Performance of the splat operator when accessing a matrix-CodePudding

Let's say we have a matrix and a vector of the indices at which we want to access that matrix:

matrix = Matrix{Float64}(undef, 5000, 4000)
point = [1244, 3353]

We can use the splat operator (...) to expand the point vector to access the matrix, or we can explicitly use point[1] and point[2] as the indices:

matrix[point...]
vs
matrix[point[1], point[2]]

The first is more elegant, but apparently also much slower:

@btime $matrix[$point...]
70.083 ns (4 allocations: 64 bytes)

@btime $matrix[$point[1], $point[2]]
1.800 ns (0 allocations: 0 bytes)

Where does this difference come from? Is there a way to make the splat operator more performant or should I simply use the other solution if I care about performance?

CodePudding user response：

Splatting a tuple seems to be the fastest of all:

julia> @btime $matrix[$point...] # point isa Array
  94.484 ns (4 allocations: 64 bytes)
0.0

julia> @btime $matrix[$point[1], $point[2]] # point isa Array
  2.889 ns (0 allocations: 0 bytes)
0.0

julia> @btime $matrix[$point...] # point isa Tuple
  2.624 ns (0 allocations: 0 bytes)
0.0

Variadic function arguments are internally passed as tuples, so it's no surprise there is the least overhead if you directly insert a tuple.