[Mesa-dev] [PATCH 2/2] nir: add lowering for gl_HelperInvocation

[Jason Ekstrand]

On Mon, Jun 11, 2018 at 3:59 PM, Roland Scheidegger <sroland at vmware.com>
wrote:

> Am 12.06.2018 um 00:32 schrieb Jason Ekstrand:
> > On Wed, Jun 6, 2018 at 7:43 AM, Rob Clark <robdclark at gmail.com
> > <mailto:robdclark at gmail.com>> wrote:
> >
> >     Signed-off-by: Rob Clark <robdclark at gmail.com
> >     <mailto:robdclark at gmail.com>>
> >     ---
> >     I can't say for sure that this will work on all drivers, but it is
> >     what the blob driver does, and it seems to make deqp happy.  I could
> >     move this to it's own pass inside ir3, but that seemed like overkill
> >
> >      src/compiler/nir/nir.h                      | 10 ++++++++++
> >      src/compiler/nir/nir_lower_system_values.c  | 17 +++++++++++++++++
> >      src/gallium/drivers/freedreno/ir3/ir3_nir.c |  1 +
> >      3 files changed, 28 insertions(+)
> >
> >     diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
> >     index 073ab4e82ea..de3d55d83af 100644
> >     --- a/src/compiler/nir/nir.h
> >     +++ b/src/compiler/nir/nir.h
> >     @@ -1963,6 +1963,16 @@ typedef struct nir_shader_compiler_options {
> >          */
> >         bool lower_base_vertex;
> >
> >     +   /**
> >     +    * If enabled, gl_HelperInvocation will be lowered as:
> >     +    *
> >     +    *   !((1 << gl_SampleID) & gl_SampleMaskIN[0]))
> >
> >
> > This only works for multi-sampling.  What about the single-sampled case?
> It doesn't make sense to me for msaa neither.
> gl_SampleID forces per-sample execution, which clearly isn't what you
> want.


Yes, we do not want gl_HelperInvocation to force per-sample shading.  It
doesn't mean we can't lower to it, it just means that the back-end needs to
be very careful that it decides whether or not to execute per-sample before
it does this lowering.


> Plus, gl_SampleMaskIN is specified to only contain bits for the
> current shader invocation, so for msaa with forced per-sample execution
> that would only contain the single bit corresponding to gl_SampleID
> anyway (that is my interpretation at least - I know hw sample mask
> inputs will probably always contains all the bits from rasterization,
> regardless of gl_SampleID), so the "1 << gl_SampleID &" part will do
> nothing at all.


I think it also relies on gl_SampleMaskIN to give 0 for helper
invocations.  This is likely true but I don't know how well-defined all
this is in the world of helper invocations.


> So I think that part is more about lowering the hw
> rasterization sample mask to gl_SampleMaskIN rather than lowering to
> gl_HelperInvocation.
> But yes, !gl_SampleMaskIN should give gl_HelperInvocation - I think all
> hw can give you raster sample mask even without msaa but I'm not
> entirely sure if it's guaranteed to work in GL with single sampling
> (similar for sample id, which should just be stuck at 0).
>

The GL spec doesn't seem to say.  It does say that gl_SampleID is
guaranteed to be 0 for single-sampled but it says nothing about
gl_SampleMaskIN.

--Jason
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20180611/52ad8bf7/attachment-0001.html>