[Beignet] Beignet not working on Dell Precision M3800

[Yichao Yu]

On Tue, May 27, 2014 at 12:24 AM, Yang, Rong R <rong.r.yang at intel.com> wrote:
> For test_copy_buffer, you can try Beignet's latest commit, it may fix it.

Yes you are right, almost all the tests in test-wrapper.py passes now
(except 3d image, more on this later).

However, it seems that there are other problems as well (as expected)
(which also seems to do sth really bad to the graphic driver :( ) but
I haven't tried to figure which test is causing that. Should I isolate
the failing case and report here (or other appropriate places) or is
there already a schedule for fixing these?

> For test_image_3d, Beignet support 3d image, but it seems this test used 2 channels float format? Beignet only support 1 channel(CL_R) and 4 channels(CL_RGBA) float format.

I.C. However, it seems that beignet returns
CL_INVALID_IMAGE_FORMAT_DESCRIPTOR for unsupported formats. Shouldn't
it return CL_IMAGE_FORMAT_NOT_SUPPORTED in such cases instead?

[1] http://cgit.freedesktop.org/beignet/tree/src/cl_mem.c#n564

>
> -----Original Message-----
> From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On Behalf Of Yichao Yu
> Sent: Tuesday, May 20, 2014 12:12 PM
> To: Yang, Rong R
> Cc: beignet at lists.freedesktop.org
> Subject: Re: [Beignet] Beignet not working on Dell Precision M3800
>
>> [Rong]: Screen frozen may be GPU hang, can you provide your tests or kernel, then I will reproduce and fix it.
>
> There seems to be multiple ones but at least here is one of them:
>
> I was using the pyopencl[1] test cases[2] (cuz I am developing the cffi version for pypy) and the two test cases that are failing in `test_wrapper.py` are `test_image_3d`[4] and `test_copy_buffer`[5].
> The 3d image test fail with invalid image format error (is 3d image not supported) and the copy buffer test hang the GPU without doint anything (the returned buffer is empty).
>
> I have created a c version for you to reproduce here[6], in which you can modify the `BUF_SIZE`. On my computer, it fails for all the numbers I have tried from 1 to 50000 so it shouldn't be the problem of too big data transfer (and .... in any cases copying 50000 numbers shouldn't take seconds :P). I haven't played with the flag for creating buffers but at least the current flags have 100% reproducibility here ....
>
> [1] https://github.com/pyopencl/pyopencl/
> [2] https://github.com/pyopencl/pyopencl/tree/master/test
> [3] https://github.com/pyopencl/pyopencl/blob/master/test/test_wrapper.py
> [4] https://github.com/pyopencl/pyopencl/blob/master/test/test_wrapper.py#L396
> [5] https://github.com/pyopencl/pyopencl/blob/master/test/test_wrapper.py#L471
> [6] https://gist.github.com/yuyichao/242fd2a812088930af91
>
>>
>> [Rong]: Yes, must hack the kernel to walk around it now, and the drm driver team is handling it now, but maybe need some time.
>
> Good to know at least someone are working on it.
>
>>
>>
>> Looking forward to full support soon.
>>
>> Yichao Yu
>>
>>>
>>> Thanks
>>>
>>
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet