Optimize expand to avoid one unecessary sdiv

Author: vchuravy

src/nditeration.jl

@@ -80,8 +80,20 @@ Base.length(range::NDRange) = length(blocks(range))
     CartesianIndex(nI)
 end
 
-Base.@propagate_inbounds function expand(ndrange::NDRange, groupidx::Integer, idx::Integer)
-    expand(ndrange, blocks(ndrange)[groupidx], workitems(ndrange)[idx])
+Base.@propagate_inbounds function expand(ndrange::NDRange{N}, groupidx::Integer, idx::Integer) where {N}
+    # This causes two sdiv operations, one for each Linear to CartesianIndex
+    # expand(ndrange, blocks(ndrange)[groupidx], workitems(ndrange)[idx])
+
+    # The formulation below saves one sdiv
+    B = blocks(ndrange)
+    W = workitems(ndrange)
+    Ind = ntuple(Val(N)) do I
+        Base.@_inline_meta
+        b = B.indices[I]
+        w = W.indices[I]
+        length(b) * length(w)
+    end
+    CartesianIndices(Ind)[(groupidx-1)* prod(size(B)) + idx]
 end
 
 Base.@propagate_inbounds function expand(ndrange::NDRange{N}, groupidx::CartesianIndex{N}, idx::Integer) where {N}