vload on Matrix{Float64}

[algorithmx]

Dear developers,

I need to access a Julia matrix M of size (4,ncol) via vload function in a performance-critical situation, where converting the matrix M to a vector by vec(M) is prohibited. How can I properly use vload to access the k-th column of M, given that this matrix is properly padded (i.e. all(M[4,:] .== 0) == true )? If I cannot do this, from where can I extent the vload function?

#feature-request

[KristofferC]

Maybe:

julia> function load_column(x, i)
           GC.@preserve x begin
               ptr = pointer(x, 4*(i-1)+1)
               SIMD.vload(Vec{4, Float64}, ptr)
           end
       end
load_column (generic function with 1 method)

julia> x = rand(4, 10)
4×10 Matrix{Float64}:
 0.105425  0.112207   0.164943  0.924851     0.962439  0.771419   0.232943  0.0477578  0.31495   0.725615
 0.583358  0.574841   0.682031  0.0163499    0.889354  0.810443   0.343777  0.372169   0.841392  0.0536699
 0.831974  0.78172    0.424055  0.000215734  0.176126  0.177913   0.842223  0.706929   0.837163  0.591129
 0.567482  0.0297791  0.552873  0.123447     0.485557  0.0712733  0.925538  0.928424   0.719696  0.236411

julia> load_column(x, 1)
<4 x Float64>[0.10542510183253728, 0.5833583593228673, 0.8319741449401951, 0.5674824353491745]

julia> load_column(x, 5)
<4 x Float64>[0.9624389422624501, 0.8893543320414525, 0.17612558720563265, 0.4855567380042497]

[algorithmx]

Maybe:

julia> function load_column(x, i)
           GC.@preserve x begin
               ptr = pointer(x, 4*(i-1)+1)
               SIMD.vload(Vec{4, Float64}, ptr)
           end
       end
load_column (generic function with 1 method)

julia> x = rand(4, 10)
4×10 Matrix{Float64}:
 0.105425  0.112207   0.164943  0.924851     0.962439  0.771419   0.232943  0.0477578  0.31495   0.725615
 0.583358  0.574841   0.682031  0.0163499    0.889354  0.810443   0.343777  0.372169   0.841392  0.0536699
 0.831974  0.78172    0.424055  0.000215734  0.176126  0.177913   0.842223  0.706929   0.837163  0.591129
 0.567482  0.0297791  0.552873  0.123447     0.485557  0.0712733  0.925538  0.928424   0.719696  0.236411

julia> load_column(x, 1)
<4 x Float64>[0.10542510183253728, 0.5833583593228673, 0.8319741449401951, 0.5674824353491745]

julia> load_column(x, 5)
<4 x Float64>[0.9624389422624501, 0.8893543320414525, 0.17612558720563265, 0.4855567380042497]

Thanks @KristofferC, problem solved! The performance difference is narrow for the two versions of my code, one with flattened matrix and the other with the original matrix plus your suggestion for load_column(x, i). Here is the code I am working on and the benchmark on a 4x10 matrix.

[algorithmx]

GC.@preserve does the job.
Reference:
https://docs.julialang.org/en/v1/base/base/#Base.GC.@preserve
https://docs.juliahub.com/ManualMemory/rywzg/0.1.8/