[Mesa-dev] [PATCH 30/53] mesa: don't always set _NEW_PROGRAM when linking

[Marek Olšák]

On Tue, Jan 3, 2017 at 11:14 PM, Timothy Arceri
<timothy.arceri at collabora.com> wrote:
> On Tue, 2017-01-03 at 22:41 +0100, Marek Olšák wrote:
>> On Tue, Jan 3, 2017 at 10:14 PM, Timothy Arceri
>> <timothy.arceri at collabora.com> wrote:
>> > On Tue, 2017-01-03 at 22:07 +0100, Marek Olšák wrote:
>> > > On Tue, Jan 3, 2017 at 9:57 PM, Timothy Arceri
>> > > <timothy.arceri at collabora.com> wrote:
>> > > > On Tue, 2017-01-03 at 20:47 +0100, Marek Olšák wrote:
>> > > > > On Tue, Jan 3, 2017 at 3:43 AM, Timothy Arceri
>> > > > > <timothy.arceri at collabora.com> wrote:
>> > > > > > We only need to set it when linking was successful and the
>> > > > > > program
>> > > > > > being linked is currently active.
>> > > > > >
>> > > > > > The programs_in_use mask is just used as a flag for now but
>> > > > > > in
>> > > > > > a following patch we will use it to update the
>> > > > > > CurrentProgram
>> > > > > > array.
>> > > > > > ---
>> > > > > >  src/mesa/main/shaderapi.c | 22 +++++++++++++++++++++-
>> > > > > >  1 file changed, 21 insertions(+), 1 deletion(-)
>> > > > > >
>> > > > > > diff --git a/src/mesa/main/shaderapi.c
>> > > > > > b/src/mesa/main/shaderapi.c
>> > > > > > index 022afab..6c03dcb 100644
>> > > > > > --- a/src/mesa/main/shaderapi.c
>> > > > > > +++ b/src/mesa/main/shaderapi.c
>> > > > > > @@ -1083,10 +1083,30 @@ _mesa_link_program(struct
>> > > > > > gl_context
>> > > > > > *ctx,
>> > > > > > struct gl_shader_program *shProg)
>> > > > > >        return;
>> > > > > >     }
>> > > > > >
>> > > > > > -   FLUSH_VERTICES(ctx, _NEW_PROGRAM);
>> > > > > > +   unsigned programs_in_use = 0;
>> > > > > > +   if (ctx->_Shader)
>> > > > > > +      for (unsigned stage = 0; stage < MESA_SHADER_STAGES;
>> > > > > > stage++) {
>> > > > > > +         if (ctx->_Shader->CurrentProgram[stage] ==
>> > > > > > shProg) {
>> > > > > > +            programs_in_use |= 1 << stage;
>> > > > > > +         }
>> > > > > > +   }
>> > > > > >
>> > > > > >     _mesa_glsl_link_shader(ctx, shProg);
>> > > > > >
>> > > > > > +   /* From section 7.3 (Program Objects) of the OpenGL 4.5
>> > > > > > spec:
>> > > > > > +    *
>> > > > > > +    *    "If LinkProgram or ProgramBinary successfully re-
>> > > > > > links a
>> > > > > > program
>> > > > > > +    *     object that is active for any shader stage, then
>> > > > > > the
>> > > > > > newly generated
>> > > > > > +    *     executable code will be installed as part of the
>> > > > > > current
>> > > > > > rendering
>> > > > > > +    *     state for all shader stages where the program is
>> > > > > > active.
>> > > > > > +    *     Additionally, the newly generated executable
>> > > > > > code is
>> > > > > > made part of
>> > > > > > +    *     the state of any program pipeline for all stages
>> > > > > > where
>> > > > > > the program
>> > > > > > +    *     is attached."
>> > > > > > +    */
>> > > > > > +   if (shProg->data->LinkStatus && programs_in_use) {
>> > > > > > +      FLUSH_VERTICES(ctx, _NEW_PROGRAM);
>> > > > > > +   }
>> > > > >
>> > > > > This doesn't seem correct. If the context has unflushed
>> > > > > vertices,
>> > > > > calling FLUSH_VERTICES after linking will use the new linked
>> > > > > program
>> > > > > instead of the previous program for which the vertices were
>> > > > > submitted.
>> > > >
>> > > > Your probably right but this doesn't make anything worse than
>> > > > it
>> > > > already is right?
>> > > >
>> > > > Before this change we where always calling FLUSH_VERTICES(ctx,
>> > > > _NEW_PROGRAM);
>> > >
>> > > The useless flagging is a different issue.
>> > >
>> > > FLUSH_VERTICES should be called before a state change, because it
>> > > flushes draw calls that the driver hasn't even seen yet. After
>> > > the
>> > > driver processes those draw calls, _NEW_PROGRAM is set. That's
>> > > how it
>> > > works.
>> > >
>> > > If you instead change a state and then call FLUSH_VERTICES, the
>> > > previous unflushed draw calls will use the new state, which is
>> > > incorrect.
>> >
>> > But the state doesn't change just from linking does it? The new
>> > programs won't be made current until glUseProgram() is called.
>>
>> The GL state changes if the program is current and re-linked. (not
>> sure if all of the internal Mesa state changes too)
>
> Right but this patch only changes things so that FLUSH_VERTICES is not
> called when the program is *not* current.

That idea is OK. The problem is the patch also moves FLUSH_VERTICES
after the _mesa_glsl_link_shader call, which breaks the behavior if
the program *is* current.

Consider this:

glUseProgram();
glBegin();
glVertex();
glEnd();
// the vbo module doesn't create a draw call in glEnd; instead, it waits
// for another glBegin call and the driver doesn't see any draw calls yet.
glDetachShader();
glAttachShader();
glLinkProgram() {
  _mesa_glsl_link_shader(); // this replaces the current program
  FLUSH_VERTICES(); // this finally creates the draw call from Begin/End,
  // but with the new program
}


The correct sequence is:

glLinkProgram() {
  FLUSH_VERTICES(); // create the draw call from Begin/End
 _mesa_glsl_link_shader(); // replace the current program
}

Marek