On test2 at an extent of 300 m: 55 chunks in place of 106, 204 thousand triangles of the terrain in place of 895 thousand (all drawn 0.82 million in place of 2.34), the GPU 1.46 ms a frame in place of 2.33, and 3.6 ms of the CPU less a frame for the reflections. The 3D view stays as it was; the range of 50 km of the tracks work area in 3D cost nothing measurable on l053. Selftest perf logs the chunks of the terrain, their triangles and the steps of their grids.
Found with an address and undefined behaviour sanitizer build, random
monkey sessions on the test sceneries and broken inputs:
- A deleted model left its pointer in the pick list of the renderer, the
bound node of the panels, the bindings of the infrastructure and the
selection: the next pick or panel frame read freed memory. The renderer
forgets it, and one erase_model does the cleaning for all the places
which deleted a model by hand.
- The plan band of a route read the vertex before the first one in the
frame a drag was let go.
- A scenery which can't be opened goes back to the launcher with the
reason in the log, a track of an unknown type or a tributary loads as
a normal track or a river with a message, and a spline with no finite
length or step gets a sane number of segments.
- Typed numbers which aren't (NaN, inf, ends 0 m or 1000 km apart, a
vector of 1e30, a diverging track part of 5 km, a curve threshold of 0,
a zero history size, radius norms below zero) are refused or clamped,
where they made the geometry run away or divide by zero.
- Folders which can't be read and file names which aren't valid in the
locale no longer throw out of the scans of sceneries, textures, sweeps,
roads and signal plates; the scans skip them.
- The launcher: Ctrl+Del with no trainset selected, a dropped vehicle
whose skin is gone and a text with a percent sign.
- A terrain mesh made for another resolution of its chunk isn't put into
the chunk, and a slot is reused only for a mesh of the same size.
- Vehicles on a track changed by the editor are moved over a copy of the
list they leave.
- The table of the keyboard shortcuts gives its stretched column a
weight, else ImGui divides 0 by 0 in the first frame of the window.
The geometry of a track made anew, ten times a second while it's
dragged, only forgot the chunks it had: the bank kept them all, so the
buffer of the section grew by the whole track and its neighbours each
time and was sent to the gpu in full again. Ten rebuilds of a test
scenery took the bank from 2 to 29 MB; a long session would run the
card out of memory. The old chunks are erased from the bank now, but
for those of a moving nose, laid again in place, and a moved traction
wire drops its old one too.
A chunk swapped for one without the extra vertex data kept the flag of
having it and the upload read from an empty array; the buffer is made
again when that changes.
- In the top view of the tracks work area, the renderer's reach is the half diagonal of the screen. It used to be three extents, which cut wide windows short.
- Cells far from the camera no longer draw only their tracks: the terrain shapes are drawn too. Models keep their own ranges.
- In that view, the heightmap streamer loads chunks as far as the corners, up to 50 km. The detail comes from the size of a pixel, not from the distance.
- The walk over the chunks around the camera is skipped while it would ask for nothing, so a wide reach costs little.
- A setting, kept in the editor settings, leaves out the ballast of the railway in the top view, so the ground under the tracks shows. It is in the View menu and next to the extent slider.
- Self-test: ortho <extent> [noballast|ballast], orthophoto off.
* Text displays: accept characters above 127, in windows-1250 or utf-8
A text display (sub-model of type "text") looks up the sub-model of each character
in a 256 entry table, and indexed it with a plain char. With char being signed, any
character above 127 gave a negative index: reading the text went outside the table,
and so did filling the table if a character sub-model had such a name. In effect the
displays could only show ASCII, and a text with national characters could crash.
The lookup is now keyed by unicode code point. Both the displayed text and the names
of the character sub-models are decoded as utf-8 if they are valid utf-8, and as
windows-1250 otherwise, so legacy (ansi) sceneries and models work next to utf-8
ones, and the two can be mixed. A short ansi text can be valid utf-8 by accident
(two upper case accented letters in a row will do); when both readings are possible
the one which finds more of its characters in the model is used.
The rest is unchanged: the first character of the text still isn't displayed, the
characters are still drawn in the translucent pass only. A missing text pointer no
longer gets dereferenced.
* Text displays: prepare the characters once for each text
The sequence of character sub-models was cached for the last text only, so
displays of several instances drawn in turn decoded their texts again on
every draw. It is now kept for each text the display shows; the texts of the
instances don't change, so a draw costs a single lookup.
Also the comments added in the renderers are now in English.
* Text displays: move the text decoding to utilities
utf8_to_utf32() and win1250_to_utf32(), with the windows-1250 table, now sit
next to win1250_to_utf8(). The existing function maps only the Polish letters
and passes other bytes above 127 through, so it can't serve the displays.
* Text displays: address the SonarCloud findings
std::array instead of the C-style arrays, the utf-8 sequence length worked
out in a function of its own instead of nested conditional operators,
std::byte for the bit operations, std::ranges::count_if and contains().
The character loop of the renderers moved to Render_Alpha_text(), which
takes the nesting back to the limit, and the static pasText is accessed
through the class.
