From c213dbfbf55b872a895bcea799610d41790f02d2 Mon Sep 17 00:00:00 2001
From: Erik Faulhaber <44124897+efaulhaber@users.noreply.github.com>
Date: Mon, 2 Dec 2024 18:02:58 +0100
Subject: [PATCH] [Proof of Concept] Use symmetry to speed up the main loop

---
 src/nhs_grid.jl | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/src/nhs_grid.jl b/src/nhs_grid.jl
index 3b5dd266..2f785a93 100644
--- a/src/nhs_grid.jl
+++ b/src/nhs_grid.jl
@@ -372,6 +372,10 @@ end
         for neighbor_ in eachindex(neighbors)
             neighbor = @inbounds neighbors[neighbor_]
 
+            # Skip half of the neighbors and use symmetry below.
+            # Note that this only works when `system_coords == neighbor_system_coords`.
+            point <= neighbor || continue
+
             # Making the following `@inbounds` yields a ~2% speedup on an NVIDIA H100.
             # But we don't know if `neighbor` (extracted from the cell list) is in bounds.
             neighbor_coords = extract_svector(neighbor_system_coords,
@@ -390,6 +394,9 @@ end
                 # Inline to avoid loss of performance
                 # compared to not using `foreach_point_neighbor`.
                 @inline f(point, neighbor, pos_diff, distance)
+                if point < neighbor
+                    @inline f(neighbor, point, -pos_diff, distance)
+                end
             end
         end
     end