[Mesa-dev] [RFC 0/3] V2: Improve GLSL support of GL_ARB_separate_shader_objects

[gregory hainaut]

On Sat, 3 Oct 2015 11:59:49 +0200
gregory hainaut <gregory.hainaut at gmail.com> wrote:

> On Sat, 03 Oct 2015 09:35:49 +0000
> Mike Lothian <mike at fireburn.co.uk> wrote:
> 
> > Would it be better to have is_interstage=0 rather than a double negative?
> > 
> Yes. I think it just need to set 1 in the constructor (forget to 
> update it by the way...) as default value. Otherwise it can be 
> renamed to is_unlinked_io (or something similar).
> 
> > 
> > > > In short, SSO allow to match by name but you can't make any hypothesis
> > > on the
> > > > previous/next stage. Therefore you must consider all inputs and output as
> > > > actives.
> > >
> > > New version based on Ian's feedbacks.
> > > * Real interstage variables of the program are still optimized
> > > * Both output and input of the program remains active
> > > * An old bug still remains if name and explicit location rendezvous are
> > > mixed
> > >
> > > Commits decription:
> > > 1/ add a not_interstage parameter that will be 1 when variable is either an
> > >    input of output of the program. I didn't know where to put the
> > > ir_set_not_interstage_io
> > >    function. Maybe it can be moved in linker.cpp.
> > > 2/ Set the interstage parameter based on the shader position in the
> > > pipeline
> > >    program. Potentially can be squased with the previous one.
> > > 3/ Don't do deadcode removal of not_interstage variable with the exception
> > >    of the built-in varyings because they don't impact location of others
> > > variables
> > >
> > > Gregory Hainaut (4):
> > >   glsl IR: add not_interstage attribute to ir_variable
> > >   glsl link: annotate not_interstage attribute of the ir_var after link
> > >     phase
> > >   glsl IR: only allow optimization of interstage variable
> > >   allow compilation on my system
> > >
> > >  configure.ac               |  6 ++++--
> > >  src/glsl/ir.h              |  6 ++++++
> > >  src/glsl/ir_optimization.h |  2 ++
> > >  src/glsl/linker.cpp        | 47
> > > ++++++++++++++++++++++++++++++++++++++++++++++
> > >  src/glsl/opt_dead_code.cpp | 37 ++++++++++++++++++++++++++++++++++++
> > >  5 files changed, 96 insertions(+), 2 deletions(-)
> > >
> > > --
> > > 2.1.4
> > >
> > > _______________________________________________
> > > mesa-dev mailing list
> > > mesa-dev at lists.freedesktop.org
> > > http://lists.freedesktop.org/mailman/listinfo/mesa-dev
> > >

I found another issue with my application (lucky me).

out SHADER
{
    vec4 t;
    vec4 c;
    flat vec4 fc;
} VSout;

in SHADER
{
    vec4 t;
    vec4 c;
    flat vec4 fc;
} PSin;


will be transformed into the following interface in SSO (in varying_matches::record) to

out SHADER
{
    flat vec4 t;
    flat vec4 c;
    flat vec4 fc;
} VSout;

in SHADER
{
    vec4 t;
    vec4 c;
    flat vec4 fc;
} PSin;

It doesn't change the interpolation/rendering as it will use the one
from the FS. However it will affect the
match_comparator function. So the sorting in
varying_matches::assign_locations will 
behave differently for some parameters. I don't know how to handle
this. Either we could adapt the sort function
to mask the flat value. Or we mustn't add the flat qualifier. Any opinions or
other solutions?