Extend support of batch_cast<...> to upcasting to a type twice as big

[serge-sans-paille]

Fix #1179

[serge-sans-paille]

@AntoinePrv : this is just some mockup to check if we're good with the API

[serge-sans-paille]

Also cc @JohanMabille & @DiamonDinoia for the API. I didn't want to support the generic conversion, say from uint8 to uint32 with larger arrays, let's start humble

[DiamonDinoia]

I would call the API extension, extend or widen because it does not just cast, it zero extend and 1 extend. I think in intrinsics terminology is called widening and convert with widening.

On the implementation:

1. using std::conditional and std::make_signed halves the number of struct you need.
   2. Proposal load/store masked #1162 introduces lower/upper :)

namespace {
template<class T> struct widen;

template<> struct widen<uint8_t>  { using type = uint16_t; };
template<> struct widen<uint16_t> { using type = uint32_t; };
template<> struct widen<uint32_t> { using type = uint64_t; };

template<class T>
using widen_t_unsigned = typename widen<T>::type;

} // namespace 

template<class T>
struct widen {
    static_assert(std::is_integral<T>::value && !std::is_same<T,bool>::value,
                  "integral non-bool type required");

    using U     = typename std::make_unsigned<T>::type;
    using UWide = detail::widen_t_unsigned<U>;

    using type = typename std::conditional<
        std::is_signed<T>::value,
        typename std::make_signed<UWide>::type,
        UWide
    >::type;
};

template<class T>
using widen_t = typename widen<T>::type;

[serge-sans-paille]

@DiamonDinoia yep, widen looks like a good name, thanks!

[AntoinePrv]

I think a trait that convert a byte size to a uint/float type could be more generic:

template <>
struct sized_uint<4>
{
    using type = uint32_t;
};

...

And then use with batch<sized_uint_t<2*sizeof(T)>> (or an alias for that).

[DiamonDinoia]

I am not in favor of this. Then we need sized_int and sized_uint and the API will need an std::conditional<std::is_signed<T>, sized_int_t<2*sizeof(T)>, sized_uint_t<2*sizeof(T)>

[AntoinePrv]

Thanks for taking the time! Looking good

[DiamonDinoia]

My only suggestion would be to merge masked_memory ops. Or merge lower/upper from it and use it here.

[serge-sans-paille]

My only suggestion would be to merge masked_memory ops. Or merge lower/upper from it and use it here.

Yup, I'll wait for it to be merged!

[DiamonDinoia]

out_buffer
and out_buffer + batch<T_out, A>::size seems clearer to me

[DiamonDinoia]

(I mean doing pointer arithmetic directly instead of using the [] and & operators)

[JohanMabille]

LGTM

[serge-sans-paille]

My only suggestion would be to merge masked_memory ops. Or merge lower/upper from it and use it here.

@DiamonDinoia : Are you fine if I split (haha) that part of your commit in a seperate one so that we don't block each other?

[DiamonDinoia]

My only suggestion would be to merge masked_memory ops. Or merge lower/upper from it and use it here.

@DiamonDinoia : Are you fine if I split (haha) that part of your commit in a seperate one so that we don't block each other?

Sure, go ahead. I'll rebase once I have time

[serge-sans-paille]

My only suggestion would be to merge masked_memory ops. Or merge lower/upper from it and use it here.

done (as in: I created a separate PR to extract that part of your patch, merged it and rebased my PR on top of it)