[Bug c++/69564] [5/6 Regression] lto and/or C++ make scimark2 LU slower
jakub at gcc dot gnu.org
gcc-bugzilla@gcc.gnu.org
Thu Feb 25 12:41:00 GMT 2016
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69564
--- Comment #12 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I don't see any difference though, neither with the fold-const.c change, nor
with the loop-invert.patch (at least on my Haswell-E, -g -Ofast x86_64, single
runs only; though, it shows the LU slowdown with C++ clearly, and also that
clang wins on SOR and Sparse matmult, we win significantly on MonteCarlo and
less significantly on FFT, C LU is comparable):
gcc trunk 20160224
Composite Score: 2482.97
FFT Mflops: 1982.24 (N=1024)
SOR Mflops: 1904.08 (100 x 100)
MonteCarlo: Mflops: 677.65
Sparse matmult Mflops: 2775.38 (N=1000, nz=5000)
LU Mflops: 5075.48 (M=100, N=100)
g++ trunk 20160224
Composite Score: 2314.35
FFT Mflops: 1986.77 (N=1024)
SOR Mflops: 1903.29 (100 x 100)
MonteCarlo: Mflops: 678.80
Sparse matmult Mflops: 2775.33 (N=1000, nz=5000)
LU Mflops: 4227.54 (M=100, N=100)
g++ trunk 20160224 + fold-const.c MIN/MAX change
Composite Score: 2331.88
FFT Mflops: 1983.28 (N=1024)
SOR Mflops: 1906.04 (100 x 100)
MonteCarlo: Mflops: 676.53
Sparse matmult Mflops: 2823.60 (N=1000, nz=5000)
LU Mflops: 4269.96 (M=100, N=100)
g++ trunk 20150224 + fold-const.c MIN/MAX change + loop-invert.patch
Composite Score: 2332.00
FFT Mflops: 1983.18 (N=1024)
SOR Mflops: 1905.64 (100 x 100)
MonteCarlo: Mflops: 674.50
Sparse matmult Mflops: 2823.55 (N=1000, nz=5000)
LU Mflops: 4273.14 (M=100, N=100)
clang 3.8
Composite Score: 2418.13
FFT Mflops: 1583.23 (N=1024)
SOR Mflops: 2130.27 (100 x 100)
MonteCarlo: Mflops: 281.80
Sparse matmult Mflops: 3026.40 (N=1000, nz=5000)
LU Mflops: 5068.95 (M=100, N=100)
clang++ 3.8
Composite Score: 2434.04
FFT Mflops: 1595.89 (N=1024)
SOR Mflops: 2131.09 (100 x 100)
MonteCarlo: Mflops: 281.63
Sparse matmult Mflops: 3001.59 (N=1000, nz=5000)
LU Mflops: 5159.98 (M=100, N=100)
More information about the Gcc-bugs
mailing list