Taronites

Too much theory for me for making decent suggestions, unfortunately.
I just could throw in nonsense vector displacement, what can’t be possible.
I really can’t do more than wait.
Wait.
Wait.

I have no time anyway, so please, keep it a bit longer unpublished

Wehehell, consider that the whole thing (fluid dynamics) manifest via vector displacement, it's not "nonsensical", but possibly redundant, hahaha. But, yeah, I think, I know what you mean.
Cool, I will try to make a really nice implementation of the new loader options!

Taron, congrats from me too!

16b is word, why you gather the bytes into dwords?

It's "Short" in BlitzMax and I just remembered seeing an old c++ code where we used DWORD buffers, because we generated floating point colors, floats being a double word (32 bit). Then in all my joy and euphoria I didn't think much, writing the happy post, haha! ...thanks, though, I still feel like a fool more often than not.
But what REALLY puzzles me, still, is why RGBA16F isn't keeping the information of the 16bit word or even floats themselves, if you draw the texture into a 32F buffer, but if you put the exact same data directly into a 32F buffer is all is well?! Somehow strange to me?!
There are still quite a few "magical" things to me left to discover.

Not sure if I understand your question correctly but maybe byte order is the key of all problems? I mean how word, dword, float are stored and where is MSB (most significant bit). There can be different implementations..

No, no, the byte order is correct. Look at the two examples I've posted. The first is with 16F texture and the second through 32F texture. The 16F shows pixelated or quantized stepping, though, I tried to smudge it down. 32F is perfect, showing the polygons, actually, hehe. No pixel stepping.

Here... I put them next to each other.
16F could possibly be half the value range of 32F, because it still is miles above only 256 levels of gray. I haven't measured it, yet, but well... it is interesting.

FOR TARON
Blitz is not dead yet
I think there is something else on the github.com/blitz-research of Mark Sibly
Some monkey2