Update docs and add descriptive error message (#128)

Author: efaulhaber

src/cell_lists/full_grid.jl

@@ -1,6 +1,7 @@
 """
-    FullGridCellList(; min_corner, max_corner, search_radius = 0.0,
-                     periodicity = false, backend = DynamicVectorOfVectors{Int32},
+    FullGridCellList(; min_corner, max_corner,
+                     search_radius = zero(eltype(min_corner)),
+                     backend = DynamicVectorOfVectors{Int32},
                      max_points_per_cell = 100)
 
 A simple cell list implementation where each (empty or non-empty) cell of a rectangular
@@ -15,12 +16,8 @@ See [`copy_neighborhood_search`](@ref) for more details.
 # Keywords
 - `min_corner`: Coordinates of the domain corner in negative coordinate directions.
 - `max_corner`: Coordinates of the domain corner in positive coordinate directions.
-- `search_radius = 0.0`: Search radius of the neighborhood search, which will determine the
-                         cell size. Use the default of `0.0` to create a template (see above).
-- `periodicity = false`: Set to `true` when using a [`PeriodicBox`](@ref) with the
-                         neighborhood search. When using [`copy_neighborhood_search`](@ref),
-                         this option can be ignored an will be set automatically depending
-                         on the periodicity of the neighborhood search.
+- `search_radius`: Search radius of the neighborhood search, which will determine the
+                   cell size. Use the default of zero to create a template (see above).
 - `backend = DynamicVectorOfVectors{Int32}`: Type of the data structure to store the actual
     cell lists. Can be
     - `Vector{Vector{Int32}}`: Scattered memory, but very memory-efficient.

src/neighborhood_search.jl

@@ -131,6 +131,8 @@ struct PeriodicBox{NDIMS, ELTYPE}
     end
 end
 
+@inline Base.eltype(::PeriodicBox{<:Any, ELTYPE}) where {ELTYPE} = ELTYPE
+
 """
     foreach_point_neighbor(f, system_coords, neighbor_coords, neighborhood_search;
                            parallelization_backend = default_backend(system_coords),

src/nhs_grid.jl

@@ -53,6 +53,13 @@ since not sorting makes our implementation a lot faster (although less paralleli
   "A Parallel SPH Implementation on Multi-Core CPUs".
   In: Computer Graphics Forum 30.1 (2011), pages 99–112.
   [doi: 10.1111/J.1467-8659.2010.01832.X](https://doi.org/10.1111/J.1467-8659.2010.01832.X)
+
+!!! note "Note"
+    The type of `search_radius` determines the type used for the internals of the
+    neighborhood search.
+    When working with single precision, the `search_radius` must be `Float32` in order
+    to avoid the use of double precision values (for example when working with GPUs).
+    When using a `periodic_box`, its type must match the type of the `search_radius`.
 """
 struct GridNeighborhoodSearch{NDIMS, US, CL, ELTYPE, PB, UB} <: AbstractNeighborhoodSearch
     cell_list       :: CL
@@ -84,6 +91,11 @@ function GridNeighborhoodSearch{NDIMS}(; search_radius = 0.0, n_points = 0,
         n_cells = ntuple(_ -> -1, Val(NDIMS))
         cell_size = ntuple(_ -> search_radius, Val(NDIMS))
     else
+        if typeof(search_radius) != eltype(periodic_box)
+            throw(ArgumentError("the `search_radius` and the `PeriodicBox` must have " *
+                                "the same element type"))
+        end
+
         # Round up search radius so that the grid fits exactly into the domain without
         # splitting any cells. This might impact performance slightly, since larger
         # cells mean that more potential neighbors are considered than necessary.