[LLVMdev] [FastPolly]: Update of Polly's performance on LLVM test-suite

[Tobias Grosser]

On 08/11/2013 05:33 PM, Star Tan wrote:
> At 2013-08-12 01:18:30,"Tobias Grosser" <tobias at grosser.es> wrote:
>> On 08/10/2013 06:59 PM, Star Tan wrote:

>
>>>   Overall, there are 198 benchmarks improved and 16 benchmarks regressed. Especially, with those recent performance-oriented patch files for ScopDetect/ScopInfo/ScopDependences/..., we have significantly reduced the compile-time overhead of Polly for a large number of benchmarks, such as:
>>>        SingleSource/Benchmarks/Misc/salsa20        -97.84%
>>>        SingleSource/Benchmarks/Polybench/linear-algebra/solvers/lu/lu        -85.01%	
>>>        MultiSource/Applications/obsequi/Obsequi        -57.12%
>>>        SingleSource/Benchmarks/Polybench/stencils/seidel-2d/seidel-2d        -50.00%
>>>        MultiSource/Benchmarks/MiBench/telecomm-gsm/telecomm-gsm        -40.09%
>>>        MultiSource/Benchmarks/mediabench/gsm/toast/toast       -39.91%
>>>        SingleSource/Benchmarks/Misc/whetstone       -39.02%
>>>        MultiSource/Benchmarks/mediabench/jpeg/jpeg-6a/cjpeg       -38.07%
>>>        MultiSource/Benchmarks/MiBench/consumer-jpeg/consumer-jpeg       -37.70%
>>
>> Very nice work!
>>
>>>   However, Polly can still lead to significant compile-time overhead for many benchmarks.
>>>   As shown on:
>>>        http://188.40.87.11:8000/db_default/v4/nts/31?compare_to=28&baseline=28
>>>   there are 11 benchmarks whose compile time are more than 2x than clang.  Furthermore, it seems that PollyDependence pass is still one of most expensive passes in Polly.
>>
>> We need to look at these on a case by case base. 2x compile time
>> increase for large programs, where Polly is just run on small parts is
>> not what we want. However, for small micro kernels (e.g. Polybench)
>> where we can significantly increase the performance of the generated
>> code, this is in fact a good baseline - especially as we did not spend
>> too much time optimising this.
>
> Yes, we should look into the compile-execution performance trade-off.
> I have summarized some benchmarks (compile-time overhead is more than 200%) as follows:
>
> SingleSource/Benchmarks/Shootout/nestedloop,
> 	compile_time(+6355.56%), execution_time(-99.21%)
> SingleSource/Benchmarks/Shootout-C++/nestedloop,
> 	compile_time(+1155.56%), execution_time(-99.23%)

This is a huge overhead on a single kernel. We should analyse how we can 
do better.

> SingleSource/Benchmarks/Polybench/linear-algebra/kernels/2mm/2mm,
> 	compile_time(+278.95%), execution_time(0%)
> SingleSource/Benchmarks/Polybench/linear-algebra/kernels/3mm/3mm,
> 	compile_time(+270.73%), execution_time(0%)
> SingleSource/Benchmarks/Polybench/linear-algebra/kernels/syrk/syrk,
> 	compile_time(+208.57%), execution_time(0%)
> SingleSource/Benchmarks/Polybench/linear-algebra/kernels/gemm/gemm,
> 	compile_time(+202.63%), execution_time(0%)
> SingleSource/Benchmarks/Polybench/stencils/seidel-2d/seidel-2d,
> 	compile_time(+1275.00%), execution_time (0%)

The Polybench kernels as committed to the test suite are currently not
properly detected by Polly. The reason is that there are different pre 
processor flags that can be used to compile polybench and for a certain 
set it is more challenging to optimize the kernels. We should probably 
first set this to use both fixed sized arrays as well as non-parametric 
loop iterators.

> MultiSource/Benchmarks/ASC_Sequoia/AMGmk/AMGmk,
> 	compile_time(+491.80%), execution_time (0%)
> SingleSource/UnitTests/Vector/multiplies,
> 	compile_time(+350.00%), execution_time(-13.64%)
> SingleSource/Benchmarks/Stanford/Puzzle,
> 	compile_time(+345.45%), execution_time(-40.91%)
> SingleSource/Regression/C/test_indvars,
> 	compile_time(+200.00%), execution_time(0%)
>
> Results show that some Polly leads to significant compile-time overhead without any execution performance improvement.
> I have reported a bug for nestedloop (http://llvm.org/bugs/show_bug.cgi?id=16843), and I would reported other bugs for those benchmarks whose compile time is significantly increased but without execution performance improvement.

Great. We can than look at them one at a time.

> Furthermore, you can view top 10 compiler passes when compiling with Polly as follows:
> https://gist.github.com/tanstar/581bcea1e4e03498f935/raw/f6a4ec4e8565f7a7bbdb924cd59fcf145caac039/Polly-top10

Interesting. This will be very helpful for further performance analysis.

>>> Even without optimization and code generation, Polly also increases the compile time for some benchmarks.
>>>   As shown on:
>>>        http://188.40.87.11:8000/db_default/v4/nts/29?compare_to=28&baseline=28
>>>   there are 10 benchmarks that require more than 10% extra compile-time overhead compared with clang.
>>
>> Having to pay at most 10% slowdown to decide if Polly should be run
>> (including all the canonicalization) is actually not bad. Especially as
>> the average on normal programs is probably a lot less.
>>
>> Still, if we should have a look into why this is happening for some of
>> the biggest slowdowns.
>>
>> Can you ping me when the server is up again. I would like to see which
>> kernels are slowed down most.
>
> The server is up now.
>
> For you information, you can view top 10 compiler passes when compiled with "polly without optimization and code generation" as follows:
>      https://gist.github.com/tanstar/40ece0e4e2bf9d052ca0/raw/9e892fe50544acca9609004941d4b3d4921cb302/Polly-NoGenOpt-top10
>
> I have checked some benchmarks. It seems that the extra compile-time overhead is mainly resutled by the following Polly passes:
>      Polly - Create polyhedral description of Scops
>      Combine redundant instructions
>      Polly - Detect static control parts (SCoPs)

If there are still cases where the scop detection is slow, we may want 
to look into them.

>      Induction Variable Simplification (Polly version)

Otherwise, I propose to look into this based on the benchmarks with the 
highest overhead.

Tobi