vbo: lower VBO_SAVE_BUFFER_SIZE to avoid large VRAM usage

The ideal case for performance is to have a single buffer for
all display list. The caveat is that large buffers are less
likely to be freed because they're refcounted: it only takes
1 user (diplay list) to keep it in VRAM.

This lowers VRAM usage when replaying the trace attached
of the trace attached to !6140 from 5.5 GB to about 1.8 GB.

Viewperf snx performance isn't affected.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/6140
Cc: mesa-stable
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20748>
(cherry picked from commit 0f5c8c3dc35561f23a5383cb05075898942ce0c5)

diff --git a/.pick_status.json b/.pick_status.json
index c9d9a23..cc18589 100644 (file)
--- a/.pick_status.json
+++ b/.pick_status.json
@@ -2938,7 +2938,7 @@
         "description": "vbo: lower VBO_SAVE_BUFFER_SIZE to avoid large VRAM usage",
         "nominated": true,
         "nomination_type": 0,
         "description": "vbo: lower VBO_SAVE_BUFFER_SIZE to avoid large VRAM usage",
         "nominated": true,
         "nomination_type": 0,

-        "resolution": 0,
+        "resolution": 1,

         "main_sha": null,
         "because_sha": null
     },
         "main_sha": null,
         "because_sha": null
     },

diff --git a/src/mesa/vbo/vbo_save.h b/src/mesa/vbo/vbo_save.h
index 78a6340..4c73d92 100644 (file)
--- a/src/mesa/vbo/vbo_save.h
+++ b/src/mesa/vbo/vbo_save.h
@@ -111,8 +111,12 @@ _vbo_save_get_stride(const struct vbo_save_vertex_list *node)
 /* Default size for the buffer holding the vertices and the indices.
  * A bigger buffer helps reducing the number of draw calls but may
  * waste memory.
 /* Default size for the buffer holding the vertices and the indices.
  * A bigger buffer helps reducing the number of draw calls but may
  * waste memory.

+ * 1MB was picked because a lower value reduces viewperf snx tests
+ * performance but larger values cause high VRAM usage (because
+ * larger buffers will be shared by more display lists which reduces
+ * the likelyhood of freeing the buffer).

  */
  */

-#define VBO_SAVE_BUFFER_SIZE (20*1024*1024)
+#define VBO_SAVE_BUFFER_SIZE (1024 * 1024)

 #define VBO_SAVE_PRIM_MODE_MASK 0x3f
 
 struct vbo_save_vertex_store {
 #define VBO_SAVE_PRIM_MODE_MASK 0x3f
 
 struct vbo_save_vertex_store {