Pages:
Author

Topic: ATTN Litecoin GPU Miners - Scrypt support for cgminer - page 11. (Read 175855 times)

sr. member
Activity: 266
Merit: 250
Thanks for your work on this Conman. =3 And thanks for the people that donated to him.

Here's where I am right now. It will mine on the Nvidia fine it seems but the ATI GPU isn't working correctly.
I'm using the 12.4 catalyst with the 2.7 SDK for the ATI GPU. I could not find a working DL link for the 2.6 SDK.



This is what "cgminer -n" says.



I've got one monitor hooked up to the first GPU (Nvidia). I've tried the gpu-map option and cannot get it to work correctly. I suspect windows thinks both adapters are hooked up to the monitor. >.>
-ck
legendary
Activity: 4088
Merit: 1631
Ruu \o/
Image:

only get ~20 kh/s

Code:
cgminer --scrypt -I 16 --shaders 800
You'll have to fine tune it with thread concurrency, shaders alone still doesn't cut it. I haven't found the optimum autotune as of yet, and I'm not sure there will be one.
full member
Activity: 204
Merit: 100
I need some help to compile CGMiner with Scrypt under Ubuntu 12.04 64 bits, 1 7970, Catalyst 12.6, AMDAPP v2.6...

What is the procedure?

Basically: git clone, checkout newscrypt, autogen, configure --enable-scrypt ... ?

Thanks!
Thiago

adl files in the adl-sdk folder

./autogen.sh
CFLAGS="-O2 -Wall -march=native -I/opt/AMDAPP/include" LDFLAGS="L/opt/AMDAPP/lib/x86_64" ./configure --enable-scrypt
make
full member
Activity: 204
Merit: 100
member
Activity: 89
Merit: 10
Image:

only get ~20 kh/s

Code:
cgminer --scrypt -I 16 --shaders 800
newbie
Activity: 40
Merit: 0
...

The good news is that this should be what the 5x/6x miners were looking for on linux. I suggest you download the latest changes from git and start experimenting with TCs that are a multiple of your shader count after setting the above variable to 100%.

Wow - I am leaving work and was planning to try the parameter at home - and now I see that you already did my homework :-)


many thanks !!!
To say the lack of progress on scrypt performance is bugging me would be an understatement in the extreme. Thank you for spotting this, but I fear even this is not the holy grail. Goddamn witchcraft....


I am excited !
git pull - make - start with the export and -D -T

 [2012-07-25 17:28:35] CL Platform vendor: Advanced Micro Devices, Inc.                   
 [2012-07-25 17:28:35] CL Platform name: AMD Accelerated Parallel Processing                   
 [2012-07-25 17:28:35] CL Platform version: OpenCL 1.2 AMD-APP (938.1)                   
 [2012-07-25 17:28:35] List of devices:                   
 [2012-07-25 17:28:35]  0       Cypress                   
 [2012-07-25 17:28:35] Selected 0: Cypress                   
 [2012-07-25 17:28:35] Long-polling activated for http://p2p:9327/long-polling                   
 [2012-07-25 17:28:35] Preferred vector width reported 4                   
 [2012-07-25 17:28:35] Max work group size reported 256                   
 [2012-07-25 17:28:35] Max mem alloc size is 536870912                   
 [2012-07-25 17:28:35] Selecting scrypt kernel                   
 [2012-07-25 17:28:35] GPU 0: selecting lookup gap of 2                   
 [2012-07-25 17:28:35] GPU 0: selecting thread concurrency of 7200                   
 [2012-07-25 17:28:35] No binary found, generating from source                   
 [2012-07-25 17:28:35] Setting worksize to 256                   
 [2012-07-25 17:28:35] cl_amd_media_ops found, setting BITALIGN                   
 [2012-07-25 17:28:35] BFI_INT patch requiring device found, patched source with BFI_INT                   
 [2012-07-25 17:28:35] CompilerOptions: -D LOOKUP_GAP=2 -D CONCURRENT_THREADS=7200 -D WORKSIZE=256 -D BITALIGN -D BFI_INT -D GOFFSET                   
 [2012-07-25 17:28:38] X-Roll-Ntime expiry set to 10

cgminer crashed but restarting with -g 1 lets it hash away with about 260 kH/s :-)


ATM it runs with
DISPLAY=:0 GPU_MAX_ALLOC_PERCENT=100 ../cgminerltc/cgminer -o http://p2p:9327 -u x -p y --scrypt --shaders 1440 -I 15 -g 1 --auto-fan --auto-gpu --gpu-engine 600-900 --gpu-memclock 1000 --temp-cutoff 85 --gpu-fan 20-70 --temp-overheat 75 --temp-target 65

-I 15 seems to be the sweet spot, only 3 kH/s less than -I 16  but 0.x% rejects instead of >10%


Thanks again !   ... and I may later try lower tc with -g 2 and expermient more :-)
-ck
legendary
Activity: 4088
Merit: 1631
Ruu \o/
Okay I've experimented some more and
export GPU_MAX_ALLOC_PERCENT=100
will obviously cause a problem, but it gives me massive scope for trying different tc, etc since I can theoretically allocate 3GB on a 7970. However the code would try to allocate the largest buffer by default which is not right with this value now, so I've committed a change (to git) that will find the nearest buffer size that is a power of 2 between the thread concurrency buffer requirement and the maximum allocable ram.

Playing some more with my 7970s I found the values it was already running were pretty close to optimal. The only thing I found useful with the increased ram available was that I am now able to increase threads a little more (5 instead of 4) which afforded another 5kh per device.

The good news is that this should be what the 5x/6x miners were looking for on linux. I suggest you download the latest changes from git and start experimenting with TCs that are a multiple of your shader count after setting the above variable to 100%.

Wow - I am leaving work and was planning to try the parameter at home - and now I see that you already did my homework :-)


many thanks !!!
To say the lack of progress on scrypt performance is bugging me would be an understatement in the extreme. Thank you for spotting this, but I fear even this is not the holy grail. Goddamn witchcraft....
newbie
Activity: 40
Merit: 0
Okay I've experimented some more and
export GPU_MAX_ALLOC_PERCENT=100
will obviously cause a problem, but it gives me massive scope for trying different tc, etc since I can theoretically allocate 3GB on a 7970. However the code would try to allocate the largest buffer by default which is not right with this value now, so I've committed a change (to git) that will find the nearest buffer size that is a power of 2 between the thread concurrency buffer requirement and the maximum allocable ram.

Playing some more with my 7970s I found the values it was already running were pretty close to optimal. The only thing I found useful with the increased ram available was that I am now able to increase threads a little more (5 instead of 4) which afforded another 5kh per device.

The good news is that this should be what the 5x/6x miners were looking for on linux. I suggest you download the latest changes from git and start experimenting with TCs that are a multiple of your shader count after setting the above variable to 100%.

Wow - I am leaving work and was planning to try the parameter at home - and now I see that you already did my homework :-)


many thanks !!!
hero member
Activity: 686
Merit: 500
-ck
legendary
Activity: 4088
Merit: 1631
Ruu \o/
Build away.

I don't have any other immediate change planned. I'm not sure this will ever be a simple start it and it will find optimum settings. Scrypt on GPU is half programming, half witchcraft.
hero member
Activity: 686
Merit: 500
Okay I've experimented some more and
export GPU_MAX_ALLOC_PERCENT=100
will obviously cause a problem, but it gives me massive scope for trying different tc, etc since I can theoretically allocate 3GB on a 7970. However the code would try to allocate the largest buffer by default which is not right with this value now, so I've committed a change (to git) that will find the nearest buffer size that is a power of 2 between the thread concurrency buffer requirement and the maximum allocable ram.

Playing some more with my 7970s I found the values it was already running were pretty close to optimal. The only thing I found useful with the increased ram available was that I am now able to increase threads a little more (5 instead of 4) which afforded another 5kh per device.

The good news is that this should be what the 5x/6x miners were looking for on linux. I suggest you download the latest changes from git and start experimenting with TCs that are a multiple of your shader count after setting the above variable to 100%.

Should I rebuild it now or are you going to commit more right now?
-ck
legendary
Activity: 4088
Merit: 1631
Ruu \o/
Okay I've experimented some more and
export GPU_MAX_ALLOC_PERCENT=100
will obviously cause a problem, but it gives me massive scope for trying different tc, etc since I can theoretically allocate 3GB on a 7970. However the code would try to allocate the largest buffer by default which is not right with this value now, so I've committed a change (to git) that will find the nearest buffer size that is a power of 2 between the thread concurrency buffer requirement and the maximum allocable ram.

Playing some more with my 7970s I found the values it was already running were pretty close to optimal. The only thing I found useful with the increased ram available was that I am now able to increase threads a little more (5 instead of 4) which afforded another 5kh per device.

The good news is that this should be what the 5x/6x miners were looking for on linux. I suggest you download the latest changes from git and start experimenting with TCs that are a multiple of your shader count after setting the above variable to 100%.
legendary
Activity: 1361
Merit: 1003
Don`t panic! Organize!
Can someone compile it w/CPU ?
My "work" machine I can use gets 15kh from gpu and close to another 15kh from 3 threads on CPU (using reaper+minerd).
cgminer uses alot less memory than reaper Smiley
-ck
legendary
Activity: 4088
Merit: 1631
Ruu \o/
AMD Buffer size problem  (cause of my performance woes with 5850)

I found an interesting thread on AMD Developer Central today: "Large buffers"  http://devgurus.amd.com/message/1282913#1282913

user drallan presents a workaround for the low maximum allocation: allocating several memory chunks and addressing everything as one block ...


maybe the thread will go on and bring more ideas

I am going to try setting "GPU_MAX_ALLOC_PERCENT" when I get at my machine after work


Most interesting....
Woah it works!... for linux people, just start with
export GPU_MAX_ALLOC_PERCENT=40

or other values and try again I was able to allocate whopping buffers on my 7970....
-ck
legendary
Activity: 4088
Merit: 1631
Ruu \o/
AMD Buffer size problem  (cause of my performance woes with 5850)

I found an interesting thread on AMD Developer Central today: "Large buffers"  http://devgurus.amd.com/message/1282913#1282913

user drallan presents a workaround for the low maximum allocation: allocating several memory chunks and addressing everything as one block ...


maybe the thread will go on and bring more ideas

I am going to try setting "GPU_MAX_ALLOC_PERCENT" when I get at my machine after work


Most interesting....
newbie
Activity: 40
Merit: 0
AMD Buffer size problem  (cause of my performance woes with 5850)

I found an interesting thread on AMD Developer Central today: "Large buffers"  http://devgurus.amd.com/message/1282913#1282913

user drallan presents a workaround for the low maximum allocation: allocating several memory chunks and addressing everything as one block ...


maybe the thread will go on and bring more ideas

I am going to try setting "GPU_MAX_ALLOC_PERCENT" when I get at my machine after work

legendary
Activity: 1204
Merit: 1000
฿itcoin: Currency of Resistance!
I need some help to compile CGMiner with Scrypt under Ubuntu 12.04 64 bits, 1 7970, Catalyst 12.6, AMDAPP v2.6...

What is the procedure?

Basically: git clone, checkout newscrypt, autogen, configure --enable-scrypt ... ?

Thanks!
Thiago
vip
Activity: 980
Merit: 1001
sent tittiez a small thank you for his contribution Smiley
-ck
legendary
Activity: 4088
Merit: 1631
Ruu \o/
There's no official binary build yet ?
Nope, only tittiez's builds. I'd still consider the software beta quality which is why there's no official release. None of this is making much sense and raper's performance seems almost fortuitous that it's coded without normal sanity checks so matching its performance is proving nothing short of a pain.
member
Activity: 98
Merit: 10
There's no official binary build yet ?
Pages:
Jump to: