perf(render3d): mesh-shader grass dispatches each tile with its own count; not yet counted as implemented

A chunk's dispatch was sized for its largest tile, so the far tiles beside a
near one ran thousands of empty invocations: 9.8 ms of grass at 4K on an RTX
3070 Ti. One dispatch per tile, sized to that tile, brings it to 5.0 ms - still
three times the chunked path's 1.7 (36 fps against 41), so GF_MESH_GRASS is not
implemented yet and the Advanced row says a coming update.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

changes/mesh-grass.md

@@ -1,9 +1,12 @@
 bump: minor
 type: feat
-**Mesh-shader grass** — on a card with `VK_EXT_mesh_shader`, `r3d_mesh_grass(on)` draws each chunk of grass
-tiles as one mesh-shader dispatch, and a blade that is culled emits no vertices at all.
+**Mesh-shader grass** — on a card with `VK_EXT_mesh_shader`, `r3d_mesh_grass(on)` draws each grass tile as one
+mesh-shader dispatch sized to that tile's blades, and a blade that is culled emits no vertices at all.
 
 - **`ludic-dev shaders`** — a variant whose first file is `*.mesh` compiles that stage as a Vulkan 1.3 mesh
   shader; its SPIR-V keeps the `.vert` name, so the manifest is unchanged.
 - **The seam** — `gpu_has_mesh()` and `gpu_draw_mesh_tasks(x, y, z)`; a mesh program's pipeline has no
   vertex input and its bindings the mesh stage. The device enables `meshShader` and `maintenance4`.
+
+Not yet faster: at 4K on an RTX 3070 Ti the grass pass takes 5.0 ms against the chunked path's 1.7, so
+`gpu_feature_implemented(GF_MESH_GRASS)` stays false and a game should not offer it as a finished setting.

packages/ludic.render3d/gpu.ludic

@@ -392,9 +392,10 @@ function gpu_caps_probe() -> void {
 # Whether the renderer actually draws a feature yet. Until a feature lands, choosing it is saved and
 # shown, and says it takes effect later. The Vulkan renderer draws the whole game (phase 37-38), and
 # DLSS super resolution and Reflex run through NVIDIA Streamline (streamline.ludic), and HDR output is an
-# HDR10 swapchain (gpu_vk_draw.ludic); whether this
+# HDR10 swapchain (gpu_vk_draw.ludic). Mesh-shader grass draws but is not yet counted: at 4K on an
+# RTX 3070 Ti its grass pass took 5.0 ms against the chunked path's 1.7. Whether this
 # machine can use one is the caps' question, not this one.
-function gpu_feature_implemented(f: int) -> bool { return f == GF_VULKAN or f == GF_DLSS or f == GF_REFLEX or f == GF_HDR or f == GF_MESH_GRASS }
+function gpu_feature_implemented(f: int) -> bool { return f == GF_VULKAN or f == GF_DLSS or f == GF_REFLEX or f == GF_HDR }
 
 # ---- vertex data --------------------------------------------------------------------
 # A Mesh is built through these and records what it is made of - which buffer feeds which

packages/ludic.render3d/grass.ludic

@@ -229,18 +229,26 @@ function grass_flush() -> void {
       var m = e - k
       if m > GRASS_CHUNK { m = GRASS_CHUNK }
       for q in 0 .. m * 4 { grass_chunk_tv[q] = grass_tv[k * 4 + q] }
-      u_f4v(gpu_uniform(p, "u_tiles"), m, grass_chunk_tv)
       if mesh {
-        # enough blade batches for the chunk's largest tile, capped as the instanced path caps a tile
+        # one dispatch per tile, as many blade batches as that tile has: sized for a chunk's largest
+        # tile, the far tiles beside a near one ran thousands of empty invocations (9.8 ms against the
+        # chunked path's 2 at 4K on an RTX 3070 Ti)
         let cells = grass_band_cells[b]
-        var most = 0
-        for q in 0 .. m { let t = f_to_int(grass_tv[(k + q) * 4 + 2]) * cells * cells;  if t > most { most = t } }
-        if most > 65535 { most = 65535 }
-        gpu_draw_mesh_tasks((most + GRASS_MESH_BLADES - 1) / GRASS_MESH_BLADES, m, 1)
+        for q in 0 .. m {
+          var total = f_to_int(grass_tv[(k + q) * 4 + 2]) * cells * cells
+          if total > 65535 { total = 65535 }
+          if total > 0 {
+            for c in 0 .. 4 { grass_chunk_tv[c] = grass_tv[(k + q) * 4 + c] }
+            u_f4v(gpu_uniform(p, "u_tiles"), 1, grass_chunk_tv)
+            gpu_draw_mesh_tasks((total + GRASS_MESH_BLADES - 1) / GRASS_MESH_BLADES, 1, 1)
+            grass_draws += 1
+          }
+        }
       } else {
+        u_f4v(gpu_uniform(p, "u_tiles"), m, grass_chunk_tv)
         gpu_draw_mesh_indirect(grass_mesh, grass_cmds, k * 20, m, 0, 0)
       }
-      grass_draws += 1
+      if not mesh { grass_draws += 1 }
       k += m
     }
   }