Mesa (staging/22.0): radv: Disable NGG for GS with suboptimal output vertex count.

[GitLab Mirror]

Module: Mesa
Branch: staging/22.0
Commit: 83f738516d4c93baa6007421f8d95396de42da36
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=83f738516d4c93baa6007421f8d95396de42da36

Author: Timur Kristóf <timur.kristof at gmail.com>
Date:   Wed Apr  6 14:15:50 2022 +0200

radv: Disable NGG for GS with suboptimal output vertex count.

When GS has a lot of output vertices: In this case,
the occupancy of NGG GS is very low because API GS invocations
can't even occupy a single Wave32 wave.
Therefore the legacy pipeline performs better here.

Cc: mesa-stable
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/6260
Signed-off-by: Timur Kristóf <timur.kristof at gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15767>
(cherry picked from commit 94706601fa2f52605d6e488f30fad9a0e2440612)

---

 .pick_status.json              |  2 +-
 src/amd/vulkan/radv_pipeline.c | 13 +++++++++++++
 2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/.pick_status.json b/.pick_status.json
index 2b9c77885fa..617a040281c 100644
--- a/.pick_status.json
+++ b/.pick_status.json
@@ -1492,7 +1492,7 @@
         "description": "radv: Disable NGG for GS with suboptimal output vertex count.",
         "nominated": true,
         "nomination_type": 0,
-        "resolution": 0,
+        "resolution": 1,
         "because_sha": null
     },
     {
diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index 8bfac4d55f7..a60ecaabc25 100644
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -3086,6 +3086,19 @@ radv_fill_shader_info(struct radv_pipeline *pipeline,
          infos[MESA_SHADER_TESS_EVAL].is_ngg = false;
       }
 
+      if (nir[MESA_SHADER_GEOMETRY] &&
+          nir[MESA_SHADER_GEOMETRY]->info.gs.vertices_out >= 9) {
+         /* GS has suboptimal number of output vertices. In this case,
+          * the occupancy of NGG GS is very low, and API GS invocations
+          * can't even occupy a single Wave32 wave.
+          * Therefore the legacy pipeline performs better here.
+          */
+         if (nir[MESA_SHADER_TESS_EVAL])
+            infos[MESA_SHADER_TESS_EVAL].is_ngg = false;
+         else
+            infos[MESA_SHADER_VERTEX].is_ngg = false;
+      }
+
       gl_shader_stage last_xfb_stage = MESA_SHADER_VERTEX;
 
       for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {