Handle arbitrary size in encoded path as well.

Themaister · Jan 13, 2024 · b890855 · b890855
1 parent 1ea6798
commit b890855
Show file tree

Hide file tree

Showing 3 changed files with 36 additions and 29 deletions.
diff --git a/assets/shaders/inc/meshlet_primitive_cull.h b/assets/shaders/inc/meshlet_primitive_cull.h
@@ -168,12 +168,12 @@ uint meshlet_get_meshlet_index()
 	return gl_WorkGroupSize.y == 8 ? gl_WorkGroupID.x : gl_WorkGroupID.y;
 }
 
-uint meshlet_get_sublet_index(uint meshlet_index, uint sublet_index)
+uint meshlet_get_sublet_index(uint sublet_index)
 {
 	if (gl_WorkGroupSize.y == 8)
-		return 8u * meshlet_index + sublet_index;
+		return sublet_index;
 	else
-		return 8u * meshlet_index + gl_WorkGroupSize.y * gl_WorkGroupID.x + sublet_index;
+		return gl_WorkGroupSize.y * gl_WorkGroupID.x + sublet_index;
 }
 
 void meshlet_emit_primitive(uvec3 prim, vec4 clip_pos, vec4 viewport)

diff --git a/tests/assets/shaders/meshlet_debug.mesh b/tests/assets/shaders/meshlet_debug.mesh
@@ -79,10 +79,11 @@ layout(set = 0, binding = 10) buffer Stats
     uint vert;
 } stats;
 
+const uint STRIDE = 4; // TODO: Spec constant or push constant
+
 void main()
 {
     uint compacted_meshlet_index = meshlet_get_meshlet_index() + 0x8000 * gl_DrawIDARB;
-    uint linear_index = gl_LocalInvocationIndex;
 
 #if defined(MESHLET_RENDER_TASK_HIERARCHICAL) && !MESHLET_RENDER_TASK_HIERARCHICAL
     CompactedDrawInfo task = mesh_payload.info;
@@ -91,35 +92,38 @@ void main()
     CompactedDrawInfo task = mesh_payload.infos[compacted_meshlet_index];
 #endif
 
-    const uint STRIDE = 4;
-    MeshletMetaRuntime meta = meshlet_metas_runtime.data[task.meshlet_index];
-    meta.stream_offset += STRIDE * gl_WorkGroupID.x;
-
     mat4 M = transforms.data[task.node_offset];
 
-    uint lane_index;
-
+    uint linear_index, sublet_index;
     if (gl_SubgroupSize == 32)
-        lane_index = gl_SubgroupInvocationID;
+    {
+        linear_index = gl_SubgroupInvocationID;
+        sublet_index = gl_SubgroupID;
+    }
     else
-        lane_index = gl_LocalInvocationID.x;
+    {
+        linear_index = gl_LocalInvocationID.x;
+        sublet_index = gl_LocalInvocationID.y;
+    }
 
-    meshlet_setup_local_invocation(uvec2(lane_index, 0));
+    meshlet_setup_local_invocation(uvec2(linear_index, sublet_index));
+    MeshletMetaRuntime meta = meshlet_metas_runtime.data[task.meshlet_index];
+    meta.stream_offset += STRIDE * meshlet_get_sublet_index(sublet_index);
 
     MeshletInfo info = meshlet_get_meshlet_info(meta.stream_offset);
     uvec3 decoded_index_buffer = uvec3(0);
     vec3 world_pos;
     vec4 clip_pos = vec4(-1.0);
 
-    if (lane_index < info.primitive_count)
-        decoded_index_buffer = meshlet_decode_index_buffer(meta.stream_offset, lane_index);
+    if (linear_index < info.primitive_count)
+        decoded_index_buffer = meshlet_decode_index_buffer(meta.stream_offset, linear_index);
 
-    if (lane_index < info.vertex_count)
+    if (linear_index < info.vertex_count)
     {
         int exponent;
         i16vec3 ipos = meshlet_decode_snorm_scaled_i16x3(
             meta.stream_offset + MESHLET_STREAM_TYPE_POSITION,
-            lane_index, exponent);
+            linear_index, exponent);
         vec3 pos = ldexp(vec3(ipos), ivec3(exponent));
         world_pos = (M * vec4(pos, 1.0)).xyz;
         clip_pos = VP * vec4(world_pos, 1.0);
@@ -135,10 +139,10 @@ void main()
         bool t_sign;
         u8vec4 nt = meshlet_decode_normal_tangent_oct8(
                 meta.stream_offset + MESHLET_STREAM_TYPE_NORMAL_TANGENT_OCT8,
-                lane_index, t_sign);
+                linear_index, t_sign);
         i16vec2 uv = meshlet_decode_snorm_scaled_i16x2(
                 meta.stream_offset + MESHLET_STREAM_TYPE_UV,
-                lane_index, exponent);
+                linear_index, exponent);
 
 #ifdef MESHLET_PRIMITIVE_CULL_SHARED_INDEX
         shared_clip_pos[out_vert_index] = clip_pos;

diff --git a/tests/assets/shaders/meshlet_debug_plain.mesh b/tests/assets/shaders/meshlet_debug_plain.mesh
@@ -119,15 +119,20 @@ void main()
     CompactedDrawInfo task = mesh_payload.infos[compacted_meshlet_index];
 #endif
 
-#if defined(MESHLET_PRIMITIVE_CULL_WAVE32) && MESHLET_PRIMITIVE_CULL_WAVE32
-    uint linear_index = gl_SubgroupInvocationID;
-    uint sublet_index = gl_SubgroupID;
-#else
-    uint linear_index = gl_LocalInvocationID.x;
-    uint sublet_index = gl_LocalInvocationID.y;
-#endif
+    uint linear_index, sublet_index;
+    if (gl_SubgroupSize == 32)
+    {
+        linear_index = gl_SubgroupInvocationID;
+        sublet_index = gl_SubgroupID;
+    }
+    else
+    {
+        linear_index = gl_LocalInvocationID.x;
+        sublet_index = gl_LocalInvocationID.y;
+    }
 
-    sublet_index = meshlet_get_sublet_index(task.meshlet_index, sublet_index);
+    meshlet_setup_local_invocation(uvec2(linear_index, sublet_index));
+    sublet_index = 8u * task.meshlet_index + meshlet_get_sublet_index(sublet_index);
     IndirectDrawMesh meshlet = indirect_commands_mesh.draws[sublet_index];
 
     mat4 M = transforms.data[task.node_offset];
@@ -137,8 +142,6 @@ void main()
     vec3 world_pos = (M * vec4(pos, 1.0)).xyz;
     vec4 clip_pos = VP * vec4(world_pos, 1.0);
 
-    meshlet_setup_local_invocation(gl_LocalInvocationID.xy);
-
     uvec3 prim = uvec3(0);
     if (linear_index < meshlet.primitive_count)
         prim = uvec3(ibo.data[meshlet.primitive_offset + linear_index]);