I think what you are seeing is Slow sparse matrix-vector product with symmetric matrices
julia> @btime mul!($c, $A, $b);
120.341 μs (0 allocations: 0 bytes)
julia> _A = Hermitian(A);
julia> @btime mul!($c, $_A, $b);
37.587 ms (4008 allocations: 187.88 KiB)