We are running a survey to help us improve the experience for all of our members. If you see the survey appear, please take the time to tell us about your experience if you can.
Is there support for 16 bit half precision floating point in RV compiler?
infocenter.arm.com/.../index.jsp
Are there any examples/sources for conversion between single and half precision (I need single <-> __fp16).
Thanks for help!
Marko
I will answer myself (hope next step is not arguing with myself :) ).
I put this: www.cellperformance.com/.../branchfree_implementation_of_h_1.html
in one uV project and get following performance:
float -> half conversion takes 919 cycles half -> float conversion takes 1070 cycles.
Hope this helps somebody else.
Reason to use 16 bit float is to reduce data transfer rate.
With compiler options
-g -O3 -Otime
the performance is much better: loat -> half conversion takes 85 cycles half -> float conversion takes 90 cycles.
:)