Changes to how the AI works the train brake on the way to a lower speed
or a stop, so that it uses few, early applications instead of a late
deep one:
- Approaching a lower speed, the deceleration may fall short of the
demand by only half of a step (fBrake_a1) before the brake is applied
harder. With a full step the demand crept up meanwhile and a long G
train got its next step many seconds too late.
- The shortfall is checked every 3 s. If it keeps growing at a handle
left alone for longer than the brake takes to respond, the AI adds a
small step at once instead of waiting until the margin calls for a
deep one. A shortfall that shrinks on its own (brake still filling,
blocks gripping harder at lower speed) does not trigger it.
- Releasing, a position short of the first one counts as giving the
minimum stage of the distributor, not less.
- Holding the speed on a descent, the predicted excess speed is taken
off over the time the brake takes to respond, not at once, so a slow
brake does not get a deep application that then takes as long to
release.
The rule "wyluzuj lokomotywe, to szybciej ruszymy" released the
locomotive's own brake whenever power was asked for and the locomotive
cylinder held more than 0.4 bar. It had no speed condition, so it also
fired in motion, at 50-75 km/h, right after an application. On a
locomotive with the EStED distributor (E186) this vented the control
reservoir; with the pipe not yet recharged the locomotive brake then
faded to nothing at the next application.
The rule is meant for starting off, so it now applies only below
1 km/h.
Corrections to the TNESt3 distributor, checked against the ESt3 data
and acceptance requirements (orange book tables 3-5, 3-7 and 3-9) and
discussed with youBy:
- The main control device settles at (control reservoir - pipe - 0.21)
times the ratio instead of using a 0.05 bar offset. The 0.21 bar is the
spring preload; without it the first stage came 0.2 bar of pipe drop
too early. BVM is derived from the same offset, so a full service drop
still gives MaxBP.
- When the valve brakes it gives at least the minimum stage of 0.28 bar
(acceptance tests 12 and 22 require 0.25-0.30 bar); when the target
drops to zero the cylinder vents completely.
- The automatic releaser vents the 15 l control reservoir from 5.0 to
0.4 bar in about 23 s (test 27: 23 +/- 3 s) instead of 5.6 s.
- 12" cylinder fill nozzle dTN is 1.62, not 1.65 (typo).
- With the AL2 relay the impulse chamber is 1 l and the cylinder nozzles
are scaled so that the fill and release times meet table 3-9; before,
the AL2 timings had been set by eye.
On a valve test bench running the acceptance procedures, the number of
results out of tolerance drops from 77 to 40. The remaining ones need
parts that are not modelled yet (accelerator venting, ZP/ZS filling,
intermediate chamber).
The condition added for holding the speed uphill near a stop pushed the
cognitive complexity of control_tractive_force() to 26, over the 25
SonarCloud allows. It now lives in is_too_fast_uphill(); the behaviour
is unchanged.
adjust_desired_speed_for_braking_test runs after the target speed was
worked out and overwrites AccDesired. In its third phase it holds the
deceleration near fAccThreshold until the speed has dropped by 8 km/h.
On a 20 per mille descent that application barely holds the speed, so
the phase can last minutes, and a stop signal coming up meanwhile is
ignored. In one real-time run an E186 with 19 container wagons started
the test 1290 m before a stop signal, lost 7 km/h by the signal and
overran it by 274 m after the emergency application.
While the test runs the AI now uses whichever of the two asks for the
harder braking: the test may brake harder than the route needs, never
softer.
Measured in real time, four runs of 32 consists each on a 20 per mille
descent, on the flat and on the descent with half the rail adhesion:
the test no longer runs within 600 m of a stop signal (before: down to
414 m on this branch, and 14 m with the overrun above in a build with
other brake changes), no overruns, stops 8-15 m short of the signal 86,
117 and 83 of 128 (before 90, 117 and 78).
Within fMaxProximityDist of a stop, going uphill, the AI holds its speed
with AccDesired = min(0, AccPreferred). control_tractive_force only cut
the power uphill when AccDesired < 0, so an AccDesired of exactly 0 kept
the controller where it was. A consist that came into that range under
power kept full power and crept up to the signal.
Holding the speed uphill on the way to a stop now asks for no speeding
up: power comes off once the train is faster than AccDesired allows.
With the current brake valves this rarely shows: on a 10 per mille climb,
four real-time runs of 32 consists each, the E186 with 25 empty 408W drew
power in 0-14 samples of the last 50 m before and 0-8 after, no overruns
either way. It surfaced once the AI predicted the ESt3 brake closer to
what the valve delivers, where the same train crept to 12 km/h from 40 m
out and overran the signal; with this change it stopped short in all runs.
* fix(ai): release what the AI was holding when the player takes over
* fix(ai): keep the light selector lookup within its table, Sonar findings of the handover
- the light selector size comes from the vehicle definition (LightsList Size)
and isn't limited, while the table holds 18 positions; a bigger selector
read past the table, and so did a table index other than the front or
rear one. Only the positions the table holds are looked at now, and the
selector stays where it is for a bad table index (cpp:S2259),
- the lambdas of the handover capture what they use explicitly (cpp:S3608),
- the chained assignments resetting the headlight state on the AI taking
over are split (cpp:S1121).
* build: doctest 2.4.11 in ref/
Single header test framework (MIT), vendored like the other libraries in
ref/ so the tests need nothing extra installed on the CI runners.
* refactor(physics): read vehicle definitions from text as well as files
TMoverParameters::LoadFIZ(path) opens the file and hands the parser over
to LoadFIZ_Parse(), which holds what LoadFIZ() did with it before. The
new LoadFIZ_Text() reads the same format from a string, so a definition
can be made up in memory, e.g. by a unit test. No change in behaviour for
files.
* build: the simulation core as a library the game links
The vehicle physics (McZapkie), the vehicle definition parser and the
helpers they use build as eu07_core, a static library the game links,
so the unit tests to come can link the very same objects instead of
compiling their own copy with their own settings.
The game's build settings (include dirs, definitions, compile options,
libraries) move from the executable to eu07_settings, an interface
target the library links publicly, so the game, the library and later
the tests all build with the same ones. What only the executable needs
stays with it: /MAP, /LARGEADDRESSAWARE, output dirs, the precompiled
header, runtime DLL copying, the crash reporter and the experimental
renderer. No change in what the game builds or links.
* test: unit tests of the simulation core, run as part of the build
eu07_unit_tests: C++ tests (doctest) of the simulation, which set up the
conditions, run the code and check the outcome. No game data, no window.
- the code under test is the core library (eu07_core), the very objects
the game links, built with the game's settings; engine_stubs.cpp stands
in for the logs, settings and scenery layers the library reaches for.
python gets linked with the settings but isn't needed to run the tests
- test vehicles are JSON definitions in tests/unit/data, with the sections
and keys of the game's vehicle definitions (tables as "rows", _keys are
comments), changed by the tests where needed and loaded from memory by
the game's own loader; an unknown section is an error rather than
silently skipped
- test::vehicle creates the mover the way TDynamicObject::Init() does and
steps it the way the simulation loop does, on a straight track of given
friction and gradient, optionally holding the brake pipe pressure the
way a coupled locomotive would
- the random numbers of the simulation are reseeded before every test
case, so a test gives the same result alone (IDE) or among the others
- building the game runs the tests at the end, once the game is linked,
and prints a one line summary ("Unit tests: 11/11 passed (80 checks)").
A failure prints the full report and fails the build, in CI and locally
alike, and removes the game binary so the next build reruns them. A
change of the tests alone relinks the game so they run (Ninja and
Makefile generators). -DWITH_TESTS=OFF leaves them out
- first tests: definitions reach the vehicle, KE service application and
release, deeper reduction brakes harder, every implemented distributor
valve charges/applies/releases, service braking stops a moving wagon
* test: stub basic_node::radius_() for GCC and Clang
The Linux build failed to link the unit tests: undefined reference to
scene::basic_node::radius_() from basic_node's vtable in engine_stubs.o.
basic_node has no key function for the Itanium C++ ABI (tooltip() is
declared plain but defined inline in the header), so GCC and Clang emit
its vtable in every unit using scenenode.h, and the vtable needs
radius_(), defined in scenenode.cpp, which the tests don't link. MSVC
emits vtables differently, so the Windows build linked.
* test: test vehicle definitions load with GCC and Clang
All the tests reading a definition failed on Linux with "unknown vehicle
definition section 0": the parsed definition was brace-initialised
(auto const json{parse(...)}), which GCC and Clang take for the
initializer list constructor of nlohmann::json, making an array holding
the definition, while MSVC copies it. Initialise it with =, and require
the definition to be a JSON object.
* test: SonarCloud issues of the unit tests
- test::definition_error for unusable test vehicle definitions instead of
std::runtime_error (S112)
- the rows of a table section written by a function of their own, less
nesting in definition::text() (S134)
- std::less<> for the section table (S6045)
- comments in the empty methods of the reseeding listener and stub (S1186)
- NOSONAR where the engine's headers fix the stubs: the extern globals
and the member function signatures (S5421, S5817, S1238)
- NOSONAR on LoadFIZ_Parse(): the cognitive complexity is the former body
of LoadFIZ(), moved as it was (S3776)
- a comment Sonar took for code reworded (S125)
* refactor(sonar): use override instead of redundant virtual (cpp:S3471)
* refactor(sonar): remove empty statements, one declaration per statement (cpp:S1116, cpp:S1659)
* refactor(sonar): drop redundant lambda return types (cpp:S3574)
* refactor(sonar): prefer using over typedef (cpp:S5416)
* refactor(sonar): use =default for trivial special members (cpp:S3490)
* refactor(sonar): use nullptr literal (cpp:S4962)
* refactor(sonar): remove redundant access specifiers (cpp:S3539)
* refactor(sonar): remove redundant casts (cpp:S1905)
* refactor(sonar): use std::scoped_lock instead of std::lock_guard (cpp:S5997)
* refactor(sonar): construct elements in place with emplace_back (cpp:S6003)
* refactor(sonar): make narrowing conversions explicit (cpp:S5276)
* refactor(sonar): mark unused parameters (cpp:S1172)
Unused names are commented out; parameters used only on some platforms
are marked [[maybe_unused]].
* refactor(sonar): declare non-mutating member functions const (cpp:S5817)
Virtual functions are left untouched so overrides keep matching.
* refactor(sonar): merge collapsible nested if statements (cpp:S1066)
Comments that preceded the inner if are kept above the merged statement.
* refactor(sonar): scope single-use variables in if init-statements (cpp:S6004)
* fix(sonar): give polymorphic base classes virtual destructors (cpp:S1235)
* refactor(sonar): use try_emplace for map insertion (cpp:S6030)
* refactor(sonar): use auto where the type is spelled out on the right (cpp:S5827)
* refactor(sonar): remove redundant parentheses (cpp:S1110)
* refactor(sonar): document intentionally empty functions (cpp:S1186)
* refactor(sonar): remove dead stores (cpp:S1854)
Side effects (e.g. LISTLINE++) are kept.
* refactor(sonar): remove unused variables, captures and private fields (cpp:S1481, cpp:S1068)
* refactor(sonar): use structured bindings for pair access (cpp:S6005)
* refactor(sonar): use range-based loops and ranges algorithms (cpp:S5566)
* refactor(sonar): split chained assignments (cpp:S1121)
* refactor(sonar): add default cases to switch statements (cpp:S3562)
* refactor(sonar): use std::size for array lengths (cpp:S7127)
* refactor(sonar): use std::ranges algorithms and std::erase(_if) (cpp:S6197)
* refactor(sonar): access static members through the class (cpp:S2209)
* refactor(sonar): add braces around nested if-else (cpp:S5261)
* refactor(sonar): avoid side effects inside typeid operands (cpp:S5279)
* refactor(sonar): drop top-level const from by-value return types (cpp:S5951)
* refactor(sonar): declare variables separately from their type definitions (cpp:S3646)
* refactor(sonar): declare loop counters in the for statement (cpp:S5955)
* refactor(sonar): rely on class template argument deduction (cpp:S6012)
* refactor(sonar): inherit base constructors instead of forwarding them (cpp:S5952)
* refactor(sonar): rename locals that shadow outer declarations (cpp:S1117)
Only local variables and definition parameters are renamed; shadowing
struct fields and enumerators in headers are left for a separate change.
* refactor(sonar): reduce enum qualification with using enum (cpp:S6177)
Applied only where no unqualified name in the block could be captured by
the introduced enumerators.
* refactor(sonar): mark converting constructors explicit (cpp:S1709)
sound_source and the gl::object conversion operators stay implicit; they
are relied upon by copy-list-initialisation and GL calls.
* refactor(sonar): initialise members in initialiser lists or declarations (cpp:S3230)
* refactor(sonar): make read-only local references and pointers const (cpp:S5350)
* refactor(sonar): take read-only pointer and reference parameters as const (cpp:S995)
TTrain command handlers keep their signature, it is fixed by the handler table.
* refactor(sonar): pass expensive-to-copy parameters by const reference (cpp:S1238)
Signatures declared in files touched by PR 206 (Driver.h, EU07.cpp) and
virtual overrides are left unchanged.
* refactor(sonar): use transparent comparators for string-keyed maps (cpp:S6045)
std::unordered_map instances and map types exposed in headers used by
other modules are left as they are.
* fix(sonar): do not use UI text as a printf format string (cpp:S5281)
Text lines containing '%' would otherwise be interpreted as format
specifiers by ImGui::TextColored.
* refactor(sonar): use std::bind_front for callbacks (cpp:S5995)
* refactor(sonar): drop const_cast on string literals passed to the Python C API (cpp:S859)
* refactor(sonar): take read-only string parameters as std::string_view (cpp:S6009)
The Windows messaging helpers keep std::string, they need c_str().
* refactor(sonar): make read-only global tables const (cpp:S5421)
Mutable global state (flags, timers, singletons) is left as is.
* refactor(sonar): replace single-case switch statements with if (cpp:S1301)
* refactor(sonar): format UUIDs and log timestamps with std::format (cpp:S6494, cpp:S5945)
* refactor(sonar): use std::array for local fixed-size arrays (cpp:S5945)
Arrays that are class members or part of headers are left unchanged.
* fix(sonar): test the bearing damage bit, not bit 0 (cpp:S872)
'DamageFlag & dtrain_bearing > 0' parsed as 'DamageFlag & (dtrain_bearing > 0)',
i.e. DamageFlag & 1, so bearing friction reacted to the wrong damage flag.
* refactor(sonar): assorted small fixes
- empty() instead of size() comparisons (cpp:S1155)
- member initialisers in declaration order (cpp:S3229)
- no redundant c_str() when a std::string is accepted (cpp:S7121)
- contains() on associative containers (cpp:S6171)
- const reference instead of a copying range-for (cpp:S6032)
- starts_with() for prefix checks (cpp:S6178)
- no std::move on temporaries (cpp:S5274)
- has_value() on std::optional (cpp:S7172)
- bool literals for bool members (cpp:S5820)
- inline only on the definition (cpp:S3548)
- explicit base access specifier (cpp:S5965)
- explicit element access instead of dereferencing an array (cpp:S7116)
- rely on the C++20 rewritten operator!= (cpp:S6186)
* refactor(sonar): more small fixes
- std::format instead of iostream manipulators in to_string/to_hex_str (cpp:S6495);
output verified identical, negative hex values keep the unsigned form
- remove an empty compound statement (cpp:S108)
- drop an always-false unsigned < 0 branch (cpp:S1768)
- explicit lambda capture (cpp:S3608)
- std::make_unique (cpp:S5950)
* chore: strip stray carriage returns left in TTrain constructor comments
* fix: keep quantize() in its own value type
The S5276 change cast the result of the generic quantize<Type_>() to
unsigned int, which truncated float/double callers (shadow angle slider,
editor rotation snapping, AI speed control steps) and broke negative values.
* fix(sonar): use integer loop counters in segment sampling (cpp:S2193)
* refactor(sonar): clean up issues reported on PR 209
- remove locals left unused after the dead-store cleanup (cpp:S1481, cpp:S1854)
- unnamed unused parameter (cpp:S1172)
- single-use declarations moved into if init-statements (cpp:S6004)
- auto for casted initialisers (cpp:S5827)
- const references and pointers where nothing is modified (cpp:S5350)
- rename structured bindings shadowed by inner declarations (cpp:S1117)
Inside the proximity range only a stop had a final approach of its own;
for a lower speed ahead that branch set nothing, so AccDesired stayed at
the preferred acceleration. With the range at 50 m, which this PR uses so
that trains stop 8-15 m short of a signal, the AI stopped braking within
the last 50 m before a speed restriction and even took power: EMUs reached
a 20 km/h restriction at 25 km/h. On master the range is 15-40 m, which
kept this mostly out of sight.
A lower speed ahead now keeps the braking curve right up to where it
begins. Stops are unchanged.
Measured in real time on current master (with #199), with a 20 km/h
restriction one coach long, two runs of 32 consists each: over 22 km/h in
the restriction on the flat 9 on master, 15 before this change, 1 after
(a 36WEa at 22.1 km/h); on the 20 per mille descent 15, 0 and 0. Stops
8-15 m short of the signal stay as they were (flat 57 and 55 of 64,
descent 44 and 42), no overruns.
- Terrain work mode shows a browser of the textures folder in place of the
node bank; textures/materials can be dragged onto the palette entries
(replacing the material), the material field and the water material field
- Ctrl+Z/Ctrl+Y for heightmap terrain changes: a brush stroke or an
operation is one step of the editor history
- Ground conversion reads only the files of the terrain directives
(.txtf with origin/rotate, or .btf without text), paints the chunks with
the materials of the triangles, drops only the shapes of those files
- Fix: ground grids converted out-of-range floats to int, so whole-scene
queries returned the triangles of a single cell (shifted chunks)
- Fix: terrain layers copied from textures already uploaded without their
data stayed grey
- Fix: ImGui ID conflict between terrain tool radios and section headers
- Fix: layers of terrain .txtf files named with the scenery/ prefix and
could not be saved
- Editor selftest commands for the terrain
