Bitwise-exact VFPU instructions #3
Description
Posting here, for extra focus (the previous discussion, on Discord and elsewhere, kind of didn't go far).
The original topic, on PPSSPP's side: hrydgard/ppsspp#21070.
The data from PPSSPP's (presumed) bitwise-exact VFPU instructions could be incorporated into vfpu-docs.
The instructions are not exhaustively tested, but there is also their source data, obtained from actual PSP (now I've collected the links):
hrydgard/ppsspp#21070 (comment)
Aside: would it make sense to host the copy of the above data on vfpu-docs?
This can be used to e.g. sharpen error bounds written in vfpu-docs (though current ones are rather good), as well as point out some details (e.g. vsin being stair-step for small x, in particular, 0 for |x|<2^(-23)). Some oddities, like peculiar error plot of vasin, or the number of trailing zeroes in vlog2 (as function of input exponent), can be mentioned too.
As mentioned, the PPSSPP's code for these is straightforward to extract into something self-contained. I have such version, as a single file (not counting the tables), in case you need it.
And, as mentioned in PPSSPP's topic, I would very much like to have the data dumps for other instructions (e.g. vdot and vdiv, see that topic for specific requests), so they can be investigated too, in case anyone wants to provide them, as I myself do not have access to physical PSP.