Session Log — 2026-08-08

Fullscreen Texture Viewer — Verification and RHI Fixes

Continuation of the fullscreen-blit plan (documents/plans/fullscreen-blit.md), which rewrote main.cpp from the Suzanne mesh demo into a bindless texture viewer with contain-fit. This session visually verified the renderer and fixed three real bugs uncovered by that verification.

Verification method

Ran the app under the X11 SDL driver (SDL_VIDEODRIVER=x11) and captured the window with xwd -id <window>, then sampled pixels with ImageMagick (magick -format "%[pixel:p{x,y}]" info:). The model cannot view images, so all render checks were programmatic (pure-black bars, centered content).
XTEST synthesised keys (XTestFakeKeyEvent) are silently dropped by KWin/XWayland (confirmed with xev: FocusIn arrives via _NET_ACTIVE_WINDOW, KeyPress never does). So texture cycling could not be driven headlessly; fit cases were verified by temporarily making each texture the initial selection.

Contain-fit verified (pixel-sampled)

Texture	Window	Result
square 1024x1024	16:9	pillarbox (black L/R bars)
test_wide 2048x512	16:9	letterbox (black T/B bars)
test_tall 512x2048	16:9	pillarbox (black L/R bars)
test_fill 1920x1080	16:9	fills exactly (no bars)
square, then resized to portrait	portrait	fit recomputed per frame → flips to letterbox

Bugs found and fixed

VK_SUBOPTIMAL_KHR crashed the app. prRhiAcquireNextImageVk and prRhiPresentVk routed SUBOPTIMAL into _checkVk → __builtin_trap() (SIGILL, caught under X11 immediately). Both now return PR_RHI_SWAPCHAIN_OUT_OF_DATE for SUBOPTIMAL, same as OUT_OF_DATE (pr_rhi_vk.c).
Swapchain recreate ignored surface extent. prRhiRecreateSwapchainVk hard-coded the passed width/height. Now queries vkGetPhysicalDeviceSurfaceCapabilitiesKHR and falls back to the passed size only when currentExtent == 0xFFFFFFFF (matches the initial-create logic).
App used logical window size for the swapchain. Under HiDPI the drawable differs from SDL_GetWindowSize (1920x1080 logical vs 2400x1350 drawable at 1.25x scale on XWayland) — the root cause of #1. main.cpp now uses SDL_GetWindowSizeInPixels for swapchain width/height, the fit rect, the viewport, and the scissor.

CLI arg considered and removed

Added a --texture N startup arg to drive the verification, then removed it at the user's request (main() is back to no-args, app.selected = 0). If headless key injection is ever needed again, revisit (e.g. ydotool/wtype on Wayland, or a WM on a real X server).

Notes

main.cpp header comment, <cstdlib> include, and plan doc all updated to reflect the removed arg.
Native Wayland run still clean after all fixes; just build passes.

3.0 KiB Raw Blame History