mattrock1988 wrote on Yesterday, 20:07:So I've been filing countless bug tickets in the github for the Mercs-port version of this game. […]
Show full quote
So I've been filing countless bug tickets in the github for the Mercs-port version of this game.
I'm going to be honest, this business of AI slopping games together really stinks. I mean this with no disrespect to the guy working on this, but either the tools being used are careless and don't actually understand intention behind fixes, or the bug reports being filed don't get playtested. It's sort of lazy and doesn't speak to quality.
I'll give an example of this. The buildings with windows have a shimmering bug, and each time I send over a request to get it fixed, I am told that they are sure *this* new build has it fixed. Nope. Tried it again, and although it gets a bit closer, it's still janky. There is too much blind trust in the tools and not enough critical thinking.
I thought serious progress was being made with all the fixes being made to the Mac version, then unexplainable issues show up in other versions, or when specific scenarios occur, and it makes the entire exercise of running this port more pain than it's worth.
I've decided to go back to playing in DOSBox. At least I know it's consistent.
The Window bug is unique. Heres everything up to now trying to sus that out:
Summary
The window bands on the Instant Action office tower still flicker for the tester after four releases of fixes (v1.2.23, v1.2.26, v1.2.28, v1.2.30). Each fix removed the flicker in the case it targeted. The latest report says it now shows on a different side of the building. It has not been reproduced in a test render here, because the tower comes out unlit in every capture so far.
• Where: the "Office Building" target (model S_1SMTWR) in Instant Action's office1 area, seen in seed 13 at about (-94,600, -90,600).
• What: red window bands on the dark walls break up as the view moves. Sometimes a band vanishes in a diagonal, stair-stepped patch, which is what you see when two surfaces at almost the same depth fight for the same pixels.
• Status: open, and listed in KNOWN_ISSUES. The next step depends on the tester's mission, side and graphics settings.
The building
S_1SMTWR is a single object with 29 polygons: four walls, a roof and 24 window bands, six on each wall. The bands are not on the walls. They stand a small distance in front of them, and that distance is different on each pair of sides.
p20–p25
• The north bands are slightly twisted: two corners sit 1 cm out and two sit 2 cm out.
• The east and west walls (p26, p27) come after their bands in the polygon list. The north and south walls (p0, p1) come before theirs.
• DOS draws without a depth buffer, painting polygons in its own sorted order, so these tiny gaps never matter there. The port uses a GPU depth buffer, so a band only 1–7 cm in front of its wall is at risk of depth fighting with it.
• Model coordinates are as stored in the PRJ; the damaged (SA1SMTWR) and collapse (SCGSMTWR) variants are hidden until the building is hit and were checked as not involved.
What was changed, release by release
Each release fixed the case it was aimed at; the tester then found the flicker somewhere the fix did not reach.
Release
Report that prompted it
Change
Verified by
v1.2.30
Video: one band at eye height breaking up in a diagonal patch
The DOS look grows each lit polygon 3 px outward and lets its span walk trim it. Grown corners moved on screen but kept their old depth, which tilted the polygon's depth plane. Grown corners now stay on their triangle's plane, as textured corners already did.
Not visually: the tower rendered unlit. Three campaign scenes re-captured identical to before.
v1.2.28
Bands still cut in and out, video plus screenshots
The v1.2.23 and v1.2.26 lift only caught overlays within 2 cm of a face. This tower's east and west bands are 7 cm out, so they were never lifted. Polygons 2–60 cm proud of a face on its seen side now get the same treatment. Cockpit parts are excluded.
Seed 13 office area rendered at 1080x720 from the side: bands reach their ends.
v1.2.26
Windows still flicker at a distance
Added a depth bias in the world vertex shader for lifted overlays (clip z pulled forward by 1.5×10⁻⁶ × w per lift level, up to 16), and extended the lift to Titanium models.
Render checks of the tower at range.
v1.2.23
Building windows cut in and out
New coplanar-lift pass: polygons lying on a larger face are moved 4 cm toward the viewer's side of that face.
Render checks; caught the north and south bands only.
Two mistakes along the way, both caught before release:
• The first v1.2.23 lift pushed the bands into the wall instead of out of it. The lift direction was flipped to match the shader's backface convention.
• The first v1.2.28 build broke the cockpit render test, because the wider rule also lifted cockpit polygons. Cockpit parts were excluded.
Checked and ruled out
• Depth buffer precision. The renderer asks for a 24-bit depth buffer (32-bit float on Apple). At 30 m with a 60 cm near plane, a 1 cm gap is about 200 times the smallest depth step, so precision alone should not cause fighting.
• The lift missing some bands. Walked through the lift code against all 24 band polygons: every one finds its wall, including the twisted north bands and the east and west bands whose walls come later in the list.
• The damaged and collapse variants. SA1SMTWR and SCGSMTWR share the tower's position but stay hidden until it is damaged. Their geometry does not overlap the wall faces.
• Something else drawn over the tower. No particles, debris or moving parts are near it in seed 13.
• Regressions from the fixes. Each release passed make check, check-shell and check-campaign, and three campaign scenes (BET1, IVA1, MED1) re-captured after v1.2.30 matched their earlier DOSBox scores exactly.
Why it is hard to reproduce here
The tester sees a lit grey tower with bright red bands against a dusk sky. Every render here shows the tower almost black, with bands that cannot be told from the walls.
• Placement. The tower is in Instant Action seed 13's office1 area, placed at (−89,500, −89,500). Several start positions were tried; some faced away, one started on a 144 m rooftop and fell, and one finally put the tower in view at ground level.
• Lighting. Seed 13 starts at 07:00 game time (the day palette). Forcing dusk (17:30) and noon with MERCS_TEST_TOD_START still left the tower unlit, so the tester's lighting comes from something not matched here: a different seed, mission or game look.
• Headless runs. The test harness renders off screen through Vulkan's software driver. It cannot show the tester's exact window, and each 10-second render takes about 1.5 minutes.
• The video. Frames show the flicker at the crosshair's height on one band, in a stair-stepped diagonal. That is the clearest clue, and it is what the v1.2.30 change targeted.
Open theories and next steps
The likeliest remaining cause is in the DOS look's polygon growth and its span-walk trimming, not in depth precision. The latest report of a "different side" fits that: the two pairs of sides differ in gap (1 cm against 7 cm) and in polygon order.
Theories still open
1. Growth on textured polygons. v1.2.30 kept grown Gouraud corners on their plane. The textured path already did, but the walk can still discard different pixels on the band and the wall when both are near the screen edge or crossing it.
2. The 1 cm sides specifically. The north and south bands sit only 1–2 cm out. The lift adds 4 cm and the bias adds a little more, but a steep viewing angle shrinks the gap on screen. If the "different side" is north or south, that points here.
3. A non-DOS game look. If the tester plays with a 3D look, the DOS growth is not involved at all, and the 3D path's handling of the lift would be the place to look.
There's going to be weird shit converting a 30 year old game to modern code/graphics stack. You're absolutely right that Dosbox is the only way to play it 100% as you remember.
For what it's worth, I appreciate the reports as I don't (currently) have the time to play the game start to finish every time something is pushed.
In Claudes defence, when it does a bug fix. it very may well cause other regressions. Could I tell it not to do that? Yes, but the rub is it'll take ~4 hours for it to do a full sweep of the game and compare frame by frame vs a copy it has in Dosbox. 2-3 of these full sweeps exhausts my work Claudes ($170 CAD plan) weekly limit. So I'm a bit limited on busting its balls.
Im confident even slop coded, it'll get sorted eventually. Heck, it's only been going on less than two months and the game is mostly the way there (in my opinion).