I have applied the kernel patch[1] as Zhigang Gong suggested off-list and it indeed fixes some failing tests (especially some out-of-resources errors). However, the test program I was using still hangs at exactly the same place (clWaitForEvents).
P.S. I have finally found a working OpenCL 1.2 implementation (intel-opencl-sdk for CPU) and it works fine with the test program so I guess it is actually a bug in beignet. [1] https://aur.archlinux.org/pkgbase/linux-beignet-fix/ On Sun, Jun 15, 2014 at 6:44 AM, Yichao Yu <yyc1...@gmail.com> wrote: > On Sun, Jun 15, 2014 at 5:42 AM, Zhigang Gong <zhigang.g...@gmail.com> wrote: >> >> >>> 在 2014年6月15日，6:02，Yichao Yu <yyc1...@gmail.com> 写道： >>> >>>> On Sat, Jun 14, 2014 at 11:51 AM, Yichao Yu <yyc1...@gmail.com> wrote: >>>> Sorry for the delay. I was busy graduating and didn't have much time >>>> in the past two weeks for testing. >>>> >>>> The gpu hang is still there and I haven't been able to make a c >>> >>> Sorry I was wrong, it seems that the hang only happens before I >>> upgrade beignet. There are still a lot of failing tests but the screen >>> does not freeze anymore. >> Which Linux kernel did you using on your previous test? Did you apply the >> kernel patch which provided by Rong in his email? If you haven't applied the >> kernel patch, you will not get the slm and barrier work correctly. And all >> related tests are known broken. > > I C. I guess that is the reason than... > > THX > >>> >>>> version of the test program. However, I have found another problem >>>> with the newly merged opencl1.2 APIs when testing sth else. >>>> >>>> The c test program to trigger the issue is here[1]. When running on my >>>> Haswell CPU, beignet hangs in clWaitForEvents with the backtrace >>>> >>>> #0 0x00007ffff78cc9d0 in __nanosleep_nocancel () from >>>> /usr/lib/libc.so.6 #1 0x00007ffff78f6c94 in usleep () from >>>> /usr/lib/libc.so.6 #2 0x00007ffff73dfc8a in clWaitForEvents >>>> (num_events=1, event_list=0x7fffffffda58) at >>>> /home/yuyichao/projects/mlinux/pkg/all/beignet-git/src/beignet/src/cl_api.c:1316 >>>> #3 0x00007ffff7bc861e in clWaitForEvents (num_events=1, >>>> event_list=0x7fffffffda58) at ocl_icd_loader.c:873 #4 >>>> 0x00000000004009aa in main () at beignet-bug2.c:34 >>>> >>>> It seems that the problem only happens for the event returned by >>>> clEnqueueBarrierWithWaitList when the wait list is not empty. I hope I >>>> am not using the api in the wrong way but I don't have another working >>>> opencl 1.2 implementation (pocl crashes on clEnqueueBarrier*...) to >>>> test it......... >>>> >>>> [1] https://gist.github.com/yuyichao/8b661d51c81f1c85466e >>>> >>>>> On Wed, Jun 4, 2014 at 7:29 AM, Yichao Yu <yyc1...@gmail.com> wrote: >>>>>> On Tue, Jun 3, 2014 at 11:15 PM, Yang, Rong R <rong.r.y...@intel.com> >>>>>> wrote: >>>>>> Printf is not a built in function OpenCL 1.1, so beignet don't support >>>>>> it now. However, beignet are supporting it, maybe you could use it soon. >>>>> >>>>> However, even if the function is not defined, shouldn't the compiler >>>>> return a error (opencl error) rather than raising a exception and >>>>> abort? >>>>> >>>>>> Yes, the patch about 3D pipe have not push now, but You can apply by >>>>>> manual and try it. >>>>> >>>>> I'm afraid I don't have time to test it soon... >>>>> >>>>>> >>>>>> -----Original Message----- >>>>>> From: Yichao Yu [mailto:yyc1...@gmail.com] >>>>>> Sent: Thursday, May 29, 2014 8:40 PM >>>>>> To: Yang, Rong R >>>>>> Cc: beignet@lists.freedesktop.org >>>>>> Subject: Re: [Beignet] Beignet not working on Dell Precision M3800 >>>>>> >>>>>>> On Thu, May 29, 2014 at 4:46 AM, Yang, Rong R <rong.r.y...@intel.com> >>>>>>> wrote: >>>>>>> I have checked this issue, it is a beignet compiler bug, should be fix >>>>>>> by patch "GBE: Change 64bit integer storage in register". >>>>>>> >>>>>>> For the first problem, I have sent some patch, can you try them? The >>>>>>> patch " HSW: Restore L3 control register to disable SLM mode." fix a 3D >>>>>>> pipe affect by Beignet bug. May be the same problem you met. >>>>>> >>>>>> I am testing using the current master >>>>>> >>>>>> c34eba71bd5a518906d6d5d3ba26e44327cab251 >>>>>> GBE: fix one illegal instruction when replace a uniform dst. >>>>>> >>>>>> So the patch u mentioned for 3D pipe doesn't seem to be included yet. >>>>>> >>>>>> Here are what I saw, >>>>>> 1, `printf("%d\n", i);` works on pocl but still crashes the compiler on >>>>>> beignet with the same error. >>>>>> 2, the c example I gave works but the original python version does >>>>>> not... Will figure out the difference once I get more time. >>>>>> 3, the interference with opengl seems to be different. The same effect I >>>>>> mentioned last time shows up when sth is running on the GPU but recovers >>>>>> afterward. However, it now gives your email a funny texture by replacing >>>>>> some of the characters with another one...[1] (o in this >>>>>> case...) I also remember seeing this problem randomly sometime before >>>>>> but it was not as reproducible... >>>>>> >>>>>> I guess I will test again once those 3d pipe fixing patches are applied. >>>>>> >>>>>> [1] http://wstaw.org/m/2014/05/29/plasma-desktopzSP722.png >>>>>> >>>>>> Yichao Yu >>>>>> >>>>>>> -----Original Message----- >>>>>>> From: Yichao Yu [mailto:yyc1...@gmail.com] >>>>>>> Sent: Wednesday, May 28, 2014 11:49 PM >>>>>>> To: Yang, Rong R >>>>>>> Cc: beignet@lists.freedesktop.org >>>>>>> Subject: Re: [Beignet] Beignet not working on Dell Precision M3800 >>>>>>> >>>>>>>> On Wed, May 28, 2014 at 11:45 AM, Yichao Yu <yyc1...@gmail.com> wrote: >>>>>>>> On Wed, May 28, 2014 at 10:39 AM, Yichao Yu <yyc1...@gmail.com> wrote: >>>>>>>>>> The second problem is that there seems to be sth wrong if I run two >>>>>>>>>> tests in series. More specifically, `test_elwise_kernel`[3], >>>>>>>>>> `test_elwise_kernel_with_option`[4] and >>>>>>>>>> `test_ranged_elwise_kernel`[5] can all pass if I run them >>>>>>>>>> individually. However, if I run them together, only the first one >>>>>>>>>> can pass... I will try to reproduce this in C... >>>>>>>>> >>>>>>>>> Sorry this is NOT what happened... I was not using the right >>>>>>>>> parameter to select the tests and there isn't any (at least no >>>>>>>>> evidence for it) interference between kernels. >>>>>>>>> The problem is rather the test_elsize_kernel_with_option and >>>>>>>>> test_ranged_elwise_kernel are not working.. >>>>>>>>> Also the failing one sometimes (~2 times in 8) hang the wm for ~10s... >>>>>>>>> will try to make a c version.... >>>>>>>> >>>>>>>> And it seems that none of them is actually working, just that when >>>>>>>> the difference is calculated using OpenCL, it always returns 0... >>>>>>>> >>>>>>>> so here[1] is the c version. The problem seems to be related to the >>>>>>>> use of get_local_size and/or get_group_id in the kernel. When I was >>>>>>>> using a simple kernel with `int i = get_global_id(0);`, everything >>>>>>>> works fine. >>>>>>> >>>>>>> I haven't applied the patch for using local memory in the kernel. Does >>>>>>> that patch affect not only local memory but also local size somehow? >>>>>>> >>>>>>>> >>>>>>>> [1] https://gist.github.com/yuyichao/242fd2a812088930af91 >>>>>>>> >>>>>>>> P.S. I was trying to use printf in the kernel and it seems to crash >>>>>>>> the compiler..... Not sure if I was using it correctly but I guess it >>>>>>>> shouldn't crash in any case... >>>>>>>> >>>>>>>> here is the error: >>>>>>>> ``` >>>>>>>> ASSERTION FAILED: it != instrinsicMap.map.end() at file >>>>>>>> /home/yuyichao/projects/mlinux/pkg/all/beignet-git/src/beignet/backen >>>>>>>> d /src/llvm/llvm_gen_backend.cpp, function void >>>>>>>> gbe::GenWriter::regAllocateCallInst(llvm::CallInst&), >>>>>>>> line 2115 [1] 28951 trace trap (core dumped) ./beignet-bug >>>>>>>> ``` >>>>>>>> >>>>>>>> with the following kernel (not sure if it is valid haven't use printf >>>>>>>> before....), >>>>>>>> >>>>>>>> ``` >>>>>>>> __kernel void fill_one(__global float *out, long n) { >>>>>>>> int i = get_global_id(0); >>>>>>>> printf("%d\n", i); >>>>>>>> if (i < n) { >>>>>>>> out[i] = 1; >>>>>>>> } >>>>>>>> } >>>>>>>> ``` >>>>>>>> (this kernel (without printf) works btw....) >>>>>>>> >>>>>>>> Yichao Yu >>>>>>>> >>>>>>>>>> >>>>>>>>>> [1] http://wstaw.org/m/2014/05/28/plasma-desktopObn722.png >>>>>>>>>> [2] http://wstaw.org/m/2014/05/28/plasma-desktopWbB722.png >>>>>>>>>> [3] >>>>>>>>>> https://github.com/pyopencl/pyopencl/blob/master/test/test_algorith >>>>>>>>>> m >>>>>>>>>> .py#L45 [4] >>>>>>>>>> https://github.com/pyopencl/pyopencl/blob/master/test/test_algorith >>>>>>>>>> m >>>>>>>>>> .py#L66 [5] >>>>>>>>>> https://github.com/pyopencl/pyopencl/blob/master/test/test_algorith >>>>>>>>>> m >>>>>>>>>> .py#L97 >>>>>>>>>> >>>>>>>>>> Yours, >>>>>>>>>> Yichao Yu >>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>>>>>>> Thanks for point out it, I have sent a patch to correct it. >>>>>>>>>> >>>>>>>>>> Seems fixed. THX. =) >>> _______________________________________________ >>> Beignet mailing list >>> Beignet@lists.freedesktop.org >>> http://lists.freedesktop.org/mailman/listinfo/beignet _______________________________________________ Beignet mailing list Beignet@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/beignet
