radeonsi: rewrite/replace gfx10_ngg_get_vertices_per_prim

Reuse si_get_input_prim (which is similar) and split it into 2 functions:
- si_get_output_prim_simplified
- si_get_num_vertices_per_output_prim

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32257>

src/gallium/drivers/radeonsi/gfx10_shader_ngg.c

@@ -8,38 +8,6 @@
 #include "si_query.h"
 #include "si_shader_internal.h"
 
-unsigned gfx10_ngg_get_vertices_per_prim(struct si_shader *shader)
-{
-   const struct si_shader_info *info = &shader->selector->info;
-
-   if (shader->selector->stage == MESA_SHADER_GEOMETRY)
-      return mesa_vertices_per_prim(info->base.gs.output_primitive);
-   else if (shader->selector->stage == MESA_SHADER_VERTEX) {
-      if (info->base.vs.blit_sgprs_amd) {
-         /* Blits always use axis-aligned rectangles with 3 vertices. */
-         return 3;
-      } else if (shader->key.ge.opt.ngg_culling & SI_NGG_CULL_VS_LINES)
-         return 2;
-      else {
-         /* The shader compiler replaces 0 with 3. The generated code will be correct regardless
-          * of the draw primitive type, but it's less efficient.
-          *
-          * Computing prim export values for non-existent vertices has no effect.
-          */
-         return 0; /* unknown */
-      }
-   } else {
-      assert(shader->selector->stage == MESA_SHADER_TESS_EVAL);
-
-      if (info->base.tess.point_mode)
-         return 1;
-      else if (info->base.tess._primitive_mode == TESS_PRIMITIVE_ISOLINES)
-         return 2;
-      else
-         return 3;
-   }
-}
-
 bool gfx10_ngg_export_prim_early(struct si_shader *shader)
 {
    struct si_shader_selector *sel = shader->selector;

src/gallium/drivers/radeonsi/si_nir_lower_abi.c

@@ -254,7 +254,7 @@ static void preload_reusable_variables(nir_builder *b, struct lower_abi_state *s
 static nir_def *get_num_vertices_per_prim(nir_builder *b, struct lower_abi_state *s)
 {
    struct si_shader_args *args = s->args;
-   unsigned num_vertices = gfx10_ngg_get_vertices_per_prim(s->shader);
+   unsigned num_vertices = si_get_num_vertices_per_output_prim(s->shader);
 
    if (num_vertices)
       return nir_imm_int(b, num_vertices);

src/gallium/drivers/radeonsi/si_shader.c

@@ -1942,7 +1942,7 @@ static void si_lower_ngg(struct si_shader *shader, nir_shader *nir)
 
       unsigned clip_plane_enable =
          SI_NGG_CULL_GET_CLIP_PLANE_ENABLE(key->ge.opt.ngg_culling);
-      unsigned num_vertices = gfx10_ngg_get_vertices_per_prim(shader);
+      unsigned num_vertices = si_get_num_vertices_per_output_prim(shader);
 
       options.num_vertices_per_primitive = num_vertices ? num_vertices : 3;
       options.early_prim_export = gfx10_ngg_export_prim_early(shader);

src/gallium/drivers/radeonsi/si_shader_internal.h

@@ -117,7 +117,6 @@ void si_get_ps_epilog_args(struct si_shader_args *args,
                            struct ac_arg *sample_mask);
 
 /* gfx10_shader_ngg.c */
-unsigned gfx10_ngg_get_vertices_per_prim(struct si_shader *shader);
 bool gfx10_ngg_export_prim_early(struct si_shader *shader);
 unsigned gfx10_ngg_get_scratch_dw_size(struct si_shader *shader);
 bool gfx10_ngg_calculate_subgroup_info(struct si_shader *shader);

src/gallium/drivers/radeonsi/si_state.h

@@ -678,6 +678,9 @@ void si_update_ps_inputs_read_or_disabled(struct si_context *sctx);
 void si_update_vrs_flat_shading(struct si_context *sctx);
 unsigned si_get_input_prim(const struct si_shader_selector *gs, const union si_shader_key *key,
                            bool return_unknown);
+unsigned si_get_output_prim_simplified(const struct si_shader_selector *sel,
+                                       const union si_shader_key *key);
+unsigned si_get_num_vertices_per_output_prim(struct si_shader *shader);
 bool si_update_ngg(struct si_context *sctx);
 void si_vs_ps_key_update_rast_prim_smooth_stipple(struct si_context *sctx);
 void si_ps_key_update_framebuffer(struct si_context *sctx);

src/gallium/drivers/radeonsi/si_state_shaders.cpp

@@ -1377,6 +1377,47 @@ unsigned si_get_input_prim(const struct si_shader_selector *gs, const union si_s
       return MESA_PRIM_TRIANGLES; /* worst case for all callers */
 }
 
+/* Return a simplified primitive type, e.g. don't return *_STRIP and *_FAN.
+ * This returns MESA_PRIM_UNKNOWN if the primitive type is not known at compile time.
+ */
+unsigned si_get_output_prim_simplified(const struct si_shader_selector *sel,
+                                       const union si_shader_key *key)
+{
+   if (sel->stage == MESA_SHADER_GEOMETRY) {
+      if (util_rast_prim_is_triangles(sel->info.base.gs.output_primitive))
+         return MESA_PRIM_TRIANGLES;
+      else if (util_prim_is_lines(sel->info.base.gs.output_primitive))
+         return MESA_PRIM_LINES;
+      else
+         return MESA_PRIM_POINTS;
+   }
+
+   if (sel->stage == MESA_SHADER_VERTEX && sel->info.base.vs.blit_sgprs_amd)
+      return SI_PRIM_RECTANGLE_LIST;
+
+   /* It's the same as the input primitive type for VS and TES. */
+   return si_get_input_prim(sel, key, true);
+}
+
+unsigned si_get_num_vertices_per_output_prim(struct si_shader *shader)
+{
+   unsigned prim = si_get_output_prim_simplified(shader->selector, &shader->key);
+
+   switch (prim) {
+   case MESA_PRIM_TRIANGLES:
+   case SI_PRIM_RECTANGLE_LIST:
+      return 3;
+   case MESA_PRIM_LINES:
+      return 2;
+   case MESA_PRIM_POINTS:
+      return 1;
+   case MESA_PRIM_UNKNOWN:
+      return 0;
+   default:
+      unreachable("unexpected prim type");
+   }
+}
+
 static unsigned si_get_vs_out_cntl(const struct si_shader_selector *sel,
                                    const struct si_shader *shader, bool ngg)
 {