[Mesa-dev] mediump / half-precision support

[Connor Abbott]

Maybe instead you could do something in the middle -- add a precision
field to glsl_type instead of adding to glsl_base_type. Then make
different precision types compare equal in most places (since it is a
hint to the compiler, rather than a full-fledged extra type) and have
a pass that propagates mediump from variables to their
uses/definitions in GLSL IR, so that non-NIR users also get the
benefit. How does that sound?

On Sun, May 15, 2016 at 3:05 PM, Rob Clark <robdclark at gmail.com> wrote:
> So I've been hacking around on a couple approaches to getting
> half-precision support going for gles (and btw, someday it might be
> interesting to have an extension to respect precision qualifiers in gl
> too)
>
> I initially started by propagating ir_variable_data::precision to
> nir_variable_data and fixing up glsl_get_bit_size() to take a
> precision argument.  Ilia recommended to extend glsl_types to have
> half-precision types instead.. so I gave that a try.  IMHO it turns
> out to be a lot of churn (type explosion, with ripple-on effects to
> builtin-functions, new glsl ir opcodes, etc).  What I have can at
> least compile correctly *some* gles shaders.  And still having some
> problems with things getting promoted to full precision when they
> should not (or missing implicit conversion on function return types,
> etc).  And builtin-functions is still missing a handling for a lot of
> the various permutations of full and half-precision types, probably
> falls over harder when there are samplers in the mix, etc... so still
> lots of typing to do..  But here is what I've go:
>
> https://github.com/freedreno/mesa/commits/wip-mediump
>
> I may resurrect the other branch and see how far I get with that,
> since I didn't get as far as to be able to compile a shader before I
> switched. I did introduce a hierarchical-visitor to derive precision
> from an ir_rvalue (which I think might be a more sane way to implement
> the rules about precision promotion).  After that I think the main
> thing was to fixup precision on generated intermediate variables (such
> as for handling inlining of fxns, etc), but I think a lot of that
> could just use the precision_visitor to figure out what to do.  For
> most part, in nir the bitsize comes from the originating ir_variable
> or glsl_struct_field for struct deref's..).  We'd have to somehow make
> glsl_to_nir clever enough to insert f2h/h2f/u2h/.. conversion
> instructions, since at the glsl ir level precision is separate from
> type.  But I kinda think this might be the easier approach.
>
> Thoughts?
>
> BR,
> -R
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev