GPU acceleration in Krita

[Dmitry Kazakov]

Hi, Andrew!

Welcome back to Krita! :)
> I'm working on a prototype using OpenCL, and I was able to get a
> *functionally* correct implementation of 'composite' algorithm
> (KoCompositeOpAlphaBase.h), which is one of the performance bottlenecks
> that I spotted. OpenCL version of 'composite' is currently sub-optimal,
> but it passes a dozen of existing unit tests.
Instead of trying to optimize the general-case KoCompositeOpAlphaBase 
composite op, you can try to optimize specific composite ops, like 
KoOptimizedCompositeOpOver{32,128} and 
KoOptimizedCompositeOpAlphaDarken{32,128}. They are used in 90% of the time

> So far the only major issue was heavy use of C++ templates and type
> traits in Krita core code. OpenCL mainly supports C language, so I had
> to write a lot of C adapters and use a preprocessor meta-magic to lower
> everything from C++ (host side) to C (device side).
Do you have any decision on how you are going to optimize data transfers 
between CPU and GPU? Some proxy object for KisPainter and 
KisPaintDevice, so that the data will not leave GPU?
> I'll focus on getting sane performance numbers out of this
> implementation, and it may change the current design quite
> significantly. Then the code will probably be good enough to be shared.
Yeah, the numbers is a good thing :)

> Andrew
>
> On Thu, Apr 6, 2017 at 5:51 PM, Boudewijn Rempt <boud at valdyas.org> wrote:
>> On Thu, 6 Apr 2017, Andrew Savonichev wrote:
>>
>>> Hi Boudewijn,
>>>
>>> Okay, I understand.
>>> I would take a couple of weeks for PoC work to have better understanding
>>> of changes need to be made and issues we may face.
>>>
>>> I will get back to you when I have more to discuss.
>> Awesome! I'm looking forward to that!
>>
>>>
>>>      - Andrew
>>>
>>>
>>> On Thu, Apr 6, 2017 at 12:01 PM, Boudewijn Rempt <boud at valdyas.org> wrote:
>>>> On Thu, 6 Apr 2017, Andrew Savonichev wrote:
>>>>
>>>>> Hello,
>>>>>
>>>>> I'd like to know what is the status of GPU usage in Krita. I know
>>>>> Krita can use OpenGL
>>>>> for rendering, but I'm thinking about offloading image processing
>>>>> algorithms to GPU.
>>>>>
>>>>>  From my understanding, many algorithms in Krita are data parallel and
>>>>> operate on entire image, what
>>>>> makes them good candidates for offloading.
>>>>>
>>>>> Is there any directions, discussions or maybe some existing work on
>>>>> enabling GPU acceleration?
>>>>> What do you think about adding this to Krita?
>>>> Right now, only the canvas uses the gpu. We have had two attempts to
>>>> use the GPU for implementing filters, once using opencl, once with
>>>> glsl. That code is so old and was so unripe, it's probably not even
>>>> useful to look at.
>>>>
>>>> Apart from filters, recomputing the layer stack on the gpu could be
>>>> worth-while, and writing brush engines that run on the gpu could be
>>>> worth-while. In both cases, the problem is getting the pixel data to
>>>> the gpu and back in the main memory.
>>>>
>>>> There are two possible approaches: store everything in main memory and
>>>> copy to the gpu memory when needed, and then when done, back, but that
>>>> is slow. The other approach would be to keep the entire layer stack in
>>>> gpu memory, but that would limit the size of images -- and people do
>>>> work on images that take gigabytes of memory.
>>>>
>>>> But, I am _very_ eager to see progress in this area and would
>>>> love to work with you to see this happen.
>>>>
>>>>
>>>> --
>>>> Boudewijn Rempt | http://www.krita.org, http://www.valdyas.org
>> --
>> Boudewijn Rempt | http://www.krita.org, http://www.valdyas.org

-- 
Dmitry Kazakov

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.kde.org/pipermail/kimageshop/attachments/20180506/83abcab3/attachment.html>