[Mesa-dev] [PATCH 1/4] mesa/teximage: use correct extension for accept stencil texture.

[Pohjolainen, Topi]

On Mon, Apr 06, 2015 at 12:05:56PM -0400, Ilia Mirkin wrote:
> On Mon, Apr 6, 2015 at 11:33 AM, Pohjolainen, Topi
> <topi.pohjolainen at intel.com> wrote:
> > On Sun, Apr 05, 2015 at 08:22:13PM +0300, Pohjolainen, Topi wrote:
> >> On Sun, Apr 05, 2015 at 08:06:50PM +0300, Pohjolainen, Topi wrote:
> >> > On Sun, Apr 05, 2015 at 08:46:16AM -0400, Ilia Mirkin wrote:
> >> > > While this change is correct, the Intel guys will yell at you, because
> >> > > they're somehow misusing this in meta for Broadwell, s.t. this will
> >> > > cause crashes when blitting stencil. IMHO that's a problem that should
> >> > > be fixed in their driver and this can go on, but... it's also not my
> >> > > driver that's crashing -- they might feel differently :)
> >> >
> >> > As far as I can tell we only do:
> >> >
> >> >    _mesa_TexParameteri(target, GL_DEPTH_STENCIL_TEXTURE_MODE,
> >> >                        GL_STENCIL_INDEX);
> >> >
> >> > which suppose to be the right thing to do - we select the stencil to be
> >> > sampled instead of depth. And this won't hit the path below. I made the
> >> > change locally and I'm now running piglit on broadwell.
> >> >
> >> > I noticed that _mesa_base_tex_format() is in turn used in
> >> >
> >> > src/mesa/drivers/common/meta_blit.c
> >> >
> >> > but we shouldn't go there with intel driver ever. On hardware older than
> >> > broadwell we don't use meta and the one used on broadwell and newer
> >> > is found in:
> >> >
> >> > src/mesa/drivers/dri/i965/brw_meta_stencil_blit.c
> >> >
> >> > But lets see what piglit says.
> >>
> >> Right you are. This is more subtle, we will hit it when we actually create
> >> a temporary texture out of the given read renderbuffer. It seems that this
> >> was hit first time when formats where adjusted and then Jason added the
> >> conditional using ARB_stencil_texturing (which is not right either).
> >>
> >> Really sorry that this is hindering your work now. I'll try to take a look
> >> at this tomorrow.
> >
> > So far I can't come up with other things than pure hacks. I'll explain
> > a little what happens in the "intel stencil meta blit". Like I said, the
> > driver creates a temporary texture out of the stencil attachment:
> >
> >    const struct gl_renderbuffer_attachment *att =
> >       &ctx->ReadBuffer->Attachment[BUFFER_STENCIL];
> >    struct gl_renderbuffer *rb = att->Renderbuffer;
> >    struct gl_texture_object *tex_obj;
> >
> >    ...
> >       if (!_mesa_meta_bind_rb_as_tex_image(ctx, rb, &blit->tempTex, &tex_obj,
> >                                           target)) {
> >
> >
> > This gets wound back to the driver, a call to
> > intel_bind_renderbuffer_tex_image() which in turn calls the core again.
> >
> >    _mesa_init_teximage_fields(ctx, image,
> >                               rb->Width, rb->Height, 1,
> >                               0, rb->InternalFormat, rb->Format);
> >
> > Here "rb->InternalFormat" is GL_STENCIL_INDEX that won't be accepted by
> > _mesa_base_tex_format() anymore without ARB_texture_stencil8. As most of
> > the texture image setting up logic takes place in the core, the boolean state
> > flag (brw_context::meta_in_progress) we have in intel driver is not much
> > help. It looks that we would need additional driver driven overriding.
> > But I don't like that at all.
> >
> > I'll look into this some more, but any suggestions are welcome.
> 
> Can you, at a high level, explain what's going on? Is stencil stored
> separately from depth? If so, you *should* support
> ARB_texture_stencil8, which will solve your issues. Alternatively, did
> you mean to create a GL_DEPTH_STENCIL image with StencilSampling
> turned on for the object?

Stencil is stored separately and I agree that supporting ARB_texture_stencil8
is the right solution of course. I was in the impression that we are missing
things in our driver preventing it simply to be turned on. And therfore I was
looking for something simple we could use in the meantime.

I could be wrong though, I just don't know all the implications turning on
ARB_texture_stencil8 has.