Marcus Müller
Marcus Müller
No reaction in a month to a claimed duplicity - closing. But please reopen if wrongly closed!
yeah I have had a gr::math fxpt benchmark from the 3.6.5 days (someone had to shift the frequency of a target, and instead of making his Bachelor thesis actually work,...
agreed, this is worth putting more time into. Also agreed, the standard library trigonometric functions are slow for the frequency shift use case – it's actually not that hard to...
Re: speed differences fxpt and volk: well, the fxpt functions are inlineable, but VOLK always needs to dispatch a function call, and for single numbers, that's probably worse than the...
Fully get your disbelief there. Let me solidify the testing.
``` +——————————————————————————————————————————————————————————————————+ | fxpt sine time: 8.9234e-03 s throughput: 2.241e+09 it/s | | fxpt sin/cos time: 2.3390e-02 s throughput: 8.551e+08 it/s | | fxpt sin/cos vec time: 2.8762e-02 s throughput:...
> @marcusmueller Do you think it would be possible to have fxpt manually calculate values that fall in between table indices? This would be the ideal solution for performance and...
ah darn, I thought it was per $\pi$, so I was wrong.
@ValeZAA we're already doing something similar (smarter, actually)
This is a bug, and it does surface in GR usage, but far as I can tell, it needs to be solved within UHD; @mbr0wn I couldn't pinpoint whether this...