Avoiding scalar indexing in cuda

[yuehhua]

Existing implementation use scalar indexing for cuarray. To avoid using scalar indexing, refactoring is done.

[ToucheSir]

The CI failure appears to be unrelated to this change? AFAICT none of the conv machinery uses maximum_dims.

[yuehhua]

maximum_dims is used in scatter operations in scatter.jl. Then, I am sure that the CI failure is unrelated to this change.

[DhairyaLGandhi]

Mostly looks good.
Why not use maximum(dims)?

[DhairyaLGandhi]

Reopening to see what's up with CI

[DhairyaLGandhi]

Yeah, seems CI is unrelated. @yuehhua happy to merge if you can address replacing this with maximum

[yuehhua]

The functionality of this function is to calculate the maximum values across indices. For example, I have a array of tuples that contains a series of indices.

julia> A = [(5, 2), (3, 4), (1, 6)];

julia> maximum_dims(dims::AbstractArray{NTuple{N, T}}) where {N,T} = Tuple(maximum(x->x[i], dims) for i = 1:N);

julia> maximum_dims(A)
(5, 6)

julia> maximum(A)
(5, 2)

The maximum value of first component across indices is 5 and the maximum value of second component is 6. So, I want maximum_dims to return me the tuple (5, 6), while maximum give me the element which has maximum value in its first component.

[DhairyaLGandhi]

Thanks all