[Mesa-dev] [PATCH] nv50/ir: Split 64-bit MAD and MUL operations

[Pierre Moreau]

On 06:24 PM - Mar 19 2016, Ilia Mirkin wrote:
> On Sat, Mar 19, 2016 at 6:15 PM, Pierre Moreau <pierre.morrow at free.fr> wrote:
> > On 06:05 PM - Mar 19 2016, Ilia Mirkin wrote:
> >> Not 100% sure, but pretty sure this is wrong. Can you provide the
> >> generated sequence of instructions in response to a 64-bit mul and
> >> mad?
> >
> > For the given mul:
> >
> >   mul u64 %r55d %r42d %r52d
> >
> > the following is generated:
> >
> >    mul u32 { $r0 $c0 } $r0 $r2
> >    mul (SUBOP:1) u32 $r1 $r1 $r3 $c0
> 
> That's not enough though... you need 4 mul's... if you have numbers
> (ab) * (cd) where a/b are the high/low of the 64-bit int, that results
> in
> 
> b*d + (a*d + b * d) * (1 << 32)
> 
> See expandIntegerMultiply in there -- it's meant for splitting a
> 32-bit multiply into 16x16 muls (which is what nv50 can do), but the
> same principle applies to splitting 64x64 into 32x32's.

Oops… that's definitely true. I'll fix that.

Pierre

> 
>   -ilia
> 
> >
> >
> > Whereas for the mad, I need to first find how to tell Nouveau to stop splitting
> > each of my mads to mul + add…
> >
> >>
> >> On Sat, Mar 19, 2016 at 5:56 PM, Pierre Moreau <pierre.morrow at free.fr> wrote:
> >> > Two 32-bit MAD or MUL operations are generated in place of the original 64-bit
> >> > operation. All operands can either be signed or unsigned, but they have to be
> >> > integers.
> >> >
> >> > Signed-off-by: Pierre Moreau <pierre.morrow at free.fr>
> >> > ---
> >> >  src/gallium/drivers/nouveau/codegen/nv50_ir_build_util.cpp | 11 ++++++++++-
> >> >  1 file changed, 10 insertions(+), 1 deletion(-)
> >> >
> >> > diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_build_util.cpp b/src/gallium/drivers/nouveau/codegen/nv50_ir_build_util.cpp
> >> > index 84ebfdb..0b37fcf 100644
> >> > --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_build_util.cpp
> >> > +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_build_util.cpp
> >> > @@ -586,6 +586,12 @@ BuildUtil::split64BitOpPostRA(Function *fn, Instruction *i,
> >> >        srcNr = 2;
> >> >        break;
> >> >     case OP_SELP: srcNr = 3; break;
> >> > +   case OP_MAD: /* fallthrough */
> >> > +   case OP_MUL:
> >> > +      if (!carry || isFloatType(i->dType) || isFloatType(i->sType))
> >> > +         return NULL;
> >> > +      srcNr = (i->op == OP_MAD) ? 3 : 2;
> >> > +      break;
> >> >     default:
> >> >        // TODO when needed
> >> >        return NULL;
> >> > @@ -600,6 +606,9 @@ BuildUtil::split64BitOpPostRA(Function *fn, Instruction *i,
> >> >
> >> >     hi->getDef(0)->reg.data.id++;
> >> >
> >> > +   if (i->op == OP_MAD || i->op == OP_MUL)
> >> > +      hi->subOp = NV50_IR_SUBOP_MUL_HIGH;
> >> > +
> >> >     for (int s = 0; s < srcNr; ++s) {
> >> >        if (lo->getSrc(s)->reg.size < 8) {
> >> >           if (s == 2)
> >> > @@ -629,7 +638,7 @@ BuildUtil::split64BitOpPostRA(Function *fn, Instruction *i,
> >> >           }
> >> >        }
> >> >     }
> >> > -   if (srcNr == 2) {
> >> > +   if (srcNr >= 2) {
> >> >        lo->setFlagsDef(1, carry);
> >> >        hi->setFlagsSrc(hi->srcCount(), carry);
> >> >     }
> >> > --
> >> > 2.7.4
> >> >
> >> > _______________________________________________
> >> > mesa-dev mailing list
> >> > mesa-dev at lists.freedesktop.org
> >> > https://lists.freedesktop.org/mailman/listinfo/mesa-dev
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20160319/8eb78c3a/attachment.sig>