[PATCH] D12246: [NVPTX] change threading intrinsics from noduplicate to convergent

[Mehdi Amini via llvm-commits]

> On Aug 31, 2015, at 9:49 PM, Mehdi Amini <mehdi.amini at apple.com> wrote:
> 
> 
>> On Aug 31, 2015, at 9:24 PM, Owen Anderson <resistor at mac.com <mailto:resistor at mac.com>> wrote:
>> 
>> 
>>> On Aug 31, 2015, at 5:57 PM, Mehdi Amini <mehdi.amini at apple.com <mailto:mehdi.amini at apple.com>> wrote:
>>> 
>>> 
>>> if (b) {
>>>     for (int i = 0; i < 100; ++i) {
>>>        // do something
>>>     __syncthreads();
>>>   }
>>> } else {
>>>    for (int i = 0; i < 100; ++i) {
>>>       __syncthreads();
>>>   }
>>> }
>>> 
>>> I think it would match your constraint, i.e. "they are be executed in under the same set of execution predicates as where the user wrote them”. 
>> 
>> I would not consider this to match the convergent constraint.  Each of the __syncthreads() calls after unswitching executes under a subset of the predicates under which the original executed.
> 
> Can you define what you mean by “predicates under which the original executed”?

I think I see what you mean: the syncthreads() call is made control dependent on b while it was not before. Your sentence made me think about something quite different.

— 
Mehdi


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150831/31cfccac/attachment.html>