ggml : use std::sort in ggml_argsort CPU implementation

[ggerganov]

Use std::sort in ggml_compute_forward_argsort_f32() instead of bubble sort.

[slaren]

Yeah std::function can be quite heavy, it is better to avoid it if possible.

[ggerganov]

In the past I've used an alloc-free alternative of std::function (based on this), but it's quite cumbersome (nvm, not relevant for this use case). Will probably just implement qsort and avoid std::sort all together.

[wishstudio]

Both qsort() and std::function involves at least one indirect function call.
Perhaps better to simply inline the std::sort() call in the switch block as it's just an one-liner:

case GGML_SORT_ORDER_ASC: std::sort(dst_data, dst_data + ne0, [src_data](int32_t a, int32_t b) { return src_data[a] < src_data[b]; }); break;

This will specialize the std::sort template and avoids indirect function call.

[slaren]

An alternative would be to do something like this:

template<enum ggml_sort_order order>
struct sort_cmp {
    const float * data;
    bool operator()(int32_t a, int32_t b) const {
        if constexpr (order == GGML_SORT_ORDER_ASC) {
            return data[a] < data[b];
        } else {
            return data[a] > data[b];
        }
    }
};
switch (order) {
    case GGML_SORT_ORDER_ASC:
        std::sort(dst_data, dst_data + ne0, sort_cmp<GGML_SORT_ORDER_ASC>{src_data});
        break;

    case GGML_SORT_ORDER_DESC:
        std::sort(dst_data, dst_data + ne0, sort_cmp<GGML_SORT_ORDER_DESC>{src_data});
        break;

    default: GGML_ABORT("invalid sort order");
}