search

UCL / openqcd-oneapi

GNU General Public License v2.0
0 stars 0 forks source link

Resolve incorrect spinor values from CUDA code #17

Closed tkoskela closed 2 years ago

tkoskela commented 2 years ago

We have generated binary files per the instructions in UCL/openqcd-oneapi#22 on csd3 using gdb and the executable qcd1 of the branch feature/library with a lattice size of 8 8 8 8. We used the rename_binary_files.sh script to name them according to the naming convention in UCL/openqcd-oneapi#22. Using the renamed files for input for the main executable in tests/cuda and tests/cuda2 results in incorrect results for spinors 2048-4095 on P100, V100 and A100 GPUs.

NOTE: The first 2048 values are therefore correct! Maybe this is an indexing issue

tkoskela commented 2 years ago

Makis is currently on leave, returning to work on Monday 21/02/2022

tkoskela commented 2 years ago

Had a discussion with Makis on 4/3/2022 He was able to provide me with his original input/output files on CSD3. Both the CUDA and the SYCL code generated by dpct report correct results with each data set (16^4, 32^4, 64^4). I've copied the files to 

/rds/project/dirac_vol2/rds-dirac-dr004/openqcd/makis-ref-data

There is very likely an error in the instructions Makis gave us on how to generate the files, but it has not yet been resolved. I think we can proceed using these files for testing, and build a more robust way of testing after the first report has been submitted at the end of March.