Re: Disabling of OpenCL by default in Calc

[Eike Rathke]

Hi Markus,

On Tuesday, 2016-06-21 02:51:17 +0200, Markus Mohrhard wrote:

> so I'd like to disable OpenCL by default in Calc. I was once again hit by
> some nasty issue in the OpenCL code that makes calc unusable for me. In my
> opinion we have way too many issues with it and it reduces the reliability
> of Calc ( who trusts his OpenCL driver enough to do his taxes with OpenCL
> enabled).

I agree, OpenCL turned out to be a can of worms, and having Calc produce
wrong results that may go unnoticed and even be not reproducible in all
cases because of different drivers is just a mess.

Let the default be disabled. Already in 5.2

  Eike

-- 
LibreOffice Calc developer. Number formatter stricken i18n transpositionizer.
GPG key "ID" 0x65632D3A - 2265 D7F3 A7B0 95CC 3918  630B 6A6C D5B7 6563 2D3A
Better use 64-bit 0x6A6CD5B765632D3A here is why: https://evil32.com/
Care about Free Software, support the FSFE https://fsfe.org/support/?erack


signature.asc
Description: PGP signature
___
LibreOffice mailing list
LibreOffice@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/libreoffice

Re: Disabling of OpenCL by default in Calc

[Markus Mohrhard]

Hey,

On Tue, Jun 21, 2016 at 9:30 AM, Stephan Bergmann 
wrote:

> On 06/21/2016 02:51 AM, Markus Mohrhard wrote:
>
>> so I'd like to disable OpenCL by default in Calc. I was once again hit
>> by some nasty issue in the OpenCL code that makes calc unusable for me.
>> In my opinion we have way too many issues with it and it reduces the
>> reliability of Calc ( who trusts his OpenCL driver enough to do his
>> taxes with OpenCL enabled).
>>
>> Unless someone does some huge refactoring of the OpenCL code and makes
>> sure that we can trust the OpenCL backend as much as our normal formula
>> code I would prefer if we disable it. In this case it is a regression
>> from 40b0b9ab7703a165295b008f47df14d2ec076fb1 but that we don't detect
>> such fundamental problems because of the different engines and the
>> different devices makes it hard to rely on the OpenCL code.
>>
>
> Can you give a short description of what the problem actually is?
>


The device selection code that runs with a new profile has become quite
slow (taking several dozen seconds) which makes Calc which makes tests
extremely slow and causes dozens of seconds delay when you create the first
formula. I have not yet completely understood what is going wrong.

But that is not the only issue. Just going to bugzilla gives several open
issues and serious ones show up from time to time. E.g.
https://bugs.documentfoundation.org/show_bug.cgi?id=99053

It is also quite obvious in all the hard to reproduce calc bugs as now in
general one of the first questions is about OpenCL as nobody is currently
trusting the OpenCL device and we have no way to check the OpenCL device.
It is a bit similar to how the OpenGL drivers have to blacklisted with the
difference that for the OpenCL ones it is more serious as it causes wrong
spreadsheet calculations.

Regards,
Markus


> [...]
>
>> It would be nice if this could be put onto the ESC agenda so that we
>> could disable it in 5.2.
>>
>
> ___
> LibreOffice mailing list
> LibreOffice@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/libreoffice
>
___
LibreOffice mailing list
LibreOffice@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/libreoffice

Re: Disabling of OpenCL by default in Calc

[Stephan Bergmann]

On 06/21/2016 02:51 AM, Markus Mohrhard wrote:

so I'd like to disable OpenCL by default in Calc. I was once again hit
by some nasty issue in the OpenCL code that makes calc unusable for me.
In my opinion we have way too many issues with it and it reduces the
reliability of Calc ( who trusts his OpenCL driver enough to do his
taxes with OpenCL enabled).

Unless someone does some huge refactoring of the OpenCL code and makes
sure that we can trust the OpenCL backend as much as our normal formula
code I would prefer if we disable it. In this case it is a regression
from 40b0b9ab7703a165295b008f47df14d2ec076fb1 but that we don't detect
such fundamental problems because of the different engines and the
different devices makes it hard to rely on the OpenCL code.


Can you give a short description of what the problem actually is?

[...]

It would be nice if this could be put onto the ESC agenda so that we
could disable it in 5.2.


___
LibreOffice mailing list
LibreOffice@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/libreoffice

Disabling of OpenCL by default in Calc

[Markus Mohrhard]

Hey guys,

so I'd like to disable OpenCL by default in Calc. I was once again hit by
some nasty issue in the OpenCL code that makes calc unusable for me. In my
opinion we have way too many issues with it and it reduces the reliability
of Calc ( who trusts his OpenCL driver enough to do his taxes with OpenCL
enabled).

Unless someone does some huge refactoring of the OpenCL code and makes sure
that we can trust the OpenCL backend as much as our normal formula code I
would prefer if we disable it. In this case it is a regression from
40b0b9ab7703a165295b008f47df14d2ec076fb1 but that we don't detect such
fundamental problems because of the different engines and the different
devices makes it hard to rely on the OpenCL code.

I would prefer to disable the OpenCL backend by default until we have a
solution to the problems. That includes a way to detect broken drivers
automatically and a way to have a few tinderboxes that are actually
executing the OpenCL code path. Right now we are not even aware if any box
is executing the tests (and I suppose none is).

It would be nice if this could be put onto the ESC agenda so that we could
disable it in 5.2.

Regards,
Markus
___
LibreOffice mailing list
LibreOffice@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/libreoffice