llvm dev - Jan 2017 - (RFC) Adjusting default loop fully unroll threshold

On Mon, Jan 30, 2017 at 3:51 PM Mehdi Amini via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> On Jan 30, 2017, at 10:49 AM, Dehao Chen via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
> Currently, loop fully unroller shares the same default threshold as loop
> dynamic unroller and partial unroller. This seems conservative because
> unlike dynamic/partial unrolling, fully unrolling will not affect
> LSD/ICache performance. In https://reviews.llvm.org/D28368, I proposed to
> double the threshold for loop fully unroller. This will change the codegen
> of several SPECCPU benchmarks:
>
> Code size:
> 447.dealII 0.50%
> 453.povray 0.42%
> 433.milc 0.20%
> 445.gobmk 0.32%
> 403.gcc 0.05%
> 464.h264ref 3.62%
>
> Compile Time:
> 447.dealII 0.22%
> 453.povray -0.16%
> 433.milc 0.09%
> 445.gobmk -2.43%
> 403.gcc 0.06%
> 464.h264ref 3.21%
>
> Performance (on intel sandybridge):
> 447.dealII +0.07%
> 453.povray +1.79%
> 433.milc +1.02%
> 445.gobmk +0.56%
> 403.gcc -0.16%
> 464.h264ref -0.41%
>
>
> Can you clarify how to read these numbers? (I’m using +xx% to indicates a
> slowdown usually, it seems you’re doing the opposite?).
>
> So considering 464.h264ref, does it mean it is 3.6% slower to compile,
> gets 3.2% larger, and 0.4% slower?
>
> Another question is about PGO integration: is it already hooked there?
> Should we have a more aggressive threshold in a hot function? (Assuming
> we’re willing to spend some binary size there but not on the cold path).
>
I would even wire the *unrolling* the other way: just suppress unrolling in
cold paths to save binary size. rolled loops seem like a generally good
thing in cold code unless they are having some larger impact (IE, the loop
itself is more expensive than the unrolled form).

>
> Thanks,
>
> —
> Mehdi
>
>
> Looks like the change has overall positive performance impact with very
> small code size/compile time overhead. Now the question is shall we make
> this change default in O2, or shall we leave it in O3. We would like to
> have more input from the community to make the decision.
>
> Thanks,
>
> Dehao
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170130/87a2c6f4/attachment.html>

On Mon, Jan 30, 2017 at 3:56 PM, Chandler Carruth <chandlerc at
google.com>
wrote:
> On Mon, Jan 30, 2017 at 3:51 PM Mehdi Amini via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
>> On Jan 30, 2017, at 10:49 AM, Dehao Chen via llvm-dev <
>> llvm-dev at lists.llvm.org> wrote:
>>
>> Currently, loop fully unroller shares the same default threshold as
loop
>> dynamic unroller and partial unroller. This seems conservative because
>> unlike dynamic/partial unrolling, fully unrolling will not affect
>> LSD/ICache performance. In https://reviews.llvm.org/D28368, I proposed
>> to double the threshold for loop fully unroller. This will change the
>> codegen of several SPECCPU benchmarks:
>>
>> Code size:
>> 447.dealII 0.50%
>> 453.povray 0.42%
>> 433.milc 0.20%
>> 445.gobmk 0.32%
>> 403.gcc 0.05%
>> 464.h264ref 3.62%
>>
>> Compile Time:
>> 447.dealII 0.22%
>> 453.povray -0.16%
>> 433.milc 0.09%
>> 445.gobmk -2.43%
>> 403.gcc 0.06%
>> 464.h264ref 3.21%
>>
>> Performance (on intel sandybridge):
>> 447.dealII +0.07%
>> 453.povray +1.79%
>> 433.milc +1.02%
>> 445.gobmk +0.56%
>> 403.gcc -0.16%
>> 464.h264ref -0.41%
>>
>>
>> Can you clarify how to read these numbers? (I’m using +xx% to indicates
a
>> slowdown usually, it seems you’re doing the opposite?).
>>
>
As this is comparing spec scores instead of run time, +xx% here means
speedup, -xx% means slowdown.

>
>> So considering 464.h264ref, does it mean it is 3.6% slower to compile,
>> gets 3.2% larger, and 0.4% slower?
>>
>
That is correct. The 0.4% slowdown is in the run-to-run noise range.

>
>> Another question is about PGO integration: is it already hooked there?
>> Should we have a more aggressive threshold in a hot function? (Assuming
>> we’re willing to spend some binary size there but not on the cold
path).
>>
>
> I would even wire the *unrolling* the other way: just suppress unrolling
> in cold paths to save binary size. rolled loops seem like a generally good
> thing in cold code unless they are having some larger impact (IE, the loop
> itself is more expensive than the unrolled form).
>

Agree that we could suppress unrolling in cold path to save code size. But
that's orthogonal with the propose here. This proposal focuses on O2
performance: shall we have different (higher) fully unroll threshold than
dynamic/partial unroll.

We can have a separate patch to further boost threshold for hot loops and
suppress unrolling for cold loops. One concern is that in order to check if
a loop is hot/cold, we will need BFI for the loop pass. In the legacy loop
pass manager, this will insert a function pass in the middle of a series of
loop passes.

Dehao

>
>
>>
>> Thanks,
>>
>> —
>> Mehdi
>>
>>
>> Looks like the change has overall positive performance impact with very
>> small code size/compile time overhead. Now the question is shall we
make
>> this change default in O2, or shall we leave it in O3. We would like to
>> have more input from the community to make the decision.
>>
>> Thanks,
>>
>> Dehao
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170130/e44929a0/attachment.html>

> On Jan 30, 2017, at 4:56 PM, Dehao Chen <dehao at google.com> wrote:
> 
> 
> 
> On Mon, Jan 30, 2017 at 3:56 PM, Chandler Carruth <chandlerc at
google.com <mailto:chandlerc at google.com>> wrote:
> On Mon, Jan 30, 2017 at 3:51 PM Mehdi Amini via llvm-dev <llvm-dev at
lists.llvm.org <mailto:llvm-dev at lists.llvm.org>> wrote:
>> On Jan 30, 2017, at 10:49 AM, Dehao Chen via llvm-dev <llvm-dev at
lists.llvm.org <mailto:llvm-dev at lists.llvm.org>> wrote:
>> 
>> Currently, loop fully unroller shares the same default threshold as
loop dynamic unroller and partial unroller. This seems conservative because
unlike dynamic/partial unrolling, fully unrolling will not affect LSD/ICache
performance. In https://reviews.llvm.org/D28368
<https://reviews.llvm.org/D28368>, I proposed to double the threshold for
loop fully unroller. This will change the codegen of several SPECCPU benchmarks:
>> 
>> Code size:
>> 447.dealII 0.50%
>> 453.povray 0.42%
>> 433.milc 0.20%
>> 445.gobmk 0.32%
>> 403.gcc 0.05%
>> 464.h264ref 3.62%
>> 
>> Compile Time:
>> 447.dealII 0.22%
>> 453.povray -0.16%
>> 433.milc 0.09%
>> 445.gobmk -2.43%
>> 403.gcc 0.06%
>> 464.h264ref 3.21%
>> 
>> Performance (on intel sandybridge):
>> 447.dealII +0.07%
>> 453.povray +1.79%
>> 433.milc +1.02%
>> 445.gobmk +0.56%
>> 403.gcc -0.16%
>> 464.h264ref -0.41%
>> 
> 
> 
> Can you clarify how to read these numbers? (I’m using +xx% to indicates a
slowdown usually, it seems you’re doing the opposite?).
> 
> As this is comparing spec scores instead of run time, +xx% here means
speedup, -xx% means slowdown.
>  
> 
> So considering 464.h264ref, does it mean it is 3.6% slower to compile, gets
3.2% larger, and 0.4% slower?
> 
> That is correct. The 0.4% slowdown is in the run-to-run noise range.
Ok, thanks for the clarifications.
What about the noise on the improvements? How reliable are you other numbers on
this aspect?
>  
> 
> Another question is about PGO integration: is it already hooked there?
Should we have a more aggressive threshold in a hot function? (Assuming we’re
willing to spend some binary size there but not on the cold path).
> 
> I would even wire the *unrolling* the other way: just suppress unrolling in
cold paths to save binary size. rolled loops seem like a generally good thing in
cold code unless they are having some larger impact (IE, the loop itself is more
expensive than the unrolled form).
> 
> 
> Agree that we could suppress unrolling in cold path to save code size. But
that's orthogonal with the propose here. This proposal focuses on O2
performance: shall we have different (higher) fully unroll threshold than
dynamic/partial unroll.
I agree that this is (to some extent) orthogonal, and it makes sense to me to
differentiate the threshold for full unroll and the dynamic/partial case.

Thanks,

— 
Mehdi

> We can have a separate patch to further boost threshold for hot loops and
suppress unrolling for cold loops. One concern is that in order to check if a
loop is hot/cold, we will need BFI for the loop pass. In the legacy loop pass
manager, this will insert a function pass in the middle of a series of loop
passes.
> 
> Dehao
>  
>  
> 
> Thanks,
> 
> — 
> Mehdi
> 
> 
>> Looks like the change has overall positive performance impact with very
small code size/compile time overhead. Now the question is shall we make this
change default in O2, or shall we leave it in O3. We would like to have more
input from the community to make the decision.
>> 
>> Thanks,
>> 
>> Dehao
>> 
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
<http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
<http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170130/7b96c055/attachment-0001.html>