Displaying 3 results from an estimated 3 matches for "non_divergence".
2015 Jan 24
2
[LLVMdev] Proposal: pragma for branch divergence
...iler to
automatically figure out which branches are divergent or not. However,
doing that precisely is extremely hard and can be expensive. Therefore, I
am proposing a compromise to have programmers provide some optimization
hints. The annotation in my mind is in the format of “#pragma clang branch
non_divergence”. Programmers can add this annotation right before a control
statement (such as if, for, and while), indicating the branch derived from
the control statement is not divergent.For example, #pragma clang branch
non_divergenceif (a > 0) { …}indicates the condition (a > 0) is uniform
across all...
2015 Jan 24
2
[LLVMdev] [cfe-dev] Proposal: pragma for branch divergence
...y figure out which branches are divergent or not. However,
> doing that precisely is extremely hard and can be expensive. Therefore, I
> am proposing a compromise to have programmers provide some optimization
> hints. The annotation in my mind is in the format of “#pragma clang branch
> non_divergence”. Programmers can add this annotation right before a control
> statement (such as if, for, and while), indicating the branch derived from
> the control statement is not divergent.For example, #pragma clang branch
> non_divergenceif (a > 0) { …}indicates the condition (a > 0) is unif...
2015 Jan 25
2
[LLVMdev] [cfe-dev] Proposal: pragma for branch divergence
...ch branches are divergent or not. However,
>> doing that precisely is extremely hard and can be expensive. Therefore, I
>> am proposing a compromise to have programmers provide some optimization
>> hints. The annotation in my mind is in the format of “#pragma clang branch
>> non_divergence”. Programmers can add this annotation right before a control
>> statement (such as if, for, and while), indicating the branch derived from
>> the control statement is not divergent.For example, #pragma clang branch
>> non_divergenceif (a > 0) { …}indicates the condition (a >...