Smart compaction goes brrrrr

shaders/compaction.slang

@@ -113,6 +113,7 @@ void scatter(
     uint32_t3 local_thread_id: SV_GroupThreadID,
     uint32_t3 global_thread_id: SV_DispatchThreadID)
 {
+    let total = index_buffer.getCount();
     let thread_index = global_thread_id.x;
     let local_thread_index = local_thread_id.x;
 
@@ -123,12 +124,12 @@ void scatter(
     let sum2 = sum_buffer[thread_index * 2 + 1] + reduced_value;
 
     // Scatter index to compact index buffer if predicate is true
-    if (count_buffer[thread_index * 2] != 0)
+    if (count_buffer[thread_index * 2] != 0 && sum1 < total)
     {
         index_buffer[sum1] = thread_index * 2;
     }
 
-    if (count_buffer[thread_index * 2 + 1] != 0)
+    if (count_buffer[thread_index * 2 + 1] != 0 && sum2 < total)
     {
         index_buffer[sum2] = thread_index * 2 + 1;
     }