User experience feedback on the language OpenCL

 
Does anyone on this forum use the OpenCL language ? I am looking for user experience feedback on this language before ordering an EA. Except for a few interesting articles, I couldn't find a topic on the forum. Personally, I have a very powerful processor with a modern graphics card, hence my interest. Whether you are a trader or a programmer, thank you for your opinions on the subject.
 
David CAUJOLLE: Does anyone on this forum use the OpenCL language ? I am looking for user experience feedback on this language before ordering an EA. Except for a few interesting articles, I couldn't find a topic on the forum. Personally, I have a very powerful processor with a modern graphics card, hence my interest. Whether you are a trader or a programmer, thank you for your opinions on the subject.

Then you did not search very well ... https://www.mql5.com/en/search#!keyword=OpenCL&module=mql5_module_forum

Besides the many posts by MetaQuotes, there several posts by normal users too and their experiences.

For example by the user @Lorentzos Roussos ... https://www.mql5.com/en/search#!keyword=OpenCL&module=mql5_module_forum&author=LoRio

But there are more. So take some time to do your research.

 

OpenCL


The forum

CodeBase

      The articles

        The blogs

          The documentation

              The market

              • many in search here
               

              If you want to use it for neural networks , and you have an nVidia gpu skip it go straight to pyTorch . 

              The main "feeling" i got is that only openCL cares about the "idea" of openCL and gpu vendors are reluctant to assign resources and money in their departments to support it.

              Recent developments in nVidia reveal why that was (sort of like someone is asking you to optimize your facilities for horses while you are developing a supersonic jet)

              If however you want to learn more and get the experience under your belt you can try it . It won't be fun at all though and just like with neural nets you will have the tendency to assume you learned something while you have actually not just because it happened to work . You need to be aware of that . (difference is neural nets are fun and frustrating this is just frustrating)

              The docs here won't help either and you will get the sense that the person who wrote the docs is not the person who deployed the library (i feel them, i would need a vacation after that too).

              There is a newer attempt by the Khronos group itself ,its called Vulkan Compute if i recall correctly and there is a high level wrapper available for c++ , i'll try and find it.


               
              Lorentzos Roussos #:
              pyTorch.

              Thanks for your feedback. It's not for developing an EA with a neural network, too complicated for me and I've heard bad feedback on that. OpenCL only interests me to simply increase the speed of my backtests.

               
              David CAUJOLLE #:Thanks for your feedback. It's not for developing an EA with a neural network, too complicated for me and I've heard bad feedback on that. OpenCL only interests me to simply increase the speed of my backtests.

              OpenCL is for programming functionality in OpenCL, not for speeding up back-tests.

               
              David CAUJOLLE #:

              Thanks for your feedback. It's not for developing an EA with a neural network, too complicated for me and I've heard bad feedback on that. OpenCL only interests me to simply increase the speed of my backtests.

              I see . 

              The first step would be to squeeze the melon as much as possible before openCL .

              So if i have an ea that has to do passes for different rsi parameters and different macd parameters i could test and see if there is a gain in loading each rsi and each macd vs calculating it at each pass (so -if the tester is not doing that already- for example if you had to test a macd setting with rsi7 rsi9 ... rsi60 etc you would remove the constant calculation of the macd which would be the same value in all these tests)

              edit : the macd and rsi are for example , there's probably caching going on for default indicators . you should substitute rsi and macd with whatever calcs you run that are repetitive across tests in each new bar or in each tick.
               
              Lorentzos Roussos #:

              If you want to use it for neural networks , and you have an nVidia gpu skip it go straight to pyTorch . 

              The main "feeling" i got is that only openCL cares about the "idea" of openCL and gpu vendors are reluctant to assign resources and money in their departments to support it.

              Recent developments in nVidia reveal why that was (sort of like someone is asking you to optimize your facilities for horses while you are developing a supersonic jet)

              If however you want to learn more and get the experience under your belt you can try it . It won't be fun at all though and just like with neural nets you will have the tendency to assume you learned something while you have actually not just because it happened to work . You need to be aware of that . (difference is neural nets are fun and frustrating this is just frustrating)

              The docs here won't help either and you will get the sense that the person who wrote the docs is not the person who deployed the library (i feel them, i would need a vacation after that too).

              There is a newer attempt by the Khronos group itself ,its called Vulkan Compute if i recall correctly and there is a high level wrapper available for c++ , i'll try and find it.


              Why is that ? This is very personal and subjective. You would better have said "I had not fun at all".

               
              Alain Verleyen #:

              Why is that ? This is very personal and subjective. You would better have said "I had not fun at all".

              touche` but i was describing my experience anyway , i was not describing the experience of a tall handsome dude from New Zealand with a porche who does not like hot dogs .

              It was implied . 😋

              The reader assumes that too . I'm describing it from my own point of view and you obviously liked working with it so , props . ps : if you liked openCL neural nets might be a breeze for you , try it.

               
              Lorentzos Roussos #:

              touche` but i was describing my experience anyway , i was not describing the experience of a tall handsome dude from New Zealand with a porche who does not like hot dogs .

              It was implied . 😋

              The reader assumes that too . I'm describing it from my own point of view and you obviously liked working with it so , props . ps : if you liked openCL neural nets might be a breeze for you , try it.

              From my experience, "implied" communication is ineffective and misleading.
               
              Alain Verleyen #:
              From my experience, "implied" communication is ineffective and misleading.

              i think you are overreacting a bit sir  😇