bcb9e644 (#109) turned a comparison into an assignment: the smoke of a
vehicle in air at 5 degrees or less, or with its engine under 45 degrees,
got the colour 128,128,128. The emitter colour is on a 0-1 scale, so the
smoke turned pure white, glowing at night, and stayed so once the engine
warmed up. Up to 26.05 the line compared and did nothing; the smoke keeps
the colour of its template again, as it did then.
Co-authored-by: maj00r <maj00r@host>
623c4f82 (#97, in 26.05.1) changed the order of the camera rotations from
Z, X, Y to X, Y, Z. The order matters only with a roll, which the cab
shake gives (Angle.z, set in drivermode.cpp), or a camera entry of the
scenery with a roll. Applied last, in the
space of the cab, the roll nodded the view up and down when looking out
of a side window instead of tilting it. The roll is multiplied in first
again, as in 26.05.
Co-authored-by: maj00r <maj00r@host>
* fix(ai): shunting AI brakes for a stop point again, the braking curve doesn't plan from a brake table left at zero
The braking curve outside the final approach plans from the consist brake table since ce87aa0e (#187, merged
with #264) whenever the brake is pneumatic. That table is worked out only for train running (Obey_train, Bank);
in shunting it stays at zero, the planned full braking falls to its floor of 0.01 m/s2, and the demand comes out
at a few hundredths of a m/s2, above the shunting brake threshold of -0.2. A shunting engine then rolled on
without braking until the final approach, too late, and ran through the signal it was to stop at (seed
480873929 of l053_poranek: ep07p-2003 at sandomierz_g4_sem_info, sm42-1096, et22-654).
With no worked out table the curve is again the constant deceleration over the distance left, as it was before
ce87aa0e and as #187 meant to keep it for shunting. Train running is unchanged.
The curve moves out of TController into vehicle/brakingcurve.{h,cpp} so that it is unit tested:
tests/unit/test_braking_curve.cpp checks a shunting engine braking to a stop point with the zeroed table, and
that a train with a worked out table gets the same demand as before.
* test(ai): pin the train braking-curve demand to the results of the code before the move
A brake table that grips harder as the train slows and a gradient, so that a wrong gravity sign, table index or a dropped term in the shaped demand shows; the expected values were worked out with the code removed from TController::adjust_desired_speed_for_braking_curve.
* fix(test): a train set across a signal doesn't pass it at danger the moment the AI first looks at it
The speed table takes in events up to a train length behind the head, so a signal the train stands across at the start of a scenario is found already behind the head, becomes elapsed on its first update, and the test mode reported it as passed at danger (calkowo_lotos: st43-189 at mac_daleko3_sem_info, manewrowo3: st43-100 at suw_szopa_a02_sem_info, both at 13.6 s with the train at a stand, the signal 20 m behind the head). A point now counts as passed only when the head of the train had it ahead once; the blackbox record of passed signals follows the same rule.
* fix(test): a signal found up to 5 m behind the head still counts as seen ahead
The distance to an event is found in steps of a metre from the position of its memory cell, so a signal just ahead of the head (a train stopped short of it when the autopilot takes over, or set right at it at the start) could come out slightly behind it, and passing it would not be reported. The trains set across a signal that the previous commit is about stood 20 m past it.
Still not reported: a signal crossed while the speed table is being traced anew in motion (a Timetable: command to a running train, the autopilot taking over in motion), as it is found behind the head when the trace is made.
* docs(test): say what the seen-ahead tolerance gives up
A train set less than 5 m past a signal at stop is still reported to pass it at danger: standing that close it can't be told from a train stopped just short of the signal, which the tolerance is there for.
---------
Co-authored-by: maj00r <maj00r@host>
Unload_Material walked the textures of a material and of its variants the
same way in both opengl renderers, and Sonar counted its declarations, set
into the long run of overrides the two headers share, as duplicated: 5.2%
of the new code, 3% allowed. The walk is gfx_renderer's now, over
IMaterial::GetAllTextures and the backend's Unload_Texture. Materials that
don't tell their textures, the Vulkan renderer's, give none, and are left
alone as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A spawned vehicle is named <spawn>:<number>, and its log went to physicslog/<spawn>:<number>.dat. On
NTFS that isn't even an error: it is the stream <number>.dat of the file <spawn>, which nothing lists,
so the log was written and no one could find it. On Linux the file was there, under a name Windows
can't open.
The name of the file of a log has < > : " / \ | ? * and the control characters replaced by _, on every
system, so the logs are named alike everywhere (spawn_ezt:12 is spawn_ezt_12.dat), and a _ before a
name Windows keeps for a device (con, nul, com1..., with an extension too). The bytes of 1250 stay.
The black box tells the log it begins by the same name.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The tapes of the user's vehicles were written by update() after the check for the writer of the
recording, which close() drops on entering the editor (F11). With "physicslog no", the speed recorder's
tape ended at F11 while the drivers' own logs of "physicslog yes" went on.
The tapes are files of their own: they are kept from the moment the recording was to be made, the file
made or not, closed or not, until the simulator closes, after the gap of the time in the editor.
Only the recording itself stops at F11, as before.
Reported in the review of #385.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The writer's thread called flush() with nothing around it, so an exception leaving write() ended the
process with std::terminate, with no message. One that happens: the file of a log was made a
std::filesystem::path of the vehicle's name, which MSVC converts in the code page of the system and
throws for bytes not of it; on a Windows with "Use Unicode UTF-8 for worldwide language support" a
vehicle named in 1250, as the sceneries have them (0xB3 for ł), isn't UTF-8.
- the file of a log is made of the name with MultiByteToWideChar without MB_ERR_INVALID_CHARS: the same
file as before for every name of the code page, U+FFFD for the bytes not of it instead of a throw.
Elsewhere the bytes are the name of the file as they were
- a log whose writing throws for any other reason (memory) is kept for the next try as one whose file
can't be written, and the other logs are written still
- the thread, flush_for() and stop() write through flush_safely(), which lets nothing out: what was taken
for a writing failing as a whole is lost and told in the log, the next writing goes on
- the messages name the file by the bytes of the name, not by path::string(), which converts back
Tests: a log which can't be written doesn't keep the others from theirs; a vehicle named in 1250 has its
log, read back by physicslog_writer::file().
Reported in the review of #385.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Its normal and specular maps were read with it and stayed in the memory: the trim passed over every image with no texture of its own, the material along with it.
An image of a material read the material with all its textures, the normal and specular maps and the variants, and only the first one was let go of: those stayed in the memory for every material scrolled past. They are let go of together now (Unload_Material). texture_manager::unload also takes the textures read and not made yet, waiting for a worker still reading one, and deletes the gl texture itself: release() would read it back from the card with gfx.resource.move only for it to be dropped.
The Refresh of the texture browser let go of the images in the middle of the frame, while panels drawn before it could still refer to them; it now marks them, and they go at the start of the next frame.
Checked on td.scn by a scripted run: 8040 images of the picker read a batch every few frames end at 2.0 GB, against 2.3-2.6 GB before this and 1.1 GB at the start; a trackbed material picked from the folder is drawn on the track after its image was let go.
The picker of the textures of a track offers every texture of the textures folder, and each tile that came into sight read its texture full size, into the graphics card and the memory, and kept it there for good. Scrolling through the list for a trackbed texture took about 2.5 MB a tile: some 900 tiles made 3.5 GB, the whole folder 28 GB, and the game went down once the memory ran out.
An image out of sight for 3 seconds is now taken out of both again (texture_manager::unload, Unload_Texture of the renderers) and read anew when it's shown once more, a few a frame as before. Only a texture the scene has never drawn is let go; those of the scene stay to the resource sweep. The texture browser lets go of its images on Refresh too. The picker no longer lists textures/previews: the 4100 previews of the node bank are images of models, not textures of anything.
Checked on td.scn by a scripted run: 8000 images of the picker read one batch a few frames after another held the game at 2-3 GB, where it went to 28 GB and fell before; images shown again after they were let go come back; a trackbed texture picked from the folder is drawn on the track after its image was let go.
Every crew is primary in the constructor and the drivers but one of a train with crews in more
vehicles are made passive only before it starts, so a log begun there left files with only
the header. With physicslog on all logs begin at once as before; one of a train waits for
update_logs.
"physicslog.trains yes" in eu07.ini writes the .dat log of each driver who
drives rail vehicles, not of the passengers nor of the road vehicles that make
"physicslog yes" hundreds of files on a big scenery: the tape of every train.
A driver made primary later begins its log then, and isn't counted while it
doesn't log. The black box goes with it when the ini says nothing of its own,
and doesn't write a second log of a driver writing its own. The doc tells the
doors' sides are the vehicle's, the driver's swapped in the rear cab.
The lines kept while files can't be written are cut to their share of 64 MB
before the lock the simulation waits on. A log written in part is cut back to
its length, so the next try doesn't write its start twice. A stop from the
writer itself is told by its id even after another thread took it to join, and
the watchdog's flush waits at most its time, the writing on a thread of its own.
In a unit driven from an unpowered car, the speed of the wheels was of the car
the user is in, whose wheels never slip; now of the motor car, signed as the
occupied vehicle's speed. The doc tells of a log per driver, and of the two
seconds of real time a crash loses.
A failed write put its lines back, ahead of those since, so a file held by
another program for a while gets its header and every line once it can be
written; up to 4 MB kept per log. Stopping is joined once, and not from the
writer's own thread. main stops the logs after an error it caught too, and the
test watchdog flushes with a time limit, so a disk that doesn't answer doesn't
keep the run from ending.
A hand-written "gfxrenderer Legacy" started the legacy renderer only because every unknown word did; with unknown words going to the full renderer it would have changed. The word is lowercased when read, as StarterNG's RenderEngineOf does, so Legacy, Simple or Full mean what they say.
The starter writes "gfxrenderer vulkan" for the Vulkan renderer of the builds
that have it. This exe doesn't know the word, and every word other than
default/experimental fell to the legacy renderer: a player who picked Vulkan
got the slowest renderer, the one whose ui crashes on NVIDIA (the arrays left
by the scene). "experimental" in a build without the NVRHI renderer ended with
"no renderer found!" and no game at all.
After the settings are read, a gfxrenderer word this build has no renderer for
becomes "default", with a line in the log saying so. The words it knows stay
as they were: full/default, legacy, simple, null, the older starters'
simpleshader and off (legacy as before), and experimental where it is built.
The legacy renderer drew the ui with the client arrays of the last scene draw
still enabled: the texture coordinates of the normal map unit and the normals,
pointing into the vertex buffer of whatever was drawn last, 48 bytes a vertex.
The ImGui backend sets up the vertex, colour and texture coordinates of the
active unit only, so the driver read the stale arrays with the indices of the
ui. Once the ui had more vertices than that buffer, the NVIDIA driver read past
its end and the game crashed; a crash dump of the editor (an area selected, a
track being laid) shows the read of vertex 15113 at a 48-byte stride in
ImGui_ImplOpenGL2_RenderDrawData.
The arrays are now disabled on every texture unit the geometry uses, with the
normal and colour arrays, inside the glPushClientAttrib already there, so the
scene gets its state back after the ui. The default and the experimental
renderers don't go through this code.
The Linux job becomes the same two-variant matrix as Windows (ENABLE_AVX2=ON/OFF), each
variant with a concurrency group of its own. The publish job downloads both Linux artifacts,
restores the executable bits in both and releases maszyna-linux-x64.tar.gz (AVX2, the old
name) next to maszyna-linux-x64-noavx2.tar.gz. The no-AVX2 executable is
eu07_<date>_<hash>_noavx2.
With ENABLE_AVX2=OFF the executable is eu07-x64_<date>_<hash>_noavx2.exe (eu07-x64_noavx2.exe
without ADD_VERSION_SUFFIX), so a starter can tell the two builds apart by the name alone and
pick this one on a CPU without AVX2. The AVX2 build keeps its name.
After master was merged the two Windows variants shared one concurrency group with
cancel-in-progress, so one cancelled the other, and the publish job looked for an artifact
name the matrix no longer produces. Each variant now has its own group, the publish job
downloads both, and the release carries maszyna-windows-x64.zip (AVX2, the old name, so the
links already handed out still work) and maszyna-windows-x64-noavx2.zip. The AVX2 variant
passes -DENABLE_AVX2=ON explicitly instead of an empty flag.
CURLOPT_PROXY_CAINFO gets the file found for CURLOPT_CAINFO, so a download through an https:// proxy on Fedora or RHEL doesn't fail on the proxy's certificate.
The Linux game bundles the libcurl of Ubuntu, which knows only Ubuntu's
/etc/ssl/certs/ca-certificates.crt. Fedora, RHEL and openSUSE keep the
certificates elsewhere, so every download of the editor (orthophoto, terrain
model, topo map) failed there with a certificate error.
The file is taken from SSL_CERT_FILE or CURL_CA_BUNDLE when set, else the
first of the files the distributions keep, and passed as CURLOPT_CAINFO.
With none found libcurl keeps its own.
The Linux job builds OpenAL Soft 1.24.2 from source, but the runner had none of the headers of its
audio backends, and OpenAL Soft's CMake leaves out every backend it can't find. The bundled
libopenal.so.1 had only OSS, Null, Wave and Loopback. Modern distributions and the Steam Linux
Runtime have no /dev/dsp, so the game logged "Failed to obtain audio device" and played nothing,
with any Steam Linux Runtime and without one.
- install libasound2-dev, libpulse-dev and libpipewire-0.3-dev before building OpenAL Soft
- configure it with ALSOFT_REQUIRE_PIPEWIRE/PULSEAUDIO/ALSA, so a missing header fails the build
instead of shipping a library without sound
The backends are loaded with dlopen at run time (ALSOFT_DLOPEN is on by default), so the bundled
library gains no new dependency. Tested by putting an OpenAL Soft 1.24.2 with these backends next to
the game from the Steam depot: under Steam Linux Runtime 4.0 the sound works.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The speed of the cab record is the vehicle's. The speedometer, and a speed recorder's tape with it, shows the
speed of the wheels as they turn, which runs away from it while they slip and drops to 0 while they are
locked in a skid. It is added at the end of the record (194 bytes), where the readers skip what they don't
know; the .dat tapes have it already, in their second column.
"physicslog yes" kept the .dat of every vehicle with a driver open for the whole run and flushed it twenty
times a second from the simulation's thread: some 580 open files on Tomaszewo. The lines are kept now and a
thread of their own writes them every two seconds, each file opened only for the moment it is written to.
The files, their lines and their text mode stay as they were.
- The black box, with the physics logs off, writes the tapes of the user's vehicles through the same
writer as it goes, not at the end. A log begins when the user is first in the vehicle and goes on
while the user looks around outside, until another vehicle is taken; back in it, the log goes on after
the gap. The vehicle's own driver keeps the log only if it is the primary one, as a passive one's clock
stands still; otherwise the driver of its train does.
- What is kept is written when the simulator closes, after the black box, at the end of a test run, and
when the test watchdog ends a run. A crash loses the last two seconds.
- On Tomaszewo with the black box alone, the one tape grows every two seconds. With "physicslog yes", the
581 logs grow and the simulator holds as many handles as without them; writing them takes the thread
about 1.9 s, which the closing waits for once.
"blackbox yes" in eu07.ini records the black box with the physics logs on or off, "blackbox no" never;
with no entry it goes with "physicslog yes" as before. With the physics logs off, the recorder keeps the
line of the physics log of the driver the user is with, every 0.05 s, and writes it to
physicslog/<vehicle>.dat when the recording closes: the speed recorder's tape of the vehicles the user
was in, without every vehicle keeping a log open and written twenty times a second.
The line and the header of the log are made by one function of the driver, which its own log writes
too, and the file is opened the same way, so the two are the same to the byte: the readers of these
logs (the old Hasler tape viewer, the starter) read them as before.
With "physicslog yes" every driver keeps its .dat open, and a large scenery (Tomaszewo: some 500 vehicles,
cars of the roads included) took all 512 streams the MSVC runtime gives: the next fopen failed and the
simulator stopped at "cannot read shader: shaders/map.vert" on opening the map. The limit is raised to
8192 at the start (the soft limit of open files up to the hard one on Linux), and a shader file that
doesn't open is told as such, with the system's reason, rather than as one that can't be read.
Sonar took the folder captured by the lambda reading it for a reference
to null, failing the quality gate on reliability. The folder is read and
the path kept by functions in the anonymous namespace, the age of the
listing declared in its if.
The manager of the models looked for the file with FileExists of a single
name, a plain check again since the banks keep the names asked for, so a
model named in another case (models/przejazdy/plyty3_L,
dynamic/pkp/st44_v2/main/2M62-0685-A) was reported missing before
TModel3d::LoadFromFile() could open it as it's named on the disk.
find_on_disk() finds the file in any case and gives back the name asked
for, the key of the bank.
The terrain directive names .txtf/.btf files in lowercase, as the parser
gives every token. The name is now taken as it is on disk, from the text
file first, so the binary file is made next to the text and the editor's
references and layers keep one name.
Measured in a case sensitive folder (3000 textures, every 20th in mixed
case, looked up the way Texture.cpp does), the walk over every part of the
path made a texture lookup 4-5 times slower than master: each name tried
and missing checked every folder of its path.
- the folders found are kept by the name they're given with, the missing
ones for a second, so a missing file checks its folder alone
- a folder named as it is on the disk is taken without its listing
- the time a folder was changed is looked at once a second at most
The lookup is now as fast as master's (about 180 us per texture on NTFS in
both) while finding the files named in another case too.
The textures, materials and sounds are kept by the names they're asked for.
Given back as named on the disk, a file in another case missed the bank and
was loaded again on every request.
- FileExists with names and extensions finds the file in any case but gives
back the name asked for; FileOnDisk gives the file to open, used where
textures, sounds and scripts are opened
- FileExists of a single name is a plain check again, the places checking
it and opening the same name stay as they were
- the texture format is told by the extension asked for again
- a folder changed shortly before it was read is read again for a name
missing from it, the time of a change being as coarse as the clock of the
file system; a listing cut short isn't kept
The parser and the loaders lowercase every name they read, while the data
made on Windows has files and folders named in any case (scenery/Braniewo,
textures/siec_trakcyjna_ph/.../STS_*.tga, sounds/sip/Gorki.wav). FileCased
matched the file name only, so a folder in another case wasn't found.
- FileCased walks every part of the path on Linux, the folder listings kept
by their lowercase and read again once the folder changed; Windows as it was
- FileExists finds the file in any case on Linux, and the variant with names
and extensions gives back the name as it's on the disk, so textures,
materials, sounds, station announcements and scripts open that name
- models, sound captions, timetables and nodebank previews open the file
as it's named on the disk; the texture format is told by the extension in
any case
The Linux job never installed libcurl's headers, so find_package(CURL QUIET)
missed it and the game was built without EU07_WITH_CURL: the editor showed
"no http client in this build" instead of the orthophoto, the terrain model
and the topo map. bundle_runtime_libs.cmake copies libcurl and its
dependencies next to the game with the other bundled libraries.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(editor): the orthophoto shown one of three ways, chosen in one place, and laid on the terrain without flicker
The settings of the orthophoto had two checkboxes ("Fit to terrain", "Covered by
objects and terrain") and three heights greyed out by turns, and the user couldn't
tell what they did. They're one choice now, each way told under it:
- over the view: drawn over everything, see-through, to compare with the map
- on the terrain, in the scene: tracks and models cover it, lit, the mouse
places objects on it
- a flat sheet in the scene: for a scenery with no terrain yet
The lift is shown only where it matters (on the terrain, in the scene), the
height is named for what it is in each way ("where there's no ground"), and the
ways in the scene are greyed out with the Better Renderer, which doesn't show
them.
On the terrain in the scene the photos are a copy of the ground's triangles 5 cm
up, which the depth buffer lost further away: they flickered with the terrain
textures. Their shapes are decals now, drawn with a polygon offset towards the
camera (the depth is reversed, so a positive one) in every pass but the shadows.
The translations hold the strings of the terrain of the download area as well.
* fix(editor): the terrain made from the NMT under one rule, also around the camera within the download area
The download area didn't bring the terrain: the NMT squares in it were fetched
only once the model had been placed (its panel opened) and the scenery had
terrain open, and nothing told why. They're fetched from the area as it is now,
in PUWG 1992 already.
Where a chunk is made from the model where there's no terrain, the rule is one
for the track laid, the area and the buttons of the panel (nmt_chunk_offset):
beside terrain made from the model it goes on at its height over the model,
otherwise (terrain made by hand, or none) at the offset of the scenery; no chunk
is left out for what lies beside it. Before, the terrain under the track laid
needed terrain in the scenery and a chunk of it beside, gave up chunks for the
session otherwise, and the model was placed only by its panel.
The terrain of the area: "The terrain of the scenery around the camera" (on by
default, kept with the editor) makes the chunks missing in the area in a square
of 9 x 9 around the camera, the nearest first, as the camera goes. Light: a chunk
at most every 0.1 s and three times as long as it took left to the editor, after
the work with the model the user asked for; the step of the terrain history the
editor keeps open is closed around each chunk, as for the terrain grown while
laying. Not the whole area at once: the chunks made wait in memory to be saved.
The log tells what holds it.
The download of the area gives way to the squares asked for: a square of the
area being downloaded is broken off as one is asked for (the terrain under the
track laid) and taken up again later.
The map no longer covers the area with green over the tiles in the cache, only
the failed ones are red.
* fix(editor): the terrain under the tracks follows them again, a brush stroke keeps its undo, the area terrain timed
From the review of the terrain made from the NMT under one rule:
- The terrain drawn to the tracks place by place (under the tracks) took the
height of the chunk beside it, carrying one offset along a run of chunks,
metres off the tracks: nmt_chunk_offset gives the offsets of the tracks there
again. The height the track laid goes on at (state.ahead) comes from the
terrain grown under it only, not from the chunks of the buttons.
- A chunk of the area made during a stroke of the terrain brush closed the
stroke's step of the history, and its undo lost the part brushed before: the
area terrain waits for the stroke to end.
- In a scenery with no name yet the terrain under the track laid was asked for
and dropped every half a second, the log filled: nothing is asked for there.
The log tells how long the chunks of the square around the camera took, and the
self-test sets the area terrain on or off for its run (nmtsettings area 0|1).
Measured on lukow.scn: 1.3-1.5 ms a chunk, 2.2 ms the longest; the longest frame
17 ms against 10 ms without it, none over 33 ms.
* fix(editor): the selftest turns the area terrain on or off for its run only, the setting kept
---------
Co-authored-by: maj00r <maj00r@host>