RFR (L): 7153771: array bound check elimination for c1

Roland Westrelin roland.westrelin at oracle.com
Mon Feb 25 07:38:17 PST 2013


> The research paper said about 55% improvement for scimark.LU. You got 18.7%. There is big effect of cpu cache size on scimark results. Could you rerun scimark with small matrix size? :
> 
> benchscripts/scimark/run
> <   scimark_args="-large"
> ---
> >   scimark_args=""


Below are small matrix scimark results.

Roland.

============================================================================
tscimark-base
  Benchmark         Samples        Mean     Stdev
  scimark                10      585.28      0.00
    Sparse               10      584.47      0.00
    LU                   10      857.41      0.00
    SOR                  10      899.27      0.00
    FFT                  10      454.81      0.00
    Monte                10      130.43      0.00
============================================================================
tscimark-abce
  Benchmark         Samples        Mean     Stdev   %Diff    P   Significant
  scimark                10      736.83      0.00   25.89 0.000          Yes
    Sparse               10      735.49      0.00   25.84 0.000          Yes
    LU                   10     1390.61      0.00   62.19 0.000          Yes
    SOR                  10      897.26      0.00   -0.22 0.000          Yes
    FFT                  10      529.91      0.00   16.51 0.000          Yes
    Monte                10      130.88      0.00    0.35 0.000          Yes
============================================================================




More information about the hotspot-compiler-dev mailing list