I don't understand why it is slower with CuStaticSharedArray

New to Julia
, , ,
4

Thanks. Such diagrams are always useful.

But regarding the code, could you provide it in a form we can just run it by copy-pasting? I.e. something of the form

using CUDA, BenchmarkTools
(...)

const Ti = Int32
const Tf = Float64

const Δ::Tf = Tf(1e-2)  # (we don't know what your df contains)
(...)

function kernel_comp_v_noshmem!(...)
    (...)
end

function kernel_comp_v!(...)
    (...)
end

v = (...)  # e.g. CUDA.rand(...)
(...)

display(@benchmark CUDA.@sync begin 
   kernel_comp_v_noshmem!!($v_temp, $v, $F; threads = block_dim, blocks = grid_dim)
end)
display(@benchmark CUDA.@sync begin 
   comp_v!($v_temp, $v, $F; threads = block_dim, blocks = grid_dim)
end)

See also (point 4 in)