cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

liwoog
Adept II

Re: RADEON HD 7970 on linux

Can one get access to the 12.3 preview for linux?

The 12.2 are still broken.

0 Likes
liwoog
Adept II

Re: RADEON HD 7970 on linux

12.3 is still broken, though it seems to freeze a little less. I guess I will try the NVIDIA GTX 680 next, the Radon HD 7970 OpenCL drivers are clearly not usable for this round of GPU cards.

29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffff8100b0f2>] ? system_call_fastpath+0x16/0x1b
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffff814f692e>] ? do_device_not_available+0xe/0x10
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffff811899c1>] ? sys_ioctl+0x81/0xa0
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffff81189444>] ? do_vfs_ioctl+0x84/0x580
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffff811892a2>] ? vfs_ioctl+0x22/0xa0
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffffa02ae93e>] ? ip_firegl_unlocked_ioctl+0xe/0x20 [fglrx]
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffffa02b8e6d>] ? firegl_ioctl+0x1ed/0x250 [fglrx]
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffffa02dc120>] ? firegl_cmmqs_CWDDE32+0x0/0x100 [fglrx]
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffffa02dc190>] ? firegl_cmmqs_CWDDE32+0x70/0x100 [fglrx]
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffffa02dd5af>] firegl_cmmqs_CWDDE_32+0x14f/0x440 [fglrx]
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffffa02a8bae>] KCL_SEMAPHORE_DownUninterruptible+0xe/0x10 [fglrx]
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffff810968b1>] down+0x41/0x50
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffff814f55a2>] __down+0x72/0xb0
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffffa02daf22>] ? firegl_trace+0x72/0x1e0 [fglrx]
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffffa02daf22>] ? firegl_trace+0x72/0x1e0 [fglrx]
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffff814f4685>] schedule_timeout+0x215/0x2e0
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffff8118ba50>] ? pollwake+0x0/0x60
29.03.2012 - 10:47:08hermesWarningkernelkernel: [<ffffffff8118ba50>] ? pollwake+0x0/0x60
29.03.2012 - 10:47:08hermesWarningkernelkernel: Call Trace:
29.03.2012 - 10:47:08hermesWarningkernelkernel: ffff880805328678 ffff880807721fd8 000000000000f4e8 ffff880805328678
29.03.2012 - 10:47:08hermesWarningkernelkernel: ffffc90000000000 ffff880807721ab8 ffffffff8118ba50 dead000000100100
29.03.2012 - 10:47:08hermesWarningkernelkernel: ffff880807721bd8 0000000000000082 0000000000000000 00000000000000db
29.03.2012 - 10:47:08hermesInformationalkernelkernel: Xorg D 0000000000000001 0 2903 2901 0x00400000
29.03.2012 - 10:47:08hermesErrorkernelkernel: "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
29.03.2012 - 10:47:08hermesErrorkernelkernel: INFO: task Xorg:2903 blocked for more than 120 seconds.
0 Likes
rick_weber
Adept II

Re: RADEON HD 7970 on linux

I have 3 7970s running in Ubuntu and the only problems I've had with them is that mapping a buffer > 256MB causes a zombie process after it hangs.

In Ubuntu, try this https://wiki.ubuntu.com/X/Troubleshooting/FglrxInteferesWithRadeonDriver#Problem:__Need_to_fully_rem...

to fully cleanse your system of the drivers and then install the ones on AMD's site. Be warned that kernel header updates break the drivers, so you'll have to do this anytime you update them.

I'm running Ubuntu 11.10, APP 2.6, and the 8.921 drivers, have a monitor connected to each card, and did the dumb tricks needed to run OpenCL over ssh. Other than that, everything is vanilla.

0 Likes
liwoog
Adept II

Re: RADEON HD 7970 on linux

I was able to get the proper computation by inserting a clFlush() between the clEnqueueNDRangeKernel() calls. 12.3 drivers. This should not be required, but fixes the problem. Would you have an explanation on the cause of the bug. This piece of code now provides the proper results (though quite slow compared to the NVIDIA). My other piece of code still ends up freezing the hardware but runs about 80% faster on the 7970 than on a GTX 580.

This works:

  // Enqueue a processing step

                    if (CL_SUCCESS != (err = clEnqueueNDRangeKernel(..)

 

  // Enqueue a barrier (For NVIDIA GPUs)

                    if (CL_SUCCESS != (err = clEnqueueBarrier(..)))

  // Make sure the previous kernel is loaded on the GPU before changing the arguments (For AMD GPUs)

                    if (CL_SUCCESS != (err = clFlush(..)))

 

  // Set the arguments for the second processing step

                    err  = clSetKernelArg(..);

                    err |= clSetKernelArg(..);

  // Enqueue a processing step

                    if (CL_SUCCESS != (err = clEnqueueNDRangeKernel(..)))

0 Likes
Marix
Adept II

Re: RADEON HD 7970 on linux

I also see these ASIC hangs on my three 7970s running on an openSUSE 12.1 in all drivers up to the 12.4 Preview with OpenCL 1.2 support. The code causing this runs fine on an AMD FirePro V7800. At least in my case the issue seems to be somewhat related to problem size. On small datasets it works, on larger datasets and the resulting increase in memory consumption and runtime the hangs happen. Worst part is that only way to get the GPU back after that is to reset the whole machine.

0 Likes
liwoog
Adept II

Re: RADEON HD 7970 on linux

Identical experience. After the clFlush fix for the kernel argument bug (see above), I can run small problems but not large ones.

0 Likes
RezaRob
Adept I

Re: RADEON HD 7970 on linux

pwvdendr wrote:

HD7970 beats Tesla by a factor 2 in speed, and still you only paid half the price. Is that worth the support?

Increase the price 45 bucks and give us smooth drivers and support.  We like that!

Even 100 bucks!  We want the thing to work!

EDIT:  I'm very sorry if you got multiple email notifications for my edits.  I just realized I could turn that off.

0 Likes
RezaRob
Adept I

Re: RADEON HD 7970 on linux

liwoog wrote:

Identical experience. After the clFlush fix for the kernel argument bug (see above), I can run small problems but not large ones.

Hi liwoog,  is it possible for you to try the steps suggested by pwvdendr here:

http://devgurus.amd.com/message/1280510#1280510

0 Likes
arsenm
Adept III

Re: RADEON HD 7970 on linux

I'm still seeing similar problems using 12.3, but they disappear using the 1.2 beta driver (1.4.1720)

0 Likes
yurtesen
Miniboss

Re: RADEON HD 7970 on linux

I seem to have a similar problem with 12.4 could you ever resolve the issue completely?

http://devgurus.amd.com/thread/159073

In my case clFlush() or clFinish() does not really help. After 8-10 enqueues of the same kernel (only with different offset). It crashes. The problem is not a 180 second time limit, I am able to run the kernel if I enqueue 3-4 times but with larger global sizes. I even tried to enqueue whole range at once and it ran in about 500 seconds with single enqueue and there were no problems.

Thanks!

0 Likes