Algorithms do not heed the maximum work-group size of the device
Open
bug
help wanted
- Dominant language
- C++
- Stars
- 1.7k
- Forks
- 340
- PR merge metrics
- No merged PRs in 30d
Description
Almost all algorithms in boost compute, which set an explicit local work-group size do not check if this size exceeds the device maximum work-group size. Instead they all set it to a default value of 128 (e.g. `algorithm/detail/binary_find.hpp` in line 82).
Thus, most of the algorithms and therefore the Boost.Compute test-cases fail for OpenCL devices with a maximum work-group size of less then 128.
So, the default value should be checked against the device's maximum work-group size, which already can be retrieved via `boost::compute::device::max_work_group_size()`.
Contributor guide
Assessment
This issue has not been assessed yet.