ac/nir/lower_ngg: add an option not to export cull distances if the shader culls them

Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35352>

src/amd/common/nir/ac_nir.h

@@ -172,6 +172,12 @@ typedef struct {
     * either against CLIP_VERTEX or POS.
     */
    uint8_t cull_clipdist_mask;
+   /* This skips exporting cull distances to increase throughput by reducing the number of pos exports.
+    * If this is set, cull_clipdist_mask must be set to cull against cull distances in the shader because
+    * the hw won't do it without the exports. The best case scenario is 100% increase in throughput
+    * (2 pos exports -> 1 pos export).
+    */
+   bool dont_export_cull_distances;
    bool write_pos_to_clipvertex;
    /* Remove clip/cull distance components that are missing in export_clipdist_mask, improving
     * throughput by up to 50% (3 pos exports -> 2 pos exports). The caller shouldn't set no-op

src/amd/common/nir/ac_nir_create_gs_copy_shader.c

@@ -120,8 +120,9 @@ ac_nir_create_gs_copy_shader(const nir_shader *gs_nir,
          if (kill_layer)
             export_outputs &= ~VARYING_BIT_LAYER;
 
-         ac_nir_export_position(&b, gfx_level, export_clipdist_mask, write_pos_to_clipvertex, pack_clip_cull_distances,
-                                !has_param_exports, force_vrs, export_outputs, &out, NULL);
+         ac_nir_export_position(&b, gfx_level, export_clipdist_mask, false, write_pos_to_clipvertex,
+                                pack_clip_cull_distances, !has_param_exports, force_vrs, export_outputs,
+                                &out, NULL);
 
          if (has_param_exports) {
             ac_nir_export_parameters(&b, param_offsets,

src/amd/common/nir/ac_nir_helpers.h

@@ -134,6 +134,7 @@ void
 ac_nir_export_position(nir_builder *b,
                        enum amd_gfx_level gfx_level,
                        uint32_t export_clipdist_mask,
+                       bool dont_export_cull_distances,
                        bool write_pos_to_clipvertex,
                        bool pack_clip_cull_distances,
                        bool no_param_export,

src/amd/common/nir/ac_nir_lower_legacy_vs.c

@@ -78,8 +78,9 @@ ac_nir_lower_legacy_vs(nir_shader *nir,
    if (kill_layer)
       export_outputs &= ~VARYING_BIT_LAYER;
 
-   ac_nir_export_position(&b, gfx_level, export_clipdist_mask, write_pos_to_clipvertex, pack_clip_cull_distances,
-                          !has_param_exports, force_vrs, export_outputs, &out, NULL);
+   ac_nir_export_position(&b, gfx_level, export_clipdist_mask, false, write_pos_to_clipvertex,
+                          pack_clip_cull_distances, !has_param_exports, force_vrs, export_outputs,
+                          &out, NULL);
 
    if (has_param_exports) {
       ac_nir_export_parameters(&b, param_offsets,

src/amd/common/nir/ac_nir_lower_ngg.c

@@ -1746,6 +1746,7 @@ ac_nir_lower_ngg_nogs(nir_shader *shader, const ac_nir_lower_ngg_options *option
 
    ac_nir_export_position(b, options->hw_info->gfx_level,
                           options->export_clipdist_mask,
+                          options->dont_export_cull_distances,
                           options->write_pos_to_clipvertex,
                           options->pack_clip_cull_distances,
                           !options->has_param_exports,

src/amd/common/nir/ac_nir_lower_ngg_gs.c

@@ -469,6 +469,7 @@ ngg_gs_emit_output(nir_builder *b, nir_def *max_num_out_vtx, nir_def *max_num_ou
 
       ac_nir_export_position(b, s->options->hw_info->gfx_level,
                              s->options->export_clipdist_mask,
+                             s->options->dont_export_cull_distances,
                              s->options->write_pos_to_clipvertex,
                              s->options->pack_clip_cull_distances,
                              !s->options->has_param_exports,

src/amd/common/nir/ac_nir_lower_ngg_mesh.c

@@ -887,7 +887,7 @@ emit_ms_vertex(nir_builder *b, nir_def *index, nir_def *row, bool exports, bool
    ms_emit_arrayed_outputs(b, index, per_vertex_outputs, s);
 
    if (exports) {
-      ac_nir_export_position(b, s->hw_info->gfx_level, s->clipdist_enable_mask, false, false,
+      ac_nir_export_position(b, s->hw_info->gfx_level, s->clipdist_enable_mask, false, false, false,
                              !s->has_param_exports, false,
                              s->per_vertex_outputs | VARYING_BIT_POS, &s->out, row);
    }

src/amd/common/nir/ac_nir_prerast_utils.c

@@ -290,6 +290,7 @@ void
 ac_nir_export_position(nir_builder *b,
                        enum amd_gfx_level gfx_level,
                        uint32_t export_clipdist_mask,
+                       bool dont_export_cull_distances,
                        bool write_pos_to_clipvertex,
                        bool pack_clip_cull_distances,
                        bool no_param_export,
@@ -302,6 +303,11 @@ ac_nir_export_position(nir_builder *b,
    unsigned exp_num = 0;
    unsigned exp_pos_offset = 0;
 
+   if (dont_export_cull_distances) {
+      export_clipdist_mask &= ~BITFIELD_RANGE(b->shader->info.clip_distance_array_size,
+                                              b->shader->info.cull_distance_array_size);
+   }
+
    uint64_t mask =
       VARYING_BIT_PSIZ |
       VARYING_BIT_EDGE |