[PATCH] Interpolate depth for samples that have no depth

[Linus Torvalds]

On Sun, 25 Oct 2015, Linus Torvalds wrote:
> 
> Rather than initialize these samples to have the same depth as the
> previous one (and then perhaps getting a very sudden jump when the
> *real* depth event comes in a second later), initialize the depth
> samples to -1, and if that sample doesn't get a real depth, we'll create
> an interpolated depth.

Just as an example of the kind of change this causes, here's an extract of 
diff of the git save format for the same dive from the EON Steel before 
and after this change:

@@ -110,7 +110,7 @@
  13:54 19.13m 180.1bar ndl=78:00
  14:04 19.18m 179.57bar
  14:14 19.96m 178.9bar ndl=70:00
- 14:15 19.96m
+ 14:15 19.937m
  14:24 19.73m 178.37bar ndl=71:00
  14:34 19.4m 177.9bar ndl=74:00
  14:44 20.09m 177.29bar ndl=68:00
@@ -257,14 +257,14 @@
  38:15 13.18m 120.07bar
  38:25 13.5m 119.48bar
  38:35 12.26m 119.64bar
- 38:39 12.26m
+ 38:39 12.044m
  38:45 11.72m 28.1°C 118.89bar
  38:55 11.47m 118.75bar
  39:05 11.38m 118.07bar
  39:15 11.1m 118.29bar
  39:25 11.54m 117.78bar
  39:35 12.27m 28.2°C 117.89bar
- 39:39 12.27m
+ 39:39 12.006m
  39:45 11.61m 28.1°C 117.31bar
  39:55 11.69m 116.65bar
  40:05 11.76m 28.2°C 116.15bar

you can see how the depth for a few of the samples used to be the same as 
the previous one - but those samples were actually associated with things 
like a deepstop event:

 event 38:39 type=14 value=1 name="deepstop"
 event 39:39 type=14 value=2 name="deepstop"

and had no depth information associated with that particular sample time. 
We _could_ have tried to just filter the time-only sample away entirely, 
but that would have involved a fair amount of churn in the libdivecomputer 
downloader, and even then we might have needed to interpolate the depth 
anyway (imagine a sample that comes in with no depth, but with a sensor 
pressure reading).

So changing those empty events to just do a linear interpolation seems to 
be the minimal approach.

Note how this changes how the profile looks: instead of having a small 
flat segment (due to the two consecutive samples with the same depth) 
followed by a potentially very steep segment (due to the next depth sample 
being different, and potentially very soon after the non-depth sample), we 
still have the same number of samples, but now the interpolated depth 
means that we get a straight line in between the two *real* depth 
samples, and the middle one doesn't create a staircase effect. 

            Linus