|
21)
Message boards :
Number crunching :
Gpu App?
(Message 1279)
Posted 26 Jul 2016 by matszpk Post: I can write something about progress in development of GPU app. I wrote tentative version of CLRunEngine (engine that can help efficiently run complex code on GPU, including a sparse and non-regular control flow). Now, I writing an utility to facilitate porting complex from original app to CLRunEngine in enough natural way. Utility is not finished so far. Conversion from original code to code accepted by my engine would be a enough complex and cumbersome to maintain in handmade way. Therefore, I am trying to write an utility that help to move original to my engine. Non-regular way of doing computations by original application and a sophisticated and sensitive code is enough big challenge for me. Very likely utility will be finished in next week or slightly later. |
|
22)
Message boards :
Number crunching :
Gpu App?
(Message 1230)
Posted 18 Jun 2016 by matszpk Post: I am still working over GPU app. I invented the engine that help to run massive number of tasks efficiently (including its irregularity) on GPU. First tests of this engine was passed correctly. Now, I am working over rewriting main code of the application to this engine. I can't write about speedup and majority of the possible potential bottlenecks. This engine will require a very little transfer between host and GPU and it will move the program control to GPU. I hope that engine allow to achieve good results under modern GPU (even if it will run quite complex code). |
|
23)
Message boards :
Number crunching :
Gpu App?
(Message 1137)
Posted 9 Apr 2016 by matszpk Post: I am developing an OpenCL application for AMD Radeon GPU's. I would like to write about progress and preliminary results of these works. Firstly: this application is difficult and cumbersome case due to two problems: nonregular compulations (in the GPU scale where tens thousands jobs scheduled independently, regular jobs are very desired) and sophisticated code of main routines that will be ported to GPU world. Secondly: I have some results that encouraged to further work. Routines of the main functions performed quite good on GPU (likes Radeon HD 7850). Main computations is doing on double precision, however timings on low-end and midrange GPU's are encouraged. The most routines ran 20-40x faster on GPU than on single core of CPU. I was testing code on my Core i5 (4GHz), and HD 7850 (1,01GHz). These results and timings need not to be reflecting final speedup of application. Workunits for GPU version can be many greater than CPU workunits for better scaling on GPU, and GPU application can need many more resources to run (mainly memory). Unfortunatelly. I don't know when a GPU version will be finished and released to tests. Current state of the works doesn't allow to give any deadline. |
(English)
(Polski)