diff --git a/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c b/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
index e008c356fe7..255f6c8f379 100644
--- a/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
+++ b/src/gallium/drivers/radeonsi/gfx10_shader_ngg.c
@@ -21,11 +21,12 @@ unsigned gfx10_ngg_get_vertices_per_prim(struct si_shader *shader)
       } else if (shader->key.ge.opt.ngg_culling & SI_NGG_CULL_LINES)
          return 2;
       else {
-         /* We always build up all three indices for the prim export
-          * independent of the primitive type. The additional garbage
-          * data shouldn't hurt. This is used by exports and streamout.
+         /* 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 3;
+         return 0; /* unknown */
       }
    } else {
       assert(shader->selector->stage == MESA_SHADER_TESS_EVAL);
diff --git a/src/gallium/drivers/radeonsi/si_nir_lower_abi.c b/src/gallium/drivers/radeonsi/si_nir_lower_abi.c
index cf1f846842f..ad57e18faa4 100644
--- a/src/gallium/drivers/radeonsi/si_nir_lower_abi.c
+++ b/src/gallium/drivers/radeonsi/si_nir_lower_abi.c
@@ -32,38 +32,6 @@ nir_def *si_nir_load_internal_binding(nir_builder *b, struct si_shader_args *arg
    return nir_load_smem_amd(b, num_components, addr, nir_imm_int(b, slot * 16));
 }
 
-static nir_def *get_num_vert_per_prim(nir_builder *b, struct si_shader *shader,
-                                          struct si_shader_args *args)
-{
-   const struct si_shader_info *info = &shader->selector->info;
-   gl_shader_stage stage = shader->selector->stage;
-
-   unsigned num_vertices;
-   if (stage == MESA_SHADER_GEOMETRY) {
-      num_vertices = mesa_vertices_per_prim(info->base.gs.output_primitive);
-   } else if (stage == MESA_SHADER_VERTEX) {
-      if (info->base.vs.blit_sgprs_amd)
-         num_vertices = 3;
-      else if (shader->key.ge.opt.ngg_culling & SI_NGG_CULL_LINES)
-         num_vertices = 2;
-      else {
-         /* Extract OUTPRIM field. */
-         nir_def *num = GET_FIELD_NIR(GS_STATE_OUTPRIM);
-         return nir_iadd_imm(b, num, 1);
-      }
-   } else {
-      assert(stage == MESA_SHADER_TESS_EVAL);
-
-      if (info->base.tess.point_mode)
-         num_vertices = 1;
-      else if (info->base.tess._primitive_mode == TESS_PRIMITIVE_ISOLINES)
-         num_vertices = 2;
-      else
-         num_vertices = 3;
-   }
-   return nir_imm_int(b, num_vertices);
-}
-
 static nir_def *build_attr_ring_desc(nir_builder *b, struct si_shader *shader,
                                          struct si_shader_args *args)
 {
@@ -411,9 +379,14 @@ static bool lower_intrinsic(nir_builder *b, nir_instr *instr, struct lower_abi_s
       replacement = nir_load_smem_amd(b, 4, addr, nir_imm_int(b, offset));
       break;
    }
-   case nir_intrinsic_load_num_vertices_per_primitive_amd:
-      replacement = get_num_vert_per_prim(b, shader, args);
+   case nir_intrinsic_load_num_vertices_per_primitive_amd: {
+      unsigned num_vertices = gfx10_ngg_get_vertices_per_prim(shader);
+      if (num_vertices)
+         replacement = nir_imm_int(b, num_vertices);
+      else
+         replacement = nir_iadd_imm(b, GET_FIELD_NIR(GS_STATE_OUTPRIM), 1);
       break;
+   }
    case nir_intrinsic_load_cull_ccw_amd:
       /* radeonsi embed cw/ccw info into front/back face enabled */
       replacement = nir_imm_false(b);
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 1aa1b037687..06d22a6cd78 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -1966,8 +1966,9 @@ 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);
 
-      options.num_vertices_per_primitive = gfx10_ngg_get_vertices_per_prim(shader);
+      options.num_vertices_per_primitive = num_vertices ? num_vertices : 3;
       options.early_prim_export = gfx10_ngg_export_prim_early(shader);
       options.passthrough = gfx10_is_ngg_passthrough(shader);
       options.use_edgeflags = gfx10_edgeflags_have_effect(shader);