Bas Zoetekouw wrote:
I don't know much about the others, but LU is likely to be this different because ccc links to cxml with Kazushige Goto's assembler BLAS. Those have always been free, and are also in ATLAS now, so at least for that benchmark, we can do as well with free software.Hi Ted!
You wrote:
It was my (perhaps ignorant?) understanding that GCC3.2 closed the gap (at
least made the gap a lot smaller) between GCC and CCC. Am I wrong? (I have
yet to test this, however, as I don't have GCC3.2 working on my alpha.)
I'm not sure if the gap has become smaller, but ccc still generated code that is a lot faster than gcc-3.2.
For example, take a look at these results from SCIbench (http://math.nist.gov/scimark2), generated on an quadruple-proc EV67 machine (running Tru64 Unix btw, not Linux):
Compaq C compiler, V6.4-014 CFLAGS = -arch ev67 -fast -O4
| Composite Score: 195.47
| FFT Mflops: 207.66 (N=1024)
| SOR Mflops: 235.00 (100 x 100)
| MonteCarlo: Mflops: 53.33
| Sparse matmult Mflops: 177.93 (N=1000, nz=5000)
| LU Mflops: 303.42 (M=100, N=100)
GNU C compiler, V3.2.1
CFLAGS = -O3 -fomit-frame-pointer -ffast-math -mcpu=ev67
| Composite Score: 137.18
| FFT Mflops: 188.23 (N=1024)
| SOR Mflops: 167.08 (100 x 100)
| MonteCarlo: Mflops: 49.71
| Sparse matmult Mflops: 163.85 (N=1000, nz=5000)
| LU Mflops: 117.03 (M=100, N=100)
Or if both LUs are compiled from code, well, then we can do better than both with free software, as Goto's BLAS blow away anything compiled.
Zeen, --
-Adam P.
GPG fingerprint: D54D 1AEE B11C CE9B A02B C5DD 526F 01E8 564E E4B6
Welcome to the best software in the world today cafe! <http://lyre.mit.edu/%7Epowell/The_Best_Stuff_In_The_World_Today_Cafe.ogg>