Michal Babej

> clang+llvm-9.0.1-aarch64-linux-gnu_kai Are you using LLVM downloaded from `https://releases.llvm.org/download.html` ? If so, that is the reason it doesn't work. These were compiled on some linux distribution (which the page doesn't...

@Oblomov WRT "SPIR-V isn't supported by pocl" - SPIR-V is actually somewhat supported by pocl (because we needed it for another project). But ATM it's only possible to use via...

@Oblomov if you mean `copyKernelFromBitcode`, it's supposed to copy the bitcode of the kernel from the bitcode of the whole `cl_program` into a new empty module. The reason this exists...

What's wrong is that you have compiled both the "platform" of pocl and the "opencl CPU device" code for ARM. Your kernel is telling you that it cannot launch ARM...

> How important is subdevice support for PoCL? Are there common use cases out there? It becomes important when using PoCL on with machines with multiple NUMA nodes (CPU sockets)....

@isuruf Yes the tests fail b/c of the subdevice support, we need an extra check in them and exit with a "skip" exit code (IIRC 77). WRT the `check_cmd_queue_for_device`: it...

@dnsampaio the relevant code is in `lib/CL/pocl_llvm_utils.cc`, function `pocl_llvm_create_context`. It'd be preferable to use the LLVM C API (`LLVMContextSetOpaquePointers()`) instead of the C++ API, but in this case it likely...

@inducer I've been unable to measure any speedup (though i admit i haven't tested extensively). The downside is: if two opencl programs which use subdevices, use the same subdevice, they...

Ahh, ok that does make sense. Perhaps we could somehow get the numa node count from hwloc and auto-enable affinity on those systems. But that's future work, right now i'm...

@jrprice yes, that's likely one issue. The kernel scheduling of current pthread driver is oriented toward lowest latency, not the highest throughput. The other issue is that everything but kernel...