Pad subpair cache.

src/cache.h

@@ -239,7 +239,7 @@ __attribute__((always_inline)) INLINE void cache_read_particles_subpair(
   /* Shift the particles positions to a local frame so single precision can be
    * used instead of double precision. */
   for (int i = 0; i < ci->count; i++) {
-    const int idx = sort_i[i].i;
+    const int idx = i;//sort_i[i].i;
     x[i] = (float)(parts[idx].x[0] - loc[0]);
     y[i] = (float)(parts[idx].x[1] - loc[1]);
     z[i] = (float)(parts[idx].x[2] - loc[2]);
@@ -250,6 +250,23 @@ __attribute__((always_inline)) INLINE void cache_read_particles_subpair(
     vz[i] = parts[idx].v[2];
   }
 
+  const double max_dx = ci->dx_max_part;
+  const float pos_padded[3] = {-(2. * ci->width[0] + max_dx),
+                               -(2. * ci->width[1] + max_dx),
+                               -(2. * ci->width[2] + max_dx)};
+  const float h_padded = ci->parts[0].h;
+
+  for (int i = ci->count; i < ci->count + VEC_SIZE; i++) {
+    x[i] = pos_padded[0];
+    y[i] = pos_padded[1];
+    z[i] = pos_padded[2];
+    h[i] = h_padded;
+
+    m[i] = 1.f;
+    vx[i] = 1.f;
+    vy[i] = 1.f;
+    vz[i] = 1.f;
+  }
 #endif
 }