Remove usages of std::call_once.
#5599
Conversation
std::call_once has a bug in some versions of glibc. A custom implementation has been provided. Usages with static once_flag have been replaced with "magic statics". Optical flow doesn't need atomicity and the once semantics were not required at all. Other usages use the custom implementation. A stale warning about antialiasing (also guarded with call once) was removed. Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
mzient
force-pushed
the
remove_call_once
branch
from
5e81b2f
to
010a519
Compare
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
| if ((has_min && is_linear_default(make_cspan(min_arg_))) || | ||
| (has_interp && is_linear_default(make_cspan(interp_type_arg_)))) { | ||
| static std::once_flag linear_default_warning_flag; | ||
| std::call_once(linear_default_warning_flag, [&]() { | ||
| DALI_WARN( | ||
| "The default behavior for LINEAR interpolation type has been changed to apply an " | ||
| "antialiasing filter. If you didn't mean to apply an antialiasing filter, please use " | ||
| "`antialias=False`"); | ||
| }); | ||
| } |
There was a problem hiding this comment.
This warning was sitting here for 2 years now. Nobody complained during that time and the behavior is hardly a "change" any more.
There was a problem hiding this comment.
I have mixed filling about including this in this particular PR. But up to you.
| std::call_once(of_initialized_, | ||
| [&]() { | ||
| optical_flow_.reset( | ||
| new optical_flow::OpticalFlowImpl(of_params_, | ||
| width, | ||
| height, | ||
| channels, | ||
| image_type, | ||
| device_id, | ||
| stream)); | ||
| optical_flow_->Init(of_params_); | ||
| }); |
There was a problem hiding this comment.
This is called from within setup. No atomicity is required, setup is never called in parallel with itself.
|
CI MESSAGE: [17296340]: BUILD STARTED |
|
CI MESSAGE: [17296340]: BUILD FAILED |
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
|
CI MESSAGE: [17298715]: BUILD STARTED |
|
CI MESSAGE: [17298715]: BUILD FAILED |
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
|
CI MESSAGE: [17313865]: BUILD STARTED |
Signed-off-by: Michał Zientkiewicz <mzient@gmail.com>
|
CI MESSAGE: [17317992]: BUILD STARTED |
|
CI MESSAGE: [17321143]: BUILD STARTED |
|
CI MESSAGE: [17317992]: BUILD PASSED |
|
Please excuse my naïve question, as I understand statics are used when the code doesn't throw, call_once where it might, right? |
I use "magic statics" when the call_once flag was also static. The compiler should take care of everything. I can double check it. |
|
Checked: magic statics do the trick on |
Category:
Other (e.g. Documentation, Tests, Configuration)
Workaround for a library bug.
Description:
Additional information:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78184
std::call_oncehas a bug in some versions of glibc. A custom implementation has been provided.Usages with static once_flag have been replaced with "magic statics". Optical flow doesn't need atomicity and the once semantics were not required at all.
Other usages use the custom implementation.
A stale warning about antialiasing (also guarded with call once) was removed.
Affected modules and functionalities:
Key points relevant for the review:
Tests:
Checklist
Documentation
DALI team only
Requirements
REQ IDs: N/A
JIRA TASK: N/A