OpenCL: internal implementation tests in MQL5 - page 19

New comment
 
WChas:

The top file is a download from mine, the bottom one from your link. Completely different size.... )

Well it's simple, I put the wrong link there. :)

I have both files. With the same size.

I gave up for now. Installed 2010 studio and AMD SDK now. They are installed correctly, will study examples tomorrow.

Going to bed now.

 
For those buying new graphics cards to use OpenCL.
I recommend that you carefully read the specs for using the new and old PCI-E bus card.
The point is that if you decide to use both video cards, the bus speed will be chosen at the slower speed, thus you could lose out on performance.
Internal tests show that the data transfer loss from GPU to RAM is quite noticeable in terms of time!
 
mql5:


Internal tests show that the loss in transferring data from GPU to RAM is quite noticeable in time!


Unfortunately, this is the worst thing that can negate the speed of GPU. For this reason, OpenCL gives an advantage only in large-scale computations where you have to load the task once and get the results a second time. That is, it is bad for parallel processes that need to load tasks with a large amount of input data from an application to GPU and receive the results when ready to go back, as well as for small tasks.
 
There is a question related to the fact that there are prerequisites for integrating the GPU into the CPU. Won't there be conflicts between OpenMP and OpenCL in the future at all levels of the terminal and tester?
 
-Alexey-: Won't there be conflicts between OpenMP and OpenCL at all levels of terminal and tester in the future?

Is OpenMP already supported in MQL5?!

If the question is too silly - don't hit too hard.

 
Supported via dll, and probably many people like me use this feature. The acceleration from one and a half to several times is quite tangible, and does not require a video card, works either in indicator or in tester.
 
You can support the devil with a dll.
 
Mathemat:
You can support the hell with a dll.

Don't give a damn. :)

But for some reason the developers chose OpenCL and not OpenMP. Not for a reason... It seems that in the next builds it will be possible to feel it in the tester, as I understand it. I'm just not sure whether I will be able to feel it at all. Tests will show. :)

 
Mathemat:
You can support the hell out of it with a dll.
The tester knows how to parallelize across processor cores. However, if the EA code already uses a dll with parallel calculations, it is kind of unnecessary since it may cause a slowdown. Local cores can be disabled in the tester but it is not clear what will happen with OpenCL programs of MCL and the OpenCL tester where, what, how. The tester will parallelize the runs or just a single run will be paralleled. I.e. the EA program will be paralleled on GPU or in the tester, or both, and how it will interact with OpenMP externally. In short, I'm already confused...
 

What does error 5102 mean and what is the danger?

Everything looks fine, the program works, but GetLastError() returns an error, no deciphering in the help so far.

1...121314151617181920212223242526...70
New comment