Hi, on linux we got some very funny results. On two separate programs that were initially not specifying workgroup size, when checked with the visual profiler seemed to have workgroups of size 1 ! not even 32 and an occupancy of 0.25. Just by specifying the workgroup size, the kernels started to run over 10 times faster ! What is that about ??? i must say i don’t really trust the opencl linux visual profiler as it gives very strange results some times.
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| work group size different work group lead to different performance | 1 | 809 | November 23, 2011 | |
| Compute Profiler - Confused by Work Group Information | 3 | 11235 | March 17, 2011 | |
| openCL --- weird behavior | 0 | 3852 | February 3, 2011 | |
| Opencl Global work size | 2 | 5460 | December 23, 2010 | |
| A new Windows 7 OpenCL app, with manual workload distribution. | 4 | 5439 | January 14, 2010 | |
| trying to understand kernel parameters and CL_INVALID_WORK_GROUP_SIZE | 8 | 3980 | February 26, 2010 | |
| help me understand `odd' performance | 5 | 1668 | June 18, 2010 | |
| Help analyzing Visual Profiler Output | 0 | 4089 | February 10, 2011 | |
| Weird profiler results in occupance analysis | 0 | 531 | May 13, 2011 | |
| Kernel runs much faster in visual profiler than cmd | 1 | 603 | August 21, 2014 |