XMVectorRound doesn't match roundf

[walbourn]

The mm_round_ps intrinsic performs ties to even (a.k.a. banker's rounding): round(0.5) = 0.0 , round(1.5) = 2.0 .

The C99 roundf function performs ties away from zero: round(0.5) = 1.0 , round(1.5) = 2.0 .

[walbourn]

Per the Intel SSE4 guide, _mm_round_ps by itself can match ceil, floor, trunc, rint, and nearbyint. You need a little more to get it to match roundf.

[walbourn]

ARM v8's FRINTN also performs banker's rounding, which I'll be using for the ARM64 implementation.

[walbourn]

The bug here is actually that the NO_INTRINSICS path on VS 2013+ should not be using roundf because all the other implementations are using round-to-nearest (even). It might be nice to offer an alternative round function that uses the roundf half-way round away from zero rules, but such a round wouldn't be a single instruction for SSE4 or ARM-NEON--filed as new issue.

[walbourn]

Fixed for DirectXMath 3.10