typedef union
{
float f[4] ;
v4sf v ;
} vector[100] ;
for( i=100; --i>=0;) {
row1[i] = vector[i].f[0];
row2[i] = vector[i].f[1];
row3[i] = vector[i].f[2];
row4[i] = vector[i].f[3];
}
Unroll by four: load four vectors, swap data around in
registers, store four vectors.
Is this the only portable way to do to a pack/unpack without asm()?
If you want fully portable at the C level without using any
conditionals, this is pretty much it. If you just don't want
to use asm(), there are intrinsics you can use.
How do I set it up differently to trigger a pack/unpack optimization?
Perhaps the auto-vectorisers aren't smart enough (yet) to
do this for you. If your goal is great performance, you
really have to write a special version for every processor;
although auto-vectorisation certainly can speed up things
quite a bit, hand-written vector code can be *much* faster.
A big part of the problem is that many vector insn sets are
very limited, or just "different".
Segher
