> Does anyone know approximately how many clock cycles it takes to copy
> the memory from the vector register to the scalar register?... At least
> compared to something like vec_madd()
It depends on the processor. On a recent G4, it takes 3 cycles to
complete the store instruction, another 6 before the data is ready
forwarding and then at least 3 cycles to complete the load. On a G5,
there is likely also a load/store alias reject going on, which can cost
you ~50 cycles. Just say no.
vec_madd can execute with a throughput of 1 vec_madd per cycle.
It is likely you can use the vector unit to do the thing that you think
you need the scalar units for. For example, to increment one float:
vector float f = vec_lde( 0, &the_float);
f = vec_add( f, (vector float) (1.0f) );
vec_ste( f, 0, &the_float );
Unless you pull up a denormal with vec_lde on G5 (and VSCR[NJ] = 0),
this should be roughly as fast as the same code for the scalar unit:
float f;
lfs f, 0, &the_float
fadds f, f, 1.0f
stfs f, 0, &the_float
You can do almost anything with the vector unit that you can do with
the scalar units. Doing scalar operations one at a time in the vector
unit doesnt add any parallelism and you have to deal with alignment in
software, but at least you didn't pay the cost to move data over to the
scalar units.
Ian
_______________________________________________
scitech mailing list | email@hidden
Help/Unsubscribe/Archives: http://www.lists.apple.com/mailman/listinfo/scitech
Do not post admin requests to the list. They will be ignored.
