[graphite] Remove checks for flag_loop_strip_mine, flag_loop_block, flag_loop_interchange in graphite_trans_bb_block.

Jack Howarth howarth@bromo.msbb.uc.edu
Fri Aug 15 14:26:00 GMT 2008


Tobias,
     I am seeing the following results with the proposed patch from...

http://gcc.gnu.org/ml/gcc-patches/2008-08/msg01000.html

applied over gcc trunk with the previous graphite patch and all of
the subsequent changes in graphite branch applied...

================================================================================
Date & Time     : 15 Aug 2008  1:33:29
Test Name       : gfortran_lin_p4_graphite123
Compile Command : gfortran -ffast-math -funroll-loops -O3 -floop-block -floop-interchange -floop-strip-mine -fgraphite %n.f90 -o %n
Benchmarks      : ac aermod air capacita channel doduc fatigue gas_dyn induct linpk mdbx nf protein rnflow test_fpu tfft
Maximum Times   :     2000.0
Target Error %  :      0.100
Minimum Repeats :    10
Maximum Repeats :   100

   Benchmark   Compile  Executable   Ave Run  Number   Estim
        Name    (secs)     (bytes)    (secs) Repeats   Err %
   ---------   -------  ----------   ------- -------  ------
          ac      5.17       10000     15.16      10  0.0275
      aermod     26.97           0     -1.00      10  0.0275
         air      4.91       10000      9.14      10  0.0699
    capacita      3.58       10000     56.19      17  0.0885
     channel      1.63       10000      3.79      10  0.0898
       doduc     11.31       10000     48.37      10  0.0342
     fatigue      5.80       10000     13.05      10  0.0421
     gas_dyn      5.35       10000     11.10      35  0.0778
      induct      9.24       10000     36.33      10  0.0151
       linpk      1.78       10000     26.37      10  0.0286
        mdbx      3.45       10000     15.65      10  0.0432
          nf      4.99       10000     30.53      18  0.0950
     protein     11.55       10000     51.00      10  0.0634
      rnflow     12.10       10000     42.72      10  0.0565
    test_fpu     11.19       10000     14.14      12  0.0962
        tfft      1.00       10000      2.86      13  0.0983

Geometric Mean Execution Time =      23.45 seconds

================================================================================
================================================================================
Date & Time     : 15 Aug 2008  2:54:19
Test Name       : gfortran_lin_p4_graphite
Compile Command : gfortran -ffast-math -funroll-loops -O3 -fgraphite %n.f90 -o %n
Benchmarks      : ac aermod air capacita channel doduc fatigue gas_dyn induct linpk mdbx nf protein rnflow test_fpu tfft
Maximum Times   :     2000.0
Target Error %  :      0.100
Minimum Repeats :    10
Maximum Repeats :   100

   Benchmark   Compile  Executable   Ave Run  Number   Estim
        Name    (secs)     (bytes)    (secs) Repeats   Err %
   ---------   -------  ----------   ------- -------  ------
          ac      2.17       10000     15.16      12  0.0286
      aermod     26.31           0     -1.00      12  0.0286
         air      4.87       10000      9.17      12  0.0910
    capacita      2.93       10000     56.51      18  0.0903
     channel      1.38       10000      3.79      10  0.0541
       doduc     11.29       10000     48.38      14  0.0983
     fatigue      4.69       10000     13.02      10  0.0552
     gas_dyn      4.71       10000     11.23      26  0.0765
      induct      8.24       10000     36.33      10  0.0201
       linpk      1.45       10000     26.40      12  0.0322
        mdbx      3.08       10000     15.66      10  0.0936
          nf      4.03       10000     30.49      10  0.0832
     protein      9.85       10000     51.04      10  0.0640
      rnflow     10.74       10000     42.70      10  0.0625
    test_fpu      9.99       10000     14.16      12  0.0739
        tfft      0.95       10000      2.86      13  0.0784

Geometric Mean Execution Time =      23.47 seconds

================================================================================

On i686-apple-darwin9. I am surprised that I'm not seeing any
major performance changes one way or another with the new
loop optimizations. Does this match what you are seeing from
the Polyhedron 2005 benchmarks?
                              Jack



More information about the Gcc-patches mailing list