[test-suite] r241675 - [LNT] Reduce I/O execution time for Polybench

Tobias Grosser tobias at grosser.es
Wed Jul 8 16:10:14 PDT 2015


On 07/08/2015 12:21 PM, Renato Golin wrote:
> Author: rengolin
> Date: Wed Jul  8 05:21:42 2015
> New Revision: 241675
>
> URL: http://llvm.org/viewvc/llvm-project?rev=241675&view=rev
> Log:
> [LNT] Reduce I/O execution time for Polybench
>
> Polybench had large execution time due to the successive call to
> fprintf as much as 4000*4000 times. For most programs, this was more
> than 1/2 of its execution time.
>
> The current solution is to transform the values into a stream of
> nibbles as a char string, and print it once for every row, ie.
> only as much as 4000 times, by using fputs instead of fprintf.
>
> Overall new execution time is 47% of previous with some as low as 5%.
> The reduction on x86_64 was 53%, on ARM was 51% and on AArch64 was 55%,
> which means most of the time was spent on I/O, not the actual benchmark.
>
> Test fdtd-apml was not converted because it's not bit identical across
> platforms, but it should be. Adding a FIXME. Other tests were not
> converted due to regressing in run time for being too simple/small even
> on our slowest targets.

Very nice. This change makes Polly really shine on polybench:

http://llvm.org/perf/db_default/v4/nts/29988?compare_to=29985

Best,
Tobias





More information about the llvm-commits mailing list