[SYCL][Doc] Fix work_group_scratch_memory example #20506
Conversation
The example in the specification was using an old API that we removed in intel#14785. Update the example and the spec wording to use `launch_config` from sycl_ext_oneapi_enqueue_functions to pass the property with the size of the scratch memory.
| local memory capacity as reported by `info::device::local_mem_size` | ||
| then the implementation must throw a synchronous exception with the | ||
| `errc::memory_allocation` error code from the kernel invocation command | ||
| (e.g. `parallel_for`). This check must take all APIs that allocation device |
There was a problem hiding this comment.
Will we have a check to confirm whether the SLM size is set for certain kernel?
Say, if a kernel calls the syclex::get_work_group_scratch_memory() in kernel body, but we do not set the syclex::work_group_scratch_size in nd_launch/parallel_for, will an exception be thrown?
I'm asking this because looks like there will be no any warning or error in current implementation.
There was a problem hiding this comment.
I think this makes sense, and it seems to me that it would be easy to implement. I think we just need a check here to see if WorkGroupMemorySize has been set by the user.
https://github.com/intel/llvm/blob/sycl/sycl/source/detail/scheduler/commands.cpp#L2474-L2481
We should be careful, though, not to throw an exception just because the user set WorkGroupMemorySize to 0. Setting the size to 0 should be allowed.
@AlexeySachkov and @steffenlarsen do you see any problem with this? I assume this code is executed synchronously by the same application thread that launches the kernel, right? Or, is this code executed asynchronously by another thread? If that is the case, the error would be an async error.
I think we could throw errc::memory_allocation in this scenario too, and we can distinguish this case from the out-of-memory case via the message string.
2 participants
The example in the specification was using an old API that we removed in #14785. Update the example and the spec wording to use
launch_configfrom sycl_ext_oneapi_enqueue_functions to pass the property with the size of the scratch memory.