Cache is written to in simulation mode

[stijnh]

The simulation runner writes to the cache file, even though that is not necessary.

When setting the cache file to read-only, the following exception is thrown:

Traceback (most recent call last):
  File "/var/scratch/sheldens/miniconda3/lib/python3.12/multiprocessing/pool.py", line 125, in worker
    result = (True, func(*args, **kwds))
                    ^^^^^^^^^^^^^^^^^^^
  File "/var/scratch/sheldens/accuracy-tuning-paper/plots/run_tournament.py", line 171, in simulate_run
    return simulator.run(threshold, strategy_name, strategy_options)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/var/scratch/sheldens/accuracy-tuning-paper/plots/run_tournament.py", line 83, in run
    results, env = kernel_tuner.tune_kernel(
                   ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/var/scratch/sheldens/miniconda3/lib/python3.12/site-packages/kernel_tuner/interface.py", line 666, in tune_kernel
    util.process_cache(cache, kernel_options, tuning_options, runner)
  File "/var/scratch/sheldens/miniconda3/lib/python3.12/site-packages/kernel_tuner/util.py", line 1161, in process_cache
    cached_data = read_cache(cache)
                  ^^^^^^^^^^^^^^^^^
  File "/var/scratch/sheldens/miniconda3/lib/python3.12/site-packages/kernel_tuner/util.py", line 1221, in read_cache
    filestr = correct_open_cache(cache, open_cache)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/var/scratch/sheldens/miniconda3/lib/python3.12/site-packages/kernel_tuner/util.py", line 1213, in correct_open_cache
    with open(cache, "w") as cachefile:
         ^^^^^^^^^^^^^^^^
PermissionError: [Errno 13] Permission denied: 'cache_cp_2048x2048_NVIDIA-A100-PCIE-40GB.json'

[benvanwerkhoven]

Especially before merging pull request #266 this is not supposed to happen and should be fixed. As of pull #266 behavior will be slightly different, which is one of the reasons it has not been merged yet.

[stijnh]

Thanks. It might make sense to wait fixing this bug until after #266 has been merged (if it still occurs afterwards).

As a temporary fix, currently I just make a scratch copy of the cache file before running the simulation runner.

[benvanwerkhoven]

I'd have to look it up but #266 completely rewrites how caches work so this fix in itself is by definition a temporary fix until #266 is merged. I do think that this should be the behavior also after #266 is merged. If you're only running simulations, you wouldn't expect Kernel Tuner to open the file for writing or appending at all, and I think it indeed should not try to.