Some strange behavior at extreme sunset angles. Note the area roughly where the soldiers are as the camera moves around.
https://streamable.com/fc7dw
That doesn't look like it's supposed to happen. Can I get a save game right there along with your settings (not just shadow settings as stuff like view distance makes a difference, too).
Known. Sort of deliberate, too - the refraction camera uses a different view matrix and there'd be work involved with working out how to convert between the two coordinate spaces. It's a good candidate for a follow-up PR.
The shadow itself is unaffected by the water normals, which looks especially strange in choppy conditions.
https://imgur.com/m4Om94K
Normals don't actually affect shadows at all. The thing that would make a difference would be a height map, and we don't have one of those for the water. It's maybe something that could be done in the future, but it might be pretty complex in practice.
It's well known that looking towards or away from the sun decimates the shadow quality. A high view distance + distant terrain makes this even more dramatic.
https://streamable.com/xd64t
And its especially noticeable effect on distant objects, such as the trees behind the fort in this before and after:
https://imgur.com/8WA0BGs
https://imgur.com/Lep52lk
These viewing angles are very common during play, and view distance is certain to be increased by future players, so I expect this to remain a conspicuous issue.
I've been considering alleviating this by adding a setting to control the maximum distance that shadows can be from the camera. This is pretty much how modern games with long view distances manage to do it (although sometimes they have prebaked lighting to fill in the gaps). Nothing else can really help except significantly increasing the shadow map resolution (which, in my testing, hasn't actually had a massive performance hit).
That's bad. Can you get me an APITrace of it being broken, please? There should be a thread somewhere explaining how. Maybe testing with an artefact from before normal offset shadows got merged, too, as that's the only thing I've done recently that might have broken it, and I'd have expected someone to have reported this already if it's been broken for a long time.
Regarding that last video, I can't help but comment on the ruinous increase in draw time with shadows enabled. I appreciate that shadows are rather costly, but going from what was an average of 100+ fps in this scene to below 60 seems unusually severe. It also appears to make the fans of my GTX 970 spin up more than Skyrim SE, which is a first for OpenMW in my testing. I hope the subject of performance continues to be examined, as it feels like there's more to be discovered.
Shadows basically require a whole extra rendering pass for each shadow map containing everything that might end up casting a shadow into the map. Even in modern games, that's a lot of extra work, and Morrowind's assets aren't especially amenable to being drawn efficiently when there are more than a few of them at once.
One problem I'm getting testing with the latest build artifact that I haven't seen talked about is some objects getting a light-coloured lip at their bases at night:
Do you know when this started? Has it been an issue for a while?
Zig-zags, pixels and general blockyness
At you end, this can be lessened with a higher shadow map resolution. In other games, there are two main solutions that actually ever get used.
The first is Percentage Closer Filtering. Basically, you do the shadow depth test on lots of nearby texels. GLSL 1.2 (what OpenMW uses) only provides a nice way of doing this for the closest 4 texels, and we already do this. Doing more is possible, but it will also make places where the shadows are already sharp look blurry.
The second is Variance Shadow Maps where as well as having a depth map, you have a depth squared map (both of which can be scaled and filtered like any other texture) and then you can work out the variance of the depth in the shrunken map and therefore estimate what ratio of samples would pass the depth test. Unfortunately, it has some major artefacts which look very bad. I've looked into it a little recently, and am beginning to think that these are more easily surpressed than I initially thought.
Either way, these are both things for a follow-up PR if ever.