Well I just made an interesting discovery. My newly modified poclbm kernel included in 2.2.4 works really well on my 6970 with sdk 2.6... I get the same hashrate with my modified poclbm kernel as I used to get with the phatk kernel on 2.4/2.5. Looks like a new default kernel may be in order.
Would others like to test this out on their various hardware?
-k poclbm
Also suggest decreasing worksize:
-k poclbm -w 64
nice Ill test it out on my various cards and let you know Thanks a lot. As I said it's only the 6970 I tested it on and I was more than surprised. Other architecture may respond differently. ok so I got upgraded to the latest CCC with sdk 2.6 on my testing Machine, its Windows 7 x64 / cgminer 2.2.4 ok so i tested the poclbm kernel as well as -w 64 on 6990 & 5870 Clocks: 6990 900/775 5870 950/180 SDK 2.5 default kernel 6990 - 405 mhash Per core 5870 - 430 Mhash SDK 2.6 -k poclbm 6690 - 380-400 Mhash Per core (noticeably more fluctuation than with 2.5 and default kernel) 5870 - 390-420 Mhash Per core (again noticeably more fluctuation in mhash) SDK 2.6 -k poclbm -w 64 6990 - Peeks at about 400 Mhash per core but usually stays below that and goes as low as 350 5870 - Same result I see 330-400+ Mhash a lot of fluctuation in mhash Not sure why this is but with the 2.5 sdk and default kernel my speeds do not fluctuate hardly at all, If I look at the console of my 6990 rig every single core is running right about 405 mhash, same with the 5870, it stays pretty stable around 430 mhash, now with the 2.6 sdk and poclbm the speed is constantly changing and only peaks where it is normally stable with 2.5 and default kernel conclusion: in my humble opinion I think it is premature to make poclbm the default kernel, this is a nice option for people who game but if your running dedicated mining rigs you cant beat the 2.5 sdk with the default kernel Thanks for that. The question was about the poclbm kernel being default, not the SDK being default. How does poclbm perform on 2.5 sdk for you? |
|
|
|
Well I just made an interesting discovery. My newly modified poclbm kernel included in 2.2.4 works really well on my 6970 with sdk 2.6... I get the same hashrate with my modified poclbm kernel as I used to get with the phatk kernel on 2.4/2.5. Looks like a new default kernel may be in order.
Would others like to test this out on their various hardware?
-k poclbm
Also suggest decreasing worksize:
-k poclbm -w 64
nice Ill test it out on my various cards and let you know Thanks a lot. As I said it's only the 6970 I tested it on and I was more than surprised. Other architecture may respond differently. |
|
|
|
|
Well I just made an interesting discovery. My newly modified poclbm kernel included in 2.2.4 works really well on my 6970 with sdk 2.6... I get the same hashrate with my modified poclbm kernel as I used to get with the phatk kernel on 2.4/2.5. Looks like a new default kernel may be in order.
Would others like to test this out on their various hardware?
-k poclbm
Also suggest decreasing worksize:
-k poclbm -w 64
|
|
|
|
Something is horribly wrong with cgminer 2.2.4 for me. It seems the newest kernel for phatk sucks monkey balls on my 5970. on cgminer 2.2.1 and kernel from last year (not sure which one) it was 359mh/s @ 805mhz / 109mhz
With the new phatk kernel as of this year, it gets around 300mh/s !!
any idea why?
Im running 11.12 drivers and 2.6sdk I believe.
Its on vista 32 ultimate.
Any idea?!
So, Ive stuck w/ cgminer 2.2.1 cuz it uses last years phatk kernel.
Repeat after me; it is not the kernel, it's the SDK. |
|
|
|
On windows, you can also specify "export GPU_USE_SYNC_OBJECTS = 1"
Only way is different, "Computer> System Properties> Advanced system settings> system environment> create a new variable" GPU_USE_SYNC_OBJECTS ", the value of the variable is -" 1 "
It just has no effect at all on Windows systems (confirmed by AMD). You could as well set MITT_ROMNEY_IS_A_DICKHEAD = 1 ... would have the same effect  Excellent. Is he? |
|
|
|
Important code committed to git tree: The longstanding generation of a zero sized binary appears to be due to the OpenCL library putting the binary in a RANDOM SLOT amongst 4 possible binary locations. Iterate over each of them after building from source till the real binary is found and use that. THOSE HAVING THE ZERO SIZED BINARY BUG PLEASE TEST THIS EXECUTABLE!http://ck.kolivas.org/apps/cgminer/temp/cgminer.exeDrop it into a clean 2.2.4 directory with no .bin files and try it please. Or try a git build if you're that way inclined. |
|
|
|
which command line options to set for optimum 7970 performance?
-I 11 for the 7970. keeping in mind, on windoz you have to go down to 9 to prevent high CPU usage. also, if you are on windoz, and you want a I above 9, and you have more then 1 card. then you will need to let cgminer have more cores if you have them. with 5 cards and 1 core, I-11 clock:1125/1000 Mhash will be under 400, and jumpy. with all cores (I have 4) then you can get the ~650Mh, but CPU usage will be ~80% all cores. I have to run mine at I-9 for now. Oh really? Suckage. Presumably that's because there is no way to 'export GPU_USE_SYNC_OBJECTS=1' on windows which fixes the CPU usage on linux. Maybe some hidden option in the catalyst control centre? According to new user ED1980 who can't post on the forums yet: On windows, you can also specify "export GPU_USE_SYNC_OBJECTS = 1"
Only way is different, "Computer> System Properties> Advanced system settings> system environment> create a new variable" GPU_USE_SYNC_OBJECTS ", the value of the variable is -" 1 "
|
|
|
|
No that didn't work either.
Thanks for trying. I'll consider this a permanent "feature" bug on cgminer on fucking windows then goddamnit... The plot thickens. When in doubt, just look to AMD for epic fail. Each compiled openCL program returns multiple binaries... and it's random which slot it ends up in. Debugging my balls off today: GPU0:
[2012-02-12 20:41:51] CL_PROGRAM_NUM_DEVICES returns 4 devices for this program
[2012-02-12 20:41:51] binary 0 size 391512
[2012-02-12 20:41:51] binary 1 size 0
[2012-02-12 20:41:51] binary 2 size 0
[2012-02-12 20:41:51] binary 3 size 0
GPU1:
[2012-02-12 20:41:53] binary 0 size 0
[2012-02-12 20:41:53] binary 1 size 0
[2012-02-12 20:41:53] binary 2 size 542460
[2012-02-12 20:41:53] binary 3 size 0
GPU2:
[2012-02-12 20:41:54] binary 0 size 0
[2012-02-12 20:41:54] binary 1 size 542460
[2012-02-12 20:41:54] binary 2 size 0
[2012-02-12 20:41:54] binary 3 size 0
GPU3:
2012-02-12 20:41:55] binary 0 size 0
[2012-02-12 20:41:55] binary 1 size 0
[2012-02-12 20:41:55] binary 2 size 0
[2012-02-12 20:41:55] binary 3 size 542460
Anyone see a problem here? Well there is one good thing... and that is I may well have a fix. It also appears to be RANDOM as to which binary it is allocated to, which is why it works on some starts and doesn't on others, so repeat attempts to start it will eventually work when a binary is made. Loading from file works fine. Now to try and make a fix. |
|
|
|
No that didn't work either.
Thanks for trying. I'll consider this a permanent "feature" bug on cgminer on fucking windows then goddamnit... EDIT: I have looked at the code in question a billion times and I simply cannot see why it's failing. I need someone with fresh eyes to look at it and tell me "oh you're just doing this wrong" but there are so few of us that work on this code, and we're all working on "competing" code, even though that's not the truth. I'd give 5 BTC to someone who found what I was doing to cause this bug and how to fix it. |
|
|
|
Con, I don't have any success in editing the .cl file for DiaKGCN, it only generates HW errors. Tried a forced submitting of these shares with --submit-stale, but that seems to not trigger a forwarding to the pool. Is there any change to output the nonce or hash that DiaKGCN produces and which is considered a HW error?
This makes me crazy, because in Phoenix it's a healthy kernel ... dunno how to cure it for CGMINER currently :-(.
Thanks,
Dia
Guess I can give up on diapolo kernel then cause neither you nor me can seem to fix it. |
|
|
|
[2012-02-11 17:00:00] binary size 0 : 0
[2012-02-11 17:00:00] OpenCL compiler generated a zero sized binary, may need to
reboot! This is totally crazy, but perhaps the command is actually non-blocking even though it's meant to be blocking and c is fast so the compiler hasn't returned yet or something stupid like that on windows. Anyway try this totally crazy exe with your regular commands (put into a fresh 2.2.4 directory): http://ck.kolivas.org/apps/cgminer/temp/cgminer.exeThis will wait until it returns a non-zero value. If it seems to just sit there doing nothing after a minute or so, just stop it. It's just hanging at /listenChannel line for like 7 minutes now. Doesn't appear to be doing anything. Sam Darn. Throw out that binary. What happens if you try with a different kernel like -k poclbm ? |
|
|
|
[2012-02-11 17:00:00] binary size 0 : 0
[2012-02-11 17:00:00] OpenCL compiler generated a zero sized binary, may need to
reboot! This is totally crazy, but perhaps the command is actually non-blocking even though it's meant to be blocking and c is fast so the compiler hasn't returned yet or something stupid like that on windows. Anyway try this totally crazy exe with your regular commands (put into a fresh 2.2.4 directory): http://ck.kolivas.org/apps/cgminer/temp/cgminer.exeThis will wait until it returns a non-zero value. If it seems to just sit there doing nothing after a minute or so, just stop it. |
|
|
|
which command line options to set for optimum 7970 performance?
-I 11 for the 7970. keeping in mind, on windoz you have to go down to 9 to prevent high CPU usage. also, if you are on windoz, and you want a I above 9, and you have more then 1 card. then you will need to let cgminer have more cores if you have them. with 5 cards and 1 core, I-11 clock:1125/1000 Mhash will be under 400, and jumpy. with all cores (I have 4) then you can get the ~650Mh, but CPU usage will be ~80% all cores. I have to run mine at I-9 for now. Oh really? Suckage. Presumably that's because there is no way to 'export GPU_USE_SYNC_OBJECTS=1' on windows which fixes the CPU usage on linux. Maybe some hidden option in the catalyst control centre? |
|
|
|
Just installed 2.2.4 on WinXP Catalyst 11.6.
It did not generate the .bin files for me. I tried multiple times with different command line options and rebooted and tried again.
I copied my renamed .bin files over from the previous install and am working now.
Thanks,
Sam
I know this has been long standing, but there is more debugging in newer versions. Just once more for my crying pleasure, can you start it with --verbose -D -T (without any bin files in the directory) and give me the full output please? Here you go. Let me know if you need anything else. I appreciate it. Thanks, Sam cgm224>cgminer --verbose -D -T No sorry, I mean add those to your regular command line options. |
|
|
|
Just installed 2.2.4 on WinXP Catalyst 11.6.
It did not generate the .bin files for me. I tried multiple times with different command line options and rebooted and tried again.
I copied my renamed .bin files over from the previous install and am working now.
Thanks,
Sam
I know this has been long standing, but there is more debugging in newer versions. Just once more for my crying pleasure, can you start it with --verbose -D -T (without any bin files in the directory) and give me the full output please? |
|
|
|
which command line options to set for optimum 7970 performance?
-I 11 for the 7970. |
|
|
|
Results of ./cgminer -D -T (with pool, user, and pass, of course) Might be a tad bit long: [2012-02-11 22:36:59] GPU 0 iAdapterIndex 0 strUDID 296:38676:4098:53248:5208 iBusNumber 1 iDeviceNumber 5 iFunctionNumber 0 iVendorID 4098 strAdapterName ATI Radeon HD 4290
[2012-02-11 22:36:59] GPU 1 iAdapterIndex 2 strUDID 512:26808:4098:8663:5208 iBusNumber 2 iDeviceNumber 0 iFunctionNumber 0 iVendorID 4098 strAdapterName ATI Radeon HD 5700 Series
[2012-02-11 22:36:59] ADL found more devices than opencl!
[2012-02-11 22:36:59] There is possibly at least one GPU that doesn't support OpenCL
[2012-02-11 22:36:59] WARNING: Number of OpenCL and ADL devices does not match!
[2012-02-11 22:36:59] Hardware monitoring may NOT match up with devices!
[2012-02-11 22:36:59] GPU 0 ATI Radeon HD 4290 hardware monitoring enabled
[2012-02-11 22:37:05] CL Platform version: OpenCL 1.0 ATI-Stream-v2.1 (145)
[2012-02-11 22:37:05] Error: Building Program (clBuildProgram)
[2012-02-11 22:37:05] sh: /bin/x86_64/clc: not found
[2012-02-11 23:24:04] Error: Building Program (clBuildProgram)
[2012-02-11 23:24:04] /tmp/OCLDOCSn7.cl(177): warning: variable "t1" was set but never used
u t1;
^
1: You have hardware monitoring on a device that can't mine. OpenCL doesn't communicate with the ATI Display Library to tell it which device is which. --gpu-reorder may help that. 2: With that older stream (2.1) it uses the binary clc as part of building kernels. The fact is this is your main problem here and it cannot build anything and is not cgminer's fault. Usually this would happen if your installation is corrupt, but the fact that you have had other binaries built fine means it's something else. The other thing that comes to mind is that you are starting cgminer as a different user and it doesn't have access to where the binary is in that user's path: eg. starting as sudo when the binary is in /usr/local or vice versa. 3. That t1 warning is a warning and will not break compilation. |
|
|
|
... but it will accept -k diakgcn provided it is loading the kernel called diakgcn120208.cl . This will allow you to at least play with the .cl file and see if you can figure out what's wrong with it perhaps.
It accepts, but don't submits shares to pool  Shows a hashrate and GPU is hot (works) but there's no info about accepted shares (as usual) and pool website also confirms that there's no worker that is mining. THIS IS NOT FOR USAGE ! THIS IS FOR DEBUGGING PURPOSES BY DIAPOLO ONLY. IT DOES NOT PRODUCE SHARES. |
|
|
|
|
I just released cgminer 2.2.4 which is the first release with respectable (read not the fastest, but good) performance for 7970.
|
|
|
|
Well ... as I said, I have no IDE setup, so currently I can't compile a version for myself. If you don't have the time to fiddle around with my commits, then I really need help in setting up an IDE in Windows. Have you got this in a readme, wiki or can you give me a brief explanation in how to do this? I worked with MS VC++ Express as a hobby some time ago ... Dia. Grab the windows binary for 2.2.4. It is undocumented, but it will accept -k diakgcn provided it is loading the kernel called diakgcn120208.cl . This will allow you to at least play with the .cl file and see if you can figure out what's wrong with it perhaps. NORMAL USERS DO NOT USE THIS. IT IS FOR DEVELOPMENT AND DEBUGGING. IT DOES NOT PRODUCE SHARES. |
|
|
|
|
Show Posts
