Re: [PATCH v2 0/5] MR53: vkd3d-shader/hlsl: transpose(), smoothstep(), and properly convert args of element-wise intrinsics.

[Zebediah Figura (＠zfigura)]

On Mon Nov 28 17:56:57 2022 +0000, Francisco Casas wrote:
> Mathematically yes, but the native compiler does a `1/(·)` first, and
> then uses a `mul_sat`.
> So it made sense to me to leave it as DIV + MUL + SAT, just in case we
> ever want to merge the MUL and the SAT.

Interesting. sm5 won't optimize a*(1/b) to a/b, or vice versa. I don't know why, but I guess if there's a difference in precision we should probably emulate native...? @Mystral, do you have thoughts on this?

Note that _sat is just a modifier which saturates the destination after performing whatever opcode. I.e. in theory this could be a single div_sat. No idea why native doesn't do that.