfunding campaign...
, but your excuse proves, I was right not dealing with you...

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

The compiler engineers at NVIDIA is probobly using EESL. Altcoin miners make their own compilers from scratach.

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Quote from: ?? on ??

Quote from: sp_ on December 28, 2015, 10:12:43 AM

If you compile for cuda 6.5 you should use the release74 version for all sourcecode files. X11 is also faster

Apparently not:
Yours: https://ottrbutt.com/miner/spx11stock7.5-12282015.png
Mine: https://ottrbutt.com/miner/wolfx11ccminer7.5-12282015.png
(both nsfw)
Mine received a bump of over 100kh/s from 7.5, yours seemed to drop a bit. As for register use, I can see why - 128 regs on yours, 24b spill stores, 20b spill loads. Mine uses 54 regs with 0b for spills/loads.

Nice findings. The compiler is different in x86 (32bit) than linux 64bit.

bensam1231

legendary

Activity: 1764

Merit: 1024

Quote from: djm34 on December 27, 2015, 12:36:29 PM

Quote from: bensam1231 on December 27, 2015, 08:40:19 AM

No one has released anything new in close to six months or so it seems. Everything that has been talked about here is old hat or the same epeen wars. Where you talk about how fast your super optimized miner is you only sell to a handful of really rich individuals.

Biggest improvement has come from Nicehash returning dividends with their paid bounty program, but that only applies to AMD.

I joined the Nvidia mining community because it was more open and forthright then AMD mining. That seems to have died out and turned back into the same shit as the AMD community. Nothing improves or changes. I still end up going back and mining Neoscrypt with ancient miners (that are definitely outdated). Donating doesn't do anything.

How would you know ? you never donated for neoscrypt, you just started mining it after I decided to release the source publicly.
Such an hypocrite Grin

You were paid by the feathercoin community. Why would I pay you for something that is already been paid for? You worked for commission. You're not in the same shoes as SP.

Conversely, I can't pay for a commission when you never offer your services besides immature and unprofessional money grabs you decide to do on a whim. Notice how when Nicehash came in here they offered real concrete bounties for set goals and they were willing to work on all of the above (and still are)? Notice how you don't do that any expect people to just throw money at you when you aren't even offering anything?

Quote from: hashbrown9000 on December 27, 2015, 02:49:26 PM

Quote

End of the game for neoscrypt as well:

https://bitcointalksearch.org/topic/m.13373431

even at 10 mBTC/GH/day neoscrypt was just barely profitable for maxwell. and it hasn't seen that rate in a while

And I just vomited a bit. Hopefully they have some Nvidia devs working for them. Good for Nicehash, shit for Nvidia miners till they get other developers working for them.

Quote from: scryptr on December 27, 2015, 05:57:01 PM

\NEOSCRYPT--

I haven't switched to dot 77+ yet because of Neoscrypt. I haven't mined any Blake algo coins either, but apparently BlakeCoin was mineable with CudaMiner back in 2013-2014.

My GTX 960 gets 340kh/s mining Neoscrypt with the latest NiceHash Miner, that makes my GTX 960 as fast as an AMD 280X with the latest NiceHash Miner. They are still using your release dot 74. If you did a hybrid compile (Pallas), NiceHash may incorporate your release dot 77+. Neoscrypt is still a top-paying algo at NiceHash, and AMD code is really being optimized there.

I may do some benchmark tests. --scryptr

Same, Neoscrypt is the goto right now. Quark and Lyra are both in the shitter at the moment and have been for a few months.

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Quote from: ?? on ??

Quote from: sp_ on December 28, 2015, 10:44:37 AM

7.5 not 7.

I was using the major version number. According to my package manager, it's 7.5.18 Tongue

Good. I will take a look at echo later. I see some code that can be removed

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

7.5 not 7.

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

If you compile for cuda 6.5 you should use the release74 version for all sourcecode files. X11 is also faster

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Release 74 (cuda 6.5) is using 64 regs. But the cuda 7.5 compiler is trashing the code

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Quote from: ?? on ??

No, no, I mean register pressure with the state itself all in regs. You can't really get good occupancy this way, right?

You need to remove the conditional code. the pseudo code is just an example.

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

No. Here is an example of a 16bit precalc approach:

Code:

_device__
static void aes_round(
const uint32_t *sharedMemory,
uint32_t x0, uint32_t x1, uint32_t x2, uint32_t x3, uint32_t k0,
uint32_t &y0, uint32_t &y1, uint32_t &y2, uint32_t &y3)
{
y0 =sharedMemory[__byte_perm(x0, x1, 0x4441)&0xffff]^
sharedMemory[(__byte_perm(x2, x3, 0x4452)&0xffff) + 256];
y1 =
...

Your table will need to be 2^16 of 32bit integers big(256KB). The precalculated 32bit value is 2 numbers from the old table xor'ed together.

now the difficult part is to reduce this to eigher:

12 bits (16KB table)
13 bits (32Kb table)
14 bits (64Kb table)

the slow approach would be something like:

Code:

if((x0&0xff)<63 && ((x2>>8)&0xff)<63)
{
   //14bit aes
y0 =sharedMemory64[__byte_perm(x0, x1, 0x4441)&0xffff]^
sharedMemory64[(__byte_perm(x2, x3, 0x4441)&0xffff) + 256];
   y1 = ...
} else
{
  //do 8bit aes
y0 = xor4_32(
sharedMemory[x0&0xff],
sharedMemory[__byte_perm(x1, 0, 0x4441) + 256],
sharedMemory[__byte_perm(x2, 0, 0x4442) + 512],
sharedMemory[__byte_perm(x3, 0, 0x4443) + 768]);

   ...
}

With this code you have a 75% probabillity of executing the fast method. (14bit lookups in shared mem 1cycle)

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

But I have another plan. Reduce the table accesses in the AES by 50% by doing lookups on 12 bits instead of 8bits.
You need 16KB of shared mem instead of 1kb. but still under the 96kb limit on the compute5.2 cards and 64kb limit on the compute 5.0
Then you could do AES with 50% less instructions and 50% less shared mem access..

But you need to rewrite/redesign the algorithm. Alot of work.

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Quote from: ?? on ??

Quote from: sp_ on December 28, 2015, 07:09:54 AM

nice. Want to share the code? Wink

I don't give away answers, you know that. But hints or exchanging ideas I'm always happy to do. There may be a better way to improve Echo than what I've done, actually, however - it will lend itself to Nvidia's architecture more - but at the cost of a slightly more complex round. Best to code for your platform, though... I might give it a go..

You should pickup a couple of cheap and used 750ti's on ebay. I have seen them below $100.

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Quote from: ?? on ??

Quote from: sp_ on December 28, 2015, 07:09:54 AM

nice. Want to share the code? Wink

I don't give away answers, you know that. But hints or exchanging ideas I'm always happy to do. There may be a better way to improve Echo than what I've done, actually, however - it will lend itself to Nvidia's architecture more - but at the cost of a slightly more complex round. Best to code for your platform, though... I might give it a go..

The bottleneck in the echo is the table based AES. But it might help to split the work in the kernal into two passes to improve the register count.
Compiling with cuda 7.5 causes a register and stack explosion.

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer