Michal Babej
Michal Babej
Hello, > the documentation you provided that you successfully built pocl in ODROID XU3 and ODROID C2 and almost passed all tests. Perhaps it's unclear, but that documentation speaks about...
> the result is that platform information is obtained, but GPU device information cannot be retrieved what is your `POCL_DEVICES` env variable ? You need to set it to `POCL_DEVICES=proxy`...
I think your build is missing `-DENABLE_LOADABLE_DRIVERS=0`. I tested building proxy & cpu on a visionfive2 board i have here. Built with: ``` mkdir ~/0/build/b_pocl_17_cpu_proxy cd ~/0/build/b_pocl_17_cpu_proxy cmake -DENABLE_ICD=0 -DENABLE_LOADABLE_DRIVERS=0...
@He32100, could you try to 1) remove the contents of pocl cache (usually `$HOME/.cache/pocl` ) 2) run the test again with `POCL_DEBUG=all` env variable, that should (hopefully) print something more...
> Program build: kernel metadata setup failed This is unfortunately an old unsolved problem with the proxy driver and program metadata. It should work if you create a `cl_program` with...
> Also, the *Rect functions do not seem to use the SVM cl_mem shadow buffer refactoring. Yes, this is a rebased & slightly cleaned up code from ~1 year ago,...
A few tests still fail: ``` The following tests FAILED: 106 - regression/barrier_between_two_for_loops_cbs (Failed) 142 - regression/test_alignment_with_dynamic_wg_114_cbs (Failed) 144 - regression/test_alignment_with_dynamic_wg_117_cbs (Failed) 146 - regression/test_alignment_with_dynamic_wg_225_cbs (Failed) 148 - regression/test_alignment_with_dynamic_wg_173_cbs (Failed)...
``` [2024-02-27 14:50:58.734061576] POCL: in fn read at line 193: *** ERROR *** | ERROR | Read error on reentrant_read(fd, req, request->req_size, &request->req_read), fd=9, reason: Bad address [2024-02-27 14:50:58.734090907] POCL:...
Hi @fodinabor, do you have time to look into the CBS failure a bit more ? Unfortunately removing the instcombine between `loop-barriers` and `barrier-tail` passes still makes `loopvec` fail (because...
The current pthread driver's scheduling is equivalent to OpenMP's dynamic scheduling. On consumer systems the difference between OpenMP/TBB/pthread seems minimal, however on machines with multiple NUMA nodes the difference can...