New converters: u16o12 and s16 #109
Conversation
s16: signed 16-bit ( -32767 -> 32767 ) u16o12: unsigned 16-bit offset 12 ( 1 -> 4095 ) representing ( -2047 -> 2047 )
* Add includes for stdlib.h and tables.h * Fix an incorrect shift in exact_unroll_8
Ran full benchmarks on rpi4b/aarch64, rpi4b/arm, nvidia-tegra/aarch64 and i7-6950x/x86_64. Output is in: wisdom/wisdom.aarch64.pi4b wisdom/wisdom.aarch64.tegra wisdom/wisdom.arm.pi4b wisdom/[email protected] This includes all current dsp implementations.
* wisdom/wisdom.aarch64.pi4b * wisdom/wisdom.aarch64.tegra
gtjoseph
force-pushed
the
dev-fa-01-converters
branch
from
49cbb19
to
4fdccd4
Compare
|
Expedited review would help a great deal since I've got 3 more branches/commits that rely on this one. :) |
|
I think we're going to need to work on a separate branch for these changes (& the associated airspy changes). There's a release coming up which needs the dev branch stabilized, and I'm not comfortable with getting all the airspy stuff merged before that; but I don't want to stall your work either. I created an |
| * Signed Scaled | ||
| * Input Value Magnitude | ||
| * ------------------------------ | ||
| * 0 <invalid> <invalid> |
There was a problem hiding this comment.
If this is raw ADC output, then 0 is probably a valid value, let's handle it.
You could either shift the whole scale by 0.5 (i.e. ADC value of 2047.5 is magnitude 0), or saturate input 0 -> magnitude 65535, or tweak the scale factor slightly such that it doesn't actually reach 65535 at both ends.
There was a problem hiding this comment.
I've been testing for that 0 for months now and haven't run across it and the definition of the offset representation specifically says it's invalid. I chose the scaling I did so that a max negative (-2047) and a max positive (2047) became 65535 on output. For the lookup table, I can add a value in the 0 slot. The calculation will simply return a rolled over value for that sample which should be no big deal.
|
Superceeded by pull request #112 |
nb: regenerated starch files *not* committed yet, in the name of minimizing merge conflicts for the moment.
Includes...