Loop skew&fusion with polly

[Nana CK]

Hi all,

I am dealing this loop with PLUTO and polly:

>>>>>>>>>>>>>>>>>>>>>>>>

for (i = 0; i < nx; i++){

  for (j = 0; j < ny; j++){

    if(i<ny-1)

      A[i+1][j] = A[i][j];

    if(j<ny-1)

      A[i][j+1] = A[i][j];

  }

}

>>>>>>>>>>>>>>>>>>>>>>>>>

in PLUTO, it can be transformed into this:

(part of the output of pluto :)

>>>>>>>>>>>>>>>>>>>>>>>>>

for (t1=2*ny-2;t1<=nx+ny-2;t1++) {

  lbp=max(1,t1-nx+1);

  ubp=ny-1;

#pragma omp parallel for private(lbv,ubv,t3)

  for (t2=lbp;t2<=ubp;t2++) {

    A[(t1-t2)][(t2-1)+1] = A[(t1-t2)][(t2-1)];;

  }

}

>>>>>>>>>>>>>>>>>>>>>>>>>>

as you can see with the "nx+ny-2"  boundary, PLUTO has skewed this loop.

I have tried with these options with polly:

>>>>>>>>>>>>>>>>>>>>>>>>>

clang -S -g -emit-llvm ../test1.cc -Xclang -disable-O0-optnone -o test1.ll

opt -S -polly-canonicalize test1.ll -o beforepolly.ll

opt -polly-allow-nonaffine -polly-opt-fusion=max -polly-tiling=false 

       -polly-detect -polly-scops -polly-simplify -polly-optree -polly-delicm -polly-simplify -polly-dependences -polly-opt-isl -polly-ast -polly-codegen beforepolly.ll

>>>>>>>>>>>>>>>>>>>>>>>>>

But I can not get the similar transform with polly. I checked the llvm IR dumped after this passes, the new loop build by polly has the same structure with the original code.

 Did I miss something ? like options or pragmas?

 

[Michael Kruse]

Polly uses ISL to optimize the loop aut applies whatever ISL
returns.Try the -polly-opt-outer-coincidence=yes option (which
translates to ISL's --schedule-outer-coincidence option) to give
higher priority to parallelism compared to locality.

Michael



Am Mi., 22. Sept. 2021 um 21:09 Uhr schrieb 'Nana CK' via Polly
Development <poll...@googlegroups.com>:

> --
> You received this message because you are subscribed to the Google Groups "Polly Development" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to polly-dev+...@googlegroups.com.
> To view this discussion on the web, visit https://groups.google.com/d/msgid/polly-dev/adaa625d-f59c-443a-8082-8bc76ea2b15cn%40googlegroups.com.



--
Tardyzentrismus verboten!

[Nana CK]

Hi Michael,

Thank you so much! It worked! The loop I posted before has been successfully fused!

But I got a new problem. It seems polly can not deal with more complex loops with just the option "the -polly-opt-outer-coincidence=yes". Is there other options need to be added?

Here is the loop I used, which can be fused to a new loop with the expected boundary "num_step+num_layer-2" by PLUTO.(and again, I can not get this fusion with polly)

Nana.

[Michael Kruse]

Am Fr., 24. Sept. 2021 um 04:54 Uhr schrieb 'Nana CK' via Polly
Development <poll...@googlegroups.com>:

> Here is the loop I used, which can be fused to a new loop with the expected boundary "num_step+num_layer-2" by PLUTO.(and again, I can not get this fusion with polly)

I don't see how you could fuse any loops to get a loop boundary of
num_step+num_layer-2.

If you want the hidden_size loop to be fused, ISL usually prefers
unrelated computations to be separated. There are a couple of related
options in ISL (such as --polly-isl-arg=--schedule-separate-components
and --polly-isl-arg=--schedule-serialize-sccs,
--polly-isl-arg=--schedule-whole-component), none of which directly
correspond to a min/max fusion option. Please refer to the ISL
documentation. There is no separate handling of fission/fusion in
Polly (other than the recently added llvm.loop.distribute.enable
metadata support).

Michael

--
Tardyzentrismus verboten!

[Nana CK]

OK !

Thanks a lot ! Have a good day !