Self test worksizes #5825
Self test worksizes #5825
Conversation
magnumripper
force-pushed
the
self-test-worksizes
branch
2 times, most recently
from
ccf5acb to
ba4cebc
Compare
|
So a build bot choked on office-opencl hanging (10 minutes no output) on its Intel CPU device. I'm not quite sure what to do with that. Edit: I used the same CPU-specific logic that Claudio added nearby. |
magnumripper
force-pushed
the
self-test-worksizes
branch
3 times, most recently
from
7e93ffb to
ff382db
Compare
|
Found two format bugs so far because of this change. Fixes are in this PR but as separate commits. |
Also with a GWS of 2x that LWS. The new figures are better at triggering bugs. If a kernel needs a lower LWS than device max. our code already handles that. We previously had it as LWS=7 GWS=49 for speed (and for checking heuristics that could bug out on non-log2 values) but that was introduced before our autotune was sped up with orders of magnitude and those heuristics has been stable for many years. Like before, the self-test will obey any given lws/gws options or environment variables (in that case even when they are past limits). Closes openwall#5822
Bug found after bumping work size during self-test.
Bug found after bumping work size during self-test.
magnumripper
force-pushed
the
self-test-worksizes
branch
from
92ec343 to
eb7682f
Compare
|
Ready for merge. Tested a lot, works fine and finds bugs. The only drawback is a tad slower self-test but it's only significant for things like |
| if (self_test_running) | ||
| global_work_size = local_work_size; | ||
| else if (!global_work_size) | ||
| global_work_size = 12 * local_work_size; |
There was a problem hiding this comment.
This looks like it'd sometimes affect more than just self-test. Could be worth mentioning in the commit message?
There was a problem hiding this comment.
I'm not sure what this code affects (not shared autotune afaik) so wouldn't know what to write. Anyway I don't think it's worth mentioning, looks like some default start value or worst case fallback.
2 participants
Self-test with default LWS at device maximum
Also with a default GWS of 2x that LWS. The new figures are better at triggering bugs. If a kernel needs a lower LWS than device max. our code already handles that.
We previously had it as LWS=7 GWS=49 for speed and for checking heuristics that could bug out on non-log2 values but that was introduced before our autotune was sped up with orders of magnitude (and the heuristics has been stable for many years).
Like before, the self-test will obey any given lws/gws options or environment variables.
Closes #5822