CCminer(SP-MOD) Modded NVIDIA Maxwell / Pascal kernels. - page 1213.

sp_

legendary

Activity: 2926

Merit: 1087

Team Black developer

Quote from: Epsylon3 on December 04, 2014, 02:41:30 AM

Quote from: djm34 on December 03, 2014, 07:37:10 AM

Quote from: sp_ on December 03, 2014, 06:49:13 AM

Your makefile use 128 registers and mine has 80 registers as default, and this will cause big differences.

actually, this should be removed from any makefile or setup (even though I didn't do it myself...), this method of allocating register is deprecated and should be replaced by __launch_bounds__ (I am paraphrasing cuda doc...)

depends of the cases, sometimes you cant change the TPB without rewriting the code (i think about sharedmem code) and you cant fine tune the max regs with the launchbound.
That is right to support all past and possible/future cards.. not for the 90% of users with 750Ti or 970/980

The problem is that when you change the number of registers in one kernal, all the other kernals needs to be recompiled, and sometimes the performance in the other kernals get worse. Use launchbounds. It's faster and bether.

sp_

legendary

Activity: 2926

Merit: 1087

Team Black developer

Quote from: Epsylon3 on December 04, 2014, 03:44:56 AM

else your echo improvement works on linux too... 2870 yesterday 2920 now we are close to the 3MH on the 750Ti ^^ I need to inspect the changes
you have also remains of a missing simd free with current commit (does not build)

The new Echo is doing 8.xx rounds of echo instead of 10 rounds. The previous version did 9.25 rounds. And the original version does 10rounds.

On the 980 we will probobly get 300KHASH more.

Less work, less power more hash.

sp_

legendary

Activity: 2926

Merit: 1087

Team Black developer

Quote from: Epsylon3 on December 04, 2014, 03:44:56 AM

else your echo improvement works on linux too... 2870 yesterday 2920 now we are close to the 3MH on the 750Ti ^^ I need to inspect the changes

2.92MHASH on the 750TI windforce black am I right? I only get 2700-2750 on my gainward ti.

sp_

legendary

Activity: 2926

Merit: 1087

Team Black developer

Not done, wait for a new checkin tonight. Multipools show high numbers , but Solomining is broken.

Epsylon3

legendary

Activity: 1484

Merit: 1082

ccminer/cpuminer developer

else your echo improvement works on linux too... 2870 yesterday 2920 now we are close to the 3MH on the 750Ti ^^ I need to inspect the changes

you have also remains of a missing simd free with current commit (does not build)

sp_

legendary

Activity: 2926

Merit: 1087

Team Black developer

Quote from: sp_ on December 03, 2014, 07:06:50 PM

Will not make a new build before I have run it through the night.

The version I checked in last night on github needs some more work...

Epsylon3

legendary

Activity: 1484

Merit: 1082

ccminer/cpuminer developer

Quote from: djm34 on December 03, 2014, 07:37:10 AM

Quote from: sp_ on December 03, 2014, 06:49:13 AM

Your makefile use 128 registers and mine has 80 registers as default, and this will cause big differences.

actually, this should be removed from any makefile or setup (even though I didn't do it myself...), this method of allocating register is deprecated and should be replaced by __launch_bounds__ (I am paraphrasing cuda doc...)

depends of the cases, sometimes you cant change the TPB without rewriting the code (i think about sharedmem code) and you cant fine tune the max regs with the launchbound.

That is right to support all past and possible/future cards.. not for the 90% of users with 750Ti or 970/980

grendel25

legendary

Activity: 2296

Merit: 1031

Quote from: Epsylon3 on December 04, 2014, 12:01:55 AM

--help or README.txt at github (or in my releases)

thx!

Epsylon3

legendary

Activity: 1484

Merit: 1082

ccminer/cpuminer developer

--help or README.txt at github (or in my releases)

grendel25

legendary

Activity: 2296

Merit: 1031

Quote from: sp_ on December 01, 2014, 01:21:46 AM

Quote from: grendel25 on December 01, 2014, 12:47:21 AM

could folks share there command line? this is mine:

Code:

ccminer.exe -q -r 3 -R 10 -a x13 --no-color -o stratum+tcp://yaamp.com:3633 -u xxx -p xxx

Thank you. Is there a 'read me' that explains the -flags? Not familiar with -q -r or -R. seems like the -q maybe 'quiets' some of the output?

sp_

legendary

Activity: 2926

Merit: 1087

Team Black developer

Will not make a new build before I have run it through the night.

sp_

legendary

Activity: 2926

Merit: 1087

Team Black developer

Today I improved the final hashing in x11 (echo). The 1mb 750 is +50 KHASH. the ti 50-100KHASH. reverted BMW to an earlier version.
My gainward 750ti is peaking at 2.750MHASH. While it was at 2650-2700 earlier.

sp_

legendary

Activity: 2926

Merit: 1087

Team Black developer

Quote from: djm34 on December 03, 2014, 07:37:10 AM

Quote from: sp_ on December 03, 2014, 06:49:13 AM

Your makefile use 128 registers and mine has 80 registers as default, and this will cause big differences.

actually, this should be removed from any makefile or setup (even though I didn't do it myself...), this method of allocating register is deprecated and should be replaced by __launch_bounds__ (I am paraphrasing cuda doc...)

Yes, most of my modded kernals use __launchbounds__.

sp_

legendary

Activity: 2926

Merit: 1087

Team Black developer

Quote from: Epsylon3 on December 03, 2014, 06:43:00 AM

Quote from: sp_ on December 01, 2014, 04:47:56 AM

I started on the merge of tpruvot 1.5.1, but there was so many files and I got lazy. I tested the latest version on compute 3.0 (gtx 650) and it crashed. I will look into it.

When you see commits changing miner.h and structures, its important to do a clean build (make clean on linux or ... Regenerate in VStudio)
I also tried to merge some of your changes, but... for the moment i had weird results on linux.
Shavite seems to be suboptimal on mine... will check that later (Conference today, the whole afternoon)

I think the best would be to refork, and just change some of the kernal code.
I will probobly revert BMW to an earlier version. It seems to go faster on 750ti without the new changes even if it spills more registers and the code is longer. But messuring is abit hard with the boost clock setting. If you push the cards to hard/get hot, they will downclock and loose performance. Different ti cards have different perfomance/voltagate/wattage.

djm34

legendary

Activity: 1400

Merit: 1050

Quote from: sp_ on December 03, 2014, 06:49:13 AM

Your makefile use 128 registers and mine has 80 registers as default, and this will cause big differences.

actually, this should be removed from any makefile or setup (even though I didn't do it myself...), this method of allocating register is deprecated and should be replaced by __launch_bounds__ (I am paraphrasing cuda doc...)

sp_

legendary

Activity: 2926

Merit: 1087

Team Black developer

Your makefile use 128 registers and mine has 80 registers as default, and this will cause big differences.

Epsylon3

legendary

Activity: 1484

Merit: 1082

ccminer/cpuminer developer

Quote from: sp_ on December 01, 2014, 04:47:56 AM

I started on the merge of tpruvot 1.5.1, but there was so many files and I got lazy. I tested the latest version on compute 3.0 (gtx 650) and it crashed. I will look into it.

When you see commits changing miner.h and structures, its important to do a clean build (make clean on linux or ... Regenerate in VStudio)

I also tried to merge some of your changes, but... for the moment i had weird results on linux.

Shavite seems to be suboptimal on mine... will check that later (Conference today, the whole afternoon)

bathrobehero

legendary

Activity: 2002

Merit: 1051

ICO? Not even once.

Quote from: SS2006 on December 02, 2014, 08:03:19 PM

SP anyway to make the miner/cmd window minimize to the tray
that would be amazing

*answered myself!
http://rbtray.sourceforge.net/

You can use the start command as in:
start /min ccminer.exe .....

You can also set the priority that way (/low /normal /high, etc).

SS2006

sr. member

Activity: 285

Merit: 250

SP anyway to make the miner/cmd window minimize to the tray
that would be amazing

*answered myself!
http://rbtray.sourceforge.net/

Handofmidas

member

Activity: 61

Merit: 10

Here's my first impression of ccminer 1.5.0. The miner shows 15.2-15.3 mh + occasional hw. Pool shows this:

Left: release14
from 12:00: release8
from 20:30: ccminer 1.5.0

So apparently some of the hash magically disappears.

Topic: CCminer(SP-MOD) Modded NVIDIA Maxwell / Pascal kernels. - page 1213. (Read 2347597 times)