On Friday, 23 November 2012 at 09:08:53 UTC, Manu wrote:
On 23 November 2012 05:03, xenon325 <[email protected]> wrote:
On Monday, 19 November 2012 at 15:48:23 UTC, Manu wrote:
This wouldn't strictly retain half precision though, it would
be
slightly higher precision since the intermediates were full
precision
(which is surely preferable?).
I would think it's actually not preferable.
Imagine you developed and tuned all the code on x86 and
everything is
fine. Then run it on ARM and suddenly all computations are
inaccurate.
I think it would always be the case that work is done in float
space,
hardware support for half applies to fast load/store into full
float
registers. You would lose precision too fast if work were done
directly in
half space.
Most CPU's also apply this principle to integer work. It's
typical to load
a byte/short into a 32/64bit register and sign extend or zero
extend, then
all integer work is done in the maximum integer precision, and
store then
truncates the top bits again. Many older FPU's work this way
too.
Then it's ok, of course. Thanks for explaining!