Error after 14sec / Gamma-ray pulsar binary search #1 on GPUs v1.18

Peter Baetz
Peter Baetz
Joined: 25 Feb 05
Posts: 9
Credit: 10794955
RAC: 2
Topic 212287


After ~ 14sec the app finish with a error:

Gamma-ray pulsar binary search #1 on GPUs v1.18 (FGRPopencl1K-ati) x86_64-pc-linux-gnu


Job ID: 330291313
Erstellt: 30 Dez 2017 15:42:17 GMT
Gesendet: 30 Dez 2017 16:09:01 GMT
Ablaufdatum: 13 Jan 2018 16:09:01 GMT
Empfangen: 30 Dez 2017 16:13:56 GMT
Serverstatus: Over
Resultat: Computation error
Clientstatus: Compute error
Endstatus: 69 (0x00000045) Unknown error code
Computer: 4867904
Run time (sec): 14.14
CPU time (sec): 0.78
Peak working set size (MB): 70.81
Peak swap size (MB): 166.58
Peak disk usage (MB): 0.02
Prüfungsstatus: Invalid
Gewährte Punkte: 0
Anwendung: Gamma-ray pulsar binary search #1 on GPUs v1.18 (FGRPopencl1K-ati) x86_64-pc-linux-gnu

Gary Roberts
Gary Roberts
Joined: 9 Feb 05
Posts: 4251
Credit: 12698771162
RAC: 25062908

Peter Baetz

Peter Baetz wrote:
Using OpenCL platform provided by: Mesa
Using OpenCL device "AMD TURKS (DRM 2.43.0 / 4.4.103-18.41-default, LLVM 6.0.0)" by: AMD
Max allocation limit: 751619276
Global mem size: 1073741824
OpenCL compiling FAILED! : -11 . Error message: error: unsupported OpenCL extension 'cl_khr_fp64' - ignoring error: unknown type name 'double2'; did you mean 'double'? error: use of type 'double' requires cl_khr_fp64 extension to be enabled
OpenCL device has no FP64 support
OpenCL compiling FAILED! : -11 . Error message: <unknown>:0:0: in function kernel_ts_2_phase_diff_sorted void (float addrspace(1)*, i32, %struct.PhotonPair addrspace(1)*, i32, i32, float, i32, %struct.Bloc_info addrspace(1)*, i32): unsupported call to function llvm.AMDGPU.barrier.local
Couldn't compile OpenCL kernels (error: 1)
initialize_ocl returned error [2006]
OCL context null
OCL queue null
Error generating generic FFT context object [5]
17:10:36 (9106): [CRITICAL]: ERROR: MAIN() returned with error '5'

Could anybody help me?

I really know very little about Mesa/LLVM as I've never tried to use it. The comments that follow are just guesses.  If I'm not stating things correctly, I'd appreciate being corrected by someone who knows more about this than I do.

Firstly, your device is seen as 'TURKS' which is relatively old - Wikipedia tells me it's something like a HD6570 or HD6670, etc.  It has just enough VRAM (1GB) to allow the scheduler to send tasks for it but in the snip above you can see that the card itself doesn't have FP64 support which means that part of the calculations requiring double precision would be performed on the CPU rather than the GPU.  In itself that doesn't stop tasks being crunched - just slows them down.

As I understand it, Mesa/LLVM is a work in progress to provide (amongst other things) support for hardware not directly supported by AMD's latest amdgpu/amdgpu-pro drivers.  The above snip makes it fairly clear that for the versions of Mesa/LLVM you are using, your card is not properly supported.

Previously, cards that did have OpenCL capability were supported by the proprietary fglrx driver which has been deprecated.  fglrx would probably allow your card to crunch (just a guess) but is incompatible with recent versions of xorg.  To use fglrx, you would need to install an old enough version of your OS to have the 1.17.x version of xorg.  Later versions won't allow fglrx to be used.

I have essentially the same problem.  I use lots of cards that work fine with fglrx so (for those cards - mainly Pitcairn series, a few Cape Verde and Tahiti) I've controllled my OS upgrades to allow fglrx to continue working.  I do this on the assumption that there will eventually be support developed for them.  Maybe they'll all die of old age before that happens :-).

For your particular situation, since your machine has a nice new Ryzen 7 CPU, shout yourself a late Christmas present and upgrade your GPU to a basic Polaris series card :-).  Something like an RX 460 (if you can find one cheaply) or an RX 560 (just a little faster - not much) would do nicely without breaking the bank :-).  I'm running versions of both that don't need PCIe power connectors so don't consume much power - probably little more than your old card.  The output is good for the power used.

Otherwise, I suspect you'll be waiting a long time for your card to be properly supported, if ever.



Peter Baetz
Peter Baetz
Joined: 25 Feb 05
Posts: 9
Credit: 10794955
RAC: 2

Thanks for your very good

Thanks for your very good comment. Yes, for some months I had updated my OS without thinking to the fglrx-problem (I did not know about it).

Your are right, and I will bought a RX560. (next year) ;-)






Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.