Half-precision floating point

RE: Half-precision floating point

Sigurd — Fri, 24 Jan 2020 16:40:11 GMT

[quote userid="82177" url="~/f/nordic-q-a/25896/half-precision-floating-point/229308"]can we do for example two fp16 operations in parallel?[/quote]

No.

[quote userid="82177" url="~/f/nordic-q-a/25896/half-precision-floating-point/229308"]So the answer is no?[/quote]

I believe fp16 is supported as a storage format only(e.g. set -mfp16-format=alternative in the makefile), but when used in any arithmetic or other operations, it will be automatically converted to fp32

RE: Half-precision floating point

Felix — Wed, 15 Jan 2020 22:01:35 GMT

So the answer is no? can we do for example two fp16 operations in parallel? thanks

RE: Half-precision floating point

Sigurd — Fri, 13 Oct 2017 13:15:53 GMT

Hi,

The nRF52832 uses the ARM Cortex M4 processor with floating-point unit (FPU). The FPU implemented supports Single-precision floating-point. This will give you more precision compared to a half-precision floating-point solution. The FPU provides floating-point computation functionality that is compliant with the IEEE 754 standard.