Benchmark question

[chuck wu]

Hello everyone,
I am trying to find a benchmark similar to PolyBench that can be compiled by PPCG but has multiple SCoPs  in a program.
Ex:
{
#pragma scop
scop1...
#pragma endscop

maybe some function calls or others can break the SCoP property

#pragma scop
scop2...
#pragma endscop
}

Are there any suggestions for benchmark or opinion for this request?

Chuck

[chuc...@gmail.com]

Hello every one,

I am trying to find a benchmark similar to PolyBench that can be compiled by PPCG but has multiple SCoPs in a program.
Ex:
{
#pragma scop
scop1...
#pragma endscop

...

function calls or others can break the SCoP property

...

#pragma scop
scop2...
#pragma endscop
}

Are there any suggestions for benchmark or opinions on this request?

Chuck

[Sven Verdoolaege]

I'm not aware of any.

It may help if you explain what you would need it for.

skimo

[Dmitry N. Mikushin]

Hi, Chunk,

In principle karpathy/llm.c and ggerganov/whisper.cpp should serve well as "benchmarks" of the described kind. There, you can find self-contained C/C++ source, where SCoP-able tensor functions are interleaved with possibly non-SCoP logic. The code is not prepared for PPCG, but worth doing. I would definitely try to work out a benchmark out of them at my earliest research opportunity.

Hope this helps,

- Dmitry.

пн, 15 апр. 2024 г. в 21:26, Sven Verdoolaege <sven.ver...@telenet.be>:

--

---
You received this message because you are subscribed to the Google Groups "isl Development" group.
To unsubscribe from this group and stop receiving emails from it, send an email to isl-developme...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/isl-development/20240415192605.GW23566MdfPADPa%40purples.home.

[chuc...@gmail.com]

> It may help if you explain what you would need it for.

I am working on finding opportunity of merging SCoPs to get more optimization space.
So, I want to find some benchmarks which have this scenario.

[chuc...@gmail.com]

> In principle karpathy/llm.c and ggerganov/whisper.cpp should serve well as "benchmarks" of the described kind. There, you can find self-contained C/C++ source, where SCoP-able tensor functions are interleaved with possibly non-SCoP logic. The code is not prepared for PPCG, but worth doing. I would definitely try to work out a benchmark out of them at my earliest research opportunity.

Thank you very much for the advice, I will give it a look.

Chuck

maem...@gmail.com 在 2024年4月16日 星期二凌晨4:27:23 [UTC+8] 的信中寫道：