jansol
jansol
Looks like the `cpu` backend already supports that extension so it's most likely just because no CUDA contributor has needed it so far.
If you have a vendor-provided OpenCL driver for the GPU you can use the proxy driver in PoCL to get both devices in the same platform.
You can't use prebuilt clinfo with the proxying PoCL. You'll have to build your own and link it directly against the pocl library instead of the OpenCL ICD loader.
The pthread driver was renamed to "cpu" a version or two ago. And for cmake you should pass `-DENABLE_HOST_CPU_DEVICES=1 -DENABLE_LLVM=1` instead.
With `POCL_DEVICES="cpu"`?
It technically affects "plain" remote too, that machine just got lucky.
#1433 should fix the RDMA CI
In case someone wants to tackle this, there are two places that need to be fixed: `lib/CL/clSetEventCallback.c`, as the callback may be called immmediately, and the more expected place: `pocl_event_updated...
#1433 should fix the RDMA CI
My educated guess is that it's trying to link against a libOpenCL that only supports OpenCL 1.x (such as Apple's) but there's not enough information here to tell for sure...