What steps will reproduce the problem?
(say we have 6-core CPU)
1. export GOMP_CPU_AFFINITY=0-5
2. run the cuda/ocelot program
What is the expected output? What do you see instead?
All CPU cores must be equally loaded.
Instead only one core is busy, others are idle.
What version of the product are you using? On what operating system?
svn revision: 2235, system: gentoo GNU/Linux amd64
Please provide any additional information below.
My own OpenMP code works fine(better than without) with affinity set.
Original issue reported on code.google.com by volyrkr on 6 Jun 2013 at 6:48
Original issue reported on code.google.com by
volyrkr
on 6 Jun 2013 at 6:48