Per-element atomicity of vector load/store and gather/scatter?

Question

Consider an array like atomic shared_array[]. What if you want to SIMD vectorize for(...) sum += shared_array[i].load(memory_order_relaxed