[Mesa-dev] [PATCH] radeonsi: Bitcast result of packf16 intrinsic to float for export intrinsic.

[Michel Dänzer]

On Don, 2012-11-29 at 08:45 -0800, Tom Stellard wrote: 
> On Tue, Nov 27, 2012 at 07:55:05PM +0100, Michel Dänzer wrote:
> > diff --git a/src/gallium/drivers/radeonsi/radeonsi_shader.c b/src/gallium/drivers/radeonsi/radeonsi_shader.c
> > index d6e37ac..d5d56c4 100644
> > --- a/src/gallium/drivers/radeonsi/radeonsi_shader.c
> > +++ b/src/gallium/drivers/radeonsi/radeonsi_shader.c
> > @@ -496,7 +496,11 @@ static void si_llvm_init_export_args(struct lp_build_tgsi_context *bld_base,
> >  						LLVMInt32TypeInContext(base->gallivm->context),
> >  						args, 2,
> >  						LLVMReadNoneAttribute);
> > -			args[chan + 7] = args[chan + 5];
> > +			args[chan + 7] = args[chan + 5] =
> > +				LLVMBuildBitCast(base->gallivm->builder,
> > +						 args[chan + 5],
> > +						 LLVMFloatTypeInContext(base->gallivm->context),
> > +						 "");
> >  		}
> 
> I think a better fix would be to change the return type of llvm.SI.packf16
> intrinsic from i32 to v2f16.  This is more correct and would allow us to
> use the intrinsic elsewhere if needed.

Makes sense.


> This will also require a new export intrinsic to handle the packed floats,
> and you'll need to add the v2f16 type to the register class definitions.

Not sure about this, though. Given the various export targets / formats,
I'm afraid this would result in a combinatorial explosion of export
intrinsics. I wonder if it might not be better to make the export
intrinsic take i32s instead (or whatever type is best for generic 32 bit
values), and have the driver bitcast as appropriate.


-- 
Earthling Michel Dänzer           |                   http://www.amd.com
Libre software enthusiast         |          Debian, X and DRI developer