LibreOffice Receives More OpenCL / GPU Support

[stiiixy]

you mean like it's already possible?

mmm, noice. Complety missed (obviously!) that feature upon release!

[Sdar]

AFAIK they already helped GIMP devs to bring some plugins to OpenCL and aren't going to stop.
Cycles is not as easy thing at all. But keep in mind. Cycles was created by someone who knows CUDA and worked with CUDA and wrote Cycles in CUDA. Then Brecht added OpenCL support for the functions that were already similiar to the ones available in CUDA. AFAIK by Brecht himself, he just transcends CUDA functions to OpenCL during compile time.
CUDA is written with NVidia architecture in mind that changed through years after they bought technology. AFAIK there is almost no good OpenCL books on Nvidia cards, alot of developers complained about it.
So yes, its a pity for myself as well but you can't make a clear example of bad OpenCL in AMD based on Cycles.
From user point of view: LuxRender works on AMD GPU, Cycles isn't. Yes, both are having troubles with yet not as mature as CUDA standart but one team is always solving it while the other doesn't give a sh*t and still writes vendor specific code. So complaining about Cycles that is optimised for Nvidia hardware is strange.

The Cycles Opencl code compiles well using OPENCL with Nvidia and intel and run pretty well. (Nvidia Opencl is a little bit slower than Nvidia Cuda in some tests but not much)

It barely work with amd unless you don't have all the features enabled, and when it works it does slower than a CPU.

Luxrender works but they're having troubles too with some added functionality cause the AMD OPENCL Compilers tends to crash everytime you tried to do something complex.

AMD knows and has admitted that there's a problem with the compiler and they're working to fix it (Yeah they're slow as hell cause it's taking years to see some improvement).

Amd said recently "16-oct-2013"#"It will be in the best interest of your time -- to migrate to GCN if you want Blender on AMD cards."# albeit it's not working yet and it didn't seems that there's going to be fixed in the next months.

So yeah it's strange to complain that amd saids years ago that they're working on it and still the amd Opencl just works whit simple code and the complex ones don't work or work too slow to be usefull, Nvidia by the way has Cuda but his Opencl Implementation works better than the amd one.

Just look some benchs and you'll realize that amd beats Nvidia Opencl performance just in simple calculations... but whit complex programs Nvidia tends to take the lead.

################################################

Amd does great GPU's but their opencl is not mature enough yet... just hope HSA got more efforts to be great.

[YoungManKlaus]

Regardless, it's only a spreadsheet, but when the OpenCL stuff is implemented, why not try? Might even start making Libreoffice network based as well (reducing the need for updating your old hardware even more).

you mean like it's already possible?

[stiiixy]

Who ever mentioned farms? I just mentioned letting a server handle the processing. I dont see it as all that particularly hard on a network with thousands of other PC's sitting arond doing nothing 95% of the cycle-time. With Intel and AMD shipping ridiculously powerful GFX chips on their CPU's now, pretty much every device would make a nice small render farm. Obviosuly prioritisiation and security would need to be managed, along with barring some machines from even processing, but that's all semantics and comes down to the design.

Regardless, it's only a spreadsheet, but when the OpenCL stuff is implemented, why not try? Might even start making Libreoffice network based as well (reducing the need for updating your old hardware even more).

[YoungManKlaus]

Say, I've got Dual AMD R 290X beasts spinning its wheels with over 5600 stream processors and and AMD FX-8350 with 32GB ready to crunch and you're complaining about off-loading distributed OpenCL first onto a distributed network of servers?

How about we scale properly on a workstation with OpenCL kernels flooding the GPGPUs first locally, then write a grid app to extend to a supercomputer [you know something the consumer and office will never buy].

Same goes for Blender, GIMP and more. Outside of a game studio/motion picture studio, no one is buying a render farm.

sure, sure, easy, you are both just forgetting something:
- someone has to pay for that super computer
- someone has to write/maintain that code
- it takes time to offload work packages to remote machines
- some people prefer to keep their data to themselves
- you can already always use remote-desktop to some powerful machine, and now even with html5

[Marc Driftmeyer]

It'd be even more impressive if they actually offloaded the core work to dedicated processing servers instead of a single local machine.

If you aren't 'interwebised', then yeah, obviously fall back to local-CL over the servers CL.

That shit was sorted out in the 70's. Some thing called Unix.

Say, I've got Dual AMD R 290X beasts spinning its wheels with over 5600 stream processors and and AMD FX-8350 with 32GB ready to crunch and you're complaining about off-loading distributed OpenCL first onto a distributed network of servers?

How about we scale properly on a workstation with OpenCL kernels flooding the GPGPUs first locally, then write a grid app to extend to a supercomputer [you know something the consumer and office will never buy].

Same goes for Blender, GIMP and more. Outside of a game studio/motion picture studio, no one is buying a render farm.

[stiiixy]

It'd be even more impressive if they actually offloaded the core work to dedicated processing servers instead of a single local machine.

If you aren't 'interwebised', then yeah, obviously fall back to local-CL over the servers CL.

That shit was sorted out in the 70's. Some thing called Unix.

[sunweb]

AFAIK they already helped GIMP devs to bring some plugins to OpenCL and aren't going to stop.
Cycles is not as easy thing at all. But keep in mind. Cycles was created by someone who knows CUDA and worked with CUDA and wrote Cycles in CUDA. Then Brecht added OpenCL support for the functions that were already similiar to the ones available in CUDA. AFAIK by Brecht himself, he just transcends CUDA functions to OpenCL during compile time.
CUDA is written with NVidia architecture in mind that changed through years after they bought technology. AFAIK there is almost no good OpenCL books on Nvidia cards, alot of developers complained about it.
So yes, its a pity for myself as well but you can't make a clear example of bad OpenCL in AMD based on Cycles.
From user point of view: LuxRender works on AMD GPU, Cycles isn't. Yes, both are having troubles with yet not as mature as CUDA standart but one team is always solving it while the other doesn't give a sh*t and still writes vendor specific code. So complaining about Cycles that is optimised for Nvidia hardware is strange.

[YoungManKlaus]

ua=42: good answer, thx
I didn't ask out of malice but because I was really interested.

[ua=42]

if this is really a thing that people can utilize. Eg. if there are really spreadsheets that are big/complex enough that a difference is noticeable?
... on the other hand I know that some businesses are crazy and implement whole business processes in spreadsheets ...

There is actually 2 scenarios where openCL support will be useful.
1. Huge spreadsheets.
2. Spreadsheets being run on netbooks, low power laptops, single core computers, old dual cores, etc.

On lower power computers that have a graphics chip that supports openCL, it will help process the data in parallel, speeding up the loading time.
Also, the effort to implement OpenCL support should also help with adding multithreading support because they have similar concerns on data being thread safe and being processed in the right order.


If you have an AMD A series, then things become interesting as it won't have the performance penalty of copying data to and from the graphics card. Which means the calc features will run allot faster.