topic Re: Integer multitude vectorization in Intel® Fortran Compiler

Integer multitude vectorization

german_kolosov — Sat, 23 Aug 2003 14:27:38 GMT

Why the compiler does not support 16-bit integer multitude vectorization?.
for example loop:
DO I = 1, N
A(I,J) = A(I,J) + C*A(I,L)
END DO
When A and C are INTEGER*2 the compiler outs "unsupported loop structure"
But MMX has PMADDWD command for such operations.

Re: Integer multitude vectorization

Steven_L_Intel1 — Sat, 23 Aug 2003 19:12:11 GMT

I don't yet know the details of what Intel Fortran's vectorization is capable of, but I do have experience with other vectorizing compilers and they wouldn't have vectorized INTEGER*2 arrays either. In fact, a lot of normal optimizations don't tend to happen when using 8 or 16-bit integers. Is there a particular reason you chose INTEGER*2? What happens if you make these INTEGER*4?

Steve

Re: Integer multitude vectorization

german_kolosov — Sat, 23 Aug 2003 20:47:57 GMT

Thank You for answer.
I have no restrictions to use INTEGER*4, but I know that MMX doesn't have any 32-bit multitude commands.
It probably causes to message from compiler: unsupported data type.