karl@mtu.UUCP (Karl Ottenstein) (05/04/85)
> Vectorization only works on certain "well structured, regular" loops. > This is nothing new, the first FORTRAN programs did not vectorize well > either, so you will either restructure to gain performance benefits of > vector C (potentially non-portable) or settle for scalar performance. > That is how most supercomputer codes are structured. It depends how much > work you are willing to put into it. The issue is not that old FORTRAN programs did not vectorize well, but rather that old compilers did not perform sufficient analysis to produce good code. The work of Kuck (et al) at Illinois and Kennedy (et al) at Rice has produced very powerful vectorizing techniques which are incorporated in several commercial translators, including [I am led to believe] the Convex compiler. It is true that Cray manuals specified that loops should look certain ways to be vectorized, but that was simply due to their stupid compilers at the time. -- Karl Ottenstein Michigan Technological University uucp: {lanl, ihnp4, glacier}!mtu!karl arpa/csnet: karl%mtu@csnet-relay
eugene@ames.UUCP (Eugene Miya) (05/10/85)
> > Vectorization only works on certain "well structured, regular" loops. > The issue is not that old FORTRAN programs did not vectorize well, but > rather that old compilers did not perform sufficient analysis to produce > good code. > > Karl Ottenstein > Michigan Technological University Pilot light: I have used several 'smart' tools. This includes software from Cray, CDC, Pacific Sierra, and soon Fujitsu. It has been my experience that even the smart tools give up easily. Perhaps, you have had different experiences. Not a flame: Nobody sent any test program to me. I ran two test programs (I prefer to not call them benchmarks) which characterized the Convex architecture. Test 1: was a test to see how smart the vectorizer restructured doubly nested and greater depth levels of DO-loops. The test involves dimensioning a reasonably large 2-D square array. Then using sub portions of the array in long slender geometries. The loops then hit the array in the long direction (column-major) and in the slender (row-major) direction. The restructuring took care of this (to a degree), there is some behavior, I am not clear on (some cases where inefficient work is being done) but I'm not yet certain why. Test 2: Memory assignment using vector scatter-gather function. This test initialized memory first in a scalar mode ( linear cost as the size of an array increased). In vector mode, (first it showed that it was a vector machine with a slight jump at 129 elements for extra startup cost) the Convex handled the scatter gather function reasonably well [Compared to the XMP/48 which I've also run the test on]. I've got other tests, but little time, now. In summary, the Convex C-1 does not appear to be a bad machine all factors considered. Now if the compiler just just be moved over to a Cray..... --eugene miya NASA Ames Research Center {hplabs,ihnp4,dual,hao,decwrl,allegra}!ames!aurora!eugene @ames-vmsb.ARPA:emiya@jup.DECNET