[Patch] Linpack-pc broken as a performance benchmark

[Hal Finkel]

----- Original Message -----
> Hi all,
> 
> SingleSource/Benchmark/Linpack is broken as a performance benchmark.
> It uses the time to execute a set of functions to determine the time
> to run another set of functions.
> 
>         {
>             time1 = second();
>             pass = pass + 1;
>             for ( i = 0 ; i < ntimes ; i++)
>             {
>                 matgen(a,lda,n,b,&norma);
>                 dgefa(a,lda,n,ipvt,&info );
>             }
>             time2 = second() - time1;
>             fprintf (stderr,"%10d times %6.2f seconds\n", ntimes,
>             0.0);
>             if (0 /*time2 > 5.0*/)
>             {
>                 pass = 0;
>             }
>             if (pass < 0)
>             {
>                 if (0 /*time2 < 0.1*/)
>                 {
>                     ntimes = ntimes * 10;
>                 }
>                 else
>                 {
>                     ntimes = ntimes * 2;
>                 }
>             }
>         }
>         while (pass < 0);
>         
>         ntimes =  5.0 * (double)ntimes / time2; // <<< BAD
> 
> 
> Later it runs:
> 
>         for (j=1 ; j<6 ; j++)
>         {
>         
>             t1 = second();
> 
>             for (i = 0; i < ntimes; i++) <<< !!!!
>             {
>                 matgen(a,lda,n,b,&norma);
>                 dgefa(a,lda,n,ipvt,&info );
>             }
> 
>             atime[0][j] = (second() - t1 - tm2)/ntimes;
> 
>             t1 = second();
>         
>             for (i = 0; i < ntimes; i++) <<< !!!!
>             {
>                 dgesl(a,lda,n,ipvt,b,0);
>             }
> 
>             atime[1][j] = (second() - t1)/ntimes;
>             total       = atime[0][j] + atime[1][j];
>             atime[2][j] = total;
>             atime[3][j] = ops/(1.0e6*total);
>             atime[4][j] = 2.0/atime[3][j];
>             atime[5][j] = total/cray;
>             atime[3][6] = atime[3][6] + atime[3][j];
>             
>             print_time(j);
>         }
> 
> 
> The attached patch fixes this by setting “ntimes” to arbitrary 10000.
> The benchmark then runs in about 30 secs on a recent mac.

30 seconds sounds good to me. However, some of the embedded folks may complain ;) -- We should probably have it depend on whether SMALL_PROBLEM_SIZE is in effect or not. (You can add a preprocessor define in the makefile when SMALL_PROBLEM_SIZE is defined, and let that set it to be a factor of 10 smaller).

 -Hal

> 
> This is going to cause a one time break in the performance numbers
> for this benchmark. But after that it can really be considered a
> “performance” benchmark.
> 
> Okay to commit?
> 
> Thanks,
> Arnold
> 
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory