Posts by [B@P] Daniel

1) Message boards : Application Code Discussion : Intel Open CL (Message 66679) Posted 6 Oct 2017 by [B@P] Daniel Post: Intel GPUs support double precision ops starting from SkyLake.
2) Message boards : Application Code Discussion : MilkyWay on ARM Mali T-628 GPU (Message 66677) Posted 2 Oct 2017 by [B@P] Daniel Post: Hi, I was able to successfully build and run MilkyWay GPU app on ARM Mali T-628, on Odroid XU4 board. First WU has just been completed and validated successfully: http://milkyway.cs.rpi.edu/milkyway/result.php?resultid=2138942517. Mali GPU needed about 100 minutes to complete it. This time could be better if app would support integrated CPU/GPU memory (the same as for Intel GPU), some time could be saved on data copying. Odroid uses Ubuntu Linux, so building procedure is the same as for x86 Linux. I had small issue with linking with libOpenCL.so because one of last system updates did something bad to it and linker started complaining about missing cl* functions in this lib. I fixed this by replacing this lib with symlink to libmali.so: cd /usr/lib/arm-linux-gnueabihf/ mv libOpenCL.so libOpenCL.so.broken ln -s libmali.so libOpenCL.so You also need proper app_info.xml to make it working. I use following one: <app_info> <app> <name>milkyway</name> <user_friendly_name>MilkyWay@Home</user_friendly_name> <non_cpu_intensive>0</non_cpu_intensive> </app> <file> <name>milkyway_separation</name> <executable/> </file> <app_version> <app_name>milkyway</app_name> <version_num>146</version_num> <platform>arm-unknown-linux-gnueabihf</platform> <avg_ncpus>0.0300000</avg_ncpus> <max_ncpus>0.0300000</max_ncpus> <plan_class>opencl_nvidia_101</plan_class> <api_version>7.6.33</api_version> <file_ref> <file_name>milkyway_separation</file_name> <main_program/> </file_ref> <coproc> <type>Mali-T628</type> <count>1.0</count> </coproc> <gpu_ram>268435456.000000</gpu_ram> <dont_throttle/> </app_version> </app_info> Support for other GPU types beside 3 major ones (Nvidia, AMG, Intel) generally works, but there are still some issues there - see https://github.com/BOINC/boinc/issues/1686. For some reason BOINC Clienr sees two instances of Mali GPU, but uses only 1st one. Recently I was running Xansons for COD on this GPU, and my BOINC Client build very long work queue. My work queue for MilkyWay so far looks good (12 tasks waiting), will see if it will remain on this level or start growing significantly.