Disable build caching when using pocl #738
Conversation
|
I gave this a quick spin (with the example in #731 (comment)) and didn't notice any negative perfomance impact on the second run, as long as the downstream package (i.e., pocl, nvidia cl) has caching enabled. |
That's a broad set to generalize over. 🙂 If they all have source -> executable caches, then sure, that'd probably be better. Nvidia has a such a cache, I think. Do you know about AMD and Intel? |
pyopencl/__init__.py
Outdated
| if _PYOPENCL_NO_CACHE: | ||
| build_descr = "uncached source build (cache disabled by user)" | ||
| else: | ||
| build_descr = "uncached source build (cache disabled for pocl)" |
There was a problem hiding this comment.
| build_descr = "uncached source build (cache disabled for pocl)" | |
| build_descr = "uncached source build (cache disabled because ICD is caching)" |
There was a problem hiding this comment.
If I add the characterize function suggested below, should I instead change this to say something like "source build via ICD cache"?
pyopencl/__init__.py
Outdated
| no_cache = _PYOPENCL_NO_CACHE | ||
| # Turn off caching when using pocl to avoid extra compile from | ||
| # get_info(BINARIES). See https://github.com/inducer/pyopencl/issues/731. | ||
| from pyopencl.characterize import get_pocl_version |
There was a problem hiding this comment.
Maybe add a characterize function on whether there is an effective source -> executable cache? (Specified to return None for "unknown", False for no, and True for yes.)
majosm
force-pushed
the
disable-build-cache-pocl
branch
from
4de0f58
to
a72c2b9
Compare
|
@inducer I'm seeing an intermittent failure in the boxtree CI (here's a failing run and a successful run for the same code). Is this something to be concerned about? |
| @@ -374,7 +374,7 @@ def _check_arg_size(function_name, num_cl_args, arg_types, devs): | |||
|
|
|||
|
|
|||
| if not cl._PYOPENCL_NO_CACHE: | |||
| invoker_cache = WriteOncePersistentDict( | |||
| invoker_cache: WriteOncePersistentDict = WriteOncePersistentDict( | |||
There was a problem hiding this comment.
Add key and value types WOPD[..., ...] (also below).
| if _PYOPENCL_NO_CACHE: | ||
| build_descr = "uncached source build (cache disabled by user)" | ||
| else: | ||
| build_descr = "source build via ICD cache" |
There was a problem hiding this comment.
| build_descr = "source build via ICD cache" | |
| build_descr = "uncached source build, assumed cached by ICD" |
Kind of, yeah. Can you reproduce it locally? |
|
Maybe it has something to do with tests being run in parallel? How good is pocl about locking its cache? |
majosm
force-pushed
the
disable-build-cache-pocl
branch
4 times, most recently
from
d0aabc5
to
e8e1c3f
Compare
majosm
force-pushed
the
disable-build-cache-pocl
branch
6 times, most recently
from
be06e3f
to
5cd5fd9
Compare
majosm
force-pushed
the
disable-build-cache-pocl
branch
from
5cd5fd9
to
2c53eb5
Compare
It turns out that AMD rocm does not appear to cache built kernels :-( (tested with rocm 5.7.1 and 6.0.3 on tioga). |
Fixes #731.