[PATCH] Divergence analysis for GPU programs

[Jingyue Wu]

Hi Justin,

Would you mind checking the NVPTX part? I got LGTM from Fernando and Diogo
for the algorithm.

Thanks,
Jingyue

On Wed, Apr 8, 2015 at 9:35 PM Jingyue Wu <jingyue at google.com> wrote:

> Thanks for your review, Diogo! As you suggested, the new patch (
> http://reviews.llvm.org/D8576) now considers laneid.
>
> Jingyue
>
> On Wed, Apr 8, 2015 at 3:31 PM Diogo Sampaio <dnsampaio at gmail.com> wrote:
>
>> Hi Jingyue,
>>
>> Very nice code and good comments. It's nice to see this analysis in
>> llvm and I believe your code is correct. There is one slight point,
>> not very common, which I faced. The use the intrinsic laneId, most of
>> the times generated by nvcc, also leads to divergence. I would change
>> the file NVPTXTargetTransformInfo.cpp either by adding
>> Intrinsic::int_ptx_read_laneid to the switch in "static bool
>> readsThreadIndex(const IntrinsicInst *II)" or just creating a method
>> for it such as...
>>
>> static bool readsLaneId(const IntrinsicInst *II) {
>>   return (II->getIntrinsicID() == Intrinsic::int_ptx_read_laneid);
>> }
>> ..... and changing the test to ...
>>     if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
>>       // Instructions that read threadIdx or laneId are obviously
>> divergent.
>>       if (readsThreadIndex(II) || readsLaneId(II))
>>         return true;
>> .
>>
>> Now, more as an discussion:
>> In the file DivergenceAnalysis.cpp, line 206
>> if (!cast<PHINode>(I)->hasConstantValue() && DV.insert(I).second)
>>       Worklist.push_back(I);
>>
>> I was wondering about the effectiveness of the method
>> PhiNode::hasConstantValue.
>> For example, in a code such as:
>> //not a smart code, I know
>> if(divergentBool){
>> a0 = b1 - c1;
>> }else{
>> c2 = -c1;
>> a1 = b1 + c2;
>> }
>> a2 = phi(a0, a1);
>>
>>
> Given the code as is, DivergenceAnalysis will consider a2 divergent.
>
> However, I wouldn't worry about it. We plan to run this analysis pass at a
> relatively late stage (definitely after InstCombine), so the code pattern
> you listed here is likely optimized away. In fact, even a PHINode with same
> incoming values is probably optimized away.
>
>
>> Will this method be good enough to say that a2 is not divergent? I
>> believe that llvm's ScalarEvolution could detect they are the same
>> value.
>>
>> And to end, there are 2 typos in your header file:
>> exeuction
>> pogram
>>
>>
>> Very nice work,
>> regards
>> --
>> Diogo Nunes Sampaio
>>
>> Portable: +33 6 59 05 75 57
>>
>> INRIA - Institut national de recherche en informatique et en automatique
>> CORSE - http://www.inria.fr/equipes/corse
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150409/11c938d8/attachment.html>