Delite benchmarking

[Matthias Klass]

Hi,

I created some benchmarks to measure Delite's performance. To do so, I run a delite app using the passes attribute. Now I am wondering on why the first run produces results which are an order of 100 slower than the following results. Have you got any explanation for this? In the optiml spec I read that there is such thing as Common subexpression elimination, which eliminates statements that are executed multiple times and reuses the result. Does this happen here? An execution time of 0,000021s for a 500x500 matrix multiplication is ... well ... unrealistic....

Thanks,

Matthias

[Arvind Sujeeth]

hey Matthias,

Usually we see significant speedups on later runs when the JIT warms up,
but the number you're reporting looks like nothing is being computed at
all. Are you doing something with the result (like printing an element)
to ensure that the computation does not get dead code eliminated? Also,
is this time just the application time reported by Delite, or something
else? And by the 'passes' attribute, do you mean the -r option in the
delite script?

cheers,
Arvind

> --
> You received this message because you are subscribed to the Google
> Groups "OptiML" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to optiml+un...@googlegroups.com.
> For more options, visit https://groups.google.com/groups/opt_out.
>
>

[Matthias Klass]

Hi,

yes, this is about the -r option, which is afterwards mapped to the delite.runs system property. In the meantime I found out that it is in deed about static code optimization. When passing something non predictable to my compute function, I don't get those effects. However, I am still wondering about the passes attribute, as I thought that this would not be influenced by any code optimization ...

Matthias