[Mesa-dev] [PATCH 2/2] radv/gfx10: correctly determine the number of vertices per primitive

[Bas Nieuwenhuizen]

On Mon, Jul 22, 2019 at 6:01 PM Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>
> On Mon, Jul 22, 2019 at 11:49 AM Samuel Pitoiset
> <samuel.pitoiset at gmail.com> wrote:
> >
> > For TES as NGG.
> >
> > Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
> > ---
> >  src/amd/vulkan/radv_nir_to_llvm.c | 17 ++++++++++++++++-
> >  1 file changed, 16 insertions(+), 1 deletion(-)
> >
> > diff --git a/src/amd/vulkan/radv_nir_to_llvm.c b/src/amd/vulkan/radv_nir_to_llvm.c
> > index 336bae28614..6e5a283f923 100644
> > --- a/src/amd/vulkan/radv_nir_to_llvm.c
> > +++ b/src/amd/vulkan/radv_nir_to_llvm.c
> > @@ -112,6 +112,7 @@ struct radv_shader_context {
> >         unsigned gs_max_out_vertices;
> >         unsigned gs_output_prim;
> >
> > +       unsigned tes_point_mode;
> >         unsigned tes_primitive_mode;
> >
> >         uint32_t tcs_patch_outputs_read;
> > @@ -3304,7 +3305,6 @@ handle_ngg_outputs_post(struct radv_shader_context *ctx)
> >  {
> >         LLVMBuilderRef builder = ctx->ac.builder;
> >         struct ac_build_if_state if_state;
> > -       unsigned num_vertices = 3;
> >         LLVMValueRef tmp;
> >
> >         assert((ctx->stage == MESA_SHADER_VERTEX ||
> > @@ -3322,6 +3322,20 @@ handle_ngg_outputs_post(struct radv_shader_context *ctx)
> >                 ac_unpack_param(&ctx->ac, ctx->gs_vtx_offset[2], 0, 16),
> >         };
> >
> > +       /* Determine the number of vertices per primitive. */
> > +       unsigned num_vertices;
> > +
> > +       if (ctx->stage == MESA_SHADER_VERTEX) {
> > +               num_vertices = 3; /* TODO: optimize for points & lines */
> > +       } else {
> > +               if (ctx->tes_point_mode)
> > +                       num_vertices = 1;
> > +               else if (ctx->tes_primitive_mode == GL_LINES)
> > +                       num_vertices = 2;
> > +               else
> > +                       num_vertices = 3;
> > +       }
> > +
> >         /* TODO: streamout */
> >
> >         /* Copy Primitive IDs from GS threads to the LDS address corresponding
> > @@ -4435,6 +4449,7 @@ LLVMModuleRef ac_translate_nir_to_llvm(struct ac_llvm_compiler *ac_llvm,
> >                                 ctx.tcs_num_inputs = util_last_bit64(shader_info->info.vs.ls_outputs_written);
> >                         ctx.tcs_num_patches = get_tcs_num_patches(&ctx);
> >                 } else if (shaders[i]->info.stage == MESA_SHADER_TESS_EVAL) {
> > +                       ctx.tes_point_mode = shaders[i]->info.tess.point_mode;
>
> Drive-by-comment without reading the full context...
>
> What if there's e.g. a GS which produces not-points? This bool will be
> set, and the logic above will say num_vertices = 1, which presumably
> is bad.

The invariant you're probably missing here is that
handle_ngg_outputs_post only gets called if there is no GS.
 (And the gs epilogue does not care about these tessellation variables).

>
>   -ilia
>
> >                         ctx.tes_primitive_mode = shaders[i]->info.tess.primitive_mode;
> >                         ctx.abi.load_tess_varyings = load_tes_input;
> >                         ctx.abi.load_tess_coord = load_tess_coord;
> > --
> > 2.22.0
> >
> > _______________________________________________
> > mesa-dev mailing list
> > mesa-dev at lists.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/mesa-dev
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev