|
does anyone actually need half precision (fp16) facilities and does any plugin actually provide support for it? I don't really wanna have it in my project and I have very good reasons:
1) there's no native support for fp16 in standard C/C++ and most CPUs provide very little or no support for it, meaning you will have to create your homegrown software-emulated fp16 type. it's not portable and it's very likely that it will be way slower than fp32 which comes with native hardware support.
2) unlike fp32 and fp64 which are universally well defined (literally every implementation out there adopts the representation defined by IEEE 754), the representation of fp16 varies on different platforms. there's the IEEE fp16, and BFloat16 (Google TPU), and who knows what else! applying arithmetic operations defined for one fp16 representation on another fp16 representation leads to garbage result. The user (or maybe even the developer) usually has absolutely no clue which fp16 representation is being used and all sorts of weird shit can happen. It might be better to avoid all this by simply not supporting fp16.
|