CCminer(SP-MOD) Modded NVIDIA Maxwell / Pascal kernels. - page 1167.

rednoW

legendary

Activity: 1510

Merit: 1003

Quote from: sp_ on February 05, 2015, 01:42:13 PM

1.5.36(sp-MOD) is available here: (5-feb-2015)

...

Post your stats here. Card name/gpu clock/memclock

gtx 750 (1470gpu/1550mem)

x11 2847 khs
quark 5610 khs
lyra 853 khs
fresh 3435 khs
x13 don't work
x14 2200 khs
x15 1920 khs

note: some of my prev records was with more oc

rednoW

legendary

Activity: 1510

Merit: 1003

Quote from: sp_ on February 05, 2015, 01:42:13 PM

1.5.36(sp-MOD) is available here: (5-feb-2015)

fixed x11

you removed "h_found[thr_id][0] = 0xffffffff;" from x11.cu but added

Code:

foundNonce = 0xffffffff;
if (foundNonce != 0xffffffff)
...

to x13.cu
x13 don't work now as far as i see ...

go6ooo1212

legendary

Activity: 1512

Merit: 1000

quarkchain.io

Hi , mate , r36 still giving me only booos on qubit..

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

1.5.36(sp-MOD) is available here: (5-feb-2015)

fixed x11
fixed anime(klaust)
faster quark, x11 etc

https://github.com/sp-hash/ccminer/releases/tag/1.5.36

The sourcecode is available here:

https://github.com/sp-hash/ccminer

Post your stats here. Card name/gpu clock/memclock

rednoW

legendary

Activity: 1510

Merit: 1003

i love it )))

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Quote from: rednoW on February 05, 2015, 06:28:51 AM

Quote from: sp_ on February 05, 2015, 06:15:30 AM

try to remove this commit. (the MyStreamSynchronize calls)

https://github.com/sp-hash/ccminer/commit/40858b1e817c4ea40cf254fe292b3a3c7ca328f2

I think quark will be faster, but your computer will be slow an unusable if you use the graphic card for other work. The syncronization code was added between release 31 and 32

int the latest versions (release 33 and up) it is also possible to increase the cpu usage for more hash. This was feature was merged from the tpruvot/ccminer fork.

--cpu-affinity set process affinity to specific cpu core(s) mask
--cpu-priority set process priority (default: 0 idle, 2 normal to 5 highest)
try --cpu-priority 5

my display is attached to built-in graphics, so this is not a problem. I will give it a try
And i always run ccminer with realtime priority. After --cpu-priority appearing i use it too.
upd: strange ... i didn't find "MyStreamSynchronize(NULL, 4, thr_id)" on line 237 or nearby in the quarkcoin.cu.
Lines
"MyStreamSynchronize(NULL, 1, thr_id)"
"MyStreamSynchronize(NULL, 2, thr_id)"
"MyStreamSynchronize(NULL, 3, thr_id)"
was found and commented.
Now building ...

yes I have moved them around, there is one more further down.. "MyStreamSynchronize(NULL, 4, thr_id)" comment out this one as well

rednoW

legendary

Activity: 1510

Merit: 1003

Quote from: sp_ on February 05, 2015, 06:15:30 AM

try to remove this commit. (the MyStreamSynchronize calls)

https://github.com/sp-hash/ccminer/commit/40858b1e817c4ea40cf254fe292b3a3c7ca328f2

I think quark will be faster, but your computer will be slow an unusable if you use the graphic card for other work. The syncronization code was added between release 31 and 32

int the latest versions (release 33 and up) it is also possible to increase the cpu usage for more hash. This was feature was merged from the tpruvot/ccminer fork.

--cpu-affinity set process affinity to specific cpu core(s) mask
--cpu-priority set process priority (default: 0 idle, 2 normal to 5 highest)

try --cpu-priority 5

my display is attached to built-in graphics, so this is not a problem. I will give it a try
And i always run ccminer with realtime priority. After --cpu-priority appearing i use it too.

upd: strange ... i didn't find "MyStreamSynchronize(NULL, 4, thr_id)" on line 237 or nearby in the quarkcoin.cu.

Lines
"MyStreamSynchronize(NULL, 1, thr_id)"
"MyStreamSynchronize(NULL, 2, thr_id)"
"MyStreamSynchronize(NULL, 3, thr_id)"
was found and commented.
Now building ...

upd2: you are a wizard Wink

A new quark record for my poor GTX750 - 5650KH, better then R31 Wink

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

try to remove this commit. (the MyStreamSynchronize calls)

https://github.com/sp-hash/ccminer/commit/40858b1e817c4ea40cf254fe292b3a3c7ca328f2

I think quark will be faster, but your computer will be slow an unusable if you use the graphic card for other work. The syncronization code was added between release 31 and 32

int the latest versions (release 33 and up) it is also possible to increase the cpu usage for more hash. This was feature was merged from the tpruvot/ccminer fork.

--cpu-affinity set process affinity to specific cpu core(s) mask
--cpu-priority set process priority (default: 0 idle, 2 normal to 5 highest)

try --cpu-priority 5

rednoW

legendary

Activity: 1510

Merit: 1003

Quote from: sp_ on February 05, 2015, 06:07:29 AM

This code will force the kernal to use 64 registers. bether on the 750ti worse on the 970/980.
What I usually do is to case on the compute version and run different configurations for different compute version.
I think many of the kernals haven't been tweaked for a while, and hash can be gained.
If you see my Bitcoin change, all I did was to change the kernal launch configutration for a 17% speedup.

here is the commit:

https://github.com/sp-hash/ccminer/commit/c79f622969393f52f3462e2c3e967777cae1d7d3

thanks for explanation
R35 after this few mods is best for me, maybe except for quark (r31 is 20-30 kh better), but it's a matter of measurement error Wink

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Quote from: rednoW on February 05, 2015, 05:51:23 AM

Quote from: sp_ on February 05, 2015, 04:40:20 AM

With vardiff I get around 25khash bether in 35 than in 33/34 , but I lost around 25 KHASH after merging the fork with klaus_t (from 32 -> 33).
Will take another look tonight.
But it could be this commit:
https://github.com/sp-hash/ccminer/commit/704e609b9003620dd29a74b2f43924209e193d24
Might work good on the 960-980 cards, but not on the 750ti.
Not many changes though.

Yes, i put back those "__global__ __launch_bounds__(256, 4)" etc and got some near measurement error gain of ~20kh

This code will force the kernal to use 64 registers. bether on the 750ti worse on the 970/980.
What I usually do is to case on the compute version and run different configurations for different compute version.
I think many of the kernals haven't been tweaked for a while, and hash can be gained.
If you see my Bitcoin change, all I did was to change the kernal launch configutration for a 17% speedup.

here is the commit:

https://github.com/sp-hash/ccminer/commit/c79f622969393f52f3462e2c3e967777cae1d7d3

rednoW

legendary

Activity: 1510

Merit: 1003

Quote from: sp_ on February 05, 2015, 04:40:20 AM

With vardiff I get around 25khash bether in 35 than in 33/34 , but I lost around 25 KHASH after merging the fork with klaus_t (from 32 -> 33).

Will take another look tonight.

But it could be this commit:

https://github.com/sp-hash/ccminer/commit/704e609b9003620dd29a74b2f43924209e193d24

Might work good on the 960-980 cards, but not on the 750ti.

Not many changes though.

Yes, i put back those "__global__ __launch_bounds__(256, 4)" etc and got some near measurement error gain of ~20kh

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Quote from: rednoW on February 05, 2015, 04:04:32 AM

upd: now it works. ~2850 kh (the same speed as in r31 & r32)

With vardiff I get around 25khash bether in 35 than in 33/34 , but I lost around 25 KHASH after merging the fork with klaus_t (from 32 -> 33).

Will take another look tonight.

But it could be this commit:

https://github.com/sp-hash/ccminer/commit/704e609b9003620dd29a74b2f43924209e193d24

Might work good on the 960-980 cards, but not on the 750ti.

Not many changes though.

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Quote from: rednoW on February 05, 2015, 04:04:32 AM

Quote from: sp_ on February 05, 2015, 04:00:36 AM

Ahh, I found the problem now
remove the line
h_found[thr_id][0] = 0xffffffff;
in x11.cu

yeah, the next line "if (h_found[thr_id][0] != 0xffffffff)" is slightly not compatible with it Wink

I added the line in order to speed test the simd kernal alone, but forgot to remove it. :/ I am away from my dev computer for a while, so will fix it on github later tonight.

rednoW

legendary

Activity: 1510

Merit: 1003

Quote from: sp_ on February 05, 2015, 04:00:36 AM

Ahh, I found the problem now

remove the line

h_found[thr_id][0] = 0xffffffff;

in x11.cu

yeah, the next line "if (h_found[thr_id][0] != 0xffffffff)" is slightly not compatible with it Wink

upd: now it works. ~2850 kh (the same speed as in r31 & r32)

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Quote from: rednoW on February 05, 2015, 03:55:55 AM

Quote from: sp_ on February 05, 2015, 03:41:57 AM

Seems to work for others. (yaamp.com)

Wich algos are broken? Did you try fresh or quark? Might be an issue with cuda 7.0 (wich is only released in beta version)

only x11 is broken. All others work fine.
I compiled with cuda 6.5 preset (as default). Cuda 7 installed is not an issue.
You should recheck a code. I didn't see any working Release 35 clients on yaamp's x11 stats ...

Ahh, I found the problem now

remove the line

h_found[thr_id][0] = 0xffffffff;

in x11.cu

rednoW

legendary

Activity: 1510

Merit: 1003

Quote from: sp_ on February 05, 2015, 03:41:57 AM

Seems to work for others. (yaamp.com)

Wich algos are broken? Did you try fresh or quark? Might be an issue with cuda 7.0 (wich is only released in beta version)

only x11 is broken. All others work fine.
I compiled with cuda 6.5 preset (as default). Cuda 7 installed is not an issue.
You should recheck a code. I didn't see any working Release 35 clients on yaamp's x11 stats ...

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Seems to work for others. (yaamp.com)

Wich algos are broken? Did you try fresh or quark? Might be an issue with cuda 7.0 (wich is only released in beta version)

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

You need cuda 6.5

rednoW

legendary

Activity: 1510

Merit: 1003

Quote from: sp_ on February 05, 2015, 01:53:12 AM

Strange, it works here.

Make sure you have the latest driver.

I might have copied the wrong exe file though. If you build it from github it should work.

just recompiled. The same. No accepts. No boos. Driver is from CUDA 7.0 pack as always 347.12
Previous releases run x11 normally

sp_

legendary

Activity: 2954

Merit: 1087

Team Black developer

Strange, it works here.

Make sure you have the latest driver.

I might have copied the wrong exe file though. If you build it from github it should work.

Topic: CCminer(SP-MOD) Modded NVIDIA Maxwell / Pascal kernels. - page 1167. (Read 2347641 times)