[Bug 92760] Add FP64 support to the i965 shader backends

bugzilla-daemon at freedesktop.org bugzilla-daemon at freedesktop.org
Mon Jan 11 23:34:18 PST 2016


https://bugs.freedesktop.org/show_bug.cgi?id=92760

--- Comment #31 from Iago Toral <itoral at igalia.com> ---
(In reply to Jason Ekstrand from comment #30)
> Created attachment 120957 [details]
> NIR indirect lowering pass
> 
> (In reply to Iago Toral from comment #29)
> > Hey Jason/Connor,
> > 
> > the lowering of trunc for doubles has some code that looks like this
> > (pseudo-code):
> > 
> > if (exponent < 0) {
> >    mask = 0x0
> > } else if (exponent > 52) {
> >    mask = 0x7fffffffffffffff;
> > } else {
> >    /* This is a 64-bit integer op, needs to be split into hi/lo 32-bit ops */
> >    mask =  (1LL << frac_bits) - 1;
> > }
> > 
> > The current implementation I have works fine using bcsel. It looks something
> > like this (again, pseudo-code):
> > 
> > mask = bcsel(exponent < 0,
> >              0x7fffffffffffffff,
> >              bcsel(exponent > 52,
> >                    0x0000000000000000,
> >                    (1LL << frac_bits) -1))
> > 
> > My problem with this is that "(1LL << frac_bits) - 1" is a 64-bit integer
> > operation that we have to implement in terms of hi/lo 32-bit integer
> > operations (at least until we support 64-bit integers), so it is really a
> > bunch of instructions. Because I use bcsel, it means that we generate code
> > for that even if exponent is not in [1..51], which is not ideal.
> 
> Right.  I would encourage you not to use if's too much because branching may
> be more expensive than bcsel depending on what paths different invocations
> take.  However, if one side of the if is overwhelmingly more likely than the
> other, then control-flow is probably a good idea.

Yeah, in this case exponents in the range 0..52 would be a lot more common than
anything else.

> > I was thinking about rewriting this as an if/else ladder instead, however, I
> > noticed that because this occurs in SSA mode I would have to deal with the
> > phi nodes etc manually and I don't see any other case where we do something
> > like that outside the NIR to SSA pass, so I wonder if this is actually a
> > good idea at all. What do you think?
> > 
> > If you think the if/else ladder is the way to go, is there any documentation
> > or
> > code references I can look at to have an idea as to how that should be
> > implemented for a lowering pass in SSA mode?
> 
> I attached a pass that I've written recently (not yet sent out for review,
> but it does work) that does exactly this.  It replaces indirect load/store
> operations with if-ladders and phi nodes (if needed).  Most of it comes down
> to using nir_insert_cf_node to insert it at the builder's cursor and then
> making sure you set the cursor to something reasonable when you're done.

Awesome, thanks a lot Jason!

-- 
You are receiving this mail because:
You are the QA Contact for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/intel-3d-bugs/attachments/20160112/ac3487a4/attachment-0001.html>


More information about the intel-3d-bugs mailing list