OpenCL: internal implementation tests in MQL5 - page 62

[Sceptic Philozoff]

joo:

It's very funny, but the OCL program runs dashingly even if there is a crawl outside the array! - and it works just fine, but it gets not just anything but 0 from an empty, unmarked blob.

Whether it's good or not is up to one's discretion. I just know about it now.

Zeros don't mean anything. The OCL program might just not compile.

Are you using an external kernel compiler (from the graphics hardware vendor)?

[Andrey Dik]

Mathemat:

Zeros don't tell you anything. The OCL program may simply not compile.

Are you using an external compiler from the graphics hardware vendor?

What do you mean it won't compile? It compiles and it works. I don't know whose OCL compiler it is.

[Ilyas]

You can write out of memory for the sake of experiment (preferably in a loop, but with more).
After hard rebooting the computer, share your impressions...

That is the reason why there is no OpenCL support in the tester so far...

At the moment we are looking for ways to solve this OpenCL problem.

[Sceptic Philozoff]

joo: What do you mean it wouldn't compile? It will compile and work. I don't know whose OCL compiler it is.

So you have a look, Andrei. I have an Intel off-line one (there's even a picture in the article). Lots of messages if something went wrong. You can edit the kernel immediately in it to make it quicker.

Very handy. Too bad there's no debugging. But it is much more convenient than in the terminal editor.

[Andrey Dik]

Mathemat:

So you have a look, Andrei. I have an Intel off-line one (there's even a picture in the article). Lots of messages if something went wrong. You can edit the kernel immediately in it to be quicker.

It is very convenient. It's a pity that there is no debugging. But it is much more convenient than in the terminal editor.

The kernel works and works without errors, according to the logic. I do not edit the kernel in ME at all, as it is useless in terms of detecting potential errors - I check and edit it in OpenCLCodeChecker. The kernel works.

I deliberately set the loop to go beyond the array and it does not cause the OCL driver to crash when the program is running; this "opening" was reported in this thread. See Ilyaz's post above.

[Sceptic Philozoff]

joo: I don't edit kernel in ME at all, as it's useless in terms of detecting potential errors - I check and edit it in OpenCLCodeChecker. The kernel works.

I agree, it's a dumb thing in ME.

My OpenCLCodeChecker doesn't work like that. Well, what the hell with it, I'll do with Intel's one somehow...

[Andrey Dik]

Mathemat:

I agree, it's a dumb thing in ME.

My OpenCLCodeChecker doesn't work like that. Well, fuck it, I'll do with the Intel one somehow...

I should try the Intel one too... Maybe it's cooler? - rhetorical question.

[Konstantin Grebenshikov]

http://www.3dnews.ru/news/629294?topblock It seems that the combination of graphics and processors (more precisely the discrete video) AMD may be significantly ahead of Intel and Nvidia in terms of performance and net speed in the future, both together and in turn, according to the above slides. The information is good for the project as a whole and for thinking about the upgrade, especially since the next version of the standard (1.2) is about to be released with the promise (according to the announcement) to simplify the software development. AMD is more suitable for OpenCL than anyone else....

[Sceptic Philozoff]

Yeah, there's a reason AMD bought ATI six years ago. Just when the Core 2 Duo was announced.

Intel is not sitting still on the IGP part either, but is still lagging behind. On ixbt experts say Intel HD 4000 will rival the current A6 in terms of graphics. But that's a big progression too.

P.S. By the way, the OpenCL part of Trinity is not all that great. It seems to be based on Cayman CE (i.e. 6xxx series, though they call it HD 7xxx for some reason) but double to float performance ratio is 16.

[Konstantin Grebenshikov]

Mathemat:

P.S. By the way, the OpenCL part of Trinity is not all that great. The core seems to be based on Cayman CE (i.e. 6xxx series, though they call it HD 7xxx for some reason), but double to float performance ratio is 16.

Most likely this will be fixed in next generation of processors. If AMD hadn't bought ATI a couple of years ago, we users wouldn't have been able to feel such a fierce struggle... The only regret that Intel competitors do not have similar technology as Hyper-threading hopefully AMD will invent it. Then I'll see what to buy if such an important event for all of us happens.